Advanced Inventory Management

Models and Algorithms
Hans Blanc
CentER Graduate Program in Business
Tilburg University
The Netherlands
Course 230261: Advanced Quantitative Logistics
January 7, 2011
Preface
This syllabus contains the lecture notes for a part of the course Advanced Quantitative Logistics. This course
was given as an elective in the OR track of the Graduate Program in Business at Tilburg University, The
Netherlands, from 2003 until 2008. The subject of these lecture notes is Advanced Inventory Management.
This concerns models and algorithms for inventory management with interaction between items or between
stocking points. The prerequisites for this topic consist of models and algorithms for single item, single
stocking point inventory systems with constant, time-varying and stochastic demand as discussed in most
introductory textbooks on operations research and management science such as Winston [80, Ch. 16, 17, Sect.
20.7] and as summarized in the first chapter of these lecture notes.
ii
Contents
Preface ii
1 Introduction to Inventory Management 1
1.1 Definitions and concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Review of single-item policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 The basic EOQ model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 The production lot size model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 The EOQ model with quantity discounts . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.4 The EOQ model with planned stockouts . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.5 Time-varying demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.6 Stochastic demand, continuous review . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.7 Stochastic demand, periodic review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3 Interactions in Inventory Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Coordinated Replenishment 20
2.1 Constant demand, indirect grouping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.1 Exact algorithmic solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1.2 Heuristic solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.1.3 Powers-of-two policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.1.4 Production lot-size problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.1.5 All-units discounts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Constant demand, direct grouping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.1 Exact solution by dynamic programming . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.2 Heuristic solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.3 Direct grouping with discounts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2.4 Indirect versus direct grouping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3 Time-varying demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.1 Exact solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.2 Heuristic algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.4 Stochastic demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.4.1 Can-order policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4.2 Periodic review policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3 Multi-echelon Inventory Systems 55
3.1 Multi-echelon structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.1.1 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.1.2 Distribution centers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2 Constant demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2.1 Linear chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2.2 Diverging chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.3 Time-varying demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.4 Stochastic demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4 Other Inventory Systems with Interactions 77
4.1 Inventory capacity constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
iii
iv CONTENTS
4.2 Repairable items . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
A Review of Probability Theory 85
A.1 Random variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
A.1.1 Moments and expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
A.1.2 Collections of random variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
A.2 Probability distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
A.2.1 Normal distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
A.2.2 Exponential distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
A.2.3 Poisson distributions and Poisson processes . . . . . . . . . . . . . . . . . . . . . . . . 88
A.2.4 Gamma and Erlang distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
A.2.5 Mixtures of Erlang distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
B Optimization methods 91
B.1 Stationary points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
B.2 Golden section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
B.3 Bisection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Bibliography 93
Index 97
List of symbols 100
Chapter 1
Introduction to Inventory
Management
Inventory management deals with ordering and stock keeping of goods for sale, production or distribution.
Inventories are idle goods waiting for use or sale. Inventories are kept in many environments, for instance, in
the mining-industry of minerals, in factories of raw materials, parts, work in progress and finished products,
and in warehouses, depots and wholesale dealers of goods for distribution, and at shops and by retailers of
goods for sale. The main reasons why inventories are held are that it is uneconomical to produce, to handle or
to transport units one by one and that consumers often do not accept a delay in the delivery of goods or only
want to buy goods that are on display or available in a shop, supermarket or department store. Inventory
theory aims to develop models and algorithms as an aid to inventory management.
In Section 1.1 inventory policies will be classified, inventory costs and concepts will be described, and the
main sources of uncertainty will be indicated. Section 1.2 contains a review of models and results for single
item inventory management. Section 1.3 provides an overview of interactions that may occur in the control
of inventories of multiple items and at multiple stocking points.
1.1 Definitions and concepts
In this section, general concepts in inventory management will be described and explained. The main sources
of uncertainty where inventory management has to deal with are
Demand: the demand for items may fluctuate from day to day (due to stochastic behavior at retailers, due
to variations in the production plan in a manufacturing environment), from month to month (due to a
seasonal pattern) and during the lifetime of a product (an upward trend in the beginning, a downward
trend towards the end);
Lead time: the total time that elapses between the reorder instant and the instant when goods are ready
for use or sale. It consists of the handling time at the supplier (the time required for order picking,
packing, and loading), the shipping time from the supplier to the stocking point and the handling time
at the stocking point (the time required for unloading, unpacking, and placing on the shelf). When
the goods still have to be produced after the reorder instant, it also includes the production time and
possibly a set-up time for the production run.
In the practical situation of uncertain (stochastic) demand and nonnegligible lead times stockout occurrences
cannot be completely avoided. For customers arriving when an item is out of stock, two cases are often
distinguished:
• any demand is backordered and the backlog is filled as soon as a replenishment is delivered; customers
are willing to wait if it is difficult to obtain the item elsewhere;
• any demand is lost; customers go elsewhere to buy the item or give up the intention of buying the item.
For some items, part of the demand may be backlogged and part may be lost. The distinction between the
two extreme cases becomes less important when stockouts occur more rarely.
1
2 INTRODUCTION TO INVENTORY MANAGEMENT
A prerequisite for applying inventory policies is a good forecasting method for future demand. A statistical
analysis is required of historical demand and lead time data. Further, an ABC classification is often carried
out. In most companies, a relatively small percentage (5%–20%) of all items account for a relatively large
percentage (55%–65%) of all sales. These are called type A items. Most effort of forecasting and inventory
control should be concentrated on this type of items. Further, there is a middle class of items consisting of
20%–30% of all items that account for 20%–40% of all sales. These type B items require less attention and
their inventory can often be controlled by standard procedures. Finally, there are type C items consisting of
50%–75% of all items and accounting for only 5%–25% of all sales which are to be controlled by simple and
safe procedures. In manufacturing environments, the classification of a raw material item may also be based
on how critical the item is for the continuation of work, beside on its value.
The three most important questions to be answered by an inventory policy are
• When to review stocks?
A distinction is made between
– periodic review policies where stocks are reviewed at fixed time intervals, the review periods;
– continuous review policies where stocks are reviewed after each transaction.
• When to order?
A distinction is made between
– periodic review policies where orders can only be placed at the periodic review instants;
– continuous review policies which use reorder points in inventory positions.
• What to order?
A distinction is made between
– policies with a fixed order quantity;
– policies with a fixed order-up-to level.
Next, we will discuss costs that may play a role when ordering and storing goods:
Ordering cost: the fixed cost of placing an order; this cost includes the cost of paperwork and accounting
associated with an order which is independent of the size of an order; if the item is made internally
rather than ordered from an external supplier, this cost is often called set-up cost and includes the
cost of labor, material and idle time associated with setting up and shutting down a machine for a
production run; if goods are ordered from another location within the same company, this cost may
include internal shipping cost.
Purchasing cost: the variable cost associated with purchasing a single unit of a good; this cost often
includes variable labor cost, variable overhead cost and raw material cost associated with producing of
handling a single unit; if goods are ordered from an external supplier, it also includes shipping cost;
the external supplier may want to stimulate larger orders to save on shipping cost by offering quantity
discounts; these cost only depend on the inventory policy in case of quantity discounts or lost sales.
Holding cost: the variable cost of holding a single unit of a good on stock during a unit time period; this
cost often includes variable opportunity cost incurred by investing capital in inventory, storage cost,
insurance cost, and cost due to possible theft, obsolescence, breakage and spoilage; the opportunity
cost is often assumed to be a certain percentage, the so called carrying charge, of the purchasing cost;
the carrying charge is strongly related to the interest rate.
Handling cost: the cost associated to the handling of goods in a warehouse; as far as this cost is proportional
to the number of items handled it does not influence the minimization of the total inventory cost if
all demand is satisfied; as far as this cost is proportional to the number of orders handled it can be
incorporated in the ordering cost; this cost is important in the design and control of warehouses.
Shipping cost: the cost associated to the transport of goods from one stocking point to another; in case of
an external supplier, the shipping cost is often included in the purchasing cost.
HANS BLANC, TILBURG UNIVERSITY 3
Stockout cost: in case of backlog of demand it is the extra cost associated to the administration and later
delivery of goods; in case of lost sales it is the opportunity cost of lost profit on unsatisfied demand; in
all cases, it may include a penalty cost for loss of future goodwill; it may also include extra cost for rush
orders or overtime work; in many cases, stockout costs are difficult to assess and are therefore replaced
by service level constraints (see below).
Management cost: the cost incurred by keeping track of inventory levels and by computing order quantities;
this cost is usually not included in inventory models but should form an incentive to choose for inventory
policies that are simple to implement.
In the stochastic demand models the following two service level constraints will be considered:
Cycle service constraint: the probability of no stockout in a reorder cycle must be at least a prescribed
probability α; the latter probability is called the cycle service level (this constraint is also called P
1
-
criterion);
Fill rate constraint: the fraction of the demand that is satisfied directly from stock must be at least a
prescribed fraction β; the latter fraction is called the target fill rate (this constraint is also called
P
2
-criterion).
The following inventory concepts will be used in the various models:
Inventory on hand: the number of units actually present at the stocking point; it is also called the physical
stock; this quantity plays a role in determining holding costs;
Net inventory (net stock): the inventory on hand minus the amount of backlog; this quantity can take
positive and negative values;
Inventory position: the net stock plus the number of units on order but not yet delivered; this quantity is
required for determining a reorder instant;
Safety stock: the average inventory position just before a delivery instant; this quantity is used as a pro-
tection against uncertainty in demand and against other irregularities like breakage and pilferage; it is
related to the service level constraint or the cost of stockouts or losses.
Some general considerations in inventory management are:
• choose a model in agreement with the availability and the reliability of data (according to the general
principle “garbage in → garbage out”);
• the robustness of a model is important: the resulting replenishment policy should not depend too
strongly on the assumptions (like the shape of the demand distribution): this requires sensitivity
analysis;
• in practice, quantities are usually rounded off (reorder cycles are chosen in whole days or weeks; order
quantities are chosen in packing units): a complicated, time-consuming algorithm for exact optimization
then has little use;
• within companies, conflicts in interests or goals may exist between the purchasing department (which
strives for quantity discounts and delivery of goods at the beginning of a season) and the logistics
department (which has to cope with large quantities at the same time, and which may be saddled with
superfluous stocks at the end of a season); such conflicts may be due to the remuneration system of a
company;
• uncertainty in demand and in lead times, and desired service levels lead to safety stocks; on the other
hand, risk of disappointing demand, technical obsolescence, going out of fashion or decay compel to
reservedness toward large stocks.
1.2 Review of single-item policies
This section contains a review of inventory policies developed for individual items at a single stocking point.
Most of this material can be found in introductory textbooks on operations research or management science,
such as Winston [80], and in more specialized textbooks on inventory management, such as Silver et al. [64]
and Axs¨ater [7]. For the case of stochastic demand, see also Tijms [65].
4 INTRODUCTION TO INVENTORY MANAGEMENT
¸
`

I(t)
t →
s+Q
Q
¯
I
s
0 T 2T T −L 2T −L
s s s
c c

Figure 1.1: Saw-tooth pattern of inventory in deterministic fluid model with lead time L < T.
1.2.1 The basic EOQ model
The basic inventory model assumes that demand occurs at a constant, known rate of D units per unit of
time. The cost of placing an order is a dollars per order. The holding cost is h dollars per unit per unit of
time. All demand must be satisfied from stock. As a consequence, the average purchasing cost per unit of
time is a policy-independent constant amount of D times the purchasing price of a unit of the good, and can
be omitted in the optimization procedure. In the absence of uncertainty in the demand and the lead time, it
is optimal to have a zero inventory level at delivery instants. Based on a fluid approximation in which goods
flow out of stock at a constant rate D, and the stock is instantaneously replenished with order quantities
Q at delivery instants, cf. Figure 1.1, the following results can be derived. The relevant inventory cost for
a given fixed order quantity Q consists of ordering costs and holding costs. Since the number of orders per
unit of time is D/Q and the average inventory level is
¯
I =
1
2
Q, it follows that the average inventory cost per
unit of time, C(Q), is
C(Q) = a
D
Q
+h
¯
I = a
D
Q
+
1
2
hQ. (1.1)
The optimal order quantity is found by differentiation of the cost function, cf. Appendix B.1:
Q

=
_
2aD
h
. (1.2)
The quantity Q

is usually referred to as the economic order quantity (EOQ). The reorder cycle (time) T,
the time between two successive replenishments, for a given order quantity Q follows as
T = Q/D. (1.3)
The average inventory cost per unit of time can also be formulated as function of the reorder cycle T:
C(T) = a
1
T
+
1
2
hDT. (1.4)
The optimal reorder cycle is
T

=
Q

D
=
_
2a
hD
. (1.5)
Example 1.1 Figure 1.2 shows the average inventory cost per unit of time both as function of the order
quantity (C(Q)) and as function of the reorder cycle (C(T)) for the case a = $5, D = 500 and h = $2; here,
Q

= 50 and T

= 0.1, while the minimum average inventory cost per unit of time is C

= 100.
In general, the minimum average inventory cost per unit of time follows by substitution of the optimal order
quantity into the cost function:
C

= aD
_
h
2aD
+
1
2
h
_
2aD
h
=
1
2

2ahD +
1
2

2ahD =

2ahD. (1.6)
HANS BLANC, TILBURG UNIVERSITY 5
0 20 40 60 80 100 120 140 160 180 200
0
50
100
150
200
250
300
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0
50
100
150
200
250
300
Q → T →
Figure 1.2: Average inventory cost as function of order quantity (left) and of reorder cycle (right).
¸
`

I(t)
t →
s+Q
Q
¯
I
s
0 T 2T T −L 2T −L
s s s
c c
Figure 1.3: Step pattern of inventory in deterministic unit-by-unit model with lead time L < T.
Observe that for the optimal order quantity (and the optimal reorder cycle) the average ordering cost is equal
to the average holding cost. The optimal number of orders per unit of time is
M

=
D
Q

=
1
T

=
_
hD
2a
. (1.7)
In general, Q

, T

, and M

are not whole numbers. Since they are interrelated, it is usually not possible
to round them all to integer values. A way to find the optimal integer order quantity is by determining the
value Q for which the cost C(Q) is equal to the cost C(Q+ 1), cf. (1.1). The equation
a
D
Q
+
1
2
hQ = a
D
Q+ 1
+
1
2
h(Q+ 1),
reduces to
Q(Q+ 1) = 2aD/h.
Similarly, the equation C(Q) = C(Q − 1) implies Q(Q − 1) = 2aD/h. This means that the integer order
quantity
˜
Q is optimal if
˜
Q(
˜
Q−1) < 2aD/h <
˜
Q(
˜
Q+ 1) or
_
˜
Q(
˜
Q−1) < Q

<
_
˜
Q(
˜
Q+ 1). (1.8)
In this way, the interval 0 < Q

<

2 is rounded to
˜
Q = 1, the interval

2 < Q

<

6 to
˜
Q = 2, the interval

6 < Q

<

12 to
˜
Q = 3, etcetera. This way of rounding corrects for the skewness of the cost function near
6 INTRODUCTION TO INVENTORY MANAGEMENT
¸
`

I(t)
t →
s+Q
s
Q
¯
I
s−Q
0 T 2T T −L 2T −L
s s s
c c
c c
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
--
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
--
-
--
-
-
-
-
-
Figure 1.4: Saw-tooth pattern of inventory in deterministic fluid model with lead time T < L < 2T.
its minimum. It should be noted that the smooth inventory pattern in Figure 1.1 is in fact an approximation
for the step pattern for discrete units as shown in Figure 1.3. The fluid approximation is better according as
larger quantities are involved (fast movers). For items that are sold or used in small quantities (slow movers)
the discrete nature of the demand and the inventory pattern should be taken into consideration.
A useful property of the EOQ formula is that the cost function (1.1) is rather flat near its minimum in many
cases, cf. Figure 1.2. This implies that the optimal order quantity Q

is rather insensitive with respect to
the cost parameters.
The above derivations do not involve a possible lead time L between the instant at which an order is placed
and the instant at which the order is delivered. If the lead time L is constant, the optimal values Q

, T

and
M

are not affected by this lead time. The lead time L only determines the instant when an order has to be
placed, namely, L units of time before the order is needed. The demand during the lead time is D L. Since
the order has to be delivered at the instant when the inventory on hand reaches the level 0 in deterministic
systems, the inventory position at reorder instants should be equal to D L. The inventory position at the
reorder instant is called the reorder point, and will be denoted by s. In deterministic systems, it is seen that
s = D L. (1.9)
Figure 1.1 includes the reorder instants and the reorder point for the case that the lead time L is smaller
than the reorder cycle T. In this case, the inventory position (dotted line) only differs from the inventory
on hand on the time intervals (mT − L, mT), m = 1, 2, . . . . Figure 1.4 shows the reorder instants and the
reorder point for the case T < L < 2T. In this case, there is always at least one order in transit so that the
inventory position is higher than the inventory on hand at every instant. In these and later figures, ◦ marks
the inventory level at a reorder instant and • marks the inventory level just before a delivery instant.
1.2.2 The production lot size model
Consider the following variant of the basic EOQ model in which units are not purchased from an external
supplier but internally manufactured. Let a denote the set-up cost of starting a production run, h the holding
cost per unit per unit of time, D the demand rate and p the production rate, that is, the number of units
that can be produced per unit of time. It is assumed that p > D. In a production environment the way in
which items or substances become available to satisfy demand may affect the optimal lot size. If the produced
items become available as a batch at the end of a production run, the optimal lot size is determined by the
optimal order quantity (1.2). As lead time one has to take L = u +Q

/p with u a possible set-up time for a
production run. The reorder point then becomes s = DL = Du +DQ

/p, cf. (1.9).
However, if the produced units become available for satisfying demand at a rate p during the production run,
the start of the production can be delayed until the instant at which the inventory reaches the level s = Du.
In this case, the maximum inventory is not Q but the inventory level at the end of the production run. To
produce Q units takes a time τ = Q/p, and during this time the inventory increases at a net rate of p − D
up to, cf. Figure 1.5,
I
max
= (p −D)τ = (p −D)Q/p. (1.10)
Hence, the average inventory level is
¯
I =
1
2
(p −D)Q/p, and the cost function becomes
C(Q) = a
D
Q
+
1
2
hQ(1 −D/p). (1.11)
HANS BLANC, TILBURG UNIVERSITY 7
¸
`

I(t)
t →
I
max
¯
I
s
s+Q
0 T 2T −u τ T +τ T −u 2T +τ 2T −u
c c c
s s s

`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
` `
` `
``
Figure 1.5: Triangular pattern of inventory on hand in deterministic production lot-size model.
The optimal production lot size becomes
Q

=
¸
2aD
h(1 −D/p)
, (1.12)
the optimal production cycle length is
T

=
Q

D
=
¸
2a
h(1 −D/p)D
, (1.13)
and the minimum average cost per unit of time is
C

=
_
2ah(1 −D/p)D. (1.14)
In fact, all results of the model with batch delivery can be translated to the model with gradual delivery
by replacing h by h(1 − D/p). Figure 1.5 shows the triangular pattern of the inventory on hand and the
saw-tooth pattern of the inventory position for a system with gradual delivery.
1.2.3 The EOQ model with quantity discounts
Some suppliers offer quantity discounts on purchasing prices to their customers to prevent the handling,
packing and shipment of small quantities or to stimulate larger sales. In this situation, it is assumed that
the holding cost is proportional to the purchasing price v, that is, h = rv, with r a carrying charge. Further,
suppose that there exists one or more () price break points V
j
, V
1
≥ V
2
≥ ≥ V

, and discount factors
d
j
, d
1
≥ d
2
≥ ≥ d

, such that a discount factor d
j
is awarded if the value of an order without discount
exceeds the price break point V
j
. Let v
0
denote the basic price per unit. Then, the price per unit is
v(Q) = v
0
, if v
0
Q < V

,
v(Q) = v
0
(1 −d
j
), if V
j
≤ v
0
Q < V
j−1
, j = 2, . . . , ,
v(Q) = v
0
(1 −d
1
), if v
0
Q ≥ V
1
.
(1.15)
The cost function becomes, including the variable purchasing cost,
C(Q) = a
D
Q
+
1
2
rv(Q)Q+v(Q)D. (1.16)
Algorithm 1.1 [EOQ with quantity discount]
Step 1: Start with the highest discount factor: m = 1.
Step 2: Determine the optimal order quantity Q

m
according to (1.2) with h = rv
0
(1 − d
m
). If Q

m
is
feasible (Q

m
≥ V
m
/v
0
), compute C(Q

m
) and goto Step 3; otherwise, compute C(V
m
/v
0
), decrease m
by 1 and repeat Step 2.
8 INTRODUCTION TO INVENTORY MANAGEMENT
20 40 60 80 100 120 140 160
4050
4060
4070
4080
4090
4100
4110
4120
4130
4140
4150
20 40 60 80 100 120 140 160
4050
4060
4070
4080
4090
4100
4110
4120
4130
4140
4150
Q → Q →
Figure 1.6: Average inventory cost as function of order quantity with and without discount.
¸
`

I(t)
t →
s+Q
Q−B
−B
s
0 T 2T τ T −L T +τ 2T −L
s s s
c c
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
``
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`` `
`
`
`
`
`
`
`
`
Figure 1.7: Saw-tooth pattern of net inventory in deterministic fluid model with planned stockouts.
Step 3: The optimal order quantity is that quantity from the set ¦Q

m
, V
m+1
/v
0
, . . . , V
1
/v
0
¦ with the
lowest cost.
Here, take d
+1
.
= 0 and note that Q

+1
is feasible in any case.
Example 1.2 Consider an item with the following data: a = $5, D = 500, v
0
= $8 and r = $0.25 so that
h = $2. Suppose that the supplier offers a discount price of v
1
= $7.92 for order quantities of 150 units or
more. This means in the above notations: = 1, d
1
= 0.01 and V
1
= 1200. Figure 1.6 (left) shows the
average inventory cost per unit of time with (lower graph) and without (upper graph) discount. According
to Algorithm 1.1 first the optimal order quantity with the discount price is computed: Q

1
= 50.25. This
order quantity is not feasible. Hence, the average cost at V
1
/v
0
is computed: C(150) = $4,125.17. Then,
the optimal order quantity without discount is computed: Q

2
= 50.00, with corresponding average cost
C(50) = $4,100.00. Hence, it is optimal to order 50 units and ignore the discount offer. However, if the
discount price would be offered for order quantities of 100 units or more, the quantity Q

1
= 50.25 is still
not feasible, but since C(100) = $4,084.00 it is optimal to order 100 units and accept the discount offer, cf.
Figure 1.6 (right).
HANS BLANC, TILBURG UNIVERSITY 9
1.2.4 The EOQ model with planned stockouts
In this section we consider the variant of the basic EOQ model in which stockouts are allowed. It is assumed
that all stockouts are backlogged instantaneously at a delivery instant. The accumulated amount of backlog
at a delivery instant will be denoted by B, and will be treated as a decision variable beside the order quantity
Q. It will be assumed that B ≤ Q (it is readily verified that this is optimal). The parameters a, h and
D have the same meaning as in Section 1.2.1. Further, b will represent the stockout cost per unit short
per unit of time. Figure 1.7 shows the saw-tooth pattern of the net inventory as function of time. When
the net inventory is positive it represents the inventory on hand. The inventory on hand is zero whenever
stockout occurs; then, the net inventory is negative and its absolute value represents the accumulated amount
of backlog. In this model, the reorder point s can be positive or negative. It is given by, cf. (1.9),
s = D L −B. (1.17)
Next, the average cost per unit of time will be determined for this model. The number of orders per unit
of time is D/Q and the reorder cycle is T = Q/D, as in the basic EOQ model. During a time τ the net
inventory is positive. This time is determined by Dτ = Q−B. Hence, a fraction τ/T of each cycle the net
inventory is positive, and the average level is
1
2
(Q − B). A fraction τ/T of each cycle the net inventory is
negative, and the average backlog is
1
2
B. Hence, the average ordering, holding and stockout cost per unit of
time is
C(Q, B) = a
D
Q
+
1
2
h(Q−B)
τ
T
+
1
2
bB
T −τ
T
.
Since τ/T = (Q−B)/Q, the average cost per unit of time can be expressed in the decision variables Q and
B:
C(Q, B) = a
D
Q
+
1
2
h
(Q−B)
2
Q
+
1
2
b
B
2
Q
. (1.18)
The partial derivative of this cost function with respect to B vanishes if h(Q − B) = bB. This means that
the optimal order quantity for a fixed total backlog per cycle is
ˆ
Q(B) =
h +b
h
B, (1.19)
with corresponding average cost per unit of time
ˆ
C(B) =
ah
h +b
D
B
+
1
2
bB. (1.20)
From this function it readily follows that the optimal backlog per cycle, the optimal order quantity and the
minimum average cost per unit of time are:
B

=
¸
2ahD
b(h +b)
, Q

=
_
2aD(h +b)
hb
, C

=
¸
2ahbD
(h +b)
. (1.21)
The actual fill rate, cf. Section 1.1, for the optimal policy is
Ψ = 1 −
B

Q

= 1 −
h
h +b
=
b
h +b
. (1.22)
Alternatively, this implies that a target fill rate β corresponds to a stockout cost of
b =
βh
1 −β
. (1.23)
In the limit as b → ∞, that is, if the stockout cost becomes very high, it follows from the above that B

→ 0,
Q


_
2aD/h and C



2ahD, which represents the basic EOQ policy, cf. Section 1.2.1.
1.2.5 Time-varying demand
In this section we consider situations in which demand fluctuates over time, but is known in advance for a
number of future periods (days, weeks, months). Time-varying demand may be due to trends or seasonal pat-
terns, and may also arise as a consequence of material requirement planning in a manufacturing environment.
10 INTRODUCTION TO INVENTORY MANAGEMENT
In the context of time-varying demand a discrete time model is considered with a finite planning horizon H
(a number of periods). Here, a single item is treated in isolation. The demand in period t is denoted as D
t
,
t = 1, . . . , H. The cost of placing an order is a dollar per order. Orders are delivered at the beginning of a
period. The demand for that period is set aside, and holding costs of h dollar per unit per period are only
charged over the excess inventory. All demand must be fulfilled and no back orders are allowed. The decision
variables are Q
t
, the order quantity for the beginning of period t (t = 1, . . . , H). The inventory level at the
end of period t is denoted by I
t
(t = 0, . . . , H). It is assumed that the initial inventory level is zero: I
0
= 0.
The inventory levels can be described by the recursive equations
I
t
= I
t−1
+Q
t
−D
t
, t = 1, . . . , H. (1.24)
The objective is to minimize the total ordering and holding costs over H periods:
C(Q
1
, . . . , Q
H
) =
H

t=1
[aδ(Q
t
) +hI
t
] , (1.25)
here, δ(.) is a dummy function: δ(0) = 0, δ(q) = 1 if q > 0. The order quantities are restricted by the
requirements that I
t
≥ 0, t = 1, . . . , H. The minimization of the cost function (1.25) is facilitated by the
observation that the optimal order quantities are zero or equal to the demand of a whole number of consecutive
periods:
Q
t
= 0 or Q
t
= D
t
+ +D
t+
, = 0, 1, . . . , H −t, for t = 1, . . . , H. (1.26)
This property follows by noting that to order only a fraction of the demand of the last period t + is not
optimal because it forces an order to be placed for the beginning of period t +. Shifting the fraction of the
demand D
t+
from the order quantity Q
t
to the order quantity Q
t+
does not change the ordering costs but
diminishes the holding costs. A further restriction on the possible optimal values of Q
t
is obtained by the
observation that it is not optimal to include the demand D
t+
in the order of period t if hD
t+
> a, because
the costs can then be reduced by placing an additional order in period t +.
Wagner & Whitin [77] have developed a dynamic programming solution for this model. Define, for t =
1, . . . , H,
• f(t): the minimum cost incurred during the periods t, . . . , H, given that the inventory level is zero at
the end of period t −1;
• C(t, ): the cost of ordering at the beginning of period t a quantity that is sufficient to fulfill the demand
of the periods t, . . . , t +:
C(t, ) = a +h

u=1
uD
t+u
, = 0, 1, . . . , H −t, t = 1, . . . , H. (1.27)
Algorithm 1.2 [Wagner-Whitin]
Step 1: Set f(H + 1) = 0 and start with t = H.
Step 2: Apply the backward recursion
f(t) = min
0≤≤H−t
¦C(t, ) +f(t + + 1)¦. (1.28)
Step 3: If t > 1 decrease t by 1 and repeat from Step 2; otherwise, stop.
The minimum total ordering and holding costs is f(1); the optimal order quantities have to be recovered
from the intermediate results of the recursion: the minimizing in (1.28) indicates the optimal number of
additional periods for which to order if t becomes a reorder instant.
The computation time of the Wagner-Whitin algorithm strongly increases with the number of periods H. In
some cases, the optimal order quantity for the first period is strongly influenced by variations in the perhaps
still uncertain demand in the last period. To overcome these difficulties, Silver & Meal [62] have developed
HANS BLANC, TILBURG UNIVERSITY 11
a heuristic for this model. They propose a forward recursion. Let τ be a reorder instant at which an order
is placed for the periods τ, . . . , τ + −1. Then, the average cost over these periods is
¯
C(τ, ) =
1

_
a +h
−1

u=1
uD
τ+u
_
, = 1, . . . , H + 1 −τ, τ = 1, . . . , H. (1.29)
The heuristic proposes to place the next order just after a local minimum of the average cost since the previous
reorder instant has been reached. The order quantity for the first period only depends on the demand in the
periods up to the period in which such a local minimum occurs.
Algorithm 1.3 [Silver-Meal heuristic]
Step 1: Start with τ = 1.
Step 2: Compute the average costs
¯
C(τ, ) for = 1, 2, . . . , until
¯
C(τ,

+1) >
¯
C(τ,

) for some

≥ 1,
or until τ +

= H + 1. Take Q
τ
= D
τ
+ +D
τ+

−1
as the order quantity for period τ.
Step 3: Take τ +

as the new value of τ (the new reorder instant). Repeat from Step 2 if τ ≤ H;
otherwise, stop.
The average cost
¯
C(τ, ) is not computed if D
τ+−1
= 0. Then, is increased until a period with nonzero
demand is found. The new average cost is then compared to the previously computed average cost.
Remark 1.1 The Silver-Meal heuristic has been tested extensively, see Baker [9]. In many cases, the total
cost of the generated order policy is less than 1% above the minimum. The heuristic works less well when
demand strongly diminishes in later periods and when there are many periods with zero demand, cf. Silver
et al. [64, Sct. 6.6.7]. Silver & Miltenburg [63] have developed a modification of the Silver-Meal heuristic for
such cases.
Another heuristic for this model is the part-period-balancing algorithm, cf. DeMatteis [22], Mendoza [52].
Let again τ be a reorder instant at which an order is placed for the periods τ, . . . , τ + . The cumulative
holding cost over these periods is
C
h
(τ, ) = h

u=1
uD
τ+u
, = 0, 1, . . . , H −τ, τ = 1, . . . , H. (1.30)
This heuristic tries to compensate ordering cost against future holding costs. A new order is placed as soon
as the cumulative holding cost since the previous reorder instant exceed the ordering cost.
Algorithm 1.4 [Part-period-balancing]
Step 1: Start with τ = 1.
Step 2: Compute the cumulative holding costs C
h
(τ, ) for = 1, 2, . . . , until C
h
(τ, ) > a for some , or
until τ + = H + 1. Take Q
τ
= D
τ
+ +D
τ+−1
as the order quantity for period τ.
Step 3: Take τ + as the new value of τ (the new reorder instant). Repeat from Step 2 if τ ≤ H;
otherwise, stop.
If the foregoing procedure generates an order in period H, the solution can be improved by shifting the
quantity D
H
to the previous reorder instant τ
p
if h(H −τ
p
)D
H
< a.
Remark 1.2 The above algorithms do not take a possible lead time into account. If there is a deterministic
lead time L > 0 the algorithms should be applied a time interval L before the beginning of the first period
in order that the first quantity Q
1
is delivered in time.
Remark 1.3 The above algorithms assume that the initial stock is zero. If the initial stock is nonzero,
say I
0
> 0, then the demand of the first few periods have to be diminished by the initial stock before the
algorithms can be applied with the modified demand. For instance, if D
1
≤ I
0
< D
1
+D
2
, the demand of the
first period can be completely satisfied by the initial stock (and, hence, Q
1
= 0), while the modified demand
for the second period becomes
˜
D
2
= D
2
− (I
0
− D
1
) > 0. The algorithms are then applied from period 2
onwards.
12 INTRODUCTION TO INVENTORY MANAGEMENT
Remark 1.4 The foregoing algorithms can all be applied with a rolling horizon. This may be useful when
the demand in periods 2, . . . , H is still subject to change, or if demand information over new periods H +
1, H + 2, . . . becomes available. In the first case, problems occur when the demand changes for periods for
which the demand has already been ordered. In general, the solutions of the models with deterministic time-
varying demand are very sensitive to small irregularities (breakage, loss) and upward changes in demand,
and a safety stock may be required to prevent stockouts under such circumstances. In the latter case, a new
solution with a new horizon is preferably determined at the first reorder instant after period 1. Blackburn
& Millen [12] have found that the Silver-Meal heuristic may outperform the Wagner-Whitin algorithm in a
rolling horizon environment because the solution found by the latter method is too sensitive to changes.
1.2.6 Stochastic demand, continuous review
This section is concerned with inventory systems with stationary stochastic demand and continuous review
policies. As in Section 1.2.1, a denotes the ordering cost per order and h denotes the holding cost per unit per
unit of time. Lead times are assumed to be constant, and equal to L units of time. Further, it is assumed that
each customer demands one unit and that all demand occurring when the item is out of stock is backlogged.
The expected demand per unit of time will be denoted by E¦D¦ and the standard deviation of the demand
per unit of time by σ¦D¦. The most commonly used continuous review policies are:
• the (s, Q) policy: whenever the inventory position drops at or below the reorder point s a fixed quantity
Q is ordered;
• the (s, S) policy: whenever the inventory position drops at or below the reorder point s an order is
placed of a size that brings the inventory position to the order-up-to level S. Policies with an order-up-to
level are also called base-stock policies.
In case of unit demand per customer, the (s, Q) and (s, S) policies are equivalent when Q = S − s. In the
analysis of continuous review policies an important role is played by the demand during the lead time. This
random variable will be denoted by D
L
. With constant lead times, the mean and the variance of D
L
are
given by
E¦D
L
¦ = LE¦D¦, σ
2
¦D
L
¦ = Lσ
2
¦D¦. (1.31)
The safety stock, defined as s −E¦D
L
¦, is needed to protect against the uncertainty in the demand during
the lead time. In case of a fill rate constraint with target β, cf. Section 1.1, the reorder point s has to be
chosen such that the actual fill rate Ψ exceeds the target fill rate:
Ψ = 1 −
E¦B¦
Q
= 1 −
E¦[D
L
−s]
+
¦ −E¦[D
L
−S]
+
¦
S −s
≥ β; (1.32)
here, the random variable B denotes the amount of backlog accumulated during a reorder cycle, [x]
+
.
=
max¦0, x¦ for all real x, [D
L
−s]
+
is the amount of backlog at the end of a reorder cycle and [D
L
−S]
+
is the
amount of backlog at the beginning of a reorder cycle. By a balance argument, the expected demand during
a reorder cycle is equal to Q = S −s. Since the expected number of orders per unit of time is E¦D¦/Q and
since the expected inventory on hand just before a delivery instant is E¦[s −D
L
]
+
and just after a delivery
instant is E¦[S −D
L
]
+
, the expected average cost per unit of time is
E¦C(s, S)¦ =
aE¦D¦
S −s
+
1
2
h[E¦[s −D
L
]
+
¦ +E¦[S −D
L
]
+
¦]. (1.33)
Although the decision variables s and S should be simultaneously determined to minimize E¦C(s, S)¦ subject
to the fill rate constraint (1.32), it is common practice to follow a sequential approach in which first the order
quantity is determined according to the EOQ formula, that is,
Q = S −s =
_
2aE¦D¦
h
, (1.34)
usually rounded to the nearest integer, and then the reorder point s is determined as the smallest integer
such that, cf. (1.32),
E¦[D
L
−s]
+
¦ −E¦[D
L
−S]
+
¦ ≤ (1 −β)Q = (1 −β)
_
2aE¦D¦
h
. (1.35)
HANS BLANC, TILBURG UNIVERSITY 13
This sequential approach performs well if σ¦D
L
¦ < Q, cf. Tijms [65]. The second term in the fill rate
constraint, E¦[D
L
−S]
+
¦, is negligible if the target fill rate is large (β > 0.9) and the coefficient of variation
of the demand during the lead time D
L
is small (σ¦D
L
¦/E¦D
L
¦ < 0.5). Zheng & Federgruen [82] describe
an efficient algorithm to search the optimal values of s and S simultaneously. See also Federgruen & Zipkin
[29] and Axs¨ater [7, Ch. 5,6].
Example 1.3 In the case that unit demand occurs according to a Poisson process at a rate of λ = E¦D¦
customers per unit of time, the random variable D
L
has a Poisson distribution with parameter λL, cf.
Appendix A.2.3. The constraint (1.35) then reads
_
_

j=s
(j −s)
(λL)
j
j!

j=S
(j −S)
(λL)
j
j!
_
_
e
−λL
≤ (1 −β)
_
2aλ
h
.
The infinite sums can be computed by the relation

j=m
(j −m)
(λL)
j
j!
e
−λL
= λL −m+
m−1

j=0
(m−j)
(λL)
j
j!
e
−λL
, m = 0, 1, 2, . . . .
The expected average cost per unit of time (1.33) becomes
E¦C(s, S)¦ =

S −s
+
1
2
h
_
_
s

j=0
(s −j)
(λL)
j
j!
+
S

j=0
(S −j)
(λL)
j
j!
_
_
e
−λL
.
Consider the following numerical example: a = $5, h = $0.05 per unit per day, and a Poisson demand rate
of λ = 10 units per day. This implies that Q = 44.72 by (1.34), with corresponding reorder cycle T = 4.47.
Table 1.1: Feasible reorder points and order-up-to levels for L = 5, β = 0.98.
s S Q E¦[S −D
L
]
+
¦ E¦[s −D
L
]
+
¦ E¦[D
L
−S]
+
¦ E¦[D
L
−s]
+
¦ E¦C(s, S)¦ Ψ
56 100 44 50.00 6.82 0.00 0.82 2.5567 0.9815
56 101 45 51.00 6.82 0.00 0.82 2.5565 0.9819
56 102 46 52.00 6.82 0.00 0.82 2.5573 0.9823
56 103 47 53.00 6.82 0.00 0.82 2.5592 0.9827
56 104 48 54.00 6.82 0.00 0.82 2.5620 0.9830
56 105 49 55.00 6.82 0.00 0.82 2.5658 0.9834
56 106 50 56.00 6.82 0.00 0.82 2.5704 0.9837
56 107 51 57.00 6.82 0.00 0.82 2.5758 0.9840
55 107 52 57.00 6.03 0.00 1.03 2.5373 0.9802
55 108 53 58.00 6.03 0.00 1.03 2.5442 0.9806
55 119 64 69.00 6.03 0.00 1.03 2.6570 0.9839
54 119 65 69.00 5.29 0.00 1.29 2.6264 0.9802
Table 1.1 concerns the case of a lead time of L = 5 days and a target fill rate of β = 0.98. It contains for
several values of the order quantity Q around Q = 44.72 the policy with the minimum value of the reorder
point s that meets the fill rate constraint. Note that the expected average cost and the actual fill rate are
rather flat functions of the order quantity Q in this region. Further, note that the expected backlog at the
beginning of a reorder cycle, E¦[D
L
−S]
+
¦, is negligible so that the expected stock level at the beginning of a
reorder cycle is E¦[S −D
L
]
+
¦ ≈ S −E¦D
L
¦ = S −50. The safety stock is s−E¦D
L
¦ = 6 in this region. The
expected average cost has a local minimum at Q = 45. However, when the order quantity becomes Q = 52
the required reorder point drops from s = 56 to s = 55 and another local minimum of C(55, 107) = $2.5373
is met which is less than C(56, 101) = $2.5565. The next local minimum is found for Q = 65 but here the
cost is higher: C(54, 119) = $2.6264.
Table 1.2 concerns the case of a lead time of L = 50 days and a target fill rate of β = 0.98. Similar remarks
can be made as those made with regard to Table 1.1. The safety stock is now s −E¦D
L
¦ = 31 in the region
of Q = 44.72. Local minima of the expected average cost are found at Q = 45, Q = 49, Q = 54, Q = 59,
14 INTRODUCTION TO INVENTORY MANAGEMENT
Table 1.2: Feasible reorder points and order-up-to levels for L = 50, β = 0.98.
s S Q E¦[S −D
L
]
+
¦ E¦[s −D
L
]
+
¦ E¦[D
L
−S]
+
¦ E¦[D
L
−s]
+
¦ E¦C(s, S)¦ Ψ
531 575 44 75.00 31.88 0.00 0.88 3.8084 0.9801
531 576 45 76.00 31.88 0.00 0.88 3.8082 0.9805
531 577 46 77.00 31.88 0.00 0.88 3.8090 0.9809
530 579 49 79.00 30.97 0.00 0.97 3.7696 0.9803
529 583 54 83.00 30.06 0.00 1.06 3.7525 0.9804
528 587 59 87.00 29.16 0.00 1.16 3.7515 0.9803
527 591 64 91.00 28.27 0.00 1.27 3.7631 0.9801
Table 1.3: Feasible reorder points and order-up-to levels for L = 5, β = 0.50.
s S Q E¦[S −D
L
]
+
¦ E¦[s −D
L
]
+
¦ E¦[D
L
−S]
+
¦ E¦[D
L
−s]
+
¦ E¦C(s, S)¦ Ψ
28 72 44 22.00 0.00 0.00 22.00 1.6865 0.5001
28 73 45 23.00 0.00 0.00 22.00 1.6862 0.5112
27 73 46 23.00 0.00 0.00 23.00 1.6620 0.5000
27 74 47 24.00 0.00 0.00 23.00 1.6639 0.5107
26 74 48 24.00 0.00 0.00 24.00 1.6417 0.5000
19 81 62 31.00 0.00 0.00 31.00 1.5815 0.5000
18 82 64 32.00 0.00 0.00 32.00 1.5813 0.5000
17 83 66 33.00 0.00 0.00 33.00 1.5826 0.5000
Q = 64, etcetera. The global minimum is C(528, 587) = $3.7515. The safety stock is here s −E¦D
L
¦ = 28.
Note that a ten-fold increase in the lead times leads to a 48% increase in cost.
Table 1.3 concerns the case of a lead time of L = 5 days and a low target fill rate of β = 0.50. Note that
the expected backlog at the beginning of a reorder cycle, E¦[D
L
−S]
+
¦, is still negligible. In this case, also
the expected stock level at the end of a reorder cycle, E¦[s − D
L
]
+
¦ is zero. The expected net stock at
the end of a reorder cycle is negative so that this case with a low target fill rate corresponds to the model
with planned stockouts, cf. Section 1.2.4. Also, the safety stock, s − E¦D
L
¦ = s − 50, is negative in these
cases. The expected average cost and the actual fill rate are more sensitive to the order quantity Q. The cost
function has a local minimum for every even order quantity Q. The global minimum is found for Q = 64:
C(18, 82) = $1.5813.
Table 1.4: Feasible reorder points and order-up-to levels for L = 1, β = 0.50.
s S Q E¦[S −D
L
]
+
¦ E¦[s −D
L
]
+
¦ E¦[D
L
−S]
+
¦ E¦[D
L
−s]
+
¦ E¦C(s, S)¦ Ψ
−12 32 44 22.00 0.00 0.00 22.00 1.6864 0.5000
−12 33 45 23.00 0.00 0.00 22.00 1.6861 0.5111
−13 33 46 23.00 0.00 0.00 23.00 1.6620 0.5000
−13 34 47 24.00 0.00 0.00 23.00 1.6638 0.5106
−14 34 48 24.00 0.00 0.00 24.00 1.6417 0.5000
−21 41 62 31.00 0.00 0.00 31.00 1.5815 0.5000
−22 42 64 32.00 0.00 0.00 32.00 1.5813 0.5000
−23 43 66 33.00 0.00 0.00 33.00 1.5826 0.5000
Table 1.4 concerns the case of a lead time of L = 1 days and a low target fill rate of β = 0.50. The main
difference with the case of L = 5 days in Table 1.3 is that the reorder point has become negative. The safety
stocks, s −E¦D
L
¦, are the same for L = 1 and L = 5 at given values of Q. In all four numerical examples,
a better order quantity is found than the quantity Q = 45 suggested by (1.34).
Remark 1.5 In case of a cycle service constraint with cycle service level α, the reorder point s has to be
chosen such that
Pr¦D
L
≤ s¦ ≥ α.
HANS BLANC, TILBURG UNIVERSITY 15
For the case of a Poisson demand process this condition reads
s

j=0
(λL)
j
j!
e
−λL
≥ α.
Note that this condition does not involve Q or S. Hence, with this service level constraint first s can be
determined as the smallest integer satisfying the above constraint, and then a value for S can be searched
that minimizes the expected average cost per unit of time (1.33).
1.2.7 Stochastic demand, periodic review
This section is concerned with inventory systems with stationary stochastic demand and periodic reviewpolicies.
The notations and assumptions are otherwise the same as in Section 1.2.6. The review period will be denoted
by R. The most commonly used periodic review policies are:
• the (R, S) policy: with intervals of R units of time an order is placed of a size that brings the inventory
position to the order-up-to level S;
• the (R, s, S) policy: with intervals of R units of time the inventory level is reviewed; whenever the
inventory position is then at or below the reorder point s an order is placed of a size that brings the
inventory position to the order-up-to level S.
The length of the review period is either assumed to be given by other management considerations or can
be chosen as the optimal reorder cycle length (1.5), in analogy to the sequential approach in the continuous
review case,
R = R

.
=
¸
2a
hE¦D¦
. (1.36)
If the review period R is much smaller than the value R

, it is preferable to use an (R, s, S) policy to avoid
many small orders. Here, we will only discuss the (R, S) policy. If an order is placed at a review instant, it
is delivered after a lead time L while the next order can only arrive after a time L + R. Hence, the current
order quantity should be sufficient to protect against the uncertainty in the demand during a lead time plus
a review period. This random variable will be denoted by D
L+R
. With constant lead times, the mean and
the variance of D
L+R
are, in analogy to (1.31), given by
E¦D
L+R
¦ = (L +R)E¦D¦, σ
2
¦D
L+R
¦ = (L +R)σ
2
¦D¦. (1.37)
When all stockouts are backlogged, and for a fixed review period R, the order-up-to level S has to be
determined such that a fraction β of all demand can be sold directly from stock (fill rate constraint):
Ψ = 1 −
E¦B¦
E¦Q¦
= 1 −
E¦[D
L+R
−S]
+
¦ −E¦[D
L
−S]
+
¦
RE¦D¦
≥ β; (1.38)
here, B stands for the accumulated backlog during a review period R, as in (1.32), and the expected value of
an order quantity is E¦Q¦ = RE¦D¦ by a balance argument. The safety stock for this policy is S−E¦D
L+R
¦.
The expected average cost per unit of time is
E¦C(R, S)¦ =
a
R
+
1
2
h[E¦[S −D
L+R
]
+
¦ +E¦[S −D
L
]
+
¦]. (1.39)
Note that there is no decision variable left to minimize this objective function if R is fixed and S is set to
satisfy the service level constraint (1.38). The expected inventory on hand is simply related to the expected
backlog through
E¦[S −D
L+R
]
+
¦ = S −(L +R)E¦D¦ +E¦[D
L+R
−S]
+
¦. (1.40)
Example 1.4 The required order-up-to level S will be determined for various distributions for the demand
during the lead time plus review period. In all cases, the review period is R = 4 weeks and the lead time is
L = 1 week. The mean demand per week is E¦D¦ = 20, so that RE¦D¦ = 80 and E¦D
L+R
¦ = 100. The
cost factors are a = $5 and h = $0.05 per unit per week, so that R = 4 is the largest integer smaller than
R

=

20 according to (1.36). The order-up-to level S is obtained by solving
E¦[D
L+R
−S]
+
¦ −E¦[D
L
−S]
+
¦ ≤ (1 −β)RE¦D¦ = 80(1 −β).
16 INTRODUCTION TO INVENTORY MANAGEMENT
Table 1.5: Order-up-to levels, fill rate, average cost and backlog at end of cycle for σ
2
¦D¦ = 125.
Target Normal Gamma
β S Ψ E¦C(R, S)¦ E¦[S −D
L+R
]
+
¦ S Ψ E¦C(R, S)¦ E¦[S −D
L+R
]
+
¦
0.90 105 0.9041 3.6918 7.67 105 0.9027 3.6946 7.78
0.95 116 0.9506 4.1487 3.95 118 0.9517 4.2465 3.86
0.96 120 0.9624 4.3260 3.01 122 0.9619 4.4263 3.05
0.97 124 0.9719 4.5062 2.25 126 0.9701 4.6098 2.39
0.98 129 0.9810 4.7380 1.52 133 0.9809 4.9382 1.53
0.99 137 0.9904 5.4193 0.77 143 0.9903 5.1493 0.77
In Table 1.5 the normal distribution is compared with the gamma distribution in a case with a low variation
in the demand: σ
2
¦D¦ = 125. This implies, cf. (1.37), that σ¦D
L+R
¦ = 25 with a coefficient of variation
C
L+R
= 0.25, and that σ¦D
L
¦ = 5

5 ≈ 11.18, with a coefficient of variation C
L
≈ 0.56. The expected
backlogs E¦[D
L+R
− S]
+
¦ and E¦[D
L
− S]
+
¦ are computed with the aid of the normal loss function, cf.
(A.21), with parameters µ = 100, σ = 25 and µ = 20, σ = 11.18, respectively. The probabilities of negative
realizations are for these normally distributed random variables 3 10
−5
and 0.037, respectively. These
expected backlogs are also computed with the aid of the gamma loss function, cf. (A.38), with parameters
ψ = 16, λ =
4
25
(hence, an Erlang distribution) and ψ = 3.2, λ =
4
25
, respectively. Observe that the normal
distribution yields lower order-up-to levels than the gamma distribution due to the lighter tail of the normal
distribution. In these cases, the expected backlog at the beginning of a cycle is negligible. In Table 1.6
Table 1.6: Order-up-to levels, fill rate, average cost and backlog at end of cycle for σ
2
¦D¦ = 1125.
Target Gamma ME (Erlang-2 and Exponential)
β S Ψ E¦C(R, S)¦ E¦[S −D
L+R
]
+
¦ S Ψ E¦C(R, S)¦ E¦[S −D
L+R
]
+
¦
0.90 185 0.9008 7.7132 8.23 184 0.9004 7.6633 8.25
0.95 231 0.9506 9.9047 4.07 229 0.9503 9.8045 4.08
0.96 245 0.9602 10.5841 3.27 243 0.9602 10.4835 3.26
0.97 264 0.9704 11.5123 2.43 261 0.9701 11.3624 2.44
0.98 290 0.9803 12.7913 1.61 287 0.9803 12.6408 1.60
0.99 333 0.9900 14.9207 0.81 329 0.9901 14.7204 0.80
the gamma distribution is compared with a mixture of two Erlang distributions in a case with a moderate
variation in the demand: σ
2
¦D¦ = 1125. This implies that σ¦D
L+R
¦ = 75 with a coefficient of variation
C
L+R
= 0.75, and that σ¦D
L
¦ = 15

5 ≈ 33.54, with a coefficient of variation C
L
≈ 1.68. The expected
backlogs are computed with the aid of the gamma loss function with parameters ψ =
16
9
, λ =
4
225
and ψ =
16
45
,
λ =
4
225
, respectively. The parameters of the mixed Erlang distribution are, cf. (A.42), K
1
= 1, K
2
= 2,
λ
1
= λ
2
= 0.0188, p = 0.1213, and, cf. (A.43), K
1
= K
2
= 1, λ
1
= 0.1779, λ
2
= 0.0221, p = 0.6368,
respectively. The corresponding loss functions are given in (A.44) and (A.46). The expected backlog at the
beginning of a cycle varies from about 0.3 for β = 0.9 to about 0.02 for β = 0.99; ignoring this quantity
in (1.38) leads to order-up-to levels that are about 2 higher than those displayed in the table. Observe
that the mixed Erlang distribution yields lower order-up-to levels than the gamma distribution. The normal
distribution does not seem appropriate in this case since the probability of negative realizations of D
L+R
is
0.09. Use of the normal distribution gives much lower order-up-to levels, e.g., S = 244 for β = 0.99.
In Table 1.7 the gamma distribution is compared with a hyperexponential distribution (H
2
) in a case with a
high variation in the demand: σ
2
¦D¦ = 8000. This implies that σ¦D
L+R
¦ = 200 with a coefficient of variation
C
L+R
= 2, and that σ¦D
L
¦ = 40

5 ≈ 89.44, with a coefficient of variation C
L
≈ 4.47. The parameters
of the gamma distributions are ψ =
1
4
, λ =
1
400
and ψ =
1
20
, λ =
1
400
, respectively. The parameters of the
hyperexponential distributions are K
1
= K
2
= 1, λ
1
= 0.03673, λ
2
= 0.00327, p = 0.7390, and K
1
= K
2
= 1,
λ
1
= 0.01964, λ
2
= 0.0036, p = 0.9447, respectively. The expected backlog at the beginning of a cycle varies
from about 1.4 for β = 0.9 to about 0.1 for β = 0.99; ignoring this quantity in (1.38) leads to order-up-to
levels that are about 50 higher than those displayed in the table (for the hyperexponential distribution the
difference decreases with increasing β). Observe that the hyperexponential distribution yields lower order-
up-to levels than the gamma distribution for β > 0.95 (tail behavior) and higher levels for β ≤ 0.95. The
normal distribution is not appropriate in this case since the probability of negative realizations of D
L+R
is
HANS BLANC, TILBURG UNIVERSITY 17
Table 1.7: Order-up-to levels, fill rate, average cost and backlog at end of cycle for σ
2
¦D¦ = 8000.
Target Gamma H
2
β S Ψ E¦C(R, S)¦ E¦[S −D
L+R
]
+
¦ S Ψ E¦C(R, S)¦ E¦[S −D
L+R
]
+
¦
0.90 636 0.9000 30.32 9.40 655 0.9000 31.27 9.40
0.95 863 0.9500 41.53 4.67 872 0.9501 41.98 4.63
0.96 938 0.9601 45.26 3.72 941 0.9600 45.40 3.69
0.97 1035 0.9700 50.08 2.79 1031 0.9701 49.88 2.75
0.98 1174 0.9800 57.00 1.85 1157 0.9800 56.15 1.82
0.99 1415 0.9900 69.03 0.92 1373 0.9900 66.93 0.90
0.3.
Table 1.8: Order-up-to levels and average cost for various R at β = 0.95.
R = 1 R = 2 R = 3 R = 4 R = 5
C
L+R
S E¦C(R, S)¦ S E¦C(R, S)¦ S E¦C(R, S)¦ S E¦C(R, S)¦ S E¦C(R, S)¦
0.25 62 6.63 80 4.55 99 4.19 118 4.25 137 4.47
0.75 168 11.95 189 10.01 210 9.75 231 9.90 251 10.18
2.00 790 43.07 814 41.29 839 41.23 863 41.53 888 42.06
Finally, it turns out that in all above discussed cases a shorter review period gives a lower expected average
cost. Table 1.8 shows the required order-up-to levels S and the expected average cost E¦C(R, S)¦ according
to gamma distributed demand at a target fill rate of β = 0.95 for the coefficients of variations in the demand
considered above and for various lengths of the review period R. The minimum cost is attained for R = 3.
Observe the tendency that a shorter review period becomes cheaper when the coefficient of variation in the
demand increases. As could be expected, the required order-up-to level is an increasing function of the review
period.
1.3 Interactions in Inventory Management
The basic models of inventory theory assume that there is no interaction between various items and various
stocking points, and consider individual items at a single stocking point. However, in many practical cases
interactions do exist, and an inventory policy may be far from optimal when such interactions are neglected.
Interactions may be due to (see, e.g., Love [49, Ch. 5]):
• joint capacity constraints:
– a common storage space used for various items;
– a common labor force and equipment for the receipt and handling of various items;
– a single budget available for the inventory of various items.
• correlated demand:
– assembly requires parts in fixed proportions;
– special offers or advertisements at a regional or national scale influence sales at many stocking
points;
– substitution effect (e.g., the same item, but of a different color);
– complementarity effect (several units of items that fit together, or make up a whole).
• interwoven cost structure:
– joint ordering cost for a family of items ordered from the same supplier;
– joint shipping cost;
– joint discount structure (on the total amount or the total quantity of an order);
18 INTRODUCTION TO INVENTORY MANAGEMENT
– joint handling cost in a warehouse.
• interwoven product structure:
– raw material is used in the production of several items;
– produced parts are used in several final products.
• interwoven distribution structure:
– an item is kept on stock at several locations (within the same company or organization);
– sold units may return to a stocking point because of customer dissatisfaction (e.g., clothing or
items bought from a mail order company) or after replacement by a spare part when repaired.
In Chapter 2, coordinated replenishment strategies will be discussed for families of items with a joint cost
structure. The considered interactions through cost structure are:
• joint ordering, production and/or transportation costs:
– items/parts stem from the same supplier or can be shipped with the same transport;
– items belong to the same product group (with limited mutual switch-over times) or are variants
of each other (only differing in packing, enclosing, etcetera);
– there is a fixed ordering cost (A), independent of the composition of the order (administration,
shipment);
– there are additional ordering costs (a
i
), depending on whether item i forms part of an order;
– the key question is: should an item be ordered at a reorder instant or not (the order quantities
are of minor importance).
• joint discount structure:
– a supplier offers discount on purchasing prices or shipping cost if the total value or volume of an
order exceeds some threshold;
– example: a contribution in shipping cost is only passed on below a certain order amount;
– the order quantities are important in this case.
Advantages of joint replenishment (or joint production) strategies may be:
• lower purchasing cost;
• lower shipping cost;
• lower handling cost;
• lower ordering cost;
• lower production cost (set-up, switch-over).
Disadvantages of joint replenishment (or joint production) strategies may be:
• higher average inventory levels (since orders may be put forward and EOQ quantities are adjusted);
• inventory management more complex;
• higher management cost (more involved reviews, more complex computations);
• higher peaks at the receipt of goods;
• reduced flexibility (e.g., in coping with unusual situations).
HANS BLANC, TILBURG UNIVERSITY 19
Chapter 3 is concerned with replenishment strategies for multi-echelon inventory systems where items are
kept in stock at various storage locations. Storage locations often form part of complex good flows (supply
chains), from raw materials via parts and semi-manufactured items to end products which reach customers
possibly via several distribution nodes: this results in multi-echelon inventory systems. Coordination of
inventory management at various stocking points is desirable (since orders of lower levels form the demand
process for higher levels, whereby demand usually occurs in larger quantities at higher echelons) to avoid
possible oscillating effects. Coordination of inventory management on various stocking points is possible as
far as they are controlled by the same company or if good agreements exist between companies (e.g., to
prevent hoarding).
In order to implement a multi-echelon replenishment strategy, proper information exchange between the
echelons is important (e.g., via EDI: electronic data interchange):
• from lower echelons upward: concerning reorder frequency, demand patterns, special actions (locally);
• from higher echelons downward: concerning production or replenishment plan: when becomes how
much available of which item; product advertisements (globally).
In multi-echelon inventory management a distinction is made between pull and push systems. In a pull
system each stocking point determines its own reorder policy. In a push system information is kept up to
date at a central level and stocks are controlled from a central point. Push systems
• require much and reliable information;
• realize cost savings by better anticipation and coordination;
• may give problems with responsibilities of local managers.
As a compromise, a central planning may be implemented that allows limited local modifications provided
that the latter are reported in time.
Chapter 4 is devoted to replenishment strategies for inventory systems with capacity constraints and with
returning, repairable items.
Chapter 2
Coordinated Replenishment
This chapter is devoted to models, algorithms and policies for inventory systems in which coordinated re-
plenishment of items may be profitable. Consecutively, models with constant, with time-varying and with
stochastic demand will be discussed. The following concepts will be used:
family: a set of items that possesses a joint cost structure (e.g., because they are ordered from the same
supplier, or because they are delivered by the same transporter);
group: a set of items from a family that have the same constant reorder cycle time.
With an individual item it is optimal to have a fixed reorder cycle; with coordinated ordering of several items
this is not necessarily the case; in general it is assumed as a simplification that a group of items has a fixed
reorder cycle (this may be suboptimal).
A distinction will be made between indirect and direct grouping of items, cf. Chakravarty [16]:
indirect grouping: the reorder cycle of a group is an integer multiple of the family reorder cycle (this
grouping aims at double saving on ordering cost, both within groups and across groups);
direct grouping: the reorder cycle of a group is not related to those of other groups within the same family
(this grouping aims at single saving on ordering cost, only within groups).
2.1 Constant demand, indirect grouping
This section is concerned with indirect grouping policies for a family of items with constant demand. The
analysis will be based on fluid approximations for the demand and the inventory level. The assumptions for
the model with constant (known) demand for a family of N items are:
• a family or joint ordering cost A, which is independent of the composition of the order;
• a supplementary ordering cost a
i
for item i, which only is incurred when this item is included in an
order;
• the demand rate D
i
for item i is a constant number of units per unit of time;
• a holding cost h
i
for item i is incurred per unit per unit of time; often, h
i
= rv
i
, with v
i
the purchasing
cost (value) of item i and r an item independent carrying charge;
• no back orders or stockouts are allowed;
• no capacity restrictions on order quantities (or on production lot sizes) exist;
• the delivery of orders (or of production lots) takes place as a whole;
• the lead time is negligible or constant and is equal for all items;
• a continuous review of stocks is applied.
20
HANS BLANC, TILBURG UNIVERSITY 21
The aim is to minimize the average ordering and holding cost per unit of time (purchasing costs are fixed
since no back orders or shortages are allowed). The decision variables could be the order quantities for the
N items, but in case of coordinated replenishment it is more suitable to consider the reorder frequencies for
the N items. For the case of indirect grouping of items the decision variables are more specifically defined
as:
• with time intervals of length T a family order is placed (the family reorder cycle);
• item i is included in every k
i
th replenishment of the family (k
i
= 1, 2, . . .).
Hence, the reorder cycle of item i has a length of T
i
= k
i
T. Items with the same reorder frequency k
i
are
said to form a group. The objective function to be minimized reads for indirect grouping:
C(T, k) =
1
T
_
A+
N

i=1
a
i
k
i
_
+
1
2
T
N

i=1
k
i
h
i
D
i
. (2.1)
Here, a restrictive assumption is that min
i
¦k
i
¦ = 1 (otherwise, a modification of the objective function is
required with respect to the term A/T since family orders are not placed every T units of time).
Remark 2.1 A trivial upper bound on the minimum cost is obtained by using the EOQ formula with
ordering cost A+a
i
for each item separately and by summing the minimum cost per item over all items. A
lower bound is obtained by ignoring the term A/T, and by minimizing the resulting separable function in
the variables T
i
= k
i
T. Hence,
N

i=1
_
2a
i
h
i
D
i
≤ C


N

i=1
_
2(A+a
i
)h
i
D
i
.
If the minimum cost C

with coordinated replenishment is close to the upper bound, coordination may not
be profitable. In such a case, the increase in management cost may be larger than the reduction in inventory
cost obtained by coordination.
The optimal family cycle for a given vector of reorder frequencies k = (k
1
, . . . , k
N
) can be derived in a similar
way as the basic EOQ-formula, cf. (1.5) and Appendix B.1:
ˆ
T(k) =
¸
¸
¸
_
2[A+

N
j=1
(a
j
/k
j
)]

N
i=1
k
i
h
i
D
i
. (2.2)
The corresponding average inventory cost per unit of time is, cf. (1.6):
ˆ
C(k) =
¸
¸
¸
¸
_2
_
_
A+
N

j=1
a
j
k
j
_
_
N

i=1
k
i
h
i
D
i
=
ˆ
T(k)
N

i=1
k
i
h
i
D
i
. (2.3)
In a similar way, it follows for a given family reorder cycle T that the frequencies
ˆ
k
i
(T) =
_
2a
i
h
i
D
i
T
2
, i = 1, . . . , N, (2.4)
would be optimal if they were integer. A rounding-off rule like (1.8) states: round
ˆ
k
i
(T) to k
i
=
˜
k
i
(T) if
k
i
(k
i
−1) <
2a
i
h
i
D
i
T
2
≤ k
i
(k
i
+ 1), i = 1, . . . , N. (2.5)
Clearly, the frequencies
˜
k
i
(T) are ordered according to ascending value of a
i
/(h
i
D
i
). The relative values of
the optimal frequencies are independent of A. The actual optimal values do depend on A through T. Further,
under this rounding the family cycle T is bounded by
_
2a
i
h
i
D
i
¸
1
˜
k
i
(T)(
˜
k
i
(T) + 1)
≤ T <
_
2a
i
h
i
D
i
¸
1
˜
k
i
(T)(
˜
k
i
(T) −1)
, i = 1, . . . , N. (2.6)
22 COORDINATED REPLENISHMENT
Remark 2.2 It is readily verified that the function
ˆ
T(k), cf. (2.2), is decreasing in each individual k
i
, i =
1, . . . , N. Hence, the maximum value of this function is at k = 1
.
= (1, 1, . . . , 1). Hence, for all feasible vectors
k it holds that
ˆ
T(k) ≤ T
max
.
=
ˆ
T(1) =
¸
¸
¸
_
2[A+

N
j=1
a
j
]

N
i=1
h
i
D
i
.
A minimum value for the function
ˆ
T(k) is obtained from the lefthand inequality in (2.6) recalling the as-
sumption that min
i
¦k
i
¦ = 1. This implies that for all feasible vectors k we have
ˆ
T(k) ≥ T
min
.
= min
i=1,...,N
_
a
i
h
i
D
i
.
Smaller family cycles may be associated with policies with min
i
¦k
i
¦ > 1.
When the family reorder cycle T and the reorder frequencies k
i
, i = 1, . . . , N, have been determined, the
order quantities follow by (1.3) as Q
i
= D
i
k
i
T, i = 1, . . . , N.
2.1.1 Exact algorithmic solution
Goyal [33] has developed an algorithm for determination of the exact optimum of the objective function (2.1)
under the assumption that min
i
¦k
i
¦ = 1. It is based on the observations that the optimal family cycle T

lies on the finite interval [T
min
, T
max
], cf. Remark 2.2, and that the inequalities (2.6) divide this interval into
a finite number of subintervals. As a consequence, the algorithm consists of finitely many steps.
Algorithm 2.1 [Goyal]
Step 1: Compute the maximum value T
max
=
ˆ
T(1) and the minimum value T
min
for the family cycle T,
cf. Remark 2.2.
Step 2: Start with the optimal (rounded) frequencies
˜
k
i
(T
max
); call this vector k
1
; compute the corre-
sponding costs
ˆ
C(k
1
).
Step 3: Determine for each item i the length of the family cycle T
c
(i) for which the value of k
i
=
˜
k
i
(T
max
)
changes to k
i
+ 1 via the bounds (2.6) of T = T
max
:
T
c
(i) =
_
2a
i
h
i
D
i
¸
1
k
i
(k
i
+ 1)
. (2.7)
Set j = 1 and start the recursion (steps 4 and 5).
Step 4: Increase j by one. Let T
c
= max
i
¦T
c
(i)¦ be the first boundary that is met when T decreases. If
T
c
≤ T
min
then goto step 6, otherwise continue with step 5.
Step 5: Let i
c
be the item for which T
c
(i
c
) = max
i
¦T
c
(i)¦. Change k
i
c
(T) to k
i
c
(T) + 1 to obtain the
vector k
j
and compute the corresponding minimum cost
ˆ
C(k
j
), cf. (2.3). Compute T
c
(i
c
) with the new
value of k
i
c
. Return to step 4.
Step 6: The algorithm stops: the optimal policy is that policy (
ˆ
T(k
j
), k
j
) for which the lowest costs
ˆ
C(k
j
) have been found.
The cost function
ˆ
C(k
j
) is not in all cases convex in j.
Example 2.1 Consider a family consisting of N = 3 items, with family ordering cost A = $6. Further data
(per week) can be found in Table 2.1, the first four columns.
The optimal individual policies, cf. Section 1.2.1, for the three items are indicated in the columns with headers
T
i
, Q
i
, and the corresponding costs under C
i
. The minimum total average cost per week without coordination
is C = $29. Note that in this academic case the reorder instants can be synchronized in cycles of 42 weeks in
which there are 21 +7 +2 = 30 family reorder instants instead of 21 +14 +6 = 41 without synchronization.
In this way,
11
42
A = $1.57 is saved on the average weekly cost.
HANS BLANC, TILBURG UNIVERSITY 23
Table 2.1: Data for Example 2.1; optimal individual and indirect grouping policies.
Item i a
i
h
i
D
i
A+a
i
h
i
D
i
T
i
Q
i
C
i
k

i
T

Q

i
˜ a
i
1 $ 3 $0.50 9 $ 9 $4.50 2 18 $ 9 2.05 18.47 $ 9.48
2 $ 3 $0.50 4 $ 9 $2.00 3 12 $ 6 2.05 8.21 $ 4.21
3 $43 $0.50 4 $49 $2.00 7 28 $14 6.16 24.63 $37.93
Next, consider coordinated replenishment by indirect grouping. Application of Goyal’s Algorithm 2.1 starts
with the computation of T
max
=
_
110/8.5 ≈ 3.597 and T
min
=
_
2/3 ≈ 0.816. Next, it determines for
T = T
max
: k
1
= 1 (
_
51/495), k
2
= 1 (
_
51/220), k
3
= 2 (
_
731/220), cf. (2.4).
The successive iterations are:
ˆ
C(1, 1, 2) = $26.52 [family cycle
ˆ
T(1, 1, 2) = 2.526]
T
c
(1) =
_
2/3 ≈ 0.816, T
c
(2) =
_
3/2 ≈ 1.225, T
c
(3) =
_
43/6 ≈ 2.677, T
c
= 2.677, i
c
= 3
ˆ
C(1, 1, 3) = $25.66 [family cycle
ˆ
T(1, 1, 3) = 2.053]
T
c
(3) =
_
43/12 ≈ 1.893, T
c
= 1.893, i
c
= 3
ˆ
C(1, 1, 4) = $25.69 [family cycle
ˆ
T(1, 1, 4) = 1.771]
T
c
(3) =
_
43/20 ≈ 1.466, T
c
= 1.466, i
c
= 3
ˆ
C(1, 1, 5) = $26.07 [family cycle
ˆ
T(1, 1, 5) = 1.580]
T
c
(3) =
_
43/30 ≈ 1.197, T
c
= 1.225, i
c
= 2
ˆ
C(1, 2, 5) = $26.58 [family cycle
ˆ
T(1, 2, 5) = 1.437]
T
c
(2) =
_
3/6 ≈ 0.707, T
c
= 1.197, i
c
= 3
ˆ
C(1, 2, 6) = $26.91 [family cycle
ˆ
T(1, 2, 6) = 1.313]
T
c
(3) =
_
43/42 ≈ 1.012, T
c
= 1.012, i
c
= 3
ˆ
C(1, 2, 7) = $27.37 [family cycle
ˆ
T(1, 2, 7) = 1.216]
T
c
(3) =
_
43/56 ≈ 0.876, T
c
= 0.876, i
c
= 3
ˆ
C(1, 2, 8) = $27.89 [family cycle
ˆ
T(1, 2, 8) = 1.138]
T
c
(3) =
_
43/72 ≈ 0.773, T
c
= 0.816, i
c
= 1
Now, T
c
= T
c
(1) = T
min
so that the iterations stop. The optimum reorder frequencies are k

= (1, 1, 3). The
optimum family cycle is T

=
ˆ
T(1, 1, 3) = 2.053. The minimum total average cost per week with coordination
is C

=
ˆ
C(1, 1, 3) = $25.66, that is, 11.5% less than without coordination (and without synchronization).
The optimum item cycles and the optimum order quantities are listed in Table 2.1 under the headers k

i
T

and Q

i
, respectively. For later reference, the table also contains the values ˜ a
i
.
=
1
2
h
i
D
i
(k

i
T

)
2
, i = 1, 2, 3,
which represent the ordering costs that give Q

i
as optimum individual order quantity, respectively.
Remark 2.3 The optimal indirect grouping policy is not found by Goyal’s algorithm if the minimum fre-
quency is larger than 1. The latter may be the case when the family ordering cost A is small with respect to
the individual ordering costs a
i
, i = 1, . . . , N. See Van Eijs [70] for an extension of the algorithm to include
strategies with minimum frequency larger than 1. Goyal’s algorithm requires a computation time, which
is strongly increasing with the number of items N in the family; as a consequence, it is only applicable to
moderately sized problems. The algorithm is, however, important for verifying the quality of heuristics.
Example 2.2 An example of a situation where the minimum frequency is larger than 1 in the optimal
indirect grouping policy has been provided by Andres & Emmons [2]. Consider a family of N = 2 items. Let
A = $1, a
1
= a
2
= $50, h
1
= h
2
= $1, D
1
= 400 and D
2
= 900. Goyal’s algorithm gives the following policy:
k

= (2, 1), T

=
ˆ
T(2, 1) = 0.299. The corresponding average cost per unit of time is C

=
ˆ
C(2, 1) = $508.33.
However, the average cost per unit of time corresponding to the frequencies k = (3, 2) is, cf. (2.1),
C(T, (3, 2)) =
1
T
[
2
3
A+
1
3
a
1
+
1
2
a
2
] +
1
2
T[3h
1
D
1
+ 2h
2
D
2
],
24 COORDINATED REPLENISHMENT
since only at four instants in each cycle of six potential reorder instants an order is placed. This cost function
is minimal for, cf. (2.2),
ˆ
T(3, 2) =
¸
2[
2
3
A+
1
3
a
1
+
1
2
a
2
]
3h
1
D
1
+ 2h
2
D
2
=
¸
2
3
127
3000
=
1
30

25.4 ≈ 0.168.
The corresponding average cost per unit of time is C = C(
ˆ
T(3, 2), (3, 2)) = 3000
ˆ
T(3, 2) = $503.98 which is
less than
ˆ
C(2, 1). Observe that the family ordering cost is much smaller than the item-dependent ordering
costs in this example.
Remark 2.4 Goyal’s algorithm is still applicable if A = 0 while a
i
> 0, i = 1, . . . , N. The minimum cost of
the coordinated replenishment problem (with integer ordering frequencies) will then be larger than (or equal
to) the sum of the minimum costs of the optimal individual policies (without integer constraints).
Exercise 2.1 What is the increase in cost if the family cycle in Example 2.1 is rounded to T = 2 weeks?
What do the order quantities become?
Exercise 2.2 A retailer orders two items from the same importer. The family ordering cost is A = $1,
while the supplementary item ordering costs are a
1
= $3 and a
2
= $8, respectively. The purchasing costs are
v
1
= $16 per unit of item 1 and v
2
= $10 per unit of item 2. The holding costs of both items per unit per
week are a carrying charge r = 0.005 times the purchasing cost. The demand for both items is assumed to be
deterministic and constant in time. The demand for item 1 is 25 units per week and the demand for item 2
is 40 units per week. First, consider the two items independently of each other and determine for both items
the optimal order quantity, the optimal reorder cycle and the minimum average ordering and holding cost per
week. How much can be saved on this cost by synchronizing the reorder instants of the two items in such
a way that they occasionally coincide? Next, determine the optimal inventory policy in the class of indirect
grouping, that is, determine the optimal family reorder cycle and the optimal reorder frequencies k
1
and k
2
with min¦k
1
, k
2
¦ = 1. Also, compute the corresponding minimum average ordering and holding cost per week.
Exercise 2.3 Consider a family of four items that are supplied by the same distribution center. The family
ordering cost is A = $10, while the supplementary item ordering costs are a
1
= $5, a
2
= $12, a
3
= $12, and
a
4
= $8. The holding costs per unit per month are h
1
= $1.50, h
2
= $2.00, h
3
= $2.00, and h
4
= $1.50. The
monthly demands are D
1
= 60, D
2
= 30, D
3
= 90, and D
4
= 80. Assume deterministic demand. Implement
Goyal’s algorithm and determine the optimal inventory policy.
2.1.2 Heuristic solutions
To avoid the long computation times required by Goyal’s Algorithm 2.1 several heuristics have been developed
to find good solutions for the indirect grouping problem. A distinction is made between one or two step
heuristics and iterative heuristics.
First we will discuss one and two step heuristics. The one-step heuristic of Goyal & Belton [35] first determines
as a reference item the item with the smallest individual cycle length. Then, the frequencies of the other
items are chosen as the optimal frequencies given this cycle length. Finally, the family cycle length is chosen
as the optimal cycle length given these frequencies.
Algorithm 2.2 [Goyal & Belton]
Step 1: Compute for every item i the value of the quotient (A+a
i
)/(h
i
D
i
). The reference item i
r
is the
item with the smallest value. This item has by definition a frequency of 1: k
i
r
.
= 1.
Step 2: Determine an initial family cycle length T
(0)
as the optimal EOQ cycle length of the reference
item i
r
, cf. (1.5), with ordering cost A+a
i
r
.
Step 3: Determine the reorder frequencies of the non-reference items i ,= i
r
by k
i
=
˜
k
i
(T
(0)
), cf. (2.5).
Step 4: Determine the family cycle length T
(1)
as the optimal cycle given the vector of frequencies:
T
(1)
=
ˆ
T(k), cf. (2.2).
The corresponding cost is
ˆ
C(k).
HANS BLANC, TILBURG UNIVERSITY 25
Remark 2.5 This algorithm is an improvement of an algorithm proposed earlier by Silver [59] which de-
termined the reference item on the basis of the values of the quotient a
i
/(h
i
D
i
). The latter is suggested by
(2.4) but it ignores the fact that the reference item, being the item with the smallest individual cycle length,
is the first responsible for the family ordering cost.
Kaspi & Rosenblatt [45] extended the foregoing heuristic because the result of this heuristic strongly depends
on the choice of the reference item. However, the item with the smallest individual cycle length may be an
item with a minor contribution to the total family inventory costs, and still it gets an important influence
on the reorder cycles of the other items. Therefore, a second step is added in which the reorder frequencies
and the family cycle can be modified independently of the reference item.
Algorithm 2.3 [Kaspi & Rosenblatt]
Step 1: Determine an initial policy (T
(1)
, k) with the aid of the heuristic of Goyal & Belton.
Step 2: Determine a better policy by first determining new reorder frequencies as k
i
=
˜
k
i
(T
(1)
) for
i = 1, . . . , N, cf. (2.5), and by then modifying the family cycle to T
(2)
=
ˆ
T(k).
Step 2 might be repeated to obtain an iterative heuristic.
Remark 2.6 A simulation study of Kaspi & Rosenblatt has revealed that the largest improvement occurs
in the first application of Step 2.
Next, we will discuss an iterative heuristic algorithm proposed by Goyal [34]. The general idea of iterative
heuristics is to choose initial reorder frequencies, and alternatingly determine a family cycle length given the
frequencies and new frequencies given the family cycle, until the policy does not change in two subsequent
iterations. Such a procedure may end at a local minimum, or may start oscillating.
The heuristic of Goyal makes use of optimal family cycles
ˇ
T
i
(k) given the reorder frequencies k, without
regard to item i, cf. (2.2),
ˇ
T
i
(k) =
¸
¸
¸
_
2[A+

N
j=1, j=i
(a
j
/k
j
)]

N
j=1, j=i
k
j
h
j
D
j
, i = 1, . . . , N. (2.8)
Algorithm 2.4 [Iterative heuristic Goyal]
Step 1: Choose as initial vector of frequencies k = 1.
Step 2: For i = 1, . . . , N, successively perform the following two operations: compute
ˇ
T
i
=
ˇ
T
i
(k) accord-
ing to (2.8) and determine
˜
k
i
(
ˇ
T
i
) with the aid of (2.5).
Step 3: Stop if the new vector k is the same as the previous one; then, compute the final family cycle
T =
ˆ
T(k) according to (2.2); otherwise, repeat from Step 2.
Alternatively, the result of a one-step heuristic can be used to obtain initial frequencies in Step 1.
Table 2.2: Data for Example 2.3; indirect grouping policies by heuristics.
Item a
i
h
i
D
i
(A+a
i
)/(h
i
D
i
)
˜
k
i
(T
(0)
)
˜
k
i
(T
(1)
)
1 $ 3 $0.50 9 2.0 1 1
2 $ 3 $0.50 4 4.5 1 1
3 $43 $0.50 4 24.5 3 3
Example 2.3 In this example, the foregoing heuristic algorithms are applied to the family of N = 3 items
described in Example 2.1 (Table 2.1). The values of (A+a
i
)/(h
i
D
i
) are displayed in Table 2.2. Clearly, the
reference item is item i
r
= 1. The initial cycle length is T
(0)
= 2, cf. Table 2.1. Then, the Algorithm 2.2 of
Goyal & Belton sets k
1
= 1 and takes k
2
and k
3
as the rounded values of
˜
k
2
(2) =
1
2

