Operations and Supply Management

INDEX
S. No Topic Page No.

Week 1
1 Introduction – (Challenges, Methodologies) 1
Week 2
Forecasting – Time series models – Simple Exponential
2 smoothing 31
3 Forecasting – Linear Models, Regression, Holt’s , seasonality 48
Forecasting – Winter’s model, causal models, Goodness of

4 forecast, Aggregate Planning, Tabular method 68
Week 3
5 Aggregate Planning, Tabular method, Linear Programming 86
6 Aggregate Planning, Transportation model 104
7 Aggregate Planning, Dynamic Programming, backordering 121

Aggregate Planning, Quadratic model, Demand and capacity
8 planning 139
Week 4
9 Inventory Models – Costs, EOQ model 154
10 Inventory – EOQ model graphs, with backordering 170

Inventory – Models for all quantity and marginal quantity
11 Discount 188
Multiple Quantity Discount, Multiple item inventory-
12 Constraint on numbers of orders 202
Multiple item inventory – Constraint on money value, space,
13 equal number of orders 217
Multiple item inventory –combining orders, production
14 consumption model 236
Inventory – Production consumption model with
15 backordering, Economic lot scheduling problem 251
16 Economic lot scheduling problem, Supply Chain inventory 267
Week 5
17 Lot sizing 282
18 Lot sizing – heuristics 304
19 Disaggregation 323
Disaggregation – time varying demand, Safety stock – ROL
20 for discrete demand distribution 343
Safety stock – ROL for normal distribution of lead time
21 demand 369
Week 6
Integrated model, ROL for normal distribution of LTD and
22 given mean 387
Safety stock reduction – delayed Product differentiation,
23 substitution. MOM 404
Sequencing and scheduling – Assumptions, objectives and
24 shop settings 421
Week 7
Single machine sequencing. Two machine flow shop –
25 Johnson’s algorithm 441
Flow shop scheduling – Three machines, Johnson’s algorithm
26 and Branch and bound algorithm 460
Flow shop scheduling – heuristics – Palmer, Campbell Dudek
27 Smith algorithm 477
28 Job shop scheduling – Gantt chart, Different dispatching rules 497
29 Job shop scheduling – Shifting bottleneck heuristic 513

Job shop scheduling – Shifting bottleneck heuristic. Line
30 Balancing 542
31 Line Balancing 563
Week 8
Location problems – p median problem, Fixed charge
32 problem 583
33 Location allocation problems in supply chain. Layout 598
34 Quantitative models for layout, Summary 614

Week 9
35 Introduction to Supply Chain Management 632
36 Location Problems 661
37 Transportation and Distribution Models 692

38 Transportation and Distribution Models(continued) 717
39 Bin Packing and Travelling Salesman Problems 740

40 Vehicle Routeing Problems 761
41 Value of Information 780

Operations and Supply Chain Management
Prof. G. Srinivasan
Department of Management Studies
Indian Institute of Technology, Madras
Lecture - 1
Introduction – (Challenges, Methodologies)
We begin this course on Operations and Supply Chain Management, we have two
instructors for this course, my colleague Rahul Marathe and I will be handling sessions
in this course.This course is about operations and supply chain management; in
operations management, we address planning and control issues in manufacturing and
services. In supply chain management, we address these issues considering more than
one organization or more than one facilities within an organization.
(Refer Slide Time: 00:59)
In this course, wewill be addressing several topics,this lecture will be on introduction to

operations and supply chain management.We would also be addressing the challenges
that a manufacturer has to face as well as alternatives, that have evolved over a period of
time to meet and address these challenges.We will also have topics on forecasting,
location decisions, production decisions.
1
Inventory models, introduction to supply chain management, value of information and

outsourcing in supply chain management, transportation decisions and information
decisions. In forecasting, we will be addressing a need for forecasting and some
quantitative models. Under the location decisions, we will have some quantitative
models on location, layout and cellular manufacturing. Under production decisions, we
will look at aggregate production, planning and control as well as scheduling.
And in inventory decisions, we will look at deterministic and probabilistic inventory

models. We will then go to the supply chain management part of the course, where we
will begin with an introduction and also highlight the importance and value of
information and outsourcing. We would then address transportation decisions and
decisions based on information.
2
So, let us begin the discussion on operations management, as I mentioned operations

management deals with planning and control issues in manufacturing and services. So,
we will address ourselves primarily to requirements of manufacturing. Today’s
manufacturing faces several challenges and some of these are shown here. These are
changing market conditions, where the market conditions seem to be changing faster
than ever before.
And more importantly, not only has the market changed from a seller’s market to a
buyer’s market, the rate of change is also very fast. Today’s customers are far less patient
and far more demanding than customers of the past. Today, there is a need to meet
various customer requirements, which we will see in a subsequent slide and in addition
to meeting the customer requirements, the manufacturing organizations will also have to
face global competition.
The number of global competitors and foreign players in the market are large and
whether the manufacturing organization is an Indian organization or a foreign company,
they have to face this global competition. It is also necessary for manufacturing
organizations to be proactive and make proactive decisions so that, they do not have the
disadvantage of having to react to changes, that competitors and others make.
In addition, there has to be increased customer focus, because the business comes from
the customers and customers are far more demanding and to some extent, far less loyal
3
today. If the manufacturing and service is not upto the expectation of the customer,
today’s customer has enough alternatives to consider and the manufacturing organization
would lose it is business.
Now, what are the customer requirements and how the customer has changed over the
years? We begin this slide with the simple comment which says that, the customer was
once satisfied with a black car. This means, we are going backwards in time to a
situation, where there were very few cars available and there were only black cars
available in the market, this was a state of manufacturing in the early 1920’s. So, the
customer was prepared to accept whatever was given and whatever was available in the
market.
Then, the customer started looking at availability which means, when the customer
wants to buy something, now it has to be available for the person to buy and this means
that, manufacturing systems have to be ready to have enough number of items produced
and available for the customer. So, availability would mean that, manufacturing systems
have to look at being capable of producing in large quantities. So, manufacturing
systems moved towards what is called high volume production, automatically high
volume production would mean less of variety.
So, in the early stages of manufacturing, the emphasis was more on volume and less on
variety. Now, once the manufacturing systems geared themselves to meet this particular
4
requirement of the customer, the customer also started looking at price. Once the
customer starts looking at price, manufacturing also has to adapt itself to produce items
that are affordable. So, a manufacturing system started concentrating on reducing the
price and cost of the item.
Now, once the availability and price requirements are met, the quality requirement
started coming and the customer was keen that, the quality provided by the manufacturer
was good enough. This also meant that, manufacturing had to ensure that, the quality of
the products was good, there was less of rejects, there was emphasis more on after sale
service and warranties, which are related to the quality of the product. Now, after the
quality requirements were met, the customer started looking at variety.
Now, when I go and buy something, do I have a variety of items or things to choose
from? So, manufacturing systems are now geared to producing more variety and when
the variety increases, it was difficult for manufacturing to simultaneously increase the
volume as well as variety and therefore, when the variety increased, the volumes came
down. So, the emphasis moved towards what are called mid volume mid variety
manufacturing.
And the manufacturing processes changed from high volume production to mid volume,
mid variety systems and the layouts also changed from what are called product layout to
what is called process layout. In process layout or functional layout, we are able to
produce in the middle volume middle variety category. Last but not the least, the
customer wants to buy frequently, earlier the customer would change the product only
when the product that the customer was using became unusable or was not working.
So, when the equipment or the product was unable, was not fit for use, the customer
would go and change the product. But today, we all of us change the products that we
use quite regularly. Now, since the customer wants to buy frequently and whenever the
customer is buying, the customer wants more features in the product, the emphasis also
shifted to what is called new product introduction and new product development.
So, today’s manufacturing has to definitely look at, being able to provide volume or
availability, should provide items at an affordable price, should provide items with very
good quality, should be able to produce a variety of products and should be able to
introduce new products.And since the rate of change is much higher, the new products
5
also have to be introduced frequently, so all these becomewhat is called the requirements
of manufacturing.
In addition to the above point,it is also observed that, today’s customer is a little
impatient in terms of placing an order and waiting for the product. So, it is absolutely
necessary that, manufacturing is able to meet the deadlines and produce and make the
products available in as short a time as possible. Today, the customer also wants more
value for the money given and therefore, factors such as qualities, warranties, after sale
services, most of these are packaged along with the price of the product. And this also
puts a little more pressure on manufacturing to take care of all these requirements
simultaneously.
Now, let us look at what are the requirements of manufacturing, so whatever we have
seen till now, a summary of that is provided in this slide. And the requirements of
manufacturing are, to make an increasing variety of products, this is because the
customer wants variety, customer wants to choose whenever he or she goes to the market
to buy something. Make it on shorter lead times; this is because the customer would buy
frequently with smaller runs and flawless quality.
Particularly here, one has to look at the adjective for flawless quality, now quality is now
something, on which there is absolutely no compromise. So, the quality has to be
exceptionally good and quality has also become some kind of a minimum requirement to
6
sustain, so manufacturing systems have to produce with flawless quality. Now, the first
bullet is with respect to the customer, the second one is with respect to the organization,
improve the rate of return on investment by automation or by automating and introducing
new technology in process and materials.
So, today there is a need to introduce new technology. Why is that such a need? The
need is because price can be reduced. So, by introducing new technology in processes
and materials, it is now possible to bring down the time to produce.When we bring down
the time to produce, automatically the cost of production comes down and the price can
be reduced.And to meet local and foreign competition, because today manufacturing has
to face local and foreign competition from other companies.
So, the second bullet that we have seen, essentially looks at requirements of
manufacturing from an internal perspective as well as a little bit from the customer,
because the price of the product has to be reduced. Third requirement is to mechanize,
but keep theschedules flexible, keep the inventories low, keep the capital cost minimum
and keep the work force contented. Mechanize says that, manufacturing has to look at
automation, to bring as much of automation as possible.
At the same time, they have to keep the schedules flexible, because the demand for the
products can be a little uncertain and cannot be forecasted or predicted with 100 percent
accuracy. So, there will be rescheduling that happens, because of demand changes and
when demand changes, the manufacturing organizations have to be prepared to do the
reschedule which means that, the schedules have to be made flexible.
The inventories have to be low so that, the price can be minimized, capital cost have to
be minimum so that, the price can be minimized and work force have to be contended -
because from a contended and happy work force, comes a good product. So, with all
these things in mind, the requirements of manufacturing have been defined by Skinner in
1985, which is roughly about 25 years ago. But, whatever we have seen in this bullet, is
still applicable today and most Indian organizations are striving to reach, what this
particular slide asks them to do.
7
Now, before we get into some more aspects of operations management, let us look at a
few events in history that have happened, which have some impact on, how
manufacturing and services and also the customer's requirements have changed. If we go
back into history, an important landmark is the Second World War, which perhaps was
the first time made people understand, that resources are scarce and scarce resources will
have to be utilized carefully.
Now, this understanding that scarce resources have to be utilized carefully, led to the
development of the field of Operations Research, where the emphasis was not only on
doing things, but doing things well. There is an objective of minimizing the cost or time
associated with carrying out the activities. Along the same time, we also had
breakthroughs in computers and computers were made available and this helped
manufacturing organizations in the areas of cost and information control.
The communication systems also started getting better and then, it is now possible to
have a personal touch and constant communication between suppliers and the customers.
More recently, we have this concept of electronic commerce, where customers or even
people need not go physically to the place of sale to buy, while they can buy from their
own houses and things have to be delivered at the door step.We also have the concept of
home offices, where a lot of work does not actually happen in the offices.
8
One of the reasons for going through this is to convey the importance of time and speed,
as well as the impact of computers, communications and information technology. We
have already seen that, the customer wants new products and customer wants them very
quickly. So, the time taken to manufacture has to come down which also means, that the
speed of manufacture has to go up.
A careful look at the points that have been described here also tells us that, the customer
right now wants things very fast and has become more and more impatient to waiting for
something to happen. So, there is a tremendous need in manufacturing, to do things at
less cost and to do things very fast.
Now, let us see how manufacturing have also adapted themselves, but before that, we
also have a couple of slides, which talk about some issues in history and how
manufacturing picked up points from these. Now, this slide talks about people who have
contributed extensively, now the first starting point of almost all of manufacturing was
the conversion from up and down motion to rotary motion, which actually helped in the
steam engine, which was a starting point of a lot of manufacturing.
Then, early effort atmanufacturing was more towards volume and scale and with
emphasis on the steel industry and there was also emphasis on cost control to increase
the profits.
9
Now, as mentioned earlier, after manufacturing looked at volume and scale,

manufacturing also looked at taking the work to the man and not the man to the work and
this idea of moving assembly line started. And about the same time, manufacturing
systems also understood to concentrate on efficiency and when the principles of
scientific management were actually brought out and used largely by the manufacturing
industry.
10
Then came the idea of scientific management, where the scheduling and other aspects
also came in.Here, there is a mention of the Gantt chart and then, we also have the time
and motion study, where manufacturing started looking at the human aspects. Now, these
slides also tell us, how over a period of time, various aspects of manufacturing starting
from being able to produce.
And then, looking at scale,looking at speed, looking at efficiency, looking at bonus and
human resources and looking at motions, all these things came into manufacturing and
they have all shaped up the evolution of manufacturing systems.
Now, let us look at, what we call as the methodologies that have happened over a period
of time to meet the customer requirements.Now, we classify them into methodologies
based on traditional approaches.
11
And then, methodologies based on process improvement then, methodologies based on

human resources and processes, and methodologies based on information systems and
decisions. Now,we will look at each one of them a little bit,so we start with traditional
approaches and as we mentioned earlier, traditional approaches to manufacturing were
based on volume and variety.
Now, this slide capturesthe relationships the x axis shows the volume and the y axis
shows the variety. So, production systems or manufacturing systems were categorized
12
based on volume and variety as shown in this picture. Very high volume and very low
variety systems were called continuous production systems shown as C, which stands for
continuous production systems. High volume and little more variety were the mass
production systems.
The middle volume middle variety, which is shown as a large portion here, is a batch
production system.And as we move on the variety and keep increasing the variety, the
volume comes down and we have what is called the job shop. And further up with very
high variety and very low volume, we have what is called the projects or a project based
approach. Now, in terms of layout, the very high volume systems, this part of the picture
or the figure have what is called line layout and also called the product layout, which was
meant for continuous high volume production.
The batch production systems which are shown in the middle have what is called process
layout or functional layout, where departments were laid out based on process
specialization. Now, these types of layouts and production systems were used for middle
volume middle variety and the focus on these were more on utilization of resources.
Quality was of importance, but quality was more towards what is called statistical quality
control with focus more on measurement and conformance to specifications.
Now, for very many years, people used these type of traditional approaches to production
systems, say till about 1960, 1970 people were using these type of approaches. Now,
13
among the traditional approaches, since a large amount of manufacturing was on the
middle volume middle variety manufacturing systems, we had more of the process
layout or functional layout that was used for manufacturing. As mentioned, in a
functional layout focus was more on utilization of resources and there was little less
focus on the product than on the process.
So, quality was also restricted more to the process and less on the product, and over a
period of time, people observed that, the ownership and responsibility was not on the
product, but was more on the process, because the focus was on utilization of resources.
It was also observed that, material move very long distances within the factor, there was
not much of responsibility, quality was also suffering, changeover times were very large
and there was a lot of inventory in the system.
As manufacturing systems continued to use the process and functional layout for batch
manufacturing, people found out that, there were lots of disadvantages of such a type of
manufacturing. The next question that came is, while line layout and process layout have
their own advantages and disadvantages, is it possible for usto borrow or to take some
advantages of the line layout and bring it into the process layout.
Now, this helped in what is called the group technology or cellular manufacturing. Now,
group technology or cellular manufacturing came somewhere in the 1960s, may be little
earlier than that, but people started using the techniques of cellular manufacturing and
14
group technology somewhere in the 1960s. Now, let ussee, what this cellular
manufacturing and group technology is.
Now, let us look at this picture which talks about, what is called a functional or
departmental specialization. Now, we have raw materials that come in, we assume A, B
and C as three families of parts, which can go to different products. Now, this layout is
based on, you can see this orange color is a set of machines, the light green is another set
of machines, light blue is another set of machines. These two are one type of machines
and these three are another type of machines, this schematic diagram shows five different
types of machines.
Within each type, the machines are similar and can perform the same functions, so the
layout is based on functional specialization. So, if we take a typical part or product or a
path family, which we call as A or B or C. For example, A would go to machine number
1 here and then, it would go to machine number 11 and then, it would go to machine
number 4 and from then, it would go to 14 then, it would go to 9 then, it would go to 10
and then, it would come out.
Similarly, you have these routes for B and C, what is also important is that, in order to
increase the utilization of the machines and in order to maximize the time, the machines
were producing the batch sizes were large. Since the production batch sizes were large,
to go from 1 to 10, the transportation was done in large batch sizes. So, production
15
batches were large in size, transportation batches were also very large in size and there
was a lot of delay in transporting from 1 to 11,11 to 4 andso on.
Because, batches were large and there were delays in transporting, it also led to
situations, where some of these machines had to wait for jobs to come. So, while this
type of a layout is very good with respect to utilization of resources, this type of layout
had it is owndisadvantages with respect to changeover times being very high,batch sizes
being very large, lot of material movement, quality suffering, delays in transportation
and idle times on machines, because machines had to wait for the products.
So, this is a schematic diagram of the functional layout and then, let us see what happens
when we reorganize this into what is called a cellular layout.
Now, this cellular layout also called group technology layout, what we have done right
now is, we have tried to look at all the machines that were used to produce this A. From
this picture,1 11, 4 14, 9 10, now you see 1 11 4 14 9 and 10, but please note that, A is
not a single product, A is a family of parts or components that can make or that can go
into one or more products. Sometimes A can also represent a product, but right now we
will assume that A represents a part family, instead of parts or components that are
required to make the full product.
16
Now, this is called a cell. This is a cell which has now 6 machines, this cell has 4
machines and this cell has 5 machines. Now, if you look at this cell, this cell has
machines which are functionally dissimilar. For example, this machine is different from
this, this machine is different from this and so on. So, a set of functionally dissimilar
machines, but similar with respect to the manufacturing process of this A or this B or this
C.
So, we try to bring functionally dissimilar machines and try to have what are called
manufacturing cells. Now, the moment these manufacturing cells are created, now
assume that, we have the facilities or the machines to produce all the parts that belong to
part family A. And if we are able to bring these machines physically together and use a
certain small amount of space then, it is possible to treat this like a factory within a
factory.
Now, this cell will have it is own capacity planning, this cell will have it is own
production planning and scheduling and control and so on. So, what was a large factory
as shown here, now with the cellular manufacturing systems and if we ensure that, there
is no movement from one cell to another, which is what we have shown here. If there is
no movement from one cell to another, each cell can be thought of as an independent
production unit and can be thought of as a factory within a factory.
So, we are able to simplify the manufacturing processes and control, and we all know
that, it is always easier to handle a much smaller system than a very large system. So,
cellular manufacturing helps in overcoming all the difficulties that were there in the
functional layout. People who started using cellular manufacturing, also have reported a
saving of about 20 percent inventory, about 10 percent space and so on.
17
Now, group technology or cellular manufacturing as we have seen, deals with organizing
machines and grouping parts to families. Now, how do we get these part families, now
these part families have to be formed and then, the machine cells, which make these part
families, will have to be created. So, it is about, first making part families and machine
cells, after which we look at the layout of machines within this cell. We look at capacity
of this cell and then, we look at scheduling and balancing the machines within these
cells.
So, these are the issues in cellular manufacturing, but largely cellular manufacturing is
about ownership and responsibility. Now, if we create this cell and then, we have about
half a dozen people or less working in this cell then, these people will have the
ownership and responsibility to carry out all the tasks within this cell. So, cellular
manufacturing brought increased ownership and responsibility to the manufacturing.
About the same time, we also had the Japanese management systems or just in time
manufacturing systems, which started with the Toyota production system, which
concentrated a lot on waste elimination and inventory reduction. Now, in order to bring
down the cost of the item, it was necessary to identify all the non value added activities.
All the non value added activities are called wastes, wastes by definition is anything that
does not add value to the product or waste is anything that is more than the minimum
amount of resources used to carry out an activity.
18
So, one of the contributions of the just in time manufacturing was, in the area of waste
elimination. Seven types of wastes were classified. They are called the seven Muda and
all the wastes were brought into one of these seven categories and there was a conscious
effort to bring down or to reduce the wastes. So, waste reduction automatically resulted
in less cost, so along with the just in time manufacturing systems, today we also have
lean manufacturing systems, where the emphasis is on ownership and responsibility and
also on cost minimization.
Now, about the same time, we also had flexible manufacturing systems come in, where
there was a lot of emphasis on automation. And flexible manufacturing systems focused
more on technological solutions and reduced the production time. And flexible
manufacturing systems because of their speed and additional capability, were able to
handle very high volumes, very high aggregate volumes, at the same time they were able
to handle variety.
So, these three come under what is called the process improvement methodologies.
There are many more but,we are looking at these three today, because these three are the
most important ones with respect to process improvement.
Now, there are other methodologies that came, that are based on human resources and
processes, there is a tremendous emphasis on total quality management. Now, quality as
mentioned earlier is some kind of a non negotiable commodity, people expect flawless
19
quality and the concept of quality moved from quality control to quality management.
Quality gained a certain structurein the organization and organizations started
concentrating on what is called total quality management.
Total quality management is a management initiative to empower all people working to

do their best. So, there is focus on people, focus on employee empowerment, it also
resulted in things like quality circles,where people who were operating and who were
doing the job, were given opportunity to discuss and solve problems in manufacturing, as
they happened. And since people who were working were able to solve problems, it
resulted in increased employee empowerment and better systems from a quality
perspective.
There was also this emphasis on business process reengineering, where the business
processes were reengineered. Business process reengineering or BPR as it is called
addresses a radical and a dramatic change in the process and many BPR issues were
related to extensive use of computers and technology. Now, somewhere since the 1990s,
there has been an increasing emphasis on quality certifications.
Largely ISO9000 certifications, where manufacturing organizations get themselves

certified for designing and carrying out a quality management system with respect to
their manufacturing processes. In addition to quality certifications, organizations also
looked at certifications related to environment management, related to health and safety,
and related to social accountability.
Now, all these initiatives help or helped the manufacturing organizations to achieve good
quality and to have a content work force, which we saw as one of the requirements of
manufacturing today.
20
There are also methodologies that are based on information systems and decisions, so we
cover some of them here. Materials requirement planning called MRP, which talks about
inventory reduction, consistency in bill of materials, lot sizing and pull systems. We are
also going to mention the little bit about enterprise resources planning and a little bit of
supply chain management that logically follows from here. We will have more details of
supply chain management, as we move along in the course.
Now, materials requirement planning essentially talks about, first it talks about the bill of
materials. If we take a product,each product is then split into various assemblies and sub
assemblies and then, into manufacturing parts and brought out items. Now, all these will
first have to be identified, before the manufacturing begins and computer systems were
used in identifying and locating them in the structure of the product.
There are also lot sizing models that came, where the demand for the various
components were met by buying from suppliers and by identifying the correct lot size to
buy from various suppliers. Now, materials requirement planning then became MRP II,
which is called manufacturing resources planning, where capacity planning was also
included.
Then, came enterprise resources planning, which are called ERP systems, there are basically
two points of view, one of which looks at ERP systems as a logical extension of MRP
systems, where additional aspects such as quality, maintenance, human resources,
21
purchase, procurement, all these were added to MRP systems.There is another school of
thought, which talks about EPR systems, as a separate area in its own right and as being
independent of MRP systems.
So, irrespective of the approach, we take one thing that is clear is, ERP systems were
large systems, ERP systems provided lot more information and sometimes decisions.
ERP systems were expensive and ERP systems required a lot of time and effort to
mature and to provide good support to manufacturing at steady state.Then, came supply
chain management, where organizations primarily wanted to integrate production and
distribution and also wanted to add the supplier into the decision making process.
So, in some sense, if we look at it as a chain with the manufacturing organization in the
middle, those to the left of it are suppliers and those who are in the right of it are
customers. So, supply chain management talks about a network of organizations or a
network of facilities, which begins with the procurement of raw material and then,
transforming this raw material into products and then, distributing these products till it
reaches the end customer.
So, supply chain management talks about the supplier, the manufacturing, the
distribution with respect to warehouse, dealers or distributors, and retailers. So, supply
chain management also is a large complex system and the success of supply chain
management systems are round the ability to use information technology and
communication to the advantage of the organization.
22
The last classification that we are going to see today, though there are many more
classifications, are what we call as business perspective. We already made a brief
mention of business process reengineering. Business process reengineering as I said is a
dramatic radical way of changing the processes, wholly from a business perspective. We
also would look at a little bit of constraint management or theory of constraints, where
the bottle neck in any process or any business is identified and all resources are
subordinated to the bottle neck.
Now, this concept of synchronous manufacturing or constraint management came from

the goal, which was a book first written by Mr Eliyahu Goldratt. And there are other
books and other material that have come from Mr Goldratt, which talks about how
organizations and businesses try to make their performance better by looking at
bottlenecks and trying to debottleneck their difficult resources.
The last thing that we would see in this presentation about business perspectives and
methodologies is, what is called agility or agile manufacturing, where several aspects
such as empowering the customer, cooperate and compete, etc are being discussed. For
example, people also look beyond organizations that compete to say, whether two
competing organizations can actually complement each other,whether two competing
organizations can get into a third business, where they can cooperate.
23
Many things such as empowering the customer, eliminating waste and producing to
customer needs, mass customization and cooperation among organizations are some
aspects that have come into agility. So, what we have seen till now are methodologies,
which are based on four different perspectives. We started with methodologies on (Refer
Slide Time: 42:44) process improvement, which is of course absolutely essential for
manufacturing, because all these methodologies aim at minimizing the time of
production and minimizing the cost of production directly.
Now, methodologies based on human resources also try and reduce the cost, but slightly
indirectly in terms of managing the human resources very well and in terms of keeping
their processes and systems in place. Similarly, methodologies based on information
systems and decisions, help in quick and correct decision making, which can reduce the
time and cost of making decisions and help the manufacturing organization to produce at
less cost quickly.
Methodologies based on business perspective also achieve the same result, but the way
they look at it is different. They look at the business entirely as a whole and try to make
changes in the business processes so that the cost of manufacture comes down. Till now
whatever we have looked at manufacturing is equally applicable to the service, though a
lot of emphasis has gone into manufacturing.
24
Now, let us look at a couple of other things, before we quickly move into the various
aspects of operations management. Now, what is the goal of the organization? Goldratt
answered this question by saying that; the goal of the firm is to make money. And in
order for the organization to make money, they also have to produce and sell a product or
provide a service in as short a manner as possible, so money is essentially time.
So, if the organization wants to make money then, the organization has to minimize the
cost of production and minimize the time it takes to produce.The time it takes to produce
in any organization increases due to delays.
So, we will now see, what are the possible delays and how these delays can be reduced,
there can be several delays, we classify delays here as delays based on raw materials,
delays based on work in process inventory and delays based on finished goods. Now, the
reason we do that is, if we look at any manufacturing system, there are a set of materials
that come in, which are raw materials.These raw materials are transformed into finished
products, so there is a manufacturing, which involves work in process inventory and
these finished products are sold and distributed, so there is a finished product delay.
So, very quickly going through these, delays in raw materials, delays happen when
orders are not placed for materials in time. We may know that, the reorder level which is
the point, at which the order has to be made, has been reached, but orders are not placed.
Now, there is a lead time which is taken by the supplier to produce, so there is a delay in
25
the lead time. There are delays in transportation, there could be delays for the raw
material to come and reach the manufacturing organization.
There are delays that are caused due to inspection, because the items that come in are
inspected and there are delays with respect to issues. So, all these delays have to be
minimized and there is a lot of emphasis on inventory management or materials
management, about which we will see in this course. Now, inventory management would
talk about, how we ensure to minimize these lead times and how do we know that, how
the reorder level has to be fixed.
So, all these questions will be answered in inventory management, but largely all these
questions will be answered in what is called purchasing management. Purchasing
management is an essential part of operations management.
Now, there are delays caused during the manufacture when the machine is busy, the job
is available, but the machine is busy, so there can be delays.There can be delay when the
machine is available, but the job is not available, there could be delays. There can be
delays with machines breaking down or not being available,there can be delays due to
large set up and change over times, delay due to large processing times due to in process
inspection, rejects transportation and movement and waiting for assembly.
26
Now, all these will have to be addressed within the manufacturing and various topics in
operations management addresses these. Machine breakdowns are addressed using
reliability and quality management, set up times reduction is a very important part of
operations. Processing times come down due to technology, inspection and quality
separate area by itself, which also handles reject and rework.
Transportation comes through material movement and waiting for assembly is reduced
by good assembly linebalancing. So, all these reasons for delays are now addressed using
methodologies that are there in operations management.
Now, delay in finished goods come because of assembly being busy, because the line is
not balanced, assembly lines are not balanced.There is excess finished goods inventory,
there is a lot of delay in transportation and selling, all these also will have to be
addressed. So, the delays actually result in increased time to produce and increased cost
of production and there is every effort made in the principles in operations management
to bring down these delays.
Operations management looks at issues like purchasing management, proper production

planning, proper scheduling, good quality, maintenance and proper distribution to ensure
that, these delays are minimized which means, the time taken to produce comes down
which also means that, the cost ofmanufacture comes down.
27
Now, we look at some topics in operations management that we will be looking at in this
course.We will begin with forecasting, forecasting is the estimation of demand that we
will be using. Now, after forecasting, we will do production planning and
disaggregation, scheduling and production control, quality, inventory management,
MRP, lot sizing and maintenance.
Now, among these topics, we are going to see forecasting, production planning,
scheduling and inventory, the ones that are shown in green color as part of this course, as
part of operations management, that is covered in this course and then, we will follow it
up with topics in supply chain management. Now, as we say that, we look at these
topics, forecasting, production planning, scheduling and inventory, we will start looking
at each one of them in detail.
Now, before we get into forecasting, let us also take a quick look at, what are the steps
and what are the ways of linkages among these activities, we will do that now.
28
So, manufacturing deals with making a product and the first thing that we need to know
is the demand for the product. So, the demand for the product is estimated through good
forecasting models. Now, once we know this demand, we also need to know the
capacity. Now, the organization should have the capacity to meet the demand of these
products and these capacities are usually in the form of regular time capacity, over time
capacity and outsourcing.
So, the organization decides, what type of these capacities that they are going to have.
So, this combination of demand and capacity results in what is called the production
plan, where the organization decides, what are the products to be produced and how
much of this capacity is going to be utilized in each period. Then, we will look at
disaggregating it and making disaggregation, where the time that is available is now
given to various products, time allotted to various products.
And then, from here we look at scheduling, where we look at time period such as shift,
week, etcetera. Now, in addition, capacity is provided by machines and these machines
have linkages with quality and with maintenance. And in order, this is where the
production actually comes and then, this gets inputs from machines as well as quality and
from maintenance. Then, the scheduling sometimes also gets inputs from the forecasting
due to rush orders, planned orders.
29
So, scheduling has to take care of this and scheduling also gets inputs from materials and
inventory, which links with both scheduling and production. Now, once all these things
are established, the product is actually made and put to use and then, based on the time,
the organization also creates new products, for which additional capacity may be
required. So, this kind of gives us a schematic representation of, how the production
processes take place and what are all the things that are related in operations
management.
So, in this course as mentioned earlier, we would do some aspects of demand

forecasting. We would do a lot of production planning, we would look at certain aspects
of scheduling and we would also look at a lot of aspects of materials and inventory. We
would look at inventory, so as mentioned we would look at these four parts and then, we
would move into the supply chain management area, where we will be looking at
distribution separately as well as information separately. So, in the next lecture,we would
start our discussion on demand forecasting.
30
Prof. G. Srinivasan
Lecture - 2
Forecasting - Time Series Models - Simple Exponential Smoothing
Today we begin our discussion on Forecasting; forecasting essentially deals with

estimation of future demand for a product. For example, they could have demand for a
particular product; let us say in the last 6 years as 25, 32, 24, 28, 26 and 27. The product
could be anything, the product could be an automobile, a product could be phones, a
product could be machinery and so on and these numbers could be in multiples of 1000
or multiples of lakhs and so on. For the sake of illustration we will take these 6 numbers,
we can assume that this is the sale in 2009, and this is the sale in 2004.
So, the data for 6 years starting from 2004 to 2009 are available, and let us say for the
sake of illustration that these are the values. Now, we want to estimate the demand for
the year 2010 say and which I have indicated by a question mark, this demand estimation
is extremely important because this demand estimation is a starting point of production
planning, as we saw in the last lecture. Only when we have the forecasted value of the
demand, we will be able to make both capacitive planning and production planning.
31
So, the whole area of production planning begins with the forecast, so we will spend time
on learning forecasting models. So, we will start with these 6 values as values of demand
and we will try to forecast for the next period, now whenever I pose such a question and
give 6 values of demand, and ask a class of students to make their own forecast. I
normally get several answers that the students give, and let me write down some of them
on this board, and then try to explain them one by one and lead to the various basics and
principles of forecasting.
So, let us consider these 6 values of forecast and let us say these have been given by 6
different people. Now, when I asked the students the basis or the reason behind which
these numbers were given, I got a variety of answers, so I will write down these answers
and then we will slowly see the relevance of each one of them, in the context of forecast.
So, the first F is equal to 27 forecast equal to 27, comes from simple average, so one of
the students believes that a simple average of these 6 numbers, adequately represents the
forecast.
And we add all this and divide it by 6, we get f is equal to 27, another student said that he
was interested only in the last four values, and this student felt that the first 2 values were
little old and therefore, he is not keen on considering these 2 values. He believes that the last
four values adequately represent what is happening and the value and the average of
32
the last four values turns out to be 26.25. A third person said that the forecast could be
28.
And the third person reasoned out this way by saying that there is a increase and then a
decrease, and then an increase and then a decrease and then an increase. And so the
estimate after the decrease happens to be the third student who said 28, started by saying
that the student looked at the last 2 values, which happened to be 26 and 27. And the
student believed that considering only the last 2 values, this seems to show an increase
and therefore, the student gave forecast as 28 which is an increase based on the last 2
values.
Another person looked at these values and said that the forecast could be 30, if we
assume that or if we consider that this data corresponds to sale of automobile. And this
person believes that the sale of automobile is related to the population in that area and is
also related to the buying power of the population. And the person believes that due to
increase in buying power of the population, the demand is expected to go up from it is
last values. And therefore, the person estimated that 30 is a very good forecast for the
year 2010.
So, this is based on other factors, now this person who gave forecast equal to 26 believed
that there was an increase and a decrease. So, it shows an increase and then a decrease,
and then an increase and then a decrease, now there is an increase and the person expects
a decrease, and believes that the value 26 comes from decrease and increase, increase
and decrease. Now, this person who gave forecast equal to 27 took the last 3 values
instead of the last 4 and said 28 plus 26 plus 27, so this is the average of the last 3.
The person, who gave 26.833, also considered the last 3 values, but believed that since
this value is the oldest among the 3 this is older, and this is the most recent value. The
person gave weights and said, this is going to have a weight age of 1, here there will be a
weight age of 2, and here there will be a weight age of 3 which is the closest. So, the
forecast would be 1 into 28 plus 2 into 26 plus 3 into 27 divided by 6 which gave 26.833,
so it is a weighted average on the last 3 values.
The last person who gave value equal to 26 felt that if you look at all these values 32
seems to be an out layer, and there could have been some assignable reason for this 32
which looks like a spike. So, the person took away this 32 and then found the average for
33
rest of them, so this is remove 32 and find the average, now let us look at each of these 8
answers, and more importantly the reasons associated with these answers, and try and
learn some aspects of forecasting.
Now, this is centred around the average which is a measure of central tendency of this
data. This is also centred around the average even though smaller or less amount of data
and more recent data has been considered. This has a slightly different assumption that
this is based on the increase that happens here, here the person has considered other
factors. Here also the person has considered an increase and decrease, this is based on the
average, this is also based on some kind of a weighted average, and this removes the out
layer.
So, let us say out of these 8 different reasons given, 5 of the reasons are centred around
the average in some form or the other. But, if we look at this data and before we start
making any forecasting, we need to understand whether this data actually represents a
constant or what is called a level or whether there is some kind of a trend that has
happened. So, if we look at these values in some sense we do not see a clear increasing
trend from this to this we also do not see any kind of a decreasing trend.
So, I am going to assume that this does not show any kind of a trend and therefore, I am
going to start developing models, only for level or what I call as a constant model, which
in some sense has been validated by 8 out of the 8 people saying that it is centred around
the average. So, essentially we are trying to find out a constant or a number which
adequately represents all these numbers, under the assumption that there is no increasing
or decreasing trend.
So, which simply means we are trying to fit something of the form F is equal to a, where
a is the constant or the level that we are trying to find out. But, then we also understand
that if this were to represent a constant, then why do we have these variations in these
values, and not the same constant that comes as a forecast. The reason behind is there is a
certain inherent variation or inherent variability in the system, which we call as some
epsilon. And this epsilon is the noise or inherent variation or variability in the system.
And this noise is expected to have parameters 0, sigma square, which means this noise
has a mean of 0, which also means that these inherent variation or statistical fluctuations
will cancel out as time passes. And these variations are small because their variance
34
sigma square is also assumed to be a very small number. So, if we make the assumption
that this represents a constant, and we are trying to estimate that constant here, now we
will try to find out various ways of estimating the constant.
So, the simplest and the best way to estimate the constant is by looking at the simple
average, which the first student actually mentioned as the forecast. So, a simple average
is the simplest and the easiest way to give a forecast for a set of data, which shows level
or which represents a constant. But, when we take the simple average we also assume a
few things, number one is we are considering all the data points, and we are not leaving
out any of the data points.
Second is we give equal weightage to all the data points, so under these two important
assumptions that we consider all the data points, and then we give equal weightage to all
the data points, we get simple average to represent the forecast of this. So, if we make a
decision based on the simple average, we would say that the forecast for 2010 is 27. It
also means that the forecast for 2011 right now is 27, because if we add a 27 and then
take the average we will get the same 27.
So, the forecast made here for every period is the same and that will be 27, now the
moment we assume that we are going to give equal weightage to all the data and
consider all the data, then the question asks whether thus data which is 6 years old is
equally important as this data which is 1 year old. So, we need to answer that question
and that makes us move towards what is called a weighted average, so we can give some
weights to each one of these, and then compute a weighted average.
For example, if we give weights w 1, w 2, w 3, w 4, w 5 and w 6 then the weighted

average will be 25 into w 1 plus 32 into w 2 plus 24 into w 3 plus 28 into w 4 plus 26
into w 5 plus 27 into w 6 divided by w 1 plus w 2 plus w 3 plus w 4 plus w 5 plus w 6.
And as long as the w’s are greater than or equal to 0, we would get an answer which is
within the range of 26 and 32, it is also a customary to define the w such that, sum of the
w’s is 1. So, that the denominator is 1 and you do not divide it by the sum of the weights.
So, a weighted average is as good as a simple average, it has the additional advantage
that we consider all the points as in simple average. But, we are able to capture the
relationship that more recent pieces of data, add more value to the forecast and therefore,
they are given higher weights or there is a possibility to give higher weights, to more
35
recent data. So, that the effect of the more recent data can be captured into the forecast,
the obvious disadvantage of a weighted forecast is, how do we determine the weights.
One person may want to give weights 1, 2, 3, 4, 5, 6 another person may want to give
weights 1 1, 2 2, 3 3 and so on. So, for different weights we have different values of the
weighted average, and it is not perhaps possible to find out a consistent set of weights or
a value that comes out of a given set of weights, as acceptable to a group of people. But,
if a group of people sit together and say that, this in our opinion are the weights w 1, w 2
to w 6, then the weighted average can be a slightly better estimator than the simple
average.
Then we move to the next idea that was shown here, where the person had taken the
average of the last 4 values. So, there are 6 values, but then somebody had decided that
the first 2 values being very old are not going to be as important as the last 4 and
therefore, the person has taken the last 4 values and computed the average. So, the
average of these 4 become 26.25, so when we leave out certain data points, and consider
only the last k data points, then we develop what is called a k period moving average, it
is called a k period moving average.
So, the 4 period moving average gives us a forecast of 26.25, a 3 period moving average
taken here gives us a forecast of 27 which is the average of 28, 26 and 27. So, moving
averages essentially are made or decisions based on moving averages, are essentially
made by assuming that the last k periods, are more relevant and the others are left out.
So, a k period moving average can be thought of as giving weights of 0 a 4 period

moving average can be thought of as giving weights 0 0, 1 1, 1 1. So, every moving
average can be thought of as a weighted average with weights given corresponding, a 3
period moving average can be thought of as giving weights 0 0 0 1 1 1 and it becomes a
weighted average.
While moving average is able to capture, the principle that more recent points are
important, the older points do not contribute much, then there are two issues that are
associated here, one is do we essentially ignore all of that, after all do not they add a
small amount of values, which perhaps could be modelled or could be taken into account
by giving slightly smaller weights.
36
But, then as we keep increasing the k for example, a 4 period moving average gave 26.25
a 6 period moving average for this data would mean that you are taking all the 6 data
points and n period moving average if there are n data points will be the simple average
etcetera. So, a moving average is related to simple average, it is also related to weighted
average, so this is another form of average, which is the moving average a 4 period
moving average is shown here, a 3 period moving average is shown here.
The next thing that we can look at is, so if we have a simple average and a weighted
average, can we have something like a weighted moving average. So, a weighted moving
average is shown here with 26.833, which was obtained by considering the last 3 points,
but giving differential weights of 1, 2 and 3. The oldest has a lesser weight of 1, the
middle one has a weight of 2, and this has a weight of 3, and the weighted 3 period
weighted moving average gave us 26.83.
So, we have seen four different forms of average, the simple arithmetic average, the
weighted average, the moving average, and the weighted moving average. All four of
them are measures of central tendency; all four of them would give us values which is
well within the range. Now, I also need to explain the reasoning behind these other three,
and also we need to try and position the other three reasons, in the broad context of
forecasting. Now, when we look at these values and when I introduce these values, I said
these would represent the demand of a particular product.
And the only thing that we have right now here is that, how the demand behaves with
respect to time. So, the behaviour of demand with respect to time has been captured in
this data, and all data where the demand is given with respect to time, come under the
category of time series models or time series forecasting. So, this is an example of time
series data, where the behaviour of the demand with respect to time is provided.
Now, when we start looking at this F is equal to 30, now here the person, who said F is
equal to 30, assumes that there is a relationship, there is some other factor which is
influencing the forecast. And that factor could be the buying power of the people and so
on, models were other factors influence the forecast are called causal models, where
there is a cause and effect relationship between the forecasted variable, which is the
independent, which is the dependent variable, and the causal variable which becomes the
independent variable.
37
So, even though this is a reasonable value the thinking process behind this value comes
from the assumption, that there are other factors. And that leads to other type of
forecasting models. Right now in this lecture, we are restricting ourselves to time series
forecasting. And much later we will see a couple of models for causal forecasting, so that
is how one can explain this other factors, now there are two other answers which need an
explanation.
Now, both of them are centred around the belief that this data is not stable and in fact
this data shows a certain increase or decrease, which essentially means that in both these
instances the person believes that there is a trend. And here also there is a trend, the rest
of them based on averages do not believe that there is a trend. So, once again time series
models have different classifications, now the first ones are called only level models or
constant models, which for example, was the assumption when somebody made this
forecast, and somebody made this forecast, and somebody made this forecast this
forecast.
Whereas, when someone gave this value which is F is equal to 28 which was based on an
increase and F is equal to 26 which was also based on increase and decrease. There is an
assumption here that the data is exhibiting a trend, a noticeable increase or decrease
which cannot be attributed to randomness or noise. So, there are trend models which we
will see a little later, after we address this particular model called the models for data that
exhibit constant.
So, the last point that requires requires certain explanation is this, now this person said if
I look at these I find that this is an outlier. So, I am removing this and then I am trying to
take the average, so the general assumption is that this person believes that this data
represents a constant or level data, person assumes that is there is no trend or anything.
Except that the person believes that this number seems to be slightly high, and perhaps
may not be adequately justified by the assumption that it is a noise.
So, in such cases it is also a customary to remove certain pieces of data, after the analyst
is fully convinced that this may not actually contribute or it might take away. So, in such
cases you remove a piece of data, and then you find the average, so these represent
different kinds of thinking, but largely centred around average, which represents that the
38
assumption behind here is that it represents a constant or a level model, and whatever
variation we have is attributed to noise or statistical fluctuation.
And we are trying to forecast F is equal to a plus epsilon where epsilon represents the
noise or the variability, and it which has the mean of 0 and it has a variance of sigma
square. Other things such as trend models, and other causal models we will see as we
move along, now let us look at one more thing, let us look at this pieces of data once
again. Now, if we look at the four models that we have seen, the arithmetic average, the
weighted average, the moving average, and the weighted moving average all of them
have advantages and perhaps a little bit of limitation.
The arithmetic average gives equal weights to all the points, which we may kind of
question because it is only customary to believe that more recent data should contain
more weight, and are more important. The weighted average has the disadvantage of not
being able to arrive at some kind of a consensus weight or an acceptable weight, moving
averages are very nice, but moving average tends to ignore or give 0 weight to earlier
points and so on.
Any weighted model has the limitation of getting the right kind of weights, so it now
boils down to another question is there a model where, we actually consider all these
points. We give progressively increasing weights to more recent data, and the weights
seem to come logically and are acceptable to everybody, now such a model is called the
exponential smoothing model.
39
So, let me write the exponential smoothing model here, it is called the exponential
smoothing model. Now, let us look at the 6 pieces of data 25, 32, 24, 28, 26, and 27, now
the exponential smoothing equation comes from. So, F t plus 1 is equal to alpha D t plus
1 minus alpha into F t.
I am going to follow this notation for sometime, where F represents the forecast and D
represents the demand, alpha is called as smoothing constant. Now, F t plus 1 is the
forecast for period t plus 1, so F t plus 1 is forecast for period t plus 1, D t is the demand
40
in period t and F is the forecast for period t. Now, in this notation this is D 1, D 2, D 3, D
4, D 5 and D 6, D 6 is the most recent demand D 1 is the oldest demand, since we know
the demand for 6 periods up to D 6, we are interested in forecasting the demand for the
seventh period. So, we are interested in finding out F 7.
Once again let me repeat that I am using this notation F t plus 1 is equal to alpha D t plus
1 minus alpha F t I am defining F t plus 1 as forecast for period t plus 1. There are other
notations possible right now it is a bit too early to describe other notations, but as we
move along at a suitable time, let me compare this with other notations and explain the
significance of this notation, but let us proceed with this notation. So, we are interested in
F 7 which is the forecast for period 7, because we know the demand up to period 6.
So, based on this F 7 is equal to alpha D 6 plus 1 minus alpha F 6, now we can write F 6
as alpha D 5 plus 1 minus alpha F 5. So, F 6 will now be written as alpha D 5 plus 1
minus alpha F 5 and proceeding further we can write F 2 is equal to alpha D 1 plus 1
minus alpha F 1. Now, we know D 1 right now we do not know alpha and we do not
know F 1, now if we assume let us assume alpha equal to 0.2 alpha as called a smoothing
constant, and alpha is a number between 0 and 1.
So, let us assume alpha equal to 0.2 we also do not know F 1, so if we know F 1 or if we
assume a value for F 1, then we can use this to compute F 2. And then we can substitute
F 2 somewhere here and find F 3, similarly use F 3 to find F 4, F 4 to find F 5 and then
once F 5 is known you can find F 6, and then once F 6 is known we can find F 7. So, let
us simply do these calculations first let the value of F 7, and then try to explain a few
things.
So, let us assume that F 1 is 27 which is the simple average, so if we use F 1 is 27, D 1 is
25, alpha is 0.2 and calculate we will get F 2. So, F 2 becomes 26.6 and F 3 is alpha D 2
plus 1 minus alpha F 2, which is 0.2 into 32 plus 0.8 into 26.6 and this on calculation will
give us F 3 equal to 27.68, now F 4 is equal to alpha D 4, F 4 is equal to alpha D 3 plus 1
minus alpha F 3.
So, F 4 is 0.2 into 24 plus point 8 into 27.68 and F 4 will become 26.944, F 5 is alpha
into D 4 plus 1 minus alpha into F 4. So, it is 0.2 into 28 plus 0.8 into 26.944 and F 5 will
become 27.1552 F 6 is 0.2 into D 6, F 6 is 0.2 into D 5 plus 1 minus alpha into F 5, so it
is 0.2 into 26 plus 0.8 into 27.1552 which would give 26.92416, and once F 6 is
41
calculated F 7 is alpha D 6 plus 1 minus alpha F 6, 0.2 into 27 plus 0.8 into 26.92416, so
F 7 will become 26.94.
So, F 7 is the forecast for period 7 based on this exponential smoothing model, we have
calculated the forecast for period 7 as 26.94, this 26.94 is calculated based on two
assumed values, one is alpha equal to 0.2 and the other is F 1 is equal to 27. Now, this
26.94 if we start comparing with these, it is very close to the simple average of 27 and it
is slightly far away from a weighted average of 26.833, if we consider a weighted
moving average on the last three values.
Now, what are the advantages of exponential smoothing and what are the areas we need
to address in exponential smoothing? So, first let us address perhaps the advantages of
exponential smoothing, now when we set out to do exponential smoothing we said that
we want a model where we use all the values. Now, we have used all the values we have
used D 1, D 2, D 3, D 4, D 5, and D 6 we have used all of them, and we also said that we
should give progressively increasing weights to more recent data. Now, let us see how
we give the progressively increasing weight to more recent data, now in order to do that.
So, let us write F 7 is equal to alpha 6 plus 1 minus alpha F 6. So, alpha D 6 plus 1 minus
alpha into alpha D 5 plus 1 minus alpha F 5, this on simplification will give alpha D 6
plus alpha into 1 minus alpha D 5 plus 1 minus alpha square F 5.
42
Now, F 5 will be written as alpha D 4 plus 1 minus alpha F 4, and if we continue t do this
we will get alpha D 6 plus alpha into 1 minus alpha D 5 plus alpha into 1 minus alpha
square D 4 plus etcetera plus alpha into 1 minus alpha to the power of 5 D 1 plus 1 minus
alpha to the power 6 F 1.
So, this is how this will expand to this if we keep inductively substituting the values for
this F. Now, if we look at this expression what we have calculated as 26.94 here, can also
be calculated in a single pass by using D 1, D 2, D 3, D 4, D 5, D 6 from these values or
from threes values, using alpha equal to 0.2 and using F 1 is equal to 27. Now, if we start
looking at these terms, each term uses the demand D and is multiplied by a number, here
it us alpha it is alpha into 1 minus alpha etcetera.
So, this can be treated as a weight associated with this demand, now I already mentioned
that when we took alpha equal to 0.2, I already mentioned that alpha is called the
smoothing constant, and alpha is a value between 0 and 1. So, 1 minus alpha is smaller
than 1, so alpha here alpha into 1 minus alpha is smaller that this weight, this is smaller
than this weight and so on. So, we have progressively decreasing weights from the more
recent data value which is 27 to the past.
So, 27 has a higher weight which is alpha which itself is less than 1, 26 has a weight
smaller than alpha, 28 has a weight smaller than this and so on. So, if we leave out the
last term the exponential smoothing is a weighted average, it is a weighted average of all
the demand points, with progressively decreasing weights as we move past or as we
move backwards in time. But, we also know that if we have a weighted average then we
have to divide by the sum of the weights.
So, let us find out the some of the weights the sum of the weights are alpha plus alpha
into 1 minus alpha plus alpha into 1 minus alpha square plus alpha into 1 minus alpha to
the power 5. Now, if you look at these sums this is one term, this is one term, this is one
term and so on, now we know that each term is multiplied by a 1 minus alpha. So, it is a
finite geometric series, some of terms have a finite geometric series is a into r to the
power n minus 1 by r minus 1, if r is greater than 1 and a into 1 minus r to the power n by
1 minus r if r is less than 1.
Now, if we make an assumption that n is large here we have n equal to 6 we have 6

demand points. So, if we make an assumption that n is large, then this will be alpha,
43
alpha into 1 minus alpha, alpha into 1 minus alpha square and so on, this will become n
or n minus 1 as a case may be. But, if n tends to infinity we get an infinite geometric
progression, with first term equal to alpha and common ratio equal to 1 minus alpha.
So, sum of an infinite geometric progression is a by 1 minus r, so that will be alpha by 1

minus 1 minus alpha which is 1. So, as the number of terms increases this will converge
to 1, and even within a very a reasonably small number of terms like 6 terms or 7 terms
or 8 terms, whatever be the value of alpha the sum will be very close to 1. And therefore,
this is like a weighted average, where the denominator which is a sum of the weights is 1
and therefore, we do not divide.
But, then we also have a term like this 1 minus alpha to the power 6 F 1, and if n is large
this term is 1 minus alpha to the power n into F 1. Now, as n is very large 1 minus alpha
to the power n tends to 0, if alpha is smaller than 1 and n is large 1 minus alpha to the
power n tends to 0 and this can be ignored. So, the exponential smoothing is like a
weighted average, which is what we wanted, we wanted to use all the terms we are using
all the demand terms.
We want progressively decreasing weights as we move to the past or progressively

increasing weights for more recent data, which is what is happening here. Some of the
weights adds up to 1, the last term tends to 0, so it is like a weighted average that we
have, and what is also interesting that given a value of alpha we get progressively
decreasing weights, which in some sense can be accepted by everybody provided
everybody accepts the value of alpha. So, over a given value of alpha we seem to be
getting a very nice and consistent set of weights which also add up to one.
So, exponential smoothing meets all other advantages or all the things that we expect
from it, which is use all data points give progressively decreasing weights to older data,
and have a consistently acceptable set of weights. So, which is what exponential
smoothing does, plus a fact that this weight is very small 1 minus alpha to the power n is
very small, even for a reasonable value of n. So, the F equal to 27 actually does not
impact the final answer significantly.
Because, the weight associated with this F 1 is practically 0 therefore, the initial value of
forecast can be taken as 27 which is this simple average, sometimes people would take
the last of the values as F 1 which also happens to be 27 in this case. But then even if we
44
take 26.5 and 26.35 or whatever as long as it represents adequately another measure of
central tendency or average the effect of this F 1 is not very high. So, in a way we can
say that F 1 equal to simple average can be accepted, but we have to spend a little more
time on this alpha. Now, first of all alpha should satisfy two conditions, alpha is greater
than 0 and alpha is less than 1. So, 0 less than or equal to alpha less than 1 and why are
we looking at this?
Now, if we write or if we come to the most basic equations which is given here from this
equation we understand that the forecast for period t plus 1 is a sum of two things, which
is a weight associated to the current demand and a weight associated to the current
forecast. And because it is a sum of two terms a weighted sum of two terms, we do not
want a negative to come here we want both of them to contribute, and we do not want a
negative term to come here.
Therefore, we will not have alpha greater than 1, the moment alpha becomes greater than
1 then we realise that this term starts contributing in a negative manner. Similarly, the
moment alpha is less than 0, then this term gives a negative contribution, since we do not
want the negative contribution from either of the terms, we have alpha between 0 and 1,
now, having kind of agreed on the fact that alpha has to be between 0 and 1.
Now, what should be the value of alpha should alpha be small or should alpha be large,
now in this example we have taken alpha is equal to 0.2. So, we have taken alpha as a
45
smaller value, and not as a larger value, now if we take alpha as a smaller value and start
substituting here, we will immediately realize that the weight or the contribution of the
demand is actually if D t and F t are comparable then the contribution of the demand
seems to be less than the contribution of the forecast.
If alpha is greater than 0.5, then the contribution of the demand will be more than the
contribution of the forecast. There it actually brings us to another question which says,
should not the contribution of demand be more than that of the forecast, which means
should alpha be closer to 1 or should alpha be closer to 0. If alpha is small and 0.2 then;
obviously, the contribution of forecast is higher, but if we look at it very carefully we
would want the contribution of forecast to be higher.
The whole forecasting comes because this demand is not extremely stable and it exhibits
a certain amount of noise, just for the sake of argument if all these numbers were 27,
then we do not forecasting and we would say that the next value is 27. So, the demand is
very stable and does not show noise or variation within, in such cases we can have a
larger alpha. So, long as there us variation and noise it is advisable to have a smaller
value of alpha, which means here, we realize that the weight associated with the demand
in this term can be less and the forecast will have a higher weight.
There is another way of explaining it if alpha is large here, then what will happen is this
F 7 will be guided by fewer terms which will contribute to F 7. For example, if alpha is
46
0.8 now this is 0.8 this is 0.16, so two terms are contributing to a weight of 0.96, so they
have almost contributed to F 7. Whereas, if alpha is 0.2 then this is 0.2 and this is 0.16,
so two terms are contributing to a weight of 0.36.
So, smaller the alpha more terms will start contributing to F 7, smaller the alpha in some
sense the weights even though there are progressively decreasing weights, this will
contribute a little more to it smaller the alpha, larger the alpha these the weight
associated with this become very, very small. So, if we want all of them to contribute
reasonably well alpha has to be slightly smaller, so this is how the exponential smoothing
model works and is dependent on the alpha and F 1.
And exponential smoothing provides us with a very good model to forecast the level,
where your able to use all the data, we are able to give progressively decreasing weights
to the data as we move past or progressively increasing weights to more recent data. So,
with this we have seen a little bit of forecasting for level or constant data, in time series
how to do forecasting when the data exhibits trend we will see in the next lecture.
47
Operation and Supply Chain Management
Prof. G. Srinivasan
Lecture - 3
Forecasting – Linear Models, Regression, Holt’s, Seasonality
We continue our discussion on Forecasting Models. In the last lecture, we developed

forecasting models for this data. This data we assumed represented a constant model and
we looked at forecasting models for this data, we looked at simple average, weighted
average, moving averages and simple exponential smoothing. And all of them, we used
to test on this model and in this model; we assumed that this data represents a constant
and we were, trying to find out that constant. The fact that each of these data does not
represent the constant comes from our belief that there is an inherent variation or noise
or something, which cannot be assigned to a specific cause or a specific reason.
Therefore, we assumed that this represents the constant and we tried to fit a model that
gives us that constant. Now today, we will look at a different type of data and we will
move to this data. Now, if we look at this data and if we ask anyone to forecast this data,
we will end up giving a forecast, which is greater than 35, because as we move along
these numbers, we realise that there is an increasing trend.
48
So, we use the term trend to show, a certain behaviour and in this case, there is an
increase, now we will not and should not fit a constant model, for this data, for example,
if we took this simple average, then the simple average the total of these 6 numbers,
happens to be 181 and the simple average will be 30.16. Now, simple average 30.16 is
always within the range of values and therefore, will be somewhere here while a
meaningful forecast would be a number, which is more than this 35.
So, we do not use the earlier models directly, we do not use models like simple average,
moving average, simple exponential smoothing etcetera directly, to give the forecast for
this. So, we have to now build different forecasting models that are capable of capturing
the trend, which this data is exhibiting. So, let us first look at very a simple model that
captures trend and then we will look at a more advanced model, that can capture trend
now before, we move to any model let us try and plot these with respect to time. So, let
us call this as time equal to 1, time equal to 2, time equal to 3 4 5 and 6 let us plot this.
The time on the x axis and the data on the y axis, so let us plot here time and then here,
we have D which is the demand or sale for this particular product. So, at time equal to 1
2 3 4 5 6 and let us have the values, we start here from 24 26 28 30 32 34 36 and so on.
So, at t equal to 1, the value is 26 which is here, t equal to 2, the value is 28 which is
here, t equal to 3, the value is 29 which is here, t equal to 4 the value is 31 which is
49
somewhere here, t equal to 5, the value is 32 which is somewhere here and t equal to 6,
the value is 35 which is somewhere here.
So, this is how the 6 points are plotted, when you plot them on this and one look at these
6 points kind of conveys to us that, there is an increasing trend and they in some sense, if
we try to draw a line, the line will go like this. Now, then when we try and draw a line,
we ask ourselves a couple of questions are we going to draw a line; obviously, all these
six points are not collinear. So, are we going to draw a line, which is going to pass
through, maximum number of points or are we going to draw a line, which kind of best
represents these 6 points by being as close to each of these 6 points as possible.
So, the answer is we do the second and we then try and fit a line or draw a line, which is
as close to these 6 points as possible. So, let me first draw the line using a free hand and
then explain a few things, let us assume that this line is going to be like this. Now, if we
draw this line, then what we are going to do is now this point is represented by this point
on the line, say this point is represented by this point, this is represented by this point,
this is represented by this point, this is represented by this point and this is represented
by this point.
Now right now though, I have drawn this line, I do not know the equation corresponding
to this line, because I have made a free hand sketch. So, I would like to now find the
equation corresponding to this line, one of the ways to do that is by, if we consider a
point like this point is going to be represented by this point on the line. And therefore,
there is going to be some kind of a gap between the actual point and the point, which is
there on this line and let us say, we call this gap as the error, we call this as the error.
Now, if we want this line to be as close to these 6 points as possible, then we will have to
draw the line in such a manner, that the errors are as small as possible. The moment we
define that the errors have to be as small as possible, we also have to observe that in
some situations, the error can be positive while in some other situation, the error can be
negative in the other direction.
So, we obviously, do not want these positives and negatives to cancel out each other,
therefore a better way to look at it is to say that, we want to draw this line, in such a
manner that the squares of these errors are minimized or essentially, we try to minimize
the what is called the error sum of squares. Try to minimize the error sum of squares,
50
which is essentially, we will try to minimize e square, where e is the error corresponding
to each point.
So, there will be 6 points here, there will be 6 error terms, now we want the sum of the
squares of these 6 errors to be minimized or if this line is going to have coordinates a
plus b t, where a and b are to be determined, then the error term will be Y minus a minus
b t is the error term. Error term specific to a particular point say the t-th point can be
written as e t is equal to Y t minus a minus b t, now sum of squares of these errors will
become sigma e t square is equal to sigma Y t minus a minus b t square.
Now, we want to minimize the error sum of square, so we minimize error sum of
squares, which gives us minimizing Y t minus a minus b t the whole square, now we
want to find out the a and b, which represents b represents the slope and a represents the
Y ordinate or the value on the Y axis for t equal to 0. So, a and b have to be found out,
such that this error sum of squares is minimized, now let us try and derive a very simple
expression to find out a and b.
So, we want to minimize Y minus a minus b t square here, I have represented it as Y t

minus a minus b t here, I am writing it as Y minus a minus b t, we could also write it as
Y t minus a minus b t. Now, this has to be minimized and a and b have to be found out,
so the condition for minimization is the first derivative equal to 0 with respect to the
51
variables. So, we try to partially differentiate this with respect to a and then with respect
to b and then try and get the values.
So, partially differentiating with respect to a we would get 2 times Y minus a minus b t
with a minus sign summation, so summation equal to 0. And partially differentiating
with respect to b we would get sigma 2 times Y minus a minus b t into minus t is equal to
0, observe that when you differentiate with respect to b, so it becomes 2 times Y minus a
minus b t and then differentiating with respect to b. So, we get a minus t, so the minus is
written here and the t is written here, now we expand this summation and we therefore,
write 2 equations the minus becomes irrelevant here, because the right hand side is 0.
So, we get an expression sigma Y is equal to n a plus b sigma t, this goes the 2 also goes,
so sigma Y is here, you take it to the other side, so sigma a is n times a where n is a
number of data points, so n a plus b sigma t b is a constant that can be taken outside, so b
sigma t is the first equation. Second equation will be sigma Y t is equal to a sigma t plus
b sigma t square, now multiply here, so Y t again this minus and this 2 goes, so you have
Y t is equal to a sigma t and then plus b sigma t square comes, because this t is multiplied
by the other t.
So, you have 2 equations and 2 unknowns a and b, we can directly solve using this, now
there are times it is also possible to simplify this to write unique expressions for a and b.
But, right now, we are not going to do that we are going to keep these equations as they
are and we are going to solve, now if we consider this data 26 28 29 31 32 35.
52
So, we look at this data 26 28 29 31 32 35, this is your Y t or Y as you may call, so this is
Y 1 Y 2 Y 3 Y 4 Y 5 and Y 6, we need sigma Y or sigma Y t is a sum of all of them,
which is 181 14 23 24 26 31 181. So, this term we have written we need b sigma t, so t is
1 2 3 4 5 6, so this is 21 3 plus 3 6 plus 4 10 15 plus 6 21, so we have this term b sigma t
then we need the term for sigma Y t.
So, you need a Y t term, so Y into t so 26 into 1 is 26 28 into 2 is 56 29 into 3 is 87 31

into 4 is 124 32 into 5 is 160 35 into 6 is 210 and the sum is 12 19 23 2 3 9 11 19 24 26 4
5 6 23 2 9 19 24 26 2 3 4 5 663 and then we have now found out this term sigma Y t now
we need to find out this term, which is sigma t square. So, we write t square as 1 4 9 16
25 36, so this total will be 5 plus 9 14 30 55 91 71 661 77 86 91, so now, these 2
simultaneous equations become sigma Y 181 is equal to n a 6 a plus b sigma t 21 b and
663 is equal to a sigma t 21 a plus 91 b.
So, we have 2 equations with 2 unknowns and we need to solve these 2 for a and b, so
the easiest thing to do is to multiply this by 2 and multiply this by 7, so that we can
cancel the a and then get the value for the b and then we go and substitute to try and get
the value for a. So, when we solve this, we get the solution a equal to 24.2667 and b is
equal to 1.6857, this means that this line is going to intersect the Y axis at a value of
24.2667 and this line of best fit has a slope of 1.6857.
53
But, then we are interested in finding a forecast for the seventh period, which actually led
us to drawing a line and finding out the coordinates of that line. So, now, for the 7th
period the forecast that we are going to make is the point corresponding to t equal to 7 on
this line, which is this point is what we are interested has the forecast for the 7th period
and that is given by a plus 7 b. So, f seven forecast for the 7th period is a plus 7 b, which
is 24.2667 into 7 times 1.6857, which is 36.066.
So, this is our forecast for the period number 7, and let us say this is going to come at
36.066, which is the forecast for period 7. So, the model that, we have just now seen is
the linear regression model, which also gives us the line of best fit given these set of
points, now the forecast for the 8th period can be given by F 8 is equal to a plus 8 b and
so on.
Now, this method is a very intuitive nice method to try and forecast, when the data
shows a certain amount of trend, now let us go back and look at this method again, what
have, we done or what are the assumptions behind this method of linear regression for
time series data with trend, what are the assumptions. Now first thing we have done is
we have considered all the data points, second is we have given equal weight age to all
the data points, even though we squared the error and tried to minimize, we did not
associate a weight with each of these.
For example we did not do minimize w t into Y minus a minus b t the whole square, we
did not do that, which means that we have given equal weight age to all the data points
and third is we have tried to minimize the error sum of squares. Now, let us go back a
little bit and try to see, whether we use the same assumptions, in any of the models for
the constant data that, we have here.
Now, for example if we make the same 3 assumptions on this, which means we fit a
model, which uses all the data points, which fit a model, which gives same weight to all
the data points and we try and fit model, which minimizes the error sum of squares.
Except that, because this data is constant data, we fit simply Y equal to C and because
this data shows trend, we fit Y is equal to a plus b t, now suppose we do this then we fit
Y equal to C and then we minimize the error sum of squares.
So, a error sum of squares will be sigma Y minus C the whole square and we want to
find out C such that, Y minus C the whole square is minimized, then the answer for C
54
will be simply sigma Y by n, which is the simple arithmetic mean. So, if we try to fit a
model Y equal to C using linear regression and minimizing error sum of squares for a
constant data, we get the simple arithmetic mean, when we make the same assumption
that, we want to consider all the data points, equal weight age to all the data points.
And minimizing error sum of squares considering trend and fitting a plus b t, we get this
linear regression model, when we were looking at this set of data the constant data. We
also said that even though, it is a good idea to use all the data points, we would rather
give a little more weightage to the recent points on little less weightage to the older
points and that thinking of ours let us towards moving averages and particularly towards
exponential smoothing, which we saw in the previous lecture.
So, if we are convinced that exponential smoothing is a very good model to do

forecasting for this type of data then the obvious question is, can we have models based
on exponential smoothing to do this data with trend? We also know that simple
exponential smoothing, that was used here was very effective, but simple exponential
smoothing, would only give us the forecast, which is within the range, just a simple
average gives us a forecast, which is within the range.
So, a simple exponential smoothing, obviously, we will not work here and simple
exponential smoothing will actually lag behind the last value, it will give you a value
closer to the mean, which means, it will lag behind the last value. So, we need a model,
which uses ideas from exponential smoothing at the same time gives different weights to
different data points with more weight age to the recent points and lesser weight age to
the old points. We will now discuss the Holt’s model, which does exactly that, it
considers all the data points, it uses ideas from exponential smoothing and it gives
progressively more weights, towards recent points.
55
In the Holt’s model, we fit F t plus 1 is equal to a t plus b t, now F t plus 1 is the forecast
for period t plus 1, a t is called the level, which represents the smoothed value up to and
including the last data, b t is the slope of the line that, we are fitting at the point t. So, a t
is the representative value of the level or the constant and then we add a slope to it. So,
that we get the forecast of the next period, so forecast of the next period will be level at
the end of the previous period or the present period and then we add a slope to it, so a t
plus b t.
Now, both a t and b t are updated for every data point using exponential smoothing, so a t
which is the level value up to an including the last point is given by alpha D t plus 1
minus alpha into a t 1 minus plus b t minus 1. Actually, this is our very familiar
exponential smoothing equation the normal exponential smoothing equation has 2 parts,
one is the demand part and the other is the forecast part alpha D t plus 1 minus alpha F t.
Now, alpha D t is the same alpha is the weight D t is the demand during the last period 1
minus alpha I had used F t in the last lecture, but I had also mentioned that there are 2
versions of it. So, now, we will use a version alpha D t plus 1 minus alpha F t minus 1, so
F t minus 1, which is the forecast made at the end of the period t minus 1 for period t. So,
in some sense D t and F t minus 1 are comparable, F t minus 1 is the estimate of D t, in
this notation F t minus 1 would mean that it is a forecast made at the end of period t
minus 1 for t.
56
So, a t minus 1 plus b t minus 1 represents, the forecast for period t, just as forecast per
period t plus 1 is represented by a t plus b t. So, here the equation is like alpha D t plus 1
minus alpha F t minus 1 right, when alpha D t plus 1 minus alpha F t where F t is for a t
minus 1 plus b t minus 1 forecast made for period t is a t minus 1 plus b t minus 1. Now,
b t, which is the slope also has 2 components, one is in some sense the component that
comes out of the level and the other is the component that comes out of the slope.
So, beta is another exponential smoothing constant like alpha, beta into a t minus a t
minus 1, if you take the present level value a t and take the previous level value a t minus
1, the difference between them in some sense represents a slope. So, that is 1 component
of the beta the b t the other component comes from the actual slope that was computed.
So, the b t also has 2 components, that are related through exponential smoothing, the a t
also has 2 components that are related, through exponential smoothing.
Now, let me explain this further, now when we did the Y equal to a plus b t through
linear regression, now these were the 6 points, the white dots are the points and we tried
to fit a line that pass through all of them and is as close to the 6 points as possible. In the
Holt’s model, we would not do that, for example, we are here at the 6th point and we are
interested in forecasting the 7th point. Now, this white thing is your D 6, the demand at
the 6 point, which is a data that we know, the level value at t equal to 6 is the equivalent
of this point, which is actually representing that.
57
The level is obtained by through exponential smoothing considering, these things in a
certain manner up to and including D 6, we have a level, which is here. So, once we
know this level plus we also need the slope at this point, in the linear regression model,
after we fit the line, the slope is the same at every point in the line, in the Holt’s model,
we are going to kind of redefine the slope at every point up to and including the last
point. So, we will do some calculations here to find the slope and therefore, a 6 plus b 6
will give us F 7, how do we find a 6 and b 6.
Now a 6 and b 6 have 2 components, now a t, now if we look at this a t has a component
alpha D t plus 1 minus alpha into a t minus 1 plus b t minus 1. So, a t has 2 components,
one that comes out of this demand, the other that comes out of this is your a t minus 1
plus b t minus 1, this plus this. So, it is use this is computed using this one and this plus
the slope here and they are related through an exponential smoothing equation.
Similarly, the slope at this point also has 2 components one is the slope at t minus 1, the
slope at this point and the other is essentially, the slope between this point and this point,
which comes from a t minus a t minus 1. So, in some sense in the Holts’ model for every
point t, we are actually trying to find out a level and a slope and we add them, so that we
get the forecast for the next period then we know that forecast for the next period and
then we know that demand point and keep iterating this till, we get to the final answer.
58
So, let us explain the computations in the Holt’s model, for the same set of data 26 28 29
31 32 35, we have already seen.
That we have get a forecast of 36.066 when we fit this line F 7 became 36.066, let us see
what happens, when we fit using the Holt’s model.
Now, the same data, we are going to use 26 28 29 31 32 35 like in all exponential
smoothing, we need some initial values and we also need to fix the values of these
smoothing constants. So, we fix alpha equal to 0.2 note that.
59
Alpha is the exponential smoothing constant, then we compute a t and beta is the
exponential smoothing constant, when we compute the slope. So, alpha is 0.2 beta is 0.3
is what we have assumed in our calculations. We also need some initial values, so the
very first value, which is 26, we assume D 1 to be 26; D 1 equal to to 26. So, we assume
that the level the first point that, we calculate here.
This is 26, so the white one, the white point represents D 1 and the other colour point
represents a 1. So, right now we are going to assume that the level value for the first
60
point is the same that is one of the assumptions. And we also need to initialise a slope
here and in order to do that here, we actually try, we know these 6 points, so with a
known 6 points the initial slope is the approximate slope of the line joining, the first
point and the last point.
So, the initial value of the slope is therefore, 35 minus 26, which is 9, 9 divided by 5,
which is 1.8, so the initial values of a 1 and b 1 are 26 and 1.8, so from our equation for
Holt’s model F 2 is equal to a 1 plus b 1, which is 27.8. Now, we need to find out a 2 by
our previous equation a t is alpha D t plus 1 minus alpha a t minus 1 plus b t minus 1, so
a 2 will be alpha D 2 plus 1 minus alpha D 1 plus 1 minus alpha into a 1 plus b 1. So,
alpha D 2 plus 1 minus alpha into a 1 plus b 0.2 alpha is 0.2 D 2 is 28, comes from here
plus 0.8 into a 1 plus b 1 is what we calculated, which is F 2.
So, 0.8 into 27.8, which becomes 27.84, so when we apply the Holt’s model, this value is
28, the white 1 is 28, which is the demand point and this 1 is 27.84, when we apply the
Holt’s model. So, we got 27.84 here the slope at this point, the slope at this point is now
calculated the slope at this point was assumed to be 3, calculated to be 3, here it is now
calculated using exponential smoothing.
So, the slope at this point is now beta into a 2 minus a 1 plus 1 minus beta into b 1, so
beta is 0.3 a 2 minus a 1 is the difference between the computed 27.84 and 26, which is
1.84 plus 1 minus beta is 0.7 into the previous slope, which is 1.8, so it now becomes
1.812 and F 3 is a 2 plus b 2, which is 29.652. Now, D 3 was 29 and F 3 is 29.652, now
we need to find out F 4, we need a 3 plus b 3, now a 3 and b 3 have to be calculated
using these.
61
So, we can now progressively calculate the rest of the numbers and this table explains
what these numbers are, so for period 4 D 4 is 31 a 4 is 31.2352, b 4 is 1.755 and this is
31.294. And if we move on, we find out finally, that F 7 is 36.32, so let me write this
36.32 here.
This is using the Holt’s model, this is from the Holt’s model, now in the Holt’s model, if
we look at this finally, a 6 is 34.592 b 6 is 1.7272, so a 6 is here, the representative value
was 34.592, b 6 was 1.7272 and finally, we got 36.32. So, the representative value here
62
was 34.592, the actual value is 35, the slope is about 1.8, but the slope has come down to
1.7272 and the final value is 36.32. Now, we observe that the final the forecast given by
Holt’s model is slightly higher than the forecast that, we got using the linear regression.
Now, it is not absolutely necessary that every time, the Holt’s model will be slightly
higher than linear regression, but it is very likely, so one of the reasons is if we look at
this data, the value keeps increasing, but then if you look at the last few numbers the
increase is slightly higher than, what it was before. So, in when we fitted the line, we
gave equal importance to all these points and therefore, the smaller increase, essentially
pulled it back a little bit and we got 36.066.
Now, because we are giving higher weightage the 35, which shows a larger increase is
tending to pull it slightly towards this 35 and therefore, the value became slightly more
than this. But, one of the advantages of Holt’s method is that, it captures idea of
exponential smoothing, which means we are giving more weight, to the more recent
point while doing the forecast and in some sense.
We are actually calculating the level and the trend the forecast at every point, through an
iterative process that involves exponential smoothing unlike linear regression, where we
compute it in one shot by giving equal importance to all the points. So, in some sense
what simple exponential smoothing did to this data, Holt’s method tries to do, to this
type of data. So, this is how we normally forecast, when the data exhibits trend, we will
now see one more aspect of forecasting.
63
So, let us assume that we are looking at this kind of a data, where the year is divided into
4 quarters and the data is given for each quarter, for the 3 years, so if we take year 1, the
quarterly data are 53 22 37 and 45. Now, when we start looking at this table, we observe
that, there seems to be a trend in this direction, for all of them the total also seems to go
up, but each quarter of the year is not showing constant kind of data, there is a variation
in this, now this thing is called seasonality.
So, each quarter represents a season and the data is showing a certain seasonality in this
case, so we need forecasting models that capture seasonality, seasonality is important
because several products exhibit seasonality, common examples, that are given are the
sale of ice creams are higher in summer than in winter. The sale of air conditioners is
higher during summer, than during winter, the sale of in some places the sale of heaters
is higher in winter than in summer, sale of sweaters are meant to be higher in winter than
in summer and so on.
So, products exhibit a certain seasonality, now can we capture that seasonality of that
product and then make the forecasting is the next question, so we will first build a simple
seasonality model and then we will also develop or explain a seasonality model that uses
ideas from exponential smoothing. So, the simple seasonality model is like this, first let
us find out the total demand for the 4 years and this is 157 173 189, so these are the total
demands in the previous years.
64
Now what we do is we try to associate a seasonality index with every season, so a
seasonality index associated with the first quarter is the proportion of the total demand
met by the first quarter. So, that will be 53 divided by 157, so 22 divided by 157, 37
divided by 157 and 45 divided by 157, now again seasonality index in the second year
will be 58 divided by 173, 25 divided by 173 and so on.
So, we compute the seasonality indices for the 3 years and for the 4 quarters. So, the
seasonality indices are, so these 453 divided by 157 is 0.3422 divided by 157 is 0.14 and
so on. So, the seasonality indices are calculated this way, now from these seasonality
indices, we can all these will have to add up to 1 and from these seasonality indices, we
can find out the average of the seasonality indices. So, the average of these seasonality
indices are given here 0.337 0.14 and 0.293, these are average across, this way.
Now, some of the averages also has to be equal to 1, but sometimes there could be some
rounding of errors in the average and therefore, this may be slightly less than one or it
may be slightly more than 1. But, many times it is exactly, it will add up to 1, if it does
not we can normalise it to 1, so that the sum of the indices is also 1.
So, now, we have captured these indices and based on this average over 3 years, we can
say that quarter 1 accounts for 33.7 percent of the total demand on an average quarter 2
accounts for 14 percent of the total demand quarter 3 accounts, for 23.3 percent of the
total demand and quarter 4 accounts for 29.3 percent of the total demand. Now, the totals
65
are 157 173 189 and the totals show a certain increasing trend or a linear trend, now we
can fit a model to forecast the demand, for the 4th year.
Now, for this particular example, there are only 3 data points, so it will be difficult for us
to fit an exponential smoothing based model or a Holt’s model, so we rather go for linear
regression and then we can fit a model of the type Y is equal to a plus b t. So, when we
fit a model of the type Y is equal to a plus b t, we get Y is equal to 141.141 plus 16 t and
the forecast of the demand for the 4th period at t equal to 4 will become 205. So, 205 will
be the forecast of the demand, for this period and since, we know that quarter 1 accounts
for 33.7 percent of 205 quarter 2 accounts for 14 percent of 205.
We simply multiply by the proportions to get the 4 forecasted values, which are 69.08
28.7 47.77 and 60.06 are the forecasted values, for the 4 season, so 62 makes forecast to
69.08 27 28.7 and so on. So, this is a very rudimentary algorithm to do seasonal
forecasting, where we de seasonalize the data, for compute the seasonality indices
average, the seasonality indices separately take the total demands, forecast the demand
and multiply by the seasonality indices to get the new forecast for all the 4 seasons.
There are only a couple of questions in this rudimentary algorithm or method and we
will see that, today before we close this lecture, first question that is commonly asked is
why cannot, I take only these 3 values and independently fit a Y equal to a plus b t and
get the forecast. Mathematically, it is possible and it is do able, if you do that, you will
even get a value, which is very close to 69.08.
Normally that is not done, because it means 2 things, it means you are treating this
quarter as independent of the other 3 quarters whereas, when you fit a seasonality model
you are capturing somewhere the dependency between or amongst the quarters. So, you
do not do this independently, it is also equivalent of saying, there are 12 data points, I am
picking up data point number 1 5 and 9 and I am trying to forecast the 13th data point,
which is not normally done.
So, we do not normally, do that we de seasonalize, it and then we bring it back second
question that is often asked is now what is this itself shows a trend 29 to 29 and 30
should I average this or should I fit another trend model right here, assuming that there is
an increasing trend. The normal answer is you average it, unless you really have several
66
years of data, to show that in a particular quarter, there is the index itself is showing a
trend.
If the index itself is showing a trend and if you are convinced you can fit a trend model,
there which also means that, if 1 index is showing an increasing trend, we are somewhere
else in some other quarter the other the it will show a decreasing trend. But, ordinarily
the average is taken de-seasonalized. So, the next model, which is called the winter’s
model where, we try and capture the level the trend and the seasonality index, using
exponential smoothing, we will see in the next lecture.
67
Prof. G. Srinivasan
Module - 01
Lecture - 04
Forecasting – Winter’s Model, Casual Models, Goodness of forecast, Aggregate
Planning, Tabular Method
In today’s lecture we continue our discussion on seasonal models; in the previous lecture
we demonstrated a basic model considering seasonality. So, we looked at this data and
this data would represent the demand over 3 years and each year is divided into 4
quarters. In the previous model, we computed a seasonality index, for example
seasonality index for the first quarter data of year one would be 53 divided by 157.
The seasonality index for this number which represents second quarter of year 1 would
be 22 by 157 and so on and we computed a seasonality matrix and then we averaged the
seasonality indices for every quarter. We also try to forecast the next year’s total demand
by using the regression or the method of least square fit. Then, we said that we could
multiply the forecasted value with each of these indices to get the forecast for the 4
seasons for the next year.
Now, what we will see is can we use exponential smoothing in some form to try and get
68
a model which can in principle forecast better in the sense that more recent data will
have more weightage. For example, when we used the method of least square to fit a line
here we gave equal weightage to all these points. Now, the question is can we use
exponential smoothing and try and build another model to do this. Now, that model is
called the Winter’s model and the Winter’s model has the following equations.
So, we use these equations for the Winters model at present, these equations may look a
little complicated, but it is easy to explain these 3 equations. Now, a represents the level
or which is the base data b represents the trend on the slope and c represents the
seasonality. Now, there are three things in a seasonal model, a level, a trend and
seasonality, for example to explain with this data the level values are like 157, 173, 189.
It actually exhibits both level and trend one could think of 157 as a level and then there is
an increasing trend. So, there is a slope which takes it to 173 and then to 189 and so on,
there is seasonality, for example the seasonality index associated with this is 53 divided
by 157.
So, now if we see that the level for a period t plus 1 has two parts exactly as in the
exponential smoothing equation alpha is a smoothing constant. So, alpha times d t by c t
d t plus 1 by c t plus 1 plus 1 minus alpha times a t plus b t, now if we are in period t and
at that time the level value is 80, then the level value for the next period will be the
earlier level plus slope. So, you have an a t plus b t which kind of represents the a t plus
69
1, similarly if there is a demand d t plus 1 in that period and there is an associated
seasonality of t plus 1.
So, demand by seasonality would also give the level value at that period, therefore the
actual level value of t plus 1 is a weighted average of this term and this term where is
multiplied by alpha and this is multiplied by 1 minus alpha. So, if you are in period t and
then we want to find out the level for period t plus 1, one can use this particular formula
to get the level at period t plus 1. Now, similarly slope at that period also has two
components now always the difference between two levels represents a slope. So, a t
plus one minus a t is one measure of slope and the slope at period t which is b t is
another measure of slope.
So, the slope at or for period t plus 1 is a weighted sum of 2 slopes, so beta times a t plus
one minus a t which is again equivalent of the slope and 1 minus beta times b t where
beta again is a exponential smoothing constant. Similarly, seasonality you can see the
subscript is t plus p plus 1 which essentially comes because if I am somewhere here, then
I am this is if this is called C 1, C 2, C 3, C 4, C 5, C 6, and C 7. So, if you are at two and
the number of periods is four t plus p plus one is c seven is what we are computing
which is essentially computed from this as well as this.
So, you will have d t plus 1 by a t plus 1 demand by the total the level which represents
the total demand this is a measure of a seasonality index. This is also the measure of
seasonality index from the previous one, so once again it is a weighted sum of two
seasonality measures, where gamma is the exponential smoothing constant. Now, after
we compute all these, we try and get the actual forecasted value which is given by this
equation. So, what we will do now is we will try and show some computations and
explain how we do the calculations for the Winters model.
So, let us take the same example and do that, so in order to do this we also need to do
some initialization so we will assume C 1 to be equal to 0.34 which actually is 53 divided
by 157, C 2 is 0.14, C 3 is 0.24 and C 4 is 0.29. So, let us initialize these three values,
now we also a need to initialize a 1 where a 1 is the level at this period. So, a 1 is
initialized to 53 divided by 0.34 which is 156, there is small rounding of error which can
which is acceptable. So, 0.34 was obtained on the division of 53 divided by 157, so it is
rounded of there, so the rounded value gives as a 156 as the level value.
70
So, we now know the level value here which is a 1, now we have to find out the level
value for a 2 at the corresponding slope b 2 and so on. So, we also use alpha equal to 0.2,
beta equal to 0.3 and gamma equal to 0.25 as the smoothing constants and we also need
to use a value of b 1 which is taken as 4 which is roughly the slope here because 157,
173. The difference is 16 over 4 periods gives us a slope of 4 once again 173, 189 gives
as a slope. So, we now have initialized all the value, so that we can do these
computations, so let me show those computations again.
So, a 2 which is level for period 2 is 0.2 into 156 plus 0.8 into 156 plus 4 which is
159.43, now how do we get those numbers a t plus 1 level for period 2 is alpha times d t
plus 1 by c t plus 1 alpha is 0.2 here. Now, d t plus 1 by c t plus 1 is 156 that we have
plus 0.81 minus alpha into a t plus b t which is a 1 plus b 1 160, so the level moves up
from 156 to 159.43 for the next period a 2. Now, b 2 is obtained from 0.3 into 159.43
minus 156 plus 0.7 into 4 which is 3.829, now this comes from this equation beta times a
2 minus a 1 159.43 which we just now calculated minus a 1 which is 156, 1 minus beta is
0.7 b t is b 1 which is the earlier slope 4.
So, you get 3.829, then we compute C 6 is equal to now essentially we are trying to look
at seasonality for somewhere here using these seasonal values. So, C 6 will be 0.25 into
22 divided by 159.43 plus 0.75 into 0.14. Now, these things come from gamma into d t
plus 1 by a t plus 1 gamma has 0.2 d t plus 1 is 22 a t plus 1 we just calculated as 159.43.
71
So, this is a measure of seasonality of this quarter which is represented by C 2, C 6, C 10,
so 22 divided by 159.43, this is a measure of seasonality plus 0.75 1 minus gamma into
0.14 which is the initialized value for C 2.
So, when we do this C 2, C 6 becomes 0.1395, now from this we can initialize f 3
forecast period 3 is 159.43 plus 3.829 into 0.24 which is 39.18. Now, let us go back to F
3 this is our F 3, this is the actual d 3 is known, so the forecast for this is 139.18 that
comes from the level at the end of 2 is 159.43. So, level plus slope would take it to
159.43 plus 3.829 which is here and that is multiplied by the seasonality which is 39.18.
Now, we have shown one full round of calculations, now once this is known we can
continue the calculations by now starting to compute a 3. So, a 3 will now become alpha
times d 3 by C 3 plus 1 minus alpha times a 2 plus b 2.
So, now we know all the values d 3 can be got from here which is 37, C 3 is 0.24, alpha
is known, 1 minus alpha, therefore is known a 2 and b 2 are calculated are available
there. So, we can calculate this a 3, similarly we can calculate b 3 as beta times a 3 minus
a 2 a 2 is known a 3 has been just calculated 1 minus beta times b 2, b 2 is known from
the previous one. So, we can calculate this one, similarly we can calculate C 7 because
gamma is known d 7 by a, this is 7. So, d 3 by a 3 is known plus 1 minus gamma C 3 is
also known, so we can calculate c seven from which we can calculate F 4.
So, we can continue with the Winters model and finally, we have a value f 13 forecast
for period 13 a 67.29 with the updated values seasonality values are also calculated
accordingly and kept updated. So, finally the forecast for period 13 is 67.29, so if we
want to forecast the value here from using Winters model we would get 67.29 here and
then we use this model to compute this and so on. Now, what are the advantages of the
Winters model as we have use exponential smoothing. The very idea that we consider all
the data points we give more importance to or higher weightage to the more recent data
points and so on.
Now, all the three important things like level slope and seasonality index are updated
using a simple exponential model where in some sense two things are captured the value
at period t and then the value at period t plus 1. So, they are captured in this each has 2
components connected by the weights and the alpha beta and gamma are the smoothing
constants. Then, we have a relationship that the forecast is level plus trend multiplied by
72
seasonality. Now, this also answers a very interesting question that we posed in the
previous lecture, now if we have a data like this, now can we simply take 53, 58, 62 and
forecast the next value can we take 22, 25, 27 and forecast.
The next value the answer to this is no from a forecasting point of view the reason being
the assumption that even though there is an increasing trend here and the increase in
trend is visible when we actually take the demand total demand for the year. It becomes
visible when we do that, if we see here we see an increasing trend, but we do not actually
see an increase in this direction because the values are very different. They are multiplied
by the corresponding seasonality index by using the Winters model what we have trying
to say here is that in some sense the 16 increase here is actually spread as 4 in each
season.
That is the reason we also got something like 67.29 which is which is sixty 2 plus some
four plus something that plus came because of the increased weights to more recent data.
So, to that extent Winters model is able to capture is able to put into each period the fact
that there is an increase and that increase is captured by deseasonalizing it. Second
advantage perhaps is if we were given the demand here let us say if the actual demand
here turned out to be 67 and then we want to make a forecast. Here, we could directly use
this formulae and get the forecast here. Where as if we are delink them and said I am
going to do this and then some something here, without doing any of these we are not
capturing the effect of these data points on a subsequent forecast.
So Winters method also helps us and trying to do the forecast with data which is not fully
complete with respect to a year. You could have only 2 or 3 pieces of data and then you
could do the four and then you could proceed. So, with this we have kind of seen
forecasting models essentially time series forecasting models for level data which i called
as a constant model then for level plus trend. Then, level plus trend plus seasonality, so
we have essentially covered three aspects of it which is level trend and seasonality. Now,
there are other things which also need to be covered in forecasting models it is also a
customary to do what are call cyclic models.
Usually cyclic models have a larger time period than the models that we have seen here
and cyclic models are usually represented using sinusoidal curves. So, we will have
models of the form y is equal to a cos 2 pi t plus b sin 2 pi t where t is the period and a
73
and b are the constants that we try and find out. Once again, we can use methods of least
square to try and get the values of a and b which are for the cyclic models.
The next thing that we will see right now is what are called causal models. In the causal
models, we try and fit something of the form y is equal to a plus b x where y is the
variable that we are actually trying to forecast. Now, it depends on another variable x
which is the independent variable where y is the dependent variable that we are trying to
forecast. The normal example given is for the grain output would depend on the amount
of rainfall, so x is the independent variable which is your rainfall for which we have the
data y is the dependent variable and then we want to forecast y for a given value of x.
So, if we have for example, six periods of data for the food grain production and we will
require 7 periods of data for the rain. So, for example, if we want to fit y equal to a pus b
x if we want to do y 7 which is the forecast for period 7 for the variable y or we can call
it F 7 which is the forecast for period 7. Then, that will be a plus 7, now we need a plus b
into x 7 where x 7 is the data for the independent variable for period 7 unlike in time
series in time series the forecast for period 7. The data we would use only up to period 6
in a causal model we also require data for the independent variable for period seven
methodology wise the fitting a plus b x and a plus b t are one and the same except that x
will replace t in the causal model.
So, methodology wise it is the same you can once again take y minus a minus b x the
74
whole square which represents the error sum of squares, we call it e square is equal to
sigma y minus a minus b x the whole squared. Now, one should also note that actually
while this is the model, we assume that the given data has a small noise or randomness
which we call as epsilon. Then, that creates a small error therefore, y minus a minus b x
is actually not 0, but some value otherwise if we start looking at y equal to a plus b x
here. When I write y minus a minus b x then one would think that it is 0, it is actually not
there is a small error component whose square we actually try to minimize.
Now, we can partially differentiate with respect to a and b which are the unknowns to try
and get the two equations which is sigma y is equal to n a plus b sigma x and sigma x y is
equal to a sigma a sigma x plus b sigma x square. Now, with these two equations we can
try and solve for a and b, now the solution is very similar to the time series. Therefore, I
am not taking a numerical example to explain the same solution once again you will have
columns for x x square x y x square and x y. Then, we can solve for this the how to solve
this has been covered in an earlier lecture in this course the only difference as i said is
having obtained a and b.
Now, if we have data for six periods for x and y, we could use data for six periods of x
and y and finally, we will get something like y equal to a plus b x. Now, for the seventh
period if we know the value of x seven if we know the value of the independent variable,
then we will be able to substitute because using these 2 we can get a and b. So, we can
substitute to get the value of y which will be the forecast for the seventh period or a
subsequent period. So, this is how we address causal models now when it is always
possible to build causal models of different types and need not be a plus b x. For
example, we can build causal models of the form a plus b x 1 plus c x 2 or we can build a
plus b x plus c x square.
We can build y equal to a x power b and so on, now when we build y equal to a plus b x
one plus c x 2 it means there are two independent variables which influence the behavior
of the dependent variable y. So, the methodology will now shift to minimizing sigma y
minus a minus b x 1 minus c x 2 the whole square. So, there are three unknowns a b and
c so partially differentiate with the respect to all the three of them to get three equations.
Here, instead of two and then we can solve those three simultaneous equations to get a b
and c and for the period to which we want to forecast we need the values of the x 1 x 2,
75
they can substitute and try and get the actual value. If we fit a model like this there is
only one independent variable, but then there is a quadratic term here. So, we will try
and minimize y minus a minus b x minus c x square once again there are three things
that we have to calculate a b and c. We will partially differentiate to get three equations,
so we will get a third equation which will says sigma x square y is equal to a sigma x
square plus b sigma x cubed.
So, there will be another term also, there will be a third term also, you can partially
differentiate it to get three equations involving three terms a b and c. So, we can fit a
quadratic and do it if we fit a causal model of the type y equal to a x power b, then what
we need to do is to take logarithm on both sides. So, get log y is equal to log a plus b log
x so this will become log y is equal to log a plus b log x. Now, log y will be replaced by
a variable y this will be another constant a plus log x become another small variable x
now we fit small y is equal to capital a plus b small x. Now, this model is very similar to
this model.
So, we need to now instead of log a we will actually end up getting this a and then from
this a we need to go back to get this small a because just its representative of log a. So, at
some point these things become a little bit of substitution in mathematics, but the basic
methodology is to try and fit something of the you know a plus b x or a plus b x plus c x
square or a plus b x 1 plus c x 2. Then, we try and minimize the error sum of squares and
get the line of best fit assuming linear.
If it is quadratic the quadratic curve of the best fit, so in almost all the causal models the
underlying methodology is to try and convert it to linear or quadratic as the case may be
and try and determine the line of best fit. Then, we proceed as I said the error sum of
squares thing is very similar to what we have seen in the basic time series forecasting
model. Now, we come to the last part of the forecasting which is to try and find out the
goodness of a forecast.
76
Now, how good is the forecast is the question that we have to ask, the reason being
depending on our assumption we actually have more than one model that we can use to
forecast like if our assumption is that the data represents a constant model. The data
represents only level then we have seen several models we have seen simple average. We
have seen moving averages, we have seen exponential smoothing all these if we have
only level if we had l plus t that is level plus trend we have seen two model. One is linear
regression line of best fit and then we saw Holt’s model and if we have l plus t plus s
which means seasonality is also included.
We saw a basic model and we saw Winters model so depending on our assumption that
the data represents only level or l plus t or l plus t plus s we could use more than one
model to try and get the forecast. So, how do we compare solutions of various models
under the same assumption if we believe that it represents only level, then we could use
any one of the three or even a weighted moving average and so on. So, we could test
with 4 or 5 models and then finally, we have to choose one.
Similarly, if the assumption is l plus t, then we could use both these models and other
models that are available in the literature and then chose one the forecast from one of
these and so on. How do we do that, now that is called measuring the goodness of a
forecast, now let me take a simple example to explain the goodness of the forecast, now
let us explain it using a very simple example.
77
So, let us take some data which is 30, 32, 33 and 35, let say this is let us assume we take
data which is 30, 32, 31 and 30. Now, we can assume that this represents only level
because at the moment we do not see a visible trend here. So, let us say we take only
level data and let us say we are going to use the total comes to 123, so simple average is
30.75 is the simple average. Now, there are several measures that we can define for the
goodness of the forecast. So, we are going to do one is called mean absolute deviation
mean squared deviation, mean percentage deviation and so on, so let me show the
calculation for some of these.
Now, let us assume we are using simple average and therefore the forecast for all the
four periods is 30.75 and forecast for a fifth period is also 30.75. So, mean absolute
deviation in this case is let us call it m a d which is mean absolute deviation mean
absolute deviation would be sum of this. So, 30 minus 30.75 plus 32 minus 30.75 plus 31
minus 30.75 plus 30 minus 30.75, now please note that we are finding out the absolute
value and not the actual. So, this is 0.75 plus 1.25 plus 0.25 plus 0.75 which is 1.5 plus
1.5 is 3. For example, we did not do 30 minus 30.75 plus 32 minus 30.75 and so on.
If we had done that value will be 0 because it is a simple average, so we take the mean
absolute deviation, now mean squared deviation is 30 minus 30.75 square plus 32 minus
30.75 square plus 31 minus square plus 30 minus square. If you are coming back to this
this is not mean because this is the total absolute deviation which is 3. So, mean absolute
78
deviation will be one by n number of point y, so 3 by 4 will give us a 0.75 as mean
absolute deviation. Now, mean squared deviation will be this is 0.75 square which is
0.5625. This is 1.25 square, so 1.5625, this is again 31 minus 30.75 is 0.25, so 0.625 plus
0.5625.
So, this would roughly be 0.56 is 1.12, so 2.68, 2.2 say 0.74, so this is roughly totals to
about 2.75 that is the mean squared deviation. So, a mean squared deviation will be this
divided by 4, now mean percentage deviation can be now the demand is 30, the forecast
is 30.75. So, the deviation is 0.75 fact we could even qualify this as mean absolute
percentage deviation. So, 0.75 is the absolute thing here, so that divided by the forecast
which is 0.75 divided by 30.75 into 100 plus 1.25 divided by 30.75 into 100 and so on.
Now, the only difference here is these two will be related because the denominator is the
forecast which is 30.75 which is common.
So, we can even do mean absolute percentage deviation with respect to demand and not
with respect to forecast so that these values can change. Now, these denominators are
become by 30 by 32 plus 2 more terms and so on, so one can define several terms like
these which basically represent the goodness of a forecast. Now, this value mean
absolute deviation has 3 that three is total absolute deviation as I said and mean absolute
deviation is 3 by 4 mean square deviation is 2.75 by 4 and so on. Now. if instead of
simple average if we had used a 2 period moving average which means the forecast
would have become 30.5 instead of 30.75, then we can calculate the mean absolute
deviation.
Now, for 30.5 forecast mean absolute deviation will become 0.5 plus 1.5, 2.5 plus 0.5
which happens to be 3 again. So, 0.5 plus 1.5 is 2, 2.5 plus 0.5 is 3. Now, mean squared
deviation will become for 30.5, it will become 0.25 mean squared deviation will become
0.5 square is 0.25, 1.5 square is 2.25, 31. So, another 0.25 plus 30, another 0.25 say this
will be about 3, so if is a 2.75 divided by 4, say this is 3 divided by 4. Therefore, this will
show a smaller value of mean squared deviation than this forecast, this way if we use
more than one model here we can choose one out of these goodness measures as the
measure that we are going to use and say for the sake of discussion we have chosen mean
squared deviation.
Now, we can evaluate the mean squared deviation for the forecast obtained by each of
79
these and then find out that model which has given as the least value of mean squared
deviation. The example is if we had used simple average the mean squared deviation is
about 2.75 by 4, if we had used a 2 period moving average, then the forecast is 30.5 and
the means square deviation is 3. Therefore, we could say that if we had used mean
squared deviation as the criteria to measure the goodness of a forecast, then the simple
average is a better forecast than the moving average.
There, one has to be a little guarded when we consistently choose mean squared
deviation as the measure of goodness. If we look it at very carefully statistically simple
average is always the measure that will minimize the error sum of squares or mean
squared deviation. The very definition of simple average comes from there, therefore if
we take this simple average, it will always be better than any other forecasting method if
we are using mean squared deviation as the criteria. So, usually mean squared deviation
is not taken as the measure for measuring the goodness of a forecast.
Normally, mean absolute deviation taken as measure for the goodness of a forecast
simply because mean squared deviation is kind of little bit biased towards a arithmetic
measurement, but another way of overcoming that is also this. Here, what we did was we
took all the arithmetic mean and then we said 30.75, we could have done something else
if we have only this data the arithmetic mean is 30, if we have this data, the arithmetic
mean is 31, we have this data the arithmetic mean is 31.
If we have all four data arithmetic mean is 30.75, now we can find the error sum of
squares as 30 minus 30 square, 32 minus 31 square, 31 minus 31 square, 30 minus 30.75
square. If we start doing that, then our basic idea that simple average will always be the
best. So, there are several issues like this that we can look at, but the very basic idea is to
look at one of the measures and use it consistently often the suggested measure is mean
absolute deviation. Now, use that measure and then if you have forecast using more than
one model evaluate it based on this measure and chose that one which gives the
minimum value.
Then, we can get into another question for example; if we have a data which is like this if
we have a data which is like 30, 34, 36 and 39. So, clearly this data shows trend, so we
could use this or this for example, if we had 30, 32, 31, 30 at the moment, we can say
that it does not show trend, but one could say that let me try and fit y equal to a plus b t b
80
is going to be small. If it does not show any trend or b is going to be 0, if it does not
show any trend can I do that the answer is not - in principle we have to check what are
the assumptions and after the assumptions,
We try and get the model you do not want to fit a model and depending on the result of
the model go back and correct our assumption oh well this may exhibit trend. So, that is
not usually the practice the other question is if we have a set of data, now can I apply a
model from here, can I apply a model from here and then chose the better one based on
the goodness the answer is actually no being the reason being the model that we chose
should come after the assumptions that we make regarding the behavior of the data. So,
it is not a good idea to try and see whether moving average is better of whether linear
regression is better after testing both the models and trying to find out the goodness that
is usually not practiced.
What is actually good is to try and make the assumption that based on the assumption
this is this exhibits level. Therefore, I am going to use one of these this exhibits level
plus trend there for I am going to use one of these and so on. So, with this we actually
come to the end of the forecasting part of this course or this lecture series. So, we kind of
started the forecasting topic by saying that there is a need to forecast there is a need to
forecast because manufacturing organizations needs to know how much they have to
produce in the subsequent periods or in the subsequent months.
So, forecast is defined as the estimate of the future demand and then we also said that the
forecast can depend on only time which lead us to time series models it can also depend
on some other independent variables which lead us to causal model. Within the time
series, we saw level plus trend level plus trend plus seasonality and we saw some causal
models and we also saw how to measure the goodness of a forecast. The next question
that comes is well after the forecasting, what do we do? Forecast has given us the
estimate of the demand for a future period. Now, what is a next step that we have to do in
order to use this data that has comes from forecast?
Now, that comes from what is called an aggregate production planning, we will now see
the basic aggregate production planning in this lecture. Then, continue a detailed
discussion in the next lecture. Now having seen forecasting model, let us now look at the
aggregate planning or the production planning problem where we use the results of
81
forecast.
Now, let me explain this problem using the spread sheet, now let us assume that there is
a single product, now there is a only one product that the company makes and that we
have made the forecast of that product which is shown here. So, in this sheet what is
shown as demand is actually the forecast of the demand for the next 12 periods which
represent the 12 months. We also right now assume by looking at this data that this
forecast is for a product which exhibits certain seasonality.
So, there is a demand which is the forecast of the demand which is shown here, now let
us assume that there is a beginning inventory of thousand units at the beginning of
January that does available at the beginning for use. Now, we are also going to assume
here that it actually takes about 10 hours to produce one unit and that there are 65 people
working. Therefore, we also assume that the organization works for 16 hours a day and
so on. So, all these assumptions finally, give us that it is possible to have 104 units which
are produced in a day using what is called a regular time production.
Now, let us spend some time on this spread sheet, now this column shows the 12 months
of the year that we have and right now. Let us look only at the beginning inventory of
January which talks about 1000 units, now demand or forecast of demand is known here
and these ads up to 30,300. Now, we also have what is a called r t day which is the
number of days available in each month for regular time production and these are 22, 18,
82
22, 19 and so on so this data is also known.
We also have what is a called o t day which means the number of days available for
overtime in each month which is given by 4, 4, 5, 4, 5 and so on. Now, if we have 22
days available in January which is shown here, now we can produce 2,288 units in
January, a maximum of 2,288 units which comes from 22 multiplied by 104. This also is
known and can be calculated once these numbers are known, once the r t days column is
known r t capacity can be calculated. We can also calculate the o t capacity column
because in 4 in a maximum of 4 o t days availed for overtime 100 and 4 into 4 which is
400 and 16 units can be produced using overtime. So, this column can also be calculated
once the numbers in this column are known.
Now, we can also calculate the total capacity in each month which is the sum of the
regular time capacity and the overtime capacity as shown here. Also the first term
represents regular time capacity second term represents the overtime capacity. So, this is
also known, now we are going to ask ourselves this question what is going to be the total
production in each month so that we try and meet the demand or forecast of the demand.
Now, as a user let us say that we start filling these values which are the total production
in each month. So, if we fill this number 2704, now the total capacity there is available to
the left is going to guide us in filling these values as total production, so total production
cannot exceed the total capacity in each month.
So, if we fill a 2,704 as the total production for the month of January now this 2,704 has
to be distributed to regular time production and overtime production. There is a regular
time production cost of hundred there is a overtime production cost of 130, since it is
cheaper to produce in regular time out of the 2,704 that we have decided 2,288 will be
produce using regular time. The balance will be produced using overtime, so the moment
the user types 2,704, here it will automatically give 2,288 to regular time which is the
maximum r t capacity and the balance will go to overtime. So, the user actually fills the
production quantities for each of these 12 months.
83
Now, the moment the 2,704 is typed by the user now there is a beginning inventory of
thousand which is available here, there is a total production of 2,704. So, we get 3,704
less 3,000 which is the demand gives us the final inventory of 704, so the ending
inventory for January is calculated this way.
Now, what are the costs there is a regular time production cost there is a overtime
production cost there is an inventory carrying cast there is a shortage cost. Now, we have
already seen that the regular time production cost is 100, so have to produce 2,288 units
84
the cost is going to be 2.288 lakhs, overtime cost is 130 to produce 416 units. It is 416
into 130, which is 0.5408 lakhs ending inventory is 704 units, there is a 20 rupees for the
inventory cost. So, 704 into 20 would give us 0.1408, since ending inventory is positive
there is no shortage, so shortage cost is 0 and the total cost is a sum of these four costs 3
costs in this case because shortage is 0.
It is 2.96 lakhs, now the ending inventory of January which is 704 will become the
beginning inventory of February which is 704 and if we decide to produce 2,288 which
is the actual capacity we have a negative ending inventory. We can calculate the r t cost
the o t cost the inventory cost is 0 because there is negative ending inventory and a
shortage cost is 0.04 because 8 units are short shortage cost is 500 into 8 is 4,000 which
is 0.04 lakhs. So, total cost for February is 2.4528, so like this the user will type the
values for the total production for the 12 months.
We can write this spread sheet in a manner that the total cost is 33.1378 lakhs, so if the
user chooses to produce 2704 in January 2,288 in February and so on, the total cost is
33.17 lakhs. The question is what should be the production quantities in these 12 months
such that the total cost is minimized. Now, this problem is called the aggregate planning
problem where given capacities and given the forecast of demands and given the various
costs in this particular spread sheet.
We have used regular time cost overtime cost, inventory cost and shortage cost, what
should be the production capacities such that the total cost is minimized. Now, this
problem is called the aggregate planning problem and we will see more about this
problem in the next lecture.
85
Prof. G. Srinivasan
Indian Institute of Technology Madras
Lecture - 5
Aggregate Planning, Tabular Method, Linear Programming
In the previous lecture, we introduced the aggregate planning problem. We started by

saying that there is a demand, we said that we assume a single item that is being
produced or a single product that is being made, the demand for 12 months is given as
3000, 3000, 2500, and so on. We also said that there are some regular time days of
production available, which are 22, 18, 22, and so on, in the 12 months, we also said that
the overtime days available are known which are 4, 4, 5, and so on.
Now, once we know the regular time days, we multiplied it by a constant and said that if
there are 22 days total of 2288 units of regular time capacity is known. So, the regular
time capacity column can be calculated by multiplying, these regular time days into a
constant, which happens to be 104. In this case 22 into 104 is 2288, it means that in one
day with the people that are available, we can produce 104 units of the item or the
product.
Similarly, if 4 overtime days are available in the month, 416 is the maximum that can be
produced using the overtime capacity, and this column that as total capacity is a sum of
86
regular time capacity and the overtime capacity. Now, as a decision maker we have to
now say, how many we are going to produce using regular time, how many we are going
to produce using overtime? So, the total production will be the sum of the regular time
production that is planned, and the overtime production that is planned.
So, if we give 2288 here and 416 here, the total automatically becomes 2704, we also
have to give numbers here, which means the planned regular time production quantities
should be less than or equal to the capacity that is available here. And similarly, the
planned overtime production quantities should be less than or equal to the capacity that is
available here, now instead of filling these two individually what the user does is the user
will try and fill a number here, which will be less than or equal to 2704 in this case we
have filled 2704.
Now, the moment we fill 2704, we can assume that the maximum possible 2288 will go
to regular time production, and the balance will go to overtime production. This is
because regular time production is assumed to be less costlier or cheaper than overtime
production, and therefore, the moment we decide on 2704 we will produce the maximum
possible through regular time, and the balance will come to the overtime. So, if we write
2704 here, which means we decide to produce 2704 the ending inventory will be
beginning inventory of 1000, which is assumed to be there plus production which is 2704
less demand, which is 3000, so it becomes 704.
87
Now, we also calculate the costs associated with producing 2704, out of which 2288 are
produced using regular time and 416 through overtime. So, we assume that the regular
time production cost is rupees 100, so we incur 2.288 lakhs as regular time production
cost. We assume that the overtime cost is 130 and to produce 416 units, we incur 0.5408
lakhs, we assume that the inventory is charged at rupees 20 per unit per month. So, to
carry an ending inventory of 704, we multiply it by 20 and incur a cost of 0.1408 lakhs,
there is no shortage cost, because the ending inventory is positive.
In this example, we are charging inventory based on inventory cost is computed based on
the ending inventory value. So, there is no shortage cost and the total cost as a sum of
these four costs which is regular time cost overtime cost inventory cost and shortage cost
and that is 2.9696 lakhs. Now, the ending inventory of January which is 704 becomes the
beginning inventory for February, which is shown here 704 is shown here, now the user
has to make a decision here as to what is the total production.
So, if the user decides to produce all possible 2288 here, out of this 1872 which is the
maximum possible will go to regular time production, and the balance will go to
overtime production, now the ending inventory will be beginning inventory 704 plus
2288 minus 3000, which is minus 8. So, for the month of February regular time cost is
1.872 lakhs, which is 1872 units into 100 rupees, overtime cost is 0.5408 lakhs which is
416 units into 130 rupees.
No, inventory cost, because ending inventory is negative, but shortage cost 8 units are
short ending inventory is minus 8 indicating that 8 units are short shortage cost is at
rupees 500. So, 8 units into 500 will give us 0.04 lakhs and the total is 2.4528 lakhs, now
we come to march with ending inventory of February is minus 8. So, beginning
inventory of March is minus 8, this minus 8 means that 8 units are back ordered and are
met in March, so what we defined as shortage cost of 500 is actually the back order cost,
which is the cost of back ordering per unit per month.
So, like this we proceed the user makes these 12 decisions here for the total production,
and the spreadsheet helps us in computing the cost, so the cost for this production plan is
33.1738 lakhs. So, the spreadsheet is an evaluative tool, where given the total production
choices made by the user the spreadsheet calculates the cost as 33.1738. So, the
important question or the next question is what are these quantities, such that the total
88
cost is minimized, so the aggregate planning problem can be defined as, given that
demand for the single product for 12 months.
Now, we are assuming one product, later we will tell you how we take care of multiple
products, so right now considering one product given the demand for a certain number of
periods, given the regular time capacity and the overtime capacity. The user has to decide
on the total production quantity, which will be split into regular time production quantity,
and overtime production quantity, using the balance equation that beginning inventory
plus production less demand is equal to ending inventory.
And ending inventory in a month becomes beginning inventory of the next month, and
the costs associated with regular time production, overtime production inventory and
shortage, what should be the total production quantities such that this cost is minimized.
Now, this is the aggregate production planning problem, now we try and answer a few
other questions the first question that we need to answer is are we going to consider only
one product or multiple products, because in practice every organization makes multiple
products.
Now, we defer that question, we answer it a little later, next question that we have to
answer is are only regular time production and overtime production are these the only
two means of production or can we have other means of production. For example, can
we consider outsourcing as a means of producing or as a means of meeting the demand?
The other question is do we have only these fore costs, or are there other costs, the third
question that we would like to answer is this in this spreadsheet, when we said that in 22
days we have a regular time production capacity of 2288.
Now, that came from a number 104 which is a number of units that we produce per day,
and if we look at these calculations, these are made out of this we assume that 65 people
are working it takes 10 hours to produce one unit. So, we have 650, 65 people are
working, each person spends 16 hours in a day, so in a day we have 65 into 16, which is
1040, and then we assume that it takes 10 hours to produce a unit therefore, 104 units are
produced in a day. Now, the next question is now the 104 comes out of this 65 which is
the number of people who are working.
Now, can we increase the number of people working in different months, so that the
regular time production capacity and the overtime production capacity can also vary, the
89
question is it going to be 104 per day right through the planning period or can we
increase or decrease that 104 by varying the number of people at our disposal. So, we try
to answer all these questions, using another spreadsheet and by considering additional
costs and also by considering varying manpower.
So, that we try and show in the other spreadsheet, which is like this same problem, we
consider here the same demands of 3000, 3000, 2500, 1500 etcetera. You can compare
this 3000 same demands, we consider 30300 is the total demand same 30300, regular
time days are the same overtime days are the same. Now, in this model we are trying to
include the workforce that is ((Refer Time: 13:00)), we explained this using another
spreadsheet, where we consider additional costs as well as varying manpower.
Now in this spreadsheet, we use the same values of the demand and these demand adds
up to 30300, we use the same values of regular time production days as in the previous
spreadsheet as well as the same values for overtime production days, as used in the
earlier spreadsheet. The change that we bring in this spreadsheet is that, we are now
going to vary the workforce. So, we are assuming that say, this represents January of a
particular year where the demand is 3000, we assume that in the previous December we
had 58 people who were working.
Now, in this spreadsheet calculation, we are going to show that this 58 is now increased
to 60, so we introduce another column which talks about hiring people. So, from 58 to 60
90
we have hired 2 people for this period, so the number of people hired is 2 we have not
laid off any one because the number has increased. So, if there is an increase in
workforce it is hiring, if there is a decrease in workforce it is a layoff, so we do not have
layoff here layoff is 0.
Now, with 60 people working for 22 days, the regular time production is 2100 regular
time capacity is 2112, now this comes from we have 60 people each works for 16 hours
in a day, so that is 960 divided by 10 is 96. So, each day we can make 96 units, so in 22
days we can make 2112 units, similarly in 4 overtime days multiplied by 96 we have
overtime capacity of 384. So, now, the regular time capacity and the overtime capacity
not only depend on the number of days, it also depends on the number of people who are
working in that month.
We also assume that once we define 60, these 60 people are going to work on all the 22
days of this month. So, we are now defining the regular time capacities and the overtime
capacities, now the user has to give a decision on the total production. So, once the user
decides say 2500 here, the first 2112 or 2112 out of 2500 will be made using regular time
production, because that is the maximum capacity, and the balance is made from
overtime production which is your 384 from overtime production.
Now, we have decided to produce 2500 in this month, our regular time capacity is 2112
our overtime capacity is 384 together the capacity is 2496, but we have decided to
produce 2500. Now, out of the RT and OT capacities we can get only 2496. So, the
balance 4 will be treated as being out sourced, so if the user enters 2500, first we will
check what is this RT capacity? So, that much is given to RT production, and then the
balance is taken then the OT capacity is seen, and then the maximum possible is given to
the OT production.
And if there is still something to be given that is given to the outsource, the reason and
the motivation is that regular time production is less costlier than overtime production,
which is less costlier than outsourcing. Therefore, maximum is given to RT production
then to OT production and then to outsourcing, now the ending inventory is calculated in
the same manner as beginning inventory. In this case beginning inventory is 1000,
production is 2500, so 3500 less demand of 3000, which comes from here and the ending
inventory is 500.
91
Now, we have various costs, we have actually 8 costs in this spreadsheet regular time
cost, overtime cost, inventory cost, shortage cost, hiring cost, layoff cost, outsourcing
cost, and under utilization cost. So, let us see all these 8 costs, now regular time cost is
kept at the same value of 100, so 2112 units will give us a regular time cost of 2.112
lakhs, overtime cost is kept at 130, so 384 units will give us an overtime cost of 0.4992
lakhs.
There is an ending inventory of 500 the inventory cost is 20, so we get 0.1 lakh, the
ending inventory is positive, therefore there is no shortage. Shortage cost is 0. In this
period we increase the workforce from 58 to 60, therefore we hired workforce, cost of
hiring is given as cost of hiring is 500 we hired 2 people, so the hiring cost is 500 into 2
people, which is given by 0.01. There is no layoff means the workforce reduces, so there
is no layoff, there is an outsourcing cost. We have out sourced 4 units the cost
outsourcing is taken as 150. So, 150 into 4 is 600, so 600 gives us 0.006. The 8th cost
which we have to describe is cost of under utilization.
Now, in this particular case the regular time capacity is 2112, and we have used up all
the 2112 of regular time capacity, if we do not use the entire regular time capacity, and if
some regular time capacity is unused then that becomes an under utilization cost, in this
case the under utilization cost is 0. Now, if you go back to few other months, where I can
show some of these costs, and if we go to month 2, February we begin with 500 ending
inventory of January becomes beginning inventory of February with 500.
Now, we define workforce to be 56 for the month of February, so there is a layoff there
is a no hiring, so February has a layoff cost that you can see here, there is a layoff cost.
Now, you also see if you look at the ending inventory column, you do not have
negatives, so at no point we have shortage all shortage costs are 0. If we look at the
month of say we look at this month, we look at the month of April, we now realize the
regular time capacity is 1824, for this month overtime capacity is 384, but we have
decided to produce only 1500.
Now, 1500 is smaller than 1824, so all the 1500 goes to regular time production, no
overtime production, no outsourcing we have not used up the entire 1824, so 324 is
available as unutilized or under utilization. So, the cost of under utilization is 324 into 2
which is 648, so the cost of under utilization is 0.0648, so there are 8 costs that we
92
calculate and then the total cost of the production plan is 32.7417 using this production
plan which has 8 costs. And more importantly, which has more options for the user, so in
this case what should be the quantities, such that this total cost is minimized.
Now, in this spreadsheet the user actually makes 2 decisions, there is a decision on the
workforce, and there is a decision on the total production, in the earlier spreadsheet the
user made only one decision, which is on the total production. So, the optimization
problem relevant to aggregate planning is given a certain decisions that the user makes
with respect to workforce and production, whichever is relevant production. Of course, is
very necessary there are situations, where we work with constant workforce, there are
situations where we work with variable workforce.
So, depending on whether it is constant or variable the workforce decisions are made. So,
the user makes production and workforce decisions, a maximum of 8 costs are
considered here, we may consider all of them we may consider a sub set of them. Then
the aggregate planning problem is what are the production and workforce decisions, such
that the total cost is minimized. We now try and address the optimization problem by
using an approach based on linear programming, so let me explain that next.
So, we try and look at what is called an LP approach to aggregate planning, so we first
assume that demand for period t is known and given, so there is a demand for period t.
The various costs are given, and the various decisions are given let us say RT is the
93
regular time production used in period t, let us call OT as the overtime production used
in period t. Let us define I t as inventory ending inventory at the end of period t, let S t be
the shortage at the end of period t, let us call some outsource of t as the number of units
out sourced.
And let U t be the under utilization in period t, let H t number of people hired in period t,
let W t be the number of people working in period t, let L t be the number of people laid
off in period t. Now, these are all where this is known D t is known, all these are
variables that we have to find out, so there are 1 2 3 4 5 6 7 8 9 variables effectively, but
there is a relationship among these variables.
Now, we also have costs associated with each one of them, which we denote by using
the lower case numbers, so let small r be the regular time production cost, if this regular
time production cost is going to be different in different periods, then we can add a small
subscript RT, which is a regular time production cost in period t. If it is going to be the
same for all the periods, then it will be r, so to generalize it we begin by saying that let
small RT be the cost of regular time production in period t.
Let small OT is the cost of overtime production in period t, let i t be the inventory cost
for period t, let s t be the shortage cost for period t. Let small OUT be the cost of
outsourcing in period t, and let small u t be the cost of under utilization in period t. Let
small h t be the cost of hiring people in period t, let small l t be the cost of laying off
people in period t. Now, if we define these costs right now i am not defining small w, but
I am defining the rest of the costs the objective function or the minimization function
will be to minimize.
94
Summed over t for t periods, if r is the r t into RT, this is the variable, this is the constant,
and this constant as assumed to be different for different periods, otherwise we can
simply put an r if it is the same for all the periods. So, this represents the regular time
production cost plus o T into O T, this represents the overtime production cost, this is the
variable this is the unit cost, plus i t into I t represents the cost of holding the inventory.
This is the variable this is the cost plus s t into S t, which is the cost of shortage, if it is
there plus h t into H t, h t is the number of people, who are hired this is the unit cost of
hiring. So, h t into H t plus l t into L t cost of layoff plus small out into outsource of t,
this is the variable, this is the cost plus cost of under utilization u t into U t. So, there are
8 costs that I have put in the linear programming formulation, these are the 8 costs.
Now, we need the constraints for any typical period I t minus 1 is the inventory at the
end of that period, which is the beginning inventory for a particular period plus the
productions, the productions come out of three ways regular time production plus RT
plus OT plus OUT of t. So, production comes in 3 forms regular time production,
overtime production, and outsourced production, so these three represent the production
beginning inventory plus production less demand D t should be equal to the ending
inventory, so should be equal to I t.
So, this is the typical balance equation, but the beginning inventory can also be negative,
if there is a shortage at the end of the earlier period, so one way is to say that now these
95
quantities are all greater than or equal to 0, these cannot be negative, but if we start
defining I t as the inventory at the end of the previous period. And if there are shortage
then I t and I t minus 1 can take negative values, which means in the linear programming
formulation I t has to be defined, as a unrestricted variable which can take positive value
or a negative value.
Now, we also know from linear programming, that every unrestricted variable can be
represented as a difference of two variables both are greater than or equal to 0, plus we
also have different costs for inventory and shortage. Therefore, it is to our advantage if
we, now start representing this as I t minus 1 minus S t minus 1 plus RT plus OT plus
OUT of t minus D t is equal to I t minus S t. So, if the beginning inventory is positive, we
will have a positive value of our I t minus 1 S t minus 1 will be 0.
If the beginning inventory were a shortage, then it has a negative value, so this will be 0,
this will be a positive value, so that minus S t minus 1 is negative. Similarly, if the
ending inventory is positive or negative I t and S t will take the corresponding values, so
this is our first equation which balances the inventories demand and production, so this is
one equation. So, we will have if there are 12 periods we will have 12 such equations,
and more importantly we will have thirteen variables for I t and S t, because we need to
initialize the I t and S t.
The first value that we assumed as thousand in the spreadsheet is an initial value given to
I naught, and S 0 was 0 in that case, so we need some initial conditions, where I t minus
1 and S t minus 1, second one is our capacity. So, regular time capacity is available, so
whatever is given as regular time production quantity, should be less than or equal to the
regular time capacity, so this is clearly written as what is the regular time capacity. From
the spreadsheet we said that if there are W t people working, then each person is capable
of producing a certain quantity per day.
So, we let us call that as some k 1 into W t if w t people are working then each person
can make k 1 per day, so this will be W t into k 1 into the number of days, which is a
constant, if k 1 is the quantity that each person produces per day, which was our 1.6 in
the previous spreadsheet example, now that has to be multiplied by the number of days.
So, we will start defining RT is the regular time production in period t, is less than or
96
equal to the number of people working into units produced per day, into another constant
which is some k 2, where k 2 is known it is RT days.
So, I can either use it as k 2 or I can use it as RT days of t,, number of regular time days
available in period t, or I can multiply this is a constant, this is a constant, so I can simply
say RT is less than or equal to some constant times W t. Now, if we get into a situation,
where RT is actually less than what is the capacity, then we have under utilization, which
we saw in the second spreadsheet. And there is an under utilization costs which is also
there, so this is rewritten as RT plus U t is equal to constant times W t, so this represents
the regular time capacity constraint.
Similarly, there is a overtime capacity constraint, so OT is less than or equal to the same
k 1 into W t into OT days of t. This W t is the number of people, who are available k 1 is
the number of units this person is producing per day, this represents the number of OT
days that are available. So, OT is less than or equal, here we do not have a under
utilization cost for OT, and this formulation we are using under utilization cost only for
RT, and we are not using it for OT.
Then we have another constraint, which is W t minus 1 plus H t minus L t is equal to W

t, w t minus 1 is the number of people who have worked in period t minus 1, who are
available at the beginning of period t it is like your inventory, you either hire or you
layoff, so plus H t minus L t is equal to W t. Now, all variables greater than or equal to 0,
now this is the linear programming formulation of this problem, now there is a hiring
cost associated with this there is a layoff cost associated with this.
Since, both this h t and l t are greater than or equal to 0, both H t, and L t will not appear
in the solution, only one of them will appear in the solution. For example, if net hiring
people is 2, you will not say that I have hired 3 and I have laid off 1, and I have a net of
2 people hired that would increase, these 3 costs you would rather say hire two and do
not layoff anyone, so this is the formulation. Now, if there are t periods, then there are as
many constraints as the number of periods, as many constraints as the number of periods,
once again number of periods, number of periods.
We also need some initial condition on W 0, also observe that in this formulation, we are
not explicitly adding the payroll cost into the objective function. We are not doing that
we assume somewhere that the payroll cost is getting reflected in the regular time cost,
97
which is includes payroll perhaps even in the under utilization cost, which includes part
of the payroll costs. That is the reason we do not have a small W t which is the cost of
actually engaging these people in the month
We also have to understand that H t and L t are not the payroll costs, but these are the
costs, these are also not the payroll cost of the additional people hired, H t and L t, which
are called this H t and L t, which are the cost of hiring and cost of layoff. H t essentially
concerns the cost of conducting interviews to hire people, the cost to train the people, it
does not include the payroll of the additionally hired people, that is taken care in the W t,
and in this case in RT and U t.
Similarly, L t is the cost of laying off, where some kind of an additional compensation is
given to the person, who is laid off, so that is your L t. So, this formulation is the
equivalent optimization or linear programming formulation, for the problem the second
spreadsheet problem that we saw earlier. The spreadsheet was an evaluative solution,
where the user made those decisions on in the spreadsheet the user made decisions on
total production and workforce.
Here what happens the decision variables are many of them, but we also know that in an
evaluative model when you make a decision for total production? You are actually
making decisions for RT OT as well as outsource of t, when we made one production
decision in the second spreadsheet, when we made a decision for W t there, we
automatically made decisions for H t L t and W t. And the moment we made this
decision for RT, we also made the decision for U t which is the under utilization, so this
is exactly the formulation for the problem, that we had there.
Now, let us go back to that spreadsheet, and see what is the optimal solution for that
particular example, when we solve that using this particular linear programming
formulation. It is a linear formulation because all variables are greater than or equal to 0,
we have taken care of that by this I t minus 1 minus S t minus 1, as well as taking care by
the H t minus L t. So, if W t is bigger than W t minus 1 H t will take a value, if W t is
smaller than W t minus 1 L t will take a value, both attract costs which are greater than or
equal to 0, so let us go back and check in the spreadsheet as to what these solutions are.
98
Now, in this linear programming formulation, we have modeled all that we saw in the
second spreadsheet, where we considered 8 different costs, which included cost of
inventory regular time, overtime, inventory, shortage, hiring, laying off, outsourcing and
under utilization. Now, if we where to model the first spreadsheet, where we considered
only 4 costs, where we considered regular time cost, overtime cost, inventory cost, and
shortage cost, which means we did not have this particular equation, where we varied the
workforce.
Here there is only one W, which is a constant which was kept at 65 for the other in this
spreadsheet, so we do not have these 4 costs and we do not have this OUT of t, there is
no outsourcing. So, this linear programming formulation can also be modified to take
care of the assumptions in the first spreadsheet, what I also wish to convey here is even
though there are 8 costs that are there, it is not absolutely necessary all the time to
consider all the 8 costs.
Sometimes, if we work with the constant workforce model, then we are going to consider
only 4 out of the 8 cost, if we leave out the outscoring and under utilization, so we are going
to look at only these 4 of the 8 costs. So, what we do now is we go back to the first
spreadsheet, we try and see the solution given in the first spreadsheet, and then we also
provide the solution in the optimization problem, where we are modifying this L P
formulation. To meet the assumptions of the first spreadsheet, where we are considering only
4 costs, we consider a constant workforce, and we do not consider the outsourcing.
99
So, we are going to consider only regular time, overtime, inventory, and shortage, now in
this spreadsheet the total cost that we see is 33.17 lakhs.
33.17 lakhs and the total production quantities that are given there are 2704, 2288, 2808,
2390, 2912, 2496, 2808, 2808, 2288, 2704, 2400, 694, these were the 12 production
quantities, which you can see here, which are shown in this spreadsheet. Now, the cost is
33.1783 lakhs, now if we solve the optimization problem for this, the optimum solution
is R 1 is 2288, R 2 is 1872, R 3 is 2288 and so on.
100
Now, the overtime values are 416 416 520 and so on, and the total cost is 33.01 lakhs,
now the evaluative model using the spreadsheet gave us a solution with 33.17, the
optimization gave us a solution with 33.01 lakhs, total production here was 2704, here
total production is 2704. This is 2288, this is 2288, this is 2808, this is also 2808, but
there is a change somewhere else in the optimal solution, so 33.17 lakhs from the
evaluative model came down to 33.01 lakhs.
Now, this is the advantage of optimization, if we try to and get this solution from the
spreadsheet, the user has to play around with these values, and for every change you can
see that this value keeps changing. For example if we had used instead of 2704, if I had
used 2600, then you see that 33.17 becomes 34.94, so the user can now play around with
these total production values and such that the value actually comes down.
Now, that is going to take enormous amount of time, if the user has to do it in a
spreadsheet the optimization gives us the best value, and when you use the spreadsheet
model or what is called as a tabular approach. You would not know, which is the
optimum value, whereas here you know what is the optimum value.
The other important thing is that, as I said this model is a very generic model, which
captures 8 different costs, sometimes you use all 8 of them, sometimes you do not use all
8 of them, you may use fewer as we did we used only 4 out of the 8. Sometimes, the
organization will say that I will use variable workforce, but I do not want to use
101
outsourcing, which means you will not have the outsourcing cost here, but you will have
the variable workforce, which means you will have the H t as well as L t.
So, depending on the specific requirements in the various organizations, we can modify
this linear programming model suitably, and then solve to get to the optimum solution.
We also made an assumption here, we did not explicitly use W t into this, I said the W t
people are paid and that gets absorbed in this RT cost, as well as in some of the U t cost,
when they are not going to be utilized. It could also get absorbed in the OT cost, if they
are used for the overtime, another way to do it is to separate it and then bring this W t
into the objective function, by saying that there is a payroll cost for W t people W into W
t.
And, if we do that then this RT, this OT and this U t will not consider the payroll
component, so this RT will only be the cost of material, the cost of consumable, the cost
of other power and few other things, which will go into R t. Normally, RT and OT
particularly the difference in these 2 costs comes from the additional money that is paid
to work overtime, so that the additional amount also has to be absorbed here in this O t.
So, the formulation can also be modified by bring the payroll cost explicitly into the
objective function, but linear programming provides just with a very generic framework
to solve this. Now, in addition organizations may have some other kind of constraints,
see here we do not have any restriction on the values, that I t and I t minus 1 can take. If
it is cheaper and economical we do not mind building a lot of inventory through this
model, because we are not restricting the value of I t though indirectly we are restricting
it because inventory has to be built only through production.
So, if I t has to go up then it means that this RT and OT and OUT of t have to go up, but
if we look at the costs very carefully, the costs of holding inventory is actually little less
compared to every other costs that we have marked. So, there can be a tendency to
increase I t, which times organizations would like to avoid by putting a restriction, that I t
should be less than or equal to some I star, I do not want to hold more than a certain
amount of inventory in this.
So, that is another way of handling, sometimes shortages should not be there, so you will
leave out the s t variable completely and say that I will have, I will meet every months
demand then and there, I will not meet the shortage. Now, we try and address the next
102
question, what we do if we have multiple products the other thing at we also see is this,
now some of these are in units, now we have to define these variables carefully.
Now, we defined some of them as units, number of units produced in regular time,
number of units produced in overtime, and so on, if we start defining them, as number of
units produced in regular time, overtime, number of units of inventory, number of units
of shortage, number of units outsourced. But, H t and L t are number of people, who are
hired and laid off, U t should ideally be the number of hours that are not utilized. So, we
need to bring these three things the number of units produced, the number of hours that
are unutilized, and number of people that are there plus in addition, we if we make
multiple products then each product would require different times to produce.
Therefore, the capacity also has to be adjusted accordingly, so in aggregate planning, we

