Professional Documents
Culture Documents
Abstract
This paper deals with a two-echelon inventory system for a non-repairable item where the system
consists of one warehouse and m identical retailers and uses continuous-review (R,Q) ordering
policy. To find an effective stocking policy for this system, a mathematical model with the objective
of minimizing the total annual inventory investment subject to constraints on the average annual order
frequency, expected number of backorders, and budget is formulated. The mathematical model of the
algorithm is proposed to solve it efficiently. A numerical example is provided at the end to illustrate
1
1. Introduction
efficiency across competitive business, to improve customer service, and to reduce inventory costs at
different locations of a supply network. One of the most important aspects of inventory management
that has vital role in supply chain operations is the distribution of goods in multi-echelon inventory
systems.
The multi-echelon inventory systems are becoming more prevalent and have begun to draw
more attentions from both practitioners and academicians in many industries and communication
networks. For instance, the military's usage of multi-echelon knowledge achievements is quite well
known and air force used a model for a multi-item, multi-echelon, multi-indenture inventory system
as a method to compute recoverable spare stock level for the F-15 weapon system (Muckstadt 1973).
Multi-echelon spare parts inventory systems have been discussed broadly in the literature.
METRIC, one of the earliest models in this topic is a multi-echelon technique for recoverable item
presented by Sherbrook (1968). The objective of this mathematical based-depot supply system model
is minimizing expected backorders subject to budget constraints with compound Poisson demand. He
also approximated outstanding orders at the retailers. According to this study, the appropriate policy
Deuermeyer and Schwarz (1981) developed a model based on an exact, single facility (R,Q)
model of Hadley and Whitin (1963) to analyze service level in a system consisting of one warehouse
and a number of identical retailers. Svoronos and Zipkin (1988) proposed several refinements of this
model in a multi-echelon inventory system. They approximated each facility as a single location and
calculated the mean and variance of the warehouse and retailer lead-time demand. Graves (1985)
presented a multi-echelon inventory model with the failures generated by the compound Poisson
process and deterministic shipment time from the repair depot to each site for a repairable item with
one-for-one replenishment. He also presented an approximation for the steady-state distribution of net
inventory level with ample serves at the repair depot and determined the average and the variance of
2
Axsäter (1990) presented a simple solution procedure for a two-echelon inventory system
with one-for-one replenishment, constant lead-time, and independent Poisson demand at retailers. He
used an inventory cost function and focused directly on evaluating the average costs. Furthermore,
Axsäter and Zhang (1996) considered a two-echelon inventory system with one warehouse and a
number of identical retailers with constant transportation time and compound Poisson demand at
retailers. They provided a simple recursive procedure for the evaluation of holding and shortage costs
at different control policies. DeBodt and Graves (1985) presented an approximate model to minimize
the expected costs including a fixed ordering cost, an echelon inventory holding cost for each level,
and a back order cost for end item in the multi-echelon inventory system controlled by continuous
review policy. Svoronos and Zipkin (1991) described a simple technique to approximate steady state
behavior in a multi-echelon with one-for-one replenishment and stochastic transit times of the parts
between locations.
Hope et al. (1997) formulated a constrained optimization model in a single location controlled
by (R,Q) policy and developed three heuristic to solve it. The objective of this model was minimizing
total inventory investment subject to constraints on order frequency and customer service. Axsäter
(2000) presented a method for exact evaluation of control policies that provide the complete
probability distributions of the retailer inventory levels in a two-echelon inventory system consisting
of one central warehouse and N retailers. This system is controlled by different continuous review
(R,Q) policies with constant transportation times and independent compound Poisson demand
processes at retailers.
Seo et al. (2001) developed an optimal reorder policy to utilize centralized stock information
for a two-echelon inventory system consisting of one warehouse and multiple retailers controlled by
continuous review batch ordering policy. Axsäter (2001) evaluated a technique where a high-demand
demand. Marklund (2002) investigated a two level distribution system consisting of one warehouse
and a number of non-identical retailers. In order to control the replenishment process at the
continuous review (R,Q) control policies. He also presented a method for exact evaluation of the
3
expected inventory holding and backorder costs for the system. Axsäter (2003) considered a two-
echelon distribution inventory system consists of a central warehouse and a number of retailers
controlled by continuous review installation stock (R,Q) policies. He presented a simple method that
uses normal approximations for the retailer demand and the demand at the warehouse in order to
Kiesmüller et al. (2004) developed analytical approximations based on asymptotic result from
continuous review (s,nQ) installation stock policies under compound renewal demand. Caglar et al.