3 and
˜
k
3
(2) =
1
2

43, cf.
(2.4), (2.5), respectively. The rounded frequencies are listed in Table 2.2. This heuristic finishes by taking
T = T
(1)
=
ˆ
T(1, 1, 3) = 2.053, and, hence, obtains the optimal policy in this case, cf. Example 2.1. Observe
that the cost contribution of the reference item is not small in comparison with that of the other items in
26 COORDINATED REPLENISHMENT
this case. The additional step of Algorithm 2.3 of Kaspi & Rosenblatt does not change the policy in this
case (T
(2)
= T
(1)
). Algorithm 2.4 starts with k = 1 and computes
ˇ
T
1
(1, 1, 1) =

26,
ˆ
k
1
(

26) =
_
2/39,
˜
k
1
(

26) = 1,
ˇ
T
2
(1, 1, 1) = 4,
ˆ
k
2
(4) =
1
4

3,
˜
k
2
(4) = 1,
ˇ
T
3
(1, 1, 1) = 4
_
3/13,
ˆ
k
3
(4
_
3/13) =
1
4
_
559/3 ≈

11.65,
˜
k
3
(4
_
3/13) = 3, which is the optimal vector of frequencies.
Exercise 2.4 A retailer purchases six items from the same regional warehouse. The joint ordering cost is
A = $10. The supplementary ordering cost a
i
, the holding cost h
i
per unit per month and the constant
demand rate D
i
per month are given in the table below, for items i = 1, . . . , 6.
item i a
i
h
i
D
i
1 $2 $0.20 2000
2 $2 $0.80 500
3 $1 $0.40 1000
4 $2 $0.80 2000
5 $2 $0.20 500
6 $1 $0.40 2000
Determine reorder policies according to the heuristic of Kaspi & Rosenblatt, and according to the iterative
heuristic of Goyal.
2.1.3 Powers-of-two policies
A subclass of the class of indirect grouping policies is formed by the so called powers-of-two policies. In such
a strategy the reorder cycle time of each item is a nonnegative integer power of two times a given basic cycle
or review period R. The objective function to be minimized becomes for this subclass of policies
C(R, k) =
1
R
_
A
k
0
+
N

i=1
a
i
k
i
_
+
1
2
R
N

i=1
k
i
h
i
D
i
; (2.9)
here, k
0
.
= min
i
¦k
i
¦ is not necessarily equal to 1, and the frequencies are restricted to k
i
∈ ¦2

; = 0, 1, 2, . . .¦,
i = 1, . . . , N. Jackson et al. [42] have developed an efficient algorithm for determining optimal reorder
frequencies given the review period R.
Algorithm 2.5 [Powers-of-two policy]
Step 1: Order and renumber the items according to ascending values of a
i
/(h
i
D
i
).
Step 2: Determine i
0
as the largest index i for which
A+

i
j=1
a
j

i
j=1
h
j
D
j

a
i
h
i
D
i
. (2.10)
Step 3: Determine k
0
as the smallest power of two that is larger than or equal to
1
R
¸
¸
¸
_
A+

i
0
j=1
a
j

i
0
j=1
h
j
D
j
. (2.11)
Step 4: Set k
i
= k
0
for i = 1, . . . , i
0
.
Step 5: Determine, for i = i
0
+ 1, . . . , N, k
i
as the smallest power of two that is larger than or equal to
1
R
_
a
i
h
i
D
i
. (2.12)
Optionally, the last three steps can be repeated for various values of R to find a suitable value of R.
HANS BLANC, TILBURG UNIVERSITY 27

C
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
25
26
27
28
29
30
31
32
R →
Figure 2.1: Minimum cost of power-of-two policies as function of the review period R.
Jackson et al. [42] have proved that the inventory cost (2.9) corresponding to the powers-of-two policy
generated by their algorithm cannot exceed the minimum cost for the coordinated replenishment problem by
more than 6%, provided that the quantity in (2.11) is larger than 1 (note that the review period R can be
decreased to satisfy this condition, if necessary). For this purpose they showed that a lower bound for the
minimum cost over any joint replenishment policy is
C

≥ C
lb
.
=
¸
¸
¸
¸
_2
_
_
A+
i
0

j=1
a
j
_
_
i
0

j=1
h
j
D
j
+
N

j=i
0
+1
_
2a
j
h
j
D
j
. (2.13)
Based on the rounding to powers of two in (2.11) and (2.12) it can be shown that for the frequencies k

determined by Algorithm 2.5 it holds that
C(R, k

) ≤ (

2 +
1

2
)
1
2
C
lb

3
4

2 C

≈ 1.06 C

, if R ≤
¸
[A+

i
0
j=1
a
j
]
_

i
0
j=1
h
j
D
j
. (2.14)
Table 2.3: Powers-of-two policies for the family of Example 2.1.
Item a
i
h
i
D
i
a
i
/(h
i
D
i
) Index LHS (2.10) k
i
(R = 1) k
i
(R = 2) k
i
(R = 4)
1 $ 3 $0.50 9 0.667 1 2.000 2 1 1
2 $ 3 $0.50 4 1.500 2 1.846 2 1 1
3 $43 $0.50 4 21.500 3 6.471 8 4 2
Example 2.4 In this example, the Powers-of-two Algorithm 2.5 is applied to the family of N = 3 items
described in Example 2.1 (Table 2.1). The values of a
i
/(h
i
D
i
) are displayed in Table 2.3. Clearly, the items
do not have to be renumbered in this case. The lefthand sides (LHS) of the inequalities (2.10) are also
displayed in Table 2.3. It is seen that i
0
= 2 is the largest index for which (2.10) holds. The frequencies k
i
,
i = 1, 2, 3, are computed by (2.11): k
1
= k
2
= k
0
, and by (2.12) for k
3
, for a review period of R = 1, R = 2
and R = 4, respectively. The lower bound (2.13) has in this example the value C
lb
= 2

39 +2

43 = $25.60.
The 6% guaranty, C(R, k

) < $27.14, holds for R ≤

1.846 ≈ 1.359. The cost C(1, (2, 2, 8)) = $25.88
is guaranteed below this bound, the cost C(2, (1, 1, 4)) = $25.88 is still below this bound, but the cost
C(4, (1, 1, 2)) = $29.38 is not. Figure 2.1 shows the minimum cost of the power-of-two policies as function
of the review period R, its lower bound and the 6% upper bound. The cost function C(R, k), cf. (2.9),
has a minimum for each vector of reorder frequencies. For moderate values of R, the cost function follows
28 COORDINATED REPLENISHMENT
a repetitive pattern with alternatingly a local minimum at $26.52, corresponding to vectors k which are
powers-of-two multiples of the vector (1, 1, 2), and a global minimum at $25.69, corresponding to vectors k
which are powers-of-two multiples of the vector (1, 1, 4). The (scaled) repetitive pattern is due to the fact
that the minimum costs corresponding to the vectors 2k, 4k, 8k, . . . , are attained at values of R which are
1
2
,
1
4
,
1
8
, . . . , of the value of R for which the minimum cost corresponding to the vector k is attained. For
larger values of R, the minimum cost becomes an increasing function. For 2.3 < R < 4.6, the best reorder
frequencies are (1, 1, 2), while for R > 4.6, the best reorder frequencies are (1, 1, 1).
Lee & Yao [48] present some properties of the cost function (2.9) and a search algorithm for the globally
optimal powers-of-two policy in which also the length of the review period R is optimized. An obvious
improvement is to replace an arbitrarily chosen review period R by the optimal cycle length given the reorder
frequencies determined by Algorithm 2.5, cf. (2.2), (2.9):
ˆ
T(k) =
¸
¸
¸
_
2[A/k
0
+

N
j=1
(a
j
/k
j
)]

N
i=1
k
i
h
i
D
i
. (2.15)
Exercise 2.5 Consider again the six-item coordinated replenishment problem described in Exercise 2.4. De-
termine optimal reorder frequencies in the class of powers-of-two policies for review periods R = 0.04,
R = 0.12, and R = 0.24 month, respectively. Compute the average monthly costs corresponding to these
policies, and compare them with the lower bound (2.13) and the upper bound (2.14) on the minimum cost.
2.1.4 Production lot-size problems
A related problem to the coordinated replenishment problem with external supplier is the production lot-size
problem for various items on a single machine. Suppose that N items are produced on the same machine.
Assume that a cyclic policy is applied, that is, there is a family production cycle of length T, and item i
is produced in every k
i
th cycle. This lot-sizing problem is solvable with the foregoing algorithms with the
following modifications, cf. Graves [37]:
• A denotes the major family set-up cost, which accounts for machine opportunity cost during a possible
family set-up time U and for family set-up labor and material cost;
• a
i
denotes the minor set-up cost for item i, i = 1, . . . , N, which accounts for machine opportunity cost
during a possible item set-up time u
i
and for item set-up labor and material cost;
• if the produced items do not become available as a batch at the end of a production run, but at a rate
p
i
for item i during the production run, replace h
i
by h
i
(1 − D
i
/p
i
), i = 1, . . . , N, in all formulas for
the coordinated replenishment problem, as in Section 1.2.2;
• the production time for item i, i = 1, . . . , N, is τ
i
= Q
i
/p
i
= k
i
TD
i
/p
i
plus possibly a set-up time u
i
,
and the maximum inventory level is I
i,max
= (p
i
−D
i
)Q
i
/p
i
, cf. (1.10);
• when the (optimal) cycle length T and the frequencies k
i
have been determined, check whether the
production schedule is feasible,
U +
N

i=1
(u
i

i
) < T, (2.16)
that is, whether the production capacity in the busiest cycle is sufficient;
• optionally, schedule items with k
i
> 1 not simultaneously in the same period (scheduling items with
k
i
> 1 such that cycle lengths for all items are constant is not always trivial or even possible).
A necessary condition for feasibility of a single machine, multi-item lot sizing problem is
N

i=1
D
i
/p
i
< 1. (2.17)
If this condition is not fulfilled, the capacity of the machine is insufficient to satisfy all demand. This condition
does not take into account any set-up times. However, if this condition is fulfilled, there exists a feasible
production schedule with all production frequencies k
i
= 1, i = 1, . . . , N, and a large enough production
cycle T such that the set-up times u
i
are negligible with respect to the production times τ
i
, i = 1, . . . , N.
HANS BLANC, TILBURG UNIVERSITY 29
Table 2.4: Production lot-size data and policies for the family of Example 2.5.
Item a
i
h
i
D
i
p
i
u
i
D
i
/p
i
k
i
τ
i
Q
i
k
i
τ
i
Q
i
1 $10 $0.40 100 400 0.01 0.250 1 0.310 124 1 0.310 124
2 $15 $0.50 150 1000 0.01 0.150 1 0.186 186 1 0.186 186
3 $50 $0.10 175 1000 0.01 0.175 2 0.434 434 2 0.434 434
4 $60 $0.10 250 2000 0.01 0.125 2 0.310 620 2 0.310 620
5 $60 $0.20 50 1000 0.01 0.050 3 0.186 186 4 0.248 248
Example 2.5 Consider a family of five items that has to be manufactured on a single machine. The family
set-up cost is A = $50 and the family set-up time is U = 0.05 week. Item-related data can be found in Table
2.4. First note that the fractions D
i
/p
i
of machine capacity required by the item i add up to 0.75, so that
condition (2.17) is satisfied. Application of Goyal’s Algorithm 2.1 or any of the heuristics described in Section
2.1.2, with h
i
replaced by h
i
(1 − D
i
/p
i
), i = 1, . . . , N, yields the production schedule with k
1
= k
2
= 1,
k
3
= k
4
= 2, k
5
= 3 and T = 1.241 week, with minimum weekly cost C = $241.79. However, this schedule
does not satisfy condition (2.16), because the total production and set-up time in a period when all five
items are produced is 1.527 week, which exceeds the length of a production period T = 1.241 week. But
by producing item 3 in odd periods and item 4 in even periods, the total production and set-up time has a
maximum of 1.207 week in odd periods and 1.083 week in even periods, which are both feasible in a production
period of T = 1.241 week. This production schedule is displayed in Figure 2.2 for a cycle of six periods. A
problem with this schedule is that the intervals between the start of production runs of item 5 are not equal
(to 3T), because the production times of items 3 and 4 are not equal. Delaying the start of the production
run of item 5 in period 4 would require an additional family set-up time U and add a family set-up cost
A = $50 per 6T weeks. A better alternative is to shorten the production run of item 5 in period 1 and to
prolong the production run of item 5 in period 4. Still an other option is to search for powers-of-two policies
which do not exhibit these kind of irregularities. Algorithm 2.5 yields, for a review period of R = 1.241 week,
the production schedule with k
1
= k
2
= 1, k
3
= k
4
= 2, k
5
= 4, and weekly cost C = $243.65 (see also Figure
2.2). This schedule is feasible if items 3 and 4 are produced in different periods and if item 5 is produced in
periods when item 4 is produced; the total production and set-up time is then maximally equal to 1.145.
Period 1
Period 2
Period 3
Period 4
Period 5
Period 6
U
U
U
U
U
U
item 1
item 1
item 1
item 1
item 1
item 1
item 2
item 2
item 2
item 2
item 2
item 2
item 3
item 3
item 3
item 4
item 4
item 4
item 5
item 5
Figure 2.2: Production schedule for five items consisting of a cycle with six periods of length T = 1.241.
An other problem which can be solved with the aid of the algorithms for the coordinated replenishment
problem will be discussed in Section 3.2.2.
Exercise 2.6 For the production schedule displayed in Figure 2.2, compute the modified production times
and lot sizes for item 5 in periods 1 and 4 such that the inventory of this item reaches the level zero at the
start of the production run. What is the increase in cost with respect to the (not realizable) minimum weekly
cost C = $241.79. What would be the actual cost if this schedule is applied with equal lot sizes in periods 1
and 4, implying that the inventory of this item does not reach the level zero at the start of the production run
in period 4?
Exercise 2.7 For the production scheduling problem described in Example 2.5, apply Algorithm 2.5 with a
review period of R = 1.19 week and show that the corresponding cost is less than that with a review period of
R = 1.241 week, while the production schedule can still be made feasible.
30 COORDINATED REPLENISHMENT
Exercise 2.8 Consider a production scheduling problem for three items and a single machine with family
set-up cost $20 and family set-up time 0.02 month. There are no item set-up times. The minor set-up cost
a
i
, the holding cost h
i
per unit per month, the demand rate D
i
per month and the production rate p
i
per
month, i = 1, 2, 3, are given in the table below.
item i a
i
h
i
D
i
p
i
1 $5 $0.40 400 1000
2 $25 $0.10 600 2000
3 $60 $0.25 50 500
Determine a production schedule with the heuristic of Kaspi & Rosenblatt and verify whether this schedule
can be made feasible. Search for a review period with which Algorithm 2.5 yields a production schedule that
can be made feasible.
2.1.5 All-units discounts
In Section 1.3 it has been noted that beside joint ordering costs also discounts may induce a company to
coordinate replenishments of items that are ordered from the same supplier. In this section, indirect grouping
for coordinated replenishments is considered for the case of joint ordering cost and a special type of discounts
known as all-units discounts. More specifically, we will consider the case of an item-independent discount
factor d, 0 < d < 1. This means that the purchasing price per unit of item i is as function of the family
reorder cycle T and the reorder frequencies k, for i = 1, . . . , N,
v
i
(T, k) = v
i0
, if

N
j=1
Q
j
v
j0
= T

N
j=1
k
j
D
j
v
j0
< V,
v
i
(T, k) = v
i0
(1 −d), if

N
j=1
Q
j
v
j0
= T

N
j=1
k
j
D
j
v
j0
≥ V ;
(2.18)
here, v
i0
stands for the basic purchasing price of item i (without discount), Q
i
= k
i
TD
i
is the order quantity
of item i at a certain reorder instant, i = 1, . . . , N, and V is a price break point for obtaining discount. Since
the purchasing costs are no longer constant, they have to be included in the objective function. The holding
costs are represented as h
i
= rv
i
(T, k), i = 1, . . . , N. Hence, the objective function becomes, with family
cycle T, ordering frequencies k and v
i
(T, k) defined by (2.18),
C(T, k) =
1
T
_
A+
N

i=1
a
i
k
i
_
+
1
2
T
N

i=1
rv
i
(T, k)D
i
k
i
+
N

i=1
v
i
(T, k)D
i
. (2.19)
Silver & Peterson developed an heuristic for indirect grouping of items under the above described discount
structure. They make the simplifying assumption that the discount is only awarded if the value of all orders
exceeds the price break point V . Under this assumption, the crucial reorder instants are those at which
only items i with k
i
= 1 are ordered. The heuristic is based on the observation that the reorder frequencies
obtained by Algorithm 2.2 do not depend on the discount factor d.
Algorithm 2.6 [Silver & Peterson]
Step 1: Apply Algorithm 2.2 (heuristic of Goyal & Belton) with h
i
= rv
i0
(1 −d), i = 1, . . . , N. Call the
resulting policy (T
(1)
d
, k). Let ¸
.
= ¦j; k
j
= 1¦. Verify whether
T
(1)
d

j∈J
D
j
v
i0
≥ V. (2.20)
If this condition is satisfied, the value (without discount) of all orders exceeds the price break point V ,
and the current policy (T
(1)
d
, k) is feasible and such that the discount is obtained. If this condition is
not satisfied, continue with the next steps.
Step 2: Increase the family cycle length T
(1)
d
to the value T
(V )
d
which is minimally required for obtaining
discount, and, hence, is determined by:
T
(V )
d

j∈J
D
j
v
i0
= V. (2.21)
Compute the corresponding cost C(T
(V )
d
, k), cf. (2.19) with v
i
(T
(V )
d
, k) = v
i0
(1 −d), i = 1, . . . , N.
HANS BLANC, TILBURG UNIVERSITY 31
Step 3: Redo Step 4 of Algorithm 2.2 with h
i
= rv
i0
, i = 1, . . . , N, to obtain cycle length T
(1)
. Compute
the corresponding cost C(T
(1)
, k), cf. (2.19) with v
i
(T
(1)
, k) = v
i0
, i = 1, . . . , N.
Step 4: Use the policy of Step 2 if C(T
(V )
d
, k) < C(T
(1)
, k), and use the policy of Step 3 otherwise.
The algorithm can be extended to situations with several discount factors and corresponding price break
points, cf. Section 1.2.3.
Remark 2.7 If there are several items i with k
i
> 1, it may be possible to divide the reorder instants of
these items such that the items in the above defined set ¸ are never ordered alone. Then, the set ¸ can
be enlarged in the foregoing algorithm, and the discount price break point is more easily reached. However,
when N is large there are many ways to divide these items. Therefore, this possibility is not considered in
the algorithm.
Example 2.6 Suppose that the holding costs in Example 2.1 stem from a carrying charge of r = 0.002 per
dollar inventory per week times a purchasing price of $250 for all three items. Further, suppose that the
supplier of these items offers an all-units discount of d = 0.04 if the value of each order exceeds V = $10,000.
This means that the discount purchasing price is $240 per item. Application of Algorithm 2.2 with h
i
= $0.48,
i = 1, 2, 3, yields the same frequencies as with h
i
= $0.50 (Example 2.3), namely k = (1, 1, 3), but a cycle
length T
(1)
d
= 2.095. The set of items with smallest order frequency is ¸ = ¦1, 2¦. The value of the smallest
orders is 250 (9 + 4) T
(1)
d
= $6,808.64, cf. (2.20). Hence, this policy is not feasible. The cycle length
that is minimally required for obtaining discount is T
(V )
d
= 3.077, cf. (2.21), and the corresponding cost is
C(T
(V )
d
, k) = $4,107.02 per week, cf. (2.19). The cycle length according to Algorithm 2.2 with h
i
= $0.50,
i = 1, 2, 3, is T
(1)
= 2.053. The corresponding cost is C(T
(1)
, k) = $4,275.66 per week. Hence, the policy
proposed by Algorithm 2.6 is k = (1, 1, 3), and T = T
(V )
d
= 3.077.
Exercise 2.9 Consider the retailer and the six items from Exercise 2.4. Suppose that the holding costs stem
from a carrying charge of r = $0.02 per dollar per month times the basic purchasing price v
0i
of item i,
i = 1, . . . , 6. Assume that the regional warehouse offers an all-units discount of d = 0.05 above the price
break point V = $20, 000 provided that the value of all orders exceeds this price break point. Determine an
inventory policy with the aid of Algorithm 2.6.
2.2 Constant demand, direct grouping
In this section we consider the same problem of a family of N items with joint ordering cost structure and
constant demand as in Section 2.1 but for the subclass of direct grouping policies. In this case, the N items
are divided into M disjunct groups (
j
, j = 1, . . . , M, and all members of group (
j
have the same reorder
cycle T
j
, j = 1, . . . , M. The reorder cycles of the various groups are determined independently of each other.
Reorder instants of different groups may accidentally coincide, but such opportunities are not taken into
account in the cost function.
The decision variables for the direct grouping problem are
• M: the number of groups;
• (
j
: the items in the jth groups, j = 1, . . . , M;
• T
j
: the group reorder cycle for items in (
j
, j = 1, . . . , M.
The objective function to be minimized is, with T
.
= (T
1
, . . . , T
M
),
C((
1
, . . . , (
M
, T) =
M

j=1
_
_
1
T
j
_
_
_
A+

i∈G
j
a
i
_
_
_
+
1
2
T
j

i∈G
j
h
i
D
i
_
_
. (2.22)
The optimal group cycles given the grouping ((
1
, . . . , (
M
) are readily determined in the standard way:
ˆ
T
j
((
j
) =
¸
¸
¸
_
2[A+

i∈G
j
a
i
]

i∈G
j
h
i
D
i
, j = 1, . . . , M. (2.23)
32 COORDINATED REPLENISHMENT
The corresponding minimum average inventory cost per unit of time for a given grouping (without accidentally
coinciding orders) is:
ˆ
C((
1
, . . . , (
M
) =
M

j=1
¸
¸
¸
¸
_2
_
_
A+

i∈G
j
a
i
_
_

i∈G
j
h
i
D
i
. (2.24)
The number of possible groupings strongly increases with the size N of the family. A reduction of the number
of groupings that have to be considered for optimization is possible due to the so called consecutiveness
property (cf. Chakravarty [14]):
• Renumber the items according to ascending value of a
i
/(h
i
D
i
).
• Then, the optimal groups (
1
, . . . , (
M
form ordered sets, that is, if item a and item b belong to the same
group, say (
j
, then item i belongs to (
j
for all i, a < i < b.
Remark 2.8 The ordering described by the consecutiveness property also appears in the optimal indirect
grouping, cf. (2.4), and in the Powers-of-two Algorithm 2.5.
Example 2.7 A collection of consecutive sets is, for instance:
(
1
= ¦1, 2, 3¦, (
2
= ¦4¦, (
3
= ¦5, 6, 7, 8¦, (
4
= ¦9, 10¦.
Put on a row, the numbers 1,. . . ,10 are counted off.
2.2.1 Exact solution by dynamic programming
After renumbering the items according to the consecutiveness property, the optimal grouping can be deter-
mined by dynamic programming, cf. Chakravarty & Goyal [17]. In each step of the recursion, an additional
item is considered. All possible groupings of this item with previously considered items satisfying the con-
secutiveness property are evaluated, while the optimal grouping of the remaining items has already been
determined. To formulate the dynamic programming recursion define:
• f(n): the minimum cost after the grouping of n items;
• : the number of items in the same group with item n.
The initial value is f(0) = 0, and the recursion reads, using one term of the minimum cost determined in
(2.24), for n = 1, . . . , N,
f(n) = min
1≤≤n
_
_
_
f(n −) +
¸
¸
¸
_
2
_
A+
n

i=n−+1
a
i
_
n

i=n−+1
h
i
D
i
_
_
_
. (2.25)
In spite of the reduction realized by the consecutiveness property, the computation time is still strongly
increasing with N.
Example 2.8 In this example, the dynamic programming approach is applied to the family of N = 3 items
described in Example 2.1. As showed in Example 2.4, the items do not have to be renumbered to satisfy the
consecutiveness property. The algorithm starts with f(0) = 0, then considers item 1:
f(1) =
_
2[A+a
1
] h
1
D
1
= $9.00.
Next, it determines whether it is better to take items 1 and 2 apart or together:
f(2) = min
_
f(1) +
_
2[A+a
2
] h
2
D
2
= $15.00.
_
2[A+a
1
+a
2
] [h
1
D
1
+h
2
D
2
] = $12.49.
It turns out that it is advantageous to join items 1 and 2. Finally, item 3 is considered:
f(3) = min
_
_
_
f(2) +
_
2[A+a
3
] h
3
D
3
= $26.49.
f(1) +
_
2[A+a
2
+a
3
] [h
2
D
2
+h
3
D
3
] = $29.40.
_
2[A+a
1
+a
2
+a
3
] [h
1
D
1
+h
2
D
2
+h
3
D
3
] = $30.58.
HANS BLANC, TILBURG UNIVERSITY 33
Hence, it follows that the optimal direct grouping of this family is a division into two groups: (
1
= ¦1, 2¦,
(
2
= ¦3¦, with minimum costs per week C

= $26.49 and with group cycles found with (2.23) as T
1
= 1.92
and T
2
= 7. Observe that the cycle length of group (
1
is smaller than the optimal individual cycle lengths
of items 1 and 2 (Table 2.1). Further, note that for this example the minimum cost with direct grouping is
higher than the minimum cost with indirect grouping (Example 2.1).
In Example 2.11 it is shown that the optimal grouping of items may follow an irregular pattern as function
of the family ordering cost A.
Exercise 2.10 Determine the optimal direct grouping of the family of two items described in Example 2.2.
2.2.2 Heuristic solution
Bastian [11] has developed an heuristic algorithm for the direct grouping of items in a family.
Algorithm 2.7 [Bastian]
Step 1: Renumber the items according to ascending value of a
i
/(h
i
D
i
).
Step 2: Start with M = N groups consisting of 1 item each.
Step 3: Determine for each pair of neighboring groups the decrease in cost (saving) when they would be
joined, cf. (2.24), that is, compute for j = 1, . . . , M −1,
¸
¸
¸
¸
_2
_
_
A+

i∈G
j
a
i
_
_

i∈G
j
h
i
D
i
+
¸
¸
¸
¸
_2
_
_
A+

i∈G
j+1
a
i
_
_

i∈G
j+1
h
i
D
i

¸
¸
¸
¸
_2
_
_
A+

i∈G
j
∪G
j+1
a
i
_
_

i∈G
j
∪G
j+1
h
i
D
i
.
Step 4: Actually join the two groups with the largest saving, decrease M by 1 and repeat from Step 3
if the largest saving is positive; otherwise, stop.
Step 5: Finally, determine the optimal reorder cycles for the established groups with the aid of (2.23).
The algorithm is readily adapted to situations in which the number of groups is prescribed; then, the algorithm
stops if the desired number of groups has been reached, possibly while there still is positive saving or after a
few steps with increasing cost.
The heuristic of Bastian is a greedy algorithm: always accept the largest saving. Note that the cost difference
in Step 3 only has to be computed in the second and later iterations as far as the group is involved that has
been formed in the previous iteration.
Example 2.9 In this example, the heuristic of Bastian is applied to the family of N = 3 items described in
Example 2.1. As noted in Example 2.8, no renumbering is required. The algorithm starts with three groups:
(
1
= ¦1¦, (
2
= ¦2¦ and (
3
= ¦3¦. The saving of joining the first two groups is 9 + 6 − 12.49 = $2.51, that
of joining the latter two groups is 6 + 14 − 20.40 = −$0.40. Since only the saving of joining the first two
groups is positive, those two groups are actually joined, and the new groups are (
1
= ¦1, 2¦, (
2
= ¦3¦. In
the final step the saving of joining these two groups is computed: 12.49 + 14 − 30.58 = −$4.09. Since this
saving is negative, the groups are not joined. In this example, the heuristic of Bastian finds the optimal
direct grouping, and the reorder cycles are the same as in Example 2.8.
Exercise 2.11 A company purchases three items a, b and c from the same supplier. With every order, a fixed
amount of $5 has to be paid. Further, item dependent ordering costs are incurred: $30 for item a, $24 for
item b, and $50 for item c. The demand for these items is deterministic and constant in time. The demand
per year is 4800 units for item a, 3600 units for item b, and 2500 for item c. The holding costs are equal for
the three items, and amount to $20 per unit per year. First, determine for each item separately the optimal
order quantity, the optimal reorder cycle and the minimum cost per year. Next, determine the optimal direct
grouping of these items with the aid of dynamic programming; also, compute the optimal reorder cycles and
the corresponding minimum cost per year. Finally, determine a direct grouping of these items with the aid of
the heuristic of Bastian.
34 COORDINATED REPLENISHMENT
2.2.3 Direct grouping with discounts
Consider the all-units discount structure described in Section 2.1.5. A discount factor d is awarded to the
purchasing price of all items in an order if the value of the order exceeds a certain price break point V . With
direct grouping the assumption is that the discount can be awarded per group.
An important observation is that the consecutiveness property remains valid. The optimal direct grouping
policy can still be determined with the aid of dynamic programming, but the recursion becomes more complex
than in Section 2.2.1, cf. Chakravarty [15]. With f(n) the minimum cost after grouping of n items, including
purchasing cost, the recursion becomes: for n = 1, . . . , N,
f(n) = min
1≤≤n
¦f(n −) +G(n, )¦; (2.26)
here, G(n, ) is the minimum cost of the group consisting of the items n − + 1, . . . , n. When determining
G(n, ), again three cases have to be distinguished (cf. Section 2.1.5):
• the globally optimal cycle length with discount; if this is not feasible then:
• the globally optimal cycle length without discount;
• and the price break point cycle length to obtain discount.
The application of the dynamic programming recursion (2.26) requires the repeated solution of these sub-
problems. Combining the concepts behind (2.22) and (2.19) the average cost of the group consisting of the
items n − + 1, . . . , n are with group reorder cycle T:
G(n, , T) =
1
T
_
A+
n

i=n−+1
a
i
_
+
1
2
T
n

i=n−+1
rv
i
(T)D
i
+
n

i=n−+1
v
i
(T)D
i
; (2.27)
here, the purchasing price as function of the cycle length is: for i = 1, . . . , N,
v
i
(T) = v
i0
, if T

n
j=n−+1
D
j
v
j0
< V,
v
i
(T) = v
i0
(1 −d), if T

n
j=n−+1
D
j
v
j0
≥ V.
(2.28)
The globally optimal cycle lengths given that the group consists of items n − + 1, . . . , n follow by (2.23).
Example 2.10 The dynamic programming approach is applied to the family of N = 3 items described
in Example 2.1 with the discount structure of Example 2.6. The algorithm starts with f(0) = 0, then
considers item 1. With h
1
= $0.48, the optimal individual cycle length is T = 2.041, and the value of the
corresponding order quantity, $4,592.79, is not feasible. The minimal cycle length for obtaining discount is
T = V/(v
10
D
1
) = 4.444, and the corresponding average cost is C = $2,171.63. With h
1
= $0.50, the optimal
individual cycle length is T = 2, and the corresponding average cost is C = $2,259.00. Hence, the minimal
cost for item 1 is
f(1) = G(1, 1) = $2,171.63.
Next, a similar analysis has to be performed for item 2 to compute G(2, 1), and for items 1 and 2 together to
determine G(2, 2). First, consider item 2. With h
2
= $0.48, the optimal individual cycle length is T = 3.062,
and the value of the corresponding order quantity, $3,061.86, is not feasible. The minimal cycle length
for obtaining discount is T = V/(v
20
D
2
) = 10, and the corresponding average cost is C = $970.50. With
h
2
= $0.50, the optimal individual cycle length is T = 3, and the corresponding average cost is C = $1,006.00.
Hence, G(2, 1) = $970.50. For items 1 and 2 together, the optimal cycle length with h
1
= h
2
= $0.48 is
T = 1.961, and the value of the corresponding order quantities, $6,373.77, is not feasible. The minimal cycle
length for obtaining discount is T = V/(v
10
D
1
+ v
20
D
2
) = 3.077, and the corresponding average cost is
C = $3,133.50. With h
1
= h
2
= $0.50, the optimal cycle length is T = 1.922, and the corresponding average
cost is C = $3,262.49. Hence, G(2, 2) = $3,133.50. Then, the second dynamic programming step becomes
f(2) = min
_
f(1) +G(2, 1) = $3,142.12.
G(2, 2) = $3,133.50.
It is seen that it is best to join items 1 and 2 in one group. For the third step, the quantities G(3, 1), G(3, 2)
and G(3, 3) have to be determined in a similar way. This is left to the reader.
Exercise 2.12 Compute f(3) for the family of three items considered in Example 2.10.
HANS BLANC, TILBURG UNIVERSITY 35

C
0 5 10 15 20 25
2300
2400
2500
2600
2700
2800
2900
3000
3100
3200
3300
individual
indirect grouping
direct grouping
A →
Figure 2.3: Minimum costs of individual, indirect grouping and direct grouping policies as functions of A.
2.2.4 Indirect versus direct grouping
Indirect grouping and direct grouping are two subclasses of the class of all possible policies for coordinated
replenishments. The two extremal policies in this class are:
• order all items separately;
• order all items together.
All other policies are called mixed policies. The subclass of indirect grouping includes joint ordering of all
items; the subclass of direct grouping includes both extremal policies. There exists no analytic way for
determining which policy (from which subclass) is optimal in which situation. Therefore, simulation studies
are required to compare the best policies in various subclasses. Here, simulation means that the parameters
of a large number of inventory problems are sampled from some probability distributions (the problems are
deterministic). It has turned out that the most important influence on the performance of the subclasses
stems from the number of items in the family, N, and the ratio between the family ordering costs and the
average individual ordering costs A/¯ a, with ¯ a
.
=
1
N

N
i=1
a
i
. In general, indirect grouping gives higher savings
with respect to separate ordering, between 30% and 70% with increasing ratio A/¯ a ≥ 1. Direct grouping is
only better if A ¸ ¯ a but then the savings are generally less than 30%. See further Van Eijs et al. [73]. See
also Silver et al. [64, Ch. 11].
Example 2.11 Figure 2.3 shows the minimum costs of individually optimal policies without synchronization,
of the optimal indirect grouping policies according to Algorithm 2.1 of Goyal, and of the optimal direct
grouping policies according to dynamic programming as discussed in Section 2.2.1, for a family of N = 8
items, of which the data are summarized in Table 2.5, as functions of the family ordering cost A. Table
2.5 also contains the individually optimal reorder cycles T
i
(0), i = 1, . . . , 8 for A = 0. The optimal direct
grouping policy consists of 8 single-item groups for A = 0; in the range from A = 0.1 to A = 1.2, items 2, 4
and 5 have joined a group, with a reorder cycle of about 0.15, and items 1 and 3 have joined a group, with a
reorder cycle of about 0.18; from A = 1.25 to A = 2.1, items 1, 2, 3, 4 and 5 have joined one group, with a
reorder cycle of about 0.16; from A = 2.2 to A = 2.7, items 2, 4, 5 and 6 have formed a group, with a reorder
cycle of about 0.13, and items 1 and 3 have split off as a group, with a reorder cycle of about 0.19; from
A = 2.8 to A = 7.3, these two groups have merged into one group, with a reorder cycle of about 0.14; from
A = 7.4 to A = 34, items 7 and 8 have formed another group, with a reorder cycle of about 0.41; from A = 35
to A = 70, item 7 has made a transition to the large group, with a reorder cycle of about 0.17, while item 8
has formed a group on its own. From A = 71 onwards, also item 8 has joined the large group. Table 2.5 also
contains the optimal reorder cycles T
i
(A), i = 1, . . . , 8 with direct grouping for A = 0, 1, 2, 2.5, 5, 10, 50, 100.
36 COORDINATED REPLENISHMENT
The optimal indirect grouping policy has reorder frequencies k
1
= k
2
= k
3
= k
4
= k
5
= k
6
= 1, k
7
= 2 and
k
8
= 4 in a broad range from A = 0 to about A = 48.4 where the reorder cycle increases from 0.143 to about
0.165; in the range from A = 48.4 to about A = 189.4, the only change is the reorder frequency of item 8:
k
8
= 3; above A = 189.4, the reorder frequency of item 7 has decreased: k
7
= 1; in the range from A = 216.7
to just over A = 1000, the reorder frequency of item 8 has further decreased: k
8
= 2; for A > 1000.2, all
reorder frequencies are equal to 1. Figure 2.3 shows that the optimal indirect grouping policy is worse than the
individually optimal policies for A smaller than about 0.8, and worse than the optimal direct grouping policy
for A smaller than about 1.6. For this family of items, direct grouping is optimal for A smaller than about
1.6, and otherwise indirect grouping is optimal (for A > 1000.2, the optimal direct grouping policy is the
same as the optimal indirect grouping policy). The difference between the individually optimal policies and
the optimal policies with coordination, both indirect and direct grouping, strongly increases with increasing
value of the family ordering cost A.
Table 2.5: Data for Example 2.11; optimal reorder cycles for direct grouping policies.
i a
i
h
i
D
i
a
i
/(h
i
D
i
) T
i
(0) T
i
(1) T
i
(2) T
i
(2.5) T
i
(5) T
i
(10) T
i
(50) T
i
(100)
1 $15 $0.20 4000 0.019 0.194 0.184 0.157 0.188 0.144 0.147 0.175 0.205
2 $20 $1.00 2000 0.010 0.141 0.146 0.157 0.134 0.144 0.147 0.175 0.205
3 $22 $0.40 3600 0.015 0.175 0.184 0.157 0.188 0.144 0.147 0.175 0.205
4 $25 $0.80 3000 0.010 0.144 0.146 0.157 0.134 0.144 0.147 0.175 0.205
5 $28 $1.00 2500 0.011 0.150 0.146 0.157 0.134 0.144 0.147 0.175 0.205
6 $30 $1.50 3200 0.006 0.112 0.114 0.115 0.134 0.144 0.147 0.175 0.205
7 $35 $1.50 500 0.047 0.306 0.310 0.314 0.316 0.327 0.412 0.175 0.205
8 $40 $2.50 100 0.160 0.566 0.573 0.580 0.583 0.600 0.412 0.849 0.205
Exercise 2.13 Consider the family of eight items described in Example 2.11. Compare the policies and
corresponding costs generated by the heuristic of Kaspi & Rosenblatt (algorithm 2.3) and the iterative heuristic
of Goyal (algorithm 2.4) for A = 5.
2.3 Time-varying demand
In this section a multi-item generalization of the model of Section 1.2.5 with time-varying deterministic
demand will be discussed. Consider a family of N items and a finite planning horizon of H periods. The
demand D
i,t
for item i in period t is assumed to be known (i = 1, . . . , N, t = 1, . . . , H). The cost factors
are constant over the planning horizon. As in Sections 2.1, 2.2, A denotes the family ordering cost, and a
i
denotes the supplementary ordering cost for item i, i = 1, . . . , N. Orders are delivered at the beginning of
a period. At the beginning of each period the demand for that period is set aside, and holding costs of h
i
dollar per unit of item i per period are charged over the excess inventories. All demand must be fulfilled and
no back orders are allowed. The decision variables are Q
i,t
, the order quantity of item i for the beginning of
period t (i = 1, . . . , N, t = 1, . . . , H). The inventory level of item i at the end of period t is denoted by I
i,t
(i = 1, . . . , N, t = 0, . . . , H). It is assumed that all initial inventory levels are zero: I
i,0
= 0, i = 1, . . . , N.
The inventory levels can be described by the recursive equations; for i = 1, . . . , N,
I
i,t
= I
i,t−1
+Q
i,t
−D
i,t
, t = 1, . . . , H. (2.29)
These inventory levels must satisfy the requirements I
i,t
≥ 0, i = 1, . . . , N, t = 1, . . . , H. The objective is to
minimize the total cost of ordering and stock-keeping over H periods:
C(Q
1
, . . . , Q
H
) =
H

t=1
_
Aδ(

N
i=1
Q
i,t
) +
N

i=1
¦a
i
δ(Q
i,t
) +h
i
I
i,t
¦
_
, (2.30)
here, Q
t
.
= (Q
1,t
, . . . , Q
N,t
) is the vector of order quantities for period t, and δ(.) is a dummy function, cf.
(1.25). As in the single-item case the minimization of the cost function is facilitated by the observation that
the optimal order quantities are zero or equal to the demand of a whole number of future periods. This
observation implies the following properties of the optimal solution. For i = 1, . . . , N, t = 1, . . . , H,
1. item i is only ordered for the beginning of period t if the inventory has reached the level zero (I
i,t−1
= 0);
HANS BLANC, TILBURG UNIVERSITY 37
2. the order quantity Q
i,t
is zero or equal to the demand in a whole number of future periods, that is, it
is restricted to the set ¦0, D
i,t
, D
i,t
+D
i,t+1
, . . . , D
i,t
+ +D
i,H
¦;
3. the inventory level I
i,t−1
also is zero or equal to the demand in a whole number of future periods, that
is, it is restricted to the same set ¦0, D
i,t
, D
i,t
+D
i,t+1
, . . . , D
i,t
+ +D
i,H
¦;
4. the demand D
i,t+
is not included in the order in period t if h
i
D
i,t+
> A+a
i
(then, to order in period
t +, if need be only item i, is cheaper than keeping D
i,t+
periods in stock).
2.3.1 Exact solution
On the basis of the foregoing properties the minimization problem can be reformulated as an integer problem.
Instead of the decision vectors Q
t
and the similarly defined state vectors I
t
we consider integer decision vectors
K
t
and integer state vectors J
t
with components: for i = 1, . . . , N, t = 1, . . . , H,
• K
i,t
: the order size for item i for the beginning of period t as number of periods demand for item i;
• J
i,t
: the inventory level of item i at the end of period t as number of periods demand for item i.
The relevant values of these quantities vary with t, and are interrelated. Define: for t = 1, . . . , H,
• o
t
: the relevant state space at the end of period t −1 consisting of vectors J
t
with 0 ≤ J
i,t
≤ H−t +1,
i = 1, . . . , N, and J
i,t
= 0 for at least one item i (only reorder instants are relevant);
• /
t
(J
t
): the relevant action space for the order at the beginning of period t given the inventory levels
J
t
; this space consists of vectors K
t
with K
i,t
= 0 if J
i,t
> 0, and 0 < K
i,t
≤ H − t + 1 if J
i,t
= 0,
i = 1, . . . , N.
Note that o
1
= ¦0¦. Further restrictions on the above spaces may be possible by property 4.
Table 2.6: State spaces (vectors) for the case N = 3, H = 4.
o
4
: 000 100 010 001 110 101 011
o
3
: 000 100 010 001 110 101 011 200 020 002 210 201 021 120 102 012
220 202 022
o
2
: 000 100 010 001 110 101 011 200 020 002 210 201 021 120 102 012
220 202 022 300 030 003 310 301 031 130 103 013
320 302 032 230 203 023 330 303 033
o
1
: 000
Example 2.12 For the case N = 3, H = 4, the state spaces are displayed in Table 2.6. No states have been
filtered out on the basis of property 4 because that step requires knowledge of the parameters of the model.
As an example of an action space, consider /
2
(2, 0, 0). Since an item is only ordered if its inventory level is
zero, we have K
1
= 0. The inventory levels of items 2 and 3 are zero, so they must be ordered. For both
items, the options are to order for 1, 2 or 3 periods. Hence, /
2
(2, 0, 0) consists of 9 possible actions.
The optimal solution can be determined with the aid of dynamic programming. To this end, define: for
t = 1, . . . , H,
• f
t
(J): the minimum cost incurred during periods t, . . . , H starting from state J ∈ o
t
at the end of
period t −1;
• C
t,
(J, K): the cost incurred during periods t, . . . , t + − 1 given that the inventory levels at the end
of period t −1 are J ∈ o
t
and given that the action K ∈ /
t
(J
t
) is taken:
C
t,
(J, K) = A+
N

i=1
_
a
i
δ(K
i
) +h
i
J
i
+K
i
−1

u=1
min¦u, ¦D
i,t+u
_
; (2.31)
here, the next reorder instant t + is determined by

.
= min
i=1,...,N
¦J
i
+K
i
¦. (2.32)
38 COORDINATED REPLENISHMENT
Note that an item i with J
i
+K
i
> is not ordered in period t + and the demand for the periods t +, . . .
lies all periods on stock.
Silver [60] has developed a generalization of the Wagner-Whitin dynamic programming solution of 1-item
problems to the case of a family of N items.
Algorithm 2.8 [Silver dynamic programming]
Step 1: Determine for each t all relevant states J
t
∈ o
t
(J
1
= 0) and all relevant actions K
t
∈ /
t
(J
t
).
Step 2: Set f
T+1
(0) = 0 and start with t = H.
Step 3: Apply the backward recursion: for J ∈ o
t
,
f
t
(J) = min
K∈A
t
(J
t
)
¦f
t+
(J +K−1) +C
t,
(J, K)¦ ; (2.33)
Step 4: If t > 1 decrease t by 1 and repeat from Step 3; otherwise, stop.
Finally, the optimal ordering scheme has to be interpreted (the integer vectors K
t
have to be translated back
to order quantities Q
t
).
Example 2.13 Consider the case N = 2, H = 3. Then for t = H = 3 there is no other choice than to order
the lacking items, and only ordering costs are involved:
f
3
(0, 0) = A+a
1
+a
2
,
f
3
(1, 0) = A+a
2
,
f
3
(0, 1) = A+a
1
.
For t = H −1 = 2 there is some more choice, depending on the inventory levels:
f
2
(0, 0) = min
_
¸
¸
_
¸
¸
_
C
2,2
((0, 0), (2, 2)) = A+a
1
+a
2
+h
1
D
1,3
+h
2
D
2,3
,
f
3
(0, 1) +C
2,1
((0, 0), (1, 2)) = f
3
(0, 1) +A+a
1
+a
2
+h
2
D
2,3
f
3
(1, 0) +C
2,1
((0, 0), (2, 1)) = f
3
(1, 0) +A+a
1
+a
2
+h
1
D
1,3
,
f
3
(0, 0) +C
2,1
((0, 0), (1, 1)) = f
3
(0, 0) +A+a
1
+a
2
;
f
2
(1, 0) = min
_
f
3
(0, 1) +C
2,1
((1, 0), (0, 2)) = f
3
(0, 1) +A+a
2
+h
2
D
2,3
,
f
3
(0, 0) +C
2,1
((1, 0), (0, 1)) = f
3
(0, 0) +A+a
2
;
f
2
(2, 0) = min
_
C
2,2
((2, 0), (0, 2)) = A+a
2
+h
1
D
1,3
+h
2
D
2,3
,
f
3
(1, 0) +C
2,1
((2, 0), (0, 1)) = f
3
(1, 0) +A+a
2
+h
1
D
1,3
;
f
2
(0, 1) = min
_
f
3
(1, 0) +C
2,1
((0, 1), (2, 0)) = f
3
(1, 0) +A+a
1
+h
1
D
1,3
,
f
3
(0, 0) +C
2,1
((0, 1), (1, 0)) = f
3
(0, 0) +A+a
1
;
f
2
(0, 2) = min
_
C
2,2
((0, 2), (2, 0)) = A+a
1
+h
1
D
1,3
+h
2
D
2,3
,
f
3
(0, 1) +C
2,1
((0, 2), (1, 0)) = f
3
(0, 1) +A+a
1
+h
2
D
2,3
.
Finally, for t = 1 there is only one relevant state but many options:
f
1
(0, 0) = min
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
C
1,3
((0, 0), (3, 3)) = A+a
1
+a
2
+h
1
D
1,2
+ 2h
1
D
1,3
+h
2
D
2,2
+ 2h
2
D
2,3
,
f
3
(0, 1) +C
1,2
((0, 0), (2, 3)) = f
3
(0, 1) +A+a
1
+a
2
+h
1
D
1,2
+h
2
D
2,2
+ 2h
2
D
2,3
,
f
3
(0, 1) +C
1,2
((0, 0), (3, 2)) = f
3
(0, 1) +A+a
1
+a
2
+h
1
D
1,2
+ 2h
1
D
1,3
+h
2
D
2,2
,
f
2
(0, 2) +C
1,1
((0, 0), (1, 3)) = f
2
(0, 2) +A+a
1
+a
2
+h
2
D
2,2
+h
2
D
2,3
,
f
2
(2, 0) +C
1,1
((0, 0), (3, 1)) = f
2
(2, 0) +A+a
1
+a
2
+h
1
D
1,2
+h
1
D
1,3
,
f
3
(0, 0) +C
1,2
((0, 0), (2, 2)) = f
3
(0, 0) +A+a
1
+a
2
+h
1
D
1,2
+h
2
D
2,2
,
f
2
(0, 1) +C
1,1
((0, 0), (1, 2)) = f
2
(0, 1) +A+a
1
+a
2
+h
2
D
2,2
,
f
2
(1, 0) +C
1,1
((0, 0), (2, 1)) = f
2
(1, 0) +A+a
1
+a
2
+h
1
D
1,2
,
f
2
(0, 0) +C
1,1
((0, 0), (1, 1)) = f
2
(0, 0) +A+a
1
+a
2
.
Note that for the action K = (3, 3) the holding costs are charged over all three periods so that the demands
D
i,3
incur a holding cost of 2h
i
, i = 1, 2. For the action K = (3, 2) the holding costs are charged over two
periods; this means that the demand D
1,3
incurs a holding cost of 2h
1
. For the action K = (3, 1) the next
reorder instant is t = 2 ( = 1) so that the holding costs are only charged over a single period. Therefore, the
demand D
1,3
incurs a holding cost of h
1
over period 1; the holding cost over period 2 is included in f
2
(2, 0).

HANS BLANC, TILBURG UNIVERSITY 39
Remark 2.9 The application of coordinated replenishment with time-varying demand with a rolling horizon
is not as straightforward as in the 1-item case, cf. Remark 1.4. The foregoing dynamic programming solution
and also the heuristics that will be discussed in the next section assume that the inventory levels of all items
before the beginning of the first period are zero. The solution of the problem over H periods may contain
intermediate reorder instants at which only a few items are ordered, and perhaps none at which all items are
ordered. This requires modifications of the algorithms. Also, cases with zero demand of some items in the
first (few) periods requires modification of the ordering cost: δ(K
i
) in (2.31) should be read as zero if the
action K
i
only concerns periods of zero demand of item i, i = 1, . . . , N.
2.3.2 Heuristic algorithms
The exact solution discussed in the previous section is very time consuming in applications with many items
and/or many periods. Therefore, several heuristics have been developed for this problem. To judge the
quality of heuristics a good and simply computable lower bound for the minimum cost is useful. The idea
to derive such a lower bound is to decompose the N-item problem into N 1-item problems. To this end, the
family ordering cost A is divided over the N items according to some weights w
i,t
(

N
i=1
w
i,t
= 1). Since for
all combinations of values of Q
i,t
and w
i,t
it holds that
δ(

N
i=1
Q
i,t
) ≥
N

i=1
w
i,t
δ(Q
i,t
), (2.34)
it follows that the minimum of the cost function (2.30) has a lower bound which is a separable function:
min
Q
i,t
H

t=1
_
Aδ(

N
i=1
Q
i,t
) +
N

i=1
¦a
i
δ(Q
i,t
) +h
i
I
i,t
¦
_

N

i=1
min
Q
i,t
H

t=1
[¦Aw
i,t
+a
i
¦δ(Q
i,t
) +h
i
I
i,t
] . (2.35)
The lower bound can be computed by solving N 1-item minimization problems. The lower bound depends on
the choice of the weights so that various lower bounds could be generated by considering different combinations
of weights.
The first heuristic to be discussed is a heuristic by Silver. It is a combination of the one-step heuristic for
coordinated replenishment with constant demand (e.g., the Goyal & Belton heuristic) and the Silver-Meal
heuristic (Algorithm 1.3) for 1-item problems with time varying demand.
Algorithm 2.9 [Silver’s heuristic]
Step 1: Determine the average demand over all periods of each item:
¯
D
i
.
=
1
H

H
t=1
D
i,t
.
Step 2: Determine an indirect grouping of the items on the basis of these average demands, that is,
determine a vector k of order frequencies according to Algorithm 2.2; divide the N items into two
groups: i ∈ (
1
if k
i
= 1 and i ∈ (
2
if k
i
> 1.
Step 3: Consider the items in (
1
as one aggregated item with
ˆ a
.
= A+

i∈G
1
a
i
,
ˆ
h
.
=

i∈G
1
h
i
,
ˆ
D
t
.
=
1
ˆ
h

i∈G
1
h
i
D
i,t
, (2.36)
and apply the Silver-Meal heuristic to the aggregated item; this yields H
1
≤ H reorder instants and
determines the order quantities for items i ∈ (
1
.
Step 4: Apply the Silver-Meal heuristic to the items in (
2
separately (with ordering costs a
i
) with
restriction to the H
1
reorder instants of (
1
.
Remember that periods with zero demand are skipped by the Silver-Meal heuristic.
Example 2.14 Consider a family consisting of N = 3 items that are ordered from the same supplier. The
family ordering costs are A = $10. The supplementary ordering costs a
i
and the holding cost h
i
can be found
in Table 2.7, just like the demand D
i,t
over H = 6 weekly periods, t = 1, . . . , 6, i = 1, 2, 3. Step 1 of Silver’s
heuristic is the computation of the average demand over all periods of each item:
¯
D
i
, i = 1, 2, 3. Step 2 is
40 COORDINATED REPLENISHMENT
Table 2.7: Demand in units per item per week; ordering and holding costs; reorder frequencies.
Item a
i
h
i
period t: 1 2 3 4 5 6
¯
D
i
A+a
i
h
i
¯
D
i
2a
i
h
i
¯
D
i
(T
(0)
)
2
k
i
1 5 0.10 D
1,t
10 0 25 40 30 15 20 7.50 3.13 2
2 2 0.30 D
2,t
60 40 60 50 30 60 50 0.80 — 1
3 12 0.60 D
3,t
20 25 20 40 25 20 25 1.47 1.00 1
the application of Algorithm 2.2: on the basis of the ratios (A+a
i
)/(h
i
¯
D
i
) the reference item is determined
as i
r
= 2. The initial family cycle becomes T
(0)
=
_
8
5
. Then, the reorder frequencies follow as k
2
= 1 and
by (2.5), as k
1
= 2 and k
3
= 1. Hence, (
1
= ¦2, 3¦ and (
2
= ¦1¦. The data for the aggregated item of group
(
1
are: ˆ a = 24,
ˆ
h = 0.90,
ˆ
h
ˆ
D
1
= 30,
ˆ
h
ˆ
D
2
= 27,
ˆ
h
ˆ
D
3
= 30,
ˆ
h
ˆ
D
4
= 39,
ˆ
h
ˆ
D
5
= 24,
ˆ
h
ˆ
D
6
= 30. The Silver-Meal
heuristic for the aggregated item yields: τ = 1,
¯
C(1, 1) = 24,
¯
C(1, 2) =
1
2
(24 + 27) >
¯
C(1, 1), hence, τ = 2,
¯
C(2, 1) = 24,
¯
C(2, 2) =
1
2
(24 + 30) >
¯
C(2, 1), hence, τ = 3,
¯
C(3, 1) = 24,
¯
C(3, 2) =
1
2
(24 + 39) >
¯
C(3, 1),
hence, τ = 4,
¯
C(4, 1) = 24,
¯
C(4, 2) =
1
2
(24 + 24) = 24 ≤
¯
C(4, 1),
¯
C(4, 3) =
1
3
(24 + 24 + 2 30) >
¯
C(4, 2),
so that τ = 6. Hence, the Silver-Meal heuristic indicates that the items in group (
1
should be ordered in
every period except period 5. Next, the Silver-Meal heuristic is applied to the item in (
2
in such a way
that period 5 cannot become a reorder instant: τ = 1,
¯
C(1, 1) = 5,
¯
C(1, 2) is skipped since D
1,2
= 0,
¯
C(1, 3) =
1
3
(5 + 2 2.5) =
10
3

¯
C(1, 1),
¯
C(1, 4) is skipped since demand for period 5 is coupled to period 4,
¯
C(1, 5) =
1
5
(5 +2 2.5 +3 4 +4 3) =
34
5
>
¯
C(1, 3), hence, τ = 4,
¯
C(4, 1) is skipped since demand for period
5 has to be ordered in period 4,
¯
C(4, 2) =
1
2
(5 + 3) = 4,
¯
C(4, 3) =
1
3
(5 + 3 + 2 1.5) =
11
3
<
¯
C(4, 2). The
ultimate order quantities are given in Table 2.8. The total cost over the 6 weeks is
5A+ 2a
1
+ 5a
2
+ 5a
3
+h
1
(2D
1,3
+D
1,5
+ 2D
1,6
) +h
2
D
2,5
+h
3
D
3,5
= $165.
Note that if the Silver-Meal heuristic would not skip period 2 with zero demand for item 1, period 3 would
become a reorder instant instead of period 4 and the total cost would be $168.50.
Table 2.8: Order quantities according to Algorithm 2.9.
period t: 1 2 3 4 5 6
Q
1,t
35 – – 85 – –
Q
2,t
60 40 60 80 – 60
Q
3,t
20 25 20 65 – 20
The next heuristic is called the coefficient method and has been proposed by Lambrechts et al. [47]. It is a
generalization of the part-period-balancing heuristic (Algorithm 1.4) for individual items. Let τ
i
denote the
last reorder instant of item i up to now, i = 1, . . . , N. At period u the heuristic computes coefficients α
i,u
as the difference between the cumulative holding cost from the last reorder instant τ
i
up to period u and the
ordering cost a
i
:
α
i,u
= h
i
u

t=τ
i
(t −τ
i
)D
i,t
−a
i
, i = 1, . . . , N. (2.37)
Note that α
i,τ
i
= −a
i
and that α
i,u
is increasing in u. The principle underlying the heuristic is to compensate
for ordering costs in a certain period by not ordering in some later periods. Orders are postponed as long
as the coefficients α
i,u
are negative or small. Since the compensation against later periods is not possible
for the last period H and, hence, the method would perform badly if the last period is to become a reorder
instant, an additional step is added for period H to see whether ordering in this period can be avoided.
Algorithm 2.10 [Coefficient method]
Step 1: Start with τ
i
= 1 for each item (i = 1, . . . , N) and take u = 2.
Step 2: Compute the coefficients α
i,u
for all items. Let (
.
= ¦i; α
i,u
> 0¦ be the set of items that are
candidates for ordering. Actually order all items i ∈ ( if

i∈C
α
i,u
> A. (2.38)
HANS BLANC, TILBURG UNIVERSITY 41
If this condition is satisfied, then the order quantities become
Q
i,τ
i
= D
i,τ
i
+ +D
i,u−1
, i ∈ (, (2.39)
and τ
i
= u for i ∈ (.
Step 3: If u < H then increase u by 1 and repeat from Step 2; otherwise, continue with the next step.
Step 4: If period H has been indicated as reorder instant by Step 2, then for all items i with Q
i,H
> 0,
• Set τ
i
equal to the last reorder instant of item i before H. Add the demand D
i,H
to the order in
period τ
i
if
h
i
(H −τ
i
)D
i,H
< a
i
. (2.40)
• For all remaining items with Q
i,H
> 0, let (
.
= ¦i; Q
i,H
> 0¦ and shift the demand D
i,H
to the
order in period τ
i
if

i∈C
h
i
(H −τ
i
)D
i,H
< A+

i∈C
a
i
. (2.41)

Table 2.9: Computations of coefficients for Algorithm 2.10.
u α
1,u
α
2,u
α
3,u
decision
1 – – – order items 1, 2 and 3
2 −5 10 3 order items 2 and 3 (13 > 10)
3 0 16 0 order item 2 (16 > 10)
4 12 13 48 order items 1, 2 and 3 (73 > 10)
5 −2 7 3 order nothing (10 ≤ 10)
6 1 43 27 order items 1, 2 and 3 (71 > 10)
Example 2.15 The coefficient method will be illustrated for the data of Example 2.14, cf. Table 2.7. The
results of the computations of the coefficients α
i,u
can be found in Table 2.9, together with the corresponding
ordering decisions. Since period H = 6 has been indicated as reorder instant by the first phase of the method
(see the left part of Table 2.10) step 4 of the algorithm has to be performed. All three items are scheduled
to be ordered in period H = 6 and period 4 is the previous reorder instant for all of them. The inequality
2h
i
D
i,6
< a
i
, cf. (2.40), is only satisfied for item 1. The demand D
1,6
is shifted to the order in period 4,
saving $2. The inequality 2h
2
D
2,6
+2h
3
D
3,6
< A+a
2
+a
3
, cf. (2.41), is not satisfied so that period 6 remains
a reorder instant for items 2 and 3. The modified ordering schedule can be found in the right part of Table
2.10. The total cost over the 6 weeks is
5A+ 2a
1
+ 5a
2
+ 4a
3
+h
1
(2D
1,3
+D
1,5
+ 2D
1,6
) +h
2
D
2,5
+h
3
(D
3,3
+D
3,5
) = $165.
Although the ordering schedule is different from that determined by Silver’s heuristic, cf. Example 2.14,
the costs are the same. Only an ordering cost of a
3
= 12 has been exchanged against a holding cost of
h
3
D
3,3
= 12.
Table 2.10: Order quantities according to Algorithm 2.10, after first phase and after final phase.
period t: 1 2 3 4 5 6 1 2 3 4 5 6
Q
1,t
35 – – 70 – 15 35 – – 85 – –
Q
2,t
60 40 60 80 – 60 60 40 60 80 – 60
Q
3,t
20 45 – 65 – 20 20 45 – 65 – 20
The last heuristic that will be discussed here has been proposed by Joneja [43]. It is called the cost-covering
heuristic and is closely related to the coefficient method. It has been built on the alleged weakness of the
coefficient method of ignoring items with a negative value of the coefficient α
i,u
, cf. (2.37), but close to zero, as
candidates for joining a family order. Such items may trigger a next order in the very near future. Therefore,
42 COORDINATED REPLENISHMENT
the cost-covering heuristic reconsiders items that were not included in a family order, and computes a second
coefficient to determine whether an item should be included as yet to a previous order:
β
i,u
= h
i

0
−τ
i
)
u

t=τ
0
D
i,t
−a
i
, u > τ
0
.
= max
1≤i≤N
τ
i
, i = 1, . . . , N. (2.42)
Observe that β
i,u
= −a
i
< 0 if τ
i
= τ
0
, that is, if item i is already included in the last family order. If τ
i
< τ
0
,
the coefficients β
i,u
represents the savings of shifting the quantity D
i,τ
0
+ + D
i,u
over τ
0
−τ
i
periods by
adding item i to the family order at τ
0
.
Algorithm 2.11 [Cost-covering heuristic]
Step 1: Start with τ
i
= 1 for each item (i = 1, . . . , N) and take u = 2.
Step 2: Compute the coefficient β
i,u
for all items. For all i with β
i,u
> 0, item i is included in the family
order in period τ
0
, and τ
i
becomes τ
0
.
Step 3: Compute the coefficient α
i,u
for all items. Let (
.
= ¦i; α
i,u
> 0¦ be the set of items that are
candidates for ordering. Actually order all items i ∈ ( if

i∈C
α
i,u
> A. (2.43)
If this condition is satisfied, then the order quantities become
Q
i,τ
i
= D
i,τ
i
+ +D
i,u−1
, i ∈ (, (2.44)
and τ
i
= u for i ∈ ( and also τ
0
= u.
Step 4: If u < H then increase u by 1 and repeat from Step 2; otherwise, stop: τ
i
remains the last
reorder instant of item i, i = 1, . . . , N.
This heuristic does not treat period H differently from other periods.
Table 2.11: Computations of coefficients for Algorithm 2.11.
u β
1,u
β
2,u
β
3,u
decision α
1,u
α
2,u
α
3,u
decision
1 – – – — – – – order items 1, 2 and 3
2 −5 −2 −12 — −5 10 3 order items 2 and 3 (13 > 10)
3 −2.5 −2 −12 no change 0 16 0 order item 2 (16 > 10)
4 8 −2 24 add items 1 and 3 −1 13 12 order items 2 and 3 (25 > 10)
5 2 −2 −12 add item 1 −2 7 3 order nothing (10 ≤ 10)
6 −5 −2 −12 — 1 43 27 order items 1, 2 and 3 (71 > 10)
Example 2.16 In this example, the cost-covering heuristic will be applied to the data of Example 2.14,
cf. Table 2.7. The results of the computations of the coefficients β
i,u
and α
i,u
can be found in Table 2.11,
together with the corresponding ordering decisions. Observe that at u = 4 the items 1 and 3 are added to
the order of period 3. This does not prevent, however, that period 4 becomes a reorder instant. At u = 5
item 1 is added to the order of period 4, and this does prevent period 5 of becoming a reorder instant. The
ordering schedule can be found in Table 2.12. The total cost over the 6 weeks is
5A+ 4a
1
+ 5a
2
+ 5a
3
+h
1
D
1,5
+h
2
D
2,5
+h
3
D
3,5
= $167.
This cost is $2 higher than the costs of the previous two heuristics.
Simulation studies have been performed in order to be able to determine the quality of the foregoing heuristics.
A large number of problems have been generated, with varying proportion A/¯ a, horizon H, average demand
and variance of demand over the H periods. Each problem has been simulated with various realizations of
the demand patterns. The problems have been solved exactly and with the heuristics.
HANS BLANC, TILBURG UNIVERSITY 43
Table 2.12: Order quantities according to Algorithm 2.11.
period t: 1 2 3 4 5 6
Q
1,t
10 – 25 70 – 15
Q
2,t
60 40 60 80 – 60
Q
3,t
20 25 20 65 – 20
The coefficient method scores best on the average relative error over all problem instances, and on computation
time. Only the maximum relative error is larger than for cost-covering heuristic. Joneja [43] has proved that
the cost of the solution generated by the cost-covering heuristic is at most three times as high as the minimum
cost. For the coefficient method, problem instances can be constructed for which the costs are an arbitrary
factor higher than the minimum cost. The heuristic of Silver turns out to perform worst, especially in case
of strongly fluctuating demand and with relatively small A. This may be related to the inflexibility of this
heuristic with respect to items that are assigned to the group (
1
and that are necessarily always jointly
ordered. The average (maximum) relative errors found are 4% (54%) for the coefficient method, 8% (41%)
for the cost-covering heuristic and 12% (300%) for the heuristic of Silver.
There exist alternative approaches; see Atkins & Iyogun [5], Iyogun [41], who also generalize the Silver-Meal
heuristic. Further, multi-pass heuristics which improve initial solutions in later steps have been developed,
cf. Kao [44]. These are generally better than single-pass heuristics but require more computation time.
Exercise 2.14 A manufacturer orders three part types from the same supplier. The joint ordering cost is
$30, the supplementary ordering costs are $10 for all three part types. The holding costs are $0.10, $0.40 and
$0.40 per unit per week for part type 1, 2 and 3, respectively. The planned demand D
i,t
for part type i in
week t has been listed in the table below for the three part types and for the next six weeks. All demand must
be satisfied from stock.
week t: 1 2 3 4 5 6
D
1,t
36 36 36 36 36 36
D
2,t
20 40 20 40 20 40
D
3,t
8 4 10 10 4 8
Determine an inventory policy for the three part types for the next six weeks with the aid of Silver’s heuristic,
with the aid of the coefficient method, and with the aid of the cost-covering heuristic, respectively. Compare
the costs associated to the three policies.
2.4 Stochastic demand
This section is devoted to coordinated replenishment problems with stochastic demand processes. The model
assumptions for a family of N items are:
• there is an infinite planning horizon;
• the ordering costs (A, a
i
, i = 1, . . . , N) are constant in time;
• the holding costs (h
i
, i = 1, . . . , N) are constant in time;
• the demand is stochastic with time independent parameters (e.g., a Poisson process for slow movers);
the expected demand per unit of time is E¦D
i
¦ and the standard deviation of the demand in a unit of
time is σ¦D
i
¦, i = 1, . . . , N;
• there is a constant lead time L for orders as a whole, or there are item-dependent constant lead times
L
i
, i = 1, . . . , N;
• any stockouts are backlogged (there is no stockout cost but a service level constraint for each item);
• there are no restrictions on order quantities.
The aim is the minimization of the expected average cost per unit of time subject to a service level constraint,
cf. Section 1.1. Some policies for coordinated replenishment for a family of items that have been considered
in the literature are (see Aksoy & Erenguc [1] and Goyal & Satir [36] for surveys):
44 COORDINATED REPLENISHMENT
• periodic review policies:
– policies with a joint review period R (with item-dependent order-up-to levels S
i
: (R, S), with
item-dependent reorder points s
i
and order-up-to levels S
i
: (R, s, S), with item-dependent reorder
points s
i
and order quantities Q
i
: (R, s, Q));
– policies with coinciding review instants (with item-dependent reorder frequencies k
i
and order-up-
to levels S
i
: (R, k, S), related to indirect grouping);
– policies with grouping and joint review periods R
j
per group (with item-dependent order-up-to
levels S
ij
: (R
j
, S
j
), related to direct grouping).
• continuous review policies:
– policies that supplement orders with other items (“can-order” policies as introduced by Ballintfy
[10], with item-dependent reorder points s
i
, can-order levels c
i
and order-up-to levels S
i
: (s, c, S),
or order quantities Q
i
: (s, c, Q));
– policies that order on the basis of the total demand (place a family order when the total demand
for all items since the last reorder instant exceeds Q, with item-dependent order-up-to levels S
i
:
(Q, S), cf. Pantumsinchai [54]).
Section 2.4.1 is devoted to a continuous review “can-order” policy under Poisson demand processes. A
heuristic for a periodic review policy related to indirect grouping will be discussed in Section 2.4.2. Litera-
ture on this subject includes Viswanathan [76] who considers an (R, s, S) policy, Wildeman et al. [78] who
apply a global optimization procedure, Eynan & Kropp [26], Van Eijs [72] who considers joint ordering and
transportation decisions, and Liu & Yuan [50] who study a system with correlated demand.
2.4.1 Can-order policies
This section is concerned with can-order policies. These are continuous review policies where the reorder
decision can be taken after each transaction.
Every item has three decision variables: for i = 1, . . . , N,
• s
i
: the reorder point for item i (the “must” order point);
• c
i
: the join-order level for item i (the “can” order level);
• S
i
: order-up-to level for item i (this level determines the order quantity).
These values are ordered as s
i
≤ c
i
< S
i
, i = 1, . . . , N. The corresponding policy is:
• order when the inventory position of an item i falls at or below its level s
i
;
• order all items j with their inventory position at or below their level c
j
up to their level S
j
.
The extremal values of the can-order level are:
• c
i
= s
i
: item i is individually ordered
(with ordering costs A+a
i
; optimal in case a
i
¸ A);
• c
i
= S
i
−1: item i is ordered at each reorder instant unless no demand has occurred
(optimal in case a
i
¸ A and h
i
D
i
relatively small).
The aim is the minimization of the expected average ordering and holding cost per unit of time subject to a
service level constraint. The objective function can be formulated as:
min
N

i=1
[M
i
(A+a
i
) +J
i
a
i
+h
i

¯
I
i
¦]; (2.45)
here, the auxiliary variables are: for i = 1, . . . , N,
• M
i
: expected number of times per year that item i causes an order (by reaching its level s
i
);
HANS BLANC, TILBURG UNIVERSITY 45
s s s s s s s s s s c
s

c

S
0 1 2 3 4 5 6 7 8 9 10
¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡
λ λ λ λ λ λ λ λ λ λ
µ µ
µ
µ
λ
Figure 2.4: Transition diagram for the Markov inventory process of the can-order model.
• J
i
: expected number of times per year that item i supplements an order (when its inventory is at or
below level c
i
);
• E¦
¯
I
i
¦: expected average inventory level of item i.
These 3N auxiliary variables M
i
, J
i
and
¯
I
i
, i = 1, . . . , N, are functions of the 3N decision variables s
i
, c
i
and S
i
, i = 1, . . . , N. The problem with 3N decision variables is too complex to solve as a whole. Therefore,
the analysis will be restricted to the case of Poisson demand processes, with demand rate λ
i
units of item i
per unit of time, i = 1, . . . , N, and a decomposition assumption is introduced: the opportunities for item i
to join an order occur according to Poisson process with rate
µ
i
=
N

j=1, j=i
M
j
, i = 1, . . . , N. (2.46)
This is a rough assumption, because the actual rates vary over time. By this assumption, the original problem
with 3N decision variables is decomposed into N problems with 3 decision variables each. The interaction
between these N problems is established via the opportunity rates µ
i
, i = 1, . . . , N.
The approximative approach of iterative solution globally consists of the following steps. First the case L = 0
is considered. When the lead time L = 0, orders are delivered immediately, and no safety stock is needed.
Hence, we can take s
i
= 0 for i = 1, . . . , N and still satisfy all demand directly from stock. Initial values for
M
i
can be obtained on the basis of individual (s
i
, S
i
) policies (in case µ
i
= 0, c
i
= s
i
= 0). Assuming that
the demand per customer is 1 unit we have S
i
= Q
i
, the economic order quantity, and
M
i
=
E¦D
i
¦
Q
i
=
¸
h
i
E¦D
i
¦
2(A+a
i
)
=
¸
h
i
λ
i
2(A+a
i
)
, i = 1, . . . , N. (2.47)
In the iteration steps, for each item successively, the opportunity rate µ
i
is estimated on the basis of the
last values of M
j
, j ,= i, cf. (2.46); then, a 1-item problem is solved and a new value of M
i
is determined,
i = 1, . . . , N. The iterations stop if the decision variables have not changed or if the cost saving is relatively
small. Next, the case L > 0 is considered. For each item separately, a safety stock is determined in such a
way that the service constraint is satisfied. To this end, the values of the levels s
i
= 0, c
i
and S
i
found with
L = 0 are equally increased until the service constraint is met.
There exist various methods for solving the 1-item problems. We will discuss an approach developed by Silver
[58]. For the ease of notation, the item index will be omitted in the following derivation. The assumptions
of Silver for solving the 1-item problems are a zero lead time (L = 0), a zero reorder point (s = 0), a Poisson
demand process with a single unit demand per customer (with demand rate λ per unit of time), and a Poisson
process of opportunities for joining an order (with rate µ per unit of time). For a single item, the expected
total cost per unit of time is of the form:
C(c, S) = M(A+a) +Ja +h
¯
I = MA+ (M +J)a +hE¦
¯
I¦, (2.48)
with M, J and E¦
¯
I¦ functions of c and S. Because both the demand and the opportunities for joining
an order occur according to Poisson processes, the inventory level process is a Markov process. The state
space is finite. Because an order is placed and immediately delivered as soon as the inventory level reaches
s = 0, the state s = 0 is an instantaneous state, and can be omitted from the state space. Figure 2.4 shows
the transition diagram of the inventory process for the case c = 4, S = 10. Let p
j
denote the stationary
probability that the inventory level is j units, j = 1, . . . , S. These state probabilities satisfy the following
46 COORDINATED REPLENISHMENT
balance equations. The states between the levels c and S − 1 can only be reached from states with 1 unit
more on stock and can only be left by the event of a demand (at rate λ):
λp
j
= λp
j+1
, j = c + 1, . . . , S −1. (2.49)
The states between the levels 1 and c can only be reached from states with 1 unit more on stock but can be
left by the event of a demand (at rate λ) and by the event of an opportunity to join an order (at rate µ):
(λ +µ)p
j
= λp
j+1
, j = 1, . . . , c. (2.50)
The state S can be reached from the states between the levels 1 and c by the event of an opportunity to join
an order (at rate µ) and from the state 1 by the event of a forced order (at rate λ, via an instantaneous visit
to state s = 0), and can only be left by the event of a demand (at rate λ):
λp
S
= λp
1

c

j=1
p
j
. (2.51)
Finally, the normalization of the probability distribution of the inventory level process implies that
S

j=1
p
j
= 1. (2.52)
With the aid of the equations (2.49) and (2.50) it is possible to express all state probabilities in terms of the
single probability p
S
:
p
j
= p
S
, j = c + 1, . . . , S; p
j
= ρ
c−j+1
p
S
, j = 1, . . . , c; (2.53)
here,
ρ
.
=
λ
λ +µ
. (2.54)
The normalization (2.52) determines the last unknown probability p
S
by substitution of (2.53):
p
S
=
_
S −c +ρ
1 −ρ
c
1 −ρ
_
−1
. (2.55)
Together, (2.53) and (2.55) specify the stationary distribution of the inventory level at an arbitrary instant.
The quantities M, J and
¯
I can now be specified as functions of c and S. Since an order is caused by a
particular item when it reaches the level s = 0, and because the latter occurs at rate λ when the inventory
is in state 1, it follows that the expected number of orders per unit of time caused by the item is
M = λp
1
= λρ
c
p
S
=
λρ
c
S −c +ρ
1−ρ
c
1−ρ
. (2.56)
Similarly, since opportunities for supplementing an order are assumed to occur at rate µ in the states 1, . . . , c,
the expected number of orders per unit of time supplemented by the item is
J = µ
c

j=1
p
j
= µρ
1 −ρ
c
1 −ρ
p
S
=
λ(1 −ρ
c
)
S −c +ρ
1−ρ
c
1−ρ
; (2.57)
here, it has been used that µρ = λ(1 −ρ) by (2.54). Together, (2.56) and (2.57) yield for the expected total
number of orders per unit of time in which the item takes a part:
M +J =
λ
S −c +ρ
1−ρ
c
1−ρ
= λp
S
. (2.58)
Finally, the expected average inventory level is

¯
I¦ =
S

j=1
jp
j
=
1
2
(S −c)(S +c + 1) +
ρ
1−ρ
[c −ρ
1−ρ
c
1−ρ
]
S −c +ρ
1−ρ
c
1−ρ
. (2.59)
HANS BLANC, TILBURG UNIVERSITY 47
Substitution of (2.56), (2.58) and (2.59) into (2.48) leads to an explicit expression of the expected total cost
as function of the can-order level c and the order-up-to level S:
C(c, S) =
λρ
c
A+λa +
1
2
h(S −c)(S +c + 1) +h
ρ
1−ρ
[c −ρ
1−ρ
c
1−ρ
]
S −c +ρ
1−ρ
c
1−ρ
. (2.60)
The minimization of this cost function proceeds in two steps. For fixed c it is a rather simple function of S.
Taking the derivative with respect to S and setting the latter equal to zero gives
d
dS
C(c, S) = −C(c, S) p
S
+h(S +
1
2
)p
S
= 0. (2.61)
Solving this quadratic equation in S formally gives
ˆ
S(c) = c −ρ
1 −ρ
c
1 −ρ
+
¸
2λ(a +ρ
c
A)
h
+
2cρ
c+1
1 −ρ