do not and try and define these as number of units produced or number of people hired,
but all the variables are now represented in a single unit, which is called man hours. So,
this formulation will be let RT be the number of man hours of regular time production,
number of man hours of overtime production, number of man hours of inventory,
number of man hours of shortage, number of man hours outsourced, number of man
hours underutilized, so it has a common unit, which is called man hours.
Therefore, some of these k 1 etcetera, will have to be redefined, which is not very
difficult to do and the moment, we start using man hours we also know that we can
handle multiple products. The word aggregate comes from the fact that we are
aggregating all these products, looking at an equivalent product. And then we try and
formulate a problem such that we compute the man hours of production, and many hours
of workforce, that are going to be implied, and these are to be found at minimum cost
now we look at further models of aggregating planning in the next lecture.
103
Prof. G. Srinivasan
Lecture - 6
Aggregate Planning, Transportation Model
In this lecture we will see the Transportation model for Aggregate Planning. In the
previous lecture, we saw the linear programming formulation for aggregate planning.
We considered demand data for 12 months or 12 periods, the linear programming

formulation had several sets of constraints. We had one set of constraints where, we
added the production to the beginning inventory, and then we subtracted the demand to
get the ending inventory. We also had another set of constraints, relating the regular time
production under utilization and capacity; and we had a third set of constraints involving
the workforce.
Now, we are going to describe a transportation model, which does almost everything that
the linear programming model does, there are a few differences which also we will see,
as we move along. We take a much smaller example, to illustrate the use of the
transportation model, and towards the end we also try and show how the transportation
model would behave, if we considered the same example involving 12 periods.
104
Now, for the sake of illustration, let us assume that we are considering 4 periods as
against 12 periods, each period say is a month. And for the sake of illustration let us
assume that, a regular time capacity and overtime capacity are the same and are known.
So, we will have regular time capacity, let us say is 2000 and overtime capacity is equal
to 500; in the last lecture we saw that it is customary to denote all these as man hours
instead of units.
So, each month we assume has a RT capacity of 2000 and OT capacity of 500, so totally
a capacity of 10000 man hours are available in the entire production period. We have 4
periods, so let us have the 4 demand values, demand is also represented in terms of man
hours. So, let us say the demand values are, 3000, 3500, 1500 and let us say it is 1500, so
let us now try and draw the transportation matrix for this.
So, the transportation matrix will look like this ((Refer Time: 03:27)), now each month
there is a supply of 2000, which comes from regular time capacity and there is a supply
of 500 that comes from overtime capacity. So, there are 4 periods, so each period has a
regular time capacity and an overtime capacity, which are the supplies. So, we will say
that 1st period 2000 comes from RT, 500 comes from overtime capacity, again 2000,
500, again 2000,500, again 2000 and 500.
As I mentioned in this particular example we are assuming that, the regular time capacity
is 2000 in all the months and overtime capacity is 500 in all the months. In practice it can
105
be different for each period, as we saw in the tabular method; the overtime capacity and
the regular time capacity depend upon the number of days that are available. And
therefore, they can be different for different periods, now we have to meet the demand of
these 4 months with these capacities.
So, we start drawing the 4 demand columns, so the 1st period demand is 3000, 2nd
period demand is 3500, 3rd period demand is 1500, and the 4th period demand is 1500.
So, the 4 period demands are written as 3000, 3500, 1500 and 1500 respectively, now we
have created the basic structure for a transportation graph. Now, we all know that we
always solve balanced transportation problems, therefore we have to find out the total
supply and the total demand.
Total supply is 2500 into 4 which is 10000, total demand is 6500 plus 3000 is 9500, so it
is customary in transportation problems to balance it, and in the event where the total
supply exceeds the total demand, we create a dummy demand column, so that the total
supply becomes equal to the total demand. In situations where the total demand is higher,
we would create a dummy supplier, so in this case the total supply is higher than the total
demand, so we create another dummy which has a dummy demand of 500, so that we
have a balanced transportation problem.
Now, we have assumed here that there is no initial inventory, so if there were an initial
inventory, then we would have one more row and the initial inventory would act as a
supply. Right now, we are assuming that at the end of the 4th period, we are not going to
have a final inventory, so if there were a final inventory, then that final inventory would
appear as another column in the transportation table. So, right now we are not assuming
initial inventory and final inventory, otherwise we will have one more row and one more
column in the transportation table, the dummy comes; so that, we balance the total
supply and the total demand.
Now, we start writing the various costs, now if we are going to use the 1st month regular
time capacity to meet the demand of the 1st month, then we assume that there is a regular
time production cost which is given by R. If the 1st month overtime capacity is used to
meet the 1st months demand, we are going to assume that it involves a overtime
production cost of O, now I call this as O and to distinguish it from 0, let us say we call
this as O.
106
If I meet the 1st months regular time production to meet the 2nd months demand, it
means I am producing it in the 1st month, I am holding it in inventory for 1 month, so
that I can meet the 2nd months demand. So, the cost becomes r plus i where I add 1
period inventory cost into this, similarly if I use this capacity to meet the 2nd months
demand, it becomes O plus i, where O is the overtime production cost per man hour,
because both demands and supplies are represented as man hours.
So, let us assume the cost are suitably represented as, so much per man hour, similarly if
I use the 1st month’s regular time to meet the 3rd month’s demand it will become r plus
2 i and this will become r plus 3 i. Similarly, this will become O plus 2 i become O plus
3 i, now if I use this supply of 2000 which is regular time that is available in the 2nd
period to meet the demand of the 1st period, which means I am producing in the 2nd
period and I am going to meet the demand of the 1st period.
Now, that I can do only if I allow back order, it means if I have to use this to meet this,
then I am not meeting the demand entirely in the 1st period, I am back ordering. And
from the 2d month’s regular time I am trying to meet the demand of the 1st month. Now,
for the present, let us assume that we are not going to allow back ordering and therefore,
it is a minimisation transportation problem. So, it is customary to put a big M, where M
is large and positive, so that we do not have this allocation and this is not allowed.
Similarly, overtime capacity of the 2nd month to meet the 1st month’s demand is not
allowed, so automatically all these will become big M, because we are not allowing back
ordering in our 1st assumption. Regular time capacity of 2nd month to meet 2nd month’s
demand will be r, this will be r plus i, this will be r plus 2 i, the r plus 2 i comes, because
I am using the 2nd month’s regular time capacity to meet the 4th month’s demand, which
means I am producing in the 2nd month. And I am keeping it in inventory for 2 periods,
therefore this becomes r plus 2 i.
Similarly this will become O, this will become O plus i, this will become O plus 2 i, now
from the 3rd month’s capacity, I cannot meet the 2nd month’s demand, because I am
assuming right now that back ordering is not allowed, therefore all these will become big
M. Now, I am using the 3rd month’s capacity to meet the 3rd month’s demand, so the
cost will be r here, r plus i here, O here, O plus i here and I cannot use the 3rd month’s
107
capacity I cannot use the 4th month’s capacity to meet the 3rd month’s demand, so this
will be big M, this will also be big M.
So, 4th month I will have r here, I will have an O here please note that these are O, so
that I can distinguish them from 0. Now, I will need to fill the cost associated with this,
in any transportation problem it is customary to fill the dummy row with cost equal to 0.
So, we put cost equal to 0 for all of these, so I am going to write this as 0 and please note
that this is o, while this is 0 and they are different. So, just to distinguish that this is 0, I
am just going to add a like a phi, so that we know that this cost is 0 and this is not O.
So, now we have completed the cost for this transportation problem, and we now need to
solve this transportation problem. Now, in earlier lectures in different courses, you would
have been introduced to how to solve the transportation problem, there are many ways of
solving it. Usually transportation problems are solved in two stages, in the first stage we try
and get a good basic feasible solution; either through the North West corner rule, or through
the minimum cost method, or through the Vogel's approximation method.
We have also seen in a different course earlier, that the Vogel's approximation method is
used usually, because it gives a better objective function value compared to the North
West corner rule. And then the second stage, we use the basic feasible solution that is
obtained to try and get to the optimal solution. Now, let us try and apply the North West
corner rule to this, North West corner rule is the simplest of the methods to get a basic
feasible solution to the transportation problem. So, in the North West corner rule, we
start with this 2000 ((Refer Time: 15:57)) and this 3000, so we allocate 2000 here, the
balance 1000 is here, so there is a 500; so we would allocate a 500 here, then there is a
remaining 500 which is there, which has to be met.
So, in this particular example, ordinarily we would put a 500 here for the North West
corner, so ordinarily we would have met this through this, but since we have a big M and
we cannot do any of these. So, we will not be able to proceed with the North West corner
rule, but now for the sake of illustration let me assume, that I am going to start with a
500 here and then, I would possibly try and improve it or make it better. In this particular
example I would not be able to, because I am not allowing back ordering, the 1st month’s
demand itself is 3000.
108
So, the 1st month’s total production available is 2500, so in this case I will not have a
solution, because I am not allowing for back ordering, so for the sake of illustration and
for the sake of getting a very good solution.
So, for the moment let me say that this demand is 2000, and let us say this demand is
3000, so that this demand is 2000 and this demand is 3000, so now what happens is your
total supply is 10000, total demand is 3 plus 2 5 plus 3 8000, so the dummy becomes
2000, so now we have balanced it. So, now if we apply the North West corner rule, we
have 2000 here, we have 2000 here, so we will use up all the 2000 here, so this will go
and this will also.
In fact, let me also make a very minor change in the data, so that we can show few more
things as we move along, so I am going to make this as 1500 instead of, this also I am
going to make 1500, so that the demands are 6000, so this will become 4000. So, I am
going to make the demands as all the 4 demands as 1500, now what happens is when I
start solving using the North West corner rule. Now, before I start solving from the
beginning, let me explain to you that the four capacities are 2500 in each month, with
2000 as regular time capacity and 500 as overtime capacity.
The 4 demands are now 1500 for me, so what I do is I apply the North West corner rule,
so I first get 1500 out of it, so this column is completed another 500 remains here, there
is a 1500 here, so I will put 500 out of this, so this is completely met. So, this will
109
become 1000, 1000 is demanded 500 is available, so I put a 500 here this is also
completely met, so I put another 500 here, so that this is met completely. Then I put a
1500 here, so this is met completely this is also met completely.
Then I go back and do a 500 and then I have a balance of 1000, so this is met, so there is
a 1000 that comes out of this ((Refer Time: 19:15)), there is a 500 that comes out of this,
there is a 2000 that comes out of this and there is a 500 that comes out of this, so it is
balanced. So, now we have got the first solution using the North West corner rule, now
let us try and see what kind of solution it is. Now, in transportation problem, if we have
M supply points and N demand points, every basic feasible solution should have M plus
N minus 1 allocations or less, but these allocations have to be independent.
If we have less than M plus N minus 1 allocations, then it represents a degenerate basic
feasible solution, so in this particular example, we have 1 to 8 supply points and 5
demand points. So, we have 13, 8 plus 5 is 13, M plus N minus 1 is 12, so the basic
feasible solution should have 12 allocations, now let us see 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
allocations we have. It is still a basic feasible solution with 11 allocations which is less
than M plus N minus 1, so it is a degenerate basic feasible solution.
And the degeneracy comes because somewhere when we made this particular allocation,
I think when we made this particular allocation, the balance of 1500 was met here and
the demand of 1500 was met here. This particular allocation did not leave behind
anything either in this row or in this column, therefore we started here with this 500 for
example, when we made this allocation of 1500, we had a balance of 500 here which
came here. So, only one row or one column was entirely filled or met by the allocation,
but at this point both the row and column were met entirely.
And therefore, both these were met and we started from here, so we had 1 allocation less,
and we also know that if we have a degenerate basic feasible solution. In the next method
which is to optimise it, in the UV method or MODI method, Modified Distribution
method, it is customary to put an epsilon to denote the place where there could be a
degenerate allocation, and then we proceed.
Now, let us assume that this is the basic feasible solution, that we have using the North
West corner rule, we have already learnt elsewhere that, if we had gone for a minimum
cost method or a Vogel's approximation method, we may have got a better basic feasible
110
solution. Right now I have not defined what the values of r and i etcetera are, so that we
will not be able to do the minimum cost method or the Vogel's method here, unless we
know the values of r, i, O and so on.
Whereas, we can do the North West corner, because North West corner depends only on
the positions and the values that we actually know, so we are fine doing a North West
corner. Now, we have to improve the solution to try and get to the optimal solution, so in
a transportation problem the improvement is done using two methods, one is called the
stepping stone method and the other is called the modified distribution method or UV
method or MODI method and so on.
Now, let us try and look at the stepping stone method, now what is the stepping stone
method do, stepping stone method believes that if this is not the optimal solution, then
there is at least 1 allocation which is currently blank, which means which is currently non
basic, which will appear in the optimal solution. So, stepping stone method will take
each one of these blank positions, and it will try and enter a plus 1 or a theta to see
whether there is a gain in the cost.
For example, ((Refer Time: 23:40)) this is a position which does not have any allocation
right now, so if this were in the optimal solution, the belief in the optimal basis, the
motivation or the belief in stepping stone is that; if I put a plus 1 here, then it will be
advantageous, then it can enter the basis. So, if I try and put a plus 1 here, now to balance
it I will have to put a minus 1 here, I have to put a plus 1 here and I have to put a minus 1
here. Now, the net gain will be plus O minus O plus i plus r plus i minus r.
111
So, let me write the net gain, so plus O that comes out of this by putting a plus 1 I am
incurring a cost of O, by taking away 1 I gain a O plus i, so minus O plus i by adding a 1
here, I am adding an r plus i; and then by taking away 1 I am taking away an r, so minus
r. So, the net cost this is O ,this is not zero, so O minus O minus i plus i plus r minus r is
0, so we observe here that if we try and put a plus 1 here, now we realise that there is not
there is not a gain the cost remains the same, so we try and do this.
So, stepping stone will try and see all the places where we do not have a variable we will
try and put a plus 1, now let us try putting a plus 1 here for example, so if we put a plus 1
here then we have to balance it with the minus 1 with the plus 1 and a minus 1. So, the
net will be r plus 2 i minus of r plus i plus r minus of r plus i, which will also be 0, so it is
not going to be an advantage doing this. But, let us try and put a plus 1 here, now what
happens I add a 0, I subtract a 0, so this will become a minus 1 here, become a plus 1
here, it will become a minus 1 here.
Now, because of degeneracy I would not be able to proceed, so I have to put an epsilon
here, so this will become plus 1 here, this will become minus 1 here and then it will
become ((Refer Time: 26:24)) this is not the thing, so plus 1 here minus 1 here plus 1
here minus 1 here, so that it gets balanced. Let me repeat it, if I put a 1 here, so to
balance it I have to put a minus 1 here, I have to put plus 1 to balance here, then I am not
able to move, because of degeneracy, so I add an epsilon here.
112
So, to balance it I would put a minus 1 in this epsilon, then I would put a plus 1 and then
I would put a minus 1, now the net will be plus 0, here I have to be very careful, because
some values involves 0 some values involves O, so this is plus 0. So, I am going to write
it as plus 0 minus 0 plus r plus i minus r plus 2 i plus r minus r plus i, so this is r plus i r
minus r plus r minus r is 0 i minus 2 i minus i is minus 3 i. So, this gives us a non zero
value.
Let me repeat it 0 and 0 gets cancelled, ((Refer Time: 27:57)) this r will cancel with this
r, this r will cancel with this r, so plus i minus 2 i is minus i minus i minus i is minus 2 i.
So, there is a gain when we do this, so what we try to do is we try to put the maximum
possible here, so it will become plus theta minus theta, plus theta minus theta, plus theta
minus theta, again we realise that theta will become epsilon, because of degeneracy. So,
this becomes theta minus epsilon, so the epsilon will shift from here to here and so on.
We know how to proceed in a transportation problem by doing this, the basic idea is we
try and solve the transportation problem. Now, what have we learnt by this, if we apply
the North West corner rule to this, we do not know the values of r i yet and we can
proceed by applying a North West corner rule, because North West corner rule does not
depend on the values of r i and so on. So, if we apply the North West corner rule and
then, we apply the stepping stone method to solve it, we realise that under the
assumptions that we have in this transportation table.
The important assumptions in this table are, we are considering only regular time
production cost overtime production cost; we are considering inventory cost, we are not
considering back ordering cost. So, we are considering only 3 costs regular time,
overtime and inventory, right now do not have a beginning inventory, we have only
capacities based on this. And if we apply North West corner rule here and then, we want
to apply the stepping stone we realise that, all the positions remember that when we put 1
here the net was 0, when we put a 1 here the net was 0, but when we put a 1 here the net
was minus i.
So, only when we put a plus 1 and compare the net gain or loss, and if there is a gain we
can proceed towards the optimality, now when the net effect is 0 it is not going to help us
in anyway. So, when we apply the stepping stone method after applying the North West
corner rule, we understand that in all the possible non dummy positions, example ((Refer
113
Time: 30:39)) this, this, this, this, this, this, this, this, this and this and this. I am not
including these, because these are not possible positions, because they have a big M,
these are not possible positions.
So, for all the possible non dummy positions, one can easily show that the net by putting
a plus 1 is only 0, so when we apply the stepping stone method under all these
assumptions, it is not necessary to evaluate it at all non dummy non basic possible
positions. Example this is a non basic non dummy position, but it is not possible, because
there is a big M, this is a non basic non dummy position it is possible, because it has a
value of O, but if we put a plus 1 here, then the net is only going to be 0.
So, the simplification is that, if we apply North West corner rule and we apply the
stepping stone method, it is enough only to consider the vacant positions in the dummy,
and try to use them as candidates for improvement. Because, every non basic non
dummy allowable position is going to have a net equal to 0, so when we apply the North
West corner rule and the stepping stone, it is enough to consider only ((Refer Time:
32:20)) this, this, this and this and proceed with the transportation.
Now, this simplification is called the Bishop’s simplification and that is useful when we
apply the North West corner rule and the stepping stone method, so that the
computations that are involved in the transportation is minimised and are simplified. But,
it is not always or absolutely necessary to use the North West corner rule followed by
stepping stone, today with more and more packages available and spreadsheets solutions
through transportation problems available.
One does not take the trouble of actually solving it like I am doing it on the board, one
could simply put it into a solver and get the solution or use a spread sheet model to try
and get the solution. But, in the earlier days when people were actually solving it by
hand, the simplification was helpful, because it reduced the number of computations that
were there. Now, let us come back and do a few more things, now one of the
assumptions that we had of course, before I proceed one can now apply the stepping
stone method or the MODI method, to try and get to the optimal solution for this
problem.
Now, let us look at some more things, now let us relax one important assumption that
instead of back ordering not allowed, let us allow back order and see what happens. So,
114
when we allow back order, we want to remove all these, so if we allow back order which
means. I can use regular time capacity of 2nd period to meet the demand of the 1st
period. which means I back order by 1 period.
So, there is a regular time production cost of r plus a back order cost of s, so the cost
becomes r plus s, 2nd periods overtime to meet this the cost becomes O plus s. Now, the
3rd periods r t capacity I am using to meet the 1st periods demand, which means I am
going to produce here and back order for 2 periods. So, this will be r plus 2 s and O plus
2 s, and this will be r plus 3 s and O plus 3 s, now this is 3rd period capacity to meet the
2nd periods demand.
So, this will be r plus s, this will be O plus s, this will be r plus 2 s, this is O plus 2 s
comes, because this is the 4th periods capacity is used to meet the 2nd period’s demand,
so there is a back order for 2 periods, so r plus 2 s O plus 2 s, this is 4th period capacity
to meet the 3rd periods demand, so it will be r plus s and it will be O plus s. Now, the
transportation model is complete, please note that your North West corner solution is
going to be the same, it is not going to change, because North West corner is only based
on positions.
So, now after from this ((Refer Time: 35:58)), under this assumption where shortages or
back orders are allowed, we can now use this starting solution to try and get to the
optimal solution of the problem. So, the next question that we have to answer is, is the
115
Bishop’s simplification still valid for example, if I put a plus 1 here what is going to
happen to me. Now, when I put a plus 1 here, it will be a plus 1, it will be a minus 1, it
will be a plus 1, it will be a minus 1, so the net change will be r plus s minus r which is s,
s plus r plus i minus r which is s plus i.
So, what happens is when I put a plus 1 here, there is a net change which is s plus i, it is
not 0, we can still proceed by considering that the net change here is s plus i, and keep
proceeding towards the optimality by applying the stepping stone method or by applying
the MODI method. The only difference is that, if we apply the stepping stone method
now, now we have to in this case there are 5 columns 5 demand points and 8, so it is 8
into 5 40.
The transportation should have 12 allocations M plus N minus 1, but in this case it has
11, but we have added an epsilon to get 12, so there are 12 basic variables, 28 non basic
variables. So, if we apply the stepping stone method, then we have to evaluate the net
increase or decrease for all the remaining 28 positions, because the simplification rule
does not work when we consider back order. If we considered back ordering then it was
enough to do it only in 4 positions, now we have to do it for all the 28 positions.
But, as I mentioned earlier, now most of the times we do not use North West corner rule
or we do not solve by hand, if we are going to use a solver to solve this problem, it
actually does not matter, whether we are going to gain advantage of the simplification
rule or we are not going to gain any advantage, because of the simplification. But, then
this is also possible now the transportation problem has 4 costs, cost of regular time, cost
of overtime, cost of inventory and cost of shortage or back order. Now, let us try and do
one more thing to it, now earlier when we did the linear programming and we did the
tabular, we said there are 8 costs were considered.
Regular time overtime as one pair, inventory and shortage was as a pair, outsourcing and
under utilization was considered as a pair, hiring and laying off was considered as a pair,
out of these four pairs which add up to 8 cost, we have now considered four cost. Now,
can we also try and add some more costs into the transportation table, now let us try and
do that. Now, if we look at these capacities, we have now only two types of capacities
which are regular time capacity and overtime capacity.
116
So, the total capacity is 10000, 2000 into 4 months plus 500 into 4 months, but when we
did the linear programming formulation, we also assumed that it is possible to outsource
and to try and get. So, outsourcing can be another capacity, so we add one more capacity
here which is the outsourcing, so let us add outsourcing as another capacity. In the
tabular method or the linear programming method, we did not have any limit on the
extent to which we can outsource.
So, in principle we can assume that, the entire demand of 6000 can be outsourced, so we
have the demand of 6000 as outsourced. So, in some sense the outsourcing capacity
becomes sigma D, which is sum of the demands, now these capacities are already 10000.
So, there is a demand 6000, so now your total capacity is 16000, the total demand is only
6000, therefore the dummy will become 10000. So, this part is sigma D which is sum of
the demands, this part is sigma S which is the sum of the regular time and overtime
capacities.
((Refer Time: 41:28)) This part is sigma S sum of the regular time and overtime
capacities, this part is sigma D, the nice thing about this modelling is that, whether
supply is greater than demand or demand is greater than supply. This total is always
sigma S plus sigma D, this is always sigma S plus sigma D, and the moment we allow
outsourcing even in the event where we have this capacity is less than the total demands,
the balance can be got out of outsourcing.
Or going back to the example that we had, where I actually made a change here, initially
I had put a higher value here and then when we were doing it, if we had put a 3000 here,
and then we realises when we were doing it 2000 could come, 500 could come, because
these had big M’s back ordering was not allowed, when I had to actually change this.
Now, in order to have a situation where, if my 1st month’s capacity is 2500 and my 1st
month’s demand is 3000, if there is no beginning inventory, then there is definitely going
to be a shortage of 500.
And now in order to meet that we have to do one of these two things, which is either
allow back ordering or allow outsource. So, if we had allowed back ordering, then we
would have something like r plus s here and I would have put some allocation here, if I
had not allowed back ordering this would have been a big M, and it will come here the
balance would be outsourced. So, at the optimum the balance would be outsourced, if I
117
am not allowing for backorder, so outsourcing helps in multiple ways and prevents
infeasibility to the transportation problem.
So, we build an outsourcing, now the nice thing as I mentioned is that, if we call this
thing as sigma S, so this part is sigma S sigma D, this is sigma D sigma S, now it does
not matter whether sigma D is bigger or sigma S is bigger. We will still be able to meet
the entire demand, because outsourcing is available as an option, now the cost of
outsourcing, so let me simply call it as o u t, o u t, o u t, o u t, so there is a cost of
outsourcing.
So, if I want to outsource and buy something from the market to meet each months
demand, then I go and buy only in that market, so I am not going to say plus i and so on.
Every month, if there is a shortfall, if there is a way by which I can go and buy it from
outside, I will assume that this is available all the time and therefore, I will buy only
when it is required, so I have outsourced as a cost that comes. Now, the most important
thing is to fill this, now ((Refer Time: 44:25)) this has to be 0, because if for some reason
I do not have to outsource and I can meet all my demand using these capacities.
Then all these capacities and excess will have to go and sit here, only then the
transportation problem is balanced, therefore I need to put cost equal to 0 here, this is
very important, so we need to put cost equal to 0. The next question is, if I have this
regular time capacity and for some reason I am not utilising this regular time capacity
fully which means, if I have 2000 and in this row, I have not allocated 2000 here, there is
some allocation that comes into the dummy.
For example, if you look at this, this is a regular time of 2000 available 1000 has been
utilised, 1000 goes to dummy, which means this 1000 is not utilised, which means it
results in a under utilisation cost, so there is an under utilisation cost here. So, there will
be a under utilisation cost here ((Refer Time: 45:37)), there will be under utilisation cost
here, there will be under utilisation cost here. And as I mentioned we are not going to put
under utilization cost for overtime capacity, consistently we did not do that in the L P
formulation also as well as in the tabular method.
So, there is no under utilisation cost for the overtime capacity, but there could be under
utilisation cost for regular time capacity, this is a example where the 2000 regular time
capacity is not utilised fully, it is used partly only. So, now this table uses 6 costs instead
118
of eight inventory, shortage, regular time, overtime, outsourcing, under utilising, so it is
able to use six costs. Now, we can solve this problem optimally, if we want to consider 6
costs, it does not consider and it is not possible to easily model hiring and lay off.
So, transportation model can be used up to six costs, we do not model hiring and lay off,
if we also want to put hiring and lay off into the problem, then we have to go to the
linear programming model, which is capable of solving it, transportation does not do
that. Transportation also does not explicitly consider what is called the set up cost
associated with the production, because every time we produce, there is a set up cost and
a set up time that has to be used. And that cost cannot be brought in to the transportation
model, the linear programming model that we saw in the earlier class, also does not
include the set up cost.
But, there is a provision or there is a way by which it can be brought into the linear
programming formulation. Now, before I wind up I can tell you that, if we go back to the
examples that we considered, both in the tabular method and in the linear programming
method. And if we modelled it as 12 periods which means, we will have 24 rows plus 1
initial inventory 25 rows plus 12 periods 12 columns, if we did not consider the
outsourcing and under utilization, we considered only this, and solved a transportation
problem that involved 25 rows, 12 columns and a dummy.
Then the solution that we would have got would be exactly the same solution that the L
P optimum we get, if we added this outsourcing as one extra row and then, we solved it
this way. And if we had solved a corresponding linear programming problem by
considering 6 costs, including outsourcing and under utilization and not considering
hiring and laying off. Then such a linear programming problem and the transportation
problem will give the same solution, but linear programming has the additional
advantage of being able to model hiring and lay off, which we would not be able to do in
the transportation problem.
So, under the same set of assumptions, wherever if the problem it is possible to solve
using transportation, then the same solution we will obtain, if we solve by linear
programming under the same assumptions. But, L P model is little more versatile than
the transportation model, if one can solve it by modelling it as transportation, it is always
preferable to do that, because from operations research we would know that the time
119
taken to solve transportation problems can be quicker than the corresponding L P's. So,
transportation is preferred, wherever it can be modelled as transportation and
transportation can handle a maximum of 6 out of the 8 costs. Now, I mentioned the
transportation cannot handle set of costs explicitly. So, the next question is how do we
model set up costs into production planning, and we will try and look at models to do
that in the next lecture.
120
Prof. G. Srinivasan
Module - 1
Lecture - 7
Aggregate Planning, Dynamic Programming, Backordering
In today’s lecture we will see another approach to solve the production planning problem
and this approach is based on dynamic programming. Let us consider this problem that
we have shown here. Again let us assume that we make a single product and we have the
demand for four periods or four months given as 80, 60, 40 and 70. There is a setup cost
that is incurred whenever there is production. For example, if we chose to produce in the
first month then in addition to the cost of production there is a cost of setup which is say
rupees 60.
So, the setup costs are given as 60, 40, 60 and 45. There is also a cost of production
which is given as 5 4 5 and 5. In addition there is an inventory cost which means one can
produce in say month one and use the production to meet the demand of month two. If
we do such a thing then we incur a certain inventory cost or inventory holding cost. We
are also going to assume that the inventory cost of 2 2 1 and 2 represent the cost of
holding the ending inventory at the end of this period.
121
For example, if we chose to produce hundred in the first month and consume 80 the
remaining 20 is assumed to be carried from the first month to the second month, and it
attracts an inventory cost of 2, which is the inventory carrying cost at the end of month
one. So, these costs represent the inventory carrying at the end of each month, there is
also a production cost which is the cost required or cost that is incurred in producing
these quantities. At the moment we do not have an upper limit on how much we produce.
For example, if it is economical and profitable one can produce the total demand is 180
plus 70, 250.
So, if it is economical and profitable, one can produce all the 250 in the first month itself
there is no limit on the production capacity at the moment. As we move along we will
also realize that it is not very difficult to incorporate the production capacity if it is there.
So, there are three costs that are involved the cost of setting up, the cost of production
and the cost of holding inventory. In our problem we have also assumed that the setup
costs are different in different periods, inventory costs are different in different periods
and production costs are also different in different periods. So, the decision or the
planning decision is to find out, how much is to be produced in these four months in each
of the four months such that the total cost of setup total cost of inventory and production
the three cost put together is minimized.
So, it is a minimization problem to minimize the cost of setup inventory and production.
Right now we are not bringing time into consideration. We have not explicitly modeled
the time taken to produce these items and the capacity that is available in terms of time.
One should also note that the setup cost 60 also involves a setup time and since we are
not explicitly modeling the time, we do not bring the effect of setup time into the
analysis.
But if there is a production in a certain month a certain amount of setup cost is incurred
if we decide to produce in this month. This model is slightly different from that the
earlier models that we have seen - the tabular approach, the linear programming
approach and the transportation approach. The difference comes from the fact that for
the first time we are including a setup cost. Previous models that we discussed did not
include setup cost explicitly. This model includes setup cost explicitly. We also said that
whenever we decide to produce in a particular month, we have to incur the setup cost.
For the purpose of modeling we are assuming a single product whose demands are so
122
much, we also assume that if it is decided to produce in months one as well as in month
two, we will incur a setup cost of 60 in month one and 40 in month two.
We are not assuming that one set is setup is going to remain the same for all the four
months. Every month if there is a production there is going to be a setup. One may also
say that this facility can be used to produce more items and products and therefore, if this
particular product is made in month one, as well as in month two, then we incur two
setup costs one for each month. The problem can be formulated and solved as an integer
programming problem with of course, a binary variable that would say that y i equal to
one, if we decide to produce in month i and y i equal to 0, if we do not produce anything
in month i. And we can have constraints that relate the decision to produce which is the y
i and x i which are the quantities that are produced in this month.
So, integer programming - binary integer programming is definitely a methodology to

solve this problem. We also observe the dynamic programming is also a very good
methodology to solve this problem, where the stage will correspond to each of these four
months. The state variable for dynamic programming would correspond to the inventory
that is available at the beginning of each period. Also note that this problem does not
have an initial inventory there is no product that is available, there is no desired ending
inventory. So, we do not wish to have any ending inventory at the end of this period,
there is no beginning inventory as well as there is no ending inventory, in this particular
example that we are seeing.
So, if we use dynamic programming to solve this particular problem, then the state
variable will be the amount of inventory that is available at the beginning of each period.
The decision variable will be the quantity that is produced in each period and the
criterion of effectiveness or the objective function will be to minimize the sum of the
setup costs, inventory costs and production cost. Now, we all know that dynamic
programming, though very useful to solve certain classes of problems has its own issues
with respect to model. So, if we consider a problem of this type, the state variable can
take very large numbers. For example, total demand for the four periods is 250.
So, if we use the tabular approach of dynamic programming to solve this, the state
variable, which is the amount of inventory available can take very large number of
values, and the tables can become lengthy and cumbersome, though dynamic
123
programming can be used to solve this.
One of the most interesting simplifications to the dynamic programming solution came
from Wagner and Whitin and it is called the Wagner and Whitin algorithm which uses
dynamic programming to solve this particular problem. The Wagner Whitin idea or
principle is as follows. Now there are four periods and we are going to assume that we
are going to make decisions on production at the beginning of each of these four periods.
Now, it is only logical that if we produce in a certain period, then we will have to meet
the demand of that period at least. Otherwise we will have to incur one more setup and
one more setup cost, therefore it is uneconomical to consider two setups in a particular
period.
So, the if there is only going to be one setup in a particular period then it implies that
what we produce with what we produce, we should be able to meet the demand of the
entire period. One other assumption important assumption is that the demands have to be
met at the beginning of each of these periods, shortages and backorders are not allowed
in this particular model. So, under that assumption if we produce something in a month,
then either the quantity that is produced should be at least the demand of that month or
with the beginning inventory and production in that month, we should be able to meet
the entire demand of that month.
So, if there is a beginning inventory in the beginning of a particular month, let us say we
124
are looking at month number two and let us say if there is a beginning inventory of say
20 in the beginning of month two, then we should produce 40 or more, but then it is also
logical to say that if there is a beginning inventory at the if there is inventory at the
beginning of this month. Then if that inventory is sufficiently large to meet the demand
of this month, then there is no need to setup and produce in this month.
So, Wagner and Whitin algorithm works around these ideas. So there are two important
points there - one is there will be production in a particular month, when only when the
beginning inventory is zero. And if there is production in a particular month then that
production quantity will be the demand of that month or the demand of that month and
the next month and the demand of three consecutive months or the production quantity is
always equal to the demands of certain number of consecutive months. The second if
there is beginning inventory in a month, then that inventory should be sufficient to meet
the demand of that month and there will be no production in that month.
So, these are the two important results of Wagner and Whitin, which I once again repeat.
If the beginning inventory is 0 or only when the beginning inventory is 0 we will decide
to produce in that month. And if we produce, the production quantity is equal to the
demand of that month or the demand of two months or demand of three months, or
demand of a certain integral number of months. If there is beginning inventory then we
will not produce and that beginning inventory should be sufficient to meet the demand of
this particular month.
The moment we accept and use these two results that came from Wagner and Whitin, the
dynamic programming solution to this problem becomes extremely simple. Now, we will
look at how to solve this using dynamic programming, under the assumption that
backordering is not allowed. So, let us go back and use the DP solution - dynamic
programming solution. So, let us start with trying to meet the demand of month one. So,
we first look at trying to meet the demand of month one, there is no beginning inventory
so we have to produce in month one. So, demand of month one is met only in one way
by deciding to produce.
So, we produce the cost equal to production quantity is equal to 80 because demand is 80
and the best way to meet this 80 is given by total cost is equal to setup cost is 60, which
is to be incurred production cost is 5 into 80 equal to 460. There is no inventory that is
125
carried because demand is 80, production is also 80, there is only one way to do it which
is the optimal way. So, we just put a star indicating that this is the best way to produce
and meet the demand of the first month.
Now, we go to demand of month two, now the month two demand can be met in two
ways one is to try and meet the first month’s demand in the most economical way and
use the previous result, which is the essence of dynamic programming where the optimal
decision up to the previous stage is computed and used for decision making in the next
stage. So, one, is to try and use you meet the month one’s demand in the best possible
way, which is 460 which comes from here plus produce what is required for month two
by incurring a setup cost of 40 plus production quantity is equal to, this is Q equal to 60
TC is equal to this.
We just write TC total cost is equal to - total cost of meeting this second month’s
demand as well as the first month’s demand, two months demand is to meet the first
month’s demand in the best possible way, which is 460 that comes from here. And then
meet the second month’s demand alone by incurring a setup cost of 40 and a production
cost of 60 into 4, 240. So, this gives a total cost of 240 plus 40 is 280, 340 plus 400, 740
is the cost that is associated with this solution. The other way to meet month one and two
demand is to produce everything in the beginning of the first month. So, that would give
us Q equal to, but it is produced in the first month.
So, all the 140 units are produced in the first month so there is a setup cost of 60 that is
incurred, all the 140 units are produce in the first month. So 140 into 5, 700 plus these 60
units are carried from first month to second month. And therefore, there is an inventory
carrying cost of this 60 into this two, this two represents the cost of carrying one unit of
inventory at the end of the first month to the second month. So, 60 into this 2 is 120 so
this becomes 700, 820 plus 60 is 880. Now, the demand of months one and two put
together can be met in two ways one which gives us a cost of 740, the other that gives us
a cost of 880. And since we are trying to minimize 740 is the best way to do this.
Now, we go to month three now. Month three can be met in three ways, one is to meet
month one in the best possible way and then produce for two and three in month two.
The other is to meet month two in the best possible way and produce for month three in
month three. And the third is to produce everything in the first month and carry for
126
months two and three. So, we have three ways of doing it. So, let us assume the first way
is we produce everything in the first month and then we carry which means, there is a
setup cost of 60 the total quantity produced is 60, 80 plus 60, 140 plus 40, 180, 180 into
5, which is 900 plus 100 units are carried at the end of the first month. So, 100 into 2
plus another 40 units are carried at the end of the second month so plus 40 into 2.
So, this gives us a total cost of 900 plus 200 is 1100, 1180 plus 60 is 1240. Second way
to do it is to produce the first months demand in the best possible way, which is 460 and
produce the demands of months two and three in the second month. So, in the second
month we incur a setup cost of 40, we produce the demands of months two and three
which is 60 plus 40 hundred in the second month so hundred into 4, 400 plus.
Now, this hundred is produced in the second month, but 40 is carried to the third month
so there is an inventory cost of 40 into 2. Now, this gives us a total cost of hundred into
4, 400, 480, 520, 580 plus 400 is 980. The third way to do this is by producing the
demand of months one and two, in the most optimal manner which is by incurring the
740 and producing the demand of month three in the third month.
So, this is to incur the 740 and then produce the demand of the third month, in the third
month by incurring a setup cost of 60 quantity produced is 40, production cost is 5, 40
into 5 there is no inventory cost here because this 40 is produced in the third month and it
is not carrying. So, total cost here is 40 into 5, 200, 260, 300 plus 700 is 1000. So, the
best way to do it is here. Now we go to month four.
127
Now, we compute the total cost for month four in four ways, one is to produce all of
these in the first month itself and carry. The second is to produce up to this optimally and
then produce the rest here, the third is to produce upto this optimally and produce the rest
here. And the fourth is to do upto this optimally and produce this.
So, we will compute the cost for all of them. So the first one will be to have all four. So,
produce 80 plus 60, 140, 180, 250. So, TC1 first way to do it is to produce all the 250 in
the first month itself by incurring a setup coast of 60, production cost of 250 into 5 unit
production cost is 5 totally 250 are produced. Now, out of the 250, 170 is carried to the
second month. So, 170 into 2 plus 110 are carried to the third month 110 into 2 plus
another 70 is carried to the forth month, so out of the 110, 40 is consumed 70 is carried
70 into 1 plus 70 and this gives us a total cost of 1250 plus 340 is 1590, 1590 plus 220 is
1810, 1880, 1940.
Now, we look at the second one where we do up to month one optimally so you incur
460, and we produce this 2, 3 and 4 in month 2. So, production quantity is 170 into
production cost is 4 so plus 170 into 4 plus out of the 170, 60 is consumed here so 110 is
carried 110 into 2. Now, out of the 110 that is carried 40 is consumed 70 is goes to the
fourth period so 70 into carrying cost of 170, and this gives us a total cost of this is 680,
740, 1140, 1160, 1260, 1330.
So, the total cost is 1330, let me repeat 170 into 4 is 680, 740, 1140, 1260, 1350, 1430.
128
Let me do it again 460 plus 680, 680, 740, 1140, 1000 plus 220 is 1360 plus 70 is 1430
setup. So, I have to include the setup time also, there is a setup cost of 40 which comes
here so that 40 has to come here. So, let me do it again from the beginning so the 460 is
carried from here there is a setup cost of 40 for this period. So, there is a setup cost of 40,
the production quantity is 170 and the cost is 4 so plus 170 into 4 plus out of the 170, 60
is consumed 110 is carried, 110 into 2 plus out of the 110 that is brought here 40 is
consumed 70 is carried, 70 into carrying cost of one is 70.
Now, the total is 460 plus 40 is 500, 500 plus 680 is 1180, 1180 plus 220 is 1400, 1470
that is the total cost associated with this. Now, in the third way of calculating it, we
produce up to this optimally incurring a cost of 740 and then we chose to produce these
two in period 3. So, it is 740 plus there is a setup cost there is a setup cost of 60. So, plus
60 there is a production quantity of 110 production cost is 5.
So, 110 into 5 plus out of these 110, 40 is consumed 70 is carried 70 into one is 70 so
this is 740 plus 60 is 800, 800 plus 550 is 1350 plus 70 is 1420. Then the forth one we
produce upto this optimally and produce the requirement of month 4 in 4. So, the optimal
way is 980 from here plus there is a setup cost of 45, there is a production quantity of 70
and production cost is 5 plus 70 into 5 there is no inventory involved.
So, 980 plus 350 is 1330 plus 45 is 1375 which is the optimal one for this, so out of these
four the cheapest is 1375. So, what are the production quantities the minimum cost is
1375, what are the production quantities? Now, the production quantity here in the forth
one production quantity is 70, so we will now say Q 4 is equal to 70.
So, when Q 4 is 70 it means we are producing 70 here, which means we have used up to
the third period, we have used the optimal solution. So, we have used this 980 which we
see has also come here. Now, go back to this 980 and find out the production quantity
which is hundred into 4, hundred is the production quantity so that hundred comes from
60 plus 40 so Q 2 is hundred. Let me explain how I got that hundred now the moment we
are here we are going to see the production quantity for at this point, this is the
production quantity hundred into 4 is the production cost. So, hundred is the production
quantity. So, I am somewhere here up to this I have solved optimally and i realize that
my production quantity is hundred which means, I am producing here 60 plus 40 so Q 2
is hundred.
129
So, we have written Q 2 equal to hundred and you also realize that 460 comes here
indirectly you also realize that here I have produced hundred. So, I have to look the
optimization here which happens at this point production quantity is 80 so Q 1 is equal to
80. Now, total cost is equal to 1375. So, we have now solved this optimally using
dynamic programming now there are four periods the optimal solution does not ask you
to produce in all of the four periods, it is enough to produce here it is enough to produce
here and it is enough to produce here. In some sense it is it is understandable because the
production quantity comes down here, so there will be a tendency to push the next
month’s production into this month, pushing this might involve a higher inventory
holding cost.
So, it is a tradeoff between inventory holding cost and additional change over or setup
cost, but if the production cost themselves change, then this will also have a bearing on
the solution. Now, in order to bring out the features of the algorithm we have considered
a problem where the inventory holding costs are different in different periods, production
costs are different in different periods, and setup costs are also different in different
periods.
Though one cannot exclude this from happening which means in organizations, we will
realize at some point when these costs are measured, one would realize that to setup for
the same product the setup cost can vary in different months. The reason could be many
sometimes even the labour associated with the setup the cost can change depending on
pay revision and additional payments made to the people. And therefore, setup costs can
change.
Inventory cost can also change due to many reasons for example, if inventory space is
added in a certain period or space is removed from the storage area, then there can be a
situation where the inventory cost can change. Production cost can also change due to
cost of labor and multiple other cause of consumable and so on. So, these things can
actually happen, but this model is also applicable when the production costs are the
same, the inventory costs are the same and the setup costs are the same across all
periods. There it becomes a tradeoff between cost of holding and additional inventory for
a particular period versus the cost of incurring one setup or one change over.
The point I am making is if all these costs are the same like 60, if all these costs are the
130
same 2, all these costs are the same say 5, then it is a tradeoff between incurring a setup
of 60 and carrying a certain quantity of the demand for the next period multiplying it by
two. Wherever it is cheaper than we would chose to setup or we would chose to produce
early, and carry the inventory. So, the Wagner Whitin algorithm actually has become,
once we apply a Wagner Whitin idea into the dynamic programming, we realize that it
has become extremely simple. All we have done is for four periods we have just
evaluated 4 plus 3 plus 2 plus 1, which is 10 possible scenarios.
Whereas, if we had not applied the Wagner Whitin idea and done using dynamic
programming then for example, the state variable can take any value from 0 to 250 and
therefore, we would have evaluated many, many more scenarios than we actually have
evaluated. So, the Wagner Whitin idea has helped in the dynamic programming solution
to this problem.
Then comes the next question do we allow backordering into the Wagner Whitin
solutions. So, we also address the backorder problem, the problem gets a little more
involved when we consider backorder. So, for example, how can we consider a
backorder, see right here we said that see here when we did not allow backordering, there
was no beginning inventory. So, the demand of the first month has to be met in the first
month itself. So, we did not consider a situation where the first month’s demand was
actually met out of producing 140 here, if backordering is allowed then such a possibility
131
exists, I do not produce anything here.
Right now when we did not consider backordering, this model will definitely have Q 1
equal to something because the first month’s demand has to be met at the end of the first
month, and there is no beginning inventory. But in the backorder case we could actually
have considered the scenario where we do not produce anything here we make 140 here
to meet these two and maybe we could have considered another scenario, where we
make 110 here to consider these. Or we could even consider a scenario of producing 180
here meet a backlog of 80 meet this month’s 60 and meet the next months 40 here and
the 70 here.
So, considering backordering throws open lot more scenarios then we can think of, but
then we also have to include a backorder cost. So, let us not spend too much time
evaluating all possible scenarios, but let us try understand one or two scenarios. So, let us
include a backorder cost which comes here. So, let us consider the backorder cost as 1, 2,
1 and 2. So, I am going to write the backorder cost as 1, 2, 1 and 2 and this quantity is
backorder cost is shown here. For example, if I chose to produce 140 here which means
this period is backordered.
So, the backorder cost will be 60 into this 1 which is 60 now as I said it throws open
several scenarios. So, let us just look at one scenario that involves backordering and let
us try and find out the cost. So, let us look at a scenario of 0 180, 0 70 it means I am
going to produce 80 plus 60 plus 40, 180 here which means I am backordering this
demand I am meeting this demand in the month itself and I am carrying to meet this
demand and then the fourth one that comes here.
So, if we do such a thing now what is our cost now? I am incurring a setup cost of 40.
So, total cost will be setup cost of 40, I produce 180 here the unit production cost is 4,
the unit production cost is 4. So, 180 into 4 now out of this production this 80 I am
actually backordering first period 80 I am backordering. So, let us say my backorder cost
is 80 into 1, this 60 I am meeting in the period itself, this 40 I am carrying to the next
period. So, the inventory cost is 40 into 2, so inventory cost is 40 into 2 plus the fourth
period I am going to produce 70 here, I incur a setup cost of 45, I incur a production cost
of 70 into 5, there is neither inventory nor backordering.
So, it is 45 plus 70 into 5 so plus 45 plus 70 into 5 this is 720 plus 350 is 1070, 1070 plus
132
80 is 1150, 1150 plus 80 is 1230, 1270 and 1315. So, this is the total cost when we
consider backordering in this case. Please note so far we have not solved the problem
optimally, we have just illustrated a possible solution that includes backorder cost and
we are able to show there, if we include backordering then the cost comes down from the
optimum 1375 to at least 1315 because 1315 is a feasible solution considering
backordering.
So, backordering is advantageous from a cost optimization sense, which is also

understandable because when we solved this problem optimally, saying that there is
going to be no backordering, which was one of the assumptions, which means we have
restricted the normal problem that could include backordering and we had solved a
restricted version of a minimization problem. So, a restricted version of the minimization
problem will only have an optimum solution higher. Now, this was a relaxed version, so
this optimum solution will be 1315 or lower.
So, we can actually solve the problem optimally using the Wagner Whitin ideas by
modifying it by saying that if my inventory is 0, I may even skip this period and
backorder it to the next period that is the only modification. But if I have an inventory I
will not skip and backorder because if I am incurring both inventory cost and backorder
cost I would not do that. So, if there is 0 inventory, I need not produce in this period I
may still skip production and add it to backorder. Now, for the same problem if we
consider backordering, we have to evaluate more scenarios than we actually have
evaluated.
In this case there are only ten scenarios, but the case with backordering has more
scenarios than we actually have. Month one we incur cost equal to setup cost of 60 plus
we produce the demand of 80 of month one incurring a production cost of 5. So, plus 80
into 5 would give us 460 as the cost for month one. Now, we move to month two now we
can meet the demands of month two in multiple ways. So, we could do 80, 60 which
means we produce 80 in the first month and 60 in the second, we could do 140 0 and we
could also do 0 140.
So, this gives us a cost equal to 740 140 0 would give us a cost equal to 880 and 0 140
would give us a cost equal to 680. So, let me explain the computation of this cost
because this involves backordering so we produce all the 140 in the second month and
133
we meet the demand of 80 of the first month and 60 of the second. So, cost to produce
this 140 there is a setup cost of 40 plus there is a production cost of 140 into 4 plus - this
80 of the first month is backordered. So, the backorder cost for the demand of the first
period is 1. So, plus 80 into 1 so this would give us 40 plus 560 which is 600 plus 80
which is 680. So, out of these possibilities this seems to be the best one and this is the
one that involves backordering.
So, with this we move to month three, where we can look at several combinations for
month three so we could use 180 0 0 we could use 80, 100 0, 80 0 100, 0 140 40, 0 180 0
and 0 0 180. Let us look at finding out the cost in two such instances so we will show the
computation for 180 0 0 the cost is 1240 for 80 100 0 it is 980. So, let us explain the
computation in detail for a couple of instances. Now, when we say it is 180 0 0 it means
in this case, we produce 180 in the first month consume the demand of 80 and then carry
over hundred to the next month. And then consume the demand of 60 and carry over the
remaining 40 to the third month.
Now, let us illustrate the computation for two instances. Let us do it for this instance as
well as this instance. Now, here is an instance where we produce 180 in the second
month to meet the demands of the three months, which means there is a backorder here
as well as there is an inventory carrying here. In this scenario we are assuming we
produce the total demand of 180 in the third month. So, there is no inventory carrying,
134
but there is backorder, but backorder for two periods. So let us first show the
computation for this so TC for this.
So, let me illustrate this with some kind of a plus sign here so TC for the plus sign will
be - I produce 180 here in the second month. So, setup cost is 40 plus production cost is
180 into 4 plus the first period demand of 80 is backordered - the cost of backordering
the demand of the first period is 1. So, plus 80 into 1 now here the backorder 80 we
consume 60 for this period. So, the balance 40 is carried over from second month to third
month so inventory at the end of period 2 is 2. So, 40 is the quantity carried into to cost
which is equal to 2 so this gives us a cost of 40 plus 720, 760 plus 80 840 plus 80 equal
to 920 which is what we find here.
Now, let us show the computation for this one. Let me illustrate this with TC+ just to
show that we are looking at this computation. Here, as I mentioned earlier there is no
inventory, but there is backordering. So, the production is done in the third period by
incurring a setup cost of 60 plus production cost is 180 into 5 plus the demand for both
the periods are backordered. So, for this period the demand is 80 per period backorder
cost is one, it is backordered for two periods. Therefore, it is 80 into 2 plus another 60
demand for this period is also backordered.
Now, backorder cost per period for month two is 2 therefore, plus 60 into 2, we should
note the difference in the way we compute the inventory cost as well as the way we
compute backorder cost. In this particular example, when we did this 180 0 0, if I were to
calculate the total cost here now out of this 180 that is produced in the first month 80 is
consumed. So, hundred is carried at the end of the first month so hundred into inventory
cost of 2 plus another 40 is carried at the end of the second month so 40 into 2. So, when
we do the inventory calculation we take the inventory at the end of that month and then
multiply by the inventory cost at the given for that period. When we do the backorder
calculation we go back here and say that the first period demand of 80 is backordered for
two periods.
Now, per period backordering cost for the first period demand is one. Therefore, 80 into 2
into 1, 1 is the cost two is the number of periods here. Similarly, 60 is the quantity
backordered its backordered for one period with the cost of two so 60 into 2. So, this gives us
60 plus 900, 960, 960 plus 160 is 1120, 1120 plus 120 is 1240. Now, out of these
135
we realize that this is the one with minimum cost then we move to month four.
Now, with month four we work out several combinations and these are shown here so we
could do 250 0 0 0 which is 1940, 0 250 0 0 1410 0 0 250 0 1660, and 0 0 0 250 which is
1815, 80 170 0 0 which is 1470 80 0 170 0 which is 1560 80 0 0 170 1635, 0 140 110 0
1360, 0 140 0 110 1315, 0 140 40 1335, 0 180 0 70 1315. So, these are the various
solutions which we evaluate here. We also realize that these four are the situations,
where all the 250 is produced in a single month, and these three cases involve
backordering because from here. We backorder this demand, from here we do both from
here all the three are being backordered.
Now, when we do this backordering, the cost computation will be - this 80 demand is
backordered for three time periods, the cost of backorder is 1. So, this will be 80 into 1
here it will be 60 into 2 into 2 because the 60 demand is backordered for two periods.
The cost of backorder in that period is 2 so this will become 60 into 2 into 2 and in this
one it will become 40 into 1 because the other 40 is backordered for one period, the cost
is also one. So, that when added with the setup cost of 45 and production cost of 350
would give us 1815.
So, these four look at the cases, where all the demands are produced in a single period.
Now, here what we do is we have this 80 and then the three cases, so if we look at this
for month one we have 80, which is the first one the remaining three are now produced in
three possible ways including backordering, there is backordering here, there is
backordering here. Now, we move to the second month and we realize that the best value
is 0 140. So, we take the 0 140 and look at the other cases that are here with 1335.
And the third one the best case is 0 180 so we have 0 180 0 70, 0 180 0 gives us this 920
plus 70. Now, we realize that the best value happens in two instances, which is 1315 here
which also happens to be the solution here that we shown compare to 1375 of the earlier
one this is the solution, without backordering. And these are solutions with backordering.
So, the 1315 that we calculated here and showed as a feasible solution is actually optimal
with the case with backorder. So, this is how we work out and solve the production
planning problem using dynamic programming. Now, there are three important things
which we have to look at before we wind up this discussion on the dynamic
programming model. The first and foremost is this model is different from the earlier
136
models because here we have considered the effect of setup. In the tabular method, in the
linear programming base method and in the transportation based method, we did not
consider setup costs.
Now, dynamic programming gives us the opportunity to consider setup costs. The
Wagner and Whitin assumption has made the solution much simpler. When we did not
consider backordering we looked only at ten cases, when we considered backordering we
did look at more cases I have not gone to the detail of calculation of each of these
numbers, but these are not difficult considering that I have shown how each of these
numbers are calculated. And also considering how I have shown how the one particular
thing is calculated.
So, one can calculate all these costs and then find out the number of calculations has
increased it is a little more than even double the thing that we have here, which means if
you are looking at n periods then this is roughly 1 plus 2 plus 3 up to N and that would
go about 2 to 3 times that number. Still the number is actually small and manageable
because the Wagner and Whitin method allows us certain conditions, where we can use
to solve this problem optimally. Then there are two other issues that we have to look at
one is should we really consider backordering, what is the impact of backordering, what
is the effect of backordering from a practical point of view.
One is for the purpose of illustration I have used that backordering cost here 1 2 1 and 2
are kind of comparable to the inventory cost of 2 2 1 and 2. In fact in this particular
numerical illustration I have even used values that are smaller than this, but in reality
cost of backorder is much higher than cost of holding inventory. Cost of backordering
has an important component, which is the loss of customer goodwill because
backordering is employed.
So, which is not a cost that can be easily measured so cost of backordering is much
higher, many times backordering also results in for example, air lifting a certain product
from one place to another place because there is a delay and already there is one period
backlog that is accumulated. So, people would like to transport it as quickly as possible
with increased cost. So, cost of backordering is much higher than cost of holding the
inventory. So, should backordering be planned many times the answer is no because
there is not very good to plan backordering, but unless there is a requirement and there
137
are certain demands which can actually be backordered and not met, which simply means
the customer is ready to postpone it because the customer has inventory already in thus
in the system.
Then in such circumstances backordering may be permitted or considered. The third part
is we assume that the demand all the two hundred and fifty could be produced in a
certain period, many times there are capacity restrictions and we would not be able to
produce all the 250. So, there can be restrictions on the maximum production that can be
made in a particular period. Now, in the next lecture we will look at some more aspects
of aggregate production planning, we will look at a couple of more models and one more
model and a couple of other issues after which we will move from aggregate planning to
materials management or inventory control.
138
Prof. G. Srinivasan
Module - 01
Lecture - 08
Aggregate Planning, Quadratic Model, Demand and Capacity Planning
So far, we have seen several models for aggregate planning. We started with the tabular
method. Then, we looked at the linear programming or L P based algorithm, we then did
the transportation based algorithm. We also did a dynamic programming based
algorithm. The tabular method was an evaluative algorithm, where the user inputs the
values of the production quantities, and the workforce as relevant and the tabular method
would evaluate a total cost function.
The user could vary these values and try and locally optimize the total cost. The linear
programming based formulation, actually considered a maximum of eight different costs.
This kind of occurred in four pairs regular time and overtime costs, inventory and
shortage costs, hiring and laying off costs, outsourcing and under utilization costs and
then there were a variety of constraints.
The transportation based algorithms essentially solving a certain restricted version of the
linear programming, the transportation model did not consider variable workforce, but
139
we showed that it could consider the six remaining costs. Then, we solve the dynamic
programming based algorithm. In the D P algorithm, we introduced a new idea that if
there is production, then there is a set up. So, there was a definite set up function or a set
of costs whenever there was a production. Now, all four of them have something in
common which is that all four in all four models, the costs was linear or proportional.
For example, in the tabular method, if we said that it is going to cost 100 rupees to
produce an item on regular time, then we said it would cost 200 rupees for 2 items. If
certain x number of items or man hours is utilized and it cost 100, then the total cost will
be 100 into x. So, the same linearity or proportionality was used across all four of them
even in the D P based algorithm.
We said that if there is an inventory cost of carrying a unit from one period to another if
x units are carried then the total inventory cost will be x into the number of the x into the
cost of carrying a unit from one period to another. So, there was this linearity or
proportionality in all of these, let me just write only the objective function of the L P
algorithm and also try and show the linearity there.
So, in the linear programming formulation, the objective was to minimize sum of - we
used a certain notation we used r into R t plus o into O t plus i into I t plus s into S t plus
h into h t plus l into L t plus into outsource of t plus u into U t. We had used eight costs
in the objective function and the objective function was to minimize the sum of eight
140
costs the lower case alphabet represents the unit costs. The upper case represents the
variable, so if capital R t is the number of regular time man hours used and r is the cost
per man hour per period.
Then, the total cost of regular time is equal into r into R t summation over t equal to 1 to t
for all periods we could consider that this r is constant across all periods. We could have
this r changing across all periods and if it changes then we would add a small subscript R
t which says this is the regular time costing period t, but if this cost is going to be same
across all periods.
Then, we will use this, similarly, we use all these values, now obviously when we have
to solve this linear programming problem, it is necessary to know the values of R, O and
I, S, H, L, OUT and U. These will have to be provided by the organization for whom or
for which an aggregate plan is being made. Now, let us ask ourselves two things. Let us
first try and question the linearity assumption in the objective function. Then, let us also
ask a question as to how does one go about finding out or computing the actual values of
these coefficients.
It is only customary to say that if we want to find out this r which is the cost of regular
time production, which means cost per man hour of regular time, how does an
organization compute this value? There are many ways by which they can do it, but the
most common way to find out what are all the components of the regular time cost. Now,
in this model we have not explicitly used the payroll, so we assume that pay roll as also a
part of this r the salary to the operator who is working plus it will have certain raw
material cost plus it will have some cost of consumable and so on.
So, all these components of this cost are now listed and then one goes back to earlier
period, say the last six months or the last one year, and then we can compute total money
spent on regular time production on each of the components of the regular time cost.
Then, we add them to get the total regular time production cost incurred over the last five
or six periods, and then when we divide it by the actual quantity or man hours employed
in regular time production. Then, we would get this r, similarly we would get this o
where all the components of the overtime costs are listed.
Then, they are evaluated or computed and that total cost divided by number of man hours
used in overtime would give us this o which is this per period per man hour cost for
141
overtime. Thus one can go about trying to compute all these costs. May be, the easiest
would be the outsourcing cost assuming that the entire product is outsourced. So, finding
the outsourcing cost would be relatively easy, inventory cost would be reasonably easy.
If the inventory cost is dominated by the interest on borrowed capital if we are going to
use the other components of inventory cost such as cost of space cost of power electricity
and other resources cost of man power cost of special facility.
Then, this it would have be largely dominated by the interest on borrowed capital or the
interest rate. Then, the other cost associated with space people facility etcetera will have
to be added. Now, computing this S is not easy because any shortage or back order cost
will have an important component which is eventual loss of customer good will by not
delivering things in time and by back order. So, that cost is not easy to compute,
similarly the cost of underutilization though computable is relatively difficult because
this is the cost of keeping facility idle or keeping a man hour of resource idle.
So, it also has to be computed carefully H and L are cost hiring and lay off which can be
calculated reasonably well. At the end of it, I wish to say is that it usually for any of
these eight costs, the common practice is to go back in time and find out what was
actually spent under this cost. Whatever expenditure could qualify under this head of say
either hiring cost or a outsourcing cost or a overtime cost and then divide it by the
number of man hours that is been used under that to get this. Therefore, it would be
difficult to have an R t and say that if we are looking at twelve periods, it would be
difficult actually to give twelve different values.
Let us say that regular time cost was so much was in the first period, so much in the
second period and so on. So, it is customary to use the same value of r across all twelve
periods and because of the way r is used and computed we have we are forced to use the
linearity or proportionality saying that if this is a per hour per man hour regular time
cost. If R t man hours are being used then r into R t is my total regular time cost, but in
practice we also know that every cost particularly in production related situations, there
is a fixed component.
Then, there is a variable component, so if we start writing this r itself as having two
components something of the form say r naught plus r 1 into r of t if this r has is of this
component which means there is a fixed cost r 0. Then, there is a variable cost per man
142
hour used, then which is which a very common occurrence because the organization has
certain fixed cost which can be apportioned to each of these costs. So, if we start defining
this r itself as a fixed cost having a fixed component and a variable component then this
will become r naught plus r 1 into R t multiplied by R t. So, this will become quadratic in
nature, so if each of this cost has a fixed component and a variable component, then the
linearity assumption is questionable.
We may have to use a model use which is quadratic in nature. Perhaps the other
disadvantage of the linear formulation the advantage of the linear formulation lies in its
simplicity. It is easier to formulate proportionality constraint is accepted and additively is
also accepted, methodology wise the simplex algorithm is well known and it is easy to
solve the linear programming problem. Then, when we computed this R and O and I and
S if in reality this R had a fixed component and a variable component, but for the ease of
computation, we have simply used one particular R which represents that R naught plus
R 1.
Then, it is likely that there could be an error in the computation of r and then the linear
programming solution can sometimes give even a higher cost. Then, what is actually
incurred by the organization because the cost that is actually incurred by the organization
has a fixed component and the variable component. Then, we have approximated it to a
single number r. The advantage of linear programming technique is its simplicity ease of
modeling and ease of solution. The disadvantage is perhaps its inability to actually
capture the reality because in reality costs have a fixed component and a variable
component.
So, considering this in mind some authors proposed a quadratic model. So a quadratic
model for linear program for aggregate planning is quite famous. It is also called the
HMMS model named after four researchers who are the authors of this work stands for
Holt, Muth, Modigliani and Simon. So, it is called the HMMS model, we will not get
into the full details of the HMMS model, but I would possibly write down only the
objective function and then explain how the HMMS model actually works. So, the
HMMS model the objective function would be minimized, so the objective function has
several costs, first the variables W t is the workforce employed in period t.
143
Here, P t is the number of hours of production in man hours in period t, W also can be
written in terms of man hours production man hours workforce man hours and I t is the
inventory at the end of period t which could also be written in terms of man hours. So,
there are three sets of decision variables a certain variables for the workforce certain
variables for the production and variables for inventory. Now, the objective function is
made out of several costs. Now, one is it has regular time payroll cost. This is like
regular time payroll cost. There is a fixed component and a variable component, now
there is an overtime cost, there is an inventory cost and then there is a workforce
changeover cost.
So, this is the inventory cost, this is the workforce changeover cost and this is the
overtime cost, this is the overtime cost expanded in a certain form. Now, we have
realized that this portion is called regular time payroll cost, there is a fixed cost and a
variable cost, this is the typical inventory cost. In this model shortages or back orders are
not allowed. So, I t minus I star, there is an ideal inventory that is defined and I t minus I
star the whole square is the penalty and that is multiplied by a C 8. Now, here again there
is a workforce changeover cost between W t and W t minus 1 and C 12 whole square.
Now, these are other costs that relate production and workforce, so they would represent
the overtime payroll cost and so on. Now, we are not going to go into too much of detail.
One form of the objective function is kept here. It is not absolutely necessary that the
objective function has to be this. This is one form of the objective function. This
objective function shows a quadratic objective function, the quadratic comes from this,
comes from here as well as comes from here. It is not linear anymore now there is a
quadratic objective function subject to a set of linear constraints.
This is a standard constraint that relates production inventory and demand. D t is the
demand in period t in terms of man hours. Here, I t is the inventory at the end of period t
I t minus 1 is inventory at the end of period t minus 1, which is inventory at the
beginning of period t, this is production in period t, this is demand in period t. So, this is
typically the inventory balance constraint, so now we have a quadratic objective
function. Let me again repeat and indicate that it is not absolutely necessary that his has
to be the objective function. A quadratic model suitably defined for a specific
organization may include additional costs also or may leave out certain costs.
144
Also, this may include backorder which is not included in this function. It may include
outsourcing which is not included in this function and so on. This is only a representative
quadratic function where some of these costs are shown as quadratic functions of the
decision variables which are essentially P t, W t and I t. So, one may look at other
objective functions also, but the objective function is quadratic, the constraints are linear,
and then we have a non negativity restriction on the variables. So, this becomes a
quadratic programming problem which can be solved in many ways. A popular way is to
actually one could think in terms of taking this constraint into the objective function
using a Lagrangean multiplier.
Then, one could solve the other one that is now associated here is quite simple in the
sense that we have a quadratic function. So, the fundamental principle of a quadratic
optimization problem is that the first derivative of a quadratic function is a linear
function. So, one can differentiate this with respect to the variables at the same time one
should be able to capture the dependency of the variables, the dependency of the
variables are here. So, even though we have P t, W t and I t as variables there is a
relationship between P t and I t which is given here.
So, it is possible to eliminate one of the variables through this constraint and then
differentiate with respect to the other variables and solve it. So, when we differentiate we
will get a linear function involving P t and W t, we also have equations here
differentiation would also give us a set of equations. So, finally, what will happen is on
differentiation and inclusion of this constraint we will have a set of linear equations to
solve. Now, as we proceed we may have some more variables than the number of
equations the reason being these I t minus 1 and I t as well as W t minus 1 and W t
would sometimes create more variables than the number of equations.
In such cases, we define some initial conditions as well as some final conditions that we
want, so we make the number of variables equal to the number of equations. Then, we
solve a certain equations to get the optimum values of P t, W t and then I t from them
using this relationship. So, the primary purpose of doing the quadratic model is only to
inform or indicate the need to consider a quadratic objective function because though
linear optimization and linear programming and linear model is very simple to model. It
is easy to solve, it has the difficulty that these coefficients, there can be some kind of an
approximations in these coefficients.
145
Even though we solve the L P optimally, a solution may not reflect what actually
happens because in reality the costs are not linear. So, if you want to capture that, then
we can bring it into quadratic model with a suitably defined objective function and I have
explained broadly the methodology to solve a quadratic optimization problem. So, in
terms of solution methodology, this is a little difficult compared to this, but this can give
results, which are little closer to what is happening because in practice costs have a fixed
component and a variable component. Therefore, can result in quadratic optimization
problems, now let us look at some more aspects of aggregate planning before we go to
the next topic if we look at the costs that we wrote, we wrote eight different costs.
So, let me consider two of those costs here one is the inventory cost and the other is a
shortage cost the other pair is hiring and lay off and then of course, regular time overtime
and outsourcing and underutilization. Now, let us consider a simple situation, where the
demands are 2500, 3000, 4000 in three periods. Now, in a typical aggregate planning
problem, we are now going to find out what is this p how much of production man hours
are we going to use in these three periods such that we minimize the total cost. Now, p
actually has two components which is regular time production and overtime production
and we could outsourcing there are three things by which we can actually produce. Now,
the total demand is 9500 over three periods, now we could do two things.
146
One is we can produce exactly 2,500 in this month produce 3,000 in this month produce
4,000 in this month. If we are able to do that it means there is going to be no inventory or
shortage at the end of this period, alternately we could do for example, the total is 9,500.
So, let us say we could produce about 3,200 here and 3,200 here and 3,100 here or 3100,
3200, 3200 which means we are trying to keep the production quantity same which is
roughly the average of the demand. If we do that the production, quantities are the same,
but there will be inventory at the end of this period and inventory at the end of the other
period sometimes there could even be shortages.
For example, if the 4,000 were here and 2,500 were there, then initially there will be a
shortage or a back order. So, by producing, by making the production quantities equal in
all the periods, we would incur inventory and shortage and therefore, inventory and
shortage costs will dominate in the cost of the aggregate plan. So, that is one type of
strategy to keep the production constant, now to keep the production constant implies
that the number of people who are employed the workforce is also kept constant. On the
other hand, if we produce 2,500, 3,000 and 4,000 every month, which means we are
varying the production quantity, which means we are varying the workforce.
This means we will incur hiring and laying off costs, but we will not incur inventory and
shortage costs that is another strategy. So, organizations can use either of these pure
strategies which is one of the pure strategy is to produce an equal quantity incur
inventory and shortage keep the workforce fixed. The other pure strategy would be to
produce the demand vary the workforce and do not incur inventory or shortage costs.
Otherwise, the organization can also look at what is called a mixed strategy, where we
have a combination of pure strategy as both the pure strategies. Sometimes, we can
adjust or increase the workforce produce more hold inventory and so on.
So, the problem which included all the eight costs the linear programming problem was
essentially trying to identify the best mixed strategy. So, one is the strategy implied a
mixed strategy can be used a mixed strategy and solve it. The other is sometimes there
can be limits or constraints of these which would also make a pure strategy not doable.
For example, in this month the demand is 4,000, now if we are following a pure strategy
where we vary the workforce and we want to produce a 4,000 here. Suppose, the regular
time production capacity is only 2,000 or 3,000, let us say the overtime production
capacity is 500.
147
Then, it would force us to go for outsourcing for another 500, here suppose the regular
time production capacity is 2,000, there is a thousand capacity there is a 500 capacity.
Then, only 3,500 is the maximum that we can reach here. So, we cannot follow the pure
strategy, it also means that somewhere we have to produce store the inventory and then
use it for subsequent periods so that we can meet the demand of the subsequent periods.
So, if we look at this things that we want to do are or things that we can play around are
one is to consider the pure strategies.
The other is to adjust the capacities, now somewhere in the discussion I mentioned that
the regular time capacity if it is 2,000. Let us say then it will be 2000 in all these months,
but can we adjust this regular time capacity during peak periods by hiring additional
workforce one of the models that we saw in the tabular method did exactly that. So, one
is adjust the capacities by adjusting the workforce adjust the capacities by overtime and
outsourcing. So, all these are strategies that organizations used to meet the demand
particularly if the demand shows a lot of variation. Otherwise, organizations will have to
go towards outsourcing and it might be difficult to get the entire product outsourced
certain critical operations. Sometimes, it is outsourced, sometimes common operations
are outsourced while the critical operations are done in house so that certain capacity is
saved by outsourcing this.
148
So, outsourcing can happen in more than one way, most of the times the entire product
can be outsourced only a certain part of it is outsourced, so to that extent that capacity
can be saved and used under regular time capacity or under overtime capacity. So,
organizations look at all of these variable workforce variable or flexible overtimes
sometimes an extra shift, sometimes an extra day and a combination of these two, a bit
of outsourcing. All these are integral part of aggregate planning to try and get as much
capacity as possible to meet the demand, particularly if the demands are fluctuated. Now,
another important thing is that being done is what is called demand planning.
Now, when we worked out the aggregate production plan, we assume that these demands
are known. These demand cannot be changed, so we have to meet this 2500 in month
one, 3,000 in month two and 4,000 in month three and let us try and workout some costs
for this and see what we do. Now, we assume that the regular time capacity is 3,000 per
month, let us assume the overtime capacity is 1,000 and any number can be outsourced.
Now, the best way to try and meet this 2500, 3,000 and 4,000, we will also assume the
regular time cost regular time cost is 100, we will say RT cost is hundred OT cost is 120
and outsourcing cost is higher than this.
So, what we will do is we will assume here that the overtime capacity is 500, so to meet
this 3,000, we will use the 3,000 regular time and 500 over time to make this 3,000, we
will use 3,000 regular time. Now, for this 2500, we will produce 3,000 regular time and
have carry have an inventory of 500, let us say here also we carry an inventory of 500.
So, the total cost associated with this plan will be 9,000 into regular time cost of 100 is
900, 1,000, this 500 overtime would cost us 500 into 120, which is 60,000. This
inventory is 1,000 units, 500 is carried here, 500 is carried here, suppose we say that the
inventory carrying cost is 10 rupees per period.
There is another 10,00, so we have a cost of 900 and 70,000 for this, now if the unit price
is 150 rupees, then, the total money or profit or sale will be 150 into 9,500 which would
give us 14,25,000 and the expenditure is 9,70,000. So, we will have 4,55,000 as the net
profit in business, now the next question is can we kind of plan it better, if we have a
way to work around with these demands and do it, very simple terms if instead of 2500,
3,000 and 4,000 which adds up to 9,500. If the demand itself where 3,200, 3,200 and
3,100, we know for sure that the cost will come down because the amount of the
overtime would be distributed.
149
Here, it will be distributed a little bit here and then the inventory cost can be saved the
500 that we used for overtime will be distributed in evenly, but the cost will be the same.
The inventory cost will come down, therefore the total cost will come down and the total
profit will go up, but what organizations try and do is also to have different ways by
which they can actually adjust this demand. Now, that is called demand planning and
how do they do that they do that by giving some price discounts. So, when price
discounts are given in certain periods say particularly during the lean period the off
season discounts are given.
So, when off season discounts are given the demand which is going to happen later in a
peak period is going to be spread out evenly so that the demand during the lean period
also goes up. So, the peak demand will come down, but the demand gets spread out, but
the total demand also increases a little bit because of the discount that is given. So, if we
look at a situation, where now after say an off season discount the demands for the
periods become 4,00, 4,500 and 4,000. If this is the demand that is given, now what we
will do is we will produce 3,000 regular time.
In each of these we will produce 500 overtime and then we will have to do outsourcing.
So, 3,500, 500 outsourced 1,000 outsourced and 500 outsourced. So, total cost will be a
regular time cost will be 9,00,000 total overtime cost will be 1,500 into 128 and 80,000.
If the outsourcing cost happens to be say 130, then we have 2000 into 130, which is
2,60,000 which gives us a total cost of 1,300 and 40,000 or 13,40,00 as the total cost as
against 900 and 70,000, here the reason being the demand is now higher.
Now, we have to look at the profit, now when we worked out the profit here we said that
the total cost is 150, let us say we give a discount and the total cost becomes rupees 145.
Then, the total money will be 145 multiplied by total demand is 13,500 which give us
1957, 500 and 13,40,000 which gives us 500, 716, 1,700 as the net profit compared to
4,55,000. Now, a quick look at these one will be first tempted to say that this is excellent
mainly because the profit has gone up, but then we also understand the profit went up not
because we only spread the demand, but we also increased the demand. As a result of
this discount of 5 that is given.
So, there is some kind of a loss because the 150 has become 145, but unless the
organization is able to increase the demand number 1 as well as spread the demand if the
150
organization is able to do these two things by offering this discount. Then, it is profitable
for the organization to do this, but on the other hand, if the demand total demand is near
about 9,500, say 10,000 is the total demand instead of 12,500. So, the total cost will be
12,500 into 145, let me do that which is one eight, 1,25,000. So, this is 18,12,500
4,71,500 is the net profit which is still more than the 4,45,000 that we had here.
Now, the organization should be able to it has brought down the cost or price by 145 by
giving a 5 rupees discount on this. Now, if the 5 rupee discount is adequately
compensated by 2 things one is a spread and balance in the demand and the other is the
increase in demand. So, if you are not able to get the increase of 3,000, then it will not be
profitable to give this discount. So, two things have to happen what is the extent of
discount that is given when is the discount when is the discount given. So, these two
things are important, so if the discount is given particularly during the lean period so that
the demand can be brought forward from the peak to the lean the peak would always
mean that we have to increase the capacity.
This means the cost will go up in these calculations we have assumed that this three
thousand is intact sometimes a regular time capacity also can be increased by increasing
the workforce. So, the cost will go up so if we are able to bring the demand the peak
demand and evenly distribute it particularly through the lean periods. It becomes
advantageous for the organization to do this and even gain a little bit of profit if the net
demand goes up because of the discount. Now, this is an important aspect which is called
demand planning and this is something that organizations now have started looking at.
So, it is actually profitable for an organization to give an off season discount provided
the demand for the period increases because of the discount and the peak demand gets
averaged over certain periods. So, these are some of the aspects that one looks at
particularly when one does aggregate planning. So, if we go back to one of the earlier
lectures where we looked at the various topics in production planning and control.
151
We said we would start with forecasting where we would look at forecast of the demand
where we answered the question how much is the demand that we have to produce and
meet. Then, we answered the next question in aggregate planning where, now if this is
the demand that we have to we have to produce and meet how we are going to do it. In
terms of regular time capacity should it be fixed can it vary with the workforce overtime
capacity what are the other strategies is there an outsourcing strategy.
What are the other strategies are we going to hold inventory or are we going to incur
some back ordering, so we looked at all those questions. So, the decision that comes out
of the aggregate planning takes input from forecasting. The output from the aggregate
planning are essentially the P t and sometimes W t w t is the workforce incurred in terms
of man hours is that I am going to employ more importantly P t has the number of man
hours of production that we are going to employ. Now, we need to actually going to do
two things one is how this P t is the number of hours of production going to be
distributed into three things regular time overtime outsourcing which the aggregate
planning has already done.
Let us say we are going to use this many hours of regular time this many hours of
overtime. Now, how are we going to use these hours to produce the various products
which product is going to be produced first what is the order in which we produce these
products. Now, that is answered in what is called disaggregation sometimes called
152
master production scheduling. Now, after we answer this disaggregation, then we need
here at this point we are going to say that well this product is going to be made in next
two periods. So, we also try and find out when these products are going to be made after
we answer that question or sometimes even parallels what are all the things that we make
individual quantities of each of these final products.
Then, these each of these final products has its own set of assemblies and sub assemblies
and bought out components. So, if the certain number of final products are to be made
what is the corresponding number of the brought out items that we should have. Now,
that is answered by what is called the bill of material as well as by what is called MRP
materials requirement planning. This tells us if we are going to produce and make the
demand for this many quantity of each of the products, what are the corresponding
bought out items and in what quantities they are demanded or required.
Then, how are we going to manage our material or inventory buying policies for these
items so that leads us to another topic called inventory management, which we will see in
the next lecture. After we study inventory management, we will go back and look at
disaggregation and then we will look at MRP and then after MRP is studied and the
inventory management is covered. We will then move into production control which is
sequencing and scheduling.
153
Prof. G. Srinivasan
Module - 01
Lecture - 09
Inventory Models - Costs, EOQ Model
In today’s lecture we study inventory models. Inventory essentially deals with materials
the procurement of materials. Materials management is to how we store the material for
subsequent use in every manufacturing organization or manufacturing situation. We buy
material and these material are usually bought as raw material. Sometimes these raw
material undergo manufacturing and become semi finished of finished products. Some of
the bought out items go directly into the assembly of the product. Earlier we have seen
aggregate production planning where the output of the aggregate planning is telling us
how much of production capacity is available for every period.
Later we would also be studying disaggregation models, where the production capacity
that is available in each period is, now allocated to the various products that are made.
So, at the end of disaggregation we would know what are the products that are going to
be made. In order to make these products we need to buy material. So, we are now going
to look at how we buy material. So, inventory essentially answers two very important
154
questions. If there is an item or material that we are going to buy there are two important
decisions that are associated with the purchase of the material.
And these two questions are how much to order and when to order. So, obviously when
the production planning is complete and the allocation of time to the various products are
made one has an idea of how much or how many products of a certain type we are going
to produce. So, the final product, whatever we are going to produce of the final product
of each of the products.
Now, requires certain raw materials or bought out items in a certain number. So, if the
demand of the various products are known. Now, this demand can be suitably multiplied
to find out the requirement of each of these items or material that goes into the final
product. Once we know the annual demand of each of the item that goes into the final
product we need to make decisions on what quantity we are going to order every time we
place an order and when exactly we are going order.
So, inventory control or inventory management essentially deals with these two aspects.
Let us also spend some time on the classification of inventory models. Broadly two types
of classification. One is called single period inventory models and multiple period
inventory models.
Single period inventory models assume that the planning horizon is a single period and
the decision is made only once. Multi period models assume that there are several
planning periods or multiple periods. And decisions are made more than once.
Sometimes these are also called static and dynamic inventory problems. Now, the
important thing that drives the inventory management is the actual demand for the item.
The moment we know that demand then only we can answer these two questions how
much to order every time we order and when to order.
So, now depending on the nature of the demand we can have broadly three classifications
one is called deterministic demand the other is probabilistic demand and within the
probabilistic demand we have two classifications. One is called inventory under risk the
other is called inventory under uncertainty. Deterministic means the annual demand is
known and known with certainty. Risk would mean that the demand is known, but the
demand is not deterministic it follows a certain distribution and the entire distribution is
known. If the distribution of demand is known then it is called problems under risk.
155
If it is assumed that the demand is not deterministic the distribution of demand is not
known but, certain important parameters such as mean and variance of the demand are
known, then it is said to be inventory models under uncertainty. And in this lecture series
on operations and supply chain management we will be looking at multiple period
deterministic models and multiple period with risk.
We would also try and look at some one or two single period models with
nondeterministic demand. So, we will first take the case of the deterministic demand and
then look at some models where the demand is known and deterministic. Then we will
move into probabilistic inventory models particularly covering situations of inventory
under risk. Then as mentioned in inventory problems we basically answer these two
questions which are how much to order and when to order.
One of the reasons we have to answer these two questions is the material that we buy or
the items that we buy for manufacturing are special items which somebody else is
producing and selling to the organization - that is manufacturing this product. So, the
vendor or the person or the supplier who actually provides this material also requires
some time to make this and then sell it.
Therefore, it is necessary to fix on a certain order quantity which is optimal such that the
person who makes it and sells it has the time to do. So, organizations do not for example
buy items the same item everyday traditionally they would. But, a slightly larger quantity
than what was required every day. So, the question of how much to buy or how much to
order came about more recently organizations have moved towards just in time
manufacturing.
And just in time purchasing where organizations tend to buy exactly what is demanded
or they buy for the demand of a very short period of time. Unlike in the older models
where you actually end up fixing an order quantity which is optimized which may result
in the quantity being the demand of a larger time period. We will first look at some
traditional models where we actually try and optimize and try and find out what is the
quantity that we have to buy.
156
Usually this thing how much to order is called the order quantity and is denoted by the
letter Q and when to order is called reorder level and is denoted by r. Now, the second
question is also important because when we place an order to the supplier the supplier
has to make that item and then give it. So, there is a time that the supplier needs to make
it and the variable called when to order is chosen such that when we place the order. And
by the time the order arrives we should have enough material with us to meet the
demand during the period between placing an order and arrival of the order.
Now, this period between placing an order and arrival of the order is called the lead time.
Very broadly defined, lead time is the time between placing an order and receiving it.
Lead time also actually has several components. Sometimes we look at identifying that
an order has to be placed and placing an order between these two certain time elapses
and then the order is placed and the order is met or supplied. And then by the time the
item is supplied and the time the item is put to use certain amount of time elapses.
Sometimes we consider all of them as the lead time, but for the purpose of our discussion
we would assume that the lead time is the time between placing an order and receiving it.
So, the when to order question becomes important because we need to place an order
such that by the time the material arrives or for the period taken between placing the
order. And the arrival of the material which is the lead time we should have sufficient
stock to meet the demand for the item during the lead time. So, generally reorder level or
157
the point at which we place the order is equal to the demand during lead time. So, that
that quantity which is called the reorder level we should be able to meet the demand
during lead time which is called lead time demand.
When the demand is deterministic and the lead time is also deterministic and known with
certainty, then the lead time demand can be computed as the product of lead time and
daily demand. But if one of them is nondeterministic, then we have to look at expected
value of the lead time demand and then try and see whether the reorder level is about the
expected value of the lead time demand or sometimes more than that. We will see those
aspects as we move along as we move towards probabilistic inventory models later.
Now, we will restrict ourselves to deterministic inventory models where the demand is
known with certainty we go back to the other question how much to order quantity there
one other reason. That we do not order the same item everyday for consumption is that
there are other costs associated with inventory. So, obviously when items are bought
there is a price or which is called the cost of the item. So, when we buy a single item we
there is a unit cost associated with that item.
So, if we order a quantity Q and if the unit cost is C then Q into C is the actual worth or
cost of the item that we have to incur when we buy that item. If we assume that the
annual demand is known and this capital d then D into c is the worth of the item or the
money spent on buying the item per year under the assumption of course. That the cost
of the item does not change during the planning period we normally assume that C
remains as C. And it does not increase or there is no effect of inflation on the cost of the
item and so on. But, in addition to the cost of the item there are few other costs that we
incur. Now those are three other types of costs that we incur.
One is called order cost or ordering cost which is denoted by the notation C naught or C
0 or C o order cost which is incurred every time an order is placed for an item. Now, this
cost is different from the actual cost of the item this does not depend or is independent of
the quantity that is being ordered. But, every time there is a decision to place an order or
make an order a certain order cost is incurred order cost has several components some of
which I will list here now. First one is the cost of people or payroll money given to the
people who work in the purchasing areas their salaries have to be factored in. And they
become part of the ordering cost because they order several items day by day.
158
So, they there is a part of the order cost that goes into meeting the salaries of the people
who are involved. Second is there is a transportation cost for the items to come from the
supplier to the organization which has been ordered which is included in with cost of
order. We can assume that there is a onetime transportation cost for bringing the items
from the supplier point to the organization third is there could be inspection costs
because these items that come in have to be inspected.
Sometimes we could have hundred percent inspection. Sometimes there could be random
sampling. But, in some form these incoming goods are inspected and there is a cost
associated with the inspection. Next is however undesirable, we could have a situation
where at the end of the inspection there is a reject. So, the rejected items have to go back
and then have to be replaced, so there is a delay there is a time element as well as there is
a cost element.
So, there could be cost of reject and rework with respect to the supplier side. Fifth one
and the perhaps another important one is called the cost of follow up. For example, an
order is placed and sometimes there could be delays and then the retime is estimated and
known. And just about the time we are going to reach the lead time which means the
item has to come then there is a lot of follow up activity going on. And some of these
follow up activities could even be senior people from this organization visiting the
supplier and making sure that the critical items that are required come in time.
So, there is a cost associated with follow up. Follow up could be over the phone it could
be over other forms which are expensive it would also mean a visit of a person going to
this place and then may be inspecting it then and there and bringing it here. So, there are
these costs that are associated with the cost of ordering. Of course, there is any delay
then each of these delays is going to into result in a higher cost.
So, all these are the components of what is called order cost or ordering cost and this
cost does not depend on – it is assumed that it does not depend on the quantity that is
ordered. For example, if you look at these components one could say that at least this
component could be could depend on the quantity that is ordered. But, then right now we
assume that the order cost is not dependent on the quantity that is ordered and is fixed for
every order.
159
Now, the other important cost that comes - that the organization has to incur after the
item arrives is called inventory holding cost or carrying cost - holding or carrying cost.
Now, this is denoted by C c inventory carrying cost of carrying inventory which is C c.
This also has many components. Now the material arrives and it arrives in a certain
quantity which is Q which we have to find out.
And then as I mentioned when we do the optimization Q, let us say, is a little more than
the daily demand which means that the material that comes is going to be used for a
certain number of days. And then an order is going to be placed the next order is going to
arrive and it is going to be used. So, when this quantity Q comes it has to be saved it has
to be stored it has to be held in control and there is a cost associated with that, now that
cost is our inventory holding cost or inventory carrying cost.
So, the first cost that we can think of is the cost of space to keep these items. So, they
have to be kept in some warehouse or some storage area and stored and kept safe and
secure and the moment we want safety and security. We need people who will guard
these items from theft and keep it safe and secured. So, the salary associated with - salary
of the people who are associated with storing the inventory is another component.
Sometimes along with space you could also have power and other accessories which are
required. So, there will be cost of power that is there in the storage area and other
accessories. Sometimes some of these items may require very specific environment such
as a dust proof environment or sometimes certain temperature which has to be
maintained for some of these items.
These are not very common, but not uncommon either that there could be some items
which require a certain temperature conditions, dust proof conditions and so on. So, we
need cost to do certain other things such as other special requirements. Now, the next
thing that can happen is suppose we order a very large quantity Q. And before we
consume this entire quantity which would say run into several days of consumption for
some reason either the technology changes or there is a better product that comes in and
so on and this item that we have bought can become obsolete and unusable.
So, if such a situation occurs there could be cost of obsolescence. So, I write this as cost
of obsolete items again however undesirable it is, there are times when the items get
160
stolen. And there is a certain pilferage cost that the organization has to occur particularly
when they are metals tend to be easily stolen if they are not properly watched.
So, however small and undesirable it is there could be a cost of pilferage. Now, over and
above all these, is an extremely important cost which is a very integral and a large
component of the cost of carrying. Now, when we buy this item say Q quantity of Q of a
particular item which costs C then Q into C is the cost of the item and the organization
has to spend this money in buying it.
Now, the organization is going to recover the money much later after all these items are
put to use the product is manufactured the product is sold and then the money is
recovered. So, when the organization spends this Q into C to, but this item the
organization actually borrows this money Q into C from some source. So, for the
borrowed money the organization has to pay interest which is called the cost of capital.
So, the most dominant cost of holding the inventory is the cost of capital though I have
written it as the sixth item or the last item this is indeed the most important of the costs.
So, this is also defined as some kind of interest rate that the organization spends which is
given by small I and therefore this C c if we consider all these costs. Then this C c is the
cost of holding the item it comes in. So, many rupees per unit per year or per period
many times it is customary and if we actually workout the calculations for the various
components one would see that this takes a very large proportion of the C c and the rest
of them are actually very small compared to the contribution of this.
So, in a situation where this dominates the rest of them it is also customary to write C c
as i into C where i is the interest percent per year and C is the unit cost of the item. So,
carrying cost can be written in two ways one is it is called C c cost of carrying where all
these costs are included. And there are situations where cost C c is written as i into C
where the interest on the borrowed capital dominates and is taken and the rest of them
are assumed to be negligible compared to the interest rate.
Now there is a fourth cost.
161
Let me write the fourth cost here and the fourth cost is called cost of back order or cost
of shortage. Now, let us assume that we are buying a certain item and based on the
availability of this item. Let us say that we have scheduled the production of a certain
product or a component let us say on a Friday. And let us assume that this item is
expected to come on Thursday. So, that on Friday morning when the production happens
the material is available for production.
And let us assume for some reason we are not getting on Thursday or Friday and when
the Friday morning when the production is about to begin the item has not come it is not
available. Therefore, we will not be able to produce it on Friday say now if the due date
for the item or for the product that is to be made on Friday, which means we will not be
able to deliver it to the customer on Friday.
So, there will be a delay now that delay is ordinarily called as shortage. We have not
been able to meet the due date or not able to meet the customer requirement for want of
material. Now, this can be handled in two ways one is to say that well I am not able to
deliver it to the customer on Friday. At the same time I will carry the unmet demand to
the next day which is say Saturday produce it and give it.
Now, such a situation is called back order where the unmet demand is carried to the next
period produced and then the demand is met. The other is called shortage or also called
lost sale which means if I am not able to meet a certain demand on Friday I have lost it
162
and I will not try to recover it by back ordering it. Very loosely both of them are called
shortage. Shortage has to be specified whether we are looking at a back ordering
situation or on a last sale situation usually when it is not specified it is assumed that it is
back ordering situation. So, there is a cost associated with this if the material does not
come in time, now such a cost is given by the notation C s cost of shortage.
Now, this also has several components the first component particularly if it is a lost sale
is the loss of profit associated with the sale because the sale is not complete because for
want of material. So, there is a loss of profit associated with that one could even go to
the extent of adding something like loss of opportunity. For example through the sale I
might get further business which is lost because I have not met that deadline or due date.
If it is a back ordering which means I carry the demand to the next period I need some
more capacity in the next period to meet the additional demand. So, cost of that capacity.
So, cost additional capacity to make this item in either case whether it is a back order or
a shortage I do not have the material. So, I would not be able to produce, so either I have
to do some rescheduling and produce some other things for which material is available.
So, if we do that then there is a cost of reschedule if we are unable to produce some other
item during the time that we have specified for this item. Then the resource is not
utilized for this period is idle for this period. So, cost of idleness of resource or
underutilization of resource.
Now, let us go back to the situation where we have not been able to produce it on Friday.
But, we try and give it on the next day we produce it on the next day, now let us assume
for a moment that we had we had actually decided scheduled it for Friday. Now, let us
assume that if we were able to make it on Friday then let us say we would have used a
truck to take the finished goods to the customer.
Now, there is a back order we are not able to produce and meet it on Friday. Let us say
we are able to do it only on Saturday, but then we do not want that delay to be carried to
the customer we do not want it to go much later. So, one of the ways we can actually
speed it up is by instead of sending it by road or by truck one could air lift the item and
send it to the customer. So, that would that would result in increased freight.
If we choose to take faster modes of transport which are expensive. So, there could be an
increased freight if in order to meet the due date we try and speed it up by sending it by
163
air instead of sending it by road. Now, we have seen six different aspects of C s, but
there is one most important one just as cost of capital is the most important one in C c
there is very important component of this which is called loss of customer good will.
Now, if there is a delay when the customer is not going to be happy with the
performance of the manufacturing organization. And if this delay continues then
definitely there is a point at which the organization would lose the customer or the
organization starts getting worried. That they are going to lose the customer and there is
a goodwill which is extremely important for the survival of any organization now that
gets hurt and there is a loss of good will.
So, organizations are very concerned about the loss of goodwill that could happen
because of delays and because of material not being available. So, the cost associated
with loosing customer goodwill is an integral part of C s. The only difficulty there is it is
extremely hard to give a numerical value to this one cannot say that it might cost
goodwill so, many rupees per day delayed. Every organization knows that this cost is
high and every organization wishes to treat this cost as high. So, that we do not
compromise on that or we do not take that for granted at the same time almost no
organization can accurately estimate what is the cost of loss of customer goodwill.
So, these are all the components that go into the inventory, now four important costs the
unit cost of the item that is bought the ordering cost the carrying cost. And the back
order or shortage cost which are described using notation C for the cost of the item C
naught or C o which is the ordering cost C c or i into C which is the carrying cost and C
s which is the shortage cost.
Now, the other three costs are different and they are additional costs they are over and
above the actual cost of the item. So, when an organization maintains - buys maintains
inventory the organization actually incurs all these four costs depending on certain
situations. So, the inventory management which talks about how much to order which is
Q which is the first question that we will look at we will, now have to try and optimize
and keep all the four costs to the minimum.
So, the objective will be to try and minimize the sum of all the associated costs. All the
four costs which are there depending on the setting or depending on the assumptions the
decision variable is how much to order which is Q. So, with this we will try and look at
164
some very basic inventory models and then extend these inventory models to slightly
more advanced situations.
So, the first inventory model that we will look at which we call as model one is called
single item continuous demand and instantaneous replenishment. Now, let us assume that
the annual demand for this item is D per year and is known. Now, let Q be the ordering
quantity that we decide to order every time we place an order. So, let us look at how the
inventory for this item behaves with respect to time.
165
So, let us draw a simple graph which is like this is time and let us assume that we have
just placed an order for Q. And we have just instantly received it instantaneous
replenishment would mean that as soon as you place the order the item arrives which
also means that lead time is 0. So, the moment the order is placed the items are received.
So, let us assume that we have received Q and we are consuming or using this Q. Now,
when we say continuous demand we assume that the demand is there the same amount of
demand is there at every instance of time. Therefore, as time goes on the demand D is
met is there for every instance of time it is met and the stock of the item comes from Q.
It is consumed and comes till it reaches 0, now this slope is D which is actually the
demand for the item to put it in simple terms more understandable terms. It will be like
saying if the annual demand is ten thousand and let us say if the company works for two
hundred and fifty days.
Then the daily demand is forty if the company works for eight hours a day then the
hourly demand is five if the hourly demand is five. The demand per minute is five by
sixty and demand per second is that divided by sixty and so on and the demand is the
same. That is that is how we should look at this continuous demand which means at
every instance of time there is a demand. And since that demand is same which is D per
year, now this will come as a straight line from Q to 0.
Now, the moment this stock position reaches zero because we have instantaneous
replenishment and we also assume no shortage. We do not want to model shortage
situations no shortage since we do not want to model shortage as soon as the demand
reaches as soon as the inventory position reaches zero. I will call this as stock or
inventory position on the y axis time on the x axis. So, as soon as the inventory position
reaches zero I place an order I place an order for another Q and there is an instantaneous
replenishment of that. So, as soon as I place an order here my stock position goes up to
this Q.
Now once again the cycle continues and then I start consuming this till I reach this. Then
again I place an order and then I start consuming. So, this is how the inventory position
will behave please note that it is the same Q which is here as well as here and. So, one
now let us call time period as capital T, now let us derive a very simple formula to find
out Q, now every time we place an order we order a quantity Q. So, number of orders per
166
year will be D by Q because every time we place an order we order for Q number of
orders per year is d by Q total order cost is equal to D by Q into C naught.
Now, C naught is this order cost. So, every time we place an order we incur a C naught
and this C naught has a unit of rupees per order. So, D by Q is the number of orders per
year C naught is rupees per order. So, D by Q into C naught will have a unit of rupees
per year spent in order cost, so this cost we have considered. Now, we have to consider
this cost holding cost or inventory cost now holding cost or inventory cost is defined as C
c and this has a unit called rupees per unit per year.
So, either year, or the same time period that we are taking about since we are looking at
year here we would also define this as year rupees per unit per year. So, this is the cost of
holding one unit rupees per unit per year. So, what is the average inventory that we have
at any point in time multiplied by C c will give us the holding cost per year and the
average inventory should be in terms of number of units that we have.
Now, there are two ways of calculating the average inventory. So, average inventory is
like saying if I start this cycle with Q then it comes down it comes down comes down to
zero at time t. So, in a particular cycle the beginning inventory is Q the ending inventory
is zero the average inventory is Q by 2 because you actually have a line here that is one
way of saying that the average inventory is Q by 2. Other way is to say that to take any
period of time find out the total inventory that you have divide it by time to get the
average inventory.
So, if you take one particular cycle the total inventory that we have held during this cycle
is the area of this triangle which is half into base into height, so half into Q into T. Now,
that much inventory is held for a period T, so half into Q into T divided by T which
gives us half into Q. So, average inventory either way is Q by 2 units average inventory
is the kind of inventory that we have at any point in time averaged over a reasonable
value. Now, the cost of holding this inventory is called total holding cost is equal to Q by
2 into C c cost of holding this also has rupees per year because this is rupees per unit per
year this is units. So, it becomes rupees per year the units get cancelled and you get
rupees per year.
Now, we look at this cost which is back order or shortage cost one of the assumptions is
that there is no shortage. So, we do not incur shortage in this particular model which
167
means the inventory position does not become negative. So, there is no shortage
assumption in this model therefore, we will not consider back order or shortage cost in
our total cost.
Now, there is this cost of the item, so cost of the item is equal to D into C D is the annual
demand C is the unit cost of the item. So, D into C rupees per year, so total cost total
annual cost T C which is the sum of the order cost the holding cost as well as the cost of
the item is given by d by Q C naught plus Q by 2 C c plus D into C. Now, in this total
cost expression as I said there are three costs cost of the item cost of ordering cost of
holding the inventory. There is no shortage or back order cost when we want to find the
variable as Q rest are all the things are known they are known parameters. So, only thing
that we need to do is to find out Q, now this total cost is a non-linear function of Q
because you find Q in the denominator here you find Q in the numerator here.
It is a single variable non-linear function, so differentiating with respect to Q and setting

it to zero will give the first derivative equal to zero will give the optimum value of Q.
Whether it is maximized or minimized will depend on the second derivative. Now, our
objective is to minimize this T C, so setting it first derivative to zero would give us d T C
by d Q equal to zero will give us minus D by Q square C naught plus C c by 2 equal to 0.
The third term is D into C which is a constant it does not contain Q. Therefore, its
derivative will be 0, so a first derivative equal to 0 would give us minus D by Q square C
naught plus C c by 2 equal to 0 which on simplification would give us Q equal to root of
2 D C naught by C c.
Now, actually speaking Q is a square root of something, so Q should be plus value or

minus value mathematically. But, since Q is the quantity that is ordered it cannot take a
negative value, therefore it is customary to write Q equal to root of 2 D C naught by C c.
And write only the positive value we do not write plus minus root of 2 D C naught by C
c only the positive value is taken; now we have to find out whether at this value of Q the
original function is a maximum or a minimum. Now, we are interested in minimizing the
total cost, so a second derivative will be this is minus D by Q square C naught. This will
become a constant, now there is a negative term and a Q square here.
168
So, on further differentiation we will get a positive term here and Q cubed will come in
the denominator. So, for a positive value of root over 2 D C naught by C c the second
derivative will be positive indicating that this quantity is actually the minimum quantity.
Now, this quantity Q is called the economic order quantity which also are called E O Q
for short. Now, the formula that we have derived is called the E O Q formula for a single
item and this formula is also called the Wilsons E O Q formula attributed to Wilson who
derived it somewhere I think in the early forties or thirties so on. Sometimes this formula
is also credited the credit goes to Harris it is also called Harris inventory formula. And
the literature talks about the contributions of both Wilson and Harris who supposedly
have independently derived or thought about the use of such a formula, to find out the
economic order quantity. Or the order quantity that minimizes the sum of the ordering
cost as well as in the carrying cost.
So, out of the four costs as I said we did not consider the back order or shortage cost we
also observe that the three remaining costs that we have got this is the order cost
component this is the inventory holding cost component. This is the item cost component
we now realise that the item cost actually does not play a part in the economic order
quantity. So, it is also a customary to leave out the item cost and derive Q simply by
considering the order cost. As well as the carrying cost alone and not including the item
cost because item cost being a constant is not going to contribute to the derivation of the
economic order quantity. So, there are only two costs which are optimized which are the
order cost as well as the carrying cost.
So, in this formula of Q equal to plus root of 2 D C naught by C c. D is known which is

the annual demand. C naught is known which is rupees per order C c is rupees per unit
per year what is very important is that D is annual demand which is demand per year.
Then C c should be defined as rupees per unit per year if D is defined as month then C c
should also be defined as, so many rupees per unit per month. So, D and C c should have
the same time period in their unit of measurement. So, if D C naught and C c are known
then the economic order quantity can be calculated. And then we can substitute this Q
into the total cost function to find out the minimum total cost that the organization would
incur. In the next lecture we will consider numerical examples and we explore a little
further on inventory models.
169
Prof. G. Srinivasan
Lecture - 10
Inventory - EOQ Model Graphs, with Backordering
In the previous lecture, we introduced the basic economic order quantity model, whose
stock verses time relationship is shown in this figure.
So, Q is the economic order quantity or the quantity that we order, every time we place
an order. So, the model works as follows, let us assume that, we have just received Q
units that has been ordered earlier. We also assume instantaneous replenishment which
means, as soon as the order replaced, the items arrive. And then we consume these items
at the rate of D till it reaches, the stock position reaches 0, and then we place another
order and instantly get this Q and this process goes on forever.
170
We also derived an expression for Q and we derived that Q is equal to root over 2 D C
naught by C c, where D is the annual demand, C naught is the order cost or cost of
placing an order and C c is the cost of holding or carrying inventor. This formula was
derived when we tried to minimize this expression; total cost is equal to D by Q into C
naught plus Q by 2 into C c. Now, this total cost is the total annual cost of inventory and
in this model, the two relevant costs are the ordering cost and the inventory carrying cost.
We do not consider shortage cost, because one of the assumptions is, no shortage or back
order is allowed which means that, this stock position will come right up to 0, it will not
go beyond that. So, it will come only up to 0, so no shortage cost is allowed; we may or
may not include the actual cost of the item. We saw in the previous lecture that, the
actual cost of the item is D into C per year, this is per year or per given time period. So,
it will be D into C if it is per year and it does not depend on Q, the ordering quantity.
Therefore, when we differentiate this, the term D into C does not contribute anything and
it is customary, therefore to leave out that term, the term that involves the actual cost of
the item and consider only the ordering cost and carrying cost and then optimize it to get
this. Now, let us workout a numerical example to understand a few more things about
this formula and the use of this formula.
Now, let us take an example, where D which is the annual demand of the item is say,
10000 per year, C naught which is the ordering cost is say, rupees 300 per order, C
171
which is the unit price of the item let us say, is rupees 20 per unit and i is the interest
rate, which let us say is 20 percent. So, C c will now be defined as, i into C will be 20
percent of rupees 20, which is rupees 4 per unit per year. Now, when we apply this
formula, we are now given D, C naught and C c, 10000, 300 and 4.
So, we apply this formula to get Q is equal to root over 2 into 10000 into 300 by 4, which
on simplification would give us 1224.74 units. We have already seen that, this Q when
we derived it, Q is root over 2 D C naught by C c. We take only the positive quantity, we
do not take the negative quantity, because Q represents an order quantity, so you get
1224.74 units as the economic order quantity in this case. Now, this means that,
whenever we order, we will order for 1224.74 units to begin with, of course we will have
questions like, how can I order for a fraction of a unit and so on.
We will answer those questions as we move along, but right now based on this formula,
we say that, every time we place an order, the order for 1224.74 units. Now, let us also
try and find out the cost associated with this, so we need to substitute Q equal to 1224.74
into this, in addition to D C naught and C c, to get this total cost. Now, instead of
substituting 1224.74, let us try and substitute this value, which is the value of the
economic order quantity. This equation for Q into this and compute this total cost as a
function that is, that does not directly use Q, let us compute this in terms of D, C naught
and C c, so let us go back and substitute.
172
Now, total cost TC is equal to D by Q into C naught plus Q by 2 into C c, now Q is given
by root of 2 D C naught by C c. So, on substitution, we will get D C naught by root over
2 D C naught C c plus root over 2 D C naught by 2 into C c. So, this on simplification
would give us root over D C naught C c divided by root 2 plus root over D C naught Q
by 2 into C c. So, Q is root over 2 D C naught C c, Q by 2, so this 2 comes here, this C c
comes here.
So, on simplification, this root over C c and this C c will cancel to give another root over
C c, so root over D C naught C c divided by root 2, because there is a root 2 here, there is
a 2 here, so we get this expression. This on further simplification will give you 2 times
root of D C naught C c by root 2, which will give us root over 2 D C naught C c. So, we
can now write the total cost in terms of only the known parameters, which are D, C
naught and C c and not in terms of the computed number, which is Q.
So, the total cost at the optimum, the minimum total cost that we will incur is root of 2
into D into C naught into C c. And in our example, when we substitute, we will get root
over 2 into 10000 into 300 into 4, which on simplification will give us 4898.98 rupees
per year. Now, this quantity is does not include the actual cost of the item, the actual cost
of the item, if we take 1 year period will be 10000 is the demand per year, say 20 is the
unit price, so 200000 will be the actual cost of the item.
So, if we say that, the total inventory cost is 4898.98 it means, we are only giving the
order cost plus carrying cost for the year and we have not included the actual cost of the
item. If we said 204898.98 it means, we have also added the actual cost of the item per
year. Now, from this formula, we also observe an very interesting thing which is, at the
optimum that is, when we substitute root over 2 D C naught by C c, at the optimum the
component of the order cost and the component of the carrying cost are equal.
So, when we do this optimization to find out the value of the economic order quantity,
we are essentially trying to find out that value of Q, for which the order cost is equal to
carrying cost. Therefore, the total cost is 2 times the order cost and carrying cost, which
comes to this expression. Now, let us also try and draw a graph that kind of depicts or
shows the economic order quantity. So, if we look at, now cost versus order quantity,
now the order cost is of the form D by Q into C naught, D and C naught are known, Q is
the unknown, so it is constant divided by Q.
173
So, when we draw this, we will get, the curve will be like this, this will be the order cost
curve, it will be a rectangular hyperbola, which will show this D by Q into C naught, this
is the order cost, carrying cost is 2 by 2 into C c, which is constant into Q. So, it is a line,
which passes through the origin, so the carrying cost will be like this. Now, we have
already said that, the economic order quantity is the point, where the order cost is equal
to the carrying cost, which we got from here, they are equal.
So, this is the economic order quantity, where the two curves intersect, this I call as Q
star, which is the economic order quantity. You can write Q equal to or you can write Q
star equal to, the best value of Q is called Q star and this is the place, where the total cost
is also minimized, because we are substituting this, which is the minimum value into the
total cost function. So, we will have, these two are equal, so this is the point, so the total
cost curve will actually look like this, so this is how the total cost curve will look like.
Now, this is the TC star, this is the root of 2 D C naught into C c, so the total cost is 2
times this and it happens for the corresponding Q star. So, this is how the cost be, this is
C c carrying cost, this is TC equal to order cost plus carrying cost. Now, let us try and
answer a few simple questions, one is when we did this calculation, we have now found
out that, the economic order quantity is 1224.74. So, the first question that we would ask
is, are we justified in treating this Q, order quantity as a continuous variable?
When we differentiated this and then we set it to 0, when we set the first derivative to 0
to find the minimum, we assume that, the Q is continuous. Now, let us let us assume that,
we took Q as a continuous variable, a assumption of Q being a continuous variable was
made considering the ease of the mathematical derivation for it. If Q were to be defined
as an integer variable then the methodology to get it would be different. Since for
convenience, we define Q as a continuous variable, we could easily set the first
derivative to 0 and get the value.
And show that, the second derivative is positive indicating that, this value is indeed a
minimum for the plus value of Q. Now, because of that, we also assume to get a number
like 1224.74, which poses the question, how can I order a fractional quantity of the item?
So, the simplest thing to do perhaps is to try and round it off to the nearest number,
which could be 1225, let us say. And let us see, what happens to the total cost, whether it
174
is very, very different from the absolute minimum, if we look at 1225. So, let us try and
do some quick computations.
So, when Q is equal to 1225, TC is equal to - all we need to do is to substitute 1225 here.
So, TC will become 10000 into 300 by 1225 plus 1225 divided by 2 into 4, which will
become... Now, when we substitute Q equal to 1225, which is the upper integer value or
the closest integer value to 1224.74 and substitute 1225 here as well as here, the total
cost is, it remains at 4898.98, which is the absolute minimum. So, between 1224.74 and
1225, there is absolutely no loss, so instead of implementing 1224.74, we could
implement 1225 if we have to keep Q as an integer.
Then, comes a next question, now can my vendor supply 1225, suppose my vendor is
willing to supply only in multiples of 100 then I may have to consider 1200 or I may
have to consider 1300, whichever is minimum. So, let us try and evaluate the TC for Q
equal to 1200, now TC will become 10000 into 300 divided by 1200. So, 300 1200 is 4,
so 10000 by 4 is 2500 plus here we have 1200 by 2 is 600, 600 into 4 is 2400, is equal to
4900.
Now, here I am specifically showing the order cost component and the carrying cost
component. Now, if Q equal to 1300, TC is equal to 10000 into 300 divided by 1300
becomes 4907.69. Let us just look at one more calculation or two more calculations, let
us say the vendor can give only 2000 and the vendor is giving at 4000. Whereas, the
175
vendor is giving in multiples of 1000 let us say then we are looking at say three numbers,
Q equal to 1000, Q equal to 2000, let us say Q equal to 4000.
Now, at Q equal to 1000 which means, the vendor is giving in multiples of 1000, here
10000 by 1000, which is 10, 10 into 300 which is 3000 plus, 1000 divided by 2 is 500,
500 into 4 is 2000, 5000. Whereas, the vendor will give us only in 2000 then this is
10000 divided by 2000, which is 5, 5 into 300 is 1500, 2000 divided by 2 is 1000, 1000
into 4 is 4000, which is 5500. Now, if the vendor would give us only 4000 then it
becomes 4000 10000 into 300 divided by 4000. So, this is two and a half, two and a half
into 300, which is 750, plus 4000 divided by 2 is 2000, 2000 into 4 is 8750.
Now, let us try and understand a few things that we have calculated here and let us try
and show some of these into this graph and try to understand little bit from this graph.
Now, the first thing we did was to round this off to the nearest numbers, we tried 1225,
when we realized that, it just is the same, the total cost is the same. So, if the supplier can
give in multiples of 1 or any integer quantity then ordering 1225 is indeed economical.
So, first let us understand that, we need not really worry about the fact that, we have a
fractional number or a decimal number coming here.
It can be a rounded off to the next highest integer and there is absolutely no change in the
total cost. Then we start looking at multiples of 100 then we looked at the two closest
numbers, which are 1200 and 1300. And we actually realize that, for 1200, the total cost
was 4900, which is hardly a rupee more than the earlier one and it is absolutely
negligible. The increase is absolutely negligible, not even of the order of 0.1 percent; it is
much less than that.
So, one can use 1200 and even if we round it off to the higher 100, which is 1300, which
is an increase of about 75. We still realize that, the total cost has just become 4907,
which is somewhat like 8 rupees more than this. So, the increase is 8 rupees on 5000,
which is 1.6 rupees on 1000 and 0.16 rupees on 100, so it is 0.16 percent increase. So,
there is 0.16 percent increase if there is 75 by 1200 into 100, which is roughly about 6
percent increase.
176
In the order quantity, is giving us 0.16 percent of increase in the total cost, which also
shows somewhere that, the total cost curve is kind of flat near the optimum, which is
what we have tried to show here, it is quite flat here at this zone. So, even if you kind of
play around with Q roughly in this area, the TC, it does not increase very significantly.
For example, even if we go to 1000 and 2000, which is the two nearest 1000s, if you
want to approximate this, you realize that for 1000, the total cost is 5000, which is about
100 rupees increase.
But, for 2000, it shows about 600 rupees increase, which is a little significant, we are
now talking of the order of 10 percent here, here you are still not talking of the order of
10 percent, you are still talking of the order of 2 percent here. And if you really increase
it to 4000, this is where you are talking of a very, very large increase of nearly 70, 80
percent. So, as long as this Q is somewhere here, as long as any ordering quantity Q is
near Q star, up to say 10 percent on either side of Q star, you realize that it is very flat
near the minimum and therefore, the effect of increase in cost is not very high.
Only when you move the ordering quantity Q significantly away from this, either this
side or this side, you realize that the total cost is very high. For example, if this is
1224.74, so let us say 4000 is somewhere here and therefore, we saw there it went up,
total cost went up. Now, when it is 4000, you also realize that, total cost went up, the
177
order cost is small, the carrying cost is large, so you see there the order cost is small
carrying cost is large.
As you move to the right of Q star, the carrying cost is going to increase; the order cost
is going to come down. For example, if you look at 1300, order cost is coming down
carrying cost is going up, 2000, order cost is coming down carrying cost is going up, less
than that 4000, order cost is coming down carrying cost is going up more. So, if the
order quantity is less then Q star like 1000, see the order cost is going up and the
carrying cost is coming down.
So, to the left of it if you realize, so somewhere here if we are then you realize that the
order cost part is higher the carrying cost part is lower, but the total cost will always be
higher, because this is the minimum. So, the one of the important lessons from the
economic order quantity formula is the realization, that it is very flat at the minimum and
therefore, one need not worry about the decimal or the fraction coming in here. One can
conveniently round it off to the nearest 100 and still not incur a significantly high cost.
Next thing that we need to look at is, if we are ordering 1224.74 units, every time we
place an order. Now, in one year, we will be ordering 10000 divided by 1224.74.
Number of orders n is equal to 10000 divided by 1224.74, this 10000 is the annual
demand, which is shown here. So, I have a demand of 10000, every time I place an order
for 1224.74, therefore I have 10000 divided by 1224.74, that many orders per year. Now,
178
this on simplification would give us something like an 8.17 orders per year. Now, this
rises another question, whether the number of orders per year should it be an integer or
should it not be an integer.
The easier thing to do is to assume that, let it be an integer, so I order 8 times in a year,
so roughly I say, I order every one and a half months. So, one and a half into 8 is 12. So,
I order roughly every one and a half months and it is also convenient that, if we make
this as 8 orders per year, the order quantity would become 10000 divided by 8, which is
1250, instead of 1224.74, 1250 is a much more convenient number. And we have already
seen here that, whether it is 1200 or 1300, we are still in the border of 4900 to 4907,
which is hardly 10 rupees on 4800 or 5 rupees.
So, it is always possible to try and have an integer number here, more for the sake of
convenience. And when we have an integer number here and if this is a very comfortable
number like 10000 then the order quantity also becomes a very comfortable number or
even if it does not become a comfortable number or it is a fraction, it can always be
rounded off to the nearest 100 on either side and so on. As such there is no sanctity about
the fact that, this has to be an integer, need not be an integer.
It simply means that, if it is 8.17 orders per year and if the year comprises of 365 then it
is like 47 days I place an order. So, every 47 days or every 46 days I place an order, plus
minus 1 day does not matter at all, the order quantity can be adjusted suitably. But, what
we learned from this is that, even though the actual results of the economic order
quantity like 1224, 4898, sometimes or many times they may not be implementable
exactly as they are.
But, the important learning is that, we need to fix our order quantities as close to the
economic order quantity as possible, subject to other constraints and considerations so
that, the actual cost that we incur is still around this 4898.98. It could be this, it could be
(Refer Slide Time: 30:14) this, sometimes it could be this, sometimes it could even be
this, but it should not be this. So, if right now we are having a vendor, who is giving us
only 4000, because of which we are incurring an 8750.
If we do the economic order quantity and understand that, it is somewhere near 5000, the
next thing we need to do is to see, whether our vendor can give this 1000 so that, we get
a cost of 5000. Otherwise, we should try another vendor, who can give 1000 and spend
179
an additional 5000 per year on this item, rather than by 4000 at a time and spend 8750 on
that. This is the most important learning from the economic order quantity formula; now
let us move to the next model.
The second model is very similar to the first model, except for relaxing one assumption,
in the earlier model or the first model; we did not allow shortages or backordering. Now,
we are trying to see, at least to begin with mathematically, what happens when we allow
the backordering. We defined four costs, we used only three of them, we also showed
that amongst the three, one of them does not affect the decision. The actual cost of the
item does not affect the decision; we left out the backorder stroke shortage cost, because
of the assumption, that there is no backorder.
Now, let us try and allow some backordering at an additional cost and see, how this
model behaves. So, what we do now is, let us assume that, we start somewhere here, so
we consume, we reach the point where the stock position is 0. We also assume
instantaneous replenishment which means, if I place an order here, I am going to get it
immediately, which is what we did in the earlier model. Now, we are going to allow a
little bit of backordering, so we do not place an order here.
We allow backordering, we build up some backorders then we place an order for Q, for
an order quantity Q, which is instantly replenished. So, it is replenished and let us say,
after the replenishment, it reaches this point, the point where we started. So, once again
180
we consume at the rate of D, like what we did, once again build some backorder and at
this point, you place the order for Q and get it replenished instantly and then it proceeds.
So now, this is time, this is quantity, now we are going to use some additional notation,
now this part, this part the quantity that we order here.
This is Q, so Q is starts from here. This is what we order. So this is our Q, the order
quantity. Now, this part, which is the backorder part is called small s, which is the
backorder and the level to which the inventory gets build up, it is called I m maximum
inventory, which is here. So obviously, Q is equal to I m plus s. Let us say this is a cycle,
which we call as capital T and the cycle repeats. Now, this cycle now has two parts, this
part is called T 1, where there is inventory and there is this part called T 2, where there is
back order and T 1 plus T 2 is equal to T.
So now, there are four costs that we will have, so let the economic order quantity be Q or
the quantity that we order is Q. So, the order cost component, so the total cost has, TC
has all the four costs, order cost plus carrying cost plus shortage cost plus item cost.
Now, Q is the variable which is the order quantity, I m is a variable, s is a variable, but
they are related, Q equal to I m plus s. Now, if we are going to compute this total cost for
a year then the item cost is going to be D into C.
So, the item cost will not contain any of the variables Q, I m or s, so like we did in the
earlier model, we leave out the item cost. So, we are going to have three costs in this
model, we had two costs in the earlier model. So, if D is the annual demand and Q is the
quantity that we order, the number of orders per year will be D by Q, like we did in the
previous model. The order cost or cost for order is C naught, so D by Q into C naught is
our total order cost per year.
Now, we need to look at the total carrying cost per year, the broad or general definition is
average inventory into the inventory carrying cost. Earlier it was Q by 2 into C c, now
we have to compute the value of the average inventory. Then if we take any cycle, say
cycle like this, the total inventory that we hold in that cycle is area under this curve,
which is half into I m into t 1. So, this much inventory is held for a period t 1 in the cycle
and this is the backorder part of the cycle, so there is no inventory held.
So, zero inventory is held for a period t 2, the average inventory of I m by 2 just as we
got Q by 2 in the earlier, the average inventory of I m by 2 is held for a period t 1. An
181
average inventory of 0 is held for a period t 2, therefore the average inventory as such, at
any point will be I m by 2 into t 1 by t 1 plus t 2. Let me repeat again, total inventory
held in a cycle is area of this triangle, half into I m into t 1. So, that much total inventory
is held for a total period of t 1 and 0 is held for a period of t 2.
So, average inventory is I m by 2 into t 1 plus 0 into t 2 by t 1 plus t 2, which simplifies

to I m by 2 into t 1 by t 1 plus t 2, this is the average inventory, this into the inventory
holding cost, which is C c. Similar manner, the average shortage or backorder is,
backorder is 0 for a period t 1, average backorder of s by 2 for a period t 2. So, 0 into t 1
plus s by 2 into t 2 divided by t 1 plus t 2, which on simplification would give s into t 2
by t 1 plus t 2 into the backorder cost, which is defined as C s.
So, let us now define C s, which is the back order cost and this is right now defined with
the same units as C c. So, this will be rupees per unit per year, C s and C c will have the
same units, rupees per unit per year.
182
Note the deference, in the earlier model this was Q by 2 into C c, now it is I m into t 1 by
2 into divided by t 1 plus t 2 and so on. Basic motivation is that, if you take a cycle, part
of the cycle which is t 1, is the time where we hold inventory and the remaining part of
cycle which is t 2, is where the backorder is there. So, the inventory portion will contain
the area under the inventory curve I m by 2 into t1 for a period t 1 plus t 2, s by 2 into t 2
for a period t 1 plus t2.
So now, this expression for TC has several variables now, it has Q which is unknown, it
has I m which is unknown, it has t 1 that is unknown, t 2 that is unknown and s that is
unknown, but the unknown variables have some relationships. We have already saw that,
I m plus s is equal to Q and t 1 plus t 2 is equal to some T, that we have not used. So, we
try and simplify this a little bit using some similar triangle property. Now, t 1 by t 1 plus
t 2 from similar triangles, is equal to I m by I m plus s, t 1 by t 1 plus t2 will be I m
divided by I m plus s.
But, I m plus s is Q, so this will become D by Q C naught plus I m by 2 into t 1 by t1

plus t 2 is I m by I m plus s, which is I m by Q. So, this will become I m square by 2 Q
into C c. Similarly, t2 by t 1 plus t2 is equal to s by s plus I m, this will give t 2 by t 1
plus Tt 2 is equal to s by Q. So, on simplification, this will give s square by there is a 2
that I have missed, here s by 2, so s square by 2 Q into C s.
183
Let me explain it again, when we calculated this portion, which is the backorder cost
portion, average backorder in a cycle is, for t 1 part of the cycle, there is no backorder,
for this part of the cycle, the total backorder is area under the curve half into s into t 2.
So, the average back order at any point in time is half into s into t 2 divided by t 1 plus t
2. So, half into s into t 2 divided by t 1 plus t 2, this multiplied by the backorder cost is C
s.
So, on simplification, t 2 by t 1 plus t 2, t 2 by t 1 plus t2 is equal to s by I m plus s,

which is s by Q. So, this portion becomes s by Q, so this becomes s square by 2 Q into C
s. So now, we are eliminated the t 1 and t 2 from this, because they are also dependent
on, they are related to I m and s, so we have eliminated these. Now, we have only three
unknowns, which are Q, I m and s and we also know that they are related, because Q is
equal to ii m plus s.
So, what we do now is, we eliminate I m by substituting I m is equal to Q minus s to get