(2004) investigated a two-echelon, multi-item spare parts inventory system and presented a
mathematical model with the objective of minimizing the system-wide inventory cost subject to
constraint on response time at each field depot. They also used a heuristic algorithm to solve it
efficiently.
Axsäter (2005) determined warehouse backorder cost and provided a newly decentralized
way with optimizing sum of expected holding and backorder costs to warehouse and retailers
regarding their reorder point in a two-echelon distribution system with installation stock (R,Q)
inventory control policy. Seifbarghi and Jokar (2006) developed an approximate cost function to find
optimal reorder points of given batch sizes in a two-echelon inventory system consisting of a
warehouse and many identical retailers with lost sales and independent Poisson demands controlled
inventory system consisting of one warehouse, several retailers, and two items with lost sale and
demand substitution. They also presented a heuristic algorithm to find cost effective based stock
policies. Al-Rifai and Rossetti (2007) investigated a two-echelon inventory system consisting of a
central warehouse and a number of identical retailers controlled by (R,Q) inventory policy for non-
repairable items. The objective function of their model is minimizing the total annual inventory
investment subject to constraint on average annual order frequency and expected number of
backorder. They solved the model by decomposing the system by echelon and location. They also
derived expressions for the inventory policy parameters and developed an iterative heuristic
4
optimization algorithm. Haji et al. (2008) considered a two-echelon inventory system consisting of
one warehouse and a number of non-identical retailers with Poisson demand in which warehouse is
facing a uniform and deterministic demand ordered by each retailer and introduced a new ordering
holding facilities organized into two levels. More specifically, the problem that is considered consists
of a two-echelon non-repairable item inventory system of one warehouse and m identical retailers.
We consider the reorder policy for a continuous-review distribution system, utilizing the total
inventory investment in both echelons subject to constraints on the average annual order frequency
and expected number of backorder. The dominant model of this system in practical applications is
based on the assumptions of unlimited warehouse and retailers' budget. These assumptions can lead to
a serious underestimating of the spare parts requirements. To deal with this dilemma, warehouse and
retailers' budget constraints are also considered. Considering a new objective function and several
constraints makes the model more applicable to real-world inventory control systems. In addition, the
solution method that is based on a parameter-tuned meta-heuristic algorithm seems more compact and
The remainder of the paper is organized as follows. In section 2, the problem is defined
explanation of the methodology proposed to solve the model is discussed in section 4. In section 5,
parameter adjustment and numerical example are given. Finally, conclusions are provided and future
2. Problem definition
and m identical retailers in which all installations use continuous review (R,Q) policy to replenish
At the beginning of a period, an outside supplier with unlimited capacity delivers the bulk of
inventory directly to the warehouse. The warehouse allocates the stock to the retailers. The demand
5
at a retailer level is either satisfied or backordered. Backorders, both at the warehouse and retailer
levels, are filled according to the first in first out (FIFO) policy. At both echelon, (R,Q) ordering
policy continuously monitors the inventory position (minus backorders plus on hand and on order
inventory) for each item. It means that as soon as the stock level declines to the reorder point R , an
order of batch size Q is placed. Note that the real lead-time at the retailer level consists of two
components: retard and delay times. Retard time is the time between placement of an order by a
retailer and the release of a batch by the warehouse (Svoronos and Zipkin, 1988). Delay time is due
to the ordering and transportation times. As a result, the effective lead-time at the retailer level is the
sum of the waiting time due to a lack of stock in the warehouse (retard) and the transportation and
ordering times (delay) .In this paper, we assume that the retard time at the retailer level is zero. In
other words, the real lead-time at the retailer level is equal to retailer's delay. Figure (1) provides a
Central
Supplier
warehouse
Retailers
In order to define the problem precisely, a set of critical assumptions are required as follows.
a. The demand process at each retailer is a Poisson process with an annual rate Dri .
d. Average annual ordering frequency at the warehouse and retailers are limited.
6
e. Expected number of backorder at the warehouse and retailers are limited.