ρ(1 −ρ
c
)(1 +ρ
c+1
)
(1 −ρ)
2
. (2.62)
The expression of which the square root is taken is in some cases positive and then describes the optimal
order-up-to level
ˆ
S(c) as function of c; otherwise it is negative, and then the cost function (2.60) has for fixed
c a boundary extreme at S = c + 1. Next, substitution of the optimal value of
ˆ
S(c) into the cost function
C(c, S) gives a function of the single variable c. The determination of the optimal value of c can be carried
out by the “Golden-Section” method, among others; see Appendix B.2. Finally, substitution of the optimal
can-order level c into the function
ˆ
S(c) gives the optimal order-up-to level S.
Remark 2.10 Application of the Golden-Section method requires the selection of an initial search interval.
The lower bound is 0. An upper bound is rather arbitrary. One could try a multiple (1.5–2) of the items’
EOQ value without coordination. Further, the Golden-Section method applies to real-valued functions. Note
that the function
ˆ
S(c) can also be evaluated for noninteger values of c. Moreover, an integer value of c does
not give an integer value of S, in general. Hence, both c and S should be rounded afterwards. Alternatively,
the Golden-Section method can be applied with rounded values, cf. Example B.1.
Once the iterations for the case L = 0 have been carried out, a safety stock is determined for each item
separately for the actual lead time L > 0. Let c
0
and S
0
denote the optimal values of the can-order level
and the order-up-to level of an item found for L = 0. Starting from s
0
= 0, c
0
and S
0
, the values of s, c
and S are equally increased until the service constraint is met. Since the differences c −s and S −c do not
change during this procedure, also the values of M and J do not change. Once the reorder point s has been
determined on the basis of the service measure, the final policy for the item is (s, c
0
+s, S
0
+s). The resulting
safety stock is equal to s −λL and gives rise to an extra holding cost of h(s −λL).
Next, the determination of the actual service level will be discussed for the two service level constraints
described in Section 1.1. In both cases, the observation will be used that the stationary distribution of the
inventory position for a (s, c
0
+s, S
0
+s) policy is related to the stationary distribution of the inventory on
hand for a (0, c
0
, S
0
) policy by
Pr¦I(s, c
0
+s, S
0
+s) = s +k¦ = Pr¦I(0, c
0
, S
0
) = k¦, k = 1, 2, . . . , S
0
, (2.63)
here, the latter distribution is given by (2.53) and (2.55). Further, the distribution of the inventory position
I
O
just before an order for the item is placed is related to the distribution of the inventory position I at
arbitrary instants by Bayes’ theorem. For the event that the item supplements an order it holds that
Pr¦I
O
= s +k¦ =
µp
k
µ(p
1
+ +p
c
0
) +λp
1
=
µp
k
λp
S
= (1 −ρ)ρ
c
0
−k
, k = 1, . . . , c
0
; (2.64)
for the event that the item causes an order to be placed it holds that
Pr¦I
O
= s¦ =
λp
1
µ(p
1
+ +p
c
0
) +λp
1
=
λp
1
λp
S
= ρ
c
0
. (2.65)
Let D
L
denote the demand during the lead time L. For a cycle service level α, consider the probability Π of
no stockout in a reorder cycle. By the law of total probability it holds that:
Π
.
= Pr¦D
L
≤ I
O
¦ =
c

m=s
Pr¦D
L
≤ m[ I
O
= m¦ Pr¦I
O
= m¦. (2.66)
48 COORDINATED REPLENISHMENT
Due to the lack of memory of the Poisson demand process it follows that D
L
is independent of I
O
, and that
D
L
has a Poisson distribution with mean λL. Hence, the probability Π is completely determined by (2.64),
(2.65) and
Π =
c
0

k=0
Pr¦I
O
= s +k¦
s+k

j=0
(λL)
j
j!
e
−λL
. (2.67)
Hence, the reorder point s can be determined for which the service level constraint Π ≥ α is met. One can
start with s = 0 and increase s by 1 until the constraint is met, or search in the neighborhood of s = λL; see
also Remark 1.5.
For a fill rate constraint with target β, consider the fraction of demand satisfied from stock (the actual fill
rate) in a reorder cycle:
Ψ = 1 −E¦B¦/E¦Q¦; (2.68)
here, E¦B¦ denotes the expected backlog that accumulates per cycle, and E¦Q¦ denotes the mean order
quantity which is equal to the expected demand per cycle since there are no lost sales. The mean order
quantity does not depend on the reorder point s, and, therefore, has to be computed only once:
E¦Q¦ =
c

m=s
(S −m) Pr¦I
O
= m¦ =
c
0

k=0
(S
0
−k) Pr¦I
O
= s +k¦ = S
0
−(1 −ρ)
c
0

k=1

c
0
−k
. (2.69)
The mean accumulated backlog during a cycle is given by
E¦B¦ = E¦[D
L
−I
O
]
+
¦ −E¦[D
L
−S]
+
¦. (2.70)
The last term, the mean backlog at the beginning of a cycle, can be computed in the standard manner,
cf. Example 1.3. For the first term, the mean backlog at the end of a cycle, it follows again due to the
independence of D
L
and I
O
that
E¦[D
L
−I
O
]
+
¦ =
c

m=s
Pr¦I
O
= m¦

j=m
(j −m)
(λL)
j
j!
e
−λL
=
c
0

k=0
Pr¦I
O
= s +k¦

j=s+k
(j −s −k)
(λL)
j
j!
e
−λL
,
(2.71)
with again Pr¦I
O
= s + k¦ given by (2.64) and (2.65). The infinite sum in (2.71) can be computed in a
similar way as in Example 1.3. Hence, the reorder point s can be determined for which the fill rate constraint
Ψ ≥ β is met, in a similar way as for the cycle service constraint. The complete heuristic for determining a
can-order policy under the assumptions imposed by Silver [58] is summarized below.
Algorithm 2.12 [Can-order policy for Poisson demand]
Step 1: Take L = 0 and s
i
= 0 for all i. Compute initial values of the mean number of orders caused by
item i, M
i
, based on deterministic demand, cf. (2.47), for i = 1, . . . , N. Start with i = 1.
Step 2: Estimate the rate µ
i
on the basis of the current values of M
j
, j ,= i, j = 1, . . . , N, cf. (2.46).
Determine the optimal c
i
and S
i
that minimize the cost function C(c, S) for item i, cf. (2.60). The new
value of M
i
follows from (2.56).
Step 3: If i < N, increase i by 1 and repeat from Step 2; otherwise, verify whether the solution has
converged; if not, set i = 1 and repeat from Step 2; otherwise, fix for each item the last found values of
c
i
and S
i
as c
(0)
i
and S
(0)
i
, and continue with Step 4.
Step 4: Consider the actual lead time L > 0. Determine for each item separately the reorder point
required for satisfying the service level constraint, that is, successively evaluate the policies (s
i
, c
(0)
i
+
s
i
, S
(0)
i
+s
i
) for s
i
= 0, 1, 2, . . . , until the constraint is met.
In the last step, either (2.67) can be used for cycle service constraints, or (2.68)–(2.71) for fill rate constraints,
together with (2.64) and (2.65).
The estimated cost corresponding to the can-order policy — under the approximative assumption that op-
portunities for joining an order occur according to a Poisson process — is, cf. (2.48),
C(s, c, S) =
N

i=1
[M
i
A+ (M
i
+J
i
)a
i
+h
i

¯
I
i
¦], (2.72)
HANS BLANC, TILBURG UNIVERSITY 49
here, M
i
, J
i
, are the appropriate item-related numbers of orders according to (2.56) and (2.57), and the
expected average inventory on hand E¦
¯
I
i
¦ is given by (omitting the item index i):

¯
I¦ =
1
2
(E¦[S −D
L
]
+
¦ +E¦[I
O
−D
L
]
+
¦). (2.73)
This quantity can be computed in a similar way as the mean backlog, cf. (2.71).
Table 2.13: Data per month and individual deterministic policies for Example 2.17.
item a
i
h
i
λ
i
Q
i
M
i
T
i
1 $10 $1 10 34.64 0.289 3.46
2 $ 5 $1 20 46.90 0.426 2.35
3 $15 $1 30 62.45 0.480 2.08
Example 2.17 Consider a family consisting of N = 3 items. The family ordering cost is A = $50. The
supplementary ordering costs a
i
, the holding costs h
i
per unit per month and the average demand per month,
E¦D
i
¦ = λ
i
, are listed in Table 2.13, i = 1, 2, 3. The lead time is L = 0.1 month. Management imposes
a target fill rate of β = 0.99 for all three items. The economic order quantities Q
i
, assuming deterministic
demand and individual ordering, and the corresponding number of orders per month, M
i
, cf. (2.47), and the
reorder cycle T
i
, i = 1, 2, 3, are also displayed in Table 2.13. The iteration starts with item 1, with opportunity
rate µ
1
= M
2
+ M
3
= 0.426 + 0.480 = 0.907. The value of the parameter ρ
1
is 10/(10 + 0.907) = 0.917, cf.
(2.54). The cost function (2.60) for L = 0 is minimized by c
1
= 13 and S
1
= 28. The new expected number of
orders caused by item 1 is M
1
= 0.144, cf. (2.56). Observe that this value is much smaller than the initial value
of 0.289. This is explained by the observation that offering opportunities for supplementing orders diminishes
the number of times an item reaches its reorder point. Then, item 2 is considered, with opportunity rate
µ
2
= M
1
+ M
3
= 0.144 + 0.480 = 0.624 and, hence, ρ
2
= 20/(20 + 0.624) = 0.970. The cost function (2.60)
for L = 0 is minimized by c
2
= 26 and S
2
= 40. The new expected number of orders caused by item 2 is
M
2
= 0.284. Next, item 3 is considered, with opportunity rate µ
3
= M
1
+ M
2
= 0.144 + 0.284 = 0.428. In
this way, the iterative procedure continues. The results are summarized in Table 2.14. After the third round
of iterations the procedure stops because the values of c
i
and S
i
, i = 1, 2, 3, after round 3 are the same as
after round 2. The final values of M
i
, i = 1, 2, 3, indicate that about 44% of all orders are caused by item 3,
36% by item 2 and 20% by item 1. The expected total number of orders per month is M
1
+M
2
+M
3
= 0.837
so that the expected time between orders is
1
0.837
= 1.195 month.
Table 2.14: Iteration steps of Algorithm 2.12 for Example 2.17.
iteration item µ
i
ρ
i
c
i
S
i
C
i
(c
i
, S
i
) M
i
1 1 0.907 0.917 13 28 28.40 0.144
2 0.624 0.970 26 40 40.68 0.284
3 0.428 0.986 28 58 58.88 0.381
2 1 0.665 0.938 14 29 29.63 0.170
2 0.551 0.973 26 41 41.32 0.296
3 0.465 0.985 28 58 58.58 0.371
3 1 0.666 0.938 14 29 29.63 0.169
2 0.540 0.974 26 41 41.41 0.298
3 0.468 0.985 28 58 58.56 0.370
In the last step of Algorithm 2.12, the policies of round 3 are adjusted to a target fill rate of β = 0.99 for
a lead time of L = 0.1 month. For i = 1, 2, 3, the actual fill rate Ψ
i
is computed according to (2.68)–(2.71)
together with (2.64), (2.65), and s
i
is increased until Ψ
i
≥ 0.99. The intermediate results are summarized in
Table 2.15.
The final values of the parameters of the can-order policy can be found in Table 2.16, together with the cor-
responding mean order quantities E¦Q
i
¦, the expected mean inventory on hand E¦
¯
I
i
¦, the mean accumulate
backlog E¦B
i
¦, the mean reorder cycle E¦T
i
¦, the mean number of orders per month M
i
and J
i
caused
respectively supplemented by item i, and the estimated monthly cost C
i
attributable to item i, i = 1, 2, 3.
Moreover, the safety stocks V
i
and the expected times τ
c
i
= (S
i
−c
i
)/λ
i
to reach the can-order level c
i
from
the order-up-to level S
i
and τ
s
i
= (S
i
− s
i
)/λ
i
to reach the reorder point s
i
from the order-up-to level S
i
50 COORDINATED REPLENISHMENT
Table 2.15: Final step of Algorithm 2.12 for Example 2.17 with target fill rates.
s
i
Ψ
1
Ψ
2
Ψ
3
0 0.9825 0.9693 0.9621
1 0.9936 0.9827 0.9742
2 0.9918 0.9844
3 0.9916
(disregarding opportunities for joining an order) are included in this table for item i, i = 1, 2, 3. Note that
these quantities are independent of the target fill rates. Observe that item 2, with the smallest supplementary
ordering cost, is fastest in reaching its can-order level, while item 3, with the smallest individual cycle time
T
i
, is fastest in reaching its reorder point (on the average, when all items simultaneously start from their
order-up-to levels).
Table 2.16: Final can-order policy for Example 2.17 with target fill rates.
item λ
i
L s
i
c
i
S
i
E¦Q
i
¦ E¦
¯
I
i
¦ E¦B
i
¦ E¦T
i
¦ M
i
J
i
C
i
V
i
τ
c
i
τ
s
i
1 1 1 15 30 23.92 17.11 0.15 2.39 0.169 0.249 $29.76 0 1.50 2.90
2 2 2 28 43 33.51 24.38 0.28 1.68 0.298 0.299 $42.29 0 0.75 2.05
3 3 3 31 61 52.55 31.95 0.44 1.75 0.370 0.201 $59.02 0 1.00 1.93
Alternatively, the policies of round 3 of Algorithm 2.12 can be adjusted to a cycle service constraint of, say,
α = 0.98 for a lead time of L = 0.1 month. For i = 1, 2, 3, the probability Π
i
is computed according to (2.67)
with (2.64) and (2.65), and s
i
is increased until Π
i
≥ 0.98. The intermediate results are summarized in Table
2.17.
Table 2.17: Final step of Algorithm 2.12 for Example 2.17 with cycle service constraint.
s
i
Π
1
Π
2
Π
3
0 0.7337 0.5520 0.3628
1 0.8901 0.6955 0.4679
2 0.9668 0.8353 0.6191
3 0.9922 0.9275 0.7680
4 0.9733 0.8789
5 0.9916 0.9451
6 0.9781
7 0.9922
The required reorder points s
1
= 3, s
2
= 5 and s
3
= 7 are larger than those for the target fill rates above.
Note that the final probabilities Π
i
all exceed 0.99 in this case.
From the foregoing discussion it can be concluded that the determination of the values of the decision
variables of a can-order policy is complex and approximative, even under strong assumptions of Poisson
demand processes and a demand of 1 unit per customer. For other demand processes (e.g., of fast movers),
modifications are needed which do not always seem tractable. Moreover, simulation is needed for a real
performance analysis of a can-order policy (without the assumption of Poisson opportunity processes for
joining orders). It has been found that the model costs of can-order policies overestimate the real costs, and
that the model service levels of can-order policies underestimate the real service levels. In some circumstances
(A/¯ a large), the assumption of Poisson opportunity processes for joining orders is demonstrably bad: the
opportunities occur more frequently towards the end of a reorder cycle; it is then better to choose c
i
= S
i
−1
and to determine (s
i
, S
i
) without the Poisson assumption. Van Eijs [71] has developed a heuristic for this
situation. It should be noted that the fact that can-order policies sometimes perform badly is mostly due
to the applied method of analysis, and not always to policy itself. However, Ignall [39] showed for a two-
item case that can-order policies are not necessarily the optimal policy for a continuous-review coordinated
replenishment problem. In Silver [61] Algorithm 2.12 has been extended to can-order policies under compound
Poisson demand processes. Federgruen et al. [28] determine the can-order parameters for the 1-item problem
HANS BLANC, TILBURG UNIVERSITY 51
in the case of a compound Poisson demand process by a policy iteration algorithm for semi-Markov decision
problems.
Exercise 2.15 Consider a single item with a Poisson demand process at a rate of 2 units per day. Assume
that this item is a member of a group in which opportunities for joining an order occur according to a Poisson
process at a rate of once every 4 days. Suppose that this item is reordered according to a can-order policy
with s = 2, c = 5 and S = 10, with negligible lead time L = 0. Determine the stationary distribution of
the inventory level process and the average inventory level. Compute the expected number of orders per day
caused by this item and the expected number of orders per day supplemented by this item.
Exercise 2.16 Show that the mean order quantity E¦Q¦ of an item, cf. (2.69), is equal to λ/(M + J), cf.
(2.58), for this item. Explain this relation.
Exercise 2.17 Give a computable expression for the mean backlog at the end of a cycle, E¦[D
L
−I
O
]
+
¦, of
an item, cf. (2.71), for the case of a reorder point of s = 0. Derive a recursion for this quantity for a reorder
point of s > 0 in terms of that for the reorder point s −1.
Exercise 2.18 Consider a family of N = 2 items, with joint ordering cost A = $40 and supplementary
ordering costs a
1
= $10, a
2
= $15, with holding costs per unit per week h
1
= $5, h
2
= $8, with Poisson
demand rates λ
1
= 1 and λ
2
= 2 items per week, and with a lead time of L = 1 week. Determine the
can-order policy according to Algorithm 2.12 at a target fill rate of β = 0.98 for both items. Compute for both
items the probability that no backlog occurs in a cycle with this policy and the mean order quantity.
Exercise 2.19 A shop sells two items which are ordered from the same wholesale dealer. The joint ordering
cost is $15, the additional ordering cost of item 1 is $12 and that of item 2 is $18. The lead time is 2 days.
The holding cost is $0.10 per unit per day for item 1 and $0.08 per unit per day for item 2. The shop uses
a can-order policy with parameters (s
1
, c
1
, S
1
) for item 1 and with parameters (s
2
, c
2
, S
2
) for item 2. First,
suppose that the demand is deterministic, 9 units per day for item 1 and 20 units per day for item 2. Compute
the optimal reorder cycle for the case that both items are always jointly ordered. Determine the values of the
can-order parameters that realize this reorder cycle in such a way that all demand is satisfied from stock.
Next, suppose that the demand is stochastic. The manager of the shop has chosen the following parameter
values: s
1
= 20, c
1
= 40, S
1
= 80, s
2
= 40, c
2
= 80, S
2
= 160. The can-order policy is applied as follows:
every evening after closing time the inventories are reviewed; if one of the items has an inventory level at or
below its reorder point, an order is placed according to the can-order parameters; ordered goods are delivered
2 evenings later. Demand which cannot be satisfied from stock is lost. At the beginning of day 1, 30 units are
on stock of item 1 and 100 units of item 2. The realizations of the demand over the next 10 days are given
in the table below.
day 1 2 3 4 5 6 7 8 9 10
i = 1 8 12 15 10 10 5 10 15 10 10
i = 2 20 25 20 15 25 40 25 20 20 15
Simulate the inventory levels of the two items over these 10 days. Compute the total ordering cost over these
10 days. Also, compute the total holding cost, charged over the remaining inventories at the end of each day
before delivery of a possible order, over these 10 days. How many units demand of the two items are lost?
2.4.2 Periodic review policies
Atkins & Iyogun [4] have proposed a heuristic for the determination of the parameters of an (R, k, S) periodic
review policy, with
• R = T: the family reorder cycle (the review period),
• k
i
(integer): the reorder frequency for item i, i = 1, . . . , N,
• S
i
: the order-up-to level for item i, i = 1, . . . , N,
for a coordinated replenishment model with family ordering cost and stationary stochastic demand. The aim
is the minimization of the expected average ordering and holding cost per unit of time subject to a prescribed
service level (a target fill rate β
i
for item i, i = 1, . . . , N).
52 COORDINATED REPLENISHMENT
Globally, the heuristic consists of the following two phases. In the first phase, the stochastic nature of the
demand is ignored, and review periods are determined based on the average demand. In the second phase,
safety stocks are determined for each item separately, such that the service level constraints for the stochastic
demands are met. More specifically, in the first phase weights w
i
, i = 1, . . . , N, are determined for the
allocation of the family ordering cost A to the N items, and then a family cycle T and reorder frequencies k
i
are determined on the basis of the expected demand E¦D
i
¦ of all items and with ordering costs w
i
A+a
i
for
item i, i = 1, . . . , N. In the second phase, for every item separately the order-up-to level S
i
is determined on
the basis of an (R
i
, S
i
) policy with review period R
i
= k
i
T fixed (from the first phase), taking into account
the stochastic nature of the demand, and with ordering cost w
i
A+a
i
for item i, i = 1, . . . , N.
Next, we elaborate the heuristic in more detail. For a given allocation of the family ordering costs A with
weights w
i
it holds that
T
i
= k
i
T =
¸
2(w
i
A+a
i
)
h
i
E¦D
i
¦
, i = 1, . . . , N. (2.74)
A reformulation of this relation leads to
w
i
=
1
2A
[h
i
E¦D
i
¦T
2
i
−2a
i
], i = 1, . . . , N. (2.75)
The heuristic chooses w
i
= 0 if k
i
> 1. Let 1
.
= ¦i [ k
i
= 1¦ be the set of items that are included in every
family order. Then it follows from the normalization of the weights and the fact that T
i
= T for i ∈ 1:
1 =
N

i=1
w
i
=
1
2A

i∈I
[h
i
E¦D
i
¦T
2
−2a
i
]. (2.76)
This normalization leads to the optimal reorder cycle for the group of items 1, cf. (2.23), which will be used
as review period:
R
2
= T
2
= 2
_
A+

i∈I
a
i
__

i∈I
h
i
E¦D
i
¦ . (2.77)
The complete heuristic for determining an (R, k, S) periodic review policy is summarized below.
Algorithm 2.13 [Atkins & Iyogun]
Step 1: Start with w
i
= 0 for all i. Order the items such that T
1
≤ T
2
≤ . . . ≤ T
N
, that is, in ascending
value of a
i
/(h
i
E¦D
i
¦), cf. (2.74). Set m = 1.
Step 2: Increase w
1
, . . . , w
m
(and correspondingly T
1
, . . . , T
m
) until T
1
= = T
m
= T
m+1
(use (2.75)
for this purpose).
Step 3: If

m
i=1
w
i
< 1 and m < N, increase m by 1 and repeat from Step 2; otherwise, define the set
1
.
= ¦1, . . . , m¦.
Step 4: Compute the family cycle R = T for this set 1 according to (2.77). Determine the weights w
i
for i = 1, . . . , m with this family cycle T and with k
i
= 1 for i = 1, . . . , m (by definition of the set 1),
according to (2.75).
Step 5: Determine k
i
for i = m + 1, . . . , N by rounding T
i
/T, computed from (2.74) with w
i
= 0,
i = m+ 1, . . . , N, to the nearest integer.
Step 6: Determine for every item separately the order-up-to level S
i
on the basis of a target fill rate β
i
at
a fixed review period R
i
= k
i
T, i = 1, . . . , N, e.g., with the assumption of gamma distributed demand
during the lead time plus review period L +R
i
.
The items i = m + 1, . . . , N do not have a share in the family ordering cost. Step 2 is to be skipped when
m = N; if m has reached the value N, then 1
.
= ¦1, . . . , N¦.
The resulting safety stocks are
V
i
= S
i
−(L +R
i
)E¦D
i
¦, i = 1, . . . , N. (2.78)
HANS BLANC, TILBURG UNIVERSITY 53
The expected average inventory levels are with this policy bounded by: for i = 1, . . . , N,

¯
I
i
¦ ≥
1
2
[S
i
−LE¦D
i
¦ +S
i
−(L +R
i
)E¦D
i
¦] = S
i
−LE¦D
i
¦ −
1
2
R
i
E¦D
i
¦ = V
i
+
1
2
R
i
E¦D
i
¦.
The expected order quantities are E¦Q
i
¦ = R
i
E¦D
i
¦, i = 1, . . . , N. The total expected cost is
E¦C¦ =
N

i=1
_
w
i
A+a
i
R
i
+h
i

¯
I
i
¦
_
=
A
T
+
1
T
N

i=1
a
i
k
i
+
N

i=1
h
i

¯
I
i
¦. (2.79)
Remark 2.11 The first phase of Algorithm 2.13 (the first five steps) can be replaced by any other heuristic
for the deterministic indirect grouping problem or by the exact Algorithm 2.1, cf. Section 2.1, using the
average demand as the deterministic demand. On the other hand, the first five steps of Algorithm 2.13 form
an alternative heuristic for the deterministic indirect grouping problem. This heuristic can often be improved
by computing the optimal cycle length given the reorder frequencies determined in Steps 4 and 5 after Step
5 (like Step 4 of the Goyal & Belton heuristic 2.2).
Table 2.18: Data per week and optimal individual policies without family ordering cost for Example 2.18.
item a
i
v
i
E¦D
i
¦ σ¦D
i
¦ T
i
Q
i
index
a $3.00 $8.00 9 9 4.17 37.50 3
b $3.00 $12.50 40 25 1.58 63.25 2
c $3.00 $3.52 4 2 9.42 37.69 4
d $3.00 $33.30 60 24 0.79 47.46 1
Example 2.18 Consider a family consisting of N = 4 items, indicated by a, b, c, d. The family ordering
cost is A = $20. The supplementary ordering costs a
i
, the purchasing prices v
i
and the averages E¦D
i
¦ and
the standard deviations σ¦D
i
¦ of the demands per week are listed in Table 2.18, i = 1, 2, 3, 4. The carrying
charge is r = $0.0048 per dollar per week, and the holding costs are h
i
= rv
i
per unit per week. The lead time
is L = 1 week. According to Algorithm 2.13, first the reorder cycles T
i
are computed with ordering cost a
i
(that is, with w
i
= 0), and deterministic demand E¦D
i
¦, cf. (2.74). These values of T
i
and the corresponding
order quantities Q
i
are shown in Table 2.18. Moreover, the items have been indexed in ascending value of T
i
:
¦1, 2, 3, 4¦ = ¦d, b, a, c¦; see the column with the header “index”. The application of the heuristic continues
as follows. First, increase w
1
until T
1
= T
2
(T
d
= T
b
= 1.58). Then, by (2.75),
w
1
= w
d
=
1
2A
[rv
d
E¦D
d
¦T
2
b
−2a
d
] = 0.45 < 1;
now 45% of the family ordering cost A = $20 is allotted to item d. Next, increase w
1
and w
2
until T
1
= T
2
= T
3
(T
d
= T
b
= T
a
= 4.17). Then,
w
1
= w
d
=
1
2A
[rv
d
E¦D
d
¦T
2
a
−2a
d
] = 4.01;
w
2
= w
b
=
1
2A
[rv
b
E¦D
b
¦T
2
a
−2a
b
] = 0.89.
Here, the procedure stops because w
1
+ w
2
> 1. The result is m = 2, and 1 = ¦1, 2¦ = ¦d, b¦. The family
cycle for this set of items 1 is, cf. (2.77):
T =
_
2[A+a
d
+a
b
]/[rv
d
E¦D
d
¦ +rv
b
E¦D
b
¦] ≈ 2.0825.
The final weights w
1
and w
2
are computed at T = 2.0825 from (2.75):
w
1
= w
d
=
1
2A
[rv
d
E¦D
d
¦T
2
−2a
d
] = 0.89;
w
2
= w
b
=
1
2A
[rv
b
E¦D
b
¦T
2
−2a
b
] = 0.11;
this means that 89% of A is allotted to item d and 11% of A is allotted to item b. No part of A is allotted to
the other items, because they only act as supplements to orders. Finally, k
3
and k
4
are computed with the
final value of T from (2.74) with w
3
= w
4
= 0:
k
3
= k
a
= T
a
/T ≈ 2; k
4
= k
c
= T
c
/T = 4.53 ≈ 5.
54 COORDINATED REPLENISHMENT
The above reorder frequencies are the same as those generated by Goyal’s Algorithm 2.1, only the reorder
cycle obtained by the heuristic of Atkins & Iyogun (2.0825) is slightly longer than optimal (2.0776). Finally,
the order-up-to levels S
i
are determined in Step 6 at a target fill rate β
i
= 0.98 and with a fixed review period
R
i
= T
i
, separately for i = 1, . . . , 4. The results are summarized in Table 2.19 for a family reorder cycle
of T = 2.0825 weeks. The column with the header “C
i
” contains the cost that is fully attributable to item
i, i = 1, . . . , 4, that is, a
i
/R
i
+h
i

¯
I
i
¦. It does not include a part of the family ordering cost which amounts
to A/T = $9.60. The expected total weekly cost is E¦C¦ = $42.31 for this policy, cf. (2.79), which is slightly
better than the expected total weekly cost with a family reorder cycle of T = 2.0776 weeks (E¦C¦ = $42.37).
This is due to the sequential approach to fix the reorder cycles first and to determine the order-up-to levels
later, cf. Example 1.3.
Table 2.19: Summary of results by Algorithm 2.10.
item index w
i
k
i
R
i
E¦Q
i
¦ S
i
V
i

¯
I
i
¦ a
i
/R
i
h
i
V
i
h
i
[E¦
¯
I
i
¦ −V
i
] C
i
a 3 0.00 2 4.17 37.48 84 37.52 56.61 0.72 1.44 0.73 2.89
b 2 0.11 1 2.08 83.30 198 74.70 117.18 1.44 4.48 2.55 8.47
c 4 0.00 5 10.41 41.65 52 6.35 27.53 0.29 0.11 0.36 0.75
d 1 0.89 1 2.08 124.95 241 56.05 119.76 1.44 8.96 10.18 20.58
Atkins & Iyogun [4] compare the heuristic (R, k, S) policy and the can-order (s, c, S) policy with a simply
computable lower bound for the minimal cost as a gauge. This lower bound, derived in Atkins & Iyogun
[3], is based on the same idea of allotting the family ordering cost A to the N items according to weights
w
i
, i = 1, . . . , N. Then, by inequality (2.34) we have for the unknown minimum cost C
min
:
C
min

N

i=1
min C
i
; (2.80)
here, C
i
is the cost associated to a corresponding 1-item problem with ordering costs w
i
A+a
i
, i = 1, . . . , N.
Hence, to compute the above lower bound, N 1-item problems have to be solved. Atkins & Iyogun [4]
conducted experiments with Poisson demand processes and stockout costs. They found that:
• the parameters of a (R, k, S) policy are easier to determine than those of a can-order policy;
• the can-order policy outperforms the periodic review (R, k, S) policy when A/¯ a is small, but otherwise
the (R, k, S) policy outperforms the can-order policy;
• in a typical case, the (R, k, S) policy is 12% and the can-order policy 28% above the lower bound (while
independent ordering is 64% above that bound).
Finally, note that the heuristic (R, k, S) policy is not dependent on the assumption of Poisson demand
processes and constant lead times, and is therefore more suitable in case of “fast movers”. Viswanathan [76]
shows that an (R, s, S) policy, in which a good common review period R is found by a stepwise search and the
reorder point s
i
and the order-up-to level S
i
are determined separately for each item i given the review period
R and using only the supplementary ordering cost a
i
, outperforms the (R, k, S) policy by a few percent in
many cases.
Exercise 2.20 Consider again the six-item coordinated replenishment problem described in Exercise 2.4.
Determine a reorder policy according to the deterministic part of the heuristic of Atkins & Iyogun.
Exercise 2.21 Apply the deterministic part of the heuristic of Atkins & Iyogun to the three-item family
considered in Example 2.1. Compare the ordering cost allotted to item i (w
i
A + a
i
) with the substitute
ordering cost ˜ a
i
in Table 2.1 for i = 1, 2, 3 and comment on the differences. Compute the order-up-to levels
for the case that the lead time is L = 1 week, that the mean demand per week is equal to the demand given in
Table 2.1, and the standard deviations of the weekly demands are σ¦D
1
¦ = 3, σ¦D
2
¦ = 2, σ¦D
3
¦ = 1, and
that the target fill rates are 90% for all three items. What are the corresponding safety stocks?
Exercise 2.22 Determine a periodic review policy according to the heuristic of Atkins & Iyogun for the
family of three items considered in Example 2.17. Compute the corresponding mean order quantities.
Exercise 2.23 Determine a periodic review policy according to the heuristic of Atkins & Iyogun for the
two-item family considered in Exercise 2.18. Compute the corresponding mean order quantities.
Chapter 3
Multi-echelon Inventory Systems
This chapter is concerned with the management of flows of goods on parts of their way from producer to
buyer. Important issues in such supply chains are the requirements to have the goods of the right quality at
the right time at the right location. The main subsystems of physical distribution are:
inventory management: this includes demand forecasting, the choice of an ordering system, which consists
of rules for determining the ordering quantities and the reorder instants;
warehouse management: this includes determining its function in the distribution channel, the choice
concerning the number of warehouses and their locations, the choice of buying or renting warehouse
space, the choice of the lay-out of the warehouse and the equipment in the warehouse, the choice of the
internal transport and material handling system, and the policies concerning the order picking system;
transport management: this includes the decision of using own transport or putting transport out to
contract, the choice of the means of transport, a route planning system, and the choice of the sizes of
the pallets to be used, and whether to buy or to rent them.
The decisions in the area of physical distribution can also be classified according to their scope:
strategic: this includes the choice of the distribution channels (directly or via distribution centers to retailers
or indirectly via wholesale dealers), the decision concerning the distribution spread (the percentage of
shops in the trade that sell the goods), the choice between a “push” or a “pull” policy, and the
determination of the desired service levels;
tactical: this includes the decisions concerning the number and the locations of the warehouses, the choice
of the handling- and storage system, the choice of a storage policy, and choices concerning the means
of transportation;
operational: this includes the decisions concerning when and how much to order or to produce, the planning
of the order picking, and the routing of the transport.
Section 3.1 contains a classification of multi-echelon structures and a discussion of the functions of a distri-
bution center. Section 3.2 is devoted to inventory policies for linear and divergent two-echelon systems with
constant demand. Section 3.3 discusses algorithms for determining pull and push policies for two-echelon sys-
tems with time-varying demand. Finally, Section 3.4 is concerned with a periodic review policy for divergent
two-echelon systems with stochastic demand.
3.1 Multi-echelon structures
In practice, there exist many echelon structures, both in the physical distribution and between departments in
a factory. Even within a company or organization several echelon structures may exist next to each other. For
instance, a chain of supermarkets may have different echelon structures for durable goods and for perishable
goods.
55
56 MULTI-ECHELON INVENTORY SYSTEMS
·
·

»

»

>
>
>
>
>.

>
>
>
>
>.

»

»
Figure 3.1: A linear (l.), a convergent (m.) and a divergent (r.) supply chain consisting of three echelons.
3.1.1 Classification
A classification of echelon structures is the following:
linear chains: all units follow the same path or route (this may occur in an environment with only sequential
working on the units); usually, they are part of more complex chains;
converging chains: they occur in factories where parts are assembled, and in distribution from various
factories to a single distribution center;
diverging chains: they occur in factories where raw materials are cut into various part types and where
semi-manufactured items are made into various end products, and in distribution from a single distri-
bution center to various retailers;
general networks: they consist of both converging and diverging parts.
It is customary to speak of high and low echelons. The lowest echelons are the ones closest to the customers
(buyers). Usually, external demand only occurs at the lowest echelons, while external orders are only placed
at the highest echelons. The goods move from the highest to the lowest echelons, while the demand works
on from the lowest to the highest echelons.
In multi-echelon inventory systems a distinction is made between pull and push systems. In pull system
each stocking point has its own ordering policy, and stocking points at higher echelons have to adjust their
inventory management to the uncontrollable demand from the lower echelons. In push system information on
the inventory positions of the stocking points at all echelons is kept centrally, and the inventory management
is done centrally. The latter requires much, reliable and up-to-date information. A cost saving is possible
by better anticipation and coordination. However, problems may arise with the responsibilities of local
managers.
3.1.2 Distribution centers
This section is devoted to the function and some aspects of distribution centers (DCs). An important
function of a DC is that it restricts the number of transport lines between a number of factories and a
number of retailers. For instance, if a group of 25 factories supply a group of 100 retailers, then direct
transport would mean 2,500 transport lines, while indirect transport via a single DC reduces the number of
transport lines to 125. Moreover, direct transport will consist of relatively low volumes, and will therefore
be relatively expensive. Further, the space for storage is generally cheaper in a DC than at the retailers
because a DC can be located in an area outside city centers. Also, transferring goods to other means of
transport (e.g., from truck, train or ship to a smaller truck or van that can reach a shopping area) may
be a function of a DC. Further, a DC may serve for regrouping and repacking of goods which may arrive
at the DC in large bulks and depart from the DC in much smaller quantities. Standard activities in a
distribution center include the receipt and the control of goods, cross-docking (transferring goods from one
means of transport to another without storage), material handling, storage, order picking, grouping of items
(per order), repacking in smaller quantities, combining of orders for transport, administration of incoming
and outgoing goods, and regular control of the inventories. Finally, Value Added Logistics (VAL) activities
HANS BLANC, TILBURG UNIVERSITY 57
may be performed in a DC. These activities aim at postponing product differentiation and lead to a larger
flexibility and lower stocks. VAL activities are most profitable if there is a large distance between a factory
and its market, as is the case with a factory in East Asia and retailers in Europe. Since VAL activities
may require costly equipment, the number of distribution centers in which they are performed will be small,
e.g., a single VAL distribution center for the whole of Europe. Value Added Logistics activities can be
performed in a distribution center owned by the producer but can also be put out to contract to logistic
service industries (this is called public warehousing). Examples of VAL activities are specific labeling (e.g.,
according to the rules of countries), supplementing manuals (in specific languages), making to measure,
mixing of ingredients, installing components or software, assembling of parts, fitting of accessories, testing,
quality control, repairing, examining of returns, reconditioning of clothes, specific repacking (e.g., according
to rules of countries), dispatching and administrating orders.
To validate the right of existence of a DC the cost savings by the reduction of the number of transport
lines and the lower holding cost have to be weighed against the cost of maintaining the DC and of the extra
handling cost in the DC.
Next, we discuss in more detail the material handling systems and the order picking systems in a warehouse.
These are often expensive systems which constitute a significant part of the costs of companies. Therefore,
efficiency is an important aspect. Ways to arrive at efficiency are by eliminating, combining and simplifying
movements in a warehouse. As far as possible it is best to work with a unit load (e.g., defined as a quantity
which can be stacked on a pallet). Further, the reliability and the safety of these systems are important.
Moreover, they should be connected to an information system so that the locations of all units can be tracked
at any time.
A rough classification of material handling systems is:
conveyors: they are relatively static in routing, allow a limited variation in products; examples of use are:
for bottles (washing, filling, closing, labelling), and for luggage at airports;
cranes: they are mobile in more directions, they are more flexible in case of disturbances, and they take
up less floor space than conveyors; examples of use are: for bottles (putting in crates), for storing in
storage carousels, and for computer controlled storage;
vehicles: they are still more mobile and have a larger range of action, they are still more flexible than cranes,
they may be manually or automatically guided; examples of use are: order picking in warehouses.
Order picking systems deal with collecting a limited number of units of some set of items for a number of
customers. It generally concerns relatively small quantities and requires much labor and/or much capital.
Examples where order picking systems are used are in mail-order companies, in factories (parts for production)
and in distribution centers. Possible objectives in optimizing the use of order picking systems are:
• minimization of the total order-picking cost per unit of time;
• minimization of the total order-picking time per unit of time;
• minimization of the average lead time to the customers;
• maximization of the average number of picked orders or items per unit of time.
Important aspects of order picking systems are:
• the lay-out of the warehouse and the division of goods over the warehouse;
• a good information system;
• a good routing system;
• the zoning of the warehouse (and the orders);
• the combining of orders (batching);
• a redivision of goods over the warehouse if some demand patterns change.
A systematic planning and design of order picking systems includes:
• a strategic planning: what type of customers, what service levels, what demand patterns?
58 MULTI-ECHELON INVENTORY SYSTEMS
¸ ¸ ¸ ¸
factory importer
wholesale
dealer
retailer
Figure 3.2: A linear supply chain consisting of four echelons.
• a classification items: poisonous, inflammable, pilferage, special sizes;
• an ABC analysis: important for the locations in a warehouse;
• the determination of the required storage capacity;
• the determination of a storage policy (what items are stored where in the warehouse? how to divide
the warehouse in to zones?);
• the determination of the degree of automatization (this concerns the actual order picking as well as the
information system);
• the choice of a material handling system;
• the choice of an information system;
• the determination of a batching and routing policy.
Hausman et al. [38] derive an optimal storage assignment for automatic warehousing systems. Wilson [79]
considers the location of stock in a warehouse as function of the order quantity and the demand of the various
products. Tomkins & White [66] is devoted to facilities planning. Ratliff & Rosenthal [55] solve a special
case of an order-picking problem as a travelling salesman problem. Other studies on order-picking systems
include Goetschalckx & Ratliff [31, 32]. Eastman [23] is concerned with material handling systems. Robeson
& House [56] discuss distribution issues.
3.2 Constant demand
In this section, models and policies will be discussed for multi-echelon systems with constant demand at the
stocking points of the lowest echelon. In general, the demand in higher echelons is formed by the orders of
lower echelons. Hence, the demand pattern in higher echelons is not continuous but stepwise! This subject
is also discussed in Silver et al. [64, Ch. 12].
3.2.1 Linear chains
We start this section with an example of the oscillation effect that may occur in supply chains when not
enough care is taken in choosing reorder policies and in exchanging information between echelons.
Example 3.1 Consider a linear supply chain consisting of a factory, an importer, a wholesale dealer and a
retailer, cf. Figure 3.2. Suppose that at all echelons the following simple ordering rule is applied. Every week
an order is placed that is immediately delivered provided that there are no stockouts at the higher echelon.
The order-up-to level is set at 110% of the demand of the previous week (S = 1.1 D) so that a safety stock is
used of 10% of the demand of the previous week (V = 0.1 D). The order quantity is the difference between
the order-up-to level and the inventory position: Q = S − I. Assume that the demand at the retailer has
been D = 100 during the last several weeks. In this stationary situation, the order-up-to level has converged
to S = 110 at all echelons, with V = 10, Q = 100. Now suppose that the demand at the retailer has increased
to D = 110 in a certain week due to a special offer. This has the following effect on the order quantities:
Retailer: D = 110 S = 121 V = 11 I = 110 −110 = 0 Q = 121 −0 = 121
Wholesale: D = 121 S = 133 V = 12 I = 110 −121 = −11 Q = 133 + 11 = 144
Importer: D = 144 S = 158 V = 14 I = 110 −144 = −34 Q = 158 + 34 = 192
Factory: D = 192 S = 211 V = 19 I = 110 −192 = −82 Q = 211 + 82 = 293
HANS BLANC, TILBURG UNIVERSITY 59
Note that the order quantities have increased by 21%, 44%, 92% and 193%, respectively, due to a 10%
increase in demand! Next, suppose that in the weeks after the special offer the demand is back at D = 100.
This has the following effect on the order quantities:
Retailer: D = 100 S = 110 V = 10 I = 0 + 121 −100 = 21 Q = 110 −21 = 89
Wholesale: D = 89 S = 98 V = 9 I = −11 + 144 −89 = 44 Q = 98 −44 = 54
Importer: D = 54 S = 59 V = 5 I = −34 + 192 −54 = 104 Q = 0 (59 −104 < 0)
Factory: D = 0 S = 0 V = 0 I = −82 + 293 = 211 Q = 0
Observe that the retailer is back in the stationary situation with the order-up-to level, but that the higher
echelons show a counterreaction with small or zero order quantities. In the factory, 193 units have been
produced above the usual quantity of 100, while only 82 are forwarded to the importer as backlog of the
previous week. So, the factory has had high production cost, perhaps with working overtime, and now ends
up with a high inventory level. The effect of a demand of D = 100 in the next week is:
Retailer: D = 100 S = 110 V = 10 I = 21 + 89 −100 = 10 Q = 110 −10 = 100
Wholesale: D = 100 S = 110 V = 10 I = 44 + 54 −100 = −2 Q = 110 + 2 = 112
Importer: D = 112 S = 123 V = 11 I = 104 + 0 −112 = −8 Q = 123 + 8 = 131
Factory: D = 131 S = 144 V = 13 I = 211 + 0 −131 = 80 Q = 144 −80 = 64
The wholesale dealer is back at S = 110, but the higher echelons exhibit an upward movement. At the
factory, this upward movement is quenched by the fact that negative order quantities are not feasible. The
next weeks, the oscillations fade away:
Retailer: D = 100 S = 110 V = 10 I = 10 + 100 −100 = 10 Q = 110 −10 = 100
Wholesale: D = 100 S = 110 V = 10 I = −2 + 112 −100 = 10 Q = 110 −10 = 100
Importer: D = 100 S = 110 V = 10 I = −8 + 131 −100 = 23 Q = 110 −23 = 87
Factory: D = 87 S = 96 V = 9 I = 80 + 64 −87 = 57 Q = 96 −57 = 39
After two more weeks, the steady state will be reached for the whole supply chain.
Possible causes of the oscillation effect may be special offers, advertisements, an announcement of a price
change, but also stochastic fluctuations and holidays. The oscillation effect can be limited by better exchange
of information between echelons, by exponential smoothing in forecasting future demand to avoid overreaction
on small changes in the demand pattern, by removing echelons from the supply chain if feasible, and by
limiting the (necessity of) safety stocks, e.g., by better quality control. For (linear) chains it is important
that ordering decisions in a stocking point consider beside the inventory position at the stocking point itself
at least the inventory positions in all lower stocking points and the demand in the lowest stocking point. This
consideration led Clark & Scarf [18] to the introduction of the concept of echelon inventory. The echelon
stock in a stocking point is defined as the number of units that are present at the stocking point itself plus
those that have already passed the stocking point but have not yet been committed to outside customers.
Hence, the echelon stock includes the inventory on hand at lower echelons and the inventory in transit to
lower echelons. As with the usual inventory concept we can distinguish for the echelon stock at a stocking
point:
echelon inventory on hand: the number of units present at the stocking point or at lower echelons, or in
transit to lower echelons; this quantity plays a role in determining holding costs;
net echelon inventory (stock): the echelon inventory on hand minus the amount of backlog at the lowest
echelon;
echelon inventory position: the net echelon inventory plus the number of units on order from a higher
echelon but not yet delivered; this quantity may be required for determining a reorder instant;
echelon safety stock: the average echelon inventory position just before a reorder instant minus the average
demand during the lead time to the stocking point; this quantity is related to the service level constraint
or the cost of stockouts or losses.
As mentioned in Section 1.3, a distinction is made in multi-echelon inventory systems between pull and push
policies. A more refined classification is:
• pure pull policies:
60 MULTI-ECHELON INVENTORY SYSTEMS
– use only local information;
– lead to increasing variability at higher echelons;
– require high safety stocks;
• pull policies based on echelon stock:
– use information from lower echelons;
– lead to limited variability;
– require limited safety stocks;
• push policies:
– requires central information;
– apply central inventory management;
– require central safety stocks;
– may apply reshuffling of stocks on echelon level.
A push policy takes inventories and costs of all echelons into account.
Next, we will consider a linear chain consisting of two echelons, and, hence, consisting of two stocking points.
The parameters of the model are:
• D: the demand per unit of time at the lower echelon;
• a
1
, a
2
: the ordering cost for the higher/lower echelon, respectively;
• h
1
, h
2
: the holding cost per unit per unit of time at the higher/lower echelon, respectively; in some
cases, h
1
= rv
1
and h
2
= rv
2
with v
1
, v
2
the purchasing cost of a unit upon arrival at the higher/lower
echelon, respectively, and r the carrying charge;
• L
1
, L
2
: the lead time to the higher/lower echelon, respectively.
The demand at the lower echelon is assumed to be deterministic and continuous in time. Also, the lead times
are deterministic. In most situations, it will hold that h
2
> h
1
because storage is usual more expensive at the
lower echelons and because the value of the goods will increase (v
2
> v
1
) when they move down the echelons
if only since more shipping cost has been invested. Further, we may have a
1
> a
2
when external ordering is
more expensive than internal ordering. However, if transport charges are included in the internal ordering
cost distances and volumes may also play a role.
The decision variables are:
• Q
1
, Q
2
: the order quantity at the higher/lower echelon, respectively;
• s
1
, s
2
: the reorder point at the higher/lower echelon, respectively.
First, the case that the lead time from the higher to the lower echelon is negligible (L
2
= 0) will be considered.
As a preliminary property, we will argue that it is optimal to have Q
1
= nQ
2
for some integer n, n ≥ 1.
We will show with the aid of an example that a policy with Q
1
/Q
2
not equal to a positive integer can be
improved upon.
Example 3.2 Suppose that a policy with Q
1
=
3
2
Q
2
is considered for a two-echelon linear chain. Then, the
quantity
1
2
Q
2
lies on stock without use. Keep Q
2
fixed, so that the cost incurred at the lower echelon is fixed.
The cost at the higher echelon in 3 periods of the lower echelon is for Q
1
=
3
2
Q
2
: 2a
1
+
3
2
h
1
Q
2
. Consider first
alternative 0: order alternatingly Q
1
= Q
2
and Q
1
= 2Q
2
. The cost at the higher echelon in 3 periods of the
lower echelon is then: 2a
1
+ h
1
Q
2
, that is a saving of
1
2
h
2
Q
2
with respect to the original policy. Hence, the
policy with Q
1
=
3
2
Q
2
cannot be optimal. But the policy of alternative 0 can be further improved upon. For
comparison, the cost at the higher echelon in 6 periods of the lower echelon is 4a
1
+ 2h
1
Q
2
. Next, consider
alternative 1: order always Q
1
= Q
2
. The cost at the higher echelon in 6 periods of the lower echelon is
6a
1
+ 0h
1
Q
2
. This is better than alternative 0 if a
1
< h
1
Q
2
. Finally, consider alternative 2: order always
Q
1
= 2Q
2
. The cost at the higher echelon in 6 periods of the lower echelon is 3a
1
+ 3h
1
Q
2
. This is better
than alternative 0 if a
1
> h
1
Q
2
. Hence, Q
1
= nQ
2
with either n = 1 or n = 2 is cheaper than alternatingly
ordering Q
1
= Q
2
and Q
1
= 2Q
2
. In this way, it can be shown that the optimal policy is of the form
Q
1
= nQ
2
, with n a positive integer.
HANS BLANC, TILBURG UNIVERSITY 61
¸
`

I
2
(t)
t →
9T
2
Q
2
0 T
2
2T
2
3T
2
4T
2
5T
2
6T
2
7T
2
8T
2
s s s s s s s s s s

¸
`

I
1
(t)
t →
Q
2
2Q
2
3Q
2
Q
1
0 T
1
2T
1
s s s
¸
`

I(t)
t →
Q
2
2Q
2
3Q
2
4Q
2
0 T
1
2T
1
s s s

Figure 3.3: Inventory on hand at lower echelon (top), at higher echelon (middle) and in system (bottom).
Next, the average cost per unit of time with a push policy with Q
1
= nQ
2
will be derived. At the lower
echelon the inventory level I
2
(t) follows a common saw-tooth pattern, cf. the upper graph in Figure 3.3, so
that the average stock at the lower echelon is found to be
1
2
Q
2
. The inventory level at the higher echelon,
however, behaves like a step function: when an order of Q
1
units is delivered, a fraction
1
n
is immediately
forwarded to the lower echelon. During the first cycle at the lower echelon the inventory level at the higher
echelon is constant and equal to
n−1
n
Q
1
= (n − 1)Q
2
. Afterwards, a second fraction
1
n
of Q
1
is forwarded
to the lower echelon, so that during the second cycle at the lower echelon the inventory level at the higher
echelon is
n−2
n
Q
1
= (n −2)Q
2
, etcetera. This continues until the end of the cycle at the higher echelon. The
middle graph of Figure 3.3 shows the inventory on hand I
1
(t) at the higher level for the case n = 4. Hence,
the average inventory level at the higher echelon is:
1
n
n−1

j=0
j
n
Q
1
=
1
2
n(n −1)
n
Q
1
n
=
1
2
(n −1)Q
2
=
1
2
(Q
1
−Q
2
). (3.1)
This implies that the average cost per unit of time with a push policy is:
C(Q
1
, Q
2
) = D
_
a
1
Q
1
+
a
2
Q
2
_
+
1
2
[h
1
(Q
1
−Q
2
) +h
2
Q
2
], (3.2)
or, with Q
1
= nQ
2
,
C(n, Q
2
) =
D
Q
2
_
a
1
n
+a
2
_
+
1
2
Q
2
[nh
1
+h
2
−h
1
]. (3.3)
Remark 3.1 An alternative derivation and interpretation of the cost function follows by noting that the
echelon inventory level I(t)
.
= I
1
(t) + I
2
(t) at the higher echelon follows a saw-tooth pattern, cf. the lower
graph in Figure 3.3, so that the average echelon inventory level is
1
2
Q
1
. Associating a holding cost h
1
to this
inventory and noting that the echelon inventory at the higher echelon includes the inventory at the lower
echelon, an additional holding cost h
2
−h
1
has to be associated to the average inventory level of
1
2
Q
2
at the
lower echelon. So, the total holding cost becomes
1
2
h
1
Q
1
+
1
2
(h
2
−h
1
)Q
2
which is the same as the last term
in (3.2) and (3.3).
62 MULTI-ECHELON INVENTORY SYSTEMS
The optimal Q
2
at a fixed value of n is:
ˆ
Q
2
(n) =
¸
2(
1
n
a
1
+a
2
)D
nh
1
+h
2
−h
1
, (3.4)
with corresponding minimum average cost per unit of time
ˆ
C(n) =
_
2(
1
n
a
1
+a
2
)(nh
1
+h
2
−h
1
)D. (3.5)
Since the square root function is monotonously increasing, the optimal real-valued minimum of this cost
function is obtained by solving the equation
d
dn
ˆ
C
2
(n) = 2D
d
dn
¦
1
n
a
1
(h
2
−h
1
) +nh
1
a
2
+a
2
(h
2
−h
1
) +a
1
h
1
¦ = 2D¦−
1
n
2
a
1
(h
2
−h
1
) +h
1
a
2
¦ = 0.
Hence, the optimal real-valued n minimizing the foregoing cost function is:
n

push
=
¸
a
1
(v
2
−v
1
)
a
2
v
1
=
¸
a
1
(h
2
−h
1
)
a
2
h
1
. (3.6)
The optimal push policy is obtained by rounding n

push
to a positive integer in such a way that the cost (3.5)
is minimal.
Alternatively, consider a pull policy for the same situation. The lower echelon determines Q

2
on the basis of
the local cost as the economic order quantity, cf. (1.2), (1.6),
Q

2
=
_
2a
2
D
h
2
, C

2
=
_
2a
2
h
2
D. (3.7)
The higher echelon must follow with Q
1
= nQ

2
for some n ≥ 1, with additional cost
C
1
(n) = a
1
D
Q
1
+
1
2
h
1
(Q
1
−Q

2
) = a
1
D
nQ

2
+
1
2
h
1
(n −1)Q

2
. (3.8)
Note that applying the EOQ formula to higher echelons does not work due to the fact that demand does not
occur continuously in time, cf. (3.1). The optimal real-valued n for given Q

2
is:
n

pull
=
1
Q

2
_
2a
1
D
h
1
=
_
a
1
h
2
a
2
h
1
, (3.9)
with corresponding cost
C

1
=
_
2a
1
h
1
D −
1
2
h
1
Q

2
=
_
2a
1
h
1
D −
1
2
h
1
h
2
_
2a
2
h
2
D. (3.10)
Again, the optimal pull policy is obtained by rounding n

pull
to a positive integer in such a way that the cost
(3.8) is minimal. For both policies it is noted that if n

= 1, the higher echelon has no stocking function and
only serves for cross-docking. Observe that both ratios n

push
, cf. (3.6), and n

pull
, cf. (3.9), are independent
of the demand rate D. The difference between these ratios is largest when h
2
−h
1
is small in comparison to
h
2
while a
1
/a
2
is large.
Example 3.3 Consider a linear chain consisting of two echelons, with parameters a
1
= $20, a
2
= $1, v
1
= $1,
v
2
= $2, D = 1000, r = $0.25, so that h
1
= $0.25, h
2
= $0.50. First, we determine the optimal push policy.
From (3.6) it follows that n

push
=

20 so that n

push
= 4 or n

push
= 5. The corresponding costs according to
(3.5) are
ˆ
C(4) =
_
500(5 + 1)(4 + 1) = 50

6 ≈ $122.47 and
ˆ
C(5) =
_
500(4 + 1)(5 + 1) = 50

6 ≈ $122.47.
Since there is no difference, both values are optimal. The corresponding optimal order quantities are
ˆ
Q

2
(4) =
40

6 ≈ 98, Q

1
= 4
ˆ
Q

2
(4) = 392, and
ˆ
Q

2
(5) =
100
3

6 ≈ 82, Q

1
= 5
ˆ
Q

2
(5) = 410, respectively.
Next, we determine the optimal pull policy. From (3.7) it follows that Q

2
= 20

10 ≈ 63 with C

2
= 10

10 ≈
31.62. Further, (3.9) leads to n

pull
=

40 so that n

pull
= 6 or n

pull
= 7. The corresponding costs according
to (3.8) are C
1
(6) = 52.91 + 39.38 = 92.29 and C
1
(7) = 45.35 + 47.25 = 92.60. Hence, n

pull
= 6 is optimal
and Q

1
= 6Q

2
= 378. The total cost with the optimal pull policy is C = 31.62 + 92.29 = $123.91 which is
not much worse than the total cost with the optimal push policy.
HANS BLANC, TILBURG UNIVERSITY 63
¸
`

I
2
(t)
t →
Q
2
s
2
0 L
2
T
2
T
2
+L
2
2T
2
2T
2
+L
2
3T
2
3T
2
+L
2
4T
2
4T
2
+L
2
c c c c c
s s s s s

¸
`

I
t
(t)
t →
Q
2
0 L
2
T
2
T
2
+L
2
2T
2
2T
2
+L
2
3T
2
3T
2
+L
2
4T
2
4T
2
+L
2
c c c c c s s s s s
¸
`

I
1
(t)
t →
Q
2
Q
1
−L
1
0
1
2
T
1
T
1
−L
1
T
1
3
2
T
1
2T
1
−L
1
2T
1
c c c s s s
¸
`

I(t)
t →
Q
1
+s
2
s
2
s
1
−L
1
0 T
1
−L
1
T
1
2T
1
−L
1
2T
1
c c c
s s s

Figure 3.4: Inventory on hand at lower echelon, in transit, at higher echelon, in system (from top to bottom).
Finally, the consequences of a nonnegligible lead time L
2
> 0 between the higher and the lower echelon will
be studied. By the usual arguments, cf. Section 1.2.1, it is still optimal to have orders arrive at the lower
echelon when the inventory level reaches zero. This implies that orders must be placed at the higher echelon
when the inventory position at the lower echelon is at the level, cf. (1.9),
s
2
= D L
2
. (3.11)
Moreover, this means that orders must arrive at the higher echelon at instants when the inventory position
at the lower echelon is at the level s
2
so that part of the arriving order can immediately be transferred to the
lower echelon. This situation is illustrated for the case Q
1
= 2Q
2
in Figure 3.4 which shows the saw-tooth
pattern of the inventory level I
2
(t) at the lower echelon, the step pattern of the inventory I
t
(t) in transit
between the echelons and of the inventory level I
1
(t) at the higher echelon, and the saw-tooth pattern of the
echelon inventory level I(t)
.
= I
1
(t) + I
t
(t) + I
2
(t). Observe that the echelon stock I(t) is replenished when
it reaches the level s
2
due to the time lag L
2
before the goods can be used to satisfy external demand at the
lower echelon. The open (closed) dots in the upper two graphs of Figure 3.4 indicate the ordering (delivery)
instants at the lower echelon, those in the lower two graphs the ordering (delivery) instants at the higher
echelon. Note that the reorder instants at the higher echelon do not correspond to a unique inventory level
of I
1
(t). However, they do correspond to a unique level of the echelon inventory position I(t). The echelon
64 MULTI-ECHELON INVENTORY SYSTEMS
reorder point at the higher echelon is
s
1
= s
2
+D L
1
= D (L
1
+L
2
). (3.12)
The average number of units in transit is
¯
I
t
=
Q
2
L
2
+ 0(T
2
−L
2
)
T
2
=
Q
2
T
2
L
2
= DL
2
, (3.13)
which is independent of the order quantities Q
1
and Q
2
. Hence, if a holding cost is incurred for units in
transit, a fixed amount is added to the cost function (3.3) and the same order quantities are optimal as those
in the corresponding case with L
2
= 0.
Example 3.4 Consider a linear chain consisting of two echelons, with parameters a
1
= $32, a
2
= $1,
h
1
= $8, h
2
= $9, D = 1800 per year, L
1
= 0.04 year and L
2
= 0.01 year. From (3.6) it follows that
n

push
= 2. The corresponding optimal order quantities are
ˆ
Q

2
(2) = 60 and Q

1
= 2
ˆ
Q

2
(2) = 120. The average
cost per year for the optimal push policy is
ˆ
C(2) = $1020, cf. (3.5). From (3.9) it follows that n

pull
= 6.
The optimal order quantity at the lower echelon is Q

2
= 20. Hence, the optimal order quantity at the higher
echelon is Q

1
= 6Q

2
= 120. The cost per year for the optimal pull policy is C = 180 +880 = $1060, which is
$40 (3.9%) more than the minimum cost. For both policies the echelon reorder point for the higher echelon
is s
1
= 1800 0.05 = 90 and the reorder point for the lower echelon is s
2
= 1800 0.01 = 18.
Exercise 3.1 Consider the following two-echelon inventory problem. A retailer has a shop in the city center
and a depot at the industrial site of the same city. Space in the city center is scarce. Therefore, the holding
cost is much higher there. For a certain item, the holding cost is $3 per unit per year at the shop and $1 per
unit per year at the depot. The internal ordering cost of the shop is $10 per order. The external ordering cost
of the depot is $80 per order. The demand for this item at the shop is 6000 units per year. This demand is
approximately deterministic and constant over time. All demand must be satisfied from stock. Lead times are
deterministic. First determine the optimal order quantity for the shop leaving the cost incurred at the depot
out of account. Compute the minimum average cost at the shop according this policy. Given this policy at
the shop, determine the optimal order quantity and the minimum average cost for the depot. Next, consider
the two echelons combined. Determine the optimal order quantities for the shop and the depot. Compute the
minimum average cost per year corresponding to this policy. Determine for both policies for both echelons the
reorder points for the case that the lead time to the depot is 10 days and the lead time from the depot to the
shop is
1
2
day (assume 300 working-days in a year).
3.2.2 Diverging chains
Consider a diverging distribution chain consisting of two echelons. N retailers order a certain item from
the same distribution center, while the distribution center orders this item from an external supplier. The
parameters of the model are:
• a
0
: the ordering cost for the distribution center;
• a
i
: the ordering cost for retailer i, i = 1, . . . , N;
• h
0
: the holding cost per unit per unit of time for the distribution center;
• h
i
: the holding cost per unit per unit of time for retailer i, i = 1, . . . , N;
• D
i
: the demand per unit of time at retailer i, i = 1, . . . , N.
The aggregated demand per unit of time at the distribution center is D
0
=

N
i=1
D
i
.
Consider the following subclass of push policies. The distribution center uses a reorder cycle of R units of
time and retailer i orders k
i
times per cycle of length R from the distribution center, i = 1, . . . , N. Then, the
order quantity Q
0
of the distribution center and the order quantity Q
i
of retailer i are determined by
Q
0
= D
0
R; Q
i
= D
i
R/k
i
, i = 1, . . . , N. (3.14)
HANS BLANC, TILBURG UNIVERSITY 65
The objective function is the total ordering and holding cost per unit of time. It can be derived in a similar
way as that for a linear chain, cf. (3.2), noting that the distribution center has similar holding costs related
to each retailer, cf. (3.1):
C(R, k) =
a
0
R
+
N

i=1
k
i
a
i
R
+
1
2
Rh
0
N

i=1
k
i
−1
k
i
D
i
+
1
2
R
N

i=1
h
i
D
i
k
i
. (3.15)
This function can be rewritten as
C(R, k) =
1
R
_
a
0
+
N

i=1
k
i
a
i
_
+
1
2
R
_
h
0
D
0
+
N

i=1
h
i
−h
0
k
i
D
i
_
. (3.16)
This cost function has the same structure as that of the indirect grouping coordinated replenishment problem,
cf. (2.1), only the role of the parameters with respect to the integers k
i
, i = 1, . . . , N, is reversed. The cost
function (3.16) can be transformed into (2.1) by the following correspondences, cf. Das & Goyal [20]. The
two-echelon diverging supply chain with N retailers is equivalent to a coordinated replenishment problem for
N + 1 items (the distribution center is related to an item labeled 0) with parameters: for i = 1, . . . , N,
T =
1
R
;
ˆ
A = 0; ˆ a
i
=
1
2
(h
i
−h
0
)D
i
; ˆ a
0
=
1
2
h
0
D
0
;
1
2
ˆ
h
i
ˆ
D
i
= a
i
;
1
2
ˆ
h
0
ˆ
D
0
= a
0
. (3.17)
Note that the reorder frequency of the distribution center is forced to be k
0
= 1. After the above transforma-
tion, the optimal policy can be determined with the aid of Goyal’s Algorithm 2.1 or a policy can be generated
by any heuristic for the indirect grouping coordinated replenishment problem, with the modification that k
0
is kept equal to 1.
Remark 3.2 It may seem odd that a coordinated replenishment problem is formulated with family ordering
cost
ˆ
A = 0. The cost function (3.16), however, does not become separable in this case, because the integer
constraints on the frequencies k
i
, i = 1, . . . , N, are essential for its validity.
It is quite difficult to determine a pull policy for a diverging chain. Each stocking point at the lowest
echelon can readily determine its optimal reorder quantity according to the EOQ-formula, but without any
coordination this will lead to irregular, acyclic demand patterns at higher echelons. This makes it cumbersome
to formulate the average inventory levels at higher echelons and to determine optimal reorder policies for
higher echelons.
Table 3.1: Data per year and a push policy for Example 3.5.
Retailer i a
i
h
i
D
i
ˆ a
i
ˆ
h
i
ˆ
D
i
ˆ a
i
/(
ˆ
h
i
ˆ
D
i
) k
i
Q
i
T
i
a $1.80 $0.50 2900 $580 $3.60 161.1 3 133 0.0458
b $2.00 $1.10 1850 $925 $4.00 231.3 3 85 0.0458
c $1.20 $0.90 2750 $1100 $2.40 458.3 4 95 0.0344
d $3.20 $0.30 1600 $160 $6.40 25.0 1 220 0.1375
e $3.10 $0.90 3200 $1280 $6.20 206.5 3 147 0.0458
f $2.70 $0.30 1400 $140 $5.40 25.9 1 193 0.1375
Example 3.5 Consider a two-echelon diverging distribution chain consisting of a distribution center and
N = 6 retailers. At the distribution center, the ordering cost is a
0
= $10 and the holding cost is h
0
= $0.10
per unit per year. The ordering cost a
i
, the holding cost h
i
per unit per year and the demand D
i
per year
at retailer i, i = 1, . . . , 6, can be found in Table 3.1. The aggregated demand per year at the distribution
center is D
0
= 13,700. The corresponding coordinated replenishment problem concerns N +1 = 7 items, has
family ordering cost
ˆ
A = $0, supplementary ordering costs ˆ a
0
= $685 and ˆ a
i
, i = 1, . . . , N, given in Table
3.1, and holding cost times demand
ˆ
h
0
ˆ
D
0
= $20 and
ˆ
h
i
ˆ
D
i
, i = 1, . . . , N, given in Table 3.1, cf. (3.17). We
determine an ordering policy for the transformed problem according to the heuristic of Kaspi & Rosenblatt
(Algorithm 2.3). First, the items are indexed in ascending value of (
ˆ
A+ ˆ a
i
)/(
ˆ
h
i
ˆ
D
i
) = ˆ a
i
/(
ˆ
h
i
ˆ
D
i
). This leads
to the reference “item” being retailer d (for the distribution center, ˆ a
0
/(
ˆ
h
0
ˆ
D
0
) = 34.3). The initial family
cycle based on this reference item is T
(0)
= 5

2 ≈ 7.071. The reorder frequencies for the other retailers,
based on this cycle length, are k
a
= k
b
= k
e
= 3, k
c
= 4, k
f
= 1 (k
0
= 1 and k
d
= 1 are kept fixed). Given
66 MULTI-ECHELON INVENTORY SYSTEMS
these frequencies, the optimal cycle length is T
(1)
= 7.270. The policy improvement step does not alter the
frequencies. Hence, the family cycle becomes T
(2)
= 7.270, so that the review period is R = 0.1375 year
(≈ 7.15 week) for the two-echelon system. The order quantity of the distribution center becomes Q
0
= 1884.
The reorder frequency k
i
and the order quantity Q
i
of retailer i, i = 1, . . . , N, are included in Table 3.1. The
corresponding average yearly cost is
ˆ
C = $601.99 which is slightly higher than the minimum cost according
to Goyal’s Algorithm 2.1: C

= $601.62.
Table 3.2: Optimal pull policies at the lower echelon for Example 3.5.
Retailer Q

i
T

i
M

i
a 144.5 0.0498 20.07
b 82.0 0.0443 22.56
c 85.6 0.0311 32.11
d 184.8 0.1155 8.66
e 148.5 0.0464 21.55
f 158.7 0.1134 8.82
Table 3.2 contains the individual optimal reorder quantity Q

i
, and the corresponding optimal cycle T

i
and
number of orders per year, M

i
, for retailer i, i = 1, . . . , 6, according to the basic EOQ models. Because
the reorder cycles are asynchronous, these optimal pull policies lead to an acyclic demand pattern at the
distribution center.
The application of push policies in divergent multi-echelon chains requires synchronization between the re-
tailers such that an order is shipped to each retailer at the instant when an order from the external supplier
arrives at the distribution center (partial cross-docking). In case of nonnegligible lead times L
0
to the distri-
bution center and L
i
from the distribution center to retailer i, i = 1, . . . , N, the distribution center should
place an order when the echelon inventory position reaches the echelon reorder point, cf. (3.12),
s
0
= D
0
L
0
+
N

i=1
D
i
L
i
=
N

i=1
D
i
(L
0
+L
i
). (3.18)
When the order arrives at the distribution center, the inventory position at retailer i will be s
i
= D
i
L
i
, i =
1, . . . , N.
Exercise 3.2 Consider a two-echelon distribution chain for a non-food product consisting of a local warehouse
and five supermarkets. At the warehouse, the ordering cost is $20 and the holding cost is $0.20 per unit per
year. The ordering cost a
2i
, the holding cost h
2i
per unit per year and the demand D
2i
per year at supermarket
i, i = 1, . . . , 5, can be found in the table below. Determine an ordering policy for the transformed problem
according to the heuristic of Kaspi & Rosenblatt and interpret the results in terms of the distribution chain.
Compare the found push policy with the individual optimal policies.
Supermarket i a
2i
h
2i
D
2i
1 $8 $0.60 2000
2 $10 $0.80 5000
3 $10 $0.50 4000
4 $12 $0.60 5000
5 $15 $0.50 4000
3.3 Time-varying demand
In this section, multi-echelon inventory systems are considered with known but time varying demand in the
stocking points at the lowest echelon of diverging chains. The optimal pull policy is easy to find for general
diverging chains by the following procedure:
Algorithm 3.1 [Pull policy]
Step 1: Apply the Wagner-Whitin Algorithm 1.2 to the stocking points at the lowest echelon.
HANS BLANC, TILBURG UNIVERSITY 67
Step 2: Aggregate the order quantities of the lowest echelon to obtain the demand process for the one
but lowest echelon.
Step 3: Apply the Wagner-Whitin algorithm to the stocking points at the one but lowest echelon.
Step 4: Continue in this way until the highest echelon is reached.
Alternatively, the Silver-Meal heuristic can be used.
The pull policy ignores cost dependencies between the echelons. Therefore, the aim is to determine a
(sub)optimal push policy that takes cost dependencies into account. An integral approach by dynamic
programming is rather difficult. Here, a sequential approach (from the lowest to the highest echelon) will be
discussed for a linear chain. The parameters of the model with a horizon of H periods are:
• D
2t
, t = 1, . . . , H: the demand per period at the lower echelon;
• a
1
, a
2
: the ordering cost at the higher/lower echelon, respectively;
• h
1
, h
2
: the holding cost per unit per period at the higher/lower echelon, respectively.
The decision variables are:
• Q
1t
, t = 1, . . . , H: the order quantity at the higher echelon in period t;
• Q
2t
, t = 1, . . . , H: the order quantity at the lower echelon in period t.
The demand at the higher echelon is D
1t
= Q
2t
, t = 1, . . . , H. Blackburn & Millen [13] have developed a
heuristic for an assembly system, that is, for a converging chain. For the case of a linear chain with two
echelons it is based on an ideal ratio n between the order quantities at the higher and at the lower echelon
derived from the constant demand model, cf. (3.6):
n = max
_
_
_
1,
¸
a
1
(h
2
−h
1
)
a
2
h
1
_
_
_
. (3.19)
Note that n is taken minimally equal to 1, but is otherwise not rounded to an integer. Motivated by cost
function (3.3) coordination between the echelons is extorted by adding a fraction
1
n
of the ordering cost a
1
to each order placed at the lower echelon and by adding (n −1)h
1
to the holding cost at the lower echelon.
Hence, the following modified cost parameters are defined for the lower echelon:
ˆ a
2
.
= a
2
+a
1
/n,
ˆ
h
2
= nh
1
+h
2
−h
1
. (3.20)
The heuristic of Blackburn & Millen [13], confined to a linear chain consisting of two echelons, proceeds as
follows.
Algorithm 3.2 [Blackburn & Millen]
Step 1: Determine the ideal ratio n according to (3.19).
Step 2: Apply the Wagner-Whitin algorithm to the lower echelon with modified costs ˆ a
2
,
ˆ
h
2
, cf. (3.20),
to determine Q
2t
, t = 1, . . . , H.
Step 3: Use D
1t
= Q
2t
, t = 1, . . . , H, and apply the Wagner-Whitin algorithm to the higher echelon
with costs a
1
, h
1
, to determine Q
1t
, t = 1, . . . , H. Here, the ratio n is abandoned.
Alternatively, the Silver-Meal heuristic can be used.
Example 3.6 Let a
1
= $20, a
2
= $4, v
1
= $20, v
2
= $30, r = $0.02/$/month, so that h
1
= $0.40,
h
2
= $0.60. The demand in the next 8 months D
2t
, t = 1, . . . , 8, is given in Table 3.3. For a push policy
according to Blackburn & Millen compute n =
_
20×10
4×20
=
1
2

10 ≈ 1.58 > 1 by (3.19) so that the modified
costs become ˆ a
2
= 4+
20
1.58
= 16.65 and
ˆ
h
2
= 0.2+1.580.4 = 0.832, cf. (3.20). Application of the Silver-Meal
heuristic at the lower echelon with the modified costs yields Q
2t
, t = 1, . . . , 8, cf. Table 3.3. The estimated
68 MULTI-ECHELON INVENTORY SYSTEMS
Table 3.3: Push policy according to Blackburn & Millen and Silver-Meal for Example 3.6.
Month 1 2 3 4 5 6 7 8
D
2t
40 10 60 40 15 45 25 25
Q
2t
50 0 60 55 0 45 25 25
Q
1t
50 0 60 100 0 0 50 0
Table 3.4: Push policy according to Blackburn & Millen and Wagner-Whitin for Example 3.6.
Month 1 2 3 4 5 6 7 8
D
2t
40 10 60 40 15 45 25 25
Q
2t
50 0 60 55 0 45 25 25
Q
1t
50 0 60 55 0 95 0 0
cost at the lower echelon based on the modified costs is 6ˆ a
2
+ (10 + 15)
ˆ
h
2
= $120.71. Application of the
Silver-Meal heuristic at the higher echelon with demand D
1t
= Q
2t
and costs a
1
, h
1
gives Q
1t
, t = 1, . . . , 8,
cf. Table 3.3. The real cost at the higher echelon is 4a
1
+ (90 + 25)h
1
= $126 and the real cost at the
lower echelon is 6a
2
+ (10 + 15)h
2
= $39. The total real cost is $165. The realized ratio between the order
frequencies is n = 6/4 = 1.50. Application of the Wagner-Whitin algorithm gives the results as displayed in
Table 3.4, with cost 4a
1
+(25+50)h
1
= $110 at the higher echelon, cost 6a
2
+(10+15)h
2
= $39 at the lower
echelon and total cost $149. Note that there are alternative optimal solutions: Q
16
= 70, Q
17
= 0, Q
18
= 25,
and Q
16
= 70, Q
17
= 25, Q
18
= 0.
Now consider a pull policy. Application of the Silver-Meal heuristic at the lower echelon with the unmodified
costs a
2
and h
2
yields Q
2t
, t = 1, . . . , 8, cf. Table 3.5. The cost at the lower echelon is 8a
2
+ 0h
2
= $32.
Application of the Silver-Meal heuristic at the higher echelon with demand D
1t
= Q
2t
and costs a
1
and h
1
gives
Q
1t
, t = 1, . . . , 8, cf. Table 3.5. The cost at the higher echelon is 4a
1
+(10+40+215+25)h
1
= $122. The total
cost is $154. The realized ratio between the order frequencies is n = 8/4 = 2. Observe that the Silver-Meal
heuristic leads in this case to a pull policy that is cheaper than the push policy obtained with this heuristic!
Application of the Wagner-Whitin algorithm gives the same results for the lower echelon but different results
for the higher echelon, cf. Table 3.6. The optimal pull policy has cost 4a
1
+ (10 + 15 + 25 + 50)h
1
= $120
at the higher echelon, and total cost $152 which is 2% higher than that of the optimal push policy. Again,
there are alternative optimal solutions.
Table 3.5: Pull policy for Example 3.6 by Silver-Meal.
Month 1 2 3 4 5 6 7 8
D
2t
40 10 60 40 15 45 25 25
Q
2t
40 10 60 40 15 45 25 25
Q
1t
50 0 115 0 0 70 0 25
Table 3.6: Pull policy for Example 3.6 by Wagner-Whitin.
Month 1 2 3 4 5 6 7 8
D
2t
40 10 60 40 15 45 25 25
Q
2t
40 10 60 40 15 45 25 25
Q
1t
50 0 60 55 0 95 0 0
Exercise 3.3 A firm owns a shop in the center of a city and a depot near the harbor. Goods are first delivered
at the depot where they are checked, repacked and kept in storage if necessary. For a certain item the demand
in the shop is deterministic but the amount varies from week to week. The demand for the next six weeks is
given in the following table.
week 1 2 3 4 5 6
demand 24 16 32 24 12 50
The ordering cost at the depot is $60, the ordering cost at the shop is $15 and mainly consists of shipping
cost. Orders are delivered at the first day of a week. Holding costs are only charged for units on stock at the
HANS BLANC, TILBURG UNIVERSITY 69
end of a week. The holding costs are $1 per unit per week at the shop and $0.20 per unit per week at the
depot. Determine a pull policy with the aid of the Silver-Meal heuristic for this two-echelon inventory system.
Determine a push policy with the aid of the Blackburn & Millen heuristic combined with the Silver-Meal
heuristic. Compute the total cost over six weeks corresponding to both policies. Show that these policies can
simply be improved upon.
3.4 Stochastic demand
This section is concerned with two-echelon diverging chains with independent, stationary stochastic demand
at the stocking points of the lower echelon. This represents a situation of a central depot or distribution
center and N local stocking points or retailers. The aim is to develop an good integral control, that is, a push
policy, cf. Sections 1.3, 3.2.1. The means to this goal are multi-echelon ordering rules, also called base-stock
control, cf. Clark & Scarf [18]. Such strategies are optimal for diverging chains provided that the demand
at the retailers is not too strongly varying, that is, there is no strong “imbalance” between the retailers.
The approach is this section is based on Van der Heijden [68]. This problem is equivalent to a hierarchical
production planning problem for a family of items with common components (ignoring production capacity
constraints).
The inventory policy is based on periodic review of the stock, applied as follows:
• the central depot uses an (R, S
0
) policy: with fixed intervals R the echelon-stock is ordered up to S
0
;
• the order of the central depot is delivered after a lead time L
0
;
• the delivery instant at the central depot is the reorder instant for the retailers;
• retailer i uses an (R, S
i
) policy: the stock is ordered up to S
i
, i = 1, . . . , N;
• the order of retailer i is delivered after a lead time L
i
, i = 1, . . . , N.
This basic procedure may require some modifications:
• if the total demand during the lead time L
0
is more than expected, then the stock in the central depot
will be too small to fully satisfy the orders of the retailers; in such a case,
– retailer i is first replenished up to the expected demand in the period L
i
+R; the remaining stock
is divided over the retailers by ratios p
i
, i = 1, . . . , N, or
– the total understock is divided over the retailers by ratios f
i
, i = 1, . . . , N;
• if the total demand during the lead time L
0
is less than expected, then the stock in the central depot
will be larger than the sum of the order quantities of the retailers; in such a case,
– the remaining stock could stay in the central depot, or
– the total overstock is divided over the retailers by ratios f
i
, i = 1, . . . , N.
In the sequel, the following model will be considered:
• there is a stockless central depot (it only serves for cross-docking and reallocation of goods);
• there is full information at the central depot to apply a push policy for fair allocation of stock over the
retailers and for lower stocks;
• no stock redistribution is possible between retailers;
• all shortages are backlogged; at each retailer, a fill rate constraint is imposed, cf. (1.38);
• there are no capacity constraints.
The parameters of the model are:
• L
0
: the lead time from the external supplier to the central depot;
• L
i
: the lead time from the central depot to retailer i, i = 1, . . . , N;
70 MULTI-ECHELON INVENTORY SYSTEMS
R R R R
L
0
L
0
L
0
L
0
— — — — —
R R R
L
1
L
1
L
1
— — — — — —
L
N
L
N
L
N
Figure 3.5: The review periods at the two echelons shifted in time (depot above, retailers below).
• R: the review period (this period is of equal length for both echelons, but it is shifted in time by an
amount L
0
for the retailers in comparison to the central depot; see Figure 3.5);
• D
i,t
: the demand at retailer i during a period t; this is a random variable with time-independent mean
µ
i
.
= E¦D
i,t
¦ and standard deviation σ
i
.
= σ¦D
i,1
¦, i = 1, . . . , N;
• h
i
: the holding cost per unit per unit of time at retailer i, i = 1, . . . , N;
• β
i
: the target fill rate at retailer i, i = 1, . . . , N.
It will be assumed that demand is independent across time periods and that demand is independent across
retailers. Since the review period is given, the ordering and shipping costs are fixed. The lead times are
assumed to be known and constant, and the approach below especially concerns situations where the lead
time to the central depot is large in comparison to the lead times to the retailers (L
0
¸ L
1
, . . . , L
N
). The
decision variables are:
• S
0
: the order-up-to level of the central depot;
• S

i
: the maximum order-up-to level of retailer i, i = 1, . . . , N;
• f
i
: the rationing fraction for allocating units to retailer i, i = 1, . . . , N.
Here, the variables are to be chosen such that

N
i=1
S

i
= S
0
and

N
i=1
f
i
= 1. The control mechanism of
the inventory policy to be discussed uses the following auxiliary random variables:
• z
i
: the inventory position of retailer i just before a reorder instant of the retailers, i = 1, . . . , N;
• S
i
: the inventory position of retailer i just after a reorder instant of the retailers, i = 1, . . . , N;
• Q: the order quantity of the central depot.
When an order arrives at the central depot, the echelon inventory position of the central depot is S
0