D by Q into C naught plus Q minus s the whole square by 2 Q into C c plus s square by 2
Q into C s. So, this way we have now eliminated I m also by substituting I m is equal to
Q minus s, so we have only two unknowns, which are Q and s. The two unknowns are,
how much to order, which is the order quantity and what is the level of backorder, which
is the back order level, so Q and s are the two unknowns.
And now, this expression does not have any dependency, all the dependencies have been
taken care of by the substitution using similar triangles as well as using Q is equal to I m
plus s. So now, we can differentiate this partially with respect to the two variables Q and
s, to try and get the optimal values of Q and s. So, let us do that, so first let us
differentiate this with respect to Q and then we differentiate this with respect to s.
184
So, dou TC by dou Q equal to 0 gives partially differentiating total cost with respect to Q
would give us, this would give us the term minus D by Q square into C naught, which is
a same term that we got in the earlier model. So, minus D by Q square into C naught
plus, now this is a term where Q appears in the numerator as well as in the denominator.
So, here we need to use the u by v differentiation, so v square will come in the
denominator v d u minus u d v by v square.
So, we take this C c by 2 outside, so plus C c by 2, u by v will give us v d u minus u d v

by v square. So, we put a Q square here, v d u, so this is Q into differentiation of this. So,
Q into 2 times Q minus s is what we get, 2 times Q minus s into 1, we differentiate this Q
you get 1, so v d u minus u d v, Q minus s the whole square into differentiation of this Q
which is 1. So, minus Q minus s the whole square is what we have here, this term Q
appears only in the denominator, so it is easy to differentiate, minus s by 2 Q square into
C s equal to 0.
This Q will give minus 1 by Q square, so minus 2 Q square s square C s v equal to 0,

partially differentiating, so s square will remain as s square, s square C s is 0. Now,
partially differentiating with respect to s dou TC by dou s equal to 0 would give us, this
term will not contribute anything, because there is no s term here, so this is a constant as
particular as partial derivative is concerned.
185
So, this term will not contribute anything, this term has s only in the numerator, so this
will become 2 times Q minus s into minus 1 divided by 2 Q square C c plus 2 s square by
2 Q, s square on differentiation would give us 2 s. So, 2 s by 2 Q into C s is equal to 0, it
may repeat, s appears in the numerator, so 2 times Q minus s into minus 1 divided by 2 Q
into C c, 2 s by 2 Q into C s equal to 0. So, this 2 and this 2 will get cancelled, when you
take this Q to the other side it goes, so we have minus Q plus s into C c plus s C s equal
to 0 or let me put it or let me simplify it differently.
I will simply take this term to the other side, so Q minus s into C c is equal to s C s, so Q
C c is equal to s into C c plus C s, from which s is equal to Q C c by C c plus C s. So, this
is the first thing that we derive, we have this, now we go back to this one and then we
have to use s is equal to Q C c by C c plus C s on to this one, to try and get the value for
Q. We will do that now, now I multiply this by 2 so that, the 2 Q square gets canceled, so
I will have minus 2 D C naught plus C c into, this is 2 Q square minus 2 Q s minus Q
square minus s square plus 2 Q s minus s square C s equal to 0.
I just multiplied this by 2 so that, I have this expression, now here I can cancel this plus 2
Q s and minus 2 Q s, plus 2 Q square minus Q square will give me Q square, so minus 2
D C naught plus C c into Q square minus s square, minus s square C s equal to 0. Now,
this will become minus 2 D C naught plus Q square C c minus s square into C c plus C s
equal to 0.
Now, we substitute for s square from here, minus 2 D C naught plus Q square C c minus
s square is Q square C c square divided by C c plus C s the whole square into C c plus C
s equal to 0. Now, this will go and this will remain, now this will simplify to minus 2 D
C naught.
186
Take this here, only here plus Q square C c into C c plus C s minus Q square C c square
by C c plus C s is equal to 0, I am leaving this as it is, I am taking this only up to this.
Now, minus 2 D C naught plus Q square C c square plus Q square C c C s minus Q
square C c square divided by C c plus C s equal to 0. Now, these 2 terms will get
canceled, so minus 2 D C naught plus Q square C c C s by C c plus C s equal to 0, from
which Q square C c C s by or Q square C c C s is equal to 2 D C naught into C c plus C
s, from which Q is equal to root over 2 D C naught into C c plus C s divided by C c into
C s.
So, when we use this model, where backorder is allowed and then we derive, we get the
expression s, the backorder quantity is Q C c by C c plus C s, the order quantity Q is 2 D
C naught by C c C s into C c plus C s. Now, we can substitute this s and Q back into the
total cost function, which is here to try and find out, what is the total cost at the optimum,
that and comparison of the model with backordering. Comparing a model with back
ordering to a model without back ordering which means, the similarities and differences
between models 1 and 2, we will see in the next lecture.
187
Prof. G. Srinivasan
Indian of Technology, Madras
Module - 1
Lecture - 11
Inventory - Models for all Quantity and Marginal Quantity Discount
We continue our discussion on inventory models. We have so far seen two models. This
is the diagram for the first model, where we consider continuous demand and no back
ordering. In the second model, we considered back ordering and we derived expressions
for the economic order quantity under both the assumptions. The economic order
quantity for this model, without back ordering is given by Q is equal to root over 2DC
naught divided by C c; where D is the annual demand, C naught is the order cost and C c
is the unit holding or carrying cost.
Here, there are two variables, one is the order quantity Q which is actually this quantity
and then there is a back ordered quantity which is called s. There are two variables. Then
the formula for economic order quantity Q, under the assumption of back ordering with
the back order cost of C s is given by Q is equal to 2DC naught divided by C c into C c
plus C s divided by C s, and s which is the back order level given by QC c by C c plus C
s. Both from this expression as well as otherwise, C c and C s have the same unit, which
188
is rupees per unit per year. Here, it is rupees per unit per year to hold one unit of
inventory. This is rupees per unit per year of one unit back ordered.
So, C c and C s will have the same unit, so that we can add them. The expression from
which we derived these two is given by total cost is equal to ordering cost plus inventory
carrying cost plus back order cost. This was simplified to D divided by Q into C naught
plus Q minus s the whole square by 2Q into C c plus s square by 2Q into C s. Now, we
will take a numerical example to try and show the computations involved in, using these
formulae Q, S and TC.
Earlier we had taken a numerical example to illustrate this, where we had assumed D,
annual demand is equal to 10000 per year. We had assumed C naught which is the order
cost as rupees 300 per order. The unit cost of the item was assumed to be rupees 20 and
the inventory holding cost i is 20 percent of 20 and therefore, C c becomes rupees 4.
Using these we had calculated the economic order quantity for the first model and we
had got a value of 1224.74.
Now, we use the same data and we also introduce a C s, which is the shortage cost and
we use shortage cost of rupees 25 per unit per year. C c is 4 per unit per year. We now
compute these values. We apply this formula to compute the economic order quantity.
So, Q is equal to 2DC naught C c plus C s by C c, C s, since I have written the formula
for both the models.
Let me call this as some Q 1, which is the economic order quantity for this system,
without back ordering, and let me call this as Q2 as the economic order quantity, for this
system, which includes and considers back order. We are now interested in computing
this Q2, we already know that this Q1 is 1224.74 which was done in an earlier lecture.
So, we compute Q2. We know all the values 2 into 10,000 into 300 into 4 plus 25
divided by 4 into 25. This on computation would give us, Q2 is equal to 1319.09 and
once we calculate this Q, 1319.09, we go back and substitute here. So this is 1319.09 into
4 by 4 plus 25 into 4 by 29 would give us s is equal to 181.9435 that would give us this
value.
Now, let us also calculate the total cost associated with this model, and try and compare
it with the total cost associated with the earlier model. The earlier model had only two
components which is the order cost and the carrying cost, and then the economic order
189
quantity was 1224.74 and the total cost in this case was found to be 4898.98. So, let us
also try and find out what happens when the back ordering is allowed and C s is
introduced.
Already we have seen that the order quantity increases from 1224.74 to 1319.09, a back
order quantity of 181.94 is allowed. Let us calculate the total cost based on this. So this
component would give us 10,000 into 300 divided by 1319.09, 2274.43 plus Q minus s is
equal to 1137.15. So, this quantity will become 1960.61 plus 313.68, and the total is
4548.72.
So, when we allow back ordering and if the back order cost is rupees 25 per unit per
year, we observe that the order quantity Q increases from the old value of 1224.74
without back ordering to 1319.09. Quantity back ordered is calculated as 181.94. The
total cost comes down from 4898.98 to 4548.72. So, the first observation is that the total
cost comes down; second observation is that, there are three components to the total cost.
The order cost component is 2274.43, carrying cost component is 1960.6 and the
shortage or back order component is 313.68.
We also observe that, this 2274.43 is exactly half of the total, which is 4548.72. This
contributes to half of it, this contributes to the other half of the total. So, in this model we
are essentially trying to find out a Q or order quantity such that the total ordering cost is
equal to the sum of the inventory carrying cost and the back order cost. In this model, we
found Q the order quantity such that the order cost component is equal to the inventory
holding cost component. So, each was equal to half of the total cost.
In this case, this is half of the total cost, these two put together will contribute half of the
total cost. Because we allowed back ordering, the order quantity went up from 1224.74
to 1319.09. So, number of orders per year came down compared to the earlier one. The
number of orders per year came down. Therefore, total order cost came down. The total
cost is two times the order cost; here also total cost is two times the order cost.
Therefore, the total cost also comes down.
Other way of describing it is, even though we order 1319.09, because of instantaneous
replenishment, we will place an order, when we were here. Then place an order for
1319.09, and as soon as the item arrives, which is instantaneous, the back ordered
190
quantity of 181.94 is immediately given. Therefore, the inventory on hand, quickly
reduces to 1137.15, which is the difference between this and this.
So, the maximum inventory that we will hold is 1137.15 as against 1224.74. Also the
period for which we will be holding the inventory, will be far less in this case, because
when we divide by 1224.74, we have a certain period. When we divide by 1319.09, the
whole period comes down. This period comes down and we also realize that this
maximum inventory is not held for this period, but it is held for a smaller period.
Therefore, the average inventory comes down and the inventory holding cost comes
down. So, the order cost comes down, the inventory holding cost comes down, but there
is a back order cost, which is not there in this model, which comes by holding the
quantity of 181, for a much smaller period.
So, the back order cost contributes to another 313.68, but as I mentioned earlier, the total
cost effectively comes down. In this model, we are trading off with order cost on one
side and inventory holding plus back order cost together will be equal to the order cost at
the optimum. This actually points us to a very interesting question. This is a model
which did not allow back ordering. This is a model that allowed back ordering. So, when
we allow back ordering, we seem to get a lesser total cost. We seem to be holding lesser
quantity, for a lesser period in inventory.
So, on the phase of it, by merely look at these numbers, one would be tempted to agree
and believe that, the model with back ordering is desirable. If we have to make a
decision only based on the total cost, then we might say, yes, the model with back
ordering has lesser total cost. This also happens because we look at it as an optimization
problem. This is a problem that allows back ordering. This is the problem that does not
allow back ordering. So, this is a restricted version of this problem, where you are
putting an additional restriction that, back ordering is not allowed.
This is a relaxed version, where back ordering is allowed, and it is only obvious that in a
minimization problem, the restricted version will have an objective function value higher
than the relaxed problem. So, the first model will have total cost higher than the second.
Then we need to answer this question, should we always go for this model or should we
go for this model?
191
Now, we can try and do a little more experiment with this model, try and show some
more computations and then we would actually realize that if we keep increasing the C s
for example, the shortage or back order cost was defined at 25 per unit per year.
Suppose, we make it 100 per unit per year, what will happen? I am not going to show the
computations, but what I am going to tell here is that if we increase this C s, then
automatically what will happen is, this cost will also increase and this will come closer to
this 4898.98. If we increase C s, the order quantity will decrease from 1319.09. The total
cost will increase and come closer to 4898.98.
As we keep on increasing this C s, this quantity will keep coming down, this will come
down, and this total cost will go up. When C s is equal to infinity, this will take a value
1224.74, this will take a value 0 and this total cost will be equal to this total cost. This Q
will be equal to that Q. That is also clear because we can rewrite this as 2D C naught by
C c into 1 plus C c by C s. So, when C s is infinity, C c by C s will become 0. So 1 plus
C c by C s is 1. Therefore, when C s is infinity this will become 2D C naught divided by
C c which is the same model.
So, only when the back order cost is kept as infinity which implies that back order is not
allowed, we will automatically get the other model. As long as the back order quantity is
anything smaller than infinity, however large it may be, this model will always give a
total cost, slightly lower than this model. But the actual way to look at it is, unless we are
absolutely sure about the correctness in the computation of C s, and we know that this is
exactly the back order cost and let us say we have included the loss of good will and all
those intangible or costs that cannot be measured easily, let us say we have measured all
the cost and if the organization is very clear that this is my back order cost and this is
exactly what I am going to incur, there is going to be no delay to the customer.
Therefore, there is no cost associated with this. If we sure about all of that, only then this
model can be used straight away. If we are not sure and if there are some more
uncertainty, where there would be a delay because of this back order, and we are not able
to estimate the cost of such delay accurately, we can either give a very large value for
this, or to be on the safe side if we do not want to have back order at the planning stage
itself, we would rather give value equal to infinity and go towards the first model.
192
The second model is very important and has to be used very carefully. Unless we are
sure of the assumptions of the second model, it is always good to go back to a situation
where the back order is not allowed at least during the planning stage. So, that is the
reason why look at two types of inventory models particularly for bought out items. Now
look at another aspect of it, which is called discount.
In order to explain the discount model, we will be looking at this model and we will not
be considering this model. When we have a single item and we buy this item, let us
assume, we are working with this data, where the annual demand is 10000, order cost is
300 and so on. We have already found that the economic order quantity is 1224.74 and
the total cost is 4898.98. When we consistently buy the same item from the same vendor
or when we are buying a larger quantity, we would ask for a discount and the vendor
would offer a discount invariably when the buying quantity is large.
Let us look at a situation where the vendor is offering a discount for this, and then we
need to evaluate, whether we will accept the discount or we will reject the offer of
discount. Let us assume that the vendor is willing to give a 2 percent discount when the
order quantity Q is greater than or equal to 2000. If we go back to the derivation of this
Q1, we first said that we would consider there were 4 costs; the order cost, the carrying
cost, the back order cost and the cost of the item. This model does not allow back order.
193
So we look at the 3 costs and wrote the expression TC is equal to D divided by Q into C
naught plus Q by 2 into C c plus D into C.
This is the ordering cost component, this is the inventory holding cost component, and
this is the cost of the item per year. Since, the variable was how much to order and the
variable did not figure in this D into C in our derivation, we left it out because D into C
is a constant. We left it out and then we derived this expression. But now if we want to
consider this order, this discount offer, for a quantity of 2000 or more. Let us consider
that we are looking at whether to accept the discount at 2000, we will then address the
situation, whether we will accept the discount at a quantity greater than 2000.
If we are interested in quantity of 2000 and we want to avail the discount, at any order
quantity which is different from 1224.74, the sum of these two D divided by Q into C
naught plus Q divided by 2 into C c which represents the total cost of ordering and
holding, will definitely be more than the 4898.98 because we have derived 1224.74 as
the optimum value, where this total function is minimum. This D into C is also a
constant. So, when we increase the order quantity 1224.74 to say 2000, this part is going
to go up. And if this were a constant it would never be economical to avail it. But then
there is a discount. Because of the discount the C now takes a new value, it is not 20
anymore but, it is 98 percent of 20 with a 2 percent discount.
So, the increase in this, if it is offset by the gain here or if the gain here is more than the
increase in this, then it would be advantageous for us to avail the discount. There is also
a very small component, which needs to be looked at; it depends on how we are going to
define this C c. We have already seen that this C c can be defined in two ways; one is as
C c or as i into C.
If we now define this C c as i into C, and if we are interested in an order quantity of

2000, at 2000 this carrying cost will also come down slightly because this C also changes
slightly. So there will be an overall increase in D by Q into C 0 plus Q by 2 into i C but,
that overall increase will also have a very small decrease because of this decrease in C.
But then this increase has to be offset by the gain because of the discount.
So, let us first look at, can we avail this discount at 2000. Before that let us answer this
question, are we going to avail? We are not going to get this discount at all, for any
quantity up to 2000 and for any quantity up to 2000; the total cost will be the lowest at
194
the economic order quantity of 1224.74. So, we should look at 2000. When we look at
2000, at 2000 T C is equal to, now D by Q is 5, 10000 divided by 2000 is 5. 5 into C
naught is 1500 plus Q by 2 is 1000, i C is 20 percent of rupees 4 into 98 percent because
at quantity of 2000 the unit price is not 20, but it is 98 percent of 20.
Therefore Q by 2 into i into C will become 1000 into Q by 2 is 1000 i is 0.2 into 20 into
0.98 plus D into C 10000 into 20 into 0.98. This on computation would give us 1500 plus
3920 plus 196000 which is 201420. This cost has to be compared with what is the cost
that we would incur if we had ordered 1224.74. At 1224.74 T C cross is equal to 4898.98
plus 20 into 10000 which is 204898.98. The total cost at economic order quantity is
204898.98. The total cost at an order quantity of 2000 and availing at 2 percent discount
would give us a total annual cost of 201420. This is lesser than 204898.98.
Therefore, it is profitable to avail the discount at 2000. We have answered the first
question as to, whether we will accept it 2000. Now, we have to answer another question
whether we will accept it at a quantity at a greater than 2000. For example, even at 3000
will get the 2 percent discount. So, is it advantageous to avail it at a quantity of 3000 and
in order to do that we should go back and do 10000 by 3000 into 300 plus 3000 by 2 into
0.2 into 20 into 0.98 plus 10000 into 20 into 0.98. So, this portion is going to be
remaining the same. But then we have to check, whether this one decreases, so that we
could avail. In order to do that, let us go back to another curve that we have already seen,
which is called the total cost curve.
195
We have already seen that the total cost curve is very flat near the economic order
quantity. Let us say this is the economic order quantity and this is 1224.74 and then at
2000 let us say we are getting this discount. We are going to get this discount. So, it
comes down a little bit. I have not actually given you the y axis cost, the y axis now is
the order cost plus carrying cost plus cost of the item. So, total cost is equal to order cost
plus carrying cost plus D into C, which is the item cost.
We realize that at 2000 it actually comes down, becomes lower than this, this is your
204898.98. This value is your 201420. This is here and then if you see carefully, the total
cost will go up. Therefore, from the very nature of this graph, we will understand that it
is not necessary for us to see whether we are going to avail it at 3000 or any value higher
than the point at which there is a discount. So, the point at which there is a discount is
called the price break point. The price break point is 2000 and therefore, we need not
evaluate it at 2001 or 2 or 3000 or any value greater than that price break point, to avail
that discount quantity.
196
Now, let us assume that the vendor is giving us another price break point and the vendor
says that he can give a 3 percent discount, if the order quantity Q is greater than or equal
to 5000. So, we have to do the same calculations. At 5000 what is the total cost? T C is
equal to D by Q into C naught, so there will be two orders 10000 divided by 5000 is 2. 2
into C naught is 600 plus Q by 2, 2500 into C c which is 0.2 into 20 into 0.97 plus 10000
into 20 into 0.97. This will become 600 plus 9700 plus 19400. So this is 204300.
To show it in this graph, let us say 5000 is somewhere here, it goes up. Say 5000 is here
so it goes up now, there is another price break, so it comes down and somewhere in the
middle, this is still minimum, this is 4898. It comes somewhere up to here and then it
keeps moving like this. This is still the optimum price break point and therefore, we will
not accept this we would accept a 2 percent discount at 2000 and we will not accept a 3
percent discount at 5000.
Wherever there is a price break, it is necessary to compare the total cost at the price
break, not at any other point. Only other part is the case where there is a price break
before the economic order quantity. Let us say there is a half a percent discount at 1000,
then in such situation, we will try and avail that half a percent discount not at 1000, but at
1224.74. So, if there is a discount at a quantity before the economic order quantity then
we will try and avail that discount at the economic order quantity. If there is a price break
after the economic order quantity, we will try and avail it, at the price break point
197
It is generally customary that price break points are for quantities, higher than the
economic order quantity, but it so happens that there is a price break point before the
OQ. For example, there is a half a percent discount at say 1000 which means, it is going
to come down here, it is going to further down, till reaches the economic order quantity
and proceed. The other minor adjustment there is that the economic order quantity can
also shift a little bit because if there is half a percent discount at 1000, then the quantity
changes from 2 into 10000 into 300 divided by 4 which is the original one which gave us
value of 1224.74.
Now if there is a half a percent discount at 1000 that we wish to avail at the economic
order quantity, then the economic order quantity will shift a little bit. It will become 2
into 10000 into 300 divided by 4 into 0.995, if there is a half a percent discount. It will
become slightly more than 1224.74. It may come to around 1224.85 or may be closed to
1225. Physically there would not be too much of a change, but if we look at it from the
computation point view, there will be a slight increase in the economic order quantity
and we will avail the discount, at the new economic order quantity.
This is how we handled the discount. The discount problem always involves computation
of a total cost, which also includes the cost of the item, because the gain in any discount
comes from the reduction in D into C, which was otherwise a constant because C was not
a function of Q. In the discount problem the item price C is of function of Q and
therefore, C will change and D into C will also start influencing the calculation of the
order quantity. The discount model that we have seen is called an all quantity discount.
By which we mean that if you order a quantity of 2000, then for every item that is
ordered, the cost is not rupees 20, but it is 98 percent of 20 which is rupees 19 and 60
paise. For each one of the 2000 units, the price is at 19.60 paise. Now we could also have
something called marginal quantity discount. So, let us look at that. Let me write the
marginal quantity discount.
198
We could have something called marginal quantity discount. In marginal quantity

discount the vendor may say that if you order for Q greater than or equal to 2000, let us
say the vendor may give us, it is 5 percent from the 2000th item and say 10 percent from
5000th item. What is the difference? The first difference is this, if we ordered 2000 items
in the all quantity discount, let us compare these two cases an all quantity discount of 2
percent versus a marginal quantity discount of 5 percent. If we order say 3000 items,
then all the 3000 items will be priced at 19.60 paise. If we order 3000 items here, the first
2000 items will be priced at 20. The next 1000 which is after the 2000th item the next
1000 are now going to be priced at 95 percent of 20, which is 19 rupees.
So, at Q equal to 3000 only the item cost per year or at time of ordering, if we order 3000
the item cost will be 3000 into 19 rupees 60 paise. Here, it is 2000 into 20 plus 1000 into
90. The gain in the marginal quantity discount is slightly lesser than, the gain in the all
quantity discount. For that reason you will see that the discount is higher in the marginal
quantity, than in the all quantity discount. Let us work out this problem with marginal
quantity discount of 5 percent from the 2000th item and 10 percent from the 5000th item.
199
The derivation will be as follows. T C is equal to D by Q C naught plus Vj minus Q

minus qj into Cj into i by 2 plus D by Q into Vj plus Q minus qj into Cj. This on
differentiation would give us, Q is equal to root over 2 D into C naught plus Vj minus qj
Cj divided by i into Cj. Let me explain this a little bit. It is also obvious that when we
differentiate this with respect to Q, we are going to get this one.
As I said, there are 3 components. There is an order cost component, there is a inventory
carrying cost component and then there is a cost of the item component. Order cost is
very clear; D divided by Q into C naught, Q is Q what we are going to order. D by Q is
the number of orders per year. D by Q into C naught is the order cost. Now, what is the
carrying cost? Carrying cost originally is Q by 2 into C c. We can write it as Q by 2 into i
into C. Can also be written as Q C by 2 into I; where C is the unit price of the item, Q is
the order quantity and Q C is the money value of the order quantity.
Now let us look at order cost, as the money value of the order quantity by 2 which is the
average into i. So, i by 2 is separated. This is the money value of the quantity order, by
the very definition of a marginal quantity, if we are looking at this case 2000 and 5
percent for a quantity above 2000. For example, if we are calculating for 3000, we would
do, 2000 into 20 plus 1000 into 19, which is given by this kind of thing. Vj is the value
for the first 2000, Q minus Qj 3000 minus 2000 which is thousand into Cj which is the
new value 19.
200
This is the expression for the money value for the item that is bought. So that into i by 2
is the inventory holding cost and this is the money value per order. This is the value the
number of orders per year is D divided by Q. So, the total annual value will be D divided
by Q into this. This on differentiation would give us this total value for the economic
order quantity.
Let us work out the same problem for the marginal discount and if we do that now, the
first one is at the economic order quantity at EOQ, Q is equal to 1224.74 and T C will be
204898.98. There is no discount. So, 4898.98 come from here. The cost of the item is
200000 so 204898.98 at Q equal to 2000. We need to substitute so you get 2 into 10000
into 300 plus 2000 is the gain.
For example, at 2000, 5 percent of 20 is 1 rupee. Vj minus qj Cj is 2000 into 1. For

example, at 2000 if I had not availed the marginal discount, I would have paid 20 rupees.
If I avail the marginal discount, I pay 19 rupees. Therefore, 2000 into balance 1, that is
Vj minus qj Cj. This quantity at 2000, Q is equal to root over 2 into 10000 into 300 plus
2000 divided by 3.8 because there is a 5 percent discount, so 4 becomes 3.8. So this on
computation gives us 3479.26.
If we really want to use this marginal discount of 5 percent from the 2000th item
onwards, the best deal we will get is, if we order 3479.26 items, out of which the first
2000 items will be priced at 20, and the balance 1479 items will be priced at 19. The
total cost associated with this will be 10000 into 300 divided by 3479.26 plus i by 2 into
money value of 3479.26. So, 0.2 by 2, I by 2 into money value of this.
So, the first 2000 is at 20. So, 20 into 2000 plus 19 into 1479.26, that is the money value
plus this money value, this will be 10000 divided by 3479.26 into 20 into 2000 plus 19
into 1479.26. All of these put together would give us 203421.28. This is the total cost
that we will incur, if we want to avail this portion of 5 percent for a quantity of 2000.This
is actually cheaper than 204898.98. We have to evaluate, what is the best deal we get, if
we try and utilize this, and then we have to compare that with 203421.28 and choose the
one which is cheaper. We will show those computations in the next lecture.
201
Prof. G. Srinivasan
Lecture - 12
Marginal Quantity Discount, Multiple Item Inventory – Constraint on Numbers of
Orders
In the previous lecture, we were discussing marginal quantity discount through a

numerical example. The example is as follows, there is a single item with an annual
demand of 10,000 the order cost is rupees 300 per order. The cost of the item is rupees
20 and the inventory holding cost is 20 percent of rupees 20. If there was no discount
then the economic order quantity Q is given by root of 2 D C naught by i into C, 2 into
annual demand into order cost by carrying cost, which would give us 2 into 10,000 into
300 by 4, which will give us 1224.74 which we have already seen in the previous lecture.
The total cost associated with this will be, root over 2 D C naught C c which would give
us 4898.98, which is a computation that we have seen in the previous lecture. Since, we
are going to look at discount models we should also add the unit cost of the item, at the
rate of rupees 20 for 10,000 items. So, T C in this case will be 4898.98 plus 200,000
which would give us 204898.98. Now, we apply the marginal discount to this problem.
Now, there is a 5 percent discount from the 2000 item and there is a 10 percent discount
202
from the 5000 item.
This means that, if we buy 2000 items then we will be paying 2000 into 20, the discount
comes for the 2000 and 1st item onwards therefore, if we buy 3000 items the 1st 2,000
will be purchased at 20 and the next 1000 will avail or will get a 5 percent discount and
will be priced at rupees 19. In a similar manner, if we buy 5000 items now the 1st 2000
will be priced at 20 the next 3000 will be priced at 19, the 10 percent discount or a price
of 18 will come for the 5000 and 1st item onwards and therefore, if we buy 6000 then the
1st 2000 will be priced the 20, the next 3000 will be priced at nineteen and the remaining
1000 will be priced at 18, which is a 10 percent discount of 20. Now, in order to find the
order quantities associated with these marginal quantity discounts are priced break, we
use this formula to compute.
203
So, the first one which is 5 percent at 2000 onwards gives us Q equal to root of 2 D C
naught plus V j minus q j c j divided by i c j. And we use this formula to compute the
order quantity considering the marginal quantity discount. Now, we use this formula to
get 2 into 10,000 into 300 plus now, we need to write the values of V j minus q j c j.
(Refer Slide Time: 00:08) Now, we are considering the quantity 2000 so this quantity
2000 would be brought ordinarily at 2000 into 20 which is 40,000. And after a 5 percent
discount the price becomes 19 so 19 into 2000 is 38,000.
So V j is 40,000 q j c j is 38,000 so the difference is 2000 divided by i into c j which is

point 2 into 19, 19 is (Refer Slide Time: 00:08) a 5 percent discount on 20 rupees. So,
this on simplification would give us Q equal to 3479.26. So, if we avail or wish to avail
this (Refer Slide Time: 00:08) 5 percent marginal discount from the 2000th item, we
need to order 3479.26 units, which also means that the 1st 2000 units will be priced at
(Refer Slide Time: 00:08) 20 and the remaining 1479.26 will be priced at 19.
So, the total cost associated with this will be D by Q into C naught plus Q by 2 into i into
c plus D by Q, which is the number of times into Q C or plus item cost. So, this will be
10,000 by 3479.26 into 300, this will be the total ordering cost. Now, this q by 2 into i c
will be, now written as i into Q c i by 2 into Q c, which is 0.2 divided by 2 into the cost
or prize of the 3479.26. which is 2000 into 20 plus 1479.26 into 19, will be the inventory
holding cost plus the item cost will now be. Now, the cost of 3479.26 items is given by
2000 into 20 plus 1479.26 into 19, this is the money value of the item every time we buy
that item and in a year the number of times we buy, is given by 10 000 divided by
204
3479.26.
Therefore, into 10,000 by 3479.26 will be the money value of the inventory. So, let us
calculate these to get 862.25 plus 0.1 into 68,105.94 plus 68,105.94 into 2.874 this would
give us so this quantity is 195748.35 the first one is 862.25 plus 6810.59, which would
give us 203421.2. Now, we observe that, if we compare these numbers with these
numbers (Refer Slide Time: 00:08) here the item cost is 200,000 because of the discount
the item cost comes down to 1, 95,748. (Refer Slide Time: 00:08)
Here the order cost will be 2449 the order cost is only 862, that is because the order
quantity has (Refer Slide Time: 00:08) increased from 1224.74 to 3479.26. So, there are
fewer orders in a year so order cost has come down the carrying cost has gone up
significantly, (Refer Slide Time: 00:08) 6810.59, because of a large quantity that we
order here. So, total cost will be 203421.2, which is less than 204898.98 so it is
advantageous at this point to avail the (Refer Slide Time: 00:08) 5 percent discount and
order 3479.26.
Now, let us look at the 10 percent discount from 5000 so Q will be equal to root of 2 into
10,000 into 300 plus. Now, we need to find out this (Refer Slide Time: 04:19) V j and q j
c j. So, if we take 5000 items and if we (Refer Slide Time: 00:08) did not have this 10
percent discount then V j will be the price associated with the 5000 items that we would
have bought. So those 5,000 items would be priced at (Refer Slide Time: 00:08) 2000
205
items at 20, which is 40,000 and 3000 items at 19 which is 57,000. So V j will be 97,000
minus for q j c j now, (Refer Slide Time: 00:08) 5000 is the quantity the price is 18 so 5
000 into 18 is 90,000 divided by 0.2 into 18. So, this would give us root of 2 into 10,000
into 7,300 divided by 3.6, this on simplification would give us 6368.2 units. So, if we
wish to avail the additional marginal discount of 10 percent, which is 2 rupees from the
5000th item then we observe that we have to order 6368.2 units.
Now, the total cost associated with this will be sum of order cost plus inventory holding
cost plus cost of the item. So, the order cost will be 10,000 divided by 6368.2 into 300,
now, this term is this represents the number of orders in a year, multiplied by the order
cost per order; this will be the annual order cost. The inventory holding cost will be 0.2,
20 percent of the average inventory so the average gives another by 2 into the money
value of the 6368.2 units. So, the 1st 2000 will be priced at 20, the next 3000 will be
priced at 19 and 1368.2 will be priced at 18. So, this is the money value of 6368.2 now,
that divided by 2 into 0.2, the 0.2 is the 20 percent inventory plus the annual item cost.
So, plus the annual item cost will be every time we place an order, we spend so much,
the number of times we place an order is given here. So, plus 10,000 divided by 6368.2
multiplied by 2000 into 20 plus 3000 into 19 plus 1368.2 into 18.
Now, let us find out this number first, before we do that we get 10,000 into 300 divided
by 6368.2 would give us: 471.09, which is the annual cost plus 0.1 into 2000 into 20 plus
3000 into 19 plus 1368.2 into 18 would give us 121627.6 so 0.1 into 121627.6 plus
10,000 divided by 6368.2, 1.57 into 121627.6. So, this term will be into 121627.6,
190992.12 this will be there is only 1.1. So, 12,162.76 plus 471.09, this will become plus
12,162.76 plus 471.09, which is 203625.96. So, this is 203625.97, 2000 20 plus 3000
into 19 plus 1368, 0.2 into 18 will give us 121627.6 into 10,000 divided by 6368.2,
190992.12 plus 12,162.76 plus 471.09; 203625.97 so we get 203625.97. Now, we have
worked out the costs for a (Refer Slide Time: 00:08) 5 percent discount at 2000 and a 10
percent discount at 5000. Without a discount it was 204898.98 now, with the 5 percent
discount it is 203421 and with the 10 percent discount 203625. This is slightly higher
than this cost, based on the computations that we have shown. Therefore, we would
choose a 5 percent discount.
206
And we would order 3479.26 and incur an annual cost of 203421.2. Now, there are two
other things that we have to see, the first thing is that this cost is not significantly higher
compared to this cost. So now, do we consider ordering a larger quantity of 6368.2
because we would have more stock with us at any point in time? General thinking in
inventory problems is not to build up a lot of inventory, and therefore this will we will
not consider because this quantity is a very large quantity. Even if we had a situation
where, this is slightly lesser than this; even if we have such a situation then it would still
be wise to go for this because the larger the ordering quantity higher will be the
inventory holding cost and more inventory will be built up in the system.
Now, we also need to understand that the order quantities have become bigger in (Refer
Slide Time: 00:08) when we looked at the marginal quantity discount, even though the
discount is available from 2000 plus onwards or 2000 and 1st item onwards. We realize
that we actually benefit from the discount, only if we order a quantity that is higher than
the price break. Therefore, a 2000 price break with marginal quantity leads us to 3479
and a 5000 price break with marginal quantity leads us to 6300. So, marginal quantity
discount has a tendency to increase the order quantity because we can avail the benefit of
the discount only when the order quantities are much higher than the price break or the
point at which or from which that discount is applicable. So, it forces us to order larger
quantities.
207
We also have to make a comparison with the all quantity discount, where we have earlier
worked out that for a 2 percent all quantity discounts, the total cost turned out to be
201420. The computations are for a 2 percent all quantity discount so the price break at
2000, at 2000 units T C would be D by Q, there will be five orders because the order
quantity is at 2000. There will be five orders, 5 into 300 plus Q by 2 into C c 1000 into
there is a 2 percent all quantity discount. So Q by 2 into carrying cost is 4 into 0.98,
which is your 2 percent discount and the item cost would be 200,000 into 0.98. So, this
would give us 1500 plus this is 3.92 so 3920 plus 196000, which would give us 201420,
which is cheaper than or less costlier than 203421.2. So, in this situation we could look at
the all quantity discount if it is available and chose this, for two reasons that the total cost
is lower and more than that the order quantity is also lesser.
208
Now, after this discussion we move towards multiple item inventories. Now, in reality
we will be looking at more than one item, though so far we have seen single item
inventory problems. So, we now look at a two item inventory problem and try to
understand, what the issues are with respect to multiple item inventories by simply
considering two items at a time. So, we now look at the 1st item which has the demand
of 10,000 per year we call it D 1 representing the first item and C 1, which is the unit
cost of the 1st item. We assume that the order cost is 300 and i is 20 percent, we already
know that Q 1 is 1224.74 and T C 1 is 4898.98.
Now, if we order 1224.74 every time, we make N 1 the number of orders per year will be
10,000 by 1224.74 which will become 8.16 orders per year. Now, let us consider a 2nd
item with demand D 2 equal to 20,000 per year, we consider the same C naught equal to
300. We use the same i equal to 20 percent we assume that the unit cost of the item is 25,
now, if we work out the order quantity then Q 2 for this item will become 1549.19,
which is root over 2 into 20,000 into 300 divided by 5, 5 is 20 percent of 25 and the total
cost associated with this will become 7745.97.
Now, for the 2nd item, with demand equal to 20,000 and order quantity equal to 1549.19
the number of orders per year N 2 will be 20,000 divided by 1549.19, which will become
12.91 orders. Now, if these two items are treated independently, then we will be making
8.16 orders for this item and 12.91 orders for the other item, which would give us a total
209
number of orders as 21.07 orders, now, we realize that 21.07 may be a bit too large when
we look at two items. And particularly, if we look at a very large number of items then
we realize that, we will be making too many orders when all these items are put together.
Therefore, we also build a restriction and a constraint that the number of orders will be
less than or equal to 15 say and see, what happens to the ordering quantities, if we put
this kind of a restriction.
So, the problem will now reduce to minimizing total cost, which is D 1 by Q 1 into C
naught or sigma D j by Q j into C naught plus i into sigma Q j C j by 2. If we take the 1st
item the order cost will be D 1 by Q 1 into C naught, for the 2nd item D 2 by Q 2 into C
naught. We have used the same C naught so C naught can be taken outside, we have used
the same i so for the 1st item it is Q 1 C 1 by 2 into i for the 2nd item it is Q 2 into i C 2
divided by 2. So, the i is taken out, so we have this. Subject to the constraint that D j by
Q j sigma is less than or equal to some N so in this case the number of order should be
less than or equal to 15.
If we do not take this constraint and we optimize this we will get the same answer here
with the 1rst item having 8.16 orders and the 2nd item having 12.91 orders. Now, the
unconstrained optimization solution would violate this constraint and therefore, this
constraint will become binding. As shown in this example, if we do not include this
constraint, the unrestricted problem will give us 21 orders, which violates 15 and
210
therefore, the constraint optimization problem has to be solved. Now, this is a constraint
optimization problem with an inequality as a constraint. Now what we will do now is, we
will try and replace this inequality with an equation and we try to understand something
through the properties of the total cost.
Now, we have already seen that if we have a single item where, we have quantity and we
have cost the total cost curve will look like this and let us say the economic order
quantity will be here. Now, we realize that if we take 1 item with order quantity 1224
and this equal to 8.16, 2nd item has 12.91. If this 21.07 orders has to be brought down to
15 orders then the order quantity for both the items will go up, which means this
constraint if we take a particular item is like saying here is Q star, now, this constraint is
like saying now we are going to increase the order quantity; so we are looking at a
constraint which looks like this. Now, the objective function is a function of this type
and we have a constraint which will try and put q greater than or equal to something.
Now, by the nature of the objective function the optimum will happen only here and will
not happen at other points. Therefore, we make an assumption which is correct that,
when the constraint is binding than the inequality can be replaced by an equation and we
can solve it. Now the advantage of replacing this inequality by an equation is that we can
use the method of Lagrangian multipliers and take this equation into the objective
function and then we solve this. But, we should also note that only when the constraint is
211
binding the inequality will become an equation.
So, we take it into the objective function and introduce a Lagrangian multiplier, which
will be like this. So, to minimize L the Lagrangian multiplier sigma D j by Q j into C
naught plus i into sigma Q j C j by 2 plus lambda, which is the Lagrangian multiplier into
sigma D j by Q j minus N. Now, partially differentiating this with respect to the k’th
item, we will have minus D k by Q k square into C naught plus i C k by 2 minus D k by
Q k square lambda equal to 0; from which Q k, general value for the k’th item will be 2
D into C naught plus lambda by i C k. Now, the effect of the constraint is the addition or
inclusion of lambda into this value.
So, the new order quantity we will be able to find out only if we know the value of
lambda, we know for the 1st item we know we will have a D k here. So, for the first item
we know its 2 into 10,000 into 300 plus lambda by point 2 into 20 but then we do not
know the value of lambda. So, in order to find the value of lambda we partially
differentiate this with respect to lambda and then we will get sigma D k by Q k is equal
to N. Now we already know this Q k is equal to this now, just to be consistent with the
rotation I am going to replace k by j. So, now we will have sigma D j by Q j. Q j is root
over 2 D into C naught plus lambda into root over i C j, so D j by Q j is equal to N.
212
So we simplify here, now we have sigma D j into root over i C j divided by root over 2 D
j into C naught plus lambda. So, D j by Q j is equal to N, so now, we simplify here to get
root over sigma root over D j C j, we have root of i by 2 times C naught plus lambda is
equal to N. Now, we take C naught plus lambda; first we square it to get i by 2 times C
naught plus lambda sigma root of D j C j square is equal to N square, from which C
naught plus lambda is equal to i by 2 N square sigma root of D j C j the whole square,
from which lambda is equal to i by 2 N square sigma root of D j C j the whole square
minus C naught. So, that is the value of lambda and now for our example, we will now
substitute lambda is equal to 0.2 by 2 into 15 square into root of D 1 C 1 plus D 2 C 2
minus 300, which on substitution, would give us 292.20.
213
And then we go back and substitute this lambda into this and if we do this substitution so
let me, Q j is equal to root of 2 D j into C naught plus lambda by i C j. Now, the new Q 1
will be root of 2 into 10,000 into 300 plus 292.20 divided by 4, which would give us
1720.76 and Q 2 will be 2176.61. Now, Q 2 can be obtained as 2 into 20,000 into 300
plus 292, which is C naught plus lambda divided by 5. So, these are the 2 order quantities
now, for these order quantities N 1 will be, the number of orders will be 10,000 divided
by 1720.76 which is 5.811; other one will be 20,000 divided by 2176.61, which is 9.188
which when added will become 14.999 or 15 orders.
Now, we also observe that the original 1224.74 Q 1 star is 1224.74 and Q 2 star without
the constraint was 1549.19. Now, we realize that the constraint has increased 1224.74 to
1720.76 and it has increased 1549.19 to 2176.61. We also observe now, that Q 1 by Q 1
star is equal to 1720.76 divided by 1224.74 which is 1720.76 divided by 1224.74, which
is 1.405. Q 2 by Q 2 star is 2176.61 divided by 1549.19, which is 2176.61 divided by
1549.19; 1.4049 which is this. And we also realize that 21.07 orders has become 15
therefore, N 1 plus N 2 divided by the constraint N will be 21.07 by 15 is 1.4046, which
is 1.405 subject to the rounding of errors.
So, we also realize that a simple formula would have helped and the Lagrangian
multiplier method leads us to the simple formula. So, if a total of 21.07 orders has to be
brought down to 15 then multiply the individual order quantities by a quantity 21.07 by
214
15 which is 1.405 so 1224.74 into 1.405 will give the new order quantity for the 1st item
and similarly, for the 2nd item. So, this is how we solve a multiple item inventory
problem with constraints on the total number of orders. But, before we wind up this
discussion on the constraint problem with total number of orders, we again go back to
this and then we look at this number of lambda equal to 292.20. Now, in order to
understand that let me write the formulation again the formulation is to minimize sigma
D j by Q j into C naught plus sigma i Q j C j by 2, subject to sigma D j by Q j less than
or equal to N.
Now the unconstrained problem here, gave us 21.07 to the left hand side which made the
constraint binding and then because we understood that when the constraint is binding
the constraint can be converted to an equation so that we could use Lagrangian
multipliers and solve it. Now, suppose we had capital N is equal to 25, which meant that
this constraint is not binding; 21.07 will be less than or equal to 25, so the obvious thing
is not to worry about this constraint and continue to order 1224.74 and 1549.19 so that
the total cost is minimized to 4898.98 plus 7745.97.
It is also obvious that when we compute the total cost for these 2 the total cost will be
higher than 4898.98 and 7745.97. So, when the constraint is satisfied as in the case of n
equal to 25, we will not use this method so we will first solve the problem independently
like we did here check whether the constraint is satisfied for N equal to 15 it is violated;
for N equal to 25 it is satisfied. When it is violated we could use the method of
Lagrangian multipliers get a lambda and compute these values or we could use these
ratios to get it ratios are also theoretically correct.
But, if N was 25 and therefore, this constraint is already satisfied, then we will not use
the method of Lagrangian multipliers, we would simply take these unconstrained values
as optimum values. In spite of that, if we go back and substitute N is equal to 25 for this,
instead of N is equal to 15 substitute 25 square now, we will realize that we will get a
negative value of lambda. Now, this negative value of lambda, when which the moment
we get a negative value of lambda we need to understand that the constraint was
actually not violated by the original values.
Therefore, there is no need to do the Lagrangian multiplier method but we simply blindly
have used this formula and we will get a negative value of lambda if we do that. Now,
215
such a negative value of lambda will now come back here and that will force the order
quantities to be lower than 124.74 and 1549.19 so that, the total will be forced to 25 now,
we should not allow that to happen. So, we have to understand this, first we solve the
problem without a constraint, check whether the constraint is violated and only when the
constraint is violated we will use the method of Lagrangian multipliers to solve this.
Now, if the constraint is satisfied as in the case of N equal to 25, we will not use this
method, we would simply use the unconstrained one and get the minimum cost of
4898.98 plus 7745.97. In spite of this, if we make the mistake of using this formula, we
would get a negative value of lambda because; this formula was derived for an equation
here. Therefore, it will give a negative value which will force this to be an equation and
it will force this to get 25 orders in a year and it will reduce these order quantities and
increase the cost to more than 4898.98 and 7745.97, which we will not use.
So, the moment we get a negative value there, we have to check the calculation and if we
are sure that you get a negative value it means, we should not have used the Lagrangian
multiplier method the original unconstrained solution itself is optimal and there is no
need to use this method to get the new values of the order quantities. Now, other aspects
of multiple item inventories such as restriction on the money value of inventory etcetera,
we will see in subsequent lectures.
216
Prof. G. Srinivasan
Lecture - 13
Multiple Item Inventory - Constraint on Money Value, Space, Equal Number of
Orders
In the last lecture, we addressed multiple item inventory problems, and we specifically
looked at a situation where there was a restriction on the total number of orders to be
placed, considering multiple items. In this lecture, we will consider another restriction or
constraint where there is a limit on the money value of inventory that is being bought or
locked up. So, let us consider the same example and then let us derive the result and then
apply it to the same numerical illustration.
So, let us consider 2 items whose demands D are 10,000 and 20,000 per year. The order
cost C naught is the same 300 per order. The inventory holding cost is 20 percent of the
unit cost for both the items and the unit price of the items are 20 and 25. We have already
computed the economic order quantity for both these items and the economic order
quantity comes to 1224.74 for the 1st item and 1549.19 for the 2nd item.
Now, the numbers of orders are also different for each of them and in the last lecture we
looked at a situation where there was a limit or restriction on the total number of orders.
217
And, at the end we also derived an expression and showed that if there is a restriction on
the total number of orders and if that constraint is binding then the order quantities of
both the items increase in the same proportion. Now, we are trying to place another
constraint which is a restriction or limit on the value of money that is being locked up in
the average inventory. Now, if the order 1224.74 of this item the average inventory is
half of this, Q by 2 which we have already seen. And, the money value of the average
inventory which we would call as Q by 2 into C, C is the unit price of the item so that
will be 1224.74 divided by 2 into 20 which is 12,247.4.
And, for the 2nd item it will be 1549.19 into 25 divided by 2 which will work out to
19,364.88. So, this amount 12,247.4 and 19,364 represent the money value of the
average inventory because we start the cycle with 1224.74. The cycle ends with 0 and we
have already computed that the average inventory is Q by 2. So, the money value of that
is so much. So, put together the money value of both these items comes out to 31,612.58.
Now, these two items put together we observed that the average money value of the
average inventory is 31,612.58. In large manufacturing organizations, there would be
hundreds and thousands of items and the money value of the average inventory can be
very high.
So, organizations place a restriction on the money value of the average inventory held at
any point in time. And then the order quantities are adjusted such that the money value
of the average inventory held at any point does not exceed a given predetermined or
given value. So, in this case if we say that these two put together the average, if money
value of the average inventory is 31,612. Now, we would pose a question and say that if
the money value of the average inventory held is restricted to 25,000 instead of 31,612,
what happens to the order quantities? So, this leads us to an optimization problem where
we try to minimize.
218
Now, D j is the annual demand for item j, Q j is the order quantity. So, D j by Q j is the
number of orders into order cost plus sigma Q j by 2 into i C j. Now, Q j by 2 is the
average inventory, i into C j is the inventory holding cost per unit per year. Here, the
demand is also in years. So, this represents the cost of ordering plus cost of carrying
subject to the condition that sigma Q j C j by 2 is less than or equal to some B. Now, Q j
C j by 2 is the money value of the average inventory. Please note, that there is no i here
whereas there is an i here. The moment there is an i here then this represents the
inventory carrying cost per year. This represents the money value of the average
inventory. So, Q j C j by 2 is less than or equal to B. Now, this is a non-linear
optimization problem because Q j comes in the denominator here, comes in the
numerator here with a linear constraint where the Q j comes in the numerator, in the
constraint.
Now, the 1st thing we need do is, to solve this problem as an unconstraint problem and
then try and solve it, and if it so happens that the solution to the unconstraint problem
satisfies this constraint then it is optimal to the constraint problem. So, we follow the
same approach and try to solve this unconstraint problem. Now, when we solve this
unconstraint problem, this unconstraint optimization problem turns out to be two
independent problems, one for each item where the optimum solution is the economic
order quantity itself.
219
We have already calculated the economic order quantity and we have said that (Refer
Slide Time: 00:54) the values are 1224.74 and 1549.19. The value of Q j C j by 2
summation for the economic order quantity we have already calculated and that is (Refer
Slide Time: 00:54) 31612.58. And now the, which is the limit on the money value of the
average inventory is 25,000. Therefore, the solution to the unconstraint problem which is
the economic order quantity while it is the constraints and the constraint is binding and
the solution to the unconstraint problem is not optimum or feasible to the constraint
problem. So, we have to use this constraint and then solve it again. Like, we did in the
previous illustration by the very nature of the total cost function, if this constraint is
binding and this has to be satisfied or if this is violated by the E O Q then when we solve
the optimization problem.
The order quantities will adjust themselves such that Q j C j by 2 will become exactly
equal to B and will not be less than B by the, very nature of the total cost curve. We have
already seen that the total cost curve is like this. The optimum is somewhere here and a
limit on the money value of the average inventory is like adding a constraint which is
like this. And therefore, the optimum will come here and the optimum will not come to
the left of the point of intersection. Therefore, when this constraint is binding it will be
an equation and it will not be an inequality. Now, the moment we understand that this is
going to become an equation then we can use the method of Lagrangian Multipliers as
we did in the earlier case and then solve it. So, the Lagrangian problem will become L is
220
equal to sigma D j by Q j C naught plus sigma Q j by 2 into i C j plus lambda into sigma
Q j C j by 2 minus B.
Now, we can partially differentiate the Lagrangian function with respect to Q and with
respect to lambda to get the solution. So, partially differentiating with respect to a Q k
and setting it equal to 0 would give us minus D k by Q k square C naught plus i C k by 2
plus lambda C k by 2 equal to 0. Now, this on simplification would give us Q j, I am
writing it for a general j, as root over 2 D j C naught into i plus lambda C j. Now, we can
compute Q j which is what we want only if we know the value of lambda because right
now in this equation all other values other than lambda are known. So, we need to find
out lambda and in order to find out lambda we partially differentiate L, the Lagrangian
function with respect to lambda. So, dou L by dou lambda equal to 0 would give us
sigma Q j C j by 2 is equal to B. So, in order to find lambda we substitute for Q j from
this and therefore we get:
Something like sigma, now Q j is root over 2 D j C naught by i plus lambda C j into C j
by 2 equal to B. So, we can take root over C naught by i plus lambda, C naught root over
C naught comes out root over i plus lambda comes out. There is a root 2 here, there is a 2
here. So, there is a root 2, so which comes under this square root. Then sigma root of D j
C j is equal to B. There is a root C j here, there is a C j here so root over this. So,
squaring we get C naught by 2 times i plus lambda into sigma root over D j C j square is
221
equal to B square, i plus lambda is equal to C naught by 2 B square sigma root over
whole square from which lambda is equal to C naught by 2 B square sigma whole square
minus i. So, this is the expression for lambda and then we can now substitute for the
values here to try and get lambda and then we can use this lambda substitute it here to
get the value of the Q j.
So, now we apply this to the numerical illustration that we have. Therefore, applying this
numerical illustration lambda is equal to 300 divided by 2 into 25,000 square. We have
now kept the B as 25,000. So, 25,000 square to root of 10,000 into 20 plus root of 20,000
into 25 the whole square minus point 2. And, this on simplification would give us
lambda equal to 0.12 and once lambda equal to 0.12 we can go back and calculate Q 1
and Q 2 out of this. So, Q 1 will be equal to root over 2 into 10,000 into 300 divided by
0.32 into 20. The 0.32 comes to out of i plus lambda. (Refer Slide Time: 09:17) So, i is
0.2 lambda is 0.12, so 0.32 into 20. So, Q 1 on computation is 968.5647 and Q 2 is
1225.148.
So, we observe that the order quantities have come down. They have become 968.5647
and 1225.148 as against (Refer Slide Time: 00:54) 1224.74 and 1549.19. We can also do
this simple calculation and say that 968.5647 into 20 divided by 2 which is Q 1 C 1 by 2
plus Q 2 C 2 by 2 will add up to exactly 25,000 which is the restriction that we have put
here. There is also an interesting rule very similar to what we did in the last time.
222
Now, we want to bring (Refer Slide Time: 00:54) a total of 31612.58 to 25,000 it is
enough to proportionately adjust the order quantities. So, we can also prove that Q 1 by
Q 1 star is equal to Q 2 by Q 2 star is equal to B by sigma Q j star C j by 2. For example:
968.5647 divided by 1224.74 is equal to 0.7908, it is 0.7908. Now, this is the same as
1225.15 by 1549.19 and it is the same as 25,000 divided by 31,612.58. Now, we realize
that the ordering quantities get proportionately adjusted and the factor is 0.7908. It is also
possible to derive through this that the ordering quantities do get proportionately
adjusted.
Now, there are two other issues that we need to look at. One is we are able to get a closed
form solution for lambda. We are able to get an expression for lambda which if we
substitute all these values we can get the correct value of lambda which would take us
there. That has come largely because the term, (Refer Slide Time: 09:17) the constraint
term is Q j C j by 2 minus B as it goes into the Lagrangian and there is also a Q j i C j by
2. So, this term and this term are more or less similar and with the Lagrangian multiplier
I get an i and a lambda. Therefore, I get an i plus lambda which goes into it and
therefore, we are able to get a solution. In the previous case where we put a restriction on
the number of orders the constraint was of the type D j by Q j equal to n. So, there is a C
naught and there was a lambda. So, we could get a C naught plus lambda term that
comes ((Refer Time: 20:01)).
It is also important to note that such a closed form solution was possible because C
naught and i (Refer Slide Time: 00:54) are assumed to be the same for both the items. If
C naught and i were not the same for both the items and if we had a C 0 1 and a C 0 2
which were distinctly different then it is not possible to straight away get this closed
form solution for this. It will become difficult to do that. Then it comes to the issue of,
are we justified in giving the same value of C naught and the same value of i for both the
items? In a way giving the same value of i can be understood and explained because i
represents the interest on the money that is borrowed to procure these items. And, we can
assume that irrespective of the item that is procured, the interest remains the same and
therefore, the carrying cost becomes i into C and i is the same for all of them.
Now, are we justified in keeping C naught the same for all of them?? In some sense the
answer is yes because of may be two different reasons. One is: it is also extremely
difficult to individually compute the C naught for various items and if we do so at the
223
end of it we also realize that they are not very different from each other. The difference
may be very small and negligible. The other way to look at it is C naught essentially, the
components of C naught are: inspection, transportation, order placing and so on. And, if
the same people are going to do the inspection then the inspection cost cannot be
different.
But, if the inspection requires different types of equipment or if the inspection requires a
difference skill set of the person inspecting then C naught can be different. Similarly, the
transportation is essentially going to be through trucks and if we assume that irrespective
of the item that is transported, the truck cost is the same. Then it is not unrealistic to
assume that C naughts the same. So, C naughts can be assumed to be the same. However,
if some of these items require special features for example, if we require an air
conditioned truck to carry something while for some other item you do not need an air
conditioned truck then the order cost can vary.
But, for the purpose of this computation we assume that order costs are the same, i is the
same and because these constraints (Refer Slide Time: 09:17) are of very similar nature
we are able to get an expression for lambda (Refer Slide Time: 14:29) which we can use.
And, at end of it we are also able to prove through this though we have not taken the
trouble of showing it that Q 1 by Q 1 star, star is the economic order quantity, this is the
new order quantity. So, (Refer Slide Time: 14:29) new order quantity by economic order
quantity follows this proportion.
So, it gets proportionately adjusted as there is a limit on the total number of items that
are there. The other aspect that needs to be looked at is also somewhere here. (Refer
Slide Time: 09:17) We said that if this constraint is violated by the economic order
quantity then this constraint will be satisfied as an equation and not satisfied as an
inequality. Now, the understanding that this is satisfied as an equation helps us use
Lagrangian Multiplier as a methodology to solve it. So, the very idea of using
Lagrangian Multipliers is centered around the fact that we are solving a constraint
optimization problem where the constraint is an equation.
So, for example if we just, for the sake of argument or for the sake of completion if we
say that instead of (Refer Slide Time: 00:54) 31,612 we solve this problem with 25,000.
Suppose, we solve this problem with 35,000 say. If the limit on the money value of the
224
average inventory was, 35,000 instead of 25,000 then what happens? If it was 35,000 and
if we had blindly use this formula to get lambda, then we would get a negative lambda
here. If we substitute B as 35,000 instead of 25,000, we would get a negative lambda. A
negative lambda means at this point, a negative lambda means that, the constraint is
actually satisfied by (Refer Slide Time: 09:17) the economic order quantity. Therefore,
one should not blindly use this lambda. One should use this lambda, only after
ascertaining that the economic order quantity violates the condition (Refer Slide Time:
09:17) Q j C j by 2 is less than or equal to B.
If we forget that and we still use the formula blindly and if we get a negative it is an
indication that, the economic order quantity actually satisfies the condition or satisfies
the upper limit on the money value of the average inventory. If we continue to make the
mistake of using a negative lambda and substituting it here into this (Refer Slide Time:
09:17) then we will realize that since lambda is negative the economic order quantity will
go up and the economic order ((Refer Time: 25:39)) or the new order quantities will be
such that, the sum of the money value will be 35,000. Because, this lambda formula
comes from satisfying this (Refer Slide Time: 09:17) as an equation. So, this will always
push to be equal to their number that we give here (Refer Slide Time: 00:54). So, the
correct procedure is to solve for the economic order quantity and only when, the
economic order quantities violate the condition (Refer Slide Time: 09:17) Q j C j by 2 is
less than or equal to B, then we make it equal to B and then use the method of
Lagrangian multipliers.
So, Lagrangian multiplier method should not be used blindly, for every value that is
given here. (Refer Slide Time: 00:54) It has to be used only in the cases where, the value
given here is less than the value given here. Otherwise, in its anxiety to make this (Refer
Slide Time: 09:17) as an equation it will end up pushing the order quantities up, which is
not desired. (Refer Slide Time: 00:54) And, if this were 35,000, then the economic order
quantity which gives 31,612 is indeed optimum and therefore, there is no question of this
constraint coming into play at all.
So, all these constraint multiple inventory problems we first solve for the economic order
quantity and only when the economic order quantity violates the condition, we use the
method of lagrangian multipliers and we solve them. Other thing is, when (Refer Slide
Time: 00:54) the constraint is binding as in this case when we compute, you find here
225
that lambda is 0.12, i is roughly 0.2. So, in all these, in this particular problem when
there is a restriction on the money value and if the constraint is binding, lambda will
usually be in the order of i, for example, we will not get lambda equal to 400 and so on.
It will be of the order of i.
In the earlier example where, there is a restriction on the number of orders lambda would
be of the order of C naught because the term here (Refer Slide Time: 09:17) is i plus
lambda. So, lambda will be of the order of i. In the earlier illustration the term was C
naught plus lambda. So, lambda would be of the order of C naught that is another little
hint that we have. So, lambda would be of the order of i. So, let us now move on to
another type of constraint problems.
Now, let us look at the 3rd type of a constraint, where we place a restriction on space.
Let us assume that there is a space restriction to store the items. Now, let us assume right
now, that the spaces required are 3 and 4 respectively. So, let me call this as S j which is
the space required to store a unit of this item and this is space required to store another
unit of this item. So, let us assume that these are in say cubic feet and this requires space
of 3, this requires space of 4.
So, what we are going to do now is the average space required for these two items. So,
we will do Q j S j by 2, which represents the average space required for each of these
items. So, Q j S j by 2 on computation would give us, 1837.125. So, this would give us
226
1837.125 and Q j S j by 2 for the other item would be 3098.38. So, this gives us 3098.38
and the total space that is required is 4935.5. So, let us assume that the space required
for, storing the average inventory value of both these items is 4935.5 and let us assume
that, we have space of only 4000 to store these items. Now, how do we solve this
problem? Now, before we actually solve the problem, let us look at one small aspect of
the whole thing. Now, here we have assumed the space requirement for the average
inventory and we have not looked at this space requirement for the entire inventory that
arrives.
So, there is this question whether, the space requirement is for storing the average
inventory or whether the space requirement is for storing the entire inventory that we
have. Now, we use average inventory simply because there are multiple items and while
some of these items will have very large inventories, some of them will have smaller
number of inventory. And therefore, we are justified in assuming that at any given point
in time, we are only going to have a certain average inventory in the system, whose
storage space is what we are looking at. Sometimes, we may take an extreme view and
say that we should have enough space, when the entire consignment comes and which
means we need Q j S j and not Q j S j by 2. So, all these numbers get doubled but then
the methodology of solving it is not affecting by that decision. So, we will proceed by
assuming that these space requirement is for the average inventory, which means on an
average we require 4935.5 cubic feet or unit subspace and let us assume that we have a
constraint of 4000. So, now we need to solve this problem because the economic order
quantity again does not satisfy this.
227
So, we will say that I now want to minimize the sum of ordering cost and carrying cost
subject to the condition that the space required by the average inventory is less than or
equal to, some R. So, as usual we solve the unconstraint version. So, the unconstraint
version gives us the economic order quantities (Refer Slide Time: 28:48) and average
space requirement of 4935.5. Now, this is higher than the available space of 4000. So, we
once again do the lagrangian and say L is equal to and once again say that this inequality
at the optimum will become an equation, by the very nature of this. Obviously a
restriction on the space is going to increase the order quantity somewhere here, please
note that this line does not exist. So, it is going to increase the order quantity.
So, lagrangian is D j by Q j C naught plus sigma Q j i C j by 2 plus lambda sigma Q j S j

by 2 minus R. So, partially differentiating with respect to Q j, we get minus D j by Q k,
let us do it with some Q for some Q k, C naught plus i C k by 2 plus lambda S k by 2
equal to 0, from which a general Q j is equal to root of 2 D C naught by i C j 2 D j C
naught i C j plus lambda S j. Now, we do not know this value of lambda. So, in order to
find lambda, we partially differentiate L with respect to lambda to get sigma Q j S j by 2
is equal to R. Now, in this particular case we, even if we substitute this Q j into this and
try to write this equation, we will not be able to get a unique expression for lambda
because of this term i C j plus lambda S j that comes in.
228
So, we do not get a closed form solution. You do not get an expression for lambda which
we can use. So, in this case we can solve for lambda only using a trial and error method.
So, we can try various values of lambda and then check for a chosen value of lambda,
you can calculate Q 1 and Q 2 and then substitute here and that value of lambda for
which l h s equal to r h s, is the best value of lambda. So, when we apply this to this
numerical example we need to try out different values of lambda and then we observed
after an extensive amount of computation, that lambda is 0.669 in this example. So,
lambda equal to 0.669. This gives us Q 1 is equal to 999.417 and Q 2 as 1250.326. So, Q
1 and Q 2 are now different.
They decrease because (Refer Slide Time: 28:48) 1224.74 becomes 999.417 because
there is a restriction on the space so the order quantity has to come down. So, 1224.74
becomes, 999.417. Let me again repeat that for the economic order quantity the space
required is 4935, which is higher than the available space. Since, the available space is
lesser a restriction on the space will only reduce the ordering quantities. So, order
quantity will reduce from (Refer Slide Time: 28:48) 1224.74 to 999.417 and from (Refer
Slide Time: 28:48) 1549.19 to 1250.326.
So, we have seen three types of constraint inventory problems till now: one is the
constraint on the number of orders, the second is on money value and the third is on
space. And, we derived three different or we solved the problem for three different
229
instances individually, we did not look at two constraints at a time. Now, suppose we had
both this constraint: number of orders, as well as money. Say, in this (Refer Slide Time:
28:48) numerical example we said that the number of orders should be 15 and the money
should be less than 25,000. Now, if we have two such constraints then how do we solve
such a problem? Now, an obvious thing to do is to try and say that, we write two
constraints (Refer Slide Time: 32:34) and we could bring two Lagrangian Multipliers and
bring them into the objective function and then solve it so that is methodologically
correct. But then let us also try and see what happens. So, if there is a restriction on the
number of orders, then the order quantity will increase.
So, this is the economic order quantity. So, order quantity increases and you have
something like this. So, the optimum will be shifted to the right, which means the order
quantity increases. If there is a limit on the money value and the economic order
quantities do not satisfy that, then that would make the order quantities decrease from
this side and the order quantity will shift to the left. So, if we have a situation where the
economic order quantity is violating (Refer Slide Time: 37:36) both this restriction and
this restriction, then we cannot have a feasible solution to the problem. So, we need not
take the trouble of writing two constraints and taking it into the lagrangian and
differentiating it. If the economic order quantity violates both these, then it is obvious
that the problem does not have a feasible solution at all.
230
(Refer Slide Time: 37:36) Now, we could think in terms of number of orders and space.
Space will behave the same way as inventory behaves. Because, if there is a restriction
on the money value, (Refer Slide Time: 28:48) the economic order quantity, the order
quantities come down. If there is a restriction on the space the order quantities come
down. (Refer Slide Time: 37:36) So, if we have a restriction on the number of orders and
then another restriction on the space and the economic order quantity violates both of
them, then once again by the same argument the solution will be infeasible because this
restriction will try to move it in this direction, (Refer Slide Time: 37:36) space will try to
move it in the other direction and therefore, you will not have a feasible situation.
So, the only thing that we could think, in terms of is if we have a restriction on the
money and if we have a restriction on the space (Refer Slide Time: 37:36) both try to
move it in the same direction. And therefore, we will have a solution, when we have
restriction on the money, as well as restriction on the space. When we have such a
situation it is actually easier to solve (Refer Slide Time: 37:36) for this 1st and then we
try and see whether this solution is satisfied by this. If so it is acceptable. If not, we have
to solve for this and the solution will automatically satisfy this. Because, solving the
problem with the money restriction is easier because we have closed form solutions, the
proportionality rule also works.
So, it is easy to solve (Refer Slide Time: 37:36) for this 1st and then check whether that
solution satisfies this. Then it is optimal to both, otherwise we have to solve for this and
it will automatically satisfy the other one. So, that is the way we take care of multiple
items. We do not prescribe to have a situation where (Refer Slide Time: 32:34) you have
more than one constraint and then you take more than one Lagrangian multiplier. The
labor required is very high and it gets highly complex. Whereas, always solving it as a
single constraint problem, one at a time is a lot easier than building two constraints into
the problem. Now, let us look at one more version of this problem. Now, particularly
when we have multiple items, we would also like to consider another aspect when we
have multiple items.
231
Now, let us assume that multiple items essentially come from the same vendor. Now, let
us assume that these two items come from the same vendor who supplies both of them.
Now, if we chose to order the economic order quantities of each of these, then we have
already found out that this will be some 8.17 orders and the other one will be roughly
about 8.16 and 12.91 orders. So, 8.16 orders and 12.91 orders per year and we also know
that, the ordering cycles do not synchronize. For example, we will be placing orders
232
every 47 days here and roughly for about 30 days or 29 days in this one, when the order
cycles do not synchronize.
So, if we look at the ordering process, the ordering process particularly it is the order cost
in particular is largely dominated by, the cost of transportation and the cost of inspection.
Now if we, since both these are coming from the same vendor, if we are able to order
them together and make the number orders equal so that the ordering frequencies are the
same. Now, both these can come in the same truck from the supplier, to the factory. The
inspection costs still have to incurred for this as well as this but the transportation cost
can be significantly reduced because both can come in same truck. If they are ordered
differently, we know that this has to come separately, this has to come separately. So, the
cost of transportation is twice, where as if we do this the cost of transportation is single.
Therefore, what we do to take advantage of multiple items that go to the same vendor?
(Refer Slide Time: 42:38) We now split the order cost into two components: one is called
the administrative cost and the other is called the truck cost.
So, if we combine two orders we still incur two administrative costs but then we incur
only one truck cost. Now, there is a saving in the total order cost when items are order
together and when they come from the same vendor. Now, is this economical or if it is
economical, can we implement it is the other question that we would like to look at.
Now, let us assume that, the order cost is made up of two components. Now, this is
called the administrative cost which we call as C 0 1 and this is called the truck cost,
which is called as C 0 2. Now, let us assume that we order both these items together.
Now let there, let the number of orders be n. So, there are n orders that are going to be
there.
233
Now, when we order them together the total order cost C naught will be 2 times C 0 1
plus C 0 2 because there is going to be (Refer Slide Time: 42:38) only one truck cost,
there will be two administrative costs. So, if we order both together and there are n
orders the total cost will be, (Refer Slide Time: 42:38) if there n orders per year, then the
total cost will be n into 2 times C 0 1 plus C 0 2. This is the total ordering cost per year.
The total carrying cost per year will be, if the demand is D j for item j, then the order
quantity Q j is D j divided by n because there are n orders per year. So, Q j will be D j by
n and the inventory holding cost is Q j by 2 into i C j so Q j by 2 into i C j, summation of
this. Now, we need to find out the best value of n. So, that can be got by differentiating d
T C by d n and setting it equal to 0. So, this would give us 2 times C 0 1 plus C 0 2
minus i C j by 2 or i sigma D j C j by n square is equal to 0. Now, this n will go in the
differentiation so you will get 2 C naught 1 plus C naught 2, i can be taken out, i by 2 can
be taken out sigma D j C j is a constant.
So, it will remain here, 1 by n will give us minus 1 by n square is equal to 0. From this,
we will have n square or this is equal to so n square will be i into sigma D j C j by 2 into
2 times C 0 1 plus C 0 2. Take this to the other side, n square goes here, this quantity
comes down. So, n square is i sigma D j C j by 2 times this and n is equal to root over i
sigma D j C j by 2 times 2 C naught 1 plus C naught 2. Now, in this D j is the annual
demand for item j, C j is the unit cost for item j. The order cost has two components,
which is the administrative cost and the truck cost. By combining the orders we incur 1
234
truck cost and 2 administrative costs and n is the equal number of orders that we have.
So, we can use this formula to get this n and then we explain this through a numerical
example in the next lecture.
235
Prof. G. Srinivasan
Lecture - 14
Multiple Item Inventory – Combining orders, Production Consumption Model.
In the last lecture, we derived an expression for the number of orders n when we
combine orders. The basic idea behind this equation for n is that, when we have multiple
items and if we ensure that, the order quantities are such that they have equal orders,
equal number of orders. If the order quantities are adjusted such that, both the items or
all the items have equal number of orders, then it is possible to save on the order cost.
Now, the order cost which was originally defined as C naught in our earlier models.
Now, has 2 components, one which is called C 0 1and the other is called C 0 2. Now, C
0 1 represents the admin cost and C 0 2 represents the truck cost.
Now, if the items are ordered such that, the number of orders per year for each item is the
same, which means the orders are combined. Then, it is possible now, to save on the
truck cost. So, there is only an admin cost for every item that is ordered. But, then we can
assume that the same truck will bring all the material from the supplier, to the plant or
the manufacturing facility. Therefore, the total order cost instead of becoming k times C
0 for each of the k items, assuming that there are k items. Now, it becomes C 0 2 plus
236
k times C 0 1 because we assume now, that the truck has enough capacity to bring all the
items together. So, there is a saving on the truck cost.
Now, this particular formula that we derived in the earlier lecture was for two items.
Therefore, we had 2 times C 0 1 here, this 2 comes from the average inventory.
Therefore, this 2 remains. But, then if in general if the number of items is k, then this can
be rewritten with k times C 0 1 plus C 0 2. Now, we take a numerical example, to show
the effect of this joint ordering. So, let us look at a situation where there are 3 items and
the demands are 3000, 5000, and 20,000.We have the truck cost as 3000 and admin cost
as 1000. So, C 0 1 which is the admin cost is 1000 and C 0 2 which is the truck cost is
3000 and then we take that, i is 20 percent, 200, 200 and 400. So, i is 20 percent and the
unit price C is: 100, 200, 400 respectively. So, we look at the 1st situation, when the 3
items are ordered separately according to their economic order quantities.
And if we do that, C naught for each item will be the sum of the admin cost which is
1000 and the truck cost which is 3000. So, C naught will be 4000 and for the 1st item Q
1 will be, root of 2 into demand 3000 into 4000 divided by 0.2 into 100 which on
computation becomes 1095.44 and total cost T C 1 which is the sum of order cost and
carrying cost for this item is 21908.9. In a similar manner, if we calculate Q2 and T c 2
we find that Q2 turns out to be 1000 and T C 2 turns out to be 40,000, Q 3 is 1414.21 and
T C 3 is 113137.08. So, individually if we order the 3 items according to their economic
237
order quantity. Now, these are the order quantities for the 3 items and the sum of order
cost and carrying cost for the 3 items are given here. So, this gives a total cost for the 3
items put together which becomes 175045.98.
Now, this is the total cost incurred if they are ordered independently and separately.
Now, if we apply this formula and assume that, they can be ordered jointly. And there is
a saving on the truck cost if we order them jointly. Applying this formula (Refer Slide
Time: 00:15) for k equal to 3. There are 3 items rest of the values are known here. We
will compute n to be equal to root of i which is 0.2 into d 1 C 1 plus d 2 C 2 plus d 3 C
33000 into 100 plus 5000 into 200 plus 20,000 into 400 divided by 2 times 3 (Refer Slide
Time: 00:15) into C 0 1 is 3 into 1000 3000 plus another 3000.
So, 6000 this on computation would give us, n equal to 12.45 and T C for each of these
items, if we calculate the total cost and the total. Now, for n equal to 12.45, we can
actually calculate the individual order cost, carrying cost and then do it. Another way of
looking at T C is, if we have n orders. For each order the ordering cost is going to be this
(Refer Slide Time: 00:15) 3000 plus 3 into 1000 so it will be 6000, so n into 6000 plus Q
by 2 into C C. So, summation Q is D so D j by 2 n Q equal to D by n. We have 12.45
orders.
So, the order quantity for each item is the annual demand divided by 12.45. So, D j by n
into Q into i C j which is the inventory carrying cost. So, this for n equal to 12.45. Now,
becomes 149398.80. So, this example tells us that, if we are incurring a (Refer Slide
Time: 00:15) 1000 rupees on the admin cost of ordering each item and there is a truck
cost of 3000. And by joining the orders and by making sure that, all 3 items have the
same number of orders per year. Now, we now have 12.45 orders for each of them and
then we realize that the total cost has actually come down from 175045 to 149398.8.
There will be a reduction in the order cost because of the saving. But, then there will be a
slight increase in the inventory carrying cost, because the order quantities will change.
But, then the sum of the order cost and carrying cost comes down significantly. Now,
what are the other issues related to this. Now, from our total cost point of view there is a
considerable saving, if we can combine orders together. Provided there is a saving in the
truck cost of 3000 every time items are combined together for example, when an item
was ordered individually, the cost was 4000. But, then when two items are combined, the
238
order cost was not 8000 but the order cost was only 5000. When three items were
combined, the order cost was not 12000.
But, then it was 6000. So, that saving helped us in reducing this considerably. Then,
there could be some other issues. For example, let us compute the order quantities for
12.45 so when you have 12.45 we realize that, the order quantity Q 1 is 3000 divided by
12.45. So, the order quantity Q 1 becomes 240.96 Q2 becomes 386.1 and Q3 becomes
1606.43 and all of them have 12.45 orders. Now, if we compare the order quantities in
the unconstrained system this was 1095 and that has come down to 240.96.1000 has
come down to 386.1 and 1414 has actually gone up to 1606. Let us now, calculate the
number of orders for these values. So, N 1 is 3000 divided by 1095.44 which is 2.738
orders, N 2 is 5000 divided by 1000 which is 5 orders and N 3 is 20,000 divided by
1414.21,14.14 orders.
So, when we did not have this constraint the number of orders was 2, 5 and 14 very large
difference. When we combined all of them now, came to 12.45 orders. So, we actually end
up making more orders for these two items. Considerably more orders because 2.738 when
we did not do this, 12.45 when we did this. Similarly, 5 when we did not have this joint
ordering and 12.45 when we had joint ordering. So, when we assume that n is the equal
number of orders for all of them. We now, tend to order more for items 1 and 2 and about
the same number for item 3. May be slightly less for item 3, 14.14 becomes 12.45.
So, since the number of orders is more here, the order cost component is also more. So,
the other question that comes is. Now, instead of assuming an equal number of order is n,
can we assume unequal number of orders? But, the number of orders for each item is a
kind of an integral multiple and then can we check whether we can have some kind of
partial saving? But, yet try and reduce this 149398 further. So, if we look at this, if we
look at this alone. This is 2.738 this is like 5 and this is 14.14. So, this is roughly 1 is to 2
is to some number like 5. This is like 1 is to 2 is to 5 that is the number of orders.
Now, instead of using 1 is to 2 is to 5 suppose, we look at 1 is to 2 is to 4 instead of

looking at 1 is to 2 is to 5.If we say that, the number of orders is roughly 1 is to 2 is to
4,2.7 into 5 is about 13.5. So, it is roughly 1 is to 2 is to 5 but then 1 is to 2 is to 4 is a
very comfortable integral multiple. So, we could think in terms of 1 is to 2 is to 4.
239
And then we say that item 1 will have n orders, item 2 will have 2 n orders and item 3
will have 4 n orders, which simply means that, for every order here, we will have 2
orders here and 2 orders here. So, this order n will synchronize with this and this. And
this 2 n will also start synchronising with this 4 n. But, then the order quantities will be
suitably defined and you can still exploit the basic idea of joint ordering, because the
number of order is only an integral multiple of this.
Now, when we start doing this what is the gain that we have. So, the gain that we would
have is this. Now, the gain will be the total cost for this system will be. So, totally there
will be n orders in a year for this, there will be 2 n orders in a year for this and there will
be 4 n orders in the year for this. So, for all these n orders so there will be a total of 4 n
orders in a year, out of which this 2 n will be synchronized within this 4 n. And this n
will also come within this 4 n.
For, example if we say that 4 n is equal to 16 and n is equal to 4 it, is like saying that the
1st item will be ordered 1 2 3 4 and then it goes, this is the next year. Now, this item will
be ordered here, 1 2 3 4 5 6 7 and there is a 8. Now, this will be ordered 1 2 3 4 5 6 like
this it will go. So, if there are 4 n orders, in all in a year. Now, for n out of these 4 n, all
the 3 will be ordered together. For another n out of this 2 n, 2 will be ordered together.
And for this 2 n out of the 4 n, there will be a single order. So, the total order cost will
240
be, for n out of these 4 n the order cost will be 6000n, because when 3 items are ordered
together, the truck cost is 3000 plus 3 times admin cost of 1000 will give 6000.
So, 6000 into n, for another n orders only these 2 will be combined. So, this will become
5000 into n and for the remaining 2 n orders this alone will go. So, this will be 4000 into
2 n so this will be the total ordering cost that we will have. Now, the inventory carrying
cost will be. If I have n item, n orders for item 1. The order quantity will be (Refer Slide
Time: 00:15) 3000 divided by n. This is Q so Q by 2 into carrying cost is C c which is i
into C which is 20 percent (Refer Slide Time: 00:15) of 100 which is 20. For the 2nd
item, the order quantity will be 5000 divided by 4 n. 5000 is the annual demand, 2 n is
the number of orders. So, 5000 by 2 n is Q and Q by 2 is 5000 divided by 4 n into C C
which is 20 percent of 200 which is 40 plus 20,000 divided by 8 n.
The 8n comes because demand is 20,000. There are 4 n orders so 20,000 divided by 4 n
is Q that divided by 2 - Q by 2 into 0.2 into 400 which is 80. So now, T C is equal to 6
plus 5, 11 plus 8 19,000 n plus this is 30,000 divided by n. This 2 gets cancelled into 10,
this 4 also gets cancelled plus 50,000 divided by n plus 200,000 divided by. So, this is
19,000, n plus 200 and 80,000 divided by n, from which the optimum value of n can be
found. By differentiating this with respect to n and setting it to 0. So 19,000 minus 200
and 80,000 by n square is equal to 0.From which n is equal to root over 200 and 80,000
divided by19,000 which would give us 3.838.
Now, this means item 1 will have 3.838 orders item 2 will have 7.676 orders and this will
get 15.352 orders. So, the order quantity for this will be 3000 divided by 3.838, which is
roughly about less than 1000. This will be 5000 divided by 7.676, 20,000 divided by
15.352 and now, when we substitute this n, into the total cost at the optimum. We now,
realize that this n for this n, the total cost is 19000, 72938.33 plus 72954.66 which is
145892. In fact, this difference, slight difference comes more because of the rounding of
errors. And at the optimum, we actually would realize that this 19,000 n will be equal to
280,000 divided by n.
These two, will have to be equal. The slight difference comes between 38 and 54 is in the
rounding off, this 8.38. Otherwise, these two values will be equal and it will be 2 times.
Now, this would ensure 1st thing we observe, is that the cost has further come down
145892 versus 149398. The other advantage that we have right now, here is that the
241
number of orders are staggered and somewhat proportional to the annual demands of
3000, 5000 and 20,000. So, roughly all of them will have order quantities in some way
or around 800 to 1200. The number of orders will be different. The average inventory
that, we hold will also come down.
So, that is the advantage of using these kind of staggered values. But, then we now, ask
ourselves another question. Now, we first worked out based on individual ordering
quantities which is here 1750, then we worked out assuming that all of them will have
equal number of orders which brought it down to 1493. Then, we worked out with
unequal number of orders and then we found a small decrease and could perhaps help us
in little more control of the order quantity. The number of orders will be different. So,
the next question is, if we do say if we do 1 is to 2 is to 4, the cost is 145892.
Now, what is that optimal combination of say k is to l is to m or what is the ratio? Is it 1

is to 2 is to 4 or is it 1 is to 2 is to 2? Or if we try 1 is to 2 is to1, or 1 is to1 is to 2? What
kind of answers would we get? And can we optimize this number further by trying
different combinations of these ratios? Very basic idea comes from the fact that, when
we have the unconstraint problem (Refer Slide Time: 04:01) 2.7, 5, 14. The first thing
that comes out is that, the number of orders kind of increase in the 3 item. So, 1 is to 1 is
to 1 gave us 1493, 1 is to 2 is to 4 would give us 1458. So, what is that ratio? What is that
k is to l is to m, such that we can bring this further as well as, we can exploit this very
idea of joint ordering.1 is to 2 is to 4 helped us in a certain way.
But, 1 is to 2 is to 4 actually increase this thing 19000, because out of the 4 n orders 2 n
orders are done individually and separately which gave us which brought this term 4000
into 2 n. Now, if we do 1 is to 1 is to 2 or if we do 1 is to 2 is to 2, then we realize that
actually this part will come down. Because, if we do 1 is to 1 is to 2, there will be 2 n
orders, out of which n orders will be joined and the other n orders will be individual. So,
we could try some of these combinations. There are some intelligent ways of trying, to
find out. What is the ratio? Is it 1 is to 1 is to 2? Or is it 1 is to 1 is to 4 and so on.
Generally multiples are advantageous 1 is to 2 is to 4 is always advantageous compared
to 1 is to 2 is to 5.
Because if we do, 1 is to 2 is to 5 then we would not be able to overlap orders. When the
number of orders is an integral multiple then it is always easier and if you have multiple
242
items, if the last one is an integral multiple of all of them, it is like an l c m of all of them.
Then it is actually advantageous, something like 1 is to 2 is to 5 or 1 is to 3 is to 5 is not
of great advantage whereas, 1 is to 1 is to 2 or 1 is to 2 is to 2 can be of certain
advantage. Actually we can go on expanding on it but let me just give you one simple
solution and leave it at that.
In fact, if we do 1 is to 1 is to 2 which means n, n and 2 n.We do 1 is to 1 is to 2 for this

particular problem, we will have n equal to 7.28 and total cost actually comes out to
145602.19, which is a slight decrease compared to 145892. The other thing that, we can
observe though it is difficult to generalize is that, once we reach this number like
1458.92, we were able to show a slight decrease. But, then we will not get too much of a
gain. It its little unlikely to have a solution which is 1375 or whatever, so the decrease is
only very minimal and one can try and look at you know, other alternatives like 1 is to 2
is to 2 or 1 is to 2 is to 3 and so on and try and see whether we could make some gains,
though 1 is to 2 is to 3 is not a great advantage compared to 1 is to 2 is to 4.
But, there are intelligent ways of trying to find out the ratios, we define something called
m j where m 1, m 2 and m 3 are the number of orders per year, as a multiple of n and
then they can try and optimize that m to a limited extent and try and optimize this cost
further. But, many times a very commonsensical approach of trying to find out these
values for example, (Refer Slide Time: 04:01) this 2.7385 and 14.14. The simple ratio of
243
these three rounded to the nearest integer, an integral multiple like 1 2 and 4 (Refer Slide
Time: 15:44) invariably gives a very good answer, though from an optimality point of
view, one would choose this compared to this. It is also easy to play around here, from
this to try and get 1 is to 1 is to 2 from 1 is to 2 is to 4.
So, these are strategies that are being used particularly, in the context of supply chain
management, the reason supply chain management is because we are now, combining
orders. So, that we can save on (Refer Slide Time: 00:15) the truck cost today, with
supply chain management and logistics becoming increasingly important. The
transportation is usually outsourced to 3rd party service providers, take up the task of
transporting from vendors to the manufacturers and since, they can combine orders from
the same vendor to different destinations or from different vendors to the same
destination and they can do some kind of a milk run and so on.
So, the whole idea of supply chain management and logistics leading as towards 3rd
party service providers, an outsourcing of the transportation function allows
organizations to now, go more for joint ordering (Refer Slide Time: 00:15). So, that the
truck cost can be saved and the total inventory cost can be brought down. Now, in such a
process sometimes, we may end up getting a large order quantities for some items and
small order quantities for some other items. Large order quantities for some items would
mean quite a lot of money, is locked up in inventory. And that also can be offset by
trying to order like (Refer Slide Time: 15:44) 1 is to 2 is to 4 and 1 is to 1 is to 2.
So, with this we have now covered certain aspects of deterministic inventory. Now, we
would look at another aspect which is production as well as consumption. So, we will get
into models, where we produce and consume. So, far in inventory we have seen, two
very basic inventory models and then we looked at discount and then we looked at
constrained inventory problems and then we looked at trying to optimize on total cost by,
saving on the truck cost. Now, let us spend some time on what are called production and
consumption models.
244
Now, let us look at a single item which is not bought out, but it is produced. Now earlier;
in the earlier inventory models we assumed that, we are going to buy the item from a
vendor and therefore, every time an order was placed there is an order cost and then the
items come in and then the items are stored. So, there is an inventory carrying cost or a
holding cost. Now, let us assume that we are not going to buy this item, instead we are
going manufacture this item internally and then consume it. So, let us assume that a
single machine or a single facility is going to make this particular part or this item. And
let us assume that the annual demand for this item is D, the rate of production is P per
year.
Now, if we have to meet the annual demand of this item, obviously our production rate
has to be higher than the demand. So, produce at P to meet the annual demand of this
item D. And since we have to meet the annual demand D, it is only fair to assume that
the production rate is higher than, the demand and therefore, P is greater than D. Now, let
us assume that P is slightly higher than, D and in fact much bigger compared to D. So, let
us assume we start producing this item so we produce this item and as we produce, we
consume this item. Therefore, the inventory is going to increase at the rate of P minus D,
because we produce at the rate of P and while producing we consume a D. So, the
inventory build-up is going to be at the rate of P minus D.
245
So, let us say we build up this inventory, up to a certain point and then we stop producing
this item and then we start consuming this item, only from the inventory that has been
built up. So, then from this, let us assume we stop production here, at this point. and then
we start consuming this item. So, we consume it at the rate of D and then the inventory
comes to 0 and we assume that the moment, the inventory comes to 0, we start producing
this item again. So, when we start producing this item again, we will again have this up
to P up to a particular point and then we use this and this is how the inventory cycle goes.
So, note that these two points are actually the same.
So, and so now, we assume that, in between during this period, when we stopped
production and then we again start production this is called t 2. This is called t 1, t 1 is
the period where we produce, which means we produce and consume from an inventory
point of view, t 2 is the period where we only consume the item from the inventory point
of view from the machines point of view t 1 is the period, where we produce this item
and once again. There is going to be another t 1 after a gap of t 2, where we produce this
item again. So, we will assume that this machine would be used to produce some other
thing, during this period t 2.
And therefore, when we come back and start making this item again on this machine
somewhere, here we need to setup this machine, which means every time we start
production of the item, we have to setup the machine to produce the item. And therefore,
there is a setup cost associated with this system. Now for the sake of uniform notation,
let us assume that, the setup cost has the notation C naught per setup, Rupees C naught
per setup. So, every time this machine is setup to produce this item, there is a C naught
per setup. This C naught, is very similar to the ordering cost C naught. Except there it is
the cost occurred internally, when the machine is being setup for the sake of uniformity
to show some other things.
We assume the same notation C naught as the setup cost, every time this machine is
setup. Now, we have to find out whenever, we produce a certain quantity Q and what is
the economic batch quantity. When we were ordering this item the Q was called
economic order quantity. Now, when we produce this Q is called economic batch
quantity, it is called E B Q it was called E O Q later. Now, let us assume that we
produced a quantity Q. Every time we setup this machine the annual demand is D.
246
Therefore, the number of times we setup this machine will be D by Q or the number of
batches we produce would be D by Q. Every time we setup we incur a setup cost of
C0so, we incur a cost of D by Q into C 0 which is a very familiar term which we have
used earlier. Now, this is the total setup cost plus there is going to be an inventory cost
associated here. Now, this is the cycle starting from 0 to a cycle T (Refer Slide Time:
33:31) which is t 1 plus t 2 another cycle. So, inventory total inventory in the cycle as
area under the triangle, average inventory will be area under the triangle, divided by the
base. So, average let us call this quantity as Im which is the maximum inventory that we
have built.
So, area of the triangle is half into I m into t, divided by another t is the average. So, the
average inventory will be I m by 2 and the inventory holding cost will be rupees C c per
unit per year. So, the total cost associated here, is the sum of the setup costs as well as
the sum of the inventory holding costs. Now, we need to write this I m as a function of
Q. Q is the only variable we need, to write this I m as a function of Q. Now, Q is the total
quantity that we produce, I m is the maximum inventory, that we actually built it. (Refer
Slide Time: 33:31) So, we also know that, as we produce at the rate of P, there is a
consumption at the rate of d.
So, this inventory is building up at the rate of P minus D. So, I m is equal to the
maximum inventory, that we have which will be P minus D into t 1.So, I m is equal to P
247
minus D into t 1, t 1 is the time period for which we produce. Inventory is building up at
the rate of P minus D. So, I m is P minus D into t 1. Now, Q is the total quantity that we
produce. Now, we produce at the rate of P for a period t 1. So, Q is equal to P into t 1
right P into t 1. Now, I now, dividing 1 by the other Q by I m or I m by Q is equal to P
minus D by P or I m is equal to Q into1 minus D by P. Now, substituting for I m, we get
D by Q C naught plus Q by 2 into 1 minus D by Pinto C c. Now, we need to find out the
best value of Q which minimizes the total cost.
So, this is got by differentiating d T C by d Q equal to 0, would give us minus D by Q

square C naught plus C c into 1 minus D by P divided by 2 equal to 0, from which Q is
equal to root over 2 D C naught by C c into 1 minus D by P. So, this is the best value of
the economic batch quantity, such that we optimize the total cost. Now, once we have
this equation, we can find out the economic batch quantity. Given values of D C naught
Cc etcetera. So, let us take a numerical illustration and show the computation of the
economic batch quantity and the associated total cost.
So, let us look at a illustration where we have D is 10,000 per year, C naught is 300 C c
is 4 and P is 20,000 per year. Now, note that D and P should have the same unit, because
we have a 1 minus D by P coming in. So, if D is per year, P is also per year. So, Q
economic batch quantity is given by root over 2 into 10,000 into 300 divided by 4 into1
minus D by P or it is 1 minus 1 by 2,1 minus 10,000 by 20, 000, so one minus one by 2.
248
So, this will give us 4 into 10,000 into 300 divided by 4, which is root over 10,000 into
300 this will be root 3 into 1000 this is 1732.05. So, every time we produce, we produce
quantity equal to1732.05.
The total cost will be D by Q 10,000 divided by 1732.05 into 300 plus Q by 2 1732.05
divided by 2 into 4 into 1 minus D by P, 1 minus half. So, this part will be equal to
1732.05 plus into Q by 2 into C c into 1 minus D by P. So, this is half. So,2 into 2, 4 and
this 4 will get cancelled plus another 1732.05 which is 3464.10. So, when we produce,
we produce a quantity of 1732.05. We also produce for a time t1. t 1 is equal to from
here t 1 is equal to (Refer Slide Time: 39:28) Q by P which is 1732.05 divided by 20,000
which is 0.0866 years. Now, this can be converted into hours to know, what is the
number of hours, that we produce. Similarly, t 2 can also be calculated and the cycle time
t (Refer Slide Time: 33:31) can also be calculated.
Now, t 2 it is very easy to write, t 2 is P 1 is equal to P, I m is equal to P minus D into t 1

I m is also equal to D into t 2. The reason being this, much inventory (Refer Slide Time:
33:31) that is built up as I m is now, consumed over a period t 2 at the rate of D. So, I m
is equal to D into t 2 from which t 2 can be calculated and the cycle time t can also be
calculated. Now, t 2 now, in order to calculate t 2, we first need to find out, I m. So, I m
is equal to Q into 1 minus D by P. So, Q into half so I m will be 866.02 that is, that is not
249
t 2. So, I m is equal to Q into 1 minus D by P. So, I m is 866. t 2. So, t 2 is equal to I m
divided by D which will be another 0.0866 years and the cycle t will be 0.1732 years.
So, in this particular example, when we draw the graph you realize that t 1 is equal to t 2.
Need not be the same for all other kind of situations. So the next model that we should be
looking at is, can we allow for some backordering here like we did in the bought out
items? Can we allow for some back ordering? Build up a backorder and then produce use
the, satisfy the backorder first and then build up the inventory. So, that would be a
production consumption model, with backordering which we will see in the next lecture.
250
Prof. G. Srinivasan
Lecture - 15
Inventory- Production Consumption Model with Backordering, Economic Lot
Scheduling Problem
In the previous lecture, we looked at models which are production and consumption
models, which means if we consider a particular item we can produce this item at the rate
of P per year, while the demand for this item is D per year, so obviously in order to meet
the demand our production rate P has to be greater than D or greater than or equal to D,
or most of the times P is greater than D. So, when P is greater than D, we produce this
item and while we produce we also consume and since, the production rate is higher than
the consumption rate, there is an inventory buildup rate at the rate of P minus D here and
we produce up to a particular time t 1. And then we stop production and consume from
the inventory which has built up.
So, we stop production here after t 1, this is time, this is inventory but this graph is for
the item and not for the facility that is producing the item. So, as far as the item is
considered inventory is built up at the rate of P minus D for period t 1. Production stops
and then this inventory is consumed at the rate of D, till the inventory position reaches 0
251
then the cycle starts all over again. So, once again from here it is again produced for a
period t 1 and again consumed for a period t 2 and so on. So, we derive an expression for
an economic batch quantity which was called E B Q very similar to E O Q, we also said
that there is a set up cost that is involved every time we start producing.
Because, if we draw a similar figure for the facility that is producing, then the facility is
busy making this item up to period t 1 and then the facility again starts making this item,
once again at this point and is busy for a period t 1. So, in some sense the facility is not
busy with respect to this item for this period t 2 and if we assume that, the facility will be
used to produce some other item then while coming back to this item; at this point in
time we need to change over. And therefore, there is a change over cost or a set up cost
which was denoted by C naught to be consistent with the earlier notation.
So, we derived expressions for the economic batch quantity. Now, we will look at what
happens to the economic batch quantity, if we allow back ordering. In the earlier version
or in the earlier model we did not allow backordering. So, we started producing as soon
as the stock reaches 0 so there is no back ordering. Now, if we allow backordering for
the same item now, this is time, this is inventory. Let us assume that the inventory is at 0
so we produce, produce as well as consume so inventory is growing at the rate of P
minus D. Now, we stop production at this point and start consuming let us say, we come
to 0 unlike the earlier model, we allow a certain backorder in this case.
So, we do not produce immediately we allow a certain backorder and then start
producing at this point so production once again begins at this point, this is parallel to
this and then once again we consume so this will be parallel to this and so on. So, now let
us define certain notation for this so let us call this as t 2 is the period where it is
produced and consumed and there is a positive inventory buildup. Production stops here
so there is a period t 3 where, the available inventory is consumed. So, at this point the
inventory becomes 0, then there is a backorder that is built-up now, that happens for a
period t 4 where, there is no production; there is consumption but there is no stock to
consume therefore, there is backorder.
And then once a backorder is built up here, we start producing here which means there is
a back order, we produce first for a period t 1 produce and consume for t 1 and also to
meet the backorder and at this point the stock becomes 0, then there is a positive
252
inventory buildup which goes on. So, the cycle essentially is from here to this point,
which we can call as T which is the cycle. So, the cycle comprises of 4 time periods t 1, t
2, t 3 and t 4; where there is production during periods t 1 plus t 2 and there is
consumption for periods t 3 and t 4. So, this production consumption model with back
ordering.
Let us also add some more notation to it, inventory is built up at the rate of P minus D
here, it is consumed at the rate of D here, it continues to be consumed at the rate of D
here and is once again built at the rate of P minus D in the other case. We will also define
I m as the maximum inventory that we have in this system and we will define s, which is
this height as the back order that we build in this system.
Now, let us try and write down the necessary total cost equation for this. There are three
costs, there is a cost of set up, there is a cost of inventory and there is cost of shortage.
So, if we produce in this cycle T, if we produce a quantity Q, let Q be the quantity that is
to be produced. So, total cost is made up of 3 terms which is a set up cost, the inventory
holding cost, as well as the backorder cost. Now, the cycle time is T and every time we
produce, we produce a quantity Q so number of times we set up will be D by Q, that is
the number of times this item is set up in a year is D by Q because D is the annual
demand. C naught by our definition is the setup cost. So, D by Q C naught is our total
annual setup cost, then we have to model the inventory holding cost, as well as the
253
backorder cost. (Refer Slide Time: 00:09) So, we said the total inventory that is held in a
cycle is the area under this curve, which is half into base into height so half into I m into
t 2 plus t 3 that is the total inventory.
So, total inventory held is half into I m into t 2 plus t 3 but in every cycle, the cycle
length is T so this much inventory is held for a period (Refer Slide Time: 00:09) t 2 plus t
3, 0 inventory is held for the period t 1 plus t 4. So, average inventory will be half into I
m into t 2 plus t 3 plus 0 into this divided by T so we get divided by another T. Average
inventory is total inventory by time so total inventory of half into I m into t 2 plus t 3 is
the total inventory held; that is held for a time period t 2 plus t 3, 0 inventory is held for
that t 3 plus t 4.
So, total inventory is half into I m into t 2 plus t 3, total time is: t 1 plus t 2 plus t 3 plus t
4 which is T so average inventory is so much. So, average inventory into C c which is the
cost of holding, this average. So, this is the inventory holding cost plus we have the back
order cost (Refer Slide Time: 00:09). So, total quantity backordered is area of this
triangle, which is half into s into t 1 plus t 4 so plus half into s into t 1 plus t 4. Now, if
we look at this cycle this is the total back order quantity so that is held for a period t 1
plus t 4 back order is 0 for a period t 2 plus t 3. So, total average back average backorder
is, half into s into t 1 into t 4 divided by t 1 plus t 2 plus t 3 t 4 which is T so divided by
T.
So, this is the total, this is the average backorder that is held in a cycle or at any point in
time so that is multiplied by C s which is the backorder cost. We should also note that
backorder cost C s and inventory holding cost C c have the same unit of rupees per unit
per year. So, C s is also rupees per unit per year backordered, this is the average
backorder quantity at any point in time. So, this is the function that we have to optimize
but then this function has then several unknowns: so Q is an unknown, I m is an
unknown t 1, t 2, t 3, t 4 and t are all unknowns, s is also an unknown.
What are known to us are D C naught C c I m is also an unknown, C c and C s are the
known quantities. Now, from this diagram we also realize that several of these unknowns
are related to each other therefore, we try and get equations relating these unknowns and
try to reduce the expression in terms two unknowns, which are Q and s as we did in
earlier models. And then try and partially differentiate with respect to Q and s. So, first
254
we have to write a whole lot of them in terms of the other known and unknown
quantities. So, from this particular (Refer Slide Time: 00:09) figure now, first thing is we
can write few equations.
Now, we have now this quantity which is I m which is the maximum inventory. I m will
be P minus D into t 2 because (Refer Slide Time: 00:09) inventory grows at the rate of P
minus D for a period t 2 and the same I m is equal to D into t 3 because it is consumed at
the rate of this so D into t 3. Now, s which is the (Refer Slide Time: 00:24) quantity that
we have here, now from here s is this quantity so s the quantity backordered is P minus D
into t 1. So, s is equal to P minus D into t 1 which is also equal to D into t 4 because this
is the consumption (Refer Slide Time: 00:09) for demand is at the rate of D for a period t
4 that is the backorder so s is equal to D into t 4. We also have Q which is the quantity
that is produced.
Now, whenever we produce, we produce for a (Refer Slide Time: 00:09) period t 1 plus t
2 so Q is equal to P into t 1 plus t 2. (Refer Slide Time: 00:09) And in a cycle the for
example, the cycle starts this is a cycle and there will one more cycle here. So, whenever
the cycle starts actually the inventory is 0 therefore, what is produced in the cycle is
consumed in the cycle. Therefore, Q is also equal to D into T; Q is also equal D into T
and we also have from similar triangles, (Refer Slide Time: 00:09) this is I m this is s so I
m by s is equal to t 3 by t 4.
So, I m by s also (Refer Slide Time: 00:09) this is t 1 this is t2 so once again from this
similar triangles I m by s is equal to t 2 by t 1, I m by s is also equal to t 2 by t 1. Now,
let us write a few simple equations from this so from this we can write now, these are all
the relationships. So, from this we can write I m plus s is equal to now, I m plus s is
equal to P minus D into t 1 plus t 2 so P minus D into t 1 plus t 2. Now, t 1 plus t 2 is
equal to Q by P so this is equal to P minus D into Q by P because from here t 1 plus t 2 is
equal to Q by P. So, I m plus s is equal to Q into 1 minus D by P now, this is a very
important equation which, we will use to relate I m s and Q so this is one important
relationship.
Now, we also have P minus D into t 2 is equal to D into t 3 so P minus D into t 2 plus P
minus D into t 1; so P minus D into t 1 plus t 2 is equal to D into t 3 plus t 4 D into t 3
plus t 4. Now, we take this D into t 1 into t 2 on the other side so P into t 1 plus t 2 is
255
equal to D into t 1 plus t 2 plus t 3 plus t 4. So, we also know that this T is t 1 plus t 2
plus t 3 plus t 4. So this is equal to D T. From this we can also write another thing so if a
by b is equal to c by d, which is equal to a plus b by c plus d, that is 1 thing. We can also
do one more thing here which is I m plus add 1 to everything. So, I m plus s by s is equal
to t 3 plus t 4 by t 4 which is equal to t 1 plus t 2 by t 1. We can actually write another
thing let us not do this, let us do here. So, this is equal to t 2 plus t 3 by t 1 plus t 4 is
equal to I m plus s.
We can write another thing from here. So I m is equal to P minus D into t 2 which is
equal to D into t 3 so this is P t 2 from this P into t 2 is equal to D into t 2 plus t 3 so P
into t 2 by T is equal to D into t 2 plus t 3 divided by T. So, we have an expression for t 2
plus t 3 divided by T here, we also need an expression for t 1 plus t 4 divided by T,
which comes from P minus D into t 1 is equal D into t 4 from which p t 1 by T is equal to
D into t 1 plus t 4 by T so we have that. Now, from this you can write t 2 plus t 3 by T is
equal to P t 2 by D T now, D T is Q, so P t 2 by Q.
Now, we also have from here, from here we can have another relationship (Refer Slide
Time: 00:09) t 2 plus t 3 by T is P t 2 by T. From here we know that t 2 is equal to (Refer
Slide Time: 07:32) I m by P minus D so this will be P into (Refer Slide Time: 07:32) t 2
is I m by P minus D so p I m by Q into P minus D. Similarly, t 1 plus t 4 by T will be
equal to P t 1 by D T which is P t 1 by Q, which is P I m now, from here t 1 is equal to t
256
1 is equal to s by P minus D. So P I P into from here P t 1 P t 1 by Q so from this (Refer
Slide Time: 07:32) t 1 is s by P minus D so P into s by Q into P minus D.
So, now we have derived all these expressions that relate all of them. So, all we now
need is to substitute them so when we substitute you get T C is equal to D by Q C
naught, (Refer Slide Time: 07:32) that comes from here. Then we have half into I m into
t 2 plus t 3 by T into C c, so half into I m into (Refer Slide Time: 07:32) t 2 plus t 3 by t
into C c so I m into C c into t 2 plus (Refer Slide Time: 19:30) t 3 by T which is P I m by
Q into P minus D. This is written as I m into I m divided by Q into 1 minus D by P,
(Refer Slide Time: 19:30) this P is brought back into a denominator as division by P so
you get 1 minus D by P plus half into (Refer Slide Time: 07:32) now, the 3rd term half
into s by T into s by t so t 1 plus t 4 by T we write from here.
So, we have s by 2 into C s plus half s into C s, s by 2 into C s into t 1 by t 4 by T which

is P s by Q into P minus D. So this is written as, once again into s by Q into 1 minus D
by P. Now, we have written the this on simplification will give us D by Q C naught plus
C c by 2 I m square by Q into 1 minus D by P plus half into C s into s square by Q into 1
minus D by P. Now, we still have three variables here which are I m Q and s. Now, these
three are also related and we have the relationship which is I m plus s (Refer Slide Time:
07:32) is equal to Q into 1 minus D by P. Now, we go back and substitute for I m as Q
into 1 minus D by P minus s.
257
So, this will become D by Q C naught plus C c by 2 into 1 minus D by P. Now, I m
(Refer Slide Time: 07:32) square will become Q into 1 minus D by P minus s the whole
square. So, this is Q into 1 minus D by P minus s the whole square, by Q plus half into C
s into s square by Q into 1 minus D by P. Now, this is written in terms of only two
variables which are Q and s. Now, we have this expression, we can partially differentiate
this expression with respect to s first and then get a equation or value for s. And then we
partially differentiate this with respect to Q and then substitute the value for s and try to
get the final one.
Then we have a lengthy derivation which can be done so when we partially differentiate
this with respect to s and set it to 0 dou T C by dou s equal to 0 would give us, these are
the terms where we will have s. Finally, we will have an expression s is equal to Q C c
into 1 minus D by P Q C c into 1 minus D by P divided by C c plus C s. And then we
have to partially differentiate this with respect to Q and Q actually comes in all the three
terms and then we need to substitute for this s into that and solve it. We will finally get,
Q is equal to root over 2 D C naught into C c plus C s by C c C s into 1 minus D by P.
So, this will be the final expression for Q as well as for s, this will give s is equal to Q C
c into 1 minus D by P C c plus C s.
Now, once we know all the values of s and Q we can calculate s and Q because all the
other things are known and once s and Q are known we can go back and calculate t 1, t 2,
t 3, t 4 (Refer Slide Time: 00:09) and the total cost and so on. So, that is how this
formula works. So, let us take a numerical example and show, what is the effect of this?
Take a small example. So, if we consider the same example that we used for the earlier
model. The earlier model (Refer Slide Time: 00:09) does not involve back ordering, this
model involves backordering so we take the same numerical illustration and then we give
a value for C s and find out happens to the total cost as well as to the economic batch
quantities.
258
So, we go back and work out the same problem as such so we consider D is equal to
10,000 D; C naught is equal to 300, C c equal to 4 and P equal to 20,000. Now, in
addition we consider C s which is the shortage cost in this case, So, C s is equal to rupees
25; equal to rupees 25. So, once we consider this the first thing we do is we need to find
out Q. So Q is equal to root over 2 into 10,000 into 300 into 29 which is C c plus C s 25
plus 4, note that, both of them have a unit of rupees 4 per unit per year rupees 25 per unit
per year. C c plus C s divided C c into C s which is 100, 25 into 4 into 1 minus D by P
which is 1 minus half 1 minus 10,000 by 20,000 which is 1 minus half.
So, this on simplification would give us a value of 1865.48, is what we get for this and
then we also have s the moment we know Q we can get s, s is equal to (Refer Slide Time:
23:57) Q C c into 1 minus D by P divided by C c plus C s. So this is 1865.48 Q C c into
4 Q C c by C c plus C s is 29 into 1 by 2, 1 minus D by P is half so into 1 by 2 so this
gives us s equal to 128.65. Now, we can calculate I m, I m which is the maximum
inventory, from this equation in I m plus s is equal to Q into 1 minus (Refer Slide Time:
07:32) D by P D by P is half. So, q into 1 minus D by P is 932.7, 932.7 minus 128 would
give us something like, 804.09.
Now, once we know Q s and I m, we can substitute here, (Refer Slide Time: 23:57) to
get the total cost or we can substitute either here or here, they are one and the same. So,
this is easy to substitute here or here to get the total cost and the total cost for this would
259
come to on substitution the total cost would come to 3216.338. Now, in order to see the
effectiveness of the back ordering, the only thing we could immediately do is to see what
happens if did not allow backorder (Refer Slide Time: 00:09) this model allowed
backordering, this model did not allow backordering.
So, we could go back and compare with numerical result of the numerical illustration, if
we did not have back ordering or is C s is equal to infinity. So, when C s is equal to
infinity this model will automatically become this model. We have anyway solved the
same numerical illustration for the 3rd model or for the other model so when we do this
we get Q is equal to 1732.05 I am just writing the value here, now this is the economic
batch quantity when we did not allow backordering or when C s is equal to infinity.
Now, in that case the maximum inventory was 866.03, I m was 866.03, in that case and
total cost in that case was 3464.10. Now, what are the, what do we gain by comparing?
The first thing that strikes us is when we have C s or when we can allow backordering
the total cost comes down 3216 versus 3464 when C s was infinity. The same result
when we go back in time to compare the first two inventory models for items that we buy
where, we showed that if we allow backordering, it actually works out cheaper per year
than the case without back ordering. The case without back ordering can be seen or
assumed to be C s is equal to infinity so the cost comes down.
Now, what happens to the total production quantity per cycle now, when we allow
backordering the quantity that will be produce is 1865.48. So for the same annual
demand T which is the cycle becomes smaller because this quantity is higher. We can
also show even though we have not shown it explicitly we can show that in this model
actually the cost is a tradeoff between this and (Refer Slide Time: 23:57) these two put
together. In fact at optimum this cost will be equal to those two costs for example, we
can easily show that now, when Q is equal to 1865.48 the number of times we order a we
produce in a year is D by Q which is 10,000 divided by 1865.48 which is 5.36 into 300 is
1608.
So, this component is 1608 (Refer Slide Time: 23:57) the total is 3216 so obviously this
is equal to this. So, what happens here is the production quantity per cycle is higher and
slightly lower, the cycle length or duration comes down slightly. Secondly there is even
though the production quantity is slightly more 1865.48 there is a backorder which is
260
also eventually cleared and then the average inventory here, actually comes down to 804
against 866 in the earlier case. The average, the I m maximum inventory that is there in
the system comes down because of the backorder and the total cost comes down because
this I m is essentially held for a slightly short period, compared to 866.03 being held.
So, inventory cost comes down, backorder cost goes up but eventually the order cost
balances (Refer Slide Time: 23:57) these two costs and the total cost come down. So,
when we have the production consumption model if we can allow backordering, then we
allow backordering. The result is very much dependent on the value of C s that we have.
So, only if we know C s accurately or if there is enough protection in the system in terms
of additional inventories and so on, then one could think in terms (Refer Slide Time:
00:09) of applying this model. But, if we are not sure about C s and if do not have the
extra question of some additional help, then it is better to assume C s to be infinity.
And then use the earlier model and not try and optimize further on the cost, by allowing
for backordering and assuming a certain C s. The strength of this model lies, or the
usability of this model lies in the exact computation or exact estimation of C s. So, if C s
can be estimated accurately and it is not going to hurt us, if we allow a backordering then
(Refer Slide Time: 00:09) this model can be used, otherwise this model has to be used.
Now, let us comeback to the old model, where we do not allow backordering which is
this model. And let us do a few more things now; let us assume that we have this for one
261
particular item. Now when we worked out this model we also said that for a particular
item, the production starts from here, stops at time period t 1 that is an inventory buildup.
Then the production for that item stops. The consumption takes place from the inventory
that has been built up. Now, if we start looking at the same graph, from the facilities
point of view, then it will look like this. Now, from zero to t 1 I am producing this item.
Right now I do not do anything then once again from here I produce this item and so on.
Now, for this period t 2, where we consume from the built up inventory, earlier I had
mentioned there it is quite likely that this will not be kept idle and this will be producing
some other item. So, let us assume that it produces a 2nd item and then so let us assume
that the 2nd item actually starts from here and the 2nd item cycle is this, so which also
means that, the 2nd items next cycle will be like this. Now, if you consider two items the
yellow is the second item and white is the 1st item. Now, once again the 2nd item is
produced up to a certain time period and then from this period to this period at the
moment our equipment can be idle.
So, what we do is we produce the 3rd item let us say, that we have here. So the third item
comes from here and say the 3rd item is produced up to this so the 3rd item is produced
up to this. So, once again the cycle now comes here may be for this period the facility is
idle and then the next cycle begins. So, if we look at this facility and three items this is
how the utilization of the facility will look like. Now, we also said that when we set this
product; this item again there is going to be a set up cost. The cost to changeover from
this to this, when we make a second cycle of yellow which is the 2nd item there will be a
cost to changeover from white to yellow. But, we also have to note that in addition to the
changeover the cost there is also a changeover time.
So, we have to assume now, that this gap now has to be for the cycle to be effective this
gap should be enough, for the time of three changeovers. Because, somewhere we have
to changeover from white to yellow, yellow to blue and then blue to white, which simply
means that, if we are able to do something like this, now, this much is the changeover
time to yellow and then the yellow cycle happens, now this much is the changeover time
for the blue, then blue cycle happens. And then this much is the changeover time once
again for white and the white setup happens. Now, if we start looking at this problem,
from the point of view of the facility that is making if I am making certain number of
262
products then I should be able to produce the cycle time for the each product in it is
desirable that the cycle time is the same.
So that, this cycle time would be the same as this cycle time we have to draw this thing
again little clearly. So, there is a small time which is there and then the yellow and then
there is some time here and then the blue and then there is some more time here and then
the white. Now, we would like to have this cycle time same, this cycle time same and so
on. So, the T has to be the same for all the products so that we need to find that T and
then there will be individual times, which you may call this is as t 1 this as some t 2, this
as some t 3 and so on. The time you produce each of these items then I should be able to
buildup inventory and then I should be able to consume and so on.
Now, can we find out such a T is the next problem, if we are given multiple items. Now,
this problem is called the economic lot scheduling problem, And then we will try and
formulate this economic lot scheduling problem and try to solve it. Now, if there are n
items that we have to produced, there are n items that we have to produce, (Refer Slide
Time: 40:08) in this example we show three items the white, the yellow, as well as the
blue. For each item there is a set up cost, there is a inventory holding cost right. So, for
example, each item will show a figure like this but the equipment will show a figures like
this.
263
Now, if we take a particular item the only condition is the cycle time T is the same for all
the items, T is the cycle time for all of them. So, if the cycle time is T, T years then the
number of times we produce is 1 by T so many times in the year we produce. So, number
of time we produce is equal to, 1 by T the production quantity is the total demand
divided by number of times you produce. So, production quantity Q will be equal to D
divided by 1 by T which is D T. So, production quantity Q is equal to D T. Right? So, for
each item, so for the first item if I find total cost is equal to set up cost plus inventory
holding cost. So, set up cost is number of times we produce which is 1 by T and every
time we produce we incur a setup cost of C naught.
So, for a particular item let us say T C 1 or T C j for the j-th item, then the number of
times we produce will be 1 by T, so C naught by T is the total setup cost that we will
have. Now, if C naught is same for all the items then we call it as C naught if the j-th
item has C naught j, then this the total setup cost for the j-th item C naught j by 2. Now,
what is the inventory holding cost? The inventory holding cost is the average inventory
into C c, average inventory from here we are as I m by 2 I m by 2 total inventory is half
into base into height so I m by 2 into t 1 plus t 2 divided by T which is I m by 2 into C c.
So, this is I m j by 2 into C c of j carrying cost of the j-th item. (Refer Slide Time: 40:08)
We also know from this I m is equal to P minus D times t 1 for this item and it is equal to
D into t 2 for this item.
264
So, from here we also know that I m is equal to Q into 1 minus D by P because P minus
D, I m is equal to P minus D into t 1 which is here which is equal to D into t 2, that we
have here. So, I m by and from similar so from this we have, I m is equal to the rate at
which the inventory grows which is and Q is equal to D into T or D into t 1 plus t 2, so
from this we would have. So, p into P minus D into t 1, so P t 1 is equal to D into t 1 plus
t 2 Q is equal to P t 1 so dividing 1 by other would give us I m is equal to Q into 1 minus
D by this is an important equation.
So, from here we know that this is C naught j by T plus I m is equal to Q into 1 minus D
by P so Q j into 1 minus D j by P j into C c j divided by 2. And we also know that Q is
equal to D T so C naught j by T plus Q j is d j T into 1 minus D j by P j into C c divided
by 2 this is your T C j for all these items. So, C naught j by T plus C c j D j into 1 minus
by 2 P j into C c this is what we have.
Now, we want to find out T such that we minimize the sum of costs for all the items, this
is for the j-th item so we want to minimize summed over j for all the items, C naught j by
T plus D j t 1 minus D j by P j C c divided by 2. So, we want to find out a T such that we
minimize this, actually we observe that what we have written here is very similar to
normally writing as D by Q into C naught plus is I m by 2 into C c that is what we have
done here. Except that we have written in terms of T. But, the T should also satisfy
another important thing that, the time that we have here in this cycle is say between this
265
start and this start should be, should take care of the production times of all the items
plus the changeover times of all the items.
So, the cycle should satisfy an additional constraint that it should be able to take care of
all the production times as well as all the setup times. So, subject to the condition sigma
if K j is the setup time for the j-h item, note that C naught j is the setup cost for the j-th
item K j is the setup time for the j-th item plus..
What is the production time for all of them? Production time production quantity is Q j Q
j, j is equal to P into t 1, t 1 is the production time. So, P j into t 1 so t 1 is equal to Q j by
P j so production time is Q j by P j and Q j is D T j so this is D t j by P j, is less than or
equal to D j into T by P j less than or equal to T. So, this is the objective function this is
the constraint. The only catch in this constraint is we have to rewrite this constraint
because the T term is here as well as here plus of course, we have a situation that T
greater than or equal to 0. Now, this problem is called the economic lot scheduling
problem. In the next lecture, we will see how we solve this problem to get the optimum
values of T.
266
Prof. G. Srinivasan
Lecture No - 16
Economic Lot Scheduling Problem, Supply Chain Inventory
In the previous lecture, we were looking at the economic lot scheduling problem; we
were looking at this graph, where we said that we are considering say 3 items. This is for
the equipment or the facility that is producing these 3 items. So, the first item is produced
for a period say t 1 and then inventory is built up and this inventory is used to meet the
demand for the rest of the period. Then we produce the second item which is shown in
yellow and we consume while we produce and then inventory is built up and then this
inventory is used to meet the demand for the rest of the cycle. Then we produce the 3rd
item which is shown in blue up to a certain point, it is not necessary that these peaks
have to be the same these are only indicative.
So, we produce this 3rd item then stop production consume this inventory, then we come
back to the 1st item. Now, this is called the cycle; this cycle is capital T and in the
economic lot scheduling problem, we assume that the cycle is the same for all the 3
items. Then, we wrote an expression for the total cost which runs like this. So, total cost
is the sum of the costs associated with all the 3 items. In general - n items. For each item
267
there is an ordering cost and there is a setup cost and there is a carrying cost. As
mentioned earlier, setup cost is the equivalent of ordering cost; ordering cost is when we
buy the item from outside and order, while setup cost is incurred when we produce.
Now, we also realize that in this figure, I have given a small gap between the ending of
production of this item and the beginning of production of the next item and this small
gap that much amount of time is used in the changeover or the setup to make this item.
And there is a cost associated with that setup, which is our setup cost. So, if we take a
particular item, the number of times we would make that item is D by Q. So, the usual
expression is D by Q into C naught. Where, C naught for item j is the changeover cost. D
j is the annual demand Q j is the order quantity or production quantity for that item plus
we also have; we have already derived Q j - I m for item j by 2 into C c of j.
Now, we also know that this I m is equal to Q into 1 minus P by D, so this is written as
plus Q j into 1 minus P j by D j into C c j divided by 2. Now, one important assumption
in the economic lot sizing problem is that the cycle times T match. So, the cycle time for
the 2nd item would also be the same T and the cycle time for the 3rd item would also be
let us say, it goes through somewhere here would also be the same T. So, since the cycle
times are the same, the production quantity Q j for item j, Q j will be D j into T, D j is
the annual demand T is the cycle time.
So, in a cycle time of T, say T years if D j is the annual demand the production quantity
is Q j is D j into T. Another way of deriving is that the number of times we place an
order is D by Q which is equal to 1 by T and therefore, Q is equal to D j into T. So, we
go back and substitute here, to get T C is equal to C naught j into T plus Q is D j into T
the D j and D j gets canceled. C naught j divided by T which is here plus this C c is
written as i C j, Q j is D j T into 1 minus P j by D j divided by 2. So, this is the cost for a
particular item under the assumption that the cycle time is T, which is the variable that
we want to find out. Now, total cost will be for all the items. So, we sum it up and say
sigma C naught j by T because we assume that this T is going to be the same for all the
items by t plus sigma i T by 2 into C j D j into 1 minus P j by D j.
So, i T by 2 into C j D j into 1 minus P j by D j. Now, this is the function that we want to
minimize and find out T that is common for all the items. Now, in addition this is where
we actually stopped in the previous lecture; in addition, we also have a constraint.
268
Constraint is like this. If I take the beginning of this, let us call this item 1 if we take the
beginning of the 1st cycle of item 1 and the beginning of the 2nd cycle of item 1, the
time duration is T, which is the same as beginning of item 2 in the 1st cycle and item 2nd
cycle for item 2.
Now, in between this time t. What we have to do is, we have to produce this item plus
this item plus this item plus we have to account for the change over from white to
yellow, which is item 1 to item 2 and from yellow to blue which is item 2 to item 3 and
blue to white which is item 1 from item 3. So, within this time period T we need to have
the production of the 3 items made as well as, the setup time for all the 3 items to be
accommodated. So let us call K j as the setup time to make item j.
Now, this is the setup time required to make the yellow item, this is the setup time
required to make the blue item and somewhere here this is a setup time that is required to
make; this is setup time to make the yellow item, this is the setup time to make the blue
item, this is the setup time to make the white item. Note that these three t times need not
be the same and secondly right now, we assume that this setup time, is time required to
make this item and does not dependent on the other item. For example, if we move from
white to yellow or if we have another situation, where we move from blue to yellow. The
time taken to set for the yellow item which is item number 2 will be the same.
So, let us call the K j as the setup time to produce or to setup for the production of item j.
So, K j is the setup time. Now, what is the production time for this item? The time we
produce is T is the cycle time here, so we produce up to a certain period which is if P j is
the production rate of this cycle. So, P j into some t 1 is what we produce in this cycle. P
1 is the production rate for item 1. So, we call this as P j into let us say some t j is the
time in which we produce item j in the cycle. Now, this should be equal to the demand of
the entire cycle because with what we produce and consume here. We produce P j t j we
also consume the demand during this period, build up this inventory and this inventory is
used to meet till the next cycle begins.
So, what we produce in 1 cycle should be equal to what we consume in that cycle. So
what we produce is P j t j and what we consume is D j into T D j is the demand annual
demand and T is the cycle. So, from which the production time t j is equal to D j by P j
into T. So, this is the production time in a cycle. This is the setup time in cycle for item j
269
so K j plus D j by P j into T is the time required for 1 item; this is the setup time required
for the item, this is the production time required for the item. So, for 1 item this is the
time required, for all the items summation K j plus summation should be less than or
equal to T which is the cycle time that we have. So, this is the constraint.
Now, this constraint also has to be rewritten in such a manner because, there is a T time
that comes here, there is also a T term that comes here. So, this can be rewritten as sigma
K j plus T into 1 minus sigma D j by P j, should be less than or equal to 0 or in other
words this is written as sigma K j, I am sorry. This is less than or equal to this, which is
rewritten as sigma K j by 1 minus sigma D j by P j is less than or equal to T. So, this is
how the constraint; this is the objective function and this is how the constraint is written,
this is the constraint. So we now, have to solve this constrained optimization problem to
get the value of T.
Now, this objective function is non-linear because the term T appears here in the
denominator and the term T appears here in the numerator and it is subject to a linear
constraint of T greater than or equal to constant. Now, there are two ways of doing it:
one is to try and use the method of lagrangian multipliers and take this constraint into the
objective function. Setup the Lagrangian function partially differentiate and try to get the
best value of T. The other which is a simpler method is, to actually solve the
unconstrained version of this problem that is; to temporarily ignore this constraint, just
solve the unconstraint version of this problem get the value of T.
And if it, happens that the T that is obtained by solving the unconstrained version
satisfies this constraint, then it is optimal to the constrained optimization problem. This is
a very simple result, you can always relax a problem by removing a certain set of
constraints and then solve a relaxed version of the problem, if it is easier to do so. And
then, the solution to the relaxed version. If it satisfies the constraint, then it is optimum to
the constraint problem.
The issue comes only when it is not optimal, then we will look at it as an when it
happens. So, we first solve the unconstrained version of this problem which is to straight
away differentiate this with respect to T.
270
So, when we differentiate this with respect to T and set it to 0 we get something like
minus sigma C naught j by T square this T to the power minus 1, so minus 1 by t square
plus sigma i by 2 into C j D j into 1 minus P j by D j equal to 0. From which, t is equal to
root over 2 C naught root over i by 2 sigma T square equal to 0, from which. So, this
goes to the other side, so we will have let me just write it down. So, this would give us,
sigma C naught j by T square is equal to sigma i by 2 into C j D j into 1 minus D j by P j,
sorry this should be D j by P j 1 minus. So, t square goes to the other side this will come
here. So, T is equal to root over sigma C naught j by sigma i by 2 into C j D j into 1
minus D j by P j. So, we get the value of T this way.
So, we compute this value of T and then we check whether the value of T here satisfies
this particular inequality or this particular constraint. If it satisfies this constraint, then it
is optimum. If it is does not satisfies this constraint, we try and realize once again from
the very nature of the function that, it will be made as an equation at the optimum and
then we will see how we handle that through a particular numerical example.
271
So, let us look at an example here, now let us look at this example where we consider
three items. So, let us consider: C naught equal to 1000, D 1 is 3000, D 2 is 5000, D 3 is
20,000. Let us say: i equal to 20 percent; C 1 is equal to 100, C 2 is equal to 200, C 3 is
equal to 400; P 1 10,000, P 2 20,000 and P 3 50,000. So, we also observe that D by P is
less than 1 so 3000 and 10,000 5 and 20 and 20 and 50. So, we would get T is, first we
need to find out this portion for each one of them and then we sum it up for each one of
the items and then we can get T.
Now, suppose we call this as some H 1 for the 1st item, so H 1 is equal to i by 2 into C 1
D 1 into 1 minus D 1 by P 1 which is 0.2 divided by 2 which is 0.1, into C 1 into D 1 is
3000 into 100 into 1 minus D 1 by P 1. D 1 by P 1 is 0.3. So, 1 minus D 1 by P 1 is 0.7
which on computation would give us, this is 3000 into 0.1 is 300. 100 into 0.7 is 70. So,
300 into 70 will give us 21,000. Similarly, we can calculate H 2 which will be 0.2
divided by 2 which is 0.1 into C 2 d 2 which is 5000 into 200 into 1 minus D 2 by P 2
which is 0.75.
So, we can calculate that and H 2 becomes 75,000. H 3 is equal to 480,000 so we have
computed: H 1, H 2 and H 3 so H 3 is 4 80,000. Now, we have to substitute all this here
to get the value of T. Now, T is equal to root of sigma C naught j by sigma H j. Now, this
is equal to root of C naught is the setup cost per item. Now, we have assumed that 3
items and there is a total setup cost of 3 into 1000; 3000 divided by sigma H j. We have
272
individually calculated the 3 values: 21,000 plus 75,000 plus 480,000 which would give
us 576,000. This would give us a value of 0.0721 years.
Now, we compute that the cycle time is 0.0721 years. Now, we have to find out whether
this value of T satisfies the constraint. And the constraint that we have to satisfy is sigma
K j by 1 sigma K j by 1 minus sigma D j by P j is less than or equal to T. We need to
check whether this constraint is satisfied by the T. Now, in order to do that we need to
first find out this sigma K j, which is the sum of the setup times. So, let us assume that
the setup time is 10 hours per setup. So, this would give us 10 hours per setup will give
us 10 by 24 by 365 so these many years. This on calculation would give us 0.00114.
So, 0.3 into 0.00114 divided by the 3 comes because of 3 items, each item has a setup
time of 10 hours. So, 3 into 0.00114 divided by 1 minus sigma D j by P j which is 1
minus D 1 by P 1 is point 3 minus D 2 by P 2 is 0.4 and D 3 by 0.25 minus 0.4. So 1
minus 0.95 is 0.05 and this would give us 0.0685 years. Now, the constraint sigma K j by
1 minus sigma D j by P j is 0.0685 which, is less than the computed value of T of 0.0721.
And therefore, if the setup time is 10 hours per setup, we can conveniently take this T to
be 0.0721.
Now, for this T equal 0.0721 we can now, calculate Q 1 is equal to D 1 into T, Q 2 equal
to D 2 into T, Q 3 is equal to D 3 into T. So, Q 1 will be 216.3 units, this will be 360.5
and this will be 1442 giving us a total cost of 83138.43. Now, if we have a situation
273
where the setup time is 10 hours per setup (Refer Slide Time: 17:07) then we would
accept the 0.0721 as the cycle time T. But, on the other hand if K j is 50 hours instead of
10 hours, then we again have to find this out, sigma K j by 1 minus sigma D j by P j.
Now, if K j is equal to 50 then sigma K j by 1 minus sigma D j by P j will become

0.0685 into 5. The reason is the denominator remains the same the numerator instead of
10 is now, 50 therefore; it is multiplied by 5 which gives us a value of 0.3425. Now,
0.3425 is greater than 0.0721. And therefore, in this case we would now, take T equal to
0.3425 and compute the rest of the numbers, from which Q 1 will be 1027.5, Q 2 will
become 1712.5 and Q 3 will become 6,850 with total cost is equal to 206039.1. We also
observe that the total cost increases when the setup time becomes 50 hours instead of 10
hours.
274
We look, at one more aspect of inventory, which let us call as supply chain; we call it
supply chain inventory. Now, the very basic inventory model, where we had this graph
for time and quantity. We assumed that there is instantaneous replenishment which
means there is 0 lead time. And then, there is continuous demand for the item. So, we
could start with a certain Q here and then we consume till we reach 0, at that instance
place an order it gets replenished. Once again we do this and then we place an order and
the cycle goes. Now, let us assume that we or this person is the retailer. There is a
demand; there is a continuous demand. So, someone is going to pick up demand every
moment. So this will be the inventory position for the retailer; so the inventory will begin
at Q. It will be consumed at the rate of D to 0. The retailer places an order, gets items
instantly it shoots up to Q once again it is consumed.
Now, let us look at the other aspect who does the retailer place the order and how does
the retailer get it? So, the first assumption is there is instantaneous replenishment. Now,
let us assume that the retailer places an order with either a distributer or a manufacturer.
So, let us call a distributor and let us try and plot the inventory position for the
distributor. So, this is again the position this is again time. Now, let us assume that the
distributor has a certain amount of stock, which is here and for the sake of our argument
we assume that there is only 1 distributor, who has warehouse. Let us say so 1 distributor
and this distributor has a certain stock. So, the stock for that person is somewhere here it
keeps flat like this.
275
So, the stock is going to come down, only when there is a demand from the retailer. So,
this person has a stock right here, at this point there is a demand for Q from the retailer.
So, this person is going to supply the Q let us say, a person supplies Q here. Now, the
retailer gets this Q and then starts consuming somewhere here. Now, as far as this person
is concerned there is no other order so this stock will remain the same up to this point.
Once again this person is going to make an order so this person is going to replenish a Q
and so on. Note that these two are not to scale this is Q; same Q is shown here as Q so
they are not to scale but essentially this quantity Q becomes this quantity in the other
graph.
So, the point that we wish to make is while the shape of the inventory position curve is
triangle this way or Sawtooth as it is called. Now, this is more like a step function it goes
on, then somewhere this person also would reach 0 stock or let us say the person reaches
0 stock here. And then, this person orders which means a person may get it somewhere
here up to this and then his cycle may continue. Now, if we start looking at inventory
planning of this person. Let us say this person is not a manufacturer whereas, this person
is a warehouse or a distributor then this person has to place an order to the manufacturer
and to get it. So, this is how the inventory moves along the supply chain now, modeling
this person is reasonably easy because the order cost is D by Q into c naught carrying
cost is Q by 2 into C c.
Now, the Q by 2 into C c comes because of the nature of this thing which is a triangle.
So, we could get total inventory divided by and so on. Now, the moment we get into this
we cannot apply the same economic order quantity formula for this person; the reason
being the inventory curve is not the Sawtooth curve it is a different kind of a curve. So,
the inventory average inventory calculation will be very different here and therefore, we
will not be able to apply the same economic order quantity formula here, for this person.
Now, how do we handle such a situation? Now to handle such a situation an important
term called Echelon inventory; was used and introduced by Clark and Scarf in their 1960
paper where, they spoke about Echelon inventory between this and this.
So, let us explain this Echelon inventory and see what it is. Now, when we say that
inventory here shows a certain behavior. Whereas, inventory here shows a different kind
of behavior, when we compute the Echelon inventory as the total inventory in the
system, which would include the inventory of this person, as well as the inventory of this
276
person. Then it was shown that the Echelon inventory more or less shows the same kind
of behavior here. So, for the Echelon inventory we could write the total cost function and
then we could differentiate it with respect to Q to get the order quantities for the retailer
as well as for the distributor.
Now, the total cost quantity will be like this. So, the total cost for both of them put
together T C will be T C will be C o w into D by Q w plus I w dash i C w dash plus C o r
D by Q r plus I r dash into i into C r dash. Where, I w dash is the Echelon inventory
given by n Q r dash by 2 and I r dash is equal to Q r by 2. Now, I have written a fairly
complicated expression here but it is quite easy to explain each of these. Now, what we
are assuming is that we call this person as r which is the retailer and we call this person
as w which could represent a warehouse or a distributor. Since, w stands for warehouse
let us call this person as a warehouse.
Now, this term is for the retailer, this term for the warehouse. Now, let us look at only
this term, which is the retailer term; now there are 2 terms in this for the retailer: one is
the order cost for the retailer other is the carrying cost for the retailer. So, the retailer’s
order cost is C o r per order D by Q into C naught is the standard thing that we know
which represents the annual ordering cost. So, D by Q into C naught is for the retailer so
D is the annual demand which is the same for both the retailer as well as for the
warehouse. So, D by Q r Q r is the order quantity for the retailer D by Q r into C naught r
order cost for the retailer plus Q by 2 into C c which is nothing but the average inventory
of the retailer into C c.
Now, I r is the generic term so I r is the echelon inventory or average inventory at the
retailer which is Q r by 2. So, Q by 2 into C c. C c is i into C i which we have already
seen i is the interest rate and C is the unit price of the item at the retailer. So, this part is
also there for the retailer. Now, for the warehouse we need to look at D by Q into C
naught plus Q by 2 into C c for the warehouse. So, D by Q again D is the annual demand
Q w is the order quantity that this person orders so Q w into C naught w order cost for
this person so this term is fine. This term is Q by 2 into C c, C c is i into C i into cost of
the unit price of the item at the warehouse. So, i C into Q by 2. Now, this is Echelon
inventory at the warehouse this will be n into Q r by 2. We assume that the order quantity
for this Q w is an integral multiple of the order quantity of this Q r. So, C dash w is the
277
worth of the Echelon inventory; unit Echelon inventory at the warehouse so this will
become n into Q r by 2.
So, what we can do now, is we can substitute and try and get the expressions for, so now we can substitute here to get T C is
equal to C naught w D. Now, (Refer Slide Time: 26:51) Q w is equal to n Q r by n Q r plus C naught r d by Q r plus (Refer
Slide Time: So, 2 times D into C 0 w by n plus C 0 r divided by i into n C w, n C w dash plus C r dash. So, 2 D n C w dash
plus C r dash is the economic order quantity for Q r. So, if we know all these values ordering cost at the warehouse, ordering
cost at the retailer, annual demand, unit price at the warehouse for the Echelon inventory and unit price at the retailer for the
inventory i, we can calculate it. The only variable that we do not know is n. Now, there are more involved ways of computing
this n but the best and the easiest thing to do is to substitute n is equal to 1 get Q r. Because, n is equal to 1 the total
expression is somewhere here; this is the total expression substitute for n equal to 1 get Q
26:51) I w dash is n Q r by 2, n Q r by 2 i c w dash plus Q r by 2 i C r dash. So now,
differentiating with respect to Q r which is the variable here. So we get and setting it to 0
minus C 0 w D by n Q r square minus C o r d by Q r square plus n i C w dash by 2 plus i
C r dash by 2 equal to 0. From which, Q r will be equal to root over 2 times D into D is
common this 2 is also a common.
278
r for n equal to 1 substitute here. Do for n equal to 2 get Q r again substitute here and
somewhere, you will see the minimum that is happening and accept it.
So, we will take a numerical example to explain this. So, let us take: D equal to 10,000,
C naught w is at the warehouse the order cost is 200, at the retailer the order cost is 300,
rate of return is 20 percent for both of them. And we assume that the unit price at the
warehouse is 7 and the unit price at the retailer is 20. So now, this would give us the
Echelon inventory’s cost is 7 and C r is at the retailer which also absorbs this 7. So, you
get C r dash the effective value at the retailer is 13; additional unit price at the retailer is
13, which is what we will substitute here. So now, on substitution we would get; we
know this value of (Refer Slide Time: 36:53) Q so we use n equal to 1 to get.. For n
equal to 1, Q is equal to root of 20,000 into (Refer Slide Time: 36:53) C 0 w by n plus
this so 500. This is 200 divided by 1 plus 300 is 500 plus 0.2 into C w dash plus C r dash
which is 20.
So, this gives us a Q value of 20,000 into 500 divided by 20 1581.14 so this Q r is equal
to 1581.14 Q w is also equal to 1581.14. Because, if here we have defined Q w as n
times (Refer Slide Time: 26:51) Q r. So now, for this the total cost will be, we will have
to go back and (Refer Slide Time: 36:53) substitute here for the total cost. So, this is D
by Q r is the same, so d by Q r is 10,000 divided by 1581.14 into (Refer Slide Time:
36:53) C 0 w by 1 plus C o r which is 500 plus (Refer Slide Time: 36:53) Q r by 2 into i
279
Q r by 2 into i n into C w dash plus C r dash. So, Q r by 2, 1581.14 by 2 into point 2 into
C w dash plus C r dash which is 20. So, this would give us 10,000 divided by 1581.4 into
500, 3161.75 plus 1581.14 into point 2 into 20 divided by 2; which is another 3161.75,
this is 6324.56.
Now, we could look at n equal to 2 (Refer Slide Time: 36:53) to get Q r is equal to root
over 20,000 which is 2 into D (Refer Slide Time: 36:53) C 0 w by 2. So, which is 100
plus 300; 400 divided by n into C w dash (Refer Slide Time: 36:53) plus C r dash so 14
plus 13; 27 into 0.2, 0.2 into 27. This would give us 20,000 into 400 divided by 0.2
divided by 27 root over 1217.16 is what we would get and for this Q w, will be 2 times
this quantity 2434.32. See what we have done here is, we have assumed that the order
quantity here (Refer Slide Time: 26:51) is 2 times the order quantity here.
So, we have substituted for n equal to 2 in this to get Q r so Q r is 1217 and Q w is

2434.32. Now, for this value of Q r and Q w are essentially (Refer Slide Time: 36:53) for
this value of Q r we will now try and find out T C. So, T C will become C 0 w (Refer
Slide Time: 36:53) into D by n into Q r so 200 into 10,000 divided by 2 into 1217.16 n
into (Refer Slide Time: 36:53) Q r plus 10,000 into C o r by Q r 300 divided by 1217.16
plus n Q r into (Refer Slide Time: 36:53) i C w dash by 2. So, 2 by 2 into 1217.16 n Q r i
Q r into 0. 2 into C w dash which is 7 plus (Refer Slide Time: 36:53) Q r by 2 into i C r
dash 1217.16 by 2 into 0.2 into 13.
280
So, we have 4 terms here. So, let me quickly calculate all these 4 terms. So, 200 into
10,000 divided by 2 divided by 1217.16 this is 821.58 plus 2464.75 plus this will get
canceled 1704.024 plus 1582.308 so this 4 will add up to 6572.76.
So now, we have got this formula and then we have tried (Refer Slide Time: 36:53) to
find Q r for n equal to 1 and n equal to 2. When we do n equal to 1 we get Q r is 1581, Q
w is also 1581. Total cost is of the 2 systems is 6324. When we put n equal to 2, Q r is
1217 it comes down but Q w gets doubled of this 2434.32. Therefore, this cost comes
down, this cost goes up. This inventory cost will go up, this inventory cost will come
down, because of the unit cost and the total is higher 6572.76.
Therefore, in this case we would accept this solution with n equal to 1. So, this is how
we handle, if we have to look at inventories (Refer Slide Time: 26:51) between retailer
as well as the warehouse or what is called Echelon inventory. Now, we will go back to
more traditional inventory models. So, far we have seen in all the inventory models that
the demand does not change with time it is the same 10,000 per year and so on. Now, the
next question is what happens if the demand for an item changes with respect to time and
is different in different time periods and is known. So, we will address such types of
inventory problems in the next lecture.
281
Prof. G. Srinivasan
Lecture - 17
Lot Sizing
Let us continue our discussion on inventory, by looking at inventory models, where the
demand varies with time. So far, in all the inventory models, we have made two
assumptions.
One is that, the planning period is infinite. So, we have an infinite planning period and
2nd is demand is the same, or constant and uniform demand. For example, when we
looked at the basic inventory model, which is the saw tooth model and when we said we
start ordering at Q. At time equal to 0, we order a Q and it is instantaneously replenished
and then it is consumed. Till the inventory reach a 0 and then another order is placed and
then this process continuous.
Now, in this we assume that this is going to go on forever and we are planning for an
infinite period. And, because we are planning for an infinite period to evaluate the total
cost we take a certain fixed period like a year and then we try and optimize the value of
Q, for which the total cost is minimized. The other important assumption is that. The
282
demand D, which is the rate at which the inventory is falling, the demand D, is the same
at every period, more importantly it is a same at every instance.
Now, what happens to inventory problems when, the demand varies with time, which is
a very practical thing that happens around; Production quantities are different in different
periods, because demands are different for different periods and once the production
quantities are different for different periods. The amount of material required to make
them would also be different. Leading to inventory problems, where the demand varies
with time. So, let us take an example and try and explain a few things about time varying
demand.
So, we take an example where we assume that, we have demand for 12 months and these
demands are: 400, 600, 1000, 800, 1200, 900, 800, 1000, 1200, 700, 600 and 800. Now,
these numbers have been taken such that, when we add all of them we get 10,000. So,
now we assume that the annual demand of 10,000 is now split into 12 monthly demands
of this. Now, several questions are there. Now, like in the earlier models (Refer Slide
Time: 00:27) are we going to have the same order quantity, or is the order quantity going
to be different? Or, like in the earlier models are we going to have the same order of
frequency, or is the order frequency going to be different?
So, we assumed here again for the sake of comparison that order cost is rupees 300 per
order. Inventory holding i is 20 percent, of unit price C, which is also rupees 20. So, that
283
C c is rupees 4 per unit, per year, which is the same number that we had assumed in our
earlier examples and illustrations and the as I mentioned the other differences, we are
considering 12 months, or 12 periods.
Now, each period is a month and our planning stops at the end of the 12 month. So, we
are not looking at the 13th month demand in this particular situation, or illustration. One
of the ways of solving this problem is by using integer programming and the motivation
for that comes as follows. Now, if we take this month’s period and this month’s period.
Now, we could either order 400, which would meet the demand of this period. Let us say
we order 500. Now what happens is, if we order 500, then we order once here for 400.
We order a quantity of 500, which is, out of which 400 is consumed at the end of the
month and there is a 100 remaining at the beginning of this month, to meet this 600.
Now, with that 100 we will not be able to meet this 600, which means somewhere in
between, we have to place an order for another 500. So, on the phase of it, if we order a
quantity which is say, in between these two demands, it’s quite likely that we end up
making one additional order. So, if we really want to save that order, it is, it makes it
easier to assume that, if we order here we would rather order a 400, or rather order a
1000 and not a number that is in between. So, this leads to a very important result and
that result is the Wagner Whitin result, which we have actually seen earlier.
So, let me repeat those two results. So, the optimum solution to this, under certain
conditions will satisfy two important properties which is, if I place an order in the
beginning of a certain period, then the order quantity will be. Either: the demand of that
period, or the demand of two period, from there, or demand of three periods from there
and so on. So, the order quantity will be a summation of the demand, for certain k
consecutive periods. 2nd important property is, I will place an order only when my stock
at the beginning of that month is 0. So, when the stock is positive, then that stock should
be enough for me to meet the demand of the entire month, at least the demand of the
entire month.
So, we look at these two basic ideas and then formulate this problem, as an integer
programming problem. One could solve it using dynamic programming also, but at the
moment let me formulate and solve an integer programming problem for this. So, the
moment we take care of these two assumptions that. The order is placed at the beginning
284
of the month and the order quantity should be the summation of k month’s demand. k
could be 1, or 2, or 3 from that point. And secondly, we place an order only when the
stock is 0. If there is stock at the beginning of that month, then there should be enough
stock to meet the requirement of that month. So, the moment we assume these two thing;
which will happen under certain conditions.
We now formulate a problem where, Y i equal to 1, if we place an order, in month i.