3. Problem formulation
The problem at hand is formulated with the objective of minimizing the total annual inventory
investment subject to constraint on average annual order frequency, expected number of backorder,
The following notations, parameters, and variables are used for mathematical formulation of
the problem:
m Number of retailers
Dw i Demand rate of the ith item at the warehouse level (in units of Qri )
Dri Demand rate of the ith item at the rth retailer (unit per year), r 1, 2,..., m
I w i ( Rw i , Qwi ) Expected on-hand inventory at the warehouse for item i (in units of Qri )
B wi ( R ri , Q ri ) Expected number of backorder at the warehouse for item i (in units of Qri )
7
N wi Average order frequency at the warehouse for item i (orders per year)
Now, the inventory control system under consideration can be formulated as the following
8
Minimize total inventory investment
Subject to:
Minimize:
k k
Min TC m C i I ri ( R r i ,Q ri ) C i Q ri I wi ( Rw i ,Qw i ) (7)
i 1 i 1
Subject to:
k
1 Dwi
k
Qi 1
Nw (8)
wi
k
1 D ri
k
Qi 1
Nr (9)
ri
B
i 1
wi ( Rw i ,Qwi ) B w (10)
B
i 1
ri ( R r i ,Q r i ) B r (11)
C Q
i 1
i wi Qr i X w (12)
k
m C i Q r i X r (13)
i 1
9
R w i Q w i i 1,..., k (14)
R r i Q ri i 1,..., k (15)
Q ri 1 i 1,..., k (16)
Q wi 1 i 1,..., k (17)
Equation (7) gives the objective function of the model and refers to the total inventory
investment of both echelons. Constraints (8) and (9) assure that average order frequencies at the
warehouse and retailers are not greater than the target order frequencies at the warehouse and retailer
levels, respectively. Constraints (10) and (11) assure that the total expected number of backorders at
the warehouse and the retailers are not greater than the corresponding target numbers. Constraint (12)
assures that total warehouse purchase is not greater than the available warehouse budget for all items.
Constraint (13) assures that total retailers' purchases is not greater than the available retailers' budget
for all items. Constraints (14) and (15) assure that when a replenishment order is received, the
outstanding backorder are satisfied. Constraints (16) and (17) show that warehouse and retailer batch
size are greater than zero. Finally, constraint (18) represents the ranges of the decision variables.
The expected on-hand inventory of the warehouse is calculated using the following equation
Qw i 1
I w i B w i ( R w i , Qw i ) R w i E[ H w i ] (19)
2
Similar equation can be used to calculate the expected on-hand inventory of retailer r. Furthermore,
Svoronos and Zipkin (1988) developed the mean and the variance of the lead-time demand at the
warehouse as:
mDri Lwi
E[ H wi ] (20)
Qri
10
Where
p 1 cos(2 p / Q ri ) (22)
p sin ( 2 p / Q ri ) (23)
Dr i
Nri (24)
Q ri
mDr i
Dw i mN ri (25)
Qri
Svoronos and Zipkin (1988) developed the mean and the variance of the lead-time demand at retailer
r as:
The warehouse backorder of the ith item is obtained via Equation (27). Similar equation can be used
to calculate the retailers' backorder quantities of the ith item (Hopp and Spearman 2001).
1
Bwi (Rw i ,Qw i ) [ (Rw i ) (Rw i Qwi )] (27)
Qw i
Where
2 (x ) 2 (x ) (x ) (x )
(x ) {( 1)[1 F ( )] f ( )} (28)
2 2
As hinted before the demand process of item i at retailer r is Poisson with an annual rate
of Dri . However, the demand at the warehouse level does not follow a Poisson process. This is due to
the fact that when the stock-out is backordered in the retailers and the demand process at each retailer
is Poisson, the arrival process of the orders at the warehouse is a superposition of the order processes
from the retailers, specifically, a superposition of independent renewal processes, each with Erlang
inter-renewal time with Qri stages and rate per state Dri (see Svoronos and Zipkin 1988).
In the next section, a search-based algorithm is developed the inventory model of the problem
at hand.
11
4. A solution procedure
The model that was formulated and introduced in the previous section belongs to nonlinear
integer-programming (NIP) problems and involves a complex shape of search space. These
characteristics lead us to use a meta-heuristic algorithm to solve such a hard problem reliably.
Genetic algorithm (GA) is one of the most efficient intelligent randomized search procedures
for solving optimization problem based on the principle of survival of the fittest in biological
evaluation and genetics, introduced first by Holland (1975). Although GA was successfully applied to
some inventory control problems, the application of GA to multi-echelon inventory systems is still
rare.
The underlying GA's idea is to generate an initial set of random solution called population.