˜
D
L
0
,
with
˜
D
L
0
.
=

N
i=1
D
i,L
0
the total demand at all retailers during the lead time L
0
. The amount that is
allocated to retailer i by the central depot at such an instant is S
i
−z
i
with
S
i
= S

i
−f
i
˜
D
L
0
, i = 1, . . . , N. (3.21)
Hence, the maximum order-up-to level of retailer i is reduced by a fraction f
i
of the total demand
˜
D
L
0
during
the lead time to the central depot L
0
, for i = 1, . . . , N. The echelon inventory position just before the central
depot places an order is S
0
− Q. The echelon inventory on hand just before the order of the central depot
arrives is:
N

j=1
z
j
= S
0
−Q−
N

i=1
D
i,L
0
= S
0
−Q−
˜
D
L
0
.
Hence, the allocation to retailer i can alternatively be described by S
i
−z
i
, with
S
i
= S

i
−f
i
_
_
S
0
−Q−
N

j=1
z
j
_
_
, i = 1, . . . , N.
The echelon inventory position just after the delivery of an order at the central depot is:
N

i=1
S
i
= S
0

˜
D
L
0
=
N

j=1
z
j
+Q.
HANS BLANC, TILBURG UNIVERSITY 71
A problem in the foregoing allocation arises if S
i
− z
i
< 0 for some retailer i = i
0
. This phenomenon is
indicated by imbalance. The cause for imbalance is
• little demand at retailer i
0
(the remaining amount of stock z
i
0
is relatively high);
• more demand at other retailers (the total demand
˜
D
L
0
is such that S
i
0
is smaller than z
i
0
, cf. (3.21)).
In case of imbalance, the allocation rule would require redistribution of stock from retailer i
0
to some other
retailers, which would lead to extra shipping cost. Let us study the occurrence of imbalance. The inventory
position at retailer i just after the order quantity Q of the central depot has been reallocated is, cf. (3.21),
S
(1)
i
= S

i
−f
i
˜
D
L
(1)
0
, i = 1, . . . , N. (3.22)
Here, it is assumed that there is no imbalance at the outset. The inventory position at retailer i just before
the next allocation instant, that is, a review period R later, is
z
(2)
i
= S

i
−f
i
˜
D
L
(1)
0
−D
i,R
, i = 1, . . . , N. (3.23)
This inventory level must be raised to the target level
S
(2)
i
= S

i
−f
i
˜
D
L
(2)
0
, i = 1, . . . , N. (3.24)
Imbalance occurs if the target level S
(2)
i
is smaller than the actual inventory level z
(2)
i
before allocation.
Hence, the imbalance due to retailer i is:

i
.
= [z
(2)
i
−S
(2)
i
]
+
= [f
i
˜
D
L
(2)
0
−f
i
˜
D
L
(1)
0
−D
i,R
]
+
, i = 1, . . . , N. (3.25)
This definition of imbalance was introduced by De Kok [21]; Zipkin [83] used another type of imbalance
measure. It is seen that the imbalance only depends on the rationing factors f
i
, and not on the order-up-to
levels S

i
, i = 1, . . . , N. To avoid the occurrence of imbalance as much as possible, the ratios f
i
will be chosen
in such a way that the total expected imbalance is minimal. To this end, Van der Heijden [68] assumes that
the random variables, cf. (3.23), (3.24),
Y
i
.
= z
(2)
i
−S
(2)
i
= f
i
˜
D
L
(2)
0
−f
i
˜
D
L
(1)
0
−D
i,R
, i = 1, . . . , N, (3.26)
approximately possess a normal distribution. Note that the random variables Y
i
, i = 1, . . . , N, can take
positive and negative values, but are not symmetric around their mean. The mean of Y
i
is independent of f
i
:
E¦Y
i
¦ = E¦z
(2)
i
−S
(2)
i
¦ = −E¦D
i,R
¦ = −Rµ
i
, i = 1, . . . , N. (3.27)
Note that the terms on the righthand side of (3.26) contain overlapping demand since the second lead time
L
(2)
0
partly coincides with the review period R at the retailers and possibly with the previous lead time L
(1)
0
.
To determine the variance of Y
i
, two cases are distinguished, cf. Figure 3.6:
• R > L
0
: then the contributions to Y
i
are:
period (0, L
0
): −f
i
˜
D
L
(1)
0
,
period (L
0
, R): −D
i,R−L
0
,
period (R, L
0
+R): f
i

j=i
D
j,L
0
−(1 −f
i
)D
i,L
0
;
• R < L
0
: then the contributions to Y
i
are:
period (0, R): −f
i
˜
D
R
,
period (R, L
0
): 0 (since L
(1)
0
and L
(2)
0
overlap in this period),
period (L
0
, L
0
+R): f
i

j=i
D
j,R
−(1 −f
i
)D
i,R
.
With M
.
= min¦R, L
0
¦, the variance of Y
i
can be written as
σ
2
¦Y
i
¦ = [R −2f
i
M] σ
2
i
+ 2f
2
i
M
N

j=1
σ
2
j
, i = 1, . . . , N. (3.28)
72 MULTI-ECHELON INVENTORY SYSTEMS
L
(1)
0
R
— — — — — — — —
(R) L
(2)
0
L
(1)
0
R
— — — — —
(R) L
(2)
0
0 0 L
0
R R L
0
L
0
+R L
0
+R
Figure 3.6: Review periods at the retailers and lead times to the depot: two cases.
The mean imbalance E¦Ω
i
¦ = E¦[Y
i
]
+
¦ is increasing with the variance σ
2
¦Y
i
¦ when Y
i
has a normal distri-
bution, i = 1, . . . , N. It follows by differentiation of (3.28) that the variance σ
2
¦Y
i
¦ is minimal with respect
to f
i
at
ˆ
f
i
=
1
2
σ
2
i
/

N
j=1
σ
2
j
, i = 1, . . . , N. However, these values of f
i
, i = 1, . . . , N, are not feasible with
respect to the constraint

N
i=1
f
i
= 1. Therefore, the constrained optimization problem has to be solved with
the Lagrange multiplier method. In order to minimize the function
N

i=1
E¦Ω
i
¦ +η
_
1 −
N

i=1
f
i
_
, (3.29)
with η the Lagrange multiplier, the following equations must be satisfied:
d
df
j
N

i=1
E¦Ω
i
¦ = η, j = 1, . . . , N;
N

i=1
f
i
= 1. (3.30)
Under the assumption of normally distributed random variables Y
i
, the total expected imbalance reads, cf.
(3.25), (3.26),
N

i=1
E¦Ω
i
¦ =
N

i=1
E¦[Y
i
]
+
¦ =
N

i=1
σ¦Y
i
¦NL(Rµ
i
/σ¦Y
i
¦); (3.31)
here, NL(.) denotes the normal loss function defined by (A.18). The derivatives of this function with respect
to the rationing factors f
i
can be determined by means of the chain rule. From (3.28) and the properties of
the normal loss function, cf. Appendix A.2.1, it follows that, for j = 1, . . . , N,
d
df
j
N

i=1
E¦Ω
i
¦ =
d
dσ¦Y
j
¦
E¦[Y
j
]
+
¦
d
df
j
σ¦Y
j
¦ =
Mf
0,1
(Rµ
j
/σ¦Y
j
¦)
σ¦Y
j
¦
_
−σ
2
j
+ 2f
j
N

i=1
σ
2
i
_
; (3.32)
here, f
0,1
(.) denotes the density of the standard normal distribution. Since σ¦Y
j
¦ depends on f
j
, cf. (3.28),
it is not possible to determine an explicit expression for the rationing fraction f
j
for which dE¦Ω
j
¦/df
j
= η
for a given value of η. Therefore, a numerical search procedure must be applied to find this root, e.g., the
bisection method for finding a root of an equation, cf. Appendix B.3, with initial search interval [
ˆ
f
j
, 1], for
j = 1, . . . , N. Further, in order to find the value of η for which the fractions f
j
sum to one another application
of the bisection method is required. In this way, the minimum total expected imbalance with respect to the
fractions f
j
, j = 1, . . . , N, can be obtained numerically. Because of this total expected imbalance minimizing
property, the above rationing rule is referred to as “Balanced Stock” rationing.
In the next step of the approach by Van der Heijden [68] it is assumed that the fractions f
i
are fixed according
to the above procedure, and the maximum order-up-to levels S

i
are determined such that the target fill rates
β
i
are reached, i = 1, . . . , N. In this step, it is assumed that by the choice of the rationing fractions the
occurrence of imbalance can be ignored. As a consequence, retailer i is allocated a positive order quantity
Q
i
= S
i
−z
i
at each order arrival instant at the central depot, with mean, cf. (3.27),
E¦Q
i
¦ = E¦S
i
−z
i
¦ = E¦D
i,R
¦ = Rµ
i
, i = 1, . . . , N. (3.33)
The mean stockout B
i
accumulated during a review period R at retailer i is, cf. (1.38),
E¦B
i
¦ = E¦[D
i,L
i
+R
−S
i
]
+
¦ −E¦[D
i,L
i
−S
i
]
+
¦, i = 1, . . . , N. (3.34)
HANS BLANC, TILBURG UNIVERSITY 73
In the present case, both D
i,L
i
+R
, D
i,L
i
and S
i
are random variables, since S
i
depends on
˜
D
L
0
, for i =
1, . . . , N, cf. (3.21). In order to reach the target fill rates, the maximum order-up-to levels S

i
have to be
determined such that, cf. (1.38),
E¦[D
i,L
i
+R
+f
i
˜
D
L
0
−S

i
]
+
¦ −E¦[D
i,L
i
+f
i
˜
D
L
0
−S

i
]
+
¦ ≤ (1 −β
i
)Rµ
i
, i = 1, . . . , N. (3.35)
Note that the smallest S

i
, i = 1, . . . , N, that satisfy the above inequalities, can be determined for each
retailer separately (if imbalance is ignored). Such a problem involves random variables
X
(1)
i
= D
i,L
i
+R
+f
i
˜
D
L
0
, X
(2)
i
= D
i,L
i
+f
i
˜
D
L
0
, i = 1, . . . , N. (3.36)
Since the lead time L
0
to the central depot occurs before L
i
+ R, it follows that
˜
D
L
0
and D
i,L
i
+R
are
independent random variables. The means of these variables are:
E¦X
(1)
i
¦ = (L
i
+R)µ
i
+f
i
L
0
N

j=1
µ
j
, E¦X
(2)
i
¦ = L
i
µ
i
+f
i
L
0
N

j=1
µ
j
, i = 1, . . . , N. (3.37)
The variances of these variables are:
σ
2
¦X
(1)
i
¦ = (L
i
+R)σ
2
i
+f
2
i
L
0
N

j=1
σ
2
j
, σ
2
¦X
(2)
i
¦ = L
i
σ
2
i
+f
2
i
L
0
N

j=1
σ
2
j
, i = 1, . . . , N. (3.38)
The distributions of X
(1)
i
and X
(2)
i
cannot readily be obtained for general probability distributions of the
demand per period at the various retailers. Therefore, Van der Heijden [68] approximates the distributions
of X
(1)
i
and X
(2)
i
by mixtures of Erlang distributions with mean and variance given by (3.37) and (3.38),
respectively, cf. Appendix A.2.5. With these approximations, the smallest maximum order-up-to levels S

i
satisfying (3.35) can be determined numerically. The computation of the decision variables for the replenish-
ment strategy is summarized below.
Algorithm 3.3 [Balanced Stock rationing]
Step 1: Compute the rationing fractions f
i
, i = 1, . . . , N, such that the expected total imbalance is
minimal, with the aid of the Lagrange multiplier method. This step requires a numerical procedure to
find the value of the Lagrange multiplier η for which

N
i=1
f
i
= 1. For each considered value of η, it
requires a numerical procedure to find the value of f
i
for which
d
df
i
E¦Ω
i
¦ = η, for each i, i = 1, . . . , N.
Step 2: Given the values of the fractions f
i
, i = 1, . . . , N, compute the smallest maximum order-up-to
levels S

i
such that the target fill rates β
i
are reached, for i = 1, . . . , N. The actual fill rates are based
on an approximation by a mixture of Erlang distributions.
The order-up-to level at the central depot follows as S
0
=

N
i=1
S

i
.
Note that there are no decision variables left to minimize the total holding cost. This issue could be addressed
in an extension of the present model with a stock-keeping central depot, cf. Van der Heijden [69]. Then,
items could be stored at the central depot in case the demand
˜
D
L
0
during the lead time to the central depot
is much less than average and if the holding cost h
0
per item per unit of time at the central depot is lower
than those at the retailers. In the present model, the total cost is fixed after application of Algorithm 3.3.
The expected average inventory on hand at retailer i is, cf. (1.39):

¯
I
i
¦ =
1
2
[E¦[S
i
−D
i,L
i
]
+
¦ +E¦[S
i
−D
i,L
i
+R
]
+
¦], i = 1, . . . , N. (3.39)
Since for any real x, [x]
+
= x + [−x]
+
, it follows with (3.21) that
E¦[S
i
−D
i,L
i
+R
]
+
¦ = E¦S
i
¦ −(L
i
+R)µ
i
+E¦[D
i,L
i
+R
+f
i
˜
D
L
0
−S

i
]
+
¦, i = 1, . . . , N, (3.40)
and
E¦[S
i
−D
i,L
i
]
+
¦ = E¦S
i
¦ −L
i
µ
i
+E¦[D
i,L
i
+f
i
˜
D
L
0
−S

i
]
+
¦, i = 1, . . . , N, (3.41)
with
E¦S
i
¦ = S

i
−f
i
L
0
N

j=1
µ
j
, i = 1, . . . , N. (3.42)
74 MULTI-ECHELON INVENTORY SYSTEMS
Hence, the expected average inventory on hand E¦
¯
I
i
¦ at retailer i can readily be computed since the expec-
tations on the right-hand sides of (3.40) and (3.41) have already been calculated when S

i
was determined
satisfying the inequalities (3.35), i = 1, . . . , N. Further, the average safety stock at retailer i is, cf. Section
1.2.7:
E¦V
i
¦ = E¦S
i
¦ −(R +L
i

i
= S

i
−f
i
L
0
N

j=1
µ
j
−(R +L
i

i
, i = 1, . . . , N. (3.43)
Finally, the average inventory in transit to retailer i is L
i
µ
i
, independent of the policy (due to the assumption
of complete backlogging).
Table 3.7: Data and results for Example 3.7 (in weeks).
Central depot Retailer 1 Retailer 2 Retailer 3
Mean demand µ
i
(1200) 100 300 800
St. deviation demand σ
i
(412) 85 50 400
Target fill rate β
i
0.95 0.90 0.70
Lead time L
i
2 1 1 1
Rationing fraction f
i
0.031 0.189 0.780
Maximum order-up-to level S

i
5094 480 1109 3505
Mean on-hand inventory E¦
¯
I
i
¦ 258 220 587
Mean stockout E¦B
i
¦ 5 30 240
Mean safety stock E¦V
i
¦ 205 55 34
Mean imbalance E¦Ω
i
¦ 4.86 0.19 1.11
Simulated mean on-hand inventory 257 219 585
Simulated actual fill rate 0.948 0.898 0.698
Simulated mean imbalance 0.5 0.4 0.0
Example 3.7 Van der Heijden [68] contains a numerical example of which the data can be found in Table
3.7. The review period is R = 1 week. The most critical retailer is retailer 1, since the coefficient of variation
σ
1

1
= 0.85 and the target fill rate β
1
are maximal over all retailers. This retailer gets the smallest rationing
fraction. Simulation indicates that the actual fill rates are slightly below the targets (in the order of 0.002).
The simulated mean imbalance turns out to be much smaller than the mean imbalance predicted by the
normal loss function in the model.
Remark 3.3 De Kok [21] introduced another rationing policy for the present multi-echelon inventory prob-
lem called Consistent Appropriate Share (CAS) rationing. In this approach, items are allocated to the
retailers based on safety stock ratios; see also Verrijdt & De Kok [74, 75]. At the order arrival instant of the
central depot, the items are allocated to the retailers such that their inventories are raised to the levels
S
i
= E¦D
i,L
i
+R
¦ +p
i
_
_
S
0

˜
D
L
0

N

j=1
E¦D
j,L
j
+R
¦
_
_
, i = 1, . . . , N;
that is, retailer i gets allocated its expected demand during the period L
i
+R plus a fraction of the remaining
amount of items, i = 1, . . . , N. These fractions are chosen proportional to the safety stocks:
p
i
= cK
i
σ¦D
i,L
i
+R
¦, i = 1, . . . , N;
with K
i
a safety factor such that the target fill rate β
i
is reached, i = 1, . . . , N, and C a normalization
constant such that

N
i=1
p
i
= 1. This allocation falls in the framework of the more general allocation rule
(3.21) by identifying
f
i
= p
i
, S

i
= E¦D
i,L
i
+R
¦ +p
i
_
_
S
0

N

j=1
E¦D
j,L
j
+R
¦
_
_
, i = 1, . . . , N.
The fraction p
i
is relatively large if σ¦D
i,L
i
+R
¦ is large, meaning high uncertainty in demand, and/or if K
i
is large, meaning a high target fill rate β
i
, i = 1, . . . , N. But the allocation to a retailer with relatively high
HANS BLANC, TILBURG UNIVERSITY 75
fraction f
i
= p
i
is most sensitive to variations in the total demand
˜
D
L
0
, cf. (3.21). Hence, most sensitivity is
directed to retailers of which the inventory is already difficult to control. This explains why CAS rationing
yields strong imbalance and erroneous results in some cases where simulation results do not agree with model
predictions, in particular with respect to the actual fill rates, even after improvements of the method for
cases when some fractions p
i
are negative, cf. Van der Heijden [68]. For instance, CAS rationing assigns a
large fraction to retailer 1 (p
1
> 0.95) in Example 3.7. The Balanced Stock rationing is better in avoiding
negative allocations and is better in predicting the actual performance of a system, while it leads to lower
mean stock levels than CAS rationing.
Table 3.8: Data and results for Example 3.8 (in weeks) with review period R = 1 week.
Depot Retailer 1 Retailer 2 Retailer 3 Retailer 4 Retailer 5
Mean demand µ
i
(800) 100 100 100 400 100
St. deviation demand σ
i
(240) 50 100 50 200 50
Target fill rate β
i
0.95 0.95 0.90 0.95 0.95
Lead time L
i
2 0.5 0.5 0.5 0.5 1.0
Rationing fraction f
i
0.078 0.116 0.078 0.650 0.078
Maximum order-up-to level S

i
3715 354 555 323 2063 419
Mean on-hand inventory E¦
¯
I
i
¦ 132 272 104 632 148
Mean stockout E¦B
i
¦ 5 5 10 20 5
Mean safety stock E¦V
i
¦ 80 219 49 422 95
Mean imbalance E¦Ω
i
¦ 0.60 7.40 0.60 1.27 0.60
Example 3.8 To study the effect of various parameter values, consider the following numerical example with
five retailers of which the data can be found in Table 3.8. The review period is R = 1 week. The parameters
of retailer 1 form the base case. Retailer 2 has a higher coefficient of variation of the demand, and is the most
critical retailer. Retailer 3 has a lower target fill rate. Retailer 4 has a higher mean demand with the same
coefficient of variation. Retailer 5 has a higher lead time. In this case, retailer 4 gets the largest rationing
fraction, then retailer 2 (in the order of decreasing standard deviation of the demand, as suggested by the
optimal solution to the unconstrained problem, cf. above (3.29)), and retailers 1, 3 and 5 get the smallest
rationing fraction which is the same for these three retailers because the rationing fractions do not depend
on the lead times and the target fill rates. To study the effect of the review period, Table 3.9 contains results
for the same two-echelon system but with a review period of R = 2 weeks. Increasing the review period shifts
the rationing fractions somewhat in favor of retailer 4 with the highest standard deviation of the demand.
Further, the predicted mean imbalance decreases considerably, but remains largest for retailer 2 with the
highest coefficient of variation of the demand.
Table 3.9: Results for Example 3.8 (in weeks) with review period R = 2 weeks.
Depot Retailer 1 Retailer 2 Retailer 3 Retailer 4 Retailer 5
Rationing fraction f
i
0.075 0.095 0.075 0.680 0.075
Maximum order-up-to level S

i
4448 444 629 403 2469 504
Mean on-hand inventory E¦
¯
I
i
¦ 179 332 143 800 189
Mean stockout E¦B
i
¦ 10 10 20 40 10
Mean safety stock E¦V
i
¦ 74 227 33 380 84
Mean imbalance E¦Ω
i
¦ 0.08 4.15 0.08 0.17 0.08
The inventory policy described in this section can be extended in several directions, cf. Van der Heijden
[68, 69]. It can be generalized to divergent multi-echelon systems with more than two echelons. It can deal
with nonstationary demand by adjusting the maximum order-up-to levels S

i
, i = 1, . . . , N (the fractions f
i
should be recomputed only if the demand or lead time characteristics change considerably). Finally, the model
can be extended with a stock keeping central depot, but then the expressions for the actual fill rates and the
procedure for determining the rationing fractions become more complicated and require more approximative
assumptions.
76 MULTI-ECHELON INVENTORY SYSTEMS
Literature on this subject includes Eppen & Schrage [24] who consider a model with penalty cost, Federgruen
[27] who applies stochastic dynamic programming, Inderfurth [40] who provides a survey on safety stocks in
divergent multi-echelon systems, Odanaka et al. [53] who consider a multi-echelon production system, Yoo
et al. [81] who are concerned with distribution requirement planning, Korugan & Gupta [46] who study a
multi-echelon system with returns, Axs¨ater & Zhang [8] and Axs¨ater [6] who provide an exact analysis of a
continuous review system with (compound) Poisson demand, and Ettl et al. [25] who solve a supply network
with base-stock control and service level constraints by the conjugate gradient method.
Exercise 3.4 Complete the details of the derivation of the variance σ
2
¦Y
i
¦, i = 1, . . . , N, in (3.28).
Exercise 3.5 Complete the details of the derivation of the derivative of the total expected imbalance with
respect to the rationing fractions in (3.32).
Exercise 3.6 Consider the results presented in Table 3.7. Explain why the mean physical stock is only 585
for retailer 3 under Balanced Stock rationing while the maximum order-up-to level for this stocking point is
3505.
Exercise 3.7 Consider the results presented in Tables 3.8, 3.9. Explain why the mean stockout seems to
double for each retailer when the review period doubles, while the maximum order-up-to levels and the mean
physical stocks only increase with a much smaller percentage.
Exercise 3.8 Consider a two-echelon system for an item with a central depot and two retailers. The lead
time to the central depot is L
0
= 2 weeks and the lead time to the retailers are L
1
= 0.1 and L
2
= 0.2 week,
respectively. The weekly demand has a mean of 20 units and a standard deviation of 5 units at both retailers.
Determine ordering policies according to Algorithm 3.3 at target fill rates of 0.90 for both retailers, for a
review period of R = 1 week as well as for a review period of R = 2 weeks.
Chapter 4
Other Inventory Systems with
Interactions
This chapter is devoted to some other inventory problems with interaction. Section 4.1 is concerned with the
inventory control of a set of items that have to share a common storage space. Section 4.2 deals with a single
item of which units may fail and then are replaced by another unit while the failed unit may be repaired
later on.
4.1 Inventory capacity constraint
In this section a set of N items is considered which are ordered independently but which have to share a
common storage space. The demand of all items is assumed to be deterministic and continuous in time. The
parameters of the model are:
• a
i
: the ordering cost for item i, i = 1, . . . , N;
• h
i
: the holding cost per unit per unit of time for item i, i = 1, . . . , N;
• D
i
: the demand per unit of time for item i, i = 1, . . . , N;
• φ
i
: the space occupied by a unit of item i, i = 1, . . . , N;
• Φ: the total inventory capacity.
The decision variables are:
• Q
i
: the order quantity for item i, i = 1, . . . , N.
The average cost per unit of time is found as a simple generalization of the single item case, cf. (1.1),
C(Q)
.
= C(Q
1
, . . . , Q
N
) =
N

i=1
_
a
i
D
i
Q
i
+
1
2
h
i
Q
i
_
. (4.1)
This cost function has to be minimized subject to the capacity constraint
N

i=1
φ
i
Q
i
≤ Φ. (4.2)
This constraint is based on the supposition that a space φ
i
Q
i
has to be reserved for the maximum inventory
level of item i, i = 1, . . . , N, since the ordering of the various items is not coordinated.
Algorithm 4.1 [Lagrange constrained optimization]
Step 1: Compute the unconstrained optimal order quantities, cf. (1.2),
Q

i
=
_
2a
i
D
i
/h
i
, i = 1, . . . , N.
77
78 OTHER INVENTORY SYSTEMS WITH INTERACTIONS
Step 2: Verify whether the capacity constraint (4.2) is satisfied. If so, stop. Otherwise, continue with
Step 3.
Step 3: Solve the constrained optimization problem with the aid of a Lagrange multiplier, that is, min-
imize the following function with respect to the vector Q and the scalar η:
f(Q, η) =
N

i=1
_
a
i
D
i
Q
i
+
1
2
h
i
Q
i
_

_
N

i=1
φ
i
Q
i
−Φ
_
. (4.3)
The last step requires a numerical search procedure (e.g., bisection).
The partial derivative of the function f(Q, η) with respect to Q
i
is
d
dQ
i
f(Q, η) = −a
i
D
i
Q
2
i
+
1
2
h
i
+ηφ
i
, i = 1, . . . , N.
This implies that the optimal order quantity for fixed η is
ˆ
Q
i
(η) =
¸
2a
i
D
i
h
i
+ 2ηφ
i
, i = 1, . . . , N. (4.4)
Further, the inventory capacity is fully used (consider the partial derivative of f(Q, η) with respect to η) so
that
N

i=1
φ
i
¸
2a
i
D
i
h
i
+ 2ηφ
i
= Φ. (4.5)
This relation implicitly determines the scalar η. The value of η has to be determined with the aid of a
numerical procedure, e.g., by the bisection method for finding a root of an equation, cf. Appendix B.3. To
this end, it is useful to have an upper and lower bound on the optimal value of η. Note that by (4.4),
η =
a
i
D
i
φ
i
Q
2
i

h
i

i
, i = 1, . . . , N.
Since Q
i
=
1
N
Φ/φ
i
, i = 1, . . . , N, is a feasible solution, it follows that the optimal value of η lies on the
interval
0 ≤ η ≤ max
i=1,...,N
_
a
i
D
i
φ
i
N
2
Φ
2

h
i

i
_
. (4.6)
Table 4.1: Data for Example 4.1 and results according to Algorithm 4.1.
product i: 1 2 3 4 5 6 7 8
a
i
$10 $10 $10 $40 $10 $10 $10 $40
h
i
$0.20 $0.80 $0.20 $0.80 $0.20 $0.80 $0.20 $0.80
φ
i
1.0 1.0 1.0 1.0 4.0 4.0 4.0 4.0
D
i
400 400 100 400 400 400 100 400
Φ ≥ 3000 Q
i
200.0 100.0 100.0 200.0 200.0 100.0 100.0 200.0
C
i
$40.00 $80.00 $20.00 $160.00 $40.00 $80.00 $20.00 $160.00
T
i
0.500 0.250 1.000 0.500 0.500 0.250 1.000 0.500
Φ = 2400 Q
i
174.7 96.3 87.3 192.7 133.5 87.3 66.8 174.7
C
i
$40.37 $80.06 $20.18 $160.11 $43.31 $80.73 $21.66 $161.47
T
i
0.437 0.241 0.873 0.482 0.334 0.218 0.668 0.437
Φ = 1800 Q
i
136.8 88.2 68.4 176.5 84.9 68.4 42.5 136.8
C
i
$42.92 $80.63 $21.46 $161.25 $55.61 $85.84 $27.80 $171.68
T
i
0.342 0.221 0.684 0.441 0.212 0.171 0.424 0.342
Φ = 1200 Q
i
91.5 71.7 45.7 143.4 49.8 45.7 24.9 91.5
C
i
$52.88 $84.47 $26.44 $168.94 $85.29 $105.75 $42.64 $211.51
T
i
0.229 0.179 0.457 0.359 0.125 0.114 0.249 0.229
HANS BLANC, TILBURG UNIVERSITY 79
Example 4.1 Table 4.1 contains data (a
i
, h
i
, φ
i
, D
i
, i = 1, . . . , 8) for a set of eight products that have
to share storage space. The unconstrained optimal order quantities require a total storage space of 3000
volume units. Table 4.1 further displays the optimal order quantities computed according to Algorithm 4.1
at storage capacity constraints of 80%, 60% and 40% of the maximally required storage space of 3000 volume
units. The table also shows the corresponding ordering and holding cost per item, C
i
.
= a
i
D
i
/Q
i
+
1
2
h
i
Q
i
,
and the reorder cycle T
i
= Q
i
/D
i
for each item, i = 1, . . . , 8. There is no clear structure in the decrease of
the order quantities: the order quantity of some items decreases more strongly when the storage capacity
decreases from 80% to 60% than when it decreases from 100% to 80%, for other items it is the other way
about. The total average cost increases from $600.00 at Φ = 3000 via $607.89 at Φ = 2400 and $647.20 at
Φ = 1800 to $777.91 at Φ = 1200.
The inventory optimization problem with a constraint on the total budget available for investment in inven-
tories can be solved with a similar algorithm as Algorithm 4.1. Dagpunar [19] discusses such an optimization
problem including a joint family ordering cost.
Exercise 4.1 Consider two items with the same ordering cost a, the same holding cost h and the same volume
φ. Only the demand rates differ: D
1
= 4D
2
. Determine the ratio between the optimal order quantities Q

1
and Q

2
both in the unconstrained and in the constrained case, with inventory capacity Φ < 3φ
_
2aD
2
/h.
Exercise 4.2 Consider four items with the same ordering cost a = $25 and the same demand rate D = 200.
The holding costs are h
1
= h
4
= $1, h
2
= h
3
= $4 per unit per unit of time, and the areas occupied per
unit are φ
1
= φ
2
= 1, φ
3
= φ
4
= 4 square meters. Units cannot be piled. The available area for storage
is Φ = 700 square meters. Determine the optimal order quantities Q

i
, i = 1, 2, 3, 4. Can you find an ad
hoc solution such that the four items are ordered according to their individual economic order quantities?
Finally, determine the optimal order quantities for the case that the available area for storage is Φ = 500
square meters.
Exercise 4.3 Prove that equation (4.5) indeed has a root in η on the interval (4.6) if Step 3 of Algorithm
4.1 is executed.
4.2 Repairable items
This section is devoted to the inventory management of spare parts that occur in machines like computers,
copiers and printers. It is assumed that a unit that fails is replaced by a spare part and that the failed units
are later repaired, if possible. Further, it is assumed that repaired units are as good as new ones. A certain
fraction of the failed units cannot be repaired. Therefore, new spare parts have to be purchased from an
external supplier from time to time. The parameters of the model are:
• D: the demand per unit of time of units for repair;
• w: the fraction of failed units that is no longer repairable (the waste rate);
• a
O
: the ordering cost for placing an order at the external supplier;
• a
R
: the set-up cost for starting a repair batch;
• h
G
: the holding cost per unit per unit of time for good (new and repaired) units;
• h
F
: the holding cost per unit per unit of time for failed, but repairable units;
• L
O
: the lead time from the external supplier;
• L
R
: the lead time of a repair batch.
The demand of units for replacement is assumed to be deterministic and continuous in time. Also, the lead
times are deterministic. In most situations, it will hold that h
G
> h
F
. The following inventory policy will be
studied, cf. Schrady [57]. The order quantity for new items is always the same, and also the repair lot size is
always the same. The decision variables are:
• Q
O
: the order quantity from the external supplier;
80 OTHER INVENTORY SYSTEMS WITH INTERACTIONS
¸
`

I
G
(t)
t →
Q
R
Q
O
0 T 2T −u T −u T −2u 2T −u
`
`
`
`
`
`
`
`
`
` s s s s
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
s s s s
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
` s s s s
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
¸
`

I
F
(t)
t →
Q
R
0 T 2T −u T −u T −2u 2T −u
s s s s s s s s s s s s
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
¸
`

I(t)
t →
(1−w)Q
R
(1−w)Q
R
+Q
O
0 T 2T −u T −u T −2u 2T −u
s s s

s s s s s s s s s s s s
Figure 4.1: Inventory on hand of good units (top), of repairable units (middle) and of usable units (bottom).
• Q
R
: the lot size for repairs;
• s
O
: the reorder point for an external order;
• s
R
: the point for starting a repair batch.
The policy is such that in between two external orders n repair batches are performed, for some integer n,
n ≥ 1. This puts a constraint on the ratio of Q
O
and Q
R
. First, the case will be considered that the lead
time of a repair batch is negligible (L
R
= 0).
Let T denote the time between the arrival of two consecutive orders of new spare parts. In a cycle of length
T, the number of incoming good units must be equal to the number of outgoing good units:
Q
O
+nQ
R
= DT. (4.7)
Further, in a cycle of length T, the number of incoming units from the external supplier must be equal to
the number of outgoing units as waste:
Q
O
= wDT. (4.8)
Together, these balance relations imply that
nQ
R
= (1 −w)DT = (1 −w)Q
O
/w. (4.9)
HANS BLANC, TILBURG UNIVERSITY 81
For the time interval u between two consecutive repair batches within a reorder cycle it must hold that
Q
R
= Du. (4.10)
As a consequence, it follows with (4.9) that the fraction of time that the system runs on repaired units is
nu
T
=
nQ
R
DT
= 1 −w. (4.11)
Figure 4.1 (upper graph) shows the inventory on hand I
G
(t) of good units for the case n = 4. This inventory
level follows a somewhat irregular saw-tooth pattern, since an external order is succeeded by n repair batches.
In between replenishments, this inventory decreases at rate D. The average inventory on hand is
1
2
Q
O
during
a fraction (T −nu)/T of a cycle and is
1
2
Q
R
during a fraction nu/T of a cycle. Hence, the average inventory
on hand of good units follows with (4.11) as:
¯
I
G
=
1
2
Q
O
(T −nu) +
1
2
Q
R
nu
T
=
1
2
wQ
O
+
1
2
(1 −w)Q
R
. (4.12)
The middle graph of Figure 4.1 shows the inventory on hand I
F
(t) of failed but repairable units. In between
repair batches, this inventory increases at rate (1 − w)D. This inventory level decreases by Q
R
at instants
when the inventory on hand I
G
(t) of good units reaches the level 0. The lower graph of Figure 4.1 shows the
total inventory on hand I(t)
.
= I
G
(t)+I
F
(t) of good and repairable units. In between external replenishments,
this inventory decreases at rate wD. This inventory level follows a saw-tooth pattern which is lifted above the
level 0. The latter is due to the fact that a number of repairable units has accumulated when the inventory
on hand I
G
(t) of good units reaches the level 0 just before the arrival of an external order. This amount is
(1−w)Du = (1−w)Q
R
, cf. (4.10). The average total inventory of good and repairable units on hand is
1
2
Q
O
plus this basic amount:
¯
I
G
+
¯
I
F
= (1 −w)Q
R
+
1
2
Q
O
. (4.13)
Hence, the average inventory of failed but repairable units simply follows as
¯
I
F
= (1 −w)Q
R
+
1
2
Q
O

1
2
wQ
O

1
2
(1 −w)Q
R
=
1
2
(1 −w)Q
R
+
1
2
(1 −w)Q
O
. (4.14)
The foregoing implies that the average cost per unit of time with this policy is:
C(Q
O
, Q
R
) = D
_
a
O
w
Q
O
+
a
R
(1 −w)
Q
R
_
+
1
2
h
G
[wQ
O
+ (1 −w)Q
R
] +
1
2
h
F
(1 −w)[Q
O
+Q
R
], (4.15)
or, with Q
R
= (1 −w)Q
O
/(nw), cf. (4.9),
C(Q
O
, n) =
wD
Q
O
[a
O
+na
R
] +
1
2
Q
O
_
h
G
_
w +
(1 −w)
2
nw
_
+h
F
(1 −w)
_
1 +
1 −w
nw
__
. (4.16)
For fixed n, the optimal external order quantity Q
O
is:
ˆ
Q
O
(n) = w
¸
2nD(a
O
+na
R
)
nw[h
G
w +h
F
(1 −w)] + (1 −w)
2
[h
G
+h
F
]
, (4.17)
with corresponding minimum average cost per unit of time
ˆ
C(n) =
¸
2wD(a
O
+na
R
)
_
h
G
w +h
F
(1 −w) +
(1 −w)
2
nw
(h
G
+h
F
)
_
. (4.18)
Since the square root function is monotonously increasing, the optimal real-valued minimum of this cost
function is obtained by solving the equation
d
dn
ˆ
C
2
(n) = 2D
d
dn
¦[
1
n
a
O
+a
R
](1 −w)
2
(h
G
+h
F
) + [a
O
+na
R
]w[h
G
w +h
F
(1 −w)]¦
= 2D¦−
1
n
2
a
O
(1 −w)
2
(h
G
+h
F
) +a
R
w[h
G
w +h
F
(1 −w)]¦ = 0.
82 OTHER INVENTORY SYSTEMS WITH INTERACTIONS
¸
`

I
G
(t)
t →
Q
R
Q
O
0 T 2T −u T −u T −2u 2T −u
`
`
`

`
`
`
`
` s s s s
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`

s s s s
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`

s s s s
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
¸
`

I
F
(t)
t →
Q
R
0 T 2T −u T −u T −2u 2T −u
´
´
s s s s s s s s s s s s ´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
Figure 4.2: Inventory on hand of good units (top) and of repairable units (bottom) with release during repair.
Hence, the optimal real-valued n minimizing the foregoing cost function is:
n

= (1 −w)
¸
a
O
(h
G
+h
F
)
a
R
w[h
G
w +h
F
(1 −w)]
. (4.19)
Note that this ratio does not depend on the demand rate D. The optimal policy is obtained by rounding n

to a positive integer in such a way that the cost (4.18) is minimal. Then, the optimal order quantities follow
with (4.17) and (4.9).
Remark 4.1 Without rounding it follows from (4.17) and (4.9) that
ˆ
Q
O
(n

) =
¸
2a
O
wD
h
G
w +h
F
(1 −w)
, Q
R
=
1 −w
n

w
ˆ
Q
O
(n

) =
_
2a
R
D
h
G
+h
F
.
The above approximately optimal values indicate that the optimal order quantity hardly depends on a
R
and
that the optimal repair lot size hardly depends on a
O
and w. In the extreme case that the holding cost for
failed units is negligible, h
F
= 0, the above quantities reduce to
n

=
1 −w
w
_
a
O
a
R
,
ˆ
Q
O
(n

) =
_
2a
O
D
h
G
, Q
R
=
1 −w
n

w
ˆ
Q
O
(n

) =
_
2a
R
D
h
G
,
that is, the quantities are approximately given by EOQ formulas independent of the waste rate w.
Example 4.2 The demand for a spare part of a certain type of copier is D = 5 per day. The fraction of units
that is irreparable is w =
1
5
. The ordering cost is a
O
= $20 and the repair set-up cost is a
R
= $20. The holding
cost per unit per day is h
G
= $0.10 for good units and h
F
= $0.01 for repairable units. From (4.19) it follows
that n

=
4
5
_
0.55
0.028
=
_
88
7
≈ 3.55. From (4.18) it is obtained that
ˆ
C(3) =
_
4 40(0.028 +
1
3
0.352) ≈ 4.822
and
ˆ
C(4) =
_
5 40(0.028 +
1
4
0.352) ≈ 4.817 so that n = 4 is slightly better. The corresponding order
quantity is
ˆ
Q
O
(4) = 41.52, cf. (4.17), and the corresponding repair lot size is Q
R
=
1−w
4w
ˆ
Q
O
(4) = 41.52, cf.
(4.9). The corresponding reorder cycle is T = Q
O
/(wD) = 41.52 days, cf. (4.8), and the corresponding repair
cycle is u = Q
R
/D = 8.30 days, cf. (4.10).
Next, consider the consequences of lead times. A lead time L
O
for external orders has the standard conse-
quence that an order must be placed L
O
units of time before the order must arrive. This reorder instant can
HANS BLANC, TILBURG UNIVERSITY 83
only be translated into a unique reorder point for the inventory position of the total number of good and
failed but repairable units, cf. Figure 4.1. In this sense, the reorder point becomes
s
O
= (1 −w)Q
R
+wDL
O
. (4.20)
The consequence of a lead time for a repair batch is more important, even if we disregard the possibility
that the item must compete with other items for repair capacity. If the repairs are performed close to the
warehouse at a rate of r units per unit of time and repaired units become available during the repair run, the
inventory patterns change to a pattern as for the production lot-size model in Section 1.2.2. This is shown
in Figure 4.2 where the upper graph contains the triangular pattern (apart from the external order) of the
inventory on hand I
G
(t) of good units and the lower graph contains the triangular pattern of the inventory
on hand I
F
(t) of failed but repairable units. However, if the repairs are performed away from the warehouse
and leave the warehouse at the beginning of the lead time L
R
, the whole graph of the inventory on hand
I
F
(t) for failed but repairable parts in Figure 4.1 has to be shifted upwards by (1 −w)DL
R
, the number of
repairable items that return to the warehouse during the lead time L
R
. The average cost per unit of time,
cf. (4.15), (4.16), increases by the amount h
F
(1 −w)DL
R
, assuming that the holding cost during the repair
lead time L
R
remains h
F
. In this case, a repair batch has to start when the inventory position of good units
reaches the reorder point s
R
= DL
R
, except if I
G
(t) + I
F
(t) < Q
R
since then an external order has to be
placed. Moreover, the reorder point in (4.20) also has to be increased by (1 −w)DL
R
to
s
O
= (1 −w)Q
R
+wDL
O
+ (1 −w)DL
R
. (4.21)
This adaptation is illustrated in Figure 4.3, where the dotted lines in the middle graph indicate the batch of
failed items that is in repair.
Example 4.3 If the lead time for orders is L
O
= 5 days and the lead time for repair batches is L
R
= 10
days in the situation of Example 4.2, the reorder point becomes s
O
= 78.22, cf. (4.21), and the starting point
for a repair batch becomes s
R
= DL
R
= 50. The reorder point lies between the minimum and maximum
inventory levels of the total number of usable units on hand ((1−w)Q
R
= 73.22 and (1−w)Q
R
+Q
O
= 114.74,
respectively; see also Figure 4.3). The starting point for a repair batch lies above the maximum level of the
inventory on hand of good units at the end of a repair cycle (Q
R
= 41.52) so that this point has to be applied
to the inventory position of the good units. It means that at some time intervals, two repair batches are
in progress. This also follows from the fact that L
R
= 10 > u = 8.30. The average total cost increases by
h
F
(1 −w)DL
R
= 0.4 to
ˆ
C(4) ≈ 5.217.
Exercise 4.4 Express the maximum inventory level of repairable units, cf. the middle graph of Figure 4.1,
in terms of Q
O
, Q
R
, and the parameters of the model.
Exercise 4.5 The demand for a spare part of a type of laser printer is D = 1 per day. The fraction of units
that is irreparable is w =
1
5
. The ordering cost is a
O
= $20 and the repair set-up cost is a
R
= $45. The
holding cost per unit per day is h
G
= $0.40 for good parts and h
F
= $0.20 for repairable parts. Assume first
that the lead times are negligible. Determine the optimal external order quantity, the optimal repair lot size
and the corresponding reorder cycle. Compute the maximum inventory levels of good parts, of repairable parts
and of all good and repairable parts together. Then, modify the results for a lead time L
O
= 5 days from the
external supplier and a lead time L
R
= 3 days for a repair batch which is transported as a whole to and from
a workshop away from the warehouse.
Exercise 4.6 Adapt the analysis of this section to the situation that the repairs are performed in the ware-
house at a rate of r units per unit of time and repaired units become available during the repair run, cf. the
inventory patterns shown in Figure 4.2.
84 OTHER INVENTORY SYSTEMS WITH INTERACTIONS
¸
`

I
G
(t)
t →
Q
R
Q
O
s
R
0 T 2T −u T −u T −2u 2T −u−L
R
`
`
`
`
`
`
`
` s s s s
c c c c c c c c c
c c c c c c c c c
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
s s s s
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
` s s s s
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
¸
`

I
F
(t)
t →
(1−w)DL
R
+Q
R
(1−w)DL
R
0 T 2T −u T −u T −2u 2T −u−L
R
s s s s s s s s s s s s c c c c c c c c c
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
´
¸
`

I(t)
t →
(1−w)DL
R
+(1−w)Q
R
+Q
O
(1−w)DL
R
+(1−w)Q
R
s
O
(1−w)DL
R
0 T 2T −u T −u T −2u 2T −L
O
s s s s s s s s s s s s
s s s
c c
c c

Figure 4.3: Inventory on hand of good units (top), of repairable units (middle) and of usable units (bottom).
Appendix A
Review of Probability Theory
A.1 Random variables
A random variable is a function defined on a sample space, that is, an assignment of a real number to each
element of the sample space Ω. The (cumulative) distribution function F
Y
(y) of a random variable Y is a
nondecreasing function which tends to 0 as y ↓ −∞ and to 1 as y ↑ ∞. It defines the probability of the event
that Y ≤ y, that is, F
Y
(y) = Pr¦Y ≤ y¦, y ∈ IR. The range of a random variable (a subset of IR) is also
called the support of its distribution function. If a distribution function F
Y
(y) is continuous for all y ∈ IR and
differentiable at almost all y ∈ IR then it is called absolutely continuous, Y is then called a continuous random
variable and f
Y
(y)
.
= F

Y
(y) is called the density of Y . Absolutely continuous distributions are determined
by their density. The integral of the density over the real axis is equal to one. If a distribution function
F
X
(x) is a jump function then the corresponding random variable X is called a discrete random variable.
Let x
1
, x
2
, . . . be the jump points of F
X
(x), that is, the values which X may assume. The function
Pr¦X = x
j
¦
.
= F
X
(x
j
+) −F
X
(x
j
−), j = 1, 2, . . . ,
is called the probability distribution of the discrete random variable X.
A.1.1 Moments and expectation
The expected value of the random variable Y with density f
Y
(y) is defined by
E¦Y ¦ =
_

−∞
y f
Y
(y) dy. (A.1)
The expected value E¦Y ¦ is also called the mean or the average of the random variable Y . If a random
variable Y has range [0, ∞) then its expected value can alternatively be formulated as:
E¦Y ¦ =
_

0
_
y
0
duf
Y
(y) dy =
_

0
_

u
f
Y
(y) dy dt =
_

0
Pr¦Y > u¦ du. (A.2)
Analogously, the mean of a discrete random variable X with range IN can be written as
E¦X¦
.
=

x=0
xPr¦X = x¦ =

x=1
x

n=1
Pr¦X = x¦ =

n=1

x=n
Pr¦X = x¦ =

n=1
Pr¦X ≥ n¦. (A.3)
The kth moment of the distribution of a random variable Y is defined as
E¦Y
k
¦ =
_

−∞
y
k
f
Y
(y) dy, k = 1, 2, . . . . (A.4)
For the special case that Y = X is a random variable with range IN this implies
E¦X
k
¦ =

x=0
x
k
Pr¦X = x¦, k = 1, 2, . . . . (A.5)
85
86 REVIEW OF PROBABILITY THEORY
The variance σ
2
¦Y ¦ of a random variable Y is defined as
σ
2
¦Y ¦
.
= E¦(Y −E¦Y ¦)
2
¦ = E¦Y
2
¦ −E
2
¦Y ¦; (A.6)
here, E
2
¦Y ¦ denotes the square of the mean E¦Y ¦. The square root σ¦Y ¦ of the variance is called the
standard deviation of the random variable Y . The coefficient of variation of the random variable Y is defined
as
C
Y
.
=
σ¦Y ¦
E¦Y ¦
=
1
E¦Y ¦
_
E¦Y
2
¦ −E
2
¦Y ¦. (A.7)
The mean and the variance of a random variable possess the following properties: for any real a and b,
E¦aY +b¦ = a E¦Y ¦ +b, σ
2
¦aY +b¦ = a
2
σ
2
¦Y ¦. (A.8)
A.1.2 Collections of random variables
Let Y
1
and Y
2
be two random variables defined on a common sample space. Their joint distribution function
Pr¦Y
1
≤ y
1
, Y
2
≤ y
2
¦, y
1
, y
2
∈ IR, represents the probability that both Y
1
≤ y
1
and Y
2
≤ y
2
. The (marginal)
distribution function of Y
1
(Y
2
) is obtained by taking the limit y
2
→ ∞ (y
1
→ ∞) of the joint distribution
function for fixed y
1
(y
2
). The density of a pair of continuous random variables is denoted by f
Y
1
,Y
2
(y
1
, y
2
).
The cross moment of two random variables Y
1
and Y
2
is defined as the expectation of the product Y
1
Y
2
:
E¦Y
1
Y
2
¦ =
_

−∞
_

−∞
y
1
y
2
f
Y
1
,Y
2
(y
1
, y
2
) dy
2
dy
1
. (A.9)
The covariance cov¦Y
1
, Y
2
¦ of two random variables Y
1
and Y
2
is defined as
cov¦Y
1
, Y
2
¦
.
= E¦(Y
1
−E¦Y
1
¦)(Y
2
−E¦Y
2
¦)¦ = E¦Y
1
Y
2
¦ −E¦Y
1
¦E¦Y
2
¦. (A.10)
The correlation coefficient ρ¦Y
1
, Y
2
¦ of two random variables Y
1
and Y
2
is defined as
ρ¦Y
1
, Y
2
¦
.
=
cov¦Y
1
, Y
2
¦
σ¦Y
1
¦σ¦Y
2
¦
. (A.11)
For any pair of random variables Y
1
and Y
2
, [ρ¦Y
1
, Y
2
¦[ ≤ 1. Moreover, ρ¦Y
1
, Y
2
¦ = ±1 only if there exist
constants a and b such that Y
2
= aY
1
+b.
The concept of joint distribution function carries over to collections of more than two random variables. Let
Y
i
, i = 1, . . . , n, be a collection of random variables defined on the same sample space. Then it generally
holds that
E
_
n

i=1
Y
i
_
=
n

i=1
E¦Y
i
¦, σ
2
_
n

i=1
Y
i
_
=
n

i=1
σ
2
¦Y
i
¦ + 2
n

i=2
i−1

j=1
cov¦Y
i
, Y
j
¦. (A.12)
Two random variables Y
1
and Y
2
are independent if for all y
1
, y
2
∈ IR,
Pr¦Y
1
≤ y
1
, Y
2
≤ y
2
¦ = Pr¦Y
1
≤ y
1
¦ Pr¦Y
2
≤ y
2
¦. (A.13)
If Y
1
and Y
2
are independent then
E¦Y
1
Y
2
¦ = E¦Y
1
¦ E¦Y
2
¦, cov¦Y
1
, Y
2
¦ = 0. (A.14)
The reverse of this statement is not true, in general. The mean and the variance of two independent random
variables Y
1
and Y
2
possess the following properties: for any real a, b and c,
E¦aY
1
+bY
2
+c¦ = a E¦Y
1
¦ +b E¦Y
2
¦ +c, σ
2
¦aY
1
+bY
2
+c¦ = a
2
σ
2
¦Y
1
¦ +b
2
σ
2
¦Y
2
¦. (A.15)
The concept of independence carries over to finite collections of random variables.
A.2 Probability distributions
This section contains an overview of some useful probability distributions and their properties.
HANS BLANC, TILBURG UNIVERSITY 87
A.2.1 Normal distributions
The normal distributions form a two-parameter family of continuous distributions with range (−∞, ∞). The
density of a normal ^(µ, σ) distribution with mean µ and variance σ
2
is, for σ > 0,
f
µ,σ
(y) =
1
σ


e

1
2
[(y−µ)/σ]
2
, −∞ < y < ∞. (A.16)
The normal distribution with mean 0 and standard deviation 1, ^(0, 1), is called the standard normal
distribution. If the random variable Z has the standard normal distribution and Y the ^(µ, σ) normal
distribution then
Y ∼ µ +σZ, Z ∼ (Y −µ)/σ. (A.17)
An important property of normal distributions is that the sum of independent normally distributed random
variables is normally distributed. For example, if D
1
∼ ^(µ
1
, σ
1
), D
2
∼ ^(µ
2
, σ
2
) and f is a constant, then
D
1
+fD
2
∼ ^(µ
1
+fµ
2
,
_
σ
2
1
+f
2
σ
2
2
).
The normal loss function is defined by
NL(y)
.
= E¦[Z −y]
+
¦ =
_

y
(u −y)f
0,1
(u) du =
1


_

y
(u −y) e

1
2
u
2
du, −∞ < y < ∞. (A.18)
This function has the properties
NL(y) = f
0,1
(y) −
y


_

y
e

1
2
u
2
du, −∞ < y < ∞, (A.19)
and by the symmetry of the density of the standard normal distribution,
NL(y) = NL(−y) −y, −∞ < y < ∞. (A.20)
For a normal ^(µ, σ) distributed random variable Y it holds that
E¦[Y −y]
+
¦ =
_

y
(u −y)f
µ,σ
(u) du = σNL([y −µ]/σ), −∞ < y < ∞. (A.21)
The normal loss function has been tabulated; see e.g. Winston [80, Sect. 17.7]. Alternatively, it is related to
the density and the distribution function (by means of integration by parts): for −∞ < y < ∞,
E¦[Y −y]
+
¦ = σ
2
f
µ,σ
(y) + (µ −y)
_

y
f
µ,σ
(u) du = σf
0,1
([y −µ]/σ) + (µ −y)
_

(y−µ)/σ
f
0,1
(v) dv. (A.22)
A.2.2 Exponential distributions
The distribution function of a nonnegative exponentially distributed random variable Y is by definition
Pr¦Y ≤ y¦ = 1 −e
−λy
, y ≥ 0, (A.23)
for some positive rate (parameter) λ. The density is λe
−λy
. The mean, the standard deviation and the
coefficient of variation of an exponentially distributed random variable Y are:
E¦Y ¦ =
1
λ
, σ¦Y ¦ =
1
λ
, C
Y
=
_
1
λ
. (A.24)
An important property of the exponential distribution is its so called lack of memory. This property is
expressed by the following formula:
Pr¦Y > u +y
¸
¸
Y > u¦ =
e
−λ(u+y)
e
−λu
= e
−λy
= Pr¦Y > y¦, y, u ≥ 0. (A.25)
It states that the conditional probability that Y exceeds the value u + y given that Y exceeds the value u
is equal to the unconditional probability that Y exceeds the value y, for every positive u and y. In other
words, if the random variable Y represents the duration of some process, and if this process is observed to
88 REVIEW OF PROBABILITY THEORY
be still in progress after some time u then the time until completion has the same distribution as it had at
the beginning of the process. As a consequence, the process can be considered as if it started anew at the
time of observation u. The future of the process after time u does not depend on its past until time u given
the observation that it is still not completed at time u.
Another important property of the exponential distribution is the fact that the minimum of a number of
independent, exponentially distributed random variables Y
j
with parameters λ
j
, j = 1, . . . , n, is exponentially
distributed with parameter the sum of the parameters of the variables Y
j
, that is, for any number n of
variables,
Pr¦min¦Y
1
, . . . , Y
n
¦ ≤ y¦ = 1 −e
−(λ
1
+···+λ
n
)y
, y ≥ 0. (A.26)
This property is readily verified by noting that, for y ≥ 0,
Pr¦min¦Y
1
, . . . , Y
n
¦ > y¦ = Pr¦Y
1
> y, . . . , Y
n
> y¦ =
n

j=1
Pr¦Y
j
> y¦ =
n

j=1
e
−λ
j
y
. (A.27)
A.2.3 Poisson distributions and Poisson processes
A random variable with a Poisson distribution has the set IN of nonnegative integers as its range. Such a
random variable is defined by
Pr¦X
a
= n¦ =
a
n
n!
e
−a
, n = 0, 1, 2, . . . , (A.28)
with a a positive parameter. Its first two moments, variance and coefficient of variation are
E¦X
a
¦ = a, E¦X
2
a
¦ = a
2
+a, σ
2
¦X
a
¦ = a, C
X
a
=
_
1/a. (A.29)
A Poisson process N(t), t ≥ 0, with rate λ, is a counting process such that the number of events in a time
interval of length t has the Poisson distribution with parameter a = λt. The distribution of the time between
successive events is exponential with mean 1/λ.
A.2.4 Gamma and Erlang distributions
The gamma distributions form a two-parameter family of continuous distributions with range (0, ∞). The
density of a gamma ((ψ, λ) distribution with shape parameter ψ and scale parameter λ is, for ψ > 0, λ > 0,
f
Y
ψ,λ
(y) =
λ
ψ
Γ(ψ)
y
ψ−1
e
−λy
, y > 0. (A.30)
The gamma function Γ(ψ) is defined by
Γ(x)
.
=
_

0
e
−u
u
x−1
du, x > 0. (A.31)
For integer-valued n it holds that
Γ(n + 1) = n!, n ∈ IN. (A.32)
The above density has a maximum at y = (ψ −1)/λ if ψ > 1 and at 0 otherwise. The moments of a gamma
distribution are
E¦Y
k
ψ,λ
¦ =
1
λ
k
k−1

j=0
(ψ +j) =
_
ψ +k −1
k
_
k!
λ
k
, k = 1, 2, . . . . (A.33)
In particular, the first two moments, variance and coefficient of variation are
E¦Y
ψ,λ
¦ =
ψ
λ
, E¦Y
2
ψ,λ
¦ =
ψ(1 +ψ)
λ
2
=
1 +ψ
ψ
E
2
¦Y
ψ,λ
¦, σ
2
¦Y
ψ,λ
¦ =
ψ
λ
2
, C
Y
ψ,λ
=
_
1
ψ
. (A.34)
The densities of gamma distributions are displayed in Figure A.1 for several values of the shape parameter
ψ. The means of all these distributions are equal to 1, that is, λ = ψ in all cases.
HANS BLANC, TILBURG UNIVERSITY 89

f
Y
ψ,ψ
(y)
0 0.5 1 1.5 2 2.5 3
0
0.5
1
1.5
1/4
12
8
4
2
3/2
1/2
1
y →
Figure A.1: Densities of gamma distributions with varying shape parameter ψ and fixed mean of 1.
Special cases of this class of distributions are the Erlang c(K, λ) distributions for which K is an integer.
With the aid of repeated integration by parts, the cumulative distribution function of the Erlang distribution
with shape parameter K can be explicitly written as
Pr¦Y
K,λ
≤ y¦ = 1 −
K−1

j=0
(λy)
j
j!
e
−λy
, y ≥ 0. (A.35)
An important property of the Erlang c(K, λ) distribution is that it emerges as the K-fold convolution of the
exponential distribution with parameter λ. This means that a random variable with an c(K, λ) distribution
can be considered as the sum of K independent, identically, exponentially distributed phases.
The gamma loss function is defined by
GL
ψ,λ
(y)
.
=
_

y
(u −y) f
Y
ψ,λ
(u) du =
λ
ψ
Γ(ψ)
_

y
(u −y) u
ψ−1
e
−λu
du, y ≥ 0. (A.36)
For Erlang distributions (with an integer shape parameter Ψ = K) this function can be simplified with the
aid of repeated integration by parts to
GL
K,λ
(y) =
λ
K
(K −1)!
_

y
(u −y) u
K−1
e
−λu
du = e
−λy
K−1

j=0
(K −j)
λ
j−1
y
j
j!
, y ≥ 0. (A.37)
For general ψ, the gamma loss function can be expressed after a single integration by parts and with the
property of gamma functions that Γ(ψ + 1) = ψΓ(ψ), as, for y ≥ 0,
GL
ψ,λ
(y) =
ψ
λ
Pr¦Y
ψ+1,λ
> y¦ −y Pr¦Y
ψ,λ
> y¦ =
ψ
λ
_

y
f
Y
ψ+1,λ
(u) du −y
_

y
f
Y
ψ,λ
(u) du. (A.38)
The latter two integrals are known as incomplete gamma functions.
A.2.5 Mixtures of Erlang distributions
A mixture of two Erlang distributions, to be denoted by /c(K
1
, λ
1
, K
2
, λ
2
, p), has five parameters and range
(0, ∞). The density of a mixture of two Erlang distributions is
f(y) = p
λ
K
1
1
y
K
1
−1
(K
1
−1)!
e
−λ
1
y
+ (1 −p)
λ
K
2
2
y
K
2
−1
(K
2
−1)!
e
−λ
2
y
, y > 0; (A.39)
90 REVIEW OF PROBABILITY THEORY
here, K
1
and K
2
are positive integers, λ
1
and λ
2
are positive scale parameters, and p, 0 < p < 1, is a weight.
The mean of a random variable Y with a mixture of two Erlang distributions is
E¦Y ¦ = p
K
1
λ
1
+ (1 −p)
K
2
λ
2
, (A.40)
and its variance is
σ
2
¦Y ¦ = p
K
1
(K
1
+ 1)
λ
2
1
+ (1 −p)
K
2
(K
2
+ 1)
λ
2
2
−E
2
¦Y ¦. (A.41)
Mixtures of two Erlang distributions are often used to approximate the distributions of random variables
with range (0, ∞) of which only the mean E¦Y ¦ and the variance σ
2
¦Y ¦ are known. For the choice of the
parameters, a distinction is made whether the coefficient of variation C = σ¦Y ¦/E¦Y ¦ is smaller than 1 or
not. If the coefficient of variation C < 1, then the parameters are chosen as follows: K
1
is the largest integer
smaller than 1/C
2
, K
2
= K
1
+ 1, and λ
1
= λ
2
and p are determined such that the mean and the variance
agree:
p =
1
1 +C
2
_
K
2
C
2

_
K
2
(1 +C
2
) −K
2
2
C
2
_
, λ
1
= λ
2
= (K
2
−p)/E¦Y ¦. (A.42)
If the coefficient of variation C > 1, then the parameters are chosen as follows: K
1
= K
2
= 1, λ
1
,= λ
2
and p
are determined in such a way that the mean and the variance agree and the third moment is equal to that
of a gamma distribution:
λ
1
=
2
E¦Y ¦
_
_
1 +
¸
C
2

1
2
1 +C
2
_
_
, λ
2
=
4
E¦Y ¦
−λ
1
, p =
λ
1
(1 −λ
2
E¦Y ¦)
λ
1
−λ
2
. (A.43)
Mixtures with K
1
= K
2
= 1 are mixtures of two exponential distributions. Such mixtures are called
hyperexponential distributions and are indicated by the symbol H
2
.
The loss function corresponding to a mixture of two Erlang distributions simply follows from (A.37) as
E¦[Y −y]
+
¦ = p e
−λ
1
y
K
1
−1

j=0
(K
1
−j)
λ
j−1
1
y
j
j!
+ (1 −p) e
−λ
2
y
K
2
−1

j=0
(K
2
−j)
λ
j−1
2
y
j
j!
, y ≥ 0. (A.44)
For the special case K
1
= K, K
2
= K + 1 and λ
1
= λ
2
= λ this loss function becomes
E¦[Y −y]
+
¦ = e
−λy
_
_
K−1

j=0
(K −j)
λ
j−1
y
j
j!
+ (1 −p)
K

j=0
λ
j−1
y
j
j!
_
_
, y ≥ 0. (A.45)
The loss function corresponding to a hyperexponential distribution (K
1
= K
2
= 1) is still simpler:
E¦[Y −y]
+
¦ =
p
λ
1
e
−λ
1
y
+
1 −p
λ
2
e
−λ
2
y
, y ≥ 0. (A.46)
Appendix B
Optimization methods
B.1 Stationary points
In inventory theory objective functions of the following form are often encountered:
g(x) =
a
x
+bx, x > 0; (B.1)
here, a and b are positive constants. The derivative of this function is
g

(x) = −
a
x
2
+b, x > 0. (B.2)
This derivative has a unique positive zero at
x = x

.
=
_
a/b. (B.3)
The second derivative of g(x) at x

is positive:
g

(x) =
2a
x
3
⇒ g

(x

) =
2b

b

a
. (B.4)
Hence, the function g(x) has a minimum at x

with the value
g(x

) = a
_
b/a +b
_
a/b = 2

ab. (B.5)
B.2 Golden section
The “Golden-Section” method is an interval elimination method for finding the minimum x

of a function
g(x) with no local minima beside the global minimum. This method assumes that g(x) is strictly decreasing
for x < x

and strictly increasing for x > x

. An initial interval (a, b) is divided into three subintervals (a, c),
(c, d) and (d, b) with relative lengths z
2
, 1 −2z
2
, z
2
, respectively, with z =
1
2
(

5 −1) ≈ 0.618. This implies
that z
2
=
1
2
(3 −

5) = 1 −z ≈ 0.382, c = a +z
2
(b −a) and d = b −z
2
(b −a) = a +z(b −a). If g(c) < g(d),
the interval (d, b) is discarded, and the procedure is repeated for the interval (a, d); otherwise, the interval
(a, c) is discarded, and the procedure is repeated for the interval (c, b). The lengths of the new intervals are
d −a = b −c = z(b −a). Hence, the procedure continues with intervals which are a factor z ≈ 0.618 smaller
with each step, until the minimum has been found with sufficient accuracy.
Example B.1 Consider an objective function of the form (B.1):
g(x) =
500
x
+
x
20
, x > 0.
As initial points we take a = 10 and b = 1000 where one of the terms of g(x) dominates the other: g(10) =
g(1000) = 50 + 0.5 = 50.5. Table B.1 shows the results of the golden section procedure with and without
rounding to integer order quantities. The procedure can be continued to end at the optimum x = 100. Note
the flatness of the objective function near its minimum.
91
92 OPTIMIZATION METHODS
Table B.1: Results of golden section procedure for the function of Example B.1.
Point c g(c) Point d g(d) Rem. interval Point c g(c) Point d g(d) Remaining interval
388 20.69 622 31.90 (10, 622) 388.15 20.70 621.85 31.90 (10.00, 621.85)
244 14.25 388 20.69 (10, 388) 243.71 14.24 388.15 20.70 (10.00, 388.15)
154 10.95 244 14.25 (10, 244) 154.44 10.96 243.71 14.24 (10.00, 243.71)
99 10.00 154 10.95 (10, 154) 99.27 10.00 154.44 10.96 (10.00, 154.44)
65 10.94 99 10.00 (65, 154) 65.17 10.93 99.27 10.00 (65.17, 154.44)
99 10.00 120 10.17 (65, 120) 99.27 10.00 120.34 10.17 (65.17, 120.34)
86 10.11 99 10.00 (86, 120) 86.24 10.11 99.27 10.00 (86.24, 120.34)
99 10.00 107 10.02 (86, 107) 99.27 10.00 107.32 10.02 (86.24, 107.32)
94 10.02 99 10.00 (94, 107) 94.29 10.02 99.27 10.00 (94.29, 107.32)
99 10.00 102 10.00 (94, 102) 99.27 10.00 102.34 10.00 (94.29, 102.34)
97 10.00 99 10.00 (97, 102) 97.37 10.00 99.27 10.00 (97.37, 102.34)
B.3 Bisection
The method of interval bisection is an alternative adaptive search method for finding the minimum x

of a
function g(x) with no local minima beside the global minimum. This method employs information about
the value of the derivative of the function g(x). It assumes that g

(x) is negative for x < x

and positive for
x > x

. The method starts with an initial interval (a, b) which should contain x

. The derivative g

(x) is
evaluated at the middle point c =
1
2
(a + b). If g

(c) > 0, the interval (c, b) is discarded, and the procedure
continues with the interval (a, c). If g

(c) < 0, the interval (a, c) is discarded, and the procedure continues
with the interval (c, b). Each application of the method halves the length of the interval in which the minimum
x

must lie. With the same initial interval, the method of interval bisection is faster than the Golden-Section
method in terms of the number of iterations. However, the computation of the derivative g

(x) may be more
involved than that of g(x).
Table B.2: Results of bisection procedure for the function of Example B.2.
Middle point c g

(c) Remaining interval Middle point c g

(c) Remaining interval
505 0.0480 (10, 505) 505.00 0.0480 (10.00, 505.00)
258 0.0425 (10, 258) 257.50 0.0425 (10.00, 257.50)
134 0.0222 (10, 134) 133.75 0.0220 (10.00, 133.75)
72 −0.0465 (72, 134) 71.88 −0.0468 (71.88, 133.75)
103 0.0029 (72, 103) 102.81 0.0027 (71.88, 102.81)
88 −0.0146 (88, 103) 87.34 −0.0155 (87.34, 102.81)
96 −0.0043 (96, 103) 95.08 −0.0053 (95.08, 102.81)
100 0.0000 98.95 −0.0011 (98.95, 102.81)
100.88 0.0009 (98.95, 100.88)
99.91 −0.0001 (99.91, 100.88)
Example B.2 Consider again the objective function of Example B.1. This function has derivative:
g

(x) = −
500
x
2
+
1
20
, x > 0.
As initial points we again take a = 10 and b = 1000 where the derivative has opposite sign: g

(10) = −4.9500,
g

(1000) = 0.0495. Table B.2 shows the results of the bisection procedure with and without rounding. Clearly,
this procedure converges faster than the golden section procedure in Table B.1.
The method of interval bisection can also be used to search a root of an equation or the zero of a function. In
this case, it is assumed that the function is continuous and that it possesses a single zero on the initial search
interval so that the sign of the function at the two end points is different. To find a root of a function, the
sign of the function at the middle point is considered. If this sign is the same as that at the left end point,
then the left half of the interval is discarded; otherwise, the right half of the interval is discarded.
Bibliography
[1] Aksoy, Y., S. Erenguc. Multi-item inventory models with coordinated replenishments: a survey, Int. J.
Production Management 8 (1988), 63–73.
[2] Andres, F., H. Emmons. A multi-product inventory system with interactive set-up costs, Management
Science 21 (1975), 1055–1063.
[3] Atkins, D.R., P.O. Iyogun. A lower bound for a class of inventory/production problems, Operations
Research Letters 6 (1987), 63–67.
[4] Atkins, D.R., P.O. Iyogun. Periodic versus ‘can-order’ policies for coordinated multi-item inventory
systems, Management Science 34 (1988), 791–796.
[5] Atkins, D.R., P.O. Iyogun. A heuristic with lower bound performance guarantee for the multi-product
dynamic lot-size problem, IIE Transactions 20 (1988), 369–373.
[6] Axs¨ater, S. Exact analysis of continuous review (R,Q) policies in two-echelon inventory systems with
compound Poisson demand, Operations Research 48 (2000), 686–696.
[7] Axs¨ater, S. Inventory Control, Springer, New York, 2006 (2nd ed.).
[8] Axs¨ater, S., W.F. Zhang. A joint replenishment policy for multi-echelon inventory control, Int. J.
Production Economics 59 (1999), 243–250.
[9] Baker, K.R. Lot-sizing procedures and a standard date set: a reconciliation of the literature, J. Manu-
facturing and Operations Management 2 (1989), 199–221.
[10] Ballintfy, J.L. On a basic class of inventory systems, Management Science 10 (1964), 287–297.
[11] Bastian, M. Joint replenishment in multi-item inventory systems, J. Oper. Res. Soc. 37 (1986), 1113–
1120.
[12] Blackburn, J., R. Millen. Heuristics lot-sizing performance in a rolling-schedule environment, Decision
Sciences 11 (1980), 691–701.
[13] Blackburn, J., R. Millen. Improved heuristics for multi-stage requirements planning systems, Manage-
ment Science 28 (1982), 44–56.
[14] Chakravarty, A.K. Multi-item inventory aggregation into groups, J. Oper. Res. Soc. 32 (1981), 19–26.
[15] Chakravarty, A.K. Joint inventory replenishments with group discounts based on invoice value, Man-
agement Science 30 (1984), 1105–1112.
[16] Chakravarty, A.K. An optimal heuristic for coordinated multi-item inventory replenishments, J. Oper.
Res. Soc. 36 (1985), 1027–1039.
[17] Chakravarty, A.K., S.K. Goyal. Multi-item inventory grouping with dependent set-up cost and group
overhead cost, Engineering Costs and Production Economics 10 (1986), 13–23.
[18] Clark, A.J., H. Scarf. Optimal policies for a multi-echelon inventory problem, Management Science 6
(1960), 475–490.
[19] Dagpunar, J.S. On the relationship between constrained and unconstrained multi-item replenishment
problems, Int. J. Oper. Prod. Management 9 (1989), 77–81.
[20] Das, C., S.K. Goyal. Economic ordering policy for deterministic two-echelon distribution systems,
Engineering Costs and Production Economics 21 (1991), 227–231.
[21] De Kok, A.G. Hierarchical production planning for consumer goods, Eur. J. Oper. Res. 45 (1990),
55–69.
93
94 BIBLIOGRAPHY
[22] DeMatteis, J. An economic lot-sizing technique I: the part-period algorithm, IBM Syst. J. 7 (1968),
30–38.
[23] Eastman, R.M. Materials Handling, Marcel Dekker, New York, 1987.
[24] Eppen, G., L. Schrage. Centralized ordering policies in a multi-warehouse system with lead times and
random demand, in: L.B. Schwarz (ed.) Multi-level Production-Inventory Control Systems: Theory and
Practice, North-Holland, Amsterdam, 1981.
[25] Ettl, M., G.E. Feigin, G.Y. Lin, D.D. Yao. A supply network model with base-stock control and service
requirements, Operations Research 48 (2000), 216–232.
[26] Eynan, A., D.H. Kropp. Periodic review and joint replenishment in stochastic demand environments,
IIE Transactions 30 (1998), 1025–1033.
[27] Federgruen, A. Centralized planning models for multi-echelon inventory systems under uncertainty, in:
S.C. Graves, A.H.G. Rinnooy Kan, P.H. Zipkin (eds.) Logistics of Production and Inventory, North-
Holland, Amsterdam, 1993.
[28] Federgruen, A., H. Groenevelt, H.C. Tijms. Coordinated replenishment in a multi-item inventory system
with compound Poisson demands, Management Science 30 (1984), 344–357.
[29] Federgruen, A., P. Zipkin. An efficient algorithm for computing optimal (s,S) policies, Operations
Research 32 (1984), 1268–1285.
[30] Fogarty, D.W., J.H. Blackstone, T.R. Hoffmann. Production & Inventory Management, South-Western
Publ. Co., Cincinnati, Ohio, 1991, 2nd ed..
[31] Goetschalckx, M., H.D. Ratliff. Order picking in an aisle, IIE Transactions 20 (1988), 53–62.
[32] Goetschalckx, M., H.D. Ratliff. Sequencing picking operations in a man-aboard order picking system,
Material Flow 4 (1988), 255–263.
[33] Goyal, S.K. Determination of optimum packaging frequency of items jointly replenished, Management
Science 21 (1974), 436–443.
[34] Goyal, S.K. Optimum ordering policy for a multi-item single supplier system, Operational Research
Quarterly 25 (1974), 293–298.
[35] Goyal, S.K., A.S. Belton. On a simple method of determining order quantities in joint replenishment
under deterministic demand, Management Science 25 (1979), 604.
[36] Goyal, S.K., A.T. Satir. Joint replenishment inventory control: deterministic and stochastic models,
Eur. J. Oper. Res. 38 (1989), 2–13.
[37] Graves, S.C. On the deterministic demand multi-product single-machine lot scheduling problem, Man-
agement Science 25 (1979), 276–280.
[38] Hausman, W.H., L.B. Schwarz, S.C. Graves. Optimal storage assignment in automatic warehousing
systems, Management Science 22 (1976), 629–638.
[39] Ignall, E. Optimal continuous review policies for two product inventory systems with joint set-up cost,
Management Science 15 (1967), 277–279.
[40] Inderfurth, K. Safety stocks in multistage inventory systems: a survey, Int. J. Production Economics
35 (1994), 321–329.
[41] Iyogun, P.O. Heuristic methods for the multi-product dynamic lot size problem, J. Oper. Res. Soc. 42
(1991), 889–894.
[42] Jackson, P., W. Maxwell, J. Muckstadt. The joint replenishment problem with a powers-of-two restric-
tion, IIE Transactions 17 (1985), 25–32.
[43] Joneja, D. The joint replenishment problem: new heuristic and worst case performance bounds, Oper-
ations Research 38 (1990), 711–723.
[44] Kao, P.C. A multi-product dynamic lot-size model with individual and joint set-up costs, Operations
Research 27 (1979), 279–289.
[45] Kaspi, M., M.J. Rosenblatt. An improvement of Silver’s algorithm for the joint replenishment problem,
IIE Transactions 15 (1983), 264–267.
HANS BLANC, TILBURG UNIVERSITY 95
[46] Korugan, A., S.M. Gupta. A multi-echelon inventory system with returns, Computers & Industrial
Engineering 35 (1998), 145–148.
[47] Lambrecht, M.R., H. Vanderveken, J. vander Eecken. On the joint replenishment problem under deter-
ministic time varying demand patterns, Onderzoeksrapport nr. 7907, K.U. Leuven, 1979.
[48] Lee, F.-C., M.-J. Yao. A global optimum search algorithm for the joint replenishment problem under
power-of-two policy, Computers & Operations research 30 (2003), 1319–1333.
[49] Love, S. Inventory Control, Mc Graw Hill, New York, 1979.
[50] Liu, L., X.M. Yuan. Coordinated replenishment in inventory systems with correlated demands, Eur. J.
Oper. Res. 123 (2000), 490–503.
[51] McGee, V.E., D.F. Pyke. Periodic production scheduling at a fastener manufacturer, Int. J. Production
Economics 46–47 (1996), 65–87.
[52] Mendoza, A. An economic lot-sizing technique II: mathematical analysis of the part-period algorithm,
IBM Syst. J. 7 (1968), 39–46.
[53] Odanaka, T., K. Yamaguchi, T. Masui. Multi-echelon inventory production system solution, Computers
& Industrial Engineering 27 (1994), 201–204.
[54] Pantumsinchai, P. A comparison of three joint ordering inventory policies, Decision Science 23 (1992),
111–127.
[55] Ratliff, H.D., A.S. Rosenthal. Order picking in a rectangular warehouse: a solvable case of the traveling
salesman problem, Operations Research 31 (1983), 507–516.
[56] Robeson, J.F., R.G. House. The Distribution Handbook, The Free Press, New York, 1985.
[57] Schrady, D.A. A deterministic inventory model for repairable items, Naval Research Logistics Quarterly
14 (1967), 391–398.
[58] Silver, E.A. A control system for coordinated inventory replenishment, Int. J. Production Res. 12
(1974), 647–671.
[59] Silver, E.A. A simple method of determining order quantities in joint replenishment under deterministic
demand, Management Science 22 (1976), 1351–1361.
[60] Silver, E.A. Coordinated replenishment of items under time-varying demand: dynamic programming
formulation, Naval Research Logistics 26 (1979), 141–151.
[61] Silver, E.A. Establishing reorder points in the (S,c,s) coordinated control system under compound
Poisson demand, Int. J. Production Res. 19 (1981), 743–750.
[62] Silver, E.A., H.C. Meal. A heuristic for selecting lot size quantities for the case of a deterministic time-
varying demand rate and discrete opportunities for replenishment, Production and Inventory Manage-
ment, 14 (1973), 64–74.
[63] Silver, E., J. Miltenburg. Two modifications of the Silver-Meal lot sizing heuristic, INFOR 22 (1984),
56–68.
[64] Silver, E.A., D.F. Pyke, R. Peterson. Inventory Management and Production Planning and Scheduling,
Wiley, New York, 1998 (3rd ed.).
[65] Tijms, H.C. Stochastic Models: an algorithmic approach. Wiley, Chichester, 1994.
[66] Tomkins, J., J. White. Facilities Planning, Wiley, New York, 1984.
[67] Van der Duyn Schouten, F.A., M.J.G. Van Eijs, R.M.J. Heuts. Coordinated replenishment systems
with discount opportunities, Int. J. Production Research 32 (1994), 2879–2896.
[68] Van der Heijden, M.C. Supply rationing in multi-echelon divergent systems, Eur. J. Oper. Res. 101
(1997), 532–549.
[69] Van der Heijden, M. Near cost-optimal inventory control policies for divergent networks under fill rate
constraints, Int. J. Production Economics 63 (2000), 161–179.
[70] Van Eijs, M.J.G. A note on the joint replenishment problem under constant demand, J. Oper. Res.
Soc. 44 (1993), 185–192.
[71] Van Eijs, M.J.G. On the determination of the control parameters of the optimal can-order policy,
Zeitschrift f¨ ur Operations Research 39 (1994), 289.
96 BIBLIOGRAPHY
[72] Van Eijs, M.J.G. Multi-item inventory systems with joint ordering and transportation decisions, Int.
J. Production Economics 35 (1994), 285–292.
[73] Van Eijs, M.J.G., R.M.J. Heuts, J.P.C. Kleijnen. Analysis and comparison of two strategies for multi-
item inventory systems with joint replenishment costs, Eur. J. Oper. Res. 59 (1992), 405–412.
[74] Verrijdt, J.H.C.M., A.G. De Kok Distribution planning for a divergent N-echelon network without
intermediate stocks under service restrictions, Int. J. Production Economics 38 (1995), 225–243.
[75] Verrijdt, J.H.C.M., A.G. De Kok Distribution planning for a divergent depotless two-echelon network
under service restrictions, Eur. J. Oper. Res. 89 (1996), 341–354.
[76] Viswanathan, S. Periodic review (s,S) policies for joint replenishment inventory systems, Management
Science 43 (1997), 1447–1454.
[77] Wagner, H.M., T.M. Within. Dynamic version of the economic lot size model, Management Science 5
(1958), 89–96.
[78] Wildeman, R.E., J.B.G. Frenk, R. Dekker. An efficient optimal solution method for the joint replen-
ishment problem, Eur. J. Oper. Res. 99 (1997), 433–444.
[79] Wilson, H.G. Order quantity, product popularity and the location of stock in a warehouse, IIE Trans-
actions 9 (1977), 230–237.
[80] Winston, W.L. Operations Research: Applications and Algorithms. PWS-Kent Publ. Co., Boston, 3rd
ed., 1994.
[81] Yoo, Y., W. Kim, J. Rhee. Efficient inventory management in multi-echelon distribution systems,
Computers & Industrial Engineering 33 (1997), 729–732.
[82] Zheng, Y.S., A. Federgruen. Finding optimal (s,S) policies is as simple as evaluating a single policy,
Operations Research 39 (1991), 654–665 (corrections: Operations Research 40 (1992), 192).
[83] Zipkin, M.C. On the imbalance of inventories in multi-echelon systems, Math. Oper. Res. 9 (1984),
402–423.
Index
(R, S) policy, 15, 52, 69
(R, k, S) policy, 44, 51, 54
(R, s, S) policy, 44, 54
(R, s, S) policy, 15
(s, c, S) policy, 44, 54
(s, Q) policy, 12
(s, S) policy, 12
ABC classification, 2, 58
Action space, 37
Advertisement, 17, 59
Assembly, 17, 56, 57, 67
Atkins & Iyogun heuristic, 52
Backlog, 1, 3, 9, 12, 15, 43, 48, 59, 69
Balance equation, 12, 15, 45
Balanced Stock rationing, 72, 73
Base-stock policy, 12, 69
Bastian heuristic, 33
Batching, 58
Bisection, 72, 78, 92
Blackburn & Millen heuristic, 67
Budget constraint, 79
Can-order level, 44
Can-order policy, 44–51, 54
Carrying charge, 2, 7, 20, 31, 53, 60
Central depot, see distribution center
Coefficient method (heuristic), 40
Coefficient of variation, 13, 16, 75, 86, 90
Complementarity effect, 17
Consecutiveness property, 32
Consistent Appropriate Share rationing, 74
Constant demand, see demand, constant
Continuous review, 2, 12, 20, 44
Converging chain, 56, 67
Conveyor, 57
Coordinated replenishment, 20–54
Correlated demand, see demand, correlated
Correlation coefficient, 86
Cost function, see objective function
Cost-covering heuristic, 42
Covariance, 86
Crane, 57
Cross-docking, 56, 62, 66, 69
Cycle service constraint, 3, 14, 47, 50
Decision variable, 9, 12, 21, 44, 60, 67, 70, 73, 77, 79
Delivery instant, 3, 4, 6, 9, 12, 63, 69
Demand, 1, 67
acyclic pattern, 66
constant, 4–9, 20–36, 58–66, 77–83
correlated, 17, 44
during lead time, 6, 12, 47, 70
during lead time plus review period, 15, 52, 71
forecasting, 2, 59
rate, 4, 6, 13, 20, 60, 64, 77, 79
seasonal pattern, 1, 9
stochastic, 1–3, 12–17, 43–54, 69–76
time varying, 9–12, 36–43, 66–68
trend, 1, 9
Density, 72, 85, 85, 87
Direct grouping, 20, 31–36, 44
Discount, 2, 3, 7, 17, 34
all-units, 30
factor, 7, 30, 34
Distribution center, 56, 64, 69
Distribution function, 15, 85, 87
absolutely continuous, 85
Distribution network, 56
Diverging chain, 56, 64, 69
Dynamic programming, 10, 32, 34, 35, 37
Echelon inventory, 59, 63, 69
net, 59
on hand, 59, 61, 70
position, 59, 63, 66, 70
Economic order quantity, 4, 7, 9, 12, 21, 45, 62, 66,
77, 79, 82
EDI, see electronic data interchange
Electronic data interchange, 19
EOQ, see economic order quantity
Erlang distribution, 16, 73, 89
Exponential distribution, 87, 88, 89
Exponential smoothing, 59
Family of items, 20, 35, 43
Family ordering cost, see ordering cost, joint
Family reorder cycle, 21–26, 30, 51, 65
Fast mover, 6
Feasible, 28
Fill rate
actual, 9, 12, 48, 49, 73
constraint, 3, 12, 15, 48, 69
target, 3, 9, 12, 15, 49, 52, 54, 70, 72, 73
Flexibility, 18
Fluid approximation, 4, 20
97
98 INDEX
Forecasting, see demand, forecasting
Gamma distribution, 16, 52, 88
Gamma function, 88, 89
Gamma loss function, 16, 89
Golden section, 47, 91
Goyal & Belton heuristic, 24, 30, 39, 53
Goyal algorithm, 22, 23, 24, 29, 35, 54, 65
Goyal iterative heuristic, 25, 36
Group of items, 20, 44
Group reorder cycle, 31, 34
Handling cost, 2, 18
joint, 18
Handling time, 1
Holding cost, 2, 4, 6, 7, 12, 20, 38, 43, 60, 61, 64, 67,
70, 73, 77, 79, 82
cumulative, 11, 40
Holidays, 59
Horizon, 10, 36, 67
rolling, 12, 39
Hyperexponential distribution, 16, 90
Imbalance, 69, 71, 73
Independence of random variables, 48, 69, 70, 73, 86
Indirect grouping, 20, 21–31, 35, 44, 65
Information
exchange between echelons, 59
Interaction in inventory management, 17, 77
Inventory, 1
capacity, 17, 58, 77
cost, 4, 21, 45, 47, 48
in transit, 6, 59, 63, 74
management, 1, 79
net, 3, 9, 14
on hand, 3, 6, 7, 12, 47, 49, 73, 81
policy, 2, 3, 20, 43, 58, 69
position, 3, 6, 7, 12, 15, 44, 47, 63, 70, 83
theory, 1, 91
Inventory level, 36
average, 4, 6, 46, 53, 81
maximum, 6, 77, 83
saw-tooth pattern, 4, 6, 8, 61, 63, 81
step pattern, 5, 61, 63
triangular pattern, 7, 83
Joint distribution, 86
Joint ordering cost, see ordering cost, joint
Joint replenishment, see coordinated replenishment
Kaspi & Rosenblatt heuristic, 25, 26, 30, 36, 65
Labor cost, 28
Lack of memory, 48, 87
Lagrange multiplier, 72, 73, 77
Lead time, 1, 6, 12, 15, 43, 47, 49, 52, 53, 57, 60, 63,
66, 69, 79, 82
Linear chain, 56, 58–64, 67–68
Lost sales, 1–3
Lot size, see production, lot size
Management cost, 3, 18, 21
Marginal distribution, 86
Markov process, 45
Material cost, 28
Material handling, 57
Material requirement planning, 9
Mean of a random variable, 12, 15, 43, 48, 49, 52, 53,
70, 71, 73, 85, 87, 90
Mixture of Erlang distributions, 16, 73, 89
Moment of a distribution, 85, 88
Multi-echelon system, 55–76
oscillation effect, 58
Net stock, see inventory, net
Normal distribution, 16, 71, 87
Normal loss function, 16, 72, 87
Number of orders per unit of time, 5, 46
Objective function, 21, 26, 30, 36, 44, 65, 77, 81, 91
Order picking, 1, 57
Order quantity, 2, 4, 9, 12, 22, 44, 58, 60, 64, 67, 70,
77, 79
mean, 15, 48, 53, 72
Order-up-to level, 2, 12, 15, 44, 51, 52, 54, 58, 70, 73
Ordering cost, 2, 4, 11, 12, 18, 40, 43, 60, 64, 67, 77,
79
joint, 17, 20, 24, 35, 36, 52–54, 65, 79
supplementary, 20, 36, 50, 53, 65
Ordering rule, see inventory, policy
Overtime work, 3
P
1
-criterion, see cycle service constraint
P
2
-criterion, see fill rate constraint
Part-period-balancing heuristic, 11, 40
Penalty cost, 3
Periodic review, 2, 15, 44, 51, 69
Physical stock, see inventory, on hand
Poisson distribution, 13, 48, 88
Poisson process, 13, 43, 45, 88
Powers-of-two policy, 26, 29
Price
break point, 7, 30, 34
change, 59
purchasing, 4, 7, 30, 34, 53
Probability distribution, 15, 73, 85, 86
Production
capacity, 28
cycle, 7, 28
lot size, 7, 28
rate, 6, 28
schedule, 28
Public warehousing, 57
Pull policy, 19, 59, 62, 66
Purchasing cost, 2, 4, 7, 18, 20, 30, 34, 60
Push policy, 19, 59, 61, 64, 67, 69
HANS BLANC, TILBURG UNIVERSITY 99
Quality control, 57, 59
Quantity discount, see discount
Random variable, 12, 15, 70, 73, 85
continuous, 85
discrete, 85
Range of a random variable, 71, 85, 85, 86
Rationing fraction, 70, 72, 73, 75
Reference item, 24, 65
Reorder cycle, 4, 12, 15, 47, 64, 79, 82
Reorder frequency, 21, 26, 30, 51, 65
Reorder instant, 1, 3, 6, 12, 30, 40, 63, 69, 82
synchronization, 22, 66
Reorder point, 2, 6, 9, 12, 44, 48, 60, 80, 83
echelon, 63, 66
Repair
batch, 79
lot size, 80, 82
rate, 83
Repairs, 18, 57, 79
Returns, 18, 57
Review instant, 2, 15, 44
Review period, 2, 15, 26, 44, 51, 52, 66, 70
Rounding off, 3, 5, 21, 62, 82
Routing, 58
Rush order, 3
Safety stock, 3, 12, 15, 49, 52, 58, 74
Sensitivity analysis, 3, 6, 12
Service level constraint, 3, 12, 14, 15, 43, 44, 47, 48,
55, 57, 59, 69, 76
Set-up cost, 2, 6, 79
family, 28
item, 28
Set-up time, 6, 28
family, 28
item, 28
Shipping cost, 2, 18, 60, 71
joint, 17
Silver dynamic programming algorithm, 38
Silver heuristic, 39
Silver-Meal heuristic, 11, 39–40, 67–69
Slow mover, 6, 43
Spare part, 79, 83
Special offer, 17, 59
Standard deviation, 12, 53, 70, 86, 87
State space, 37, 45
Stochastic demand, see demand, stochastic
Stockout, 1, 12, 15, 43, 47, 58, 72
cost, 3, 9, 59
planned, 9, 14
Storage capacity, see inventory, capacity
Substitution effect, 17
Supply chain, 19, 55
Synchronization, see reorder instant, synchronization
Target fill rate, see fill rate, target
Time varying demand, see demand, time varying
Value Added Logistics, 56
Value of a good, 60
Variance, 12, 15, 71, 73, 86, 90
Vehicle, 57
Wagner-Whitin algorithm, 10, 38, 66–68
Warehouse, 2, 55, 58
Waste, 79
Zoning, 58
List of symbols
This list only contains the most important nota-
tions. Some symbols may locally have a different
meaning. Random variables are indicated by “r.v.”.
.
= is by definition
≈ is approximately equal to
[. . . ]
+
is max¦0, . . .¦
1 vector of ones
a individual item ordering cost
A family ordering cost
B backlog per cycle
c can-order level
c vector of can-order levels
C(. . .) cost function
C
Y
coefficient of variation of r.v. Y
d discount factor
D demand per unit of time (fixed or r.v.)
D
L
demand during lead time (r.v.)
D
L+R
demand during lead time and review period
(r.v.)
E¦. . .¦ mean of a r.v.
( set of items forming a group
I
{.}
indicator function
h holding cost
H planning horizon
I inventory level
I
O
inventory level at reorder instant
k reorder frequency
k vector of reorder frequencies
L lead time (fixed or r.v.)
M number of orders per unit of time
n ratio between order quantities
N number of items in a family (Ch. 2, Sect. 4.1)
N number of stocking points at an echelon (Ch. 3)
Q order quantity (fixed or r.v.)
p production rate
r carrying charge
R review period
s reorder point
s vector of reorder points
S order-up-to level
S vector of order-up-to levels
T reorder cycle
u set-up time
U family set-up time
v purchasing price
V price break point
w waste rate
α cycle service level
β target fill rate
Π actual probability of no backlog
σ¦. . .¦ standard deviation of a r.v.
τ production time
φ volume of items
Φ inventory capacity
Ψ actual fill rate
100