When we say place an order, we mean place an order at the beginning of the month. We
still assume instantaneous replenishment. So, the with the quantity that arrives at that
time. We will be able to meet the demand of that month. Now, let X i be, the quantity
ordered. So, Y i is a binary variable: so it is a 0 1 variable. We either place an order, or
we do not place an order. X i is the quantity that is ordered. So, the objective function
will be to minimize, the sum of the ordering cost as well as the inventory holding costs.
So, we first write the objective function, we write part of the objective function. Then we
write the constraints and then we write the remaining portion of the objective function
So, objective function will be to minimise the order cost plus carrying cost. So, minimise
sigma, i equal to 1 to 12, 300 Y i. Because, every time when an order is placed, if Y i
equal to 1 then we incur 300, which is the order cost. So, this is the total order cost
component. Plus we also need to have, the inventory holding cost for the items. Now
285
here, what we are going to do is, we have calculated that C c is rupees 4 per unit per
year.
Our period is a month. So, C c has to be written as, so many rupees per unit per month.
So, this will become 4 by 12, or 1 by 3, per unit per month. So, the inventory holding
cost will be plus 1 by 3, into. Now, we are going to charge the inventory, we are going to
make another assumption that, we are going to charge the inventory, for the ending
inventory in that period. Normally, inventory can be charged in two ways, one is you
charge it on average inventory, which means beginning inventory plus ending inventory
divided by 2, or you could charge it only on ending inventory.
So, let us assume we are going to charge it only on ending inventory. So, let us call, I 1
plus, I 1, I 2, I 3 etcetera, as the ending inventories. So, this is 1 by 3 into, I 1, plus I 2
plus, I 12. Again we make an assumption that the beginning inventory is 0. We do not
have any inventory, which means we have to place an order here. It is also fair to assume
that I 12 also will be 0. So, it is enough to stop this at I 11, because we are trying to
minimise the cost. So, at the end of the 12th period, or 12th month we are not going to
leave behind an inventory. Now, the cost will be higher if we do so. Only for the sake of
completion, we write this I 12 and let the solution say, that I 12 is 0. There is no
inventory left at the end of the year.
So, this is the objective function. But right now, I have not defined the I 1 I 2 here,
because I am going to eliminate the I 1 and I 2 at some point. Now, we have to look at
the constraints. So the constraint, (Refer Slide Time: 02:38) the 1st month constraint will
be, X 1, is the quantity that I order. So, it should be X 1 greater than or equal to D 1,
because it should be greater than or equal to 400. So, that I meet the 400 demand. Now,
the other way of looking at X 1 greater than equal to D 1 is, I am just writing it here. This
is not an additional constraint. It is going to replace this. If I 1 is the ending inventory at
the end of the period 1, then I 1 is X 1 minus, D 1. Or, when we convert this inequality
into an equation, you can write X 1 minus, I 1 is equal to D 1, which would give us X 1
minus, D 1 is equal to I 1.
So, this is written as, from this we will know that I 1 is equal to X 1 minus, D 1. So, I am
just going to write here as I 1 is equal to X 1 minus, D 1. Later we will substitute for I 1
here. Now, if I take the 1st two months the quantity that I order in two months is X 1
286
plus X 2. So, what I order in X 1 plus, X 2 now should be greater than, the sum of these
two demands which is 1000. I should be able to meet them. So, I will have my 2nd
constraint, which is X 1 plus, X 2 should be greater than or equal to D 1 plus, D 2.
So, if I 2 is the ending inventory at the end of period 2, then I 2 is obviously the
difference between D 1, plus D 2. That has been ordered in the two months, I am sorry X
1, plus X 2, which has been ordered in two months and D 1, plus D 2, which has been
consumed for two months. The difference is I 2. So, I 2 will become, from here X 1 plus
X 2, minus D 1, minus D 2.
So, the moment we understand these two constraints, the rest of the constraints are easy.
The 3rd constraint will be, X 1 plus, X 2 plus, X 3, is greater than or equal to D 1 plus, D
2 plus, D 3. From which I 3 will be X 1 plus, X 2 plus, X 3 minus, D 1 minus, D 2
minus, D 3. So like this, we have 12 constraints, for the 12 time periods. This can also be
generalised and written in a single constraint, which is of the form sigma X j, j equal to 1
to I, is greater than or equal to, sigma D j, j equal to 1 to I for I equal to 1 to 12, which
means if we take the 1st month I, then this will be X 1 greater than or equal to D 1.
Take the 2nd month i equal to 2. So, j summed from 1 to 2. So, X 1 plus, X 2 is greater
than or equal to D 1 plus, D 2 and so on. So, there are 12 constraints, which are like this.
Then we also need to link the X with the Y. X is the quantity, that is ordered and we can
order an X j, or X i in the i-th month. Only if we choose to order, which means the
corresponding Y i has to be 1. If X i is greater than 0. So, we need to relate the X and Y,
which is done by the usual constraint, which is X i is less than or equal to M into Y i. So,
when Y i is 0, which means I choose not to order in this month. Then M into Y i will be
0. This will force X i to be automatically 0. If I chose to order in a month, where Y i
equal to 1. Then X i could be either 0, or non 0, which is what we want.
So, there are again 12 more constraints like this, which takes care of this part of it. Then
of course, we have already defined Y i to be binary 0, 1 and we may define X i to be an
integer, or we can define X i to be a continuous variable. So, let us assume we simply
define X i also to be an integer, because by the Wagner and Whitin idea, we will have all
X i’s to be integers because these demands are all integers. So now, the constraints are all
written in terms of X and Y. But, the objective function has this I term. So, this I 1 has to
be replaced in terms of X. So, I 1 was X 1 minus, D 1. I 2 is X 1 plus, X 2 minus,
287
D 1 minus, D 2, I 3 is X 1 plus, X 2 plus, X 3 minus, D 1 minus, D 2 minus, D 3 and so
on.
So, when we substitute this into the, I 1 plus, I 2 plus, I 3. What we will get is, 12 times
X 1, plus 11 times X 2, plus 10 times X 3, etcetera, up to X 12. Because, for I 1 you are
going to have X 1 minus, D 1, I 2 also has going to have a X 1 plus, X 2. So, it is going
to have an X 1 term. I 3 also will have an X 1 term. So, all the 12 will have X 1 12 times.
So, this will become 12, X 1. Now, X 2 is going to appear from I 2, I 3 up to I 12. So, X
2 will appear 11 times, X 3 will appear 10 times and so on. Minus, once again we will
have 12 D 1 minus, or plus 11 D 2, plus 10 D 3, etcetera, plus D 12. The reason is this: D
1 is also appearing in all the 12 I’s. So, you will have minus 12 D 1. D 2 is appearing in
11 terms, from I 2 to I 12, so this will also appear.
So now, we have eliminated the I’s, the ending inventories. The I variables we have
eliminated. And, it is also a customary in any linear programming, or integer
programming, that if you have a constant term in the objective function. You remove it
and then you solve it and later, you can add that constant into it. So, this is the constant
term. So, we remove this term and then we solve this. So, we can now solve this integer,
linear integer programming problem, using a solver or otherwise, to get the optimum
values of X and Y.
288
So, there are 12 terms here, 300 into Y 1 plus, Y 2 up to Y 12. Plus another terms here
that represent the inventory holding cost. There are 12 constraints, representing the
demand. So, that the demand in every period is met. There is no shortage. Therefore, no
shortage cost is added in the objective function. This model does not allow shortage, or
backordering, Plus of course, the linking constraints between X and Y. So, when we
solve this optimally, we will get a solution which is this.
So, the solution to this will be Y 1, equal to Y 3, equal to Y 5, equal to Y 6, equal to Y 8,

equal to Y 9, equal to Y 11, equal to 1. So, the solution says that we will order in months:
1, 3, 5, 6, 8, 9 and 11. So, 7 out of the 12 months, we will order. What are the order
quantities? Of course, the solutions would also throw the excess as a variable which will
be there. But, the moment we know the Y’s we also know the X’s, because I am placing
an order here.
289
I am placing an order here at Y 3, placing an order at Y 5, placing an order at Y 6,

placing an order at Y 8, placing an order at Y 9 and I am placing an order at Y 11.
So, obviously my order quantities are X 1 equal to 1000. Because, I am not ordering in
this month, I am not ordering here. So, the order quantity here is (Refer Slide Time:
20:12) the sum of these two. X 3 is 1800, X 5 is 1200, X 6 is 1700. So, 1700 comes
because I have to order for this 900 plus another 800, so 1700. X 8 is 1000, X 9 is 1900
and X 11 is 1400.
290
Now, what is the total cost associated with this system? Total cost would be given by the
objective function value here. This (Refer Slide Time: 17:07) would be the objective
function value. Plus of course, that constant has to be added. So, what we will do now is,
we will compute the total cost. So, there are 7 orders here. So, 7 into 300, the inventory
holding cost is plus 1 by 3 into. Now (Refer Slide Time: 20:12) here we order 1000, at
the end of this month we carry 600. So, 600 is carried. Now here it is 0.
So, we order 1800, so we carry another 800 here. Now, here (Refer Slide Time: 20:12)
we order 1200. We consume all the 1200 here. Here we order 1700 we consume 900. So,
we carry another 800 here, so plus another 800. We order 1000 we consume all 1000.
We order 1900 we consume 1200. So, balance of 700 is carried, which is consumed at
the end of this period. So, we order 1400, 600 is consumed here. 800 is what is carried,
so 800. It is also interesting and obvious to note that there are 7 periods where there is an
order. There are 5 periods where there is inventory holding, making up for the 12.
So, total cost will be 2100 plus, 6 plus 8, 14, plus 8, 22, 29 plus 8, 37, 3700 by 3 which
is, so 2100 plus 1233.33, which is 3333.33. So, while writing this formulation I had
mentioned that the objective function has two terms. One is the order cost term and the
other is the inventory carrying cost term. (Refer Slide Time: 17:07) So, while writing the
formulation I mentioned that the, inventory cost can be calculated either based on the
ending inventory, or based on the average inventory.
Now, this expression was written, when we wrote it for the ending inventory in mind. So,
we defined I 1, I 2 up to I 12 as the ending inventory in the 12 months and then we
simplified to 12 X 1 plus, 11 X 2 plus etcetera, minus a constant. And then, we removed
the constant, because in any mathematical programming problem. When you optimise a
function, if there is a constant in the objective function, you can remove it and then you
can optimize only the terms containing the variables.
Now, what we will do is we will try and see what happens if, we write this inventory
(Refer Slide Time: 17:07) term, the carrying cost term. Considering the average
inventory and not considering the ending inventory. Now, does the solution change, if we
do that? So, let us go back and try to first compute this term (Refer Slide Time: 17:07) or
write the equivalent of this term, if we consider average inventory and not the
291
ending inventory. So, let us write that portion here. So, to begin with we will assume
that, I 1, I 2 up to I 12, are the 12 values of the ending inventory.
So, 12 values of the ending inventory are I 1, I 2 up to I 12. Now, if we take the 1st
period, the beginning inventory is 0. So, the order quantity is X 1. So, beginning
inventory now will be 0 plus X 1, which is X 1 in the 1st period. Now, at the end of the
period, the ending inventory at the end of period 1 is I 1. Now, we order X 2, in the
beginning of period two. So, beginning inventory will be I 1 plus X 2.
Now, for the 3rd period, the ending inventory will be I 3. But, the beginning inventory
will be I 2 plus X 3, because I 2 is the inventory that is available at the end of period 2.
And at the beginning of period 3 we are adding an X 3. So, I 2 plus X 3, is the beginning
inventory for period 3. I 3 is the ending inventory for period 3. So, this way we can write
all the beginning inventories and ending inventory term and for the 12th month
beginning inventory will be I 11, plus X 11.
So, the average inventory for month 1 will be, X 1 plus I 1 divided by 2. The average
inventory here will be, I 1 plus X 2, plus I 2 divided by 2 and one more term, it will be I
2, plus X 3, plus I 3 divided by 2 and for the 12th period. It will be I 11, plus X 11, plus I
12. The 12th period: the beginning inventory will be I 11, plus X 12, because X 12 is the
quantity that we order. So, this will become I 11, plus X 12. So, this will become I 11,
plus X 12, plus I 12 divided by 2.
292
So, if we sum all the average inventories, we will get the I 1 plus 2 appears here, I 1 plus
2 also appears here. Similarly, I 2 divided by 2 appears here, I 2 divided by 2 appears
here. So, when we take only the I 1, I 2 up to I 12 terms, we will have I 1, because I 1
divided by 2 plus another I 1 divided by 2. So, I 1 plus I 2 plus up to I 11, the I 11 will
have a term from the previous, apart from the previous term and apart from here, I 12
does not have a succeeding term because that is the last period.
So, up to I 11, plus we will have I 12 divided by 2. And then, we are going to have, X 1
by 2, plus X 2, by 2 plus X 3 by 2 up to X 12 by 2. So, this will be plus, X 1 plus, X 2
plus etcetera, plus X 12 divided by 2. Now, what we need to do is to have, to write this I
12 divided by 2, in terms of first write I 12 and then substitute for I 12 to see what
happens. Now, what is I 12? I 12 is the inventory at the end of the 12th period, beginning
inventory is 0 here.
So, I 12 is in principle sigma X 1, plus X 2, or sum of X 1 plus, X 2 up to X 12, which is

what we have procured. Minus D 1 plus D 2 to D 12 which is what we have consumed.
So, I 12 will now be equal to, I 12 is X 1 plus X 2, plus X 12 minus, D 1 plus D 2 up to
D 12. So, let me write this as sigma X and let me write this as sigma D. So, I 12 divided
by 2, is equal to sigma X divided by 2, minus sigma D divided by 2. So, this will become
I 1. So, the sum of the average inventory will be, I 1 plus I 2 up to I 11, plus this I 12
divided by 2 is sigma X by 2, minus sigma D by 2. This is another sigma X by 2. So, this
will become, plus sigma X minus sigma D. Because, there is a sigma X by 2 term that
comes out of this. This is another sigma X by 2 term. So, they add up to give you sigma
X minus sigma D.
Now, we need to go back. We can actually look at it in two ways. But, the more
mathematical way is to try and define what are these I 1, I 2 up to I 11 and then see what
they are. Now, what is I 1 plus I 2? I 1 is X 1 minus D 1, I 2 is X 1 plus X 2 minus D 1
minus D 2 and so on, then plus sigma X minus sigma D. Now this, now you get I 1 plus I
2 plus I 11, which is what I am writing. I 1 is X 1 minus D 1, I 2 is X 1 plus X 2 minus D
1 minus D 2.
Now, the total inventory is this term, this term is what I am writing here. Now, this term
is I 1 plus I 2 up to I 11, plus I 12 by 2. Now, this portion I 12 by 2, is written as this. So,
this portion is up to. So, this portion is, this is here plus I 12, let me write from here. So,
293
that there is no confusion. So, up to I 11 I am writing from here. Now, this portion is I 12
by 2, plus sigma X by 2. Now, I 12 by 2 is sigma X by 2, minus sigma D by 2 plus sigma
X by 2. So, there is a by 2 that comes here.
So, this will become X 1 minus D 1, plus X 1 minus X 2 minus D 1 minus D 2. We get
up to 11 terms. So, the 12th term will not be there. So, the 11th term will be X 1 plus X 2
to X 11, minus D 1 minus D 2 up to minus D 11. So, this on simplification will give this,
X 1 will come 11 times. So, this will give me 11 X 1, plus 10 X 2, plus 9 X 3 etcetera, up
to X 11. Now, this one will become, minus 11 D 1, minus 10 D 2, minus 9 D 3 and so on,
minus D 11. Once again this term will repeat.
So, plus sigma X, minus sigma D divided by 2. Now, as far as we are concerned, minus
11 D 1 minus 10 D 2 minus 9 D 3 up to minus D 11 is a constant. This is also a constant.
So, this is 11 X 1, plus 10 X 2, plus 9 X 3 up to X 11. Sigma X is, X 1 plus, X 2 plus, X
3 up to X 12. So, when we add it we get 12 X 1, plus 11 X 2 etcetera, plus X 12, minus a
big constant K.
So, we have to write, if we write average inventory term and bring it back to this (Refer
Slide Time: 17:07) expression we should write, 1 by 3 into 12 X 1, plus 11 X 2, plus 10
X 3, plus X 12 minus some very big K constant. And, based on the same idea we can
remove this constant and we can solve it. So, as far as the optimization is concerned,
even when we look at average inventory, which is taken as beginning inventory plus
ending inventory divided by 2. We get the same optimization problem (Refer Slide Time:
17:07) except that the constant that we are subtracting is slightly different. There is
another way of looking at same thing. The reason perhaps is that, if you see this. Initially
we started with ending inventory as I 1 plus, I 2 plus, I 3 up to I 12 is what we had.
Now, ordinarily we will have, I 12 will be 0. Because, beginning inventory is 0, pushing

an I 12 will lead to higher cost. So, I 12 will be 0. So, it is enough if we do up to I 11 and
by the same reason we will realise, that sigma X will be equal to sigma D, because when
the initial inventory is 0. We will not buy and then finally, keep it as final inventory. So,
this term will essentially become a constant. So, average inventory would also lead us to
the same value for ending inventory. So, we have already seen the optimum solution to
294
it. The problem does not change. So, this (Refer Slide Time: 19:32) is the solution for it.
The only change is in our calculating the total cost.
Now, when we consider ending inventory, the total cost was computed for order cost
plus ending; carrying cost of the ending inventory. Now, what we have to do is, we have
to compute this cost again. Considering the average inventory and there will be 12 terms,
instead of 5 terms. The 5 terms came here, because we ordered for 7 periods, which
meant that 5 periods is demand was carried. So, when we computed based on ending
inventory, we got 5 terms here. Now, when we compute it based on average inventory
we will have 12 terms. So, let us compute those 12 terms again and see what happens to
the total inventory cost. So, there are 7 orders. Order cost is taken as 300.
So, 7 into 300 plus 1 by 3, 1 by 3 is our C c, (Refer Slide Time: 02:38) 4 per unit per year
becomes, 4 divided by 12, per unit per month, so 1 by 3 per unit per month. So now,
what do we do? We start the 1st month by ordering 1000. So, 1st month beginning
inventory is 1000. So, let me write down the points what we order, using a different
colour perhaps yellow.
295
So, we will have 1000 ordered here. Y 3 is 1800. This is 2100; I am sorry 1200, 1700,
1000, 1900, 1400. So, 1st month beginning inventory is 1000, ending inventory is 600.
So, 1000 plus 600 divided by 2, is 800. 2nd month beginning (Refer Slide Time: 38:42)
inventory is 600, ending inventory is 0. So, average inventory is 300. 3rd month
beginning inventory is 1800, ending inventory is 800. So, 2600 divided by 2 is 1300. 4th
month beginning inventory is 800, ending inventory is 0. So, you get 400. 5th month
1200, ending inventory is 0, so we get 600. 6th month, 1700 plus 800 divided by 2; so
296
1250. 7th month will be 400. 8th month will be 500. 9th month is 1900, plus 700 divided
by 2, 1300. 10th month is 350. 11th month is 1100 and 12th month is 400.
Now, 400 comes, 1100 comes because we order 1400. At the end of this month ending
inventory is 800. 1400 plus 800 is 2200 divided by 2. So, this will become, 2100 plus,
now let us add only the 50 terms first. 1250 plus 350 is 1600. 2400, 2700, 4000, 4400,
5000, 5400, 5900, 6200, 7200, 8300 plus 400 is, 8700 divided by 3.
So, this is 2900 plus 2100, which is 5000 will be the optimum, the value of the objective
function at the optimum, if we consider the average inventory. So, important
assumptions as to, for a quick recap the important assumptions are, that the Wagner and
Whitin assumption that, we have placed an order, when the inventory is 0. And, when we
place an order, we place an order for that period or some of demands of certain k number
of periods. When there is inventory at the beginning of the period, we will not place an
order and that inventory should be, enough to meet at least this month’s demand and in
principle to meet the demand of an integral number of months, starting from now.
So, under those two assumptions, we formulated this problem. The problem could also
be solved by dynamic programming. But we solved by integer programming by defining
Y i as minus. Now, let us make a very quick comparison with what we know. Now, let us
go back and (Refer Slide Time: 38:42) take this 10,000, as the annual demand. The
important deviation of this problem from the economic order quantity problem is that.
The demands are now given over 12 periods.
Whereas, there is annual demand of 10,000 was given as an annual demand and the
demand is the same at every instance of time. Now, demand varies here. The original e o
q, we assumed in infinite horizon. Now, there is a finite horizon of 12 time periods.
Nevertheless, let us make a comparison. Now, just in case we have taken this 10,000 as
the annual demand and then use the economic order quantity formula, to find out the
economic order quantity and the associated total cost. What happens? So, when we do
that.
297
So, we could write D equal to 10,000, C naught equal to 300, C c equal to 4 per unit per
year. Now, Q is equal to, root over 2 D C naught by C c, which we have already
calculated in an earlier lecture, 1224.74 and then we also know that the total cost for this
is, root over 2 D C naught C c, which will give us 4898.98.
Now, we have to first understand. This is an optimal solution to an integer programming

problem, under a certain assumptions, which gives us 5000. The best value of economic
order quantity, assuming that this is an annual demand of 10,000, is 4898.98 and it is
slightly less than that 5000. It is very clear as to why this was marginally higher. That is
because; this problem is further constraint, by many reasons.
Now there is an, this is (Refer Slide Time: 39:16) has a finite period, this has an infinite
period. That has demand that varies with time. Here you have demand that does not vary
with time. So, in principle that is a constraint problem. So, it gives us 5000, but then why
should we not use this, if this is usable? Now, let us go and see why we need this and
why we would not be able to use this. Now, to do that let us go back and find out. If the
order quantity is 1224.74, then the number of orders per year n will be 10,000 divided by
1224.74 and we quickly compute that to be, 8.2.
So, let us call it 8.2 orders, actually 8.17. Let us call it as 8.2 orders. Now, there are
going to be 8.2 orders over 12 months. So we, let us call the ordering frequency per
month. How many orders is 12 divided by 8.2, which is roughly about 1.5. Let us
298
approximate it to about 1.5 orders per month, 8.2 orders over 12 months then, we have
an order every 1.5 months, per order.
Now for example, we are going to order, let us say 1224.74 lets kind of approximate it to
1225, which is not a very wrong approximation. So, we 1st order 1225. Now, let us try
and see what we do (Refer Slide Time: 38:42) with this. So, we can meet the 1st month’s
demand of 400 through it. So, at the end of the 1st month, we will have 825. But, then
somewhere in between, we after 0.5 months we will have an another 1225 coming in.
So, the (Refer Slide Time: 38:42) 2nd month’s 600, we can actually meet. But,
somewhere in between we will get the 1225 that comes in. So, at the end of it we will
have, 225 plus 1225. So, we will have 1225, plus 225, which is 1500. So, at the end of
the 2nd month, we will have 1000 every one and half months we will place an order. So,
we will have 1500. This is, beginning of the 1st month. So, end of 1st month, end of the
2nd month, we will have 1500.
Now, meet this 1000 here. (Refer Slide Time: 38:42) So, we would have 500 at the end
of this month. Now, half way, the next order is going to come half of, now the 1st order
has actually come here. The 2nd order will come at this point. So, 500 plus 1250, will be
1750 will be the order. So, we will have 1750 here, at the end of 3. That will be the
inventory at the end of 3 and so on.
So, actually we would realise that, we will be able to meet the entire demand by doing
this. Actually, we will be able to do that in this case, if we continue this kind of
calculation. We will be actually able to show that, we can meet the 12 months demand by
suitably ordering this. There are no issues on that. The only place, why? Even though this
will be a feasible solution, if we show, this will be a feasible solution, with cost equal to
4898.98. But, then the optimal solution here (Refer Slide Time: 39:16) will show, will
ask us not to order, somewhere in the middle here (Refer Slide Time: 38:42) and so on.
The optimum solution will ask us to order, at these points, with the slightly higher 5000.
So, we need to explain, why that difference happens. Essentially the difference happens
because when we made this calculation. This calculation was based on a total cost, T c is
equal to D by Q, C naught plus Q by 2 C c. Now, the average inventory there was taken
as Q by 2, that comes from here, which was based on the assumption, that the demand is
299
D in every instance. Now, the moment we look at problem with time varying demand
this way, (Refer Slide Time: 38:42) the demand is not D at every instance. The demand
is different. Therefore, the correctness of Q by 2, as representing the average inventory
comes into question. So, Q by 2, is not a correct estimator of this.
Therefore, if we use this Q by 2 instead and then optimize, the optimization portion is
right differentiating and setting it to 0 and getting it as right. But, Q by 2 into C c, is not
the correct way to represent the average inventory and because the average inventory
was represented in an incorrect way here, it ended up giving us a 4898.98. Whereas, the
best way to do it, is the way we have done it in this formulation (Refer Slide Time:
39:16) to get this answer, to get a value of 5000. So, using the economic ordering
quantity per say and implementing it, even though it is implementable, as in this example
it is implementable, it is fine.
But, then the correctness of the average inventory will come into the question and when
we actually correct this term, for the solution, which is 1225, every one and a half
months and if we correct the term while doing this calculation. We correct this term
when we do this calculation, then you realise for example, that the average inventory
would not be Q by 2. It will be something slightly more, which would I eventually give
us a number; there is slightly more than 5000. So, we need to use this method (Refer
Slide Time: 39:16) when we actually work with problems on, time varying demand and
not approximate the economic order quantity. Now, another area where the economic
order quantity can suffer is this. Now, suppose instead of the, this is the total demand let
it be 10,000, but then if we say that the 1st periods demand is not 400, but for the sake of
argument.
300
If the 1st period’s demand is itself is 2000. Now, the rest of the demands are such that,
the total annual demand is 10000, which means, the demand for the remaining 11
periods on summation, is actually 8000. So, that the total demand becomes 10000. Then
instead of using this model, which is a Wagner Whitin algorithm, if we had gone by the
economic order quantity and said we would use it, then the economic order quantity tells
you to order 1224.74, the same thing based on the same 10000.
So, we initially begin with 1224.74 and then our next order is going to happen after one
and a half months, because of this 8.2 and 1.5 months. So, if we assume that in the first
day of January, you have ordered 1225 and their next order you are going to make
around 15th of the February, which is one and a half months from now. So, you have to
meet the one and the half months demand, with the 1225, that you have.
Now, you would not be able to do that, because the 1st month’s demand itself is 2000.
So, we will be forced to have shortage and then we have not included shortage in doing
this. So, for another reason you cannot directly use the economic order quantity formula
and then say that I will suitably modify it to take care of time varying demand. In an
example like this, where it is not you know, you do not have a real peak. One month
taking a demand of 2000, while the other 11 months having a demand of 8000. Clearly
shows there is a peak here. Whereas, if we look at these numbers, you see that they are
301
reasonably spread out, around the average value of 1000, or slightly less, the average
value of 833. It is kind of spread out, even though you have a 1200 somewhere here.
But, if one of the demand shows a peak, then implementing a modification of the
economic order quantity, would lead us to shortage. Therefore, this approach (Refer
Slide Time: 43:19) is not recommended, while this approach (Refer Slide Time: 38:11)
is recommended. Now we do, having seen this, let us look at just a couple of more
instances of this.
One is, this problem is formulated and solved as a integer programming problem, where
the Y i’s were binary variables and X i, were defined as integers. I also mention the little
earlier that, it is not necessary really to keep them as integers, because the demands
themselves are integers. But, the Y i‘s have to be binary.
Now, this would require an integer programming solver and if we are doing it for large
time periods, integer programming may take a little longer than, we normally anticipate
and we may, require more sophisticated solvers. So, the question is? Can we try and
solve this problem heuristically and not really worry about the optimal solution, but try
and get a good near optimal solution. 2nd issue that comes is, now this integer
programming formulation can be made. If we have a finite number of periods: if we have
a situation, where we do not want finite time period.
302
If we look at practice, about demand of a particular item and let us say the forecasting
happens continuously. And then there is a rolling kind of a forecast, where initially the
forecast is made for 1 year and then after a quarter, after 3 months the forecast is made
for another 1 year. Therefore, the forecast is rolling and in some sense there is no end to
this number. So if this, there is no end to this number. We cannot make that formulation,
because that formulation is for a finite time period. So, the question is how do we extend
this to look at situations, where we do not have finite time periods?
Now, can we solve that optimally? The answer is no, because we need to know how
many periods. All these are summations. So, you need to know how many periods you
require. So, in order to handle situations, which are more practical, where the time period
is not fixed and defined. We need to look at heuristics, which give good solutions
comparable to this optimum solution. So, we look at a couple of such heuristics, in the
next lecture.
303
Prof. G. Srinivasan
Lecture - 18
Lot Sizing – Heuristics
In the last lecture, we looked at how to order when the demand varies with time. So, this
was a example that we took where we took the annual demand to be 10,000 but the annul
demand is not distributed equally to the 12 time periods. But, it changes with respect to
each of these time periods. We saw a mathematical programming formulation for that
problem and we also saw, why and how we may not be able to use the well known
economic order quantity result in solving this type of a problem. We also ended by
saying that, if we have a finite time horizon as it is in this case, integer programming can
give us an optimal solution and dynamic programming also can give us an optimal
solution.
But then integer programming is hard and difficult and we require a solver. More
importantly, in practice the time periods are not finite and time periods can extend. And
we also have rolling forecast for the products which in turn gives us some kind of an
estimate of the demand for each of the item that rolls over a time period. So, we have to
solve a problem where this is not fixed, then this goes on and on and then we concluded
304
in the last lecture, that we could use Heuristics to solve this problem. Some other data
that we used while doing the integer programming is that.
The order cost is rupees 300 per order and the carrying cost C naught is 300 per order.
And the carrying cost C c was rupees 4 per unit per year which becomes 4 by 12 or
rupees 1 by 3 per unit per month. So, if you want to look at Heuristic solutions, if you
these Heuristic solutions do not guarantee optimality, they try and guarantee a solution
which is close to the optimum. And since, the basic problem is a minimization problem a
Heuristic solution would have an objective function value higher or equal to that of the
optimal solution. So, the easiest thing to do is to if there are 12 periods.
Now, only for the sake of illustration of the Heuristic, we will take the same example so
that they will be able to compare the final solution given by the Heuristic as well as the
optimum. But, then we will also show how this Heuristic can be extended to meet the
case where the planning period is not finite and keeps rolling. So, the easiest Heuristic is
if we are looking at 12 months, the simplest thing to do is make an order every month.
That is called lot for lot Heuristic.
So, simplest thing is to make an order every month. If the demand time bucket is 1
month for which the demand is known, place an order exactly for that. So, if we do that
we end up placing an order in each of these 12 months. So, total cost will be 12 times
order cost because there is an order every month, 12 into 300 plus we are also computing
305
the average inventory that we have in each month so plus 1 by 3 which is the inventory
holding cost. The average inventory for January will be 400 divided by 2 because you
begin with 400 and end with 0. For February it is 600 divided by 2 so, average inventory
for each month is demand by 2.
So, total sum of the average inventories over the 12 months is the total demand being
10,000. The average inventory, sum of the average inventory will be 10,000 divided by 2
which is 5000 so, 1 by 3 into 5000. So, this would give as a total cost of 3600 plus
1666.66. This would give us 5266.66. We also saw in the last lecture, that the optimal
solution to this did not have an order in each of these months, it had 7 orders in the 12
periods and the optimal value or the best value was 5000 and this Heuristic is giving us a
value a 5266. Now, this Heuristic is very nice because it is very easy to implement.
Every month you look at the demand and you place an order to meet the demand of that
month but it could be slightly costly because you end up making more orders than the
optimum. The optimum had 7, now you have 12.
Now, this Heuristic can also handle time period, if it rose beyond that because it is a very
simple Heuristic where every month you just place an order. The second Heuristic that
we will see is called part period balancing. If we try and see what we actually optimized,
in our mathematical programming model, what we optimized or try to minimize is a sum
of the order cost and the carrying cost. So, the objective function value is total cost is
equal to order cost plus carrying cost. Now, we want to make decisions regarding the
purchase such that, we wish to optimize or minimize the sum of the order cost as well as
the carrying cost.
We would also, if we see the optimum solution carefully we would also have realized
that in periods where the demands were very high, you make order in that month and do
not combine it with the next month and so on. Now, the basic idea is that, if carrying an
item for an extra month for example, (Refer Slide Time: 00:18) if the tradeoff is between
placing an order here for 400 and placing an order here for 1000. So, when we place an
order here for 400 it automatically means we have to place another order here for the 600
so there will be 2 order costs. And individually we will be carrying 200 here and 300
here.
For example, if it is cheaper to place 1 order here and then find out the sum of the
306
inventories that we will carry and if it is cheaper, then obviously we would like to order
1000 here, rather than ordering a 400 and a 600. Or in other words, if the cost of carrying
the inventory of this in here also which is the additional thing that we do when we
combine, if that cost of holding that inventory is actually cheaper than one order cost
then we would end up ordering a 1000 here. So, if the cost of holding the 600 here
exceeds the order cost then it is better to order separately one here and one here. So, it is
a tradeoff between the inventory carrying cost of a particular month versus one order
cost. That is the essential or central idea in the part period balancing Heuristic.
So, first thing what we do is (Refer Slide Time: 00:18) we look at this period and say
order quantity Q is equal to 400. So, when order quantity Q is equal to 400 which means
we are ordering only the first months demand. Then average inventory that we have is
equal to 200 and the inventory holding cost is 200 divided by 3. Now, we try and see
whether we can order 1000 here. So now, if Q equal to 1000 here which means we are
trying to order for these two demands in this place. Average inventory will be equal to,
1st month (Refer Slide Time: 00:18) we begin with 1000 and we end with 600 so the
average inventory is 800 for the 1st month.
Second month we would begin with 600 and end with 0. So, there will be another 300
here which will be and the carrying cost is 1100 divided by 3. So here, 200 divided by 3
is 66.66. 1100 divided by 3 is 3 threes are 9 20 366.66, 20, 3 366.66. Now, right here we
observe that the carrying cost of bringing these two together is now exceeding an order
cost of 300. So, the moment the order cost is exceeded we stop that and then we have to
decide whether we chose this 66 or whether we chose this 366.66.
So, common sense tells us that we choose 366.66 because it is closer to 300 than the 66.
So, when the carrying cost exceeds the order cost, stop the computation. And then, look
at the last two values of the carrying cost, whichever is closer to the order cost you take
it. So, now we would place, (Refer Slide Time: 00:18) we pick this and we say that in the
1st order we would order for 1000 which is bringing these two things together. So, let me
explain the 2nd computation also and then you move to 2.
307
It means, right now we have finished ordering up to this. So, we will 1st look at Q equal to
1000. So, Q equal to 1000, average inventory is 1000 divided by 2 which is 500. So, the
carrying cost is 500 divided by 3 which is 166.66. (Refer Slide Time: 11:20) Now, we
consider 1000 plus another 800. So, we consider 1800 here. Now, the average inventory will
be, 1st period here we begin with 1800, we end with 800. So, 1800 plus 800 divided by 2
which is 1300 which comes here. This is 1300 for the 1st period and for the 2nd period we
start with 800 and we end with 0. So, plus another 400 so this is 1700 so, 1700 divided by 3
3 5s are 15, 566.66. Now, this 566.66 has exceeded the order cost of 300.
308
Therefore, we stop here and then we try and see which one is closer to 300. So, 166.66 is
closer to 300, therefore we place an order here, which means your 2nd order is only for
1000. And now, your problem starts from here where you will 1st start looking at 800
and then you will look at Q equal to 2000 and so on. So, like that we can proceed till we
finish orders for all the 12 periods and if we do so we would get a solution which is like
this. So, we would get, when we complete this.
So, when we complete the part period balancing Heuristic our order quantities will be:
1000, another 1000, 800, 1200, 900, 800, 1000, 1200, 1300 and 800. (Refer Slide Time:
03:14) So, the part period balancing Heuristic, two of the things that we saw. We saw
order quantity of 1000 and another order quantity of 1000 which I have shown here as
1000 and 1000. So, this would say or would ask us to order: 1, 2, 3, 4, 5, 6, 7, 10 times it
would ask as to order in the 12 months.
309
So this, these two are coupled together for a 1000. 1000 remains 800, 1200, 900, 800,
1000, 1200, there is another bunching here 1300 and another 800.
So, what is the cost associated with part period balancing? So P P B, cost is equal to,
there are ten orders so, 10 into 300 plus 1 by 3 into (Refer Slide Time: 14:47) So, here
1st period is 1000, 1000 plus 600 divided by 2. So, this is 800 for the 1st period, 300 for
the 2nd, 3rd will be 500, 4th will be 400, 5th will be 600, 6th will be 450, 400, 500, 600.
(Refer Slide Time: 14:47) Now, here there is a bunching. So, we would start with 1300
310
finish with 600. So, 1900 divided by 2 which is 950. This will be 300, we begin with 600
end with 0 so average is 300 and here the average is 400 so, plus 300 plus 400.
So, this is 3000 plus 1 by 3 into 8, 11, 16, 20, 26, 30, 35, 41, 44, 48, 4800 plus 450 plus
950 is 1400. So, this is 3000 plus 6200 divided by 3, this is 3000 plus 2s are 6, 0, 6 are
18, 66.66, so this is 5066.66. So, you can see the difference, now the optimum was 5000.
Now, part period balancing is giving a solution with 5066.66. Lot for lot was much more
expensive. In some sense, lot for lot did not try to optimize anything. It just said do it as
it is or take it as it comes. In some sense there is no optimization here, there it is only an
evaluative algorithm where, if I make an order every period what is going to be my cost?
Now, here there is a certain amount of optimization because we found at least by

bunching (Refer Slide Time: 14:47) these two we are able to save one order cost, which
means the cost of carrying this 600 along with this 400, that is far more important for
another period seem to be lesser than one additional ordering cost. So, there was a
bunching here. Similarly, there was another bunching here. So P P B is able to give a
better solution in this instance when compared to lot for lot. And we could generalize and
say that part period balancing would anyway give a better solution than lot for lot
because it is trying to optimize something somewhere.
Now, the third one that we will see is the most popular Heuristic. It is called Silver-Meal
Heuristic, after the paper from Silver and Meal who are the researchers who made that
Heuristic. It is somewhat similar to part period balancing but slightly different. We
explain the difference using a certain computation. Now, let us start with like we did in
part period balancing. So, we begin with Q equal to 400. (Refer Slide Time: 14:47) Let
us assume that we are, let me remove this and let me also remove this.
So, we start first with this, so we start with Q equal to 400. So, when we place an order
we have to incur an order cost of 300. There is a carrying cost of, we begin with 400 we
end with 0 so, average inventory is 200. So, 200 divided by 3 which is 66.66. So, total
cost is 266.66, let me keep it there. (Refer Slide Time: 14:47) Then we consider
combining these two together so, Q equal to 1000. So, when we combine these two
together we still try and have 366.66, we will make that change. The order cost is 300,
carrying cost is 66.66 so, total cost is 366.66. Now, when Q equal to 1000 we still try and
order 1000 in a single order. So, order cost is 300, carrying cost we have already
311
computed. We have just kept it here, carrying cost is 366.66.
So, this is 366.66, total cost is 666.66. Now, it is not fair to compare this 366.66 with
666.66 because this 366.66 is the cost for the 400, while 666.66 is the cost for 1000 but
essentially it is a cost to meet the demand for 2 periods. So, what we do is we try and
compare what is called a per period average cost. So, we know write another one called
per period, will be equal to this is 366.66. This is 666.66 divided by 2 which will give us
333.33. So, now we realize that the per period cost is coming down. So, we try and see if
we can optimize further on the per period cost.
So, we look at the 3rd possibility which is to try an order, (Refer Slide Time: 14:47) 400
plus 600 plus 1000 so try an order 2000. So, order cost is again 300, carrying cost we
first have to find out the average inventory. So, to find the average inventory, 1st month
we start with 2000 and we end with 1600. So, 2000 plus 1600 divided by 2 is 1800. So,
this is 1 by 3 into 1800 for the 1st month. (Refer Slide Time: 14:47) 2nd month we begin
with 1600 and end with 1000 so, 2600 divided by 2 is 1300. 3rd month we would begin
with 1000 and end with 0 so, average inventory will be 500.
So, this alone will be 1800 plus 1300 is 3100. 3100 plus 500 is 3600, 3600 divided by 3
is 1200. So, total cost will be 1200 plus another 300 which is 1500 and the per period
will be 1500 divided by 3 periods which would give us 500. Now, we realize that the per
period cost was 366.66 when we considered 1 period, 333.33 when we consider 2
periods and 500 for 3 periods. So, it is started, it went down and it came up. Therefore,
this is the place where it is giving us a local benefit the best out of the 3. Some kind of a
local optimum is happening here so, the decision is to try an order for 1 and 2 together,
combined these 2 together.
312
Now, we have to repeat what we have did. The reason I said, it similar to part period
balancing but different is that, both of them essentially try to tradeoff between the order
cost and the carrying cost. Part period balancing explicitly tries to see if the additional
carrying cost exceeds the order cost. Silver-Meal tries to do the combine cost for
multiple periods, average asset and see that period for which the average is showing a
local minimum or the average is optimize. Now, we have to continue like what we did
here (Refer Slide Time: 24:41) starting from this. So, we will start with Q equal to 1000,
Q equal to 1800 and so on till we find yet another local optimum and then we proceed.
313
Now, if we do the Silver-Meal Heuristic, continue to do it for all the 12 periods, the
solution that we will have is, so let me write down the solutions for the Silver-Meal. So,
the solution for the Silver-Meal will be 1000, 1800, 2100, 800, 1000, 1900 and 1400. Let
us go back and make this comparison once again.
So, if we do Silver-Meal, let me use a different color, so when we use Silver-Meal we

order these two together. Then comes 1800, we end up ordering these two together
whereas, in part period balancing they came as two different orders. Then we do 2100,
we end up ordering these two together. Then, we have an 800 which is here. Then we
have 1000 which is here, then we have 1900 which comes here and then we have 1400
which actually comes here. I think in part period balancing the combination happened
like this.
So, we realize that the solution is slightly different in Silver-Meal compare to part period
balancing in fact, we could even show. Take it exactly here to show how a Silver-Meal is
optimizing, combining these two together whereas, part period balancing is trying to look
at these two separately. (Refer Slide Time: 03:14) We have the part period balancing
calculation which said that, go for 1000 here. So, let us just do the Silver-Meal
calculation to show that it is actually moving to one more. So, let us just start from here
and then look at this.
314
So, we know look at the case two when Q equal to 1000. Order cost is 300, carrying cost
is 1 by 3 into 500 which is 166.66. So, T C is equal to 466.66 and T C per period is equal
to 466.66. (Refer Slide Time: 26:26) Now, in Silver-Meal we try and combine this 1000
and 800 together so, Q equal to 1800. Order cost is 300, carrying cost is 1 by 3 into
(Refer Slide Time: 26:26) 1800 plus 800 divided by 2 which is 1300 plus 400 so, 1700.
So, 1 by 3 into 1700 which is 5100 divided by 3 so, 5100 divided by 3 is 700 by 3 1700
by 3 is 566.66 so, T C is 866.66. Now, T C per period is 433.33.
So, Silver-Meal straight away tells you that based on the Silver-Meal way of
optimization, combining 1000 and 800 is cheaper than doing 1000 separately. Whereas,
part period balancing based on its own way of optimization is a now says somewhere
here that, if you do 1000 it is here (Refer Slide Time: 03:14) 1800 is here 1000 is
cheaper. (Refer Slide Time: 26:26) Now, when we do 1000 plus 800 plus 1200 we would
realize that the part period cost is higher so, Silver-Meal would have combined 1000 and
800 whereas, part period balancing separated them from 1000 and 800.
So, let us try and do the total cost for Silver-Meal. So, total cost for Silver-Meal is equal
to, Silver-Meal is giving us 7 orders as against (Refer Slide Time: 13:48) part period
balancing had 1 2 3 4 5 6 7 8 9 10 orders. So, 7 into 300 so, Silver-Meal order cost is 7
into 300 plus 1 by 3 into so, for the 1st month I am combining this 1000 so 1st month I
begin with 1000 and I end with 600. So, the first month will be 800, 1st month is 800
315
2nd month is 300. Now, 3rd and 4th I am combining. (Refer Slide Time: 26:26) So, 1800
plus 800 divided by 2 is 1300 plus another 400 so, 1300 plus 400.
Similarly, 5 and 6 I am combining so, I begin here with 2100 and I end with 900, 3000
divided by 2 is 1500, 1500 plus 450 is 1950 so, 1500 plus 450. (Refer Slide Time: 26:26)
Keeping this separately, this is 400, this is 500 so, plus 400 and 500. (Refer Slide Time:
26:26) Combining these 2 so 1900 I begin with plus 700 is 2600 so it is 1300 is the
average, 1300 another 350. So, 1300 350, the last one again I am combining. (Refer Slide
Time: 26:26) So, start with 1400 and end with 800 so 2200 divided by 2 is 1100 plus
another 400. So, 1100 plus 400 so, this is 2100 plus 1 by 3 into 8 plus 3 11, 24, 28, 43,
47, 52, 65, 76 plus 4 80, 8000 plus 450 plus 350 which is another 800.
So, this is 2100 plus 1 by 3 into 8800 which is 2100 plus 2933.33 which would give us
5033.33. For this particular example, there is a slight gain when we do Silver-Meal
compare to when we do part period balancing. In some sense, both Silver-Meal and part
period balancing give solutions which are very close to the optimum, the optimum is
5000. If I have the problem is also such that the optimum 5000 was not very far away
from the EOQ is 4898.98, though we discussed the issues involved in using or
approximating the EOQ in the earlier lecture.
So, the optimum is 5000, part period balancing is giving a solution that has a cost of
5066.66, Silver-Meal is giving cost of 5033.33. Both are very good Heuristics, Silver-
Meal is used is much more popular and used little, perhaps a little more than part period
balancing. But, invariably both of them are used for a given problem and then the best
one is chosen depending on what solution they give to the particular instance of the
problem. Both of them are intuitive in their own way, the difference is very subtle but
both of them are centered around optimizing the total cost. The only difference is Silver-
Meal computes the total cost and then divides it by the periods to optimize it.
Part period balancing essentially looks at the moment the inventory cost exceeds the
order cost, which one is closer. So, mostly in that decision to find out which one is
nearer, part period balancing solution is different from Silver-Meal solution. We did see
a situation (Refer Slide Time: 03:14) here where 166 566 because order cost was 300 this
was chosen, while Silver-Meal is way of optimizing allowed us to combine both of these.
So, both the methods can be used and both of them are used extensively. Two other
316
issues before we close this and move to another topic.
The first issue is, how do we measure the goodness of a Heuristic? There is a Heuristic
there are more than one Heuristic in fact, there are many more Heuristics there are other
Heuristics also which we are not covering in this lecture series, there are few more
Heuristics to do that. Then how do we measure the goodness of the Heuristic is number 1
and number 2 is are there instances where the Heuristic can behave slightly poor, are
there problem instances where a particular Heuristic may not give a very good solution.
So, people address both these issues we are not going to dwell into much more detail on
these. I will make a very brief mention of both of these.
May be, in some sense it is also in fact the Heuristic is complete when we are able to say
what is called or able to define what is called the worst case performance of the
Heuristic. Now, if L naught or the C naught in some sense let us call it L naught as the
optimum or o as the optimum (Refer Slide Time: 15:05) o in this case is 5000, h is the
Heuristic solution. In still in part period balancing it is 5066.66. Now, based on this
numerical example one would say that h divided by o, which is Heuristic divided by the
optimum would be of the order of 1.000 something because you are dividing 5066
divided by 5000 whereas, the same as 5033 divided by 5000 would be little smaller for
the Silver-Meal.
So, h by o the smaller h by o is the better the Heuristics. (Refer Slide Time: 26:26) one is
to try and look at numerical simulations like this. Take examples and show find out
average h by o and so on. The other is also to derive some kind of theoretical
expressions. After all, the basic idea is known so depending on how large or small this
value is with respect to the previous one. Part period balancing is going to decide
whether this is going to be separate or whether this is going to be clubbed because it is
going to check the closeness to 3000.
So, it is possible to do what is called a worst case analysis and show that for example,
whatever be the problem, whatever be the number of time periods part period balancing
will be within 1 plus x of the optimum or Silver-Meal will be 1 plus x of the optimum,
actually between the two researchers have found that part period balancing is little better
amenable to that kind of an analysis. So, there are some results which talks about the
goodness of the part period balancing. Now, there is another question that can also come
317
is, how do these Heuristics behave? (Refer Slide Time: 26:26)
If for example, there are lots of peaks and valleys in this data, we did briefly mention it
in the earlier lecture the impact of the peak. We did give an example where if the 1st
month’s demand itself is 2000 out of the 10,000, then rounding off economic order
quantity is not going to help us. The other issue is if there are time periods where the
demand is 0 for certain items, how does a Silver-Meal behave? How would part period
balancing behave? Would something behave better than the other consistently if there are
0 demands in certain periods?
So, some of these questions have also been addressed in the later step and there are some
problems instances where, Silver-Meal does not behave extremely well when there are
too many 0s coming in the middle and particularly if those 0s are constitutive and so on.
But, leaving out all those exceptional situations both are very effective Heuristics to
solve this problem and both these Heuristics are not bound by the number of periods. By
the very definition, the very nature, the way the Heuristics run we are not bound by a
finite number of periods. (Refer Slide Time: 26:26) For example, if we are right here let
us assume we are addressing a practical problem and we are right here after 10 months.
But, by that time we would have forecast for the next 6 or 7 months whereas, when we
are solving this particular instance where we have to close this. And let us say for some
reason the Silver-Meal optimizes up to here because we are looking at a finite time
period this has to go into one order. But then, if you have a rolling time period by the
time you come here, you would have another four 5 pieces of data for the next few
months. And then, you can continue to apply part period balancing or Silver-Meal and
you will not be bound by having one order exclusively for this. You may be able to
combine this and the next one and so.
So, in fact in the situation where the demands are rolling and you do not have a finite
time period so, like these work much better than integer programming from a
practicability point of view because integer programming is going to bind you to a finite
number of time periods. So, these Heuristics are used extensively to solve. So, let us try
and introduce one more problem in this lecture and then let us see how we are able to
handle this.
318
Now, let us revisit the Economic Lot scheduling problem. So, let me write the Economic
Lot Scheduling problem again. Now, when we define the Economic Lot scheduling
problem, we did something like this. We said, we looked at a particular machine or a
facility that is making multiple products. If you take a particular product or an item it
produces at the rate P, the demand is at the rate D. So, what you normally do is you
produce for a certain period, consume while you produce so, inventory is built up at the
rate of P minus D. And then with the built up inventory you consume it for a certain
number of period and the cycle begins again.
So, let us say that if we are going to do this we consume here and let us say the cycle
begins again. We also said that in between this period when we do not produce this item
say we may produce another item which we are going to show by the yellow color which
is item number 2. And we also said that there has to be a certain changeover time
between the end of this and the beginning of this. So, let us say we have a small
changeover time here and then the yellow begins. Let us assume the yellow is produced
up to this period.
And then, inventory is built up. Now, let us say yellow is also consumed up to a certain
time period and then in the remaining period we could again set it up for another product
which is say the blue colure product. And then, the blue color is produced say up to this
and then, the blue color production is over. Let us say the blue is produced up to this. So,
319
the blue color production is, let me say blue is produced up to this. So, the blue color
production is up to here then we changeover and go back to the white color product
which is the original product and then the cycle begins.
The other thing in the Economic Lot scheduling problem we said that, between this and
this the consumption the cycle time is T and then we try to find out what is the value of T
that optimizes the changeover cost plus the inventory holding cost for all the periods
taken. Now, in some sense when we did worked on the Economic Lot scheduling
problem we are not unduly varied about the order in which we are going to make this
white, yellow and blue products. Though, the example that I am giving right now follows
the order: white then yellow and then blue.
Let us assume these three colors represent three distinct products. Once we solve the
Economic Lot scheduling problem whether we make it in the order white, yellow, blue
or white, blue, yellow it is the same because the cycle time is the same for all. The
reason why it is not order dependent is because the time to changeover or setup as we
call which is this time to changeover to yellow, this much time to changeover to the blue,
this much time to changeover to the yellow and this much time to changeover to the
white, depends only on the succeeding job and does not depend on the procedural job.
This is not defined as changeover between white and yellow. This is not defined as
changeover between yellow and blue. It is defined as changeover to yellow, changeover
to blue irrespective of what the earlier product is. If the changeover times were sequence
dependent which means it depends both on the current job as well as the next job to be
done then the problem gets far more complicated because sequencing them to in the
order that minimizes the sum of changeover times. When the changeover times are
sequence dependent becomes a traveling salesman problem which by itself is a very
difficult problem to solve.
So, in the Economic Lot scheduling problem we do not assume at least at the moment
that the changeover times are sequence dependent. So, changeover times are sequence
independent and then the Economic Lot scheduling problem essentially tries to minimize
total cost which is a sum of order cost plus carrying cost. Now, let us look at a certain
variation of this problem where we are going to make some more simplistic assumptions
and then we are going to solve another type of a problem here. Let me introduce this that
320
problem in this lecture.
Now, let us assume while we started, when we started doing this we said, now I am
building up inventory for this at the rate of P minus D and then when I consume this
which is shown by this white dotted line, I have built up inventory here. Similarly, when
I start this yellow I am building up inventory here and then I stop producing the yellow
and I consume the yellow for a while. Now, what am I doing to meet the demand of the
yellow and blue for this period and for this period. In the same graph if you see that I
start producing when the inventory of white is 0, I produce and consume so why it is
completely taken care of.
If my entire planning for yellow is going to start only here or here I set it up and then I
produce. Then this part is ok because to meet this demand of the yellow product during
this period I have built up this inventory. Now, what have I done to meet the demand
during this period? So, if we start the problem here at some time equal to 0, we definitely
need some inventory of yellow and blue on hand so that by the time we stop producing
this or by the time we start producing this yellow, that inventory should be enough to
meet the demand of that period.
Similarly, that inventory of blue should be enough to meet the demand of that period. So,
the problems will begin, will be redefined as I have a certain initial inventory of all the
items may be: the white, the yellow and the blue. I have inventory of all the items so
when I have some inventory of all the items now let us assume that we are going to make
certain number of items. So, let us say that I have n items or n products which I have to
make. I have a certain inventory of each of them so I j is the inventory that I have of each
of them on hand. D j is the demand for item j that I have, P j is the rate at which I can
produce the item.
Now, I want to define what is the cycle length T which is a same decision variable as in
the Economic Lot scheduling problem. What is the cycle length T such that I minimize a
total cost, this is cycle length such that I minimize a total cost right, such that I minimize
a total cost. Now, ordinarily this total cost is going to be order cost plus carrying cost.
Now, this problem in some sense will become the Economic Lot scheduling problem if
the D j’s are the same for every period for each item right. If we take annual demand for
item a, it is a same its one number which means you are not using the case of time
321
varying demand. Demand is the same at all periods.
And, if we consider the order cost as well as the carrying cost. The only difference
between this and the Economic Lot scheduling problem is the presence of I j which
essentially means that we are ensuring that there is no shortage, right up to here. One
way of looking at the economic lot scheduling problem is we are looking at it at some
kind of real time, which can start from any point and there is sufficient inventory to meet
this the point that I raised right now to meet the demand of yellow for this period and the
demand of blue for this period.
The modification here is if I have a certain amount of inventory my initial inventory is

also going to help me in some way to stretch the T, the cycle. It will help me in defining
the cycle. And then, in some sense I would like to have longer cycles so, I would like to
have longer cycles. When we have longer cycles obviously the order cost is going to
come down and the carrying cost is going to go up. So, from a purely inventory point of
view or just in time manufacturing point of view, we would like to have small lot
production which simply means that we do not want longer T but then, let us try and
look at this problem at least to begin with that how long can we make our cycle. Then we
start defining how we play around to make that cycle as small as it can be.
So, let us first define the problem to try and have longer cycles so that the shortages are
not there as well as the order cost or setup cost component because its manufacturing. It
is not order cost, it will be setup cost. So, this will be setup cost though we have used the
same C naught as the noted. So, the problem will now boil down to trying to maximize
the cycle length, not to try and maximize the total cost will also show how the moment
we assume that we want to maximize the cycle length this objective automatically moves
to maximizing the cycle length. For a given amount of inventory, what is the maximum
cycle length that I can achieve? And then, we modify the problem by saying if I play
around with this inventory, can I maximize my cycle time further? And third is if the
demand is time varying, how do we bring that, modify that into this problem. So, these
problems are called disaggregation problems and these problems we will try and address
them in the next lecture.
322
Prof. G. Srinivasan
Lecture - 19
Disaggregation
In this lecture, we address the Disaggregation problem; the disaggregation problem was
introduced in the last lecture. Disaggregation problem is also a certain variation of the
economic lot scheduling problem; now one way of looking at the disaggregation problem
is like this, when we solve the aggregate planning problem, we used a variable called P t
which is the production time that is available in every period. Now, the next step is to try
and see how we use that production time P t to produce various products. Now, in
aggregate planning, when we studied aggregate planning, we assumed that there is an
aggregate product, which adequately represents all the products that are made in the
organization, and variable such as P t that is production time that is available in the
period, which is a decision variable of an aggregate plan.
W t work force that is used in period t, which is also a decision variable, I t inventory at
the end of period t, they were all defined in terms of a single unit called man-hours or
person-hours. So that, all of them can be measured consistently using a single unit. So,
when we say P t, we represent P t in terms of man-hours or person-hours available not in
323
terms of number of units produced. Because, at the disaggregation level we are going to
break this P t into several products and therefore, it is convenient to define P t in man-
hours rather than in number of units.
So, let us define the disaggregation problem using an example, so let us assume that
there are four products, which we call A, B, C, D and demand for these four products are
400, 600, 800 and 700 for A, B, C and D respectively. Inventory available is 200, 400,
500 and 300, and we also assume that P available is 2500 man-hours. In the first simple
assumptions that we have made is that, the demand is also given in man-hours and the
total demand is 1000, 1800, 2500, which is equal to P t that is a simple assumption that
we have made.
So, we are not talking of either underutilizing this P t or demand being in excess of P t
and not being able to meet the demand. So, we are looking at a situation where, the total
demand is equal to the total capacity and demand is also given in man-hours. Similarly,
the beginning inventory or inventory available is also given in man-hours, so all of them
are in consistent unit called man-hours.
Now, let us go back and draw a very familiar figure that we had, which we have already
seen, let us assume this is time and let us assume we are talking of producing some of
these products. So, earlier we had used different colors to represent different products, so
if we are looking at the production consumption model, which means we consume where
324
we produce. And we also in such a situation know that P has to be greater than D, which
is true in all the cases, the D’s are like this and P is like this, so P is greater than D.
So, what we are doing is, we are saying that we produce a first item here and then we
consume the first item and then we setup the second item and then we produce a second
item let us say up to this and then we consume the second item. And then we again setup
the third item and say we produce the third item and consume the third item, and then we
go back and let us say there is a fourth item that we have. So, we setup the fourth item
and then start consuming the fourth item and then we come back to the first item here
and do it, which means we consume up to this point, so we are looking at something like
this.
Now, we looked at this figure and we kept the cycle time T same in the economic lot
scheduling problem, and in the economic lot scheduling problem, the objective was to try
and minimize the sum of the setup cost as well as the inventory cost. Now, here what we are
trying to do is this, now if I am starting at time 0, let us say I have inventory of 0 of this item
I can produce and consume. So, I produce and consume, I buildup inventory up to this point
and then I am going to consume out of the inventory till I begin the next set.
Whether, if we take the second, third and fourth items, there production is going to start
only ((Refer Time: 05:47)) here, here and here, which means we should have enough
inventory to consume when we are not producing. So, till the time we start producing
items two, three and four, we need enough inventory to meet the demand of that item.
And that is the reason we have these initial inventories which are here, which we will be
using before we start producing that particular item.
And once we start producing that item, we consume while production, buildup some
more inventory and then use it once the production stops. Now, in this example I have
used one, two, three, four and this example assumes that the order in which, we are going
to take up these items for production is one, two, three, four. We have also mentioned
that the changeovers are not sequence dependent, so any order is acceptable to us, as
long as the changeover time, is only a setup time to produce.
In the economic lot scheduling problem, the setup time figured explicitly in the
constraint, the setup figured in the objective function. Now, here we now have different
levels of initial inventory, I think we also know that, while the first item can be produced
325
at time equal to 0, the second, third and fourth items, which means the items that we are
going to produce as second, third and fourth. For example, we could start with A and
then D, and then C, and then B, so whatever we produce as the second item, the
inventory should be sufficient to meet the demand till we start producing that item.
So, we have to choose A, B, C, D or rank them or sequence them in an order such that,
before they we start the production of any one of A, B, C, D, we should have enough
inventory to meet the demand from time period 0, up to the time period at which we are
going to start the production of any of them. Now, using this inventory I, we define a
ratio called r and this ratio tells us, the amount of time buffer that we have. The inventory
is now made into a time buffer, the man-hours it is already in time units, but then r is
represented as, so many months of inventory that we actually have.
So, when we divide this 200 divided by 400 which is 0.5, so this is 200 is equivalent of
0.5 months of inventory of this item, this is equivalent of 0.66 months of inventory of B.
So, let me write A, B, C, D, so ((Refer Time: 08:42)) this is A, this is B, this is C and this
is D, so this is 5 by 8, which is 0.625 months and this is 3 by 7, which is 0.428 months.
So, from this r, r tells us the number of months of inventory that we of these or the
inventory that is available here is capable of meeting, so many months demand.
So, now, this r also gives us an indication of the order in which we can produce A, B, C,
D, because the item whose inventory we are going to exhaust first is D. So, D has the
smallest value of the ratio r, then it is going to be A and then it is going to be C and then
it is going to be B. So, a good order in which they can be produced is D, A, C and B,
because that is the order and which, we will exhaust the inventory if we do not produce
them. I have to put it differently, we have to start the production of D, before 0.28
months, before 0.428 months, we have to start production of A before 0.5 months and so
on.
So, now let us start defining the variables, let t D, t A, t C and t B be the start time of
production of D A, C B respectively. So, it is like this is a time graph, let us say I am
starting ((Refer Time: 10:53)) t D here, I am starting t A here, I am starting t B here, I am
starting t C here, t B here and then once again I start t D and so on. Now, this is my cycle
of production which is capital T, so I start from t D and one more up to this is my cycle.
So, when I start producing D here at time t D and I start producing A at t A, then it
326
means I produce item D for a period t A minus t D. So, t A minus t D, first thing I have
to ensure is these productions t D, t A, t C and t B should start, before I exhaust these
inventories.
So, the first constraint will be t A is less than 0.5, t B is less than 0.66, t C is less than
0.625 and t D is less than 0.428, then we are producing this item that is t D first, for a
time period t A minus t D. So, t A minus t D is the time period for which we produce the
item D, and that is produced at a rate of P 2500. So, that production should be able to
meet the demand of the item D for the full cycle t D, so this should be greater than or
equal to T into 700, should be greater than or equal to 700 T.
Now, this item A is produced for a period t C minus t A, so t C minus t A into 2500 is
greater than or equal to, this is producing item A, so this is 400. t B minus t C into 2500
is greater than or equal to 800 T. And if this is t B, if we start producing T, we start
producing D again here, so this is not t D, but this is T plus t D, where the cycle has
come here. Earlier when I wrote t D, I was trying to say that we are going to start
producing D again, but the movement we want to write the time at which we start
producing D again, it becomes t D plus another T.
So, we are producing this item D up to t D plus T, so this is t D plus T minus t B into 2
1500 is greater than or equal to 600 T. Of course, we have all T j, T greater than or equal
327
to 0. So, these are the constraints that we actually have and then we have to write a
suitable objective function.
Now, the usual objective function is to minimize some of the setup cost plus inventory
cost, this is the very standard way of doing setup cost plus inventory cost. Now, one of
the assumptions that we have here is that, we are going to assume for the sake of
simplicity that, when we say that we are producing item D up to this period and then we
are starting item A at t A and up to t C. So, we are going to make an implicit small
assumption that the changeover times are negligible.
So, we do not have large changeover times, otherwise the changeover time will come
into the constraint plus another K, if that is the changeover time which is that. Now, the
movement we look at this constraint, what we are trying to say through this constraint is
that, these four ensure that the starting times of A, B, C, D are well within the r, which
means we should be able to start the production of A, B, C, D, before they are exhaust.
But, when we write these constraints, we are able to write these constraints largely
because the total 2500 is equal to exactly this ((Refer Time: 16:17)). So, the cycle can be
so defined that, when we actually complete the cycle, we have consumed for the cycle
based out of what we have produced, which kind of means that, these initial inventories
are going to decide the t A, t B, t C, t D, but they do not play a major role in this.
328
So, the general assumption is this, initial inventories will be used up, at the same time
will not be very different at any point in time. The aggregate or sum of these inventories
will not be very different at any point in time, because at steady state the way these
constraints are written, we are able to meet the demand of the cycle based on what we
produced in the shorter period of the cycle. So, the aggregate inventories will remain as
they are, so at any point in time the average inventory will not deviate or will be more or
less the same constraint.
So, if we make that assumption that the average inventory is more or less the same
constant, then the inventory cost will become a constant. Again assuming that the
holding cost is the same for each item C c is the same for each item, so the question is to
minimize the setup cost. Now, the total setup cost per year or per given time period is the
product of the individual setup cost into the number of setups per year.
329
So, individual setup cost if we call that as C naught, and number of setups per year each
cycle is going to have four setups, so number of setups per year is four times, the number
of cycles in a year. So, we want to minimize C naught into four times into number of
cycles in a year, which means you are effectively minimizing the number of cycles in a
year.
And if you want to minimize the number of cycles, it is like saying I want to actually
maximize the cycle length capital T, so maximum the cycle length T, minimum number
330
of cycles in a given period, so this problem becomes this. So, this problem becomes a
linear programming problem, because the objective function is a linear function of the
decision variable. And all constraints are linear and then you have an explicit non
negativity restriction on the variable.
If we had not made this assumption that the inventory holding cost is a constant, then we
would get a non-linear term in the objective function, where this T will appear in the
numerator here, and the T will appear in the denominator in the inventory. Then you will
have a problem with a non-linear objective function and a set of linear constraints, so the
solution methodology becomes a little more complicated. So, let us first take a look at
this as a linear programming problem, and then let us say that if we solve this linear
programming problem, we get an optimal solution which will be like this.
The optimum solution to our example is t D is equal to 0, t A is equal to 0.2453, t C is

equal to 0.3855 and t B is equal to 0.666 and capital T is equal to 0.8763, all these are in
months. So, this cycle will extend for about 0.8763 months and then the cycle will repeat
and so on.
331
Now, let us also try and look at the inventory positions during this time period, so what
we do is we first keep item D we are producing it, in the order item D, A, C and B. So, at
time t equal to 0, the initial inventories are 300 for D, 200 for A, 800 for C and 600 for
B, 500 for C and 400 for B. Now, what we are doing is we start production of D at time
equal to 0, and we produce D for 0.2453 months, so let us look at what are the inventory
positions at t equal to 0.2453 months.
Now, this period we are producing D, at the same time we are consuming D while we
produce, so we are producing at the rate of 2500 man-hours, we are consuming at the rate
of 700 man-hours. So, the net inventory buildup is at the rate of one 1800 is the rate at
which it is built up, and it is built up for a period 0.2453. So, the inventory at that
position is 1800 into 0.2453 plus 300, which will be 741.54, which approximates to 742.
Now, during this 0.2453 months, we are not producing A, we are only consuming A out
of the 200 that we have.
So, we would have consumed 0.2453 into 400, which is roughly about 100 and so the
inventory on hand will be about 100, so this is 102 on simplification. Similarly, we
consume only C, so the inventory will be 500 minus 0.2453 into 800, which will be
roughly of the order of 300 and the value is 304, and the last one can be calculated as
253. Now, we look at what happens at 0.3855, now here we produce and consume, here
we consume out of 742, here we consume out of 304, here we consume out of 253.
332
So, quickly doing these calculations, we will have 643 here ((Refer Time: 23:50)), 396
here, 192 here and 169 here, now at time equal to 0.666, the inventories are 447 here,
284 here, 668 here and 0 here. And at t equal to 0.8763, we will get 300, 200, 500 and
400 again, let me just show only one of these calculations. Now, at 0.666, this is the
inventory position of D, now between 0.666 and 0.8763, we are going to produce B, so B
is inventory is going to go up, so B starts at 0.
So, we produce at 2500, we consume while production at the rate of 600, so 1900 is a
rate at which inventory is building up, and that inventory is building up for a period this
minus this, which is 0.2103, periods it is building up. So, 0.2103 into 1900 is 399.57
which is rounded off to 400, now if we take this D just for the illustration, we only
consume from this inventory. So, we consume D at the rate of demand is 700, so 700 into
0.2103 is the consumption is 147 and this is 447, so balance inventory is 300.
So, we realize that at the end of the first cycle or at the beginning of the second cycle, the
inventory position comes back to the same 300, 200, 500 and 400, and that is largely
because of the assumption that some of the demands is 2500, while production is 2500.
So, if we take a cycle of 0.8763, we have produced 2500, we have consumed 2500, so
inventories will stay as they are, now we also observe that the total inventory at this point
is 1400, ((Refer Time: 26:29)) this is also 1400, if we add this, it should be actually
1400.
But, it has become 1401, because of rounding off, so we could take this also as 1400, So,
this is also 1400 and this is understandable, because at every point as I mentioned we
have exactly the total consumption and total production will match. Therefore, the
aggregate inventory will remain the same.
So, this would also validate our assumption that we left out the inventory cost as a
constant and than we did this of course, another way of looking it is we left out the
inventory cost, and then we optimized on that T and we observe that the sum of the
inventories is a constant. So, the basic idea of this problem is to have a cycle of 0.8763
and proceed. Now, there are two other interesting aspects to this problem, now this
formulation essentially us to maximize T, which means we want the cycles to be as
lengthy or as long as possible.
333
So, the question is are we conflicting on our assumption, that if we want the cycle length
to be long, then does it conflict with the basic ideas of shorter production runs, and just
in time manufacturing. Just in time manufacturing requires that, we have very short
production run lengths, and we are able to changeover very quickly, in a way we are not
violating it and we will see why we are not violating it. After we look at one more
modification of this, now let us go back to this we will address the other problem of
conflict between just in time manufacturing a little later.
So, we go back and look at this problem, now let us assume for a moment that our
objective is only to maximize T and then if we want to maximize T further what do we
want to do? Now, one first thing we observe is that, if we start changing these inventory
values, the r values will change, therefore these things will change and therefore, it will
either shorten or lengthen the cycle. But, let us not change the inventory values, let us
keep the inventory values as they are, let us keep these also as they are, the r values also
as they are.
Now, if the cycle can be lengthened, then the first indication is here that B inventory is 0,
when actually start producing this item, so the first thing I can do is, here I produce time
equal to 0. So, when I produce a time equal to 0, I have an inventory of 300, but then end
also have an inventory of 300, so I do not need to actually have this inventory of 300 at
all, if I have 0 here, I would still be able to do it, so this will be 0 ((Refer Time: 30:16)),
this will be 442, this will be 343, this will be 147 and this will be 0.
So, the next question is, if I ideally I can stretch my cycle further, T further if I produce
an item when the inventory comes to 0. So, the question is can I keep a 0 here and can I
distribute this 300 man-hours to other items, in such a manner that every time I start
producing only when my inventory reaches 0. So, here you are producing when the
inventory is 300, you are producing when the inventory is 102, you are producing when
the inventory is 192 and then you are producing when the inventory is 0. So, thing is can
I adjust it indirectly without explicitly changing it here such that, I can stretch the cycle.
So, this would mean that, we actually go through we look at this problem all over again
and then say that, now let us assume I have 1 cycle where I go through and at the end of
that cycle I am going to adjust the inventory levels, I am going to keep the 1400 constant.
But, at the end of the first cycle, I will kind of adjust the inventories to all the four items
334
such that, in what is called the steady state cycle that is going to follow, I would have an
inventory of 0 here when I produce D.
I will have an inventory of 0 when I produce A, I will have an inventory of 0 when I

produce C and I will have an inventory of 0 when I produce. Now, how do I reallocate it
through at the same linear programming problem, at the end of the first cycle is the next
formulation that we will see.
So, in the next formulation what we would do is we call t D dash, t A dash, t C dash, t D
dash as the start time, in the first cycle and t D, t A, t C, t D in the steady state cycle. So,
it is like saying I have a t D dash, t A dash, t C dash, t B dash, t D, t A, t C, t B, t D plus t
t A plus t and so on, t D, t A, t C, t B and t B in the real cycle. So, your actual cycle starts
from t D to t D plus T, which is this as well as t A, t A plus T and so on. So, let me just
write these constraints, so you will first have t D dash is less than or equal to you will
have the same maximize T.
And then you will have t D dash is less than or equal to 0.428, t A dash less than or equal
to 0, t B dash less than or equal to 0.666 and t C dash less than or equal to 0.625. So, the
first cycle obviously, has to begin before these things run out of items, then we get into
this, the first cycle should be such that, we also have three hundred plus, now I am
looking at item D which is my first item. So, I want to stretch my t D to such an extent
335
that, I am able to meet that demand up to this point, using the 300 that I have, plus what I
produce here.
So, 300 plus 2500 into t A dash minus t D dash is greater than or equal to 700 t D. Now
this constraint is going to tell me that, I am going to start producing t D in the steady
state cycle or the actual cycle somewhere here. But, I have already produced a certain
quantity for this period, so my production is 2500 into this length that should be equal
and 300 inventory that I have. So, that should be capable of meeting my demand up to
this point, when I actually start producing D again.
So, like that I write four constraints, which I write very quickly 200 plus 2500 into t C
dash minus t A dash is greater than or equal to 400 t A, 500 plus 2500 into t B dash
minus t C dash is greater than or equal to 800 t C and 400 plus 2500 into t D minus t B
dash is greater than or equal to 600 t B. Now, the fourth constraint is important, because
we are now looking at this, now if I look at item B, I am going to in the first cycle I am
going to produce B up to this period t D minus t B dash at the rate of 2500. So, this the
production in the first cycle, inventory that available with us is another 400.
So, with this I should be able to the meet the demand starting from here, up to the period
t B where I start producing again, so write these four constraints that will try to stretch
these things to the extent possible. Plus now I go back and write the steady state cycle
equations, now this is the period where I produce item D at steady state. So, now, I go
back to 2500 into t A minus t D is greater than or equal to 700 into T, because now at
steady state when I start making this item, now this ((Refer Time: 38:26)) is the
production, this is the demand period; so production should be greater than or equal to
the demand.
Similarly, I write four more equations, now t C minus t A, 2500 into t C minus t A is
greater than or equal to 400 T, 2500 into T t B minus t C is greater than or equal to 800 T
and 2500 into t D plus T minus t B is greater than or equal to 600 T and then we also
have all the t j dash t j T greater than or equal to 0. So, this gives us a much slightly
larger linear programming problem, which can be solved optimally to try and get the
maximum cycle length that is possible by playing around, or adjusting the inventories.
336
Now, when we solve this problem optimally, so let me write the optimal solution to this
problem somewhere here, t D dash is 0, t A dash is 0.1686, t C dash 0.3218, t B dash
0.666, t D 1.031, t A 1.475, t C 1.7005 and t B 2.1875 and capital T 1.52179 that is the
capital T cycle which goes. Now, the first thing that is happened is, the cycle time T has
increased from 0.8763 to 1.52179 almost doubled. And then if we go back and draw this
kind of a chart here, the first cycle ends and the steady cycle begins at 1.031 that is when
the steady cycle begins.
Now, D is produced again the second cycle will be 1.031 plus 1.521, which is roughly
2.552, so 2.552 D is produced again, D is produced again. Now, it is also possible for us
to draw a similar chart here, to understand the inventory positions and let me just quickly
give you the values.
337
So, time equal to 0, it is 300, 200, 500, 400 at time equal to 0.1686, it is 603 133 365,
299 we go through this, and point that is most important for us is 1.031, it is 0 170 536
694, 1.475 767 0 195 and 438, 1.7005 597 512 0 292, 2.1975 256 317 828 0 and 2.553 it
is 0 170 or 171 536 690. There are some minor rounding of errors, this is 2.552 is written
as the starting time of the next cycle of D, the last 3rd digit is the rounding off here, it is
given as 553.
Similarly, the 170 and 171 is due to round off, now what you observe out of this, the
movement the steady cycle begins, we produce D when the inventory is 0 Now, at this
time when we actually start producing A the inventory has come to 0, when we start
producing the next item C the inventory has come to 0. And when we start producing the
next item B the inventory has come to 0, whereas when we went here, which was the
steady cycle in the earlier case when we started producing D the inventory was not 0, but
300 and so on, only here it was 0.
So, by adjusting the inventories, in such a manner that at steady state we are getting 0,
we are able to stretch our T as much as we can. Now, once again back to the beginning
of the next cycle, it is 171 536 694, now let us do another quick calculation, which is
fairly obvious. Now, the sum of these is 1400. Now, inventory here is 1400.
So, inventory is not going to change, the total inventory is going to remain the same
largely because the total demand is 2500, total production is also 2500. So, what we have
338
effectively done is through the second problem, we have redistributed this 1400, which
was at 200, 400, 500, 300. Now, that has been redistributed to 0 170 536 694 to add up to
1400, so that the cycle time T can be made as large as possible. So, there are two aspects
to this problem, one is obviously, the cycle T depends on the total amount of inventory
that you have number 1.
If total demand is equal to total production, then the cycle time T number 1 is going to
depend on the total amount of inventory that we have, as secondly, and more
importantly, it also depends on the way in which the existing inventory is distributed, so
the existing inventory can be distributed in such a manner that your T exceeds. When I
say distribution one has to quickly understand that, it is 1400 man-hours of inventory, it
is not substituting A with D, it is about saying these many man-hours are used or
available there, so it can be redistributed.
So, distributing is helped us achieve a larger value of cycle time. Now, then we get into
another question, now would we use this model or would we use this model or would we
use a model that is somewhere in between. Now, the advantage of this model is that, if
my objective is to maximize capital T which is the cycle length, then I have really
stretched it to the limit, where I am going to have exactly 0 inventory, while I produce.
Whereas, here I certainly had a certain buffer or a cushion for at least three items when I
started production.
So, if the demand is purely deterministic and it is not going to change at all, then one can
take the risk of redistributing it, in such a manner that you achieve maximum cycle time.
And every time you reach 0, you produced or every time you produce when you reach
inventory, but reality is very different from these, we have made very simplistic
assumptions in both the models.
339
the first one is of course, is that the total P 2500 is sum of the demands, and we have also
made an assumption that this demand is going to be constant for every period, and it is
going to be continuous. Because, when we made these calculations, we have subtracted,
we have just multiplied the time by the demand, so it means the demand is stated as
continuous and so on. The movement the demand is not continuous and the demand start
showing some variations, then this model has to be very carefully looked at, because we
may get into shortage.
Now, we do not want shortage in reality, so either we have to put a certain safety stock
over this model, or essentially we use something similar to this, so that we stretch
ourselves only at the last point and for all other places, we have some cushion that we
have. So, it effectively boils down to playing around with the available inventory, in such
a manner that if we want a certain cushion to handle variation and uncertainty, it is
actually good to distribute it in a particular way. But, then if the demand is purely
deterministic, then we can try and stretch it to the extent possible that comes out of this.
Now, let us answer the next question, now both the linear programming models the
objective was to maximize T, which is to maximize the cycle time and make the cycle
time as large as possible. When we raised a question, whether it would conflict with the
principles of smaller run manufacturing and so on, the answer actually comes here. We
340
know by now that the T depends not only on these numbers, but also depends on the way
these numbers are distributed.
So, automatically if we want the cycle time to be smaller, then keep the total amount of
initial inventory on hand keep it smaller. If this 1400 total becomes 400, then
automatically T is going to move from 0.87 to somewhere like 0.5 or 0.25 or what, which
is like a weekly cycle. So, the answer still lies in playing around with these inventories,
or if organizations follow the principle of zero inventory, by which they want to have
lesser and lesser inventory.
Automatically the mechanics of the problem or optimization of the problem will lead us
to solutions, which have smaller values of cycle times. Another dimension to it is also to
see the tradeoff between inventory holding cost and changeover cost. And if we
changeover times are smaller, then automatically more changeovers are possible and
therefore, we also be able to make more variety, and then keep the cycle time as small as
possible.
So, even though the objective is to maximize T, in some sense we are not trying to do it
at the expense of today’s thinking, the model essentially tries to maximize T, but the
model still takes us to something that is practical and something that is the requirement
of today, which is to minimize the cycle. So, that we are able to have shorter runs of
larger variety, the answer lies in the amount of inventory that we keep on hand.
So, automatically if these values are smaller, the r values will becomes smaller and the
cycle time will be smaller. Similarly, if the changeover times and changeover costs are
smaller, then it is possible to have more changeovers bringing more variety, and produce
at shorter runs and smaller T. Another relaxation that we have to look at is, now can we
handle a situation where this demand varies with time right now, this has been both the
models have assumed that this demand is the same in every period; and this demand is
uniform.
Now, can we change these modules to meet situations where, these demand can be
different in different months, and this production capacities are also different in different
months. So, aggregate planning could give us solutions where the production capacity P
available is different in different months. So, we solve such a problem using a slight
341
approximation and a heuristic solution, which essentially leads us to another linear
programming problem, and we will see that model in the next lecture.
342
Prof. G. Srinivasan
Lecture - 20
Disaggregation – Time Varying Demand, Safety Stock – ROL for Discrete Demand
Distribution
In this lecture, we continue the discussion on disaggregation models. We have already

seen 2 models in disaggregation, we will look at the third model, where we consider that
the demand varies in each period and the production capacity also varies in each period.
So, let us explain that using an example, let us say there are 4 products A B C D. So,
these are the inventory positions that are available for the 4 products and these
inventories are given in man hours or person hours. And then there is a demand for these
products. In the previous example, we assume that the demands 400, 600, 800 and 700
for the 4 products in all the periods, the demand did not change and in the previous
model, we assume that the production capacity P is 2500, which happened to be the sum
of the demands.
Now, in this model we are going to assume, that the demand is going to change in every
period. So, we call this as demand 1 and then D 2 or demand 2, will be something like
this 600, 600, 600 and 500 say demand D 3 will be 300, 800, 500, 600 and so on. So, we
could have any number of months of demand that is known and most of the times these
343
demands are actually forecasted values of the demand for each of these products. We
also assume that P 1 production capacity available, in the first period or 1st month is
2500 and we would assume that P 2, this also change so this could be 2300 and let us say
P 3 could be 2,400.
Now, the first thing that we did, in the earlier example was to calculate this r, r is the
period up to which we need not produce and use the existing inventory to meet the
demand of the item. For example, for product A or item A the inventory, current
inventory is 200 and the first month’s demand is 400. So, with this 200 inventory, we can
meet 0.5 of the demand of the 1st month. So, 0.5 months we can use this inventory, we
need not produce. But, we have to make sure that we produce before 0.5 months
otherwise, we will be short of material or item. Similarly, inventory is 400 demand is
600. So, 4 by 6 is, 2 by 3, which is 0.666.
Inventory is 500 1st month’s demand is 800. So, 5 by 8 is 0.625 and 3 by 7 is 0.428,
based on these ratios, we also said in the earlier examples, that it is preferable and good
to produce in the order of increasing ratios which means, we produce D first, then
produce A and then C and then B so that before we have a stock out, we unable to
produce these items. There is another thing to note here for example, if this were 500
instead of 200. How do we calculate this r? Because it was 200 it happened to be less
than the first month’s demand. So, we said half of the 1st month can be meet with this
400. In case this was 500, then we would say that out of this 500 the first, some 400 can
be used to meet the 1st month demand and with the balance 100, we could meet 1 6th of
the 2nd month’s demand. So, this r would become 1.1666, 1 for the 1st month and 1 by 6
is 0.166.
So, we can calculate this r even proportionately or correctly even if, when the inventory
is higher than the 1st month’s demand. So, we also know that we have kind of decided
on the order in which we will make it, which is D A C and then B. In the earlier models,
we assumed that, all these values are the same as well as our each of these values the
demand does not change with time. So, what we have to do now, is we have to first find
out a cycle for the first cycle, essentially we need to find a first cycle. And then in order
to find out the first cycle, we are going to kind of pick part of this data and then try to
look at a demand up to a certain period.
344
For example, based on the experience of the two models, that we have seen earlier, we
know that the cycle time is more than this 0.666 which is the largest of the ratios. The
reason is that, in both the models our objective was to make T, the cycle time as large as
possible and that would invariably happen when the last item, which means the item that
has the highest r value. The inventory position comes to 0 and then we start producing
that item. So, the T which is the cycle length is bigger than the maximum of the r.
So, a quick thumb rule would be that, the cycle time T would be bigger than r and
roughly of the order of r plus 1 by r, this is just a thumb rule and this is not a optimal in
any way, with some kind of a guideline that the maximum cycle length T, will be of the
order of r plus 1 by r. In this thumb rule of r plus 1 by r, we have to take the maximum
value of r, which happens to be 0.666, 0.666 is 2 by 3. So, r plus 1 by r is 2 by 3 plus 3
by 2, which happens to be 13 by 6 which is 2.166. So, r plus 1 by r gives us a value of
2.166. So, what we do now is, we try and take the demand of each of these items for
2.166 periods and we also try and take the production for 2.166 periods.
So, let us write that so the 4 items A B C D and let us say equivalent demand for 2.166
period, (Refer Slide Time: 00:10) for A would be 400 plus 600 for 2 periods plus 0.166
of 300 which is 50. So, 1050, 1050 divided by 2.166 is 484 now, 484.63 we could even
make it as 485. Now, this represents the equivalent or average monthly demand or
average demand per period, (Refer Slide Time: 00:10) if we take 2.166 periods and
345
average it out. So, this value of 485 is 400 plus 600 plus 0.166 into 300 divided by 2.166.
Now, for B it will be 600 plus 600 plus 0.66 into 800, which is 132 plus 1200 divided by
2.166, which is 615.32.
So, this will be 615, for item C it will be 800 plus 600 plus 500 into 0.166 which is 83
plus 1400 divided by 2.166 which is 685. And for D it will be 700 plus 500, 1200 plus
0.166 into 600 divided by 2.166, would give us 600. Now, the P value will be 2500 plus
2300 plus 0.166 into 2400, would give us 2400. So, what we have effectively done here
is to try and find out, what is the equivalent demand or average demand per period, if we
consider 2.166 periods. Now, the production also has been subsequently averaged for
2.166 periods. Now, we could also get into a situation where, the average monthly
production the equivalent term which is 2400, may not be equal to the sum of these 4.
For example, sum of the average demands is 2385 which is less than 2400.
In such situation, if we utilize all the 2400 production capacity there will be a small
inventory buildup, if this happens to be more than the 2400, then obviously the demand
being more than production, at the end of the cycle, we would have consumed something
from these values (Refer Slide Time: 00:10). That is precisely the reason, why these
things are required, they serve 2 purposes, they serve one purpose where they allow us to
have enough inventory, till the production of the item begins. And the second purpose is
when the demand, some of the demands exceeds the production capacity at the end of
the cycle, the total will be less than this total, but then this also acts as a cushion.
As long as at steady state or at the end of several cycles, if sigma P is equal to sigma D
production capacity is equal to the demands, then at the end at any point, the inventory in
the system will be a sum of these. So, these act also as a cushion to ensure to take care of
two things, variation in the demand, initial inventory available till we produce the item.
And there is a third aspect which is in situations, where demand is more than production,
some part of this inventory will be consumed. But, at the end of it, it will level itself
because finally, the average production will be equal to the average demand so it serves
multiple purposes. So, it is also not necessary that these should add up to the 2400.
The question is do we have to compute r again considering the 4 demands of 485, 615, 685
and 600? The answer is we do not compute r again, because these demands represent the
weighted or average demand. Over a period of 2.16 months whereas, (Refer Slide
346
Time: 00:10) the inventory that is available which is 200, 400, 500 and 300 is capable of
meeting the demands of 0.5, 0.66, 0.625 and 0.428 months with respect to this data.
Therefore, we retain the same r values and then we try and solve a linear programming
problem that would try and maximize the cycle time or the run time. So, we will now
sort A B C D in the order of r and we will do D first, followed by A and then C and then
B. So, we now go back and formulate our familiar problem.
Where we try to maximize the cycle time T, so the objective will be to maximize the
cycle time T.
347
Now, we will start producing them, in the order D A C and B.
And then we will say let t D, t A, t B, t C represent, the start of production, of D A B C

respectively. (Refer Slide Time: 14:36) Then we have to use these constraints t D less than or
equal to 0.428, t A less than or equal to 0.5, t C less than or equal to 0.625 and t B less than
or equal to 0.666. Then we produce the item D for a period t A minus t D. So, t A minus t D
into 2400 is the production rate. So, t A minus t D into 2400 should be greater than or equal
to, (Refer Slide Time: 14:36) the demand of A is 400 and demand of
348
D is 600 so that will be greater than or equal to 600 T. Similarly, t C minus t A into 2400
should be greater than or equal to (Refer Slide Time: 14:36) 485 T, t B minus t C into
2400 is greater than or equal to (Refer Slide Time: 14:36) 685 T.
And t D plus t minus t B into 2400 is greater than or equal to 615 T, t D, t A, t B t C, T

greater than or equal to 0. So, this is our linear programming formulation. We have seen
this formulation before; this is the first of the disaggregation models. Now, the
explanation for this comes from the fact that first, we start producing (Refer Slide Time:
14:36) D at t D and then we stop producing D at t A so D is produced for the period t A
minus t D.
So, this is the total production and the cycle time is T. So, this should be capable of
meeting the demand of the entire cycle so (Refer Slide Time: 14:36) that is 600 into T.
Similar explanations can be given for the production, this is for the production of A, this
is for the production of C and production of B.
Will start at t B and then again the next cycle of D will happen and therefore, this is the
entire cycle, which is T that goes up to this. So, t D plus T is the time at which D’s
production starts again. So, t D plus T minus t B this much is the length of time of
production of B. (Refer Slide Time: 14:44) So, this is the total production of B, should be
greater than or equal to the total demand of D during this period. We will follow model
1, we will not follow model 2, we will not follow the transient model. We will use the
349
steady state model and try, which means we also will not try and adjust (Refer Slide
Time: 00:10) these inventories in or in turn adjust the r j’s and have a transient model.
We will not do that; we will follow a steady state model. So, we try to maximize the T.
So, when we solve this linear programming problem. We would get t is equal to 0.903
months and we will have values which will be t A equal to 0.2258 so t A is equal to
0.2258, t D will be 0, t B and t C. So, t C is 0.4082, 0.666, 0.4082 and this will be 0.666.
So, this is the solution once again. For ease of implementation, we will take T is equal to
0.9 instead of 0.903. And then we will say that we will implement the first cycle for a
period of 0.9 months. So, when we do that we implement the first cycle for a period of
0.9 months. And in these 0.9 months, we will start producing D for a period of 0.2258.
Then we will start producing A, which is between this and this. And then we start
producing B which is the time period between this and this. And then the next production
of D will start at time equal to point 9.
So, we will implement the first cycle by doing this way. Now, at the end of the cycle we
will now, want to take stock of the inventory positions as well as the demand positions.
Now, what can happen is ordinarily, (Refer Slide Time: 00:10) if this demand was
deterministic which means this demand did not change. If this demand remained at 400,
600, 800 and 700, then (Refer Slide Time: 17:50) we would have produced D A C B in a
certain way. And we would have consumed exactly (Refer Slide Time: 00:10) 90 percent
350
of 400, 600, 800 and 700 in this period. We also had this initial inventory to begin with.
So, there is an initial inventory, there would be some production for each of these items.
And then there would be some consumption for each of these items.
So, if we assume that the demand was exactly deterministic. Then we can go back and
find out, then we would have consumed 90 percent of the demand of all of these. So, the
unmet demand once again for A B C D, for A B C and D, the unmet demand for A B C
D would be 40, which is 10 percent of 400, 40, 60, 80 and 70. The second month’s
demand would be 600, 600, 600, 500 and then D 3 would be 300, 800, 500 and 600. So,
the data will now, move to something like this, where 10 percent of this demand has to
be meet. Now, we are making an assumption right now that the demand is not going to
change here (Refer Slide Time: 00:10).
Now, if for some reason within this 0.9 months, the demand had changed a little bit
(Refer Slide Time: 00:10). Then we could have consumed some other number, other than
90 percent of this. And then these data also will change accordingly. So, if there is a
change because of the change in demand during this period then we can update the 10
percent of the demand that is yet to be met. But, for the sake of this problem, we will
assume that the demand is deterministic. And therefore, exactly 90 percent would have
been consumed.
Now, again there is also a possibility that these demands (Refer Slide Time: 00:10) can
351
change. In fact, we could even have updated the demand for the 4th month. We would
assume that, by that time we would have an idea of the demand of the 4th month so we
could have updated, so even some of these numbers may change a little bit, depending on
the changes in the demand. Now, at time equal to 0.9, we will take stock and we will
update these numbers. We will also update the inventory position at time equal to 0.9.
Now, that inventory position would depend (Refer Slide Time: 00:10) on the initial
inventory plus the production, less the demand will be the final inventory. I will just
show the computation for the item D that, we would be doing first.
So, from this D is produced for a period of 0.2258 months. So, total production of D will
be 0.2258 into 2400 so that would become so this is, this will be the production of D. So,
the inventory of D will be, the starting inventory of 300 plus production 0.2258 into 2400
less demand of D, which is 700 into 0.9. So, this will be the inventory position of D and
this works out to 209.82, 209.82.
352
Now, like this we can update the inventories. So, this is updated to 209 and the rest of the
numbers are 276, 427 and 581. Now, once again these numbers have been computed,
assuming that the demand is deterministic. But, if there are some changes, then these
numbers will also correspondingly change. So, now we are at time equal to 0.9 and then
we can start doing this one more time by trying to understand, how many months can this
inventory of 276? How many months demand can this 276 meet? So now, it will meet
0.1 month which would be 40. So, the balance 236 and the rest of them will be 236 by
600.
So, this will be 0.1 plus 236 by 600. So, this will meet so for this r will be 0.49. So, for
this r will be 0.49 rest of the r’s will be 0.49, 0.72, 0.94 and 0.378. Now, using this
particular data now, we have to again find r plus 1 by r like we did earlier. So, we will
have the new value of r plus 1 by r (Refer Slide Time: 07:04) and then for that new value
of r plus 1 by r. (Refer Slide Time: 17:50) We will now, have to adjust and get these
equivalent demands, as well as equivalent production. Now, the equivalent production
numbers will be we have used for 90 percent of the times.
353
So this will be 10 percent so remaining production will be 250 here, then 2300 for this
month and 2400 for this month. Now, for the new value of r plus 1 by r, we now have to
find out the equivalent demands and the equivalent production like what we did. And
then solve another linear programming problem, to get the length of the second cycle.
The length of the second cycle need not be 0.903, it can be some other number. So, like
this, we continue trying to have cycles of different lengths. So, this approach may not be
the most optimal or the best approach but this is practical and this can help us in actually
trying to solve the disaggregation problem with time varying demand.
Now, we also have these inventories, (Refer Slide Time: 00:10) we do not try to follow
the transient model. We do not try to push these inventories to 0, because each cycle is
now, going to be different. Also when we do not push this to 0 and use these inventories
as they are. We will be able to absorb fluctuations in demand, over these months and
then these demands get periodically updated and so on. So, that is how the
disaggregation problem with time varying productions and demands, is actually solved in
practice, using linear programming technique.
354
So, far we have seen several models that involved inventory of items. We started with
very basic inventory models, where items were bought out. We looked at models where
we allowed backorder deliberately. We also had models of where, we avail discount,
where we looked at constraint inventory problems for multiple items. And then we
looked at production consumption models, then we looked at time varying demand
models. We looked at economic lot scheduling problem, which have production and
inventory. And then we also looked at the disaggregation problems, where we try to
maximize the cycle length with certain restriction on the inventory.
Now, in all these models that we have seen the demand and more importantly is the
demand was taken to be deterministic. The monthly demands were known many times,
the demand was the same, it was constant and continuous. There were some models
where the demands changed at different periods. But, nevertheless the demands were
known. So, let us go back to the basic model and try to relax one or two of the
assumptions, to see, what kind of models such relaxation of assumptions lead us to. So,
we go to the very basic inventory model, where we buy an item. So, we have this as the
time axis, we order a quantity Q. So, let us assume at the beginning so this is time, this is
inventory.
So, we begin with stock equal to Q and then we consume this Q at the rate of D till it
comes to 0 and then we place an order here. So the stock comes up here back to Q. Once
355
again, we consume it at the rate of D, you can take these 2 to be parallel and then it
reaches 0 and then an order is placed and this goes on now, this is called the Saw tooth
model of inventory. Now, this is T the length of the cycle. Now, we made certain
assumptions here the first assumption of course is that, the demand is continuous and the
demand is the same D per year at every instance. Second and the most important
assumption is that of instantaneous replenishment, which means the moment, we place
the order we get the items. Now, in reality these 2 things are not the way we have
assumed.
So, first thing we need to look at is, the demand need not be deterministic and
continuous. We will come to that a little later but then there would always be a time
between placing an order and receiving the item. Now, that time is called the Lead time.
And then in this model we have assumed that, the lead time is zero. So, the moment the
order is placed, the items arrive. Now, we look at several cases of what happens to the
lead time. So, the first one is, let us say lead time is not zero. But, lead time is a known
value say, let us say let us look at a situation where lead time is equal to 1 week. Now, let
us go back to the numerical example that we have used earlier.
So, we have used a numerical example where, the demand for the item is 10,000, the
order cost C naught is 300, the carrying cost is rupees 4 per unit per year, this gave us the
economic order quantity Q as root over 2 D C naught by C c which is 1224.74 and we
would take total cost as 4898.98 which is the sum of the order cost, as well as the
inventory carrying costs. Now, let us assume that we are not exactly implementing
1224.74. Let us say we are implementing a more comfortable number of 1250, which
would also give us exactly 8 orders per year. Now, if we say that this 10,000 is actually
based out of, let us say we work for 50 weeks in a year.
356
And therefore, we would say weekly demand is equal to 200. So, that the total demand is
10,000 per year. Now, if we say that our lead time is 1 week, then what we do is, we will
not place an order when the stock is equal to 0. We will look at, since they it is going to
take 1 week for us to get the items. And we know that the weekly demand is 200. Let us
assume that the 1 week stays and there is no change, it is exactly 1 week and every week
the demand is exactly 200. So, in such a situation, we would place the order when our
stock position reaches 200 so that exactly in 1 week the 200 will be consumed. The new
order will arrive (Refer Slide Time: 28:56), so stock position will be zero exactly when
the order arrives and its get replenished to 1250.
357
So, let me write this as Q equal to 1250 and then let us say, that this is the position where
we have stock equal to 200. Let me use a different color so this is the position where
stock is equal to 200 so this is 200. So, if this stock comes to 200 so at this place at this
point, we will place the order. And then we start consuming so after a week the stock
will come to zero. And exactly at that time, the new consignment will arrive.
So, the point at which, we place the order is called the Reorder level, which is
abbreviated to R O L. And the reorder level is equal to the lead time demand in the case
358
of the lead time and the demand, being deterministic. If lead time is not 1 week but if the
lead time is exactly 2 weeks, (Refer Slide Time: 35:08) then instead of placing the order
at 200 we would place the order at 400. More importantly the definition of reorder level
is, the stock position at which we place the order. Now, we need to compute the reorder
level, under certain assumptions of the lead time, the demand and the lead time demand,
lead time demand is called demand during the lead time. So, if the lead time is exactly 1
week, lead time demand is 200 in this particular example.
Now, if the lead time demands, let us say, let us look at a situation where the lead time is
exactly 1 week but the demand is not deterministic, it is not exactly 200, it could be 200
plus minus something. So, in certain now, what happens is. There are exactly 8 cycles in
this year (Refer Slide Time: 37:19) because what we do right now, is we assume that, the
moment we compute the reorder level. (Refer Slide Time: 37.33) Let us say we are now,
going to compute the reorder level. Now, the moment we compute the reorder level, we
would place and, as soon as the stock comes to the reorder level, we place an order.
Now, we assume that with the quantity which is called the reorder level, which is with
us, we will be able to meet the lead time demand, if we are able to meet the lead time
demand, there is inventory before the next cycle, next order, next consignment comes. If
you are not able to meet the lead time demand with the reorder level quantity, there could
be some shortage or some backorder. Now but on an average even though the demand
fluctuates, let us say every week. But, on an average, the demand is going to be 200
every week. So, the expected demand for the year is going to be 10,000 and since (Refer
Slide Time: 28:56) we are going to have 8 cycle of 1250 ordered, then we will have
exactly 8 cycles where we will be ordering 1250 in each cycle. But, then the reorder level
has to be found out so we have separated two things, we have separated the order
quantity and the reorder level.
So, in this particular way of analyzing the problem, we define the order quantity
separately based on economic order quantity model or a value that is suitably rounded
from the economic order quantity which is 1250. So, this model would be like saying
that every time we order, we place, we place order for 1250. We expect to order 8 times
an area and we place an order when, the order quantity is, when the stock that we have is
200, or if we compute the reorder level differently now, whenever we reach the reorder
level, we will try, we will place an order. Now, say the lead time is exactly 1 week (Refer
359
Slide Time: 28:56). But, let us look at a case where the (Refer Slide Time: 35:48)
demand is not exactly 200. But, the demand follows a certain distribution.
So now, we would say that the demand can be anything from 100 to 300. So, demand
can be 100 with the probability 0.1, demand can be 150 with probability 0.2, demand can
be 200 with probability 0.4, demand can be 250 with probability 0.3 or 0.2 and demand
is 300 with probability 0.1. So, let us say this is the weekly demand and we will take this
as lead time demand because lead time is exactly 1 week. Now, we have to find out the
reorder level that optimizes our total cost. Now, this is the demand during the lead time
and lead time is exactly 1 week. So, let us assume that we have placed an order, when we
have reached a certain reorder level, which we are going to calculate using this data.
So, once we our stock reaches that reorder level, we place an order and we believe that
the reorder level quantity which we have, which is the stock on hand when we place the
order, is capable of meeting the lead time demand. And we want to find out that reorder
level. Now, if we keep now, the expected value of lead time demand, will be 100 into 0.1
plus 150 into 0.2 plus 200 into 0.4 plus 250 into 0.2 plus 300 into 0.1. Now, this will add
up to 200 because it is 0.1, 0.2, 0.4, 0.2, 0.1 and the gap is that, is symmetric about the
mean so 200 is the expected value of the lead time demand.
Even before, we proceed further; we could ask ourselves how did we get these numbers?
How do we get these probabilities? These probabilities can be thought of as we go back
360
in time collect data, historical data has to, how the lead time demand has behaved over
several past cycles? Let us say someone has collected such a data for 10 past cycles and
we have realize that the lead time demand has been 100 in 1 out of the 10 times 150 in 2
out of the 10 times 200, 4 out of the 10 times and so on. So, a proportion automatically
represents the probability of occurrence under the assumption, that future will replicate
the past.
So, that is how we get these numbers and the expected value is 200. Now, if we take the
reorder level to be expected value of lead time demand. (Refer Slide Time: 35:48) In this
case we took exactly equal to lead time demand, because lead time is deterministic,
demand is deterministic, if we take expected value of lead time demand based on this
thing. Now, about keep this roughly about 70 percent of the times, we will be able to
meet the demand and 30 percent of the times, we will not be able to meet the demand
because of 200. So, the thing is we will be able to meet the demand more times then we
will, than the situation where we will not be able to meet the demand.
But, then if we look at from a symmetric point of view, say roughly 50 percent of the
times we will be able to meet at, 50 percent of the times you will not be able to meet it. If
this is not 100, 150, 200, If this were 100, 1 naught 1, 1 naught 2, 1 naught 3 up to 300,
then we will say 50 percent of the times, we would have met 50 percent of them we
would not met. So, whenever we are meeting the demand there will be excess inventory,
where we are not able to meet the demand. We incur a backorder cost or a shortage cost
we assume backorder, which means as soon as the stock arrives, we meet the demand
and then we continue.
Now, when we have excess inventory, we incur inventory cost. When we have shortage
or backorder, we incur backorder cost. So, backorder cost is always higher than the
inventory cost. Therefore, we want to avoid backorder. And therefore, we end up having
this reorder level always as greater than the expected value of lead time demand. If C s is
greater than C c, C s is shortage cost, C c is inventory cost, shortage cost is higher than
inventory cost. So, reorder level is always greater than or equal to the expected value of
lead time demand.
361
So, reorder level is now, equal to expected value of lead time demand plus safety stock.
This S S is called safety stock so we could say either we want to compute the reorder
level or we would say we want to compute the safety stock. Now, the simplest thing to
do here is, to say that we want to find out the reorder level. So, we can look at all these 5
cases, where reorder level is equal to 100, reorder level is equal to 150, 200, 250 and
300.
So, let us do those calculations. So, let us have a situation where reorder level is equal to
362
100 so when reorder level is equal to 100 which means when the stock level is going to
be 100. We are going to place an order and the demand is either 100 or 150 or 200 or 250
or 300. Therefore, there is going to be no inventory at the end of the period because
whatever is the demand, we have to consume the 100. So, there will be no inventory in
the period so then what will happen is, there will be only shortage. So, expected value of
shortage in a cycle will be, if the demand is 100 there is no shortage. If the demand is
150 shortage is 50. So, 50 into 0.2 so 50 into 0.2 is 10 plus 100 into 0.4 is 40 plus 150
into 0.2 is 30 plus 200 into 0.1 which is 20, this is the expected shortage.
Let me explain, the computation again, if the reorder level is 100 which means our stock
is 100. If the demand is 100, then we are able to meet all the demand, no problems but
when the demand is 150 (Refer Slide Time: 44:27) we will meet only 100 out of the 150
so shortage will be 50 which happens with the probability of 0.2 so 50 into 0.2 is 10.
Similarly, 100 into 0.4 is 40 and so on. So, expected shortage is 100, there is no expected
inventory out of this. Now, total cost will be order cost plus carrying cost plus shortage
cost. Now, order cost is going to be constant, because we are going to order 1250 eight
times. (Refer Slide Time: 28:56) So, 8 into 300 is 2400, we have assumed order cost as
300 so 8 into 300 is 2400.
Now, what is the inventory carrying cost? The cycle inventory that we will carry is Q by
2 that is the expected inventory, normally expected inventory is Q by 2. Now, in this case
what will also happen is the actual cycle inventory will be Q by 2 plus the safety stock
that we have. So, in this case we will have Q by 2 minus 100 because there is an
expected shortage of 100 every time so Q by 2 minus S S. So, it will become 1250 minus
safety stock of 1250 divided by 2 Q by 2 is 625 minus 100, would give us 525 that is the
average inventory into C c, C c will be 4 per unit per year plus expected shortage is 100.
Now, let us take the shortage cost in this example, as rupees 2.5 per unit backordered.
So, C s is taken as 2.5 per unit backordered.
So, 100 is the expected shortage cost in a cycle, there are 8 expected cycles in a year. So,
into 8 into per unit is 2.5. So, this is 2400 plus 525 into 4 is 2100 plus 800 into 2.5 is
2000, this is 6500 is the total cost. Now, let us do one more computation if R O L is
equal to 150, if R O L is 150, first we calculate safety stock, is equal to minus 50. Now,
this comes because expected value of lead time demand is 200. So, from this definition
(Refer Slide Time: 44:27) R O L is equal to expected value of lead time demand plus
363
safety stock. Now, for the sake of computation we are showing the safety stock, is minus
50.
So, now once again expected inventory will be now, the R O L is 150. (Refer Slide Time:
44:27) So, we will have inventory, if the demand happens to be 100. So, there will be an
inventory of 50 in this cycle, there will be an inventory of 50 this cycle, with the
probability of 0.1 so expected inventory is 5. Now, expected shortage will be (Refer
Slide Time: 44:27) R O L is 150 so there will be shortage if the demand is 200 or 250 or
300. So, this is 50 will be the shortage if the demand is 200 so 50 into 0.4 which is 20
plus 100 into 0.2 which is another 20 plus 150 into 0.1 which is 15 so expected shortage
is 55.So, total cost will be, order cost plus carrying cost plus shortage cost so order cost
is 2400 plus carrying cost is Q by 2 minus safety stock as a general rule. So, 625 minus
50 is 575 into 4 plus 55 into 2.5 into 8. So, this would give us 2400 plus 2300 plus 1100
this is equal to 5800.
364
(Refer Slide Time: 44:27) So similarly, if you we compute for 200, 250 and 300. If we
do those computations we get, if R O L is 200 we get, we would get total cost equal to
5300. If R O L is equal to 250, our total cost is 5200 and if R O L is 300, our total cost is
5300 per year. Now, based on all the total costs that we have, the total costs are 6500
(Refer Slide Time: 45:07), 5800 (Refer Slide Time: 50:43) 5300, 5200 and 5300. If we
go by minimum total cost, then we would look at R O L is equal to 250. So, if we go by
minimum total cost our R O L is 250 and our total annual expected cost is 5200, which is
the cheapest.
Now, if we have reorder level of 250, then we will have a safety stock of 250 minus 200
which is 50, safety stock of 50. And we also realize that this 50 safety stock is something
that we will need not end up consuming at all. Because if we take every week or every
cycle, the expected demand is 200 but then we have an additional 50 which is the safety
stock. So we have a 250 so in some weeks the demand can be more than 200. Some
weeks the demand can less than 200 but when we average out all the demand, the
demand is going to be 200 per week. But, then the safety stock of 50 is going to be with
us all the time. And we will not consume this safety stock at all. Now, this safety stock of
50 to put it in a different way this safety stock of 50 would incur, a cost of 50 into 4
which is 200 that we have.
Now, this cost for 1250 (Refer Slide Time: 28:56) would actually give us 5000. So, we
365
have this 5200 which is there, plus a very small shortage cost that is also involved. So,
roughly the figure is close to that 5200 that we actually have, plus if we calculate the for
(Refer Slide Time: 28:56) 1250, if we calculate D by Q C naught plus Q by 2 C c, then
the total cost will become 8 times 300.
Which is 2400 plus Q by 2 into C c which is 625 into 4 so 2400 plus 2500 will be 4900.
The additional inventory cost is 200 which is 5100. Now, the expected shortage is, if
reorder level is 250. The shortage is 50 (Refer Slide Time: 44:27) so 50 into 0.1 is 5. So,
expected shortage cost will be 5 is the unit short 8 times a year, into 2.5, which is 40 into
2.5 is 100. So, the order cost is 2400. The normal inventory safety inventory carrying
cost which is Q by 2 into C c is 2500. (Refer Slide Time: 53:15) The excess 50 safety
stock incurs a cost of rupees 4 and note that we do not put safety stock by 2 here we put
Q by 2 in the calculation because of the average inventory. The safety stock is not going
to be consumed at all.
So, the safety stock of 50 will be carried forever and it would incur an additional cost of
200 plus a shortage cost of 100. So, 4900 plus the extra 300 is the expected additional
cost, if the demand instead of being exactly (Refer Slide Time: 44:27) 200 every week,
follows a distribution like this so the effect of that is an increased inventory of 50 which
is our safety stock and an increased cost of expected increased cost of 300 per year.
(Refer Slide Time: 44:27) Now, with an additional safety stock of 50 and a reorder level
366
of 250 which means here, then every time or every cycle, we will be able to meet the
demand 80 percent of the times. So, the demand in a cycle is met 80 percent of the times
that is called service level is, 80 percent. It is the number of percentage of times, we will
be 90 percent of the times. So, this is 0.1, 0.3, 0.7, 0.9.
So, we are able to meet the demand in 90 percent of the cycles, we will be able to meet
the demand. There is also another quantity called fill rate, which is like this. Now, if we
have a (Refer Slide Time: 44:27) reorder level of 250. If the demand is 100, 150, 200,
250 we will be able to meet all the demand during the lead time. So, 90 percent of the
times, we are able to meet all the demand. (Refer Slide Time: 44:27) And if the demand
happens to be 300, we will actually meet 250 out of the 300, 50 is the only thing that is
short. So, we will be able to meet 250 out of the 300 which is 5 by 6, 250 by 300 is 5 by
6 which is 83.33 percent of the times, we could meet and that is going to happen 0.1 of
the time.
So, the fill rate will be 98.33 percent of the demand I will be able to meet. 90 percent of
the times I will be able to meet 100 percent of the demand 83.3 percent of the demand, I
will be able to meet with the probability of 0.1 so if the fill rate is like 98.3. So, the fill
rate is higher than the service level, many times we are concerned about the service level,
which is our ability to meet all the demand. There are times; we also look at the fill rate,
where we are able to meet part of the demand. So, if the demand follows a discrete
367
distribution like this now, we have found the way by which we can calculate the safety
stock as well as the reorder level. Now, what are those computations, if the demand
follows different distributions or the demand follows a continuous distribution such
models, we will see in the next lecture.
368
Prof. G. Srinivasan
Lecture No. - 21
Safety Stock – ROL for Normal Distribution of Lead Time Demand
In the previous lecture, we looked at probabilistic inventory model.
Where, we took the same numerical illustration and then we also spoke about the reorder
level and the demand not being deterministic during the lead time. We took an example
where the annual demand is 10000 and then if we assume that we work for 50 weeks in a
year, then the average weekly demand is 200 and then we said that the demand is
actually not 200 per week, every week. But, it can vary. And then we also said that, the
lead time was 1 week in this example. And therefore, the demand during the lead time,
was assumed to follow this discrete distribution, with an expected value of 200.
So, expected value was 200.We also define the reorder level, which is the level at which
we will place the next order and we also define the safety stock to be the difference
between the reorder level and the expected value of the lead time demand. And we
showed the total cost computations for reorder level equal to 100, 150, 200 and 250 and
300 and then at the end we said that, reorder level of 250 was optimum.
369
So, reorder level was found to be 250 which was optimum, which had the least annual
cost of rupees 5200. We also explained, how we got that 5200. Now, and we compared it
with the basic inventory model. So, if we looked at the basic inventory model which is
our familiar Saw tooth model, so, the basic inventory model with continuous demand
when we optimized, would give us Q equal to root over2 D C naught by C c. D is 10000
C naught is 300 C c is i into C which is 20 percent of rupees 20 which is 4. So, this
would give us 1224.74, we also said that we would assume an order quantity of 1250
which would give us a total cost of D by Q C naught plus Q by 2 Cc, which would give
us a total cost of, this would give us 8 orders per year.
So, D by Q into C naught will be 2400 plus Q by 2 C c will be 2500, which would give
us a total cost of 4900 here. In this computation (Refer Slide Time: 00:19) we also
assumed that the shortage cost was rupees 2.5 per unit backordered and we explained
that this total cost of 5200, was actually could be split into 3 components which is 4900,
which is the cost of the safety inventory of Q by 2 into C c plus D by Q into C naught,
which came from here. The reorder level being 250 (Refer Slide Time: 00:19) and the
expected lead time demand being 200. We have an excess stock or safety stock of 50,
which is continuously carried. So, this 50 is carried at the rate of rupees 4 per unit per
year.
370
So, that would give us an additional 200 and then we also said that the shortage would
give us an additional cost of 100, which would give us a total of 5200 per year. So, we
could explain all of these. Before we proceed further, let us look at another thing where,
when we develop the basic inventory model, we also made an assumption to the basic
inventory model that backordering could be allowed and then we derived model which
allowed backorder explicitly. So, let us go to that model and try and compare and
contrast this particular model and this and the result with the deterministic case when
backordering is allowed. (Refer Slide Time: 01:51)
So, when we look at this problem and then when we look at the deterministic inventory
where, we allowed backordering. In this case, where we would order a Q and then the
demand is continuous it comes up to 0. Assuming instantaneous replenishment we do not
place an order here, we allow certain amount ofbackorder to build and we call that
backorder quantity as s. And then once we reach a backorder quantity of s, we place an
order and because of instantaneous replenishment we would get that Q instantly.
Instantly we replenish this s, which is the backorder that has been built and therefore, the
top inventory is called I m maximum inventory that we hold and that is equal to Q minus
s and this cycle proceeds.
We have, already derived an expression for this and that would be like, the economic order
quantity Q there is given by root over 2 D C naught by C c into C c plus C s by C s.
371
And we worked out a numerical example, where (Refer Slide Time: 01:51) we took C s
to be rupees 25 per unit per year and therefore, our Q in that case was root over, 2
into10,000 into 300 by 4 into C c plus C s by C s, which is 29 by 25 and this came to
1319.09. The shortage that we allowed, this s backorder quantity that we allowed sis
given by Q C c by C c plus C s which comes to 181.94, while T C the total cost for this,
would become D by Q into C naught plus Q minus s the whole square by 2 Q into C c
plus s square by 2 Q into C s which was 4548.59.
Now, let us compare this model, which is a deterministic demand model with this model
where the demand is probabilistic. Let us also try and understand the differences and try
to explain them. (Refer Slide Time: 00:19) Now, this model when we optimize (Refer
Slide Time: 01:51) the total expected annual cost is 5200 which included the cost of
ordering, cost of holding the inventory, including the buffer and the expected shortage
cost and the components were 4900, 200 and 100 The very minor approximation was
there, because if we had ordered 1224.74 then, the total cost would work out to 4898.98.
When we change the order quantity to 1250 for the sack of convenience, the very
marginal rounding off error which was rounded off to 4900.
So, that is how this 4900 comes, cost of excess inventory is 200 the safety and expected
shortage cost is 100. The first difference that we see is that in this model, the cost is
actually lesser than 4898.98. Whereas, (Refer Slide Time: 01:51) in this model the cost is
more than the value of 4900, that is the first difference. Second difference of course, is in
even though we use the same notation C s, here C s is defined as (Refer Slide Time:
01:51) rupees per unit per year whereas, the C s that we used when we did those
calculation (Refer Slide Time: 00:19) was rupees per unit backordered. This is not per
year, this is unit backordered.
And therefore, when we work this out we multiplied this by another D by Q. Because
this would be for a cycle, expected shortage in a cycle into the shortage or backorder cost
and then the number of cycles per year. Whereas, when we derived this, we have
conveniently defined C c and C s to have the same unit. So, that we could have C c plus
C s by C s coming here, we could also have C c by C c plus C s coming here. So, C s is
defined here as rupees per unit per year. Next is how do we explain that, the cost is
higher here, cost is lower here in fact we even went ahead and said that as long as we
have a C s - when C s is equal to infinity, then this would approach the basic model.
372
And as long as, we have any C s this model will be totally cheaper the total cost would
be less than that of the basic model. The order quantity Q will be slightly higher, because
this is 1 plus something and there will be a small s and as this C s keeps increasing, as C
s keeps increasing C c plus C s by C s keeps decreasing. And therefore, this as C s equal
to infinity it will eventually come to the old model, s will become 0 and T C will
approach the other T C. Now, where is the difference? The difference is as follows; in
this model the expected shortage cost was calculated, only when the demand exceeded
the reorder level. So, there is an expected lead time demand, then there is an additional
buffer.
And when the, actual demand exceeded that, then the shortage came into the picture.
Whereas, this is still a deterministic case, where the demand is not going to exceed, it is
going to be the same D. There is no additional buffer. Because there is no additional
buffer, we do not have (Refer Slide Time: 01:51) this contribution of 200. This 200
contribution (Refer Slide Time: 00:19) in this model came because R O L was fixed at
250 expected L T D is 200, lead time demand is 200 and the differenceis 50, which is the
additional inventory that is carrying.
Therefore, this component this 200 component does not exist in this model.This model
actually has only 2 components which is this component and this component. So, when
these two put together is 5000 we still seem to be getting a lesser value here, which we
will try and explain now. (Refer Slide Time: 01:51) Actually this C s value is taken or
slightly different here its rupees 25 per unit per year,(Refer Slide Time: 00:19) here it is
actually 2.5 per unit backordered, which we generally multiply by the number of cycles.
So, if the E O Q is 1250, we have 8 cycles in a year and this can be taken as 2.5 into 8
which is 20, which could be compared with 25 so, let us say they are comparable.
The real reason, why this value comes down is because here by building a certain
backorder, by building a certain backorder and by incurring a certain backorder cost, we
realize actually two things. We realize that, what is the other alternative for not building
this backorder? The other alternative for not building this backorder, is to add it to the
order quantity, do not allow the backorder, which means an additional quantity is going
to be carried for the entire cycle. Now, the moment we build this backorder, we are
actually going to carry lesser quantity in that cycle. So, there is a tradeoff between the
373
inventory holding cost and this backorder cost, the inventory is not held for the entire
cycle, the inventory is held for a smaller period.
The other, advantages as soon as Q, the economic quantity is received we instantly fill
the backorder. Therefore, the maximum inventory that we hold will also come down. So,
a lesser quantity of inventory is held for a lesser period so, inventory cost there will come
down but, then the backorder cost is going to go up. So, if we find out that the inventory
cost, backorder cost tradeoff is a gain, particularly when because of the backordering we
are able to increase the length of the cycle. And therefore, decrease the number of orders
per year and therefore, the order cost also comes down.
So, there is a gain in the order cost, there is a gain in the inventory holding cost, there is a
loss in the backorder cost. So, essentially again the same 4898.98. We are now, dividing
it adding another component, which is going to increase the cost a little bit but, we are
going to gain from the order cost and from the carrying cost. And then we are able to
show that by introducing this s, we are actually gaining a little bit. And therefore, this
cost comes down. So, that is the explanation why this cost is even less than 4898.98
(Refer Slide Time: 01:51) whereas, this is more than 5200. But, the more important thing
to note is that, here we do not look at case where the demand exceeds D.
The equivalent of D here (Refer Slide Time: 00:19) is 200 per week whereas, this model
is centered around the possibility, that D is greater than 200 and it can even go up to 250
or 300. So, this is a probabilistic inventory model where demand can exceed the expected
value with certain probabilities, here the demand does not exceeds the demand is D all
the time with the probability of 1. So, this is the purely a deterministic model but, then
one needs to understand the difference between this model, as well as the other model.
Now, let us proceed further to understand a few more things (Refer Slide Time: 00:19) in
this model. Now, towards the end of the last lecture, we also introduced the service level
and we said that by ordering (Refer Slide Time: 01:51) by keeping a reorder level of 250,
which means (Refer Slide Time: 00:19) we will be able to meet the demand in every
cycle as long as the demand is less than or equal to 250, which means we will be able to
do it at a 90 percent service level. So, 0.1 plus 0.2, 0.3, 0.7 at 90 percent service level.
374
We also, in those computations found out that if R O L was 300, the total cost was 5300.
So, the organization faces the next question (Refer Slide Time: 00:19) which is, do I
actually fix it at 250? Because it is cheaper by 100 rupees or do I fix it at 300, what is the
advantage of my 300? The advantage of 300 is that I will never encounter a shortage. If
my reorder level is 300, this data shows that my lead time demand does not exceed 300.
So, I will have a 100 percent service level, if I fix my reorder level at 300 (Refer Slide
Time: 01:51) at an additional cost of rupees 100.
So, now should an organization, look at this cost of 5200 and decide on 250 or should it
look at trying to give a service level of 100 percent (Refer Slide Time: 00:19) and with
an additional cost of 100, which makes it 5300. Now, the answer comes in two ways,
(Refer Slide Time: 01:51) one is this 5200 after all was computed
375
assuming that C s is rupees 2.5 per unit backorder, which means that the organization is
pretty sure of what is the effect of backordering, it is under the assumption that this cost
is not going to go up. For example, if this cost is going to go up and its going to become
3 per unit backordered say, we could go through the computations and for all we know
we might find that 300 is the best value. Because as the backorder cost increases, you
know that your buffer stock has to increase for the same service level, buffer stock has to
increase which means that it will move from 250 to 300.
So, if the organization is very sure about the C s then (Refer Slide Time: 01:51) one
could go back and look at choosing 250, ahead of 300. Otherwise, it is safe to say that
well I know that my demand is not going to exceed 300. Therefore, I would rather go for
a R O L of 300, which gives me a safety stock of 100 with an additional cost of rupees
100 per year. So, the organization does not most of the time make a decision only based
on cost, it also makes a decision based on the expected service level. So, there are two
things on which the organization makes a decision, one is total cost and the other is
service level (Refer Slide Time: 18:55).
Alternately, if this organization says I am not happy with the 90 percent service level. I
want a 100 percent service level or I want a 95 percent service level, then this would
give us only 90 percent service level. If we want 95 percent service level, the next jump
is here, which would give us actually 100 and then we would end up choosing 300 as the
376
reorder level, incurring a cost of 5300. Now, let us look at it another way. Now,let us
assume that this organization wants a 95 percent service level, or let us even say that this
organization wants a 100 percent service level based on this data.
Now, let us assume we are looking at 95 percent service level is what this organization
wants, the most economical value is 250 which gives us 90 percent. But, then we have to
go to 300, if we want 95 percent service level. Now, with this assumption it would even
give us a 100 percent service level. Now, what is the assumption? The assumption is that
after all if we go back and see how we got this data? We said that we would possibly go
back to the last 10 ordering cycles, let us say and then based on some data we would
have said that in the last 10 ordering cycles, the lead time demand happen to be 100, 1
out of 10 times 150, 2 out of 10 times 200, 4 out of 10 times and so on. So, the
proportion of times the lead time demand was 100, 150, 200, 250, 300.
Now, that became the probabilities at which we got this. Now, when we make this
analysis and say that if we want 95 percent service level, then we moved from 250 to
300. Now, we are making an assumption that this demand is not going to exceed 300 at
all. But, then we may encounter a situation where, this demand can also exceed 300. So,
we look at that case, we have to look at another case. Now, let us assume in the last 10
cycles the demands were 100, 150, 200, 250 and 300 now, we did not have a case where
the demand was 220 or 225.
Now, can we look at a demand of 220 or 225 in the future? How do we handle that? So,
we try and address both these cases right now. So, one is to say that, if the demand for
example exceeds this, what do we do? The other is we make a slight approximation and
say that, can wehandle this 300 and so on. Now, just for the sake of explanation and
make it very simple. To make it very simple, let us assume that it is 100, 150, 200, 250,
300 all with equal probabilities of 0.2, just for the sake ofmaking it simpler let all of
them have same probability of 0.2.
377
Now, in such a case we could possibly assume that it is a uniform distribution, with
between 200 plus minus100. So, we could say that the demand could be anything
between 200 and 300 with equal probability.
Now, if we look at such a case then our diagram will look like this, say this is 200, this is
300. Now, we could keep reorder level anywhere, this is the expected value is at 250.
Now, if we want 95 percent service level,then we want our reorder level to be such that,
we could use R or we could use small r as the case may be. Such that we want this area
378
to be 95 percent of the total area so,100 percent is covered at 100. So, 95 percent will be
covered at 95 therefore, our reorder level in this casewill be 295 if we wanted a 95
percent service level.
So, that is easy to do when we have a uniform distribution between two given values like
200 plus minus 100. If we want a 90 percent service level then it will become 290 and so on.
Now, let us look at the other thing that if we actually plot this, we realize that at 200 the
value is the highest and as you move to the left, the values are lower and as you move to the
right the probabilities are also lower. So, if we try and plot it we realize that it is actually can
be modeled close to a normal distribution. Now, when we start modeling it as a normal
distribution, the only difference is that the demand can even exceed this 300.
So, now what happens if the demand exceeds 300 or it can be anything from minus
infinity to plus infinity, assuming it is normally distributed? So, if you want to look at the
cases where the demand can actually exceed 300, then we need to fit a corresponding
continuous distribution and then do the analysis. So,let us we can have a situation where
the demand can even be 400, how do we handle such a thing? Now,let us model this or
approximate this to a normal distribution and then let us see what we do in order to find
out the reorder level.
We assume, that the lead time demand is now, normally distributed with mu equal to
200(Refer Slide Time: 24:07) which is the same expected value of this distribution. So,
379
100 into 0.1 plus 150 into 0.2 plus 200 into 0.4, 250 into 0.2 plus 300 into 0.1 gives an
expected value of 200. So, we assume that the normal distribution has mu equal to 200,
for a meaningful comparison of results between you assuming a discrete distribution as
well as assuming a normal distribution we now, try and compute the standard deviation
of that.
So, the standard deviation of that would come to sigma equal to root of 3000 for this
which we now approximate it to 50. So now, we assume that the normal distribution lead
time demand follows a normal distribution with mu equal to 200 and sigma equal to 50.
This figure is slightly more than 50 it comes to 54.77 but, for the sake of computation
weassume that it is 50. Now, with this if we define a service level of 95 percent now, let
us observe that in the earlier case (Refer Slide Time: 24:07).
when we fix the reorder level at 250 here, we said that the service level is 90 percent.
Because there is a 10 percent probability or a 0.1 probability, that the demand may
exceed the reorder level of 250. Now, for a 90 percent service level with reorder level is
equal to 250, we found out that the total cost was 5200. Now, in this case let us assume
that we are interested in a service level of 95 percent. So, with the service level of 95
percent now, we have to find out the corresponding safety stock assuming the normal
distribution. (Refer Slide Time: 28:42)
380
So, here when we then when we fix the reorder level at 250, the service level was 90 and
the safety stock was 50 which is 50 more than which is a difference between the reorder
level of 250 and the expected value of lead time demand of 200. Now, in this case we
need to find out the service level, given the service level we need to find out the safety
stock. Now, to do that we go back to the standard normal table and then find out that
value of z, for which the area under the standard normal table, from the left most this is
0.95 because 0.95 comes from a 95 percent service level so, that happens at z is equal to
1.645.
So, this happens at z is equal to 1.645. Now, the safety stock is given by z into sigma or
R O L minus mu by sigma is equal to z so, mu plus z sigma will be R O L. So, z sigma is
the safety stock so, safety stock is z into sigma which is 1.645 into the assumed value of
50 which gives us a safety stock of 82.25. So, this gives us a safety stock of 82.25 and
the reorder level therefore, will be 282.25 in this case. Now, when we have this reorder
level equal to 282.25 then we can calculate the total cost now, the total cost will have
three components. Now, total cost will beD by Q into C naught plus Q by 2 plus SS into
C c plus S bar of x into D by Q into C s.
So, let me explain this expression now, total cost has three components this is the order
cost component D by Q is the number of orders into order cost. This is the carrying cost
component now, Q by 2 is the average inventory in the system which is multiplied by the
cost of holding. Now, this safety stock as mentioned earlier is carried continuously
therefore, there is no average there, the same safety stock will come here C c plus S bar
of x into D by Q into C s now, having a service level of 95 percent now, there is a 5
percent probability that in a cycle we will not be able to meet the demand.
381
So we now, we have to find out the expected shortage multiply it with D by Q which is
the number of cycles in a year and multiply with C s, C s is the cost of backordering
which is taken as rupees 2.5 per unit backordered. Now, if we do all these three, then we
will be able to find out the total costand then we will be able to compare this total cost
with the optimal with the cost of 5200 (Refer Slide Time: 17:41) that we obtained in the
earlier case.
So, there are three dimensions or three aspects that we will have to consider so, the first
one is the reorder level and the 2nd one is the safety stock, the 3rd one is the service
level. We have seen the relationship among all of them (Refer Slide Time: 26:56). Now,
here we defined a service level of 95 percent, for which we found out a safety stock of
82.25 which gave us a reorder level of 282.25. Now, once this safety stock is computed
now, from this safety stock we can find out the expected shortage and then we can
multiply to find out the shortage cost multiply it with the given shortage cost and find it.
Another way, of looking at the whole thing is, given a shortage cost or backorder cost as
the case may be given. The shortage cost now, what is the correct value of the reorder
level and the service level. So, let us do that assuming that our given shortage cost (Refer
Slide Time: 28:42) is 2.5 per unit short or 2.5 per unit backordered. Now, in order to find
out the best value of r the reorder level in this case. We need to differentiate this with
respect to the reorder level and then try and get the best value of the reorder level as in
382
this case now, if we try to do that we actually observe that this can be rewritten as D by
Q into C naught plus Q by 2 into C c.
Now, safety stock is r minus expected value of lead time demand. For example, if the
safety stock is 82.25, then reorder level became 282.25. Because expected value of lead
time demand which is mu is 200 so, 200 plus 82.25 gives this. So, safety stock can
always be written as r minus mu or r minus E of x into C c plus S bar of x into D by Q
into C s. Now, we realize that r comes in this expression now, r is hidden somewhere in
this expression. The reorder level is hidden somewhere in this expression because this
now, can again be rewritten as D by Q into C naught plus Q by 2 plus r minus mu into C
c plus.. What is s bar of x?
S bar of x is the expected shortage per cycle which is simply integral r to infinity x minus
r f x d x into D by Q into C s. Now, this term is like the expected shortage now, the
shortage occurs when the demand exceeds the reorder level. So, the extent of shortage is
x minus r multiplied by the probability of occurrence which is the given by f x by d x and
summed over which is given by the integral sign. Now, this term is equivalent to
something here (Refer Slide Time: 28:42). Now, when we did this calculation if we say
that the reorder level is 250, then we say that shortage can occur shortage quantity will
be 50 with a probability of 0.1.Therefore, the expected shortage would be 5, when we did
the calculation for R O L equal to 200. We would have said shortage can be 50 this
difference into 0.2, plus this difference 100 into 0.1.
So, that summation is actually the integration here x minus r is the extent to which the
demand exceeds the reorder level multiplied by the probability of occurrence. Therefore,
we have this expression. Now, if we want to find out the best value of r let us partially
differentiate this and set it to 0 so, we would get C c that comes from here now, here we
will have minus D C s by Q and then we also need this integration that will come. So, r
to infinity f x d x equal to 0, from which we would get this expression integral r to
infinity f x d x will be equal to Q C c by D C s(Refer Slide Time: 26:56).
Now, if we actually observe it very carefully, integral r to infinity f x d x is the