Each individual in the population is called a chromosome that is represented as a candidate solution to
a problem and consists of a number of genes .The chromosomes must go through a successive set of
solution called generation and the fitness function is used to evaluate all individuals. Crossover and
mutation operators are used to create new chromosomes called offspring. A new population is created
with respect to fitness value of chromosomes. The population of solution become better and better
Generally, the proposed GA woks according to the scheme that is described and given in
2. Set generation = 0
4. Evaluate all individuals of the current generation via the fitness function
12
-Make random mutation
6. Output
4.2 Chromosome
One of the most important factors for successful implementation of GA is designing a more
suitable chromosomal structure. In this article, the chromosomal solution consists of a matrix with
four rows and k columns. The first and the second rows show replenishment batch sizes and reorder
points of all items at the warehouse level, respectively. The third and the fourth rows show
replenishment batch sizes and reorder points of all items at the retailers, in turns. Furthermore, each
column shows the replenishment batch sizes and the reorder points at the warehouse and the
replenishment batch sizes and the reorder points of the retailers for each item. These lead to a 4×k
matrix as the chromosomal structure of all items depicted in Figure (2). Note that due to the constraint
13
4.3. Initial population
random generation covers almost all of the search space, a high-quality solution, obtained from
another heuristic technique as an initial population, might help a GA to find better solution.
4.4 Evaluation
applications, the objective function of the optimization model at hand is considered a fitness function.
However, as explained before, the inventory model of this research has 10 constraints on the average
annual order frequencies, the expected number of backorders, and the budget. These characteristics
make the probability of a generated chromosome being feasible very low. In order to promote this
chance, a penalty function is defined to be a positive and known sum of squared violation of each
constraint. As a result, the penalty and the fitness function in which violation per constraint is denoted
by Ei is defined as follows:
10
i 1
Ei2 The chromosome feasible region
Penalty function=
(29)
The initial requirement described in step 1 of Algorithm (1) consists of two parts:
D r i , Lw i , L ri , and C i . These values are determined by the user and depend on the system ability
14
2. GA data: consists of the probability of performing crossover called crossover rate denoted by pc and
the probability of performing mutation called mutation rate denoted by p m . The number of
chromosomes that is called population size and is denoted by Pop_size plays the main role in the run
At this stage, parents are selected using one of the most popular selection methods called
roulette-wheel. In this method, the parents are chosen based on the probability distribution of their
fitness value and copied into the mating pool. Thus, the chance of selecting the best individuals
becomes higher.
4.7. Crossover
After parent selections, several pairs of chromosomes are selected randomly from mating pool
by predetermined crossover rate ( pc ) and are mixed to produce offspring. In a crossover operation,
some of the genes in the first selected parent are replaced with the corresponding genes in the other
parent. In the proposed GA, at first a binary chromosome is created for the parents under
consideration. Then, the matrix components (genes) of the two selected parents that correspond to
zero values of the binary chromosome are replaced with each other. Those genes that correspond to
value one do not change. As an example, the crossover operation of six items is performed as given in
Figure (3).
4.8. Mutation
genes by a predetermined mutation rate of p m . For the mutation operation of this research, first a
random chromosome whose components are between 0 and 1 is created and applied to the selected
parent. Then, the parent genes that correspond to values less than p m are mutated within the
15
boundaries of their corresponding variable. The other genes of the offspring are exactly the same as its
15 43 5 83 9 30 7 355 121 18 21 62
21 382 61 12 7 5 41 82 33 121 41 74
62 31 191 23 82 9 3 90 48 44 371 11
93 28 341 44 10 14 1 10 51 37 45 23
Parent1 Parent2
1 0 0 1 1 0
0 0 1 0 1 0
0 0 0 1 1 0
1 0 0 0 1 1
Binary chromosome
7 43 5 18 21 30 15 355 121 83 9 62
21 382 33 12 41 5 41 82 61 121 7 74
62 31 191 44 371 9 3 90 48 23 82 11
1 28 341 44 45 23 93 10 51 37 10 14
Offspring 1 Offspring 2
0 . 36 0 . 38 0 . 61 0 . 64 0 .5 0 . 84
0 . 11 0 . 77 0 . 93 0 . 31 0 . 49 0 . 24
0 . 83 0 . 29 0 . 28 0 . 01 0 . 38 0 . 71
0 . 51 0 . 65 0 . 41 1 0 . 59 0 . 27
15 43 5 83 9 30 15 43 5 83 9 30
21 382 61 12 7 5 33 382 61 12 7 168
62 31 191 23 82 9 62 31 191 1 82 9
93 28 341 44 10 14 93 28 341 44 10 14
Parent Offspring
16
4.9 New population
During this phase, the fitness function value of all members consisting parents and offspring
are evaluated. After evaluation, the chromosomes with better fitness scores are selected to create new
population. In order to achieve better solution, the fittest chromosomes must be preserved at the end
of this phase. Note that the number of selected chromosomes must be equal to Pop_size.