Preface
This syllabus contains the lecture notes for a part of the course Advanced Quantitative Logistics. This course was given as an elective in the OR track of the Graduate Program in Business at Tilburg University, The Netherlands, from 2003 until 2008. The subject of these lecture notes is Advanced Inventory Management. This concerns models and algorithms for inventory management with interaction between items or between stocking points. The prerequisites for this topic consist of models and algorithms for single item, single stocking point inventory systems with constant, time-varying and stochastic demand as discussed in most introductory textbooks on operations research and management science such as Winston [80, Ch. 16, 17, Sect. 20.7] and as summarized in the first chapter of these lecture notes.

ii

Contents
Preface 1 Introduction to Inventory Management 1.1 Definitions and concepts . . . . . . . . . . . . . . 1.2 Review of single-item policies . . . . . . . . . . . 1.2.1 The basic EOQ model . . . . . . . . . . . 1.2.2 The production lot size model . . . . . . . 1.2.3 The EOQ model with quantity discounts 1.2.4 The EOQ model with planned stockouts . 1.2.5 Time-varying demand . . . . . . . . . . . 1.2.6 Stochastic demand, continuous review . . 1.2.7 Stochastic demand, periodic review . . . . 1.3 Interactions in Inventory Management . . . . . . 2 Coordinated Replenishment 2.1 Constant demand, indirect grouping . . . . . . . 2.1.1 Exact algorithmic solution . . . . . . . . . 2.1.2 Heuristic solutions . . . . . . . . . . . . . 2.1.3 Powers-of-two policies . . . . . . . . . . . 2.1.4 Production lot-size problems . . . . . . . 2.1.5 All-units discounts . . . . . . . . . . . . . 2.2 Constant demand, direct grouping . . . . . . . . 2.2.1 Exact solution by dynamic programming 2.2.2 Heuristic solution . . . . . . . . . . . . . . 2.2.3 Direct grouping with discounts . . . . . . 2.2.4 Indirect versus direct grouping . . . . . . 2.3 Time-varying demand . . . . . . . . . . . . . . . 2.3.1 Exact solution . . . . . . . . . . . . . . . 2.3.2 Heuristic algorithms . . . . . . . . . . . . 2.4 Stochastic demand . . . . . . . . . . . . . . . . . 2.4.1 Can-order policies . . . . . . . . . . . . . 2.4.2 Periodic review policies . . . . . . . . . . 3 Multi-echelon Inventory Systems 3.1 Multi-echelon structures . . . . . 3.1.1 Classification . . . . . . . 3.1.2 Distribution centers . . . 3.2 Constant demand . . . . . . . . . 3.2.1 Linear chains . . . . . . . 3.2.2 Diverging chains . . . . . 3.3 Time-varying demand . . . . . . 3.4 Stochastic demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii 1 1 3 4 6 7 9 9 12 15 17 20 20 22 24 26 28 30 31 32 33 34 35 36 37 39 43 44 51 55 55 56 56 58 58 64 66 69 77 77

4 Other Inventory Systems with Interactions 4.1 Inventory capacity constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

. .1 Random variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bibliography Index List of symbols . . . . . . . . . . .2 Golden section . . . . . . . . . . . . . . . . . . . processes . . . . . . . . . . . . B. . . . . . . . . .2 Probability distributions . . . . . . . . . . . . . . . . . . . . . . . . .2. . . . . . . . . . . . . . . .2. . . . . . . .5 Mixtures of Erlang distributions B Optimization methods B. . . . .1 Normal distributions . . . . . . . . . . . . . . . . . . . . . .2 Collections of random variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 CONTENTS Repairable items . . . . . . . . . . . . . . . . . . . . . . . . . . .2. . .2 Exponential distributions . . . . . . . . . . . . . . . . A. . . . . . . . . . . . . . . . . .iv 4.3 Poisson distributions and Poisson A. . . . . . . . . . . A. .1 Moments and expectation . . . . . . . . . . . . . . . . . . . . . . A. . . . . . . . . . . . . B. . . A. . . . . . . . . . . . . . . .1.3 Bisection . . . . . . . . . . . . . . . .2. . . . . . . . . . . . . . . 79 85 85 85 86 86 87 87 88 88 89 91 91 91 92 93 97 100 A Review of Probability Theory A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . . .4 Gamma and Erlang distributions A. . .1 Stationary points . . . . . . . . . . . . . . . . . . . . . A. . . . . . . . . .2. . . . A. . . .

from month to month (due to a seasonal pattern) and during the lifetime of a product (an upward trend in the beginning. 1 . production or distribution. general concepts in inventory management will be described and explained. • any demand is lost. in the mining-industry of minerals. Inventories are idle goods waiting for use or sale. parts. Section 1. 1. customers are willing to wait if it is difficult to obtain the item elsewhere. The main reasons why inventories are held are that it is uneconomical to produce. and in warehouses. It consists of the handling time at the supplier (the time required for order picking. For customers arriving when an item is out of stock. The distinction between the two extreme cases becomes less important when stockouts occur more rarely. inventory costs and concepts will be described.2 contains a review of models and results for single item inventory management. and loading). Inventory theory aims to develop models and algorithms as an aid to inventory management. part of the demand may be backlogged and part may be lost. the shipping time from the supplier to the stocking point and the handling time at the stocking point (the time required for unloading. In the practical situation of uncertain (stochastic) demand and nonnegligible lead times stockout occurrences cannot be completely avoided. it also includes the production time and possibly a set-up time for the production run. and at shops and by retailers of goods for sale. and the main sources of uncertainty will be indicated. work in progress and finished products. to handle or to transport units one by one and that consumers often do not accept a delay in the delivery of goods or only want to buy goods that are on display or available in a shop. Inventories are kept in many environments. unpacking. Section 1. due to variations in the production plan in a manufacturing environment).1 inventory policies will be classified.3 provides an overview of interactions that may occur in the control of inventories of multiple items and at multiple stocking points.Chapter 1 Introduction to Inventory Management Inventory management deals with ordering and stock keeping of goods for sale.1 Definitions and concepts In this section. a downward trend towards the end). customers go elsewhere to buy the item or give up the intention of buying the item. in factories of raw materials. In Section 1. supermarket or department store. two cases are often distinguished: • any demand is backordered and the backlog is filled as soon as a replenishment is delivered. Lead time: the total time that elapses between the reorder instant and the instant when goods are ready for use or sale. for instance. When the goods still have to be produced after the reorder instant. For some items. packing. and placing on the shelf). The main sources of uncertainty where inventory management has to deal with are Demand: the demand for items may fluctuate from day to day (due to stochastic behavior at retailers. depots and wholesale dealers of goods for distribution.

storage cost. we will discuss costs that may play a role when ordering and storing goods: Ordering cost: the fixed cost of placing an order. the so called carrying charge.2 INTRODUCTION TO INVENTORY MANAGEMENT A prerequisite for applying inventory policies is a good forecasting method for future demand. variable overhead cost and raw material cost associated with producing of handling a single unit. Further. beside on its value. this cost often includes variable labor cost. if the item is made internally rather than ordered from an external supplier. there is a middle class of items consisting of 20%–30% of all items that account for 20%–40% of all sales. Next. this cost is important in the design and control of warehouses. this cost is often called set-up cost and includes the cost of labor. In most companies. these cost only depend on the inventory policy in case of quantity discounts or lost sales. These are called type A items. in case of an external supplier. it also includes shipping cost. – policies with a fixed order-up-to level. of the purchasing cost. Purchasing cost: the variable cost associated with purchasing a single unit of a good. the classification of a raw material item may also be based on how critical the item is for the continuation of work. Most effort of forecasting and inventory control should be concentrated on this type of items. there are type C items consisting of 50%–75% of all items and accounting for only 5%–25% of all sales which are to be controlled by simple and safe procedures. obsolescence. The three most important questions to be answered by an inventory policy are • When to review stocks? A distinction is made between – periodic review policies where stocks are reviewed at fixed time intervals. Shipping cost: the cost associated to the transport of goods from one stocking point to another. a relatively small percentage (5%–20%) of all items account for a relatively large percentage (55%–65%) of all sales. Handling cost: the cost associated to the handling of goods in a warehouse. material and idle time associated with setting up and shutting down a machine for a production run. Further. These type B items require less attention and their inventory can often be controlled by standard procedures. this cost often includes variable opportunity cost incurred by investing capital in inventory. an ABC classification is often carried out. the shipping cost is often included in the purchasing cost. as far as this cost is proportional to the number of items handled it does not influence the minimization of the total inventory cost if all demand is satisfied. the external supplier may want to stimulate larger orders to save on shipping cost by offering quantity discounts. insurance cost. this cost includes the cost of paperwork and accounting associated with an order which is independent of the size of an order. breakage and spoilage. Holding cost: the variable cost of holding a single unit of a good on stock during a unit time period. • When to order? A distinction is made between – periodic review policies where orders can only be placed at the periodic review instants. and cost due to possible theft. • What to order? A distinction is made between – policies with a fixed order quantity. A statistical analysis is required of historical demand and lead time data. if goods are ordered from another location within the same company. if goods are ordered from an external supplier. the carrying charge is strongly related to the interest rate. Finally. – continuous review policies where stocks are reviewed after each transaction. this cost may include internal shipping cost. – continuous review policies which use reorder points in inventory positions. In manufacturing environments. . the opportunity cost is often assumed to be a certain percentage. as far as this cost is proportional to the number of orders handled it can be incorporated in the ordering cost. the review periods.

HANS BLANC. • within companies. In the stochastic demand models the following two service level constraints will be considered: Cycle service constraint: the probability of no stockout in a reorder cycle must be at least a prescribed probability α. Some general considerations in inventory management are: • choose a model in agreement with the availability and the reliability of data (according to the general principle “garbage in → garbage out”). order quantities are chosen in packing units): a complicated. Most of this material can be found in introductory textbooks on operations research or management science. time-consuming algorithm for exact optimization then has little use. and in more specialized textbooks on inventory management. this quantity plays a role in determining holding costs. this quantity is used as a protection against uncertainty in demand and against other irregularities like breakage and pilferage. it may include a penalty cost for loss of future goodwill.2 Review of single-item policies This section contains a review of inventory policies developed for individual items at a single stocking point. Fill rate constraint: the fraction of the demand that is satisfied directly from stock must be at least a prescribed fraction β. such as Winston [80]. it is also called the physical stock. Net inventory (net stock): the inventory on hand minus the amount of backlog. Management cost: the cost incurred by keeping track of inventory levels and by computing order quantities. in all cases. 1. such conflicts may be due to the remuneration system of a company. the latter probability is called the cycle service level (this constraint is also called P1 criterion). it is related to the service level constraint or the cost of stockouts or losses. stockout costs are difficult to assess and are therefore replaced by service level constraints (see below). • in practice. • uncertainty in demand and in lead times. on the other hand. TILBURG UNIVERSITY 3 Stockout cost: in case of backlog of demand it is the extra cost associated to the administration and later delivery of goods. For the case of stochastic demand. technical obsolescence. going out of fashion or decay compel to reservedness toward large stocks. conflicts in interests or goals may exist between the purchasing department (which strives for quantity discounts and delivery of goods at the beginning of a season) and the logistics department (which has to cope with large quantities at the same time. [64] and Axs¨ter [7]. in case of lost sales it is the opportunity cost of lost profit on unsatisfied demand. in many cases. and which may be saddled with superfluous stocks at the end of a season). The following inventory concepts will be used in the various models: Inventory on hand: the number of units actually present at the stocking point. the latter fraction is called the target fill rate (this constraint is also called P2 -criterion). a . • the robustness of a model is important: the resulting replenishment policy should not depend too strongly on the assumptions (like the shape of the demand distribution): this requires sensitivity analysis. quantities are usually rounded off (reorder cycles are chosen in whole days or weeks. this quantity can take positive and negative values. and desired service levels lead to safety stocks. Inventory position: the net stock plus the number of units on order but not yet delivered. such as Silver et al. this quantity is required for determining a reorder instant. this cost is usually not included in inventory models but should form an incentive to choose for inventory policies that are simple to implement. see also Tijms [65]. it may also include extra cost for rush orders or overtime work. risk of disappointing demand. Safety stock: the average inventory position just before a delivery instant.

Figure 1. In the absence of uncertainty in the demand and the lead time.1: Saw-tooth pattern of inventory in deterministic fluid model with lead time L < T .2) The quantity Q∗ is usually referred to as the economic order quantity (EOQ). 2 2 h (1.5) 1 + 1 hDT. cf. The average inventory cost per unit of time can also be formulated as function of the reorder cycle T : C(T ) = a The optimal reorder cycle is T∗ = Q∗ = D 2a .6) .2. and the stock is instantaneously replenished with order quantities Q at delivery instants. hD (1. it is optimal to have a zero inventory level at delivery instants. it follows that the average inventory cost per 2 unit of time.1: Q∗ = 2aD . The reorder cycle (time) T . the average purchasing cost per unit of time is a policy-independent constant amount of D times the purchasing price of a unit of the good. the time between two successive replenishments.1 The basic EOQ model The basic inventory model assumes that demand occurs at a constant. (1. and can be omitted in the optimization procedure.2 shows the average inventory cost per unit of time both as function of the order quantity (C(Q)) and as function of the reorder cycle (C(T )) for the case a = $5. while the minimum average inventory cost per unit of time is C ∗ = 100.4) (1. Q∗ = 50 and T ∗ = 0.3) Example 1.1. Appendix B. The holding cost is h dollars per unit per unit of time. All demand must be satisfied from stock. 1. Based on a fluid approximation in which goods flow out of stock at a constant rate D. for a given order quantity Q follows as T = Q/D. here. The relevant inventory cost for a given fixed order quantity Q consists of ordering costs and holding costs.1. the following results can be derived.1 Figure 1.4 s+Q T Q ↑ I(t) ¯ I s INTRODUCTION TO INVENTORY MANAGEMENT rr r rr rr rr rr r rr rr rr rr rr r rr r rr rr rs 0 rrc rr T −L rr s T rrc rr rr s 2T − L 2T t→ E Figure 1. D = 500 and h = $2. cf. known rate of D units per unit of time. the minimum average inventory cost per unit of time follows by substitution of the optimal order quantity into the cost function: C ∗ = aD h + 1h 2aD 2 √ √ √ 2aD = 1 2ahD + 1 2ahD = 2ahD. In general. h (1. Since the number of orders per ¯ unit of time is D/Q and the average inventory level is I = 1 Q. The cost of placing an order is a dollars per order. 2 T (1.1) Q Q 2 The optimal order quantity is found by differentiation of the cost function. is D D ¯ C(Q) = a + hI = a + 1 hQ. As a consequence. C(Q).

2a (1.2 0.15 0.7) In general.3: Step pattern of inventory in deterministic unit-by-unit model with lead time L < T . cf. the interval ∗ ˜ = 3. Observe that for the optimal order quantity (and the optimal reorder cycle) the average ordering cost is equal to the average holding cost.35 0. TILBURG UNIVERSITY 5 300 300 250 250 200 200 150 150 100 100 50 50 0 0 20 40 60 80 100 120 140 160 180 200 0 0 0. Q Q+1 2 . Similarly. The equation a reduces to Q(Q + 1) = 2aD/h. Since they are interrelated.25 0.3 0.8) Q(Q − 1) < 2aD/h < Q(Q + 1) or √ √ √ ∗ ∗ ˜ ˜ In the √ this way. the equation C(Q) = C(Q − 1) implies Q(Q − 1) = 2aD/h. √ interval 0 < Q < 2 is rounded to Q = 1. etcetera.4 Q→ T → Figure 1. (1. Q∗ . it is usually not possible to round them all to integer values.HANS BLANC. T s+Q Q ↑ I(t) ¯ I s c s 0 T −L c s T 2T − L s 2T t→ E Figure 1. This way of rounding corrects for the skewness of the cost function near 6 < Q < 12 to Q D 1 D + 2 hQ = a + 1 h(Q + 1). and M ∗ are not whole numbers.1 0. This means that the integer order ˜ quantity Q is optimal if ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ Q(Q − 1) < Q∗ < Q(Q + 1).1). The optimal number of orders per unit of time is M∗ = D 1 = ∗ = ∗ Q T hD . (1. A way to find the optimal integer order quantity is by determining the value Q for which the cost C(Q) is equal to the cost C(Q + 1). the interval 2 < Q < 6 to Q = 2. T ∗ .2: Average inventory cost as function of order quantity (left) and of reorder cycle (right).05 0.

However. . In this case. and will be denoted by s. In deterministic systems. A useful property of the EOQ formula is that the cost function (1.11) . In these and later figures. the optimal values Q∗ . and the cost function becomes 2 C(Q) = a D 1 + hQ(1 − D/p). It should be noted that the smooth inventory pattern in Figure 1.2 The production lot size model Consider the following variant of the basic EOQ model in which units are not purchased from an external supplier but internally manufactured. cf. h the holding cost per unit per unit of time. If the lead time L is constant. cf.3. T ∗ and M ∗ are not affected by this lead time.2). To produce Q units takes a time τ = Q/p. . Figure 1. the number of units that can be produced per unit of time.9) Figure 1. The demand during the lead time is D · L. the inventory position (dotted line) only differs from the inventory on hand on the time intervals (mT − L. its minimum. that is. there is always at least one order in transit so that the inventory position is higher than the inventory on hand at every instant.1 includes the reorder instants and the reorder point for the case that the lead time L is smaller than the reorder cycle T . Figure 1. Since the order has to be delivered at the instant when the inventory on hand reaches the level 0 in deterministic systems. . L units of time before the order is needed. Imax = (p − D)τ = (p − D)Q/p. the start of the production can be delayed until the instant at which the inventory reaches the level s = Du. This implies that the optimal order quantity Q∗ is rather insensitive with respect to the cost parameters. . (1.1 is in fact an approximation for the step pattern for discrete units as shown in Figure 1.4 shows the reorder instants and the reorder point for the case T < L < 2T . (1. Q 2 (1.2. namely.4: Saw-tooth pattern of inventory in deterministic fluid model with lead time T < L < 2T . In this case. 2. If the produced items become available as a batch at the end of a production run. The lead time L only determines the instant when an order has to be placed.6 T s+Q ↑ I(t) s Q ¯ I s−Q INTRODUCTION TO INVENTORY MANAGEMENT c ˆˆˆ ˆˆ s ˆ 0 c ˆˆˆ ˆˆˆ ˆˆˆ ˆˆˆ ˆc ˆ ˆ ˆ ˆs 2T − L ˆˆˆ ˆˆˆ ˆˆˆ ˆˆ E t→ ˆˆˆ ˆc ˆ ˆ ˆ ˆs T T −L 2T Figure 1. It is assumed that p > D. and during this time the inventory increases at a net rate of p − D up to. Let a denote the set-up cost of starting a production run. The fluid approximation is better according as larger quantities are involved (fast movers). cf. it is seen that s = D · L. As lead time one has to take L = u + Q∗ /p with u a possible set-up time for a production run.10) 1 ¯ Hence. if the produced units become available for satisfying demand at a rate p during the production run. mT ).5.1) is rather flat near its minimum in many cases. 1. (1. In a production environment the way in which items or substances become available to satisfy demand may affect the optimal lot size. The reorder point then becomes s = DL = Du + DQ∗ /p. the average inventory level is I = (p − D)Q/p. In this case. ◦ marks the inventory level at a reorder instant and • marks the inventory level just before a delivery instant.9). the optimal lot size is determined by the optimal order quantity (1. For items that are sold or used in small quantities (slow movers) the discrete nature of the demand and the inventory pattern should be taken into consideration. the inventory position at reorder instants should be equal to D · L. The inventory position at the reorder instant is called the reorder point. The above derivations do not involve a possible lead time L between the instant at which an order is placed and the instant at which the order is delivered. D the demand rate and p the production rate. m = 1. Figure 1. the maximum inventory is not Q but the inventory level at the end of the production run.2.

Further. . and discount factors dj . The optimal production lot size becomes Q∗ = the optimal production cycle length is T∗ = Q∗ = D 2a . (1.16) Algorithm 1. the price per unit is v(Q) = v0 .5 shows the triangular pattern of the inventory on hand and the saw-tooth pattern of the inventory position for a system with gradual delivery.3 The EOQ model with quantity discounts Some suppliers offer quantity discounts on purchasing prices to their customers to prevent the handling. . otherwise. all results of the model with batch delivery can be translated to the model with gradual delivery by replacing h by h(1 − D/p). including the variable purchasing cost. if v0 Q < V . packing and shipment of small quantities or to stimulate larger sales.1 [EOQ with quantity discount] Step 1: Start with the highest discount factor: m = 1. decrease m m m by 1 and repeat Step 2. v(Q) = v0 (1 − dj ).HANS BLANC. v(Q) = v0 (1 − d1 ). . V1 ≥ V2 ≥ · · · ≥ V . Step 2: Determine the optimal order quantity Q∗ according to (1.12) and the minimum average cost per unit of time is C∗ = 2ah(1 − D/p)D. with r a carrying charge. compute C(Q∗ ) and goto Step 3.13) 2aD . suppose that there exists one or more ( ) price break points Vj . h = rv. that is. Then.2) with h = rv0 (1 − dm ).2. compute C(Vm /v0 ). if Vj ≤ v0 Q < Vj−1 .15) The cost function becomes. d1 ≥ d2 ≥ · · · ≥ d . TILBURG UNIVERSITY T 7 s+Q Imax ↑ I(t) ¯ I c s   ¡ ¡ ¡ ¡ ¡ ¡ ¡ s ¡ ¡         c  ¡ ¡ T −u T ¡ ¡ ¡ ¡ ¡ s ¡ ¡        ¡  c ¡  ¡ 2T −u 2T ¡ ¡ ¡ ¡ s ¡ ¡ E 2T +τ −u 0 τ T +τ t→ Figure 1. such that a discount factor dj is awarded if the value of an order without discount exceeds the price break point Vj . h(1 − D/p) (1. (1. Figure 1. it is assumed that the holding cost is proportional to the purchasing price v. If Q∗ is m m feasible (Q∗ ≥ Vm /v0 ). if v0 Q ≥ V1 .5: Triangular pattern of inventory on hand in deterministic production lot-size model. C(Q) = a D 1 + rv(Q)Q + v(Q)D. h(1 − D/p)D (1. In this situation. 1. Let v0 denote the basic price per unit. Q 2 (1. .14) In fact. . . j = 2.

7: Saw-tooth pattern of net inventory in deterministic fluid model with planned stockouts.125. Hence. cf. Figure 1.17. . This 1 order quantity is not feasible.25 so that h = $2. the quantity Q∗ = 50.2 Consider an item with the following data: a = $5.084. but since C(100) = $4.100.00 it is optimal to order 100 units and accept the discount offer. Example 1. = 0 and note that Q∗+1 is feasible in any case. D = 500.6 (left) shows the average inventory cost per unit of time with (lower graph) and without (upper graph) discount. V1 /v0 } with the m lowest cost. However. Figure 1. This means in the above notations: = 1. the optimal order quantity without discount is computed: Q∗ = 50. if the discount price would be offered for order quantities of 100 units or more. Hence. it is optimal to order 50 units and ignore the discount offer. v0 = $8 and r = $0. with corresponding average cost 2 C(50) = $4. d1 = 0. take d +1 . s+Q T Q−B ↑ I(t)                            0 s −B   τ  T −L   c  T T +τ 2T −L  E c   2T t→ s  s s Figure 1.25. Suppose that the supplier offers a discount price of v1 = $7.8 INTRODUCTION TO INVENTORY MANAGEMENT 4150 4150 4140 4140 4130 4130 4120 4120 4110 4110 4100 4100 4090 4090 4080 4080 4070 4070 4060 4060 4050 20 40 60 80 100 120 140 160 4050 20 40 60 80 100 120 140 160 Q→ Q→ Figure 1. .1 first the optimal order quantity with the discount price is computed: Q∗ = 50.01 and V1 = 1200.25 is still 1 not feasible. Step 3: The optimal order quantity is that quantity from the set {Q∗ . Then. the average cost at V1 /v0 is computed: C(150) = $4. According to Algorithm 1. Vm+1 /v0 . .00.00. Here.6: Average inventory cost as function of order quantity with and without discount.6 (right). . .92 for order quantities of 150 units or more.

23) In the limit as b → ∞. Q∗ → 2aD/h and C ∗ → 2ahD. that√ if the stockout cost becomes very high. This time is determined by Dτ = Q − B. the average cost per unit of time will be determined for this model. then. Section 1. and will be treated as a decision variable beside the order quantity Q.4 The EOQ model with planned stockouts In this section we consider the variant of the basic EOQ model in which stockouts are allowed. TILBURG UNIVERSITY 9 1. h and D have the same meaning as in Section 1.5 Time-varying demand In this section we consider situations in which demand fluctuates over time. holding and stockout cost per unit of time is D τ T −τ C(Q. weeks.21) The actual fill rate. Further. for the optimal policy is Ψ=1− h b B∗ =1− = . hb C∗ = 2ahbD .1. During a time τ the net inventory is positive. h+b B (1. Section 1. 1−β (1. A fraction τ /T of each cycle the net inventory is 1 negative. but is known in advance for a number of future periods (days. and the average level is 2 (Q − B). It will be assumed that B ≤ Q (it is readily verified that this is optimal). 1. the optimal order quantity and the minimum average cost per unit of time are: B∗ = 2ahD . Hence. it follows from the above that B ∗ → 0.17) Next. The inventory on hand is zero whenever stockout occurs. and the average backlog is 2 B. cf. months). The number of orders per unit of time is D/Q and the reorder cycle is T = Q/D.2.2. this implies that a target fill rate β corresponds to a stockout cost of b= βh . When the net inventory is positive it represents the inventory on hand. In this model. (1. It is given by.1. a fraction τ /T of each cycle the net 1 inventory is positive. as in the basic EOQ model. This means that the optimal order quantity for a fixed total backlog per cycle is h+b ˆ Q(B) = B. Time-varying demand may be due to trends or seasonal patterns.18) 2 2 Q Q Q The partial derivative of this cost function with respect to B vanishes if h(Q − B) = bB. which represents the basic EOQ policy. the average cost per unit of time can be expressed in the decision variables Q and B: D (Q − B)2 B2 C(Q. Figure 1. B) = a + 1 h + 1b . The parameters a.19) From this function it readily follows that the optimal backlog per cycle. b will represent the stockout cost per unit short per unit of time.2. (1. is.22) Alternatively. 2 2 Q T T Since τ /T = (Q − B)/Q. . the reorder point s can be positive or negative. the net inventory is negative and its absolute value represents the accumulated amount of backlog. B) = a + 1 h(Q − B) + 1 bB .1.2. cf.20) (1. The accumulated amount of backlog at a delivery instant will be denoted by B. h with corresponding average cost per unit of time ah D 1 ˆ C(B) = + 2 bB.7 shows the saw-tooth pattern of the net inventory as function of time.HANS BLANC. It is assumed that all stockouts are backlogged instantaneously at a delivery instant. b(h + b) Q∗ = 2aD(h + b) . the average ordering.9). s = D · L − B. Q∗ h+b h+b (1. (1. and may also arise as a consequence of material requirement planning in a manufacturing environment. (h + b) (1. cf. Hence.

.24) The objective is to minimize the total ordering and holding costs over H periods: H C(Q1 . . . . given that the inventory level is zero at the end of period t − 1. The computation time of the Wagner-Whitin algorithm strongly increases with the number of periods H. stop. All demand must be fulfilled and no back orders are allowed. . the optimal order quantity for the first period is strongly influenced by variations in the perhaps still uncertain demand in the last period. t = 1. The order quantities are restricted by the requirements that It ≥ 0. . . Step 2: Apply the backward recursion f (t) = 0≤ ≤H−t min {C(t. . . for t = 1. . H. . H − t. The decision variables are Qt . The inventory levels can be described by the recursive equations It = It−1 + Qt − Dt . a single item is treated in isolation. H. . . . (1. .) is a dummy function: δ(0) = 0. . ) = a + h u=1 uDt+u . = 0. t = 1. . 1. Silver & Meal [62] have developed .27) Algorithm 1. t = 1. The inventory level at the end of period t is denoted by It (t = 0. . . (1. . . . . . . H). . H. δ(q) = 1 if q > 0. . otherwise. • f (t): the minimum cost incurred during the periods t. H. . To overcome these difficulties. The demand in period t is denoted as Dt . . the order quantity for the beginning of period t (t = 1. . . Define.26) This property follows by noting that to order only a fraction of the demand of the last period t + is not optimal because it forces an order to be placed for the beginning of period t + . the optimal order quantities have to be recovered from the intermediate results of the recursion: the minimizing in (1. δ(.2 [Wagner-Whitin] Step 1: Set f (H + 1) = 0 and start with t = H. . because the costs can then be reduced by placing an additional order in period t + . H. Wagner & Whitin [77] have developed a dynamic programming solution for this model. It is assumed that the initial inventory level is zero: I0 = 0. . QH ) = t=1 [aδ(Qt ) + hIt ] . Here. . (1. 1. . In some cases. (1.25) here.10 INTRODUCTION TO INVENTORY MANAGEMENT In the context of time-varying demand a discrete time model is considered with a finite planning horizon H (a number of periods). and holding costs of h dollar per unit per period are only charged over the excess inventory. • C(t. . The cost of placing an order is a dollar per order. . . . . . The demand for that period is set aside. = 0. H. . ) + f (t + + 1)}. The minimum total ordering and holding costs is f (1). . . H − t. The minimization of the cost function (1.28) Step 3: If t > 1 decrease t by 1 and repeat from Step 2. A further restriction on the possible optimal values of Qt is obtained by the observation that it is not optimal to include the demand Dt+ in the order of period t if h Dt+ > a. Orders are delivered at the beginning of a period. . . Shifting the fraction of the demand Dt+ from the order quantity Qt to the order quantity Qt+ does not change the ordering costs but diminishes the holding costs. t = 1. . . . . t + : C(t. .25) is facilitated by the observation that the optimal order quantities are zero or equal to the demand of a whole number of consecutive periods: Qt = 0 or Qt = Dt + · · · + Dt+ .28) indicates the optimal number of additional periods for which to order if t becomes a reorder instant. ): the cost of ordering at the beginning of period t a quantity that is sufficient to fulfill the demand of the periods t. (1. . H). H. for t = 1.

HANS BLANC. Step 3: Take τ + otherwise. 1. stop. The heuristic works less well when demand strongly diminishes in later periods and when there are many periods with zero demand. The new average cost is then compared to the previously computed average cost. Then. the average cost over these periods is 1 ¯ C(τ. H − τ. until C(τ. cf. Remark 1. H + 1 − τ. u=1 = 1. They propose a forward recursion. the demand of the first period can be completely satisfied by the initial stock (and. ) for = 1. . as the new value of τ (the new reorder instant). . Algorithm 1. .7]. τ = 1. ¯ The average cost C(τ. ) > a for some . DeMatteis [22].2 The above algorithms do not take a possible lead time into account. ∗ + 1) > C(τ. If there is a deterministic lead time L > 0 the algorithms should be applied a time interval L before the beginning of the first period in order that the first quantity Q1 is delivered in time. Repeat from Step 2 if τ ≤ H.29) The heuristic proposes to place the next order just after a local minimum of the average cost since the previous reorder instant has been reached. τ + . Another heuristic for this model is the part-period-balancing algorithm. (1.6. The cumulative holding cost over these periods is Ch (τ. If the initial stock is nonzero. . . . Silver et al. . ∗ ∗ ≥ 1. ) = a+h −1 uDτ +u . is increased until a period with nonzero demand is found. say I0 > 0. τ + − 1. . Silver & Miltenburg [63] have developed a modification of the Silver-Meal heuristic for such cases. 2. [64. . Take Qτ = Dτ + · · · + Dτ + ∗ −1 as the order quantity for period τ . . The order quantity for the first period only depends on the demand in the periods up to the period in which such a local minimum occurs. . . ) is not computed if Dτ + −1 = 0. the total cost of the generated order policy is less than 1% above the minimum. . . . Sct. .1 The Silver-Meal heuristic has been tested extensively. the solution can be improved by shifting the quantity DH to the previous reorder instant τp if h(H − τp )DH < a. Let again τ be a reorder instant at which an order is placed for the periods τ. Repeat from Step 2 if τ ≤ H. 6. while the modified demand ˜ for the second period becomes D2 = D2 − (I0 − D1 ) > 0. . TILBURG UNIVERSITY 11 a heuristic for this model. . ∗ ) for some ∗ or until τ + = H + 1.30) This heuristic tries to compensate ordering cost against future holding costs. . Remark 1.3 The above algorithms assume that the initial stock is zero. if D1 ≤ I0 < D1 + D2 . In many cases. as the new value of τ (the new reorder instant). H. τ = 1. Then. then the demand of the first few periods have to be diminished by the initial stock before the algorithms can be applied with the modified demand. . cf. see Baker [9]. . Mendoza [52]. Algorithm 1. . . Remark 1. If the foregoing procedure generates an order in period H. The algorithms are then applied from period 2 onwards. ) = h u=1 uDτ +u . . ¯ ¯ ¯ Step 2: Compute the average costs C(τ. Take Qτ = Dτ + · · · + Dτ + −1 as the order quantity for period τ .3 [Silver-Meal heuristic] Step 1: Start with τ = 1. Let τ be a reorder instant at which an order is placed for the periods τ. . = 0. .4 [Part-period-balancing] Step 1: Start with τ = 1. Step 3: Take τ + otherwise. Step 2: Compute the cumulative holding costs Ch (τ. hence. . H. Q1 = 0). For instance. . ) for = 1. . (1. . 2. A new order is placed as soon as the cumulative holding cost since the previous reorder instant exceed the ordering cost. . or until τ + = H + 1. . stop. until Ch (τ. .

E{[DL − s]+ } − E{[DL − S]+ } ≤ (1 − β)Q = (1 − β) 2aE{D} .2. and a safety stock may be required to prevent stockouts under such circumstances. Q=S−s= 2aE{D} . Q) policy: whenever the inventory position drops at or below the reorder point s a fixed quantity Q is ordered.4 The foregoing algorithms can all be applied with a rolling horizon.33) Although the decision variables s and S should be simultaneously determined to minimize E{C(s. the random variable B denotes the amount of backlog accumulated during a reorder cycle. .1.12 INTRODUCTION TO INVENTORY MANAGEMENT Remark 1. S) policy: whenever the inventory position drops at or below the reorder point s an order is placed of a size that brings the inventory position to the order-up-to level S. With constant lead times. the expected average cost per unit of time is E{C(s. becomes available. cf.34) usually rounded to the nearest integer. [DL − s]+ is the amount of backlog at the end of a reorder cycle and [DL − S]+ is the amount of backlog at the beginning of a reorder cycle. S−s (1. . (1. continuous review This section is concerned with inventory systems with stationary stochastic demand and continuous review policies. . the mean and the variance of DL are given by E{DL } = LE{D}. Lead times are assumed to be constant.32). In general. the (s. σ 2 {DL } = Lσ 2 {D}. [x]+ = max{0. H + 2. Section 1.6 Stochastic demand.35) . In the analysis of continuous review policies an important role is played by the demand during the lead time. (1. or if demand information over new periods H + 1. This may be useful when the demand in periods 2.1. cf. H is still subject to change. As in Section 1. In case of unit demand per customer. and equal to L units of time. . . it is common practice to follow a sequential approach in which first the order quantity is determined according to the EOQ formula. a denotes the ordering cost per order and h denotes the holding cost per unit per unit of time.32). defined as s − E{DL }. Further. h (1. here. In case of a fill rate constraint with target β. This random variable will be denoted by DL . h (1.32) . The most commonly used continuous review policies are: • the (s. a new solution with a new horizon is preferably determined at the first reorder instant after period 1.2. Policies with an order-up-to level are also called base-stock policies. In the latter case. Q) and (s. The expected demand per unit of time will be denoted by E{D} and the standard deviation of the demand per unit of time by σ{D}. By a balance argument. the expected demand during a reorder cycle is equal to Q = S − s. S)} = aE{D} 1 + 2 h[E{[s − DL ]+ } + E{[S − DL ]+ }]. Q S−s (1. S)} subject to the fill rate constraint (1. problems occur when the demand changes for periods for which the demand has already been ordered. 1. the solutions of the models with deterministic timevarying demand are very sensitive to small irregularities (breakage. . it is assumed that each customer demands one unit and that all demand occurring when the item is out of stock is backlogged. is needed to protect against the uncertainty in the demand during the lead time. x} for all real x. and then the reorder point s is determined as the smallest integer such that.31) The safety stock. . the reorder point s has to be chosen such that the actual fill rate Ψ exceeds the target fill rate: Ψ=1− E{B} E{[DL − s]+ } − E{[DL − S]+ } =1− ≥ β. S) policies are equivalent when Q = S − s. that is. Blackburn & Millen [12] have found that the Silver-Meal heuristic may outperform the Wagner-Whitin algorithm in a rolling horizon environment because the solution found by the latter method is too sensitive to changes. Since the expected number of orders per unit of time is E{D}/Q and since the expected inventory on hand just before a delivery instant is E{[s − DL ]+ and just after a delivery instant is E{[S − DL ]+ . loss) and upward changes in demand. In the first case. • the (s.

00 6. j! m = 0.9806 0. S−s 2 j! j! j=0 j=0 Consider the following numerical example: a = $5.00 1.5442 64 69. 5. .00 6. TILBURG UNIVERSITY 13 This sequential approach performs well if σ{DL } < Q. Further.82 0.00 1.82 2.00 6.03 0.98.5592 48 54. It contains for several values of the order quantity Q around Q = 44.00 0.98. Q = 49.82 2. Similar remarks can be made as those made with regard to Table 1. note that the expected backlog at the beginning of a reorder cycle.03 2.9815 0.82 0.82 2.00 0.5). S)} 44 50. 107) = $2.9837 0.82 2.00 6.82 2.00 1.5704 51 57.9827 0.82 0. β = 0.5373 is met which is less than C(56.00 0. Table 1. The expected average cost has a local minimum at Q = 45.03 0.82 0. Note that the expected average cost and the actual fill rate are rather flat functions of the order quantity Q in this region.9834 0. The safety stock is now s − E{DL } = 31 in the region of Q = 44.3 In the case that unit demand occurs according to a Poisson process at a rate of λ = E{D} customers per unit of time. The safety stock is s − E{DL } = 6 in this region.9839 0.29 0.03 2. Q = 54.00 0.82 2.82 2.5658 50 56. However.00 0.82 0.2 concerns the case of a lead time of L = 50 days and a target fill rate of β = 0.72. e ≤ (1 − β) j! j! h j=s j=S The infinite sums can be computed by the relation ∞ (j − m) j=m (λL)j −λL e = λL − m + j! m−1 (m − j) j=0 (λL)j −λL e . See also Federgruen & Zipkin [29] and Axs¨ter [7.82 0.98.1.00 5.9840 0.03 0.00 1.00 6.47.82 0. Tijms [65].2. .00 6. Local minima of the expected average cost are found at Q = 45.9819 0. cf.00 6. E{[DL − S]+ }. .6264. 119) = $2.82 2.33) becomes   s S (λL)j aλ (λL)j  −λL E{C(s.9823 0. 1.5573 47 53. Table 1.5758 52 57.5565 46 52.9802 Table 1.00 0. This implies that Q = 44.6].34).82 0.9830 0. s 56 56 56 56 56 56 56 56 55 55 55 54 S 100 101 102 103 104 105 106 107 107 108 119 119 Q E{[S − DL ]+ } E{[s − DL ]+ } E{[DL − S]+ } E{[DL − s]+ } E{C(s.00 6.5567 45 51.1 concerns the case of a lead time of L = 5 days and a target fill rate of β = 0.5565. The constraint (1. Q = 59. E{[DL − S]+ }.29 2.5373 53 58. h = $0. The second term in the fill rate constraint. .5620 49 55. and a Poisson demand rate of λ = 10 units per day.HANS BLANC. Zheng & Federgruen [82] describe an efficient algorithm to search the optimal values of s and S simultaneously. when the order quantity becomes Q = 52 the required reorder point drops from s = 56 to s = 55 and another local minimum of C(55. The next local minimum is found for Q = 65 but here the cost is higher: C(54. is negligible so that the expected stock level at the beginning of a reorder cycle is E{[S − DL ]+ } ≈ S − E{DL } = S − 50.6264 Ψ 0. is negligible if the target fill rate is large (β > 0.00 0.6570 65 69. the random variable DL has a Poisson distribution with parameter λL. with corresponding reorder cycle T = 4. . 101) = $2.72 the policy with the minimum value of the reorder point s that meets the fill rate constraint. The expected average cost per unit of time (1.72 by (1. S)} = + 1 h  (s − j) + (S − j) e .05 per unit per day. Appendix A. Ch.00 6. cf.1: Feasible reorder points and order-up-to levels for L = 5.3.9) and the coefficient of variation of the demand during the lead time DL is small (σ{DL }/E{DL } < 0.03 2. 2.00 0.35) then reads   ∞ ∞ (λL)j  −λL (λL)j 2aλ  (j − s) − (j − S) . a Example 1.00 6.9802 0.00 6.

Note that the expected backlog at the beginning of a reorder cycle.00 0.5813. .00 32.00 0. β = 0. E{[s − DL ]+ } is zero.00 0.00 28.7631 Ψ 0.98. also the expected stock level at the end of a reorder cycle.00 0.5826 Ψ 0.6638 48 24.27 3. Note that a ten-fold increase in the lead times leads to a 48% increase in cost. The expected net stock at the end of a reorder cycle is negative so that this case with a low target fill rate corresponds to the model with planned stockouts.00 0. The global minimum is found for Q = 64: C(18.6861 46 23.00 0.00 0.88 0.00 0.00 30.00 32.5813 66 33.00 33.88 0. s 531 531 531 530 529 528 527 S 575 576 577 579 583 587 591 Q E{[S − DL ]+ } E{[s − DL ]+ } E{[DL − S]+ } E{[DL − s]+ } E{C(s.6417 62 31. Remark 1.5000 0. the safety stock.3: Feasible reorder points and order-up-to levels for L = 5. The safety stocks.5111 0.5000 0.88 0.14 INTRODUCTION TO INVENTORY MANAGEMENT Table 1. a better order quantity is found than the quantity Q = 45 suggested by (1.9805 0.00 0.7515 64 91.00 23.4 concerns the case of a lead time of L = 1 days and a low target fill rate of β = 0.6865 45 23.00 1.00 0.00 31.00 1. In this case.00 0.5107 0.00 0.00 0. is negative in these cases. S)} 44 22. s − E{DL } = s − 50.5000 Q = 64. Section 1.3 concerns the case of a lead time of L = 5 days and a low target fill rate of β = 0.00 1.06 3.2: Feasible reorder points and order-up-to levels for L = 50. S)} 44 22. Table 1.00 0.00 24.4.06 0.3 is that the reorder point has become negative.00 0.00 23.5815 64 32.5 In case of a cycle service constraint with cycle service level α. The global minimum is C(528.00 0. The main difference with the case of L = 5 days in Table 1.9809 0.00 1. The expected average cost and the actual fill rate are more sensitive to the order quantity Q. is still negligible.00 1.5000 0.00 23.5000 0.6639 48 24. Table 1.8082 46 77.00 0.9804 0.00 0.00 31.5813 66 33. β = 0.7696 54 83.50.9801 0.00 30.00 1.00 0. S)} 44 75.00 23.4: Feasible reorder points and order-up-to levels for L = 1.00 0. The cost function has a local minimum for every even order quantity Q.00 0. The safety stock is here s − E{DL } = 28. etcetera.5826 Ψ 0.00 0.00 1.6620 47 24.00 22.00 1. 82) = $1.00 1.00 29.00 22.00 1.8084 45 76.50.97 3.6862 46 23.00 0.00 1.9803 0. Also.5000 0.00 0.00 22.5815 64 32.5001 0.00 1. s −12 −12 −13 −13 −14 −21 −22 −23 S 32 33 33 34 34 41 42 43 Q E{[S − DL ]+ } E{[s − DL ]+ } E{[DL − S]+ } E{[DL − s]+ } E{C(s.34).00 33.00 1.5106 0.50.50.00 1.7525 59 87.00 0.00 0.00 1. 587) = $3.00 0.2. In all four numerical examples.00 0.5000 Table 1.9801 Table 1. s − E{DL }.8090 49 79.00 0.00 0.00 1.5000 0.00 24.00 31.00 0.00 0.00 1. the reorder point s has to be chosen such that Pr{DL ≤ s} ≥ α.97 0.16 0.00 0.7515.00 1.88 3.00 0.00 0.6864 45 23.00 31.5000 0. E{[DL − S]+ }. s 28 28 27 27 26 19 18 17 S 72 73 73 74 74 81 82 83 Q E{[S − DL ]+ } E{[s − DL ]+ } E{[DL − S]+ } E{[DL − s]+ } E{C(s. are the same for L = 1 and L = 5 at given values of Q.5000 0.16 3.00 0.88 3.00 1.27 0. cf. β = 0.5112 0.9803 0.5000 0.00 22.6417 62 31.88 3.00 31.6620 47 24.

E{Q} RE{D} (1. TILBURG UNIVERSITY For the case of a Poisson demand process this condition reads s j=0 15 (λL)j −λL e ≥ α. In all cases. it is delivered after a lead time L while the next order can only arrive after a time L + R.33).2. we will only discuss the (R. so that RE{D} = 80 and E{DL+R } = 100.HANS BLANC. σ 2 {DL+R } = (L + R)σ 2 {D}. The expected average cost per unit of time is E{C(R. in analogy to (1. If an order is placed at a review instant.36) hE{D} If the review period R is much smaller than the value R∗ .5). S)} = a + 1 h[E{[S − DL+R ]+ } + E{[S − DL ]+ }]. (1. S) policy: with intervals of R units of time the inventory level is reviewed. with this service level constraint first s can be determined as the smallest integer satisfying the above constraint. and the expected value of an order quantity is E{Q} = RE{D} by a balance argument. S) policy.2.31). The cost factors are a = $5 and h = $0.6. the mean and the variance of DL+R are.40) Example 1. the current order quantity should be sufficient to protect against the uncertainty in the demand during a lead time plus a review period. s. With constant lead times. in analogy to the sequential approach in the continuous review case. s.05 per unit per week. the review period is R = 4 weeks and the lead time is L = 1 week. Hence. Hence. (1. and for a fixed review period R. The most commonly used periodic review policies are: • the (R. The review period will be denoted by R. This random variable will be denoted by DL+R . The mean demand per week is E{D} = 20.4 The required order-up-to level S will be determined for various distributions for the demand during the lead time plus review period. S) policy: with intervals of R units of time an order is placed of a size that brings the inventory position to the order-up-to level S. (1. The notations and assumptions are otherwise the same as in Section 1. The safety stock for this policy is S −E{DL+R }. as in (1. R = R∗ = . S) policy to avoid many small orders. The order-up-to level S is obtained by solving E{[DL+R − S]+ } − E{[DL − S]+ } ≤ (1 − β)RE{D} = 80(1 − β). it is preferable to use an (R. Here. given by E{DL+R } = (L + R)E{D}.36).39) Note that there is no decision variable left to minimize this objective function if R is fixed and S is set to satisfy the service level constraint (1. The expected inventory on hand is simply related to the expected backlog through E{[S − DL+R ]+ } = S − (L + R)E{D} + E{[DL+R − S]+ }. whenever the inventory position is then at or below the reorder point s an order is placed of a size that brings the inventory position to the order-up-to level S.38).7 Stochastic demand. and then a value for S can be searched that minimizes the expected average cost per unit of time (1. j! Note that this condition does not involve Q or S. R 2 (1. periodic review This section is concerned with inventory systems with stationary stochastic demand and periodic reviewpolicies. • the (R. 2a .38) here. . the order-up-to level S has to be determined such that a fraction β of all demand can be sold directly from stock (fill rate constraint): Ψ=1− E{B} E{[DL+R − S]+ } − E{[DL − S]+ } =1− ≥ β.37) When all stockouts are backlogged.32). 1. so that R = 4 is the largest integer smaller than √ R∗ = 20 according to (1. B stands for the accumulated backlog during a review period R. The length of the review period is either assumed to be given by other management considerations or can be chosen as the optimal reorder cycle length (1.

95 (tail behavior) and higher levels for β ≤ 0. S = 244 for β = 0.03673.56. The parameters 1 1 1 1 of the gamma distributions are ψ = 4 . and K1 = K2 = 1.5123 2.75.0036.9803 12. This√ implies.3 for β = 0. λ2 = 0.23 0. Observe that the mixed Erlang distribution yields lower order-up-to levels than the gamma distribution.44 0.97 0.9 to about 0.9382 1. cf.9704 11. The expected backlog at the beginning of a cycle varies from about 1.7204 0.8045 4. K1 = 1. Use of the normal distribution gives much lower order-up-to levels. The parameters of the mixed Erlang distribution are. Target β 0. respectively.0188. These expected backlogs are also computed with the aid of the gamma loss function.2.4263 3.7132 8.1779. λ1 = 0. with parameters µ = 100. (A.81 S 184 229 243 261 287 329 ME (Erlang-2 and Exponential) Ψ E{C(R. S)} E{[S − DL+R ]+ } 0.9 to about 0.25 0.6368.60 0. cf.3624 2.99. cf.99 S 105 116 120 124 129 137 Normal Ψ E{C(R.47.5: Order-up-to levels. σ = 25 and µ = 20. (A.67 0.38) leads to order-up-to levels that are about 50 higher than those displayed in the table (for the hyperexponential distribution the difference decreases with increasing β). respectively. with a coefficient of variation CL ≈ 1. p = 0. The probabilities of negative realizations are for these normally distributed random variables 3 × 10−5 and 0.00327. λ = 400 and ψ = 20 .037.9047 4.80 the gamma distribution is compared with a mixture of two Erlang distributions in a case with a moderate variation in the demand: σ 2 {D} = 1125.6 Table 1.05 0.01 0.1493 0.9041 3. In these cases. λ = 400 . with parameters 4 4 ψ = 16. λ = 25 (hence. 9 45 4 λ = 225 .9619 4.9719 4.9900 14. λ1 = 0.68.98 0.39 0.. p = 0.1487 3.99 S 185 231 245 264 290 333 Gamma Ψ E{C(R. Observe that the normal distribution yields lower order-up-to levels than the gamma distribution due to the lighter tail of the normal distribution.01964.25.25 0.44.95 0. cf. p = 0.5062 2.6: Order-up-to levels. respectively.95. λ = 25 .27 0. In Table 1. λ1 = 0.95 0.2465 3.77 S 105 118 122 126 133 143 Gamma Ψ E{C(R. λ2 = 0. with a coefficient of variation CL ≈ 4. K1 = K2 = 1. and.9809 4.9008 7.26 0. S)} E{[S − DL+R ]+ } 0.9624 4.42). ignoring this quantity in (1. λ1 = λ2 = 0.61 0.6918 7.53 0.52 0.7380 1. the expected backlog at the beginning of a cycle is negligible. The normal distribution is not appropriate in this case since the probability of negative realizations of DL+R is .9506 4. p = 0.90 0.9701 4. average cost and backlog at end of cycle for σ 2 {D} = 1125.9602 10.97 0. This implies that σ{DL+R } = 200 with a coefficient of variation √ CL+R = 2. The expected 4 backlogs are computed with the aid of the gamma loss function with parameters ψ = 16 .99.3260 3. cf.46).21).54. The parameters of the hyperexponential distributions are K1 = K2 = 1.9207 0.4 for β = 0. The corresponding loss functions are given in (A.07 0. K2 = 2. (A.99.90 0.43 0. In Table 1.6098 2.9503 9. S)} E{[S − DL+R ]+ } 0.9447.18. The expected backlog at the beginning of a cycle varies from about 0.4835 3.02 for β = 0.78 0. e.9901 14. This implies that σ{DL+R } = 75 with a coefficient of variation √ CL+R = 0. Target β 0. Observe that the hyperexponential distribution yields lower orderup-to levels than the gamma distribution for β > 0. fill rate.9602 10.16 INTRODUCTION TO INVENTORY MANAGEMENT Table 1.08 0. σ = 11. λ = 225 and ψ = 16 . respectively.44) and (A.9506 9. (A.77 In Table 1.4193 0. respectively.98 0. respectively. an Erlang distribution) and ψ = 3.0221.43).37).38) leads to order-up-to levels that are about 2 higher than those displayed in the table.5841 3.6633 8.9904 5.18.6408 1. respectively.9701 11.96 0.7 the gamma distribution is compared with a hyperexponential distribution (H2 ) in a case with a high variation in the demand: σ 2 {D} = 8000.9803 12.1213. The expected backlogs E{[DL+R − S]+ } and E{[DL − S]+ } are computed with the aid of the normal loss function.5 the normal distribution is compared with the gamma distribution in a case with a low variation in the demand: σ 2 {D} = 125. The normal distribution does not seem appropriate in this case since the probability of negative realizations of DL+R is 0. and that σ{DL } = 40 5 ≈ 89.9027 3. fill rate.g.9004 7. λ2 = 0.9903 5. that σ{DL+R } = 25 with a coefficient of variation CL+R = 0.38). and that σ{DL } = 5 5 ≈ 11.6946 7.95 0.86 0.7913 1. average cost and backlog at end of cycle for σ 2 {D} = 125. ignoring this quantity in (1.1 for β = 0.9810 4.09. and that σ{DL } = 15 5 ≈ 33. S)} E{[S − DL+R ]+ } 0. with a coefficient of variation CL ≈ 0.7390. (1.9517 4.96 0.

9600 45. but of a different color).3 Interactions in Inventory Management The basic models of inventory theory assume that there is no interaction between various items and various stocking points. 17 Target Gamma H2 β S Ψ E{C(R.9800 56.90 251 10. and consider individual items at a single stocking point.9900 69.75 168 11.63 0.3.26 3.93 0. S)} according to gamma distributed demand at a target fill rate of β = 0.25 62 6.9500 41.88 2. – complementarity effect (several units of items that fit together.9000 31.HANS BLANC. and an inventory policy may be far from optimal when such interactions are neglected. e. – joint discount structure (on the total amount or the total quantity of an order). However.98 4. 1. in many practical cases interactions do exist.15 1.01 210 9. S)} S E{C(R. As could be expected.40 655 0.47 0. – a common labor force and equipment for the receipt and handling of various items.95 863 0.00 790 43.99 1415 0.00 1.90 0.08 2.85 1157 0. S)} S E{C(R.23 863 41. Interactions may be due to (see.32 9. S)} E{[S − DL+R ]+ } 0. fill rate.g. – a single budget available for the inventory of various items. TILBURG UNIVERSITY Table 1.25 137 4..96 938 0.7: Order-up-to levels. The minimum cost is attained for R = 3. S)} S E{C(R.9900 66.82 0.9700 50.27 9.55 99 4. R=1 R=2 R=3 R=4 R=5 CL+R S E{C(R.07 814 41.9800 57.06 Finally. S)} S E{C(R.90 636 0. • interwoven cost structure: – joint ordering cost for a family of items ordered from the same supplier.69 0.92 1373 0.g.9701 49.75 231 9. Table 1.19 118 4.67 872 0.98 1174 0. Ch.40 0. S)} E{[S − DL+R ]+ } S Ψ E{C(R.9000 30.95 189 10. Love [49.03 0. the required order-up-to level is an increasing function of the review period.95.75 0.8: Order-up-to levels and average cost for various R at β = 0.95 for the coefficients of variations in the demand considered above and for various lengths of the review period R. S)} 0. .97 1035 0. Observe the tendency that a shorter review period becomes cheaper when the coefficient of variation in the demand increases.9501 41.79 1031 0. Table 1.53 4.. – joint shipping cost.72 941 0.29 839 41. or make up a whole). – substitution effect (e.53 888 42.40 3. 5]): • joint capacity constraints: – a common storage space used for various items. the same item.8 shows the required order-up-to levels S and the expected average cost E{C(R. • correlated demand: – assembly requires parts in fixed proportions.63 80 4. average cost and backlog at end of cycle for σ 2 {D} = 8000. it turns out that in all above discussed cases a shorter review period gives a lower expected average cost. – special offers or advertisements at a regional or national scale influence sales at many stocking points.9601 45.18 2.

– produced parts are used in several final products. in coping with unusual situations). coordinated replenishment strategies will be discussed for families of items with a joint cost structure.g. • interwoven product structure: INTRODUCTION TO INVENTORY MANAGEMENT – raw material is used in the production of several items. etcetera).g. more complex computations). Disadvantages of joint replenishment (or joint production) strategies may be: • higher average inventory levels (since orders may be put forward and EOQ quantities are adjusted). In Chapter 2. Advantages of joint replenishment (or joint production) strategies may be: • lower purchasing cost. – items belong to the same product group (with limited mutual switch-over times) or are variants of each other (only differing in packing. switch-over). shipment). • reduced flexibility (e.18 – joint handling cost in a warehouse. • lower production cost (set-up. enclosing. The considered interactions through cost structure are: • joint ordering.. – the order quantities are important in this case. – the key question is: should an item be ordered at a reorder instant or not (the order quantities are of minor importance). • lower shipping cost. • lower handling cost. independent of the composition of the order (administration. • inventory management more complex. • lower ordering cost. – example: a contribution in shipping cost is only passed on below a certain order amount. depending on whether item i forms part of an order. • higher peaks at the receipt of goods. – there are additional ordering costs (ai ). • joint discount structure: – a supplier offers discount on purchasing prices or shipping cost if the total value or volume of an order exceeds some threshold. clothing or items bought from a mail order company) or after replacement by a spare part when repaired. – there is a fixed ordering cost (A). • interwoven distribution structure: – an item is kept on stock at several locations (within the same company or organization). – sold units may return to a stocking point because of customer dissatisfaction (e.. production and/or transportation costs: – items/parts stem from the same supplier or can be shipped with the same transport. . • higher management cost (more involved reviews.

from raw materials via parts and semi-manufactured items to end products which reach customers possibly via several distribution nodes: this results in multi-echelon inventory systems. repairable items.. product advertisements (globally). whereby demand usually occurs in larger quantities at higher echelons) to avoid possible oscillating effects. .g.HANS BLANC.g. demand patterns.. via EDI: electronic data interchange): • from lower echelons upward: concerning reorder frequency. In multi-echelon inventory management a distinction is made between pull and push systems. a central planning may be implemented that allows limited local modifications provided that the latter are reported in time. As a compromise. In a push system information is kept up to date at a central level and stocks are controlled from a central point. special actions (locally). proper information exchange between the echelons is important (e. • may give problems with responsibilities of local managers. TILBURG UNIVERSITY 19 Chapter 3 is concerned with replenishment strategies for multi-echelon inventory systems where items are kept in stock at various storage locations. In order to implement a multi-echelon replenishment strategy. Storage locations often form part of complex good flows (supply chains). to prevent hoarding). In a pull system each stocking point determines its own reorder policy. Chapter 4 is devoted to replenishment strategies for inventory systems with capacity constraints and with returning. Coordination of inventory management on various stocking points is possible as far as they are controlled by the same company or if good agreements exist between companies (e. • from higher echelons downward: concerning production or replenishment plan: when becomes how much available of which item. Coordination of inventory management at various stocking points is desirable (since orders of lower levels form the demand process for higher levels. • realize cost savings by better anticipation and coordination. Push systems • require much and reliable information.

The following concepts will be used: family: a set of items that possesses a joint cost structure (e. • the delivery of orders (or of production lots) takes place as a whole. • no back orders or stockouts are allowed. • no capacity restrictions on order quantities (or on production lot sizes) exist.1 Constant demand. • a continuous review of stocks is applied. • the lead time is negligible or constant and is equal for all items. with vi the purchasing cost (value) of item i and r an item independent carrying charge. The analysis will be based on fluid approximations for the demand and the inventory level. algorithms and policies for inventory systems in which coordinated replenishment of items may be profitable. models with constant. group: a set of items from a family that have the same constant reorder cycle time. with coordinated ordering of several items this is not necessarily the case. • a supplementary ordering cost ai for item i.Chapter 2 Coordinated Replenishment This chapter is devoted to models. often. with time-varying and with stochastic demand will be discussed. or because they are delivered by the same transporter). • a holding cost hi for item i is incurred per unit per unit of time. A distinction will be made between indirect and direct grouping of items. because they are ordered from the same supplier. only within groups). direct grouping: the reorder cycle of a group is not related to those of other groups within the same family (this grouping aims at single saving on ordering cost. cf. which is independent of the composition of the order. which only is incurred when this item is included in an order. hi = rvi . Chakravarty [16]: indirect grouping: the reorder cycle of a group is an integer multiple of the family reorder cycle (this grouping aims at double saving on ordering cost. With an individual item it is optimal to have a fixed reorder cycle. in general it is assumed as a simplification that a group of items has a fixed reorder cycle (this may be suboptimal). Consecutively. The assumptions for the model with constant (known) demand for a family of N items are: • a family or joint ordering cost A.g. • the demand rate Di for item i is a constant number of units per unit of time. 2. both within groups and across groups). 20 .. indirect grouping This section is concerned with indirect grouping policies for a family of items with constant demand.

(2. N. • item i is included in every ki th replenishment of the family (ki = 1. k) = 1 ai A+ + 1T 2 T ki i=1 N N ki hi Di . a restrictive assumption is that mini {ki } = 1 (otherwise. N N 2ai hi Di ≤ C ∗ ≤ i=1 i=1 2(A + ai )hi Di . . . .6):  ˆ C(k) = 2 A + j=1 N  aj  kj N N ˆ ki hi Di = T (k) i=1 i=1 ki hi Di . The relative values of the optimal frequencies are independent of A. the frequencies ki (T ) are ordered according to ascending value of ai /(hi Di ). . In such a case. .). The objective function to be minimized reads for indirect grouping: C(T. (2. The optimal family cycle for a given vector of reorder frequencies k = (k1 . cf. For the case of indirect grouping of items the decision variables are more specifically defined as: • with time intervals of length T a family order is placed (the family reorder cycle). Remark 2. the reorder cycle of item i has a length of Ti = ki T . it follows for a given family reorder cycle T that the frequencies ˆ ki (T ) = 2ai . (2. hi Di T 2 i = 1. coordination may not be profitable. . and by minimizing the resulting separable function in the variables Ti = ki T . i=1 (2. Hence. .HANS BLANC. . ˜i (T )(ki (T ) − 1) ˜ k i = 1. 2. . hi D i T 2 i = 1. Further.2) The corresponding average inventory cost per unit of time is.1 A trivial upper bound on the minimum cost is obtained by using the EOQ formula with ordering cost A + ai for each item separately and by summing the minimum cost per item over all items.4) ˆ ˜ would be optimal if they were integer. (1. kN ) can be derived in a similar way as the basic EOQ-formula. . under this rounding the family cycle T is bounded by 2ai hi Di 1 ≤T < ˜i (T )(ki (T ) + 1) ˜ k 2ai hi Di 1 . .3) In a similar way. . . . TILBURG UNIVERSITY 21 The aim is to minimize the average ordering and holding cost per unit of time (purchasing costs are fixed since no back orders or shortages are allowed). Items with the same reorder frequency ki are said to form a group. . .6) . The actual optimal values do depend on A through T .5) and Appendix B. N. (2. Hence.8) states: round ki (T ) to ki = ki (T ) if ki (ki − 1) < 2ai ≤ ki (ki + 1). N. but in case of coordinated replenishment it is more suitable to consider the reorder frequencies for the N items.1: ˆ T (k) = 2[A + N j=1 (aj /kj )] N i=1 ki hi Di . A rounding-off rule like (1. . cf. a modification of the objective function is required with respect to the term A/T since family orders are not placed every T units of time). If the minimum cost C ∗ with coordinated replenishment is close to the upper bound. . The decision variables could be the order quantities for the N items. .5) ˜ Clearly. (1. the increase in management cost may be larger than the reduction in inventory cost obtained by coordination. A lower bound is obtained by ignoring the term A/T .1) Here. (2.

ˆ A minimum value for the function T (k) is obtained from the lefthand inequality in (2.. Hence. Step 4: Increase j by one. If Tc ≤ Tmin then goto step 6. have been determined. . ki (ki + 1) (2. Change kic (T ) to kic (T ) + 1 to obtain the ˆ vector kj and compute the corresponding minimum cost C(kj ). N .2. . and that the inequalities (2. and the corresponding costs under Ci . . Remark 2.2 It is readily verified that the function T (k). call this vector k1 . Note that in this academic case the reorder instants can be synchronized in cycles of 42 weeks in which there are 21 + 7 + 2 = 30 family reorder instants instead of 21 + 14 + 6 = 41 without synchronization. . 42 A = $1.1. Example 2.2). i = 1.3) as Qi = Di ki T. 1). 11 In this way. Step 5: Let ic be the item for which Tc (ic ) = maxi {Tc (i)}. Tmax ].7) Set j = 1 and start the recursion (steps 4 and 5). cf.1. This implies that for all feasible vectors k we have . Further data (per week) can be found in Table 2. is decreasing in each individual ki . otherwise continue with step 5. (2. . the first four columns.. The minimum total average cost per week without coordination is C = $29. ˆ T (k) ≥ Tmin = min ai .2. . . cf. . 1. ˆ Step 6: The algorithm stops: the optimal policy is that policy (T (kj ). sponding costs C(k ˜ Step 3: Determine for each item i the length of the family cycle Tc (i) for which the value of ki = ki (Tmax ) changes to ki + 1 via the bounds (2.1.6) divide this interval into a finite number of subintervals. i = 1.1 Exact algorithmic solution Goyal [33] has developed an algorithm for determination of the exact optimum of the objective function (2. 1. Compute Tc (ic ) with the new value of kic . Hence. .1) under the assumption that mini {ki } = 1.. N . the algorithm consists of finitely many steps. Return to step 4. Algorithm 2. Remark 2.3). . i = . kj ) for which the lowest costs ˆ C(kj ) have been found.N Smaller family cycles may be associated with policies with mini {ki } > 1. compute the correˆ 1 ). for the three items are indicated in the columns with headers Ti . . cf.57 is saved on the average weekly cost. . (2.2. for all feasible vectors k it holds that . N .6) recalling the assumption that mini {ki } = 1. As a consequence. When the family reorder cycle T and the reorder frequencies ki . ˜ Step 2: Start with the optimal (rounded) frequencies ki (Tmax ). . . Qi . ˆ The cost function C(kj ) is not in all cases convex in j.22 COORDINATED REPLENISHMENT ˆ Remark 2. The optimal individual policies. the maximum value of this function is at k = 1 = (1. . cf.6) of T = Tmax : Tc (i) = 2ai hi Di 1 . Section 1.. It is based on the observations that the optimal family cycle T ∗ lies on the finite interval [Tmin . . cf. 2. ˆ ˆ T (k) ≤ Tmax = T (1) = 2[A + N i=1 N j=1 aj ] hi D i . the order quantities follow by (1.1 [Goyal] ˆ Step 1: Compute the maximum value Tmax = T (1) and the minimum value Tmin for the family cycle T . with family ordering cost A = $6. . Let Tc = maxi {Tc (i)} be the first boundary that is met when T decreases. hi D i i=1.1 Consider a family consisting of N = 3 items.

1 starts with the computation of Tmax = 110/8. Example 2. . 1. 6) = $26. ic = 3 ˆ ˆ C(1. 3) = 2. 2) is.197. 2) = $26. Item i 1 2 3 ai $3 $3 $43 hi $0.2 An example of a situation where the minimum frequency is larger than 1 in the optimal indirect grouping policy has been provided by Andres & Emmons [2].597 and Tmin = 2/3 ≈ 0.89 [family cycle T (1. C(T. Next. D1 = 400 and D2 = 900. For later reference.1: Data for Example 2. The optimum reorder frequencies are k∗ = (1. 1. ic = 2 ˆ ˆ C(1.466. 1). it is only applicable to moderately sized problems.773.16 23 Q∗ i 18.138] Tc (3) = 43/72 ≈ 0. The algorithm is. 3) = $25.00 $2. a1 = a2 = $50. 2. optimal individual and indirect grouping policies. Tc = 1.580] Tc (3) = 43/30 ≈ 1. 5) = 1. (2. The corresponding average cost per unit of time is C ∗ = C(2. 1. 1) = $508.05 6.053. 4) = $25. N .37 [family cycle T (1. Consider a family of N = 2 items. respectively.50 $2. The successive iterations are: ˆ ˆ C(1. cf. it determines for T = Tmax : k1 = 1 ( 51/495).876. 2. 2)) = 1 2 [ A + 1 a1 + 1 a2 ] + 1 T [3h1 D1 + 2h2 D2 ].225. Tc = 1. .876. 2. 3. .50 Di 9 4 4 A + ai $9 $9 $49 hi Di $4.58 [family cycle T (1. 1.893. 8) = 1. 1 ∗ ∗ ∗ 2 ˜ and Qi . however.5 ≈ 3. Application of Goyal’s Algorithm 2. i = 1.69 [family cycle T (1.HANS BLANC. h1 = h2 = $1. Tc = 1. See Van Eijs [70] for an extension of the algorithm to include strategies with minimum frequency larger than 1. The minimum total average cost per week with coordination ∗ ˆ 1. 1. 1. important for verifying the quality of heuristics. 7) = $27. 5) = $26. is C = C(1. 2.677. the table also contains the values ai = 2 hi Di (ki T ) . 2) = 2. k2 = 1 ( 51/220). 2. 2. 1.1. 8) = $27. Tc (3) = 43/6 ≈ 2. (3. The latter may be the case when the family ordering cost A is small with respect to the individual ordering costs ai . consider coordinated replenishment by indirect grouping.313] Tc (3) = 43/42 ≈ 1. ic = 1 Now. (2.63 ai ˜ $ 9.893. Tc = 1.526] Tc (1) = 2/3 ≈ 0. ∗ The optimum item cycles and the optimum order quantities are listed in Table 2.52 [family cycle T (1.707.771] Tc (3) = 43/20 ≈ 1. 3) = $25. T ∗ = T (2.47 8.21 $37.225.216] Tc (3) = 43/56 ≈ 0. ic = 3 ˆ ˆ C(1. the average cost per unit of time corresponding to the frequencies k = (3. Tc = 0. ic = 3 ˆ ˆ C(1. Goyal’s algorithm gives the following policy: ˆ ˆ k∗ = (2. 1) = 0.466. Goyal’s algorithm requires a computation time. as a consequence. ic = 3 ˆ ˆ C(1.299. Tc = 1.05 2. respectively. 5) = 1.4).437] Tc (2) = 3/6 ≈ 0.91 [family cycle T (1.197.00 Ti 2 3 7 Qi 18 12 28 Ci $9 $6 $14 ∗ ki T ∗ 2.21 24.053] Tc (3) = 43/12 ≈ 1.1 under the headers ki T ∗ . 2.50 $0. which is strongly increasing with the number of items N in the family. 3 2 2 T 3 . 2. 5) = $26.3 The optimal indirect grouping policy is not found by Goyal’s algorithm if the minimum frequency is larger than 1. Tc = 0.66 [family cycle T (1.5% less than without coordination (and without synchronization). ic = 3 ˆ ˆ C(1. The ˆ optimum family cycle is T ∗ = T (1. 4) = 1. 6) = 1.816. TILBURG UNIVERSITY Table 2. Tc (2) = 3/2 ≈ 1. 1.33.07 [family cycle T (1. . k3 = 2 ( 731/220). However. Tc = 2. i = 1.48 $ 4. Let A = $1. that is. 11. cf.012. 1.677.816.1).93 Next.50 $0. Tc = Tc (1) = Tmin so that the iterations stop. ic = 3 ˆ ˆ C(1. i Remark 2. 2. 3) = 2. 1. 7) = 1.012.816. 3). which represent the ordering costs that give Q∗ as optimum individual order quantity.66.