probability of occurrence of shortage, 1 minus that will be the service level.Therefore, if
we do this calculation or the numbers that we actually have. Now, (Refer Slide Time:
26:56).Q is equal to 1250 C c is 4 demand is 10000, C s is 2.5. So, when we do this
383
multiplication we would get 0.2. Now, this gives us that if we fix the shortage cost as
2.5(Refer Slide Time: 28:42) if the shortage cost is given as 2.5 the best service level
that we would look at under this assumption is 80 percent. So, this would give us a
service level of 80 percent.
Now, for a service level of 80 percent we could go back so, we could now go back to
service level is 80 percent. So, when we have service level 80 percent, then we once
again go back to this one to this. Standard normal curve to find out what is the value of z
for which the area under the standard normal curve is 0.8 and if we do that, we get z
equal to 0.84, z is equal to 0.84. Now, based on that the safety stock will now, be z sigma
which is 0.84 into 50 which is 42. So, now let us compare three situations or scenarios.
Now, the first one when we were here (Refer Slide Time: 28:42) when we said the
service level is 90 percent, assuming this discrete distribution safety stock was 50. Now,
when we moved to the normal distribution for a chosen service level of 95 percent, we
found out that the safety stock was 82.25. Now, if we want to make a meaningful
comparison (Refer Slide Time: 28:42) now, here service level was 90 percent. So, let us
now look at the other one, if service level is 90 percent. Once again we can go back to
the standard normal distribution and find out what is the value of Z now, Z is equal to
1.28 which would give us safety stock equal to 1.28 into 50 which is 64.
384
So now, you see here that the safety stock actually increases a little bit (Refer Slide
Time: 28:42) when we compare the situation that is here. If we fix the service level at 95,
safety stock is 82.25. If we now, do not fix the service level but, assume the shortage cost
and then compute the best service level, then that service level is 80 percent for the 2.5
which gives us a safety stock of 42. Now, we have to find out the total cost now, total
cost as I said has three components which is D by Q into C naught Q by 2 plus SS into C
c plus S bar of x into D by Q into C s now, let us do this calculation because we have
found out the best value of safety stock using this.
So, this would give us D by Q 10000 by an assumed value of 1250 which is 8, 8 into 300
is 2400. Q by2 is 1250 by 2 which is 625, 625 plus 42 is 667 into 4, 2668 plus we need to
find those out. Now, in all this we know D by Q we know C s but, we do not know this S
bar of x. So, this will become S bar of x into 8 into 2.5(Refer Slide Time: 17:41). Now, if
we remember carefully we have 3 terms here and these 3 terms are very similar to this
term, that we have from where we calculated 5200 for the discrete distribution. Now, this
4900 comes (Refer Slide Time: 17:41) out of this 2400 plus another 2500 out of this so,
that is 4900. Now, the balance is 168 which is equivalent of this (Refer Slide Time:
17:41) and then we are going to have one more term here which is the shortage cost term
which is equivalent of this term.
So, we really need to find out, what is the value of S bar of x corresponding to this 80
percent service level. That involves the little more detailed look at the normal
distribution, which also means that we have to find out the y and so on. So, I am just
going to give that figure here now, that s bar of x that we have when we have this normal
distribution is given by 0.112 so, this will be 2400 plus 2668 plus 0.112 into 20 and this
on multiplication gives us 0.112 into 50.
385
So, S bar of x is 0.112 that comes from the normal distribution that has to multiplied
with the sigma to get 0.112 into 50, which is roughly about 6 as the expected shortage.
(Refer Slide Time: 28:42) Whereas, in the other one we saw that the expected shortage
was roughly 5 (Refer Slide Time: 28:42) because, 50 into 0.1 was 5. Now, here the
expected shortage is 6. So, this gives us a total cost of 5180 now, the comparable cost
here is 5180 as (Refer Slide Time: 17:41) against 5200 which was the optimal when we
assumed this kind of a distribution.
But, then we also have to look at one thing we should not actually make a comparison
between these two values. Because of several assumptions one of which of course, is the
normal assumption, the assuming that the normal distribution has the same mu and sigma
roughly of this.We also know that the actual value is 54.77 which was approximated to
50 so, we have another 5 units that could come out of this.Whichwill add to the safety
stock so, plus 5 into 420 would give us about the same 5200.
But, then the way we look at this problem. Assuming the normal distribution and the
mathematics of it is very different from looking at (Refer Slide Time: 28:42) how we
solved it using the discrete distribution.The advantage of following this approach is that,
this gives us the flexibility to model any value of lead time demand, which could be
within this number (Refer Slide Time: 28:42) as well as which could be outside of the
range of these numbers.
386
Prof. G. Srinivisan
Module No - 1
Lecture No - 22
Integrated Model, ROL for Normal Distribution of LTD and Given Mean
In today’s lecture, we will develop an integrated model where, in a single integrated

model we try to find both Q and r where Q is the order quantity and r is the reorder level.
In the earlier models that, we saw in the previous lectures we had computed Q based on
the economic order quantity equation and then, r was computed based on the lead time
demand. Now, we will try and build an integrated model to see whether, there is any
slight change in the value of Q that is computed using an integrated model or is there a
slight difference in the cost computations, if we build an integrated model involving Q
and r.
Now, let us go back to the basics of inventory where, we draw the usual time versus
inventory graph, now let us say we start at a certain level; which right now we are not
defining it either as Q or as I m and then, we start consuming. Now, we are going to
387
assume that the demand D, annual demand is D the aggregate annual demand is known
but, at every instance the demand is probabilistic; it is not deterministic as in the earlier
examples. So, we start consuming this inventory and let us say we use a curved line to
indicate that, the demand is nondeterministic, so the inventory drops like this.
Now, let us assume that there is a reorder level r which we will determine this is our
reorder level r. So when the inventory reaches this r we place an order, now the lead time
is the time between placing the order and getting it and let us assume that we get the
quantity, when we get the fresh consignment the stock is here. So the stock gets updated
by a quantity Q and let us say it comes here, then once again we start consuming based
on this and again when we reach the reorder level let us say we place another order.
But then, let us assume in this cycle by the time the order comes the inventory has come
down to this position. Now once again the quantity arrives there is a backorder here
quantity arrives so, this stock is updated by a certain quantity Q. So, we can assume that,
this is the Q and this is roughly the same as this and once again the cycle continues and
so on. So, the only change that we have made is because of the probabilistic nature of the
demand, there can be some cycles where the ending inventory is positive when the stock
arrives and there can be some cycles; where the ending inventory is negative when the
stock arrives.
This is different from the earlier deterministic models where, we consistently assumed
that the stock will be 0 or the inventory will be 0 when the stock arrives or in the case
with backordering; there will be backordering every time the stock arrives. So we could
have cycles like this now we have two variables: which is which are Q and r. Q is the
quantity which means this quantity that, we are going to order which we will get after a
lead time and r which is the reorder level, which is the point at which we place the order.
388
So, we will try and get both using an integrated model. So now, we write the total cost
expression for this, so the total cost is made up of three parts: the annual demand is D
with the aggregate annual demand is known. So, the number of orders per year will be D
by Q into C naught which is the order cost, this is a familiar term D by Q C naught
represents the annual order cost that we will have. Now, there is an inventory component
there is a shortage or backorder component so, the inventory component will be given by
average inventory into C c plus average shortage into C s. C c is always defined as rupees
per unit per year and therefore, average inventory is in units.
Now, it again depends on how we are going to define C s we need to either multiply with
the D by Q or not. If C s is defined as rupees per unit per year then, we do not multiply by
D by Q but, if C s is defined as rupees per unit short or rupees per unit then, we have to
do that because expected value of the shortage. Instead of saying average shortage let us
say let us use expected ending inventory or expected average inventory plus expected
shortage into C s.
389
Now, in this case we are going to define C s as rupees per unit short so C c will be
defined as, rupees per unit per year and C s will be defined as rupees per unit short. So
the E shortage is the expected shortage in a cycle that, into rupees per unit short into D by
Q will give us the expected cost per year because, there will be D by Q cycles in a year.
Now, E of s or E of shortage is the expected shortage in a cycle D by Q is the number of
cycles per year and C s is rupees per unit short.
So, all these when multiplied would give us expected shortage cost per year. Now we
have to start defining this as well as this, now in order to do that, now let us call the
expected ending inventory as E of ending inventory, let us call it as E of ending
inventory. Now, expected ending inventory or expected beginning inventory will be
expected value of ending inventory plus Q because, the any cycle if there is an expected
value of the ending inventory for example, if there is something here then the cycle
begins with an addition of Q.
So, expected average inventory will be beginning inventory plus ending inventory
divided by 2 which will be this is the ending, this is the beginning inventory plus ending
inventory divided by 2 which will be expected value of ending inventory plus Q by 2. So,
when we substitute we should now start writing this, now we call this expected shortage
390
as S bar of x as the expected shortage. And then, the expected ending inventory will be;
what is the expected ending inventory? The expected ending inventory is the difference
between the reorder level and the lead time demand.
Because, when we reach the reorder level we place an order and when the there is a lead
time and there is a lead time demand. So, we expect that we will meet the lead time
demand using this reorder level so, r minus expected value of lead time demand, will
become the expected ending inventory, because there could be some cycles where there is
positive inventory, there could be some cycles where there is negative inventory.
That would depend on the actual lead time demand but, we are not modelling the actual
lead time demand we are modelling the expected ending inventory. Therefore, expected
ending inventory will be r minus expected value of lead time demand. So, expected
ending inventory will be r minus f of x, where, f of x represents the expected value of
lead time demand or we could use slightly a different notation we could not call f of x
here we would say r minus expected value of lead time demand.
So now, substituting in all of them we would get T C is equal to D by Q C naught plus Q

by 2 plus r minus, let us also call it f of x so, r minus f of x into i C plus S bar of x C s
391
into D by Q. Now, partially differentiating this with respect to Q and r we can get the
optimum values of Q and r which are the 2 variants. So, partially differentiating with
respect to Q so dou T C by dou Q equal to 0 would give us minus D by Q square C
naught plus i C by 2. Because, the expected value of the lead time demand does not
depend on the order quantity, so it does not have any effect. So, plus i C by 2 minus S bar
of x C s D by Q square equal to 0, so this on simplification will give us Q equal to root of
2 D C naught plus S bar of x C s divided by i C. So, this is the equation for the order
quantity in the integrated model, so this is our first equation.
Now, differentiating with respect to r would give us - this has no effect this also has no
effect, so there is a i C which is here; from here i C plus dou of S bar of x by dou r into C
s D by Q equal to 0. This is what we get because S bar of x represents the expected
shortage, so this will depend on the reorder level because, shortage happens when lead
time demand exceeds the reorder level. So, S bar of x can be defined as integral r to
infinity x minus r f x dx. Now, when we differentiate this with respect to r we would get a
minus term here, which comes from the x minus r
So differentiating and substituting this we would finally, get the expression integral r to
infinity f x dx will be equal to i C Q by D C s. The minus term comes from here from the
partial derivative we get a minus, so minus goes to the other side i C Q by D C s would
give this. So, this is our second equation which is this. And our 3rd of course, is the
expression for s bar of x, so we have three expressions which come here. This derivation
is quite similar to the earlier one, if we see very carefully it is very similar because, we
still have that same term i C Q by D C s, which we wrote in the previous lecture as Q C c
by D C s, Which represented 1 minus integral which is kind of represented area under the
curve, so in the some sense in the previous expression we got the value of r directly from
this.
This is the well known equation for economic order quantity; with the only change that s
bar of x comes explicitly into this function. Earlier we did not have this portion so we had
Q equal to 2 D C naught by i C when we separated Q and r we got Q equal to 2 D c
naught by i C. Now because, we brought r into the picture we are getting another
392
expression plus S bar of x C s. So, in some sense this derivation is very similar to the
earlier one and then, when we want to use it how do we compute the values of Q and r;
which is a purpose of this derivation. So we explain the computation of Q and r through a
numerical example, so let us go to the numerical example here.
So we try and use the same example, for the purpose of comparison: so we assume D
equal to 10,000 per year, so we are going to assume c naught equal to 300, i C equal to
rupees 4, C s is equal to rupees 2.5 per unit backordered. Then, we want to find out Q and
we also assume, that the lead time demand follows, a uniform distribution between; 150
to 250 with 200 as the expected value. In the earlier discrete case we took from 100 to
300 and we followed a discrete distribution. Now, we are going to assume that it follows
a uniform distribution between; 150 to 250 with 200 as the expected value. We had
already seen how the uniform distribution will look like; so it is like saying this is how
the lead time demand will look like. So, this is from 150 to 250 with equal probability
with the mean at 200.
So we now want to find out Q and r as well as T C. So, first in order to find (Refer Slide
Time: 10:41) Q we know 2 D C naught plus S bar of x C s by i C, we do not know the
term S bar of x. So, if we know the term s bar of x we can find out Q. Now, let us go back
393
and see what we require to find out S bar of x. To find out S bar of x we need integral of r
infinity x minus r f x d x. So if we know r we can find S bar of x because, f x is known
so, integral x minus r f x d x. So, in order to find S bar of x we need r and now you want
to find out; how do we find out r? So, in order to find out r you have once again r to
infinity f x d x is i c Q by D C s, which means you need Q. So, essentially to find Q you
need S bar of x; to find S bar of x you need r and in order to find r you need Q. So what
we do is we start with the simple assumption that S bar of x is 0 and then get Q.
Now, use this Q to get r use this r to get S bar of x come back and recompute Q and use
the kind of an iterative algorithm, which will affectively get Q and r to converge into
their ultimate optimum values. So, first begin with S bar of x equal to 0, so iteration 1 s
bar of x equal to 0; would mean Q equal to (Refer Slide Time: 10:41) 2 into D into C
naught by i c. So, Q is equal to 1224.74 we have computed the value of Q, Q equal to
1224.74. We now, use this value of Q in order to get r so the corresponding equation is
integral r to infinity so integral r to 250 in this case it is a uniform distribution; it does not
go up to infinity. It has an upper limit of 250 or to infinity f x d x, so f x is 1 by 100 so 1
by 100 d x is equal to Q C c by D C s, so 1224.74 into 4 divided by 10,000 into 2.5.
So, this 1 by 100 will come out so this is 250 minus r is equal to, so we would get 250
minus r is equal to this quantity multiplied by 100. So, this is 1224. So this on
simplification would give us point 1959. 1224.74 into, gives us 0.1959 or 0.196. Now,
this is 0.196 so 250 minus r will be into 100 which would give us 19.6 from which r is
equal to 250 minus 19.6 which would give us 230.4.
Now, that we have the value of r we can now use that r here, to get S bar of x, so S bar of
x will become integral r to infinity x minus r f x d x. So integral r is 230.4 to 250 or to
infinity x minus 230.4 by 100 d x, so this would give us x minus 230.4 the whole square
divided by 2 into 100. So this is on the upper limit it is 250 on the lower limit it is 230.4;
so the contribution will be 0 from the lower limit from the upper limit it will be 19.6 the
whole square by 200, which would give us 1.918 as the expected value of s.
394
Now, we have to use this 1.918 into this equation to go back and compute Q again we
show that somewhere here, so the new computation of Q will be root over 20,000 which
is 2 D, C naught plus s bar of x into C s which is 1.918 into 2.5 divided by 4 which would
give us 1234.5. Now we can use this 1234.5 into this equation and compute r again and
once again use this r to compute S bar of x and keep iterating till the values of Q r and S
bar of x converge.
So let us assume that Q has converged to 1234.5 let us assume that r has converged to
230 (Refer Slide Time: 17:28) and S bar of x is converged 1.9. So, finally, we can
compute the total cost from this equation, from this equation and for these chosen values
the total cost T C will be D by Q into C naught which is 10,000 by 1234.5 into 300 plus
Q by 2; 1234.5 divided by 2 plus r minus expected value of f of x. let us say r is 230.4
(Refer Slide Time: 17:28) expected value is 200, so plus 30.4 into C c 4 plus S bar of x
lets say is 1.918 which is say 1.9 into C s into D which is 1.9 into 2.5 into 10,000 divided
by 1234.5.
So, this on computation would give us this is 2430.13, which is the order cost component,
the carrying cost component will be 2590.6 plus shortage cost component will be 38.48
and this would give us a total cost of 5059.2. For the same illustration when we had
395
deterministic demand our total cost was about 4,900. Now, when there is the demand is
nondeterministic and follows a uniform distribution (Refer Slide Time: 17:28) we get
about 30 extra units as the safety stock or the buffer and then, we have an expected
shortage of about 2 units in every cycle.
So, this is the effect of having the integrated model, now integrated model has given us
values which are very close to what we had independently found out. For example, when
we use the normal E O Q model we had 1224.74; here we have 1234.5. We also know
that in practice the slight difference of about 10 units is not going to make a big impact
many times vendors may not be able to give accurate quantities such as 1234.5 etcetera, it
is also uneconomical to do that.
So, we would choose a convenient number like 1250 which we used for earlier
illustrations; so there is not much of a change here. Similarly, there was not much of a
change here 230.4 would gives us a service level of about 80 percent and so on. So, this
is also a comparable and the total cost is also comparable. So, either one could use the
integrated model or one could simply separate the order quantity computation Q and the
reorder computation r separately. And then, depending on the distribution of the lead time
demand calculate the reorder level and use it so that the overall cost is minimized.
396
Now, let us come back to one more aspect of probabilistic inventory models - something
that we had already seen and then, something that is going to have a large implication in
the way we order. In an earlier illustration, we had calculated the safety stock assuming a
normal distribution. S o, let us go back to that example, so we first said that, the annual
demand is 10,000 but, we also said that, the lead time is 1 week and we also assume that
weekly demand follows a normal distribution with mean equal to 200 and sigma equal to
25 .
Now, we worked out a certain computation and said that, if the lead time is exactly
1week and if the weekly demand follows a normal distribution like this. If we want a
service level of 95 percent what is our safety stock? This calculation we did in the earlier
lecture so, what we did was we went to the normal distribution and find out the z value at
which the area was 0.95 and the z value at which the area is 0.95 happened to be for z
equal to 1.645. And then, we said z sigma is safety stock so, 1.645 into 25 which would
give us about 40 units as the safety stock, there was 41.125 units as the safety stock.
Now, let us look at the situation to what happens when lead time is equal to 2 weeks. So
now, when lead time is equal to 2 weeks of course, in this case safety stock is 41.25, so
reorder level in the case of 1 week is reorder level is equal to expected value of lead time
397
demand plus safety stock which is 200 plus 41.25 which is 241.125 so, this is the reorder
level here. Now, when lead time is equal to 2 weeks expected value of lead time demand
is 400 because, weekly demand is has an average of 200. Now we need to calculate the
standard deviation of the demand (Refer Slide Time: 29:25) during the lead time, now
standard deviations are not additive but, variances are additive.
So, what we do is variance of the lead time for the (Refer Slide Time: 29:25) first week is
25 square, variance of lead time for the 2nd week is 25 square so, total variance is 2 into
25 square. So, total variance, variance of lead time demand will be 2 into 25 square from
which standard deviation of lead time demand will be root 2 into 25. So, if the weekly
demand is D and the lead time is L weeks then, the standard deviation of the lead time
demand is root over L into D. So, this on computation would give us 25 into root 2 which
is 35.36, so standard deviation of lead time demand is 35.36 and now once again for a 95
percent service level we would have z is equal to 1.645.
So, for a service level of 95 percent, we would have z is equal to 1.645 which means
safety stock is 1 point 645 into 35.36, this would give us 58.16 and reorder level is equal
to 458.16. Now, let us try and understand what is the effect or what has happened when
the lead time move from 1 week to 2 weeks. Now this is 241.125, this is 458.16, now the
400 comes as the contribution of the expected value of lead time demand and the safety
stock which was 41.125 here, has now increased to 58.16 here. So, as lead time increases
we end up carrying more safety stock, so it is necessary to keep the lead time small or
alternately, if we had not adjusted the safety stock and brought it up to 58 but, in the 2
week case if we had still kept the safety stock at 41.125.
398
If we had done that, then what would have happened if in the 2 week case if we had
retained at 41.125 then, the corresponding z will be 41.125 divided by 35.36 which is
1.16. And for z equal to 1.16 the area from minus infinity to z in the normal distribution
would now become 0.877, so area equal to 0.877, telling as that our service level will
only be 87.7 percent, if we had kept 41.125 and the lead time increased from 1 week to 2
weeks.
So, if there is an increase in the lead time then, there has to be an increase in the safety
stock to maintain the same service level. If we maintain the same safety stock then,
service level will come down. So that is the first lesson that we learnt. In both the cases it
happened that while the demand followed the normal distribution the lead time was still
deterministic. The lead time was deterministic with 1 week and the lead time was
deterministic in this case with 2 weeks.
399
Now, if the lead time also follows a certain distribution then what happens, now let us
call this as now demand during a week; so demand follows, let us call this as mu demand
as the mean and sigma demand as the standard deviation of the demand. And now, lead
time also follows a distribution. So, let us call mu L as the expected value of the lead time
distribution and sigma L as the standard deviation of the distribution of lead time. So, in
this particular example both demand as well as lead time are following certain
distribution.
Now, in such cases the expected value of the lead time demand; expected value of the
lead time demand is equal to expected value of demand into expected value of lead time
which is mu D into mu L. The standard deviation of lead time demand is now, given by
root over mu L sigma D square plus mu D square sigma L square mu D square sigma L
square. Now, let us understand the effect of this and when we apply it to the same
numerical illustration.
Now, we have seen a case here where, (Refer Slide Time: 31:48) demand follows a
normal distribution with mean 200 per week (Refer Slide Time: 29:25) and standard
deviation 25 per week. So, let us say mu demand is 200 per week and standard deviation
of demand is 25 per week when we looked at this case (Refer Slide Time: 31:48) mu lead
400
time is 2 weeks but, standard deviation of lead time was 0. Now, let us look at a case
where mu lead time is 2 weeks and standard deviation of lead time is also 2 weeks.
Now what happens to our expected values, now the expected value of lead time demand
is simply the product of mu D and mu L, so it is a product of 200 and 2 which is 400
which was exactly the same here (Refer Slide Time: 31:48) expected value of lead time
demand was 400. Now the only difference that happens is in the computation of the
standard deviation of the lead time demand. So, this becomes root over mu L which is 2
into sigma D square which is 25 square plus mu D square which is 200 square into sigma
L square which is 4.
So this becomes root over 625 into 2 1250 plus 200 square is 40000 160000 plus 160000.
So, this will become 160000 plus 1250 root over 401.6, so standard deviation of this
would become 401.6. So, standard deviation is given by the formula, mu L into sigma D
square plus mu D square into sigma l square. So, we do have this D square term that
keeps coming in, so this is roughly of the order of; this is roughly of the order of mu D
into sigma L.
So, our mu D is of the order of 200 sigma L is of the order of 2, so this is of the order of 2
into 200, which is 400. So the standard deviation of lead time demand becomes 401.6 in
this case. Now, what happens is your safety stock should be z sigma now for a 95 percent
service level we have calculated z equal to 1.645. So, z sigma which is the safety stock is
401.6 into 1.645 which is 660.56 units.
So now, the same scenario here (Refer Slide Time: 31:48) that we saw when the lead
time did not show variation which means sigma lead time was 0, gave us a safety stock of
58.16. The moment we have the lead time also showing variation with the sigma and
sigma equal to 2 we realize that, the safety stock has moved up to 660.56 units. So, the
safety stock itself is about three weeks of inventory is held as safety stock and we already
know that, such a safety stock is never going to be consumed off. So, so much of worth
of inventory is going to remain as safety stock at 95 percent service level.
401
So, the lesson is this, now if we have to choose between suppliers, the ideal situation is
something like this (Refer Slide Time: 29:25) where the lead time is small, the lead time
is deterministic, the lead time does not show any variation. So, once the lead time is small
all we do is simply have a very small amount of safety stock. 2nd is the lead time is
slightly (Refer Slide Time: 31:48) larger but, it is still deterministic and does not show
any variation.
Now, in this case the safety moves up a little bit from 41 to 58 the 3rd scenario is when
the lead time shows variation automatically, the safety stock shoots up. For example, if
we went ahead here and computed the case where lead time is actually 4 weeks, (Refer
Slide Time: 31:48) now, let us look at a case when lead time is 4 weeks but, then we are
not looking at any variation
So lead time is 4 weeks; now sigma lead time is 50 root over 4 into 25 which is 50, safety
stock is equal to 1.645 into 50 which is roughly 164 is 82 is roughly 82. So, if we have a
situation where (Refer Slide Time: 31:48) lead time is 4 weeks but, does not show
variation safety stock is eighty 2. But, if the lead time is 2 weeks (Refer Slide Time:
37:55) but, shows a variation of another 2 the safety stock is 401, safety stock is 660.56.
So, it is always advisable to go for a supplier, who has less variation in lead time and lead
402
time may be higher but, then the supplier is very sure about delivering it in a certain
amount of time.
So, the most critical thing is to understand the variation in lead time and the effect of such
a variation on the safety stock. So, it is always good to chose, if you have to choose
between 2 suppliers who even has the higher lead time but, definitely sticks to it, versus
another person who has variation in lead time. So, variation in lead time is not desirable it
is always good to go for a supplier who has no variation in lead time but, who can stick to
the suggested lead time. So, a variation in lead time is going to play a major part in our
supplying in our choice of suppliers and in the purchasing policy of the organizations.
We will see some more aspects of this affect of lead time particularly, in the context of
supply field and we will do that in the next lecture.
403
Prof. G. Srinivasan
Lecture - 23
Safety Stock Reduction – Delayed Product Differentiation, Substitution MOM
In the previous lecture, we looked at some safety stock models and we also studied the
impact of lead time on the safety stock. We assume that the demand during the lead time
is going to be probabilistic and we have looked at several cases, where the demand
follows a discrete distribution or the demand follows a continuous distribution. We also
looked at lead time and we looked at two cases, where the lead time is deterministic and
the lead time is probabilistic and followed a distribution. And we also saw that, this
would lead us into more safety stock than the case of deterministic inventory.
And one of the things that we said at the end is, it also has an impact on supplier
selection. Ideally, we would like to choose a supplier, who has a deterministic lead time
rather than choosing a supplier, whose lead time performance follows a distribution. The
moment the lead time is not deterministic and shows variation, we are fully aware that
the safety stock is very high. So, the first impact or the first application or use of these
concepts lies in supplier selection, where we have to be careful in choosing a supplier,
the suppliers should not show variation in lead times. We will now see a couple of more
404
areas in supply chain management, where this theory of safety stock versus probabilistic
demand has some application. So, we look at the second one, which is called
aggregation.
Now, let us look at a case, where there are five retail outlets and the demand for a
particular item, let us say is 50 and sigma of demand is equal to 10. We could also use s
to represent the standard deviation of the demand, but let us use sigma to represent the
standard deviation of the demand, which is 10 units. Now, if we have this situation and if
the lead time is say 2 weeks and these are weekly demands. And if the lead time is 2
weeks then demand during the lead time is equal to 50 into 200 and sigma during lead
time is equal to root 2 into 10, which is 14.14.
Now, each of these retail store wishes to maintain 95 percent service level and the
demand distribution is normal. Then, we need to go into the standard normal curve to
find out the z value, for which the area is 0.95 and that value is 1.645, is the value of z,
for which it is 0.95. So, the safety stock that each retail store will be having is z sigma,
which is 1.645 into 14.14, which would give us 23.26 and the five retail store will have a
total safety stock of 23.26 into 5, which is about 116.25.
Now, if we aggregate these five retails into a single aggregated unit then when you
aggregated that the, when you look at the case of aggregation and aggregate all of them
together, so mu will be 5 times 50, which is 250, sigma will be root 5 into 10, which is
405
22.36. And once again, if lead time is 2 weeks then demand during lead time will be 500,
which is the same as the demand of 100 over five retailers, but the standard deviation of
lead time demand would now be root 2 into 22.36, which will become 31.62.
And once again, if we want to maintain 95 percent service level, z will become 1.645
and safety stock will become 31.62 into 1.645, which is 52.01. So now, you see the
impact of aggregation, so when we are able to aggregate all of them and bring it together
in the same place then we realize that the safety stock is actually 52.01, instead of
116.25. The place we actually gained was in the fact that, the standard deviation of lead
time demand is 31.62 here for the aggregated, while it is 14.14 and its going to
effectively get multiplied by 5.
So, it is at root over, which gives us the saving in the safety stock, because moment we
aggregated the positives and negatives, somewhere cancel out each other at steady state
and then it is enough to have a slightly lesser or lower safety stock to take care of this.
But then there are several ways of achieving this, from a very practical point of view if
we look at five retail stores aggregated into a single one then there are some advantages
of having five retail stores in five different locations.
And having all aggregating them into a single location will have different effects, for
example, the demand can change when it is aggregated. But then one of the ways of
handling today to enable this is, through information sharing, where if these five retailers
even though they are located in five different places, they are able to share information.
And in situations, where the actual demand exceeds the reorder level, if the other one
that has positive stock can come and help the retail, which is going to run into shortage,
one can achieve the benefits of aggregation.
Of course, the time taken to transport between one place to another would also have an
impact, but assuming that it is possible to transport things quickly, one can gain the
advantage of a lesser safety stock when you aggregate. So, it becomes a tradeoff at
sometimes between the inventory holding of the excess inventory versus the cost of
transportation between one place to another. So, in several instances in supply chain
management, the tradeoff is between holding inventory and additional transportation
cost.
406
The same idea can be applied to warehouses, now instead of five retailers, one could
think of five small warehouses versus a single large warehouse and essentially the
concepts are the same. So, through information sharing, one entity can help the other in
situations, where the actual demand exceeds the reorder level and once again, tradeoff
between transportation cost and excess inventory. The second aspect that happens is
called component commonality.
Common components, now suppose we look at manufacture of an electronic item, say it

could be a computer, an assembly, largely assembly dominated, but manufacture of as
electronic item. And let us assume that, about 10 components that we are considering,
nowadays electronic items come with a lot of variety and since there are lots and lots of
variety, each variety essentially constitutes a particular product. So, let us assume that,
there are 10 different types of components, 10 different components, 10 different
components go into the final assembly.
Let us also assume that, the lead time demand shows a distribution with mean equal to
4000 and standard deviation is equal to 2000 and lead time is equal to, let us say 1 week.
Now, in order to maintain 95 percent service level then your safety stock is z sigma,
which is 1.645 into 2000, which becomes 3290 and for 10 different components, the total
will become 32900 items. Now, if we combine an aggregate or have n common
407
components as it is called then we would get mu equal to 40000, sigma will be equal to
root 10 into 2000, so this will become 6324.55.
And once again, for lead time equal to 1 week and 95 percent service level, z equal to
1.645 and z sigma is equal to 6324.5 into 1.645 is 10403.8. So, once again we observe
that, the safety inventory that we have in the case of common components is much lesser
than the safety inventory that we have in the case of different components. So,
organizations now try to have a common component in several of their assemblies.
And the uniqueness of the variety comes towards the end, where the unique components
which are few, will differentiate products from one another and bring variety into the
system. Now, this idea which is being increasingly used is called delayed differentiation,
where up to a very large time in the assembly, the components are very similar. And after
which the differentiation actually happens and the longer it takes to differentiate the
variety, the better it is from a safety stock point of view.
As long as we are able to aggregate and as long as we are able to have common
components that keep happening, it is always possibly to do that. Another way, by which
organizations handle this, is also to try and have what are called substitutable products.
The same product for example, can come in two different specifications and they may
meet the same purpose. So, when we have a situation, where the actual demand is for, let
us say the two different things are A and B.
408
A good example is A and B could be hard disks of different capacities, so when we have
to make A and have to make B, when we make the product that has component A and
then the product that has component B. Now, if A and B are treated separately as two
different components then we get into this part but then there can be situations, where A
and B need not be of the same cost. In fact, one of the assumptions that we made here
when we compare these two is that, the actual cost of the item is comparable.
Because, here there are 10 different components and total of 32900, 3290 of each
component is in safety, whereas here there is only one common component with 10403.
Now, if the cost of this common component is the same as these ten or is the is the same
as average of these ten then these two figures are directly comparable, otherwise the
costs are comparable, now when we look at A and B, which are the same component, but
two different specifications.
Now, by the same idea we understand very quickly that, if A and B are substituted for
each other or if A is a component with the higher specification and B is a component
with the lower specification, but A can substitute B. While we may not want B to
substitute A, but A can substitute B, we know that, when if we are able to do this then
we are able to achieve what has been told here. The safety stock will come down when
we are able to substitute which means, we are aggregating the demand for A and B.
But then we are not sure that, A and B will have the same cost, B being a component
with lesser specification, can be a little less costlier than A. But then organizations may
still like to do it, not always but sometimes, because in some sense, the additional costs
that is incurred or lost by substituting A for B will be compensated by the fact that, when
you combine both of them together, the safety stock comes down and the overall cost of
maintaining the safety stock will come down.
So, today in supply chain management, we have all these things, which are very
common. This aggregation, common components and substitution of components,
delayed differentiation and postponement which means, to try and delay the
differentiation of different varieties of the same product as much as possible so that, up
to the point where we delay, we can do all the three of them. Essentially, the principles
are the same, the principle is that, you are able to aggregate a demand and once you
409
aggregate a demand, the standard deviation comes down, so we are able to use the same
principle and get it.
This superimposed by the additional principle that lead time is not probabilistic, the
moment the lead time becomes probabilistic, all these safety stocks jump up and they
become order of magnitudes higher than what they actually are. So, if one wants to
exploit this idea and try to gain by the fact that, my safety stock is coming down if I
aggregate something or if I substitute something or if I delay the product differentiation
then I have to ensure that, I do not get into this territory.
So, the first thing that is required is excellent supplier relationship, where the supplier is
able to give items in time which means, we are looking at deterministic lead time. The
moment supplier starts showing variation and lead time is non deterministic then this
safety stock shoots up. So, if we are in this territory then even though this can give us a
slight advantage, the total cost would be very high, because in any case, safety stock will
be high.
So, the first thing that organizations have to do in order to benefit from things like
aggregation, component commonality, delayed differentiation is to ensure that, lead
times are deterministic and work with the suppliers to get deterministic lead time. So,
this is the impact of safety stock in the context of supply chain management, and in the
performance of organization, Now let us look at another topic.
410
Now, let us go back to basic inventory problems, now in the basic inventory model, we
have looked at two types of demands. We first looked at a case, where annual demand is
given, annual demand is continuous or demand is continuous and we looked at case,
where demand is time varying. We could use the economic order quantity formula to get
the order quantities here, we could use a lot sizing heuristic like Silver-Meal to try and
get the order quantities here.
But, both of them require, the first thing both of them require is the D which is the
annual demand in this case and the same D given as periods in this case. So, let us get
into the question of, how do we calculate this annual demand and once this annual
demand is given or demand is given in any form, for a particular item or a set of items,
how do we do it. And in inventory we looked at, most of the times we looked at a single
item and on some occasions, we looked at multiple items.
Whereas, in forecasting or aggregate planning, we were not looking at an item level, we

were looking more at a product level or at an aggregate level. So, how do we find out the
item wise demand from the forecasted value, which is product wise. So, there has to be a
relationship between the product and the individual items and that relationship is called
the bill of materials, also called BOM. We use the simple schematic diagram to represent
some ideas of bill of material and how we are able to move from the product wise
forecasted demand to item wise demand.
So, use a simple schematic diagram in reality; these diagrams are much bigger and far
more complex. So, let us assume that, we make a product P and for the sake of
illustration, let us say that, P is made up of two components, two assemblies A and B,
both are assemblies and then these assemblies go to make the final product P. So, let us
also assume that, A is made up of, A and B are sub assemblies, which finally make P.
Let us say A is made of C and D, where D is a bought out item and C is a machined -
something that comes from the supplier, undergoes manufacturing and then it becomes a
component C, which goes to the assembly. So, D is bought out, so it comes from the
supplier, let say C is actually made out of E, which also comes from supplier, which we
call some S 2 as another supplier. And only for the sake of illustration, B let us say is
also made out of D, but it is machined or let us say, for the sake of completion, B comes
out of E.
411
E comes from the supplier S 2 and then E is machined and then it becomes B, which
goes to the assembly. So, it is a very simple diagram, in reality if you take any
manufactured finished product, this bill of material would be a much bigger network,
which could sometimes involve thousands of components and assemblies and sub
assemblies. Just for the sake of understanding the concept and for the sake illustration, I
have used a very simple figure here.
Now, let us assume that, this month we want to make 400 of P, this 400 could either be a
forecast or it could be a demand, which we have to make, a firm demand that we have to
make. Now, let us assume again for the sake of discussion, that it requires 1 of A and it
requires 2 of B to make 1 of P, so automatically, 400 of P would involve 400 of A. So, it
would involve 400 of A and it would involve 800 of B. Now, once again let us assume
that, this would require 2 of C and 1 of D which means, to make 400, this would mean
800 of C and 400 of D.
Now, once again we would say that, in order to make 1 of C, I need 2 of E, so I need
1600 of E and here for example, I say to make 1 of B, I need 1 of E, so I need 800 of E.
So, at the end, I would go back and say that, if I want 400 of P, I need to buy 1600 of E
from the supplier, I need 400 of D from the supplier and I need 800 of E from the
supplier. Now, let us also try and bring some lead time into the picture, now there are
several lead times or several times taken.
So, let us assume that, it takes about 1 week to - for example, if I want P today, the
assembly takes 1 week to do this and if I want to make A, let us say the assembly takes 2
weeks and to make B, the assembly takes 1 week. We will later convert these weeks into
days, but right now let us look at weeks and then D comes directly from the supplier and
within 2 weeks, it will become A. Now, let us say E comes from the supplier and let us
say, it takes 1 week to become C. So now, let us go back and represent the same thing
pictorially in a different manner.
412
So, let us assume that, we have something like this and let us say we are talking of
weeks. So, here I want to make 400 of P, so if I want to make 400 of P then I need 400
of A here, here I need 400 of A, because it is going to take 1 week, here I need 800 of B.
Now, if I need 400 of A here then 2 weeks before that, I need 800 of C and 400 of D and
if I need 800 of C then 1 week before, I need 1600 of E, so here I need 1600 of E. Now,
if I need 800 of B here, 1 week before, I need 800 of E, so I need 800 of E.
If I need 800 of B then 1 week before I need 800 of B, 1 week before, so I need 800 of E
somewhere here, so here I need 800 of E. So now, if I need 400 of P at the end of this month
then I need, now what are the things I have to go to the vendor and buy. So, things that I go
to the vendor are E and D at the end of it. So, I need to have 800 of E, I need to have 400 of
D and I need to have 1600 of E, these are the three things that I need here.
So now, this has been translated to this and then I know that, if need 400 of P here, this is
what I need. Now, let us go back and say that, we are now changing this 400 of P at the
end of the month, let us say and we change that to, we say that, this month is going to
have 4 weeks. And then we say that, we are going to have 200 100 50 and 50 as the
demand for 4 weeks. So now, let us draw this diagram again considering that, let us say I
need 200 P here, if I need 400 P here, I need 800 E.
So, if I need 200 P here, I need 400 E here, I need 200 D here and I need 800 E here,
now if need 100 P here, this is the 4 th week of the month, this is the 3 rd week of the
413
month. So, if I need 100 P here, then I need, for this 100 P, I need 200 E here, I need 400
D here and I need 100 D here and 400 E here. Now, if go back and say that, I need
another 50 P here, now this 50 P would mean, 100 E here, 50 D here and 200 E here.
Now, another 50 P here would mean that, we need 100 E here, 50 D here and 200 E here,
so now, the moment we shift this demand into weeks, we are going to have this as the
requirement for the bought out items. So now, I can summarize it and say that, the same
thing I am representing here. So, I am saying, if this is the plan for P in 1 month then I
summarize here and say, at this point I need 400 E, at this point I need 200 E, at this
point I need 900 E, at this point I need 500 E, at this point I need 200 E, at this point I
need 200 E.
I go back and say, somewhere here I need 200 D, I need 100 D, I need 50 D and I need
50 D, so I can sum it up and say, this is my requirement for this. So now, all I need to do
is, the moment I get this structure right, I need to keep inputting values here, which are
my expected production quantities of P, let us say in every week. Now, the week or the
time period can be changed, the algorithm will remain the same, now weeks can become
days.
So, if this for next 4 days then this whole thing will become day and the whole thing will
shift here and say, one day before I need so much, another day before I need so much,
another day before I need so much and so on. So, depending on how we define the time
period, which is either a week or a day, the algorithm will remain the same and the way
we compute the quantities will also remain the same. So then all we need is, if we have
something like this and on a time scale we are able to say, so much of P required it in all
these times.
We can back calculate and say, this much of D and E we require here but then in reality,
P is not dependent only on D and E, but P would be dependent on some 100 other items
or 500 other items, but it does not matter, but the algorithm and the way to compute is
the same and needless to say that, this can be done extremely on a computer. So, when
for the first time, manufacturing organizations started computerizing this whole thing,
bill of material and then translating the, this is the bill of material in a pictorial form, you
could write the bill of material also on a file.
414
Now, it is it is easy to write a program that will translate this into this, now the program
that started this thinking process, which is to translate this to this is called the materials
requirement planning or it is called the MRP. So, the essential function of the MRP is to
incorporate the bill of material and it is ability to look at quantities of individual demands
when the aggregate demands are fed into the MRP system. Now, the additional thing that
can be done is, this can be extended for the next, assuming that the period is a week.
This can be extended for the next month or next 2 months, it can either work on a rolling
calendar or rolling time frame or one can do right at the beginning of the year, what is
the demand for the 50 weeks or 52 weeks and then bought it out. Now, if we are working
only with P, continuing on this numerical illustration, if we are working only with P,
which depends on D and E then we have one calendar for E and another calendar for D.
So, once we have these two calendars, we can separately use a lot sizing model here like
a Silver-Meal heuristic or part period balancing, to try and find out what quantities we
have to order here at each point. Not necessary that we have to order 200 200 500 900
200 and 400 because then the supplier lead time comes into the picture that is, 1. So, we
need to order before the supplier lead time, giving enough time for the supplier to deliver
plus bunching would also save, because every one order would make us incur an order
cost.
So, if it is cheaper for example, to order 400 here, instead of a 200 and 200, because we
will save one order cost incur an extra inventory carrying, which is the general principle
of lot sizing heuristic such as Silver-Meal. So, one could apply heuristics like Silver-
Meal to try and find out, exactly at what point, how much we order, so that can be done
for each of these items. Now, the next thing that we look at is, it is not that the
organization produces only P, so it could producing another product say, Q, which is
another product.
415
And then this Q would also depend on for example, let us say this would depend on two
sub assemblies R and S. And just to make the illustration simpler, let us say R would
depend on E and S would depend on D, so one would say that, if you want make 200 of
Q or if you want to make... Before we do that, if you want to make 1 of Q, let us say we
need 1 of R and say 2 of S and then we would say 1 of R would want 1 of E and each
would involve 2 of D.
So, if we want make 200 of Q, now 200 of Q will become 200 of R and 400 of S and this
would mean, 200 of E and 800 of D. So, once again if we go back and write these lead
times as we did earlier then we can draw another diagram like this for Q, as we did for P.
And once we are able to draw another diagram for Q, now from that diagram, we can
draw separate time calendars for D and E, which would depend on the amount of Q that
we want to produce, the total as well as the split up of that week wise.
So, if the organization is making 10 different products then we can get 10 different
calendars for E and D, not only for E and D, every item that is bought out. Now, we can
merge these 10 calendars of D and E to make only one calendar for each of these items
and then apply the lot sizing heuristic so that, you are able to order these effective. The
other alternative in the pre MRP days was, since this was time consuming activity even
when done manually, it was always easy to go back and do a single multiplication.
416
And say, totally in the year this much P I am producing, totally in the year this much Q I
expect to produce. And I know that, for every P I need this much of E and this much of
D, for every Q I need this much of E and this much of D. Simply multiply by the
requirement and aggregate it, we would get that single D, which is the annual demand for
that item and one could use an economic order quantity there. The only catch or
difficulty there was, if this annual P is distributed unequally over the months and
unequally over the weeks then the economic order quantity can sometimes result in
shortages.
But, the moment they are distributed unequally over the 50 weeks and then if we have a
system like this in place, which is computerized and MRP system that is computerized,
most MRP systems are and then we have a MRP system in place then the system would
provide us the kind of calendars that we could require for each of the item. And once we
are able to get this aggregated calendar for E, this was for P, similarly there could be for
Q, some other item and so on, aggregate all of them then we could place individual
orders for E.
Then, gets through the next complication, we have seen basic inventory models and lot
sizing models, but in this lecture series, we have not seen lot sizing models and discount.
Whereas, we saw discount in the regular inventory model, we did not see discount in the
lot sizing models then we could build a lot sizing model that takes care of discount. We
could build whatever analysis or whatever additional models that we saw in the normal
regular inventory with annual demand, we could bring all those dimensions into the lot
sizing.
So, things like discount can be brought in, things like multiple vendors can be brought
in, things like aggregating items to a single vendor and that can be brought in. So, when
items are aggregated to a single vendor, we could apply multiple item inventory models,
where by ordering items together, we can save on the truck cost. So, all these models can
be brought in to make effectively the entire cost of the materials management system as
less as possible.
At the same time, keeping in mind service levels, keeping in mind that we do not tie up
too much of inventory in the system, because today’s manufacturing does not require
organizations to tie up so much of inventory in the system. So, minimum amount of
417
inventory is certainly require to do and to find out that minimum and to exploit ideas
from safety inventory, that we saw ideas like substitution, postponement, delayed
differentiation, aggregation.
Use all of these to try and benefit from minimizing the total cost of regular inventory as
well as the total cost of maintaining the safety inventory. Now, before we proceed, we
take a quick recap of, what are all the things we have covered till now in this lecture
series and try and relate all of them very quickly and then move on to the next topic in
this lecture series.
So, the first topic that we saw was forecasting, which essentially helped us in getting an
estimate of the demand for the product. The kind of forecast that we saw were aggregate
forecasts for example, we were not trying to look at forecasting models that would
forecast the demand of the product tomorrow. The time series models that we saw,
which we concentrated, were looking at periods like 1 month and so on. So, forecasting
was done at a reasonably aggregate level and from forecasting, we saw models for
aggregate planning.
Now, aggregate planning would give two outputs, one is the production capacity that we
are going to have in terms of regular time and overtime. If there is some outsourced
capacity then that is going to come in, what is the production capacity that we are going
to have in a certain period and let us say, the period that we were looking at, is like a
418
month. If the other output from the aggregate plan is the work force but then let us not
dwell too much on work force changes.
Let us assume that, work force is a slightly different thing and let us look at the
production capacity that is available in each period in terms of man hours aggregate to
meet the aggregate demand of the product. So, till this aggregate planning, we did not
come to individual products. Now, at the end of aggregate planning, we looked at two
things, we looked at economic lot scheduling problem and we also looked at
disaggregation models.
In all these disaggregation in particular, we said if a certain amount of man hour is or

time is available, how do we divide this time to make various products. So, the first time
we actually came at a product level or an item level was in disaggregation. So, if a
certain amount of capacity is available in a month, how much of that is going to be used
to make product A, how much is used to make product B, C, D and E, is what we saw in
this disaggregation.
Then, after we cover disaggregation, somewhere in between we started inventory, though

it does not come sequentially in the system, inventory is somewhere here, but inventory
provides with our management of materials. So, if we want to produce item A at this
point in time, when do we order, how much do we order so that, we have that much of
the item available with us. So, here in inventory we saw couple of things, we started by
assuming that each item has a certain annual demand and then how the material is
ordered and how they are brought in so that, they can be produced.
At the same time, we also saw models when this annual demand is broken to months,
broken to weeks, broken to periods, where the demand was be different in each period.
And then later we saw the bill of material and how we are able to relate inventory to the
final product. Item to product is what we saw right now, where certain production of a
certain amount of product P is now broken down item wise for D and E, item wise as
well as time wise.
So, this actually helped us to try and relate the item to the product, so this is exactly what
we are right now and we have seen all these topics. And then we need to move to another
topic, where we know that, at the end of disaggregation, we are going to make or
produce a certain product or an item in the next say, 1 week or so or the next 2 days or so
419
whatever is the cycle time that we are talking out, we are going to make this product. We
also assume that, through a good materials management system or inventory planning
system, the material to make, these are also available.
Now, we need to carry this material and time into individual machines or resources and
then try and say, how much of time each machine is going to spend on each of these
products or in general, each of these jobs. Now, that leads us to a topic called sequencing
and scheduling, where given a set of machines and machine resources and the set of jobs
and activities to be performed and these products have to be made. Now, how are we
going to utilize these machines to make these jobs or how are we going to ensure that,
these jobs and all the individual requirement of these jobs are going to be met within a
certain time on a given set of resources, which are the machines. So, we would go into
detail on sequencing and scheduling in the next lecture.
420
Prof. G. Srinivasan
Lecture - 24
Sequencing and Scheduling - Assumptions, Objectives and Shop Settings
In this lecture, we consider the Sequencing and Scheduling problem. The sequencing and
scheduling problem talks about, a set of jobs or a set of tasks, that have to be performed
on given machines or a given set of resources or facilities, in it is simplest term the
problem deals with performing a certain given jobs on a set of given machines. The jobs
have to be carried out on the machines, such that certain objectives are met, certain
constraints are satisfied. Now, the scheduling and sequencing is to try and meet certain
objectives, we will get into detail about these objectives, but simple objectives would be
like to try and finish the jobs before they have to be delivered.
Now, there is a difference between scheduling and sequencing. Sequencing essentially

refers to the order, in which a sequence represents an order or a list or an order, in which
the jobs are sent. The schedule talks about the time table or gets into detail saying in
between this time and this time, this particular job is getting processed on a particular
machine. So, we need to understand the essential difference between the two, while
421
sequence represents the order, in which things have to be done, schedule represents the
timing or time table of the activities.
In certain sequencing and scheduling problems, we will use the word sequencing and in
certain other problems, we will use the word scheduling, when the time table
automatically follows the order, in which the jobs are sent then we use the term
sequence, where sequence also includes the schedule. In situations where the time table
has to be explicitly stated, we use the word schedule, as we move along, we will see
places where, we use the word sequence and places where, we use the word schedule.
Now, sequencing and scheduling problems in the context of what we are looking at is
essentially about performing a set of jobs on a set of machines, now there are certain
assumptions in sequencing and scheduling.
Students: ((Refer Time: 03:17))
So, we start with the assumptions and these assumptions are in a given set of machines
are all available, machines are available continuously all the time, there is no machine
interruption or machine down time or anything like that, machines are available as they
are. The jobs or tasks that have to be performed, the jobs that are available all the jobs
that or tasks that have to be performed are available at the start of the scheduling period
and are continuously available.
So, at time t equal to 0, we assume that all the jobs that have to be processed are
available, which also means at this point in time, we are making an assumption that the
jobs do not come late. There could be situations where, we might start the sequencing
and scheduling process with the available jobs and as we move along some other jobs
may arrive and then they may join the system, at present we are not looking at that
assumption.
The assumption is that all the jobs are available at time t equal to 0, once a job is started
on a machine, we are going to assume that the job will be completed and only then
another job will be taken up for processing, on a particular machine. So, we could call
that as, no job splitting or the first assumption of course is no job splitting, which means
that, if a particular job or a task requires say 10 minutes on a particular machine. We do
422
not and there are 2 machines that are capable of doing, it we do not split it into first 6 on
1 machine and the next 4 on the other.
What we assume here is once we pick a machine to do that job or a task, the entire task is
completed on that machine, so that is an assumption, then we also have no job
interruption. Now this assumption means that, if I have taken up a particular job on a
machine and let us say it requires 10 minutes, once I take it up, I complete this job and
only then I pick up another job for processing, I am going to assume that in between if
there is a more important job that has come.
I will not stop processing on this keep it out take the more important one finish it and
then get back to this, that assumption we that kind of a thing, we will not do and that is
an assumption, which means once taken up a job will be completed and only then the
next job will be taken up for processing, so these are some of the simple assumptions.
Next one is the processing times are known and they are deterministic, we are not
assuming that, processing times follow any distribution or anything like that, processing
times are known and deterministic.
Setup times, we also are aware that, if we take up a job for processing on a machine, then
we need to setup the machine to process a particular job. So, we are going to assume that
setup times are small and setup times are included in the processing times etcetera, so
there is also an assumption about the setup times. So, these are some simple assumptions,
which lead to very basic simple scheduling problems, as we relax some of these
assumptions scheduling problems become more involved and little more complicated.
And in this lecture series most of the times, we are going to make all these assumptions
and then we are going to proceed to develop some algorithms, which will solve
sequencing and scheduling problems. Now having seen some of the assumptions, we will
now see some of the objectives that are there in sequencing and scheduling.
423
Now the first objective and the most popular objective that, we will be looking at is
called Makespan in simple terms Makespan represents the span of time, that is required
to make a set of jobs or to process or manufacture a set of jobs. So, if a certain set of jobs
have to be done, on a certain set of machines, now there is a point t, let us say a point at
which we start, which we call as t equal to 0 and then at some point all the jobs are
completed.
Now that point from t equal to 0 to that point is the amount of time that, we have
consumed, to do or perform all the tasks or to do all the jobs. Now that is called the
Makespan, that is the amount of time that we have taken from t equal to 0, to complete
all the jobs, now that Makespan we want to keep it as small as possible. So, that all the
jobs are completed at the earliest, so the first objective is to do is to is to look at the
Makespan and then try to minimize the Makespan, we want to have the Makespan as
small as possible.
The second objective is called the flow time, now if we have n jobs to be done, on a set
of machines, now all the jobs are available at time equal to 0, they may start processing
at time equal to 0 or they may start processing later. But, each and every one of these
jobs gets completed at a certain point, the completion indicates, the time at which the last
operation or the last the machine that the job has to visit last has been visited and the
operation has been completed on that machine.
424
So, if we call completion time for the j-th job as C j, the time at which the j-th job is
completed that is called C j, then automatically the Makespan is the maximum of this C j
values. If we have ten jobs, each job has a completion time C 1 C 2 up to C 10, it is not
necessary that C 2 has to be bigger than C 1 and so on. Now, at some time all of them are
over, so the maximum of C 1 to C 10 or C 1 to C n in general indicates the Makespan
and we want to minimize the Makespan, which is essentially to minimize the maximum
of C j, j equal to 1 to n.
There are n jobs each is going to have a completion time, so C j is the completion of the
j-th job, the maximum of the completion times is the Makespan and we want to minimize
the maximum of C j. Now, each job is completed at time C j, so the completion time is
called C j, let me write flow time, as well as completion time, completion time is C j,
now here what we want to do is all of them start at time equal to 0. So, in some sense, we
want all the jobs to be completed individually, but then we want the completion times of
each one of the job to be as small as possible.
So, instead of minimizing the maximum of the C j’s, here we end up minimizing the total
of the C j minimize sigma C j j equal to 1 to n, so the completion time of job j is C j.
Now, we want to minimize the sum of all the completion times, which is an indication
that each job in some sense has spent as less a time as possible in the system, at the same
time has got all the tasks completed and is ready to leave. Now, flow time as such by the
very name it suggests the amount of time the job flows in the system.
So, if all the jobs are available at time equal to 0, then the completion time represents the
flow time, if all of them are available at time equal to 0 and for example, a particular job
finishes in 10 hours, then 10 hours is the time, at which it is there in the system. So, its
completion time is 10 hours, it is flow time is also 10 hours, but on the other hand, if we
relax one of the assumptions, that all jobs are available at time equal to 0 or the job is
available at time equal to 0, but the job is taken up for processing first say at time equal
to 2.
And then it finishes at time equal to 10, the completion time is still 10, but the flow time
is 8, because it started at 2 and finished at 10. So, we would like to in this particular
objective, we could define it as some of the completion times that, we try to minimize, so
we minimize the mean completion time or we minimize the total completion time, the
425
mean completion time or average completion time is the total completion time divided
by n, which is the number of jobs.
There is another reason to look at this particular objective, if we start looking at

inventories in the system, inventories in any manufacturing system can be categorized
into 3 inventories. One is called the raw material inventory, the other is called work in
progress or work in process inventory and then the finished goods inventory, now when
we talk of sequencing and scheduling problems, we are talking of a manufacturing
facility or a plant or a shop where jobs come get processed and they leave the system.
So, we are talking about work in process inventory, when we look at sequencing and
scheduling problems, so the time at which each job spends within the manufacturing
facility or within the shop floor as it is called represents the amount of inventory that the
shop floor is having. So, by minimizing the sum of completion times, we essentially try
to minimize the work in process inventory inside the shop floor.
So, this is an objective, which primarily aims at minimizing the inventory in the system,
inventory as in work in process inventory in the manufacturing shop floor. Then we look
at a third objective, where we start looking at the jobs from a different angle. Now these
jobs and tasks are carried out and they are carried out, they are finished and then they
ultimately go to a customer, these jobs are there, because a customer has ordered, these
jobs.
So, let us say that each job j, when the job is taken up for processing, there is a time,
which is called the due date, which indicates the time, at which this job is expected to be
finished and delivered. So, we would say that, there is a due date associated with each
job j, which is called D j, now this due date can be internally created or this due date can
be externally created. Now, when a particular task is taken up for manufacture and
supply to a customer, there is a due date, which, at which time this is to be delivered to
the customer.
Now from manufacturing, it might take some more time for the finished goods to move
and ultimately reach the customer, so from the customer due date, we could always
derive the manufacturing due date or the due date or time, at which we expect this job to
be completed by this manufacturing facility. So, in all our discussion due date would
426
simply mean, the time at which this job is expected to be finished by the manufacturing
facility.
So, each job has it is due date, now obviously, we want to meet the due dates, because
not meeting the due date is going to create a lot of trouble for manufacturing, any delay
the customer is not willing to wait, we have already seen that the customer is extremely
demanding. And therefore, it is the job of manufacturing to try and finish all their
activities within time, so the primary objective is to avoid delays. And then if we start
looking at delay can be defined as, if the completion time C j exceeds the due date D j
then there is a delay.
Now what we do is we define a very generic term called lateness, where lateness is how
late the job is, so lateness is defined as C j minus D j, C j is the time, at which the job is
completed D j is the time, at which it is due. So, C j minus D j is called lateness and it is
used general term L j is used to represent lateness, now by the very definition of L j, L j
can be positive or L j can be negative. Now L j becomes positive, when C j is greater
than D j, which means there is actually a delay, now if a job is completed before the
time, it is due, then C j minus D j becomes negative and it is early.
So, lateness has 2 components, which is called earliness and tardiness, earliness is when
the job is completed ahead or before the due date, tardiness is when the job is completed
after the due date. So, tardiness would mean that C j minus D j is positive, to begin with
we would say that, we are if we are able to complete the job early, but then we do not
want the job to be tardy.
So, there is a tardiness T j associated with jobs that exceed the due dates, so using the
definition of lateness, tardiness T j is max of 0 comma C j minus D j, now this is not
difficult to follow, because if C j minus D j is negative and the job is early, then the
maximum value is 0, therefore, the tardiness is 0. If C j minus D j is positive, the job is
late and tardy then this will take a positive value and therefore, T j will be equal to that,
so each job, now has a T j.
Now this T j can be either 0 or this T j can be positive. Now, we are concerned about this
T j being positive, and that positive T j represents the delay or the extent, to which the
job goes behind the schedule. So, now, the objective would be to try and minimize, the
total the sum of the individual tardiness, so the 3rd objective is to minimize sigma T j,
427
which is sum of the tardiness, so minimize T j and note that this T j can only be 0 or it
can be a positive quantity, T j cannot be a negative quantity by the very definition here.
So, the third objective that we can think of in sequencing and scheduling is to try and
minimize the sum of the tardiness of the individual jobs, which is commonly mentioned
as minimizing total tardiness. Then we move to a 4th objective, now let us assume that,
we can some jobs are tardy, we are unable to complete them in time some jobs are tardy,
let us look at two situations.
One situation where, there are 3 jobs that are tardy and the sum of the tardiness are 2 3
and 3, which means they are behind by let us say 2 days, 3 days and 3 days respectively,
there is another situation, where a 2 jobs that are tardy, let us say the situation is 4 and 4,
the 2 jobs both are 4 and 4. Now, if we go by minimizing the total tardiness, then we
would calculate that the total tardiness is 8 but then 3 jobs are tardy 2 jobs are tardy.
Now, it depends on the situation, if these 3 jobs go to 3 different customers, then in some
sense 3 customers are unhappy, because the jobs are late, whereas if these goes to 2
customers then two customers are unhappy. So, essentially in addition to worrying or
being concerned about the delays, we should also look at the number of jobs, that are late
or that are tardy. So, number of jobs that are tardy also becomes a meaningful objective
and that at time, which depends on the number of customers, who are getting affected,
because of the tardy jobs that are there in the scheduling system.
428
So, the fourth objective will be to minimize the number of tardy jobs, so N j represents
the N j equal to 1, if T j greater than 0, which means this job is tardy, if it has a positive
tardiness. And then we try and minimize N j, the number of tardy jobs or number of jobs
that go behind the schedule, going back to the same example, sometimes we might even
look at a situation like this.
There are 2 jobs that are tardy, but the tardiness is 4 and 5 in terms of total tardiness, this
is preferred, but in terms of minimizing the number of tardy jobs, this would be
429
preferred. So, it depends on the situation, it depends on the number of different
customers to whom it is going, now let us look at another situation like this. Now, I have
one situation with 2 and 6 and I have another situation, where there are 2 tardy jobs with
3 and 5.
Once again based on total tardiness is 8, we could choose any one of them based on
minimizing the number of tardy jobs, this has 2 jobs this has 2 jobs, so let us say, we
would prefer these 2 schedules to this, which has 3 jobs, which are tardy. Now, once
again between the 2, if you see very carefully, here is the situation, where one job is late
or tardy by 2 days, another job is late and tardy by 6 days, now this is a situation, where
it is 3 and 5.
In fact, let us make it 4 and 4, which also adds up to 8, now based on minimizing number
of tardy jobs, we could take this or we could take this, but then which one would we
prefer. In a way, we would prefer, this because there are two customers, who are going to
be delayed again two customers, let us say who are going to be delayed assuming that,
these jobs go to different customers, but then this person is going to be extremely happy,
because the extent of delay is very, very high. So, in addition to minimizing the total
tardiness and minimizing the number of tardy jobs, we should also try and reduce this
number to the extent plus, so the objective also moves to trying to find out.
430
So, we go to the fifth objective, which actually tries to look at the maximum tardiness,
which is max over j T j and then we try to minimize max over j T j. So, minimizing the
maximum tardiness also becomes another objective. Then we look at the sixth objective,
write from 3 4 and 5, we assumed that, it was to complete a job early. So, we complete
the job early, let it go to the finished goods inventor area, let it stay there wait till the due
date is reached and let it be shipped to the customer depending on the due date.
Today manufacturing organisations, more and more tend to follow, just in time
manufacture, so if a job is completed ahead of schedule or early you cannot immediately
ship, it to the customer or send it to the customer, because the customer does not want it.
If the customer is practising just in time the customer would say, if you are early by 2
days, do not send me, today send it to me exactly on my due date, therefore earliness is
also not a desirable thing.
So, the objective now shifts to computing, both earliness and tardiness, so here it is to
minimize E j plus T j, E j is earliness of j, which has been defined earlier, but not written
here. So, earliness of j E j will now be E j will be max of 0 D j minus C j not over j max
of D j minus C j, so you try and minimize both earliness, as well as tardiness. Now, we
also know that, for any given job j, it will have either a positive earliness or it will have a
positive tardiness or it will have 0, which is completed exactly at the due date, earliness
is 0 tardiness is also 0.
Now, earliness represents the additional cost of holding finished goods inventory, the job
is completed ahead of schedule, it is going to stay in finished goods inventory and then it
is going to be transported. So, it represents some kind of an inventory cost, T j represents
some kind of a cost associated with a late delivery and the cost associated with late
delivery, essentially have 2 important components. One is when the job is completed
later than the due date, what organisations normally try to do is to try and use different
modes of transport.
So, that the transportation time can be reduced, so many times instead of sending
something by road people may end up sending it by air, so there is an increased
transportation cost, because of this tardiness T j. The second and the most important cost
is the loss of customer goodwill, which gets reflected in the customers business being
431
affected and then at some point eventually, the customer is so unhappy, that the customer
decides to leave.
So, loss of customer goodwill is a very large component of the cost associated with
tardiness and the difficult part is that, the loss of the cost associated with loss of customer
goodwill is not easy to measure and compute. So, T j is right now, T j represents time as
far as our objective function is concerned, but there is a cost associated with that T j, now
when we try and add E j and T j, when E j and T j are actually time measures, it is to add
them as they are.
But, if we are going to add the cost associated with them, we will quickly understand that
though both earliness and tardiness are not desirable the cost associated with earliness is
a little cheaper than, the cost associated with tardiness. So, you do not want to add both
of them instead, you want to do a weighted objective function, which is to minimize over
j something like an alpha E j plus the beta T j, where alpha and beta are the weights.
But, beta is bigger than alpha, because the difficulties associated with tardiness are far
more than the difficulties associated with earliness, but also the moment we start doing
this alpha and beta then we get into issues of however, we are going to define the weights
alpha and beta. Then we move to the seventh objective, which also deals with earliness
and tardiness, but then looks at it slightly differently, now let us look at this, let us first
look at this definition of minimize E j plus T j, under the assumption that, both earliness
and tardiness are equally undesirable.
So, both alpha and beta the weights become 1, so you minimize earliness plus tardiness,
now let us go back and go to these definitions, so what we do here is minimize summed
over j, E j is max of 0 comma D j minus C j and tardiness is max of 0 comma C j minus
D j. So, essentially either earliness or tardiness is nothing but the absolute value of C j
minus D j, if it is 0, it is not going to contribute, if the term inside is positive, which
means completion time is after the due date, then it is tardiness the absolute value will be
positive.
If C j minus D j is negative, which means it is early the absolute value would still be
positive, so this adequately represents this, so this is nothing but if there is a due date D j
associated with job j, then the objective is to try and minimize the sum of the absolute
value or absolute difference between C j and D j. Now, if all due dates are equal, if all D
432
j equal to d, which means all the jobs have the same due date, which means these are all
jobs that have to be sent by today evening or by a Friday evening say then with this
becomes minimize C j minus d. If we make a further assumption here that, we want to
give larger penalties for larger deviations then this can be approximated to minimize,
now when we square the differences, we give a larger penalty for a larger deviation.
For example, If due date is equal to 10 or if D equal to 10 D equal to 10, if we have two
completion times, say which is 8 and 8 then some of the absolute differences would still
be 10 minus 8 plus 10 minus 8, which is 4. If we have 2 difference, 2 due dates, which is
9 and 7, some of the absolute differences would be 1 plus 3, which is 4, but if you do the
square, then this will become 2 square plus 2 square, which is 4, which is 8, here it will
become 1 square plus 3 square, which is 10. So, we would prefer this to this, because this
gives larger penalty for larger deviation, so if we want to give larger penalty, for larger
deviation instead of using the absolute value, you could use the square.
433
So, this is called minimizing mean square deviation of completion times, minimize mean
square deviation of completion times, about a given common due date. So, right now we
have seen about 7 different objectives, there are other objectives, sometimes the
organizations would also like to maximize the utilization of machines. So, maximize
machine utilization is another objective that we could think of, now if we look at these 7
objectives that, we have right now seen.
434
Now you realize that the first 2, do not involve due dates and the rest of them involve,
due dates in some form, so the remaining five, 3 to 7 are called external objectives or
external measures, where the role of the customer comes in the form of a due date. The
first two are internal measures, where the organization within itself is trying to optimize
a few things like, early completion less inventory and so on, now these are more
important from a customer point of view ok. Now, having seen the objectives and the
assumptions, we will also try and spend some time on the type of scheduling models and
scheduling and sequencing models that we are going to see. So, we would see three types
of scheduling and sequencing problems in this lecture series.
One is called single machine scheduling, we have flow shop scheduling and we will see
job shop scheduling. Actually we will be seeing single machine sequencing problems,
we will be actually seeing flow shop sequencing problems and we will see job shop
scheduling problems. Earlier, I had mentioned that, when the order alone is enough to
represent or describe the entire thing, then we uses the word sequence, but then if the
order and the timetable have to be mentioned then we use the word schedule. Now, we
will also try and understand the difference between each of them, what is the single
machine problem, what is a job shop problem, what is the flow shop problem.
435
So, single machine problem as the name suggests has a single machine, which you could
call as M as 1 machine and there are several set of jobs, which could be J 1 J 2 J n and
these jobs have to be processed on the machine. Sometimes in single machine problems,
we could have actually, 2 machine, now it is called parallel processes, so if we have
more than one machine in a single machine problem, it may again repeat, if we could
have more than one machine in what are called single machine systems.
Then those problems are called single machine with identical or with parallel processes,
which means it is like the simple example from a non-manufacturing context is it is like
someone going to a to a place to pay electricity bill or to book a railway ticket. So, there
could be 2 counters, so these 2 counters would be like 2 machines, so in manufacturing
for example, this could be a particular machine, which is some kind of a pressed and then
there could be another press, which is there, so you could have parallel machines.
In the parallel machine case a job can go to either of them, it can either go to, let us now
call look at it in the parallel machine case, so same machine, I could have M 1 as well as
M 2. But, then each of these jobs can go to either M 1 or M 2 and they can complete all
that is required in either M 1 or in M 2 and then the further assumption is that, in this
case, it will go to either M 1 or M 2, it will not visit both, it will be allowed to go to only
one of them, complete all the operations and leave.
436
If these 2 machines are such that, the processing time taken, for a particular job J 1
whether it goes to M 1 or whether it goes to M 2, if the processing is the same, then it is
called identical parallel process. If for example, this takes ten minutes, it can go to M 1
or it can go to M 2, it takes 10 minutes on M 1, it takes only 8 minutes on M 2, to do the
entire thing. Even then we could send it to M 1, even though it takes a little longer, due
to optimizing other objectives, but such situations are called parallel machine scheduling
with non identical processing.
If you use the word identical processing, it means the processing time is the same, on
either of the machines or both the machines and if you use the word non identical, it
means the processing times are different. So, single machine problems can be seen either
as the single resource or as a multiple resource with identical and non identical machines,
now flow shop is like this.
You have a set of jobs, which you call as J 1 to J n and there are a set of machines, which
are called M 1 to M m 1 2 3 and so on, so this is called a 3 machine flow shop, which
could represent 3 different processes in a manufacturing process. It could represent some
kind of a cutting, welding and painting, in sheet metal manufacturing or it could
represent some kind of a raw material grinding, mixing and then heating, in other process
type manufacturing and so on.
437
So, depending on the manufacturing situations, we would have all these jobs requiring a
set of machines or facilities not only that, they require these facilities in the same order.
So, whether it is job 1 or job 2 or job 8, it will first have to go to M 1 get something done
from then it goes to M 2 get something done, then it goes to M 3 and so on. So, in a flow
shop it is customary to say that the machines or facilities are already arranged in the
sequence, in which the processing is required.
So, whatever is the first machine that all the jobs visit is called M 1, whatever is the
second machine that all the jobs visit is called M 2 and all the jobs will therefore, visit M
1, first then M 2 then M 3 and then the last machine and then they go out. Now such a
system is called a flow shop, we could have sometimes, we could have parallel processes
here and so on, then it becomes flow shop with parallel processes, it is also a called a
flexible flow shop and so on. But, the basic flow shop problem assumes that, there is
only processor in the sequence to represent 1 particular machine then we move to the last
one, which is called the job shop.
So, this could be M 1 this could be M 2, this could be M 3, this could be M 4, now in a
job shop each job has a certain route, which it will take for example, say job 1 would first
come to M 1 from there, it may go to M 3 from there, it may go to M 4, it may then come
to M 2 and then it may leave. You could have job 2 that starts with M 2 from there it may
go to M 3, from there it may go to M 1 and from there it may leave you could
438
have a job 3, which starts with M 1, from there it may go to M 4, it may go to M 2, it
may go back to M 1 and it may leave.
So, like this we could have a different route for each of the jobs and when, we actually
define the routes, we did a couple of things, if you look carefully at say job J 2, J 2 starts
with M 2 goes to M 3 goes to M 1 and goes out it does not visit M 4, it is not necessary
for all the jobs to visit all the machines. Whereas, in a flow shop it is absolutely
necessary for all the jobs to visit all the machines, the order is very different for each job
here in the flow shop the order has to be the same.
If you look at J 3 very carefully, it comes into M 1 goes to M 4 goes to M 2 goes to M 1

again and goes out, it does not visit M 3, but it visits M 1 twice, now such a thing is
called a revisit. So, in a job shop a revisit is also allowed, in a flow shop revisit is usually
not allowed, so job shop is the most flexible thing is the is kind of an unconstrained kind
of a system where, but then in spite of all of that the route for each job is very clearly
defined and each job will follow, it is route.
And there is a process in time associated with each one of it, for example, J 3 on M 1 will
have a processing time J 3 on M 4 will have a given processing time and so on. So, we
have now kind of described, the three basic shops that, we are going to see as mentioned,
there are other shops, there are things like an open shop where, a job can be done on any
machine, that is available here. And then it takes a certain time to do then I also
mentioned about flexible shops, where you have parallel machines in a flow shop and so
on.
So, the and there are many other types of shops, which people have studied over, so
many years, but as far as this lecture series is concerned, we would be looking at single
machine, we would be looking at flow shop, and we would be looking at job shop. We
will also be seeing certain models in this lecture series.
439
So, we would look at single machine sequencing and we would look at models that deal
with makespan completion time, in flow shop, we will see models for minimizing
makespan and then in the job shop, we will see models that minimize almost all
objectives. We would also be looking at algorithms, which can give optimal solutions,
which means the best value of the objective function, as well as we will be doing
heuristic solutions or solutions, which are essentially non-optimal, but close to optimal,
we will see some of these algorithms in detail, in the next lecture.
440
Prof. G. Srinivasan
Lecture - 25
Single Machine Sequencing. Two Machine Flow Shop – Johnson’s Algorithm
In the lecture, we continue our discussion on Sequencing and Scheduling. In the previous
lecture, we had seen the assumptions, the objectives.
and different types of sequencing and scheduling problems. We will now address each
one of the problems in detail, and see how we can optimize or minimize the chosen
objective. So, the first problem that we will be looking at is a single machine sequencing
problem. We take a numerical example, to explain few ideas in single machine
sequencing. So, we consider situation with 4 jobs, which we call J 1, J 2, J 3, and J 4. So
let the processing times be 6 10 8 and 6 units of time, these units of time can be minutes,
can be hours and so on, so it is customary to keep them as minutes.
Now, we want to try and sequence them send them on a single machine, where the single
machine is here and we want to send them, for processing on the single machine, so that
we try and optimize or minimize a chosen objective. Now, there are 4 jobs J 1 J 2 J 3 and
J 4 and these 4 jobs can be arranged in 4 factorial ways, in general, if we have n jobs,
they can be arranged in n factorial ways. Now, depending on the objective, one or more
441
of these n factorial sequences will turn out to be optimum. Now let us look at minimizing
makespan.
Now, first let us explain the computation of completion times and then we try and look at
each of the objectives, now if this is a machine and we try and send them in the given
order say J 1 J 2 J 3 and J 4, now J 1 enters at time equal to 0. So, the completion time
for J 1 J 2 J 3 and J 4, J 1 will enter at time equal to 0 and come out, at time equal to 6, so
J 1 comes out at time equal to 6. One of the assumptions is that the machine can process
only one job at a time. So, the machine becomes available at time equal to 6 and then J 2
goes and comes out at 6 plus 10 equal to 16 time units J 2 comes out.
Now at time equal to 16 the machine is free. So, J 3 enters, so 16 plus 8, 24 is the time J
3 comes out, so at time equal to 24 the machine is again free J 4 enters takes another 6
units of time comes out at time equal to 30. So, if we consider the sequence 1 2 3 4,
which is the order, in which the jobs are going inside, now the jobs come out at time 6,
16, 24 and 30. Now, as I said there are two important assumptions one is the machine is
not idle at all and a machine can process only one job at a time.
So, we also make an assumption that the machine is not idle or the machine is not kept
idle, when there are jobs waiting in front of it. Now if we look at Makespan as the
objective Makespan is the time, at which the last of the jobs is completed. So, Makespan
for this sequence is 30, for a given sequence J 1 J 2 J 3 J 4, the sum of completion time is
6 plus 16 plus 24 plus 30, which is 22 plus 426 46 plus 30 76. So, sum of completion
times equal to 76 and mean completion time equal to 76 divided by 4, which is 19 is the
mean completion time.
So, given a sequence J 1 J 2 J 3 J 4, we can now calculate the Makespan, we can

calculate the mean completion time, let us assume that, the due dates for these 4 jobs are
10 15 22 and 30. So, if we follow the sequence J 1 J 2 J 3 and J 4, the completion times
are 6 16 24 and 30, we will assume that the due dates are 10 15 22 and 20 for these four
jobs. So, the completion times are 6 16 24 and 30, now job one comes out at time equal
to 6, it is due date or time, at which its due is 10, therefore it is early, job 2 comes out at
equal to 16, it is the due date is 15.
So, the tardiness is one here the completion time is more than the due date, so tardiness
is one here the tardiness is 2 and here the tardiness is 10, so total tardiness is equal to 13,
442
for this So, what we have right now done is for a given sequence, we have explained,
how we calculate the 3 objectives, once the completion times are calculated all the
objectives now depend on the completion times, some of them depend on the due date.
So, the given process times, you would call this as p j process time of job, we have now
given the processing time p j of job j, we have now computed here the completion time
of job j, once the completion time and due dates are known, we can calculate the
objectives. Now, your optimization problem is what is the sequence that minimizes
Makespan? What is the sequence that minimizes, mean completion time or sum of
completion times? What is the sequence that minimizes total tardiness and so on, we will
now address these objectives in a little more detail.
Now, as long as we do not keep the machine idle and as long as the machine can process
only one job at a time, whatever order we sent these jobs, the last of the completion times
is going to be the sum of the processing times, which is 30. Whether we send it in the
order 1 2 3 4 or whether we send it in the order 2 1 3 4 or any one of the 4 factorial ways
by which, we can send this job, the Makespan is going to remain the same.
In fact, the assumption that the machine will not be kept idle, when there is a job waiting
in front of it, would actually help the Makespan or alternately, if we are looking at
Makespan as the objective then does not makes sense for us to keep the machine idle,
when there are jobs waiting in front of it. So, as soon as the jobs, all the jobs are
available at time equal to 0, so the machine starts processing 1 job after another and in
whatever order, we send these jobs, the total Makespan is going to be 30, so Makespan is
equal to 30.
443
So, minimizing Makespan on a single machine is not a very difficult optimization

problem, any order any one of the n factorial sequences will give us the same Makespan,
which is equal to sum of the processing times. Then we move to the total completion
time or mean completion time, now we ask ourselves the question, if we send the jobs in
this order 1 2 3 4, the sum of completion time is 76 and the mean completion time is 19,
what is the order, in which we have to send these jobs, such that sum of completion times
is minimized.
Mean completion time is sum of completion times divided by the number of jobs, so
whether we minimize sigma C T or whether we minimize M C T, it means the same,
because you are dividing by n and in this example, we are dividing by 4. So, what we
want do is if we want to minimize this 76, now let us go and find out, how we got this
76, 76 was got by the sum of 6 plus 16 plus 24 plus 30.
Now how did we get this 6 16 24 and 30, now the moment, we chose an order 1 2 3 4, 6
is the processing time of the first job 16 became processing sum of the processing times
of these two, 24 becomes sum of the processing time of these three and 30 becomes sum
of the processing times of these four. So, essentially we want to for the order sequence,
such that sum of these numbers are minimized, now when we want to minimize the sum
of these 4 numbers, it only makes sense to minimize all the individual numbers.
444
Because the first job that we pick, which is, which has processing time 6 finds, it is place
in here, here, here and here, so whatever job, we fix as the first job, it is processing time
is going to contribute to all the 4 completion times, whatever we pick as the second job
in this example 10. Now, it is contribution is going to be from here and here, whatever
we pick as the 3rd job, it is contribution will be from here and here, whatever is the 4 job
will contribute here.
So, if we want to minimize the sum of completion times, we wish to minimize all these 4
of them and whatever, we are going to chose here is going to contribute 4 times,
whatever we are going to fix here is going to contribute 3 times and so on. Therefore it is
only correct that, we use the smallest of them first, which contributes in all the 4
completion times, then next smallest comes second, which contributes in 3 of the
completion times and so on.
Therefore, we arrange the jobs in increasing order of processing times and in general non
decreasing order of processing time and this is called shortest processing time also called
as S P T. So, arrange the jobs in the increasing or non decreasing order of processing
times to get the sequence, now the processing times are 6 10 8 and 6. So, we arrange
them in the order J 1 J 4 J 3 and J 2, now the moment, we order them in J 1 J 4 J 3 and J
2, the completion times are C 1 has a processing time of 6.
So, this is complete at time equal to 6, at time equal to 6, the machine takes up J 4 takes
another 6 units of time do finishes at 12, at 12 it takes up J 3 takes another 8 units of time
finishes at 20 and at 20, it takes up the last job, which is J 2, which has a processing time
of 10 and this will finish at time equal to 30. Now, sum of completion times is 6 plus 12
18, 18 plus 20 is 38, 38 plus 30, 68 and mean completion time is 17, in this example, the
optimum value of mean completion time is 17.
So, the shortest processing time rule minimizes sigma C T or M C T mean completion
time, this is an extremely important result, which finds it is way into other sequencing
and scheduling problems. More importantly this is introduced us to the rule, which is
called shortest processing time rule, where we try and arrange the job to form a
sequence, such that the processing times are in increasing order or non decreasing order
as in this example.
445
Now, we look at the third objective, which is minimize sigma T j, we have already
defined T j, T j is defined as the tardiness associated with the completion times and due
dates, so tardiness T j of job j, so T j is maximum over 0 coma C j minus D j. Now, C j is
the completion time or time, at which job j is completed, D j is the due date associated
with job j, now if the job completes later than the due date, then it is tardy in such case C
j minus D j will be positive.
So, maximum of 0 comma C j minus D j will be positive and tardiness will be positive, if
on the other hand, the job completes ahead of the due date, then C j minus D j will be
negative, because C j the completion time will be smaller than the due date. So, when the
job is early, which means C j is smaller than D j, then C j minus D j becomes negative on
the tardiness associated with job j becomes 0.
So, tardiness of job j is maximum of 0 comma C j minus D j and total tardiness is given
here, now the tardiness has two terms, which is the completion time term, as well as the
due date term. Since we want to minimize the total tardiness, we would like to begin
with we would like to arrange or sequence the jobs, in non decreasing or increasing due
dates and hope that such a sequence will turn out to be optimum.
So, first evaluate that, so the rule here that, we try to use, this called E D D, which is call
earliest due date, where jobs are arranged in the order of increasing or non decreasing
value of the due date. Now, let us apply the earliest due date rule to the example problem
and first compute the total tardiness.
446
So, now, based on earliest due date rule, we first sort the jobs according to the due dates,
so the dues dates given are 10 15 22 and 20, so job 1 comes first job 2 comes 2nd job 4
comes, 3rd and job 3 comes 4th. So, based on the earliest due date rule the order will be J
first 1 J 2 2nd J 4 3rd and J 3 will be the 4th. Now, the moment, we find out this order the
completion times are J 1 starts at time equal to 0 and finishes at times equal to 6, J 2
starts at time equal to 6 finishes at time equal to 16, J 4 starts at 16 and finishes at 22 and
J 3 finishes at 30 the last of the jobs always finishes at the Makespan.
Now, due dates that we have here, so the due dates are for job 1 the due date is 10, so due
date is 10 15 20 and 22, now let us find out the tardiness associated with job j, now here,
the completion time is 6, the due date is 10, therefore it is early. Due date is 15
completion time is 16, so tardiness is 1, due date is 20 completion time is 22, so tardiness
is 2 and due date is 22 completion time is 30, so tardiness is 8, so sigma T j for this
example is 11, so total tardiness is 11.
So, it is very customary to use the earliest due date rule to try and minimize the total
tardiness, for example if we had used, we showed some calculation here. So, when we
use the rule J 1 J 2 J 3 and J 4, we found that the total tardiness was 13, the earliest due
date rule gives us a total tardiness equal to 11. Now, only difficulty here is that while,
here we can say confidently that the shortest processing time will always give the
447
minimum value of sigma C T or M C T we cannot say that E D D rule is optimal to
minimize the total tardiness, so E D D is not optimum always.
Now, there are examples, through which we can show that the E D D rule is not always
optimum, though in our example, we have calculated a total tardiness of 11, which
actually turns out to be optimum for this particular illustration. But, E D D need not give
us the optimum solution always, then we get into the next question of, can we get to the
optimum solution always and if, so how do we get to the optimum solution always.
Because, the 2 earlier objectives of Makespan and total completion time, we have said
that for the Makespan any sequence is optimum, so I could possibly write here and say
any sequence or every one of the sequences is optimal, which will gives us a Makespan
of 30. As far as minimizing sum of completion times, we could say that shortest
processing time rule minimizes sum of completion times, earliest due date rule tries to
minimize the total tardiness, but it need not give us the optimum solution always.
Now, in this case, if we need a method to find out the optimum solution always then that
method would either explicitly or implicitly, it has to evaluate all the 4 factorial or in
general n factorial sequences that are possible and then only it can gives an optimum
solution. Sometimes, it evaluates explicitly, sometimes, it evaluates implicitly, because n
factorial becomes very large, when n becomes large for example, if we look at n equal to
20, the number of computations is extremely large, when we want to enumerate n
factorial.
So, we would look at methods that would implicitly evaluates a part of n factorial, but
tell us there, it has a evaluated that and then by evaluating fewer than n factorial, it could
give us the optimum solution. So, several branch and bound algorithms, try to do that by
implicitly evaluating some of the solutions, but n factorial is very large and these
sequencing and scheduling problems become hard problems.
Particularly, when the problem size increases, when we, so far we have not found an
algorithm that runs in other than n factorial explicitly or implicitly, because the
computational effort to find out n factorial becomes exponential and very large as n
increases. Therefore in these class of problems, it is customary to look at thumb rules or
what are called dispatching rules, to try and get the best solution.
448
Earliest due date is one such popular dispatching rule, though earliest due date does not
guarantee the optimum solution, in the case of the single machine tardiness
minimization. Now, we have seen three objectives with respect to single machine
problems and at the end of it two very important rules have come out, one is called the S
P T rule the shortest processing time rule, the other is called E D D rule the earliest due
date rule, now these 2 rules are also going to be used in later scheduling problems, such
as job shop scheduling.
While the S P T provides an optimum solution, to the problem of minimizing sum of

completion times, earliest due date is an extremely popular rule to address problem that
have due date related objectives or due date related measures, E D D does not guarantee
the optimum solution. Now, there are a couple of other things, which we would look at
for example, instead of minimizing total tardiness, if we try to minimize total lateness,
where lateness is simply L j is C j minis D j, if we want to do that, then we could realize
that, if we want to minimize sigma C j minus D j.
D j’s are known, so it is enough to minimize sigma C j and we already know that S P T
minimizes sigma C j, so S P T also minimizes mean lateness or total lateness. And there
are other rules, which would say that, if S P T or E D D gives only one tardy job then it
minimizes the number of tardy jobs and so on. So, there are some interesting rules and
algorithms, which are available, but then we have seen a small sub set of it and at the end
of it I would say here that, at the our discussion on single machine problems. We realize
the two important rules have come out, one is called the S P T rule, the other is called the
E D D rule or the earliest due date rule, now we spend a few minutes on minimizing in
parallel processors.
449
On parallel processors, now if we assume that, there are two machines instead of 1 and
they are identical. Now if we want to minimize the Makespan, we have 4 jobs with 6 10
8 and 6, so with 6 10 8 and 6. Here again, one can show that, the problem of minimizing
Makespan becomes hard and difficult a simple rule that is often used is L P T, which
means arrange them in the non decreasing order of non increasing order or decreasing
order of processing times, to get 10 8 6 and 6 and starting locating them to the 2
machines this way.
So, 10 goes here 8 comes here, the next 6 will come here and this 6 will come here, if
there is a 5th job, it will go here, the 6th 7th 8th and so on, now here the Makespan is
equal to 16, here the Makespan is 14. So, the actual time, at which all the jobs are over is
the maximum of these 2, which happens to be 16, so one would say that, if we apply this
rule then the Makespan that, we get is 16. So, there are other rules and algorithms, to try
and minimize other objective, such as mean completion time tardiness and so on. But,
has I have mentioned the end of discussion on single machine problems, we learnt two
important rules, one is called the shortage processing time rule and the other is called the
earliest due date rule.
450
So, now we get into the next set of problems and sequencing and scheduling, which is
called flow shop sequencing called flow shop sequencing, we explain it through a small
example. As we have already mentioned, a flow shop is a shop, where say we have n
jobs J 1 up to J n and then we have M machines, which we would call as M 1 M 2 and M
m an important condition in a flow shop is that, all the jobs will have to visit all the
machines, in the same sequence.
So, J 1 requires processing on M 1 M 2 M 3 up to M m, it also requires processing in the

same order except that, the processing times are different - the same with J 2, same with
J 3 and so on. So, we it is customary to say that the machines are already arranged, in the
order, in which all the jobs visit them, so we could say, we would call the machine,
which all the jobs visit first as M 1, a machine which all the jobs visit second as M 2 and
so on.
So, the machines are already arranged in the order of visit of the jobs, so this is a typical
flow shop problem and then we try to look at, what we do for objectives such as
Makespan mean completion time and so on. In this lecture series will concentrate only
on the Makespan for flow shop sequencing, now we explain it through a small example,
we first address, what is called the 2 machine flow shop problem, where there are n jobs
and there are only 2 machines.
451
So, we take a small example, to explain the n job 2 machine problem, now this is an
example, where there are 5 jobs, which we call J 1 J 2 J 3 J 4 and J 5 and there are 2
machines M 1 and M 2. Now, we want to find out, the order or sequence, which
minimizes the Makespan, so there are 5 jobs, so there are maximum of 5 factorial
sequences, in which we can arrange these 5 jobs. Now, we will check about this 5
factorial after a while, but then we right now assume that, there are 5 jobs and they can
be arranged in 5 factorial ways.
And we right, now assume that one of the 5 factorial ways, by which we can send the
jobs is going to be optimal, for the Makespan. So, first let us try and compute the
Makespan for a given sequence and then we try and optimize the sequence, so to make it
very easy, we try and compute the Makespan, for the sequence J 1, J 2, J 3, J 4 and J 5.
452
Now, first for a given sequence, let us find out the completion times, now these are
completion times on machines M 1 and M 2, so we first send job J 1, so job J 1 goes first.
So, machine both the machines are free and their available at time equal to 0, so J 1
requires 10 units of time on M 1, so it will start at 0 and comes out time equal to 10
remember that, these are completion times. So, it starts at 0 and comes out at time equal
to 10, so at time equal to 10, it goes to M 2, M 2 is free, at M 2 it requires another 6
units, so it comes out at time equal to 16.
Now at 10 M 1 is free and therefore, the next job goes to M 1, so right now we are
evaluating it for a sequence J 2, so J 2 comes in next, machine M 1 is available at time
equal to 10, therefore J 2 starts at time equal to 10 and finishes at time equal to 16. Now,
in this case exactly at time equal to 16, J 2 finishes at M 1, now M 2 is also free at time
equal to 16, so J 2 straight way takes M 2 at time equal to 16, requires another 12 units
of time, so at time equal to 28, J 2 finishes at M 2 and J 2 goes out of the system.
Now, once again at time equal to 16 M 1 is free. So, M 1 can takes its next job, which
happens to be J 3, because we are evaluating a sequence 1 2 3 4 5, so it starts at 16 and
finishes at time equal to 24, 16 it requires a further 8, so 16 plus 8 24 8 comes out. Now,
at time t equal to 24, J 3 comes out of M 1 and J 3 wants use M 2, but M 2 is free only at
time equal to 28, so J 3 has to wait for 4 units of time before, it could go to M 2.
453
So, right now we assume that, there is enough space and buffer between M 1 and M 2, so
that the person, who is working can take J 3 out of M 1 keep it in the buffer space and
then take up the next job. So, the next job can enter at time equal to 24 into M 1, but to
finish this computation, now we say that the J 3 can take M 2 only at time equal to 28, so
at 28, it takes on it goes to M 2, so 28 plus 9 is 37. Now J 4 can enter at 24, because at 24
it is over, the person has taken the job and kept it outside M 1 is free.
So, J 4 starts at time equal to 24 takes a further 8, here finishes at time equal to 32, now J
4 again has to wait for 5 more units, so that M 2 becomes free at 37 and then it goes to M
2 at 37 takes 10 more units finishes at 47 and then it comes out. Now, at 32 J 5 can enter
M 1, so J 5 enters M 1 at time equal to 32, once again, we assume that, there is enough
space in between for the person to take it and keep it, so that M 1 can be released for it is
next job, so it starts at 32 and finishes at 44 - 32 plus 12 is 44.
And then it has to wait for 3 more time unit to get into M 2, starts at 47 finishes at 47
plus 5, 52 and time equal to 52 all the jobs are over and therefore, for this given sequence
the Makespan is 52. Now, the question is what is the sequence or what is the order, in
which we can send this, such that the Makespan is minimized? We also realize
something interesting from this particular table. If we see very carefully 44, which is the
time, at which M 1 completes all the jobs is actually, the sum of the processing times
here 10 16 24 32 plus 12, 44.
So, M 1 is not idle, till it finishes everything then M 1 becomes idle in this case for 8
more time units or 8 more minutes, other thing that, we observe is M 2 cannot start at
time equal to 0, because of the flow shop condition. So, M 2 can now start here at 10 and
then we also realize that, there were some delays that happened here, so there was a
delay of 4 units here, there is a delay of 5 units here, there is a delays of three units here,
so 4 plus 5 9 plus 3 12.
Now, we look at this 18 27 37 plus 5, 42 plus the other 10 was 52. 18 plus 9 27 37 42 10
52, but then it is started at 10, 10 plus 42 is 52, M 2 also was not idle. Now sequencing
and scheduling problems become important. Because, this number, at which it completes
the Makespan largely depends on a couple of things. It depends on 2 very important
things here, which is called job waiting for the machine and machine waiting for the job.
454
Now, in this case, we had situations, where the job was waiting for the machine, this job
was over at time equal to 24, but then the job had to wait till 28, here the job was over at
32, it had to wait till 37. Now, you look at this case machine is waiting for a job the
machine was available at 0, but then the machine has to wait for this job it has to wait for
10 minutes. So, it is usually the delays that happen, when the job has to wait for the
machine and the machine has to wait, for the job that creates, this kind of an issue, where
the objective changes.
So, the problem that we are looking at is given a sequence, we can find out the
Makespan, what is the sequence that minimizes the Makespan? Now, a very popular
algorithm is called the Johnson’s algorithm was developed by Johnson in the year 1954,
which tries to minimize the Makespan on 2 machines. Now, Johnson not only gave a
very simple algorithm, but he actually gave a very elegant proof for the algorithm.
So, we will see the algorithm alone in this lecture series, we will not see the proof, but
then we see how intuitive Johnson’s algorithm is when, it comes to solving a 2 machine
flow shop sequencing problem. So, Johnson’s algorithm works like this, a very simple
way to do it is there are 5 jobs.
So, create a small table with 5 positions, now find out the smallest number in this table, the
smallest number is 5 and 5 happens to be for job 5 on machine M 2, so we look at this and
the rule says, if the smallest number is on M 1, then come from the left and if the
455
smallest number is on M 2 come from the right. So, smallest number, which is 5 is on M
2, so we come from the right in this table and since that happens, for J 5, now we come
from the right and write J 5 here as the last job.
So, once again look at this smallest number, the smallest number is 5 and check whether,
it is on M 1 or whether, it is on M 2 and for, which job it is, so in the case the smallest
number is on M 2 and it is for J 5. Since it is on M 2 come from the right and write that
job in the first available position from the right, so first available position from the right
is here, so J 5 goes here, now remove this job from the list.
Now, look at the rest of the table and try to find out the smallest number, the next
smallest number is 6 and it happens to be on M 2 for J 1 and also happens to be on M 1
for J 2. So, when we have a situation like this, where the smallest number is coming on
M 2 as well as M 1, we may be tempted to believe right now that there is a tie, because 6
is appearing in 2 places and 1 case, it is on M 2, other case it is on M 1.
The fact that, it is on M 2 in one case and on M 1 in the other indicates that, it is actually
not a tie and we can actually break it, arbitrarily there are no problem at all, if it happens
to be on M 2 as well as on M 1 for 2 different jobs. So, we could take either this or we
could take this, now let us take J 2 on M 1 as the smallest number, so J 2 on M 1 means,
since it is on M 1 the rule says come from the left and take the first available position
from the left. So, it is on J 2 on M 1, because it is M 1 come from the left, take the first
available position, so from the left take the first available position and put J 2.
Now, remove J 2 from the table temporarily, now once again find the smallest number
has to be is 6, it is on M 2 and it is for J 1, since it is on M 2 come from the right and
take the first available position, so from the right you take the first available position and
put J 1 here. Now, at this point we realize, what we said earlier, now we had the
minimum coming here, for J 1 on M 2 and the minimum coming here, which is J 2 on M
1, we first took this and we put J 2 here and then we ended up writing J 1 here.
If we are taken it the other way, we would have taken this first then we would have
written J 1 here and then J 2 would have come here, the sequence has not changed,
because the smallest number happened to be on M 1 for a particular job and on M 2 for
some other job. So, now, remove J 1 also, so 3 jobs have been put in the table, 3 jobs
456
have gone out, so what we have remaining are J 3 and J 4, now find the smallest number,
which happens to be 8, which is on M 1, which I for J 3 as well as, it is for J 4.
Now here there is a tie whether, we take this J 3 on M 1 or whether we take J 4 on M 1,

so as long as the smallest number happens to be on 2 different machines, there are only 2
machines, happen to be on 2 different machines, there is no tie, the sequence does not
change or the sequence is not affected. But, if there is a tie like this, when the smallest
number happens to be on M 1, now we would either take J 3 or take J 4, so first let us
take J 3.
So, since it is on M 1 come from the left and take the first available position, so we take J
3 here, now remove this J 3 there is only 1 job remaining and there is only 1 position
vacant, so put J 4 here. Now, we had a tie, we had a tie between J 3 and J 4, both of them
have a smallest number and it was on M 1, we resolve the tie by taking J 3 first, now
what would have happen, if we had taken J 4 first, if we had taken J 4 first we would
have got the sequence J 2 J 4 J 3 J 1 and J 5.
So, Johnson’s algorithm in this for this particular example gives 2 sequences, now 2
sequences happened, because of a tie occurring here and we had one sequence, which we
resolved by putting, J 3 first and then J 4 and the other sequence by putting J 4 first and
then J 3. Now, let us try and evaluate the Makespan, for both the sequences will first do
that for this sequence, so the sequence that we are taking is J 2 J 3 J 4 J 1 and J 5 and
completion times on M 1 and M 2.
So, J 2 goes first J 2 finishes starts at 0 finishes at 6, now J 2 goes to M 2 at 6 takes

another 12 units, so finishes at 18. J 3 goes next J 3 can start on M 1 at 6 needs another 8
units of time finishes at 14, but M 2 is available only at 18. So it takes another 9, 18 plus
9, 27. Now, J 4 can start at 14 on M 1 takes another 8 units, so 14 plus 8, 22. Now it has
to wait till 27, for M 2 to be free takes at M 2 at 27, takes another 10 units of time
finishes at 37.
Now J 1 can starts at 22 takes another 10 units of time finishes at 32, 22 plus 10. It has to
wait for another 5 units for 37, so takes M 2 at 37 plus 6, 43 it comes out. J 5 is the last
job, it can start at time equal to 32 takes another 12 units of time, so finishes at time
equal to 44. And now, J 5 comes out at time equal to 44 M 2 is already free at 43, but J 5
can take M 2 only at time equal to 44, takes another 5 units finishes at 49 and now all
457
jobs are over at time equal to 49, which is the optimum Makespan according to
Johnson’s algorithm.
Now here we can see both, we can a situation, where this job is coming out at 14, but the
machine is available at 18 so here is the case, where the job is waiting for the machine to
be free. Here the job comes out at 44, but the machine is already free at 43, but the
machine has to wait for the job to come, so there is a machine waiting time of, one which
again increases the Makespan.
So, the Makespan increases due to two things one is job waiting for the machine the
other is machine is waiting for the job. It is only these 2 delays that makes sequencing
and scheduling problems important and scheduling and sequencing problems difficult.
So, based on the Johnson’s algorithm, where we got the sequence, we got the Makespan
to be 49, but we also got another sequence here, so we will quickly try and compute the
Makespan for the other sequence.
So, here we are again computing cycle times, we have M 1 completion times, we have M
1 M 2, the sequence is J 2 J 4 J 3 J 1 and J 5, now for J 2, it is the same, so it is 6 and 18.
Next we have J 4, so J 4 comes here. So, J 4 takes M 1 at 6 finishes at 14 has to wait for
4 units of time, so goes to M 2 at 18 and finishes at 28. Now, J 3 can start at 14 takes
another eight units of time, so 14 plus 8, 22, waits for 6 more units goes to M 2 at 28 and
then plus another 9 finishes at 37.
Now J 1 can go to M 1 at 22 takes 10 units of time, finishes at 32 waits for 5 more units
of time to go to M 2 at 37 takes another 6 units and comes out at 43. Now J 5 goes to M
1 at time equal to 32 takes 12 more units finishes at 44. Now, M 2 is waiting, so it goes
to M 2 at 44 takes another 5 units of time and finishes at 49. So this sequence also gives
us the Makespan of 49. So Johnson’s sequences, because of tie is if we have more than
one sequence all the sequences are optimal and all of them give the same amount of
Makespan.
Now, Johnson’s algorithm gives the optimum Makespan, for a general m jobs, but 2
machine flow shop and Johnson’s algorithm is based on a very simple rule, that if you
have n jobs create a table with n slots, find out the smallest number in the table. Now, if
this smallest number is on M 1, then come from the left and put the corresponding job in
458
the first available position, if the smallest number happens to be on M 2, then come from
the right and put the corresponding job in the first available position.
Once a job is entered into the table temporarily eliminate or remove the job and once
again find the smallest number, repeat this till all jobs go into this, if there is a tie, can be
broken arbitrarily. But, ties can result in more than one sequence a tie will happen, if the
smallest number is on a particular machine say M 1 and more than 1 job has that.
Similarly, tie will happen when, it is on M 2 and more than one job has it, but if the same
smallest time happens to be on M 1 and M 2, for 2 different jobs then it is not a time.
So, get as many solutions as there are ties all of them are optimum with the same value
of Makespan, now Johnson’s algorithm is seen is the starting point of research in
sequencing and scheduling. Because, what appears to be a complicated problem with the
worst case possibility of n factorial Johnson came off with the very, very elegant
algorithm, which could solve the 2 machine case.
Now Johnson’s algorithm triggered a whole lot of research in sequencing and scheduling
and for well over 55 plus years 100 and 1000s of people have been working on
sequencing and scheduling problems. How, Johnson’s algorithm can actually be
extended, to a very special case of the 3 machine problems, now Johnson’s extension to
the three machine, we will see in the next lecture.
459
Operation and supply chain management
Prof. G. Srinivasan
Lecture - 26
Flow Shop Scheduling – Three Machines, Johnson’s Algorithm and Branch and
Bound Algorithm
In the previous lecture, we saw the Johnson’s algorithm, which could provide an
optimum solution to a n job, 2 machine, flow shop, sequencing problem to minimize
Makespan. We also showed a numerical illustration to explain the various steps in the
Johnson’s algorithm. Now, the next obvious question is, can the Johnson’s algorithm
works for a 3 machine problem, other things being equal, n jobs 3 machines flow shop
sequencing problem to minimize makespan. Now, Johnson actually proved that, under
certain conditions, the Johnson’s algorithm can be extended to solve the 3 machine
problem optimally.
460
So, Johnson gave two conditions, which are this. If maximum of the processing times is
less than or equal to minimum of the processing time, or less than or equal to minimum
of the... Now, this condition says that, the maximum of the processing time of the jobs on
machine number 2, which is here, the maximum of the processing time. Now, we are
considering a 4 job 3 machine problem flow shop sequencing problem to minimize
makespan.
So, the largest of the values for machine 2, which is 10 should be less than or equal to the
smallest of the values on machine 3, which happens to be another 10, so 10. Or the
largest on machine 2, which is 10 should be less than or equal to the smallest on machine
1, which is 3. Now, out of the two inequalites we observe that, for this particular
problem, the first one is satisfied, the second one is not satisfied. But, it is also enough if
one of them is satisfied, therefore the Johnson condition holds for this problem.
So, when the Johnson condition holds for this problem, how do we get the optimum
solution? What we do here is, we create an equivalent 2 machine problem, because the
basic Johnson’s algorithm works for 2 machines. So, we create an equivalent 2 machine
problem with the 4 jobs J 1, J 2, J 3 and J 4 on 2 equivalent machines and these two
equivalent machines are called M 1 plus M 2 and M 2 plus M 3. So, the first column
here, which is called M 1 plus M 2 is actually the sum of the processing times on
machines 1 and 2.
461
So, it becomes 3 plus 8, 11, 12 plus 9, 21, 8 plus 6, 14, and 12 plus 10, 22, second
column of the equivalent Johnson 2 machine problem is M 2 plus M 3 which means, we
add the processing times on machines 2 and 3 to get 8 plus 10, 18, 9 plus 12, 21, 13 plus
6, 19 and 10 plus 16, 26. Now, we have an equivalent Johnson problem with 2 machines,
so we apply the Johnson’s algorithm on this data and try and get the optimum sequence
that minimizes makespan for the equivalent 2 machine problem.
So, once again as we have done in the past, we write a small table with 4 positions,
which are for the 4 jobs. Take the smallest number available in this, which happens to be
11 and it happens for job 1, therefore come it happens to be for job 1 on this first
machine. So, come from the left hand side and place job 1 in the first available position.
Now, strike of job 1. Now try to find out the smallest number that is available, which is
14 here, which is for job 3 and on first equivalent machine.
So, come from the left and put J 3 here, now once again remove this, now find out the
smallest number, which happens to be 21, which is for J 2, both on M 1 and M 2, so there
is tie. So, let us go by this 21, which happens to be on the first machine, therefore J 2 will
come here, which automatically means that, the left over job J 4 will come to this side.
Now, since there was a tie, the tie could be broken other way, so when we break the tie
the other way, the sequence would have been J 1 J 3, choosing this 21 would have put J 2
here and J 4 here. Now, for both the sequences, let us try and find out the makespan with
the original data, because the original data is the problem that we are looking at. So, we
look at the sequence J 1 J 3 J 2 J 4.
462
And we look only at completion times on M 1 M 2 M 3, J 1 J 3 J 2 and J 4, so J 1 goes

first, comes out of machine 1 at time equal to 3, requires 3 minutes. At time equal to 3, it
goes to machine 2, it requires a further 8 minutes processing, so comes out at time equal
to 11. Once again enter M 3 at 11 and comes out at 21, because it requires a further 10
minutes on M 3. J 3 on M 1, J 3 starts at time equal to 3, requires another 8, so 3 plus 8,
11. Exactly at 11, M 2 is free, so it takes another 6 units, 11 plus 6, 17.
Now, it has come out of M 2, but M 3 is free only at time J equal to 21, therefore it visits
M 3 at time equal to 21 and requires another 13 time units, so 21 plus 13 is 34. Now, J 2
starts at time equal to 11, finish at time equal to 23. Now, at time equal to 23, M 2 is
available, so requires another 9 units on M 2, so 23 plus 9, 32, but has to wait for M 3,
because M 3 is available at 34, so 34 plus 12 is 46. Now, J 4 enters, so M 1 is available at
time equal to 23, so J 4 requires another 12 units, so 23 plus 12, 35.
At 35, M 2 is free, it requires another 10 time units, so finishes at 45. Now it has to wait
for M 3, because M 3 is available only at 46. So, enters or visits M 3 at 46, takes another
16 time units and comes out at time equal to 62. So, the makespan for this sequence will
be 62 we have, we also have an alternate sequence, which is J 1 J 3 J 4 and J 2.
463
So, let us try and compute the makespan for the other sequence J 1 J 3 J 4 and J 2, so J 1
J 3 J 4 and J 2. Since J 1 and J 3 are the first two jobs, the completion times will be the
same up to this, so we can write 3, 11, 21, 11, 17, 34, so J 4 enters at time equal to 11 on
M 1 and finishes at 23. Now, at 23, M 2 is available, so takes another 10 units, so
finishes at 33 and has to wait till 34, because M 3 is available only at 34. So, it visits M 3
at 34 requires another 16 units, so 34 plus 16 is 50, comes out at 50.
Now, J 2 enters M 1 at time equal to 23, so 23 plus 12, 35, it comes out, so at 35, M 2 is
available, so takes another 9 units, 35 plus 9 is 44, but has to wait till 50. So, at 50, it
visits M 3, takes another 12 units and comes out at time equal to 62. So once again the
makespan for the second sequence is also 62. So, there are two alternate solutions, both
of which are optimum with the same makespan of 62. So, this is how the Johnson’s
algorithm for the 2 machine problem can be extended to 3 machine case, provided the
data satisfies these conditions.
So obviously, the next question that would come is, what happens if this data does not
satisfy the condition, is there another way to try and get to the optimum solution? Now,
what happens is, the moment we do not satisfy this condition, when we realize that the
Johnson’s extension does not give the optimum solutions to the problem or optimum
makespan to the problem.
464
So, we have to resort, if we really want to find the optimum then we have to resort to
what are called the Branch and Bound algorithms, which are computationally a little
more demanding, but nevertheless give us optimal solution to the problem. So, we look
at another illustration, through which we explain the Branch and Bound algorithm to
minimize the makespan on multiple processes in a flow shop.
Now, let us consider this numerical example and try and see, whether first it satisfies the
Johnson’s conditions and if so use it, if not try and look at a Branch and Bound algorithm
to solve. Now, the Johnson condition says, maximum processing time on machine 2,
which happens to be 98, should be less than or equal to the minimum of this. So,
minimum of this happens to be 1, minimum of this happens to be 8, so the Johnson
condition is clearly not satisfied.
So, we cannot apply this to try and get the optimum solutions or the optimum makespan
to the problem. Now, let us ask yourselves a few more questions, now by looking at this
data, can we try and estimate the makespan, try and have what is called a lower estimate
to the makespan or a lower bound to the makespan. Now, what is the lower bound to the
makespan, now let us add the processing times on M 1, M 2 and M 3. So, this is 200, if
we add on M 1, on M 2 is 237 and 129 on M 3, so sum of processing times on 3
machines are 200, 237 and 129.
465
Now, the maximum of these three is 237 and we could say definitely, that the makespan
cannot be smaller than 237. Because, all the jobs have to go through machine 2 and the
total work load on machine 2 itself is 237, therefore the makespan can only be more
than, greater than or equal to 237. So, the maximum of the sum of processing times on
the machines give us a lower bound to the makespan, so we could say, lower bound is
237.
Now, we can actually make these lower bound slightly better or we can tighten the lower
bound. How do we do that? Now, if we take the first machine, now the total load or total
processing on the first machine itself is 200. But then all the jobs are completed only
when the last job entering machine 1 also finishes on machine 2 and also finishes on
machine 3. So, the machine 1 alone would require 200, so there will be an additional
thing that will go into the makespan, because the last of the jobs entering M 1 has to
finish M 2 and M 3.
So, the minimum possible increase after this 200 is the minimum of the sum of
processing times on M 2 plus M 3. So now, we realize, that this is 11 plus 82 is 93, this is
92 plus 8 is 100, this is 66, this is 107. So, the minimum increase that will happen has to
be 66, because one of these four jobs has to be the last job on the machine. And if job 1
happens to be the last job then there is an additional 93.
If job 2 happens to be the last job there is an additional 100, if job 3 happens to be the
last job there is an additional 66, job 4 happens to be the last there is an additional 107,
so definitely the makespan will go up by 66. So, now the lower bound on the makespan
is 200 plus 66, which is sigma P i 1, sum of the processing times on machine 1 plus
minimum over i minimum over the jobs P i 2 plus P i 3, sum of processing times on
machines 2 and machine 3, so this becomes 266.
So, one could say that, the makespan has to be 266 or more, now this is the lower bound
that is computed based on machine 1. Now, let us compute a lower bound based on
machine 2. Now, machine 2 by itself has a load of 237, but the first job entering the flow
shop cannot start processing on machine 2, unless it has finished machine 1. So, the
earliest time machine 2 can begin is the minimum of these numbers, so the earliest
machine 2 can begin is 1.
466
So, lower bound is again 1, which is the earliest machine 2 can begin plus machine 2
requires at least 237 time units. So, earliest machine 2 can finish everything is 1 plus 237
and it is still not over, because the last job on machine 2 has to visit machine 3. So, there
is an additional component with the last job on machine 2 visiting machine 3. And
therefore, at least a minimum of another 8 will have to be incurred on machine 3 before
all the jobs are complete, so 1 plus 237 plus 8, which is 246.
So, to write this algebraically, we will say, minimum over i minimum over jobs P i 1 plus
sigma P i 2 plus minimum over i P i 3, is another lower estimate of the makespan, so the
makespan cannot be less than 246, it has to be 246 or more. Now, lower bound based on
machine 3, let us call these as LB 1, LB 2 and LB 3, lower bound based on machine 3 is
this. Machine 3 by itself has a time of 129, but machine 3 cannot begin till at least one
job has finished on M 1 and M 2.
So, the earliest machine 3 can begin is the minimum of the sum of these two, so this is
88, this is 126, this is 124 and this is 99. So, minimum, the earliest M 3 can begin is 88 if
J 1 happens to be the first job. So, earliest it can begin is 88 and it requires at least 129
then the last job is over, so the earliest based on M 3 would be 88 plus 129, so 217. Now,
to write it in our notation, this will be minimum over i P i 1 plus P i 2 plus sigma P i 3
over i.
So now, we have calculated three lower bounds, one based on each machine, so based on
M 1, we could say the 266 is a lower bound to the makespan, based on M 2 246 and
based on M 3 217. Now, out of the three, we could say now that, the makespan can only
be the 246 or more, so the actual lower bound is the maximum of the three lower bounds
that we have. And then we say that, the lower bound right now is 266, for a makespan is
266, which is the biggest of the three numbers.
467
So, we start what is called branch and bound tree by saying that, the lower bound is 266,
now what we do is, we try and branch in this tree by trying to fix each of the jobs as the
first job in the sequence. So, there are four jobs, so there are four possible branches, the
four possible branches will have J 1 or job 1 as the first job, job 2 as the second, job 3 as
the third, four jobs J 1 J 2 J 3 J 4, each as first job in the sequence. So, this sequence will
have J 1 as the first job, this sequence will have J 2 as the first job and so on.
So now, we try and find out the lower bound for each of these four loads of the branch
and bound tree. So, let us first compute this, so for this, we actually compute three
numbers, once again based on these three things which means that, if we fix J 1 as the
first job then on M 1, all the processing will be over at time equal to 200 and this lower
bound will take the minimum of P i 2 plus P i 3 over i. Since J 1 is the first job, J 1
cannot be the last job, because a very basic idea of this lower bound is that, the total
processing time on M 1 plus the last jobs processing time on M 2 plus M 3.
So, since the J 1 is fixed as the first job, J 1 cannot be the last job, therefore to compute
this minimum, we would only look at the minimum of this (Refer Slide Time: 23:59) this
and this, we will leave out this 93. So, minimum of this, this, this, still remains as 66, so
200 plus 66 is the first computation. Now, for the second computation, J 1 is the first job,
so J 1 takes 77 time units on machine 1. So, M 2 can start only at time equal to 77, so this
468
will become 77 plus sum of processing times on machine 2, which is 237 plus the last on
M 3.
We will leave out J 1, because J 1 is already the first job, so J 1 cannot be the last job, so
the last job can be only J 2 or J 3 or J 4. So, out of J 2 or J 3 or J 4, we take the minimum,
so out of 8, 30 and 39, we take the minimum, which again happens to be 8. So 77 plus
237 plus 8. Now on M 3, J 1 is the first, so earliest this can begin is 77 plus 11, which is
88 and M 3 by itself has a time of 129. So, the maximum of these three numbers will
now be the lower bound if job 1 is fixed as the first job.
And we calculate that, so this is 266, this is 312, 237, 245, 252, 322, so this is 322 and
this will be 217. So, the maximum of the three values is 322, therefore we say that, if we
start with J 1, the lower bound is 322. Now, let us do this computations with J 2 as the
first job, now in J 2 is the first job, on M 1 the total time required is 200. Now, J 2 cannot
be the last job, because it is the first job, so the last can only be J 1 or J 3 or J 4. So the
minimum is 93, again 66 and 107, so this will remain as 200 plus 66.
Now, the lower bound based on M 2, J 2 is the first job, so earliest M 2 can begin is 34
plus 200 and 37, which is the sum of processing times on this. So, 34 plus 237, since J 2
is the first job, now it cannot be the last job, so the last job can only be the 1, 3 or 4. So,
take the minimum of 1, 3 or 4, which happens to be 9, so we have plus another 9. Now, if
J 2 happens to be the first job, the earliest M 3 can begin is the sum of 34 plus 92, which
is 126 plus another 129, so 126 plus 129.
So, this is 266, this is 237 plus 9 is 246, 280. This is 126 plus 129 is 255, so maximum of
the three is 280 and therefore, 280 is the lower bound when J 2 is fixed as the first job.
So, let us now do that, if J 3 is fixed as the first job, now when J 3 is fixed as the first job,
M 1 by itself will require 200. The last of the jobs cannot be J 3, because J 3 is fixed as
the first job. So, the last can only be J 1, J 2 or J 4. So, sum of 11 plus 82 93, 92 plus 8
100, 98 plus 9 107, so minimum is 93 here, so we will get 200 plus 93.
Now, if J 3 happens to be the first job, the earliest M 2 can begin is 88, because J 3 has to
finish on M 1 and only then it can enter M 2. So, earliest it can enter is 88 plus another
237, now J 3 cannot be the last job, so we have to take the minimum of this, (Refer Slide
Time: 29:16) this and this, which happens to be 8, so 88 plus 237 plus 8. Now, once
469
again if J 3 is the first job, earliest M 3 can begin is after it has finished this and this,
which is 88 plus 36, 124 plus another 129.
So, 124 plus 129, this is 293, this is 245, 253, 333 and this is 253, so the highest of them
is 333, so the lower bound becomes 333. Now, let us look at J 4 as the first job and
complete this. So, M 1 takes a time equal to 200, J 4 now cannot be the last job, last job
can only be J 1 J 2 and J 3, so minimum of 93, 100 and 66, so you will get the same 200
plus 66 here. Now, if J 4 is the first job, the earliest M 2 can begin is 1, so 1 plus 237 is
238, once again J 4 cannot be the last job.
So, minimum of this, this and this, which is 8, so 1 plus 237 plus 8, which is 246 and if J
4 is the first job, earliest M 3 can begin is 99, because J 4 has to finish 1 plus 98. So, 99
plus 129, which is 228, so this is 266, this is 246, this is 228, so lower bounds remains as
266. So, what we have observed now after this calculation is that, if we fix job J 1 as the
first job, the makespan is going to be 322 or more. If we fix J 2, makespan is going to be
280 or more, if we fix J 3, 333 or more and if we fix J 4, it is 266 or more.
Since we want to minimize the makespan, we proceed in this direction, because this is
the smallest number 266. So, there is a possibility that, we get a smaller number as we
proceed from here. So, we proceed from this node right now, we still have to evaluate
this, this and this, we will see that as we move along. But then we proceed from the
node, which has the smallest value of the lower bound. So, from this node we proceed,
now we have already fixed job 4 as the first job.
So, we have to now fix the second job, so job 1 is the second job, job 2 can be the second
job, job 3 can be the second job. Now, let us evaluate this, if job 4 is the first job and job
1 is the second job, now what we have to do is, now we have what is called a partial
sequence, so according to this, the partial sequence is J 4 and then J 1. Now, let us first
evaluate the completion times for this partial sequence, which is J 4 and J 1, let us first
evaluate the completion times on this.
So, J 4 begins here, this is completion time, so J 4 is 1, it finishes at 1, 1 plus 98, 99 and
99 plus 9 is 108. Now, J 1 is here say, J 1 starts at 1 and finishes at 78, now we are
evaluating this for the partial sequence J 4 J 1. Now, at 99, it goes it visits M 2 takes
another 11 units, so comes out at 110. So, at 110, M 3 is free, takes another 82 units,
470
comes out at 192, so we keep this partial sequence with us to help us make the lower
bound better.
Now, what we do, now we have to find out the three lower bound corresponding to the 3
machines. So, corresponding to M 1, the minimum time is 200, already 4 and 1 are fixed,
2 and 3 are remaining, so 2 and 3, one of them only can be the last job. So, we take the
minimum of 100 and 66, so the first lower bound still remains as 200 plus 66. Now,
when we look at M 2, the computation will be very interesting, so we would have...
Now, at this point, when we consider the partial sequence J 4 J 1, now M 2 is available
for the remaining two jobs, which are J 2 and J 3 at time equal to 110.
So, the minimum M 2 can complete all the jobs is 110 plus 92 plus 36, so 110 plus 92
plus 36 is 110 plus 92 is 202, 202 plus 36 is 238. Now, the earliest M 2 can complete is
238, now two jobs are remaining J 2 and J 3, one of them has to be the last job. So, it
would require an additional 8, which is the minimum of this, so 238 plus 8. And we will
see the same thing happening for M 3 also, now based on these partial sequence, M 3 is
available for the remaining jobs J 2 and J 3 at time equal to 192, so at time equal to 192,
M 3 can start looking at J 2 and J 3.
So now, from 192, J 2 and J 3 plus 8, 200 plus 30, 230, so 192 plus 8 plus 32, 130, so the
lower bound is 266, 246, 130, so lower bound remains as 266. Now, let us show one
more computation here with J 4 and J 2 as the two jobs. So we will do for J 4 and J 2.
Now, let us quickly do the completion times, so J 4 will be again 199 and 108, coming
from this 199 and 108. Now, J 2 enters at 1, finishes at 35, so J 2 finishes at 35, enters at
99 requires another 92, so 99 plus 92 is 191, so this is 191 and M 3 is free at 191 takes
another 8 units, so 199 it comes out.
Now, using this partial sequence, we will now go back and try and find out the lower
bound. So, as far as M 1 is concerned, M 1 will finish at 200, jobs 4 and 2 are fixed, so 1
and 3 are available. So, 1 and 3, once again 3 is the minimum with 66, so the first lower
bound will be 200 plus 66. Now, M 2 is available at 191 for jobs 1 and 3, so 191 plus 1
and 3, 191, 192, 202, 238, so at 238, all the jobs can be over earliest in M 2. Then once
again, it has to go for another minimum over 1 and 3, which is another 30.
So, this will be 191 plus sum of processing times of 1 and 3, which is 11 plus 36, which
is 47 plus 30. And as far as M 3 is concerned, M 3 is available at 199 to do 1 and 3, so
471
199 plus 82 plus 30, which is 112, so 199 plus 112. Now, this is 266, now this is 191 plus
30 is 221, 228, 268, and this is 199 plus 112, 211, plus 100 is 311. So, the lower bound
becomes 311 when we do this.
Now let us calculate the other lower bound for J 4 and J 3 also. Now, to do that, we have
J 4 and J 3 here, so we will quickly do the completion times 199 and 108, for J 3 it is 88,
so 1 plus 88 is 89, starts at 99 requires another 36, so 135, 99 plus 36. At 135, M 3 is
free, so takes another 30, 135 plus 30 is 165, so again based on the lower bounds, so as
far as M 1 is concerned, it will take 200. Now, 4 and 3 are already fixed, so 1 and 2 are
available, so minimum over, this is 93, this is 100, so 200 plus 93, which is 293.
The earliest M 2 is available is 135 for 1 and 2, so 135 plus (Refer Slide Time: 41:05)
this plus this is the earliest M 2 is over. Then the last job has to do one more operation,
so minimum over 82 and 8, which is 8, so 135 plus 11 plus 92 plus 8 and M 3 available
at 165 for 1 and 2. So, it has to do these two plus another 90, so 165 plus 90, 255, this is
293, so this is 100, 235 246, so the lower bound remains at 293. So now, we have all
these then we still need to proceed to get to the solution.
So now, the smallest available number is still 266, so we can proceed from this to try and
get some more solutions. So, we have already fixed two positions, so we could fix the
third position now. So, J 4 and J 1 have been fixed, so now J 2 and J 3 are remaining, so
we will say, we fix J 2 as the third job, we fix J 3 as the third job in the other sequence.
So, we fix 3 positions now, so J 4 J 1 and J 2, so we can continue in the same way to get
the lower bound for J 4 J 1 and J 2.
But, we observed that, there are only four jobs, we have already fix three positions, so
the left over job has to automatically take the fourth position. So, instead of computing a
lower bound, we compute the makespan for a full sequence, which is J 4 J 1 J 2 and J 3.
So, J 4 J 1 J 2 and J 3, the other alternate sequence is J 4 J 1 J 3 and J 2, so the other
alternate sequence is J 4 J 1 J 3 and J 2, because here we fix J 3 as the third job, so the
left over job, which is J 2 will be the fourth job.
So, branching from here, we are able to get two full sequences, because in all these
partial sequence, there is only one more job to go. So, we can evaluate this sequence or
we can evaluate this sequence, we can evaluate either of them. So, let us first evaluate
this sequence and try to learn a few things then we will come back to this sequence. So,
472
we evaluate the full sequence here, we already have for J 4 and J 1. So, for J 4 and J 1,
we already have here, so we write 199, 108, 78, 110, and 192.
Now, we look at J 3, now J 3 enters at 78 takes another 88, so 166. So at 166, M 2 is free,
so it takes another 36 – 166, 172, 202. At 202, again M 3 is free, takes another 30, so
232, now J 2 enters at 166 plus 34 is 200, waits till 202 for M 2, takes another 92, so 202
plus 92 is 294. At 294, M 3 is free, so goes to M 3 and comes out at 302, now 302 is the
makespan associated with this sequence. Now, we have a feasible solution with
makespan equal to 302, 4 1 3 2 is a valid sequence, it is one out of the four factorial
possible sequences for this problem.
So, it is feasible and has a makespan of 302, so the moment we have a feasible solution
with a makespan of 302, we also understand that, the optimum now can be 302 or less.
So, the feasible solution provides an upper bound to the optimum, so it provides an upper
bound to the optimum and therefore, the optimum now is 302 or less. Now, with this
additional information, we now observe that, if we look at this node from here, with
fixing J 1 as the first job, we understand from here, that if we proceed down, the
makespan that we will get can only be greater than or equal to 322.
So, we are now not interested in any solution, which has makespan 302 or more, because
you want to minimize the makespan and we already have a solution with 302. So, there is
no need for us now to proceed from here, because we are only going to get solutions,
which are poorer than 302. So, we decide not to proceed from this, so we say that, we
fathom this node, so this node is fathomed, because the lower bound here is greater than
or equal to the current based upper bound of 302.
So, 302 acts as an upper bound, so we fathom a node when the lower bound in that node
is greater than or equal to current based upper bound. Now, similarly the 333 is also
fathomed for the same reason, that lower bound is greater than or equal to the current
based upper bound, but we cannot do that to this 280. Now, this 311 can also be fathom,
we normally use fathom or with an f, they mean the same. So, we could put an f and say
that, it is fathomed for the same reason, that the lower bound is bigger than the current
based upper bound.
We cannot proceed from here, because we have already reached the full sequence, so this
node is fathomed by upper bound or by feasibility, because we have a feasible solution
473
there, we cannot proceed any more. So, now we realize that, doing this computation has
helped us actually indirectly, because we now do not have to proceed and evaluate
sequences from here, we do not have to proceed from here.
But, there are still some nodes that are active which means, we can move from here, we
can move from this 293, we can evaluate this. So, we now try and evaluate this. So we
already know the sequence for J 4 J 1, so you get 1, 99, 108, 78, 110, and 192. Now, J 2
enters at 78, takes another 34, so 112 is when J 2 finishes here. At 112, machine 2 is
available at 110 itself, so takes another 92 units, so 112 plus 92 is 204. At 204, M 3 is
available, takes another 8 units, so at 212, this leaves.
Now, J 3 is the last job, J 3 enters M 1 at 112, 112 plus 88 is 200, waits till 204 and then
visits M 2 at 204, requires another 36 time units, 204 plus 36 is 240. M 3 available at
240, so visits M 3, takes another 30 units and comes out at 270. Now, the makespan
associated with this sequence is 270. Now this 270 is going to help us in many ways.
Now, this node as I said, is fathomed by feasibility, this 270 is an upper bound and the
moment we have solution with makespan equal to 270, it is better than the current best
solution of 302, so we update.
So, we keep the best solution now as 270, now the moment we have a 270 solution, we
are not interested in any solution, which are more than 270. So, we now start looking at
all these nodes, now this tells us that, if we move from this direction, it is going to be 280
or more. So, we do not have to move from here, because we already have solution of
270, so this is again fathomed with lower bound greater than or equal to upper bound.
Only difference is that, the upper bound has been updated, so from 302, it has become
270, therefore you can fathom this.
Now, proceeding here, which is one that is available, 293 is also be fathomed with lower
bound greater than upper bound. So now, we are realized, that there is no active node in
this network, so we cannot proceed any more from this network, because all the nodes
have been fathomed. They have been fathomed either, because they give a feasible
solution or they have been fathomed, because the lower bound associated with that load
is greater than or equal to the current best upper bound.
So, there is no more node to proceed, the algorithm terminates, the Branch and Bound
algorithm terminates and the best feasible solution at this point, which is the solution
474
with 270 is the optimum solution. So, the optimum solution is J 4 J 1 J 2 J 3 with 270 as
the optimum makespan. So, this is how the Branch and Bound algorithm works, so we
branch by creating sequences and partial sequences, we evaluate lower bounds.
And then as we move down the tree, we evaluate feasible solutions and upper bounds
and try to fathom the nodes based on two things in this case, based on a feasible solution
and based on the condition that lower bounds is greater than the upper bound. Now, this
is called, is like an Implicit Enumeration algorithm, we have actually evaluated all the
four factorial, but not explicitly. We have explicitly evaluated only two out of the four
factorial, but based on these, the way we worked out the branch and bound tree, we have
implicitly evaluated the remaining.
And we have said that, the remaining cannot give a solution, which is better than 270, so
it is an implicit enumeration kind of an algorithm, but in the very worst case, we may
have to evaluate all the n factorial, which are there. So, this is how the Branch and Bound
algorithm works for solving the makespan minimization problem in the flow shop. And
other thing that we could have done to cut down the computation is this, somewhere here
when we fix two jobs J 4 and J 1, we evaluated the lower bound and we proceed here.
When we fix three jobs, we said there is only one vacant position, so the fourth job will
automatically come there and therefore, we got a feasible solution here. In Branch and
Bound algorithm, it is customary to actually write here to evaluate the feasible solution.
Because, when we fix these two, there are only two more positions remaining with two
more jobs remaining and they can be filled only in two ways.
So, right here we could have said, this will give rise to two feasible solutions, which is J
4 J 1 J 2 J 3, J 4 J 1 J 3 J 2 which means, right here we could have evaluated this without
doing this and without doing this 266. So, that way you can speed up the Branch and
Bound algorithm, when you have only two more positions remaining, you do not
evaluate the lower bound, instead you fill it in two ways and evaluate two feasible
solutions.
So, this is how the Branch and Bound algorithm works in order to minimize the
makespan. This can be extended to general n job m machine based on the idea, we have
3 machines, so we calculated 3 lower bounds and choose the best. If you have m
475
machines, you will choose m such calculations for every node and choose the lower
bound. Now, in case, obviously we understand that the Branch and Bounds algorithm is
time consuming. And as the problem size increases, there are times the Branch and
Bounds algorithm is unable to give the optimum solutions within a reasonable
computational effort. So, in such cases, what do we do is the next question that we have
to address, we will address that question in the next lecture.
476
Prof. G. Srinivasan
Lecture - 27
Flow Shop Scheduling - Heuristics - Palmer, Campbell Dudek Smith Algorithm
In the last two lectures we have been addressing the n job, m machine, flow shop
scheduling problem to minimize make span. We looked at the Johnson's algorithm,
which would solve the two machine problem optimally, we also looked at the extension
of the Johnson's algorithm, where the 3 machine problem the make span can be solved
optimally, provided the data on the processing times satisfy certain conditions. Now, we
also looked at a branch and bound algorithm, which we could use to solve the general m
machine problem optimally in particular we showed an example with three machines.
477
We also mentioned that, the branch and bound in the worst case if there are n jobs, then
the branch and bound could evaluate all the n factorial possible permutations or n
factorial possible sequences, before it comes or before it arrives at the optimum solution.
So, in the worst case the branch and bound algorithm can do n factorial, which means
that it will or can do an exponential number of computations. So, when it does an
exponential number of computations, the computing time taken to solve the problem
optimally becomes very large, as the size of the problem increases.
So, in such cases it is customary to resort to what are called heuristics or heuristic
solutions, which essentially do not guarantee optimum. So, in some sense they do not
guaranty optimum, so they can be termed as non optimal methods, but for certain
instances the heuristic solutions can give the optimum solutions themselves. The
heuristics are capable of giving the optimum in certain instances, but they do not
guarantee optimal solutions on all occasions.
So, we will now look at some heuristics to solve the general flow shop scheduling
problem to minimize make span; but before that let us also look at one aspect which is
regarding this n factorial. Now, if we consider a two machine flow shop, which means
we have M 1 here, and we have M 2 here. And let us say we have only two jobs J 1 and J
2 which are here, so since there are two jobs, there are two sequences possible, which is
J 1 followed by J 2 and J 2 followed by J 1.
478
Now, let us assume that the processing times for J 1 are 6 and 5 which means 6 on M 1
and 5 on M 2, and processing time for J 2 are 7 and 8, 7 on M 1, 8 on M 2. We have
already seen the computation of the completion times, and the make span, but let us
revisit that aspect once more. So, if we consider the sequence J 1, J 2 which means job J
1 goes first, and then followed by J 2 we automatically compute the sequence like this.
We look at completion times on M 1 and M 2, let me say completion times here for J 1
for J 2. So, we would say that J 1 first goes to M 1 at time equal to 0 finishes at 6, then J
1 goes to M 2 finishes at 11, J 2 enters M 1 at time equal to 6 finishes at 13, and then this
is available at 11. So, J 2 can take M 2 at 13, 13 plus 8 21 this we are very familiar with.
We have made an important assumption here that, the moment we decide that the
sequences J 1 followed by J 2 that sequences maintained on all the machines, both M 1
and M 2 we have maintained that sequence in our computation. For example, let us look
at a scenario where we decide that the sequence on M 1 is J 1 followed by J 2, and the
sequence on M 2 is J 2 followed by J 1. So, let me show those calculations.
So, we will first look at M 1 and then we look at M 2 let us assume on M 1 the sequence
is J 1 followed by J 2. So, the sequence is J 1 followed by J 2 on M 1, which means that
in M 1 we first start with J 1 takes 6 time units, so this comes out at time equal to 6, now
J 2 enters M 1 at time equal to 6 therefore, will come out at time equal to 13. Now, let us
here on M 2 the sequence is J 2 followed by J 1, so let me write that on M 2 it is J 2
479
followed by J 1.
So, this would mean that J 1 will not immediately go to M 2 even though M 2 is free
because the sequence on M 2 is J 2 followed by J 1. So, M 2 will wait for J 2 to be over,
so J 2 can access M 2 at time equal to 13, and then J 2 requires another 8 units at 21 J 2
will come out. Since the sequence on M 2 is J 2 followed by J 1, now J 1 can enter M 2
after J 2 is completed, so J 1 even though it has completed here at 6 will wait till 21 and
then it would take another 5 units and then at 26 J 1 comes out.
The next span therefore, for this order is 26 compared to the make span here, which is
21, now if we see very carefully the flow shop assumption is not violated the flow shop
assumption means that every job will visit the machines in the same order. So, even if we
look at this schedule or time table J 1 first visits M 1, and then J 1 visits M 2, J 2 visits M
1, J 2 visits M 2. So, the flow shop assumption is not violated, so in principle, in the flow
shop, if the flow shop has m machines, in each machine the n jobs can be sequenced in n
factorial ways.
So, over each machine it can be done in n factorial ways and since there are m machines
it can be done the total sequence is possible actually n factorial to the power m. But, it is
not common and customary to use this; it is fairly obvious making a comparison between
this and this. If we retain the same sequence for all the machines, on many occasions we
would get better value of the performance measure, such as make span or completion
time and so on.
For example, the make span is 21, the make span is 26 sum of completion times is 33,
sum of completion times is 47. So, if there were a due date related measure, once again
this would perhaps show a higher tardiness than this; so for what are called regular
performance measures, where the performance measure or the objective function will
increase when completion time is increased. It is better not to follow this, and to follow
this where the sequence of jobs is the same on all the machines.
This is a situation where the order or sequence of job is different on different machines,
even in a flow shop. So, it is advantageous to have a situation or to model it in such a
manner that the sequence of visit is the same on all the machines therefore, we do not
consider all the n factorial to the power m; instead we consider only n factorial
sequences. So, 1 out of the n factorial will be optimal for the first machine, and the same
480
1 out of the n factorial now is chosen, such that it is optimal when we consider all the
machines.
So, we reduced the feasible solution space from n factorial to the power m to n factorial,
so those sequences where or situation where, we consider the same sequence for all the
machines, such as this is called a permutation flow shop. So, unless otherwise stated,
flow shop would mean a permutation flow shop and we will not look at non permutation
flow shops, where the order or visit on the machines are different. So, far we have seen
only permutation flow shops in this lecture series and we will continue to look only at
permutation flow shops in this lecture series.
Now, let us go back to the heuristics to solve the problem. So now, let us consider this 5
job, 3 machine flow shop sequencing problem to minimize make span, and as I have just
mentioned, we will consider only permutation sequences and we will not look at non
permutation sequences. So, 1 out of the n factorial will be optimal in this case, now we
observe first of all we have to check whether it satisfies the Johnson condition, if it
satisfies the Johnson condition then we could get the optimal solution straightaway using
a simple algorithm.
Now, minimum on the first machine is 13, minimum on the third machine is 11,
maximum on the second machine is 20. So, maximum on the second machine is greater
than the minimum on the first, as well as minimum on the third. Therefore, the Johnson
481
condition is not satisfied. So, we cannot apply the Johnson condition to try and get the
optimal solution to this problem, we may not get the optimum solution if we continue
using the Johnson condition and solve it. So, let us look at some other heuristic solutions
to this problem.
Now, the first heuristic that we will look at will try and find out a weighted sum for each
of these jobs. So, we will try and give some weights to each of these machines, and then
we try and find a weighted sum for every job, so let us begin with giving a weight of 0 to
the machine in the middle, giving a weight of plus 2 here and giving weight of minus 2
for this machine. And then let us try and find out a weighted sum of processing times for
each of the jobs.
So, let us call that as say let us call W 1 as the weight for job 1, so for job 1 the weight
will be 16 into minus 2, which is minus 32 plus 18 into 0 plus 12 into 224. So, 24 minus
32 is minus 8, now W 2 will be 14 into minus 2 which is minus 28 minus 28 plus 22 is
minus 6. W 3 will be 13 into minus 2 minus 26 plus 15 into 230 which will be 4. W 4
will be 19 into minus 2 - minus 38 plus 19 into 2 plus 38, 0 and W 5 will be 15 into
minus 2 is minus 30 plus 16 into 2, 32, so it will be plus 2.
So, we have now found a weight associated with each job, based on the weights that we
have arbitrarily assigned to the machines. So, it is very clear that if these weights are
changed these numbers will also change, but right now let us assume a standard weight
of 0 for the middle plus 2 here and minus 2 here. And then we have computed the
weights, now sort the jobs in decreasing order of the weight, non increasing order or
descending order of the weights.
So, the maximum is W 3, so J 3 comes first the next 1 is W 5, so J 5 comes second the
third 1 is 0. So, J 4 comes third the next one is minus 6, so J 2 comes here and the last is
minus 5 J 1 comes here, so this is the sequence that we get, if we compute a weight for
each job, based on a weight given to each machine. Now, for this sequence let us try and
find out the make span, so we write the completion times for this sequence, we have M
1,M2,M3,J3,J5,J4,J2,J1.
So, J 3 starts at 0 finishes at 13 now goes to M 2 at 13 takes another 20, so finishes at 33 now
goes to M 3 at 33 takes another 15 and finishes at 48. Now, J 5 can begin at 10 equal to 13
on M 1 takes another 15, so finishes at 28 has to wait till 33, so that M 2 is free
482
takes another 16. So, finishes at 49, M 3 is already available at 48, so starts on M 3 at 49
takes another 16, so finishes at 65, J 4 can start on 20 on M 1 at 28. So, 28 plus 19 it
finishes at 47.
M 2 is available only at 49, so it waits till 49 visits M 2 takes another 15 units finishes at
64, M 3 is available only at 65. So, goes to M 3 at 65 takes another 19 and finishes at 84.
J 2 starts at 47 on M 1 takes another 14, so finishes at 61 waits till 64 for M 2 starts at 64
on M 2 takes another 10 for 74 waits still 84 for M 3 starts at 84 takes another 11, so
finishes at 95.
Now, J 1 starts at 61 on M 1 takes another 16, so finishes at 77, M 2 is already available