4.10. Termination
The population of solutions improves from generation to generation based on the theory of
survival of the fittest. GA obtains a near optimal solution when the fittest member of the population
satisfies a termination condition. In this research, the algorithm stops when the fitness function values
In the next section, a numerical example is given to illustrate the application of the proposed
5. Numerical illustration
Consider an inventory control system of this research that consists of six items with the
item Dr i L wi Lr i Ci
17
In order to determine the best value for GA parameters (the crossover & mutation rates and
the population size) that result in a better solution, a fine-tuning procedure is followed. In this
procedure the proposed GA runs 150 times with different values of pc , p m , and Pop_size. Then the
fitness function is recorded in each run. Table (2) shows the results of the first 25 runs.
To investigate the relationship between the fitness value and the parameters, a regression
analysis using the SPSS software is then employed. This relationship is obtained as
Next, the LINGO software is employed to the following optimization problem in which the best GA
Subject to:
0 pc 1 (31)
0 pm 1 (32)
As a result, the combination of the crossover rate of 0.52, the mutation rate of 0.18, and the population
Employing the proposed GA with the obtained values of the parameters and after one hundred
generations, the algorithm converges to the following solution with a total inventory investment of
$9016.42:
11 9 48 6 14 35
5 2 40 2 18 7
6 1 4 1 2 1
6 0 1 21 6 10
Furthermore, the graph of the convergence path of the GA is presented in Figure ( 5).
18
Table (2): Numerical results of 25 runs
19
Fitness value
Number of Generations
In this research, a new model for a two-echelon inventory system was developed to find the
continuous review inventory control parameters that minimize the total annual inventory investment
subject to constraints on the average annual order frequency, the expected number of backorder, and
budgets. This model is closer and hence more applicable to real-world inventory problems than the
other existing ones. We also explored the use of a GA to solve this problem efficiently. Future
research may extend the model by considering different assumptions or different solving methods .In
addition, many variations in GA operators, initialization method, fitness definitions, and replacement
strategies are obviously possible and might lead to more efficient solutions.
References
Al-Rifai MH, Rossetti MD, 2007. An efficient heuristic optimization algorithm for a two echelon (R,
Axsäter S, 1990. Simple solution procedures for a class of two echelon inventory problems.
20
Axsäter S, Zhang W, 1996. Recursive evaluation of order-up-to-S policies for two-echelon inventory
systems with compound Poisson demand. Naval Research Logistics 43: 15 1-1 57.
Axsäter S, 2000. Exact analysis of continuous review (R,Q) policies in two-echelon inventory systems
Axsäter S, 2005. A simple decisions rule for decentralized two-echelon inventory control.
Caglar D, Chung-Lun L, Simchi-Levi D, 2004. Two-echelon spare parts inventory system subject to a
Debodt MA, Graves SC, 1985. Continuous review policies for a multi-echelon inventory problem
Deuermeyer BL, Schwarz LB, 1981. A model for the analysis of system service level in warehouse-
retailer distribution Systems: The identical retailer case. TIMS Studies in the Management
Graves SC, 1985. A multi-echelon inventory model for a repairable item with one-for-one
Hadley G, Whitin TM, 1963. Analysis of inventory Systems. Prentice-Hall, Inc., Englewood Cliffs,
NJ.
Haji R, Neghab MP, Baboli A, 2008. Introducing a new ordering policy in a two-echelon inventory
system with Poisson demand. International Journal of Production Economics 117: 212–218.
Holland JH, 1975. Adaption in natural and artificial systems. University of Michigan Press, Ann
Hopp WJ, Spearman ML, Zhang RQ, 1997. Easily implementable inventory control policies.
Hopp WJ, Spearman ML, 2001. Factory Physics, second ed., McGraw-Hill, New York.
21
Jokar MA, Zangeneh S, 2006. Developing a model for a two-echelon two-item inventory system with
lost sale and demand substitution. Management of Innovation and Technology 2: 926-930.
Kiesmüller GP, de Kok TG, Smits SR, van Laarhoven PJM, 2004 .Evaluation of divergent N-echelon
Marklund J, 2002. Centralized inventory control in a two-level distribution system with Poisson
Muckstadt JA, 1973. A model for a multi-item, multi-echelon, multi-indenture inventory system.
Seifbarghi M, Jokar MR, 2006. Cost evaluation of a two-echelon inventory system with lost sales and
Seo Y, Jung S, Hahm J, 2001.Optimal reorder decision utilizing centralized stock information in a
Sherbrooke CC, 1968. METRIC: A multi-echelon technique for recoverable item control. Operations
22