The family ordering cost is A = $1. D2 = 30. . a3 = $12. consider the two items independently of each other and determine for both items the optimal order quantity. . 2) = $503. This cost function is minimal for. cf. (2. that is. 2. i = 1.24 COORDINATED REPLENISHMENT since only at four instants in each cycle of six potential reorder instants an order is placed. The demand for both items is assumed to be deterministic and constant in time.50.5). The one-step heuristic of Goyal & Belton [35] first determines as a reference item the item with the smallest individual cycle length.1. Also. Algorithm 2. The minimum cost of the coordinated replenishment problem (with integer ordering frequencies) will then be larger than (or equal to) the sum of the minimum costs of the optimal individual policies (without integer constraints). the optimal reorder cycle and the minimum average ordering and holding cost per week.2).2 A retailer orders two items from the same importer.00. and h4 = $1. First. ˆ ˆ The corresponding average cost per unit of time is C = C(T (3. Then.168. k2 } = 1. How much can be saved on this cost by synchronizing the reorder instants of the two items in such a way that they occasionally coincide? Next. determine the optimal inventory policy in the class of indirect grouping. .4 Goyal’s algorithm is still applicable if A = 0 while ai > 0.1 What is the increase in cost if the family cycle in Example 2. (2. Assume deterministic demand. The purchasing costs are v1 = $16 per unit of item 1 and v2 = $10 per unit of item 2. (1.1 several heuristics have been developed to find good solutions for the indirect grouping problem. ˜ Step 3: Determine the reorder frequencies of the non-reference items i = ir by ki = ki (T (0) ). cf. respectively. h3 = $2. Exercise 2. The holding costs of both items per unit per week are a carrying charge r = 0.3 Consider a family of four items that are supplied by the same distribution center.1 is rounded to T = 2 weeks? What do the order quantities become? Exercise 2. the frequencies of the other items are chosen as the optimal frequencies given this cycle length. determine the optimal family reorder cycle and the optimal reorder frequencies k1 and k2 with min{k1 . item with the smallest value.2 Heuristic solutions To avoid the long computation times required by Goyal’s Algorithm 2. (2. Finally.2). The holding costs per unit per month are h1 = $1.2 [Goyal & Belton] Step 1: Compute for every item i the value of the quotient (A + ai )/(hi Di ). and D4 = 80. Observe that the family ordering cost is much smaller than the item-dependent ordering less than C(2. This item has by definition a frequency of 1: kir = 1. Step 4: Determine the family cycle length T (1) as the optimal cycle given the vector of frequencies: ˆ T (1) = T (k). ˆ T (3. . cf. Step 2: Determine an initial family cycle length T (0) as the optimal EOQ cycle length of the reference item ir .5). The monthly demands are D1 = 60.4 ≈ 0.98 which is ˆ 1). N . . The reference item ir is the . The family ordering cost is A = $10. The demand for item 1 is 25 units per week and the demand for item 2 is 40 units per week. D3 = 90. ˆ The corresponding cost is C(k). Remark 2. and a4 = $8. 2) = 2[ 2 A + 1 a1 + 1 a2 ] 3 3 2 = 3h1 D1 + 2h2 D2 2 3 · 127 = 3000 1 30 √ 25. A distinction is made between one or two step heuristics and iterative heuristics. h2 = $2. costs in this example. the family cycle length is chosen as the optimal cycle length given these frequencies. Implement Goyal’s algorithm and determine the optimal inventory policy. cf. compute the corresponding minimum average ordering and holding cost per week. Exercise 2. First we will discuss one and two step heuristics.00. with ordering cost A + air . a2 = $12. 2)) = 3000 · T (3. (3. while the supplementary item ordering costs are a1 = $3 and a2 = $8.50. while the supplementary item ordering costs are a1 = $5.005 times the purchasing cost. 2).

5 24. then. The general idea of iterative heuristics is to choose initial reorder frequencies. .053. i = 1. However. repeat from Step 2.8) and determine ki (Ti ) with the aid of (2. Then.50 $0. cf. compute the final family cycle ˆ T = T (k) according to (2. the Algorithm 2.3. The values of (A + ai )/(hi Di ) are displayed in Table 2. Algorithm 2. cf. without regard to item i.2). and.5).2.5).2. hence. cf. . . Table 2. otherwise. and by then modifying the family cycle to T = T Step 2 might be repeated to obtain an iterative heuristic. until the policy does not change in two subsequent iterations. (2.4).2).6 A simulation study of Kaspi & Rosenblatt has revealed that the largest improvement occurs in the first application of Step 2.2 of √ √ 1 ˜ ˜ Goyal & Belton sets k1 = 1 and takes k2 and k3 as the rounded values of k2 (2) = 2 3 and k3 (2) = 1 43. 3) = 2. is the first responsible for the family ordering cost. N. obtains the optimal policy in this case.3 In this example. Such a procedure may end at a local minimum. the result of a one-step heuristic can be used to obtain initial frequencies in Step 1.50 $0.1. the foregoing heuristic algorithms are applied to the family of N = 3 items described in Example 2. cf. the item with the smallest individual cycle length may be an item with a minor contribution to the total family inventory costs.HANS BLANC. The rounded frequencies are listed in Table 2. cf. Clearly. successively perform the following two operations: compute Ti = Ti (k) accord˜ ˇ ing to (2. . Alternatively. . The latter is suggested by (2. .5 ˜ ki (T (0) ) 1 1 3 ˜ ki (T (1) ) 1 1 3 Example 2.4 [Iterative heuristic Goyal] Step 1: Choose as initial vector of frequencies k = 1. we will discuss an iterative heuristic algorithm proposed by Goyal [34].1 (Table 2. Step 3: Stop if the new vector k is the same as the previous one. indirect grouping policies by heuristics. TILBURG UNIVERSITY 25 Remark 2. a second step is added in which the reorder frequencies and the family cycle can be modified independently of the reference item. N . ˇ ˇ Step 2: For i = 1. .5). respectively. and alternatingly determine a family cycle length given the frequencies and new frequencies given the family cycle.4) but it ignores the fact that the reference item. . 2 (2.3 [Kaspi & Rosenblatt] Step 1: Determine an initial policy (T (1) .5 This algorithm is an improvement of an algorithm proposed earlier by Silver [59] which determined the reference item on the basis of the values of the quotient ai /(hi Di ). .0 4. k) with the aid of the heuristic of Goyal & Belton.50 Di 9 4 4 (A + ai )/(hi Di ) 2. . Table 2. Example 2. or may start oscillating. j=i k j hj D j . Next. N . and still it gets an important influence on the reorder cycles of the other items.2: Data for Example 2.8) Algorithm 2. j=i (aj /kj )] N j=1. i = 1. being the item with the smallest individual cycle length. Therefore. 1. Kaspi & Rosenblatt [45] extended the foregoing heuristic because the result of this heuristic strongly depends on the choice of the reference item. (2. Remark 2. the reference item is item ir = 1.1. ˜ Step 2: Determine a better policy by first determining new reorder frequencies as ki = ki (T (1) ) for (2) ˆ(k). Observe that the cost contribution of the reference item is not small in comparison with that of the other items in . (2. Item 1 2 3 ai $3 $3 $43 hi $0.1). . (2. The initial cycle length is T (0) = 2. ˇ Ti (k) = 2[A + N j=1. ˇ The heuristic of Goyal makes use of optimal family cycles Ti (k) given the reorder frequencies k. . This heuristic finishes by taking ˆ T = T (1) = T (1.

11) Optionally. . 2. . 1. 2. here.80 $0. Exercise 2. .26 COORDINATED REPLENISHMENT this case. N .5 [Powers-of-two policy] Step 1: Order and renumber the items according to ascending values of ai /(hi Di ). . T3 (1.4 starts with k = 1 and computes T1 (1. k0 = mini {ki } is not necessarily equal to 1.9) . The joint ordering cost is A = $10. 1) = 4. ki as the smallest power of two that is larger than or equal to 1 R ai . The additional step of Algorithm 2. i0 . = 0. 1. the holding cost hi per unit per month and the constant demand rate Di per month are given in the table below. √ √ 1 ˜ ˆ ˜ ˆ ˇ ˇ k1 ( 26) = 1. . hi Di (2.12) A+ i0 j=1 i0 j=1 aj hj D j . N .4 A retailer purchases six items from the same regional warehouse. Step 5: Determine. . k2 (4) = 1.1. Jackson et al. 1) = 4 3/13. (2. item i 1 2 3 4 5 6 ai $2 $2 $1 $2 $2 $1 hi $0. [42] have developed an efficient algorithm for determining optimal reorder frequencies given the review period R. The objective function to be minimized becomes for this subclass of policies C(R.40 $0.80 $0. The supplementary ordering cost ai . .10) Step 3: Determine k0 as the smallest power of two that is larger than or equal to 1 R Step 4: Set ki = k0 for i = 1. In such a strategy the reorder cycle time of each item is a nonnegative integer power of two times a given basic cycle or review period R. R k0 i=1 ki i=1 N N (2. hi Di (2. k3 (4 3/13) = 3. for i = i0 + 1.20 $0. .3 of Kaspi & Rosenblatt does not change the policy in this √ √ ˆ ˇ case (T (2) = T (1) ). 1. k) = 1 A ai 1 + 2R + ki hi Di . . . . Algorithm 2.}. . Step 2: Determine i0 as the largest index i for which A+ i j=1 i j=1 aj hj D j ≥ ai .40 Di 2000 500 1000 2000 500 2000 Determine reorder policies according to the heuristic of Kaspi & Rosenblatt. . . 1) = 26. . the last three steps can be repeated for various values of R to find a suitable value of R. 1.3 Powers-of-two policies A subclass of the class of indirect grouping policies is formed by the so called powers-of-two policies. . 6. . . k1 ( 26) = 2/39. and according to the iterative heuristic of Goyal. k3 (4 3/13) = 1 559/3 ≈ 4 √ ˜ 11. Algorithm 2. and the frequencies are restricted to ki ∈ {2 . i = 1. which is the optimal vector of frequencies. k2 (4) = 4 3.20 $0. . T2 (1. .65. for items i = 1.

(2. [42] have proved that the inventory cost (2. i = 1.9). if necessary). Jackson et al. C(R. For moderate values of R.1 shows the minimum cost of the power-of-two policies as function of the review period R.3.5 3 3.1 (Table 2. (1. TILBURG UNIVERSITY 27 32 ↑ C 31 30 29 28 27 26 25 0 0. its lower bound and the 6% upper bound. √ The 6% guaranty. Figure 2.13) Based on the rounding to powers of two in (2. It is seen that i0 = 2 is the largest index for which (2.000 1. 8)) = $25.1: Minimum cost of power-of-two policies as function of the review period R. k∗ ) ≤ ( 2 + 1 √ ) 1 Clb 2 2 ≤ 3 4 √ 2 C ∗ ≈ 1. respectively. the cost C(2.13) has in this example the value Clb = 2 39 + 2 43 = $25.88 is guaranteed below this bound. 3.3: Powers-of-two policies for the family of Example 2.12) it can be shown that for the frequencies k∗ determined by Algorithm 2.1. The lefthand sides (LHS) of the inequalities (2. are computed by (2.HANS BLANC. provided that the quantity in (2.06 C ∗ .12) for k3 . 2)) = $29.9) corresponding to the powers-of-two policy generated by their algorithm cannot exceed the minimum cost for the coordinated replenishment problem by more than 6%. 1. 2. the items do not have to be renumbered in this case. if R ≤ [A + i0 j=1 aj ] i0 j=1 hj Dj . The frequencies ki .11): k1 = k2 = k0 . The lower bound (2.5 it holds that √ C(R. For this purpose they showed that a lower bound for the minimum cost over any joint replenishment policy is  . The cost function C(R.10) are also displayed in Table 2.5 2 2. 2. Item 1 2 3 ai $3 $3 $43 hi $0. 4)) = $25.38 is not. cf.5 1 1.359. The cost C(1.50 $0. 1.3. R = 2 R √ and R = 4.500 Index 1 2 3 LHS (2.11) is larger than 1 (note that the review period R can be decreased to satisfy this condition. C ∗ ≥ Clb = 2 A + j=1 i0  aj  i0 N hj D j + j=1 j=i0 +1 2aj hj Dj .4 In this example.667 1.471 ki (R = 1) 2 2 8 ki (R = 2) 1 1 4 ki (R = 4) 1 1 2 Example 2.10) 2.1). for a review period of √ = 1.50 $0.5 4 4. k∗ ) < $27. but the cost C(4.50 Di 9 4 4 ai /(hi Di ) 0.5 R→ 5 Figure 2. the Powers-of-two Algorithm 2. (1.846 6.11) and (2.846 ≈ 1. k). The values of ai /(hi Di ) are displayed in Table 2.10) holds. and by (2.14.88 is still below this bound. holds for R ≤ 1. (2. the cost function follows .5 is applied to the family of N = 3 items described in Example 2. Clearly. (2.14) Table 2.60. (2.500 21. has a minimum for each vector of reorder frequencies.

. 1). corresponding to vectors k which are powers-of-two multiples of the vector (1. • ai denotes the minor set-up cost for item i. . .5 Consider again the six-item coordinated replenishment problem described in Exercise 2. A necessary condition for feasibility of a single machine. 2. and R = 0. . as in Section 1. . • when the (optimal) cycle length T and the frequencies ki have been determined. the best reorder frequencies are (1. 1. if this condition is fulfilled. (2. Determine optimal reorder frequencies in the class of powers-of-two policies for review periods R = 0.9): ˆ T (k) = 2[A/k0 + N i=1 N j=1 (aj /kj )] ki hi Di .2). . i = 1. which accounts for machine opportunity cost during a possible item set-up time ui and for item set-up labor and material cost. whether the production capacity in the busiest cycle is sufficient. .1.69. in all formulas for the coordinated replenishment problem. Lee & Yao [48] present some properties of the cost function (2. (1.15) Exercise 2. i=1 (2. However.6.10). . 2). . the best reorder frequencies are (1. . Suppose that N items are produced on the same machine. N U+ i=1 (ui + τi ) < T. 8k. . check whether the production schedule is feasible. that is. . . . . For 2.24 month. . . there exists a feasible production schedule with all production frequencies ki = 1. . This condition does not take into account any set-up times. 1. Compute the average monthly costs corresponding to these policies. cf. N . i = 1. and item i is produced in every ki th cycle. multi-item lot sizing problem is N Di /pi < 1.2.3 < R < 4. (2. . • the production time for item i.4. . which accounts for machine opportunity cost during a possible family set-up time U and for family set-up labor and material cost. the minimum cost becomes an increasing function.04. 1. schedule items with ki > 1 not simultaneously in the same period (scheduling items with ki > 1 such that cycle lengths for all items are constant is not always trivial or even possible).5. 2). . are attained at values of R which are 1 1 1 2 .4 Production lot-size problems A related problem to the coordinated replenishment problem with external supplier is the production lot-size problem for various items on a single machine. and the maximum inventory level is Ii. respectively. i = 1. replace hi by hi (1 − Di /pi ). For larger values of R. R = 0. (2. .max = (pi − Di )Qi /pi . Graves [37]: • A denotes the major family set-up cost. N .52.28 COORDINATED REPLENISHMENT a repetitive pattern with alternatingly a local minimum at $26. is τi = Qi /pi = ki T Di /pi plus possibly a set-up time ui . N . 1. corresponding to vectors k which are powers-of-two multiples of the vector (1. N . and a large enough production cycle T such that the set-up times ui are negligible with respect to the production times τi . and a global minimum at $25. This lot-sizing problem is solvable with the foregoing algorithms with the following modifications. (2. there is a family production cycle of length T . • optionally.2. 8 . and compare them with the lower bound (2. N . but at a rate pi for item i during the production run. of the value of R for which the minimum cost corresponding to the vector k is attained. . i = 1. cf. 4 .13) and the upper bound (2. . . .17) If this condition is not fulfilled.16) that is. . i = 1. . An obvious improvement is to replace an arbitrarily chosen review period R by the optimal cycle length given the reorder frequencies determined by Algorithm 2. The (scaled) repetitive pattern is due to the fact that the minimum costs corresponding to the vectors 2k.12. .14) on the minimum cost. while for R > 4. Assume that a cyclic policy is applied. the capacity of the machine is insufficient to satisfy all demand.6. cf.9) and a search algorithm for the globally optimal powers-of-two policy in which also the length of the review period R is optimized. 4k. 4). • if the produced items do not become available as a batch at the end of a production run.

Application of Goyal’s Algorithm 2.1. . Period 1 Period 2 Period 3 Period 4 Period 5 Period 6 U U U U U U item 1 item 1 item 1 item 1 item 1 item 1 item 2 item 2 item 2 item 2 item 2 item 2 item 3 item 4 item 3 item 4 item 3 item 4 item 5 item 5 Figure 2. .10 $0. with hi replaced by hi (1 − Di /pi ). this schedule does not satisfy condition (2. Delaying the start of the production run of item 5 in period 4 would require an additional family set-up time U and add a family set-up cost A = $50 per 6T weeks.241 week.16).175 0.5 Consider a family of five items that has to be manufactured on a single machine.79. k3 = k4 = 2.50 $0. Item-related data can be found in Table 2.01 0.5 with a review period of R = 1.79. Exercise 2.7 For the production scheduling problem described in Example 2.248 Qi 124 186 434 620 248 29 Example 2.250 0. k3 = k4 = 2. TILBURG UNIVERSITY Table 2.145. Still an other option is to search for powers-of-two policies which do not exhibit these kind of irregularities.241.75.01 Di /pi 0. and weekly cost C = $243.40 $0. i = 1.2.150 0. Item 1 2 3 4 5 ai $10 $15 $50 $60 $60 hi $0. First note that the fractions Di /pi of machine capacity required by the item i add up to 0.1 or any of the heuristics described in Section 2. A problem with this schedule is that the intervals between the start of production runs of item 5 are not equal (to 3T ).19 week and show that the corresponding cost is less than that with a review period of R = 1.2. This production schedule is displayed in Figure 2. the total production and set-up time is then maximally equal to 1.10 $0. k5 = 4.5 yields. k5 = 3 and T = 1.17) is satisfied.125 0. . because the production times of items 3 and 4 are not equal.186 0. compute the modified production times and lot sizes for item 5 in periods 1 and 4 such that the inventory of this item reaches the level zero at the start of the production run.5.20 Di 100 150 175 250 50 pi 400 1000 1000 2000 1000 ui 0.2).310 0. implying that the inventory of this item does not reach the level zero at the start of the production run in period 4? Exercise 2.434 0.050 ki 1 1 2 2 3 τi 0. yields the production schedule with k1 = k2 = 1.01 0.241 week. so that condition (2.5.241 week. N . But by producing item 3 in odd periods and item 4 in even periods. which are both feasible in a production period of T = 1. while the production schedule can still be made feasible. because the total production and set-up time in a period when all five items are produced is 1.310 0.01 0.4. Algorithm 2.01 0. The family set-up cost is A = $50 and the family set-up time is U = 0. for a review period of R = 1. This schedule is feasible if items 3 and 4 are produced in different periods and if item 5 is produced in periods when item 4 is produced.2 for a cycle of six periods. What would be the actual cost if this schedule is applied with equal lot sizes in periods 1 and 4. the total production and set-up time has a maximum of 1.186 Qi 124 186 434 620 186 ki 1 1 2 2 4 τi 0.083 week in even periods. However.434 0.2.186 0.310 0.4: Production lot-size data and policies for the family of Example 2. which exceeds the length of a production period T = 1.527 week. A better alternative is to shorten the production run of item 5 in period 1 and to prolong the production run of item 5 in period 4.2. with minimum weekly cost C = $241.241 week.6 For the production schedule displayed in Figure 2.HANS BLANC. . An other problem which can be solved with the aid of the algorithms for the coordinated replenishment problem will be discussed in Section 3. . the production schedule with k1 = k2 = 1. apply Algorithm 2.241 week.310 0.05 week.207 week in odd periods and 1. What is the increase in cost with respect to the (not realizable) minimum weekly cost C = $241.65 (see also Figure 2.2: Production schedule for five items consisting of a cycle with six periods of length T = 1.

Verify whether Td (1) j∈J Dj vi0 ≥ V.3 it has been noted that beside joint ordering costs also discounts may induce a company to coordinate replenishments of items that are ordered from the same supplier. i = 1. k) is feasible and such that the discount is obtained.10 $0. if if N j=1 N j=1 Qj vj0 = T Qj vj0 = T N j=1 N j=1 kj Dj vj0 < V. i = 1. . 0 < d < 1. (2. . . Call the .21) (V ) Td Compute the corresponding cost C(Td (V ) (V ) j∈J Dj vi0 = V. kj Dj vj0 ≥ V . There are no item set-up times.2 (heuristic of Goyal & Belton) with hi = rvi0 (1 − d).30 COORDINATED REPLENISHMENT Exercise 2. If this condition is not satisfied. Algorithm 2.25 Di 400 600 50 pi 1000 2000 500 Determine a production schedule with the heuristic of Kaspi & Rosenblatt and verify whether this schedule can be made feasible. kj = 1}. is determined by: (1) to the value Td (V ) which is minimally required for obtaining (2. 3.02 month. k) = vi0 (1 − d). N . N .18).5 All-units discounts In Section 1. i = 1.8 Consider a production scheduling problem for three items and a single machine with family set-up cost $20 and family set-up time 0. k) = vi0 . k). hence.2 do not depend on the discount factor d. This means that the purchasing price per unit of item i is as function of the family reorder cycle T and the reorder frequencies k. the holding cost hi per unit per month. . Qi = ki T Di is the order quantity of item i at a certain reorder instant. k)Di ki + i=1 i=1 vi (T. The holding costs are represented as hi = rvi (T. indirect grouping for coordinated replenishments is considered for the case of joint ordering cost and a special type of discounts known as all-units discounts. the value (without discount) of all orders exceeds the price break point V .20) If this condition is satisfied. . cf. vi (T. Step 2: Increase the family cycle length Td discount. (2. In this section. They make the simplifying assumption that the discount is only awarded if the value of all orders exceeds the price break point V . .40 $0. . they have to be included in the objective function. 2. . for i = 1. k) defined by (2.5 yields a production schedule that can be made feasible. ordering frequencies k and vi (T. . C(T. 2. k) = 1 ai 1 A+ + 2T T ki i=1 N N N rvi (T. (1) resulting policy (Td . i = 1. . . (2.19) with vi (Td . k). are given in the table below. . . . More specifically. k)Di . . Search for a review period with which Algorithm 2. . the crucial reorder instants are those at which only items i with ki = 1 are ordered.1. . The minor set-up cost ai .6 [Silver & Peterson] Step 1: Apply Algorithm 2. vi0 stands for the basic purchasing price of item i (without discount). continue with the next steps. Under this assumption. N . Since the purchasing costs are no longer constant. i = 1. with family cycle T . The heuristic is based on the observation that the reorder frequencies obtained by Algorithm 2. and. Let J = {j. Hence. . k).18) here. the demand rate Di per month and the production rate pi per month. (1) and the current policy (Td . N . we will consider the case of an item-independent discount factor d. . (2. . N . . and V is a price break point for obtaining discount. vi (T. the objective function becomes.19) Silver & Peterson developed an heuristic for indirect grouping of items under the above described discount structure. . k) = vi0 (1 − d). item i 1 2 3 ai $5 $25 $60 hi $0.

. k) < C(T (1) . This means that the discount purchasing price is $240 per item. and T = Td = 3. The objective function to be minimized is. . . Therefore. k) = $4. The decision variables for the direct grouping problem are • M : the number of groups. j = 1.7 If there are several items i with ki > 1. Application of Algorithm 2.21). the policy (V ) proposed by Algorithm 2. .66 per week. The value of the smallest (1) orders is 250 · (9 + 4) · Td = $6.053. . . j = 1. T) = A+ ai + 2 Tj  Tj  j=1 i∈Gj  hi Di  . .077. . . . N . cf.22) The optimal group cycles given the grouping (G1 . k) = $4. . to obtain cycle length T (1) . GM ) are readily determined in the standard way: ˆ Tj (Gj ) = 2[A + i∈Gj i∈Gj ai ] hi Di . 2}. cf. direct grouping In this section we consider the same problem of a family of N items with joint ordering cost structure and constant demand as in Section 2.2 Constant demand. this possibility is not considered in the algorithm. .002 per dollar inventory per week times a purchasing price of $250 for all three items. . GM . i = 1. .64. and the corresponding cost is (V ) C(Td . The algorithm can be extended to situations with several discount factors and corresponding price break points. . TM ).808. The set of items with smallest order frequency is J = {1. .2 with hi = $0. M . j = 1.02 per week. Determine an inventory policy with the aid of Algorithm 2.2 with hi = rvi0 . M. . . .6. . Section 1.095.275.20). k) = vi0 . (2. 1. .1 but for the subclass of direct grouping policies. 3. Reorder instants of different groups may accidentally coincide. (2.19) with vi (T (1) . Hence. i = 1. . The corresponding cost is C(T (1) . it may be possible to divide the reorder instants of these items such that the items in the above defined set J are never ordered alone.1 stem from a carrying charge of r = 0. . . but such opportunities are not taken into account in the cost function. 2. .48. TILBURG UNIVERSITY 31 Step 3: Redo Step 4 of Algorithm 2. . 6. i = 1. N . Then. i∈Gj (2. Compute the corresponding cost C(T (1) . . the N items are divided into M disjunct groups Gj . . Assume that the regional warehouse offers an all-units discount of d = 0. .107.19). M . . the set J can be enlarged in the foregoing algorithm. • Gj : the items in the jth groups. • Tj : the group reorder cycle for items in Gj . and the discount price break point is more easily reached. k). 2. with T = (T1 .05 above the price break point V = $20.3).077. (2.    M  1  1  C(G1 . 3).6 Suppose that the holding costs in Example 2.23) .50. Exercise 2. Suppose that the holding costs stem from a carrying charge of r = $0. 000 provided that the value of all orders exceeds this price break point. 1. . is T (1) = 2. (2.3. . cf. when N is large there are many ways to divide these items. . namely k = (1. .2. cf. The reorder cycles of the various groups are determined independently of each other. . and use the policy of Step 3 otherwise. but a cycle (1) length Td = 2. (2. Remark 2. However. . . Step 4: Use the policy of Step 2 if C(Td (V ) . cf. j = 1. The cycle length (V ) that is minimally required for obtaining discount is Td = 3. 2. k).2 with hi = $0. Example 2. M . 3. suppose that the supplier of these items offers an all-units discount of d = 0. . Hence. The cycle length according to Algorithm 2. j = 1.6 is k = (1.04 if the value of each order exceeds V = $10. . i = 1. . 3).4. . Further.9 Consider the retailer and the six items from Exercise 2.02 per dollar per month times the basic purchasing price v0i of item i.HANS BLANC. and all members of group Gj have the same reorder cycle Tj . . this policy is not feasible. yields the same frequencies as with hi = $0. . In this case. M . . i = 1.50 (Example 2. .000.

2. . . Example 2. Example 2. . while the optimal grouping of the remaining items has already been determined.4). N .49. 10}. The algorithm starts with f (0) = 0.8 The ordering described by the consecutiveness property also appears in the optimal indirect grouping. for instance: G1 = {1. Remark 2. G3 = {5. f (n) = min f (n − ) + 2 A + (2. . 2. • : the number of items in the same group with item n.1 Exact solution by dynamic programming After renumbering the items according to the consecutiveness property. a < i < b. In each step of the recursion.00. Finally. Next. the numbers 1.2. an additional item is considered. . . . Chakravarty & Goyal [17]. G2 = {4}.5.32 COORDINATED REPLENISHMENT The corresponding minimum average inventory cost per unit of time for a given grouping (without accidentally coinciding orders) is:   M ˆ C(G1 . To formulate the dynamic programming recursion define: • f (n): the minimum cost after the grouping of n items. 3}. All possible groupings of this item with previously considered items satisfying the consecutiveness property are evaluated. and in the Powers-of-two Algorithm 2. then item i belongs to Gj for all i.7 A collection of consecutive sets is.40. 6. G4 = {9.4. A reduction of the number of groupings that have to be considered for optimization is possible due to the so called consecutiveness property (cf. .   n n   ai hi D i .00. 7. The initial value is f (0) = 0. the optimal groups G1 . that is. say Gj . it determines whether it is better to take items 1 and 2 apart or together: f (2) = min f (1) + 2[A + a2 ] h2 D2 = $15.1. GM ) = j=1 2 A + i∈Gj ai  i∈Gj hi Di . the computation time is still strongly increasing with N . if item a and item b belong to the same group. using one term of the minimum cost determined in (2. . Chakravarty [14]): • Renumber the items according to ascending value of ai /(hi Di ).24) The number of possible groupings strongly increases with the size N of the family. item 3 is considered:   f (2) + 2[A + a3 ] h3 D3 = $26. the dynamic programming approach is applied to the family of N = 3 items described in Example 2. . (2. Put on a row. cf. As showed in Example 2.10 are counted off. .8 In this example. the optimal grouping can be determined by dynamic programming. cf. . and the recursion reads. • Then.25)  1≤ ≤n  i=n− +1 i=n− +1 In spite of the reduction realized by the consecutiveness property. for n = 1.58. It turns out that it is advantageous to join items 1 and 2. 2[A + a1 + a2 ] [h1 D1 + h2 D2 ] = $12.  2[A + a1 + a2 + a3 ] [h1 D1 + h2 D2 + h3 D3 ] = $30..49. . the items do not have to be renumbered to satisfy the consecutiveness property. . .24). 8}. then considers item 1: f (1) = 2[A + a1 ] h1 D1 = $9. GM form ordered sets. (2. f (3) = min f (1) + 2[A + a2 + a3 ] [h2 D2 + h3 D3 ] = $29. .

with minimum costs per week C ∗ = $26. the heuristic of Bastian finds the optimal direct grouping.8. (2.51. stop.7 [Bastian] Step 1: Renumber the items according to ascending value of ai /(hi Di ). G2 = {2} and G3 = {3}.23) as T1 = 1. no renumbering is required. In this example. Step 3: Determine for each pair of neighboring groups the decrease in cost (saving) when they would be joined. note that for this example the minimum cost with direct grouping is higher than the minimum cost with indirect grouping (Example 2. compute the optimal reorder cycles and the corresponding minimum cost per year. the heuristic of Bastian is applied to the family of N = 3 items described in Example 2. 3600 units for item b. . The heuristic of Bastian is a greedy algorithm: always accept the largest saving. determine the optimal reorder cycles for the established groups with the aid of (2. . First. b and c from the same supplier. The algorithm starts with three groups: G1 = {1}. M − 1.2. it follows that the optimal direct grouping of this family is a division into two groups: G1 = {1.1. a fixed amount of $5 has to be paid. . Since only the saving of joining the first two groups is positive. . item dependent ordering costs are incurred: $30 for item a. 2}. . those two groups are actually joined. the groups are not joined. Further.58 = −$4. G2 = {3}. The demand per year is 4800 units for item a. and $50 for item c.09. otherwise. With every order. The saving of joining the first two groups is 9 + 6 − 12.23).49 and with group cycles found with (2. Since this saving is negative. the algorithm stops if the desired number of groups has been reached.  2 A + i∈Gj  ai  i∈Gj  hi Di + 2 A + i∈Gj+1  ai  i∈Gj+1  hi D i − 2 A + i∈Gj ∪Gj+1  ai  i∈Gj ∪Gj+1 hi D i . The algorithm is readily adapted to situations in which the number of groups is prescribed.11 A company purchases three items a. 2}. Step 5: Finally.2 Heuristic solution Bastian [11] has developed an heuristic algorithm for the direct grouping of items in a family.9 In this example.HANS BLANC.49 = $2.1). determine the optimal direct grouping of these items with the aid of dynamic programming. The holding costs are equal for the three items. decrease M by 1 and repeat from Step 3 if the largest saving is positive.92 and T2 = 7.8. Note that the cost difference in Step 3 only has to be computed in the second and later iterations as far as the group is involved that has been formed in the previous iteration.24). Algorithm 2.11 it is shown that the optimal grouping of items may follow an irregular pattern as function of the family ordering cost A. determine for each item separately the optimal order quantity. $24 for item b. and the reorder cycles are the same as in Example 2. compute for j = 1.40.10 Determine the optimal direct grouping of the family of two items described in Example 2. Example 2.2. Exercise 2.1). TILBURG UNIVERSITY 33 Hence.40 = −$0. Observe that the cycle length of group G1 is smaller than the optimal individual cycle lengths of items 1 and 2 (Table 2. possibly while there still is positive saving or after a few steps with increasing cost. determine a direct grouping of these items with the aid of the heuristic of Bastian. The demand for these items is deterministic and constant in time. then. the optimal reorder cycle and the minimum cost per year. and the new groups are G1 = {1. Next. In the final step the saving of joining these two groups is computed: 12. 2. Finally. that of joining the latter two groups is 6 + 14 − 20. cf. Exercise 2.49 + 14 − 30. G2 = {3}. Further. that is. Step 4: Actually join the two groups with the largest saving. As noted in Example 2. and amount to $20 per unit per year. In Example 2. and 2500 for item c. also. Step 2: Start with M = N groups consisting of 1 item each.

142. G(n. .6. including purchasing cost.50. For items 1 and 2 together. This is left to the reader. is not feasible.061. 1). the optimal individual cycle length is T = 2.1. $4. Then. 1) = $2. With h1 = $0. An important observation is that the consecutiveness property remains valid.961. With direct grouping the assumption is that the discount can be awarded per group.922.77. G(3. the optimal cycle length with h1 = h2 = $0. and the corresponding average cost is C = $1. 1≤ ≤n (2. vi (T ) = vi0 .27) here. and the corresponding average cost is C = $2.34 COORDINATED REPLENISHMENT 2.444. is not feasible. . Next. ) is the minimum cost of the group consisting of the items n − + 1. With h1 = $0. The application of the dynamic programming recursion (2. if this is not feasible then: • the globally optimal cycle length without discount. ). is not feasible.48 is T = 1. With f (n) the minimum cost after grouping of n items. and the corresponding average cost is C = $3.133.22) and (2.48. Example 2. the minimal cost for item 1 is f (1) = G(1. )}. T ) = 1 T n n n A+ i=n− +1 ai + 1T 2 i=n− +1 rvi (T )Di + i=n− +1 vi (T )Di . the quantities G(3. .50.86.006. . and the corresponding average cost is C = $2. It is seen that it is best to join items 1 and 2 in one group. 2) = $3.077.1. . the recursion becomes: for n = 1. the optimal individual cycle length is T = 3. vi (T ) = vi0 (1 − d). • and the price break point cycle length to obtain discount.19) the average cost of the group consisting of the items n − + 1. the optimal individual cycle length is T = 3.259. G(2. and the value of the corresponding order quantity. A discount factor d is awarded to the purchasing price of all items in an order if the value of the order exceeds a certain price break point V . With h2 = $0. Combining the concepts behind (2. .12. With h1 = h2 = $0.1. 2).49. $6. 2) = $3.5. n are with group reorder cycle T : G(n. .50. the optimal cycle length is T = 1.50. N . and the corresponding average cost is C = $970. G(2.171. the second dynamic programming step becomes f (2) = min f (1) + G(2.79. The optimal direct grouping policy can still be determined with the aid of dynamic programming. Hence. (2. and the corresponding average cost is C = $3.5): • the globally optimal cycle length with discount. and the value of the corresponding order quantity. and for items 1 and 2 together to determine G(2. cf.262. When determining G(n. n. The minimal cycle length for obtaining discount is T = V /(v20 D2 ) = 10.48. Dj vj0 ≥ V. 1). N . consider item 2. . Section 2.50. 3) have to be determined in a similar way. Exercise 2. .10 The dynamic programming approach is applied to the family of N = 3 items described in Example 2. if T if T n j=n− +1 n j=n− +1 Dj vj0 < V.171. With h2 = $0. . and the value of the corresponding order quantities. Hence. . 2) and G(3. .133.23).041.2. . The minimal cycle length for obtaining discount is T = V /(v10 D1 ) = 4. f (n) = min {f (n − ) + G(n. 1) = $3.63.12 Compute f (3) for the family of three items considered in Example 2. The minimal cycle length for obtaining discount is T = V /(v10 D1 + v20 D2 ) = 3. $3.062.592. but the recursion becomes more complex than in Section 2. . the optimal individual cycle length is T = 2.50.1 with the discount structure of Example 2.00.3 Direct grouping with discounts Consider the all-units discount structure described in Section 2.50. n follow by (2.373.133. 1) = $970.50. . .10. (2.28) The globally optimal cycle lengths given that the group consists of items n − + 1. again three cases have to be distinguished (cf. For the third step. a similar analysis has to be performed for item 2 to compute G(2. then considers item 1.26) here. . . The algorithm starts with f (0) = 0.00. . . Chakravarty [15]. .26) requires the repeated solution of these subproblems. G(2.63. Hence. First.2. . the purchasing price as function of the cycle length is: for i = 1.

• order all items together. 8 with direct grouping for A = 0. [73]. items 2.2. and items 1 and 3 have split off as a group. 2. 2. the subclass of direct grouping includes both extremal policies. while item 8 has formed a group on its own. Table 2.1 of Goyal.13.41. i = 1. Ch. items 2. In general. N . It has turned out that the most important influence on the performance of the subclasses stems from the number of items in the family. with a reorder cycle of about 0.5 also contains the individually optimal reorder cycles Ti (0). .25 to A = 2. Direct grouping is a only better if A a but then the savings are generally less than 30%.19.17. 4 and 5 have joined a group.3. The optimal direct grouping policy consists of 8 single-item groups for A = 0. with a reorder cycle of about 0. between 30% and 70% with increasing ratio A/¯ ≥ 1. for a family of N = 8 items. 5 and 6 have formed a group.4 Indirect versus direct grouping Indirect grouping and direct grouping are two subclasses of the class of all possible policies for coordinated replenishments.3: Minimum costs of individual.5. 50.14.HANS BLANC. with a = N i=1 ai . simulation means that the parameters of a large number of inventory problems are sampled from some probability distributions (the problems are deterministic). items 1. with a reorder cycle of about 0. these two groups have merged into one group.5. Therefore.11 Figure 2. All other policies are called mixed policies. From A = 71 onwards. of which the data are summarized in Table 2. indirect grouping and direct grouping policies as functions of A. 1 N average individual ordering costs A/¯. 3.4 to A = 34.2. . with a reorder cycle of about 0.15. with a reorder cycle of about 0. 5. . items 7 and 8 have formed another group. and the ratio between the family ordering costs and the . and items 1 and 3 have joined a group.2.1. also item 8 has joined the large group. 4 and 5 have joined one group.5 also contains the optimal reorder cycles Ti (A). Example 2. from A = 2. with a reorder cycle of about 0. 100. 8 for A = 0. There exists no analytic way for determining which policy (from which subclass) is optimal in which situation.18. 4. simulation studies are required to compare the best policies in various subclasses. with a reorder cycle of about 0. with a reorder cycle of about 0. See further Van Eijs et al. 11].1. . .1 to A = 1. 1.3 shows the minimum costs of individually optimal policies without synchronization. from A = 35 to A = 70. TILBURG UNIVERSITY 35 3300 ↑ C 3200 3100 individual 3000 2900 2800 2700 2600 direct grouping 2500 indirect grouping 2400 2300 0 5 10 15 20 A→ 25 Figure 2. as functions of the family ordering cost A. The two extremal policies in this class are: • order all items separately. 2. See ¯ also Silver et al. from A = 2. .7. from A = 1. 10. . [64. in the range from A = 0. The subclass of indirect grouping includes joint ordering of all items. indirect grouping gives higher savings a ¯ with respect to separate ordering. . of the optimal indirect grouping policies according to Algorithm 2. and of the optimal direct grouping policies according to dynamic programming as discussed in Section 2.2 to A = 2. Here. from A = 7. 2. i = 1.8 to A = 7. . item 7 has made a transition to the large group.16. Table 2.

144 0.019 0.134 0.160 Ti (0) 0. H. the only change is the reorder frequency of item 8: k8 = 3.00 $1. . N .115 0.144 0.205 0. . above A = 189.13 Consider the family of eight items described in Example 2.3 Time-varying demand In this section a multi-item generalization of the model of Section 1.134 0.175 0. The demand Di. .310 0. .147 0. .t } .583 Ti (5) 0. . The decision variables are Qi.25). . for i = 1. and ai denotes the supplementary ordering cost for item i.6. k7 = 2 and k8 = 4 in a broad range from A = 0 to about A = 48. in the range from A = 48. . . t = 1.157 0.t−1 = 0).175 0. The inventory levels can be described by the recursive equations. both indirect and direct grouping. . in the range from A = 216. QN. i = 1.580 Ti (2.010 0.205 0.t − Di.t .316 0.188 0.165. This observation implies the following properties of the optimal solution. 2. . . .4) for A = 5.5 with time-varying deterministic demand will be discussed.2.205 0.314 0.t ≥ 0. . .t ) is the vector of order quantities for period t. It is assumed that all initial inventory levels are zero: Ii.175 0. here. Consider a family of N items and a finite planning horizon of H periods. N. and holding costs of hi dollar per unit of item i per period are charged over the excess inventories. . Compare the policies and corresponding costs generated by the heuristic of Kaspi & Rosenblatt (algorithm 2. t = 0. (2. H). . The cost factors are constant over the planning horizon. N.184 0. .175 0.175 0. N.5) 0. the reorder frequency of item 8 has further decreased: k8 = 2.80 $1. N .144 0. .134 0.205 Exercise 2.184 0.50 Di 4000 2000 3600 3000 2500 3200 500 100 ai /(hi Di ) 0.6.205 0. . direct grouping is optimal for A smaller than about 1.3 shows that the optimal indirect grouping policy is worse than the individually optimal policies for A smaller than about 0.t = Ii. .147 0.50 $2. .150 0. . cf.4 where the reorder cycle increases from 0. .4. .144 0.5: Data for Example 2. . .) is a dummy function.36 COORDINATED REPLENISHMENT The optimal indirect grouping policy has reorder frequencies k1 = k2 = k3 = k4 = k5 = k6 = 1.188 0. As in Sections 2. (2. .147 0. For this family of items.143 to about 0.147 0.146 0.194 0. . .114 0. 1.3) and the iterative heuristic of Goyal (algorithm 2. . .573 Ti (2) 0. . t = 1. N. N . . H).015 0. . and δ(. t = 1.1.00 $0. i = 1. strongly increases with increasing value of the family ordering cost A. . . As in the single-item case the minimization of the cost function is facilitated by the observation that the optimal order quantities are zero or equal to the demand of a whole number of future periods.849 Ti (100) 0.157 0. Figure 2.11. . and worse than the optimal direct grouping policy for A smaller than about 1.047 0. N. .11. i = 1.175 0.t−1 + Qi. .4 to about A = 189. item i is only ordered for the beginning of period t if the inventory has reached the level zero (Ii. The objective is to minimize the total cost of ordering and stock-keeping over H periods: H C(Q1 .157 0. . the optimal direct grouping policy is the same as the optimal indirect grouping policy). all reorder frequencies are equal to 1. A denotes the family ordering cost. i 1 2 3 4 5 6 7 8 ai $15 $20 $22 $25 $28 $30 $35 $40 hi $0. . t = 1. t = 1.7 to just over A = 1000.4. Table 2. . H. . optimal reorder cycles for direct grouping policies.t . . The difference between the individually optimal policies and the optimal policies with coordination. The inventory level of item i at the end of period t is denoted by Ii.2. . . .144 0. For i = 1.20 $1.50 $1. .175 0. Qt = (Q1. . . the reorder frequency of item 7 has decreased: k7 = 1. Ii.134 0. .412 0.566 Ti (1) 0.175 0. . Orders are delivered at the beginning of a period.2.144 0. the order quantity of item i for the beginning of period t (i = 1.t (i = 1. . for A > 1000.205 0.010 0.141 0. .144 0.327 0. .40 $0.2.29) These inventory levels must satisfy the requirements Ii. H).t ) + hi Ii.112 0. and otherwise indirect grouping is optimal (for A > 1000.8. .157 0. . . All demand must be fulfilled and no back orders are allowed.306 0.t ) + i=1 {ai δ(Qi.147 0. 2. H. At the beginning of each period the demand for that period is set aside. . .205 0.147 0.157 0.146 0.146 0. .006 0.0 = 0.t for item i in period t is assumed to be known (i = 1.t .600 Ti (10) 0.30) . .011 0.205 0. . (1.412 Ti (50) 0. QH ) = t=1 Aδ( N i=1 N Qi. .

N. . Hence. . (2. • St : the relevant state space at the end of period t − 1 consisting of vectors Jt with 0 ≤ Ji. . .t+ is not included in the order in period t if hi Di. .t = 0 if Ji. the options are to order for 1. 0). t = 1. • Ji. H = 4. Di. K) = A + i=1 ai δ(Ki ) + hi u=1 min{u. and are interrelated. we have K1 = 0. 2 or 3 periods. = min {Ji + Ki }. Define: for t = 1. (J.t−1 also is zero or equal to the demand in a whole number of future periods. N .t is zero or equal to the demand in a whole number of future periods. . the demand Di. the inventory level Ii. 0. • Ki. . • At (Jt ): the relevant action space for the order at the beginning of period t given the inventory levels Jt . . . 2. H. • Ct. it is restricted to the same set {0. . and Ji. }Di. the next reorder instant t + is determined by .t+ > A + ai (then. . The relevant values of these quantities vary with t. the order quantity Qi. . The inventory levels of items 2 and 3 are zero. . Since an item is only ordered if its inventory level is zero.t : the inventory level of item i at the end of period t as number of periods demand for item i.31) here. Instead of the decision vectors Qt and the similarly defined state vectors It we consider integer decision vectors Kt and integer state vectors Jt with components: for i = 1. t + − 1 given that the inventory levels at the end of period t − 1 are J ∈ St and given that the action K ∈ At (Jt ) is taken: N Ji +Ki −1 Ct. .3. .6: State spaces (vectors) for the case N = 3.12 For the case N = 3. Di. define: for t = 1. that is. H. • ft (J): the minimum cost incurred during periods t. Di. K): the cost incurred during periods t. TILBURG UNIVERSITY 37 2. . consider A2 (2. Further restrictions on the above spaces may be possible by property 4.t ≤ H − t + 1 if Ji. . 0. i = 1. if need be only item i. .6.t = 0 for at least one item i (only reorder instants are relevant). .t > 0. . . Di.t .t . . Di. .t+1 . i=1. 3. . Note that S1 = {0}. so they must be ordered. to order in period t + . Di. . .t + Di. it is restricted to the set {0.t+ periods in stock).. H. . .t : the order size for item i for the beginning of period t as number of periods demand for item i.H }.t + · · · + Di. The optimal solution can be determined with the aid of dynamic programming. . S4 : S3 : S2 : 000 000 220 000 220 320 000 100 100 202 100 202 302 010 010 022 010 022 032 001 001 001 300 230 110 110 110 030 203 101 101 101 003 023 011 011 011 310 330 200 200 301 303 020 020 031 033 002 002 130 210 210 103 201 201 013 021 021 120 120 102 102 012 012 S1 : Example 2. is cheaper than keeping Di..t+u . that is. H starting from state J ∈ St at the end of period t − 1. H = 4.t + · · · + Di. and 0 < Ki.t+1 . . N . A2 (2. i = 1. To this end. . 4. . (J. Table 2. . As an example of an action space. . the state spaces are displayed in Table 2.. .N (2.H }. this space consists of vectors Kt with Ki.t + Di. For both items. .1 Exact solution On the basis of the foregoing properties the minimization problem can be reformulated as an integer problem. .t ≤ H − t + 1. No states have been filtered out on the basis of property 4 because that step requires knowledge of the parameters of the model..t = 0. . . 0) consists of 9 possible actions.32) .HANS BLANC. .

3 . Note that for the action K = (3. 2)) = f3 (0. 0). 3)) = f3 (0. 1)) = f2 (0. 1)) = f2 (1. 1) + C1. 1) the next reorder instant is t = 2 ( = 1) so that the holding costs are only charged over a single period.2 ((0. 0) + C2.   f3 (0. the demand D1. 0) + A + a2 + h1 D1. (3.     f2 (0. 0) + A + a1 + a2 + h1 D1. 1) + C1. (0. 1) + C1.1 ((0. (3. 0) + A + a1 + a2 .1 ((2.2 + 2h2 D2. 0) + A + a1 + a2 . Then for t = H = 3 there is no other choice than to order the lacking items.3 . 1) = A + a1 . 2)) = f3 (0. 2)) = f3 (0. Step 3: Apply the backward recursion: for J ∈ St . otherwise.2 ((0. f3 (0. Therefore. 1).3 ((0. 0) + A + a1 . 0) + A + a1 + a2 + h1 D1. 3) the holding costs are charged over all three periods so that the demands Di. H = 3. (2.2 + h2 D2. 0).1 ((0. 0) = min  f3 (1.3 + h2 D2. (2. (1. 0).     f3 (0. depending on the inventory levels:   C2. and only ordering costs are involved: f3 (0. 0) + A + a2 .1 ((0. 1)) = f3 (1.3 . .3 f2 (0.3 .1 ((0. 1)) = f2 (2.1 ((0. 1) + C2. For t = H − 1 = 2 there is some more choice. 0). (J. (3.2 ((0.  f2 (2. 0). 3)) = A + a1 + a2 + h1 D1. 2) + A + a1 + a2 + h2 D2. (2. . 0). (2. 3)) = f2 (0. (1. 0)) = f3 (0. 1) + A + a1 + a2 + h2 D2. 0). Step 2: Set fT +1 (0) = 0 and start with t = H. 0) = A + a2 . C2.2 ((0. (1. (0. lies all periods on stock. 2)) = A + a2 + h1 D1. the optimal ordering scheme has to be interpreted (the integer vectors Kt have to be translated back to order quantities Qt ).3 incurs a holding cost of h1 over period 1. 1) + A + a1 + a2 + h1 D1. stop. 2) + C1. (1. For the action K = (3. f2 (0.3 . f3 (1.3 incur a holding cost of 2hi .3 . 1)) = f3 (0.3 .13 Consider the case N = 2. Silver [60] has developed a generalization of the Wagner-Whitin dynamic programming solution of 1-item problems to the case of a family of N items. 2) the holding costs are charged over two periods.1 ((0. i = 1.2 + h2 D2. 0). 0). 2.3 .3 . ft (J) = K∈At (Jt ) min {ft+ (J + K − 1) + Ct.3 + h2 D2. f2 (1. 0) + C2.2 + 2h2 D2. f3 (1.1 ((0.2 + h2 D2. f2 (2. 0) = min f3 (0. 1) + A + a1 + h2 D2.1 ((0. 1) + A + a1 + a2 + h1 D1. 0) + C2. C2. 1). 0) + C2. f3 (0.     f2 (0. 2)) = A + a1 + a2 + h1 D1. 1)) = f3 (1. (0.2 .3 . f2 (0. 0) + C1.33) Step 4: If t > 1 decrease t by 1 and repeat from Step 3.2 ((0. 0) + C2.38 COORDINATED REPLENISHMENT Note that an item i with Ji + Ki > is not ordered in period t + and the demand for the periods t + . for t = 1 there is only one relevant state but many options:   C1. 1) = min f3 (0. 0) + A + a1 + h1 D1. 0) + C1.2 + 2h1 D1. K)} .3 + h2 D2. 0). 1)) = f3 (0. 1) + C2.2 + 2h1 D1. (1.8 [Silver dynamic programming] Step 1: Determine for each t all relevant states Jt ∈ St (J1 = 0) and all relevant actions Kt ∈ At (Jt ). (1.2 . (1. .    f2 (0. 0).1 ((0.1 ((0.   f3 (0.2 .2 ((2. Finally.3 .1 ((0.1 ((1. 0) + A + a1 + a2 + h1 D1. Algorithm 2. 0) = A + a1 + a2 . 1) + A + a2 + h2 D2. 2)) = f3 (0. 0) = min f3 (1.2 + h1 D1. 2) = min f3 (0. 0). the holding cost over period 2 is included in f2 (2. (2.     f2 (1. (0. 0) + C1. 2)) = f2 (0. 2).3 + h2 D2. 0). 2). f1 (0. . 0).    f3 (0. 0) = min   f3 (0. 0). (2. (2. 0) + A + a1 + a2 + h1 D1. Example 2. 0)) = A + a1 + h1 D1.3 + h2 D2.1 ((1. 1) + A + a1 + a2 + h2 D2.2 . 0).3 . (2. 0) + C2. 0) + C1. 0)) = f3 (0. 0). 1) + C2.3 incurs a holding cost of 2h1 . this means that the demand D1. 0)) = f3 (1. 0). Finally. For the action K = (3.

Step 4: Apply the Silver-Meal heuristic to the items in G2 separately (with ordering costs ai ) with restriction to the H1 reorder instants of G1 . divide the N items into two groups: i ∈ G1 if ki = 1 and i ∈ G2 if ki > 1. 3. Since for all combinations of values of Qi. the N family ordering cost A is divided over the N items according to some weights wi.t ) + hi Ii.t over H = 6 weekly periods.t = 1). .HANS BLANC.. 6. The lower bound depends on the choice of the weights so that various lower bounds could be generated by considering different combinations of weights. N .t ). Example 2. just like the demand Di. i = 1. Step 1 of Silver’s ¯ heuristic is the computation of the average demand over all periods of each item: Di .t ] . ˆ . . . i = 1. The family ordering costs are A = $10. Step 1: Determine the average demand over all periods of each item: Di = 1 H H t=1 Di.4.t ( i=1 wi.7. The idea to derive such a lower bound is to decompose the N -item problem into N 1-item problems.3) for 1-item problems with time varying demand. Step 3: Consider the items in G1 as one aggregated item with .31) should be read as zero if the action Ki only concerns periods of zero demand of item i. h= i∈G1 hi . . TILBURG UNIVERSITY 39 Remark 2. (2.2 Heuristic algorithms The exact solution discussed in the previous section is very time consuming in applications with many items and/or many periods. 2. i∈G1 (2. Algorithm 2. This requires modifications of the algorithms.t .14 Consider a family consisting of N = 3 items that are ordered from the same supplier.t ) + i=1 {ai δ(Qi. i = 1.t t=1 [{Awi. 2. The first heuristic to be discussed is a heuristic by Silver. determine a vector k of order frequencies according to Algorithm 2.t t=1 Aδ( N i=1 N N H Qi. It is a combination of the one-step heuristic for coordinated replenishment with constant demand (e. Therefore.t it holds that δ( N i=1 N Qi.g.3.t ) + hi Ii. cases with zero demand of some items in the first (few) periods requires modification of the ordering cost: δ(Ki ) in (2.30) has a lower bound which is a separable function: H min Qi.t + ai }δ(Qi. 1 Dt = ˆ h hi Di. Step 2 is .t . ˆ . The foregoing dynamic programming solution and also the heuristics that will be discussed in the next section assume that the inventory levels of all items before the beginning of the first period are zero.9 The application of coordinated replenishment with time-varying demand with a rolling horizon is not as straightforward as in the 1-item case. that is. the Goyal & Belton heuristic) and the Silver-Meal heuristic (Algorithm 1. cf.36) and apply the Silver-Meal heuristic to the aggregated item. (2.34) it follows that the minimum of the cost function (2. t = 1. To judge the quality of heuristics a good and simply computable lower bound for the minimum cost is useful. a=A+ ˆ i∈G1 ai . Also.35) The lower bound can be computed by solving N 1-item minimization problems. To this end. . 3. Step 2: Determine an indirect grouping of the items on the basis of these average demands. The supplementary ordering costs ai and the holding cost hi can be found in Table 2. this yields H1 ≤ H reorder instants and determines the order quantities for items i ∈ G1 . and perhaps none at which all items are ordered.9 [Silver’s heuristic] ¯ . The solution of the problem over H periods may contain intermediate reorder instants at which only a few items are ordered.t } ≥ i=1 min Qi.t δ(Qi.t ) ≥ i=1 wi. . Remember that periods with zero demand are skipped by the Silver-Meal heuristic. 2. . Remark 1. . several heuristics have been developed for this problem.2.t and wi.

4) is skipped since demand for period 5 is coupled to period 4. 2) = 2 (24 + 24) = 24 ≤ C(4.t 1 10 60 20 2 0 40 25 3 25 60 20 4 40 50 40 5 30 30 25 6 15 60 20 ¯ Di 20 50 25 A+ai ¯ hi Di 2ai ¯ hi Di (T (0) )2 ki 2 1 1 7. Table 2. .37) Note that αi. The Silver-Meal ˆ 1 ¯ ¯ ¯ heuristic for the aggregated item yields: τ = 1. 1). Hence.2: on the basis of the ratios (A + ai )/(hi Di ) the reference item is determined as ir = 2. G1 = {2. The data for the aggregated item of group ˆ ˆˆ ˆˆ ˆˆ ˆˆ ˆˆ ˆˆ G1 are: a = 24.40 COORDINATED REPLENISHMENT Table 2.38) .5) = 10 ≤ C(1. Let τi denote the last reorder instant of item i up to now. ordering and holding costs. The 2 3 3 ultimate order quantities are given in Table 2. Next. N ) and take u = 2.τi = −ai and that αi. hence.u for all items.7: Demand in units per item per week.60 period t: D1.8: Order quantities according to Algorithm 2. 5) = 5 (5 + 2 · 2. 1 1 ¯ ¯ ¯ ¯ ¯ hence. Actually order all items i ∈ C if αi. reorder frequencies. C(1. 3). the Silver-Meal heuristic indicates that the items in group G1 should be ordered in every period except period 5. C(4. τ = 2. 2). Then. C(1. h = 0.5).13 — 1.10 [Coefficient method] Step 1: Start with τi = 1 for each item (i = 1. It is a generalization of the part-period-balancing heuristic (Algorithm 1. αi. hD1 = 30. C(2. 1) = 24. At period u the heuristic computes coefficients αi. N . Step 2: Compute the coefficients αi. hence.t Q3. . τ = 4. hD4 = 39. Algorithm 2.5 + h3 D3.u is increasing in u. . i∈C (2. .u are negative or small.t − ai . the Silver-Meal heuristic is applied to the item in G2 in such a way ¯ ¯ that period 5 cannot become a reorder instant: τ = 1. 1) is skipped since demand for period 5 ¯ ¯ ¯ 5 has to be ordered in period 4.t D2.u > A.5) = 11 < C(4. 3) = 3 (5 + 2 · 2. Hence.5 = $165. 1). hD3 = 30. C(1.50 0. the method would perform badly if the last period is to become a reorder instant. 3} and G2 = {1}. τ = 3.2 = 0.6 ) + h2 D2. hD5 = 24. hD2 = 27. C(4. C(4. hence.8. C(3. . Item 1 2 3 ai 5 2 12 hi 0. 1) = 24. . C(1.5 + 2D1. the reorder frequencies follow as k2 = 1 and 5 by (2. 3) = 1 (5 + 3 + 2 · 1. The initial family cycle becomes T (0) = 8 .9. 1) = 24.00 ¯ the application of Algorithm 2. Orders are postponed as long as the coefficients αi. hence. . 3 1 ¯ ¯ ¯ C(1.5 + 3 · 4 + 4 · 3) = 34 > C(1. 1) = 24. 1) = 5. 2) = 2 (24 + 27) > C(1. C(3. period t: Q1. Since the compensation against later periods is not possible for the last period H and. Let C = {i. C(4.t 1 35 60 20 2 – 40 25 3 – 60 20 4 85 80 65 5 – – – 6 – 60 20 The next heuristic is called the coefficient method and has been proposed by Lambrechts et al. C(1. N. hD6 = 30. 1 1 ¯ ¯ ¯ ¯ ¯ ¯ C(2. 2) is skipped since D1. so that τ = 6.t D3. [47]. 2) = 2 (24 + 30) > C(2.50. i = 1.u > 0} be the set of items that are candidates for ordering. . . 1). 1).3 + D1. 1 ¯ ¯ ¯ C(1. . C(4. τ = 4. . 2) = 2 (24 + 39) > C(3. 1). period 3 would become a reorder instant instead of period 4 and the total cost would be $168. . 2).30 0.47 3. Note that if the Silver-Meal heuristic would not skip period 2 with zero demand for item 1. i = 1.u as the difference between the cumulative holding cost from the last reorder instant τi up to period u and the ordering cost ai : u αi.4) for individual items.10 0. as k1 = 2 and k3 = 1.80 1.t Q2. C(4. . 3) = 3 (24 + 24 + 2 · 30) > C(4. 2) = 1 (5 + 3) = 4.u = hi t=τi (t − τi )Di.90. (2. an additional step is added for period H to see whether ordering in this period can be avoided. The total cost over the 6 weeks is 5A + 2a1 + 5a2 + 5a3 + h1 (2D1. The principle underlying the heuristic is to compensate for ordering costs in a certain period by not ordering in some later periods.

H to the order in period τi if hi (H − τi )Di.3 + D3. TILBURG UNIVERSITY If this condition is satisfied.5 + h3 (D3. Add the demand Di.15 The coefficient method will be illustrated for the data of Example 2.u can be found in Table 2. i ∈ C.10) step 4 of the algorithm has to be performed.HANS BLANC. 2 and 3 order items 2 and 3 (13 > 10) order item 2 (16 > 10) order items 1.H < ai . It is called the cost-covering heuristic and is closely related to the coefficient method. (2. (2.τi + · · · + Di. The demand D1. together with the corresponding ordering decisions. but close to zero. (2. cf. The total cost over the 6 weeks is 5A + 2a1 + 5a2 + 4a3 + h1 (2D1. the costs are the same. The results of the computations of the coefficients αi.H > 0} and shift the demand Di.5 + 2D1.H > 0. Example 2. cf.3 = 12.u – −5 0 12 −2 1 α2. Although the ordering schedule is different from that determined by Silver’s heuristic.t Q2. cf. Therefore.t Q3.39) Step 3: If u < H then increase u by 1 and repeat from Step 2. All three items are scheduled to be ordered in period H = 6 and period 4 is the previous reorder instant for all of them. otherwise.u−1 . (2. then the order quantities become Qi.10. Such items may trigger a next order in the very near future.6 ) + h2 D2.H < A + ai .τi = Di.6 + 2h3 D3.u – 3 0 48 3 27 decision order items 1. as candidates for joining a family order. is only satisfied for item 1. is not satisfied so that period 6 remains a reorder instant for items 2 and 3. The inequality 2h2 D2. cf.3 + D1. period t: Q1.10. cf. Table 2.14. Step 4: If period H has been indicated as reorder instant by Step 2.H > 0. let C = {i. (2.6 < ai . The modified ordering schedule can be found in the right part of Table 2.41) i∈C i∈C Table 2.u . 2 and 3 (71 > 10) Example 2.t 1 35 60 20 2 – 40 45 3 – 60 – 4 70 80 65 5 – – – 6 15 60 20 1 35 60 20 2 – 40 45 3 – 60 – 4 85 80 65 5 – – – 6 – 60 20 The last heuristic that will be discussed here has been proposed by Joneja [43]. then for all items i with Qi. u 1 2 3 4 5 6 α1.37).7. 2 and 3 (73 > 10) order nothing (10 ≤ 10) order items 1. .u – 10 16 13 7 43 α3. 41 (2. after first phase and after final phase. Table 2.5 ) = $165. The inequality 2hi Di. saving $2.14.40) .10. continue with the next step.40). Only an ordering cost of a3 = 12 has been exchanged against a holding cost of h3 D3.H to the order in period τi if hi (H − τi )Di. Qi.9.6 < A + a2 + a3 .41). It has been built on the alleged weakness of the coefficient method of ignoring items with a negative value of the coefficient αi.6 is shifted to the order in period 4. and τi = u for i ∈ C. • Set τi equal to the last reorder instant of item i before H.10: Order quantities according to Algorithm 2. • For all remaining items with Qi. Since period H = 6 has been indicated as reorder instant by the first phase of the method (see the left part of Table 2.9: Computations of coefficients for Algorithm 2.

u represents the savings of shifting the quantity Di.44) Example 2.12.42) Observe that βi. if item i is already included in the last family order. and this does prevent period 5 of becoming a reorder instant. .11: Computations of coefficients for Algorithm 2. (2. that period 4 becomes a reorder instant. (2.u = hi (τ0 − τi ) t=τ0 Di.5 + h3 D3. This heuristic does not treat period H differently from other periods.u = −ai < 0 if τi = τ0 . The total cost over the 6 weeks is 5A + 4a1 + 5a2 + 5a3 + h1 D1.u for all items. .u and αi.42 COORDINATED REPLENISHMENT the cost-covering heuristic reconsiders items that were not included in a family order.5 + h2 D2. i∈C (2.u > 0.5 8 2 −5 β2. . N ) and take u = 2.τi = Di. .11. This does not prevent. together with the corresponding ordering decisions.u > A.τi + · · · + Di.u−1 . αi.14. .u – −5 0 −1 −2 1 α2. The problems have been solved exactly and with the heuristics. Simulation studies have been performed in order to be able to determine the quality of the foregoing heuristics. Algorithm 2. . The ordering schedule can be found in Table 2. Let C = {i. .11. and τi = u for i ∈ C and also τ0 = u. Step 4: If u < H then increase u by 1 and repeat from Step 2.t − ai . average demand a and variance of demand over the H periods. . Actually order all items i ∈ C if αi. 1≤i≤N i = 1. . At u = 5 item 1 is added to the order of period 4. however. Table 2. cf. u 1 2 3 4 5 6 β1.τ0 + · · · + Di. with varying proportion A/¯. Step 3: Compute the coefficient αi. and τi becomes τ0 . 2 and 3 order items 2 and 3 (13 > 10) order item 2 (16 > 10) order items 2 and 3 (25 > 10) order nothing (10 ≤ 10) order items 1.u for all items. Each problem has been simulated with various realizations of the demand patterns.u – −5 −2. This cost is $2 higher than the costs of the previous two heuristics.16 In this example. The results of the computations of the coefficients βi.7. and computes a second coefficient to determine whether an item should be included as yet to a previous order: u βi. that is.u – −2 −2 −2 −2 −2 β3. . .u over τ0 − τi periods by adding item i to the family order at τ0 . the cost-covering heuristic will be applied to the data of Example 2. stop: τi remains the last reorder instant of item i. Observe that at u = 4 the items 1 and 3 are added to the order of period 3. Step 2: Compute the coefficient βi. N.5 = $167. N . . then the order quantities become Qi. .u – 3 0 12 3 27 decision order items 1. otherwise. A large number of problems have been generated. i = 1.u – −12 −12 24 −12 −12 decision — — no change add items 1 and 3 add item 1 — α1. For all i with βi.u can be found in Table 2. . .11 [Cost-covering heuristic] Step 1: Start with τi = 1 for each item (i = 1. u > τ0 = max τi .u > 0} be the set of items that are candidates for ordering. horizon H.43) If this condition is satisfied. the coefficients βi. 2 and 3 (71 > 10) i ∈ C.u – 10 16 13 7 43 α3. If τi < τ0 . item i is included in the family order in period τ0 . Table 2.

HANS BLANC. . i = 1. There exist alternative approaches. The model assumptions for a family of N items are: • there is an infinite planning horizon. The planned demand Di. i = 1. .t 1 10 60 20 2 – 40 25 3 25 60 20 4 70 80 65 5 – – – 6 15 60 20 43 The coefficient method scores best on the average relative error over all problem instances.11. and with the aid of the cost-covering heuristic. The joint ordering cost is $30. • there are no restrictions on order quantities. especially in case of strongly fluctuating demand and with relatively small A. The heuristic of Silver turns out to perform worst. • the ordering costs (A. who also generalize the Silver-Meal heuristic. • the holding costs (hi . ai .4 Stochastic demand This section is devoted to coordinated replenishment problems with stochastic demand processes. . . problem instances can be constructed for which the costs are an arbitrary factor higher than the minimum cost.t D2. . This may be related to the inflexibility of this heuristic with respect to items that are assigned to the group G1 and that are necessarily always jointly ordered. . i = 1. . see Atkins & Iyogun [5]. The aim is the minimization of the expected average cost per unit of time subject to a service level constraint. i = 1. and on computation time. cf. cf. N . N .1. . . the supplementary ordering costs are $10 for all three part types.t for part type i in week t has been listed in the table below for the three part types and for the next six weeks.t 1 36 20 8 2 36 40 4 3 36 20 10 4 36 40 10 5 36 20 4 6 36 40 8 Determine an inventory policy for the three part types for the next six weeks with the aid of Silver’s heuristic. N ) are constant in time.g. 8% (41%) for the cost-covering heuristic and 12% (300%) for the heuristic of Silver. 2 and 3. $0.40 and $0. with the aid of the coefficient method. multi-pass heuristics which improve initial solutions in later steps have been developed. . Further. .14 A manufacturer orders three part types from the same supplier. Iyogun [41]. These are generally better than single-pass heuristics but require more computation time. • there is a constant lead time L for orders as a whole.10.. The holding costs are $0. a Poisson process for slow movers). Compare the costs associated to the three policies. period t: Q1. TILBURG UNIVERSITY Table 2. The average (maximum) relative errors found are 4% (54%) for the coefficient method.t Q2. Exercise 2.t Q3. Only the maximum relative error is larger than for cost-covering heuristic.12: Order quantities according to Algorithm 2. Joneja [43] has proved that the cost of the solution generated by the cost-covering heuristic is at most three times as high as the minimum cost. . • any stockouts are backlogged (there is no stockout cost but a service level constraint for each item). All demand must be satisfied from stock. . . or there are item-dependent constant lead times Li . respectively. Section 1. week t: D1. Kao [44].t D3. For the coefficient method. respectively. the expected demand per unit of time is E{Di } and the standard deviation of the demand in a unit of time is σ{Di }. . Some policies for coordinated replenishment for a family of items that have been considered in the literature are (see Aksoy & Erenguc [1] and Goyal & Satir [36] for surveys): . • the demand is stochastic with time independent parameters (e. . 2. N ) are constant in time.40 per unit per week for part type 1.

• ci = Si − 1: item i is ordered at each reorder instant unless no demand has occurred (optimal in case ai A and hi Di relatively small). A heuristic for a periodic review policy related to indirect grouping will be discussed in Section 2. s. Sj ). • order all items j with their inventory position at or below their level cj up to their level Sj . . • ci : the join-order level for item i (the “can” order level). The corresponding policy is: • order when the inventory position of an item i falls at or below its level si .4. – policies with grouping and joint review periods Rj per group (with item-dependent order-up-to levels Sij : (Rj . • Si : order-up-to level for item i (this level determines the order quantity). . The extremal values of the can-order level are: • ci = si : item i is individually ordered (with ordering costs A + ai . the auxiliary variables are: for i = 1. The aim is the minimization of the expected average ordering and holding cost per unit of time subject to a service level constraint. S). . with item-dependent reorder points si . with item-dependent reorder points si and order quantities Qi : (R. – policies that order on the basis of the total demand (place a family order when the total demand for all items since the last reorder instant exceeds Q.1 Can-order policies This section is concerned with can-order policies. Pantumsinchai [54]). S) policy. • si : the reorder point for item i (the “must” order point). . with item-dependent reorder points si and order-up-to levels Si : (R. . s. S).1 is devoted to a continuous review “can-order” policy under Poisson demand processes. Eynan & Kropp [26]. Q)). with item-dependent order-up-to levels Si : (Q.4. c. related to indirect grouping). N . or order quantities Qi : (s. cf. – policies with coinciding review instants (with item-dependent reorder frequencies ki and order-upto levels Si : (R. N . optimal in case ai A). Wildeman et al. • Mi : expected number of times per year that item i causes an order (by reaching its level si ). 2. Section 2. . S). . . s. can-order levels ci and order-up-to levels Si : (s. (2. • continuous review policies: – policies that supplement orders with other items (“can-order” policies as introduced by Ballintfy [10]. . The objective function can be formulated as: N min i=1 ¯ [Mi (A + ai ) + Ji ai + hi E{Ii }]. . k. . N . and Liu & Yuan [50] who study a system with correlated demand.44 • periodic review policies: COORDINATED REPLENISHMENT – policies with a joint review period R (with item-dependent order-up-to levels Si : (R. Literature on this subject includes Viswanathan [76] who considers an (R. related to direct grouping).4. These values are ordered as si ≤ ci < Si . S). i = 1. Van Eijs [72] who considers joint ordering and transportation decisions. S).45) here. These are continuous review policies where the reorder decision can be taken after each transaction. Q)). . Every item has three decision variables: for i = 1. . c.2. [78] who apply a global optimization procedure.

There exist various methods for solving the 1-item problems. with demand rate λi units of item i per unit of time. The approximative approach of iterative solution globally consists of the following steps. . i = 1. . These state probabilities satisfy the following . . (2. and can be omitted from the state space. the economic order quantity. For the ease of notation. Let pj denote the stationary probability that the inventory level is j units. We will discuss an approach developed by Silver [58]. i = 1. . the expected total cost per unit of time is of the form: ¯ ¯ C(c. ci and Si found with L = 0 are equally increased until the service constraint is met. . j = i. . Ji and Ii . . S. . are functions of the 3N decision variables si . N . The iterations stop if the decision variables have not changed or if the cost saving is relatively small. N .48) ¯ with M . . Assuming that the demand per customer is 1 unit we have Si = Qi . First the case L = 0 is considered. . For each item separately. and a Poisson process of opportunities for joining an order (with rate µ per unit of time). The state space is finite. The problem with 3N decision variables is too complex to solve as a whole. J and E{I} functions of c and S. a Poisson demand process with a single unit demand per customer (with demand rate λ per unit of time). the state s = 0 is an instantaneous state. . Because both the demand and the opportunities for joining an order occur according to Poisson processes. . we can take si = 0 for i = 1. and no safety stock is needed. the analysis will be restricted to the case of Poisson demand processes. N . (2. To this end. (2. Si ) policies (in case µi = 0. . . N. TILBURG UNIVERSITY 45 µ s c' λ 0 s' λ 1 µ s' λ 2 µ s' λ 3 c s' λ 4 µ s' λ 5 λ s' λ 6 s' λ 7 s' λ 8 S s' j s λ B 10 9 Figure 2. N . . S) = M (A + a) + Ja + hI = M A + (M + J)a + hE{I}. The assumptions of Silver for solving the 1-item problems are a zero lead time (L = 0). Therefore. N. . the opportunity rate µi is estimated on the basis of the last values of Mj . When the lead time L = 0. and a decomposition assumption is introduced: the opportunities for item i to join an order occur according to Poisson process with rate N µi = j=1. Initial values for Mi can be obtained on the basis of individual (si . Hence.46).HANS BLANC. • Ji : expected number of times per year that item i supplements an order (when its inventory is at or below level ci ). . i = 1. . S = 10. . and Mi = E{Di } = Qi hi E{Di } = 2(A + ai ) h i λi .4: Transition diagram for the Markov inventory process of the can-order model. the values of the levels si = 0. the original problem with 3N decision variables is decomposed into N problems with 3 decision variables each. . Figure 2. .46) This is a rough assumption. i = 1. The interaction between these N problems is established via the opportunity rates µi . the inventory level process is a Markov process. Because an order is placed and immediately delivered as soon as the inventory level reaches s = 0. i = 1. j=i Mj . . By this assumption. a zero reorder point (s = 0). ¯ • E{Ii }: expected average inventory level of item i. . i = 1. 2(A + ai ) i = 1. . a 1-item problem is solved and a new value of Mi is determined. ci and Si . a safety stock is determined in such a way that the service constraint is satisfied. ci = si = 0). the item index will be omitted in the following derivation. N . orders are delivered immediately. . Next. because the actual rates vary over time. cf. for each item successively. then. For a single item. the case L > 0 is considered. . N and still satisfy all demand directly from stock. .47) In the iteration steps. . (2. . . . . .4 shows the transition diagram of the inventory process for the case c = 4. . . . ¯ These 3N auxiliary variables Mi . j = 1.

. S − c + ρ 1−ρ 1−ρ (2. c.52) With the aid of the equations (2.54). j = 1.46 COORDINATED REPLENISHMENT balance equations. c. S − 1. λ .49) and (2. . (2.51) Finally. (2.57) here. . and can only be left by the event of a demand (at rate λ): c λpS = λp1 + µ j=1 pj . ρ= j = c + 1. (2.50) it is possible to express all state probabilities in terms of the single probability pS : pj = pS . and because the latter occurs at rate λ when the inventory is in state 1. (2. . j = c + 1. since opportunities for supplementing an order are assumed to occur at rate µ in the states 1. .49) The states between the levels 1 and c can only be reached from states with 1 unit more on stock but can be left by the event of a demand (at rate λ) and by the event of an opportunity to join an order (at rate µ): (λ + µ)pj = λpj+1 . here. the expected number of orders per unit of time supplemented by the item is c J =µ j=1 pj = µρ λ(1 − ρc ) 1 − ρc pS = c . . it follows that the expected number of orders per unit of time caused by the item is M = λp1 = λρc pS = λρc c . . . .55) specify the stationary distribution of the inventory level at an arbitrary instant. j=1 (2. . .53) and (2.56) Similarly. the expected average inventory level is S λ c = λpS . (2.54) The normalization (2. .53): pS = S − c + ρ 1 − ρc 1−ρ .57) yield for the expected total number of orders per unit of time in which the item takes a part: M +J = Finally. pj = ρc−j+1 pS .53) (2. The states between the levels c and S − 1 can only be reached from states with 1 unit more on stock and can only be left by the event of a demand (at rate λ): λpj = λpj+1 .52) determines the last unknown probability pS by substitution of (2. . . Together. 1−ρ S − c + ρ 1−ρ 1−ρ (2. (2. S − c + ρ 1−ρ 1−ρ (2. via an instantaneous visit to state s = 0). S. J and I can now be specified as functions of c and S. . . . . λ+µ −1 j = 1.59) . (2.56) and (2. . the normalization of the probability distribution of the inventory level process implies that S pj = 1. .55) Together. c. it has been used that µρ = λ(1 − ρ) by (2. . Since an order is caused by a particular item when it reaches the level s = 0. (2.58) ¯ E{I} = j=1 jpj = 1 2 (S − c)(S + c + 1) + S−c+ρ ρ 1−ρ [c 1−ρc 1−ρ − ρ 1−ρ ] 1−ρ c . ¯ The quantities M .50) The state S can be reached from the states between the levels 1 and c by the event of an opportunity to join an order (at rate µ) and from the state 1 by the event of a forced order (at rate λ.