at 74, so starts at 77, 77 plus 18 is 95 it comes out at 95 M 3 is available. So, starts at 95
takes another 12 and finishes at 107, so make span associated with this sequence is 107,
this is the make span associated with this sequence.
So, this is how this algorithm works, so very simple algorithm, so one pass algorithm
where you define weights for the machines. Right now the weights are arbitrary, but
there is a logic in this which we will see. Right now the weights are arbitrary, but for any
given weight you can calculate an index or a weight for every job, compute that index
and then sort the jobs in decreasing or non increasing value of the indices to get the
sequence like this and evaluate the make span of this sequence and that gives a heuristic
solution.
So, this heuristic algorithm will evaluate only one sequence it could evaluate 2 if for
example, the some of these indices are equal. So, you could have 2 or more if some of
these indices are equal, but in general it evaluates only one sequence if all the values of
these each of these indices is unique. So, it gives a solution with make span equal to 107,
now 1 thing that we know is that this 107 need not be optimal, with the bit of luck it
could be optimum. But, then there is no guaranty that the 107 that we obtain is indeed
optimal.
So, we need to find out how good it is, is it good enough or is it not good enough, now in
order to do that we also try and find out lower bounds for the make span, we have
already seen lower bounds for the make span based on each of the machines. So, we
apply those lower bounds to see, what we get as a lower bound for the make span.
483
So, lower bound based on M 1 is equal to.
Now, let us first find out sum of processing times on all the 3 machines 16, 30, 43, 62
plus 15, 77. 18, 28, 48, 63 plus 16, 79. 12, 23, 38, 57 plus 16 is 73. So, the lower bound
based on each of these is M 1 all require 77 time units, but then the last job on M 1 has to
go to M 2 and M 3. So, the minimum additional time that we require is the minimum of
these sums 30, 21, 35, 34, 32, 21 is the minimum, so minimum additional time required
is 77 plus 21 ((Refer Time: 22:06)).
484
So, lower bound based on M 1 is 77 plus 21 which is 98, now lower bound based on M 2
will be the earliest we can start processing on M 2 is the minimum of these which is 13.
Because, no job can directly start on M 2, it has to go to M 1 and then only it can start M
2, so the earliest M 2 can begin is the minimum of these numbers which is 13, M 2 itself
requires another 79. And then the last job on M 2 also has to go to M 3, so it requires at
least an addition which is the minimum of these which happens to be 11.
So, the lower bond will be 13 plus 79 plus 11, so this will be 13 plus 79 plus 11 which is
79 plus 11 is 90, 90 plus 13 is 103 is a lower bound based on M 2. Now, lower bound
based on M 3 is no job can start on M 3 directly, so any job that the first job on M 3
should have visited M 1 and M 2. So, the earliest M 3 can begin is these 2 sum to be
minimum, so this is 34, 24, 33, 34, 31, 24 is the minimum plus it requires a 73, so lower
bound on M 3 is equal to 24 plus 73 which is 97.
Now, out of the 3 lower bounds the maximum one is the best bound, so 103 is a very
good lower bound that we have got from using the 3 machines. So, now we have 2 pieces
of information, we know definitely based on the lower bound that the make span cannot
be less than 103. So, the make span can only be 103 or more, now we also know from
here that since we have the optimum make span can be 103 or more based on the lower
bound.
Based on this we already have a feasible sequence, with the make span of 107 -
therefore, the optimum make span can only be 107 or less. So, the goodness will be H
minus L by L into 100 percent, now we assume the worst case the optimum is the lower
bound, which is the worst case. And therefore, if the optimum where 103 then the
goodness of this is 107 minus 103 divided by 103 into 100 percent, so this is 4 by 103
into 100 percent. So, this is let me say approximately 4 percent or less than 4 percent, 3
point something percent we would say 4 percent.
So, if we are willing to accept a solution which is say worst case is 4 percent above the
optimum, the reason the worst case 4 percent comes because you do not know the if this
optimum is more than L. Actually this should be H minus O by O, heuristic minus
optimum by optimum, since we do not know the optimum we substituted with a lower
bound. So, the optimum is 103 or more therefore, the actual goodness will be 4 or less
because this can increase and H minus L can actually decrease.
485
So, in the worst case it can only be 3 plus something, so if we are willing to accept a
solution which is within 4 percent, then we can comfortably take this and we need not
worry about trying to get to the optimum. But, on the other hand we would like solutions
which are better than this, do we have sequences which have make span less than 106. In
such a case then we can look at other heuristic algorithms, and try to see if we can get
solutions, which are less than 106 make span.
Now, this heuristic is called the Palmers heuristic or Palmers algorithm, and was
developed by Palmer in the year 1965. Even though these values of minus 2, 0 and plus 2
appear arbitrary, there is a rationale in choosing these values, if we see very carefully
these sequences J 3, J 5, J 4, J 2 and J 1. They were based on a decreasing order of the
index, now J 3 had the highest value of the index, while J 1 had the lowest value of the
index.
Now, if you see these J 1 is it is very clear it got the lowest value because this 16 is
bigger than 12, and J 3 had a higher value positive value because this 13 is smaller than
15. Now, for a 3 machine problem we have put a 0 here, so whatever is there on second
machine does not matter, only what is there on the first and third machines matter. So, if
a job has a higher processing time on the third machine or on the last machine, and a
lower processing time on the first machine like J 3 it will have a higher value of the
index.
Another job which has a higher value on M 1 and a lower value on M 3, will have a
lower value of the index and therefore, figure later in the sequence. J 3 which has a lower
value on M 1 and higher value on M 3, will have a greater than 0 and a higher value of
the index and therefore, we will figure early in the sequence. Now, this is very, very
close to the Johnson idea that if we go back and see the rule, the machine that has the
smallest number on M 1 will appear from the left, which means it will appear early in the
sequence while jobs that have smaller values on M 2 will appear from the right, which
means it will appear later in the sequence.
That essence is captured by the weights, in principle that essence is captured by the
weight. So the principle should be the middle one has a certain weight, the 1's to the
right generalizing it to a 5 machine or a 6 machine problem, the 1 that comes to the right
should have positive and increasing weights. And those to the left should have negative
486
and decreasing as - then the absolute weight will be high, but because of the negative it
will be decreasing.
Then we fit this algorithm into the Johnson principle which works, but then there are a
couple of limitations here. Because, we put a 0, then the second machine is not
contributing at all. And then there are a few other questions like should this be 0 or
should this be some constant, some other thing other than 0. Now, should these be the
same all these questions can come, but essentially the principle behind computing this
index is based on the Johnson idea.
Now, what happens when I have an even number, if I have 4 machines what do I do? So
the suggestion was minus 3, minus 1, plus 1 and plus 3 there is no middle machine. So,
the two middle 1's will get minus 1 and plus 1 and then a plus 3 on this side and a minus
3 on the other side. So, there were several improvements to this algorithm. But the most
important contribution or two things associated with this algorithm is that it is a very
early algorithm - one of the earliest. And that the very idea of the weight on the index
captures the essence behind the Johnson's algorithm.
Now, let us go to another algorithm let us simply retain this solution somewhere and let
us write this solution here which is J 3, J 5, J 4, J 2, J 1 with make span equal to 107.
Now, let us go back to another algorithm and see what we do. Let us also write that our
lower bound is 103 that we have, now when we started, we check the Johnson condition
487
and we said that this does not satisfy the Johnson condition, so that it definitely gives the
optimum solution, still we could have used the same algorithm.
Because, if it had satisfied the Johnson condition, then we would have created an
equivalent 2 machine problem, and then we would have solved a 2 machine problem
optimally using Johnson's algorithm. Now, it does not satisfy the Johnson condition, but
nothing prevents us from applying that rule to see what we get, so let us try that.
So, let us now creates an equivalent 2 machine Johnson problem by adding, so this
becomes M 1 plus M 2, this becomes M 2 plus M 3, J 1, J 2, J 3, J 4 and J 5. So, this
becomes 34 18 plus 16 18, 18 plus 12 30 24 21 33 and 35 34, 34 and 31 32, so now, we
make the Johnson sequence for this, the smallest one is 21. So, J 2 goes here this goes the
next smallest 1 is 30, so J 1 comes here 30 the next smallest 1 is 31, so J 5 comes here
the next smallest 1 is 33. So, J 3 comes here 33 is on the first machine, so it comes from
the left and the only 1 available is J 4.
So, when we apply the extended Johnson's algorithm for the 3 machine into this,
knowing fully well that it will not or may not give the optimum solution, we get a
sequence like this. Now, let us find out the completion time for this sequence, so CT is
here the sequence is J 5 first J 3, J 4, J 1, J 2, now the make span associated with this is J
5 starts at 0 and finishes at 15, at 15 it goes to M 2 comes out at 31 15 plus 16 31, at 31 it
goes to M 3 takes another 16 and comes out at 47.
488
Now, J 3 is the next 1, so J 3 starts at 15 on M 1. This 15 takes another 13, so comes out
at 28 has to wait till 31 for M 2 to be free, takes M 2 at 31 finishes at 51, 31 plus 20 by
which time M 3 is free. So, 51 plus 15 is 66. Now J 4 comes third M 1 is available at 28,
so starts at 28 finishes at 28 plus 19 which is 47 waits till 51 for M 2 to be free, goes to
M 2 at 51 takes another 15 and finishes at 66. Now, exactly at 66 M 3 is free, so it takes
M 3 and then takes another 19 and finishes at 85.
J 1 can start at 47 because M 1 is free at 47 takes another 16, so a 47 plus 16 is 63 - waits

till 66 for M 2 to be free, goes to M 2 at 66 takes another 18 and then comes out at 84,
but M 3 is available only at 85. So, goes to M 3 at 85 takes another 12 units, so comes
out at 97. Now J 2 is the last job it can go to M 1 at 63, starts at 63 takes another 14, so
finishes at 77 waits till 84 for M 2 takes another 10, so 94 waits till 97 on M 3 takes
another 11. So, gives us a solution with 108, so the make span associated with this
sequence is 108. Right?
So, what we have done right now is we have simply applied the Johnson's extension
algorithm, to a 3 machine problem. And created an equivalent 2 machine problem, with
M 1 plus M 2 and with M 2 plus M 3, now if we go back and revisit what we did in the
palmer algorithm, we essentially put a 0 weight on M 2 which means we looked only at
M 1 and M 3. So, nothing prevents us from actually using the palmer idea again into it,
and take only M 1 and M 3 and try and get an equivalent 2 machine Johnson problem
and apply the Johnson's algorithm on it.
489
So, right now we will look at only with M 1 and M 3 let us formulate another 2 machine
problem here only with M 1 and M 3 let us formulate a 2 machine problem with this. So,
once again J 1, J 2, J 3, J 4 and J 5, so we get 16 12, 14 11, 13 15, 19 19, 15 16, now let
us make another junction the smallest one is 11. So, J 2 on the second machine, so comes
from here J 2, the next smallest is J 1 once again from the second machine, so J 1 comes
here.
The third smallest is J 3 which is on the first machine, so J 3 is here the next smallest is J
5 which is here and then J 4 comes here. So, let us quickly find out the make span for this
sequence, so we write M 1, M 2, M 3, J 3, J 5, J 4, J 1, J 2, so very quickly writing the
completion times. So, J 3 is 13, 13 plus 20, 33, 33 plus 15, 48, J 5 starts at 13 finishes at
28 waits till 33 finishes at 49, M 3 is already available 49 plus 16 is 65. Then we have J 4
starts at 28 plus 19 finishes at 47, waits till 49 takes another 15 which is 64, waits till 65
takes another 19 and finishes at 84.
J 1 starts at 47 takes 16 finishes at 63, waits till 64 takes another 18, so 82, again waits
till 84 takes another 12 which is 96. Now, J 2 starts at 63 takes another 14 finishes at 77,
waits till 82 takes another 10, 92 waits till 96 takes another 11 which is 107, so what we
have done in this algorithm is, now we get another solution whose make span is 107. So,
between these two solutions, we could pick the one which is better which has 107, so we
essentially did two solutions we evaluated in this algorithm.
490
So, one of which is for a 3 machine, the first one we did was a Johnson problem with M
1 plus M 2 and M 2 plus M 3 based on the Johnson extension. The other is we did only
M 1 and M 3, and we took the best out of the two and therefore, we get another solution
where J 3, J 5, J 4, J 1, J 2 were 107 again. So, we did not get a solution better than
palmer, but then we got the same solution as palmer based on this, but we got a different
sequence in this.
Now, for three machines we got 2 solutions and we pick the best, now can we generalize
it to m machines. So, if we have m machines, we call them M 1, M 2, M m, now we
create m minus 1 sequences for three machines we created two sequences for general m
machines we create m minus 1 sequences. So, let us call sequence S 1 as M 1 and M m
take the first machine, take the last machine do a Johnson out of that, which is this, the
second 1 S 2 will be M 1 plus M 2 and M m minus 1 plus M m.
Take the first 2 add take the last 2 add do a Johnson out of it, the third sequence will be
M 1 plus M 2 plus M 3 and M m minus 2 plus M m minus 1 plus M mM take the first 3
and add take the last 3 and add and do it. So, as we keep doing the last sequence S m
minus 1 sequence will be M 1 plus M 2 plus M m minus 1, this will be M 2 plus M 3
plus M m. So, take the first m minus 1, take the last m minus 1 starting from 2 and get it,
we do not find the m'th sequence because that would mean you add all of them both
sides we will get the same, so we do not do that.
491
So, for a general m machine problem you generalize it to find m minus 1 sequences, so if
I have 10 machine problem, irrespective of the number of jobs I will evaluate 9
sequences, based on these 9. Somewhere if each of these has a tie in the Johnson, you
could get more sequences, but if you have m machines we will end up doing m minus 1
sequences, and take the best out of that. So, in this case in a 3 machine case we are 2
sequences, and we took the best, if you are 8 machines then you would have 7 sequences
and you would take the best.
Now, this still a heuristic algorithm, now this heuristic algorithm is commonly called as a
CDS algorithm stands for Campbell Dudek and Smith who are the three authors, and this
came in 1977. And a very popular algorithm for use to find out the make span for a
general m machine flow shop sequencing problem. It is also generally based on
experimentation, it is observed that this algorithm gives better solutions than the
Palmer’s in general. Palmer’s would still evaluate only one sequence CDS would
evaluate m minus 1 sequences, and choose the best.
Now, let us look at one more heuristic algorithm, and then we to see whether we can get
a solution, which is better than 107. So, the goodness is once again 107 minus 103 by
103, so we have not made too much of an improvement for this particular example the
CDS gives the same value of the make span as that of the palmer.
Now, we look at one more heuristic very quickly and what we do here is that we start
492
with, let us find out the sum of processing time of each of the jobs, so 16 plus 18, 34 plus
12 46, 14 plus 10, 24 plus 11 35, 13 plus 20, 33 plus 15, 48, 19 plus 15, 34 plus 19, 53,
15 plus 16, 31 plus 16, 47. So, now we arrange the jobs in decreasing order of the total
processing time, decreasing would always mean non increasing order of this. So, the
highest is J 3 the next is J 5 I am sorry the highest is J 4, the highest is J 4 with 53, the
next is J 3 the 3'rd is J 5 the 4'th is J 1 and the 5'th is J 2.
Now, we take the first two of these which is J 4 and J 3, so take, so step 1 take J 4 and J
3, now J 4 and J 3 can be permuted in two ways, which is J 3, J 4 and J 4, J 3. It can be
permuted in two ways J 3 and J 4, when I say take the first two if we are taking J 4 and J
3 not necessarily in that order, J 4 and J 3 not J 3 following J 4. Say if we take only J 4
and J 3, there are 2 jobs and they can be permuted in two ways, now find out the make
span for the partial sequence J 3, J 4 and J 4, J 3 let us just do that.
So, for J 3, J 4 it will be 13, 33, 48, J 4 is 13 plus 19, 32, 33 plus 15, 48, 48 plus 19, 67
for J 4, J 3 it is 19 plus 15, 34, 34 plus 19 is 53, 13 plus 19, 32, 34 plus 20, 54, 54 plus 15
69. Now, this partial sequence has a make span of 67 this partial sequence has a make
span of 69, so choose this sequence, so now, choose J 3, J 4 step 2. So, step 2 choose J 3,
J 4, now take the next job here which is J 5, now J 5 can be squeezed in three ways J 5
can come here, J 5 can come here, J 5 can come here. So, you could have three sequences
which are J 5, J 3, J 4,- J 3, J 5, J 4,- J 3, J 4, J 5.
493
Now, find out the make span for all these 3 partial sequences like we did here, and
choose the best. So, one of them will be the best at the moment we are not going to show
the computation, but one of them will be the best, if for some reason let us say that this is
the best for the purpose of argument, after the computation. Then look at the next one J 1
and squeeze it in four positions it could come here, it could come here, it could come
here, it could come here.
So, then evaluate four more sequences choose the best, and take the last one and squeeze
it in five positions, and then evaluate the best out of these five. So, at the end we will
have five sequences and we will choose the best out of these five sequences.
Now, if we continue with the algorithm then we finally, get a solution 3, 4, 5, 1, 2, so we

finally, get a solution 3, 4, 5, 1, 2. Let us compute the make span for this.
494
So, we would have M 1, M 2, M 3, J 3, J 4, J 5, J 1 and J 2, now let us quickly compute

the make span. So, first 1 is J 3, so 13, 33, 48 followed by J 4 13 plus 19, 32 waits till 33
finishes at 48 goes at 48 finishes at 67. J 5 comes next 15. So, 32 plus 15, 47 waits till 48
takes another 16, 64, waits till 67 takes another 16 and finishes at 83. Then comes J 1
starts at 47 plus 16, 63 waits till 64, 64 plus 18, 82 waits till 83 takes another 12 and
finishes at 95. J 2 starts at 63 takes another 14 finishes at 77, waits till 82 takes another
10, 92 waits till 95 takes another 11 and finishes at 106.
495
So, the make span here is 106 which is better than the 107 that we got, so we have a
make span of 106. ((Refer Time: 54:12)) Now, this algorithm is an insertion algorithm
and this insertion algorithm for make span is called the NEH algorithm, stands for
Nawaz Enscore and Ham algorithm, which came in the year 1984. Now, this insertion
algorithm is found to work very well for make span minimization in a flow shop, and it
has been shown to give on an average slightly better results than the Campbell Dudek
and Smith algorithm.
Now, if you have m machines, and n jobs CDS will evaluate m minus 1 sequences and
will choose the best, NEH in the end will evaluate n sequences and choose the best, and
usually non flow shop problems n is greater than m. So, NEH algorithm would actually
evaluate more sequences, than the CDS algorithm either because of that or otherwise it is
able to perform slightly better than the CDS algorithm, when we consider the make span
minimization.
Now, that we have a solution with 106 then the goodness will become 106 minus 103 by
103 into 100 percent. Now, this is of the order of 3 percent whereas, with 107 it was of
the order of 4 slightly less than 3, this will be the slightly less than 4 in the earlier case
and slightly less than 3 in this case. So, if we are happy with the solution which is within
3 percent of the optimal, we can take the NEH and use it.
If not we proceed for other methods with the hope that we can get, we may get or we
wish to get a solution less than 1 naught 6. But, then it depends on where the optimum is
and for this particular problem instance, optimum is 106 and therefore, NEH has given us
the optimum, but then we do not have a way to know that. So, we would still say that
NEH has given as a solution, which is within 3 percent of the optimum. So, we have now
seen some heuristics to solve the make span minimization problem in a flow shop, earlier
in our discussion on sequencing and scheduling, we started with single machine flow
shop and then job shop. So, we will look at some aspects of scheduling in a job shop in
the next lecture.
496
Prof. G. Srinivisan
Lecture – 28
Job Shop Scheduling - Gantt Chart, Different Dispatching Rules
In the last lecture, we saw some heuristics for the floor shop scheduling problem. In this
lecture we will address the job shop scheduling problem. So, as we mentioned earlier, in
a job shop there are several jobs, let us say there are n jobs, and there are m machines.
Unlike a floor shop each job has a unique route or a unique order of visit of the
machines. For example, if there are 3 machines one particular job may come here, first
then visit this, then visit this and leave, while some other job may start with this, visit
this come back to this and then leave.
So, in a job shop each job has a pre specified route or order of visit of the machine, it is
also not absolutely necessary that all the jobs will visit all the machines. A job can visit a
subset of the existing set of machines. So, let us explain the job shop scheduling problem
using an example. So in this example, we are going to look at three jobs and three
machines. So, let me call the three jobs as J 1, J 2 and J 3, so let us consider this
particular example, where there are three jobs and three machines. Each job has it is own
pre specified route, which means, job 1 will first will visit machine 1, then visit machine
497
3, and then visit machine 2. So, like job one first visits machine 1 then visits machine 3
then visits machine 2 and goes out, another job J 2 may first visit machine 2, and then
machine 1, and then machine 3.
For example, J 2 will first visit machine 2, and then it visits machine 1. What I have
shown in yellow essentially stands for J 2, first visits machine 2 then visits machine 1,
then visits machine 3 and then leaves. Now, the number shown in the brackets are the
processing times for the particular job on the particular machine. Now, let us try and get
a make span for a particular schedule, and let us draw a schedule to solve this particular
problem.
So, what we do first is we make a chart where, we are now going to show all the 3
machines on this chart. So, let us say this is machine M 1, this is machine M 2, and this
would be machine M 3 on this chart. Now what we do is we let each job go and stand in
front of it is first machine. So, J 1 will go and stand in front of M 1, so J 1 will be here. J
2 will go and stand in front of M 2, so J 2 will be here. J 3 also will go and stand in front
of M 1. J 3 will be here.
Now, we are at time equal to 0. All the jobs are available and each job is now waiting or
standing in front of it is first machine to be processed. So, if we look at machine M 2 at
time equal to 0, job J 2 is standing in front of M 2 there is only one job waiting in front
of M 2. So, job J 2 can be loaded on M 2, so it starts processing on M 2 at time equal to
498
0, finishes at time equal to 6 because it requires 6 time units, and at time equal to 6 it will
leave M 2 and come to M 1 for further processing, which we show here as J 2 coming
here at time equal to 6.
Now, at time equal to 0, M 3 does not have any job in front of it, so M 3 will remain idle,
now we have to look at M 1 there are two jobs waiting in front of M 1 at time equal to 0.
We have to choose only one of them for processing, because a machine can do only job
at a time, now J 1 requires 7 units of time, J 3 requires 8 units of time. So, we would be
inclined ordinarily to choose J 3 ahead of J 1 or J 1 ahead of J 3 because J 1 has lesser
processing time compared to J 3.
So, we now follow a rule and say that rule choose job with Smallest Processing Time or
use SPT rule. We have already seen the SPT rule. So, let us assume that if we have to
choose between two jobs or among a set of jobs that are waiting, we would pick a job
which has the smallest processing time. So, between J 1 and J 3 here J 1 requires 7 units
of time, J 3 requires 8 units of time, so we choose J 1, so J 1 will start at 0 finish at time
equal to 7.
And at time equal to 7, we will now move to M 3, so it will come here. So J 1 at time
equal to 7. Now, we have to look at two issues at this point. Now in this particular
example we had two jobs waiting J 1 and J 3, and we decided to choose one amongst the
jobs waiting based on shortest processing time. So, when we apply this rule there is no
499
tie because J 1 requires 7 units, J 2 requires eight units clearly J 1 is chosen based on the
rule.
But, if we had a situation where this was also a 7, then two jobs would have been waiting
and both of them have the same processing time therefore, if we apply the rule there is a
tie. So, if there is tie, we need a tie breaking rule also, and we need to define a tie
breaking rule also. Now right now there is no tie, so the tie will happen when you apply
your first rule, and after applying the first rule you still have more than one candidate. In
which case we need to define a tie breaking rule and use the tie breaking rule to choose,
the jobs that are in the tie.
Second issue is this, at here a t equal to 0 we realize that J 1 and J 3 are waiting, J 1
requires 7, J 3 requires 8. So, we picked J 1 ahead of J 3 at the same time we also have an
additional information that J 2 is going to come at 6 here, but then one look at the table
we realize that J 2 actually requires only four units of time. So, the question is will we
leave out J 2 in the choice, and choose between J 1 and J 3 because they are waiting at
time equal to 0, J 2 has lesser processing time. But, in order to take J 2 ahead of J 1 or J 3
the machine will have to be idle for some more time.
So, we always make a decision, but we will not keep a machine idle if there are jobs
waiting in front of it. So, another rule is do not keep a machine idle when there are jobs
waiting in front of it therefore, we will not consider that J 2 is going to come after 6
minutes when we make a decision here at time equal to 0. So, at time equal to 0 we will
see what is available and then make a decision accordingly, now we go back and see that
up to time equal to 6 or up to time equal to 7, machine 1 is busy. Machine 2 is busy up to
time equal to 6, machine 3 does not have anything except that time equal to 7 it can start
J 1.
So, the clock shifts to time equal to 6 now, so at time equal to 6. M 1 is busy at time
equal to 6; M 2 is free, but there is no job waiting, at time equal to 6. M 3 is still idle, and
there is no job in front of it. So, the clock shifts to time equal to 7 where on M 1 we have
J 3 waiting we also have J 2 waiting, so J 3 requires 8 units of time, J 2 requires 4 units
of time. So, based on the shortest processing time rule, we would choose J 2 ahead of J 3
even though J 3 came earlier, the rule that we use consistently is to choose the one with
the smallest processing time.
500
So, we would chose J 2, so we start J 2 at 7 takes another 4 units, so at 11 J 2 is
completed. J 2 will go to M 3 at time equal to 11. Now, once again at time equal to 7, M
2 does not have any job in front of it. Now at time equal to 7 M 3 has J 1 in front of it, so
it starts J 1. So, J 1 on M 3 is another 8, so 7 plus 8 15, so J 1 finishes on M 3 at 15 and
goes to M 2 at 15, so J 1 comes at 15.
Now, the clock moves to time equal to time equal to 11. Only J 3 is waiting in front of M
1. So, we start J 3, J 3 requires 8 units, so 11 plus 8, 19, so at time equal to 19 J 3
completes on M 1 and comes to M 2. So at 19 J 3 comes here, now the clock moves to
time equal to 15 at time equal to 15, J 2 is available here. So, we take J 2 for processing J
2 on M 3 is another 12, so 15 plus 12, 27.
So, J 2 finishes at 27 on M 3, so J 2 completes at 27, so at time equal to 27 J 2 has

finished it is last job on M 3. Now, once again we are at time equal to 15, J 1 is available
here, and as I mentioned earlier we will not look at the presence of J 3 at 19 because we
follow this rule that we will not keep a machine idle waiting for something to happen.
So, at this point we will take J 1. So J 1 on M 2 is another 10, so 15 plus 10, 25. So J 1
has completed processing on M 2 at 25, so J 1 finishes at 25.
Now, once again we have J 3 has come here at 19, now we are at 25, J 3 is waiting on M
2 takes another 8, so 25 plus 8, 33. So J 3 finishes at 33 here. And then at 33, J 3 comes
to, so J 3 comes at 33, so at 33 J 3 is the only 1 available here. It takes another 7 units of
501
time. So, starts at 33 and finishes at 40. So J 3 finishes at 40, so all the jobs are
completed at time equal to 40, so the make span for this fixed schedule is 40. So now, we
can write that make span is 40, total completion time is 27 plus 25, 52 plus 40, 92.
Now, if each job has a due date and if the due date for the jobs are say 25, 30 and 35 then
this is completion time. So, this is tardy with tardiness equal to 2. This is early this is
tardy with tardiness equal to 5, so sigma T j total tardiness will be 7, number of tardy
jobs N j will be 2. Because, this is tardy, this is also tardy, maximum tardiness which is
usually called L max will be equal to 5 because the due date is 35 completion time is 40.
Now, if we want to see the utilization of these machines, so till time equal to 40 the shop
is busy, so this takes 19 out of 40. So, utilization of M 1 can be shown as 19 by 40,
utilization here would be the time taken on M 2 which is 10 plus 6, 16 plus 8, 24 by 40,
the utilization on M 3 is 8 plus 12, 20 plus 7, 27 by 40. If we want the idle time, this is
idle for 21 time units, this is idle for 16 time units, and this is idle for 13 time units.
So the moment we get a chart like this we will be able to evaluate all the objectives. The
objectives that we have evaluated are make span, total completion time or mean
completion time and for assumed due dates the total tardiness. We can also look at total
earliness. Only this job is early, so earliness is 5, total tardiness is 7 number of tardy jobs
2, number of early job 1, maximum tardiness is 5, utilizations, idle times. So, all possible
objectives we can evaluate once we draw this chart. Now this chart is called the Gantt
chart named after Henry Gantt who first proposed a chart or a pictorial representation of
a schedule.
502
So, the Gantt chart can be drawn for a given rule like the SPT rule. Now these rules on
which we draw the Gantt chart are these rules, using which we will choose one job out of
the available jobs for processing are called dispatching rules. It is very customary, and
very commonly used practice that we use dispatching rules to try and draw the Gantt
chart. Once the dispatching rule is decided, the Gantt chart will be the same if more than
one person works on the Gantt chart.
So, we could have we first define a dispatching rule, and then if necessary we have to
define a tie braking rule, as mentioned earlier. Right in the beginning if both J 1 and J 3
they had 7 and 8. If both of them had 7 and 7, then the shortest processing time rule
would not be able to get a unique job from the jobs waiting, so there will be a tie. Now,
those ties are broken by what are called tie breaking rules, a commonly used tie breaking
rule could be first in first out or lowest subscript, in this case J 1 would be seen as a job
with the lowest subscript and so on.
So, depending on the dispatching rule and the tie breaking rule, the Gantt chart is
defined, and Gantt chart is an extremely efficient and intuitive way to draw a schedule.
But, the Gantt this schedule need not be optimum with respect to make span or with
respect to any other performance measure, Gantt chart is essentially an evaluative tool
given a dispatching rule, we are now able to show pictorially how the schedules will
look like, with absolutely no guarantee for optimality.
503
If the dispatching rule changes, then this Gantt chart will also change. We also look at
when we used two things consistently. The first thing was the rule, which we now
defined as a dispatching rule, the second principle that we used consistently was that we
would not keep a machine idle, if there are jobs waiting front of it, with the hope that
some other job that comes later if that is processed. We may be able to optimize the
objectives. We will not follow that principle which we wrote here which we also used
here.
For example, we did not consider this J 2 when we made a decision to choose J 1. Here,
now schedules where this particular rule is used are called non delay schedules. We will
not deliberately create a delay, so that we can optimize on the performance measure for
that, and in general non delay schedules perform better than delay schedules where
delays are intentionally introduced. From a practical point of view non delay schedules
are important and are used.
Now, as mentioned earlier, depending on the dispatching rule the Gantt chart can be
created, and there are several dispatching rules that are available which one can use to
create Gantt charts. So, let us see some dispatching rules which are commonly practiced,
we also have categories of dispatching rules.
So, the first categories are processing time based, now the rule that we used to create this
Gantt chart is called shortest processing time rule. So, the first rule that is based on short
504
term processing time is the Shortest Processing Time rule also called SPT rule, now if we
can draw a chart using this SPT rule, we can also draw a chart using the LPT rule. For
example, we chose between J 1 and J 3 which was J 1 first here based on SPT, but based
on LPT, we could have chosen J 3 and the schedule would have looked very different.
So, if SPT is a good rule LPT by itself is not a bad rule. If SPT will try and give a good
make span, it does not mean that LPT will give very long make span, LPT in it is own
way will also try to get a good value of the make span. So, SPT is also a rule that is often
used, now we could consider instead of SPT we could have considered for example, we
have to choose between J 1 and J 3, now let me look at what is the total processing time
on J 1, 7 plus 8 15 plus 10, 25, 6 plus 4 10 plus 12, 22.
So, total processing time I am sorry 8 plus 8, 16 plus 7, 23, 8 plus 7, 15 plus 10, 25, so
total processing times are 25 and 23. So, if we use what is called STPT Shortest value of
Total Processing Time, we could have picked J 3 ahead of J 1, so STPT is another rule
that we can think of shortest total processing time. Similarly, if STPT is a rule then LTPT
can be another rule, choose the job which has the largest value of the total processing
time.
Many times, we also do another thing for example, let us say we were here, so in this
chart we are not able to encounter that situation except in the front. But, let us have a
situation where, somewhere in the middle let us assume that we are in a particular
machine. Let us say we are here, and a particular job has just finished processing, let us
say J 2 has just finished processing here, now let us assume that J 1 has come at this
point, and J 3 has come at this point.
Now, which means we have to apply a dispatching rule to choose between J 1 and J 3 to
choose one of them. Now, let us assume that in this case J 1 has a total processing time
of 25, but when J 1 has come here it has finished 7 and J 1 requires a remaining 18
processing time, including the machine that I have. Whereas, if I look at J 2 when it
comes there, let us say this is machine M 2, now in J 1 has come to M 2, J 1 has already
finished M 1.
And including M 2, J 1 has 18 units of time to go, when J 3 comes here on M 2. J 3 has a
total of 23 it has already finished 16 it has 7 to go including the current processing time.
505
So, this has 7 to go including the current processing time, this has 18 to go including the
current processing time, this has a total of 25, but let us assume it is waiting for M 3 after
it has finished M 1. So, it has finished 7. So the remaining processing time for J 1
including the current machine is 18 whereas, when J 3 has come to M 3 it has already
finished 16, so the remaining one including the current is 7.
So, you could choose a job based on Shortest Remaining Processing Time, where
remaining includes the current processing time. So, you could have SRPT and you could
have LRPT Longest Remaining Processing Time where remaining includes the current
processing. So, these are some examples of processing time based dispatching rules, now
other rules based on number of operations, in this particular example we have chosen
three jobs, and all three jobs visited all three machines.
It is not necessary for all the jobs to visit all the machines, now suppose we had a
situation where this one was not there. Then J 1 visits three machines, J 3 visits two
machines. So, J 1 requires three operations, J 3 requires two operations, so one could
choose J 3 based on minimum total operations, so you could have minimum total
operations minimum N o P Minimum Number of operations, you could use maximum
number of operations as two different dispatching rules.
Similarly, we could use minimum number of remaining operations for example, when we
went back to the earlier situation. When we are looking at visit of say J 1 and J 3 on M 3,
when J 1 visits M 3 it requires two more operations and whereas, when J 3 visits M 3 it
requires one more operation. So, you could look at minimum remaining number of
operations, and maximum remaining number of operations these are not processing time
based, but these are number of operations based methods.
Then we can look at due date based methods. One of course, is the earliest due date, the
second is called slack. Slack equal to due date minus current time minus remaining
processing time. For example, write here to choose between J 1 and J 3 current time is 0,
due date for J 1 is 30, so due date 30 minus current time 0, 30 minus remaining
processing time is 7 plus 8, 15 plus 10, 25. So, slack is 5 for J 1, now for J 3 due date is
35 current time is 0, total time is 23, so slack is 12.
So, based on minimum slack J 1 can be taken, slack represents the extra time that is
available, which acts like a slack or a buffer or a cushion to try and meet the due date.
506
So, that job which has the least slack will have to be taken first, so you could look at
least slack rule, another commonly used rule is called minimum slack by remaining
number of operations, I have a slack of 5 I may have only one more operation to go. I
may have another situation where I have a slack of 8, but I have two more operations to
go. So, 8 by 4 will be 2 or 8 by 2 will be 4 versus 5, so the other one can be taken.
So, we compute the slack and then we see how many more operations are to go, so
divide the slack by remaining number of operations, and then chose the minimum of
slack by remaining number of operations. Then we have a set of rules, which are very
common rules which are neither or none of the above one would say, neither processing
time based nor due date based general rules. So, first is First Come First Served or First
In First Out, they mean the same First In First Out, Last In First Out, the one that came
first will be taken the one that is came last will be taken.
So, first in first out last in first out one could use some kind of a random rule, just pick
the job randomly out of the jobs that are available. So, these are examples of popularly
used dispatching rules. Sometimes we could have other combination rules where we
could think in terms of some alpha into SPT plus some beta into LPT, where we could
have weights, we could have weighted combinations and so on. We could have priorities
we could have all of them.
So, a simple example would be, we could use alpha into remaining processing time plus
beta into remaining number of operations or beta into total processing time that can be a
combination rule. So, we could have rules like this. Now out of these, so many rules it is
also common practice to essentially look at SPT, first in first out, and EDD to single
them out as three extremely popular dispatching rules.
Now, these three rules are very common easily understandable, and easily implementable
when we know that certain number of jobs are here, it is very intuitive that we choose the
one which has the smallest processing time. Similarly, intuitively we choose the one that
came in first, and if we have an idea of the due date we would intuitively choose the job or
activity that has the earliest due date. So, when we really want to put this to use, the Gantt
chart based schedule generation to be put to use in a floor shop or in a shop floor.
507
The most convenient thing to do is to use one of these three rules, and then draw the
Gantt chart and make a schedule. So, if this schedule goes to an operator who is working
on a machine, say let us say we are looking at the operator who is working on say M 3.
Now, based on this schedule we will tell the operator that at time equal to 7, J 1 will
come put J 1 in, at time equal to 11, J 2 will come put J 2 as soon as you finish J 1. And
then J 3 will come later and put it here. Now in this particular schedule there is no
dispatching rule on this machine. Whereas, if you are looking at the operator on M 1, we
have to tell the operator that this is your schedule, even though J 1 and J 3 are coming
first you take J 1 based on SPT, then you will use J 2 even though it came later and then
you use J 3.
That is one way of explaining it or giving this, the other is to say when you have to make
a decision choose based on the one that has the shortest processing time. So, it is
extremely easy for the operator to implement, in practice it there could be situations
where, there can be slight delays. For example, J 1 that requires 7 units of time may not
actually reach M 3 at 7, it may reach M 3 at 8 or little later than 7 and so on, but then the
operator also will not be very specifically bound by these numbers.
But, the operator will be very specifically bound by the rule using which he or she can
pick a job and put it on the machine. So, when we use very common rules like SPT or
EDD or first in first out the, the Gantt chart becomes usable and any one can understand
508
the Gantt chart and appreciate the consistency with what is actually happening in
practice. But when we make these then this conflict comes as to the next question is, we
also mentioned that this particular Gantt chart based on the SPT rule, need not give us
the optimum solution all the time.
Now, we have listed about 6 plus 4, 10, 13, 16, 17 rules. There are actually hundreds of
rules that are there, and today with the amount of computerization and use of information
technology devices, it is possible to write a program that will generate the Gantt chart. It
is also possible for the user to define a dispatching rule, which is either one of those
listed here or another dispatching role. And it is easy to get the corresponding Gantt
chart very, very quickly.
Now the question is if SPT Gantt chart give us the solution with 40 and let us assume
that we are using slack by remaining number of operations as another rule or we are
using an 0.5 times RPT plus 0.5 times TPT as a dispatching rule, and we are able to get a
solution with 38 say, make span of 38. Then from an optimization point of view 38 is
better than 40, but from implementation point of view 40 would be a little easier to
implement on the shop floor. Because, it is easy to understand and use, so the question is
do we emphasize on advanced or more complicated or more involved rules.
If they can provide better make span or are we content and happy with commonly used
rules like SPT FIFO and EDD, because they are easy to implement and the person
implementing will not make a mistake in implementation. So, the answer to that comes
in the amount of training, and the amount of education that is provided to the people who
do the job. The people who do the job can understand and appreciate, the role of these
dispatching rules, and the fact that a slightly involved dispatching rule like minimum of
slack over remaining number of operations, can give a solution which is lesser than this.
And if people are aware educated and are ready to use, then it will be an advantage to use
involved dispatching rules, as long as the performance measure is optimized. But, if on
the other hand the organization is comfortable with the use of simpler and well known
dispatching rules like SPT or EDD or FIFO, which is more commonsensical than the
organization could use, such rules and try and try and schedule the jobs in the job shop
that is being looked at.
509
The other advantage as I have mentioned with the Gantt chart based scheduling, and the
Gantt chart based dispatching rule is that these are evaluative in nature, with a single
Gantt chart schedule, we can evaluate each and every one of the objectives. The positive
thing is the efficient pictorial representation, easy to understand, ability to evaluate a
variety of objectives. They are not, so positive thing it is the fact that, the it is simply
dependent on a particular rule, and it does not depend on a particular objective function.
Therefore, while it has the ability to evaluate all objective functions, it perhaps also has
the lack of ability to optimize a particular objective function. Because, it was based on a
particular objective function therefore, the ability to guarantee optimality is also very
less, and job shop scheduling is a very hard problem. If we look at now in this particular
example, there were 3 machines and each machine has all the three jobs visiting, so
every machine if all the n jobs visit every machine.
So, for each machine the n jobs can be arranged in n factorial ways, and since there are
m machines we can have a total of n factorial to the power m possible alternatives in
schedules. So, n factorial to the power m is a very large number and job shop scheduling
problems are known to be very hard problems, so though dispatching rule is a very
convenient way or method to solve job shop scheduling. It’s inability to give optimal
solutions is actually not very good.
510
But, over a period of time people have been using dispatching rules, to try and get Gantt
charts. And with more and more computerization happening, and perhaps more and more
dispatching rules available, one could program or use these dispatching rules into a
software. And then try out several alternative dispatching rules, to try and optimize the
performance measures, like a make span or a tardiness or whatever, and then use it
accordingly by training the people.
Now, we have also looked at one aspect of job shop scheduling which is called a static
job shop, a static job shop assumes that the number of jobs that we are going to do is
known. So, in this example there are only three jobs, so the planning period or the
planning horizon will end, as soon as all the three jobs are completed therefore, from a
Gantt chart schedule it is possible to compute a make span, because the point at which all
the three jobs are over will be the make span.
In real time job shops are not static in nature, jobs keep coming jobs keep arriving, so
when jobs keep arriving then we can make certain variance to it. Now, and we also made
an assumption here that all the jobs are available at time equal to 0, which is not a valid
assumption in practice. So, in practice jobs keep coming dynamically, so we have to
move from what are called static job shops to dyamic job shops, where the jobs can
come at any time.
511
And the time at which the jobs are going to enter the system is not known a priori. When
the time at which the jobs enter the system is known a priori, then it becomes static job
shop with release times, which is usually called r j, where r j is the release time for job j.
So in dynamic job shops, we do not know what time the jobs are going to be released
into the system. So, in dynamic job shops there is also no end to the jobs coming in, so
there is no make span concept in a dynamic job shop.
Dynamic job shops are essentially concerned with other measures like, flow time or
completion time, tardiness, number of tardy jobs and so on. Dynamic job shops do not
consider make span, we could also model job shops with the release times or no release
times which is called r j. So, this gives us a complete view of certain aspects of what is
called static job shop scheduling. Most static job shop scheduling is concerned with the
use of dispatching rules, based on which we draw Gantt charts and then we evaluate all
the objectives.
The limitation of course, is the lack of optimality or the inability to give the optimum
solution to either make span or any objective for a static problem. Static problem implies
number of jobs are known, the inability to give the optimum solution. Now, if we want to
focus on the optimum solution, then we need to look at other methods such as
mathematical programming and so on. And the problem also becomes very hard to try
and get to the optimal solution.
So, the next question that comes is while dispatching rules are very good to try to give
solutions, which are implementable, which are easy to understand. Are there other
heuristic methods or other methods, which may not be based on dispatching rules, but
can they also be used to try and get solutions, to the job shop scheduling problem. For
example, can we have other heuristic methods other than dispatching rule based method,
which may give us a value less than 40. One such method is called the shifting
bottleneck heuristic, which follows an entirely different approach to solving the job shop
scheduling problem. So, we will look at the shifting bottleneck heuristic in the next
lecture.
512
Prof. G. Srinivasan
Lecture - 29
Job Shop Scheduling – Shifting Bottleneck Heuristic
In today’s lecture, we will look at the Shifting Bottleneck Heuristic, to try and minimize
the Makespan in the job shop scheduling scenario. In the previous lecture, we had found
a solution using the SPT rule, which gave us a Makespan of 40.
513
Now, we will apply the shifting bottleneck heuristic to the same example, to see whether
we can get a better solution or to understand, how the shifting bottleneck heuristic
works. Now, in order to understand the shifting bottleneck heuristic, let us first draw a
network, which represents the job shop scheduling problem.
Now, let us look at this network, this network captures a part of the information that is
given here. Now, there is a start and there is a finish. Instead of looking at the job shop
scheduling problem through a Gantt chart, which is what we did in the previous lecture,
514
let us try and look at it through this network. So, there are three jobs to be done, and
these three jobs represent three paths in this network, so this would represent the job 1
first visits machine 1, if we see the route of job 1 it is machine 1, machine 3 and machine
2.
So, job 1 on machine 1 followed by machine 3 followed by machine 2. So if we trace the

path of job 1, then the arrow will have to be like this. Similarly, job 2 first goes to
machine 2, machine 1 and machine 3. So job 2 first goes to machine 2 machine 1 and
machine 3, and therefore the directions for job 2, flow of job 2 will be like this, and
similarly for job 3 it will be like this.
We also have a certain processing time associated with each node, and each node
represents an operation which is the visit of a job on a machine. So, if a node has a i
comma j, then j is the job and i is the machine, so 1, 1 has value 7. This has 8, this has 10
6, 4 and 12, and this is 8, 8 and 7. So, left to itself, each job will go through the
machines. The order of visit of these machines are captured, but then we also have a
condition that a machine can process only one job at a time. So, if we take a particular
machine, now the machine figures here, machine 1 figures here as well as machine 1
figures here.
So, machine 1 would now take the jobs according to a particular sequence, and
depending on that the time taken will change. So, if we take machine 1 and show this we
could have for example, between jobs 1 and 2 we could have something like this, where
at the moment we do not know, whether job 1 is going to go first on machine 1 or job 2
is going to go first on machine 1. So, if job 1 goes first, and then job 2 goes on machine 1
then something like job 1 will go first then job 2 will go on machine 1, and the flow will
continue.
If job 2 is going first and job 1 is coming later, it is like saying job 2 will do it first, then
the arrow will move towards job 1 and then it will proceed. So, at the moment we do not
know that if we consider these two jobs on this particular machine, we do not know the
order in which they are going to be processed by the machine. And depending on the
order we know that one of these two will be active, while the other will not be active,
similarly we have to consider for machine 1 jobs 2 and 3, and also for machine 1 jobs 1
and 3.
515
So, there are actually for every pair of jobs there are two arcs, one of them will finally be
active depending on the solution, similarly we have to do for machine 2. So, machine 2
is here machine 2 is here, and machine 2 is here, so we would have something like this,
something like this and something like this, depending on the order in which the machine
is going to process the jobs.
As I mentioned, for every pair in the final solution, only one of the arcs will be active,
now similarly for machine 3 there are, machine 3 figures in 3 different places here, so we
will have one set like this, we will have one set like this, and we will have the third set
like this. So, this completes the network, where these type of arcs will not change the
directions will not change, and in all the places where we have two arcs, depending on
the solution only one of them will be active in the final solution.
Now, if we draw such a network, and our decision now is to find out if we take machine
1 out of these 3 sets of arcs, 6 arcs that we have written, which of them are going to be
active and which of them are not going to be active. To put it differently, when we
decide which of them are going to be active, and which of them are not going to be
active, we automatically are deciding the order or sequence in which this particular
machine is going to take up the jobs.
So, if we are able to get the set of active arcs out of these bunched arcs, in an intelligent
manner such that the longest path is reduced, then the Makespan is automatically
reduced, the longest path on this arc indicates the Makespan of a given feasible schedule.
Let us explain that aspect by first looking at the SPT solution showing, how we are able
to capture that 40 on this network, and then showing how the longest path represents the
Makespan of a given schedule.
516
So, we have already seen this Gantt chart or bar chart in a previous lecture, where the
Makespan is 40. More importantly for the SPT schedule, M 1 is taking the jobs in the
sequence J 1, J 2 and J 3. So, let us try and match that here were M 1 is taking them in
order J 1, J 2 and J 3, so M 1 is first taking it in this order, so this is done first, then this
is done, then this is done, so M 1 is taking first J 1 then J 2 then J 3.
So, let me just show this with a different color, and put a dash here to indicate that out of
these two arcs that are connecting these two nodes, this arc this particular arc with an
arrow mark here, indicates that J 1 is done first and then J 2 is done on M 1. Similarly,
since 3 is done after 2 between these two arcs, now this arc will be active the arrow mark
here will be active, which is now shown by this dotted arc.
So, for the SPT sequence on M 1, the two dotted arcs are now replacing this pair, this
pair, and this pair as far as this pair is concerned both of them are inactive, because we
will not put 1 to 3 the sequence in, which it happens is J 1, and then J 2, and then J 3. So,
automatically when we put one arc here and another arc here it takes care of the fact that
3 is done after one and the route moves like this. So, both these will not be used, so only
2 out of the 6 will become active according to a feasible schedule.
517
Now, let us look at M 2 on the SPT chart and realize that the order is J 2, J 1 and then J
3. So on M 2 first it is doing J 2, then it is doing J 1 and then it is doing J 3. So, let us
replace this with this arc here, and then this arc, which is this. Similarly on M 3 it is
again J 1, J 2 and J 3. So, on M 3 it is first J 1, then it is J 2, and then it is J 3, so the order
will be and then J 3. So, now, this network has become slightly clumsy, because I have
overwritten the colored arrows, but then we have to find out the longest path on this
network, considering only the colored arrows as well as these kind of arrows.
So, to do that quickly, let us draw this network again with the relevant arcs and then try
to find out the longest path, so we have s 1 1, 3 - 1, 2 1, F - 2 2, 1 2, 3 2, 1 3, 2 3, 3 3 and
F. Plus of course, 1 1 to 1 2, and then 1 3, then 3, 1, 2, 3, so 3, 1, 3, 2, 3, 3 and as far as 2
is concerned 2 1, 2 3, 2 2, 2 1 and 2 3, so 2 2, 2 1, 2 3. So, this is the arc that we have,
now we have to find out the longest path on this, we also have to write down the process
times 7, 8, 10, 6, 4, 12, 8, 8, and 7.
So, we will now start with from here to here, so this node the value is 7, now this node
the value is 6, now the longest path to reach this node is 7 plus 4, and then you come
down here, so the value is 11. And if we see very carefully these are exactly the values
of 7 and 11, which represents the start times, so 0 is here, so you realize 0, 7 and 11 plus
another 8 of course, will happen, which is 19. Then we have to move to this node, this is
518
15, 7 plus 8, 15, which also we can see in that chart, now we come here, so this is 11, 7
plus 4, 11.
Now, we have to write in such a manner, that we take it after we include this, so this will
become 7 plus 4, 11. 11 plus this 8 will become 19. So, now, we have come up to this
then in order to look at this, this will be 15 plus 10 25, this will be 6 plus 10 16, so we
will look at this at 25. Now, we have evaluated this node, so this is 15 plus 12 27, now
this will be 25 plus 8 33, and this will become 33 plus 7 40, and this is 0, so this will
become 40.
So, now we realize that the longest path on this network has a length equal to 40, which
happens to be the Makespan of the sequence, so we have essentially mapped the SPT
solution, on to this arc to try and get the Makespan corresponding to the SPT solution.
That is not what we intend to do, what we have tried to show is given a Gantt chart, from
which the order of visit of the jobs can be ascertained, we have now mapped the Gantt
chart on to this and chosen the relevant arcs corresponding to a given solution, and we
have evaluated the Makespan to show both of these are the same, but that is not what we
intend to do.
What we intend to do, is to try and find out what is the best way that we choose the 2 out
of these 6. SPT made us choose this and this, now is there a another way by which we
can choose, another set of 2 arcs out of this.
Similarly, another set of 2 arcs on the yellow and another 2 arcs on the other, which is
machine 2, so that at the end of it the longest path is less than 40. So, if we are able to get
the longest path as less than 40, then we have achieved a solution that has lesser
makespan than the solution that has been shown here. The other learning also is that the
important thing though is the computation of the makespan, and try to minimize the
makespan, but the decision that has to be taken.
What SPT rule did is to try and find out a certain order in which the jobs are going to be
processed on the machines, for every such given order there will be an makespan. So,
what we try to do is we try to get that order by looking at these arcs, and try to find out
what is a best arc that we use, each arc that we use represents an order in which the jobs
are going to be processed on the machines.
519
So, let us now try and see now having understood this path, let us now try and see is
there a best way or what is the way that we are going to adopt to choose the best 2 out of
the 3. So, what I will do now is I will remove these, because these belong to SPT, so now
we have to find out, if we take machine 1 out of these 6 arcs which 2 we are going to
choose. Similarly, for machine 2 and similarly for machine 3, and once we have chosen
the longest path will automatically give us a makespan. Now, in order to understand what
is the best way to choose, these arcs we try and look at another sub problem, which I will
explain. Now, and then use results from that sub problem to determine, what is the
correct set of arcs that we need to choose, so let us go to the other sub problem.
And this sub problem is called 1 r j L max problem, it means 1 stands for a single
machine, r j stands for release times of jobs, and L max stands for maximum tardiness.
So, the sub problem that we will look at is a single machine problem with release times,
it means all time are not available at time all jobs are not available at time equal to 0, but
the times in which the jobs are available is known. So, r j is the time at which job j is
available, and L max is to minimize the maximum tardiness on this.
So, let us take an example, try to understand the 1 r j L max problem, how do we solve
that and then see how we use results from 1 r j L max, and superimpose it on this
network, so that we are able to get the correct set of arcs, that would try and give lesser
makespan. So, let us take this example for 1 r j L max and let us look at problem like
520
this, where there are 4 jobs on a single machine. The processing times are 4, 2, 6, and 5,
the release times are 0, 1, 3, and 5 and the due dates are 8, 12, 11, and 10, so if we
consider a sequence 1, 2, 3, 4, then the completion times are 4 the job can enter at time 0.
So, it can come out at 4 by which time job 2 is available, so 4 plus 2 6 job 3 is available
once again at 3, so 6 plus 6, 12 and job 4 is available from 5, so 12 plus 5 17, these are
the completion times 4 plus 2 6 plus 6, 12 plus 5, 17. We have not encountered a
situation, where we have to the machine has to wait for the job arrival. That has not
happened in this particular sequence.
If we consider a sequence 2, 4, 1, 3 job 2 starts at 1, so finishes at 3, because it arrives at

1 takes 2 more units finishes at 3, 4 arrives at 5 takes 5 more units finishes at 10, 1 is
available at 0. So, 10 plus 4 14, 3 is available at 3 14 plus 6 20, so these will be the
completion times. Now, the due dates will be 12 here, for 4 it is 10, for 1 it is 8, for 3 it is
11, so these are completion times these are due dates, so this job is early, this job is
neither early nor tardy, this job is tardy with tardiness equal to 6, this job is tardy with
tardiness equal to 9. So, maximum tardiness is L max, so L max is equal to 9, there are 4
jobs in the worst case you can have 4 factorial sequences, so L max associated with the
sequence 2, 4, 1, 3 is 9 what is the sequence that minimizes the L max is the problem.
521
So, now let us look at this problem and then go back and see what we do, so let us apply
a branch and bound algorithm to do that. So, let us start here with the starting tree, and
then let us say the job 1 is the first job, job 2 is the second first job, job 3 as first job, and
job 4 as first job. Now, when we take job 1 as first job, we want to find out a lower
bound on L max, so this can be written as 1 dash dash dash or 1 star star star, where 3
other jobs will have to be now put in a certain sequence. So, we now start with job 1 as
the first job, so job 1 is ready at time equal to 4.
522
So, job 1 completion time is 4 due date is 8, now let us try to arrange the rest of the jobs
using an earliest due date sequence. So, if we look at the rest of the jobs using an earliest
due date sequence, then the tentative sequence will be 4, then 3, and then 2, that is a
tentative sequence, if we use earliest due date for the remaining jobs. Now, the tentative
completion times tentative completion times will be machine is available at 4, but this is
going to be there it is going to come only at 5.
So, if we want to find out the completion time, then the completion time will become 5
plus 5 10, but we are interested in finding a lower bound to L max, therefore we will use
what is called a preemptive EDD rule. Now, in the actual sequence, if we want to find
the actual sequence for 4, it will start at 5 and finish at 10, and the machine is idle for that
one time unit, because it is over at 4, but job 4 is available only at 5. Now, that additional
1, it can have an impact in increasing the maximum tardiness, but in a preemptive EDD
rule, what we assume is that we allow job preemption.
So, we now see that if there is going to be a delay, because of a mismatch between this
completion time and this arrival time, we look at the next available job or tentative job in
the sequence, which is 3. Now, 3 is available at 3, so it is like saying I will do one minute
of processing of job 3, take it out at time equal to 4, then I will enter this one and so this
one will, now start at 5, and finish at 10, so completion time here will be 10.
Now, for 3 already I have done 1 unit of processing time, so now 3 will now go back and
complete the remaining 5 units of processing time, finish at 15 by which time job 2 is
available and then this will be 17. If on the other hand I had not used a preemptive EDD
rule then this would still have been 10, but this would have been 16 and this would have
been 18. So, preemptive EDD rule allows me to try to minimize my completion time
only for the purpose of the lower bound computation.
When I am actually evaluating L max for a given sequence. Then I will not use a
preemptive rule, because one of my assumptions is that job preemption is not allowed,
only for the purpose of computing a lower bound, I can use a preemptive EDD rule and
allow preemption. So, that the lower estimate of L max is completed, a lower bound is a
lower estimate of L max is completed, now due dates are 8, 10, 11, and 12, so this has
maximum tardiness, so lower bound on L max is equal to 5.
523
Now, let me explain the computation of the other one, now we start with job 2 as first.
Here, we have to fix you cannot for the fixed jobs, you cannot apply the preemptive
EDD, you can apply the preemptive EDD, only for the jobs that are not fixed in the
partial sequence. So, 2 is fixed in the partial sequence, so 2 starts at n equal to 1 and
finishes at time equal to 3. The next one will be 1, the next 3 jobs will be 1 3, 1 4 and 3
will be the next 3 jobs, which are not shown in white, at time equal to 3 jobs is available,
524
so this will be finished at 7, at 7 job 4 is available, so this will be at 12, and at 12 job 3 is
available this will be at 18.
Now, the due dates are 12 for this 8 10 and 11, so this is going to contribute to L max, so
L max is equal to 7, so lower bound is equal to 7. Now, here we have to look at can job 3
be the first job in the sequence, now if we see this problem very carefully, if job 3 is the
first job in the sequence. Then definitely job 3 can start at 3 finish at 3 plus 6 equal to 9,
and for 3 minutes machine has to be idle, but then in those 3 minutes idle, if I could
easily have put job 2 in, and taken it out.
So, typically is there is a job here, as the first job 3 and there is some other job, which
can be completed before this job can be started as the first job. Then the L max
associated with this will always be higher than the L max associated with this, because
this can be done ahead. So, see carefully if I have to do job 3 first, I have to wait for time
equal to 3, so 3 minutes the machine is idle, and then after that I have to do 2, but
common sense tells me that in those 3 minutes I can finish 2, because 2 starts at 1 and
finishes at 3.
So, if there is a job like J 3, whose start is going to be delayed and in that delay, if I can
finish some other job completely, then the lower bound here will always be more than
the lower bound here. So, I do not calculate this at the moment I simply keep this vacant,
because this is enough for me, so I do not do this. Similarly, for 4, because for 4 I have to
wait for 5 more minute, in which time I could finish as well as I could have finished this,
so I will not do for a 4.
525
So, now, I branch again on here, I branch with 1 2, 1 3, and I can actually branch on 1 4,
so I could branch on 1 4 also, so let me just keep 1 4 also here. Now, we can continue
this idea of the preemptive EDD, let me just explain the preemptive EDD for 1 2.
So, I have 1 and 2 which are fixed, so I completion time is I start at 0, I finish at 4 by
which time this is available. So, 4 plus 2 6 now based on a preemptive EDD I will do 4
and 3, because the remaining due dates are 11 and 10, so I will do 4 and 3, now at time
526
equal to 6 this is available, so this will become 11. And then this will become 17, so the
due dates are for 1 and 2, 8, 12, 10, and 11, so lower bound becomes 6.
Now, if I do 1 and 3, which is this node, so at time equal to 4 this is available, so 4 plus 6
10, now these are the 2 remaining things, so I have 4 and 2. So, at 10 job 4 is available,
so this will become 10 plus 5, 15, at 15 job 2 is available, so this will become 17. Now,
the due dates are 8, 11, 10 and 12, so L max remains at 5, lower bound remains at 5.
Now, there is actually a way by which we need not do 1 by 4, but let us try and do 1 4,
and understand may be we need not do it.
527
So, we look at 1 4 here, now when we do 1 4 this is at time equal to 4, so this has to wait
for another 1 minute, so takes it at 5 and finishes at 10. Then by EDD rule, I have 3 and
2, so at 10 this is available, so 16 and 18, now the due dates are 8, 10, 11 and 12, 11 for 3
12 for 2, L max is here, so L max is 6, so lower bound is equal to 6.
It is actually possible to show that we need not calculate this, because we have a 5 and
this is going to result in at least one more, but in any case we have calculated this, so we
keep this lower bound equal to 6. Now, we can proceed from here, and create 2 branches,
528
so I have 1 3 2 4 and 1 3 4 2, these are the 2 sequences that I can think of, because there
are only 2 more jobs remaining, so these are feasible sequences.
So, in general one would say even here we need not have evaluated this lower bound, we
could have straight away said at 1 3, there are only 2 more jobs remaining, which can be
fixed in 2 ways, so we could create 2 more sequences. Now, remember these are full
sequences, so we will not use the preemptive EDD rule, we will simply evaluate the L
max for these 2 sequences, assuming that job preemption is not allowed.
So, let us evaluate L max for these 2 sequences, so the first sequence is 1 3 2 4, so one
completion time is 4, 3 is available 4 plus 6 10, at 10 2 is available 10 plus 2 12, and at
12 4 is available 12 plus 5 17. Now, the due dates are 8, for 3 it is 11, for 2 it is 12, for 4
it is 10, so L max is equal to 7 for this sequence.
529
Now, let us look at the other one 1 3 4 2, so at time equal to 10, 4 is available, 15 at time
equal to 15, 2 is available 15 plus 2 is 17, and the due dates are 10 and 12. So, L max is
5, L max is 5, L is 5, so L max is 5. So, we have L max equal to 5, and this is optimal or
optimum, because with this 5, I can this is fathom by feasibility we cannot proceed. We
have a feasible solution. This is fathomed by bound, we have already seen this idea in the
branch and bound for flow shop sequencing. So here the lower bound is higher than the
current best upper bound.
So, we need not proceed and in any case these 2 we will not do, therefore we can go
back and say that 1 3 4 2 with L max equal to 5 is indeed optimum. So, this is a branch
and bound algorithm to solve the 1 r j L max problem optimally, but then 1 r j L max
problem is a hard problem. It is a difficult problem, and may in the worst case involve
computation of all the n factorial possible sequences.
530
Now, we took an example with 4 jobs, so with these 4 jobs we may encounter a situation
not in this case, but if there are n jobs, these 4 jobs could give us 4 factorial possible
sequences, but then for another problem with n jobs. We may encounter a situation
where we would be evaluating all the n factorial possible sequences, so 1 r j L max with
the branch and bound can take us through an exponential number of computations, and it
is a hard problem.
But, then using the preemptive EDD based lower bound which we demonstrated here,
and also applying certain dominance conditions, which we explained through this
example. The dominance condition ensured that we need not evaluate this or we need not
evaluate this, simply because by the time this job is ready as the first job, some other job
can be completed entirely, based on such a dominant rule, the branch and bound
algorithm is found to perform very efficiently, even for a reasonable job size.
For example, with job sizes are like 30 or 40 or 50, one can write a very efficient code,
which will solve this problem within a reasonable amount of computation time, when we
do experiments or when we do an empirical style, though in the worst case 1 r j L max is
harder. So, we still use this branch and bound algorithm to solve 1 r j L max, and try to
get the optimum solution which shows the, which minimizes the maximum tardiness, on
a single machine with release times.
531
Now, we have seen the 1 r j L max problem, and now we have to go back and use this 1 r
j L max on the job shop scheduling problem, so that we try and get a good schedule at
the end of the application of the shifting bottleneck heuristic.
Now we will do that right now, so let us come back to this network, and then we
understand that in this network all these arcs not going to change. But, wherever we have
these paths for example, 1 1, 1 2 and 1 3 we said that out of these 6 arcs that are shown 2
of them will be in the solution, and which 2 will have to be decided.
Now, let us look at this problem first ((Refer Time: 43:52)) and let us look at the load for
each of these jobs, which means if I start looking at this path to completion, temporarily
ignoring the these kind of arcs. If we start looking at only these paths, now this path is 7
plus 8, 15 plus 10, 25, which represents the total processing time for job 1 which is a
simple addition of these 3.
532
So, job 1 requires 25 time units, job 2 requires 22 time units, job 3 requires 23 time units.
So automatically the lower bound on the Makespan is the maximum of the job
processing times, so 25 is a lower bound to the Makespan. Let us also try and look at
machine times on the 3 machines that are required, so M 1 requires 7 plus 4, 11 plus 8,
19, M 2 requires 10 plus 6, 16 plus 8, 24, and M 3 requires 8 plus 12, 20 plus 7, 27. So M
3 with a load of 27 represents a slightly tighter lower bound on the Makespan. Now, we
can say that since M 3 alone requires 27; obviously, the Makespan has to be greater or
equal to 27, so we start with a lower bound on the Makespan as 27.
533
So, lower bound on the Makespan was 27, which is for M 3, and then we identify M 3 as
the first bottleneck machine, because it has the highest load amongst the 3 machines.
Now, we apply the 1 r j L max algorithm on machine M 3, and try to look at the 4 jobs
with machine M 3 as the bottleneck machine. So, let us now write the 1 r j L max data
for machine M 3, which J 1, J 2, J 3,- J 3, because we have only 3 jobs.
So, we have to write p j processing time on machine 3, r j release time or time at which
the job is available, and d j which is the due date. Now, writing the processing times is
easy, because we can see from here on M 3, J 1 requires 8, on M 3 J 2 requires 12, and
on M 3 J 3 requires 7. Now, let us go back and find out the r j, which is the release time
of the job or the earliest one can expect the job to be made available for M 3. So, if we
look at J 1, J 1 earliest it can be made available for M 3 is 7, because it has to be finish
one M 1, only then it has to come on M 3.
So, earliest release time for J 1 on M 3 is 7, earliest release time for J 2 on M 3 is 10, 6
plus 4, and earliest on J 3 again 8 plus 8 which is 16. Now, let us find out the due dates,
if we assume that the Makespan is 27, which is the lower bound on it. Now, M 3 is the
last machine for job J 2, so the due date job J 2 will be 27, which is the lower bound on
the Makespan, similarly due date for this will also be 27, but now on M 1, M 3 is not the
last machine.
So, if the Makespan has to be 27, then the due date for M 3 has to be 17, because it
requires another 10 unit of time at least, so the maximum you can stretch the due date
will be 17 for J 1. So, now, our data is complete when we have to do the 1 r j L max on
this, so let us do show the computation of 1 r j L max considering these 3 jobs on M 3.
534
So, we again start the branch and bound tree by saying, job 1 is the first one job 2 is the
first one job 3 is the first job, so let us try and find out the L max for this.
So, this will be 1 x x, but based on due dates it can be 1 2 3 or it can be 1 3 2, so let us
take the case 1 2 3. Now, completion times, now this is available on 10 at time equal to
7, so 7 plus 8, 15. Now at 15 J 2 is available, 15 plus 12, 27, and at 27 J 3 is available 27
plus 7, 34. Now, due date is 17 here, for 2 also 27, for 3 also 27, there is only one l, so L
max is 7, so lower bound is 7. In fact, if we have done 1 3 and 2 we actually realize that
535
this 15, but at 15 this is not available, so L max would go up only, so this is better than
the other one, since we are using a lower bound let us look at 7 as the lower bound.
Now, we look at 2 as the first job, so 2 r j is 10, completion time is 22 then based on due
dates it is 1 and 3, so at 22 it is available, 22 plus 8, 30, 30 plus 7, 37. Now, due dates are
17, 27, 17 and 27, so this is 10, this is 13, so lower bound is 13.
536
Now, let us look at 3 1 and 2 based on due dates 1 and 2, now completion time this is
available at 16, so 23, so at 23 this is available 23 plus 8, 31 plus 12, 43. Due dates are
27, 17, 27, so this is 16, now L max is 16.
So, lower bound is 16, so we have to proceed from here, so 2 as the first, 3 as the first, so
it gives us a sequence 1 2 3, 1 3 2 as the 2 possible sequences, so we will quickly
evaluate the L max for each of these. So, for 1 2 and 3 completion times are 7 plus 8, 15,
at 15 this is available 15 plus 12, 27, 34 due dates are 16, 27, 27, so L max is equal to 7.
Now, for 1 3 and 2, so one is available at 7, 7 plus 8 15, this is available at 16, so 16 plus
7, 23, and then at 23 is available 23 plus 12, 35, due dates are 16, 27, 27, so L max is
equal to 8.
So, with L max is equal to 7, actually we could have fathomed it, because with L max
equal to 7, we can fathom this, we can fathom this. And since, it is a branch and bound
algorithm, the value that you can get here will be greater than or equal to 7. And since,
we have a solution with 7 we need not have evaluated this also, but we know that L max
equal to 7 is optimal for this. So, now, we could go and say, that if we look at M 3 based
on 1 r j L max, we realize that J 1 J 2 J 3 is the optimum sequence using the branch and
bound algorithm.
537
And therefore, now we can go back to M 3, machine 3 is here, machine 3 is here and
machine 3 is here. Now, we say that we are going to do J 1, J 2 and J 3, therefore we
replace these 2 sets of arcs by one arc which is this J 1 to J 2. We replace these 2 arcs
with one arc here, and then we leave out these 2 arcs, because they are not used at all, so
we have this. So, now, we have identified for M 3 what is the correct set of arcs, if we
now go and find out the longest path on this network, the longest path on this network
would be 27 plus 7 which is 34, because the 7 will be added to the longest path on this.
538
Now we have two more machines machine 1 and machine 2, machine 1 with a load of 7
plus 4, 11 plus 8, 19, and machine 2 with a load of 6 plus 10, 16 plus 8, 24. Now,
machine 2 becomes the next bottleneck with 24, and then we solve a 1 r j L max on
machine 2 as we did here. And then we try to a with Makespan updated to 34, because
the longest path is 34, when we did for M 3 we got for 27, now we have updated the
Makespan to 34.
So, we solve a 1 r j L max on M 2 with Makespan equal to 34, and then we suitably
define the r j’s and the d j’s for M 2 and so on.
539
Then the Makespan can increase, may increase if, so keep that and then solve it again for
M 1 to get the best route for M 1. So, in this we will get the best route of jobs for M 2, as
well as the best route of jobs for M 1 which means we would have chosen the best out of
these set as well as the best out of the other set.
And once, we complete for all the machines, we can now find out the longest path, which
will be a feasible solution, which will give us the Makespan. Now, the computations on a
540
machines M 2 and M 1 and application of this branch and bound idea on them. To
further continue on this problem, we will see in the next lecture.
541
Prof. G. Srinivasan
Lecture - 30
Job Shop Scheduling – Shifting Bottleneck Heuristic. Line Balancing
We continue the discussion on the Shifting Bottleneck Heuristic. Let me begin with a
quick recap of what we saw in the earlier lecture, so we are trying to solve the problem
of minimizing Makespan in a job shop, we have already seen dispatching rule based
solutions.
542
For example, we have seen this Gantt chart, which has been prepared with shortest
processing time as the dispatching rule, so if we use shortest processing time as the
dispatching rule for this problem instance, we get a Makespan of 40. We are now trying
to see another approach, which helps us to solve the Makespan minimization problem on
the job shop, we also said that in a job shop as given in the example problem, each job
has a definite route, so J 1 will go through M 1 M 3 M 2 and so on.
543
In this example all the jobs visit all the machines, so a feasible schedule will require that
each machine, now treats the job visit in a certain order. For example, if we go back to
the SPT solution, in this solution the jobs visit M 1 in the order J 1 J 2 J 3, M 2 in the
order J 2 J 1 J 3, and M 3 in the order J 1 J 2 J 3. So, one other way of describing a
schedule is to determine the order in which the jobs are going to visit the machines, the
in any case if we see here job 1 has to visit all the 3 machines, but we are interested in
taking each machine and the order in which the jobs come in.
That is the variable part, which we are actually try into optimize, if you are able to get
the correct order, in the sense at the moment if the order on M 1 is J 1 J 2 J 3, order on M
2 is J 2 J 1 J 3, an order on M 3 is J 1 J 2 J 3 the Makespan is 40. So, the question can be
posed differently as what is the order such that this Makespan is minimized, also we have
to ensure that this order is satisfied, that J 1 will first visit M 1, and then M 3, and then M
2, which has been satisfied by the Gantt chart schedule. We also now have a network
representation, which we have seen now the definite orders for the jobs, for example, J 1
will visit M 1 M 3 M 2 is Shown here as J 1.
Here, i j j represents the job, and i represents the machine, so J 1 first visits M 1 then M 3
and then M 2 and finishes, so this part has been shown through these fixed arrows in this
network and through dark lines here. What are shown through dotted lines is the decision
that we have to make, for example since a machine can process only one job at a time.
544
So, if we take machine 1, then machine 1 figures here for J 1, here for J 2, and here for J
3, so what we do is we draw 2 arcs, which are dotted, joining 1 to this in this direction
and the other in the other direction.
So, for every pair we have that, so since we have 3 jobs, we actually have 1 to 2, 1 to 3
and 2 to 3 each has 2 arcs. So, we have 6 dotted arcs, the ones related to machine 2 are
shown in yellow, and the ones related to machine 3 are shown in a kind of a dark
brownish pink color. So, we now have to find out, if we take out of if we take machine 1
alone, out of these 6 arcs, we actually find in the end only 2 arcs will be active, which 2
are active is what we have to find out.
If we want to map the SPT sequence on this, and for example if we take M 2, then the
order is J 1 J 2 J 1 and J 3, which means on M 2 it will be 2 to 1 and 1 to three. So, on M
2 it will be first 2 to 1, which means 2 to 1 and then 1 to 3, so out of these 6 arcs which
are in yellow, the 2 to 1 here 2 to 1 and then 1 to 3 will be active. Now, we wish to find
out for each machine, which will be the active arcs out of the dotted arcs. So, we set out
to do that, and then we also related this aspect to solving a 1 r j L max problem on the
bottleneck machine, we also identified first M 3 as the first bottleneck machine by first
calculating the loads on each of these machines.
So, M 1 has a load of 7 plus 4, 11 plus 8 19, M 2 has a load of 10 plus 6, 16 plus 8, 24,
and M 3 has a load of 8 plus 12, 20 plus 7, 27. So, we took M 3 as the first bottleneck
545
machine, and then we solved a 1 r j L max problem, which means problem of
minimizing the maximum tardiness on a single machine with release times.
So, 1 r j L max on M 3, so we solved a 1 r j L max problem, where one represents single

processor, r j means released times for job j, and L max means maximum tardiness that
we wish to minimize. So, we solved this problem on M 3 and then we got the sequence J
1 J 2 J 3 which minimized this, so we got the sequence J 1 J 2 J 3 on M 3 with L max
equal to 7. Now, we have to put the solution or superimpose the solution on this network,
so we have solved for the third machine, so we have solved for this set of 6 arcs. And
then we have J 1 J 2 J 3 coming in, which means on M 3 we will first do J 1, and then we
will do J 2, and then we will do J 3.
546
So, on M 3 we will first do J 1, so this is 3 1 would mean job 1 on machine 3, first we do

this then we do from this, so this arc becomes active, and then from here we will do this,
so this arc becomes active. The rest of the dotted pink arcs are now removed, because
they are no longer required, so we will remove this we would also remove all this, so we
just have these 2 arcs coming in to the picture.
Now, if we find the let us mark it again with the thicker line, now the longest path of the
on the network gives once again a lower bound to the Makespan. Earlier the longest path
on this network was the sum of the job processing times 7 plus 8 15 plus 10, 25, 6 plus 4
10 plus 12. 22. 8 plus 8, 16 plus 7, 23. But now with the inclusion of these 2 arcs, if we
compute the longest path on this network would become 34. So let us just show the
computation of a longest path on this network.
So, let us say we start with 0 for this, so this comes at 7 it depends on how we write it,
because it is activity on node network, so we could say that this one finishes at 7 from
here. So, this would mean 7 plus 8, 15, 15 plus 10, 25, now what we do is this is 6, this is
6 plus 4, 10. Now, this one will be 15 plus 12, 27. 10 plus 12, 22, so 27, now this will be
8, 8 plus 8, 16, now this will be 16 plus 7, 23, 27 plus 7, 34 and the finish will be 34.
Now, the lower bound on the Makespan increases and it becomes 34 - 27 plus 7 more
importantly it is the longest path on the network. Now, we look at the next bottleneck,
547
which is M 2 and then we try and solve the 1 r j L max problem on M 2, on M 2 with 34
which is the lower bound of the Makespan, so that is updated as the final due date.
So, with due date equal to 34 we solve 1 r j L max on M 2, so we have J 1 J 2 and J 3,

and we have the processing times p j, p j on M 2. So, J 1 on M 2 is 10, J 2 on M 2 is 6,
and J 3 on M 2 is 8, then we write r j and d j, r j is the release time at which it is
available, and d j is the due date. Now, if we look at M 2 and we look at J 1, so J 1 on M
2 is the third operation, so earliest J 1 is available on M 2 is 7 plus 8 15.
So, 15 is the time earliest that J 1 is available on M 2, because at J 1 has to finish

processing on M 1 and M 3, and the earliest it can be available is 7 plus 8, so r j is 15
here. Now, as far as J 2 is concerned J 2 on M 2 is the first operation. Therefore the
earliest it is available is 0, as far as J 3 is concerned M 2 is the second machine, and
therefore the earliest J 3 can be available at M 2 is at time equal to 8.
Now, as far as the due dates are concerned J 1 on M 2 is the last operation, so it is
enough if it is available at time equal to 34. As far as J 2 is concerned J 2 on M 2 is the
first operation. So, the latest due date the farthest, we can think of is 12 plus 4, 16 it
definitely requires at least 16 more units to finish, so the due date can be 34 minus 16
which is 18. And as far as J 3 is concerned it is once again the last operation, J 3 on M 2
is last, but one operation. So, at least 7 more units are required, therefore the due date can
be 34 minus 7 which is 27. So now, we have to solve a 1 r j L max on M 2.
548
So, we start with the starting node and then we put J 1 as the first job, we start another
sequence with J 2, and we start another sequence with J 3 as the first job, so when we do
this let us start writing the completion times. Now, when we put J 1 as the first job, J 1 is
available only at time equal to 15, so a completion time is 25 for J 1, and then by the
preemptive EDD rule, we would based on the due dates we have J 2 and then J 3. Now, J
2 is available at time equal to 0, so J 2 can finish at 31, because it is available at 25 plus
another 6 is 31 by 31 J 3 is also available, which is available at 8.
So, 31 plus 8 39, due dates are 34 18 and 27, so here L max is 13 here L max is 12, so
lower bound on L max is equal to 13. Now, when we start with J 2, so J 2 is available at
time equal to 0, so it will finish at 6, and then J 3, and then J 1 by the EDD rule, now this
is available at 8 this is available only at 15. So, you have to definitely wait till 8, so 8
plus 8 16, now this is available at 15 16 plus 10 26, now the due dates are 18 27 and 34,
so lower bound on L max is equal to 0, because all of them are within the due dates.
And we also realize that actually what we have done is we have evaluated a feasible
solution, with 2 3 and 1 with L max equal to 0, this is the feasible solution with J 2 first,
then J 3 and then J 1 with L max equal to 0. Therefore, we can stop the algorithm here
itself, for a maximum tardiness problem if you have a solution with maximum tardiness
equal to 0, a feasible solution with 0, then it is optimal. So, we could stop right here, and
say that the sequence on J 2 on M 2 is J 2, J 3, and J 1, now we superimpose this on M 2.
549
So, M 2 is shown with the yellow color, so these 6 arcs are there, so the sequences J 2, J
3, and J 1, so first on M 2, we will first do J 2, then we will do J 3 and then we will do J
1. So, we have to update this, so first we do J 2, so we do this first, then we do J 3, so we
come to 3, and after that we go to 1 through this, so this is the sequence that will happen
on machine M 2.
Now, with this sequence we now need to update the lower bound, so the longest path on
this network will give us an updated value of the lower bound, so that; obviously, the
lower bound now can only be 34 or more, because we are adding arcs on this network.
Therefore, the longest path cannot decrease, so now, we have to once again find what are
these node labels, so that we find out the lower bound on the Makespan. So this is at 7
this is again at 15, so this is at 27, now this is at 8, now this one is 6 plus 8, 14, but then
we have 8 plus 8, 16, so this stays.
Now, this will become 16 plus 8, 16 plus 10, 26 whereas, earlier it was 15 plus 10, 25, so
this will become 26, otherwise 15 plus 12, 27, 27 plus 7, 34, 16 plus 7, 23, 34, so the
lower bound remains at 34 with this. The only change in the label is the change here, that
this is 16 plus another 10, 26, whereas here it was 15 plus 10 25 earlier, so the lower
bound does not change with the inclusion of these 2 arcs, but only one of the node labels
changes. So, now, we look at the third machine which is M 1, we have already seen M 3
and M 2, so we now need to do a 1 r j L max on M 1. So, we do a 1 r j L max on M 1
with the same due date as 34, because the longest path in this network right now is 34.
550
We now solve the 1 r j L max problem on the third machine, which is machine M 1, we
have already solved the 1 r j L max problems for M 3 and then M 2. So, now from this
network, we observe that the processing times, for the 3 jobs J 1 J 2 J 3 on M 1 are 7 for
J 2 it is 4 and for J 3 it is 8.
Now, we have to compute the released times r j as well as the due dates d j, so if we take
job J 1 on M 1, now the earliest this is available is time equal to 0, so r j will be 0 for
this. Now, here the earliest this is made available is after it goes though this machine,
551
therefore r j will be 6 and as far as J 3 on M 1 is concerned the earliest it is available is at
time equal to 0. Now, we look at the due dates, the due date remains as 34, now when we
look at this the longest path that connects will be 7 plus 8, 15 plus 10, 27, so 34, so 34
minus 7 is 27, 27 minus 12 is 15, 15 minus 8 is 7, so due date for this will have to be 7,
because the longest path is 7 plus 8 15 plus 12 27 plus 7 34.
Now, as far as this is concerned the longest path would be from here, so 34 minus 7, 27,
27 minus 12 will be 15 for this to come, so due date will be 15. And for J 3 which is
here, the 34 minus 10 plus 8 would give us 17, 34 minus 15 would give us 19, so this
will become 17. So let me compute this again. Now, as far as J 3 on M 1 is concerned
this is the one, so 7 minus 8, 15 would give us 19 from this side it will be 10 minus 8, 18
therefore, at time equal to 16 this should be available, so that 16 plus 18 becomes 34.
So, this is not 17 this becomes 16. So let me again explain the computation of the due
dates, J 1 on M 1 is here and the longest path is 7 plus 8 15 plus 12, 27 plus 7 34, which
gives us this 34. So, in order to meet this head line of 34 this should be over by time
equal to 7 plus 12, 19 plus 8, 27, so 34 minus the remaining processing time of 27 gives
us a due date of 7. Now, as far as this is concerned once again in the due date earliest or
latest that this has to be completed comes from 12 plus 7, it has remaining 19 time units
of processing, and since it has to finish at 34 the due date will be 15. Now, as far as J 3 is
concerned we are here, so again we go through the path, so 34 is the final at the moment
the Makespan is 34, so 10 plus 8, 18 more time units of processing is required, so the due
date becomes 16 for this.
552
Now, with this data we now try and solve a 1 r j L max problem on machine 1, so we
now begin with J 1 as the first job in the sequence, J 2 as the second job first job, and J 3
as the first job in the sequence. So, when we look at J 1 as the first job in the sequence,
now it is available at time equal to 0, therefore the completion time will be 7, following
the earliest due dates sequence we try to get J 2 inside. So, J 2 is available at 6, now the
machine is available at 7, so this will become 7 plus 4, 11, by which time J 3 is available,
so plus 8 19, so these are the completion times, the due dates are 7, 15 and 16, so only
one job is tardy.
So, from this we find that L max is 3, we also observe that we get a feasible solution,
when we actually apply this preemptive EDD rule, we get a feasible solution with L max
equal to 3. Now, with J 2 as the first job in the sequence, so J 2 starts at 6 completion
time is 10, now going by the preemptive due date J 1 comes next, so J 1 starts at 10 and
finishes at 17 by which time J 3 is available. So, J 3 will take 17 plus 8, 25, so now, the
due dates are 7, 15 for the for job 2 the due date is 15, for job 1 the due date is 7, and for
job 3 the due date is 16 So, this gives us a solution with L max equal to 9.
Now, this job this is the sequence this is J 2 J 1 J 3, this is J 1 J 2 J 3, so the sequence J 2
J 1 J 3 will give us a solution with L max equal to 9. Now, this job is early this job is
tardy with due date equal to 7 and completion time equal to 17, so L max is equal to 10
from this, and for J 3 due date is 16 completion time is 25, so tardiness is 9, so maximum
553
tardiness is 10. Now, if we consider the third one with J 3 as the first job, preemptive
EDD would give us the order J 3 followed by the lowest one J 1 and then J 2.
So, we get J 1 J 2, so here the completion times will be starts at 0 and finishes at 8, J 1 is
available at time equal to 0, so finishes at 15 by which time J 2 is available. So this
finishes at 19. Now, due dates are 16 for J 3, 7 for J 1, and 15 for J 2, now we realize that
J 3 is ahead of the due date this is behind the due date by 8 units, this is behind by 6
units, so we get L max equal to 8. Now, when in these 3 nodes what we have done is we
have started with J 1 J 2 and J 3, and we are able to get feasible solutions by applying the
preemptive earliest due date rule.
We may also a compute the lower bound and proceed, but finally, when we solve this
problem, we will get this solution with minimum L max equal to 3. So, 1 r j L max
problem talks about the sequence J 1, J 2, J 3 on M 1 with L max equal to 3, now we go
back and try to map this on this network, so on M 1 the sequence is J 1 J 2 J 3.
So, J 1 first then J 2 and then J 3, so we map this, so J 1 to J 2 will come here, and then
we will do J 2 to J 3, which will now come here, so now, this is the final network, where
on M 1 J 1 followed by J 2 followed by J 3. Now, we have to compute the longest path,
and update it now we do the computation of the node labels, so we have 7 plus 8, 15 and
this one will become 8 plus 8, 16 plus 10, 26. Now, this will become 7 plus 4, 11 instead
of 6 plus 4, 10, so this will become 7 plus 4, 11, now this node will become 11, 11 plus 8
554
19 this will become 19, so this will become 19 plus 8, 27, 27 plus 7, 34, but 8, 27 plus 10
will become 37 here.
Now, this one will become 11 plus 12, 23, 15 plus 12, 27, so finally the longest duration
will be the maximum of 37, 27 and 34, and we will get a solution with 37. Now, the
Makespan is 37 and we have a feasible solution, which Makespan equal to 37. You may
also observe that earlier the Makespan was 34, we computed L max equal to 3, so the L
max gets added and 34, finally becomes 37 which is the Makespan for the sequence
obtained using the shifting bottleneck heuristic. Now, one way of representing it is on
this network, the other way of representing it is on the familiar Gantt chart, and now we
will represent this schedule on the familiar Gantt chart, which we are used to.
So, we will, now represent it on the Gantt chart, this will be M 1, this is M 2, and this is
M 3, so the sequence on M 1 from this comes from 1 1 1 2 and 1 3, so the sequence is J 1
J 2 J 3. Now, on M 2 we realize that this is first, so J 2 J 3 and then J 1, so J 2 J 3 and J 1
and on M 3 it is again 1 2 and 3, so J 1 J 2 and J 3. Now, we have to map this on the
Gantt chart, so we look at J 1, J 1 first goes for processing on M 1, so J 1 on M 1 finishes
at 7, and at 7, J 1 goes to M 3.
Now, J 1 is actually the first job on M 3, so J 1 has come at time equal to 7, and finish at
15, and at 15, J 1 will go to M 2. J 1 comes here at 15. J 1 is the third job to be processed
on M 2, so we will just leave J 1 here, and we will take up J 1 after J 2 and J 3 have been
555
processed. Now, we look at this J 2’s first operation is on M 2, so J 2 will start at time
equal to 0 and finish at 6, and at 6 J 2 will come to M 1.
Now, J 2 is the second job to be done on M 1, and J 1 is the first J 1 is already over J 2
has come at 6, so it will take up J 2, so J 2 on M 1 will be 7 plus 4, 11, and J 2 will now
move at time equal to 11 to M 3, so J 2 will come here J 2 will come at 11. Now, as far
as M 3 is concerned J 1 is already over J 2 is the second job, so J 2 is ready, so J 2 on M
3 J 2 starts at 15 takes 12 units of time from here, and therefore 15 plus 12, 27. J 2 will
finish and J 2 completes all the 3 activities at time equal to 27. Now, let us look at J 3. J
3’s first operation is one M 1, so J 3 will be taken up at time equal to 11 and it will go
from 11 to 19.
So, J 3 will finish at 19, and J 3 will come to M 2 at time equal to 19, now J 3 is the
second job J 2 is already done, so J 3 will start at 19, so J 3 on M 2 will be 19 plus 8, 27.
So J 3 will finish at 27, and at 27, J 3 will now go to M 3. So J 3 will come here at 27.
Now, J 1 J 2 are already completed J 3 is the only one available, so J 3 on M 3 is another
7, so J 3 will start at 27 and J 3 will finish at 34, so J 3 will finish everything at time
equal to 34.
Now, we look J 1. J 1 has actually come here at time equal to 15, so J 1 on M 2 is another
10, so this can start at 27 only takes 10 more time units and finishes at 37. So J 1 will
finish at 37, which is indeed the Makespan. And this 37 is the same as this 37 that we
have computed, so we can do either of these, now from the shifting bottleneck solution,
we can map it on to the Gantt chart like we did. If we get a Gantt chart solution, it is also
a possible to map it on to this network and the longest path on this network will be the
same as the longest of the bars on the Gantt chart and for this problem the Makespan is
37.
Now, we have to check whether this Makespan is optimal or does the shifting bottleneck
heuristic guarantee optimality, the answer is the shifting bottleneck heuristic does not
guarantee optimality. What it tries to do is either when we look at it through this network
or through the Gantt chart, it takes one machine at a time and tries to find an optimum
sequence not for the Makespan problem, but for the 1 r j L max problem.
556
And then it adds on this L max into the existing Makespan, and updates the existing
Makespan. So, on the one hand it solves an optimization problem for a particular
machine, but the order in which the machines are taken will is now based on the loads.
Say here, we took the machines in the order M 3 followed by M 2 followed by M 1, this
order need not be the optimal order, therefore the shifting bottleneck heuristic does not
guarantee the optimum solution. Even though, it solves an optimization problem at each
stage, we also have to observe that 1 r j L max problem is actually a difficult problem to
557
solve, but luckily the branch and bound that we used is able to use solves problems of
reasonable number of jobs optimally.
Therefore, we are convinced that we can use the 1 r j L max on a single machine and
apply it sequentially on the machines. The thumb rule that we used here we first did it on
M 3, then on M 2, and then on M 1, because our choice was based on the work load on
each of these machines. Now, this particular order of applying the 1 r j L max need not
be optimum, so shifting bottleneck heuristic does not guarantee the optimal solution, but
it provides an excellent alternative to a dispatching rule based approach to solve the
problem, because any dispatching rule based solution would give us a certain Gantt
chart.
Now, this Gantt chart on one hand computes the Makespan on the other, also tells us the
order in which the jobs are going to be processed on each machine. Now, this order can
always be mapped on to the network, and the longest path would give us. Alternately the
shifting bottleneck heuristic solves a 1 r j L max problem on each machine, and then tries
to find out this order on each machine and maps it on to the network.
So, this is a very powerful heuristic does not guarantee optimality and is completely
different from dispatching rule based approach. Even though in principle both the
dispatching rule based approach as well as the shifting bottleneck heuristic, try to find
out what is the best sequence of jobs on each of these machines. Now, this method
became very popular largely because of it is ability to give optimum solutions in many
instances, and close to optimum solutions in many instances as well.
In fact, if one looks at the job shop scheduling literature there are 3 important or difficult
problems, which are called the Fisher and Thompson problems and particularly the
second of the Fisher and Thompson problem. For which people knew the optimum
Makespan, but we are not able to get that comfortably using dispatching rules. The
shifting bottleneck heuristic, for that particular problem instance gave the optimum
solution.
So, the shifting bottleneck heuristic is important for many reasons, one it is able to provide
very good solutions, even though not optimal on all occasions, it is able to give optimum on
many instances and close to optimum on very many instances. It provides an
558
alternative to a dispatching rule based approach to solve a job shop scheduling problem,
because this sequence 1 2 3 or 2 3 1, 1 2 3.
Suppose, we take a shortest processing time based Gantt chart, then it gives a certain
sequence for each machine that is obtained through a dispatching rule. Here, the
corresponding sequence is obtained by solving a 1 r j L max or by solving an
optimization problem, therefore it can give better solutions than using a single
dispatching rule.
So, this is used effectively used to solve job shop scheduling problems to minimize
Makespan.
Now in the line balancing problem, essentially we use the line balancing idea in
assembly, as and also these ideas can be used in manufacturing. Now, what is the
assembly line balancing problem? Now let us assume that we are trying to assemble a
product. Now, this assembly involves assembling or fixing several components and sub
assemblies into a main assembly, so first and foremost in the assembly problem, there are
components that are brought into the assembly and assembly of each one of them is now
called a task.
So, an assembly would involve 10 or 12 or 15 tasks, where 15 different subassemblies

and components are assembled into a final product or final assembly. Now, each task
559
also has a processing or a task time, so there is a time associated with each of these tasks.
Now, many times if there are 10 or 12 parts and sub subassemblies that go the assembly,
there has to be an order in which these have to be assembled.
Therefore, there are precedence relationships, which may say that I cannot assemble or I
cannot do task 3 unless I finish task 2. So, an assembly line balancing problem the data
that are required to do this are the number of tasks, the time on for each of these tasks
and the precedence relationship of the assembly.
So, let me take an example to explain all these, and then try and use the same example to
solve the line balancing problem. So this network shows an assembly line balancing
problem with 11 tasks, which are shown as 1 to 11 as nodes of the network. Now, the
task times are shown as the arcs, and these are 4, 2, 3, 6, 1, 6, 8, 2, 5, 6 and 4. Now the
number shown in yellow are the time taken to do each of these tasks. And they are shown
as node labels, associated with the node, the precedence is captured through the arcs. For
example from this network one can say that we can perform task 2, after we have
performed task 1, task 1 is a preceding activity for task 2.
Similar, task 1 should precede task 4, similarly 3 and 5 should precede 6, which means 6
cannot be done till 3 and 5 are completed 6 cannot be taken up and so on, so the precedence
relationship are captured in the form of these arcs on this network. Now, without loss of
generality we can assume that the order 1 2 3 4 5 6 7 8 9 10 11 is feasible,
560
so there is a feasible solution to this problem. For example, we will not have a situation,
where one has to precede 2 two has to precede 3 and 3 has to precede one, so we will not
have that.
So, we would say that 1 2 3 4 5 6 7 8 9 10 11 is feasible, which also means that if we

have one operator sitting and doing all the assembly of 1 to 1 to 11, then the operator can
come with a finished product or an assembly. And if we assume that these tasks times
are in minutes, then if one person does it and does it in the order 1 2 3 4 5 6 7 8 9 10 11,
then the total time that will be taken is 4 plus 6 10 12 15 16 22 30 32 38 43 plus 4, 47
units of time. Every 47 units we will see that an output comes. Many times what happens
is these assemblies, may also require a certain skills like delicate assembly or testing
after the assembly is completed and so on.
So, many times if we want to really increase the output of this we need more operators
sitting. If 2 operators sit, and if both the operators actually assembled everything. Then
we would get 2 pieces in 47 time units, which means effectively one piece in 23 and a
half time units, but that would involve for example, creating if activity number 6 is a
testing activity. And that would involve creating 2 test benches with the same testing
equipment, so if we are going to have a situation where we have several operators sitting,
and each one of them doing the entire thing.
Then the output can be 47 divided by the number of operators, but many times such a
thing would not be feasible, because that would mean duplicating a lot of resources, that
are required for each one of them. Therefore, it is customary in a assembly line
balancing, that one operator is given 2 or 3 tasks or certain number of tasks, and the
assembly moves to the next operator, who carries out certain number of tasks and so on.
So, that each operator would have a work bench and there will be a time taken by each
operator in the work bench. Now, if we assume that we are going to have 11 work
benches, where task 1 is done in work bench 1, task 2 in bench 2, task 3 in bench 3, and
so on. And if, we put 11 operators one for each and assuming that each operator has the
skill to do that activity, then if we start the line with 11 like this.
If, we start the line here at the end of 47 time units one piece will come out, the first
piece will come out, but then the second piece would have started by them. And at steady
state the output of this with 11 desks and benches or workstations, and 11 operators will
561
actually be the maximum of the time which is 8 time units. So, if we create a situation
with 11 work benches, and one operator doing each, so the output can be one in 8
minutes, as long as we have we do not have parallel workstations.
So, if we have one workstation sequential, and each workstation is carrying out one
tasks, so with 11 workstations, if we proceed we will have a an output every 8 minutes
and that is call the cycle time, so the cycle time in this case will be 8 minutes. If, we have
11 workstations, but then we have only one operator who first goes to workstation 1
finishes, the assembly takes it goes to workstation 2, finishes the assembly and so on,
then the cycle time will be 47 minutes.
So, as long as we do not have parallel stations, the realistic cycle time that we can
achieve in this case is anything between 8 minutes and 47 minutes. So, the problem that
we wish to solve is: given a cycle time T minutes, what is the minimum number of
workstations that we require, what are the tasks that have to be carried out in each of
these workstations or what is the allocation of the task to each workstation. Such that the
workstation time is less than or equal to capital T, we have minimum number of
workstations that is required.
And the precedence relationship that we have described is not violated by the
unidirectional flow that we want to have amongst the workstations. Now, that problem is
called the line balancing problem, and that problem is also generally solved using
heuristic algorithms. So we will see some aspects of line balancing and the heuristic
algorithm in the next lecture.
562
Prof. G. Srinivasan
Lecture - 31
Line Balancing
In this lecture, we continue the discussion on Line Balancing. So, this network represents
the activities in a typical line balancing situation. So there are 11 activities that have to
be performed, one could assume that 11 components are subassemblies are assembled to
form a final product. Now, there is an activity time associated with each of these
activities, which are given in yellow, the 4 units, 2 units, etcetera; units could be in
minutes or seconds depending on the assembly. There is also a precedence relationship,
which is also captured in this network.
For example, activity number 2 or say component number 2 cannot be assembled unless
one is assembled, so 1 precedes 2. Similarly, 1 precedes 4, 4 precedes 5, 3 and 5 precede
6 and so on, which means we cannot do activity 6 unless both 3 and 5 are completed. As
we have mentioned in the previous lecture, if there is only one person who is actually
assembling everything, then this person would take the sum of all the processing times
which is, 12, 15, 16, 22, 30, 32, 47 units, to assemble to each final product.
563
So, if one person is sitting and doing everything it would take 47 time units, if two
people do it, then at steady state it will be 47 divided by 2, every 23 and a half time units
1 could get an output and so on. But, in general it does not happen that way, because
some of these could involve equipment and machinery and testing devices, which we
may not duplicate so much. So, you would rather have different workstations or benches
to which these activities are assigned.
So, if we imagine a situation where there are the 11 activities, and there are 11
workstations or benches say 1, 2, 3, 4 up to 11. We also assume in this example as an
every line balancing problem, there the order 1, 2, 3, 4 up to 11 is feasible, so if we have
11 benches or workstations, which say 11 operators, and each person carries out 1
activity. Then the first unit will come out at time equal to 47, but at steady state the
output of the system will be the bottleneck, which will be the maximum of the individual
assembly times, which would be 8 units.
At the same time we also define something called T called cycle time, which is the
expected output from the system. For example, in this, if we say that, if we say T equal to
12, then the demand is such that we would expect an assembly to come out every 12 time
units. So, once we define this T, the problem is one of trying to find out, how many
workstations we require - minimum workstations such that, each task or activity is
assigned to only 1 workstation. The time taken in a workstation does not exceed the
564
cycle time 12, and the precedence relationships among the activities are satisfied, so this
is called the line balancing problem.
So, given a set of activities that have to be performed, given their activity times and the
precedence relationships, for a given cycle time T what is a minimum number of
workstations it is required? How these tasks are assigned to these workstations such that,
the total time taken in a workstation is less than or equal to T. Each tasks is assigned to
only one workstation and the precedence relationships are satisfied; now this is called the
assembly line balancing problem. So, let us first show a formulation of the line balancing
problem and then we would look at a couple of solutions to the line balancing problem.
Now, there are 11 tasks, so we could say that there can be a maximum of 11 workstations
or 11 benches that we could have 1 for each. But, then that would give us a feasible
solution with 11 workstations or 11 benches, where each activity goes to 1 workstation.
And the cycle time would still be less than or equal to 12, in this case because the cycle
time will be 8. Since, the cycle time is less than or equal to 12, it is possible to add
activities to workstations and thereby try and minimize the number of workstations.
So, we now start defining the first variable as S j equal to 1, if workstation j is chosen,
we initially assume that a maximum of 11 stations can be used, there are n activities we
would assume that a maximum of n stations can actually provide us a feasible solution.
And therefore, we would like to minimize the number of stations, so here we would
565
define j equal to 1 to 11 for this problem or j equal to 1 to n for a general problem. So,
we would say that a maximum of 11 stations are there, how can we bring down the
number of stations.
Now, if we can solve it with 5 workstations, then the solution would take 5 out of the 11
values will take 1 and the remaining 6 out of the 11 values will take 0, saying that these
stations are not created. So, S j equal to 1, if station j is chosen or created, and let X i j
equal to 1, if activity i is allotted to station j; now we start writing the constraints, each
activity should go to only 1 station. So, we would have the constraint sigma X i j
summed over j equal to 1 for every activity i, each station the activities allotted to a
station, a sum of the activity time of the activities allotted to a station, should not exceed
T.
So, this would be sigma T i is the activity time for activity i, so T i into X i j summed
over i should be less than or equal to T for every j, j represents the workstation. So, the
activities that are allotted to a workstation, the sum of the activity times of those
activities should not exceed the cycle time T. More importantly we can allot an activity
to a workstation, only when the workstation is chosen. So, we redefine it by saying that,
if there are activities allotted to a chosen workstation, then the sum of times of those
activities should not exceed T.
So, the left hand side tells us, the sum of the activity times of those activities that are
allotted to a station, now this is the cycle time, but then we could also add an S j, where S
j equal to 1 when the station is chosen. This also means this constraint would also ensure
this is T into S j, so we will write is as T into S j for every j. So, only when S j equal to 1
that workstation is chosen and to which activities can be assigned. So, you cannot assign
activities to a workstation that is not chosen, and for every chosen workstation the sum of
the activity times should be less than or equal to T.
This constraint also helps us in linking X i j variables with S j variable, we still have not
come to the objective function, we will do that, right now this links the X i j variable
with the S j variable. Then we have this precedence constraint, so let us take 1
precedence which is say 1 and 2, so activity 2 can be assigned only after activity 1 is
assigned, it means 1 precedes 2.
566
Now, this would mean that, for example we have X 1 1, which means activity 1 is
assigned to station 1, activity 2 can be assigned to station 1, 2, 3, 4 up to 11. Now, this
implies that, if 1 and 2 are assigned to the same station 1, the operator will first assemble
component 1 and then assemble component 2. So, we this is modeled as if 1 goes to 1,
then X 1 1 should be less than or equal to X 2 1 plus X 2 2 plus X 2 11, this would mean
X 1 1 plus X 2 1, X 1 2 plus X 2 2 component 1 goes activity 1 is in workstation 1. Then
activity 2 can go to workstation 1 or workstation 2 or 3 or 4 and so on.
But, then if for example, activity 1 is allotted to workstation 2, then activity 2 should not
be allotted to workstation 1, it can be allotted to 2, 3, 4 etcetera till 11. So, this will be
modeled as X 1 2 will be X 2 2 plus X 2 3 plus etcetera plus X 2 11, similarly if activity
1 goes to workstation 3, then we have X 2 3 plus etcetera plus X 2 11. So, like this the
last set for this pair would be X 1 11 is less than or equal to X 2 11, so these 11
constraints capture the precedence relationship between 1 and 2.
Similarly, to capture the precedence relationship between 1 and 4, we will have 11 other
constraints, so we have to first find out what are all the, how many precedence
relationships we have, so here we have the number of arcs essentially represent, the
number of precedence relationships. So, we have 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 and
13 precedence relationships, each precedence relationship is between a pair of activities
and then we have to have 11 constraints for each of the precedence relationships.
567
Giving us 13 into 11, 143 constraints to handle this, we are just shown the 11 constraints
for one particular precedence relationship. Now, there are actually better ways of writing
the precedence relationships, but they are not going to emphasize on that, the purpose
here is only to show how the formulation looks like. There are slightly more efficient
ways of reducing this 143 constraints into a lesser number of constraints, which capture
the same thing.
But, what we have shown here is, the reason or the rationale behind the precedence
constraints, where if there is a relationship between 1 and 2, the workstation to which
activity 2 is assigned should be higher numbered workstation or equal to than the
workstation to which 1 is assigned. So, if 1 is assigned to workstation 1, then 2 can be
assigned to 1, 2 up to 11, but if 1 is assigned to workstation 2, then 2 can be assigned
only up to 2 to 11, so we cannot have X 2 1 assignment there.
So, we cannot this would prevent activity 1 from being assigned to workstation 2 and it
will it will have activity 2 assigned to workstation 1, it will prevent that, because if
activity 2 is assigned to workstation 1, through this all of these will be 0. Therefore, we
will have 1 less than equal to 0 and this will prevent activity 2 from being assigned to
workstation 1. So, we have 11 such constraints which model this precedence relationship
and then we have 13 such precedence relationships, then the objective function is to
minimize the number of workstations.
So, minimize sigma S j, j equal to 1 to 11 will be the objective function associated with
this of course, all X i j is and S j are binary variables, so they are either 0 or 1. So,
optimally solving this 0 1 problem, can give us the optimum solution to the line
balancing problem. But, then we will look at two very simple heuristics to solve this line
balancing problem. One is a heuristic solution, which is based on shortest processing
time based heuristic.
568
So, we would start by saying workstation activity time, so we would say to begin with,
we can do activity 1 only, because all the rest of them have precedence relationships
which are related to activity 1, so we start with activity 1, which goes to workstation
number 1. We assign activity 1 and activity 1 has processing time equal to 4. Now, that
we have assign activity 1 to workstation 1, 2 and 4 are available now, because activity 1
has been allotted. So, we have let us say 2 and 4 with us, now 2 requires the time of 2
units, 4 requires the time of 6 units.
So, we would assign based on SPT rule we would assign activity 2 and then it takes 2
more units, so activity 2 has been assigned and since, activity 2 has been assigned we
have activity 3 that is available now. So, 3 and 4 are available, when the 3 and 4 are
available, we can try and assign one of than, 3 requires 3 units, 4 requires 6 units, so once
again based on SPT, so 3 will go here, so plus another 3 units. So, since 3 is assigned we
cannot add 6 into the list, because 4 is yet to be assigned.
So, we have only 4 that is available, so we try and put 4 here, but then we realize that the
cycle time exceeds 12 as 4 plus 2, 6 plus 3, 9 plus another 6 is 15, so cycle time exceeds
12. So, we create a second workstation and try and put activity 4 here, this goes at time
equal to 6. Now, when activity 4 is allotted, 5 is the only one that is remaining, so you
put 5 comes here, 5 is the only one that remaining, so 5 goes here ((Refer Time: 19:09))
plus 1 and once 5 is allotted 6 can be brought in, so 6 comes here.
569
Now, 6 requires 6 time units, so we try and put 6 here and we realize that it exceeds the
value of 12, so create a third workstation with activity 6 with time 6, 6 goes now; the
moment 6 is allotted we could bring in 8, 7, 8 and 9. So, 7, 8 and 9 with 7 has time equal
to 8, 8 has time equal to 2, 9 has time equal to 5, so we choose 8 based on SPT rule. So 6
and 8 plus 2, so 8 goes now when 8 goes, 10 can come in. So, we have 7, 9 and 10
available, 7 requires 8 time units, 10 requires 6 time units, 9 requires 5 time units. Now
the smallest is 9 with 5 time units.
We try and put it here, but we are not able to, because it exceeds the cycle time. So,
create a 4th workstation with 9 here and time equal to 5. So 9 goes, then we have 7 and
10, ((Refer Time: 20:52)) this is 8 and this is 6, so the smallest number is 6. So, we take
10 plus 6, so 10 goes, we can still we cannot bring 11 into it, because 7 has not yet been
done. So, now, 7 is the only one that is available. So 7 requires 8 time units adding an 8
here will exceed the cycle time. Create one more workstation with 7 and time equal to 8,
so 7 goes.
Now, 11 is the only one that is remaining 11 has 4 time units required 4, so add 11, 7 and
11, 8 plus 4, so now all the activities are assigned, and we get a solution like this. Now,
the cycle times in each of these are 4 plus 2, 6 plus 3 9, 6 plus 1 7, 6 plus 2 8, 5 plus 6, 11
and 8 plus 4 12. So, based on the shortest processing time rule, we now have a solution
with 5 workstations and these are the allotments to it. Now, the goodness of a solution is
measured, using two, a primary measure and a secondary measure. Now the primary
measure is called line efficiency.
570
Line efficiency is we now have 5 lines with the cycle time of 12 and the maximum is
also 12, so 5 lines with 12 cycle time is about 60 time units, and the total of these 47
time units are now made in 5 stations each with 12 in 60. So, line efficiency is 47 by 60
or in general it is sigma T j or T i sum of the activity times divided by N into T, where N
is the number of workstations chosen and capital T is the cycle time. So, this is 47 by 6,
so dividing 6 7's are 42, 50, 6, 8 are 48, so line efficiency is 78.33 percent. We also
realize that in the line efficiency equation the numerator is a constant and actual variable
comes only in the denominator with this N, because this is the constant and the N and T,
depending on capital T, the N gets defined.
So, the actual measure is dependent only on the denominator and not on the numerator,
numerator is a constant. Now there is a secondary measure, which also tells us how
balanced the line is within these 5 workstations. For example, this is going to take 9 units
7 units 8 units 11 units 12 units, now when even though the allowed cycle time is 12, if
some stations take less time and some stations take 12, there is some kind of an
imbalance. Now, the operator who is working on the workstation, which takes 12 units is
going to be occupied all the time, whereas the operator who works on this is going to use
only 7 time units, for every 12 time units.
Or will be producing excess and building inventory within the line, which cannot be
consumed, because the slowest person is going to take 12, either this person has to stop
571
working for every 5 time units, or this person will end up building some inventory. There
are many things that matter, because the input to this is the 9 that comes from this
workstation. So, we all need to ensure that the workstations are sufficiently balanced, so
there is very little difference between these 5 numbers, the closer these 5 numbers are
either inventory will be less or the idle time will be less.
If these numbers vary as in this example, you will find that either the person has to be
idle or the person will end up creating inventory which cannot be consumed. So, there is
a measure called smoothness index, which tries to in this example the smoothness index
will be the actual cycle time is 12, so 12 minus 9, which is 3 square 12 minus 7, 5 square
plus 4 square plus 1 square plus 0. So, in general the smoothness index formula will be
sigma T minus sigma T i X i j whole square, so this represents the time in a workstation,
this represents a cycle time; so square and sum as a smoothness index. Larger the line
efficiency more efficient is the line, smaller the smoothness index better it is for the line,
so the solution to a line balancing problem is measured on these two factors, which are
line efficiency and smoothness index.
Now, let us look at another solution to this problem, which is this heuristic based on
longest path in the network.
572
Now, if we take activity 1, the longest path is going to be 4 plus 6, 10 plus 1, 11 plus 6,
17 plus 8, 25 plus 4, 29 will be the longest path for 1, once again 4, so you could reach 6
using this which would give us 5 or using this which would give us 7. So, 4 plus 7, 11
plus 6, 17 plus 8, 8 and 5, 25 plus 4, 29. For activity 2 the longest path will be 5 plus 6,
11 plus 8, 19 plus 4, 23. For activity 3 the longest path will be 3 plus 6, 9 plus 8, 17 plus
4, 21. For activity 4 the longest path will be 7 plus 6, 13 plus 8, 21 plus 4, 25.
For 5 the longest path will be 1 plus 6, 7 plus 8, 15 plus 4, 19. For 6 the longest path will
be 6 plus 8, 14 plus 4, 18. For 7 it will be 12. For 8 it will be 12; for 9 it will be 9; for 10
it will be 10 and for 11 it will be 4. So, these are the longest paths for the 11 activities
that we have, now we will sort the activities in the decreasing order of the longest path.
And the arranged order will become 1 first with 29, 4 with 25, 2 with 23, 3 with 21, 5
with 19 6 with 18, 7 with 12, 8 with 12, 10, 9 and 11. So, this is the order of arranging
the activities based on the longest path, which is there in this network. Now, we use this
information and develop another heuristic. Now such a heuristic would start like this.
573
So, once again we have workstation activity and time, so we start with workstation 1 and
activity 1 is assigned to workstation 1 with time equal to 4, now the moment activity 1 is
assigned we have 2 and 4 that are available. Now, earlier we chose that activity based on
shortest processing time, now we will choose the activity based on which one appears
early in this order. So, between 2 and 4 you realize that 4 actually appears earlier than 2,
so we choose, so activity 1 followed by activity 4, so 4 goes 4 requires time equal to 6,
so 4 plus 6.
Now, once 4 is over, then you can take 5 into it now go back and see between 2 and 5
which one goes first, so 2 goes first, but then we realize that if we add 2 we have to
create another workstation, because this requires or we cannot take 2 here. So, between 2
and 5, 2 comes first, so 2 requires time unit of 2. So, we add time unit of 2 here, so this
goes activity 2 plus time unit equal to 2.
Now, once we have activity 2 competed, then we have 3; let us here that comes in. Now
between 3 and 5, 3 comes first in this. So we take 3 requires 3 units, we cannot add 3 into
it, because it exceeds the cycle time of 12. Create a 2nd workstation with 3, activity 3
and with time equal to 3 units. So, once activity 3 is allotted we cannot allot 6, because 5
is still to be allotted. So we allot 5, so 5 is the only one. So, we do not look at this order,
we look at this order only when we have more than 1 activity available for allocation.
So, 5 is the only one, so it takes plus 1 time unit, so now 6 is available. 6 is the only one
574
available and 6 can be added here, so plus 6. 6 also takes 6 time units once activity 6 is
allotted, we have 7, 8 and 9 that are available. So, we have 7, 8 and 9, we have 3
activities that are available, now we look at this order first it is 7, and then it is 8, so we
take 7, 7 requires 8 time units 7 goes. We cannot add 11, because we still need to do 8
and 9, now between 8 and 9 8 comes first 8 requires 2 units, so allocate 8 here, 8 requires
2 units; 8 goes.
Now 10 comes first, so now 10 comes here next, now between 9 and 10 we look at this
order and realize that 10 comes ahead of 9, now we cannot add 10 there, so create a 4th
workstation and then put 10 here with time equal to 6. So 10 goes. So, that brings us 11
into the picture, so between 9 and 11 9 comes ahead, so 9 plus 5, so 11 is the only one
that remains, so 11 requires 4 time units. So, 11 cannot be fit into this, so create a 5th
workstation states activity 11 and has cycle time equal to 4.
So, this has cycle time equal to 12, 10, 10, 11 and 4, now this will have the same line
efficiency, because there are 5 stations with maximum time equal to 12 - same line
efficiency. But, this will have a larger smoothness index, because this would contribute
12 minus 4 the whole square. So, this from a smoothness index point of view, this is not
very desirable, because it is going to create a lot of variation, so either a lot of inventory
or a lot of idleness into the system.
Now, we can actually do another thing here which is this, now the next effort can be in
both these solutions, we have a total of 47 time units for allocation and by choosing 5
workstations and having a maximum time of 12, we have 60 time units available, so the
line efficiency was 47 by 60. Now, if we have 5 workstations and instead of cycle time
of 12, if we can have the cycle time of 11, then not only will the line efficiency go up,
but the smoothness index will also come down.
So, the next issue will be given the number of workstations can be reduce the cycle times
slightly further, so one could think in terms of solving this problem again, and using
some heuristic to see whether we can actually get a cycle time of 11 instead of a cycle
time of 12. Now, if we try to do this here, can we get a cycle time of 11 in this, let us try
and do this once again with a cycle time equal to 11. And let us do it little quickly, so
workstation 1 would get activities 1 and 4 with cycle time 4 plus 6.
So, 1 and 4 are allotted, then we have 2 and 5 remaining, now in this order 2 comes first
575
2 requires 2 time units, so create a second workstation with activity 2 and time 2. So 2
goes 5 remains, 3 comes in. Now, between 3 and 5, 3 requires, 3 comes first, so 3 can be
added here plus 3, 3 goes. Now, 5 is the only thing that is remaining. If 5 is the only
activity remaining ordinarily we would try and put 5 here, because this is the current
workstation, but we also realize that the 5 can actually go here, because both 1 and 4 are
satisfied.
So, you could add 5 here ((Refer Time: 39:24)) to give a plus 1, so 5 also goes which
brings us 6 into the picture. 6 requires 6 units of time, so 6 can go here plus 6, so 6 is
over. Then we have 7, 8 and 9; out of these 7 comes first with time equal to 8, so 3rd
workstation 7 comes first with time equal to 8; 7 goes. Then we have 8 and 9, 8 comes
next, so 8 comes next plus 2, so 8 goes then 10 comes in, now between 10 and 9, 10 is
first.
So, create a forth workstation with 10, 10 has time equal to 6 and goes. Then comes 9, 9
has time equal to 5 and then we need to create a 5th workstation with 11 and time equals
to 4. Now, both the line efficiency and the smoothness index will be better for this
solution, because line efficiency will be higher, because it will become 47 by 55 instead
of 47 by 60. Smoothness index would also be slightly better, because the loads on the
workstation are 11, 11, 10, 11, 4, so smoothness index will also be better here compared
to this one.
So, we could if we end up from the solution if we realize, that the line efficiency is
slightly lower, then we could think in terms of for the same number of workstations that I
have in this solution, can I bring down this T by 1 and get a better line efficiency. So, we
can work on that till we cannot improve the line efficiency any more. Now there are a
few other advantages of using these heuristic solutions, now if we go back and try to
solve this optimally in the way that has been given here.
576
For this particular problem, we will have 11 variables for S j, because we maximum of
11 S are available, S 1 to S 11 and then we have another 121 variables for the X i j. And
then we have many constraints, this will be for every i, so this will be 11 constraints, this
would also be for 11 constraints and this set will be 11 for each precedence and there are
13 precedence, so 143 constraints. So, if we solve it this way even though as I mentioned
there are better ways of representing this, if we solve it as it is we are talking of 132
variables.
And 143 plus 22, which is 165 constraints. Now we solve a slightly large binary problem
with 132 variables and 165 constraints, now let us try and map, let us say this solution
into this formulation and see what happens. Now, this feasible solution has 5
workstations.
577
So, if I try to map this here, for this I will have a solution which this is like S 1 equal to S
2 equal to S 3 equal to S 4 equal to S 5 equal to 1, so 5 workstations. And I will have X 1
1, X 2 1, X 3 1, X 4 2, X 5 2, X 6 3, X 8 3, X 9 4, X 10 4, X 7 5, X 11 5 equal to 1 that
will be the solution, and we map it to this. Now, what does this mean, this has given me a
feasible solution with 5 workstations, so a feasible solution is an upper bound to a
minimization problem.
And therefore, since I have a feasible solution with 5 workstations, I am not interested in
feasible solutions with optimum solution, cannot be 6 or 7 or 8 or 9 or 10 or 11, so the
optimum solution can have 5 workstations or less. So, what I will do is, I will now try
and solve to find out, if there is an optimum solution with 4 workstations, because
already I have a feasible solution with 5 workstations. So, I will try and see, if I can have
an optimum solution with 4 workstations. So, if I try and solve it with 4 workstations,
then you can see the number of variables and constraints that I will have.
578
Now, if I solve for 4 workstations there will be only 4 S j values, and there will be 44 X i
j values, so X i j will now take only 44 variables, because each of the 11 activities can be
assigned to each of these 4 workstations. Now, the constraints X i j, this will be 11
constraints, this will be only 4 constraints and here, for every precedence relationship
there will be only 4 constraints into 13. So, I end of having 48 variables and 52 plus 11,
63 plus 467 constraints and if the optimum solution is 4, it will give me the optimum
solution.
So, getting a heuristic solution first, actually helps us in getting the optimum solution in a
slightly better way, otherwise we would be solving an optimization problem with 132
variables and 165 constraints. And because of the heuristic solution that we have, it is
enough now that the binary programming to get the optimum solution will have 48
variables and 67, 52 plus 11 63 plus 4, 67 constraints, so it is enough to formulate and
solve that problem. Now, in case the optimum solution does not have 4 benches, but has
5 benches then the optimum would give us infeasible, because we have worked with 4
benches.
579
Then we already know that we have a solution with 5 benches, and therefore this is 5
benches are 5 workstations, and therefore this is optimum. If the optimum had a solution
with 4 workstations, then it will show that solution, then we will have to try and see
whether a solution with 3 is possible. Now, in order to approach that we will look at it
from a lower bound point of view, because total time required to assemble is 47 and if
we have 12 as a cycle time, then the minimum number of workstations which is a lower
bound is 47 by 12 upper integer value, which would give us 4.
So, there cannot be an optimum solution with 3 for this case, so the only case that we
have to look at is can we have a solution with 4 workstations, which means it is enough
to solve this problem with 48 variables and 67 constraints. And if we actually do that, we
get an optimum solution to this problem which will look like this, so the optimum
solution is like this.
580
So, the optimum solution has 4 workstations, so workstation 1, activity and time, so 1, 2
and 4 with 4 plus 6 plus 2, station 2 is 3, 5, 6, 8 with 3 plus 1 plus 6 plus 2. Station 3 is
10 and 9 or 9 and 10 with 6 plus 5 and station 4 is 7 and 11 with 8 plus 4, so this is the
optimum solution to this particular problem. Where we have 1, 2 and 4, once again
precedence are satisfied, 3, 5, 6 and 8 precedence are satisfied, 9 and 10 precedence are
satisfied, 7 and 11 precedence are satisfied.
So, the cycle times here will be 4 plus 6, 10 plus 2, 12, 12, 11, 12 the total equal to 47,
line efficiency is equal to 47 by 4 into 12, which is 47 by 48, which is the best that is
achievable. And smoothness index, this is about 98 percent the smoothness index will be
12 minus 12, the whole square plus 12 minus 12 the whole square plus 12 minus 11
whole square plus 12 minus 12 whole square, which will be 1, which is the smallest that
is possible. So, this is how we solve the assembly line balancing problem, so we first try
and find out given a cycle time, what is a minimum number of workstations that are
possible.
Now, if we get to the optimum solution, we first in order to make the optimum solution a
little simpler, in terms of number of variables, we first try and get a heuristic solution like
what we have shown here. So, in for 11 activities we got a heuristic solution with 5
workstations, and therefore we try and solve the optimum solution case with 4
workstations. So, that if there is a solution it will show, if the optimum is 5 then it will
581
show in feasibility or alternatively we could actually do it with 5. So, the number of
variables and number of constraints will increase slightly and then if the optimum is 5 it
will show up as 5, if the optimum is 4 it will show up at 4 and so on.
In fact, if the optimum were 3, which is not possible in this case, then we solved it for 4
stations, if we could get a solution with 3 it would still show the 3, but in this case it is
not possible, because the total is 47 and the cycle time is 12, so minimum of 4
workstations are required. So, the main problem is to minimize the number of
workstations for a given cycle time, the next problem if we were to look at only heuristic
solutions is to try and get invariably the heuristic solution will over estimate, the number
of benches slightly.
And in such a case the problem will then be for a given number of workstations, can I
reduce the cycle time further, so that the line efficiency can go up, the smoothness index
can come down. But, the basic problem is always 1 that minimizes the number of
workstations for a given cycle time. So, with this we come to the end of certain
production control related decisions, we have seen some aspects like scheduling, as well
as line balancing.
Now, after this we will start looking at some aspects of location decisions, how do we
locate plants from a manufacturing and supply chain point of view. And layout decisions
as to how, what are the algorithms and methods by which we can have relative location
of facilities inside a given location. So, these aspects related to location decisions, we
will see in the next lecture.
582
Prof. G. Srinivasan
Lecture - 32
Location Problems - P Median Problem, Fixed Charge Problem
In today’s lecture, we address location and layout. Location essentially deals with, where
we want to locate the facilities and layout deals with, how we locate the other facilities
inside, once the location is chosen. So, we are going to see two things, one is called
location, which answers the larger question of, where do we locate a facility and layout,
how do we relatively locate facilities inside a chosen location. So we ask two questions,
one is on location and the other is on layout.
Now, location decisions are important as well as the layout decisions are. Location
decisions are strategic in nature and location decisions mostly are qualitative in nature,
even through several quantitative models are available for location. The decision on
location would also depend on many qualitative factors, some of which we will see now.
The first important decision is raw material, if we are looking at locating a
manufacturing plant or a facility, which is raw material intensive, then it is customary to
locate these near the places, where we have raw material is available.
583
Sometimes, location decisions are also based on labor availability. Areas where labor is
abundantly available and labor is not very costly. Sometimes, location of facilities also
happens near the demand areas, particularly retail stores, banks, etcetera are all located
near the demand area. Sometimes, locations are also made based on transportation
decisions for example, near a sea port, depending on the amount of material that come in
and have to be transported.
Sometimes, the location decisions also depend on the policies of the government.
Governments can give subsidy to locating certain types of industry in certain areas.
Certain areas in the state are declared as areas that are suitable or that are allocated for
location of manufacturing plants and people who choose to locate plant there, would get
certain benefits and subsidies on tax, etcetera. So, these are other things that force the
manufacturing and other organizations to locate plants in certain areas, depending on the
policies of the government.
Now, the location decisions also become slightly different when we want to locate public
facilities such as a police station or a fire station, etcetera. In such cases, the reach
becomes a dominant factor and we try to locate them in such a way that, the time or
distance between the place, where these central facilities are located and the farthest
customer is minimized. So, it will be like a mini max problem, where we try to minimize
the maximum distance between the location and any one of the customers. Now, in
addition to these qualitative factors, there are other quantitative models that are available
for location. And in this lecture series, we would concentrate a little more on the
quantitative models that are available for location as well as for layout.
584
So, the first model that we will see is called the p median model and this model
originated from location and layout, though this has several useful applications. Now, the
p median model is like this, and suppose we have 8 points, say we call them 1 to 8. Now,
from the location point of view, what we are interested is, is there are 8 points and these
8 points represent demand areas or 8 customer demand areas, where we do locate the
facilities such that, we are able to meet the demand with minimum distance of
transportation.
Now, as far as the p median problem is concerned, to begin with we may say that, there
are 8 points and the coordinates of these 8 points are known. So, we can also have a
distance matrix d i j between points i and j, so d i j will represent the distance between i
and j. Now, let us assume that, these 8 points represent 8 demand points or areas where
the products are required. And we wish to locate, let us say p facilities, let us say 2
facilities such that, we produce in these two facilities and we meet the demand in all the
8 facilities with minimum transportation distance.
Example, if we decide to have say 5 and 3 as the two facilities, if we decide to do that
then we say we are going to locate facilities in these two points. And we will say that, we
will find out, to which facility the remaining six points are to be attached such that, there
is minimum distance. So, if we fix 3 and 5 then we will attach 1 to 5, because 1 to 5 is
assumed to be lesser than 1 to 3, 2 will go to 3, because 2 to 3 is lesser than 2 to 5. Let us
585
say 4 will go to 5, 7 will go to 5, 8 will also go to 5 and let us say, 6 goes to 3, because 6
3 is lesser than 6 5.
So, if we fix 3 and 5 as the two places, where we are going to locate the facilities then
based on minimum distance we may say that, 1 4 7 and 8 will go to 5 and 2 and 6 will go
to 3. Now, at the moment, we are not looking at, what are the demands in these eight
places or we may assume that, the demands are equal. We are not looking at, when we
locate a facility in 5 and attach 1 4 7 8 as well as 5 to itself, we assume that, this facility
will have enough capacity to meets the demands of all the points that are attached to it.
So, this being a first model, we are not looking explicitly at what is the demand and what
is the capacity or alternatively we assume that, capacity is enough to meet the demand of
all the points that are assigned to it and therefore, the decision is based only on minimum
distance. Now, these two are called the median points and since we are talking of p
medians, this problem is called the p median problem. Now, this problem can be
formulated as a binary integer programming problem, which we will see right now.
So, we will assume now that, we will have X j j equal to 1 if point j is a median and X i j
equal to 1 if a non median i is attached to a median j, so X j j equal to 0, otherwise
automatically follows. Now, since we have defined X j j equal to 1 if point j is a median,
for this feasible solution, we would have X 3 3 equal to 1 and X 5 5 equal to 1. Now,
586
there are 8 possible values of X j j and if there are m general points, there are m possible
values for X j j.
So, the first constraint would be that sigma X j j, j equal to 1 to n is equal to p, so in our
example, j equal to 1 to 8, X j j equal to 2 and this feasible solution has X 3 3 equal to 1,
X 5 5 equal to 1, so for p equal to 2, this will be satisfied. The next constraint is the non
medians 1 2 4 7 6 and 8 will have to attached to some median. So, we write this as a
constraint, which is like this, sigma X i j, i equal to 1 to n equal to 1, this is j equal to 1 to
n for every i.
Now, there are 8 points here, now each point is attached to only one other point, which
could be itself or it could be some other point. So, as far as this solution is concerned, X
1 5 is equal to 1, X 2 3 is equal to 1, X 3 3 is equal to 1, X 4 5 is equal to 1, X 6 3 is
equal to 1, X 7 5 equal to 1 and X 8 5 equal to 1. So, if I take a particular i then for that i,
it has to be attached to only one j, which could be itself or it could be some other point.
Then we have another constraint, which is called X i j, should be less than or equal to X j
j for every i j. This means, that a non median i can be attached only to a median, this tells
that, every point whether it is a median or a non median, is attached to only one point. If
it is a median then it is attached to itself, which comes out of this. If the point is a non
median then it can be attached only to another median, but it is attached only to one point
comes from here. And from here it says that, if it is a non median, it has to be attached
only to a median for example, if we take this solution, 1 is a non median.
If 3 and 5 are declared as medians, we cannot have X 1 2 equal to 1, because X i j as less