Example B. Moreover. Hence. For fixed c it is a rather simple function of S. Once the reorder point s has been determined on the basis of the service measure. m=s (2. Taking the derivative with respect to S and setting the latter equal to zero gives d C(c. . S) gives a function of the single variable c. S0 + s).5–2) of the items’ EOQ value without coordination.2.60) The minimization of this cost function proceeds in two steps. µ(p1 + · · · + pc0 ) + λp1 λpS k = 1.1. For the event that the item supplements an order it holds that Pr{IO = s + k} = µpk µpk = = (1 − ρ)ρc0 −k .63) here. . Let c0 and S0 denote the optimal values of the can-order level and the order-up-to level of an item found for L = 0. S0 + s) policy is related to the stationary distribution of the inventory on hand for a (0.48) leads to an explicit expression of the expected total cost as function of the can-order level c and the order-up-to level S: C(c. S0 ) policy by Pr{I(s. Next. c0 + s.65) Let DL denote the demand during the lead time L.53) and (2. and then the cost function (2. otherwise it is negative. c0 and S0 . Finally. S0 .60) has for fixed ˆ c a boundary extreme at S = c + 1. c0 . Since the differences c − s and S − c do not change during this procedure. One could try a multiple (1.55). 2. (2. . c0 . (2. Note ˆ that the function S(c) can also be evaluated for noninteger values of c. An upper bound is rather arbitrary. S) = ρ λρc A + λa + 1 h(S − c)(S + c + 1) + h 1−ρ [c − ρ 1−ρ ] 2 1−ρ c S − c + ρ 1−ρ 1−ρ c . cf. the determination of the actual service level will be discussed for the two service level constraints described in Section 1. c0 .66) . among others. see Appendix B.HANS BLANC. S) = −C(c.1.59) into (2. the latter distribution is given by (2. k = 1. Further.61) ρ(1 − ρc )(1 + ρc+1 ) 2λ(a + ρc A) 2cρc+1 + − . also the values of M and J do not change. Once the iterations for the case L = 0 have been carried out. c0 + s. By the law of total probability it holds that: .64) for the event that the item causes an order to be placed it holds that Pr{IO = s} = (2. the distribution of the inventory position IO just before an order for the item is placed is related to the distribution of the inventory position I at arbitrary instants by Bayes’ theorem. c0 + s. In both cases.62) The expression of which the square root is taken is in some cases positive and then describes the optimal ˆ order-up-to level S(c) as function of c. 2 dS Solving this quadratic equation in S formally gives 1−ρ ˆ + S(c) = c − ρ 1−ρ c (2. Further. µ(p1 + · · · + pc0 ) + λp1 λpS λp1 λp1 = = ρc 0 . . (2. the final policy for the item is (s. For a cycle service level α. (2. h 1−ρ (1 − ρ)2 (2. Alternatively. TILBURG UNIVERSITY 47 Substitution of (2. consider the probability Π of no stockout in a reorder cycle. . an integer value of c does not give an integer value of S. The lower bound is 0. the observation will be used that the stationary distribution of the inventory position for a (s. S0 ) = k}.56). . The resulting safety stock is equal to s − λL and gives rise to an extra holding cost of h(s − λL). Starting from s0 = 0. in general. c and S are equally increased until the service constraint is met. Next. .58) and (2.10 Application of the Golden-Section method requires the selection of an initial search interval. S0 + s) = s + k} = Pr{I(0. Π = Pr{DL ≤ IO } = c Pr{DL ≤ m | IO = m} Pr{IO = m}. S) pS + h(S + 1 )pS = 0. The determination of the optimal value of c can be carried out by the “Golden-Section” method. both c and S should be rounded afterwards. substitution of the optimal value of S(c) into the cost function C(c. . the Golden-Section method can be applied with rounded values. a safety stock is determined for each item separately for the actual lead time L > 0. substitution of the optimal ˆ can-order level c into the function S(c) gives the optimal order-up-to level S. Remark 2. the values of s. the Golden-Section method applies to real-valued functions.

see also Remark 1.12 [Can-order policy for Poisson demand] Step 1: Take L = 0 and si = 0 for all i. cf. Determine the optimal ci and Si that minimize the cost function C(c.71) with again Pr{IO = s + k} given by (2. ci + (0) si .56).64).48 COORDINATED REPLENISHMENT Due to the lack of memory of the Poisson demand process it follows that DL is independent of IO . . Algorithm 2. the reorder point s can be determined for which the service level constraint Π ≥ α is met. Si + si ) for si = 0. in a similar way as for the cycle service constraint. . The mean order quantity does not depend on the reorder point s. either (2.64) and (2. The estimated cost corresponding to the can-order policy — under the approximative assumption that opportunities for joining an order occur according to a Poisson process — is. and that DL has a Poisson distribution with mean λL.69) The mean accumulated backlog during a cycle is given by E{B} = E{[DL − IO ]+ } − E{[DL − S]+ }. cf. successively evaluate the policies (si . Compute initial values of the mean number of orders caused by item i.71) for fill rate constraints. S) for item i. the probability Π is completely determined by (2. otherwise. 1. the mean backlog at the end of a cycle. c. consider the fraction of demand satisfied from stock (the actual fill rate) in a reorder cycle: Ψ = 1 − E{B}/E{Q}. (2.65) and c0 s+k Π= k=0 Pr{IO = s + k} j=0 (λL)j −λL e . j = i.67) can be used for cycle service constraints. (2. cf. or (2. (2. j = 1.46). . j! (2. One can start with s = 0 and increase s by 1 until the constraint is met. until the constraint is met. j! (2. . Start with i = 1. it follows again due to the independence of DL and IO that c ∞ E{[DL − IO ]+ } = m=s Pr{IO = m} j=m (j − m) (λL)j −λL e = j! c0 ∞ Pr{IO = s + k} k=0 j=s+k (j − s − k) (λL)j −λL e .65). For the first term. Hence. (2.65).3. 2. .71) can be computed in a similar way as in Example 1. . increase i by 1 and repeat from Step 2. that is. cf. The infinite sum in (2. verify whether the solution has converged.3. can be computed in the standard manner. Step 3: If i < N . and continue with Step 4. . Step 4: Consider the actual lead time L > 0.5.68)–(2. cf. For a fill rate constraint with target β. . Mi . the mean backlog at the beginning of a cycle. The new value of Mi follows from (2. (2. . (2. based on deterministic demand. Determine for each item separately the reorder point (0) required for satisfying the service level constraint. together with (2. . N . fix for each item the last found values of (0) (0) ci and Si as ci and Si . S) = i=1 ¯ [Mi A + (Mi + Ji )ai + hi E{Ii }]. Example 1. (2. (2. N C(s. therefore.72) . and E{Q} denotes the mean order quantity which is equal to the expected demand per cycle since there are no lost sales. or search in the neighborhood of s = λL. the reorder point s can be determined for which the fill rate constraint Ψ ≥ β is met. set i = 1 and repeat from Step 2. for i = 1.64) and (2. E{B} denotes the expected backlog that accumulates per cycle. if not. In the last step. Step 2: Estimate the rate µi on the basis of the current values of Mj . .70) The last term.60).47). N . has to be computed only once: c c0 c0 E{Q} = m=s (S − m) Pr{IO = m} = (S0 − k) Pr{IO = s + k} = S0 − (1 − ρ) k=0 k=1 kρc0 −k . and.48).67) Hence. Hence.68) here. (2. . The complete heuristic for determining a can-order policy under the assumptions imposed by Silver [58] is summarized below. otherwise.

12. 2.90 62.480 Ti 3. 3. 3. (2.465 0.837 1 so that the expected time between orders is 0. and si is increased until Ψi ≥ 0. ρ2 = 20/(20 + 0. and the reorder cycle Ti .58 29. Table 2. the safety stocks Vi and the expected times τci = (Si − ci )/λi to reach the can-order level ci from the order-up-to level Si and τsi = (Si − si )/λi to reach the reorder point si from the order-up-to level Si .917.907.540 0.480 = 0.624 0.1 month.17. (2. assuming deterministic demand and individual ordering.16. 2.99. Moreover. The family ordering cost is A = $50.371 0. Ji .35 2. Si ) 28.985 0.970.907 0.57).985 ci 13 26 28 14 26 28 14 26 28 Si 28 40 58 29 41 58 29 41 58 Ci (ci .938 0.426 + 0.54).428. 2. with opportunity rate µ1 = M2 + M3 = 0. The cost function (2.32 58. are also displayed in Table 2. i = 1. the mean reorder cycle E{Ti }.60) for L = 0 is minimized by c1 = 13 and S1 = 28.68 58.144 0. The lead time is L = 0. (2.381 0. are listed in Table 2. After the third round of iterations the procedure stops because the values of ci and Si .428 0.970 0.284.64). the iterative procedure continues. the actual fill rate Ψi is computed according to (2.837 = 1.426 0.12 for Example 2.551 0. The new expected number of orders caused by item 2 is M2 = 0.63 41. 3. In this way.624) = 0. 3. The final values of Mi . cf. cf. with opportunity rate µ3 = M1 + M2 = 0. i = 1. the holding costs hi per unit per month and the average demand per month. TILBURG UNIVERSITY 49 here. Management imposes a target fill rate of β = 0. 2.64 46. The economic order quantities Qi . (2. and the estimated monthly cost Ci attributable to item i.71). item 3 is considered. Next. the mean number of orders per month Mi and Ji caused respectively supplemented by item i. Mi .99 for all three items.917 0.169 0.99 for a lead time of L = 0. together with the cor¯ responding mean order quantities E{Qi }. Then.68)–(2.144 + 0.15. 2.45 Mi 0.14: Iteration steps of Algorithm 2. cf. item 2 is considered. The cost function (2.88 29. the mean accumulate backlog E{Bi }.08 (2. the policies of round 3 are adjusted to a target fill rate of β = 0.665 0.17 Consider a family consisting of N = 3 items. (2.289 0.938 0.468 ρi 0.73) Example 2. i = 1. item 1 2 3 ai $10 $5 $15 hi $1 $1 $1 λi 10 20 30 Qi 34.17. The new expected number of orders caused by item 1 is M1 = 0. The results are summarized in Table 2. and the corresponding number of orders per month. iteration 1 item 1 2 3 1 2 3 1 2 3 µi 0.296 0. after round 3 are the same as after round 2.974 0. E{Di } = λi . This is explained by the observation that offering opportunities for supplementing orders diminishes the number of times an item reaches its reorder point. and the ¯ expected average inventory on hand E{Ii } is given by (omitting the item index i): ¯ E{I} = 1 (E{[S − DL ]+ } + E{[IO − DL ]+ }).63 41.47).1 month. 2 This quantity can be computed in a similar way as the mean backlog. The final values of the parameters of the can-order policy can be found in Table 2. Observe that this value is much smaller than the initial value of 0. The intermediate results are summarized in Table 2. Table 2. 3.56).41 58. indicate that about 44% of all orders are caused by item 3. 2.284 0.65).170 0. The value of the parameter ρ1 is 10/(10 + 0.370 2 3 In the last step of Algorithm 2.71) together with (2.56 Mi 0.60) for L = 0 is minimized by c2 = 26 and S2 = 40. i = 1.144. 3. The supplementary ordering costs ai .144 + 0.195 month. are the appropriate item-related numbers of orders according to (2.HANS BLANC. Mi .40 40. For i = 1.13.289. the expected mean inventory on hand E{Ii }.666 0.973 0.298 0.986 0.907) = 0.13. i = 1.14.284 = 0.13: Data per month and individual deterministic policies for Example 2. 36% by item 2 and 20% by item 1.624 and. with opportunity rate µ2 = M1 + M3 = 0. The iteration starts with item 1.46 2.56) and (2.480 = 0. hence. The expected total number of orders per month is M1 + M2 + M3 = 0. cf.

2.17 with target fill rates.11 0. 3.92 17.9916 Π3 0. From the foregoing discussion it can be concluded that the determination of the values of the decision variables of a can-order policy is complex and approximative.8353 0.5520 0. [28] determine the can-order parameters for the 1-item problem .9693 0.9781 0.9825 0.169 0.50 0. Observe that item 2.90 2. Note that the final probabilities Πi all exceed 0.7680 0. For i = 1.38 0.9844 0. Van Eijs [71] has developed a heuristic for this situation.51 24. 2. Table 2.201 $59.39 0.05 1. s2 = 5 and s3 = 7 are larger than those for the target fill rates above.9275 0. even under strong assumptions of Poisson demand processes and a demand of 1 unit per customer.00 τ si 2.75 0. The intermediate results are summarized in Table 2.9733 0.9916 (disregarding opportunities for joining an order) are included in this table for item i.298 0. and that the model service levels of can-order policies underestimate the real service levels. It should be noted that the fact that can-order policies sometimes perform badly is mostly due to the applied method of analysis.95 0.28 1. In Silver [61] Algorithm 2.98. Note that these quantities are independent of the target fill rates.9451 0.3628 0. with the smallest supplementary ordering cost.6191 0. of fast movers).12 can be adjusted to a cycle service constraint of. is fastest in reaching its reorder point (on the average.17 with cycle service constraint.50 COORDINATED REPLENISHMENT Table 2.9936 Ψ2 0.8789 0. the assumption of Poisson opportunity processes for joining orders is demonstrably bad: the a opportunities occur more frequently towards the end of a reorder cycle.68 0.76 0 2 28 43 33.64) and (2. It has been found that the model costs of can-order policies overestimate the real costs.7337 0. Table 2.370 0. Moreover.8901 0.67) with (2.9668 0.9922 Π2 0.9621 0.. i = 1. α = 0. ¯ si ci Si E{Qi } E{Ii } E{Bi } E{Ti } Mi Ji Ci Vi 1 15 30 23.93 Alternatively.9918 Ψ3 0.6955 0. 3. is fastest in reaching its can-order level.12 for Example 2.16: Final can-order policy for Example 2. si 0 1 2 3 4 5 6 7 Π1 0. Si ) without the Poisson assumption.15 2.9742 0. modifications are needed which do not always seem tractable.12 for Example 2.29 0 3 31 61 52.12 has been extended to can-order policies under compound Poisson demand processes. Federgruen et al. and si is increased until Πi ≥ 0.299 $42.249 $29. For other demand processes (e.98 for a lead time of L = 0.02 0 item 1 2 3 λi L 1 2 3 τci 1. it is then better to choose ci = Si − 1 and to determine (si .15: Final step of Algorithm 2.55 31. Ignall [39] showed for a twoitem case that can-order policies are not necessarily the optimal policy for a continuous-review coordinated replenishment problem. with the smallest individual cycle time Ti . si 0 1 2 3 Ψ1 0.17: Final step of Algorithm 2.1 month.4679 0. say.17 with target fill rates. the probability Πi is computed according to (2.99 in this case. while item 3.44 1. In some circumstances (A/¯ large). and not always to policy itself.75 1.9922 The required reorder points s1 = 3.65).9827 0. simulation is needed for a real performance analysis of a can-order policy (without the assumption of Poisson opportunity processes for joining orders).g. when all items simultaneously start from their order-up-to levels). However. the policies of round 3 of Algorithm 2.17.

The joint ordering cost is $15. . Next. How many units demand of the two items are lost? 2. S1 = 80. .58).10 per unit per day for item 1 and $0. c2 . S) periodic review policy.71). The can-order policy is applied as follows: every evening after closing time the inventories are reviewed. . Suppose that this item is reordered according to a can-order policy with s = 2. N . for this item. c = 5 and S = 10. The shop uses a can-order policy with parameters (s1 . TILBURG UNIVERSITY 51 in the case of a compound Poisson demand process by a policy iteration algorithm for semi-Markov decision problems. • Si : the order-up-to level for item i.69). The realizations of the demand over the next 10 days are given in the table below. .HANS BLANC. cf. S2 = 160. charged over the remaining inventories at the end of each day before delivery of a possible order. . for a coordinated replenishment model with family ordering cost and stationary stochastic demand. with • R = T : the family reorder cycle (the review period). with Poisson demand rates λ1 = 1 and λ2 = 2 items per week.98 for both items. . . 9 units per day for item 1 and 20 units per day for item 2. The lead time is 2 days. i = 1. Compute for both items the probability that no backlog occurs in a cycle with this policy and the mean order quantity. . S2 ) for item 2. suppose that the demand is deterministic. First. 30 units are on stock of item 1 and 100 units of item 2. Determine the can-order policy according to Algorithm 2. Compute the total ordering cost over these 10 days. with joint ordering cost A = $40 and supplementary ordering costs a1 = $10. is equal to λ/(M + J). i = 1. Exercise 2. a2 = $15. compute the total holding cost. . S1 ) for item 1 and with parameters (s2 . .16 Show that the mean order quantity E{Q} of an item. (2.08 per unit per day for item 2. . if one of the items has an inventory level at or below its reorder point. ordered goods are delivered 2 evenings later. Compute the expected number of orders per day caused by this item and the expected number of orders per day supplemented by this item. Assume that this item is a member of a group in which opportunities for joining an order occur according to a Poisson process at a rate of once every 4 days. . and with a lead time of L = 1 week. h2 = $8. Compute the optimal reorder cycle for the case that both items are always jointly ordered. Exercise 2. with holding costs per unit per week h1 = $5. i = 1. E{[DL − IO ]+ }. c2 = 80. an order is placed according to the can-order parameters.15 Consider a single item with a Poisson demand process at a rate of 2 units per day. Derive a recursion for this quantity for a reorder point of s > 0 in terms of that for the reorder point s − 1.12 at a target fill rate of β = 0. The holding cost is $0.2 Periodic review policies Atkins & Iyogun [4] have proposed a heuristic for the determination of the parameters of an (R.17 Give a computable expression for the mean backlog at the end of a cycle.18 Consider a family of N = 2 items. N ). • ki (integer): the reorder frequency for item i. (2. cf. . day i=1 i=2 1 8 20 2 12 25 3 15 20 4 10 15 5 10 25 6 5 40 7 10 25 8 15 20 9 10 20 10 10 15 Simulate the inventory levels of the two items over these 10 days. N . c1 . Also. c1 = 40. Exercise 2. of an item. the additional ordering cost of item 1 is $12 and that of item 2 is $18. Demand which cannot be satisfied from stock is lost. suppose that the demand is stochastic. for the case of a reorder point of s = 0.4. k. Exercise 2.19 A shop sells two items which are ordered from the same wholesale dealer. Explain this relation. At the beginning of day 1. Exercise 2. with negligible lead time L = 0. (2. over these 10 days. cf. The aim is the minimization of the expected average ordering and holding cost per unit of time subject to a prescribed service level (a target fill rate βi for item i. s2 = 40. The manager of the shop has chosen the following parameter values: s1 = 20. Determine the stationary distribution of the inventory level process and the average inventory level. Determine the values of the can-order parameters that realize this reorder cycle in such a way that all demand is satisfied from stock.

. . define the set . . (2. Step 2 is to be skipped when . .75) for this purpose). taking into account the stochastic nature of the demand. . m (by definition of the set I). . . . that is. . In the first phase. . which will be used as review period: R2 = T 2 = 2 A + i∈I ai i∈I hi E{Di } . . cf.74) 1= i=1 wi = 1 2A [hi E{Di }T 2 − 2ai ]. . are determined for the allocation of the family ordering cost A to the N items. . N by rounding Ti /T . . to the nearest integer. safety stocks are determined for each item separately. Step 5: Determine ki for i = m + 1. The items i = m + 1. The resulting safety stocks are Vi = Si − (L + Ri )E{Di }. . cf. i = 1. Step 4: Compute the family cycle R = T for this set I according to (2. . . . In the second phase. N }. . (2.75). . Si ) policy with review period Ri = ki T fixed (from the first phase). . ≤ TN . then I = {1. .74) with wi = 0. m with this family cycle T and with ki = 1 for i = 1. m}. . I = {1. i = 1. N. N do not have a share in the family ordering cost. . . . . Order the items such that T1 ≤ T2 ≤ . increase m by 1 and repeat from Step 2. Tm ) until T1 = · · · = Tm = Tm+1 (use (2. Step 2: Increase w1 .77) The complete heuristic for determining an (R.52 COORDINATED REPLENISHMENT Globally. . . . N .75) 2A . .76) This normalization leads to the optimal reorder cycle for the group of items I. according to (2. and review periods are determined based on the average demand. . . . hi E{Di } i = 1. otherwise. such that the service level constraints for the stochastic demands are met. . k. (2. N . we elaborate the heuristic in more detail. . Let I = {i | ki = 1} be the set of items that are included in every family order. N. . if m has reached the value N . in ascending value of ai /(hi E{Di }). N. Step 6: Determine for every item separately the order-up-to level Si on the basis of a target fill rate βi at a fixed review period Ri = ki T . . . i = 1. Then it follows from the normalization of the weights and the fact that Ti = T for i ∈ I: wi = N 2(wi A + ai ) . i = 1. . . N . .78) m . i∈I (2. . . . i = 1. The heuristic chooses wi = 0 if ki > 1. Determine the weights wi for i = 1. .13 [Atkins & Iyogun] Step 1: Start with wi = 0 for all i. . (2. . .. the stochastic nature of the demand is ignored. Step 3: If i=1 wi < 1 and m < N . . . (2. with the assumption of gamma distributed demand during the lead time plus review period L + Ri . and with ordering cost wi A + ai for item i.23). N . N . S) periodic review policy is summarized below. . for every item separately the order-up-to level Si is determined on the basis of an (Ri . i = 1. . . Next. . More specifically. Algorithm 2. wm (and correspondingly T1 . . . . In the second phase. m = N . . . . (2. e. computed from (2.77).g. . . For a given allocation of the family ordering costs A with weights wi it holds that Ti = ki T = A reformulation of this relation leads to 1 [hi E{Di }Ti2 − 2ai ]. the heuristic consists of the following two phases. and then a family cycle T and reorder frequencies ki are determined on the basis of the expected demand E{Di } of all items and with ordering costs wi A + ai for item i. . . Set m = 1. i = m + 1. in the first phase weights wi . .74).

TILBURG UNIVERSITY The expected average inventory levels are with this policy bounded by: for i = 1.50 63.01. ¯ E{Ii } ≥ 1 2 53 [Si − LE{Di } + Si − (L + Ri )E{Di }] = Si − LE{Di } − 1 2 Ri E{Di } = Vi + 1 2 Ri E{Di }.0048 per dollar per week.50 $3.52 $33. The family ordering cost is A = $20. Here.74).58). Finally.46 index 3 2 4 1 Example 2.69 47.13 form an alternative heuristic for the deterministic indirect grouping problem. and I = {1.45 < 1. The lead time is L = 1 week. w1 = wd = 1 2A 2 [rvd E{Dd }Tb − 2ad ] = 0. No part of A is allotted to the other items.75). 3. These values of Ti and the corresponding order quantities Qi are shown in Table 2. first the reorder cycles Ti are computed with ordering cost ai (that is. N .00 $3. . Then. i = 1. increase w1 and w2 until T1 = T2 = T3 (Td = Tb = Ta = 4. k3 and k4 are computed with the final value of T from (2. the purchasing prices vi and the averages E{Di } and the standard deviations σ{Di } of the demands per week are listed in Table 2.79) Remark 2.0825 from (2. 2} = {d.18 Consider a family consisting of N = 4 items. 4} = {d. N .00 $3. 3. c. now 45% of the family ordering cost A = $20 is allotted to item d. . This heuristic can often be improved by computing the optimal cycle length given the reorder frequencies determined in Steps 4 and 5 after Step 5 (like Step 4 of the Goyal & Belton heuristic 2. cf. the first five steps of Algorithm 2. this means that 89% of A is allotted to item d and 11% of A is allotted to item b.17 1. On the other hand. The family cycle for this set of items I is.HANS BLANC. . b.18. The total expected cost is N E{C} = i=1 wi A + ai A 1 ¯ + hi E{Ii } = + Ri T T N i=1 ai ¯ + hi E{Ii }. (2. Moreover. because they only act as supplements to orders. The supplementary ordering costs ai . w1 = wd = w2 = wb = 1 2A 1 2A 2 [rvd E{Dd }Ta − 2ad ] = 4.18.00 vi $8. ki i=1 N (2. The expected order quantities are E{Qi } = Ri E{Di }.75): w1 = wd = w2 = wb = 1 2A 1 2A [rvd E{Dd }T 2 − 2ad ] = 0. using the average demand as the deterministic demand. . cf.25 37. by (2. the items have been indexed in ascending value of Ti : {1. The final weights w1 and w2 are computed at T = 2. According to Algorithm 2. 2. The carrying charge is r = $0.30 E{Di } 9 40 4 60 σ{Di } 9 25 2 24 Ti 4.13 (the first five steps) can be replaced by any other heuristic for the deterministic indirect grouping problem or by the exact Algorithm 2. 4. The result is m = 2.11.13. and the holding costs are hi = rvi per unit per week. increase w1 until T1 = T2 (Td = Tb = 1.77): T = 2[A + ad + ab ]/[rvd E{Dd } + rvb E{Db }] ≈ 2. and deterministic demand E{Di }.0825. d. The application of the heuristic continues as follows. item a b c d ai $3. .17). . .58 9. Section 2.00 $3. First. b}.11 The first phase of Algorithm 2. 2 [rvb E{Db }Ta − 2ab ] = 0.1. with wi = 0). a. cf. .1. [rvb E{Db }T 2 − 2ab ] = 0. Table 2.74) with w3 = w4 = 0: k3 = ka = Ta /T ≈ 2.53 ≈ 5. . 2. Then. the procedure stops because w1 + w2 > 1. b. (2.89.00 $12.18. see the column with the header “index”.42 0. k4 = kc = Tc /T = 4. Next.89.18: Data per week and optimal individual policies without family ordering cost for Example 2. i = 1.79 Qi 37.2). c}. indicated by a.

s. N 1-item problems have to be solved. Compute the corresponding mean order quantities. The column with the header “Ci ” contains the cost that is fully attributable to item ¯ i. k.44 4.52 56. to compute the above lower bound.0825 weeks. Hence. and that the target fill rates are 90% for all three items. which is slightly better than the expected total weekly cost with a family reorder cycle of T = 2. the order-up-to levels Si are determined in Step 6 at a target fill rate βi = 0.89 8. k. in which a good common review period R is found by a stepwise search and the reorder point si and the order-up-to level Si are determined separately for each item i given the review period R and using only the supplementary ordering cost ai .95 241 56. Atkins & Iyogun [4] conducted experiments with Poisson demand processes and stockout costs. Compute the corresponding mean order quantities.21 Apply the deterministic part of the heuristic of Atkins & Iyogun to the three-item family considered in Example 2. They found that: • the parameters of a (R. . Table 2. k.11 0. i = 1. 2. . It does not include a part of the family ordering cost which amounts to A/T = $9.61 0. The expected total weekly cost is E{C} = $42. • the can-order policy outperforms the periodic review (R.35 27. . k. 4. Determine a reorder policy according to the deterministic part of the heuristic of Atkins & Iyogun. the (R. S) policy with a simply computable lower bound for the minimal cost as a gauge. .1.18 1. Finally.48 84 37. Viswanathan [76] shows that an (R.53 0.19: Summary of results by Algorithm ¯ Ri E{Qi } Si Vi E{Ii } ai /Ri 4. cf.0776 weeks (E{C} = $42. (2.37).80) here. .17 37. c. . S) policy is 12% and the can-order policy 28% above the lower bound (while independent ordering is 64% above that bound). i = 1.76 1.54 COORDINATED REPLENISHMENT The above reorder frequencies are the same as those generated by Goyal’s Algorithm 2. S) policy when A/¯ is small.05 119.10. • in a typical case.34) we have for the unknown minimum cost Cmin : N Cmin ≥ i=1 min Ci .36 10.70 117. k.72 2. N .1.96 ¯ hi [E{Ii } − Vi ] 0. Compare the ordering cost allotted to item i (wi A + ai ) with the substitute ordering cost ai in Table 2. only the reorder cycle obtained by the heuristic of Atkins & Iyogun (2.08 124.4. S) policy. Then.19 for a family reorder cycle of T = 2. σ{D3 } = 1. separately for i = 1. by inequality (2.3.89 ki 2 1 5 1 Atkins & Iyogun [4] compare the heuristic (R.1. and is therefore more suitable in case of “fast movers”.58 item a b c d index 3 2 4 1 wi 0.20 Consider again the six-item coordinated replenishment problem described in Exercise 2. . Exercise 2. N . derived in Atkins & Iyogun [3].18. that the mean demand per week is equal to the demand given in Table 2. (2.23 Determine a periodic review policy according to the heuristic of Atkins & Iyogun for the two-item family considered in Exercise 2. This lower bound.00 0. S) policy by a few percent in many cases.00 0. but otherwise a the (R.79). . is based on the same idea of allotting the family ordering cost A to the N items according to weights wi . Finally. .31 for this policy.44 2. This is due to the sequential approach to fix the reorder cycles first and to determine the order-up-to levels later. σ{D2 } = 2. Example 1. hi Vi 1.98 and with a fixed review period Ri = Ti .30 198 74. .0825) is slightly longer than optimal (2.11 8. and the standard deviations of the weekly demands are σ{D1 } = 3. Ci is the cost associated to a corresponding 1-item problem with ordering costs wi A + ai . Exercise 2. k.29 2. Exercise 2. S) policy is not dependent on the assumption of Poisson demand processes and constant lead times. Compute the order-up-to levels ˜ for the case that the lead time is L = 1 week. note that the heuristic (R. that is. k. What are the corresponding safety stocks? Exercise 2. .44 10. .17. . The results are summarized in Table 2.73 2.22 Determine a periodic review policy according to the heuristic of Atkins & Iyogun for the family of three items considered in Example 2.60. 3 and comment on the differences. . outperforms the (R. .47 0.48 0.41 41. 4.18 Ci 2. S) policy are easier to determine than those of a can-order policy. . S) policy and the can-order (s.55 0.1 for i = 1. i = 1. ai /Ri + hi E{Ii }.75 20. . cf.0776). S) policy outperforms the can-order policy.08 83.65 52 6.

3 discusses algorithms for determining pull and push policies for two-echelon systems with time-varying demand.4 is concerned with a periodic review policy for divergent two-echelon systems with stochastic demand. transport management: this includes the decision of using own transport or putting transport out to contract. there exist many echelon structures. and whether to buy or to rent them.2 is devoted to inventory policies for linear and divergent two-echelon systems with constant demand. the choice of the internal transport and material handling system. Section 3. and the policies concerning the order picking system. the choice of the handling. Finally. Section 3. and the routing of the transport. The decisions in the area of physical distribution can also be classified according to their scope: strategic: this includes the choice of the distribution channels (directly or via distribution centers to retailers or indirectly via wholesale dealers). the choice of the means of transport. warehouse management: this includes determining its function in the distribution channel. the choice of buying or renting warehouse space.Chapter 3 Multi-echelon Inventory Systems This chapter is concerned with the management of flows of goods on parts of their way from producer to buyer.and storage system. tactical: this includes the decisions concerning the number and the locations of the warehouses. 3. the planning of the order picking.1 Multi-echelon structures In practice. the choice of an ordering system. the decision concerning the distribution spread (the percentage of shops in the trade that sell the goods). which consists of rules for determining the ordering quantities and the reorder instants. Even within a company or organization several echelon structures may exist next to each other. a chain of supermarkets may have different echelon structures for durable goods and for perishable goods. the choice concerning the number of warehouses and their locations. both in the physical distribution and between departments in a factory. operational: this includes the decisions concerning when and how much to order or to produce. the choice between a “push” or a “pull” policy. Important issues in such supply chains are the requirements to have the goods of the right quality at the right time at the right location. For instance. Section 3. and the determination of the desired service levels. the choice of the lay-out of the warehouse and the equipment in the warehouse. and choices concerning the means of transportation. a route planning system. The main subsystems of physical distribution are: inventory management: this includes demand forecasting. and the choice of the sizes of the pallets to be used. the choice of a storage policy. Section 3. 55 .1 contains a classification of multi-echelon structures and a discussion of the functions of a distribution center.

and will therefore be relatively expensive.1.). Further. train or ship to a smaller truck or van that can reach a shopping area) may be a function of a DC. while the demand works on from the lowest to the highest echelons. 3. direct transport will consist of relatively low volumes. and in distribution from various factories to a single distribution center.. The latter requires much.) supply chain consisting of three echelons. general networks: they consist of both converging and diverging parts. In multi-echelon inventory systems a distinction is made between pull and push systems. combining of orders for transport.g. Finally. and the inventory management is done centrally. and regular control of the inventories. The lowest echelons are the ones closest to the customers (buyers). storage. a convergent (m. and stocking points at higher echelons have to adjust their inventory management to the uncontrollable demand from the lower echelons.2 Distribution centers This section is devoted to the function and some aspects of distribution centers (DCs). administration of incoming and outgoing goods. converging chains: they occur in factories where parts are assembled. In pull system each stocking point has its own ordering policy. a DC may serve for regrouping and repacking of goods which may arrive at the DC in large bulks and depart from the DC in much smaller quantities. However. Usually. In push system information on the inventory positions of the stocking points at all echelons is kept centrally. reliable and up-to-date information. then direct transport would mean 2.1. cross-docking (transferring goods from one means of transport to another without storage). while indirect transport via a single DC reduces the number of transport lines to 125. grouping of items (per order). from truck. It is customary to speak of high and low echelons. Standard activities in a distribution center include the receipt and the control of goods. Value Added Logistics (VAL) activities . external demand only occurs at the lowest echelons. An important function of a DC is that it restricts the number of transport lines between a number of factories and a number of retailers. Also.1 Classification A classification of echelon structures is the following: linear chains: all units follow the same path or route (this may occur in an environment with only sequential working on the units). 3. the space for storage is generally cheaper in a DC than at the retailers because a DC can be located in an area outside city centers. transferring goods to other means of transport (e.56 MULTI-ECHELON INVENTORY SYSTEMS d c d   ‚ © d     d d   ‚   d ©   ¨ ¨¨ ¨ % ¨ rr r r r j c rr r r j r ¨¨ % ¨ ¨¨   d   d   © d ‚   d   d ©   ‚ d Figure 3. The goods move from the highest to the lowest echelons. For instance. if a group of 25 factories supply a group of 100 retailers. material handling. repacking in smaller quantities. they are part of more complex chains. usually. while external orders are only placed at the highest echelons. Moreover.) and a divergent (r. and in distribution from a single distribution center to various retailers.500 transport lines. diverging chains: they occur in factories where raw materials are cut into various part types and where semi-manufactured items are made into various end products. order picking. A cost saving is possible by better anticipation and coordination.1: A linear (l. Further. problems may arise with the responsibilities of local managers.

HANS BLANC. and they take up less floor space than conveyors. • a good routing system. examples of use are: for bottles (putting in crates). efficiency is an important aspect. and for luggage at airports. As far as possible it is best to work with a unit load (e. what demand patterns? . These are often expensive systems which constitute a significant part of the costs of companies. • a good information system. making to measure. Examples of VAL activities are specific labeling (e. To validate the right of existence of a DC the cost savings by the reduction of the number of transport lines and the lower holding cost have to be weighed against the cost of maintaining the DC and of the extra handling cost in the DC. a single VAL distribution center for the whole of Europe. they are more flexible in case of disturbances. they may be manually or automatically guided. It generally concerns relatively small quantities and requires much labor and/or much capital. combining and simplifying movements in a warehouse. fitting of accessories. Possible objectives in optimizing the use of order picking systems are: • minimization of the total order-picking cost per unit of time. the number of distribution centers in which they are performed will be small. as is the case with a factory in East Asia and retailers in Europe.g. filling. examples of use are: for bottles (washing.g. A rough classification of material handling systems is: conveyors: they are relatively static in routing. Further.g.. installing components or software. Next. they should be connected to an information system so that the locations of all units can be tracked at any time. vehicles: they are still more mobile and have a larger range of action. supplementing manuals (in specific languages). assembling of parts. • the combining of orders (batching). • the zoning of the warehouse (and the orders). dispatching and administrating orders. allow a limited variation in products. and for computer controlled storage. what service levels. in factories (parts for production) and in distribution centers. repairing. the reliability and the safety of these systems are important. mixing of ingredients. according to rules of countries). • maximization of the average number of picked orders or items per unit of time. according to the rules of countries). e. These activities aim at postponing product differentiation and lead to a larger flexibility and lower stocks.. Order picking systems deal with collecting a limited number of units of some set of items for a number of customers. specific repacking (e. quality control. • minimization of the total order-picking time per unit of time.g. closing. for storing in storage carousels. Examples where order picking systems are used are in mail-order companies. Important aspects of order picking systems are: • the lay-out of the warehouse and the division of goods over the warehouse. Ways to arrive at efficiency are by eliminating. Value Added Logistics activities can be performed in a distribution center owned by the producer but can also be put out to contract to logistic service industries (this is called public warehousing). A systematic planning and design of order picking systems includes: • a strategic planning: what type of customers. testing. defined as a quantity which can be stacked on a pallet). TILBURG UNIVERSITY 57 may be performed in a DC. examining of returns. Therefore. Moreover.. reconditioning of clothes. Since VAL activities may require costly equipment. they are still more flexible than cranes.. labelling). cranes: they are mobile in more directions. VAL activities are most profitable if there is a large distance between a factory and its market. we discuss in more detail the material handling systems and the order picking systems in a warehouse. • a redivision of goods over the warehouse if some demand patterns change. • minimization of the average lead time to the customers. examples of use are: order picking in warehouses.

Figure 3. [64. • an ABC analysis: important for the locations in a warehouse. Hausman et al. Robeson & House [56] discuss distribution issues. the order-up-to level has converged to S = 110 at all echelons. In this stationary situation.2: A linear supply chain consisting of four echelons. In general. 32]. The order quantity is the difference between the order-up-to level and the inventory position: Q = S − I. The order-up-to level is set at 110% of the demand of the previous week (S = 1. pilferage. Eastman [23] is concerned with material handling systems. 3. cf. • the choice of an information system. Q = 100. • the determination of the required storage capacity.1 Linear chains We start this section with an example of the oscillation effect that may occur in supply chains when not enough care is taken in choosing reorder policies and in exchanging information between echelons. Wilson [79] considers the location of stock in a warehouse as function of the order quantity and the demand of the various products.2 Constant demand In this section. • the determination of a storage policy (what items are stored where in the warehouse? how to divide the warehouse in to zones?). the demand pattern in higher echelons is not continuous but stepwise! This subject is also discussed in Silver et al. with V = 10. 3. Tomkins & White [66] is devoted to facilities planning.2.1 Consider a linear supply chain consisting of a factory. a wholesale dealer and a retailer.58 MULTI-ECHELON INVENTORY SYSTEMS factory E importer E wholesale dealer E retailer E Figure 3. Now suppose that the demand at the retailer has increased to D = 110 in a certain week due to a special offer. 12]. • a classification items: poisonous.1 · D) so that a safety stock is used of 10% of the demand of the previous week (V = 0.1 · D). • the determination of the degree of automatization (this concerns the actual order picking as well as the information system).2. Example 3. [38] derive an optimal storage assignment for automatic warehousing systems. Ratliff & Rosenthal [55] solve a special case of an order-picking problem as a travelling salesman problem. Every week an order is placed that is immediately delivered provided that there are no stockouts at the higher echelon. • the determination of a batching and routing policy. Other studies on order-picking systems include Goetschalckx & Ratliff [31. Hence. inflammable. an importer. Ch. Suppose that at all echelons the following simple ordering rule is applied. • the choice of a material handling system. models and policies will be discussed for multi-echelon systems with constant demand at the stocking points of the lowest echelon. special sizes. This has the following effect on the order quantities: Retailer: Wholesale: Importer: Factory: D D D D = 110 = 121 = 144 = 192 S S S S = 121 = 133 = 158 = 211 V V V V = 11 = 12 = 14 = 19 I I I I = 110 − 110 = 0 = 110 − 121 = −11 = 110 − 144 = −34 = 110 − 192 = −82 Q = 121 − 0 = 121 Q = 133 + 11 = 144 Q = 158 + 34 = 192 Q = 211 + 82 = 293 . the demand in higher echelons is formed by the orders of lower echelons. Assume that the demand at the retailer has been D = 100 during the last several weeks.

echelon inventory position: the net echelon inventory plus the number of units on order from a higher echelon but not yet delivered. this quantity is related to the service level constraint or the cost of stockouts or losses. by removing echelons from the supply chain if feasible. echelon safety stock: the average echelon inventory position just before a reorder instant minus the average demand during the lead time to the stocking point. The next weeks. The effect of a demand of D = 100 in the next week is: Retailer: Wholesale: Importer: Factory: D D D D = 100 = 100 = 112 = 131 S S S S = 110 = 110 = 123 = 144 V V V V = 10 = 10 = 11 = 13 I I I I = 21 + 89 − 100 = 10 = 44 + 54 − 100 = −2 = 104 + 0 − 112 = −8 = 211 + 0 − 131 = 80 Q = 110 − 10 = 100 Q = 110 + 2 = 112 Q = 123 + 8 = 131 Q = 144 − 80 = 64 The wholesale dealer is back at S = 110. The echelon stock in a stocking point is defined as the number of units that are present at the stocking point itself plus those that have already passed the stocking point but have not yet been committed to outside customers. So. The oscillation effect can be limited by better exchange of information between echelons. 193 units have been produced above the usual quantity of 100. but the higher echelons exhibit an upward movement. this quantity may be required for determining a reorder instant.3. 44%. and by limiting the (necessity of) safety stocks. advertisements. due to a 10% increase in demand! Next. At the factory. This consideration led Clark & Scarf [18] to the introduction of the concept of echelon inventory. suppose that in the weeks after the special offer the demand is back at D = 100.g. but that the higher echelons show a counterreaction with small or zero order quantities. this quantity plays a role in determining holding costs. by better quality control. Possible causes of the oscillation effect may be special offers. a distinction is made in multi-echelon inventory systems between pull and push policies. As with the usual inventory concept we can distinguish for the echelon stock at a stocking point: echelon inventory on hand: the number of units present at the stocking point or at lower echelons. the oscillations fade away: Retailer: Wholesale: Importer: Factory: D D D D = 100 = 100 = 100 = 87 S S S S = 110 = 110 = 110 = 96 V V V V = 10 = 10 = 10 =9 I I I I = 10 + 100 − 100 = 10 = −2 + 112 − 100 = 10 = −8 + 131 − 100 = 23 = 80 + 64 − 87 = 57 Q = 110 − 10 = 100 Q = 110 − 10 = 100 Q = 110 − 23 = 87 Q = 96 − 57 = 39 After two more weeks. net echelon inventory (stock): the echelon inventory on hand minus the amount of backlog at the lowest echelon. In the factory. the echelon stock includes the inventory on hand at lower echelons and the inventory in transit to lower echelons. or in transit to lower echelons. A more refined classification is: • pure pull policies: . 92% and 193%. perhaps with working overtime. the steady state will be reached for the whole supply chain. the factory has had high production cost. This has the following effect on the order quantities: Retailer: Wholesale: Importer: Factory: D D D D = 100 = 89 = 54 =0 S S S S = 110 = 98 = 59 =0 V V V V = 10 =9 =5 =0 I I I I = 0 + 121 − 100 = 21 = −11 + 144 − 89 = 44 = −34 + 192 − 54 = 104 = −82 + 293 = 211 Q = 110 − 21 = 89 Q = 98 − 44 = 54 Q = 0 (59 − 104 < 0) Q=0 Observe that the retailer is back in the stationary situation with the order-up-to level. As mentioned in Section 1. this upward movement is quenched by the fact that negative order quantities are not feasible. and now ends up with a high inventory level. an announcement of a price change..HANS BLANC. but also stochastic fluctuations and holidays. Hence. For (linear) chains it is important that ordering decisions in a stocking point consider beside the inventory position at the stocking point itself at least the inventory positions in all lower stocking points and the demand in the lowest stocking point. e. by exponential smoothing in forecasting future demand to avoid overreaction on small changes in the demand pattern. while only 82 are forwarded to the importer as backlog of the previous week. respectively. TILBURG UNIVERSITY 59 Note that the order quantities have increased by 21%.

For comparison. the 3 policy with Q1 = 2 Q2 cannot be optimal. and. respectively. consider alternative 1: order always Q1 = Q2 . if transport charges are included in the internal ordering cost distances and volumes may also play a role. consider alternative 2: order always Q1 = 2Q2 . • h1 . we will argue that it is optimal to have Q1 = nQ2 for some integer n. • L1 . n ≥ 1. the lead times are deterministic. Hence. The cost at the higher echelon in 3 periods of the 1 lower echelon is then: 2a1 + h1 Q2 . Then. • a1 . – lead to limited variability. a2 : the ordering cost for the higher/lower echelon. We will show with the aid of an example that a policy with Q1 /Q2 not equal to a positive integer can be improved upon. Further. the cost at the higher echelon in 6 periods of the lower echelon is 4a1 + 2h1 Q2 . L2 : the lead time to the higher/lower echelon. The parameters of the model are: • D: the demand per unit of time at the lower echelon. h2 : the holding cost per unit per unit of time at the higher/lower echelon. respectively. so that the cost incurred at the lower echelon is fixed. Example 3. – lead to increasing variability at higher echelons. apply central inventory management. The decision variables are: • Q1 . respectively. Next. hence. However. consisting of two stocking points. respectively. v2 the purchasing cost of a unit upon arrival at the higher/lower echelon. As a preliminary property. require central safety stocks.2 Suppose that a policy with Q1 = 3 Q2 is considered for a two-echelon linear chain. First. • pull policies based on echelon stock: – use information from lower echelons. Keep Q2 fixed. and r the carrying charge. Q1 = nQ2 with either n = 1 or n = 2 is cheaper than alternatingly ordering Q1 = Q2 and Q1 = 2Q2 . 3 The cost at the higher echelon in 3 periods of the lower echelon is for Q1 = 3 Q2 : 2a1 + 2 h1 Q2 . h1 = rv1 and h2 = rv2 with v1 . it can be shown that the optimal policy is of the form Q1 = nQ2 .60 – use only local information. Q2 : the order quantity at the higher/lower echelon. But the policy of alternative 0 can be further improved upon. In this way. The cost at the higher echelon in 6 periods of the lower echelon is 6a1 + 0h1 Q2 . MULTI-ECHELON INVENTORY SYSTEMS A push policy takes inventories and costs of all echelons into account. This is better than alternative 0 if a1 < h1 Q2 . the 2 1 quantity 2 Q2 lies on stock without use. in some cases. Hence. with n a positive integer. may apply reshuffling of stocks on echelon level. . we will consider a linear chain consisting of two echelons. s2 : the reorder point at the higher/lower echelon. we may have a1 > a2 when external ordering is more expensive than internal ordering. This is better than alternative 0 if a1 > h1 Q2 . Consider first 2 alternative 0: order alternatingly Q1 = Q2 and Q1 = 2Q2 . the case that the lead time from the higher to the lower echelon is negligible (L2 = 0) will be considered. In most situations. The demand at the lower echelon is assumed to be deterministic and continuous in time. • push policies: – – – – requires central information. respectively. it will hold that h2 > h1 because storage is usual more expensive at the lower echelons and because the value of the goods will increase (v2 > v1 ) when they move down the echelons if only since more shipping cost has been invested. The cost at the higher echelon in 6 periods of the lower echelon is 3a1 + 3h1 Q2 . Also. • s1 . respectively. Finally. – require high safety stocks. Next. that is a saving of 2 h2 Q2 with respect to the original policy. – require limited safety stocks.

(3.1 An alternative derivation and interpretation of the cost function follows by noting that the . The n middle graph of Figure 3.1) This implies that the average cost per unit of time with a push policy is: C(Q1 . behaves like a step function: when an order of Q1 units is delivered. So. Hence. Afterwards. Q2 ) = D or. so that during the second cycle at the lower echelon the inventory level at the higher echelon is n−2 Q1 = (n − 2)Q2 . Q2 ) = a1 a2 + + Q1 Q2 1 2 [h1 (Q1 − Q2 ) + h2 Q2 ]. with Q1 = nQ2 . the total holding cost becomes 2 h1 Q1 + 1 (h2 − h1 )Q2 which is the same as the last term 2 in (3.3) Remark 3. . Associating a holding cost h1 to this inventory and noting that the echelon inventory at the higher echelon includes the inventory at the lower echelon. an additional holding cost h2 − h1 has to be associated to the average inventory level of 1 Q2 at the 2 1 lower echelon. 2 1 however. C(n. (3.3: Inventory on hand at lower echelon (top). This continues until the end of the cycle at the higher echelon. the lower 1 graph in Figure 3. etcetera.3. echelon inventory level I(t) = I1 (t) + I2 (t) at the higher echelon follows a saw-tooth pattern. Next. cf. The inventory level at the higher echelon.HANS BLANC.3 shows the inventory on hand I1 (t) at the higher level for the case n = 4. at higher echelon (middle) and in system (bottom). a second fraction n of Q1 is forwarded n to the lower echelon.2) and (3. the average cost per unit of time with a push policy with Q1 = nQ2 will be derived. 2 n n (3. TILBURG UNIVERSITY Q2 T ↑ I2 (t) 61 rr r r r r r r r r r r rr rrr rrr rrr rrr rrr rrr rrr rrr r s rs rs rs rs rs rs rs rs rs E r 0 T2 2T2 3T2 4T2 5T2 6T2 7T2 8T2 t→ 9T2 Q1 T ↑ I1 (t) 3Q2 2Q2 Q2 s 0 ↑ 4Q2 T I(t) 3Q2 2Q2 Q2 s T1 s 2T1 t→ E rr rr rr rr rr rr rr rr rr r rr rr rr rr rr rr r s r 0 r rr s T1 r rr r rr s 2T1 t→ E Figure 3. so that the average stock at the lower echelon is found to be 1 Q2 .2) D a1 + a2 + Q2 n 1 2 Q2 [nh1 + h2 − h1 ]. At the lower echelon the inventory level I2 (t) follows a common saw-tooth pattern. a fraction n is immediately forwarded to the lower echelon.3. cf.3). the average inventory level at the higher echelon is: 1 n n−1 j=0 j Q1 = n 1 2 n(n − 1) Q1 1 = 2 (n − 1)Q2 = 1 (Q1 − Q2 ). so that the average echelon inventory level is 2 Q1 . the upper graph in Figure 3. During the first cycle at the lower echelon the inventory level at the higher 1 echelon is constant and equal to n−1 Q1 = (n − 1)Q2 .

h2 = $0.7) it follows that Q∗ = 20 10 ≈ 63 with C2 = 10 10 ≈ 2 √ 31.2).7) The higher echelon must follow with Q1 = nQ∗ for some n ≥ 1. and n∗ . Q∗ = 5Q∗ (5) = 410. The optimal real-valued n for given Q∗ is: 2 n∗ = pull with corresponding cost ∗ C1 = 1 Q∗ 2 2a1 D = h1 a1 h2 . and Q∗ (5) = 100 6 ≈ 82.47 and C(5) = 500(4 + 1)(5 + 1) = 50 6 ≈ $122. h2 ∗ C2 = 2a2 h2 D.8) are C1 (6) = 52.6) it follows that n∗ 20 so that n∗ push = push = 4 or npush = 5. ˆ Since there is no difference. For both policies it is noted that if n∗ = 1. From (3. 2 (3.4) with corresponding minimum average cost per unit of time ˆ C(n) = 1 2( n a1 + a2 )(nh1 + h2 − h1 )D. First. (1.1). respectively.38 = 92.50. a2 h1 (3. n∗ = 6 is optimal pull and Q∗ = 6Q∗ = 378. (3. with parameters a1 = $20. cf.29 and C1 (7) = 45. Q1 = 4Q2 2 1 2 3 √ √ ∗ Next.62 + 92. the higher echelon has no stocking function and only serves for cross-docking. The corresponding costs according to √ √ ˆ ˆ (3.62 The optimal Q2 at a fixed value of n is: ˆ Q2 (n) = MULTI-ECHELON INVENTORY SYSTEMS 1 2( n a1 + a2 )D . √ ∗ From (3. (1. so that h1 = $0.9) leads to n∗ = 40 so that n∗ = 6 or n∗ = 7. the optimal real-valued n minimizing the foregoing cost function is: n∗ push = a1 (v2 − v1 ) = a 2 v1 a1 (h2 − h1 ) . (3.3 Consider a linear chain consisting of two echelons.8) Note that applying the EOQ formula to higher echelons does not work due to the fact that demand does not occur continuously in time. with additional cost 2 C1 (n) = a1 D + Q1 1 2 h1 (Q1 − Q∗ ) = a1 2 D + nQ∗ 2 1 2 h1 (n − 1)Q∗ .60. the optimal real-valued minimum of this cost function is obtained by solving the equation d dn 1 d 1 ˆ C 2 (n) = 2D dn { n a1 (h2 − h1 ) + nh1 a2 + a2 (h2 − h1 ) + a1 h1 } = 2D{− n2 a1 (h2 − h1 ) + h1 a2 } = 0. are independent push pull of the demand rate D.25. a 2 h1 (3.5) are C(4) = 500(5 + 1)(4 + 1) = 50 6 ≈ $122. ˆ ˆ 40 6 ≈ 98. consider a pull policy for the same situation. Hence.47. cf. (3.25.9) 2a1 h1 D − 1 2 h1 Q∗ = 2 2a1 h1 D − 1 2 h1 h2 2a2 h2 D.91 + 39. (3. (3. both values are optimal. nh1 + h2 − h1 (3.29 = $123. cf. The total cost with the optimal pull policy is C = 31. the optimal pull policy is obtained by rounding n∗ to a positive integer in such a way that the cost pull (3. (3.5) is minimal.62. (3. Q∗ = 2 2a2 D . The corresponding optimal order quantities are Q∗ (4) = 2 √ √ ∗ ˆ ∗ (4) = 392. we determine the optimal push policy.6). Further. r = $0.91 which is 1 2 not much worse than the total cost with the optimal push policy. Example 3.6). D = 1000. The difference between these ratios is largest when h2 − h1 is small in comparison to h2 while a1 /a2 is large. The corresponding costs according pull pull pull to (3.10) Again. .35 + 47.25 = 92. we determine the optimal pull policy. Observe that both ratios n∗ . The lower echelon determines Q∗ on the basis of 2 the local cost as the economic order quantity.8) is minimal.6) The optimal push policy is obtained by rounding n∗ push to a positive integer in such a way that the cost (3. Alternatively. a2 = $1.9). v1 = $1. Hence. v2 = $2.5) Since the square root function is monotonously increasing. cf.

11) Moreover.2. Observe that the echelon stock I(t) is replenished when it reaches the level s2 due to the time lag L2 before the goods can be used to satisfy external demand at the lower echelon. in transit. s2 = D · L 2 . This implies that orders must be placed at the higher echelon when the inventory position at the lower echelon is at the level.4: Inventory on hand at lower echelon. (3. Section 1. this means that orders must arrive at the higher echelon at instants when the inventory position at the lower echelon is at the level s2 so that part of the arriving order can immediately be transferred to the lower echelon. By the usual arguments. (1. the consequences of a nonnegligible lead time L2 > 0 between the higher and the lower echelon will be studied. the step pattern of the inventory It (t) in transit between the echelons and of the inventory level I1 (t) at the higher echelon.HANS BLANC. they do correspond to a unique level of the echelon inventory position I(t). it is still optimal to have orders arrive at the lower echelon when the inventory level reaches zero. echelon inventory level I(t) = I1 (t) + It (t) + I2 (t).1. Note that the reorder instants at the higher echelon do not correspond to a unique inventory level of I1 (t). cf. those in the lower two graphs the ordering (delivery) instants at the higher echelon. The echelon .9). and the saw-tooth pattern of the . cf.4 which shows the saw-tooth pattern of the inventory level I2 (t) at the lower echelon. However. in system (from top to bottom). at higher echelon. This situation is illustrated for the case Q1 = 2Q2 in Figure 3. TILBURG UNIVERSITY Q2 T 63 ↑ I2 (t) rr s2 rr rr c rr rs 0 L2 rr rr r rr c r s r rr r rr r rr c r s r 2T2 2T2 +L2 rr rr r rr c r s r 3T2 3T2 +L2 r r rr rr c r s E r 4T2 4T2 +L2 T2 T2 +L2 t→ ↑ It (t) Q2 T c 0 Q1 T ↑ I1 (t) Q2 s L2 c s c s c s t→ c s E T2 T2 +L2 2T2 2T2 +L2 3T2 3T2 +L2 4T2 4T2 +L2 c −L1 Q1 +s2 T ↑ I(t) s 0 1 T 2 1 c T1 − L1 s T1 3 T 2 1 c 2T1 − L1 s 2T1 t→ E r rr r r rr r rr rr r r rr r rr r rr r r c s1 r s2 −L1 rr rr rs r 0 rc rr rr s T1 rr rc rr 2T1 − L1 rr s E 2T1 t→ T1 − L1 Figure 3. Finally.4 indicate the ordering (delivery) instants at the lower echelon. The open (closed) dots in the upper two graphs of Figure 3.

From (3. Compute the minimum average cost at the shop according this policy. the order quantity Q0 of the distribution center and the order quantity Qi of retailer i are determined by Q0 = D0 R. Example 3. • h0 : the holding cost per unit per unit of time for the distribution center. Lead times are deterministic. D = 1800 per year. T2 T2 (3. . • hi : the holding cost per unit per unit of time for retailer i. . h2 = $9. the holding cost is much higher there. if a holding cost is incurred for units in transit. Next. N . The cost per year for the optimal pull policy is C = 180 + 880 = $1060. i = 1. The average ˆ cost per year for the optimal push policy is C(2) = $1020.3) and the same order quantities are optimal as those in the corresponding case with L2 = 0. . i = 1.5).6) it follows that ∗ ˆ∗ ˆ∗ n∗ push = 2. Compute the minimum average cost per year corresponding to this policy. which is 1 2 $40 (3.1 Consider the following two-echelon inventory problem. . The internal ordering cost of the shop is $10 per order. • Di : the demand per unit of time at retailer i. First determine the optimal order quantity for the shop leaving the cost incurred at the depot out of account. while the distribution center orders this item from an external supplier. From (3.04 year and L2 = 0.2 Diverging chains Consider a diverging distribution chain consisting of two echelons. Hence. pull The optimal order quantity at the lower echelon is Q∗ = 20. . Qi = Di R/ki . The external ordering cost of the depot is $80 per order. cf. Given this policy at the shop. .4 Consider a linear chain consisting of two echelons. . . .64 reorder point at the higher echelon is MULTI-ECHELON INVENTORY SYSTEMS s1 = s2 + D · L1 = D · (L1 + L2 ). .12) (3. The parameters of the model are: • a0 : the ordering cost for the distribution center. i = 1. . Determine the optimal order quantities for the shop and the depot. . For a certain item. This demand is approximately deterministic and constant over time. the optimal order quantity at the higher 2 echelon is Q∗ = 6Q∗ = 120. . The distribution center uses a reorder cycle of R units of time and retailer i orders ki times per cycle of length R from the distribution center.2. i = 1. • ai : the ordering cost for retailer i. . The aggregated demand per unit of time at the distribution center is D0 = i=1 Di . a2 = $1. The demand for this item at the shop is 6000 units per year.13) which is independent of the order quantities Q1 and Q2 . N retailers order a certain item from the same distribution center. Determine for both policies for both echelons the reorder points for the case that the lead time to the depot is 10 days and the lead time from the depot to the 1 shop is 2 day (assume 300 working-days in a year).01 year. A retailer has a shop in the city center and a depot at the industrial site of the same city.9%) more than the minimum cost. Consider the following subclass of push policies. Therefore. For both policies the echelon reorder point for the higher echelon is s1 = 1800 · 0. Hence.14) N .05 = 90 and the reorder point for the lower echelon is s2 = 1800 · 0. Space in the city center is scarce. (3. N . . (3.01 = 18. N . The corresponding optimal order quantities are Q2 (2) = 60 and Q1 = 2Q2 (2) = 120. a fixed amount is added to the cost function (3. N . N.9) it follows that n∗ = 6. . . with parameters a1 = $32. the holding cost is $3 per unit per year at the shop and $1 per unit per year at the depot. consider the two echelons combined. determine the optimal order quantity and the minimum average cost for the depot. 3. L1 = 0. . h1 = $8. Exercise 3. . All demand must be satisfied from stock. The average number of units in transit is Q2 L2 + 0(T2 − L2 ) Q2 ¯ It = = L2 = DL2 . i = 1. . Then.

HANS BLANC, TILBURG UNIVERSITY

65

The objective function is the total ordering and holding cost per unit of time. It can be derived in a similar way as that for a linear chain, cf. (3.2), noting that the distribution center has similar holding costs related to each retailer, cf. (3.1): C(R, k) = This function can be rewritten as C(R, k) = 1 hi − h0 1 a0 + ki ai + 2 R h0 D0 + Di . R ki i=1 i=1
N N

a0 ki ai ki − 1 hi D i + + 1 Rh0 Di + 1 R . 2 2 R R ki ki i=1 i=1 i=1

N

N

N

(3.15)

(3.16)

This cost function has the same structure as that of the indirect grouping coordinated replenishment problem, cf. (2.1), only the role of the parameters with respect to the integers ki , i = 1, . . . , N , is reversed. The cost function (3.16) can be transformed into (2.1) by the following correspondences, cf. Das & Goyal [20]. The two-echelon diverging supply chain with N retailers is equivalent to a coordinated replenishment problem for N + 1 items (the distribution center is related to an item labeled 0) with parameters: for i = 1, . . . , N , T = 1 ; R ˆ A = 0;
1 ai = 2 (hi − h0 )Di ; ˆ

a 0 = 1 h0 D 0 ; ˆ 2

1ˆ ˆ 2 hi Di

= ai ;

1ˆ ˆ 2 h0 D 0

= a0 .

(3.17)

Note that the reorder frequency of the distribution center is forced to be k0 = 1. After the above transformation, the optimal policy can be determined with the aid of Goyal’s Algorithm 2.1 or a policy can be generated by any heuristic for the indirect grouping coordinated replenishment problem, with the modification that k0 is kept equal to 1. Remark 3.2 It may seem odd that a coordinated replenishment problem is formulated with family ordering ˆ cost A = 0. The cost function (3.16), however, does not become separable in this case, because the integer constraints on the frequencies ki , i = 1, . . . , N , are essential for its validity. It is quite difficult to determine a pull policy for a diverging chain. Each stocking point at the lowest echelon can readily determine its optimal reorder quantity according to the EOQ-formula, but without any coordination this will lead to irregular, acyclic demand patterns at higher echelons. This makes it cumbersome to formulate the average inventory levels at higher echelons and to determine optimal reorder policies for higher echelons. Table 3.1: Data per year and a push policy for Example 3.5. Retailer i a b c d e f ai $1.80 $2.00 $1.20 $3.20 $3.10 $2.70 hi $0.50 $1.10 $0.90 $0.30 $0.90 $0.30 Di 2900 1850 2750 1600 3200 1400 ai ˆ $580 $925 $1100 $160 $1280 $140 ˆ ˆ hi Di $3.60 $4.00 $2.40 $6.40 $6.20 $5.40 ai /(hi Di ) ˆ ˆ ˆ 161.1 231.3 458.3 25.0 206.5 25.9 ki 3 3 4 1 3 1 Qi 133 85 95 220 147 193 Ti 0.0458 0.0458 0.0344 0.1375 0.0458 0.1375

Example 3.5 Consider a two-echelon diverging distribution chain consisting of a distribution center and N = 6 retailers. At the distribution center, the ordering cost is a0 = $10 and the holding cost is h0 = $0.10 per unit per year. The ordering cost ai , the holding cost hi per unit per year and the demand Di per year at retailer i, i = 1, . . . , 6, can be found in Table 3.1. The aggregated demand per year at the distribution center is D0 = 13,700. The corresponding coordinated replenishment problem concerns N + 1 = 7 items, has ˆ family ordering cost A = $0, supplementary ordering costs a0 = $685 and ai , i = 1, . . . , N , given in Table ˆ ˆ ˆ ˆ ˆ ˆ 3.1, and holding cost times demand h0 D0 = $20 and hi Di , i = 1, . . . , N , given in Table 3.1, cf. (3.17). We determine an ordering policy for the transformed problem according to the heuristic of Kaspi & Rosenblatt ˆ ˆ ˆ ˆ (Algorithm 2.3). First, the items are indexed in ascending value of (A + ai )/(hi Di ) = ai /(hi Di ). This leads ˆ ˆ ˆ ˆ 0 D0 ) = 34.3). The initial family ˆ to the reference “item” being retailer d (for the distribution center, a0 /(h ˆ √ cycle based on this reference item is T (0) = 5 2 ≈ 7.071. The reorder frequencies for the other retailers, based on this cycle length, are ka = kb = ke = 3, kc = 4, kf = 1 (k0 = 1 and kd = 1 are kept fixed). Given

66

MULTI-ECHELON INVENTORY SYSTEMS

these frequencies, the optimal cycle length is T (1) = 7.270. The policy improvement step does not alter the frequencies. Hence, the family cycle becomes T (2) = 7.270, so that the review period is R = 0.1375 year (≈ 7.15 week) for the two-echelon system. The order quantity of the distribution center becomes Q0 = 1884. The reorder frequency ki and the order quantity Qi of retailer i, i = 1, . . . , N , are included in Table 3.1. The ˆ corresponding average yearly cost is C = $601.99 which is slightly higher than the minimum cost according to Goyal’s Algorithm 2.1: C ∗ = $601.62. Table 3.2: Optimal pull policies at the lower echelon for Example 3.5. Retailer a b c d e f Q∗ i 144.5 82.0 85.6 184.8 148.5 158.7 Ti∗ 0.0498 0.0443 0.0311 0.1155 0.0464 0.1134 Mi∗ 20.07 22.56 32.11 8.66 21.55 8.82

Table 3.2 contains the individual optimal reorder quantity Q∗ , and the corresponding optimal cycle Ti∗ and i number of orders per year, Mi∗ , for retailer i, i = 1, . . . , 6, according to the basic EOQ models. Because the reorder cycles are asynchronous, these optimal pull policies lead to an acyclic demand pattern at the distribution center. The application of push policies in divergent multi-echelon chains requires synchronization between the retailers such that an order is shipped to each retailer at the instant when an order from the external supplier arrives at the distribution center (partial cross-docking). In case of nonnegligible lead times L0 to the distribution center and Li from the distribution center to retailer i, i = 1, . . . , N , the distribution center should place an order when the echelon inventory position reaches the echelon reorder point, cf. (3.12),
N N

s0 = D0 L0 +
i=1

Di Li =
i=1

Di (L0 + Li ).

(3.18)

When the order arrives at the distribution center, the inventory position at retailer i will be si = Di Li , i = 1, . . . , N . Exercise 3.2 Consider a two-echelon distribution chain for a non-food product consisting of a local warehouse and five supermarkets. At the warehouse, the ordering cost is $20 and the holding cost is $0.20 per unit per year. The ordering cost a2i , the holding cost h2i per unit per year and the demand D2i per year at supermarket i, i = 1, . . . , 5, can be found in the table below. Determine an ordering policy for the transformed problem according to the heuristic of Kaspi & Rosenblatt and interpret the results in terms of the distribution chain. Compare the found push policy with the individual optimal policies. Supermarket i 1 2 3 4 5 a2i $8 $10 $10 $12 $15 h2i $0.60 $0.80 $0.50 $0.60 $0.50 D2i 2000 5000 4000 5000 4000

3.3

Time-varying demand

In this section, multi-echelon inventory systems are considered with known but time varying demand in the stocking points at the lowest echelon of diverging chains. The optimal pull policy is easy to find for general diverging chains by the following procedure: Algorithm 3.1 [Pull policy] Step 1: Apply the Wagner-Whitin Algorithm 1.2 to the stocking points at the lowest echelon.

HANS BLANC, TILBURG UNIVERSITY

67

Step 2: Aggregate the order quantities of the lowest echelon to obtain the demand process for the one but lowest echelon. Step 3: Apply the Wagner-Whitin algorithm to the stocking points at the one but lowest echelon. Step 4: Continue in this way until the highest echelon is reached. Alternatively, the Silver-Meal heuristic can be used. The pull policy ignores cost dependencies between the echelons. Therefore, the aim is to determine a (sub)optimal push policy that takes cost dependencies into account. An integral approach by dynamic programming is rather difficult. Here, a sequential approach (from the lowest to the highest echelon) will be discussed for a linear chain. The parameters of the model with a horizon of H periods are: • D2t , t = 1, . . . , H: the demand per period at the lower echelon; • a1 , a2 : the ordering cost at the higher/lower echelon, respectively; • h1 , h2 : the holding cost per unit per period at the higher/lower echelon, respectively. The decision variables are: • Q1t , t = 1, . . . , H: the order quantity at the higher echelon in period t; • Q2t , t = 1, . . . , H: the order quantity at the lower echelon in period t. The demand at the higher echelon is D1t = Q2t , t = 1, . . . , H. Blackburn & Millen [13] have developed a heuristic for an assembly system, that is, for a converging chain. For the case of a linear chain with two echelons it is based on an ideal ratio n between the order quantities at the higher and at the lower echelon derived from the constant demand model, cf. (3.6):    a1 (h2 − h1 )  n = max 1, . (3.19)   a2 h1 Note that n is taken minimally equal to 1, but is otherwise not rounded to an integer. Motivated by cost 1 function (3.3) coordination between the echelons is extorted by adding a fraction n of the ordering cost a1 to each order placed at the lower echelon and by adding (n − 1)h1 to the holding cost at the lower echelon. Hence, the following modified cost parameters are defined for the lower echelon: . a2 = a2 + a1 /n, ˆ ˆ h2 = nh1 + h2 − h1 . (3.20)

The heuristic of Blackburn & Millen [13], confined to a linear chain consisting of two echelons, proceeds as follows. Algorithm 3.2 [Blackburn & Millen] Step 1: Determine the ideal ratio n according to (3.19). Step 2: Apply the Wagner-Whitin algorithm to the lower echelon with modified costs a2 , h2 , cf. (3.20), ˆ ˆ to determine Q2t , t = 1, . . . , H. Step 3: Use D1t = Q2t , t = 1, . . . , H, and apply the Wagner-Whitin algorithm to the higher echelon with costs a1 , h1 , to determine Q1t , t = 1, . . . , H. Here, the ratio n is abandoned. Alternatively, the Silver-Meal heuristic can be used. Example 3.6 Let a1 = $20, a2 = $4, v1 = $20, v2 = $30, r = $0.02/$/month, so that h1 = $0.40, h2 = $0.60. The demand in the next 8 months D2t , t = 1, . . . , 8, is given in Table 3.3. For a push policy √ according to Blackburn & Millen compute n = 20×10 = 1 10 ≈ 1.58 > 1 by (3.19) so that the modified 4×20 2 20 ˆ costs become a2 = 4+ 1.58 = 16.65 and h2 = 0.2+1.58×0.4 = 0.832, cf. (3.20). Application of the Silver-Meal ˆ heuristic at the lower echelon with the modified costs yields Q2t , t = 1, . . . , 8, cf. Table 3.3. The estimated

. and Q16 = 70. t = 1. 8.71.3 A firm owns a shop in the center of a city and a depot near the harbor. The realized ratio between the order frequencies is n = 6/4 = 1. . .6. cf.4. h1 gives Q1t . The cost at the lower echelon is 8a2 + 0h2 = $32. Observe that the Silver-Meal heuristic leads in this case to a pull policy that is cheaper than the push policy obtained with this heuristic! Application of the Wagner-Whitin algorithm gives the same results for the lower echelon but different results for the higher echelon.6. . Table 3. Q17 = 25. . .68 MULTI-ECHELON INVENTORY SYSTEMS Table 3. .3: Push policy according to Blackburn & Millen and Silver-Meal for Example 3. Goods are first delivered at the depot where they are checked. cost 6a2 + (10 + 15)h2 = $39 at the lower echelon and total cost $149.6. Month D2t Q2t Q1t 1 40 40 50 2 10 10 0 3 60 60 115 4 40 40 0 5 15 15 0 6 45 45 70 7 25 25 0 8 25 25 25 Table 3. Application of the a Silver-Meal heuristic at the higher echelon with demand D1t = Q2t and costs a1 . week demand 1 24 2 16 3 32 4 24 5 12 6 50 The ordering cost at the depot is $60. Month D2t Q2t Q1t 1 40 40 50 2 10 10 0 3 60 60 60 4 40 40 55 5 15 15 0 6 45 45 95 7 25 25 0 8 25 25 0 Exercise 3. cf.6: Pull policy for Example 3. . 8. The total real cost is $165. Application of the Wagner-Whitin algorithm gives the results as displayed in Table 3. .3. Q18 = 25.50. The total cost is $154. t = 1. Month D2t Q2t Q1t 1 40 50 50 2 10 0 0 3 60 60 60 4 40 55 100 5 15 0 0 6 45 45 0 7 25 25 50 8 25 25 0 Table 3. Table 3. cf. Month D2t Q2t Q1t 1 40 50 50 2 10 0 0 3 60 60 60 4 40 55 55 5 15 0 0 6 45 45 95 7 25 25 0 8 25 25 0 ˆ cost at the lower echelon based on the modified costs is 6ˆ2 + (10 + 15)h2 = $120. . Table 3. Now consider a pull policy. 8.5: Pull policy for Example 3. For a certain item the demand in the shop is deterministic but the amount varies from week to week. with cost 4a1 + (25 + 50)h1 = $110 at the higher echelon. The optimal pull policy has cost 4a1 + (10 + 15 + 25 + 50)h1 = $120 at the higher echelon. and total cost $152 which is 2% higher than that of the optimal push policy. Note that there are alternative optimal solutions: Q16 = 70. The demand for the next six weeks is given in the following table. Q18 = 0. the ordering cost at the shop is $15 and mainly consists of shipping cost. The realized ratio between the order frequencies is n = 8/4 = 2. Application of the Silver-Meal heuristic at the higher echelon with demand D1t = Q2t and costs a1 and h1 gives Q1t . Q17 = 0. Application of the Silver-Meal heuristic at the lower echelon with the unmodified costs a2 and h2 yields Q2t . The cost at the higher echelon is 4a1 +(10+40+2×15+25)h1 = $122. Again. Table 3. .5.4: Push policy according to Blackburn & Millen and Wagner-Whitin for Example 3. Holding costs are only charged for units on stock at the .6 by Silver-Meal. there are alternative optimal solutions. . repacked and kept in storage if necessary. t = 1. Table 3.6 by Wagner-Whitin. cf. Orders are delivered at the first day of a week. The real cost at the higher echelon is 4a1 + (90 + 25)h1 = $126 and the real cost at the lower echelon is 6a2 + (10 + 15)h2 = $39.5.

HANS BLANC, TILBURG UNIVERSITY

69

end of a week. The holding costs are $1 per unit per week at the shop and $0.20 per unit per week at the depot. Determine a pull policy with the aid of the Silver-Meal heuristic for this two-echelon inventory system. Determine a push policy with the aid of the Blackburn & Millen heuristic combined with the Silver-Meal heuristic. Compute the total cost over six weeks corresponding to both policies. Show that these policies can simply be improved upon.

3.4

Stochastic demand

This section is concerned with two-echelon diverging chains with independent, stationary stochastic demand at the stocking points of the lower echelon. This represents a situation of a central depot or distribution center and N local stocking points or retailers. The aim is to develop an good integral control, that is, a push policy, cf. Sections 1.3, 3.2.1. The means to this goal are multi-echelon ordering rules, also called base-stock control, cf. Clark & Scarf [18]. Such strategies are optimal for diverging chains provided that the demand at the retailers is not too strongly varying, that is, there is no strong “imbalance” between the retailers. The approach is this section is based on Van der Heijden [68]. This problem is equivalent to a hierarchical production planning problem for a family of items with common components (ignoring production capacity constraints). The inventory policy is based on periodic review of the stock, applied as follows: • the central depot uses an (R, S0 ) policy: with fixed intervals R the echelon-stock is ordered up to S0 ; • the order of the central depot is delivered after a lead time L0 ; • the delivery instant at the central depot is the reorder instant for the retailers; • retailer i uses an (R, Si ) policy: the stock is ordered up to Si , i = 1, . . . , N ; • the order of retailer i is delivered after a lead time Li , i = 1, . . . , N . This basic procedure may require some modifications: • if the total demand during the lead time L0 is more than expected, then the stock in the central depot will be too small to fully satisfy the orders of the retailers; in such a case, – retailer i is first replenished up to the expected demand in the period Li + R; the remaining stock is divided over the retailers by ratios pi , i = 1, . . . , N , or – the total understock is divided over the retailers by ratios fi , i = 1, . . . , N ; • if the total demand during the lead time L0 is less than expected, then the stock in the central depot will be larger than the sum of the order quantities of the retailers; in such a case, – the remaining stock could stay in the central depot, or – the total overstock is divided over the retailers by ratios fi , i = 1, . . . , N . In the sequel, the following model will be considered: • there is a stockless central depot (it only serves for cross-docking and reallocation of goods); • there is full information at the central depot to apply a push policy for fair allocation of stock over the retailers and for lower stocks; • no stock redistribution is possible between retailers; • all shortages are backlogged; at each retailer, a fill rate constraint is imposed, cf. (1.38); • there are no capacity constraints. The parameters of the model are: • L0 : the lead time from the external supplier to the central depot; • Li : the lead time from the central depot to retailer i, i = 1, . . . , N ;

70 R L0 ————— —— L1 LN R —— L1 LN L0 R

MULTI-ECHELON INVENTORY SYSTEMS R L0 R —— L1 LN R L0 R

Figure 3.5: The review periods at the two echelons shifted in time (depot above, retailers below). • R: the review period (this period is of equal length for both echelons, but it is shifted in time by an amount L0 for the retailers in comparison to the central depot; see Figure 3.5); • Di,t : the demand at retailer i during a period t; this is a random variable with time-independent mean . . µi = E{Di,t } and standard deviation σi = σ{Di,1 }, i = 1, . . . , N ; • hi : the holding cost per unit per unit of time at retailer i, i = 1, . . . , N ; • βi : the target fill rate at retailer i, i = 1, . . . , N . It will be assumed that demand is independent across time periods and that demand is independent across retailers. Since the review period is given, the ordering and shipping costs are fixed. The lead times are assumed to be known and constant, and the approach below especially concerns situations where the lead time to the central depot is large in comparison to the lead times to the retailers (L0 L1 , . . . , LN ). The decision variables are: • S0 : the order-up-to level of the central depot;
∗ • Si : the maximum order-up-to level of retailer i, i = 1, . . . , N ;

• fi : the rationing fraction for allocating units to retailer i, i = 1, . . . , N .
∗ Here, the variables are to be chosen such that i=1 Si = S0 and i=1 fi = 1. The control mechanism of the inventory policy to be discussed uses the following auxiliary random variables: N N

• zi : the inventory position of retailer i just before a reorder instant of the retailers, i = 1, . . . , N ; • Si : the inventory position of retailer i just after a reorder instant of the retailers, i = 1, . . . , N ; • Q: the order quantity of the central depot. ˜ When an order arrives at the central depot, the echelon inventory position of the central depot is S0 − DL0 , . N ˜L = with D 0 i=1 Di,L0 the total demand at all retailers during the lead time L0 . The amount that is allocated to retailer i by the central depot at such an instant is Si − zi with
∗ ˜ Si = Si − fi DL0 ,

i = 1, . . . , N.

(3.21)

˜ Hence, the maximum order-up-to level of retailer i is reduced by a fraction fi of the total demand DL0 during the lead time to the central depot L0 , for i = 1, . . . , N . The echelon inventory position just before the central depot places an order is S0 − Q. The echelon inventory on hand just before the order of the central depot arrives is:
N N

zj = S0 − Q −
j=1 i=1

˜ Di,L0 = S0 − Q − DL0 .

Hence, the allocation to retailer i can alternatively be described by Si − zi , with  
N ∗ Si = Si − fi S0 − Q − j=1

zj  ,

i = 1, . . . , N.

The echelon inventory position just after the delivery of an order at the central depot is:
N N

˜ Si = S0 − DL0 =
i=1 j=1

zj + Q.

HANS BLANC, TILBURG UNIVERSITY

71

A problem in the foregoing allocation arises if Si − zi < 0 for some retailer i = i0 . This phenomenon is indicated by imbalance. The cause for imbalance is • little demand at retailer i0 (the remaining amount of stock zi0 is relatively high); ˜ • more demand at other retailers (the total demand DL0 is such that Si0 is smaller than zi0 , cf. (3.21)). In case of imbalance, the allocation rule would require redistribution of stock from retailer i0 to some other retailers, which would lead to extra shipping cost. Let us study the occurrence of imbalance. The inventory position at retailer i just after the order quantity Q of the central depot has been reallocated is, cf. (3.21), Si
(1) ∗ ˜ = Si − fi DL(1) ,
0

i = 1, . . . , N.

(3.22)

Here, it is assumed that there is no imbalance at the outset. The inventory position at retailer i just before the next allocation instant, that is, a review period R later, is zi
(2) ∗ ˜ = Si − fi DL(1) − Di,R ,
0

i = 1, . . . , N.

(3.23)

This inventory level must be raised to the target level Si
(2) ∗ ˜ = Si − fi DL(2) ,
0

i = 1, . . . , N.
(2)

(3.24) before allocation. (3.25)

Imbalance occurs if the target level Si is smaller than the actual inventory level zi Hence, the imbalance due to retailer i is: . (2) (2) ˜ ˜ Ωi = [zi − Si ]+ = [fi DL(2) − fi DL(1) − Di,R ]+ ,
0 0

(2)

i = 1, . . . , N.

This definition of imbalance was introduced by De Kok [21]; Zipkin [83] used another type of imbalance measure. It is seen that the imbalance only depends on the rationing factors fi , and not on the order-up-to ∗ levels Si , i = 1, . . . , N . To avoid the occurrence of imbalance as much as possible, the ratios fi will be chosen in such a way that the total expected imbalance is minimal. To this end, Van der Heijden [68] assumes that the random variables, cf. (3.23), (3.24), . (2) (2) ˜ ˜ Yi = zi − Si = fi DL(2) − fi DL(1) − Di,R ,
0 0

i = 1, . . . , N,

(3.26)

approximately possess a normal distribution. Note that the random variables Yi , i = 1, . . . , N , can take positive and negative values, but are not symmetric around their mean. The mean of Yi is independent of fi : E{Yi } = E{zi
(2)

− Si } = −E{Di,R } = −Rµi ,

(2)

i = 1, . . . , N.

(3.27)

Note that the terms on the righthand side of (3.26) contain overlapping demand since the second lead time (1) (2) L0 partly coincides with the review period R at the retailers and possibly with the previous lead time L0 . To determine the variance of Yi , two cases are distinguished, cf. Figure 3.6: • R > L0 : then the contributions to Yi are: ˜ period (0, L0 ): − fi DL(1) , 0 period (L0 , R): − Di,R−L0 , period (R, L0 + R): fi j=i Dj,L0 − (1 − fi )Di,L0 ; • R < L0 : then the contributions to Yi are: ˜ period (0, R): − fi DR , (1) (2) period (R, L0 ): 0 (since L0 and L0 overlap in this period), period (L0 , L0 + R): fi j=i Dj,R − (1 − fi )Di,R . . With M = min{R, L0 }, the variance of Yi can be written as
N 2 σ 2 {Yi } = [R − 2fi M ] σi + 2fi2 M j=1 2 σj ,

i = 1, . . . , N.

(3.28)

. with initial search interval [fj . Therefore.) denotes the normal loss function defined by (A. (3. N . cf.72 MULTI-ECHELON INVENTORY SYSTEMS L0 (1) R L0 R (2) L0 (1) R L0 (2) (R) ———————— 0 L0 (R) ————— L0 + R 0 R L0 L0 + R Figure 3.18). . . . Since σ{Yj } depends on fj ..25). . for j = 1. . NL(. d dfj N E{Ωi } = i=1 d d M f0. . i = 1. cf. N . (3. (3.1 (Rµj /σ{Yj }) 2 E{[Yj ]+ } σ{Yj } = −σj + 2fj dσ{Yj } dfj σ{Yj } N 2 σi .Li − Si ]+ }. Appendix B. j = 1. As a consequence. cf.29) with η the Lagrange multiplier. . (3. In this way. N. i = 1.38). a numerical search procedure must be applied to find this root. . are not feasible with respect to the constraint i=1 fi = 1. Further. In the next step of the approach by Van der Heijden [68] it is assumed that the fractions fi are fixed according ∗ to the above procedure. . . The mean imbalance E{Ωi } = E{[Yi ]+ } is increasing with the variance σ 2 {Yi } when Yi has a normal distribution. . it follows that. Therefore. However. Because of this total expected imbalance minimizing property. (3. it is assumed that by the choice of the rationing fractions the occurrence of imbalance can be ignored. N . N . cf. the minimum total expected imbalance with respect to the fractions fj . i=1 i=1 fi = 1. e. . . .28). N . N . The derivatives of this function with respect to the rationing factors fi can be determined by means of the chain rule. . the following equations must be satisfied: d dfj N N E{Ωi } = η.27). (3. . In this step.Li +R − Si ]+ } − E{[Di.33) The mean stockout Bi accumulated during a review period R at retailer i is. cf. N N N E{Ωi } = i=1 i=1 E{[Yi ]+ } = i=1 σ{Yi }NL(Rµi /σ{Yi }). the constrained optimization problem has to be solved with the Lagrange multiplier method. . (3.26). f0. . . N. . these values of fi . i = 1. and the maximum order-up-to levels Si are determined such that the target fill rates βi are reached. for j = 1. . . i = 1.R } = Rµi . E{Qi } = E{Si − zi } = E{Di.3. . Appendix A. the ˆ bisection method for finding a root of an equation. N . j = 1.1. N . It follows by differentiation of (3.g. .1 (. (1. . . .31) here. E{Bi } = E{[Di. From (3. . i=1 (3. retailer i is allocated a positive order quantity Qi = Si − zi at each order arrival instant at the central depot. it is not possible to determine an explicit expression for the rationing fraction fj for which dE{Ωj }/dfj = η for a given value of η. i = 1. . (3. cf. can be obtained numerically.6: Review periods at the retailers and lead times to the depot: two cases. i = 1. . . . in order to find the value of η for which the fractions fj sum to one another application of the bisection method is required. In order to minimize the function N N N E{Ωi } + η 1 − i=1 i=1 fi .2. .30) Under the assumption of normally distributed random variables Yi . with mean. . . . . 1].34) . the above rationing rule is referred to as “Balanced Stock” rationing. the total expected imbalance reads.28) and the properties of the normal loss function.28) that the variance σ 2 {Yi } is minimal with respect N 1 2 2 ˆ to fi at fi = 2 σi / j=1 σj . (3.) denotes the density of the standard normal distribution.32) here. . .

(3.Li +R + fi DL0 − Si ]+ }. (3. . . ∗ ∗ ˜ ˜ E{[Di. . i = 1. .41) and with ∗ ˜ E{[Si − Di. . . . i = 1.40) (3. that satisfy the above inequalities.42) .37) The variances of these variables are: 2 σ 2 {Xi } = (Li + R)σi + fi2 L0 j=1 (1) (2) (1) N 2 σj . .HANS BLANC. (3. compute the smallest maximum order-up-to ∗ levels Si such that the target fill rates βi are reached. N . . Van der Heijden [68] approximates the distributions (1) (2) of Xi and Xi by mixtures of Erlang distributions with mean and variance given by (3.35) can be determined numerically.Li + fi DL0 − Si ]+ } ≤ (1 − βi )Rµi . i = 1. (3. Step 2: Given the values of the fractions fi .39): ¯ E{Ii } = 1 2 [E{[Si − Di. i = 1. . . cf. For each considered value of η. the smallest maximum order-up-to levels Si satisfying (3. The computation of the decision variables for the replenishment strategy is summarized below. 2 σ 2 {Xi } = Li σi + fi2 L0 j=1 (2) N 2 σj .21).Li +R . .39) Since for any real x. . . for i = ∗ 1. N. The order-up-to level at the central depot follows as S0 = N i=1 ∗ Si .5. TILBURG UNIVERSITY 73 ˜ In the present case. The means of these variables are: E{Xi } = (Li + R)µi + fi L0 j=1 (1) N µj . i = 1. . . cf. (1. N. N ∗ E{Si } = Si − fi L0 j=1 i = 1. .3. . Then. E{Xi } = Li µi + fi L0 j=1 (2) N µj . . Algorithm 3. N . . . i = 1.21) that ∗ ˜ E{[Si − Di. both Di. cf. N . . i = 1. i = 1.Li +R + fi DL0 − Si ]+ } − E{[Di. . (3.36) ˜ Since the lead time L0 to the central depot occurs before Li + R.Li + fi DL0 − Si ]+ }. cf. Appendix A. (3. This step requires a numerical procedure to N find the value of the Lagrange multiplier η for which i=1 fi = 1. the maximum order-up-to levels Si have to be determined such that. it d requires a numerical procedure to find the value of fi for which dfi E{Ωi } = η. Therefore. . [x]+ = x + [−x]+ . . i = 1. N.Li +R ]+ }]. Van der Heijden [69]. (1. for i = 1. Such a problem involves random variables Xi (1) ˜ = Di. . N . Note that there are no decision variables left to minimize the total holding cost.35) ∗ Note that the smallest Si . N. . (3. . N. . .Li and Si are random variables. such that the expected total imbalance is minimal. cf. . . With these approximations. for each i. . N. . . . since Si depends on DL0 . N . .37) and (3. This issue could be addressed in an extension of the present model with a stock-keeping central depot. with the aid of the Lagrange multiplier method. .Li ]+ } + E{[Si − Di.Li +R + fi DL0 . i = 1.38) The distributions of Xi and Xi cannot readily be obtained for general probability distributions of the demand per period at the various retailers. N. .2. µj . ˜ items could be stored at the central depot in case the demand DL0 during the lead time to the central depot is much less than average and if the holding cost h0 per item per unit of time at the central depot is lower than those at the retailers. N. The actual fill rates are based on an approximation by a mixture of Erlang distributions.Li +R are independent random variables. . Di. . .Li ]+ } = E{Si } − Li µi + E{[Di. In the present model.3 [Balanced Stock rationing] Step 1: Compute the rationing fractions fi .Li + fi DL0 . . (3. .Li +R ]+ } = E{Si } − (Li + R)µi + E{[Di. . . . . . it follows with (3. . . N . Xi (2) ˜ = Di. . ∗ respectively. The expected average inventory on hand at retailer i is. the total cost is fixed after application of Algorithm 3. In order to reach the target fill rates. can be determined for each retailer separately (if imbalance is ignored). i = 1. .38). . it follows that DL0 and Di.38). . .

. . . N . . Further.189 1109 220 30 55 0. the expected average inventory on hand E{Ii } at retailer i can readily be computed since the expec∗ tations on the right-hand sides of (3.948 0. . see also Verrijdt & De Kok [74. But the allocation to a retailer with relatively high .43) Finally. 75]. .2. Section 1. the items are allocated to the retailers such that their inventories are raised to the levels   N ˜ Si = E{Di.5 Retailer 2 300 50 0.11 585 0. . the average inventory in transit to retailer i is Li µi . . . These fractions are chosen proportional to the safety stocks: pi = cKi σ{Di. meaning high uncertainty in demand.0 Example 3. independent of the policy (due to the assumption of complete backlogging). . This retailer gets the smallest rationing fraction. Mean demand µi St. The fraction pi is relatively large if σ{Di. . N. Simulation indicates that the actual fill rates are slightly below the targets (in the order of 0. .90 1 0. since the coefficient of variation σ1 /µ1 = 0. N .898 0.7: Data and results for Example 3. . N . . . The simulated mean imbalance turns out to be much smaller than the mean imbalance predicted by the normal loss function in the model. . .85 and the target fill rate β1 are maximal over all retailers. . that is.21) by identifying   N fi = pi . .95 1 0.7 Van der Heijden [68] contains a numerical example of which the data can be found in Table 3.698 0. .3 De Kok [21] introduced another rationing policy for the present multi-echelon inventory problem called Consistent Appropriate Share (CAS) rationing. The review period is R = 1 week. . At the order arrival instant of the central depot.002). with Ki a safety factor such that the target fill rate βi is reached. i = 1.86 257 0.Lj +R } . i = 1. . . This allocation falls in the framework of the more general allocation rule (3. N . (3. .7 (in weeks). In this approach.Lj +R } . retailer i gets allocated its expected demand during the period Li + R plus a fraction of the remaining amount of items. i = 1. i = 1.40) and (3. . deviation demand σi Target fill rate βi Lead time Li Rationing fraction fi ∗ Maximum order-up-to level Si ¯ Mean on-hand inventory E{Ii } Mean stockout E{Bi } Mean safety stock E{Vi } Mean imbalance E{Ωi } Simulated mean on-hand inventory Simulated actual fill rate Simulated mean imbalance Central depot (1200) (412) 2 5094 Retailer 1 100 85 0. N .41) have already been calculated when Si was determined satisfying the inequalities (3. ∗ Si = E{Di.4 Retailer 3 800 400 0.780 3505 587 240 34 1. Remark 3. i = 1.19 219 0. items are allocated to the retailers based on safety stock ratios.7: N ∗ E{Vi } = E{Si } − (R + Li )µi = Si − fi L0 j=1 µj − (R + Li )µi . . i = 1. N. N .Li +R } is large. Table 3.031 480 258 5 205 4. . and C a normalization N constant such that i=1 pi = 1. .70 1 0.74 MULTI-ECHELON INVENTORY SYSTEMS ¯ Hence. cf. i = 1. .7.Li +R } + pi S0 − j=1 E{Dj. . . The most critical retailer is retailer 1. . and/or if Ki is large.Li +R }. i = 1.35). meaning a high target fill rate βi . the average safety stock at retailer i is.Li +R } + pi S0 − DL0 − j=1 E{Dj.

5 0.9: Results for Example 3. The Balanced Stock rationing is better in avoiding negative allocations and is better in predicting the actual performance of a system. TILBURG UNIVERSITY 75 ˜ fraction fi = pi is most sensitive to variations in the total demand DL0 . but remains largest for retailer 2 with the highest coefficient of variation of the demand.7.075 444 179 10 74 0. the model can be extended with a stock keeping central depot.15 Retailer 3 0.650 2063 632 20 422 1. 69].078 323 104 10 49 0. Retailer 2 has a higher coefficient of variation of the demand. and is the most critical retailer. in particular with respect to the actual fill rates. Further.8 To study the effect of various parameter values. cf. The parameters of retailer 1 form the base case. The review period is R = 1 week. Table 3. .HANS BLANC. N (the fractions fi should be recomputed only if the demand or lead time characteristics change considerably). For instance.08 Retailer 2 0. Mean demand µi St. but then the expressions for the actual fill rates and the procedure for determining the rationing fractions become more complicated and require more approximative assumptions.0 0. Retailer 4 has a higher mean demand with the same coefficient of variation.095 629 332 10 227 4.95) in Example 3. .9 contains results for the same two-echelon system but with a review period of R = 2 weeks.27 Retailer 5 100 50 0. cf. as suggested by the optimal solution to the unconstrained problem. Van der Heijden [68]. even after improvements of the method for cases when some fractions pi are negative.21).60 Example 3. deviation demand σi Target fill rate βi Lead time Li Rationing fraction fi ∗ Maximum order-up-to level Si ¯i } Mean on-hand inventory E{I Mean stockout E{Bi } Mean safety stock E{Vi } Mean imbalance E{Ωi } Depot (800) (240) 2 3715 Retailer 1 100 50 0. retailer 4 gets the largest rationing fraction. cf. i = 1.8: Data and results for Example 3. In this case.60 Retailer 2 100 100 0. It can deal ∗ with nonstationary demand by adjusting the maximum order-up-to levels Si . .95 0.5 0. Retailer 3 has a lower target fill rate.08 Retailer 4 0. cf. Finally.8 (in weeks) with review period R = 1 week.8 (in weeks) with review period R = 2 weeks. while it leads to lower mean stock levels than CAS rationing. This explains why CAS rationing yields strong imbalance and erroneous results in some cases where simulation results do not agree with model predictions. and retailers 1. . 3 and 5 get the smallest rationing fraction which is the same for these three retailers because the rationing fractions do not depend on the lead times and the target fill rates.078 354 132 5 80 0. Retailer 5 has a higher lead time.40 Retailer 3 100 50 0. (3. To study the effect of the review period.075 403 143 20 33 0.95 0.60 Retailer 4 400 200 0.075 504 189 10 84 0. most sensitivity is directed to retailers of which the inventory is already difficult to control. Hence.95 0.08 The inventory policy described in this section can be extended in several directions. the predicted mean imbalance decreases considerably.078 419 148 5 95 0.116 555 272 5 219 7. above (3.17 Retailer 5 0. Depot Rationing fraction fi ∗ Maximum order-up-to level Si ¯i } Mean on-hand inventory E{I Mean stockout E{Bi } Mean safety stock E{Vi } Mean imbalance E{Ωi } 4448 Retailer 1 0. consider the following numerical example with five retailers of which the data can be found in Table 3. then retailer 2 (in the order of decreasing standard deviation of the demand. It can be generalized to divergent multi-echelon systems with more than two echelons.5 0.8.5 0.680 2469 800 40 380 0. Increasing the review period shifts the rationing fractions somewhat in favor of retailer 4 with the highest standard deviation of the demand. Van der Heijden [68.90 0. Table 3. CAS rationing assigns a large fraction to retailer 1 (p1 > 0.95 1. . Table 3.29)).

2 week.90 for both retailers. Exercise 3.7 Consider the results presented in Tables 3. Exercise 3. . . The lead time to the central depot is L0 = 2 weeks and the lead time to the retailers are L1 = 0.8.28). Exercise 3.9. i = 1. N .6 Consider the results presented in Table 3. . for a review period of R = 1 week as well as for a review period of R = 2 weeks. Yoo et al. Federgruen [27] who applies stochastic dynamic programming. 3. Exercise 3. respectively. Exercise 3. The weekly demand has a mean of 20 units and a standard deviation of 5 units at both retailers. [25] who solve a supply network with base-stock control and service level constraints by the conjugate gradient method. while the maximum order-up-to levels and the mean physical stocks only increase with a much smaller percentage.4 Complete the details of the derivation of the variance σ 2 {Yi }. in (3.8 Consider a two-echelon system for an item with a central depot and two retailers.3 at target fill rates of 0.76 MULTI-ECHELON INVENTORY SYSTEMS Literature on this subject includes Eppen & Schrage [24] who consider a model with penalty cost.32).1 and L2 = 0. Inderfurth [40] who provides a survey on safety stocks in divergent multi-echelon systems.7. Axs¨ter & Zhang [8] and Axs¨ter [6] who provide an exact analysis of a a a continuous review system with (compound) Poisson demand. . Explain why the mean physical stock is only 585 for retailer 3 under Balanced Stock rationing while the maximum order-up-to level for this stocking point is 3505. Odanaka et al. and Ettl et al. Explain why the mean stockout seems to double for each retailer when the review period doubles. [53] who consider a multi-echelon production system. Korugan & Gupta [46] who study a multi-echelon system with returns. . [81] who are concerned with distribution requirement planning. Determine ordering policies according to Algorithm 3.5 Complete the details of the derivation of the derivative of the total expected imbalance with respect to the rationing fractions in (3.

• Φ: the total inventory capacity. . i = 1. (1.1 [Lagrange constrained optimization] Step 1: Compute the unconstrained optimal order quantities. . . since the ordering of the various items is not coordinated. QN ) = N ai i=1 Di 1 + 2 hi Qi . .1) This cost function has to be minimized subject to the capacity constraint N φi Qi ≤ Φ. Section 4. i=1 (4. . i = 1. . C(Q) = C(Q1 . N . (1.2) This constraint is based on the supposition that a space φi Qi has to be reserved for the maximum inventory level of item i. Algorithm 4. . . N . N . . .2). . • hi : the holding cost per unit per unit of time for item i. i = 1. • φi : the space occupied by a unit of item i. . The parameters of the model are: • ai : the ordering cost for item i. . . N . Q∗ = i 2ai Di /hi . . N. The demand of all items is assumed to be deterministic and continuous in time. i = 1.Chapter 4 Other Inventory Systems with Interactions This chapter is devoted to some other inventory problems with interaction. . cf. .1 is concerned with the inventory control of a set of items that have to share a common storage space. . . . 4. . . i = 1. Section 4. 77 i = 1. N .2 deals with a single item of which units may fail and then are replaced by another unit while the failed unit may be repaired later on. The decision variables are: • Qi : the order quantity for item i. . . . .1 Inventory capacity constraint In this section a set of N items is considered which are ordered independently but which have to share a common storage space. cf. Qi (4. . N . . The average cost per unit of time is found as a simple generalization of the single item case. . . • Di : the demand per unit of time for item i. . . i = 1. .1). .

product i: ai hi φi Di Qi Ci Ti Qi Ci Ti Qi Ci Ti Qi Ci Ti 1 $10 $0.0 100 100.06 0. the inventory capacity is fully used (consider the partial derivative of f (Q.64 0.437 136. N.68 0.00 0.2 $80.73 0.80 4. . e.37 0.000 87. stop. (4.342 91.249 8 $40 $0.441 143.51 0. N.4 $85.0 400 100.218 68.5 $27.31 0. η) with respect to Qi is d Di 1 f (Q. i = 1.0 $80. Otherwise. . that is. .000 66.20 4.80 0.46 0.84 0.8 $42.3 $80.0 $40.684 45.N Table 4. .29 0. .92 0.500 174. .500 174.3 $20. (4.0 $40.5 $161.0 $160..1: Data for Example 4. Appendix B. . η) with respect to η) so that N 2ai Di φi = Φ.66 0.20 1.0 400 200.342 91.00 0.00 1.668 42.3.00 0.5 $211. bisection).241 88.0 $80. is a feasible solution.4).8 $85. If so.. .7 $84.11 0.500 192. .78 OTHER INVENTORY SYSTEMS WITH INTERACTIONS Step 2: Verify whether the capacity constraint (4. .6) i=1.424 24.873 68. (4.1 and results according to Algorithm 4. cf. dQi Qi This implies that the optimal order quantity for fixed η is ˆ Qi (η) = 2ai Di .114 7 $10 $0.4) i = 1.0 400 200.8 $21.0 400 200.5 $43.0 400 100.20 1.75 0.7 $105.0 100 100. η) = i=1 ai Di + 1 hi Qi + η 2 Qi N φi Qi − Φ .00 0. The partial derivative of the function f (Q.0 $160.7 $161. ai Di hi − . continue with Step 3.80 1.3 $80. Note that by (4. . Step 3: Solve the constrained optimization problem with the aid of a Lagrange multiplier.8 $171. N.482 176.44 0.4 $168.80 1. Further. .9 $55.7 $160. φi Q2 2φi i i = 1. it is useful to have an upper and lower bound on the optimal value of η.47 0. i=1 (4.g...179 3 $10 $0.61 0.00 0.437 136.g. by the bisection method for finding a root of an equation.20 4.00 0.80 4.229 2 $10 $0.221 71.63 0.7 $26.94 0.. To this end.250 96.250 87. . N . . η) = −ai 2 + 2 hi + ηφi .0 $20.500 133. it follows that the optimal value of η lies on the 0 ≤ η ≤ max ai Di φ i N 2 hi − Φ2 2φi .5 $52.5) hi + 2ηφi i=1 This relation implicitly determines the scalar η.88 0.334 84.125 6 $10 $0.7 $40. The value of η has to be determined with the aid of a numerical procedure.9 $42.4 $21.171 45.457 4 $40 $0.0 $20.2) is satisfied.229 Φ ≥ 3000 Φ = 2400 Φ = 1800 Φ = 1200 . minimize the following function with respect to the vector Q and the scalar η: N f (Q.25 0.3) The last step requires a numerical search procedure (e.47 0.359 5 $10 $0. . hi + 2ηφi i = 1. .212 49.00 1.. η= Since Qi = interval 1 N Φ/φi .1.18 0.0 400 200.

. The following inventory policy will be studied. Therefore. with inventory capacity Φ < 3φ 2aD2 /h. The unconstrained optimal order quantities require a total storage space of 3000 volume units. 2 and the reorder cycle Ti = Qi /Di for each item. for other items it is the other way about.1 Table 4. Determine the optimal order quantities Q∗ . The decision variables are: • QO : the order quantity from the external supplier. φi . • LR : the lead time of a repair batch. In most situations. but repairable units. and also the repair lot size is always the same. Exercise 4. The holding costs are h1 = h4 = $1. There is no clear structure in the decrease of the order quantities: the order quantity of some items decreases more strongly when the storage capacity decreases from 80% to 60% than when it decreases from 100% to 80%.3 Prove that equation (4. • aO : the ordering cost for placing an order at the external supplier. It is assumed that a unit that fails is replaced by a spare part and that the failed units are later repaired. 4. Schrady [57]. it is assumed that repaired units are as good as new ones. Only the demand rates differ: D1 = 4D2 . cf. Di . φ3 = φ4 = 4 square meters. Exercise 4. .6) if Step 3 of Algorithm 4. • aR : the set-up cost for starting a repair batch. • w: the fraction of failed units that is no longer repairable (the waste rate). Table 4. i = 1. 2.91 at Φ = 1200. • hF : the holding cost per unit per unit of time for failed. TILBURG UNIVERSITY 79 Example 4. The order quantity for new items is always the same. new spare parts have to be purchased from an external supplier from time to time. • hG : the holding cost per unit per unit of time for good (new and repaired) units. Dagpunar [19] discusses such an optimization problem including a joint family ordering cost. hi . Units cannot be piled. 3. A certain fraction of the failed units cannot be repaired. the same holding cost h and the same volume φ. 8. and the areas occupied per unit are φ1 = φ2 = 1. if possible. The inventory optimization problem with a constraint on the total budget available for investment in inventories can be solved with a similar algorithm as Algorithm 4.20 at Φ = 1800 to $777. Ci = ai Di /Qi + 1 hi Qi .00 at Φ = 3000 via $607.1 Consider two items with the same ordering cost a.1 contains data (ai .89 at Φ = 2400 and $647. The parameters of the model are: • D: the demand per unit of time of units for repair.5) indeed has a root in η on the interval (4.1 is executed. Determine the ratio between the optimal order quantities Q∗ 1 and Q∗ both in the unconstrained and in the constrained case.2 Consider four items with the same ordering cost a = $25 and the same demand rate D = 200. . 60% and 40% of the maximally required storage space of 3000 volume . . Further.HANS BLANC. The available area for storage is Φ = 700 square meters. The total average cost increases from $600. h2 = h3 = $4 per unit per unit of time.1 at storage capacity constraints of 80%. units. i = 1. . it will hold that hG > hF . The table also shows the corresponding ordering and holding cost per item. determine the optimal order quantities for the case that the available area for storage is Φ = 500 square meters. the lead times are deterministic. 2 Exercise 4. . 8) for a set of eight products that have to share storage space.1 further displays the optimal order quantities computed according to Algorithm 4. The demand of units for replacement is assumed to be deterministic and continuous in time. . copiers and printers. 4. i = 1. • LO : the lead time from the external supplier.2 Repairable items This section is devoted to the inventory management of spare parts that occur in machines like computers. Can you find an ad i hoc solution such that the four items are ordered according to their individual economic order quantities? Finally. . Also. .1.

for some integer n. • sO : the reorder point for an external order. the number of incoming units from the external supplier must be equal to the number of outgoing units as waste: QO = wDT. the number of incoming good units must be equal to the number of outgoing good units: QO + nQR = DT. in a cycle of length T . the case will be considered that the lead time of a repair batch is negligible (LR = 0). these balance relations imply that nQR = (1 − w)DT = (1 − w)QO /w. (4. n ≥ 1. (4.8) Together. (4. The policy is such that in between two external orders n repair batches are performed.1: Inventory on hand of good units (top). In a cycle of length T .80 ↑ QO T IG (t) QR OTHER INVENTORY SYSTEMS WITH INTERACTIONS e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e es es e 0 e es es e e e e e e es e es e e e es e T e es e es e e es e e e es e e e e E es e 2T t→ −u T −2u T −u 2T −u T ↑ IF (t)                        QR                                           s s s s s s s s s s s s E  −u 0 T −2u T −u T 2T −u 2T t→ (1−w)QR T +QO ↑ I(t) r rr r r rr r rr rr r r rr r rr r rr r r r rr (1−w)QR rr r rrs rr rr r r rs rr r rs s −u s 0 s s s T −2u s T −u s T s s s s 2T −u s 2T t→ E Figure 4. Let T denote the time between the arrival of two consecutive orders of new spare parts. • QR : the lot size for repairs.7) Further. • sR : the point for starting a repair batch. This puts a constraint on the ratio of QO and QR .9) . of repairable units (middle) and of usable units (bottom). First.

this inventory decreases at rate wD. In between external replenishments. cf.HANS BLANC. . TILBURG UNIVERSITY For the time interval u between two consecutive repair batches within a reorder cycle it must hold that QR = Du.15) or.9) that the fraction of time that the system runs on repaired units is nu nQR = = 1 − w. cf. 2 2 T (4.18) Since the square root function is monotonously increasing. nw (4. This inventory level decreases by QR at instants when the inventory on hand IG (t) of good units reaches the level 0.10) As a consequence. since an external order is succeeded by n repair batches. QR ) = D aO w aR (1 − w) + + 1 hG [wQO + (1 − w)QR ] + 1 hF (1 − w)[QO + QR ]. this inventory increases at rate (1 − w)D. This amount is (1 − w)Du = (1 − w)QR . the average inventory 2 on hand of good units follows with (4. (4.9). it follows with (4. The lower graph of Figure 4. the optimal external order quantity QO is: ˆ QO (n) = w 2nD(aO + naR ) .13) Hence.1 shows the . This inventory level follows a saw-tooth pattern which is lifted above the level 0. The latter is due to the fact that a number of repairable units has accumulated when the inventory on hand IG (t) of good units reaches the level 0 just before the arrival of an external order. The average inventory on hand is 1 QO during 2 a fraction (T − nu)/T of a cycle and is 1 QR during a fraction nu/T of a cycle.14) The foregoing implies that the average cost per unit of time with this policy is: C(QO . (4. the average inventory of failed but repairable units simply follows as 1 ¯ IF = (1 − w)QR + 2 QO − 1 wQO − 1 (1 − w)QR = 1 (1 − w)QR + 1 (1 − w)QO . The average total inventory of good and repairable units on hand is 1 QO 2 plus this basic amount: 1 ¯ ¯ IG + IF = (1 − w)QR + 2 QO . (4.1 (upper graph) shows the inventory on hand IG (t) of good units for the case n = 4. total inventory on hand I(t) = IG (t)+IF (t) of good and repairable units. In between repair batches.1 shows the inventory on hand IF (t) of failed but repairable units.11) Figure 4. nw[hG w + hF (1 − w)] + (1 − w)2 [hG + hF ] (4. 2 2 QO QR (4. n) = (1 − w)2 wD [aO + naR ] + 1 QO hG w + 2 QO nw + hF (1 − w) 1 + 1−w nw . T DT (4. Hence. 81 (4. this inventory decreases at rate D.11) as: ¯ IG = 1 2 QO (T − nu) + 1 QR nu 2 = 1 wQO + 1 (1 − w)QR . (4. the optimal real-valued minimum of this cost function is obtained by solving the equation d ˆ2 dn C (n) d 1 = 2D dn {[ n aO + aR ](1 − w)2 (hG + hF ) + [aO + naR ]w[hG w + hF (1 − w)]} 1 = 2D{− n2 aO (1 − w)2 (hG + hF ) + aR w[hG w + hF (1 − w)]} = 0.10). C(QO . In between replenishments.12) The middle graph of Figure 4.16) For fixed n. 2 2 2 2 (4. This inventory level follows a somewhat irregular saw-tooth pattern. with QR = (1 − w)QO /(nw).17) with corresponding minimum average cost per unit of time ˆ C(n) = 2wD(aO + naR ) hG w + hF (1 − w) + (1 − w)2 (hG + hF ) .

10 for good units and hF = $0. cf.028 + 1 · 0.17). hG QR = 1−w ˆ QO (n∗ ) = n∗ w 2aR D . the above quantities reduce to n∗ = 1−w w aO . The corresponding reorder cycle is T = QO /(wD) = 41. and the corresponding repair lot size is QR = 1−w QO (4) = 41. The ordering cost is aO = $20 and the repair set-up cost is aR = $20. Example 4.55 = 88 ≈ 3.1 Without rounding it follows from (4.17) and (4. and the corresponding repair cycle is u = QR /D = 8. (4. hG + hF The above approximately optimal values indicate that the optimal order quantity hardly depends on aR and that the optimal repair lot size hardly depends on aO and w.9) that ˆ QO (n∗ ) = 2aO wD . cf.352) ≈ 4. Hence. (4. hG that is.17) and (4. cf.18) it is obtained that C(3) = 4 · 40(0.2: Inventory on hand of good units (top) and of repairable units (bottom) with release during repair. (4. aR ˆ QO (n∗ ) = 2aO D .2 The demand for a spare part of a certain type of copier is D = 5 per day.9).8). aR w[hG w + hF (1 − w)] (4. Then.19) Note that this ratio does not depend on the demand rate D. the quantities are approximately given by EOQ formulas independent of the waste rate w. In the extreme case that the holding cost for failed units is negligible.30 days. cf. Next. The holding 5 cost per unit per day is hG = $0. hG w + hF (1 − w) QR = 1−w ˆ QO (n∗ ) = n∗ w 2aR D .352) ≈ 4. Remark 4.817 so that n = 4 is slightly better. This reorder instant can . The fraction of units that is irreparable is w = 1 . A lead time LO for external orders has the standard consequence that an order must be placed LO units of time before the order must arrive.82 ↑ QO T IG (t) QR OTHER INVENTORY SYSTEMS WITH INTERACTIONS e e e e e e e ££e ££e £ e ££e ££e e e e e e £ e e e £ es e£ s e −u ↑ IF (t) 0 ££e e e e e £ e£ es £ e e £ es e£ e £ es e£ £ e e £ es e£ £ e e e es e T e £ es e£ £ e ££e e £ es e£ £ e ££e e £ es e£ £ e ££e e £ e£ es £ e ££e e e e e E es e 2T t→ T −2u T −u 2T −u T            QR                                         s  s s s s s  s s s s s  s E −u 0 T −2u T −u T 2T −u 2T t→ Figure 4.028 7 3 ˆ and C(4) = 5 · 40(0. hF = 0.822 5 0. The corresponding order 4 ˆ ˆ quantity is QO (4) = 41. From (4. 4w (4. consider the consequences of lead times.18) is minimal.52.9). The optimal policy is obtained by rounding n∗ to a positive integer in such a way that the cost (4. the optimal real-valued n minimizing the foregoing cost function is: n∗ = (1 − w) aO (hG + hF ) .19) it follows ˆ that n∗ = 4 0.028 + 1 · 0.55.52 days.10). From (4.52.01 for repairable units. the optimal order quantities follow with (4.

Moreover. the whole graph of the inventory on hand IF (t) for failed but repairable parts in Figure 4.HANS BLANC.30. The ordering cost is aO = $20 and the repair set-up cost is aR = $45. The fraction of units 1 that is irreparable is w = 5 .21) This adaptation is illustrated in Figure 4. Compute the maximum inventory levels of good parts. However. Figure 4. Then. the reorder point in (4.4 to C(4) ≈ 5. Exercise 4.15).52) so that this point has to be applied to the inventory position of the good units.22. cf. the optimal repair lot size and the corresponding reorder cycle. The holding cost per unit per day is hG = $0. QR . Exercise 4.1.20) The consequence of a lead time for a repair batch is more important. This is shown in Figure 4. and the starting point for a repair batch becomes sR = DLR = 50. cf. (4. the inventory patterns shown in Figure 4. the reorder point becomes sO = (1 − w)QR + wDLO . In this case. . the reorder point becomes sO = 78.4 Express the maximum inventory level of repairable units. Assume first that the lead times are negligible.22 and (1−w)QR +QO = 114. cf.20 for repairable parts.20) also has to be increased by (1 − w)DLR to sO = (1 − w)QR + wDLO + (1 − w)DLR .1.2. where the dotted lines in the middle graph indicate the batch of failed items that is in repair.3. The average total cost increases by ˆ hF (1 − w)DLR = 0.16).2 where the upper graph contains the triangular pattern (apart from the external order) of the inventory on hand IG (t) of good units and the lower graph contains the triangular pattern of the inventory on hand IF (t) of failed but repairable units.2. cf. Example 4. the middle graph of Figure 4. the inventory patterns change to a pattern as for the production lot-size model in Section 1. of repairable parts and of all good and repairable parts together. if the repairs are performed away from the warehouse and leave the warehouse at the beginning of the lead time LR . (4. even if we disregard the possibility that the item must compete with other items for repair capacity.3 If the lead time for orders is LO = 5 days and the lead time for repair batches is LR = 10 days in the situation of Example 4.40 for good parts and hF = $0. cf.5 The demand for a spare part of a type of laser printer is D = 1 per day. respectively. In this sense. The starting point for a repair batch lies above the maximum level of the inventory on hand of good units at the end of a repair cycle (QR = 41. This also follows from the fact that LR = 10 > u = 8. except if IG (t) + IF (t) < QR since then an external order has to be placed.6 Adapt the analysis of this section to the situation that the repairs are performed in the warehouse at a rate of r units per unit of time and repaired units become available during the repair run. (4. assuming that the holding cost during the repair lead time LR remains hF . a repair batch has to start when the inventory position of good units reaches the reorder point sR = DLR . Determine the optimal external order quantity.3). and the parameters of the model. modify the results for a lead time LO = 5 days from the external supplier and a lead time LR = 3 days for a repair batch which is transported as a whole to and from a workshop away from the warehouse. It means that at some time intervals. two repair batches are in progress. the number of repairable items that return to the warehouse during the lead time LR . (4. (4. The reorder point lies between the minimum and maximum inventory levels of the total number of usable units on hand ((1−w)QR = 73. in terms of QO . If the repairs are performed close to the warehouse at a rate of r units per unit of time and repaired units become available during the repair run. TILBURG UNIVERSITY 83 only be translated into a unique reorder point for the inventory position of the total number of good and failed but repairable units.217. Exercise 4.21).2.1 has to be shifted upwards by (1 − w)DLR . increases by the amount hF (1 − w)DLR . The average cost per unit of time. see also Figure 4.74.2.

84 OTHER INVENTORY SYSTEMS WITH INTERACTIONS ↑ QO T IG (t) QR sR e e e e e e ec e e e ec e e e ec e e e ec e e e e e e ec e e e ec e e ec e e ec e e e e e e e es −u e e es 0 e e e e e c es c es c es c es e e e es T e e e e e e e c es c es c es c es 2T −u−LR e e e es 2T ec e e c E t→ T −2u T −u               (1−w)DLR          +QR                                        (1−w)DLR  s −u (1−w)DLR T +(1−w)QR +QO ↑ I(t) sO (1−w)DLR +(1−w)QR T ↑ IF (t) s 0 c s c s c s c s s T c s c s c s c s s 2T c E t→ T −2u T −u 2T −u−LR rr r rr rr r rr r rr rr rr rr r r rr rr r rc rr rs rr rs rc rr rr rs (1−w)DLR s −u s 0 s s s c s s T s s s c 2T −LO s s 2T E t→ T −2u T −u Figure 4. of repairable units (middle) and of usable units (bottom).3: Inventory on hand of good units (top). .

∞) then its expected value can alternatively be formulated as: ∞ y ∞ ∞ ∞ E{Y } = 0 0 du fY (y) dy = 0 u fY (y) dy dt = 0 Pr{Y > u} du. 2. . the values which X may assume. variable and fY (y) = FY (y) is called the density of Y . E{X} = ∞ ∞ x ∞ ∞ ∞ x Pr{X = x} = x=0 x=1 n=1 Pr{X = x} = n=1 x=n Pr{X = x} = n=1 Pr{X ≥ n}. that is. If a distribution function FY (y) is continuous for all y ∈ IR and differentiable at almost all y ∈ IR then it is called absolutely continuous.1) The expected value E{Y } is also called the mean or the average of the random variable Y . If a distribution function FX (x) is a jump function then the corresponding random variable X is called a discrete random variable. The function . . . . Let x1 . . the mean of a discrete random variable X with range IN can be written as . . . 2. x2 . The range of a random variable (a subset of IR) is also called the support of its distribution function. The integral of the density over the real axis is equal to one. (A.4) For the special case that Y = X is a random variable with range IN this implies ∞ E{X k } = x=0 xk Pr{X = x}.1. .Appendix A Review of Probability Theory A. y ∈ IR.1 Random variables A random variable is a function defined on a sample space. . k = 1. The (cumulative) distribution function FY (y) of a random variable Y is a nondecreasing function which tends to 0 as y ↓ −∞ and to 1 as y ↑ ∞.3) The kth moment of the distribution of a random variable Y is defined as E{Y k } = ∞ −∞ y k fY (y) dy. . Absolutely continuous distributions are determined by their density. . j = 1. . .1 Moments and expectation ∞ The expected value of the random variable Y with density fY (y) is defined by E{Y } = −∞ y fY (y) dy. (A. be the jump points of FX (x).2) Analogously. Pr{X = xj } = FX (xj +) − FX (xj −). an assignment of a real number to each element of the sample space Ω. . (A. A. that is. . It defines the probability of the event that Y ≤ y. (A. If a random variable Y has range [0. Y is then called a continuous random . k = 1.5) 85 . (A. that is. FY (y) = Pr{Y ≤ y}. is called the probability distribution of the discrete random variable X. 2.

1.15) The concept of independence carries over to finite collections of random variables. Y2 ≤ y2 } = Pr{Y1 ≤ y1 } Pr{Y2 ≤ y2 }. (A. Y2 } = 0. Their joint distribution function Pr{Y1 ≤ y1 .Y2 (y1 . . σ{Y } CY = = E{Y 2 } − E 2 {Y }. Y2 } of two random variables Y1 and Y2 is defined as . If Y1 and Y2 are independent then E{Y1 Y2 } = E{Y1 } E{Y2 }. (A. The density of a pair of continuous random variables is denoted by fY1 . σ 2 {aY1 + bY2 + c} = a2 σ 2 {Y1 } + b2 σ 2 {Y2 }. be a collection of random variables defined on the same sample space. Pr{Y1 ≤ y1 . y2 ).Y2 (y1 . Y2 } of two random variables Y1 and Y2 is defined as .11) (A. cov{Y1 . n. ρ{Y1 .13) The reverse of this statement is not true. Let Yi . (A. in general. i = 1. Y2 } = E{(Y1 − E{Y1 })(Y2 − E{Y2 })} = E{Y1 Y2 } − E{Y1 }E{Y2 }. cov{Y1 . A. Y2 } = ±1 only if there exist constants a and b such that Y2 = aY1 + b. σ 2 {Y } = E{(Y − E{Y })2 } = E{Y 2 } − E 2 {Y }. y1 . (A.8) A. σ{Y1 }σ{Y2 } (A. Y2 ≤ y2 }. The concept of joint distribution function carries over to collections of more than two random variables. σ 2 {aY + b} = a2 σ 2 {Y }. (A. y2 ) dy2 dy1 . E{aY + b} = a E{Y } + b. The coefficient of variation of the random variable Y is defined as 1 . y2 ∈ IR. Y2 } = . (A.86 The variance σ 2 {Y } of a random variable Y is defined as REVIEW OF PROBABILITY THEORY . The mean and the variance of two independent random variables Y1 and Y2 possess the following properties: for any real a. Y2 }| ≤ 1.10) For any pair of random variables Y1 and Y2 . . The cross moment of two random variables Y1 and Y2 is defined as the expectation of the product Y1 Y2 : ∞ ∞ −∞ E{Y1 Y2 } = −∞ y1 y2 fY1 .2 Collections of random variables Let Y1 and Y2 be two random variables defined on a common sample space. Yj }. E 2 {Y } denotes the square of the mean E{Y }.14) (A. Y2 } ρ{Y1 . σ2 i=1 Yi = i=1 σ 2 {Yi } + 2 i=2 j=1 cov{Yi . . E{aY1 + bY2 + c} = a E{Y1 } + b E{Y2 } + c.9) The covariance cov{Y1 . y2 ∈ IR. .7) E{Y } E{Y } The mean and the variance of a random variable possess the following properties: for any real a and b. . Then it generally holds that n n n n n i−1 E i=1 Yi = i=1 E{Yi }. represents the probability that both Y1 ≤ y1 and Y2 ≤ y2 . |ρ{Y1 .12) Two random variables Y1 and Y2 are independent if for all y1 . The (marginal) distribution function of Y1 (Y2 ) is obtained by taking the limit y2 → ∞ (y1 → ∞) of the joint distribution function for fixed y1 (y2 ). (A.6) here. b and c. The square root σ{Y } of the variance is called the standard deviation of the random variable Y . The correlation coefficient ρ{Y1 . cov{Y1 . Moreover.2 Probability distributions This section contains an overview of some useful probability distributions and their properties.

σ (u) du = σNL([y − µ]/σ). Z ∼ (Y − µ)/σ. TILBURG UNIVERSITY 87 A. NL(y) = E{[Z − y]+ } = y ∞ 1 (u − y)f0.7]. σ) distributed random variable Y it holds that E{[Y − y]+ } = y ∞ (u − y)fµ.HANS BLANC.2. − ∞ < y < ∞.σ (u) du = σf0. N (0.2 Exponential distributions Pr{Y ≤ y} = 1 − e−λy .g. σ 2π − ∞ < y < ∞. y ≥ 0. then 2 2 D1 + f D2 ∼ N (µ1 + f µ2 .1 (y) − √ 2π ∞ y e− 2 u du.24) An important property of the exponential distribution is its so called lack of memory. σ2 ) and f is a constant. u ≥ 0. σ) normal distribution then Y ∼ µ + σZ. λ CY = 1 . if the random variable Y represents the duration of some process.σ (y) + (µ − y) y ∞ ∞ fµ. λ σ{Y } = 1 . for σ > 0. − ∞ < y < ∞. and if this process is observed to . 1 2 − ∞ < y < ∞. Sect. E{[Y − y]+ } = σ 2 fµ. for every positive u and y. is called the standard normal distribution. σ) distribution with mean µ and variance σ 2 is. σ1 + f 2 σ2 ). (A.σ (y) = 1 2 1 √ e− 2 [(y−µ)/σ] .19) and by the symmetry of the density of the standard normal distribution.1 ([y − µ]/σ) + (µ − y) (y−µ)/σ f0. This property is expressed by the following formula: Pr{Y > u + y Y > u} = e−λ(u+y) = e−λy = Pr{Y > y}. (A.1 Normal distributions The normal distributions form a two-parameter family of continuous distributions with range (−∞. Alternatively. NL(y) = NL(−y) − y.25) It states that the conditional probability that Y exceeds the value u + y given that Y exceeds the value u is equal to the unconditional probability that Y exceeds the value y. (A. The mean. if D1 ∼ N (µ1 .18) This function has the properties y NL(y) = f0.23) The distribution function of a nonnegative exponentially distributed random variable Y is by definition for some positive rate (parameter) λ. the standard deviation and the coefficient of variation of an exponentially distributed random variable Y are: E{Y } = 1 . 1 2 − ∞ < y < ∞. (A. λ (A. 17.20) For a normal N (µ. D2 ∼ N (µ2 . The normal loss function is defined by . see e. (A.1 (v) dv. (A. The density is λe−λy . The density of a normal N (µ. e−λu y.22) A. fµ.16) The normal distribution with mean 0 and standard deviation 1. Winston [80. it is related to the density and the distribution function (by means of integration by parts): for −∞ < y < ∞. (A.17) An important property of normal distributions is that the sum of independent normally distributed random variables is normally distributed. For example. σ1 ). (A. 1).21) The normal loss function has been tabulated. (A.2. ∞). In other words.1 (u) du = √ 2π ∞ y (u − y) e− 2 u du. If the random variable Z has the standard normal distribution and Y the N (µ.

2.30) e−u ux−1 du. The moments of a gamma distribution are k−1 1 ψ + k − 1 k! k . . n ∈ IN. As a consequence. .λ } = ψ . Pr{min{Y1 . Yn > y} = j=1 Pr{Yj > y} = j=1 e−λj y . . . . ∞).31) For integer-valued n it holds that Γ(n + 1) = n!. λ2 CYψ.26) This property is readily verified by noting that. . for ψ > 0. Its first two moments. . The distribution of the time between successive events is exponential with mean 1/λ. . (A. The future of the process after time u does not depend on its past until time u given the observation that it is still not completed at time u. . t ≥ 0. n! n = 0. n n Pr{min{Y1 . variance and coefficient of variation are E{Xa } = a. (A. λ) distribution with shape parameter ψ and scale parameter λ is.λ } = ψ(1 + ψ) 1+ψ 2 = E {Yψ. (A.28) with a a positive parameter. . exponentially distributed random variables Yj with parameters λj . . Γ(ψ) y > 0. The density of a gamma G(ψ. . Γ(x) = 0 ∞ λψ ψ−1 −λy y e . . . variance and coefficient of variation are E{Yψ. Another important property of the exponential distribution is the fact that the minimum of a number of independent. . Yn } > y} = Pr{Y1 > y.29) A Poisson process N (t).2. Such a random variable is defined by Pr{Xa = n} = an −a e . . is a counting process such that the number of events in a time interval of length t has the Poisson distribution with parameter a = λt. with rate λ.λ (y) = The gamma function Γ(ψ) is defined by . 2. λ > 0. . j = 1. . . the process can be considered as if it started anew at the time of observation u. ψ (A. σ 2 {Xa } = a. .88 REVIEW OF PROBABILITY THEORY be still in progress after some time u then the time until completion has the same distribution as it had at the beginning of the process. that is. .33) E{Yψ. (A. The means of all these distributions are equal to 1.1 for several values of the shape parameter ψ. (A. that is. λ 2 E{Yψ. for y ≥ 0. A. k = 1. 2 E{Xa } = a2 + a. Yn } ≤ y} = 1 − e−(λ1 +···+λn )y . .2. . for any number n of variables. 1.λ } = k (ψ + j) = λ j=0 λk k In particular. (A. (A. . is exponentially distributed with parameter the sum of the parameters of the variables Yj .λ } = ψ .λ = 1 .3 Poisson distributions and Poisson processes A random variable with a Poisson distribution has the set IN of nonnegative integers as its range. .32) The above density has a maximum at y = (ψ − 1)/λ if ψ > 1 and at 0 otherwise. λ = ψ in all cases. n.27) A. the first two moments.λ }.34) The densities of gamma distributions are displayed in Figure A. (A. x > 0. y ≥ 0.4 Gamma and Erlang distributions The gamma distributions form a two-parameter family of continuous distributions with range (0. CXa = 1/a. fYψ. 2 λ ψ σ 2 {Yψ.

exponentially distributed phases.λ (u) du = λψ Γ(ψ) ∞ y (u − y) uψ−1 e−λu du. TILBURG UNIVERSITY 89 1.38) The latter two integrals are known as incomplete gamma functions.HANS BLANC.5 y→ 3 Figure A. (A. the gamma loss function can be expressed after a single integration by parts and with the property of gamma functions that Γ(ψ + 1) = ψΓ(ψ).5 Mixtures of Erlang distributions A mixture of two Erlang distributions. (K1 − 1)! (K2 − 1)! y > 0.λ (u) du. to be denoted by ME(K1 . (A. The gamma loss function is defined by .5 ↑ fYψ.39) .2.λ > y} − y Pr{Yψ. y ≥ 0. (A.λ (y) = y ∞ (u − y) fYψ.35) An important property of the Erlang E(K. j! y ≥ 0.λ (y) = ψ ψ Pr{Yψ+1. Special cases of this class of distributions are the Erlang E(K.37) For general ψ. has five parameters and range (0. j! y ≥ 0. With the aid of repeated integration by parts. p). GLψ. The density of a mixture of two Erlang distributions is f (y) = p λK2 y K2 −1 −λ2 y λK1 y K1 −1 −λ1 y 1 e + (1 − p) 2 e .ψ (y) 8 12 1 4 3/2 0.5 1 2 1/4 1/2 0 0 0.36) For Erlang distributions (with an integer shape parameter Ψ = K) this function can be simplified with the aid of repeated integration by parts to λK GLK. K2 . (A. λ2 .λ (u) du − y y fYψ.1: Densities of gamma distributions with varying shape parameter ψ and fixed mean of 1. (A.5 2 2. This means that a random variable with an E(K. λ1 . λ) distribution is that it emerges as the K-fold convolution of the exponential distribution with parameter λ. for y ≥ 0.λ > y} = λ λ ∞ y ∞ fYψ+1. ∞). the cumulative distribution function of the Erlang distribution with shape parameter K can be explicitly written as K−1 Pr{YK. identically.λ ≤ y} = 1 − j=0 (λy)j −λy e . A. λ) distribution can be considered as the sum of K independent. GLψ. λ) distributions for which K is an integer.5 1 1. as.λ (y) = (K − 1)! ∞ y K−1 (u − y) u K−1 −λu e du = e −λy j=0 (K − j) λj−1 y j .

λ2 λ2 1 2 (A. λ1 and λ2 are positive scale parameters.40) K1 (K1 + 1) K2 (K2 + 1) + (1 − p) − E 2 {Y }. (A. (A. If the coefficient of variation C < 1. λ2 = − λ1 . (A.46) . j! y ≥ 0. y ≥ 0. K2 = K + 1 and λ1 = λ2 = λ this loss function becomes   K−1 K j−1 j j−1 j λ y λ y  (K − j) E{[Y − y]+ } = e−λy  .90 REVIEW OF PROBABILITY THEORY here. Such mixtures are called hyperexponential distributions and are indicated by the symbol H2 . λ1 = λ2 = (K2 − p)/E{Y }. The mean of a random variable Y with a mixture of two Erlang distributions is E{Y } = p and its variance is σ 2 {Y } = p K1 K2 + (1 − p) . λ1 = λ2 and p are determined in such a way that the mean and the variance agree and the third moment is equal to that of a gamma distribution:   C2 − 1 2  4 λ1 (1 − λ2 E{Y }) 2  λ1 = 1+ . then the parameters are chosen as follows: K1 is the largest integer smaller than 1/C 2 . a distinction is made whether the coefficient of variation C = σ{Y }/E{Y } is smaller than 1 or not. λ1 λ2 (A. + (1 − p) j! j! j=0 j=0 The loss function corresponding to a hyperexponential distribution (K1 = K2 = 1) is still simpler: E{[Y − y]+ } = p −λ1 y 1 − p −λ2 y e + e . (A.41) Mixtures of two Erlang distributions are often used to approximate the distributions of random variables with range (0.44) For the special case K1 = K.45) (A. The loss function corresponding to a mixture of two Erlang distributions simply follows from (A. ∞) of which only the mean E{Y } and the variance σ 2 {Y } are known.42) 1 + C2 If the coefficient of variation C > 1. p= . K1 and K2 are positive integers. 0 < p < 1. and p. K2 = K1 + 1.37) as K1 −1 E{[Y − y]+ } = p e−λ1 y j=0 (K1 − j) λj−1 y j 1 + (1 − p) e−λ2 y j! K2 −1 (K2 − j) j=0 λj−1 y j 2 . For the choice of the parameters. and λ1 = λ2 and p are determined such that the mean and the variance agree: 1 2 p= K2 C 2 − K2 (1 + C 2 ) − K2 C 2 .43) E{Y } 1 + C2 E{Y } λ1 − λ2 Mixtures with K1 = K2 = 1 are mixtures of two exponential distributions. is a weight. then the parameters are chosen as follows: K1 = K2 = 1. λ1 λ2 y ≥ 0.

1 Stationary points g(x) = a + bx. z 2 . c = a + z 2 (b − a) and d = b − z 2 (b − a) = a + z(b − a). x a + b. d). respectively. (B.1) In inventory theory objective functions of the following form are often encountered: here. and the procedure is repeated for the interval (c. Example B.2 Golden section The “Golden-Section” method is an interval elimination method for finding the minimum x∗ of a function g(x) with no local minima beside the global minimum. the interval (a. x = x∗ = The second derivative of g(x) at x∗ is positive: g (x) = 2a x3 √ 2b b ⇒ g (x∗ ) = √ . x 20 x > 0. d) and (d. the procedure continues with intervals which are a factor z ≈ 0. 1 − 2z 2 . 91 . the interval (d.1 shows the results of the golden section procedure with and without rounding to integer order quantities.618. If g(c) < g(d). Table B.4) a/b.5. (B. √ (c. Hence.618 smaller with each step. b) is divided into three subintervals (a. c).382. This implies 2 √ 1 that z 2 = 2 (3 − 5) = 1 − z ≈ 0.5) B. x2 x > 0. with z = 1 ( 5 − 1) ≈ 0. This method assumes that g(x) is strictly decreasing for x < x∗ and strictly increasing for x > x∗ . b) is discarded. The lengths of the new intervals are d − a = b − c = z(b − a).5 = 50.2) Hence. As initial points we take a = 10 and b = 1000 where one of the terms of g(x) dominates the other: g(10) = g(1000) = 50 + 0. until the minimum has been found with sufficient accuracy. (B. a √ a/b = 2 ab. Note the flatness of the objective function near its minimum. b) with relative lengths z 2 .1): g(x) = 500 x + . (B. b). otherwise. c) is discarded. An initial interval (a. and the procedure is repeated for the interval (a. the function g(x) has a minimum at x∗ with the value g(x∗ ) = a b/a + b (B. a and b are positive constants.Appendix B Optimization methods B.3) x > 0. The derivative of this function is g (x) = − This derivative has a unique positive zero at . The procedure can be continued to end at the optimum x = 100.1 Consider an objective function of the form (B.

154.96 10. interval (10.75) (71.08 98. 103) (96.00 10. 243.15) (10.50) (10. b) which should contain x∗ .24 10.17 10.81 87.24 10.00 10.37 g(c) 20. b).2: Results of bisection procedure for the function of Example B.00 10. the sign of the function at the middle point is considered. The derivative g (x) is evaluated at the middle point c = 1 (a + b). If g (c) > 0.50 133.88.00 10.91.00 10. This method employs information about the value of the derivative of the function g(x).95 100.34) (86.00 10.24. 107. this procedure converges faster than the golden section procedure in Table B.11 10.00 Rem.00.69 14.0468 0.00 10.3 Bisection The method of interval bisection is an alternative adaptive search method for finding the minimum x∗ of a function g(x) with no local minima beside the global minimum. As initial points we again take a = 10 and b = 1000 where the derivative has opposite sign: g (10) = −4.00 10.27 g(d) 31.0053 −0. 2 x 20 x > 0. 244) (10.88 102.34. and the procedure 2 continues with the interval (a.27 102.34 95. the right half of the interval is discarded. 388. The method of interval bisection can also be used to search a root of an equation or the zero of a function. 102.2 shows the results of the bisection procedure with and without rounding. the interval (c.37.27 120. b) is discarded. 133.0027 −0.34) (97. This function has derivative: g (x) = − 1 500 + .24 99. 134) (72. 133. 622) (10.34 99.0220 −0.00 Point d 621.00 10.0011 0.29.27 94.0425 0.1. It assumes that g (x) is negative for x < x∗ and positive for x > x∗ . 120) (86.71) (10.17 99. 120) (86.90 20.0001 Remaining interval (10.00.17 10.15 243. Point c 388 244 154 99 65 99 86 99 94 99 97 g(c) 20. However.1: Results of golden section procedure for the function of Example B.34 99. 103) (88.00.34) (86.70 14.29 99. 120.17. 621.71 154.85 388.0425 0.00.85) (10. 107) (94.02 10. Table B.1.24. 154) (65. 107. With the same initial interval.88 99. 102) (97.00 10. and the procedure continues with the interval (c. .00.00 10.2 Consider again the objective function of Example B. 100.96 10. the computation of the derivative g (x) may be more involved than that of g(x). 107) (94.75 71. 154) (65.02 10.0029 −0.00.32) (94. then the left half of the interval is discarded. 103) Middle point c 505.93 10.00 10.0495.00) (10.70 14. the interval (a.27 86. Clearly.81) (98.9500.25 10. 100. 102.0000 Remaining interval (10.00 Point d 622 388 244 154 99 120 99 107 99 102 99 g(d) 31.75) (71. If g (c) < 0. g (1000) = 0.44 99. In this case.25 10. otherwise.00 Remaining interval (10.32) (94.0146 −0.0480 0. 102. 505. 134) (72. 154.44 99.02 10. 388) (10.08. Each application of the method halves the length of the interval in which the minimum x∗ must lie. 102.69 14.00 10. c).11 10.0155 −0.95 10.91 g (c) 0.00 10.88) Example B.27 65. 257.15 243. 505) (10.88.71 154.88) (99.0043 0. Table B.17.44) (65.81) (95. The method starts with an initial interval (a. c) is discarded. 102.00 257.94 10. 102. To find a root of a function.0009 −0.92 OPTIMIZATION METHODS Table B. 102) Point c 388.1.0222 −0.27 97.00 10.81) (87. the method of interval bisection is faster than the Golden-Section method in terms of the number of iterations.95.95 10.34) B.29.44) (65. If this sign is the same as that at the left end point.32 99.0480 0. it is assumed that the function is continuous and that it possesses a single zero on the initial search interval so that the sign of the function at the two end points is different.81) (98. 120. 258) (10.27 107.02 10.00 10.00.90 20.2.0465 0.95. Middle point c 505 258 134 72 103 88 96 100 g (c) 0.

O. H. Oper. 243–250.. 45 (1990).S. A. 199–221. [9] Baker. [20] Das. K. R. J. Management Science 30 (1984).K. Iyogun. A. Oper.. Operations Research Letters 6 (1987). [6] Axs¨ter. M. Millen. A. 13–23. Management 9 (1989).R. [21] De Kok. 63–73.K. J. 55–69. An optimal heuristic for coordinated multi-item inventory replenishments.). P. [19] Dagpunar. 369–373. P. A joint replenishment policy for multi-echelon inventory control. A lower bound for a class of inventory/production problems. D. IIE Transactions 20 (1988). Management Science 10 (1964). A. 2006 (2nd ed.. 37 (1986).. Oper. J. 36 (1985). Zhang. Springer. a [8] Axs¨ter. Res. A heuristic with lower bound performance guarantee for the multi-product dynamic lot-size problem. J. 32 (1981). Management Science 6 (1960). [10] Ballintfy. S. Optimal policies for a multi-echelon inventory problem. H. [3] Atkins. Emmons. [2] Andres. Int. J. D. Management Science 21 (1975). [7] Axs¨ter. A.. [17] Chakravarty. [16] Chakravarty. Scarf.. Inventory Control. [14] Chakravarty. 227–231. Y.. Engineering Costs and Production Economics 21 (1991). [15] Chakravarty. 19–26.. 475–490. Millen. 287–297.R. Exact analysis of continuous review (R. Lot-sizing procedures and a standard date set: a reconciliation of the literature. D. Res. 1113– 1120. C. 791–796. J. J. Engineering Costs and Production Economics 10 (1986). A multi-product inventory system with interactive set-up costs. Hierarchical production planning for consumer goods. 93 . Improved heuristics for multi-stage requirements planning systems. J. Res. S. S. 44–56. S. Goyal.L. 1027–1039. 77–81. 1055–1063. Operations Research 48 (2000).R.K.Q) policies in two-echelon inventory systems with a compound Poisson demand.F. Res. Heuristics lot-sizing performance in a rolling-schedule environment.G. Multi-item inventory grouping with dependent set-up cost and group overhead cost. W. [12] Blackburn. S. Iyogun. Oper. Oper. Periodic versus ‘can-order’ policies for coordinated multi-item inventory systems.K.R. Int. Eur. Int. Production Management 8 (1988). a Production Economics 59 (1999).O.K. Manufacturing and Operations Management 2 (1989). Management Science 28 (1982). Goyal. R. 691–701. Decision Sciences 11 (1980). Erenguc. Prod. Joint replenishment in multi-item inventory systems. J.K. [18] Clark. Management Science 34 (1988). [13] Blackburn. 686–696.J. Iyogun.. On a basic class of inventory systems. Economic ordering policy for deterministic two-echelon distribution systems. J. S. A.Bibliography [1] Aksoy. [5] Atkins. Soc. F. Multi-item inventory models with coordinated replenishments: a survey. J. [4] Atkins. On the relationship between constrained and unconstrained multi-item replenishment problems. P.. Soc. [11] Bastian. Joint inventory replenishments with group discounts based on invoice value. Multi-item inventory aggregation into groups.O.. New York. 63–67. J. 1105–1112. Soc.

E. W. On a simple method of determining order quantities in joint replenishment under deterministic demand. Management Science 25 (1979). 53–62. Graves.. On the deterministic demand multi-product single-machine lot scheduling problem.94 BIBLIOGRAPHY [22] DeMatteis. [29] Federgruen..C. 7 (1968). S. Production Economics 35 (1994). 321–329. P. Res. S. J. 1993. G. [42] Jackson.. Ratliff. A. Coordinated replenishment in a multi-item inventory system with compound Poisson demands. M. [24] Eppen. IIE Transactions 30 (1998). An efficient algorithm for computing optimal (s.M. IBM Syst. [40] Inderfurth. Materials Handling. Schrage. Centralized ordering policies in a multi-warehouse system with lead times and random demand. The joint replenishment problem: new heuristic and worst case performance bounds.) Multi-level Production-Inventory Control Systems: Theory and Practice. 279–289. Blackstone. H. [31] Goetschalckx. IIE Transactions 20 (1988).B. Rosenblatt. [37] Graves.K.R. Management Science 21 (1974). 2–13. Sequencing picking operations in a man-aboard order picking system. Satir. Heuristic methods for the multi-product dynamic lot size problem. J. Soc. Optimum ordering policy for a multi-item single supplier system. Belton.T.C. Zipkin (eds. Marcel Dekker. Management Science 30 (1984). W. Material Flow 4 (1988). 1025–1033. Optimal storage assignment in automatic warehousing systems. NorthHolland..D.. An economic lot-sizing technique I: the part-period algorithm. 889–894. A. [41] Iyogun. Zipkin.G. [25] Ettl. 711–723. 604. Kropp. L. Amsterdam. M. [35] Goyal. The joint replenishment problem with a powers-of-two restriction. Oper. D. M. Co. A. 293–298. IIE Transactions 15 (1983).D. Operations Research 38 (1990).K. G.H. . P. Hoffmann. D. A multi-product dynamic lot-size model with individual and joint set-up costs. Management Science 22 (1976). [30] Fogarty. [39] Ignall. M.D.H.J. Periodic review and joint replenishment in stochastic demand environments. Maxwell. New York.C... S. A.K. in: L. P. Order picking in an aisle. Ratliff.H. 344–357. L. Groenevelt. A. 277–279.C. [43] Joneja. Management Science 25 (1979). Joint replenishment inventory control: deterministic and stochastic models. Muckstadt. 30–38. Cincinnati. Optimal continuous review policies for two product inventory systems with joint set-up cost. 1991. 436–443. P. Yao. 216–232.S. 276–280. Eur. A. [45] Kaspi. S. Graves. North-Holland. [34] Goyal. [38] Hausman. 255–263. T. Centralized planning models for multi-echelon inventory systems under uncertainty. Res. S. J. J.) Logistics of Production and Inventory. [27] Federgruen. [28] Federgruen. 2nd ed. Operational Research Quarterly 25 (1974). Operations Research 32 (1984). R. Schwarz. IIE Transactions 17 (1985). Int. D. Rinnooy Kan.O.W.Y. D. 629–638. 1981. K.E. S. [32] Goetschalckx.. [26] Eynan. M. J. P. in: S. 25–32. Oper.. Lin.K. 38 (1989). [23] Eastman. 1987.. Determination of optimum packaging frequency of items jointly replenished. Ohio. [33] Goyal.. Amsterdam.. An improvement of Silver’s algorithm for the joint replenishment problem. [44] Kao. Feigin. 1268–1285. Production & Inventory Management.S) policies.H. Safety stocks in multistage inventory systems: a survey. A... G.B. H.H. Management Science 15 (1967). Tijms.C. H. Schwarz (ed. H. A supply network model with base-stock control and service requirements. South-Western Publ. J. [36] Goyal. Operations Research 27 (1979). 264–267.. 42 (1991). J. Operations Research 48 (2000).

G. 532–549.. Coordinated replenishment in inventory systems with correlated demands. Yuan. P. 289. A. [54] Pantumsinchai. J.A. 507–516. 7907. Pyke. J.. New York. Meal. [51] McGee. An economic lot-sizing technique II: mathematical analysis of the part-period algorithm. House.J. Decision Science 23 (1992). Naval Research Logistics Quarterly 14 (1967). Mc Graw Hill.).A. J. 1994.G. Establishing reorder points in the (S. X. 743–750.. E. [58] Silver. J. T.A. M. 1984. Rosenthal. A simple method of determining order quantities in joint replenishment under deterministic demand. TILBURG UNIVERSITY 95 [46] Korugan..M. Computers & Industrial Engineering 27 (1994). Oper.E. [48] Lee. A. [52] Mendoza. J. 12 (1974). Wiley. E. 490–503. L.F. Res. [61] Silver.A. [63] Silver. D.s) coordinated control system under compound Poisson demand. M.M. 141–151. u . J.A. INFOR 22 (1984). J. D. Eur. J. D. New York. A global optimum search algorithm for the joint replenishment problem under power-of-two policy. R. Eur. IBM Syst.. M. 647–671. [57] Schrady. Soc. vander Eecken. [62] Silver. [66] Tomkins... Int.. Heuts.. F. Naval Research Logistics 26 (1979). 101 (1997). White. 201–204. Wiley. F. S. Facilities Planning. [70] Van Eijs.G. Masui. Miltenburg. A deterministic inventory model for repairable items. A control system for coordinated inventory replenishment. Two modifications of the Silver-Meal lot sizing heuristic. Wiley. A.A.U. 111–127.HANS BLANC. Multi-echelon inventory production system solution. The Distribution Handbook. 1979. Coordinated replenishment systems with discount opportunities. Oper.. [59] Silver. M.S. 1319–1333. Near cost-optimal inventory control policies for divergent networks under fill rate constraints. [56] Robeson. A note on the joint replenishment problem under constant demand.J.C. Gupta. 65–87. 1351–1361. Production Research 32 (1994).J. Int. Inventory Control. Pyke. 391–398. J.G. 1985.A. T. Res.F. 1979. E. V. Production Economics 46–47 (1996). Supply rationing in multi-echelon divergent systems. On the joint replenishment problem under deterministic time varying demand patterns. S. [53] Odanaka.F. [69] Van der Heijden. E. E. M. J.J. Onderzoeksrapport nr. 185–192. M.C.-C. H. Leuven. Res. [71] Van Eijs.A. E. Van Eijs. Production Economics 63 (2000). K. K. 1998 (3rd ed.. Yao. Stochastic Models: an algorithmic approach. 123 (2000). J. Int.c. Chichester. 7 (1968).. M. 64–74. A heuristic for selecting lot size quantities for the case of a deterministic timevarying demand rate and discrete opportunities for replenishment.D. Order picking in a rectangular warehouse: a solvable case of the traveling salesman problem. 44 (1993). Periodic production scheduling at a fastener manufacturer. [64] Silver. A multi-echelon inventory system with returns. [49] Love. [55] Ratliff.-J. 161–179.C. H. Oper. Peterson. E. [47] Lambrecht. J. Management Science 22 (1976). H. New York. 2879–2896. 14 (1973). [50] Liu. 39–46. [60] Silver.. 56–68. Yamaguchi. H. The Free Press. New York. A comparison of three joint ordering inventory policies. R. R. Production Res. [68] Van der Heijden. [65] Tijms. Vanderveken. Zeitschrift f¨r Operations Research 39 (1994). Inventory Management and Production Planning and Scheduling.M. Computers & Operations research 30 (2003). J. Computers & Industrial Engineering 35 (1998). Coordinated replenishment of items under time-varying demand: dynamic programming formulation.R. Operations Research 31 (1983). Production Res. [67] Van der Duyn Schouten. Int. J. On the determination of the control parameters of the optimal can-order policy. Production and Inventory Management. 145–148. 19 (1981). Int.

S) policies for joint replenishment inventory systems. .G. PWS-Kent Publ. A. product popularity and the location of stock in a warehouse.C.J.C. 99 (1997). 402–423. J. Dekker.E. 341–354. W. Kim. 89–96. R. 3rd ed. [75] Verrijdt. J. R. Res. T. [81] Yoo. Management Science 43 (1997). [82] Zheng.M.M. Oper. J.C..H. Heuts. 433–444. Management Science 5 (1958). An efficient optimal solution method for the joint replenishment problem. 59 (1992). J. Operations Research: Applications and Algorithms.. Frenk. M. Rhee. 285–292.C. Math. S. Eur.H. Eur.P.M. On the imbalance of inventories in multi-echelon systems.96 BIBLIOGRAPHY [72] Van Eijs. Finding optimal (s. Y.G. 230–237.. 654–665 (corrections: Operations Research 40 (1992). 729–732. [73] Van Eijs. Analysis and comparison of two strategies for multiitem inventory systems with joint replenishment costs.G. Res. A. [77] Wagner. Dynamic version of the economic lot size model. [80] Winston. 9 (1984). Kleijnen.. [79] Wilson. Boston. [83] Zipkin.B. Efficient inventory management in multi-echelon distribution systems.M. [78] Wildeman. J.. 1994. J.. Computers & Industrial Engineering 33 (1997). Production Economics 38 (1995). H.G. Operations Research 39 (1991). Order quantity. Res. 192). 89 (1996). W.. Co. De Kok Distribution planning for a divergent N -echelon network without intermediate stocks under service restrictions. Int. [76] Viswanathan. Multi-item inventory systems with joint ordering and transportation decisions. 1447–1454. Periodic review (s.. De Kok Distribution planning for a divergent depotless two-echelon network under service restrictions. Oper. Res. 405–412. Within.S) policies is as simple as evaluating a single policy. M.G.. Int. Production Economics 35 (1994).J.J. Eur.L. Y.S. J. Federgruen. IIE Transactions 9 (1977). Oper. R. J. Oper. A. 225–243. H.G. [74] Verrijdt. J. J. M.M.

9. 15. 12. see distribution center Coefficient method (heuristic). 69 Bastian heuristic. 17. 12–17. 47. 54 (R. 13. 70 Economic order quantity. 59. 69 net. 44. 59. 51. 2. 44 Converging chain. 63. 20. 59. 72. 21. 16. 70. 58 Bisection. 48. 9. 72. 12. 60. 69 Dynamic programming. S) policy. 43 Family ordering cost. 20 . 12. 15. 12. 43–54. 21–26. 17. Q) policy. 43. 82 EDI. 4. 58–66. s. joint Family reorder cycle. 70 during lead time plus review period. 45. 69 Balance equation. 30. 13. 6 Feasible. 67 Conveyor. 51. see objective function Cost-covering heuristic. 9. 3. 12. see electronic data interchange Electronic data interchange. 69 Cycle service constraint. 20. 12. 86 Cost function. 89 Exponential smoothing. 77. 57. 66. 85. S) policy. 12. see demand. 89 Exponential distribution. 42 Covariance. see demand. 78. 86. 70 position. 35. 57 Cross-docking. 74 Constant demand. 77. S) policy. 9. 30. 90 Complementarity effect. 45 Balanced Stock rationing. 44–51. 59 rate. 3. 31–36. 7. k. 72. 44 during lead time. 7. 1. 15. 9 Density. 33 Batching. 69 Distribution function. 32 Consistent Appropriate Share rationing. 87. 9–12. 54 Carrying charge. 37 Echelon inventory. 75. 61. 88.Index (R. 20. 12. 66 constant. 64. c. 44. 1. 87 Direct grouping. 12. correlated Correlation coefficient. 69–76 time varying. 71 forecasting. 36–43. S) policy. 30 factor. 44. 34. 62. 15 (s. 66. 69 target. 40 Coefficient of variation. 56. 7. 85 Distribution network. 20–36. 4. 59. 20–54 Correlated demand. 56. 12 ABC classification. 2. 73 Base-stock policy. 79 Can-order level. 12. 54. 79 Delivery instant. 15. 59 on hand. 48. 50 Decision variable. 4. 17 Consecutiveness property. 85. 65 Fast mover. 6. 31. 2. 85. 56. 69 97 Demand. 63. 1. 14. see ordering cost. 9 stochastic. 72. 73. 19 EOQ. 4. 17. 15. 70. 86 Crane. 56 Diverging chain. S) policy. 3. 34 Distribution center. 9. 59 Assembly. 7. 77. 64. 54 (R. 49. 48. 79. 16. 2. 44 Discount. 35. 12 (s. 67. 15. 17. 53. 77–83 correlated. see economic order quantity Erlang distribution. 60 Central depot. 3. 6. 44. 18 Fluid approximation. 67 Atkins & Iyogun heuristic. 87 absolutely continuous. 49. 69 (R. s. 62. 9. 54 (s. 73 Flexibility. 3. 52. 67 Budget constraint. 73. 59 Family of items. 2. 67 acyclic pattern. 64. 10. 6. 15. 3. 79 seasonal pattern. 32. 57 Coordinated replenishment. 44 Can-order policy. 58 Action space. 52. 56. 37 Advertisement. constant Continuous review. 12. 4–9. 34 all-units. 60. 1–3. 92 Blackburn & Millen heuristic. 20. 63. 21. 52. 52 Backlog. 1. 66–68 trend. 66. 56. 28 Fill rate actual. S) policy. 47. 73 constraint. 20.

24. 77. 36.98 Forecasting. 24. 61. 28 Labor cost. 77 Public warehousing. 43. 4. 29 Price break point. 79 mean. 6. 48. 67 rolling. 2. 59 Interaction in inventory management. 59 purchasing. 69 . 51. 18 joint. 20. 57 Material requirement planning. 52. 44 Group reorder cycle. 30. 6. 88 Multi-echelon system. 64. see fill rate constraint Part-period-balancing heuristic. 35. 49. 90 Imbalance. 20. 19. 44. 1. 79 net. 58–64. 65 Information exchange between echelons. 48. 90 Mixture of Erlang distributions. 74 management. joint Production Joint replenishment. 57. 22. 4. 17. 26. 12. 28 cycle. 30. 61. 54. 2. 19. 46. 65 Goyal iterative heuristic. 34 Handling cost. 69. 53. 83 theory. 24. 13. forecasting Gamma distribution. see cycle service constraint P2 -criterion. 23. 1. 40 Penalty cost. 7. 86 Joint ordering cost. 64. 85. 70. 87 Number of orders per unit of time. 44. 6. 6. 36 Group of items. 20. 7. 1 Holding cost. 61. 12. see inventory. 45 Material cost. 26. 12. 67–68 Push policy. 28 Lagrange multiplier. 1–3 Lot size. 44. 15. see inventory. 7. 88 Powers-of-two policy. 73 Ordering cost. 30. 69 Physical stock. 73. 6. 25. 48. 83 Lost sales. 15. 15. 64. lot size INDEX Management cost. 3 Periodic review. 1 capacity. 71. 25. 60. 17. 17. 2. 3. 16. 91 Inventory level. 50. 48. 73. 81. 18 Handling time. 4. 49. 38. 39 Hyperexponential distribution. 87. 9. 21. 88 Poisson process. 36. 44. 3 P1 -criterion. 4. 63. 87 schedule. 21 Marginal distribution. 79. 26. 67. 67. 2. 6. 28 Kaspi & Rosenblatt heuristic. 49. Pull policy. 3. 21–31. 58. 77 cost. 34. 67. 3. 72. 70. 79 supplementary. see inventory. 11. 73 Independence of random variables. 18. see production. 52–54. 88. 70. 52. 7. 22. 1. 63. 83 saw-tooth pattern. 81 step pattern. 12. 9. 4. 63. 72 Order-up-to level. 7. 89 Golden section. 60. 44. 63 triangular pattern. 58. 58. 3. 62. 43. 29. 34. 89 Gamma loss function. 5. 4. 18. 59. 15. 47. 7. 47. 45. 54. 81 policy. 20. 57 Order quantity. 73. 79. 65. 47. 56. 45. 35. 69. 2. 43. 51. 70. 12. 20. 16. 55–76 oscillation effect. 65 Ordering rule. 63. 82 cumulative. 30. 15. 36. 64. 53 Goyal algorithm. 73. 86 Markov process. 2. 30. 58. 47. 8. 2. 7. 6. on hand Poisson distribution. 2. 3. 31. see coordinated replenishment capacity. 15. 43. 40. 91 Order picking. 44. 48. 60. 77. 72. 82 Purchasing cost. 20. 28 Material handling. 53 Joint distribution. 85. 89 Moment of a distribution. 21. 58 Net stock. 36. 86 Indirect grouping. 69. 77. 7. 70. 73. 65. 87 Normal loss function. 18. 35. 77. 67. 81 maximum. 30. 12. 16. 77 Inventory. see ordering cost. 10. 9 Mean of a random variable. net Normal distribution. 16. 60. policy Overtime work. 71. 40 Holidays. 60 Linear chain. 77. 36 average. 28 rate. 1. 20. 16. 59 Horizon. 70. 43. 12. 85. 65 lot size. 52. 46 Objective function. 53. 34 change. 5. 53. 36. 52. 53. 6. 11. 28 Lack of memory. 14 on hand. 4. 12. 39. 59. 6. 4. 73. 15. 73. 13. 12. 86 Probability distribution. 16. 48 in transit. 7. 43. 11. 61. 91 Goyal & Belton heuristic. see demand. 59. 69 position. 79 joint. 47. 53. 57 Lead time. 66 66. 88 Gamma function. 71.

83 echelon. 15. 69. 51. 79 Zoning. 1. 12. 70. 43 Spare part. 76 Set-up cost. 17 Silver dynamic programming algorithm. 74 Sensitivity analysis. 58 Waste. 39 Silver-Meal heuristic. 3.HANS BLANC. 15. 87 State space. 6. 83 Special offer. 79 lot size. 12. 45 Stochastic demand. 71. 66–68 Warehouse. 26. 11. 22. 58. 73. 3. 37. 47. 57 Review instant. 14. 15. 73. 83 Repairs. 85 continuous. 1. 85. 79. 47. 66 Repair batch. 79 family. 75 Reference item. 28 Set-up time. 58 Rush order. 9. 47. 59 Standard deviation. 52. 2. capacity Substitution effect. 70 Rounding off. 55. 12. see inventory. 2. 44. 24. 71 joint. 6. 3. 39–40. 15. 66. 4. 2. 57. 3. 58 99 . 82 Reorder frequency. synchronization Target fill rate. 85 discrete. 72 cost. 63. 28 family. 12. 70. 12. 57. 60. target Time varying demand. 59 Quantity discount. 48. 71. 59. 30. 82 rate. 90 Vehicle. 79. 52. 9. 43. 18. 65 Reorder cycle. 3. 86. 48. 12. 55 Synchronization. 60. 57 Wagner-Whitin algorithm. 38 Silver heuristic. 86. 28 item. 6. see demand. 2. see demand. 79 Returns. 19. 67–69 Slow mover. 5. 18. 49. 12. 82 Routing. 2. 69. 55. 44. time varying Value Added Logistics. 15. 58. 59 planned. 21. 86 Rationing fraction. 60 Variance. 40. 43. 80. 57. 80. 2. 15. 72. 44. 6. 17 Supply chain. 64. 12. 21. 3 Safety stock. 85 Range of a random variable. 12 Service level constraint. 30. 82 synchronization. 9. 28 Shipping cost. 62. see discount Random variable. 18. 66 Reorder point. 6. 44 Review period. 53. 17. 15. TILBURG UNIVERSITY Quality control. 85. 63. 6. 28 item. 14 Storage capacity. see reorder instant. see fill rate. 70. stochastic Stockout. 10. 38. 51. 3. 73. 56 Value of a good. 12. 15. 26. 65 Reorder instant.

. .) cost function CY coefficient of variation of r.) E{. . Sect.”.} standard deviation of a r.List of symbols This list only contains the most important notations. Some symbols may locally have a different meaning. . .v.1) N number of stocking points at an echelon (Ch. .v. . .v. Y d discount factor D demand per unit of time (fixed or r. .v.} 1 vector of ones a individual item ordering cost A family ordering cost B backlog per cycle c can-order level c vector of can-order levels C(.v.) M number of orders per unit of time 100 .v. τ production time φ volume of items Φ inventory capacity Ψ actual fill rate is max{0. . G set of items forming a group I{.v.) DL demand during lead time (r.} indicator function h holding cost H planning horizon I inventory level IO inventory level at reorder instant k reorder frequency k vector of reorder frequencies L lead time (fixed or r.v.) DL+R demand during lead time and review period (r. 4. 3) Q order quantity (fixed or r. .} mean of a r. = is by definition ≈ is approximately equal to [.v. Random variables are indicated by “r. ] + n ratio between order quantities N number of items in a family (Ch. 2. .) p production rate r carrying charge R review period s reorder point s vector of reorder points S order-up-to level S vector of order-up-to levels T reorder cycle u set-up time U family set-up time v purchasing price V price break point w waste rate α cycle service level β target fill rate Π actual probability of no backlog σ{.

Sign up to vote on this title
UsefulNot useful