than or equal to X j j. So, since 3 and 5 are medians, X 3 3 equal to 1, X 5 5 equal to 1,
but X 2 2 is equal to 0. Since X 2 2 equal to 0, we cannot have X 1 2 equal to 1, X 1 2
has to be 0, because X i j is less than or equal to X j j. Therefore, a non median point 1
can only be attached to either 3 or to 5, so it will be attach, so either X 1 3 will be equal
to 1 or X 1 5 equal to 1, so that way for a non median point 1, this will be satisfied, this
would also be satisfied.
This will ensure that, a non median point 1 does not get attached to more than one point
or more than one median. This will ensure that, it will get attached to only one median
and cannot get attach to a non median point, so all these are taken care of, by this set of
constraints. So, the objective function now is to minimize d i j X i j, where the moment
587
there is a link X i j, there is a distance that is going to be travelled. So, we want to
minimize the sum of the distances, so obviously d j j is 0, distance between a point and
itself is 0, so we want to minimize the sum of d i j X i j.
So, in this example with 8 points and 2 medians, there will be 6 distances whose, sum we
try to minimize. Or essentially, what we try to minimize is, if there are, these are the two
median points and if we take the remaining 6 non median points and assign them to the
median with minimum distance, we are automatically minimizing the d i j X i j. So, the
objective function double sigma d i j X i j would take care of, minimizing the distance
between the non median points and the median points under the assumption that, every
non median point is attached to only one median point and that distance is taken.
So, as a binary integer programming, so X i j equal to 0 1, so if there are n median points,

there are n square variable, n square decision variables. Now, there are n constraints here,
one for each point, there is one constraint and there are n square constraints that come
from here for every X i j. So, this will be 1 plus n plus n square constraints in this
formulation. This formulation can also be seen as, given 2 seed points, the problem is to
allocate the rest of the non medians or non seed points.
The medians acts as seed points, so given 2 medians or p medians, the problem of it is to
find out, how the non medians are allocated to the medians such that, the distance is
minimized. So, the thumb rule is, allocate or assign each non median to the closest
median. So, given the median points, getting this other rest of the solution is very easy
and it is not very difficult at all. The real issue in the p median is, how do I find the
correct median points?
Now, if there are 8 points and if we want 2 medians, we are looking at 8 C 2 ways of
actually creating the medians. And for each median set that contains p medians, the
remaining n minus p non median points are allocated to closest median. So, the problem
is about, essentially enumerating if necessary, all the n C p ways of getting p medians
out of n points.
Now, this problem is also a difficult problem to solve from a binary integer programming
point of view and many times, heuristic solutions are given to solve this problem. So, we
will look at one heuristic solution to this problem and also we will compare that heuristic
solution with an optimum solution to this problem.
588
So, let us take 6 points, instead of 8 and the distance matrix among these six points is
given here. Now, we can assume symmetry, so distance between 1 and 3 is 18 distance
between 3 and 1 is also 18. So, let us try and solve this problem for p equal to 2, now if
we take for example, 1 and 2 as the two median points then the non median points 3 4 5
and 6 will have to be allocated to the two median points 1 and 2. So, for 3, the minimum
distance 3 to 1 is 18, 3 to 2 is 22, so the two median points are here.
The non median point 3, 3 to 1 is 18, 3 to 2 is 22, so 3 goes to this with value 18. 4 to 1 is 14,
4 to 2 is 18, so 4 goes to 1 with 14. 5 to 1 is 16, 5 to 2 is 30. So, 5 also goes here with
16. 6 to 1 is 12, 6 to 2 is 26, so 6 also goes here with 12. So, we get the solution with
facility located in 1, will meet the requirements of 1 3 4 5 and 6, facility created at 2 will
meet the requirements of 2 only and the total distance will be 8 plus 4, 12, 18 20 2 6, so
total distance will be 60, if we look at this solution.
Now, let us look at another solution for p equal 2, now we have not given the coordinates
of the 6 points, but we are given the distance matrix among the six points, but let if we
assume that, there are some six points, which are like this. And we are interested in two
medians, if we choose the two medians here then we realize that, we may not get a very
good solution. Whereas, if we choose one of these as median and one of the other three
as another median then we automatically get a solution, which is slightly better than
choosing two medians that are near each other.
589
So, a simple thumb rule is to choose medians that are far away from each other so that,
they can attract other points, which are nearer to them. So, using this, we can look at the
largest number in this matrix, which happens to be 32 for 3 and 4, so choose 3 and 4 as
the two medians. So, the two medians based on a thumb rule will be X 3 3 equal to 1 and
X 4 4 equal to 1. So, thumb rule is to choose two points, which are farthest away as two
medians.
Now, when we do this, we have the two medians 3 and 4, the non medians are 1 2 5 and
6, so 1 to 3 1 to 4, so 1 will go with 4 with distance equal to 14. 2, 2 to 3 and 2 to 4, so 2
will also go to 4 with distance equal to 18. 5, 5 to 3 is 20, 5 to 4 is 20, so let us put 5 here
with 5 to 3 as 20 and 6, 6 to 3 is 22, 6 to 4 is 22. So, let us put 6 here, so 6 is 22. So 18
plus 14 is 42 62 64, distance equal to 74. So, we would now end up locating two
medians, which are at 3 and 4, the location at 3 will meet the requirements of 3 5 and 6,
the location at 4, would meet the requirements of 1 2 and 4.
Now, if we compare this arbitrary solution here with this, we realize that, this seems to
be better than this. So, the thumb rule base solution did not work very well for this
particular type of data. Now, if we try and solve this problem as a binary integer
programming problem with, as I said n equal to 6, so there are 36 decision variables.
And then we have 36 plus 6, 42 plus 1, 43 constraints we would get the optimum
solution X 1 1 equal to 1 and X 2 2 equal to 1. So, the optimum solution has 1 and 2 as
590
median points and we have already calculated the assignment of the non median points
with 1 and 2 as median points and the solution is shown here. So, with 1 and 2 as median
points, the optimum solution is 3 4 5 and 6, all of them get attached to 1 and 2 is by is
itself. So, point 2 will meet the requirement of customer number 2, which is itself and
point 1 will meet the requirements of 1 3 4 5 and 6 with distance traveled equal to 60.
So, the p median problem also can be solved, let us say for 3 groups, in which case we
would put X j j equal to 3. So, 3 out of the 6 will become medians and the remaining non
medians will then be attached to the chosen medians. Now, in that case, there are can be
6 C 3 ways of trying to fix the medians, after which fixing the non medians becomes
easy, because each non median will get attached to that median point, with which it has
the least distance.
But then if we want this thumb rule based algorithm to find out based on a thumb rule,
three different median. It is customary to take the largest distance and those two points
as the first two medians. Now, going by what we saw here, all the three medians have to
be away from each other, for them to attract points that are nearby. So, the third median
that we should choose should be such that, it is far away from the first two chosen
medians. So, generally another thumb rule is used, where if points i and j have distance
greater than the average distance of this then they qualified to be medians.
So, using this thumb rule, we can chose three points such that, their mutual distances are
all greater than the average, that way three median points are chosen and the non
medians can go and get attached to the medians. In any case, solving this problem
optimally for p equal to 3, would give us the best allocation for a three median problem
out of this given data. Now, there are two more aspects that have to be looked at in the p
median problem.
Now, one of which is, the p median problem is also seen as a way to group points or to
group entities, where distances among these points or similarities among these entities
are given. Another way of looking at it is, I have 6 points here, I want to make them into
two groups or two clusters as they are called such that, the distances are minimized.
Now, we would say after solving a p median problem that, points 1 3 4 5 6 will become
one cluster and point 2 will become another cluster.
591
Now, if distances are not the criteria for clustering and some similarities are the criteria
for clustering then the p median problem now will become one of maximizing s i j X i j,
where s i j would be the similarity and the rest of the constraints will remain the same.
So, p median can be seen as some kind of an optimization problem to cluster points or
entities, that is another application in use of the p median problem. Second is, as I
mentioned in the beginning, when we looked at say these six points, assuming that these
are the six points, we are not looking explicitly as what are the demands.
We are also assuming that, if we have these two points as medians, say and locate
facilities, the cost of locating the facility at any median point is the same. And therefore,
it is not going to influence the choice of the location, the choice of the location is merely
influenced only by the distance between the median and the non medians, which also
means that, the cost of location in any of the six points is the same. The demands are not
explicitly taken, which is also can be assumed that, the demands are the same in all the
six points.
Capacities are not explicitly taken which means that, we have assumed that, capacity is a
either infinite or each median when a plant is located there, has the capacity to handle the
requirement of all the six points. So, p median can be seen as a location model with
certain assumptions that, the cost of locating in every facility is the same, demands are
equal and when a facility is located, it will have the capacity to meet the requirements of
all the customers or all the points.
Each point here acts like a customer, so obviously the next location model will be like, if
I have some demand points and if I have some potential location points, where do I
locate, with some restrictions on capacity, with some different requirements at different
points and with different costs of locating the facility. Now, such a problem is called the
fixed charge problem or it is called a location allocation problem that we will see now.
592
Now, in a fixed charge problem or a location allocation problem, we may say that, say
we have three potential locations, where we want to locate a facility and let us say, there
are 8 customer points that are there. Now, let us say that, if we choose to locate a facility
in point i, we choose to locate a facility in point i, i equal to 1 to 3, there is a fixed cost of
locating a facility, which is given by f i. We could also say, that there is a capacity which
is C i, which is the maximum that this facility can produce.
Now, there are say 8 demand points and there is a demand D j in point j, which has to be
met. Now, the products are made manufactured in these facilities and if let us assume a
single product for the sack of convenience then there is a unit cost of transportation
between i and j, which is C i j. Let us assume that, this is upper case C, which represents
the capacity and this is the lower case C i j, which represents the unit cost of
transportation from i to j.
Now, the question is, if I want to locate say 2 facilities such that, I have enough capacity
to meet the demands and I am able to meet the demands at minimum cost of locating the
facility and distributing it or transporting it from the chosen facilities to the demand
points.
593
(Refer Slide Time: 34.35)
Now, for this problem, the formulation is as follows, Y i equal to 1 if location i is chosen
and X i j is the quantity transported from i to j. So, the objective function has two costs,
costs of location and costs of distribution. Now, this distribution is also called allocation,
where in the located facilities, how much of this capacity that is available in the located
facility, is allocated to meet the demands of each of these points.
So, this problem is also called location allocation problem. It is also called a fixed charge
problem, because this problem explicitly assumes that, there is a fixed cost or a fixed
charge of locating facilities. So, it is called fixed charge problem, it is also a called
location allocation problem. So, the two components of the objective function are the
fixed cost, which are the location cost and the transportation cost, which are the
allocation costs.
So, the location costs are sigma f i Y i, i equal to 1 to m, there are m possible places, Y i
equal to 1 if it is chosen and therefore, there is a fixed cost, plus there is an allocation
cost sigma C i j X i j, i equal to 1 to m, j equal to 1 to n. So, here, n is equal to 8 and m is
equal to 3, so whatever is allocated from i to j, there is a transportation cost. Now, we can
have a constraint, if we say that, we want to allocate a exactly 2 facilities then we would
have a constraint sigma Y i is equal to 2, i equal to 1 to m.
Now, with every facility, there is a capacity C i, so whatever goes out of the facility,
should be less than or equal to it is capacity. So, if I take this facility then X i j
594
summation over j, whatever goes out of this to the various destination points should be
less than or equal to it is capacity. So, this should be less than or equal to C i Y i, the C i
is the capacity and Y i comes, because only when Y i is equal to 1, this capacity is
available when the facility is located there.
So, if this is not chosen then nothing can out of it, because the right hand side will
becomes 0, all the corresponding left hand sides will have to be 0. Now, as far as each
demand point is concerned, whatever it receives should be greater than or equal to the
demand there. So, sigma X i j summed over i is greater than or equal to D j and then we
will have Y i and X i j - Y i is binary and X i j greater than or equal to 0, X i j need not
be declared as a binary variable.
So, this is the basic location allocation problem or a fixed charge problem. Many times it
is not necessary to have this constraint, whereas this is an additional further binding on
this. Sometimes, unless we are very sure about this constraint, this constraint can actually
harm the optimization. For example, if sum of all these demands, let us say is 10000 and
let us say the capacities are 4000 each, so automatically we need to locate in all the three
places.
If we put a restriction that we need only 2 then it will be becoming infeasible, because
the capacities will not able to meet the demands. So, unless we are very sure about
capacities as well as demands, or unless we have some very binding constraints that we
do not want more than a sort of number then we do not use this. We use this, only when
we have some other additional restrictions that we want locate exactly a certain number
or we do not want locate more than a certain number, otherwise this can be left out and
this is enough.
It will automatically ensure that, minimum number of facilities will be located, because
locating an additional facility is going to increase this f i Y i. So, the problem will
automatically find out the correct number of facilities to be opened such that, this cost is
minimized. Now, there is another variation to this problem or variant to this problem,
now if we look at this problem, the way this is formulated. Now, we can have a situation,
where both these facilities are opened and both these facilities transport to this.
So, this does not assume that, every customer is attached to only one facility, which was
the assumption in p median. Every customer or every non median point was attached to
595
only one median point; here we can have situations, where both two facilities can
actually supply to one particular customer. Many times this is desirable, but sometimes it
is also necessary to have each customer being supplied from a dedicated facility or each
customer is attached purely or entirely to only one facility. So, we would bring that also
into the optimization.
The model slightly changes, where Y i equal to 1 if facility i is located are chosen, X i j
equal to 1 if demand j or demand at j is met entirely from i. So, the objective function
will be to minimize sigma f i Y i plus sigma C i j X i j D j, because C i j is the unit cost
of transportation, the entire D j is transported, so D j into X i j will be the total quantity
that is transported. Now, these two constraints will be, for every facility that is chosen,
summation over j D j X i j is less than or equal to C i Y i.
So, if a facility is located here let us say, wherever it supplies, it should be able to meet
the entire demand of the customers to which there is supply from this place. So, C i Y i is
the actually supply, because that supply exist only when Y i equal to 1. Now, if this one
is supplying to these three customers, when we should be able meet to D 1 plus D 2 plus
D 3, which is given by sigma D j X i j, it is less than or equal to C i Y i. Then you will
have sigma X i j equal to 1, for every j summed over i, because every demand point will
receive the entire thing only from one supplier.
596
So, sigma X i j equal to 1 for every j and in this case, Y i X i j both are binary, so this is
the variation or variant of this model, where all the demand of a customer is met from
only one supply point. And this is a more generalized version, where a supply point can
meet the demands, part demands or full demands of more than one customer. So, here in
this case, we can model a situation, where this will supply to this and this or more
importantly, this can get from this and this. Whereas, in the second model, this can get
only from one of the chosen points, so these are the two variations in the fixed charge
problem or location allocation problem. We explain this further using a numerical
illustration and we will see that in a next lecture.
597
Prof. G. Srinivasan
Lecture – 33
Location Allocation Problems in Supply Chain. Layout
In the previous lecture, we saw the formulations for the fixed charge problem, where we
looked at locating certain facilities and allocating the capacity of those facilities, to meet
the requirement at some demand points. We saw 2 versions of the problem, one version
is where a demand point can receive, from two facilities the other version where, every
demand point is met from only one facility. So, we will take a numerical example to
show, how this fixed charge problem works, now let us consider 3 locations and the
fixed cost to locate in these would be.
So, the 3 facilities the fixed cost could be taken as five million, 4 million and 4.5 million,
now the capacities or in these 3 facilities could be taken as, so they are taken as 1 million
800000 and 1.25 million are taken as the capacities. Now, will assume that, there are 8
demand points and the demand in 8 are 200000, for 1 to 4 and 250000, for 5 to 8.
598
We will also assume that, there is a unit transportation cost from the 3 proposed sides to
the 8 demand points and these are given as 4 5 5 4 4 4.2, so these are the unit costs of
transportation, so if we use the formulation that, we saw in the previous lecture. So, the
formulation will have an objective function of the type minimize sigma f i Y i plus C i j
X i j, so the f i Y i will be 5 million into Y 1 plus 4 million into Y 2 plus 4.5 million into
Y 3, the C i j X i j will be 4 into X 1 1 plus 5 into X 1 2 up to 5 into X 1 8 and so on.
We also have the constraints, we have 3 constraints for the capacity of these 3 that come
in, there will be 8 constraints of the demand, so there will be 3 constraints for the
capacity, the capacity constraint would be like this summed over X i j, j is less than or
equal to C i Y i. So, if facility i is chosen then, there is a capacity C i and things can be
transported from i, so this will be the capacity constraint, this will be the demand
constraint, for every j summed over i will be greater than or equal to D j, which is the
demand for that item. So, we can actually solve this problem, optimally and I will give
the optimum solution to it.
599
So, the optimum solution would be Y 2 equal to Y 3 equal to 1 with X 2 2 X 2 3 X 3 1 3

4 is equal to 200000 X 2 8 X 3 6 X 3 7 is equal to 250000 X 2 5 equal to 150000 X 1 5
equal to 100000. So, if you look at this solution, the optimum solution has picked Y 2
and Y 3, so it has picked this as well as this, so there is a cost of 9.4 plus 4.5, 8.5 million
will be the fixed cost and the locations are now these 2 are chosen, so this and this are
chosen.
So, the capacities are 800 plus 1 2 5 0, so 2 0 5 0 0 0 0, so available capacity equal to 2 0

5 0 0, the requirement is 200000 into 4, which is 800000 plus 250000 into 4, which is
one million, so the requirement will be 1 8 0. So, total requirement is 1.8 million
whereas, the total capacity is 2 0 5 0 0 0 0, so we have enough capacity to meet the
demands of these 8 places and let us see how the capacities located to the various
demands.
So, we have X 2 2 equal to 200000, so this will give 200, I am just writing 200, but it is
200000 X 2 3 is 200, so there is another 200 here, X 3 1 is 200, so X 3 1 is 200, X 3 4 is
200, X 2 8 is 250, X 3 6 is 250 and X 3 7 is 250, 2 5 is 150 and X 3 5 should be 100000.
So, this should not be 1 2, this is X 3 5, X 3 5 should be 100000 1 5 would not come
here, because facility one is not chosen only 2 and 3 are chosen, so the transportation can
happen only from facilities 2 and 3.
600
Now, we realize that, if we see this capacity, this capacity is 800000, this capacity is 1 2
5 0 0 0 0, now we realize that, this 800 capacity is utilized like this 200, 200 remember
these are in 1000s, so 200 plus 200 400, 550 plus 250 is 800, 800000 is utilized. Here, it
is 200, 200, 400 plus 100, 500, 500 plus 1000 is 1500 plus 500 is 1000, so there is a
excess capacity that is left here, so the demand for all those is also met by this solution.
So, the cost corresponding to this solution has 2 components, one is the fixed cost
component, the other is a variable cost component, the fixed cost component is a sum of
these 2, which is 8.5 million. The variable cost component will be the multiplication of
these and the final value, when we find out is total cost at the optimum is equal to 1 5 5 1
5 0 0 0 comes as the optimum cost of this.
Now, there are this is the solution that, we get if we formulate this problem as an integer
programming problem and solve now sometimes heuristic solutions are also sought for
this problem, where the heuristic solution can be of 2 types. One is to try and find out in
terms of cost that is to try and locally optimize this portion, which is the total fixed cost
that is required, which means to choose facilities that have lower fixed cost that together
can meet the total demand.
So, which would give us these 2 facilities and once these 2 facilities are chosen, this one
is not chosen, take only these 2 and solve a transportation problem with the capacity of
these facilities as the supplies and the demand taken there and we would get a solution to
that. Another way of solving it is, to try and solve a transportation problem first, looking
at all of these and then, we will try and get some facilities being used, some facilities not
being used. We may in this case get a solution, where if we solve it as a transportation
problem first we may still have some allocations from here, here and here.
Which means, if we solve the transportation problem optimally, it might ask us to send
something from this, to send something from this and to send something from this, it
may happen. In this particular example, it may not happen, because these 2 put together
have enough capacity to meet and all the costs here are greater than or equal to the cost
except this one. So, it may still not happen, but then if we have different types of data
and if we solve a transportation problem, it might give some allocations here, some
allocations from here, and some allocations from here.
601
But, then it will be costly to create all the 3 in all the 3 places, because there is going to
be both excess capacity and very high cost, then from the transportation solution, we can
do some local optimization, to try and reduce this cost further. But, for smaller sized
problem such as this, which involves only 3 potential locations and 8 demand points, it is
always easy to solve this problem optimally, using any solver and try to get the best
solution to the problem.
Now, the second variation of this, is a case where, a demand met entirely by one facility
and not by more than one facility, if we see this particular example, the demand of this is
met only from here, only from here, only from here, only from here, 5 is met both from 2
and 3. Otherwise 6 8 7 out of the 8 demands, the demand points the demands are met
from supply from only one facility and not from 2, except demand 0.5, which gets 150
from location 2 and 100 from location 3.
So, if we put an additional restriction that, the entire demand is to be met, only from 1
supply, we have seen the formulation for that, we have said C i j D j X i j, X i j equal to
1, if facility I meets the requirement of demand point j. And we also said in this case that,
this should be D j X i j should be less than or equal to C i Y i and X i j should be equal to
1, so it comes only from one facility. We have already see this formulation in the
previous lecture, we will now show only the solution to it, so in this case, what will
happen is since, there is additional capacity, this entire demand will be met from here.
602
So, the solution would now become, the only change in the solution would be that, X 3 5
will be equal to 1 and the entire demand of 5 will be met only from this, so there will be,
so this will have enough capacity to meet. This 800000 will be spent as 200 plus 200 plus
250 about 150 will not be used here, this entire 150 will move here to get 200 200 250
250 and 250, it would still be within the 1 2 5 0 0 0 0.
So, the solution there will be still X 3 5 equal to 1, which means all the demand of 5 will
be entirely met from supply 3, because the when, we consider dedicated facilities, entire
demand is met from 1 supply X i j will be equal to 1, if supply I meets the entire demand
for j. So, the solution would be X 3 1 equal to 1, X 2 2 equal to 1, X 2 3 equal to 1, X 3 4
equal to 1, X 3 5 equal to 1, X 3 6 equal to 1, X 3 7 equal to 1 and X 2 8 equal to 1.
So, there will be an additional cost and that additional cost will be this 100 and 50000
into 0.4 will be the additional cost that, we will incur, if we solve this problem with
dedicated facilities. So, this is a numerical example, which kind of tells us how the fixed
charge problem works, there is a location component, there is an allocation component to
the fixed charge problem, there are still a couple of interesting issues in solving the fixed
charge problem.
One is this f i represents the fixed cost of setting of the facility, whereas C i j X i j
represents the cost of transporting items from the chosen facilities to the demand points.
Now, there is also a D j, which comes here D j is the demand during period j, now what
is the period that, we are talking about now this is like a onetime cost of allocating or
locating a facility whereas, this is incurred every time there is transportation. So, the best
way to define f i and C i j is like this, f i is should be defined as some kind of a
equivalent cost per year in locating it.
For example, if this facility has a realistic life span or a period of use of say 10 years and
whatever is the total f i should be brought down to equivalent 1 year, now d j should
represent the demand at point j for 1 year and C i j is the unit cost of transportation. So,
that both these actually represent cost corresponding to a certain time period and
corresponding to the same time period. So, if D j is demand per month in point j then f i
should adequately represent, the equivalent monthly cost, if we were to locate this
facility in this place i.
603
So, one should not say that f i is the cost that is incurred over a 5 year life of the plant
whereas, D j is a monthly demand, the assumption of course, what we make is there is a
single item or D j represents the demand of an equivalent single item. Another
assumption that we make is that, this D j is not going to change with time, so D j
represents the demand at point j of an equivalent item, for the equivalent period.
So, both f i and D j particularly, these 2 parameters one is the cost parameter for location
and the other is the demand parameter, the one has to make sure that, they are factored
suitably to represent the period, for which this cost is computed. So, it is customary to
compute it as an annual cost, so f i is often taken as the equivalent amount per year and
D j is taken as the demand of an equivalent product for 1 year.
So, that we are consistent with the notation and the use of the algorithms to solve this
problem, now we can actually expand this fixed charge problem in the context of a
supply chain to consider intermediate facilities also. So, when we looked at the basic
fixed charge problem, we said there are plants or production facilities and there are
demand points or distribution points, now we can expand this to consider intermediate
facilities.
So, there could be factories here, there would be warehouses here or distribution centers
and then there would be retailer or customer or customer, so we would have multiple
locations for these factories. Now, there could be different possible locations for the
604
distribution centers and then, there could be several customers, which are here, now as
far as this. So, we could say that, there is a fixed cost f i of locating a factory here, there
could be a fixed cost of g j of locating a warehouse in this place.
So, then we would and there will be a capacity of C i here and there could be a capacity
of some R j here and there will be a demand of D k from customer k. So now, the
decision variables will become X i j are the quantities transported between the factories
and the distribution centers and Y j k are the quantities that are transported between the
distribution centers and the retailers. So, there is a certain fixed cost of locating the
facilities and there is the variable transportation cost.
So, the problem will be to minimize f i Y i, now let us not call it Y i, let us call it as X Y
Z i, so Z i equal to 1, if we locate a factory in place i plus g j say R j, where R j equal to
1, if we locate a warehouse in place j plus C i j X i j, one more summation. C i j is the
unit cost of transportation between i and j and X i j is the quantity transported plus C j k
Y j k, where C j k is the unit cost of transportation from this to this and so on. So, this
will be the objective function that has the location cost, as well as the transportation cost.
Now, the first thing that we require is this, so what goes out, so sigma X i j, X i j from a
particular i summed over all j, so from this whatever goes into this and there is a second
summation also here, because 2 facilities are involved. So, whatever goes out, if this
should be less than or equal to its capacity C i, so sigma X i j summed over all j should
605
be less than or equal to C i into Z i, Z i comes in only, when the facility is chosen I can
move something out of it.
So, whatever that goes out is less than or equal to C i Z i, now as far as these warehouses
are concerned, whatever it receives should also be first two things will have to happen, it
can receive and some quantity X i j only when, the warehouse is opened, otherwise it
cannot receive. So, as far as every warehouse is concerned, whatever it receives X i j
summed over i, whatever it receives, one way is to write it as should be less than or equal
to big m into R j.
So, R j is 1 only when there is a warehouse that is opened here, so only when a
warehouse is opened, it can receive something, if the warehouse is not opened then this
R j will be 0, Big M is large and positive, therefore it cannot receive it. If this warehouse
is opened then whatever it can receive right now an infinite amount of quantity, but if
there is a capacity with respect to this, which is, now I defined here.
Now, R j is a variable, now let me write it R j equal to 1, if warehouse j is opened and let
S j be the capacity of warehouse j, so one other way is to say that instead of saying big
m. So, whatever it receives should be less than or equal to S j R j. R j is the variable S j is
a known capacity that it is there, now similarly whatever leaves this warehouse can leave
only from a warehouse that is opened, so sigma Y j k summed over k should also be less
than or equal to S j into R j.
Now, whatever leaves it can leave only from a opened warehouse and whatever leaves
should be within the capacity of the warehouse, so Y j k summed over all k, so whatever
leaves from a chosen warehouse should be less than or equal to the capacity of the
warehouse and whatever reaches the customer. So, sigma Y j k summed over j should be
greater than or equal to D k should be greater than or equal to the demand of the
customer.
Now, we will have Z i R j as binary and X i j and Y j k greater than or equal to 0, which
represent the transportation quantities, so this is a generic formulation of a multi facility
or a multi level location allocation problem, where we consider 3 levels here in the
production distribution network. And we say 2 stage this is one stage, this part is another
stage or essentially, there are 3 levels and 2 sets of interactions between the factory and
the warehouse and between warehouse and the retailer.
606
Now, one can expand it by including one more level in between, which could be
factories warehouses separately distribution centers, as well as retailer and the location
will be restricted to factories warehouses and distribution centers. So, depending on how
many levels are there in the supply chain the formulation can get bigger and bigger, right
now, the formulation talks about only one item or a single item that is being transported.
So, D k is defined only as demand at customer k or at retailer, k which is for a single

item, for an equivalent item, then we can add number of items into it and this demand
will now have a second dimension, which is at customer k, for a particular item l. In
which case the production capacities will now have to be defined, for producing each of
the items, it would not be C i, it will be C i l. Then, we would also have time as another
dimension, where the demand can be given for various time periods, so D will now
become D k l m demand at customer side k for item l in time period m.
In which case, the capacities may be defined as different, for different time periods, but
same capacity for each time period, but for different items, so this would have one more
subscript that comes. Now, these capacities, which are our S j’s ,now can also be defined
item wise or can be kept as a generic upper limit on the capacity of the warehouse.
Sometimes the item capacities come into the picture, if different items are of different
sizes or if different items and products require some kind of a specialized either
equipment or infrastructure to be preserved and kept, in which case the capacity will now
become item specific in addition to the capacity in the place. So there will be one more
dimension that would come in. So, we add the number of dimensions the problem
becomes far more bigger and complicated in terms of number of variables and in terms
of number of constraints. Right now the cost or the cost of location and the cost of
distribution, sometimes when we consider multi period problems, we would even look at
cost of holding inventory and cost of not meeting the demand in that period.
So, we could add terms that are for inventory cost, as well as shortage or back order cost,
if we consider multiple time periods in this formulation. So, the formulation can become
extremely big and large, which would be would become a little difficult to solve
particularly, when we are considering a large number of facilities. We are considering
multiple items and we are considering multiple periods. So, in such cases getting the
optimum solution can be bit difficult and people would resort to using heuristic solutions
to solve this problem. Nevertheless, this is the broad frame work with, which we
607
formulate and solve location allocation problems or fixed charge problems in the
contacts of the supply chain, we now move on to the next part, which is about the layout.
Now, so far in location, we also spoke about where to locate a facility, location decisions
are one of the most important decisions, for the success of any organization. Where does
one locate the factories, where does one locate the central distribution points etcetera are
strategic decisions, which actually dedicate the success of the supply chain in
manufacturing. Now, from location, we move to layout in talks always about relative
location of facilities, so the word relative is very important, when it comes to layout.
So, layout is relative location of facilities inside a chosen plant or location, essentially it
is like, we have decided to locate a factory somewhere, now within that certain
departments have to be located. So, layout normally talks about location of facilities or
departments within a chosen plant, so there could be 4 or 5 departments and how does
one relatively locate, these 4 or 5 things inside. Now, why does a question of relative
location come into the picture, now if for example, we have 4 departments that have to
be located within a certain place or within a certain factory?
And suppose, we locate department 1 here, department 2 here, department 3 here and
department 4 here, now there are 4 places, where these 4 departments can be located,
now this is called location 1 or is usually called site 1, site 2 site 3, site 4. So, in this
608
example, we are talking about locating department 1 in site 1, department 2 in site 2,
department 3 in site 3, department 4 in site 4.
We could have another situation where, department one is located in site 4, department 2
is located in site 3, department 3 is located in site 2, department 4 is located in site 1, this
is another possibility. Now, why are we concentrating, so much on the relative location
the reason is within this is let us say the overall facility and 4 departments are located,
what we assume in all layout problems is that, there will be a material movement, among
the departments.
So, let us call some w i j, as the amount of material movement from department i to j and
this material moves within various places, now there is also a distance between sites k
and l, which we call as d k l. So, if we have this kind of a layout, then the total material
movement will be this is usually given in tons, this is in kilo meter, so this is a weight
measure, this is a distance measure. So, the material movement has a unit of ton kilo
meter or weight into distance measure, so if we have a, this kind of a layout.
Then the total material movement will be W 1 2 from 1 to 2, there is a material

movement, into d 1 2 w 1 2 into d 3 4, because department 1 is located in site 4
department 2 is located in site 3, so w 1 2 amount of material moves from 1 to 2 and
moves a distance of d 4 3, from 4 to 3 it moves. We will assume by symmetry that d 4 3
609
is a same as d 3 4, similarly whatever moves from 2 to 1 will also move the distance d 3
4, so we will have w 2 1 into d 3 4.
Now, between n 1 and 3, there is going to be a w 1 3 d 2 4 plus w 3 1 d 2 4 moves from

here and whatever moves from here, similarly w 1 4 d 1 4 plus w 4 1 d 1 4, this is the
interaction between these 2. Now, the interaction between these two will be w 2 3 d 2 3
plus w 3 2 d 2 3 interaction between these 2 the interaction between these 2 will be w 2 4
d 2 4 plus w 4 2 d 2 4 and lastly the interaction here will be w 3 4 d 1 2 plus w 4 3 d 1 2.
But, if on the other hand, we do not follow this layout, but instead we follow some other
layout like this, where say 1 goes to 1, 2 goes to 2, 3 goes to 3 and 4 goes to 4.
This is site 1 site 2 site 3 site 4, then the weight into distance ton kilo meter measure will
become w 1 2 d 1 2 plus w 2 1 d 1 2, w 1 3 d 1 3, w 3 1 d 1 3, w 1 4 d 1 4, w 4 1 d 1 4, w
2 3 d 2 3, w 3 2 d 2 3, w 2 4 d 2 4, w 4 2 d 2 4, w 3 4 d 3 4, w 4 3 d 3 4. So, if we follow
this layout, this is the total ton kilo meter, if we follow this layout, this is a total ton kilo
meter, now whichever is cheaper or less costly, we would use it. So the problem of
relatively locating them with respect to each other becomes significant.
So, in layout we normally have 2 types of layout models, some of which are called
qualitative layout models and we have quantitative layout models. Now, in a in a typical
quantitative layout model, such as what we have seen, we will have a w matrix, which is
the material movement matrix and then, we will have distance matrix. And then, we
would like to try various combinations of putting them into various sites and then
evaluating the total ton into ton kilo meter or weight into distance and trying to find out,
which is the best.
So, quantitative models concentrate on that qualitative models on the other hand
approach the problem in a slightly different, way which we will see right now. So, if we
have 4 departments. Let us say, we have 4 departments and they have to be placed in
certain places relatively, now what we do in a qualitative model is we try to define a
letter or an index, which captures the relative importance of these facilities being located
to close to each other.
610
Ordinarily, we use terms like A E I O and U, sometimes X is also used, where A means it
is absolutely essential, E would means essential, I is important O is ordinarily ordinary
closeness desirable, U is unimportant and X is not at all required or extremely
unimportant. Now, what we do is there are only let us say, there are 4 facilities
departments that have to be allocated, so what we do is we try and create a matrix, which
is made up of A E I O U and ask the experts to define, how important it is to keep these,
these two facilities close to each other for example.
We would have 4 departments 1 2 3 4 1 2 3 4, now somebody might say might fill an O

here, somebody might fill an A here, somebody might fill an X here, somebody might fill
an E here, somebody might fill an X here, somebody might fill an A here. So, here we
are not quantifying the relative importance through, the weight or material movement
and the distance. Here we ask the experts to say based on what they have seen and based
on their view as to how important, it is 2 keep these 2 facilities close to each other.
So, they would mark and an A or E or A I or A U inside this matrix, sometimes this is
symmetric sometimes, it need not be symmetric, in general it is not a wrong thing to
assume that it is symmetric. But, if we go purely by the w i j’s, there can be situations,
where the direction of flow of material would be from 1 to 2 and not from 2 to 1 and
therefore, they may not be symmetric.
611
But, by and large symmetry is assumed, so if someone marks and A or E or I or O or U
or X then, out of these A is the biggest thing. So if there is an A, you want these 2 to be
very close to each other, if there is a U or an X, you do not wanted to be close at all, you
do not want them, you want them to be far away. Now, based on these, we could have
some kind of relative location of these departments, now we look at this matrix and try
and find out there is a U between the highest is a there is an A between 1 and 3.
Now, we might also assign some kind of areas to these 4 departments, suppose we say
that after all these departments require space and let us say the department 1 requires 20
units of space, department 2 requires 18 units of space, department 3 requires 40 units of
space, department 4 requires 30 units of space. Now, we realize that 1 and 3, there is an -
so one is 20 units of space, so we could try and have a rectangle as close to as a square,
whose area is 20, so we could think of a 5 by 4.
So, we could have facility a here with 5 by 4 here, so this is a with 5 by 4 and then a 1 to
3, this is department 1 department 1 with area 25 by 4 department 3 is the next best, so
here it is 40, we could think in terms of 8 by 5 rectangle, which is close to a square. So,
we could think in terms, this is 5 and this is 4, so we could have another 8 by 5, which is
facility 3, so this is 8 by 5, this is 4 by 5, let me use a different color to show that, these
are facilities, so 1 and 3.
Now, there is another between 3 and 4, which is here, so 3 and 4 has an so 4 requires 30,
so we could think in terms of five into 6. So, there is a 5 here, we could think in terms of
now we could try and put. And now 3 and 4 as an A, so we could think in terms of the 4
coming here or 4 can come here or 4 can come here, if we want this entire thing to be a
very compact square or a rectangle then, we would put department number 4 somewhere
here with 30.
So, we could have 5 into 6, so department 4 will come here, so this is department 4 with
5 and 6, now the only one that remains is 2, so 2 as an O E and X, so E is the next best 2
to 3 is E. So, we would think in terms of putting 2 somewhere here, 2 is 18, so 2 could be
a 6 into 3 or it would be a 5 into 4 minus a small quantity, so we could think in terms of 2
coming somewhere here, because 2 to 4 is also an X 2 1 to 2 is O. So, we could do this or
we could think in terms of putting 2 above here, which could also be possible, because 2
to 3 is E 1 to 2 is also O.
612
So, putting it above would bring 1 and 2 closer to each other, so typically 1 would say
that this is how, we would have the layout for departments 1 2 3 and 4, now there is a
certain amount of subjectivity in the whole thing. But, there is still a certain reason and a
scientific way of relatively placing them, to take care of some of these, there are better
versions of the algorithm that, we just now described.
But, most qualitative algorithms are around a basic principle of trying to generate a
matrix, which captures how close, we want the departments to be and then, essentially
trying to have a layout of these, which fits as close to possible, as what as what is
represented in this matrix with regard to some of the quantitative models for layout, we
will see them in the next lecture.
613
Prof. G. Srinivasan
Lecture - 34
Quantitative Models for Layout, Summary
In this lecture, we look at Quantitative Models for Layout. In the previous lecture, we
saw one qualitative model, where the layout was made based on a qualitative assessment
of the proximity or nearness of the departments to be located. So, in this lecture, we will
look at some quantitative models for layout, we have already seen that layout is relative
placement or relative location of facilities or departments, now let us assume that four
departments have to be placed.
614
And let us call these four departments as A B C and D let us also assume that there are 4
sites which we call as 1 2 3 and 4. Now, these 4 departments there is going to be either
some material movement or some people movement or some movement that happens
among these departments. So, which we call as w i j, where w i j is the amount of
movement from i to j, and let us describe a w i j matrix between A B C and D.
Now, we have assumed this matrix to be symmetric, this says that movement from A to
B is 3 units, usually it is in tons or some kind of weight measurement, A to C is 7 units A
to D is 4 units and so on. We have assumed symmetry though in many cases it need not
be symmetric because the direction of flow of material inside a factory or a department
can be different from i to j and j to i. So, in reality this need not be symmetric, but just to
make the computations easy, we are assuming a symmetric matrix here which also means
that from B to A the weight is 3 and so on.
We also create a distance matrix amongst these 4 sites, which we call as 1 2 3 4, so the
distance matrix is given as. Now, initially we assume that all these 4 are of the same size,
and all these 4 are placed equidistant from this point therefore, if this distance is treated
as 1, this would also be 1. And distance between this and this, we do not take the
Euclidean distance, we take the rectangular distance and therefore, the distance will be 1
plus 1 which is 2.
615
So, there are 4 sites and we create what is called a d k l matrix, where d k l is the
distance between sites k and l. So, the 4 sites are written as 1 2 3 4, and the distance will
be 1 to 1 the distance is 0 or it can be a dash 1 to 2 is 1, 1 to 3 is 1, 1 to 4 is 2, 1 to 4 is 2
because we take the rectangular distance. So, this plus this we do not take the crow
flying or Euclidean distance 2 to 3 is 2, 2 to 4 is 1 and 3 to 4 is also 1, so; obviously, this
1 is also symmetric, so we have 1 1 2 2 1 and 1. Now, the problem is to locate the
facilities to sites such that, the product of the weight and distance is minimized.
So, mathematical formulation to this problem will be like this. Let X i k equal to 1 if
department i is assigned to site k. So, now, each department has to be assigned to only
one site, we assume an equal number of departments and sites, so sigma X i k equal to 1
summed over all k for every department i. So, every department i goes to only one site,
every site gets only one department X i k summed over i equal to 1 for all k, k is the site.
So, every site gets exactly only one department allocated to it and X i k is equal to 0 1,
which is a binary variable. So, either department i if department i goes to site k, then the
value is 1, if department i does not go to site k the value is 0, now these are the
constraints the objective function is to minimize the product of the w i j’s and d k l’s. So,
the objective function will be of the type, minimize w i j d k l X i k X j l this is summed
over i j k l.
616
Now, let me explain this objective function. If facility i goes to site k, and facility j goes
to site l, then the product X i k into X j l will be 1 otherwise it will be 0. So, if facility i
goes to site k and facility j goes to site l, then there is a material movement of w i j
between facilities i and j, which travels a distance d k l from sites k to l. Therefore, the
material travel which is the product of w’s and the d’s will become w i j, which is the
material movement from i to j, this much material travels a distance k to l if facility i
goes to site k and facility j goes to site l.
So, we have this objective function which has four summation terms over i j k and l, the
most important thing about this objective function is that, this objective function has a
product of two variables X i k X j l summed over all possibilities. So, this objective is not
linear because it has a product term that comes it, it is also customary to write this
instead of w i j d k l as C i j k l X i k X j l, where C i j k l represents the product of w i j
into d k l. So, the problem is to minimize w i j d k l into X i k into X j l subject to these
two sets of constraints and a binary restriction.
So, for a 4 by 4 problem we have 4 constraints here 4 constraints here, which means we
have 8 constraints and we have 4 into 4 16 variables. So, there are 16 binary variables
and there are 8 constraints for a 4 department, 4 site problem, in general for a n
department, n sites problem we have 2 n constraints and we will have n square variables.
Now this problem which has this product term or a quadratic term in the objective
function, and otherwise has assignment constraints, this is very similar to the assignment
problem in operations research. So, it has a quadratic objective function, and it has the
assignment constraints it is called the quadratic assignment problem.
617
And we can solve this formulation optimally to get the best layout for the 4 departments
and the 4 sites, given the w i j’s and given the d k l’s, w i j represent the amount of
material movement from the department i to j, d k l represents the distance travelled or
distance between sites k and l. We solve this problem optimally to get the solution 1 4 3
2 with Z equal to 64, so let me explain first the notation and how we get this 1 4 3 2.
So, in this notation now this represents the site to which facility i is assigned, so this
would mean that facility 1 is assigned to site 1. So, X 1 1 equal to 1, facility 2 is assigned
to site 4, X 2 4 equal to 1, facility 3 is assigned to site 3, X 3 3 equal to 1, and facility 4
is assigned to site 2, X 4 2 equal to 1. So, this is what this notation explains, and let us
show this computation of this 64, so now, there will be 6 interactions, interaction
between facilities 1 and 2, 1 and 3, 1 and 4, 2 and 3, 2 and 4 and 3 and 4.
So, facilities 1 and 2 the w is 3 the material movement between 1 and 2 is 3, 1 and 2 are
allocated to sites 1 and 4 respectively. So, distances between site 1 and 4 is 2, so we get 3
into 2 plus interactions between facilities between 1 and 3. So facility 1 is here facility 3
is here the weight is 7 into 1 is allocated to site 1, 3 is allocated to site 3. So, distance
between sites 1 and 3 is 1, so 1 into 1 plus facilities 1 and 4 have weight equal to 4,
facility 1 is allocated to site 1, facility 4 is allocated to site 2.
So, distance between 1 and 2 is 1 facilities 2 and 3 have weight equal to 6, 2 is allotted to
4, 3 is allotted to 3. So, distance between 3 and 4 is 1 plus facilities 2 and 4, so 2 and 4
618
have weight equal to 5, facility 2 is allocated to site 4, facility 4 is allocated to site 2
distance between 2 and 4 is 1. 3 and 4 the weight is 2, 3 is allocated to 3, 4 is allocated to
2. So, between 2 and 3 the distance is 2. So the sum of all these is 6 plus 7, 13 plus 4, 17
23 plus 5, 28 plus 4, 32 and due to symmetry, we have another two.
Actually we should be multiplying this interaction of when we take this pair, we should
look at interaction between 1 and 2 and distance between 1 and 4. And interaction
between 2 and 1 and distance between 4 and 1 by symmetry these 2 are equal, so what
we compute here is 32 should be multiplied by 2 for the purpose of the objective
function. Therefore, we get a solution with 64. So, the solution of course, can be found
out using a solver or by using several known algorithms to solve the quadratic
assignment problem.
And there are very efficient algorithms, which work on the principle of branch and
bound to solve the quadratic assignment problem. Nevertheless the quadratic assignment
problem is a difficult problem to solve, and as the number of as n increases, as the
number of departments and the number of sites increases, the computational effort to
solve this problem optimally is of exponential order. And it becomes difficult to solve
the quadratic assignment problem optimally, because the computational time will be
very high.
So, people have resorted to using heuristic algorithms which try to give the best solution,
a near optimal solution or sometimes slightly more than the optimal in terms of objective
function value. Because, this problem is a minimization problem, so we look at one
simple heuristic which people have been using to solve the quadratic assignment
problem, and that is shown here.
619
Following this notation we could start with any feasible solution and the easiest is to do
1 2 3 4. Now, please note this notation once again, this notation is it represents the site to
which facility i is assigned, so here the site to which the facility 1 is assigned, facility 1
is assigned to site 1, facility 2 is assigned to site 4, facility 3 is assigned to site 3, facility
4 is assigned to site 2. So, we could start with the easiest 1 which is 1 2 3 4, we have
already explained how to calculate the objective function value, given a vector like this
is which is what we explain here.
So, we can follow the same computation to try and get the objective function value, if we
start with 1 2 3 4 which is the most intuitive solution to begin with. So, this gives us Z
equal to 78, now we could follow some kind of a pair wise exchange heuristic, where we
exchange two at a time to get a series of solutions. So, we could start with 2 1 3 4 and on
computation we get Z equal to 64. We could do 4 2 3 1 which has Z equal to 74, we
could have 1 3 2 4 with z equal to 74, we could have 1 4 3 2 with Z equal to 64 and we
could have 1 2 4 3 with Z equal to 78.
Now, these are all some of these solutions which we could get by exchanging some of
the facilities for example, exchanging these two would give this. Now, from here we can
exchange 4 and 2 to get this, we can use this and then exchange 2 and 3 to get this and so
on. So, we could do a lot of possible exchanges to try and get better and better solutions,
620
and after a series of exchanges we realize that we could either pick this solution or this
solution, as the one which has minimum value of the objective function.
And the algorithm can terminate with the best minimum possible value that it can give,
at the moment here if we are done only this we know that we have a solution with 64,
but we do not know that it is the optimum solution. Right now, since we have already
seen that Z equal to 64 is optimum, we could say that this exchange algorithm is able to
give the optimum solution for this problem instance.
But, many times particularly with as n increases the exchange algorithm is unable to give
the optimum solution in many instances, never the less it is a good heuristic to solve the
quadratic assignment problem. Now, let us come back to one of the assumptions of the
quadratic assignment problem which is this, now first of all we have an equal number of
facilities and sites that is the first assumption that we have. The second assumption that
we have is also that, all the departments which are here require the same area they do not
require different areas.
And therefore, and all the sites are of the same area and therefore, any department can go
to any site. Now, if we want to apply this basic idea to practice where in a manufacturing
shop floor, we have different departments and they have to be placed relatively closed to
each other. So, that the w into d is minimized w into d represents the amount of material
movement that goes within the shop floor.
Now, in such cases the first thing is that the areas required by the various departments
can be different. And since, if the areas are different how do we adopt or how do we
modify this idea of a quadratic assignment problem suitably or how do we modify the
idea of the exchange heuristic suitably to get a good solution, which can be implemented
on the shop floor.
621
Now, that is answered by a very popular algorithm called the craft algorithm, now which
was created by Armour and Buffa. Somewhere in the early 60's, and CRAFT algorithm
stands for Computerized Relative Allocation of Facilities Technique. So CRAFT stands
for computerized relative allocation of facilities technique. And CRAFT is a method by
which, we also bring the areas of the various departments and then try to optimize the ton
kilometer or material distance in a layout. So, let us explain the CRAFT algorithm
through an example.
622
Now, let us assume that we have an initial layout which is like this, now let us begin
with an initial layout, where there are 4 departments A B C D, the material movement
among these departments is given by the same w i j that we have. And this is the initial
layout, where A requires 24 into 10, 240 square units of area, B requires 16 into 10
square units of area, C requires 10 into 26 and D requires 10 into 14 square units of area.
So, now, what we can do is now this is an initial layout that we create.
Now, we want to see whether this is the best layout or can we have other layouts better
than this. So, to begin with the first thing we do, is we keep this point as some kind of
reference point or origin, and then we try to find out the centroids of A B C and D. So
centroid for A will now become, so, this will become 12 comma 15 because this is the
midpoint, so this distance is 12 this is 10 plus 5. So, for A it is 12 comma 15 for B it will
be 24 plus 8, 32 comma 15 for C it will be 13 comma 5, and for D it will be 33 comma 5.
So, now based on these centroids we can now create a distance, so we can now call these
centroids as 1 2 3 4, and then we can compute a distance matrix here based on the
centroids. So, the w into d for this given layout will be 3 times this distance plus 7 times
the distance between this centroid and this centroid plus 4 times the distance between
this and this plus 6 times the distance between this and this plus 5 times distance
between this and this and 2 times distance between this and this.
Now, distance between for example, this and this will be absolute value of 32 minus 33
which is 1 plus absolute value of 15 minus 5, 10. So, it is taken as 11, so right through
rectangular distances are use Euclidean distances are not used, so distance between B and
D these centroids of B and centroids of D will be 1 plus 10 which will be 11. So, w into d
for a given layout can be calculated, and this comes to about 530 for the given layout.
Now, the next thing that we have to do is in order to see whether we can have a better
layout than this, what we will do now is, we will create a pair wise exchange by initially
assuming that A and B will exchange positions, which means A and B’s centroids are
exchanged, we could think of A and C exchanging positions, so A and C’s centroid get
exchanged, we could think of B and C exchanging. So, there centroids will exchange and
so on.
623
So, since there are 4 there are actually 6 possible exchanges, but out of the 6 possible
exchanges, we will consider only 5 of them, we will consider those exchanges where
there is a common boundary between the 2 facilities. For example, we will consider A
and B, A and C we will not consider A and D because they do not have any common
boundary or area. We will consider B and C there is a common area C and D C and D.
So, 5 changes can be considered by simply changing the centroids from A to B then this
will if we are looking at changing A and B, then this will become A’s centroid, this will
become B’s centroid. So, like this 5 possibilities exist, and then the best out of these 5 are
actually chosen, and then we realize that the best exchange happens between A and C
with value equal to 481. So, we have considered 5 exchanges that is between A and B A
and C B and C B and D C and D, by exchanging the centroids.
We have not considered A and D, because they do not have a common border, and out of
the 5 we realize that exchanging A and C gives us a good value of 4 81. But, then there
is a small approximation that we have assumed, for the purpose of this interchange when
we interchanged A and C we said A’s centroid will become 13 comma 5, and C’s
centroid would become 12 comma 15. But we actually cannot exchange the centroids
like that because the areas are different.
So, when the areas are different the centroids will change, so what we do now is we try
to incorporate that, and then we actually exchange A and C to see what happens. Now,
that is shown here, so when we actually exchange A and C. B will remain as it is, so B
will be here, D will also be here. Now, we have to exchange A and C in a way the
advantage of choosing departments that have a common border is by saying that when I
exchange A and C I have to put A here and I have to put C here.
Now, the area that is available for C can accommodate A, whereas, the area that is
available for A cannot accommodate C. So, what we do now is we try to put the smaller
area into the bigger area, and keep the rest of the area for the bigger one, so what we will
do now is, we will try and put A here now this is 41 A is 24 by 10. So, I will take another
24 for A this is 40, so 24 plus 4 38 plus 240, and C will now come here, now C loses it is
rectangular area.
Now, it is actually made up of two rectangles which are like this, now I go back and
update the centroids. Now, you see that A’s centroid is 2 plus 12, 14 comma 5, 14 comes
624
because there is a 2 here, and half of this 24 is 12 14 comma 5, these centroid will
remain as 32 comma 15 D’s centroid will remain as 33 comma 5 C’s centroid will now
have to be redefined, and that is computed as, now C is now taken as 2 rectangles. And
we use the usual formula of a 1 x 1 plus a 2 x 2 by a 1 plus a 2 and b 1 y 1 plus b 2 y 2
by b 1 plus b 2.
Now, when we use that this into centroid of this plus this into centroid of this divided by
the area, will give us the actual centroid of C. We may even have a situation, where the
centroid can go outside of the figure depending on the shapes that we actually get and
therefore, in this case C’s centroid will now become 11.15 and 14.23. If we see carefully,
the actual exchange has not happened 13 comma 5 has become as 14 comma 5, and 12
comma 15 has become 11.15 and 14.23.
Now, we have another layout which is like this, now when this layout we now try to
make exchanges. Now, we can try and exchange A to B, we can do A to C, we can do A
to D, we can do B to C, we will not do C to D, but we can do B to D. So, we can try 5
more exchanges by simply first exchanging the centroids, and then finding out of
something is better. And after that actually making the exchange and updating the
centroid.
So, this process can be continued till we actually get the best solution, and in this case
the best solution would give us. Actually we realize that this would be the best solution,
we are unable to get further improvement with this and therefore, we could actually
terminate with this layout for this particular departments C B A and D. So, this would be
the final output for the CRAFT algorithm, and we would now understand that we could
have C and then B and then A and D, which are actually there.
Now, the algorithm would terminate with this kind of a layout, but then we also realize
that having this kind of a layout is a little cumbersome. Because, C does not have a shape
which is completely rectangular, and C has a different kind of shape, now for all
purposes of control it is actually good to have all A B C D having rectangular shapes,
and rectangle closure to a square. So, when it comes to actually implementing it we
could think in terms of you know making some very small changes, and then checking
out whether can we add some more area here and bring it, so that we can expand see a
little bit.
625
So, we could have C occupying something like this, and then maybe we could think in
terms of D being here, and A being here. Rather than have C come up to this, but then
one has to understand that if we put C and B together we are looking at 26 plus 16 which
is 42 instead of the 40 that is available. So, then we explore and see whether we could get
some more area here for the department, and not have the entire thing as the rectangle.
But, then each department actually has a rectangular area for itself.
Now, this is how the CRAFT algorithm works. CRAFT algorithm has a couple of
advantages one is CRAFT is able to consider different area requirements for different
departments, which was not explicitly considered by the quadratic assignment problem.
Second CRAFT is a simple algorithm, it is a heuristic algorithm it is centered around pair
wise exchange of areas of the various departments. So, it does not guarantee the optimal
solution always, it is a heuristic solution.
Sometimes by the very nature of updating the centroids, we could get you know shapes
like this. But, as I mentioned common practice is to convert them into manageable
rectangles, so that a layout like this can be finally achieved even though the craft layout
is like this. Now, CRAFT is a very significant contribution to the literature in operations
management and layout because CRAFT also had a computer program written in the
early 60's, which could actually generate and one could print a layout, using the
computer program for the CRAFT.
626
For this reason CRAFT became an extremely popular algorithm, even though CRAFT
has certain simple limitations. Like the fact that it is not an optimal algorithm, it is
essentially a heuristic algorithm, and sometimes it would give us some kind of
complicated shapes like this. CRAFT also can be modified and has been modified to
include three way exchanges, and several modifications to CRAFT have also happened.
In this lecture series we have touched upon very basic algorithms in the area of location
and layout, location and layout as an area is extremely important and as mentioned takes
care of strategic decisions, in the context of manufacturing and supply chain. There are
several algorithms both qualitative and quantitative that researchers, and practitioners
have created over the last 50 years or so which will try and meet the requirements of
location and layout.
But, then we have addressed very basic algorithms, which capture the very essence of the
ideas in facilities location, and facilities layout. So, let us now have a quick recap of what we
have actually covered, and are there few other things which can be touched upon before, we
kind of wind up this part of operations management in this lecture series. Now, the other
things that one could look at is in terms of layout, what we have actually done here is we
have tried to model what is called a functional layout into a given area.
Now, when we say that there are 4 departments A B C D, we assume that there is
departmental specialization. And these departments have interactions amongst them, so a
typical functional layout, will be laid out this way, where A B C D would be four
departments in a functional layout. Much later if we see the advances, the different types
of layouts are would be the line layout, and the functional layout. Line layout is also
called the product layout, where according to product the machines are laid out.
Functional layout is called the process layout, so the machines are laid out according to
the manufacturing process.
627
Now, much later people moved on to what is called cellular layout or cellular
manufacturing, about which we saw briefly in the introductory lectures to this lecture
series, where machines different machines are grouped together, in order to process a set
of parts which are called part families, so in the cellular manufacturing and cellular
layout, a set of parts which can be made by a set of machines, the parts would be similar
the machines would be dissimilar functionally.
But, they will be capable of making everything required for these parts are all grouped,
and each machine in part group is called a cell. So, the layout move towards what is
called cellular manufacturing, where the cells where created and cells were also kept
separate. The difference is in this context of cellular manufacturing, we do not consider
each cell like a department here because material movement among the cells is far
minimal compared to the material movement, among the departments in a functional
layout.
So, ordinarily we do not try to optimize the ton kilometer distance amongst the various
cells because the very principle of cellular manufacturing the material movement among
the cells is very, very less. So, the basic ideas in cellular manufacturing is more to do
with cell formation than to do with how to layout the cells; however, depending on the
unidirectional flow in the cell, depending on some intracellular movements, depending
on the flows of various parts that go in the cell.
628
Cell layout is also an important aspect to be touched upon, though we do not attempt to
go deeper into cell layout in this lecture series. So, let us come back from the beginning
and see what are the topics that we have actually covered in this lecture series.
So, we started off with an introduction to, we started off with an introduction to
manufacturing and operations management. We also described what was the basic
challenge that today’s manufacturing organizations face, and challenges that are there
faced by organizations, which are part of any supply chain. And then we moved into
basics of operations management, we started with forecasting and we explained a lot of
time series models for forecasting.
And then from forecasting we moved to aggregate planning, where using the demand for
the products, we made decisions on how much of regular time, overtime capacity that is
being used. And from aggregate planning we moved to disaggregation, which gave us
how the time that comes from an aggregate plan is going to be used to make different
products. Then we also saw a little bit of lot sizing here, and then we moved to inventory
problems, where we saw deterministic inventory, order quantities.
And then we saw the reorder levels, and the safety stocks, and we discussed some
probabilistic inventory models, which also spoke about the safety stock that is required.
We then moved into production control, within the inventory we also saw inventory with
time varying demand, we saw some aspects of production control, we saw flow shop
629
scheduling, we saw job shop scheduling, we also saw line balancing. There are two other
aspects of production control, which we have not touched upon in this lecture series
which we would possibly in different courses under NPTEL.
And those two aspects of production control are called JIT or Just in Time
manufacturing, which also deals with kanban controlled manufacturing systems. And the
other is called theory of constraints or synchronous manufacturing, where the production
control mechanism is called the DBR mechanism or Drum Buffer Rope mechanism. So,
these two parts of production control, we have not touched in this lecture series.
Then we saw some aspects of layout and location, we saw the fixed charge problem or
location allocation problem, which is also called the network design problem. And we
have seen some aspects of layout, such as the craft algorithm and the quadratic
assignment problem. So, what we have seen, so far are from the context of supply chain
we could call major decisions in supply chain as location decisions, production
decisions, inventory decisions, distribution decisions and information related decisions.
So, from a context of supply chain we have seen location decisions here, where we spoke
about different aspects of location and layout, we seen production decisions. Production
planning decisions, as well as production control decisions, and we have seen aspects of
inventory, so we have seen inventory decisions.
630
And what we need to see are distribution decisions, and information related decisions in
the context of a supply chain. So, we will spend some time on the distribution as well as
information related decisions, in the context of supply chain management in the
subsequent lectures in this lecture series.
631
Prof. G. Srinivasan
Lecture - 35
Introduction to Supply Chain Management
In this lecture, we formally define Supply Chain Management, introduce the various
concepts, and discuss relevant issues in supply chain management. What we have seen
so far in the earlier 34 lectures are principles and concepts in operations management.
Essentially, things related to what happens within an organization, even though there
were instances, where we have seen models, were multiple organizations or the role of
multiple organizations, where captured. Now, we define supply chain management, and
also try to position what we have seen in the previous lectures, in the overall context of
supply chain management.
We have already seen the slide in an earlier lecture. Let us quickly recap the importance
of this slide, now this slide talks about the requirements of manufacturing. Let me read
it, make an increasing variety of products, on shorter lead times with smaller runs and
flawless quality, improve return on investment by automating and introducing new
technology in process and materials so, that price can be reduced to meet local and
foreign competition, mechanize keep schedules flexible, keep inventories low, keep
632
capital cost minimum, and keep the work force contended. This is a definition given by
skinner in 1985. If we look at this definition very carefully, there is a role for
manufacturing, there is a role to increase the variety of products, there is a need to
produce them in very short lead times.
There is also a need to automate and bringing new technology in materials, to meet
competition, and there is a requirement to keep the schedules flexible, and to keep the
work force contended. So, most of these relate to principles and requirements of
manufacturing, but there is also a role or different organizations or multiple
organizations, with whom the manufacturing firm interacts to make these variety of
products.
So, in order to be successful in the business, it is important to meet these requirements of

manufacturing, and also do this with the partners, who are suppliers who are customers
and who are other stake holders. So, that the overall performance of the organization, as
well as its partners is made better, and this chain of interaction between the organization
which makes these products, as well as the suppliers and customers is broadly the supply
chain that we are talking about.
Researchers and practitioners have been talking about the requirements of manufacturing
from the context of meeting customer expectation, from the context of reducing the cost
of production, and the context of maximizing the profit. If required some time and over a
633
period of time, several methodologies have evolved to meet the requirements of
manufacturing, four important aspects and concepts quality delivery and flexibility.
And it is also important to understand that these are not to be traded off against one
another, but to be simultaneously prioritized, which means that like if an organization
focuses on cost cutting, it cannot do it at the expensive reduced flexibility or reduced
quality. One has to keep the other three at the same level and try and increase the four or
try and increase all of them to the extent possible, but make sure that the performance
with respect to any of these parameters does not come down, because more effort and
emphasis is placed on one of them.
Broadly, let us look at some methodologies that have evolved in order to address these
metrics or performance indicators of manufacturing or of a business. Total quality
management evolved to achieve flawless quality, and to motivate people within the
organization towards producing quality goods and services, as well as to maintain the
overall quality in the organization. So, one of the metrics is quality and total quality
management evolved as a methodology to meet the requirements of quality. Flexible
manufacturing systems evolved to achieve quick response and flexibility.
Agile manufacturing, lean manufacturing and the forms of just in time manufacturing
also look at minimizing the waste, so that the total cost of manufacturing is reduced. And
supply chain management evolved over a period of time to deliver products, quickly and
maintain low inventories. So, in some sense if an organization would be practicing or
would be concentrating on all the four methodologies, that we talk about TQM, FMS,
Agile, Lean manufacturing, and Supply Chain Management, so that cost flexibility
quality and delivery are simultaneously achieved.
634
Now, let us move to formal definition of supply chain. Several definitions for supply
chain exist, one of them we are going to concentrate on, and this definition is as follows.
Supply chain is a network of facilities and distribution options that performs the function
of procurement of materials, transformation of these materials into intermediate and
finished products and distribution of these finished products to the customers.
So, let me repeat it again, it is a network of facilities and distribution options that
performs three tasks. One is procurement, second is manufacturing or transformation of
these materials into products, and third is distribution, which is to deliver the
manufactured products to the ultimate or end customer. This is the definition given by
Lee and Billington, and is widely used in text books and in other research material.
635
Now, we will quickly see who are the players in the supply chain. So one could think of
3 different or 4 different types of players in the supply chain. Now, there are a set of
suppliers, which is shown first here, and then there is the manufacturing or the
organization, which is central which is making the product that we are talking about. And
then the produced goods are sent to customers through a channel, so it could be a
distribution channel and from there it goes to the final customer. We could actually
combine both these into a single thing called customer or distribution.
So, to fit in with the earlier definition of procurement, transformation of the products
into manufactured goods, and to customers who are there. Now, the manufacturing that
we show here, we also have written that the manufacturing has internal suppliers and
customers. The transformation of the purchased item or products into manufactured
products takes place in stages. And if we look at these stages we would realize that each
stage acts as an internal supplier - customer within the overall manufacturing, which is
shown in as internal suppliers and customers.
And then it moves to external customers, so from external suppliers, it moves to the
internal suppliers and customers, and then it moves to the external customers and then it
reaches the final customer. Now, we also see here the supply chain cuts across
enterprises; there are lots of organizations many organizations who come under these
suppliers. Even though, we have shown this as a single chain which just connects 3 or 4
636
entities, when we say suppliers here we mean a whole lot of suppliers that supply items
to this manufacturing organization, so if we talk of a supply chain with respect to a
single manufacturing organization.
Now, there is one organization here, there are many organizations here who are the
suppliers, and there are many organizations here who may the final customers, and there
would be few who would be in the distribution channel. So, if we are very specific with
respect to this manufacturing organization, then you will have several nodes here as
suppliers, who will be sending items to this.
And from here it will move towards several external customers, it will move this way to
many people who are here, and then it will move this way to that. One can also observe
that if we look at a overall supply chain, where these set of supplier may even be
producing to more than one in this and so on; but ordinarily when we look at supply
chain management, from the context of an organization, then we assume that there is a
single organization here .
We have also seen that within this manufacturing, there are several stages where each
stage will act as a supplier and customer to the next stage, as well as the previous stage
respectively. So, supply chain cuts across enterprises, and an important thing to note here
that it is not one company or organization here, it is not one distribution mechanism here
or one customer here. This has several companies, several organizations that come here,
and therefore supply chain is a network of suppliers, if we go back to the definition it is a
network of facilities and distribution options.
Now, what are the reasons for us to look at supply chain holistically and supply chain,
which is a network of organizations together and not concentrate on how an individual
organization optimizes or increases the performance of the particular manufacturing unit.
Now, the reason to look at supply chain holistically or as a network of these facilities and
organizations are the following.
Now, the first one is shrinking product life cycles, traditionally product life cycle were
longer, any product that came into the market survived in the market for a long time, but
with increase in computers and electronic goods. We realize that the product life cycles
are shorter, and therefore it is necessary to bring out a new products in time, to bring out
new products before the existing products become obsolete or to increase product variety
637
by bringing in new products, which would eventually make the older products
progressively obsolete.
So, product life cycles are shorter, and there is a need to bring new products quickly.
Now, when there is a need to bring new products quickly, we go back to previous slide
we realize that if this entity which is the manufacturing has to bring out new products
quickly. Then there is a definite role for this entity which has a set of suppliers, as well
as there is a role for this set of entity who are the in between the ultimate customer and
manufacturer.
Now, suppliers have to be ready to supply the items that are needed for these new
products quickly, which means they will have to integrate and coordinate with the
manufacturer. And once these manufactured products are ready for dispatch or to be sent
to the customer, now the distribution channels have to be ready to do it quickly. So, there
is the need to bring these various organizations together, and collectively plan and
execute this activity, so that the profits for all of them are increased and maximized.
In a similar manner, we also have shrinking time window for delivery, today once the
order is placed the customer expectation is that the delivery will be quick. So,
organizations cannot take additional time to deliver, and therefore have to produce things
faster which also necessitates the same amount of thinking, where we have to interact
with the suppliers, as well as the customers.
638
Non shrinking lead times, one would also observe that even though a technology through
a flexible manufacturing and automation, brings down the time to manufacture a
product, the rate at which manufacturing times or processing times reduce is not as high
as the rate in which product life cycles change. So, we call this as non shrinking lead
times, while we have shrinking time window for delivery. The rate of shrinkage here is
much faster or higher than the shrinkage here, which means that one needs to take care
of these aspects as well and make decisions holistically.
Also, there is ever increasing product variety which brings in newer people here, and
which also brings in newer set of customers here. So, it is not that this set of people, who
are in the suppliers is a finite set of suppliers it also keeps changing as the product
volume and variety increases. Similarly the set of final customers here also changes as
the product volume and variety increases. And therefore, there is a need to integrate all
these into a single entity called the supply chain, and do the planning and coordination
together.
So, that the overall profits of the chain or the overall profits of all the members of the
chain increases and the performance gets better. Now, what are the additional problems
that we need to look at? First and foremost is demand uncertainty, now the once one we
get into multiple products, here we realize that as the product volumes and variety
increases. We have already seen in earlier lectures that today’s challenge is to meet the
increased volume, and variety particularly increased variety.
So, as increased as variety increases, we also observe that the demand uncertainty also
increases for a given variety. When, we aggregate the variants of a given product or we
when we aggregate the demands of the various varieties of a similar product, it may be
possible to have a forecast error which is smaller or it may be possible to have the total
demand showing less variance. But, the moment we have more products and more
variety, the demand variation increases, and the uncertainty in the demand, therefore
increases. Now, this has to be addressed, and this can be addressed only when
organizations come closer to each other.
639
Now, the order to order lead time also increases, and there is also a responsiveness
squeeze which makes organizations produce, and deliver things much faster. And there
is also a constant need to improve the profits as well as to increase, the cash flow within
the organization. So, these are reasons for which we need organizations to come
together, to do the planning and coordination together, so that the overall objective is
met for all of them.
640
Now, what are the areas for improvement in a supply chain, the several areas of
improvement in a supply chain? Now, if we are positioning ourselves as this entity,
which is manufacturing or the organization, which is going to interact with a lot of
suppliers as well as with a lot of customers. The important question is how do I improve
my inbound logistics? Now in order to manufacture we need to get the material from
several suppliers.
Now, this process of getting the material from suppliers to the organization is called
inbound logistics. Inbound because they come into the organization and logistics is the
area by which we look at transportation, and other issues related to transporting items
from one place to another. So, the inbound logistics have to be better, so that the right
amount of bought out items are made available to manufacturing. The next question that
we should ask is how will the structure of the supply chain have an impact on the overall
performance?
Now, most of the times we use this serial chain which we show here, through one
supplier manufacturing and then the customer, but then we have other structures. We
have network structures and other structures in supply chain which will have an impact
on the overall performance. Now, what are the planning tools that can bring different
organizations together? Now do we use information technology to bring them together?
But, to what extent and how do we do that and what are the advantages of having
information technology brought in? What are the things that we can share between
companies and organizations to make the planning much better?
So, these questions come when organizations talk together or work together. The same
supplier may be producing to the competitor of the manufacturing organization, and how
do we take care of that dimension, when we do the overall planning also. Now, are there
strategies to handle product variety and short product life cycle, and how do we do it
through organizations coming together is an aspect that has to be looked at when we look
at supply chain management.
Now, there are three aspects we spoke about in a supply chain, the three important things
are procurement - buying items from suppliers, transformation or converting these items
into finished products, and distribution which is to send the finished products to the
ultimate customers. So, those three aspects are now written as inbound logistics, in house
641
logistics, and outbound logistics, as I mentioned inbound means items that come into the
organization, which means they come from the suppliers.
So, inbound logistics is handled through supplier development and supply management.
How do we create new suppliers? How do we develop the suppliers or ensure that the
suppliers are able to give us the right quantity of items at the right time? In house
logistics is the whole gamut of operations management that we have seen through the
earlier lectures. Most of what we saw in the earlier lectures have to do with how an
organization handles things, which are directly within the scope of the organization.
Example, we saw forecasting, we saw aggregate planning, we saw scheduling, and we
saw all other aspects.
So, all these aspects, so if we now go back to the previous lectures in this course, we
understand that many of the things that we have seen, so far now, deal with what comes
under in house logistics or what happens within a manufacturing enterprise. So, there
could be scheduling, then master scheduling, materials requirement planning, layout
inventory management within the organization. So, all these come under what are called
in house logistics. Outbound logistics talks about, how do I distribute or how do I send
the finished goods from manufacturing, to the end customer.
So, here the aspects are warehousing and distribution management, transportation some
understanding of how we can optimize the costs with respect to distribution, and
transportation, we will see some aspects of this as we move along in the remaining
lectures in the course. And we have seen a little bit of outbound logistics, when we have
seen location models that talk about how to locate warehouses, and finally send the
products through these ware houses into the final customers. We will also be seeing
some more aspects of warehouse location, and other location problems in the remaining
lectures in this course.
642
Now, this picture tells us the different layers or levels in the supply chain. You can see
four of them, you can see suppliers, you can see manufacturing, you can see distribution,
and you can see the final customer. So, these are four layers in the supply chain, as I also
mentioned, if we are going to look at one organization or the organization, which is
doing the manufacturing. There will be several suppliers, there will be few distribution
channels, there could be few warehouses, and there will be many customers through
which the product or products will flow.
Now, the material or product will flow from left to right, so from the suppliers to
manufacturing, from manufacturing to distribution, and from distribution to the
customers. Cash flow will happen in the opposite direction. The customers, who buy the
product, will give cash to the distribution. Distribution would provide cash to
manufacturing and manufacturing, would finally provide cash to the suppliers.
So, the cash will flow in the opposite direction compared to the material flow.
Information flow which we were talking about for the first time will happen in both
directions. We have already seen that the supply chain is successful or the overall
business of an organization can be made better, through interacting with the suppliers
and the customers, this we call as the supply chain management.
So, when the manufacturing starts interacting with the suppliers, there is a two way flow
of information between the manufacturing and the suppliers. Manufacturing talks about
643
how much they want from the suppliers, what quantities, what items, and then the
supplier talks to manufacturing by not only providing the material, which happens in this
direction, but also providing information on the availability of these with the respect to
the demands of the manufacturing.
So, there is a two way information flow between suppliers and manufacturing, and there
is a two way information flow between manufacturing and distribution, and then between
distribution and customers. So, information flows in both directions, unlike material flow
and cash flow, which flow in one direction and opposite to each other. Now, the biggest
challenge is while we know that we could we could get these two entities to talk to each
other, supplier, manufacturing or the other two manufacturing distribution, and
distribution to customer. There is information flow, how do we integrate them or bring
all of them together, so that there is information flow right across the supply chain is an
important challenge with respect to supply chain management.
Now, when we look at supply chain what are the broad issues in supply chain? Broad
issues in supply chain as usual can be categorized into strategic issues, tactical issues and
operating or operational issues. Strategic issue are long issues that have long term impact
or long term with respect to the time frame of the impact, tactical issues are usually
medium term, and operational issues are very short term.
644
So, the supply chain issues can be categorized into strategic, tactical and operational.
Some of the strategic issues are design of the supply chain, partnering, location and so
on. So, design of the supply chain would talk about locating the various facilities,
perhaps different products made in different countries, different items outsourced are
bought from organizations which are located in different countries and so on.
And partnering, who are the partners, who are the suppliers, how do we select them, how
do we develop them or how do we get them to meet our requirements and so on. So,
these are some of the strategic issues, some of the tactical issues in a supply chain or
what is the inventory policy across members of the supply chain, what are the purchasing
policies across members of the supply chain.
What are the production policies, transportation policies, quality policies etcetera?
Operating issues within each of them would be quality control, what are the quality
aspects, production planning scheduling etcetera come under the operating issues. Now,
each of these issues are handled through different methodologies, and these solution
methodologies are given under each of these issues.
So, strategic issues are handled using what is called a network design approach, tactical
issues are usually handled using simulation based approaches. And operational issues are
usually handled using heuristic or rough cut or rule of thumb approaches. So depending
on the nature of the issue a different solution approaches can also be used to address
these issues.
645
Then we move on to classification of supply chain decision making, the important

classification given by Ram Ganeshan and Terry Harrison would broadly classify supply
chain decisions into location decisions, production decisions, inventory decisions and
transportation decisions. We also, they have also included information decisions as the
fifth dimension, because information is the one that connects all the organizations, which
are physically located, or physically away from each other.
So, we will look at five types of decisions in a supply chain - location decisions,
production decisions, inventory decisions, transportation decisions and information
decisions. Now, within an organization if you see we have already seen some aspects of
production decisions, inventory decisions, and location decisions in the earlier lectures,
where we have looked at production planning, scheduling, inventory, and location.
We are yet to see some of the transportation decisions or models related to transportation
as well as some aspects of information decisions, which we will see as we move along in
this course. Location decisions primarily talk about the geographic placement of
production facilities, where do I place my production facilities.
646
Where do I place my warehouses, where do place my stocking points, and who are my
sourcing points, who are the companies for or organizations from which we are going to
buy items. And finally, these are solved using optimization that considers production
costs, taxes, because geographic placement across different countries, geographic
placement across different states within a country. Involve different taxes duties tariffs
distribution costs would change depending on the distance, and production capabilities
would change with different plans.
So, we could think of a complex optimization routine that considers production costs
taxes, duties, tariffs, distribution, cost production, capabilities or capacities etcetera to
finally, arrive at location decisions.
647
Production decisions are about which product to produce and which plants will produce
these products. So production decisions would mean that the plants where production
would take place are being chosen. And then the next level decision is to assign products
to these plants, at first, of course to find out which products to produce and then assign
these products to the plants.
And then how do we allocate suppliers to the plants, and how do we allocate plants to
distribution centers and distribution centers to customers. Now, important solution
methods include production scheduling, construction of master production schedule,
machine scheduling, equipment maintenance work load balancing, and quality control.
So, these are mostly things that are related to within an organization or in house once the
decision to produce the products on the plants are made, and some of these we have
already seen in the earlier lectures in this course.
Inventory decisions deal with three types of inventories, raw material inventory, semi-
finished or work in progress inventory, and finished goods inventory. Now, we bring in a
fourth dimension which is in process inventory between locations, because in a supply
chain we are going to look at interactions between various organizations.
648
(Refer Slide Time: 32:21).
So, there could be inventory which is in transit or in process between locations also. Now
inventory decisions are important to buffer against uncertainty. Now, the moment there is
an uncertainty, one of the ways to reduce the uncertainty is to handle inventory.
Traditionally inventory would mean physical items which are present. But, today
inventory can be modeled as physical items, inventory could be time, inventory could be
cash. The idea is to reduce the buffer or excess inventory, which we hold to handle
uncertainty or to address the effect of uncertainty.
So, inventory decisions are about trying to hold minimum inventory, in terms of either
units or time or cash to buffer against uncertainty. The solution methods would include
control policies optimal level of order quantities, economic order formula and different
order quantity or production quantity formulae, that we have seen in the previous
lectures. Reorder levels computations that we have seen, safety stock levels and
computations that we have seen, at each location based on given service levels.
649
Transportation decisions are the tradeoff between cost of transportation with the indirect
cost of inventory. Now, when the goods are made in the manufacturing facilities, now
they have to be transported to the end customer - now many times they are transported
through a distribution network or a distribution channel though in some instances direct
sales to customers also exist.
So, when we have this transportation through distribution channels, the important
question that arises is where do I locate the inventory? This physical inventory has to be
located somewhere. And the question is how do we or where do we locate this inventory
because there is a cost of transportation, and there is a cost of holding inventory. So,
when inventory is held in a central place, we would realize that it is able to handle
uncertainty better, but the cost of transportation would increase, because every time it
has to be transported from a central place to different customers.
If on the other hand this inventory is not held in a central place, but it is decentralized
and held in different places, then the cost of transportation may come down, but the
amount of inventory that has to be held particularly to handle uncertainty will be higher.
So, either way you will we will realize that between the two costs, which is cost of
holding inventory as well as cost of transportation, one of them will be high and the
other will be low and therefore, the question would be what would be the transportation
decisions, such that the cost of trading off the inventory and transportation is maintained
650
well, or there is a good tradeoff between the two costs, or the sum of the two costs would
be minimum. Now, what are the factors that would affect the level of inventory as well
as the extent of transportation? Customer service levels are the first one, because the
moment we have customer service levels and say that we have to meet the service level,
higher the service level more the buffer inventory that is needed which we have seen
earlier in this course. Geographic locations would determine the cost of transportation,
from the place where the inventory is currently held to the place, where the inventory
moves. Desired buffer and inventory levels will depend on the customer service levels,
and also observe the transportation costs are higher and they contribute to more than 30
percent of the overall logistics costs.
And good transportation decisions can actually bring down, the total transportation cost
other decisions would be shipment sizes routing and scheduling of equipment or product
are some of the factors in transportation decisions.
Now, we will briefly talk about something called the bullwhip effect, which is very
specific to supply chain management, and we would look at this bullwhip effect in a little
more detail in subsequent lectures. The bullwhip effect talks about the amount of
inventory in the various stages of the supply chain, and generalizes that as we move to
beginning of the supply chain or leftmost side of the supply chain, the overall inventory
will be higher.
651
Bullwhip effect also talks about the fact that while the overall inventory would also be
higher, it also talks about the fact that the variation in inventory would also be higher. A
simple explanation to the bullwhip effect is that if there is a customer demand which is
D, then the distributors will hold a demand will hold the quantity of slightly more than
D, to take care of the demand fluctuation at the customer level.
So, the amount of inventory that these people would like to hold, or would like to have is
a little more than D, and they do not produce, so they will buy from the people who
produce. And because their requirement is little more than D, and the demand that they
give to the producer is a little higher than their D, that they have, and the producer ends
up producing a little more than what is demanded by the distributor.
So, therefore, the pressure is on the tier one suppliers, and tier two suppliers, to meet this
requirement of the producer and they also end up increasing their demands a little bit, so
that the uncertainty effect is captured. So, the uncertainty in demands of the customer
finally, increases the demand on tier two suppliers, and more than that the amount of
inventory that average inventory that the tier two supplier holds is very high.
So, this is called the bullwhip effect, where the average inventory held by various players
in the supply chain increases as we move towards the left or as we move towards the
supplier direction. The average inventory held will increase, and not only would the
average inventory increase the variation would also increase, fluctuations will also
increase. While, there will be times, this will be small, this will be large, and this could
be small, this could be small, something else would be large, but the effect when we look
at the average. As we move towards the left the average inventory will increase, and the
inventory will fluctuate along the supply chain, this is called the bullwhip effect and we
would later see methods or ways to reduce the bull whip effect. And one of the ways to
reduce is by coordination, by sharing information and by bringing, these people closer to
each other from an information point of view. Now what causes the bull whip.
652
Effect? Some of these are more number of layers, higher the average inventory. Delay or
time taken to move from one entity to another and the rate of change of the fluctuations,
so more the fluctuations higher the inventory. Now, each because this happens, because
each layer in the chain updates the forecast in varying patterns, places orders at different
times there could also be price fluctuations which would determine the amount of
inventory they would wish to have.
And these price fluctuations happen, because of promotion, product promotions, and
there are times the rationing of supply also increases the inventory. Now, when there is
rationing, which means when total supply is less than total demand, all the demand is not
going to be met. And therefore, people would increase their demand notionally, so that if
they are getting a proportionate increase of supply, then it makes thing better.
So, the moment there are instances where total supply is less than total demand, and
rationing happens which means only a proportion of the demand is going to be met every
player would automatically increase their demand. So, that the actual quantity that they
get can increase, so all these result in excess inventory in the system, which causes what
is called bullwhip effect.
653
The several ways of avoiding the bullwhip effect. Some of which are use point of sale
data, rather than forecasts, electronic data interchange, and use the internet to help in
sharing information between or among the players in the supply chain. So, what are the
data that can be shared sales data, capacity data, inventory data, they all can be shared
which would effectively bring down the lead time, and will reduce the costs of ordering,
and therefore it would bring down the bullwhip effect.
654
Now, how do we plan better and how do we keep inventories low? Now some of the
points are here - getting help from reducing lead times. Now, the moment we bring down
the lead time which is either the external lead time or internal lead time, now what is the
external lead time? External lead time is the time taken for items to come from the
supplier to the manufacturing, which is external to the organization. Internal lead time is
the time taken for manufacturing, which happens within an organization.
So, when we try to reduce the lead time, automatically the inventory that we hold will
come down. Now, enabling or getting enabled through IT solutions, now IT solutions
would help us reduce, this lead time, because more information is available and accurate
information is available. And this can be done both internally and externally. Internal IT
solutions talk about management information systems, they talk about ERP solutions
within an organization.
And external IT solutions is how we get the ERP’s of the partnering organizations to
share information directly, or through the net and so on. So IT solutions internally and
externally help them. Build market and business intelligence ideas and have better
manufacturing planning systems like Just in Time manufacturing, which talks about
reduced inventory and different kinds of planning, to have the reduced inventory. So,
these are ways by which organizations can plan better and reduce the inventory, so that
the bullwhip effect is also reduced. Now, what are the planning issues in supply chain, so
that organizations come together and plan?
655
So, supply chain structure impacts the lead time and the planning. How the supply chain
is structured, how many entities or organizations are there at each layer, and the extent to
which these organizations cooperate with each other is about the structure, which would
help in effective planning. Forecasting and demand management are key requirements
for supply chain. So we should have good forecasting systems, which are reliable and
show less variation. And we should depend more on demands than on forecasts,
constantly look for new means of cutting lead times, exploit data visibility provided by
ERP systems or use information technology to help in the planning.
656
It is also necessary to understand that there is not only one type of supply chain, there
could be different types of supply chain, so that is listed by the heading which says one
size does not fit all. The same supply chain cannot work for all organizations or different
organizations, so there are two classifications broad classifications that have been made.
One is called an efficient supply chain, and the other is called a responsive supply chain,
and if you see carefully the products are now classified into two very broad types.
So, one needs to understand the profile of the product, before we look at what kind of
supply chain strategies we are going to have for the product. So products are broadly
classified as functional products and innovative products. And their supply chains are
called efficient supply chain and responsive supply chain, functional products have
efficient supply chain, and responsive supply chains are for innovative products. Now,
how does one classify a product into a functional product or an innovative product is the
next question, because only after that classification we talk about efficient supply chains
and responsive supply chains.
Now, there is a very important and article from how about business review written by
Marshall Fisher in 1997, which is titled "what is the right supply chain for your product".
And there the author provides a framework to be used, which can be used to classify
products as functional and innovative or talks about what are the broad characteristics of
functional products, and what are broad characteristics of innovative products?
657
Against several criteria and some of these criteria are product life cycle, contribution
margin, variety average forecast error, stock out forced end of season markdown, lead
time, for made to order. Now, if you look at product life cycles, functional products have
longer life cycles, suggested figure would be greater than 2 years, while innovative
products have shorter life cycles. Contribution margins are lower for functional products,
larger for innovative products, variety is low for functional products, high for innovative
products forecast errors are lower for functional products, which means we get better
forecast for these products.
Whereas, forecast errors are higher for innovative products, which means the forecast
accuracy is relatively lower, for innovative products, average stock out is very less for
functional products, stock outs are very large for innovative products. End of the season
markdowns are higher, which means more products are available for forced sale or
through promotions discounts and so on.
And lead time for made to order is large for functional products, and very small for
innovative products. So, through this broad classification it is possible to identify
weather a product fits into the functional category, or it fits into the innovative category,
and depending on the functional product and innovative product. We talk about efficient
supply chain and responsive supply chain, with efficient supply chain focusing more on
cost minimization, and responsive supply chain focusing more on customer delivery.
Both the aspects of cost minimization and customer delivery are important to both the
supply chains, but then we understand that, as we move towards a different type of a
supply chain which is either efficient or responsive there will be a lot more emphasis on
cost minimization, in an efficient supply chain and delivery or customer delivery in the
responsive supply chain.
658
Several other strategies can be used to design efficient supply chains, some of these
strategies are continuous replenishment program using EDI links, which is electronic
data interchange links. To be able to capture point of sale data for accurately updating
the forecast, and we would have also understand the higher the forecast accuracy better
the forecast less will be the inventory. The important aspect of getting the organizations
together is to reduce the total cost of the supply chain, so that individual organizations
benefit by that cost of the supply chain will come down if the inventory at the various
levels of the supply chain are regulated.
Now, to understand that, we saw some aspects of the bullwhip effect. So, if the inventory
in the supply chain reduces, then the cost would reduce, and inventory will reduce if
forecasts are made better and more accurate. And therefore, there is a lot of emphasis on
capturing point of sale data to accurately update the forecast, partner with organizations,
which is essential aspect of supply chain, both on the left side or inbound side or the
supplier side and the right side or the outbound side or the customer side.
Now, integrate systems between organizations to benefit from data visibility, and data
sharing, and develop robust inventory control mechanisms to fix reorder points and
reorder levels, so that the overall inventory comes down. So, what we have seen in this
lecture is how organizations should come together, and work together and form a supply
659
chain, so that the total cost of the supply chain reduces and the delivery from the supply
chain is made better.
How we can do that is through information technology and by understanding the various
aspects in terms of location, production inventory and transportation decisions. We also
saw the need for doing this through shrinking product life cycles, and the constant need
to reduce the cost. And we also saw the two major classifications as functional
depending on the type of the product as an efficient supply chain, and a responsive
supply chain. Other aspects of supply chain management, we will see in the next lecture.
660
Prof. G. Srinivasan
Lecture - 36
Location Problems
In this lecture, we continue the discussion on supply chain management. In the previous
lecture, we introduced two types of supply chain, which are called the efficient supply
chain and the responsive supply chain. The efficient supply chain is to be designed for
what are called functional products and the responsive supply chain is for the innovative
products. We have also seen some ways, by which we can classify products into
functional and innovative. And we also mentioned that the efficient supply chain for
functional products concentrates more on cost minimization, while the responsive supply
chain for innovative products would concentrate a lot more on service and delivery.
Now, we will see some aspects of, how we design efficient supply chains or what are the
things that go into creating an efficient supply chain. Some of these are continuous
replenishment programs using EDI links for information sharing, capturing point of sale
data for updating the forecast, invest in supply chain partnerships on the inbound and out
bound side. Integrate the MRP or planning systems to benefit from improved data
661
visibility and develop robust inventory control mechanisms to accurately fix reorder
points and order levels.
We saw all these five points in the previous lecture and the common thread that goes
along these five points is the reduction in inventory. Now, we will quickly see how that
happens, now when we talk of continuous replenishment program through information
sharing, we immediately talk about reducing the inventory in the partnering entities.
Periodic replenishment systems involve higher inventory than continuous replenishment
system.
And when we have a continuous replenishment system using information sharing, the
buffer also will come down. Accurately updating the forecast would immediately help in
reducing the variation in the forecast and therefore, reduce the extent of extra inventory
that organizations would hold. Invest in partnerships on the inbound and outbound would
also reduce both transportation cost as well as inventory cost. And integrating the
various systems would automatically reduce the buffers in terms of inventory that
various entities in the supply chain would have.
So, we can now see the thread in all these points, which is the reduction in the inventory
and reduction in the transportation cost, which would reduce the total cost. Now, this fits
into our description of efficient supply chains concentrating more on cost minimization.
Now, we will see how responsive supply chains react and what are the things we need to
do there - accept uncertainty in demand, and large forecasting errors as reality. We have
already seen that, responsive chains are for innovative products and innovative products
show higher variation or uncertainty in demand and also the forecast for these cannot be
accurate leading to higher forecast errors.
662
So, we need to accept this and devise systems, which will take care of these, which also
implies that, there would be slightly higher inventory in these kind of systems to the
extent possible. And because the profits margins are higher, the overall profits can be
increased, even if the amount of inventory is a little high. But, what is also needed here
is, improve responsiveness by cutting down lead times, use postponement strategies and
delayed product differentiation strategies and deploy modular design and product
platform strategies.
Now, in all these we will realize that, the focus is more on reducing the time taken to
produce as well as to be able to produce a variety, both postponement and delayed
product differentiation are ways, by which... If there are variety of products, the variety
actually happens towards the end of production process and for most of the production
process, the product would be the same upto the point, where the product differentiates.
Now, the advantage of doing that is to reduce the inventory that is there, so delayed
differentiation that would mean, the inventory upto the point, where the products looks
similar.
The inventories are smaller and the inventories change when the product becomes different.
But, more than the change of inventory, it is also easy to manage or bring down the time to
produce through standardization, because the product would be the same upto a very
significant point of the production cycle. So, the emphasis on the responsive
663
supply chain design is more on cutting down the lead time or reducing the time taken to
bring the product to the customer.
Postponement strategies, packaging postponement is saving in transportation and to be

able to handle different requirements. Some examples are also given, postponing the
assembly, postponing is trying to do it in the end and not to try and do it in the middle.
And manufacturing postponement talks about, unless the firm orders are received, delay
the manufacturing so that, the amount of inventory in the system is also reduced.
664
We come to the last slide in this presentation, which is called supply chain planning
framework, where we show the strategic tactical and operational decisions in the supply
chain through a planning framework. And we have already seen several aspects of this,
supply chain location and network design come under strategic decisions, we have seen
this in an earlier slide. Supply planning, demand planning, forecasting, operations
planning are all tactical decisions, while materials management, distribution, execution,
logistics, they all come under operational decisions.
We also know that, the strategic decisions are long term decisions, whose effect of
strategic decisions are there for several years. Tactical decisions are medium term, they
are there for several months or several weeks, while operational decisions are short term
and they are there for several days of activity. So, when we start designing a supply chain
and we start talking of organizations coming together, now these are the various aspects
that have to be looked at.
The location decisions which are strategic in nature, the production decisions which are
tactical in nature, the inventory and distribution decisions, some of the production
decisions are also operational in nature, inventory decisions are both tactical and
operational in nature and distribution decisions are more operational in nature. And the
information decision or information sharing or information technology brings all these
aspects together and binds them. So that, both inventories that exist in the system and
time taken to do an activity come down.
If we realize that, the organization would make more profit by reducing these two
dimensions, which is the extent of inventory in the system and the time it takes to make
the final product and get it to the customer. So, these two aspects are crucial and these
two have to be brought down so that, the overall cost in the supply chain comes down.
There are differences that we have seen between the two types of the supply chain, one
primarily concentrating on cost reduction and the other concentrating primarily on
delivery and reduced time to market.
In spite of these differences, if you see that if we bring these two costs together, the cost
of inventory and the loss of profit by delaying the time taken to produce or the gain by
reducing the time taken to produce, organizations can benefit and make a lot of profit
through proper relationship, which is the essence of supply chain management. We just
665
go back to one of the slides that we have seen earlier, which talks about the major supply
chain decisions, which I have indicated five of them.
Location decisions, production decisions, inventory decisions, transportation decisions

and information decisions. Now, we go back and see, how much of these we have
covered in the course and what are the things that are remaining to be done in this course.
When we started this course with a lot of emphasis on operations management, as I
mentioned we have spent a lot of time on production decisions. Starting from
forecasting, understanding forecasting completely to the extent possible, understanding
the various models in forecasting.
And then we looked at aggregate planning, we saw several models for aggregate
planning and then much later we did scheduling and sequencing, we did little bit of
MRP, so all of these come under production decisions. We have seen several inventory
models for what are called cycle inventory and safety inventory which means, models
that relate the order quantities and models that relate the reorder levels and safety stocks.
So, we have spent some time covering the inventory decisions and we have spent some
time addressing the production decisions in this course.
Very briefly, we have looked at some aspects of location decisions and we are yet to look at
some aspects of transportation decisions and the role of information technology on the
supply chain. So, what we will do now is revisit some of the location decisions that we
666
have seen and continue to discuss some location models that talk about locating facilities
and warehouses. And then we will move on to transportation decisions and models for
transportation and then we will look at information systems and the role of information
system in effectively making the supply chain performance better.
The models that we have actually have seen, related to location earlier in this course are
three models that we have seen.
667
So, we have seen the p median model, we have seen the fixed charge problem and we
have seen location of multiple facilities or location of warehouses, earlier in this course
through the location models. Now, we will see few more location models and try to link
them with this.
So, the most basic model when it comes to locating warehouses and retailers for a single
product is an assignment problem, where we minimize d i j X i j, where X i j equal to 1 if
the retailer j gets supply from warehouse i and gets all the supply from warehouse i. The
model that we show here is to minimize d i j X i j.
668
Here we assume that, we would have several warehouses and a large number of
customers or retailers. So, we call this as warehouse and we call this as retailers, now we
are going to assume, when we look at this formulation, that there is a demand D j from
each of these. Now, there is a supply from each one of these S i, we are going to assign
each of these retailers to this. For example, we may have a solution, where if we have
something like this then we would say that, retailers 1, 2 and 4 get their supplies from
supply S i.
Now, at present we need not consider the D j’s and S i’s explicitly, we would simply say,
that there are certain numbers of retailers and certain number of warehouses. And when
we actually do up to this, the problem will be to minimize the distance, so there is a
distance d i j that we would like to minimize. We also would like to use all these supplies
and then we have to say that, each of these retailers will get their supplies only from one
particular warehouse.
There is a demand for D j for that item from each of these, there is also a restriction that
a particular warehouse will not supply to more than a certain number of retailers. So, this
would still be related to a typical assignment problem, even though a typical assignment
problem in the OR literature would talk about sigma X i j is equal to 1, sigma X i j is
equal to 1. Now, this can be expanded to a simple assignment problem, even when we
put a restriction that this it will supply to a maximum of certain numbers.
669
Now, upto this is either directly an assignment problem or can be modelled as an
assignment problem. Now, this constraint also becomes important, when we bring a
certain supply to each one of these which means, there is a capacity here and then we
need to do this. Now, this is not exactly a transportation problem, the difference comes,
because in the transportation problem, there is a supply quantity and there is a demand
quantity.
Transportation problem does not put an explicit restriction that all the demands have to
come through only one supply. Transportation problem would allow that the demand of
this can be met from here as well as here. So, the moment we put an additional
restriction to the transportation problem that, the demand of this should be met only from
one of them, it becomes what is called a generalized assignment problem. So,
generalized assignment problem, which can be solved through binary programming, X i j
takes values 0 or 1 and this can be solved as a 0 1 or binary optimization problem.
So, the one of the models is, which given a set of warehouses or given a set of supply
points, given a set of retailers or given a set of demand points with known demand D j.
Now, which warehouse will supply to which retailer such that, the capacity is considered
and all the demand of retailers has to be met from the warehouse. And there is an
additional restriction that, for reasons of control, that this does not supply to more than a
certain number of retailers.
670
So, this is a standard problem where we model, how the demand of each thing can be
met. The second model would be known number warehouses and retailers, single
product and a retailer can get from more than one source. So, the second problem is a
simpler version of this problem, so we would still have this, we would still say that, this
constraint is modified.
Now, this constraint will be modified to sigma X i j is greater than or equal to D j for the
retailer j and sigma X i j is less than or equal to S i and this does not exist, this does not
exist. X i j is not binary any more, X i j is the amount of demand or quantity that is met
from transporting from i to j and it becomes a straight forward transportation problem,
this does not involve dedication. For example, here in model 2, we do not assume that,
all the demand is going to be met from only one, you may have a situation, where the
demand of this can be met from both these.
671
So, the model 2 is actually a simpler version than model 1 in terms of assumptions,
where a retailer can also can get from more than one supply and in terms of solution,
because it becomes a straight forward transportation problem, which is simpler to solve
than the generalized assignment problem. Now, having seen those two models, we will
now come back to first the p median
and then the fixed charge and then we will go to the model where we talk about, which is
called the fixed charge transportation problem and I will explain the difference between
672
them. So, though I have shown this as model 3 in this PPT, after the model 2, we need to
look at the p median as well as the fixed charge problem. We have seen p median and
fixed charge in an earlier lecture, so I will just provide a very quick recap of both of
these before we go to the next one, which is called the fixed charge transportation
problem.
P median, one of the ways of introducing the p median problem is to say that, there are
points here, let us say there are 6 points or there are 6 customers. And now, let us say, we
want to locate say 2 factories or known number of factories or p factories or p medians.
Now, which of these six will be medians such that if for example these two are the
medians. Now, first let me describe this problem mathematically then I will describe this
problem with respect to the context, in which we are looking at it.
Now, there are 6 points and there are n points, there is a distance matrix d i j that is
known which means, the distance between each of them is known. We call these as 1 2 3
4 5 6, distance between a point and itself is 0, so the d i j matrix is known, a p median is
to choose p out of these n points. There are n points, to choose p out of these n points,
which act as medians and the remaining n minus p, each of these point is attached to a
median.
For example, if we choose these two as median then these four are the non median points
and each of the non median gets attached to a median point. For example, this might get
attached to this and this might get attached to this, this would be attached to this, this
would be attached to this such that, sum of these distances is minimized. There will be 6
distances, now these two points attach to themselves, so their distance is 0, these four
points are going to get attached to a median, so there will be 4 distances.
Now, which are the two medians, the best two medians and to which medians, the non
medians get attached such that, sum of the distance, the four distances is minimized is
called the p median problem. Now, what is the relevance of the p median problem in the
context of, what we are looking at? Now, we could say that, here are 6 customers, now
we want to establish say 2 factories, p equal to 2; we want to establish two factories such
that, we are able to meet the customer demands based on distance.
Now, we want to find out the best location out of these six, the best two locations out of
these six such that, the remaining four are allocated into the two locations and the
673
distance is minimized. So, that is the relevance of the p median problem in the context
of, what we are looking at. So, there are 6 customers and we want to or who are located
in six different places, the d i j matrix is known and if p equal to 2, which out of these
two six locations, we would locate a factory.
And these two located factories will cater to the requirement of the four remaining
points. Now, which of these four remaining points are attached to the factory is given by
the p median problem. So, we have seen ways to solve the p median problem, which is
our next location model. Now, the fixed charge problem, automatically it can be seen as
logical extent of the p median problem, it can also be seen in a slightly different context.
Now, when we extend this p median problem and then we start saying, in the p median
problem we are going to assume, that whether I made this one as a factory or this one as
a factory, the cost of making this as a factory is the same across all the points. Now, in
the fixed charge, I am going give an f i, which is the fixed cost of making a factory from
this location, so the problem changes to f i Y i plus d i j X i j. So, there is going to be a
fixed cost, now the other way is, there can also be a capacity of this, which I may call as
C i or capacity of i.
Now, in the p median, we have not included cost and capacities, now suppose I solve a p
median and for example, let us say these are the two best locations and these two are the
things then at the end of p median I would say, locate factories here based on distance.
The individual cost is going to be same, now if the demands are D 1 plus D 6 plus D 5
then it is necessary to have a capacity of D 1 plus D 6 plus D 5 here and a capacity of D
2 plus D 4 here.
Now, in the fixed charge we do it slightly differently, we start defining what the
capacities are going to be and then decide, which are the two best locations and how the
demands of rest of them are going to be handled by this. Fixed charge does not, again
there are two ways of fixed charge, where one of the ways is to say that, if there is a
factory located here and this one does not get a factory then all the demand of this should
come only from one of these, either this or this, which is p median kind of a assumption.
But, in practice we could say, part of the demand can come here, part of the demand can
come here, depending on the capacities of these two places. So, fixed charge can be seen
as a logical extension of p median, where there is a fixed cost of creating, which is
674
different for different points. As well as, the capacities are predetermined in the fixed
charge, whereas in some ways, they are post determined in the p median, they are already
given to us.
And the difference is, fixed charge does not necessarily restrict, that all demand should
only come from one of the chosen factories. Other way to look at the fixed charge
problem is to say, here are some potential sites for the factories and here are the
customers. So, in this way of introducing the fixed charge problem, we are not going to
locate factories in any customer locations. Now, these specific sites where factories are
located are different from customer sites, therefore this as a graph is a bipartite graph.
So, there is no connection between or amongst these or amongst these, now again the d i
j’s are now like this and so on. So, here the d i j will be a square matrix, there are 6
points, we will get a 6 into 6 d i j. Now here, if there are 4 sites and 6 customers, the d i j
matrix will be a 4 into 6 matrix then we want to find out, which two of these we select,
there is an f i, there is an C i and there is a D j. So, we have seen some aspects of p
median and also we have seen some aspect of fixed charge problem in an earlier lecture.
Now, with this background we move to model 3, which talks about warehouses and
retailers, we will visit this location of warehouses later.
So, this model talks about warehouses and retailers, where we have a known number of
warehouses and retailers. We are talking of a single product, we also say that a retailer
675
can get items from more than one warehouse and there is a fixed and variable cost of
transportation. So, the important change is that, there is a fixed and variable cost of
transportation. So, let me go back and explain this through this network, now here we
have a set of warehouses and then we have a set of retailers.
Now, the question is, there is a supply S i in each one of these, this is not a location
problem per say, the decision here is not where to locate. In a p median the decision is
where to locate, in a fixed charge the decision is where to locate, now here we assume
that, there are a set of warehouses and there are set of retailers. For the sake of
illustration, there are 4 warehouses and there are 7 retailers, now each one is connected
to every other, so it means, every warehouse can supply to every retailer.
The normal transportation problem, which was our model 2, talks about there is an S i,
there is a D j, how much do I send. Assumption is, this can be met even partly from here
and partly from here, the additional dimension in this problem if you look at the
objective function, the old transportation problem will have C i j X i j. And now, we also
have another d i j, here we have C i j Y i j and d i j X i j, so there are two terms in the
fixed charge transportation problem, whereas in the old transportation problem, there
was only one term, which is d i j X i j, X i j is the quantity transported from i to j.
Now, we introduce a Y i j and say that, if for example, I transport a quantity X 1 1 from
the first supplier to the first demand. Now, the cost of transportation will be d 1 1 and X
676
1 1, but there is an additional cost, which is called a fixed charge of doing this. Now, this
fixed charge would be like an additional toll that I may have to pay if I choose to
transport from this to this. So, there are two decisions, first of all whether I choose to
transport from i to j and then if I choose to transport from i to j, what is the quantity that I
transport from i to j.
Whether I choose to transport is given by the Y i j and the quantity that I transport is
given by the X i j. So, the constraints will be that, each sigma X i j is greater than or
equal to D j, whatever comes into this will meet this demand. And as far as the other one,
sigma X i j is less than or equal to S i and Y i j, where we have shown it as two different
constraints.
We have shown two constraints, which is X i j is less than equal to S i and then said X i j
is less than equal to M Y i j. They can be brought together into a single constraint, which
will say that, this supply or capacity of S i is available, only when to meet the
requirement of a j partly, only when we choose the root i j, so only if Y i j is 1, this in
whole or part is available for this. So, this is the constraint, Y i j will be binary and X i j
will be a quantity, so this is called fixed charge transportation problem.
See, it is different from the fixed charge problem, the difference comes from the fact that,
in a fixed charge transportation problem, there is a charge fixed charge f i j between i and
j, whereas in a fixed charge problem, there is an f i associated with this i. Now, in
677
the fixed charge problem, the fixed charge is for a node and in a fixed charge
transportation problem, the fixed charge is for an arc or an edge that connects i and j,
here it is for a node or vertex, which is i.
So, there is a cost associated here and there is a cost associated here, this cost should be
seen as the cost of setting up a facility. And this cost should be seen as a fixed cost of
transporting, which is like a toll that is paid when you move items from i to j. So, that is
our model 3, which talks about warehouses and retailers.
Then, we move to the 4 th model, which is the capacitated plant location model, which is
actually the fixed charge problem. So, f i y i, so this is what is shown there as model 4,
which is the fixed charge problem.
678
Now, you see the objective function is sigma f i y i plus sigma C i j X i j or d i j X i j, if

this is a generic expression, which talks about the unit cost of transporting from i to j, we
could use d i j X i j if the unit cost is proportional to the distance and the proportionality
is 1. For example, it costs 1 rupee to transport the unit distance then you can use d i j X i
j or C i j X i j. So, we have already seen some aspects of the plant location or a fixed
charge problem earlier in this course.
679
Now, we look at locating plant, capacitated plant location model with a single sourcing,
so as model 5, what we see is something like this, the same fixed charge problem here.
So, when we go back to this network which I draw again, these are the possible sites,
these are the customers. So, there is an f i associated with this side, there is a D j
associated with this customer, now all that now we have to f i, so y i equal to 1. If this
site is chosen, so the objective function will minimize, this is the cost of location, this is
the cost of allocation, so it is also called location allocation problem. Because, we talk of
locating facilities as well as we talk of allocating from the capacity of these facilities
allocating to meet the demand, it is called allocation problem.
So, all the demand should be from only a single place, that is why we call capacitated
plant location with single sourcing which means, the entire demand of this will be met
from only one of these and that one has to be chosen. So, you will have a constraint
which will say, sigma X i j equal to 1 which means, in this model, X i j is binary, y i is
also binary, y i is 0 or 1 if we choose this, if it is chosen is 1, if it is not chosen it is 0.
So, whatever is chosen will incur a fixed cost, which is f i y i, now the demand of all of
these have to be met. So, X i j equal to 1 if the demand of this j is going to be met from a
chosen i. Since X i j equal to 1 will ensure that, the demand of each of these is met from
only one of these. And the second thing that we need to look at is, sigma D j X i j is less
than or equal to K i y i or C i y i.
680
So, not only this gets the entire demand from one of them, this should get it is demand
from a chosen site. So, if this site is chosen then y i will be 1, only then it will get from
this, if this is not chosen this cannot supply to this and if this is chosen, this has a
capacity of C i y i. So, whatever you get from this capacity, should also be within what
this supplies, so this C i y i is the capacity of the chosen one and this is given to several
D j X i j.
So, the capacity is greater than the demand, so the single constraint takes care of all of
them. The objective function changes, because X i j is binary, so D j X i j is the demand,
so C i j into this, so cost into demand. So, this is called locating plants capacitated model,
because K i or C i talks about the capacity of this when it is chosen, so it is capacitated.
Single sourcing, because all the demand from here is met from only one supply, so that
is our model 5, which is capacitated plant location with single sourcing.
Sixth model is locating plants and warehouses simultaneously, now this model is exactly
the same model that we have seen here earlier, where we have spoken about locating
plants and warehouses simultaneously which means, we are looking at multiple stages.
681
So, there can be a stage where we have plants or factories, there will be a stage where we
will have warehouses and there will be a set of customers. So, the question here is, we
can put a restriction on the number of factories that we wish to locate, we can put a
restriction on the number of warehouses that we wish to locate. So, we will have
multiple decisions f i y i, so f i would be a fixed cost here, y i will be the decision here
and then we also have another one f e y e.
So, this will be a fixed cost f e and then we would possibly have another y e or z,
whatever variable we can use, so if we choose some of these. So, there is a fixed cost of
location here, there is a fixed cost of location here. Now, the moment there is a fixed cost
of location here, there is a capacity, so things have to be produced and transported only
to chosen facilities, because there is cost of creating this. Then these are transported to
the chosen facilities and from there, they are transported to meet the requirements of
these customers.
So, here again, we can have a dedicated or a single sourcing model or a multiple
sourcing, where a chosen warehouse can supply to either or the other way, a demand is
met only completely from one warehouse or it can be met from multiple warehouses.
Now, there is a distribution cost here, there is a distribution cost here, there is a fixed
location cost here, there is a fixed location cost here.
682
Therefore, you we see 4 terms or 5 terms and this depending on the number of layers and
if there is one more layer here, there is a warehouse distribution as well as the customer
then you have one more transportation cost. The location decisions usually concern only
this and this, how many plants to locate, how many warehouses to locate.
Rest of them are transportation cost and linking constraints, which will link the capacity
as well as the demand, which is shown here 4 layers, so there are 3 types of distribution
costs.
683
We could also look at more advanced location problems, particularly for multiple
products, so far the models that we have seen, are largely for single products. And now,
these can also be expanded to looking at multiple products and fixed toll.
So, we get much more numeric formulation.
Which is shown here with C i j k X i j k, where k represents a product or a market i to j

which simply means, from the plant to the customer. So here, we are not looking at
684
multiple layers in the supply chain, we are talking about a set of factories and a set of
customers and multiple products or multiple markets.
There is a location and there is an allocation and there are some linking constraints for
each one of them.
Now, we will look at a couple of more aspects, which are called warehouse aggregation
and safety stock and service levels in this. Now, what exactly do we mean by warehouse
685
aggregation. Now, when we look at a model like this, when we look at all over location
and allocation models.
Now, let us see, there are sometimes when we have, let us assume there is a warehouse
here and there are several customers here. Now, there are times we wish to make another
decision which says, now can we actually have two warehouses. For example, can we
have one warehouse here and we have one warehouse here. Now, if we replace this
central warehouse with these two warehouses we realize that, the distance travelled will
be less.
Because, here we are looking at these six distances, whereas here you are looking
essentially three distances, if this merges with this, we are looking at 3 plus 3, we will
look at 4 distances and it will be much smaller than the sum of the distance here. But
now, what are the advantages and the disadvantages, now the advantages and
disadvantages, advantages in reduced transportation cost, the disadvantage could be more
inventory or buffer needs to be kept here as well as here.
Whereas, when we centralize them and bring them together, the inventory that we will
have there, will be much less. So, there is this traditional trade off between a centralized
warehouse, where the inventory is kept, the safety stock is less, transportation costs are
more. Decentralized warehouses, where the transportation costs are less and the
686
inventories would be more. Now, we look at some of these, we look at a single
warehouse and multiple retailers.
Now, one could simply easily calculate the economic order cost for each retailer for n
retailers.
So, if we have n retailers, using our traditional formulae that we have used, the economic
order quantity Q will be equal to root over 2 D C naught by C c, where this is the
demand for the item, this is the order cost, this is the inventory carrying cost. So, for each
687
of them for a single item, this will be the economic order quantity and the total cost at
the economic order quantity will be root over 2 D C naught C c.
Then, if we have it for N different retailers, the total cost there will be, for n different
retailers N into root of 2 D C naught C c. Now, if bring all of these together into a central
place then the total demand will become N into D and therefore, the economic order
quantity will be root over 2 N D C naught by C c. And the total cost T C 2 if we may call
this, is equal to root over 2 N D C naught C c, now this is smaller less than T C 1.
Because, N into root of 2 D C naught C c will be bigger than root over 2 times N D C
naught c c and now there is a saving in the total cost, if we bring it down together. And
the same thing can be seen as either warehouse aggregation or can be seen as inventory
aggregation for a set of N retailers and so on. So, the saving will come down and this
saving can easily be shown as 1 over root N by N or 1 over 1 by root N or 1 over root N
by N will, now be the saving when we try and bring the warehouses together.
688
The other aspect that we can see, is also the saving in safety stock and in service levels,
which we have actually seen in an earlier lecture on inventory. Now, we also know that,
if each one of these has each one of these retailers, there is a lead time demand
distribution.
And if sigma is the standard deviation of the lead time demand and then we have a
service level SL and based on this service level, we go back to the normal distribution
and then find out a z from the normal distribution. Now, z sigma is the safety stock that
689
will be maintained, now for N people, we will have N into z sigma in each of these
places. Whereas, when we pool them together, we have already shown that, the standard
deviation when combined together is actually root N into sigma.
And the same safety stock, the total amount of safety stock maintained if we bring them
together will be z into root N into sigma, whereas if we maintain them in different
places, it will be z into N into sigma. Very similar to N into root over D C naught C c
and root N over root of 2 D C naught C c, similarly N into z sigma versus root N into z
sigma. Once again it is possible to show that, the saving is of the order of 1 minus root N
by N or 1 minus root 1 by root N.
So, the advantage of doing this from an inventory point of view is to try and aggregate it
or try and make decisions, which are for everybody together, which is an essential
dimension of supply chain. Because, if the individual players make their own decisions
or if the individuals have their own warehouses, the amount of safety stock will be higher
and the total cost of inventory will be higher. So, by centralizing and by bringing them
together, the total cost of inventory can be brought down.
But, we have already seen that, in some ways the cost of transportation will increase
when we try, the cost of transportation will be lower if we have different warehouses,
rather than a centralized warehousing. The important aspect is, for these organizations to
come together and to be able to do centralized ordering or to be able to take from a
690
centralized warehouse. So, the inventory cost may go up a little bit when you centralize
it and take it from a centralized warehouse, but the total cost can come down.
Much later we will see that, by proper information sharing, we can actually bring down
even the cost of transportation by aggregating or when we aggregate the demand. So,
with this, we also look at some models for warehouse and models for warehouse
aggregation. And what we have seen now along with, what we have seen earlier in terms
of the p median problem, the fixed charge problem and location of factories and
warehouses, would give us a good picture of the location decisions with respect to a
supply chain and the ability to bring down the total cost through aggregation. Rest of the
aspects such as transportation decisions and modes for transportation and distribution, we
will see in the next lecture.
691
Operation and supply chain management
Prof. G. Srinivasan
Lecture - 37
Transportation and Distribution Models
In this lecture, we look at Transportation and Distribution Models in a supply chain. We

spend some time on understanding the broad issues related to transportation and
distribution. And then we address a few mathematical models and provide solutions to
them through numerical examples and illustrations. The supply chain decisions we have
seen can be categorized into location decisions, production decisions, inventory
decisions, transportation or distribution decisions. Information related decisions also can
be added to this, so transportation and distribution decisions are among the most
important decisions in planning and executing a supply chain.
Distribution essentially talks about moving the manufactured items to the end customer,
in an earlier lecture we defined supply chain as a network of facilities and distribution
options that procure a set of parts or products, transform them into finished goods and
distributes them to the customers. So, you can find the role and relevance of distribution
in the supply chain by it is presence in the very basic definition of the supply chain.
692
Distribution is important, because for two reasons, one is the manufactured products
should reach the customer in the right quantity at the right time.
While achieving customer service by distributing them in time, it is also important for
the organization or the supply chain to ensure that, the cost of distribution is suitably
managed and the cost of distribution is minimized or is as small as possible. So now, we
will see some aspects of the costs related to distribution directly and indirectly, the most
important cost is the vehicle related cost.
Because, it is manufactured, products are transported from one place to another, we

definitely need a mode of transportation, which involves different kinds of vehicles.
Most frequently used is the road transportation, though sea transportation and
transportation by air and mixed modes of transportation are also prevalently used.
Whatever be the mode of transportation, there is the vehicle that carries these items from
one place to another and there is vehicle related cost.
Most of the times, this cost is proportional to the distance and there are times, there is
also a fixed cost in addition to the vehicle related cost, which could be the cost of owning
the fleet of vehicles or the cost of hiring the some of these vehicles or trip specific cost
such as toll and so on. So, there are some fixed costs, which are also there in addition to
the vehicle related cost. These costs are not proportional to the distance of movement of
the products, there are also costs related to the quantity that is being transported.
693
The vehicles have fixed capacity and depending on the quantity that is being transported,
we would be a requiring certain number of vehicles. So, the number of vehicles used will
first depend on the quantity that is transported, it would also depend on the volume or the
size of the product that is being transported. In addition, there are also loading and
unloading costs, where the items have to be loaded into the vehicles or the trucks and
they have to be unloaded, when these vehicles reach the customer destinations, so there
are costs related to loading and unloading.
The fourth one that we see here is inventory costs and there is always the cost of holding
the inventory at any point between the factory or the place where it is manufactured and
the customer destination. At any point in between these two, the inventory is held and
there is a cost of holding this inventory. We have already seen that, the cost of holding
inventory can be treated as the interest that is to be paid on the money that is invested in
making the product.
We have also seen some aspects of the tradeoff between inventory cost and distribution,
if we centralize the distribution and keep it in the central warehouse, the safety stock or
excess or extra inventory that is to be held will be less, but the transportation cost will be
more. If we locate warehouses and distribution points nearer to the customers then the
transportation cost will be less, but the cost of holding the extra inventory or buffer
inventory will be more.
So, there is always this tradeoff between inventory cost and transportation cost, so these
are some of the costs that are related to distribution. Now, depending on the nature of the
product that is being transported, the distribution network has to be designed. Some of
these products may require what is called point to point transportation which means, a
truck or a vehicle will go from the warehouse to the customer and come back to the
warehouse will not visit any other customer.
So, whatever is the customer demand, met by this way and most of these instances, either
the product requires such a type of transportation or the demands would be closer to full
truck loads, then the mode of distribution would be a point to point distribution. Other
examples of modes of distribution that we will see, as we move along are called milk run
and cross docking. When we see milk run, milk run is about transporting from a place,
from a warehouse, the truck would visit multiple customers or more than one customer
694
and deliver and then come back to the place where it start, so that is a different type of
distribution system.
The third one which we will see is called cross docking, where we actually do not have a
warehouse, items can come from the factory, come to a warehouse or come to a place,
which does not physically store the items. When we say use the word warehouse, most of
the times we mean that, items come, they stay physically in the warehouse and then they
are moved from the warehouse to the customers or retailers. In cross docking, they do
not stay in the warehouse, there is a place where trucks arrive from the factories and
items are transported to other set of trucks in a central place.
And that place is where the cross docking takes place and the items that are transported
will now move towards the retailers or the customers. So, the distribution mechanism
does not store the inventory inside the warehouse, so like this we could have different
types of distribution networks in practice. Some of the aspects are the ownership
structure of the distribution network also has an impact; choice of the distribution
network has an impact.
Also check whether distribution requires an exclusive strategy and last but not the least,
product, price, commoditization and criticality also have an impact on the type of
distribution system, that are preferred by the customers. So, this slide essentially tells us
an adequate care has to be taken and a lot of thought has to be go in, when we decide the
695
type of distribution network and that would depend on the customer, that would depend
on the product and that would depend on the costs as well, here are some examples that I
just now mentioned.
So, direct shipment it means, there is no intermediate warehouse, it goes directly from
the factory to the customer. It is a very simple method, but it becomes effective when
close to truck loads are set, if the demands are such that, we do not have close to truck
load then direct shipment is not advantageous. Because, there is always a fixed cost of
hiring or owning a truck and transporting items through the truck and that cost will
increase.
And therefore, direct shipment will not be a cost saving proposition or it will not be
economic, therefore the cost can be slightly higher. It becomes effective, only when we
have close to truck load that is being sent from the factory directly to the retailer or the
customer. In such cases, it is also advantageous to use milk run, therefore the truck will
start from the factory, use a direct shipment mode and go to multiple customers and it
will come back.
Here, we have better truck utilization, the cost become lower and there is a little more
control. Next model is to ship via an intermediate distribution centre, so there is a
distribution centre that stores the inventory and transfers the items and from there, it goes
to the end customers. We could also have distribution centre across docking, where it is
696
not exactly physically stored in the distribution centre, but cross docking takes place and
this kind of a design can also be with or without milk run. So, if it has milk run then it
goes to multiple retailers from a single supply or multiple supply to a single retailer and
it will be designed that way.
It is also important that, we should look at inventory costs. Distribution modes with
higher transportation cost is justified only when they result in significant inventory
reduction. We have already seen that, there is a strong relationship between the
transportation cost and distribution cost. There will be situations and modes, where the
transportation cost would appear lower, but that would result in inventory being
distributed in several places, increasing the cost of holding the inventory.
And particularly when we consider the buffer or safety inventory that is required to take
care of uncertainties, the cost of holding them will be very high. On the other hand, some
modes will require a higher transportation cost and that would result in lesser inventory
cost. So, a distribution mode or a channel that would result in minimum sum of
transportation and inventory cost and is able to maintain the service level is the most
desirable thing.
We have seen in an earlier lecture, that if we have N warehouses and if we centralized,

we could save of the order of 1 minus 1 by root N if we aggregate. But, when we use
those models, we should also look at the combined effect of the inventory cost as well as
697
the distribution cost. It also leads to this inventory aggregation a very desirable thing.
Aggregate means to add, so inventory aggregation is about adding the inventory or
keeping the inventory together in a certain place.
So, inventory aggregation talks about storing in few warehouses which means, the
transportation cost will be high and as I had mentioned, the cost of buffer inventory will
be low when we use this model.
Here again, we also have to look at tradeoff between inventory cost and facility cost,
because when we setup few warehouses, the cost of setting up the warehouses has to be
looked at and the capacity of these warehouses also has to be looked.
698
Now, from a customer’s point of view, tailoring the transportation based on customer
distance, size or product demand, allows a supply chain to achieve responsiveness and
cost reduction. So, at the end of the day, we also have to make sure that, customer
expectations are met, service levels are met and responsiveness which is a measure of
customer expectation or ability to meet the customer expectation is achieved. So, all
these factors are to be considered when distribution channels and transportation modes
are being considered for distribution.
Finally, all these will have to be linked to the very basic definition of the supply chain,
which talks about procurement, manufacturing or transforming them into finished
products and distributing them. To be able to meet the customer demand and maintain
service levels and to be able to do it at minimum cost is the very purpose of the supply
chain. And the minimum cost is a cost that is aggregated or added considering all the
players in the supply chain. So, with that in mind, we have to look at all aspects of
decision making in the supply chain including distribution decisions.
699
Now, we move to some mathematical models, which are available, which can be studied,
which we are going look at, which help us in making distribution decisions in the context
of the supply chain. Some of these are familiar and some of these we would have seen
earlier in this course as well as in some other courses. But, I will address them for the
sake of completion and for the importance that, they have in the context of distribution in
the supply chain. The first problem that we will be looking at, is the transportation
problem.
700
Now, we have seen this transportation problem earlier, we saw some aspects of
transportation modeling in the previous lecture, where I had described the transportation
problem as the first and the most basic problem, that we have to study in distribution. I
have also explained transportation in the little more detail in the operations research
course, where we have several lectures that address the transportation problem. I would
still take an example now and explain some of these principles that are used in solving
the transportation problem.
So, a numerical illustration to explain the transportation problem is like this, we explain
the transportation problem using this example. Now, we can assume that, there are 3
supply points or we will assume that, these three are warehouses. There is a single item
which is being transported, these are the requirement points or the destination points. In
the context of supply chain, these can be seen as retailers or customers, now the single
item has to be transported from these three warehouses to these three retailers or
customers.
We have given an example, where the total supply is equal to the total demand and both
are equal to 130. Now, this number represents what is called the unit cost of
transportation which means, it represents the cost of transporting one item from this to
this, is represented by this four. So, we assume that, the costs are proportional or linear,
so if we transport 1 item, it is going cost 4, if we transport 20 items from this to this, it
would cost as 80.
Therefore, we are trying to minimize the total cost of transportation such that, all these
items from which are available here, are transported to the customers or the destinations.
701
Now, the formulation of the transportation problem is as follows, X i j is the quantity

transported from i to j. So, if X 1 1 is transported from 1 to 1, next 1 2 is transported
from 1 to 2 and so on, the total cost of transportation will be 4 into X 1 1 plus 6 into X 1
2 plus 8 into X 1 3 and so on. So, we need to minimize this, so we minimize 4 into X 1 1
plus 6 into X 1 2 and so on, so we minimize 9 terms, which are there from X 1 1 to X 3
3.
And this is generally written as minimize sigma C i j X i j, i equal to 1 to 3, j equal to 1

to 3, there are 3 suppliers and 3 customers and it is not necessary that, we should have an
equal number of suppliers and customers always. Now, we also have to make sure that,
whatever goes out of this, does not exceed 30 40 and 60 respectively, and whatever that
enter here, do not exceed 20 60 and 50 respectively. So, that gives us sigma X i j less
than or equal to a i which means, X 1 1 plus X 2 1 or X 1 1 plus X 1 2 plus X 1 3 is less
than or equal to 30.
So, X 1 1 plus X 1 2 plus X 1 3 less than or equal to 30, we will have two more
constraints, one for this and one for this. The general expression is this, where a i is the
availability in place i, similarly what is required, this is a minimum amount of items that
are required. So, the amount that come into this is X 1 1 plus X 2 1 plus X 3 1, so we will
have the other constraint of the type X 1 1 plus X 2 1 plus X 3 1 is greater than or equal
to 20.
702
So, there are three constraints like this, one is for the first one, second one and the third
one, this will be X 1 2 plus X 2 2 plus X 3 2 greater than or equal to 60 and so on. So,
this in general is sigma X i j greater than or equal to b j summed over i here and summed
over j here and X i j is greater than or equal to 0. So, this completes the mathematical
formulation of the transportation problem, now this transportation problem with m
supplied points and n demand or destination points, will have m into n variables.
And in this model, we will have m plus n constraints and this X i j also need not be
integer, X i j can take continuous values. We have seen elsewhere in the other course in
the operations research course that, even if we solve with X i j as continuous variables,
we will get integer valued solutions as long as the a i and the b j’s are integers, which is
what we have here. In the earlier course we have also seen, how to solve a balanced
transportation problem.
Now, this is a balanced transportation problem, because the total supply is 130 and that is
equal to the total demand. So, if you have a solver, if you have an access to a solver then
you can easily solve this problem as a linear programming problem, which has 9
variables as I mentioned, i equal to 1 to 3, j equal to 1 to 3 and 6 constraints. And an LP
solution would straight away give us the quantities that can be or have to be transported
from each of these i’s to each of these j’s.
Now, the transportation problem is a very important problem in operations research and
it is a well known problem and there are better ways of actually solving it, which are
computationally a little faster when we compare the LP way of solving it. So, if you have
an access to a solver, you could easily solve it using the linear programming approach
and the solver will give the solution very quickly for a problem of this size, which has 9
variables and 6 constraints.
But, if we want to study other ways through which we solve this, assuming that we do
not have access to a solver. Now, there are very efficient ways of solving it, which would
be in terms of computational time and effort, a little superior to solving the linear
programming. A more detailed version of this is given elsewhere in the other course, but
I will very quickly out line one method to solve this problem.
703
So, what we can do is, we would assume that, when we have to transport, after all we are
interested in minimizing the total cost of transportation, so what we will do, is to try and
transport according to the minimum cost. So, we first find out, what is the minimum cost
of transportation, so that can happen here as well as here, so we could take any one of
them. So, at this if we consider this, the demand 20 and the supply is 60, so we will
transport only 20 here, which to meet the entire demand.
So, when we do that, this demand is completely met, this is met and this 60, the
availability becomes 40, now we look at the next smallest cost, which happens to be here
or here. So, let us take this one, we can take any one of them, the supply is 40 and the
requirement is 50, so we can take all the 40 here to meet this requirement, this goes and
this becomes 10. Now, we take the next minimum, which happens to be here, supply is
30, demand is 60, so we provide 30 here, this is completely exhausted and then this
becomes 30.
And we realize that, we have to use this 30 and this 10 so that, the supplies and demands
are met. So now, if we add this from row wise, we will get 30 40 and 60, and column
wise we have met. Now, this is possible only if the problem is balanced which means,
total supply is equal to the total demand. So, we now have a solution and at the moment
we do not know, whether it is the optimum or the best solution, but it is expected to be a
704
good solution, because we have made allocations progressively based on cost, by
considering the least cost one first.
So, this method has a generic name, it is called minimum cost method, now the total cost
of transportation would be 6 into 30 is 180 240 80 240 and 120 with the total cost of 860.
Now, there are some interesting things about it, we also realize that, every time we make
an allocation here, we are either completely using up what is available here or
completely meeting what is required here. Because, every time we update the supply and
the demand, therefore every time we make an allocation, we are either completely
utilizing what is available or completely meeting what is required.
Therefore, we would be able to do it in, there are 3 supplies and 3 demands, so using 5
allocations, generally we will be able to do it, unless one allocation does both of utilizing
the entire supply as well as meeting the entire demand. Such problems are called
degenerate problems and how to handle degenerate problems, I have explained in the
operations research course. So, we will now look at a solution, which will have 5
allocations if there are 3 plus 3 items.
We will be able to do it in 5 allocations, because the problem is balanced and the last
allocation will automatically exhaust the supply as well as meet the demand. So,
generally if we have m supplies and n destinations, we talk of m plus n minus 1
allocations, with which we will be able to complete this. So, the first stage of this
solution is a minimum cost method that I have spoken about, now this gives us a solution
with the cost of 860.
Many times when we solve the transportation problem by hand as I have illustrated here,
we would be using another method called Vogel’s approximation method, which on an
average, can give slightly better solutions than the minimum cost method. I have also
explained the Vogel’s approximation method in much more detail in the course on
operations research. But, right now for the purpose of this discussion, we would use the
minimum cost solution and then try and see, whether this solution can be made better or
can be improved, now we do that in a in a different way.
705
Now, if this solution is not the one that has the minimum cost then there is some other
solution that has a cost lower than this and such a solution is expected to have at least
one different allocation. So, right now there are 4 places, where there is no allocation, so
we check, if any one of the presently unallocated position can actually get an allocation.
So, if we try this, so we have 30 here, so if we try to put 1 to make an allocation here
then we have to reduce 1 here to meet this requirement of 30 40 60 20 60 and 50.
So, if we put a plus 1 here then I have to put a minus 1 to meet this then I have to put a
plus 1 to meet this and I have to put a minus 1 to meet this as well as to meet this, this is
called a loop. So, when I put a plus 1 here then my cost will go up by 4, it will reduce by
6, because I have taken away 1, it will go up by 1 by 8, because of this plus 1 and it will
go with the minus 4, because I have removed 1. So, the net cost of putting 1 here is, 4
minus 6, which is minus 2 plus 8, 10 minus 4, 6, so it is plus 6 and by putting 1 here, my
cost is only going to increase and it is not going to decrease.
In a similar manner, if I try and put 1 here in this position, so I will get a plus 1 minus 1
plus 1 and minus 1. So, the cost will be a plus 8 minus 12 plus 8 minus 6 and this is
minus 2, so there is a gain by putting a plus 1 here, so 8 minus 12 plus 8 minus 12, so
there is a gain in this, so I get a minus 2. So now, when I try to put a plus 1 in the other
unallocated position, which is this position, so I will get a plus 1 minus 1 plus 1 minus 1,
so I will get a plus 6 minus 6 plus 12 minus 4, which is equal to 8.
706
So, the cost will increase and if I put 1 here, it will be plus 7 minus 6 plus 12 minus 8,
which will again be positive. So, 7 minus 6 1, 1 plus 12, 13 minus 8 is plus 5. So now,
we observe that, putting something here, adding 1 here can reduce the cost, so we try to
add as much as we can so that, the cost will come down. So, the maximum that we can
do here is 10, so that this will become 0, this will become 40 and this will become 20.
Now we realize that, the total cost of this solution is 120, 6 into 20 plus 80 plus 240 plus
80 plus 320, the cost is 840 and that 840 comes from the fact that, we found out that the
gain is 2 per unit and then we have added 10 units here, so the gain is 20 and therefore,
860 becomes 840. Now, this is the next solution with the cost of 840 and now we need to
check, whether this solution can be improved further, so we try and add 1 to all
unallocated positions.
So, when I do this here, I get plus 4 minus 6 plus 8 minus 4, so 4 minus 6 is minus 2,
plus 8 10, 10 minus 4 is 6, 4 minus 6 is minus 2, minus 2 plus 8 is 6, 6 minus 4 is 2
which is positive. So, I will get a plus 2 here, now when I put something here, this will
become a plus 1, this will become minus 1 plus 1 minus 1 plus 1 minus 1 and the loop is
closed, so 6 minus 6 0, 0 plus 8 is 8, 8 minus 6 2, 2 plus 8 10, 10 minus 4 is 6. So, I get
plus 6, when I put a plus 1 here, this is the loop, 7 minus 6 is 1, 1 plus 8, 9, 9 minus 6 is 3
and when I put here, this is the loop, 12 minus 8 4, 4 plus 6 10, 10 minus 8, 2.
707
So, all of them will only increase the cost, the cost does not decrease, so this solution is
the best solution or the optimum solution. The method which we have just now seen,
which took the previous solution from the minimum cost method to the optimum is
called stepping stone method and the stepping stone method will provide the optimum
solution from a starting solution. The minimum cost method solution is called a starting
solution, it is also a basic feasible solution and the stepping stone method can help us to
get the optimum solution.
Many times when transportation problems are solved by hand or even when efficient
programs are written, another method called the modified distribution method or the u v
method is usually used, instead of the stepping stone method. I have explained the u v
method in detail in the operations research course. Now, for the sake of illustrating this
from a point of view of distribution, I have explained it using a combination of minimum
cost method and the stepping stone method.
Both the minimum cost method and the stepping stone method are very intuitive
methods, that give us good starting solutions as well as finds the best solution from the
starting solution. Other methods such as Vogel’s approximation and u v method, which
are widely used, are also intuitive. But, they are also computationally a little intensive
and a little different and might also require some principles of operations research and
duality, etcetera to explain how they work.
Now, for the purpose of making it simpler and for illustrating it, I have used a
combination of the minimum cost method as well as the stepping stone method. So, this
is how we solve the transportation problem optimally, one of the ways of solving it
optimally. And this is the final solution, where this 30 that is present here, 20 will go to
this customer, 10 will go to this customer, this 40 will entirely go to this customer and
out of this 60, 20 will go here and 40 will go here.
So, both these people will get their items from two suppliers, this person will get it from
these two, the other one will get it from these two, this person will get only from one
person. Similarly, this person will be supplying to two people, this person supplies to one
person, this person supplies to two. So, this is how the basic transportation problem
works, some of the important aspects in the transportation problem is, the assumptions
708
that there is a unit cost of transportation and that the total cost of transportation is
actually a product of the unit cost and the quantity.
For example, when we computed this 840, we said 6 into 20 plus 8 into 10 and so on, so
if 20 units are transported, the cost is 120. Now, this linearity, proportionality, additivity,
assumption makes the problem a linear programming problem here, but in practice we
also observe that, the actual cost of transportation need not be the per unit cost multiplied
by the quantity.
When we actually look at truck loads that move, now the truck has a certain capacity and
irrespective of what quantity that goes into the truck, as long as it is within the capacity
of the truck then there is the cost of transportation, is actually the distance between the
source and destination multiplied by the cost of traveling a unit distance and need not be
dependent on the quantity that is being transported. In addition, the truck would have a
fixed cost that would be there for a hiring the truck and utilizing it or the cost associated
with owning the truck and if there are some tolls then additional cost is being paid.
So, when we looked at the cost structure, which is very different from the cost structure
that we are looking at here, that would lead us to different problems, which can be
solved using techniques and principles of operational research. We would be seeing
some of them, at the same time we would also be extending this to multiple stages.
709
So, we will next see the multistage transportation problem which is here. So now, we
explain a multistage transportation problem, which is quite common in the context of
supply chain.
So, we could think in terms of 3 supplies or 3 factories, for the purpose of illustration, we
have 2 warehouses or 2 intermediate points, where they are stored. And then for the sake
of illustration, we have 3 customers, the supplies are 100 80 and 60, and the requirements
are 90 70 and 80. Now, we are going to assume that, the items from here move like this,
so if something goes from here to here, it would either go through this or it would go
through this.
So, it would go through one of the warehouses and then it would reach the ultimate
customer. The first model is, we are not going to restrict the capacity of the warehouse,
we are going to assume, that the capacity of the warehouse is infinity. Or as far as this
problem is concerned, the capacity of the warehouse is greater than or equal to 240
which means, we are allowing the possibility of all the items to go through this and then
go there or the other possibility of all the items going through the second warehouse and
then reaching the customer.
Now, when the capacity of these warehouses is less than 240 then it mandates that, some
of the items will have to go through the second warehouse. So, the first model that we
will see, will not consider capacity, it will just leave it like this. Now, we could again
710
thing in terms of a similar formulation to the one that we have seen here. So, then we
need to look at defining two sets of variables.
So, we could think in terms of C i j X i j plus double sigma C j k y j k, now here Y j k is

the quantity transported, j to k. So, here j is the intermediate storage which we call as the
warehouse, i is the production facility which could be a factory and k are the customers
who are here. So, we solve a two stage transportation problem, transportation in the first
stage involves the X i j variables, transportation in the second stage would involve the Y
j k variables.
Now, the constraints also become a little different from this, so whatever goes out of this
should be less than or equal to the capacity. So, X i j less than or equal to a i which is
fine, whatever that reaches here should be greater than or equal to the requirements. So,
X j k summed over j should be greater than or equal to b k, now k stands for the
customer, b k is the demand of the customer, so whatever reaches the customer, is greater
than or equal to b k.
Now, in the event of a balanced problem that is, total supply is 240, total demand is 240,
exactly 90 70 and 80 will reach and all the 100s 80 and 60 will be exhausted. If it is not a
balanced problem then depending on, if this is higher than this then all the demand will
be met. Even in that case, more than the demand will not be supplied, because the cost
will go up. It is a cost minimization problem, as long as these costs are greater than or
711
equal to 0, there will always be a solution, where only the exact requirement is met and
some of them will not move from this.
If we have a situation where the total supply is less than the total demand then all the
total supplies will be exhausted, some of them will not get their demands. So, in such
cases, we have to carefully define the problem, particularly if there is a situation, where
the total supply is less than the total demand, if we put an explicit condition that it is
greater than or equal to b k then it will show infeasibility, because it will not be able to
meet all the demands.
So, this formulation has to be written accordingly, particularly for unbalanced problems,
so it is actually customary to make it balanced by adding a dummy. If the total supply is
less than the total demand then it is customary to add a dummy here and create a notional
one so that, we balance it and then at the end say, that whatever goes out of it, actually
does not go out of it. So, when we formulate it as a linear programming in this case, one
needs to be a little careful to give due consideration to unbalanced problems.
Because, there can be situations, where total supply can be less than or equal to the total
demand. So, there is also another set of another constrains which links this, because as
far as this is concerned, whatever that goes out, should be less than or equal to whatever
that comes in. So, you have a set of intermediate constraints, which will tell you that, for
all these intermediate facilities, sigma X j k whatever that goes out, is less than or equal
to whatever that comes in, sigma X i j (Refer Slide Time: 45:37).
So, y j k is what goes out, so in this definition y j k is what goes out, so both X i j and y j
k are greater than or equal to 0. So, this two stage problem the way we have defined here,
so there will be 6 X i j variables, there will be 6 y j k variables and then there will be 3
constraints here, there will be 3 constraints here, there will be 2 constraints here, so there
will be 8 constraints and 12 variables. One can solve the linear programming formulation
this way and then try to get the optimal solution.
712
The optimal solution is shown here with X 1 1 equal to 70 and X 1 2 equal to 30, so X 1
1 will be 70, X 1 2 will be 30 from here, 2 2 is 80, 3 2 is 60.
So, 2 to 2 is 80, 3 to 2 is 60, (Refer Slide Time: 46:44) y 1 1 is 20, y 2 1 is 70, and 2 3 is
80. So, this is 150, so 80 plus 30, 110 170, so this is 20, so this should be 70 and this
should be 20. So, this is the final solution the 70 that comes here goes here, 30 comes
here, 80 comes here, 60 would come here and they get distributed as 20 70 and 80
713
respectively. So, if we have capacity constraints then we put another capacity restriction
on this.
And that will tell us that, sigma X i j is less than or equal to capacity of j summed over i
and the solution can change depending on the capacity constraint that we have
introduced. So, one could solve the linear programming formulation here, it would still
be an LP, one could do something else, if there is no capacity here. Then it is only true
that, if I have to transport this from here to this or this or this, I will route it through an
intermediate warehouse. It could be any one of them, I could route it through the
intermediate warehouse such that, the distance travelled is minimized.
So, what we can actually do in this case, is to try and find the shortest path from each of
these three to each of these three, passing through these nodes. So, try to find the shortest
path and the shortest distance, so if we find the shortest path and the shortest distance,
we will now have the 9 distances.
714
And we can now directly formulate a 3 by 3 transportation problem with 100 80 60 and
90 70 80 with the value of C i j representing the shortage distance from this. So, 1 to 1
what is the shortage distance, either this distance or this distance, so the minimum of
them we can write here. So, if we actually compute this and then a multistage
transportation problem can be looked upon as a single stage transportation problem,
provided there are no capacities in the intermediate stages.
So, that is how we solve a typical multistage transportation problem, here we have
shown the linear programming formulation to the multistage transportation problem.
715
The next problem that we will look at is called the fixed charge transportation problem.
We have already seen the formulation of the fix charge transportation problem in the
previous lecture, now we will explain the fixed charge transportation problem.
And then use a numerical illustration to explain the fix charge transportation problem
and we will see this in the next lecture.
716
Prof. G. Srinivasan
Lecture – 38
Transportation and Distribution Models (Continued)
In this lecture, we see the fixed charge transportation problem. We have already seen this
formulation of the fixed charge problem in the previous lecture, where I have explained
that, this problem becomes significant when there is a fixed charge of transporting from a
given source to a given destination. We have also seen that, the toll that is paid while
moving items from i to j could represent the fixed charge between i and j. There could
also be other fixed costs that are associated with travelling from a particular place to
some other place. So, the fixed charge transportation problem has two components, C i j
X i j plus f i j Y i j, the objective function is to minimize the two costs.
717
So, minimize C i j X i j plus f i j Y i j, so the decision variables X i j represents the

quantity transported from i to j and Y i j represents, if arc i j is chosen. So, Y i j is a
binary variable, it is equal to 0 or 1, it is equal to 1 if we choose to transport from i to j
and it is 0 otherwise, X i j is the quantity that is transported from i to j. On the face of it,
it appears that, this objective function is similar to that of the multistage transportation
problem, but it is not so.
In the multistage transportation problem that we saw in the previous lecture, we had an i,
j and k, so there were three entities. Now, we have only i and j, we have two entities, so
it is very similar to the regular transportation problem, where there are only two entities a
set of supply points, which may be warehouses and a set of destination points, which
could be retailers or customers. So, there is only one stage and there are two entities and
they the material flows from one set of entities to another set of entities.
Typically, material we could think in terms of three supply points and three demand
points and items flow this way. Now, there is an a i, which is the supply available here
and there is a b j, which is the requirement for this. Now, there is a f i j, which is a fixed
cost and a C i j, which is the unit cost of transportation, if we move from i to j. So, we
have typical constraints, which will be, whatever goes out of this should be less than or
equal to a i.
718
So, sigma X i j summed over j is less than or equal to a i, whatever that reaches here
should be greater than or equal to b j, so sigma X i j summed over i is greater than or
equal to b j. So, once again the same caution that, the balanced problem would have total
supply equal to total demand, in which case we can even change this with the equations.
All the items will be transported and all the items will be received, we will not be
receiving more than what is demanded, because the C i j’s and the f i j’s are not, none of
them are negative.
Therefore, there will be a solution, where all these are consumed and all these are taken,
now if we have a situation, where the total supply exceeds the total demand then all the
demand will be met. More than the demand will not be given, because if we have to
transport more than the demand, the total cost will only increase. And some of these
supplies will not be utilised fully, so the same structure will hold. Only if we have a
situation, where the total demand is more than the total supply then this formulation
would give us infeasibility, because it would not be able to meet all the demand.
In such cases, what we do is, we balance it by creating another supply and then say at the
end after solving, that whatever goes out from this supply actually does not go and
therefore, correspondingly some of them will get less than what is demand. So, we need
to observe that, particularly when we have a situation, where the total demand exceeds
total supply, we have to be little careful in using this formulation. But, when we choose
to use this, we have to create another dummy supply or a non existing supply.
Now, we also have to relate the Y i j variables to the X i j variables, so X i j is less than
or equal to M into Y i j which means that, if I am transporting from a particular i to j
then Ii have to choose that arc first and then I transport which means, when I choose that
arc, I incur the fixed cost of transportation. So, I incur the f i j, so only when I choose, I
will be able to set and this big M is a very large number. So, when I choose, I can send
as much I can and if I do not choose this, I cannot send.
So, when Y i j is 0, X i j will automatically be 0, so we will now have if there are m

supply points and n destination points, there are m into n X i j variables and there are m
into n Y i j variables. So, it has 2 times m n variables, where this m n X i j variables are
continuous variables, they need not be integers, the other m into n are binary variables,
719
these Y i j’s are binaries. So, we have a problem that has binary variables as well as
continuous variables.
Now, there will be m constraints that relates the supply, there will be n constraints that
relate the demand and then there will be m into n constraints that link the supplies to the
demands. So, this fixed charge transportation problem, which has an additional fixed
charge on the arc, will now have 2 into m n variables, where m n are binary and the other
m n are continuous.
And it has m plus n plus m n constrains, where these m and n are the supply demand and
these are the linking constraints that link the X i j to the Y i j, so Y i j’s are binary and X
i j are greater than or equal to 0. Now, we can solve this as a optimization problem like a
mixed linear integer programming problem, where some of these variables are binary
variables, some of these variables are continuous variables. Now, we can solve this to get
the optimum solution to this problem.
So, let me look at a 3 by 3 problem, that we have already seen and try and add a fixed
charge to it.
720
So, we look at this 3 by 3 problem, we look at the same problem that we saw in the
previous lecture to explain the transportation, except we also add a fixed charge to it. So,
the fixed charge is now shown here, that the fixed charge is 100 120 110, so 100 120 110
80 120 60 120 80 and 60. So, the values shown here are the f i j values and the values
shown here are the C i j values. Now, if all the f i j values are 0 which means, there is no
fixed charge, the problem will become, automatically it will become a transportation
problem.
So, if all of them are 0 then there is no need to have this set of constraints, so it will
become a transportation problem. Now, for the purpose of illustration, let us look at the
optimum solution to the transportation problem without the fixed charge and then let us
see, how the solution looks like. We set X 1 2 is 20, X 1 3 is 10, 40 20 40, 2 3 is 40, 3 1
is 20 and this will be 40. So, this is the optimum solution to the transportation problem
without the fixed charge.
So, the C i j X i j component of this is 20 into 6, 120 plus 80, 200, 200 plus 240 is 440,
440 plus 80 is 520, 520 plus 320 is 840 that we saw. Now, if we have to look at this
solution and implement the fixed cost associated with the transportation then we would
be incurring a 120 here, because we have used this. We will incur 110 here, because we
have used this, we would incur 60 we have used this, we would incur another 120 and we
would incur another 80.
721
We have already seen that, as a transportation problem, there will be five allocations
here, because the problem is balanced and non degenerate. So, this would give us a cost
of 490, so the total cost will be 840 plus 490, which is 1330, now this is a feasible
solution to the fixed cost transportation problem. So, what we have done is, we have not
actually solved the fixed charge transportation problem, we have actually solved the
transportation problem without the fixed charge.
And then we have added the fixed charge to it, to try and get a solution, whose cost is
1330, but if we solve this formulation of the fixed charge transportation problem, the
formulation that we have written here. And for this example as I said, there will be 9 plus
9, 18 variables, there will be 3 plus 3 6 constraints plus another 18 constraints, so 24
constraints and 18 variables. If we solve this problem optimally, our solution would be X
1 1 is 20, so let me write it with a different colour, so X 1 1 will be 20, X 1 3 is 10, X 2 3
is 40, X 3 2 is 60, this is the solution.
So, 20 10 40 and 60 is the optimum solution to the fixed charge transportation problem,
so let us try and evaluate the costs associated with it. The cost of transportation is 20 into
480, the fixed cost is 100, the cost of transportation is 10 into 8, 80, the fixed cost is 110,
the cost of transportation is 40 into 6, 240, fixed cost is 60, cost of transportation is 60
into 8, 480 and this cost is 80. So, now this will be ((Refer Time: 14:00)) 880 and this
will be 350, so the total cost will be 1230.
So, the cost will be 1230, now the solution that is shown here, this is the optimum
solution to the fixed charge transportation problem by solving this formulation. Now, we
compare this solution to the other solution, where we evaluated the total cost by
considering the fixed cost. Now, 1330 was the cost here, while the cost is 1230 here,
there is a saving of 100.
We also realize that, there are only four variables in the solution here, this cost has come
down, this cost has marginally gone up, but this cost has come down, the very fact that,
there are only four variables in the solution implies that, we are going to have four arcs in
the solution and only four fixed costs. Whereas, this would force us to have five and then
there were five fixed costs and then we realized, there are lot of saving there, which was
resulted roughly in the saving of 100. Though this has gone up a little bit, this has gone
down significantly to give us a saving of 100.
722
So, fixed charge transportation problem, if we have the fixed charge assumption and
there is a cost associated with it then we need to look at this formulation and then solve
it. So, the next question that comes is, can we have another method like our minimum
cost method or the stepping stone method, to try and actually solve the fixed charge
transportation problem without the formulation. A very simple thumb rule would be, now
to define C dash i j is equal to C i j plus f i j by m i j. So, let me explain this thumb rule,
now let us look at the same 3 by 3 problem.
Now, the unit cost of transportation from 1 to 1 is given by 4 and the fixed charge is 100,
now the supply that is available here is 30 and the demand is 20. So, if we consider this
position, so C dash i j will be C i j, which is 4, plus f i j is 100 divided by m i j is the
minimum of this and this, minimum of the supply and the demand, so minimum of 30
and 20, which is 20. So, this would give us a number equal to 9, so I am going to write
this number here, this number is 9.
So, like this we can calculate this number, which captures both C i j as well as f i j into to
get a new number, which is called C dash i j, which could represent some kind of an
equivalent cost, considering both the C i j’s and the f i j’s. The motivation comes from
the fact that, this fixed charge, if we are going to use this to transport, we are incurring
an additional charge of 100. Now, this additional charge of 100 is used in transporting
from this to this.
723
The maximum that can be transported from this to this, is the minimum of the two
numbers 30 and 20, which is 20. Therefore, this 100 should be used in transporting a
maximum of 20 from this to this, which is the minimum of this 30 and this 20. So, per
unit, the apportionment of the fixed cost can be 100 by the minimum of these two, which
is 100 by 20, which is 5 and therefore, C dash i j becomes C i j plus f i j by m i j. So, this
way, we can complete the calculations for C dash i j and then the values now become 9.
So, the values become 9 10 11.66 10 10 and 7.5 10 9.33 13.2.
724
So, let me again explain this 13.2, 13.2 would come as 12 plus 60 divided by the
minimum of these two, so 12 plus 60 by 50, 12 plus 1.2, which is 13.2. Now, we can
solve a transportation problem with 30 40 and 60 and 20 60 and 50. So, this becomes the
normal transportation problem and when we solve that transportation problem, we would
get a solution, which is 20 10 40 and 60. So, we will get a solution like this and the cost
associated with this will be different, it will not be 1230, it will be different.
But, once we get this optimum solution, we now have to go back and map this optimum
solution here and then try to find out the fixed cost as well as the transportation cost.
Transportation cost will become 880, fixed cost will become 350 and the total will be
1230. If we do 9 into 20 plus 10 into 11.66 and so on, that will be different, so it will not
be 1230, we have to do that separately and try and get the cost to the original problem.
But, this thumb rule need not give us the optimum solution all the time, in this example it
has given us the optimum solution.
There could be instances, where this thumb rule may give a cost that is actually higher
than the cost given by the optimum solution, which is solved using the integer
programming. That is because from an operation research point of view, the fixed charge
transportation problem is a hard problem. And therefore, we do not have effective thumb
rule based methods, which can give the optimum solution in all the instances, in some
instances it may give the optimum solution, as it did in our example.
So, if we really want the optimum solution to the fixed charge transportation problem
then we can solve this or use this and use a solver to solve this, which again necessitates
the availability of a solver. And if we do not have a solver, we could think in terms of
other methods such as a Branch and Bound algorithm and so on, such methods exist in
the literature, they can be used to get the optimum solutions. At the same time, if we are
not very keen on the exact optimum solution.
But, if we are content with the good solution, which can be slightly higher than the
optimum in some instances, but a good performance on an average, one could look at
this thumb rule. And then create an equivalent transportation problem from the fixed
charge transportation problem then solve it and get the solution, so we can do that as
well.
725
The next problem that we will look at, is what is called a point to point one warehouse
one retailer problem.
Simple point to point one warehouse one retailer problem, now let us assume, we explain
it through a numerical example. So, we will assume now that, we want to transport 41
tons of items from this warehouse to this retailer. Now, in order to transport these 41
tons, we can do that using different types of trucks. So, let us assume, we are looking at
726
two types of trucks, one truck can carry 16 tons per truck and the other can carry 9 tons
per truck.
Now, the cost of hiring and using this is 10 and the cost of hiring and using this per truck
is 7. So, the question boils down to, how many of these trucks do we take, how many of
these trucks do we take such that, we are able to transport the 41 tons and we are able to
do it with minimum cost. So, the problem would be, let X 1 be the number of 16 ton
trucks used, let X 2 be the number of 9 ton truck used. So, the total cost will now be 10 X
1 plus 9 X 2 or 10 X 1 plus 7 X 2 plus is the total cost, which has to be minimized.
Subject to, if I am using X 1 trucks of this type I can carry 16 X 1 items, if I use X 2 type
trucks of this type I can carry 9 X 2, is greater than or equal to 41. So, the problem is to
find out X 1 and X 2, that minimizes 10 X 1 plus 7 X 2, subject to 16 X 1 plus 9 X 2
greater than or equal to 41, X 1 X 2 greater than or equal to 0 and integer. Now, the
integer is extremely important, because we can only hire an integer number of trucks. So,
it becomes a very simple linear integer programming problem.
And if we are going to consider only point to point from one supply to one destination or
one warehouse to one retailer, there is going to be only one constraint. So, this is a single
constraint integer programming problem, all integer problem, because both X 1 and X 2
will have to be integers. So, there is an objective function that is to be minimized, there is
a single constraint and there is a all integer. Now, this problem is called a single
constraint knapsack problem.
The standard single constraint knapsack problem is a maximization problem with a less
than or equal to constraint and variables greater than or equal to 0. Now, this version of
the knapsack problem is a minimization problem with the greater than or equal to
constraint, nevertheless it is also an knapsack problem, so one need not look at it as a
different problem. The original knapsack problem talks about putting items into a
knapsack such that, the value that we put in is maximized, X j is the number of units of
different items that can be put in.
Now, this coefficient could represent, there is a single constraint, which could either be a
volume constraint or a weight constraint, not both. If we have both then it becomes a two
constraint problem, this is a single constraint and if it is a volume constraint, this would
represent the volume of each item and the volume that we put in should be less than or
727
equal to the volume available. So, it is a maximization problem with a single constraint
of a less than or equal to type.
This version of the knapsack is a minimization problem with the single constraint, which
is greater than or equal to type. Now, the same problem can become a little different if
we have different warehouses and different retailers or if we have one warehouse and
multiple retailers R 1 R 2 R 3 and so on. This person may require 41, this person may
require 52, this person may require 37 and so on. So, again we may have to now define
X 1 1 and X 1 2 as the number of trucks of 16 ton and 9 ton, X 2 1 and X 2 2, X 3 1 and
X 3 2 and so on.
As long as there is no availability restriction on the number of 16 ton trucks and the 9 ton
trucks, the problem can be separated into several single constrained knapsack problem,
one for each arc. But, if there is a capacity restriction or the availability restriction on the
number of 16 ton trucks and 9 ton trucks available then we would get then it would
simply become 10 X 1 1 plus 7 X 1 2, there will be another 2 1, 2 2 and all that.
And then there will be a constraint, which will say X 1 1 plus X 1 2 plus X 1 3 plus X 1 n
will be less than or equal to the available number of 16 ton trucks. So, we could have
another set of availability constraints and the problem becomes very different, it is not a
single constraint knapsack problem anymore. Now, let us restrict our discussion only
with the single retailer or a single customer and therefore, we will have the single
constraint knapsack problem.
Now, if we again we have a solver, one can easily solve this using the solver, there is
only a single constraint. If we have only two variables and a single constraint, one can
even do an enumeration and try and solve it. But then there are better ways of actually
solving this using some kind of a Branch and Bound algorithm.
728
Now, if we use only 16 ton trucks then we will require X 1 equal to 3, which is upper
integer of 41 by 16, so we will need 3 trucks here and the cost will become 30. If we use
only 9 ton trucks then we would need 5 such trucks, so I would have 5 and then the cost
will become 7 into 5, which is 35. So, we could simply think in terms of an enumerative
algorithm, so we can simply enumerate and say X 1 X 2, so X 1 is 0, X 2 is 5, cost is 35.
X 1 is 1 so then we have 1 here, so this would give us a capacity of 16, so from here, we
need a capacity of 25, so we will have 3 here.
So, 1 into 10 is 10, plus 21 is 31, if I have 2 of this, so I have 32, out of this I require only
9, so 2 comma 1, so I will have 27 and I will 3 comma 0, I will have 30. So, only four
enumerations are possible in this case and the best solution is here, which is use two 16
ton truck and one 9 ton truck to get this 41 and with the total cost of 27. But then we can
also understand that, the enumerative method that we have shown here, is not
computationally efficient and can become intractable when some of these numbers
become higher.
Therefore, the single constraint knapsack problem also becomes a hard problem and
some efficient method such as Branch and Bound algorithms are also available, which
essentially do the complete enumeration. But, they do what is called implicit
enumeration, they do not do a complete enumeration explicitly. Now, by creating some
729
bounds, it is possible to reduce the number of computations here to a meaningful
number.
And then to try and get the best solution using Branch and Bound algorithms or implicit
enumeration algorithms, which reduce the number of explicit enumerations and tries to
give us the optimum solution. I have actually outline the Branch and Bound algorithm to
solve a single constraint knapsack problem with a maximization objective and less than
equal to constraint in the advanced operations research course, where we look at the
cutting stock problem.
A very quick note on this, we can also do this, now what we can do is, this is a
minimization problem. So, this is a minimization problem, if we can solve it as a linear
programming problem, an LP solution to it which means, we are ignoring the integer
restriction. Since it is a minimization problem, we can actually sort the variables in terms
of the ratio 10 by 16 and 7 by 9, 10 by 16 is like 0.625, 7 by 9 is 70, so 9 7's are 63, 70,
0.77.
730
So, LP solution would give us X 1 equal to 41 by 16, which is 2.90, so 16 5’s are 80, 100
6 are 96, 40, 2.5625 and cost equal to 25.625. So, this will be the LP optimum, it would
choose, because there is only one constraint, the LP will choose only one variable and it
will choose that variable, which has the smaller coefficient. So, it will choose X 1 and
give it a value 41 by 16, so 25.625 will act as a lower bound to this. So, we can start the
problem by saying X 1 is equal to 2.5625 with lower bound equal to 25.625, which
makes it 26.
Now, from this, since X 1 equal to 2.5625, we can branch of with X 1 equal to 0, X 1
equal to 1 and X 1 equal to 2 and then when X 1 equal to 0, we will solve another LP.
So, when X 1 is equal to 0, this would we are solving an LP that minimizes 7 X 2,
subject to 9 X 2 equal to 41. So, X 2 will be 41 by 9 and objective function value Z will
be 41 by 9 into 7, so 41 into 7 by 9, 287 by 9 which is 9 3’s are 27, 180 9 8's are 72,
31.88.
Now, we look at this one, now the feasible solution for this will be X 1 equal to 0, X 2
equal to 5 with Z equal to 35, now this gives us an upper bound for this. Now, when X 1
equal to 1, we will now have a solution X 1 equal to 1, X 2 equal to 25 by 9, 25 by 9 is 9
2’s are 18, 70, 9, 7’s are 63, 2.77. Now, Z value will be 10 plus 25 into 7 by 9, 175 by 9,
this is 10 plus 9 1’s are 9, 85, 9 9’s are 81, 40, 9 4’s are 36, 19.44, 29.44 is the lower
731
bound and the corresponding feasible solution is X 1 equal to 1, X 2 equal to 3 with Z
equal to 21 plus 10, 31.
Now, X 1 equal to 2 would give us X 1 equal to 2, X 2 equal to 1 with Z equal to 27, so

this is integer feasible and this integer feasible is lower than both the lower bounds and
therefore, it is optimum. Now, the Branch and Bound algorithm that we have seen,
borrows ideas from lower bounds and upper bounds to between LP’s and IP’s. So, it
involve certain operations research ideas and right now, I am just explaining it to show
that, it is actually optimum.
A little more detail and depth and a detailed way of representation of this as I already
mentioned, is considering a maximization problem with less than or equal to, which I
have explained in further detail in the advanced operation and research course, where we
address the cutting stock problem. Now, this Branch and Bound method is another
method, which is slightly different, similar and different to the enumeration method that
we saw here and gives the optimum solution of 27.
But, as I mentioned, if we expand this problem to multiple retailers and more

importantly, keep a capacity restriction on the number of trucks available of different
types. In practice, it need not be only two types, it can be multiple types, the problem
becomes much more complex and it has to be solved using integer program.
732
The next model that we can see is one warehouse and multiple retailer, which is a logical
expansion of this.
But, the problem is slightly different here, there is a single warehouse, there are multiple
retailers and so on. Now, we will assume that, there is a truck, which goes from here
point to point, it goes to the retailer and comes back and then it can go from here to
another retailer and come back and the third retailer and come back and so on. So, let us
assume that, there is a retailer j, there is a distance of d j, which the truck covers. Let us
733
say, both ways put together, there is a distance d j, which is this into (Refer Slide Time:
40:23) this plus this into this is d j.
Now, we will assume that, there is a single truck which is going to transport, so whatever
quantity that is transported from here to here, is less than or equal to a single truck load
and or, but we also do point to point which means, from here it goes it will come back
and then it will go to here or some other place and come back. Now, if there are j equal
to 1 to n retailers, so we will have d j values d 1 to d n, now the total time required by the
truck by a single, if a single truck has to do it, the total time required is the sum of d 1 to
d n.
But, if this truck is available only, now we will call this time as using this notation, we
would call this time as t 1 to t n. Instead of distance, we call them as time t j, j equal to 1
to n and the truck is available only for a time of capital T then the question is, how many
trips should this truck take or how many trucks do I need such that, I meet the demand of
all of these t to n. So now, that problem is called a bin packing problem, it is called a one
dimensional bin packing problem or it is called a bin packing problem.
And the problem is very similar to saying that, I have a bins of height T and then I need
to put items into this bin, I have n items, each item has a height, the j th item has a height
t j. So, i have to put these items one above the other, it is a one dimensional bin, so I
need to put them one above the other such that, I do not exhaust the height of the bin.
How many bins do I need so that, I can pack all these n items, where the j th item has t j,
is a one dimensional bin packing problem.
Problem can also be explained in different ways, I have several sticks of length capital T,
now I need to make smaller sticks n number of sticks with each length equal to t j. How
do I break these longer sticks to get these lengths and I have to use the smallest number
of such long sticks with length T, how. What is the minimum number of the longer stick
with length T that I require, through which I can break them and get these t j n lengths,
where the j th smaller stick has a length t j, that is another way of looking at this.
The third way to look at this is, given n numbers t 1 to t n, form minimum number of
groups such that, the sum of the numbers in each group is less than or equal to capital T,
all these is the description of what is called one dimensional bin packing problem. So,
the formulation of the one dimensional bin packing problem will be like this.
734
Now, we are going to assume that, there are n items with values t 1 to t n and there is a
bin with length T, where we need to put it. Now, we want to minimize the number of
bins, so we will assume now that, Y 1 to Y n represents the number of bins that we have.
It is fairly obvious, that if each of these is smaller than T, which is not a bad assumption
then we can take, there are since there are n items, we can take n bins and put each item
into a bin.
So, that would give us a solution with n bins, but that is not what we want, we want to
have the number of bins to be as small as possible. So, what we do is, we would assume
right now that, we have n bins and then try to minimize the number of bins. So, we
would say, minimize sigma Y i, i equal to 1 to n, Y i equal to 1 if bin i is chosen. So, by
minimizing the number of sigma Y i, I essentially try and minimize the number of bins
that I choose, so the objective function would be this.
Now, what are the constraints, each item should go to only one bin and before that, I
need to introduce another variable X i j equal to 1, if item i goes to bin j. So, I will now
say, sigma X i j equal to 1 summed over the bins j or let me look at it slightly differently,
I have used subscript i for bin, so I would use subscript j for the item, so item j goes to
bin i. So, summed over all the bins, X i j equal to 1 for every item j, so every item j will
go to exactly one bin, which is given by this constraint.
735
Now, as far as the bin is concerned, sigma X i j summed over j, there are the items go to
the particular bin and t j now is the length of the j th item, so t j X i j is the length
associated with this bin. Now, this should be less than or equal to T that we have, which
is the overall bin length and then this should be less than or equal to T into Y i for every
i. So, Y i is equal to 1 when the i th bin is chosen and when the bin is chosen, it gives us
a capacity of T and then whatever that goes into the bin, should be less than or equal to
this, so this is one way of doing it.
The other way of doing it is, what is shown in this slide and these are explicitly written
as two different constraints. So, we could simply say, t j X j less than or equal to T which
means that, we need to have a linking constraint between the X i j’s and the Y i’s.
736
Now, we could also write X i j is less than or equal to M into Y i which means, only
when the bin is chosen, I can put something into the bin. Now, this has more constraints,
so this constraint can be eliminated by simply putting (Refer Slide Time: 48:00) T into Y
i. Now, this is the formulation of the bin packing problem, so both Y i and X i j’s are
binaries. Now, one can solve this bin packing problem to try and get the optimum
solution, but the way we have formulated for example, if we have 10 items then we are
talking of...
We are now also defining Y i, so Y will take 10 values and X i j will take 100 values, so
we will have 110 variables. And this formulation will be, this will give us 10 constraints,
this would give us 10 constraints, so there will be 20 constraints. So, if we actually solve
a problem with 10 items using this formulation, we would have 110 variables and 20
constraints, we can do that. Alternately, we could think in terms of some other heuristic
or thumb rule based algorithms, which would say that, we may be able to pack this
instead of a maximum of 10, we will be able to pack it in say, 4 or 5 bins.
Then, if we restrict the number of bins from 1 to 10, we can bring it down to 1 to 5 then
the number of variables and constraints will reduce. So, it is customary to look at a
heuristic solution to the problem then find out a certain number of bins and then solve the
optimization by taking advantage of the fact that, a heuristic solution would actually
737
result in lesser number of variables than constraints. So, we will look at some of the
heuristic solutions to the bin packing problem in the next lecture.
738
Prof. G. Srinivasan
Lecture - 39
Bin Packing and Travelling Salesman Problems
In this lecture, we continue the discussion on the bin packing problem. The application
to transportation model comes from, assuming that we have a central warehouse and we
have to transport to a large number retailers or customers and there is a point to point
transportation, which means a truck will go from here to this and then it will come back
to the starting point and then it will go to the next customer after picking up material
from here and so, on.
So, the distance are time the truck takes from the warehouse, to customer j can be given
as d j or C j or t j, if it is the time taken, so this time would include the loading the
unloading as well as the transportation time. Now given n customers, where customer j
requires a time of t j and if a vehicle is available for T time units, the problem is to find
out, which of these customers are orders will be allotted to each of these vehicles. Such
that, the number of vehicle is minimized and the sum of the t j’s allotted to each vehicle
will be less than or equal to T.
739
As mentioned in the previous lecture, this reduces to what is called, a one dimensional
Bin packing problem, where the one dimension is time here, time required by each of the
customers, to be compared with time available with each of the trucks. So, the problem
is to find out the minimum number of trucks, now we explain it through a numerical
illustration.
So, we would, so we assume that, there are 10 requirements and the time, for each of
these requirements are given in these 10 numbers and then the total time availability is
80. So, the problem is to find out the minimum number of trucks, such that we are able to
meet all the demand, so now t j is given here, now we have to find out, which of these
items, now we can treat each one of these as items that goes into a bin, the bin’s length is
80 items have lengths 11, 73, 13, 37 etcetera.
So, which of these items will go to which of these bins, such that we minimize the
number of bin’s, as well as ensure that the length of items that go into a bin does not
exceed the bin length. So we have seen the formulation in the previous lecture. So we
now assume that, Y j equal to 1, if bin j is chosen and X i j equal to 1, if item, i goes to
bin j. Now, we know that with 10 bin’s, we can definitely meet all the requirements, so
we are going to assume that, there are smaller sub set of Y 1 to Y 10, which we would
like to minimize.
740
So, the objective is to minimize sigma Y j, where Y j equal to 1, if bin j is chosen. This
depends on the notation, now we are using j for the bin and i for the item, therefore
summation Y j, j equal to 1 to 10. Now, each bin each item is to be associated to only
one bin, so sigma X i j summed over j is equal to 1, for every i, which means each item, i
will go to exactly one bin.
And then for each bin t i X i j is less than or equal to 80 Y j. Now X i j equal to 1, if item
i goes to bin j. So for a for a particular bin, sum of the X i j’s is the number of items that
are attached to that bin. Sigma t i X i j is the length of sum of the lengths of the items
attached to that bin. So, this is the sum of the lengths of the items attached to bin j and
this should be less than equal to 80 and we can assign i to j only when it chosen, only
when Y j is 1.
Therefore, this constraint less than equal to 80 Y j or capital T Y j will ensure that, the
not only items will be assigned to bins that are chosen, it will also ensure that when
items are assigned a sum of the length’s does not exceed this 80. And then we have Y j
and X i j as binary variables, now this is a binary integer programming formulation, this
can be solved optimally using an I P solver.
But, right now as I mentioned for 10 items, we will have 10 variables here and 100
variables here, so we will have 110 variables and then there will be 10 constraints here,
there will be 10 constraints here, so there will be 20 constraints. Now, this constraint is
for every item this constraint is for every bin, so there will be 110 variables and 20
constraints, so in general, if there are n items, we would have n square plus n variables
and 2 n constraints.
Now let us we can solve this optimally and when we solve this optimally, we would get a
solution, which would be like this. Y 1 equal to 1, Y 1 equal to Y 2 equal to Y 3 equal to
Y 4 equal to 1, which means 4 bins are chosen. So, the item with 73 goes to this bin
items with 11, 13, 37 and 19 go into this bin. The item with 51 and 29 go into this bin
and the item 17, 39 and 23 go to this bin. So the solution will be this will be item number
1, so X 1 1, X 3 1, X 4 1 and X 10 1 and so on.
Now, the sum of these is eleven plus 19 is 30, 43, so 80, 73, 51 plus 29 is 80. 17 plus 39
is 56, 56 plus 23 is 79. So we find a solution with 280’s and 79 and a 73. Now, let us
assume, we want to solve this using heuristics, several heuristics are available, we will
741
see 1 or 2 of them, so most of these heuristics first sort these items in the decreasing
order.
So, when these items are sorted in decreasing order, we get 73 51 39 37 23 29 23 19 17

13 and 11, so the we first apply a very commonsensical heuristic, so we take the first
item and then we observe, we create a bin, so we say bin 1 and put the first item 73 here.
Now, we take the second item and see, if that would go into the bin 51 will not go into
the bin, because the length of 80 will be exceeded, so create the second bin, so bin 2 has
51.
Now take the 3rd item, which is 39 and see whether, it will go into this bin first it would
not go and then see, whether it can go into this bin 51 plus 39 does not go into this bin.
So create a 3rd bin, bin 3 with 39. Now, take the 4th item with 37, so it would not go into
this bin, because the length will become 110, it would not go into this bin, because the
length will become 88, it will go into this bin, so 39 plus 37 with length equal to 76.
Now take the next item 29, see it will go to this bin it would not go, then try this bin, so it
fits in 51 plus 29 is equal to 80. Now look at the 6 item with 23, it would not go into this
bin, it would not go into this, it would not go into this, create a 4th bin with 23. Now take
the next 1, 19, so 73 it would not go, again it would not go, again it would not go, so 23
plus 19. Now, take the item with 17, it is still would not go because this will be 90, it
742
would not go into this bin, it would not go into this bin, it will go into this bin 23 plus 19
is 42 42 plus 17 is 59, so it will go into this bin.
Now, look at the 13, it would not go into this bin, it would not go into this bin, it would
not go here, now here it is 23 plus 19, 42, 42 plus 17 is 59, 59 nine plus 13 is 72, so it
will go into this bin. And then take the last one, it will not go here, it would not go here,
it would not go here, now let us see this is 72, so it would not go here, so create a 5th bin
with 11. So, now, this method gives us 5 bins, whereas the optimum gives us 4 bins, it is
a heuristic method, it does not guarantee optimum all the time, therefore it would do that.
Now this, this method gives us it is a minimization problem, so a heuristic algorithm

would give, what is a feasible solution to the problem will be an upper bound to a
minimization problem, so the number of bins 5 is on upper bound to the optimum value,
which is 4. Now, this is a very popular heuristic, this is called the next best decreasing
heuristic, next fit decreasing heuristic, it is called the next fit decreasing heuristic,
important thing, it is a decreasing heuristic, because the items are already arranged in
decreasing or non increasing order, therefore, it is a decreasing heuristic.
In fact, this is called the first fit decreasing heuristic, this is called the first fit decreasing
heuristic, which means every item, we see, we take it, we see whether it will go to the
first bin, if not we will see, whether it will go to the next bin and then it will go to the 3rd
and then so on. So, whichever it fits first it will be put there, so it is called a first fit
decreasing heuristic, now we have other heuristics some of them are called best fit
decreasing heuristic and so on.
So, if there is more than 1, where it can fit, now at that point, we try and fit it into the 1
where it fills up to the maximum, that is called a best fit decreasing and then the worst fit
decreasing would be, if there is a choice, put in the bin, where it would actually fill the
minimum after filling. So, like that we can develop different kinds of heuristics for this
particular problem, now one of the advantages of using this heuristic is now here, we
have a solution with 5 bins.
So, we know that, there is a solution with 5 bins and once, we know that, there is a
solution with 5 bins or 5 trucks in this example; we are only interested in solutions,
which are 4 or lower, because we already have a solution with 5. So, using a very simple
743
heuristic like this, we get 5 bins and if we still want to solve it optimally and we want to
check whether, 4 is possible then we have to come back.
And solve this optimization problem, but the nice thing about, this is now you be do not
have to define j equal to 1 to 10, it is enough to define j equal to 1 to 4, because we
already have a solution with 5, so it is enough to define j equal to 1 to 4. And now see
that simply define Y j equal to 1, for j equal to 1 to 4 and X i j equal to 1, if item I goes to
bin j, so we will have 4 plus 40, which is 44 variables. And then we will have 14
constraints n square plus n no here it will change now, this one be n square plus n.
Now, here the number of variables Y j will be 4, X i j will be 40, so we will have 44
variables and in terms of constraints, this is for every i, so there are 10 constraints, this is
for every j, there are 4 constraints, so there are 44 variables and 14 constraints, whereas
in the old formulation, if we use 10 bins then we would have 110 variables and 20
constraints.
So, here are a set of problems, where it is actually advantageous to solve it, using a
heuristic first and then based on the heuristic solution, if we approach the optimization,
we observe that the number of variables and constraints in the optimization will come
down.
744
So, that is one advantage of using this heuristic. At the same time, if we are all right with
the heuristic solution, then we could say that well, this solution would be acceptable to
us. But, if we actually look at this solution and then we can tweak this solution, now that
is not part of the best fit decreasing heuristic, the best fit decreasing heuristic would stop
here, because it is interested in minimizing the number of bins and this heuristic gives 5
bins.
But, then we could do many things, if you look at the loads, it is 73, 80, 76, now this will
become 72 and this is 11, now we can do a few more things by adjusting some of them
by a little adjustment one can try and actually, reduce the number of bins from 5 to 4.
But, if we were to implement this solution particularly for our truck example, then we
observe that one truck is going to be used for 73 minutes another for 80, 76, 72 and 11.
So when we start applying this heuristic to the situation that, we are talking about, we are
also interested in some kind of a load balance among the trucks.
Because, it is also hard to say that, one truck will be utilized only for 11 minutes, while
few trucks are utilized for 80 76 73 and so on, so we also need to have a secondary load
balancing objective, if we were looking at solving this kind of a problem. We can do few
more things, we can actually relate this problem to a traditional machine scheduling
problem, in fact, we can try and do a few more things.
745
Now, if we want to solve with the better load balance, for 5 bins, we can try something
else, we can now look at something like bin 1 bin 2 bin 3 bin 4 and bin 5, now we will
go back and put this 73 here. Now the next will not go to this bin we will put a 51, we
will put a 39, we will put a 37 and then we will put a 29 and then the next one, we can
put a 23, so it will come here, so 29 plus 23 is 52, now 19 will come here, because if we
see, where we assign for example here.
Now at this point, the first 5 go to the first 5 bins, the 6th one is 23, so 51 plus 23 is 74,
39 plus 23 is 62, 37 plus 23 is 60 and 29 plus 23 is 52, so the best place to put that 23 is
here, because the loads will get balanced. Even though, we can put that 23 here, which
will be closer to 80, but then we realize that the loads will be will be unbalanced and then
there will be a load imbalance, therefore, the best thing to do is to put that 23 here, so
that the loads are also balanced as they can be, so we put this 23 here to get 52.
Now the next one is 19, now if we look at the loads are 73, 51, 39, 37 and 52, so this is
the smallest, so we get better load balance by putting this 19 here. Similarly, the 17 will
come here, the 13 will come here, now this 11 cannot come here, because 73 plus 11 is
84. Now we check this is 64, so this can take that 11. This is 56, this is also 56, this is 52,
so this will finally, take that 11 and now the loads for 5 bins will be 73 64 56 56 and 63.
So now, we have a solution, if we were to use this 5 bins, then this is a solution with
better load balance than this, so depending on the problem and depending on more than
one objective, we can use suitable heuristics. Now, the same problem, we can try
something more and see that, if we have 4 trucks or bins, what happens to and if we start
allocating this 10, what happens to the loads. So, 73 will come here 51 will come here,
39 will come here and 37 will come here, so automatically from a load balance point of
view, 29 will go here, because this is the smallest load 37 plus 9, 46 plus 20, 66.
Now, 23 will come here, 39 plus this will give a 62. Now 19, this is 51, so 19 will come
here, 51 plus 19 is 70. 17 cannot come here, because it will exceed, now to put this 17,
17 also will not come, 17 can come here, 51 plus 19 is 70. So, it will 7 plus 17 will
exceed, 39 plus 23 is 62, so 62 can handle is this 17, this is 66, so this will handle get the
17 and this will get this 13. And then we do not have a place to put this 11, because this
the load will exceed 80, if we add 11 here, also it will exceed 80, here the load will be 62
plus 17, 79, it will exceed 80, this will become 66 plus 13, 79, it will again exceed 80.
746
But, then we have to put this 11 somewhere, so we can put this 11 here and then say that,
this is 73, this is 81, this is 12 plus 7, 19, 79, this is 12 plus 7, 19, 79. Now we realize
that, the optimum solution was able to give us, 80 73 80 79, the heuristic solution gave us
73 81 79 79. So, using this heuristic solution, now we observe that, we would not be able
to squeeze all the orders in 80 time units, if we have only 4 trucks or if we use this
heuristic.
And if we want to minimize the number of trucks, such that, I can squeeze everything
into 80, the answer is 5 whereas, the optimum is able to squeeze all the 10 items into 4
trucks or 4 bins, such that the length is 80, here it says, it requires an additional length of
81. Now, we can either solve this optimally, this is easily the best thing that, we would
like to have, but in the absence of solving the integer programming optimally and if we
rely on heuristic solution then we have to understand here, there are times, we would get
even the optimum solution through some of the heuristics, we can try several heuristics.
Now, we will able to minimize the number of bins for a given length and then within
that, we may be able to get a load balance and we may also say given certain number of
bin, what is the length of a bin that, I require. Now, it happens to be 81, what is the
minimum length that, I require, now this problem is equivalent to the machine
scheduling problem, it is called a parallel machine scheduling problem. Now, for this
particular example, if I force 4 bins and then apply a load balancing algorithm, it says
that, i may require 81 minutes on one of the trucks or one of the bins whereas, the
optimum is able to do it, within 80 minutes. Right?
747
We next move to the travelling salesman problem. Now here what we do is we assume
that, there is a central warehouse and then we have a set of customers. Now we - this is
like the equivalence of where we look at milk run. So, we start a truck from the
warehouse and then we meet the requirements of all these customers in one trip, which
means that even, if each of these have a requirement.
The sum of all the requirements is less than or equal to the capacity of the truck or
theoretically or mathematically, we can assume that the truck has a very large capacity,
which is close to infinity and can handle the requirements of all these customers.
Practically, we would say that, we know these requirements and the truck capacity is
higher than the requirements of all of these, in such a case that truck will start from the
warehouse, visit each customer once and only once deliver and it will come back to the
warehouse, now that is called the travelling salesman problem.
Now, if you assume one warehouse and 6 customers, we have 7 points, we can assume
that, we will have a 7 by 7 distance matrix and then starting from W, we want to visit
each customer once and only once and then we come back to the starting point. Now, we
have already seen several aspects of the TSP, in the advanced operations research course,
where I have given the formulation for, the travelling salesman problem. And also I have
explained many 2 or 3 different types of branch and bound algorithms to solve the
travelling salesman problem.
748
And I have also looked at a lot of heuristics extensively in that course, where we have
studied the travelling salesman problem in detail. What we will do now is will take
another look at the formulation of the T S P. And then we would also look at some
heuristics and I would also try and give you, some of the optimum solutions to the
travelling salesman problem and make it a little brief, because a more elaborate treatment
is available already in the other course.
Now, we have already seen this formulation of the travelling salesman problem. Now X i
j equal to 1, if the salesman visits city j immediately after visiting city i. Now d i j is the
distance matrix. Now, again in the context of the travelling salesman problem, we will
assume d i i that is distance between a point and itself to be infinity and not 0. So, a
distance between a point and itself, so large that, we does not visit the point immediately
after visiting it, for example, if put d 1 1, d 2 2 etcetera to be infinity or very large value,
the variables X 1 1, X 2 2, X 3 3 etcetera, will not be in the solution.
Now this will ensure that, if salesman is in a particular city X i j equal to 1 summed over
J, which means, if the person is in a particular city, the person has to visit only one city
immediately. So, that will be taken care of by this and if this person is in a particular city
he or she should have come from 1 city, so that is taken care of this. But, then we need
the extra constraints otherwise the T S P will become an assignment problem or if we
have 5 points, including the warehouse.
749
It may give solutions like this, which means it goes from say 1 to 2 and then 2 to 1 and
then the salesperson suddenly goes 3 to 4, 4 to 5, 5 to 3. Now they will satisfy this, so
this will be like solution to the assignment problem. Now, things like this are called
subtours and these have to be eliminated. The solution to the T S P is always a full tour,
which means you will have a solution like 1 3 4 2 5 1, this is called a full tour 1 to 2, 2 to
1, 3 to 4, 4 to 5, 5 to 3 is called a subtour.
750
So, we need this plus we have, what are called subtour elimination constraints. Now, if
there is a subtour let us say, there is a first, we define u i minus u j plus n X i j less than
equal to n minus 1 and then we define it for i equal to 1 to n minus 1, j equal to 2, 3 up to
n and i not equal to j.
Now, if there is a subtour, let us say let us say, there is a subtour 1 to 2 and 2 to 1 or first
let us assume that, there is a subtour, which is there is a subtour, which is 3 to 4 and then
let us say for a 5 city problem, there is another subtour that is 1 to 5. So, a subtour 3 to 4
will be eliminated here.
751
Because, there will be a constraint which will be say u 3 minus u 4 plus 5 X 3 4 is less
than or equal to 4 and u 4 minus u 3 plus 5 X 4 3 less than or equal to 4, there will be a
constraint like this. Where I have just written u i minus u j plus n X i j less than equal to
n minus 1, now if this, if such a subtour exist then these 2 constraints have to be satisfied.
Now, we realize that adding vertically, you are you will get, if both are equal to 1, you
will get 10 less than equal to 8, which is not true, therefore this constraint will ensure that
subtours involving say a 2 city subtour or whatever 3 4 4 3 are automatically eliminated.
The same thing will hold, for any for example, if we have 3 4 5 then there will be u 3
minus u 4 plus something u 4 minus u 5 plus something and u 5 minus u 3 plus
something, so vertically adding it will give us an inconsistent set.
Therefore, all these subtours involving this and this are eliminated, now if there are
subtours that are involve city 1, the reason one is important is because of this i equal to 1
2 to n minus 1 j 2 to 3 to something. So, if there is a subtour involving city 1, then that
subtour apparently 1 will get a feeling that, it will not be eliminated by this, for example,
if we start writing for this, then you will get a thing like.
752
u 2 minus u 1 plus 5 X 2 1 is less than or equal to 4, if I add u 1 minus, u i minus u j, so u

1 minus u 2 plus 5 X 1 2 less than equal to 4, but then you do not have a 2 1 plus j to i, j
is not defined from for 1, so you would not get a second constraint. But, the way we
understand this is if there is a subtour that involves 1, then there has to be another
subtour at least one more subtour, which does not involve 1 and since all subtours not
involving 1 are eliminated, automatically subtours involving 1 will also be eliminated.
Now, this will not violate a full tour, because a full tour will have something like this 1 to
3, 3 to 4, 4 to 2, 2 to 5, 5 to 1, now that 5 to 1 constraint does not exist in the formulation,
therefore you will have only 4 constraints here, you would not have 5. Now, the u i’s and
u j’s can be, so defined, such that these set of constraints will be satisfied, so these set of
constraints are called subtour elimination constraints and they are used extensively, if we
have to solve a T S P optimal.
But I having said that this T S P is usually not solved as a mixed I P. because this is a
mixed I P, you realize now that u i’s are greater than or equal to 0, now if there are n
cities then we have n square variables for X i j. So, there will be n square variables for X
i j and n variables for u i, so we will have n square plus n variables, where these are
binary and these are continuous.
And in terms of number of constraints, now we will have n constraints here, we will have n
constraints here, we will roughly have about n C 2 constraints slightly less than that
753
may be n C 2 minus n constraints and so on. So, we will have of the order of n C 2
constraints here, so this will still become a large problem, when the number of cities
increases and this would roughly increase of the order of n square, so we will have a
large number of constraints. And this is a binary formulation plus a mixed formulation,
because this these set of variables are continuous, so we normally do not use this
formulation to solve the travelling salesman problem, instead we use branch and bound
algorithms.
We use branch and bound algorithms, there are several types of branch and bound
algorithms that are available; the more popular 1 is the 1 by Little and 3 other authors.
Now, this is based on the penalty approach, I have given a detail description of the
Little’s branch and bound algorithm in the advanced operations course and hence I am
not repeating the Little’s algorithm once again here. So, the reader the viewer could go to
the other lecture and take a look at the Little’s algorithm, through which we can solve
this. Now, several heuristic algorithms are very popular for the branch and bound, for the
travelling salesman problem and for the purpose of completion, let me just explain 1 or 2
heuristic algorithms here. So, we take a 5 city T S P and explain through this.
754
So, a 5 city T S P with distance matrix is dash 10 8 9 7 10 dash 10 5 6 8 10 dash 8 9 9 5 8 6,

now here we have to assume that, there is a warehouse and there are 4 other places that, we
have to visit. So, one could be the warehouse 2 3 4 and 5 could be the customers. In fact, in a
T S P does not matter, because the solution 1 2 3 4 5 1 is the same as a solution 2 3 4 5 1 2,
therefore, it does not matter, where the truck actually starts.
But, then it is customary to say then it starts from the warehouse and it completes one
round by visiting each and every customer once and only once and then it come back to
the starting point. Now, the simplest heuristic is called the nearest neighbor, if I assume
that, I am in city 1, now I go to the nearest neighbor, whose is 5 from here, I go to 5, now
from 5, I go back and realize, I can go to 2 or I can go to 4, so I go to 2. Now, I come to
2 and the nearest person is 4, now I have only 3 remaining, so 3 and 1.
Now, the length of this will be 1 to 5 is 7 plus 5 to 2 is 6 plus 2 to 4 is 5 plus 4 to 3 is 8

plus 3 to 1 is also 8, so this is 13 18 plus 8, 26 plus 8, 34. We do not know whether, this
is the best solution or we could make this solution better. So, we simply expand this by
trying that, we start with 2 first, so I start with 2, so I go to 4, this is the minimum from 4,
I realize this is the minimum, but I have already visited 2, so I go to 5 and from 5, I go to
1 and then 3 and then 2.
So, this will be 5 plus 6 plus 7 plus 8 plus 10, so this will be 11, 18, 26 plus 10, 36. Like
this, we can do a nearest neighbor with each of the cities as the starting city and then we
755
can proceed and then choose the best out of the these solutions, that we have. In this
example, we can show that, this is actually the optimum solution and the nearest
neighbor is able to give the optimum solution right away. Then we start from 1, this is
because of the nature of the numbers that, we have in this particular example.
Another very popular heuristic is called the 3 opt heuristic, the 3 opt heuristic essentially
does the 3 way exchange, so what we do is if you have a T S P let us say a large T S P.
We have a large T S P. Now, let us assume this is a solution, so 1 2 3 4 5 6 7 8 1 this
need not be the optimum or the best solution, this is a T S P with 8 nodes or 8 cities, now
I take a solution like this and then I take one of these as a b, another pair is p q and the
third pair is u v.
Now, we realize, there are eight nodes, so a b p q u v can be chosen in many ways. We
would also realize this is a b this is p q, but then I left a node and then I chosen u v here.
Now, using this, I can actually create some feasible solutions, for example I can remove
this, I can remove this and I can remove this, now I can do this, I can add these 2, I can
add this and I can retain this, this will be another feasible solution. For example, I can
remove this, I can do this, this could be another solution. So like this we have several
possible solutions.
So, if we have a a b p q u v, then we can actually get 7 more solutions to this, we can
have a b p u q v a, p b q u v a, p b u q v a, q u b p v a, q u p b v a, u q b p v a, u q p b v a,
756
so where every a b p q u v chosen, we can have 7 solutions. Now, if there are n, we take
a starting solution, so there are n vertices, there are n edges in the solution, so a b, p q, u
v represent 3 out of the n, so they can be chosen in n C 3 ways and then each of these a
b, p q, u v, we can write 7 more solutions from this.
So, if we are given a T S P with say 5 vertices, then 5 C 3 is 5 X 2, which is 10, 10 into
7, we can get 70 solutions out of this. The only thing is because we have 5, then we
realize that these 70 units are not unique, there will be overlap, because when we have a
5 by 5 to chose a b p q u v, then some of them have to overlap. So, a lot of those 70 will
overlap, but then we can still chose the best of them and we can proceed. Whereas, if we
have a problem with 10 vertices, (Refer Slide Time: 47:29)
And now, we realize that we a b p q u v itself can be chosen in 10 C 3 ways, 10 into nine
into 8 by 6 another 7, so for every solution will be able to this, so we have 800 and 12 7s
are 84, we have 840 solutions, which we can get, from a single solution and then we can
keep improving this. So, 3 opt is a very popular method, one can write a simple computer
program, which would do the 3 opt heuristic and then we could get the best values of the
solution.
But then when we apply these ideas particularly to trucks and distribution and so on, we
also need to look at a little beyond this kind of a T S P and then add a few more aspects
to the T S P. Now, this T S P is symmetric, which means distance between i to j is the
same as j to i, but then when we actually apply these on practical problems, we
757
understand that actually the distances are not the same, because there could be one ways
and there could be different highways, you need to take when, you go from one place to
another, so we have T S P, which are not symmetric.
We also observe that particularly in critical distributions like say mail motor and other
such distributions, one would observe that actually, even if the routes are the same, the
time taken to go from one place to another will depend on the time of the day. When the
travel is made the travel is made early in the day one can reach a little faster and when
we move into peak hours, then we have, we go much slower.
So, there is a speed dimension that come in, there is a distance the speed from, which the
time is to be computed and the speed will depend on the time of the day, because most of
these also at the end of the day result, in problems were certain delivery times have to be
met and so on. So, we have to take into accounts speed of the vehicle also, we also could
have windows, time windows, where delivery to a customer has to be made either, before
a certain time or after a certain time or within a time interval.
So, all these result in handling the travelling salesman problem very effectively, in terms
of the nature of the objective function, in terms of additional constraints, that we may
have, also in terms of getting the data depending on dynamically updating and getting the
data so that the correct data is used and the corresponding T S P are formulated and
solved. The next thing that, we will have to look at is called the vehicle routing problem,
758
where we would have a central warehouse, we would have a set of customers, now we
realize that the demand of these customers is such that, we may require more than one
trip, the vehicle capacity cannot handle the demand of all of them. So, we get into a
situation, of which customer goes into, which trip and then how the trips are made; now
that problem is called the vehicle routing problem.
A solution to vehicle routing problem could be, one trip will be like this and the other
trip would be like this, but in general, which cities or customers will go to which vehicle,
such that we minimize the total distance. We use minimum number of vehicles and then
we are able to meet the demand is called a vehicle routing problem, we will see various
aspects of the vehicle routing problem, in the next lecture.
759
Prof. G. Srinivasan
Lecture - 40
Vehicle Routeing Problems
In this lecture, we address the Vehicle Routeing Problem. The vehicle routeing problem
occurs in distribution, when we have to distribute from a central warehouse to multiple
customers. Let us say there is a warehouse and the 6 retailers or 6 customers and we have
to distribute. Now, each person has a certain requirement, which is given by q j, is the
requirement of the material by customer j and we will assume that trucks take this
material, from the warehouse, distribute to the customers and returns back to the
warehouse.
This is an example of the milk run that, we spoke about in an earlier lecture, now if these
q j’s are small and such that all the 6 demands put together as less than one truck load, it
is economical for the truck to start from the warehouse, visit each of these customers
once and only once and come back to the warehouse, because the truck capacity will be
is assumed to be greater than or equal to the sum of the requirements.
In such case, it becomes a traveling salesman problem with 7 nodes the warehouse plus 6
nodes. Many times that does not happen, the individual q j’s will add up to a number,
760
which is more than a single truck load, it is customary to use the notation Q, for the truck
capacity. So, if sigma q j is greater than Q, then we need multiple trips, we either need
multiple trips or we need multiple trucks or multiple vehicles, to do it, if time is a
constraint.
If for example, we need 2 trips and we have only one truck, then we have the issue, of
which of these customers or cities, it is customary in vehicle routeing problem, to say
these customer nodes as cities and the warehouse as a depot. So, which are the cities that
will go in the first trip and which are the cities that, we will go in the second trip or the
next trip. And then in what order would the vehicle, visit the customers, who are
assigned to a trip is a vehicle routeing problem.
We may have 1 vehicle and have multiple trips, where for example, in the first trip, it
might go to 1 and then 2 and then 3 and back to the warehouse and in the second trip, it
might do 6, 5, 4 and back to the warehouse. Now let us assume that, we start at time
equal to 0 and let us it takes some time to finish the first trip and comeback, let us say it
starts it takes about, 4 hours to finish the first trip. Then the second trip can start only
after 4 hours and by the time, we reach the last customer on the second trip the time may
be delayed; so that would necessitate multiple vehicles.
So, whether we talk of multiple vehicles or multiple trips, the problem is the same as
long, as we are not looking at time windows, which means as long, as we are not looking
at, times either before which or after which or in between, which we need to visit the
customer. So, we use the terms trip and vehicle interchangeably, but wherever it is
required, we talk about vehicles, it is also customary to say that in this case, even though
there are 2 trips are possible, we would say there are 2 vehicles.
And the first vehicle will visit 1, 2, 3 and back and the second will visit 6, 5, 4 and back,
they may do it sequentially, they may do it parallelly, but for the purpose of argument or
discussion, we would say that there are 2 vehicles. So, a vehicle routeing problem
essentially has more than one vehicle and if the vehicle routeing problem has a single
vehicle, then it becomes a traveling salesman problem. So, T S P can always be seen, as
a case of a vehicle routing problem, where there is only one vehicle.
Now, we have already seen some methods to solve the traveling salesman problem, now
we will look at some ways to solve the vehicle routing problem. The vehicle routing
761
problem is widely used in almost all kinds of distribution. It is used extensively in
pickup problems, employee pickup, school children pickup, mail pickup, it is also used
in a lot of delivery problems and so on. So, we will now see formulation for the vehicle
routing problem and then we will explain it to an example and then we will also see
some heuristic algorithms to solve the vehicle routing problem. So, several formulations
exist for the vehicle routing problem.
Now we will give one formulation which is this minimize C i j X i j.
762
Now, we have this formulation for the V R P, we will try and explain this formulation in
2 stages, we will first explain this part of the formulation and then we will add this part
of the formulation. Now, what we do in this formulation is this.
Now suppose, we have a 6 customers and we call them 1 2 3 4 5 6, now what we do is

the matrix will now have then there is a depot, so we will have the depot and 6
customers, so the distance matrix is a 7 by 7 matrix. So, we will now call a 7, which is
called the start node, then we will have an 8, which is called the destination node. Now
what we do is each of these 6 customers are connected to itself. So, there will be a 1 to 2
a 2 to 1 and then 1 to 3, 1 to 4, 1 to 5, 1 to 6 and then you will have these arcs.
So, each one is essentially connected to each of the others, so this will be complete and
so on, 2 and 5 will be connected like that, we connect all these 6 and then we complete
the network by connecting 7 all these are directed arcs. So, there are connections and
then 7 is connected to all the nodes and all nodes are also connected to 8, this
formulation is by Achutan and Caccetta.
So, we compete the network by adding these 6 customers here having them in the
middle, connecting each of them with the other and then this 7 and 8 act as some kind of
source and sink assuming that, the warehouse is now split into 2 entities. One from
where the vehicles go and the other, the where the vehicles at the end come into, so both
actually represent the warehouse, but then they are like a source and sink on this
763
network.
So, each of these distance between i to j on this is given from the distance matrix depot
to each one is written here and each one of them to depot is also written here, so
minimizing C i j X i j is on the entire network, so it is on all arcs in the network, so we
would still have variables like X 7 1 X 7 2 X 7 3 and so on. X 7 we will not have
variables like, X 1 7 2 7 from 7, it will only be 7 1 7 2 7 3 7 4 7 5 7 6, there will not be
arcs that go into 7.
Similarly, there will not be arcs that, that go out of 8 you will have all arcs come into 8 1
8 2 8 3 8 4 8 5 8 6 8, there is no 7 8, so that is how the network is drawn, so summed
over all arcs. Now, what happens is if I take a particular node 1, now few things can
happen to that node, one is 1 the node 1 or any node say let us say node 3, I can reach
node 3, either from the source, which means it is a first node that the vehicle comes or I
can reach it from some other node.
So, X i j plus X n plus 1 j is equal to 1 for all with the j from 1 to 6, so if I take any of
these nodes, I reach that node either directly from the source or through some other node,
which is given by this set of constraints. Similarly, if I take a node from a typical node, I
can go either directly to the destination or to some other node, so X i j plus i n plus 2 is
equal to 1. So, either I go to the destination or I go to some other of these nodes, from
here, I do not go back to 7 here, this is not defined for 8, this is defined this is not defined
for 7.
Here, the i j is only up to this 6, 6, 7 comes here, 8 comes here, so n is equal to 6, so n

plus 1 is 7 N plus 2 is 8, so this represents the n plus 2 node. Now, both these will
represent that, if there are 2 vehicles that, I am considering, then capital V is equal to 2,
which means exactly 2 of these arcs out of these 6, there are 6 arcs, exactly 2 of these
arcs will take the value 1, which means this will force from 7, I will directly visit 1 2 out
of this 6 directly.
Similarly, X i n plus 2 equal to V, again V is equal to 2, which is the number of vehicles,

so this will force 2 out of these 6 arcs to be equal to 1, which means, from 2 out of these
6 arcs, we will reach directly to the destination. It may also happen that 1 vehicle may
just go from 7 to 1 to 8, while the other vehicle could do 7 2 4 3 6 5 8, it can happen, but
it will ensure that, 2 arcs go out of this, 2 arcs reach this and X i j equal to 0 1
764
minimizing the total distance would give us the distance that, we travel in the vehicle
routeing problem.
So, this is a very nice formulation, when it comes to solving a V R P, more as a flow
problem, solving it as a as a multiple path problem, from source to destination with some
constraints and restrictions on the path that, we actually have here. So, this represents the
formulation for the V R P. Now this would have for example, if we have the warehouse
and then we have 6 customers or 6 destinations, now this X i j will have 36 variables.
I equal to 1 to 6, j equal to 1 to 6, it will have 36 variables, then X 7 1 to 7 6 will be 6

variables and 1 8 to 6 8 will be another 6 variables, so 36 plus 6 plus 6, so if I have N
customers, it will have N square plus 2, N variables, where N square X i j variables will
be there and another 2 N variables formulas. The number of constraints, this will be N
constraints this will be N constraints, this will be 1 constraint and this will be 1
constraint, so we will have 2 N plus 1 constraints.
So, this will have N square plus 2 N variables and then it will have 2 N plus 1, 2 N plus 2
constraints and you can solve it, as a binary I P to get the solution to this particular
problem. And it is always possible to define only one path here and solve the T S P, so
once again the T S P will also have N square plus 2 N variables and it will have 2 N plus
2 constraints. So, it is a more precise formulation to solve the T S P instead of having the
other set of constraints, so essentially, we have N square plus 2 N variables and the 2 N
765
plus 2 constraints to solve the T S P also, this can also be used to solve the V R P. Now
we have to explain some of these, so a typical V R P is not only about, finding the
vehicle to which each city or customer is assigned and the order of visit of the customers
by the vehicles, there are few other constraints that, we may like to have.
Now, this constraint is called this is a restriction on the number of customers that each
vehicle visits, so this is a restriction on the number of customers that, we want the
vehicle to visit, so we say each vehicle will visit less than or equal to l customers. Now,
we could have restriction on capacity, in fact, if we remember right, the motivation to the
vehicle routeing problem comes from the fact that, there is capacity restriction, so each
customer would require a certain demand, which I said, which I used small q i as a
notation.
So, we could either use capital q i or we could have small q i and this W is actually the
capacity of the vehicle, we could use W or we could use Q, to represent the capacity of
the vehicle. Because, if we do not have the capacity constraints, then it is always
possible, if the vehicle has infinite capacity or if the vehicle has more capacity, than
demanded, it is always possible to solve it as a T S P.
And then get minimum distance, because the minimum distance for a T S P with 7 nodes
will always be less than or equal to the distance that, we travel when we use multiple
vehicles or multiple trips as the case may be, if distances follow triangle inequality.
766
Generally Euclidean distance follow triangle inequality and in practice, we would know
that d i j plus d j k is greater than or equal to d i k triangle, inequality is followed and
therefore, a T S P will always give in terms of distance minimum distance compared to
the V R P.
But, what can also happen is that, we might want to do parallel movement of this, we
want to reach the customers within a certain time and so on, therefore we might say that
the 2 vehicles individually may travel, if it is a T S P, then it travels a certain distance.
Now, if we solve a 2 vehicle V R P the distance will be higher, but then from a time
point of view both the vehicles could start at the same time and by the time, they will
finish and comeback to the depot they can do it with sooner time.
So, that is another motivation to go for a V R P type of an implementation rather than a T

S P, even if we do not have capacity constraints, but if we have capacity constraints, this
will necessarily force us to use multiple vehicles and then we have to solve the V R P.
Now, these set of constraints are called distance constraints, now we do not want a
situation where, for example, if I have my total capacity of a vehicle, let us say is 50 and
then and let us say the total capacity of the vehicle is 50 and the total demand is 55.
So, when we actually solve the problem as a 2, this would necessitate 2 vehicles and if
we solve a 2 vehicle V R P, we might get into a situation, where one vehicle does about,
45 distance and the other vehicle does about 10 distance. So, we like to have a certain
balance in the in terms of distance traveled, so we would like to say that each vehicle
does not travel more than a capital T, which is the distance.
So, we could have either all 3 types of constraints, which would be number of customer
constraint, capacity constraint and distance constraint or we could have a subset of these
depending on our actual requirement. So, this would also all of these would require of
the order of N square constraints, because now we realize that y i and y j are some other
continuous variables at we have introduced. X i j is the same L is known. Similarly y N
plus 2 and y N plus 1 are another set of continuous variables that have been defined.
Similarly, u i to these are very similar to our u i minus u j plus N X i j less than equal to
N minus 1 that we have seen, so necessarily they take care of making sure that within a
certain allocation, it does not exceed a certain number. So, again we have something like
N square plus N square plus N square, so put together there will be about, 3 N square N
767
square constraints, for each type. The only addition is a set of y i and y j, so that will be
for in this example with 6 customers, there will be 8 y variables, there will be another 8 u
variables and there is another 8 V variables.
The other catch of course, is that the y i’s the u’s and the V’s are continuous, so when we
add this the problem does, now it moves away from a 0 1 or a binary I P to a mixed I P,
where some of these variables are continuous. The y i variables are continuous, the u i
and the V i whereas, the X i j’s are binary. But, we could still use this formulation very
effectively to solve V R P’s provided, we actually have a solver. If we do not have a
solving facility then we use heuristics or thumb rule based algorithms. Now let us take a
a small example and then show, how we solve this.
So, we have a problem with a warehouse and 6 cities, this is the distance matrix, 0 is the
warehouse or depot 1 to 6 are the 6 cities or 6 customers, the matrix is right now
assumed to be symmetric. So, distance between 0 to 1 is the is the same as the distance
between 1 to 0, 0 represents as I said the depot or the warehouse, in a previous lecture, I
had also explained that in many practical situations this matrix need not be symmetric,
due to multiple reasons.
Both distance and time to travel can change depending on the time of the day depending
on the route that is taken and so on. But, just to illustrate the V R P, we would include a a
symmetric matrix, so I will just give you some solutions here. So, if we apply the T S P
768
formulation.
We would get the solution, 7 5 1 3 6 2 4 8 representing the solution, but distance equal to
134, which is the optimum solution to the T S P, which means we use this formulation
and then we put V equal to 1, we do not use any of these, when it is a T S P, we simply
use this formulation and put 1 V equal to 1, which means one vehicle is good. So, we get
that solution with 134 on this, now if we ignore the load constraint all these, we ignore,
we simply solve for 2 vehicles and then if you solve for 2 vehicles, we have a solution,
which is D 4 3 1 2 6 D and D 5 D, this gives us a distance equal to 140.
Now, when I use D 5 D, what also happens is we would have something like this, we
would, if the solution to this, will give something like 7 4 3 1 2 6 7 and 7 4 3 1 2 6 8 and
7 5 8. So, it will be like in the network that, we had here, 1 path will start from the source
just go to 5 and then go to the destination, so if we look at it from this formulation point
of view the solution will be 7 4 4 3 3 1 1 2 2 6 6 8 7 5 5 8.
So, this will be the solution with 140, it is also obvious that, I have to increase the
number of vehicles, force, it to certain number of vehicles, we realize that the distance
actually increases, which is expected. If we have 3 vehicles, we would have D 3 1 2 6 D
D 4 D and D 5 D with 150, so when I use a capital D, I am using this D as a depot, so as
I said, if we use that formulation on this, we would have 7 to 3, 3 to 1, 1 to 2, 2 to 6, 6 to
8, 7 to 4, 4 to 8 and 7 to 5, 5 to 8.
769
Now, the total distance will be 150 in this you can see that D increases, now if we add
the constraints vehicle capacity Q equal to 15 and the 6 individual capacities are 4 6 3 5 3
6 and if we solve this formulation, for 2 vehicles. Now, we realize that the total is 4 plus
6 10 13 18 21 27, so we need minimum of 2 vehicles, so if we solve for 2 vehicles, we
will get the solution D 1 3 4 D and D 5 2 6 D with total distance total distance D equal to
148.
So, similarly, we can solve this, for various combinations, we could this, we have right
now looked at we have looked at this right now. For example, we could go back and then
we could add some other constraints also. We could now here, we have a situation,
where it visits 3 customers, now we want to restrict it to visiting 2 customers, then we
could bring this and put l is equal to 2 and add this, now we will realize with 2 visits, it
may not be possible with 2 vehicles.
Then, we have to increase the number of vehicles to 3 and we can solve it, for 3 vehicles,
we can look at this as well as this, we will get a different solution, the number of
constraints will get added and so on. So, depending on our requirements, we can actually
do this, but then as I said this is how we solve the vehicle routeing problem optimally
using this formulation. One of the important requirements, when we look at distribution
is that in addition to all these.
we have situation, which is called time windows, for example if we look at the solution
770
D 1 3 4 D and D 5 2 6 D, now let us assume that the solution D 1 3 4 D and D 5 2 6 D,
now the distance traveled will be D to 1 is 20 use a different color. So, D to 1 is 20 plus 1
to 3 is 18 plus 3 to 4 is 20 plus 4 to D is 16 and here D to 5 is 12, 5 to 2 is 22, plus 2 to 6
is 21, plus 6 to D is 19, so this takes distance equal to 38, 58 plus 16 is 74. This has 23,
24, 44 plus 19 is 63, so D to 1 is 21, to 3 is 18, 3 to 4 is 20 plus 16, so 38, 58 plus 16, 74.
So, we get 137 here. So D 5 2 6 D, D to 5 is 12 plus 5 to 2 is 26, 5 to 2 is 22 plus 2 to 6 is

21 plus 6 to D 40, 40 plus 22, 62 plus 12, 74. So we get the total of 148. Now, for a
moment instead of assuming these as distances, let us assume, these also as time taken
and we have trying to solve the problem, such that the total time taken is minimized.
Now, we have if both these vehicle start at time equal to 0 and both these vehicles will
come back to the depot at time equal to 74, because it, so happens that both of them are
equal.
Now, if for, now it will reach the customer number 4 at time equal to 38 plus 20, 58.
Here it will reach customer number 6 at 34, 35 plus 35, now if we have a requirement
that, we need to reach customer 4, before 40 say, then it becomes a problem with time
windows. If we have a situation, where we should visit customer 5 after 20, right now it
will visit customer 5 at time equal to 12, of course for this instance, we could simply go
back and switch the sequence, we can do D 4 3 1 etcetera.
But, the point is if we have a larger pool of customers and we have larger number of
vehicles, we and more constraints, we may not be able to do that. Then we get into
problem with time windows, where there is an explicit restriction of visiting a customer,
either before a certain time or after a certain time or in between a certain time. Now, time
windows problems are not easy to solve, using this kind of formulation, because this
kind of formulation is more a path type formulation, which talks about, which node to
visit after another note.
Now, there in time windows, we need formulations, where we have the time of arrival as
an explicit variable in the formulation, so other formulations exists to solve time
windows related V R P’s. This is not, this is widely used, when we actually do not have
time, windows and we solve the problem more as the problem to find out, which way,
which customer goes to which vehicle and in which order and the subject to all of these.
This does not explicitly consider time of arrival at a customer as a variable and therefore,
771
would not solve the time windows based problems.
Several formulations exist, in fact, V R P has multiple formulation, depending on, how
we wish to solve them, we could use different types of formula, there are other types of
formulations exists, where we could use them for certain bounds and we could also use
them to use some other relaxation based procedure. But, right now for the purpose of
discussion with respect to distribution in a supply chain, we would look at this
formulation, which helps us also model constraints on capacity, constraints on distance,
as well as constraints on the number of customers.
Now, as I said all this is possible, if we have solvers, which can solve this problem for
us, if we have a situation, where we have about 50 or 60 customers, then we are talking
of, here we have 50 square, here we have here, we have if we have 50 customers. We are
talking about the order of 50 square, which is 2500 variables and then we also have
another 50 plus 50 and so on. The problem also gets unwieldy and becomes difficult to
solve with large number of customers.
So, we then require heuristics or thumb rule based methods to solve and there are the
most popular heuristic is called the savings based approach. Once again I have addressed
heuristics, the Clarke and Wright savings based approach, as well as the improvement
over the Clarke and Wright, which is called the Holmes and Parker method, I have
addressed them in the advanced operations research course. But, one again for the
purpose of completion of this, I will do only the Clarke and Wright solution to this
problem and we will leave the Holmes and Parker for the viewer to view it in other
lectures. Now the Clarke and Wright algorithm is based on a very simple idea
772
that I have a warehouse W or a depot as I might call it. We call it a depot and there are 2
customers 1 and 2 or in general i and j. Now if a vehicle goes from the depot, it goes to
the customer and then it meets the customer requirement and it comes back. It once again
goes to this customer and comes back, the distance traveled will be 2 times d naught i
plus d naught j, d 0 is the origin or the depot i is the customer i, so 2 times d 0 i plus 2
times d 0 j.
But, if there is enough capacity and we move from the depot to customer i and then to
customer j and then we come back to the depot, the distance traveled would be d naught i
plus d naught j plus d i j. Now, if distances satisfy triangle inequality, we know that this
is smaller than this, therefore, there is a saving, if we move instead of going to customer i
and back and then customer j and back, if we combine, there is a saving, which is called
S i j, which is the difference between this and this is bigger this is smaller.
So, this on subtraction would give us d naught i plus d naught j minus d i j, so given a d i
j matrix, such as this given a d i j matrix, such as this, it is possible for us to find out an s
i j. Now, please note that this matrix is a 7 by 7 matrix with 0 representing the depot and
1 to 6 representing the 6 customers the savings matrix will be a 6 by 6 matrix and that
will be only for customer pairs i and j, because the 0 is hidden inside the saving. So, now,
what we do is we can compute the savings matrix, for the given distance matrix, let me
write this savings.
773
So, the savings would be 1 to 2, the saving is 16, 1 to 3, the saving 16, 2 to 6, the saving
is 16, 2 to 4 it is 14, three to 6 is 12, 1 to 6 is 11, 3 to 4 is 11, 2 to 5 is 8 1 to 4 is 6, 1 to 5
is 6, 5 to 6 is 5, 3 to 5 is 4, 4 to 6 is 3, 2 to 3 is 0 and 4 to 5 is minus 2. So, this is our
savings matrix, then let us assume Q equal to 15 and small q is equal to 4 6 3 5 3 6 and
vehicle capacity is 15. Now, when we compute the savings matrix, now I have not shown
the savings matrix as a square matrix, I have shown it pair wise, so there are 6 customers,
so these there are 6 C 2, which is 15 pairs, it is also symmetric.
So, saving between 1 to 2 is the same as 2 to 1, so I have just shown the 15 values here.
We just observe 2 things, that when we do this, particularly here, we now realize that 2
to 3, there is a saving of 0, which means the depot 2 and 3 lie on a straight line. Now,
here the saving is minus 2 would indicate the numbers actually do not satisfy triangle
inequality and there is a little violation in these numbers, these numbers are not realistic
and practical from a triangle inequality point of view. But, it actually does not matter,
when it comes to solving this particular illustration. Now, the other thing, I have done is
I have already written these pairs or the savings in this sorted order in the decreasing or
non increasing order, they have been written. So, the Clarke and Wright algorithm will
begin by looking at the highest saving and then it would save 1 and 2.
774
So, it starts with vehicle 1 puts 1 and 2 together to have a saving of 16 and if we put 1
and 2 together, we use 10 units of the demand, the vehicle has a capacity of 15, now the
requirement for 1 and 2 is 10, so we use of 10 out of the 15. We now, take the next
saving, which is 1 and 3 now, one is already here one is at the end of it, so 1 and 3 can
give us an additional saving of 16, now we are adding this 3 inside, so 4 plus 6 10 plus 3
13 is so we put this 13.
We have another saving of 16 here and then we have to add 3 here, the reason is the
saving comes out of 1 and 3, therefore you we should not put here, if we put here, then it
becomes 2 and 3, which would be a 0. So, you put 1 and 3 here, so we have now have a
saving of 32 and we have used 13 out of the 15 capacity, then we look at 2 and 6, so we
can try and add 6 here, but we cannot do it, because the capacity will be exceeded.
Similarly we cannot do 2 4, we cannot do 3 6, we cannot do 1 6, we cannot do 1 3, we

cannot do 2 5, we cannot do 1 4, we cannot do 1 5 and the next thing, we can do is 5 and
6, but we have to put them in a separate vehicle. So, we start with vehicle V 2 or vehicle
2, we put 5 and 6 together, to have a saving of 5 and we used up the capacity of 9, 5 and
6 is three plus 6, which is 9. Then we look at 3 and 5, we cannot do that, because both
are assigned look at 4 and 6, which has 3, so 6 to 4 will come here, so saving is plus 3
and 4 has another 5, so 9 plus 5, 14, 13, so total saving is 40.
And we can also show, how this saving comes, we can calculate this d 3 1 2 d and d 5 6 4
775
d, now d 3 1 2 d will give me, d to 3 is 14, 3 to 1 is 18, 1 to 2 is 22 plus 2 to d is another
18, so 14 plus 18 32 54 plus 18 is 72 and the other 1 d to 5 is 12, 5 to 6 is 26, 6 to 4 is 32
and 4 to d is 16. So, 12 plus 26, 38, 40, 70, 86, so this gives us 158, is the total distance
that, we travel, the saving is 40 and if we realize that, if we are actually gone from the
depot to 1 and back depot to 2 and back 3 and back and so on, the distance we would
have traveled is 20 38 52 68 60 99, 99 into 2 is 198. So 198 minus saving 40 gives us
158.
Now if we do this through the Clarke and Wright this is called the savings based
algorithm or it is called Clarke and Wright algorithm. For this particular example Clarke
and Wright gives us a solution with distance of 158 and a saving if 40. So, one of the
limitations of Clarke and Wright is that, once we start this 1 and 2 and 1 and 3 by fixing
these 2 and by creating the first vehicle, we have got a saving of 32 here.
But, this would prevent us from looking at all of these and the next thing, we can look at
is somewhere here, now the question is if we follow a different order of taking can, we
make it better. Now, we can, so the improvement to Clarke and Wright is called Holmes
and Parker method and as I have mentioned Holmes and parker method is
computationally, more intensive than Clarke and Wright method. It is an extension of
Clarke and Wright method with more computations, it is like a branching algorithm, it
will which will eventually terminate by evaluating all possible paths.
But, we can also terminate it sooner or quicker depending on our requirements. I have
explained the Holmes and Parker solution, in detail in the advance operations research
course under the topic vehicle routeing, where for the same illustration using the Holmes
and parker method, we obtained a saving of 50 instead of saving of 40. So, a saving of 50
would actually give us a solution with distance of 198. Now, this is 99, so 99 into 2 is
198 minus saving of 50 would give us a solution with distance equal to 148, which
actually happens to be the optimum that, we have shown here 74 plus 74, which is 148
Holmes and parker is capable of giving us this solution.
776
So, let me go back here and check, we our savings are 1 to 3 and 3 to 4 and then 5 to 2
and 2 to 6, so 2 to 6 and 2 to 5. So 16 plus 16, 32 plus 8, 40 plus 10, 50 is possible.
Holmes and Parker method is capable of getting this solution. Clarke and Wright is
unable to because Clarke and Wright is like a greedy heuristic, it simply sorts it in this
order and takes them only in this order and does not try or does not make does not take it
to consideration other ways of choosing them.
So, better methods than Clarke and Wright, such as Holmes and Parker exist, but these
methods require more computation, than the Clarke and Wright method. So with this we
come to the end of our discussion on vehicle routeing problems, for this course. We also
end our discussion on transportation decisions or transportation modeling. A very quick
recap, we started with the basic transportation problem, where we transport a unit a
single commodity between a set of source nodes and a set of destination nodes.
And then we extended the basic transportation problem to multistage transportation and
we also extended the transportation to do, what is called fixed charge transportation.
Then we looked at from a given point to another destination, we looked at a model,
where we could have different types of trucks. And then we try to find out the number of
trucks that are needed of each time to do this transportation, so that gave us a single
constraint knapsack problem and we saw, ways to solve that.
Then we looked at point to point transportation and then modeled it more like a bin
777
packing problem and then we looked at milk runs and first looked at the traveling
salesman problem formulation. And then we extended it to the vehicle routeing problem
formulation. So, with this we wind up our discussion on distribution models and a supply
chain and then we look at some aspects of information, in the next lecture.
778
Prof. G. Srinivasan
Lecture - 41
Value of Information
In this lecture, we look at the Value of Information in the context of supply chain. Earlier
when we looked at five aspects of supply chain, we had categorized them into location
and location based decisions. Production, production based decisions, transportation
distribution based decisions, inventory, inventory based decisions. And we also
mentioned at that point, that information, and information based decisions become the
fifth decision making aspect in supply chain management. In fact, the whole idea of
supply chain management is centered around getting good information, and the impact
and role of information, information systems, decision support systems, ERP systems
etcetera, in the supply chain decision making process.
It is not uncommon or it is nor incorrect, if we say that the two important pillars in
supply chain management, are communication and information technology. If we look at
what separates supply chain management, from traditional way of looking at it through
procurement, manufacturing, operations management, and distribution is the speed. The
779
speed is the one that distinguishes supply chain management, from traditional way of
looking at these aspects.
And that speed is essentially given through communication, as well as through

Information Technology. There are additional advantages, other than speed we also have
some other advantages, which we will see as we move along, but the most important
aspect is the speed with which data is obtained. Speed with which, we are able to
communicate either through the net or through telephones, and other media of
communication.
So, the two pillars are essentially communication and information technology.
Information technology also helps us in the following: which is data, forecast, statistical
analysis, optimization and information. One of the advantages of Information
Technology is the fact that we can get data, and we can get very good data and we can
also get recent data, data that is far more important. Simply because, it is the most recent
data, now this data helps us in good forecasts.
And we have already seen that good forecasting systems are essential. Good forecasting
systems ensure that we are able to produce what is demanded. Particularly either if we
produced demand, and demand comes from good data or we produce a little bit to
forecast, and the forecast is made on good and recent data. Once we are able to build
good forecasting models that forecast the demand, we are able to produce what we can
sell, and sell what we can produce which essentially reduces the inventory in the system,
as well as reduces the total cost in this system.
Data also helps in statistical analysis, data also helps us in being able to fit trends, data
also helps in customizing products, and making known to the customer, products, which
suit his or her buying as well as thinking preferences. All these are carried out through
statistical analysis, in a separate upcoming field of analytics or data analytics as it is
called, where we have large scope for Information Technology, data, forecasting,
optimization, and statistical analysis.
Optimization helps in trying to get the best decisions from a given situation; once again
the success of an optimization would depend on the success of the numbers or data that
go into the optimization process or optimization program. So, once again it boils down to
getting good data, and good data can be obtained through good information systems, and
780
Information Technology. Therefore, contributes extensively to the success of a supply
chain.
Now, we look at further detail on what is the value of information, and what we can get
through good data and good information. First one is the reduction of uncertainty,
uncertainty is reduced because, correct information is available at the right time. We
have already seen that uncertainty relates to inventory. Wherever there is uncertainty
there is excess inventory, and inventory is one of the ways and most commonly used way
to reduce uncertainty in manufacturing systems.
More the uncertainty higher the inventory, and if we are able to reduce the uncertainty
through good information systems that are able to provide information and data, then the
inventory holding will come down. We have already seen that inventory can be
measured in terms of units, in terms of money and in terms of time, and all the three are
important. And if we are able to reduce the cost or reduce the time to deliver, we
automatically improve the performance.
So, the first value of information is to reduce uncertainty in the system, by providing
good data and information at the right time, and this helps us in reducing the inventory.
Data also helps us to produce to demand rather than produce to forecast. Now when
production systems were designed to produce to forecast, invariably there was excess
production. Because, there was a forecast error which is the difference between what is
781
the actual demand is, and what the estimated is the forecast is an estimated demand,
while the actual demand would be different from the forecast.
Now, when we start using pull systems which means we produce to demand, which also
means that we are able to sell whatever we produce. When we produce to forecast, if
there is a change in demand then two things would happen, fact that we would not be
able to meet the actual demand. Because, we would have produced something else or
produced in some other quantity, and then the excess that has been produced we would
not be able to sell or we would be able to sell in subsequent periods, which result in
higher cost of inventory or cost of shortage.
So, manufacturing systems move towards pull systems which means the demand will
pull the manufacturing, and the production. So, we end up producing what can be sold
rather than trying to sell what has been produced, manufacturing always has a lead time.
Therefore, if we have to produce what can be sold and if the demand is known at a
certain time period, the product is realized or made after a certain time period, which is
the manufacturing lead time, and we will be able to sell what has been produced.
Therefore, it is necessary to capture the actual demand and not capture the estimate of the
demand. So, information systems help us in getting actual demand and help in
implementing pull systems, and sell what has been produced. Information also helps in
centralized versus decentralized inventory. We have already seen earlier in the course,
that if we centralize the inventory or keeping it in one place, the buffer inventory or the
additional inventory will come down. But, the transportation cost will increase because
the customers will be located far away.
If on the other hand, we locate the inventory close to the customer, the cost of
transportation will reduce, but the cost of holding extra inventory or excess inventory
will now be increased. Now, we could through information system we can try and get the
best of both, now we will be able to physically locate the inventories close to the
customer points. Reduce the excess inventory or the safety inventory which is there, and
in the event of the demand being higher than, the available stock in the warehouses that
are closed to the customer.
We will be able to pull stock from other warehouses, which have excess stock there will
be a slight increase in the transportation cost. But, information systems and Information
782
Technology would help us quickly understand and know, where how much of stock is
available. And therefore, it will be able to get the stock in time and give to the customer,
the fourth aspect is what is called vendor managed inventories. Now traditionally each of
these manufacturing systems, they maintain their own inventory.
They have their own EOQ computations; they find their order quantities and the reorder
levels. And when the stock reach the reorder level, they place orders with the suppliers,
now vendor manage inventory is a very interesting concept, now if you imagine a super
market or a retail store which sells a variety of products. Now, they will traditionally
they will be maintaining their own inventory, so they would know their reorder levels
and the points at which they will have to place the order.
In a vendor managed inventory, the retail store does not mange the inventory or does not
worry about the reorder level. Now, the reorder level is communicated to the vendor.
Now how does the vendor or the supplier know that the reorder level is reached? Now
the information system in the store will now talk to the information system with the
vendor. So, whenever there is a sale happening in the store, the stock is updated in the
store and it gets reflected and the vendor is able to track it.
Now, through Information Technology when the stock reaches the reorder level at the
store, the vendor understands that the stock has reached the reorder level in the store.
And the vendor automatically generates an order or the information system at the vendor,
generates the order resulting in vendor managed inventories. Now, the store does not
worry about the inventory positions, the responsibility of managing it is with the vendor.
And such a thing is enabled through Information Technology.
The next thing that we can see is what are called reverse auctions, now reverse auctions
is like our tendering. So, when we tender for something we will choose the person, who,
actually, reverse action is the opposite of in an action, we try and sell the product and
whoever bids maximum for that gets the product, pays the money and gets the product.
In a reverse action we would, it is like tendering, so we would like to put forth our
requirements, and then people bid for those requirements online through information
systems, and the person who actually has the lowest bid now gets it.
So, reverse auctions also help in procuring or buying material because, we try to pick the
one who has the lowest bid for it, from home we can buy at the lowest cost. Now, this is
783
also enabled through information systems, and technology and through online processes.
Last, but not the least the status of the product or status of the order is now known to the
customer, so when customer places an order with an organization for a product, now the
organization triggers the work order, and manufacturing executes the work order.
Now, the customer through Information Technology, through internet be being able to
login to some systems is now able to understand or view the status of the order. And that
helps in the customer knowing exactly, how much more time is needed for the order to
be finished and for the product to be delivered. Now, all these also put extra
responsibility and owners on the manufacturing system, to gear up and produce and
deliver in time.
So, in essence the information system is able to capture and provide the right kind of
data to the right kind of person so that uncertainties in the system are reduced, and the
organization is able to produce and deliver things in time, and essentially make profits
through this process. Therefore, Information Technology is an important component of
supply chain management.
Now, how do organizations maintain or create information systems and Information

Technology. Traditionally, we have had MRP systems and MRP II systems within
manufacturing. MRP systems were called Materials Requirement Planning systems, and
MRP 2 systems were called Manufacturing Resources Planning systems. Now, both
784
these systems essentially were restricted to the scope of manufacturing, MRP or
materials requirement planning talked about the bill of materials.
And the level of each bought out item or component or manufactured item in the product,
and through the bill of material it was able to generate, how many quantity or in what
quantity each item has to be bought, to meet the demand of the product. And then the
output of the MRP systems, were used to identify the vendors, and sometimes used
through lot sizing algorithms and methods, to try and bunch or combine orders to a
vendor, and increase the order quantity so that certain other benefits can be accrued.
MRP 2 systems also called Manufacturing Resources Planning systems were seen as
systems that were one level above the MRP systems. And also handled additional
information on availability and on little bit of human aspects of manufacturing, so MRP
and MRP 2 systems were essentially created, to act as information systems within the
scope of manufacturing.
Now, ERP systems or Enterprise Resource Planning systems came much later, and
started looking at the entire organization and ERP systems could map the various
functions within the organization or within the organization as a whole. Some of the ERP
modules, one could think of are finance modules, logistics modules, order fulfillment
modules, manufacturing modules, human resources modules, and supplier management
modules. Each of these modules could also have sub modules such as maintenance,
quality and, so on.
Now, the advantage of this large system is it is ability to capture the linkages amongst all
of these. For example, procurement is largely related to the finance, logistics is largely
related to manufacturing because, the finished goods that come out of manufacturing,
will now have to be distributed or sent to the customer. Procurement also is related to
manufacturing, as well as to marketing, so that the demand of the products are known, as
well as the capacities and availabilities from the manufacturing are also known.
Similarly, human resources and manufacturing are equally important, so that the human
resources available or known to manufacturing through these systems. So, if we carefully
look at the organization as a whole, with it is multiple functions there is a strong
relationship among each one of these functions, each of these functions within itself
carries out a whole lot of activities. But, importantly there is a strong relationship among
785
these functions, and a lot of data and information has to move from one system to
another.
Now, all these are done through the enterprise resources planning systems or ERP
systems as they are called. Now, there are two ways of looking at ERP systems, one
approach is to look at ERP as a logical extension of MRP and MRP 2, while the other
approach looks at ERP as a completely different system. Different from the scope of
ERP as well as MRP, then whatever way we look at essentially what we try to do, is to
understand and importantly capture the relationships, among these different functions of
management of an organization and use the right kind of data from each one of these
functions to make decisions on other functions that would impact the performance of the
supply chain.
Now, we spend a little more time on information system and supply chain management,
now we classify information systems into three, one is called within firm information
systems, inter firm information systems and supply chain information systems.
Now, this kind of classification is only indicative it is not exhaustive, and we will see
some aspects of this classification. Now, within firm we could think in terms of MIS,
DSS and ERP. MIS stands for Management Information Systems, DSS stands for
Decision Support Systems, and ERP stands for Enterprise Resources Planning systems.
786
MIS started off a little early. Management Information Systems, started little early and
they had cretin limited scope within the organization.
DSS also called Decision Support Systems also had limited scope, and DSS would
essentially help the organization in decision making, while MIS provided the information
that is necessary for decision making. Now, DSS modules also could incorporate some
optimization kind of modules, so that some decisions could come out of these computer
systems. Just as mentioned just now ERP systems are, systems that link and integrate the
various functions of management, functions within an organization and try and use data
from one function to help make decisions in other functions.
So, these three are good examples of within firm information systems, but supply chain
management is about a network of facilities or a network of organizations, and it is
spreads across organizations. Therefore, we will look at inter firm information systems
and it is role, now some of these aspects are EDI Electronic Data Interchange, the use of
the internet just in time manufacturing systems, and cross docking systems. We will look
at each of these very briefly, electronic data interchange is the ability transfer data from
one organization through the other, and it is used extensively.
Internet is well known, and is a way to see and view information public can view
information, as well as organizations can view information that is available in a different
place. Now, these two help in supply chain management extensively. We did speak about
vendor managed inventory that where the supplier manages the inventory of the
customer. Now, that is possible either when data comes from the customer to the
supplier, through EDI systems or through internet where the supplier is able to view the
data that is present in the customer.
So, that is the way by which these two enabling mechanisms help in vendor managed
inventory, and other aspects of supply chain management. The other two are
applications, where we have inter firm information systems, the role of inter firm
information systems, in effective implementation. Now, JIT stands for just in time
manufacturing, and one of the important aspects of just in time manufacturing is also just
in time purchasing, which means the organizations purchase the items or the products
from their suppliers.
787
And purchase in quantities that are just in time, which means the quantity that is
purchased is enough to meet the requirements of a very short period, predetermined
period which could be a day, which could be a shift, which could be a week which is
predetermined by the system. Now, if we assume that we are going to buy daily, so for
an organization to make the product typically today manufacturing organizations talk
about, make daily, sell daily we also look at buy daily.
So, let us assume that the organization buys daily, converts them into product and sells
them daily. Now, in order to buy daily, the supplier should be in a position to deliver or
to provide the items that the organization can buy daily, now in order to do that the
information system of the supplier, should talk to the information system of the
organization or the customer or someone who is going to buy from the supplier.
Therefore, we see the role of interactions between information systems of different firms.
In a cross docking environment which we explained earlier, there is a space in which

warehouses send trucks to come to that place. And these trucks are there, and the cross
docking area has additional trucks, and the inventory is shifted from one set of trucks to
another set of trucks. Inventory does not stay physically in the cross talking area, and
then the stock is sent to the customers, once again the timing, the quantities that have to
be shifted all of these can now be exchanged through information systems.
And there is enormous scope for inter firm information systems to help in effective
supply chain implementation. Now, in addition we have what are called supply chain
information systems, where we talk about decision making for the supply chain for the
entire supply chain, which means a chain of manufacturing organizations, chain of
organizations that act the supplier customer etcetera.
So, effective decisions forecasting for the product that would be useful to the entire
supply chain, and effective decision making in planning that would be useful to the
entire supply chain. So, we will need software that would do that, information systems to
support, create, maintain, and execute such software also have to be created. And then
we talk about supply chain information systems, which take which uses ideas from inter
firm as well as within firm.
788
Now, we spend a little bit of time on coordination. Coordination generally means people
work together, they coordinate and work together towards a common task. Now,
coordination is an important concept in supply chain, supply chain is coordinated if the
decisions made by the members of the supply chain, maximizes the total profit of the
supply chain or minimizes the total cost in the supply chain. A good way of explaining
coordination is like this.
Now, suppose there is a organization which orders for a certain item to a supplier. Now
let us assume that this supplier is ordering it and getting it from another person. So, now,
we will call these as A, B and C. Now A is the organization which is going order to B, B
is the supplier, and B orders from C. Now, A can independently workout the economic
order quantity, and the total cost associated with the order, now B will also have to order
to C and therefore, B will separately workout the economic order quantity.
Because B’s ordering cost may be higher and B’s inventory holding cost may be
different, now if both A and B work on their economic order quantities, the total cost
will be the total cost for A and the total cost for B put together. It is always possible for
A and B to come together such that they order a common quantity, so that the total cost
for a plus B put together can come down. Now, such a thing is called coordination, now
when we do this coordination two things can happen.
789
Now, through this coordination the cost for A can come down as well as cost for B can
come down and the total cost can come down. There can be a situation where, the cost
for one of them may increase and the cost for other may come down and therefore, that
organization which is A or B, for whom the cost after the coordination increases, while
the total cost comes down. Now, that organization will be hesitant to implement it
because, its cost would increase.
In such cases organizations should come together to absorb the increased cost of one
person in the interest of both the organizations coming together. So, that the total cost
reduces. Now such a thing can be done through discounts to share the difference in
individual cost, so that the total cost comes down. Now, lack of coordination results in
what is called the bullwhip effect, now we have already seen what is bullwhip effect
very briefly.
Bullwhip effect is the result of increased inventory in the supply chain, as we move from
right to left, which means as we move from customer to supplier there is the average
inventory in the system would increase if there is no coordination. The average inventory
in the system increases simply because, of the uncertainty that is communicated in the
ordering. So, every entity will actually order a little more than what is required, and
therefore, the variation in the order increases.
And therefore, the total inventory in the system increases, as we move from right to left
or as we move from customer to manufacturer. Now, that is called bullwhip effect, and
the effect of bullwhip is that there is total increased total inventory in the system that
results in increased total cost of the system, and this cost increases due to lack of
coordination. Now, there is an important thing in supply chain called the beer game,
which is usually played to help people understand the effect or the role of bullwhip
effect, and the increased cost of the supply chain if there is no coordination.
We would spend a little bit of time on the beer game in this lecture after a few slides,
now we also have to understand obstacles to coordination. Now, while we know that if
organizations come together, and make decisions together that can bring down the total
cost of the supply chain, then why do organizations are why are they not doing it? Now,
today there is a conscious effort for organizations to come together.
790
Traditionally they were not being they were unable to do it for multiple reasons, one of
which was that information was not available. And information was not shared; there
was hesitation in sharing information among organizations. Now, today organizations do
not hesitate to share a lot of data, and the means of sharing the data which is through IT
has also advanced considerably. So, there is a lot of data sharing and there is a lot of joint
decision making, which result in increased coordination.
But, in spite of that there can be some obstacles to coordination and organizations have
to understand that, and make sure that they do not fall into the trap of these obstacles.
Now, four obstacles to coordination are listed here, one is incentive obstacles. Many
times through discounts or through other incentives when we make purchasing decisions
we end up buying large quantity, much more than what is required simply because, of
certain incentives and discounts that we would get.
Now, these things increase the cash that is being spent or the money that is being locked
up in the inventory and therefore, it increases the total cost in the system. So, when we
stock a lot of items in the system, we are going to make fewer orders per year and
therefore, the orders cannot be synchronized or coordinated. Sometimes, organizations
also are joined together, place orders together, either organizations themselves place
orders together or within an organization orders for different items are placed together so
that there can be a saving of transportation or truck cost. In this course we have seen
791
such models when we studied inventory. Now, if we have discounts and incentives for
either certain items or if certain organizations wish to avail and benefit from these
discounts, then joint ordering becomes extremely difficult. In a similar manner many
obstacles happen through these incentives. Second is lack of information sharing. We
have already seen that information sharing and good information helps in reducing
uncertainty and therefore, reduces inventory.
Now, organizations have to come together shun their inhibitions, shun their fear and start
sharing data. So, that they work in a coordinated manner and the total cost comes down.
Large orders and large lead time; large lead times are also obstacles because, whenever
there is a lead time there is an inventory, there is an uncertainty and there is an inventory.
We have also seen the effect of lead time, in when we studied inventory through some
examples.
We not only studied the effect of large lead time, we also studied the effect of variation
in the lead time. If the lead time shows variation or if there is a standard deviation of lead
time, then we found out through our computations that the total inventory in the system
is very high. Now, organizations have to be coordinated to reduce lead times, reduce
variation in the lead times work together, share information and data bring down the lead
times, so that the total cost comes down.
The last one is called rationing and the effect of rationing, now rationing happens when
total demand for a product is more than the supply. So, if we assume a single supplier
and let us say four customers, who demand this product and if the total demand is higher
than the total supply, then certainly all of them will not get their demand. Then rationing
policies come into the picture. Now, the common thing that one would think of is to
proportionately reduce the quantity that is transported.
Many times that does not happen largely because, of the importance of the customer,
because of the quantities ordered and, so many other factors. Rationing is a very separate
area of importance from both theory as well as practice, one of the problems that happen
when somebody knows that there is going to be rationing and particularly when we know
that the proportion we are going to get will depend on the demand that is being ordered,
organizations will deliberately hike up the demand.
792
So, that a certain proportion of the increased demand would come, which would mean
the organization would get more items from the supplier. Now, such a thing also reduces
or prevents coordination from happening and organizations suffer through this, so these
are some of the aspects of coordination and obstacles to coordination.
Now, it is also important that coordination is carried out through Information

Technology, and there is a strong role for Information Technology in coordination. Now,
we have listed three points which are - sharing point of sale data, the important aspect is
when we share point of sale data with the suppliers in particular, the vendor can manage
the inventory. The vendor knows exactly when the re order point is going to be reached,
and the vendor can also plan, so that the reorder point can effectively be reduced by
proper coordination.
Collaborative forecasting and planning: Organizations come together, and instead of

individually doing forecasting and planning, come together and do the forecasting
planning and other systems in a collaborative manner. And the third one is to look at
vendor managed inventories, we did speak about vendor managed inventories where the
organization does not worry about the ordering, through coordination, through
information sharing, through sharing point of sale data. The supplier is able to keep track
of the inventory position of the various products of the customer.
793
And as and when the reorder level reached for these products, the customer will trigger
or will create an order and will deliver the item to the customer. Now, this is called
vendor managed inventories, there are several examples in the literature for vendor
managed inventories, between very famous super markets, and very famous
manufacturing organizations.
Now, we also talk about E business which is called Electronic business. Now E business
is the execution of business transactions through the internet. Now, this is a phenomenon
that has been in existence for about 20 years or more. Now, one could say comparatively
recent but, extremely important and flourishing. Each one of us would have used E
business in many ways today, using E business to order books, today using E business to
buy items, today using E business and E portals to book tickets. These are all extremely
popular and famous examples of using electronic business.
Some of the aspects are providing information across the supply chain, prices and
contracts amongst prices of the products, telling the customers about the prices, so that
the buy price changes contracts. Customer can place orders; customers can track orders.
Now, the delivery aspect as well as the payment aspect can be ensured through E
business strategies. Payments are also famous today we are able to pay most of our bills
electronically, so there is a very large role for E business and E commerce in the supply
chain.
794
Some other things that E business systems can do are direct sales to customers. Examples
of computer sale which happen directly to the customer. Another important aspect is a 24
hour access, where these systems are available right through and not restricted by
timings. E businesses provides personalized services by keeping track of the buying
patterns of the customers, so when you place an order for a certain book immediately the
website will show other similar books or books by the same author and, so on.
E business systems also help us in flexible pricing, depending on the stock positions of
the books, depending on other interest and, so on. Particularly airline pricing - to be able
to track the prices of air tickets, all these are possible through E business systems, lots of
examples in PC industry, book industry and, so on. But, E business also has an important
aspect in the supply chain management.
795
Now, we spend a few minutes on this beer game. The beer game is a game that was
created somewhere in the 1960's by a university. Where they made people understand the
effect of a supply chain by modeling the supply chain through a classroom game. Now,
over the years this game which started off as the classroom game, also has had it is
electronic version, and several versions of this game is also available. The beer game can
be played in many ways.
It started off with the beer game because, the first time the game was played the product
that was used for illustrative purposes, was beer. And therefore, the name beer game has
stayed for well over 50 years. This game essentially captures or through a game it
explains the various aspects in a supply chain, and also helps in understanding the basic
principles of the supply chain, including the bullwhip effect. Now, what we are going to
see are results from beer game that have been played in classroom on two consecutive
batches of students. Now, in a class containing about 60 students one could divide them,
into 7 or 8 groups you see 8 groups which are here.
796
Now, each group will have about 8 students and each group will represent an
organization. And these students will now be divided - some of them will be customers,
some of them will be warehouses or distribution points, and some of them will be
manufacturing points. Now, if there are 9 students in a group or 9 people in a group we
would divide them into 3 groups of 3 each, and each of these the 3 will act as a customer
the other 3 will act as a distribution and the third will act as manufacturer.
Sometimes, we could even have 4 levels in the supply chain or 4 entities in the supply
chain resulting in 3 levels, and it can be manufacturing, warehouse, distribution and
customer. So, what we do is one of them acts as the customer or as the retailer and then
the end customer comes in. So the demands are generated for the classroom. There is a
single product that is assumed, and demands are generated. So, this slide would tell us
the demands for each of these groups.
So, the demand 100 plus minus 20 would mean that we generate a random number
uniformly distributed between 80 and 120. So, you would see 100 plus minus 20 and you
will see 100 plus minus 100, so you realize that the demand shows more variance for
groups 5, 6, 7 and 8 and lesser variance from 1, 2, 3 and 4. Now, what we also do in this
case is that somebody represents the market, and they come and give the demand to the
retailer. The retailer has some initial stock, and the retailer will meet that demand from
that stock, and place an order with the warehouse or the distributor.
797
Now, when the lead time is zero, the warehouse or distributor immediately gives the
order, and then the warehouse places an order with manufacturing and, so on. And
notionally the product is made from the manufacturer given to the warehouse, and then
to the retailer, and then to the customer. So, each of these persons updates their
beginning inventory demand, what is supplied as well as their ending inventory. Now,
for each of them there is an inventory carrying cost, which is taken as one unit or one
rupee and there is a shortage cost which is taken as 20 units or 20 rupees.
Now, we also give different lead times sometimes the lead time zero means that when
the retailer comes to the warehouse, the warehouse immediately tries to give the item
that is available in the warehouse. If the lead time is 1, it means the warehouse will take
one more week to deliver. The third dimension that we bring into this game is for groups
1 to 6, we do not have information sharing, for group 7 and 8 we have information
sharing. So, for group 7 and 8 when the customer demand is told, the customer demand
is told to everybody.
And they can also discuss it, and then they can make their own decisions with respect to
how much they wish to buy, so that is how information sharing in modeled. So, if you
see the beer game setting that I have described, there is a variation in demand which you
can see the mean demand or the average demand is the same. There is a variation in lead
time, and then structurally there is a variation where some groups play without
information sharing, which means the customer demand is not known to other stages in
the supply chain, but it is known only to the retailer. And only the retailer demand is
known to the warehouse, and the warehouse demand is known to the manufacturer.
798
We now show some of the results of the beer game that is played in class, and the results
of the 8 groups are shown here. The column 1 talks about the 8 groups, the retailers, the
distributors, and the manufacturer or the producer, column 2 simply talks about number
of periods for which they played, it is observed that one group is a little slow and played
only for about 7 periods. Column is 3 and 4 are important from the point of view of
understanding the supply chain.
799
Now, we realize that groups 1 to 6 did not have information sharing while group 7 and 8
had information sharing.
And now when we come back and see that the average inventory in the system
particularly when we look at group number 7 which is here, 207, 155, 172 we look at
group number 7, we observe that the average inventory in the system is far less than the
one that is in group 4 or that is in group 5 and, so on. Now, we realize that information
sharing helps us in reducing or understanding that the average inventory in the system
comes down.
Now, we should also keep in mind that since this is a game that is played in classroom,
one cannot make lot of accurate comparisons. But, one can only try and understand some
meaningful inferences that come out which I will try to do right now. So, I am trying to
explain that the average inventory in the system comes down when we have information
sharing. Where even between group 7 and group 8, there is no difference in the problem
setting, but then we realized that there are differences, in the numbers that they have,
largely because, of the way they have played the game.
800
Now, if we look at other groups for example, if we look at group 1 and 2, demand
patterns are the same, but lead times are different.
Now, one could understand that when the demand patterns are the same, but the lead
times are different, the way the inventories behave is very, very different. Now, you see
that group 2, the inventories are not comparable, the average inventories are not
comparable, the average inventory is very high here, little low here, and much lower here
801
which shows some kind of bullwhip effect that is happening. Here in group 1 the average
inventories are reasonably comparable.
For example, if you look at group 3, group 3 has the higher variation then groups 1 and
2, and now we see that the average inventory in group 4 is much higher than that of
groups 1 and 2. So, more variation in the system increases the average inventory, now
lead time would the variation in demand is not very high, but there is a lead time and the
lead time distributes the inventory unevenly in the supply chain, which can be seen in
group number 2’s performance.
So, at the end we could back and summarize the results by saying that when we
information sharing, the inventory in the system comes down compared to groups or
systems, which had no information sharing. Now, when the product variation is higher
we will now see that the total inventory in the system is also higher, and when the
variation is higher we would also see that the inventory is higher compared to systems
which have lesser variation.
Now, when we have lead times, lead times essentially try and redistribute the inventory
among the players in the supply chain, and a combination or variation in lead time as
well as variation in demand shows the bullwhip effect. Now, let me show another sample
of beer game that is played in a class, where there are 10 groups and then we also look at
the similar pattern, where 8 groups have no information sharing and 2 groups have
information sharing.
802
And the lead times are 0, 0, 0 to 1, 1, 0, 1, 2 and, so on, we have deliberately used 2
week lead time. So, that if you place an order you will get that order only after two
weeks and, so on and this time we also increase the demand, as well as we change the
variation in the demand.
Now, the results of this game is also shown and there are 10 groups, so they are shown
over three slides. And then one could understand here, if we look at the inventories here
between, we observe that as we move along groups 1 and 2 there is a large variation, so
803
if we see now that the cost also increases here between 1 and 2. The average inventory in
the system here remains increases a little bit. So, with variation we could see that the
average inventory increases a little bit. Now, between groups 1 and 3, there is a clear
difference in lead time. So when we look at groups 1 and 3 the maximum inventory
increases, which is an effect of the difference in lead times. So, when there is increased
lead time the maximum inventory in the system will also increase.
Now, we will look at groups 9 and 10 we realize that the average inventories 196 which
is much lower, than other average inventories in other places without information sharing
like 214, 242, 296 and, so on. So, that we are able to show here the total cost in the
system also behaves in a certain manner, this is of the order of 5000, 4900, 4100
compared to 7599, 8612 and 6000. And one can see that group 8 did not have
information sharing, group 9 have information sharing and therefore, is able to
coordinate data. So, what are all things that we have described in theory and practice, we
are able to capture it through a classroom game which is called the beer game.
804
So, the summary of results of the beer game would be - information sharing helps in
better performance in terms of inventories. We observe that the less variance of demand
less is the ending inventory and less is the total cost. So, variance of demand increases
the inventory as well as total cost. Information sharing helps and high variance of
demand increases the inventory. Bullwhip effect is seen, where inventory positions of
two teams are known, inventory fluctuates, reaches a high and then comes down. The
fact that the inventory reaches the high end keeps coming down is the essential
component of the bullwhip effect.
805
So, this is the this is how we explain or use the beer game in class as well as in other
places to make people understand the effect of supply chain, and the importance of
coordination and information sharing. Now, with this we come to the end of this session
on role of information as well as information sharing. We also come to the end of this
course on operations and supply chain management.
Now, there are 41 lectures and in somewhere during the middle of the course, around the
end of the 34th lecture, I had given a summary of what we had done, up to the operations
management portion. Or looking at it from a supply chain point of view I had given a
summary of the important supply chain ideas such as, production decisions and inventory
decisions. Now, from lectures 35 to 41 we looked at transportation and distribution, as
well as we looked at some aspects of information system.
We also looked at some location decisions, when we did location and layout as well as
location and layout in the context of supply chain, and also when we started looking at
supply chain in much more detail. So, the earlier lectures in the course concentrate a lot
on operations management, things that happen within the firm in terms of managing
production, and in terms of managing inventories.
The later lectures talk about the role of distribution, the role of location as well as the
role of information. So, with this I would conclude this course on operations and supply
chain management. I hope I have covered the various aspects of operations management,
and some aspects of supply chain management in this course and I hope that this course
is useful to all it is viewers.
Think you.
806
THIS BOOK IS NOT FOR SALE
NOR COMMERCIAL USE
(044) 2257 5905/08 nptel.ac.in swayam.gov.in

Operations and Supply Management

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Operations and Supply Management

Uploaded by

Copyright:

Available Formats

INDEX

S. No Topic Page No.

Forecasting – Winter’s model, causal models, Goodness of

7 Aggregate Planning, Dynamic Programming, backordering 121

10 Inventory – EOQ model graphs, with backordering 170

29 Job shop scheduling – Shifting bottleneck heuristic 513

34 Quantitative models for layout, Summary 614

37 Transportation and Distribution Models 692

39 Bin Packing and Travelling Salesman Problems 740

41 Value of Information 780

(Refer Slide Time: 00:59)

In this course, wewill be addressing several topics,this lecture will be on introduction to

Inventory models, introduction to supply chain management, value of information and

And in inventory decisions, we will look at deterministic and probabilistic inventory

So, let us begin the discussion on operations management, as I mentioned operations

(Refer Slide Time: 05:02)

(Refer Slide Time: 11:01)

(Refer Slide Time: 17:25)

Now, as mentioned earlier, after manufacturing looked at volume and scale,

(Refer Slide Time: 18:43)

(Refer Slide Time: 19:33)

And then, methodologies based on process improvement then, methodologies based on

(Refer Slide Time: 20:16)

(Refer Slide Time: 21:37)

(Refer Slide Time: 24:11)

(Refer Slide Time: 24:36)

(Refer Slide Time: 27:28)

(Refer Slide Time: 34:10)

Total quality management is a management initiative to empower all people working to

Largely ISO9000 certifications, where manufacturing organizations get themselves

Now, this concept of synchronous manufacturing or constraint management came from

(Refer Slide Time: 44:02)

(Refer Slide Time: 44:50)

(Refer Slide Time: 46:44)

(Refer Slide Time: 48:10)

Operations management looks at issues like purchasing management, proper production

So, in this course as mentioned earlier, we would do some aspects of demand

(Refer Slide Time: 00:17)

Today we begin our discussion on Forecasting; forecasting essentially deals with

(Refer Slide Time: 02:39)

For example, if we give weights w 1, w 2, w 3, w 4, w 5 and w 6 then the weighted

So, a k period moving average can be thought of as giving weights of 0 a 4 period

(Refer Slide Time: 29:39)

(Refer Slide Time: 37:55)

Now, if we make an assumption that n is large here we have n equal to 6 we have 6

So, sum of an infinite geometric progression is a by 1 minus r, so that will be alpha by 1

We want progressively decreasing weights as we move to the past or progressively

(Refer Slide Time: 47:37)

(Refer Slide Time: 50:51)

(Refer Slide Time: 00:16)

We continue our discussion on Forecasting Models. In the last lecture, we developed

(Refer Slide Time: 03:33)

(Refer Slide Time: 10:27)

So, we want to minimize Y minus a minus b t square here, I have represented it as Y t

So, you need a Y t term, so Y into t so 26 into 1 is 26 28 into 2 is 56 29 into 3 is 87 31

So, if we are convinced that exponential smoothing is a very good model to do

(Refer Slide Time: 29:22)

(Refer Slide Time: 32:32)

(Refer Slide Time: 32:38)

(Refer Slide Time: 32:51)

(Refer Slide Time: 34:02)

(Refer Slide Time: 37:47)

(Refer Slide Time: 45:56)

(Refer Slide Time: 00:16)