S.C. Graves et al., Eds., Handbooks in OR & MS, Vol.

4
© 1993 Elsevier Science Publishers B.V. All rights reserved.

Chapter 1

Single-Product, Single-Location Models

Hau L. Lee
Department o f lndustrial Eng., Stanford University, Stans~brd, CA 94305-4024, U.S.A.

Steven Nahmias
Decision and Information Sciences, Santa Clara University, Santa Clara, CA 9505.3,
U.S.A.

1. Introduction

Inventory management is one area in which operations research has had a

significant impact. At the present time, the total value of inventories in the
United States is close to one trillion dollars (Economic Report of the Presi-
dent, February, 1989). Mathematical models form the basis for many of the
inventory control systems in use today, whether they be inventories of raw
materials, spare parts, cash, or finished goods.
Most mathematical inventory models are designed to address two fundamen-
tal issues: when should a replenishment order be placed, and how much should
the order quantity be. The complexity of the model depends on the assump-
tions that one makes about the demand, cost structure, and physical charac-
teristics of the system. In some cases, the operating characteristics of a given
policy are analyzed.
The objective of virtually all inventory control models is to minimize costs.
In most cases, minimizing costs will result in the same control policy as that
obtained by maximizing profits. When uncertainty is present the traditional
approach has been to minimize expected costs. The use of the expected value
operator is justified from the law of large numbers: as the number of planning
periods grows, the arithmetic average of the actual costs incurred will be close
to the expected costs. Since most inventory control problems are ongoing, this
approach makes sense.
Most inventory control problems in the real world involve multiple products.
For example, spare parts systems for military applications may require manage-
ment of hundreds of thousands of different items. However, it is offen true that
single product models are able to capture the essential elements of the
problem, so that it is not necessary to explicitly include the interaction of
different items into the formulation. Furthermore, multiple product models are
4 H.L. Lee, S. Nahmias

orten too unwieldy to be of much use when the number of products is very
large. For this reason, single product models dominate the literature and are
used most frequently in practice.
This chapter is concerned with mathematical models for controlling the
inventory of a single product. Other chapters in this volume will be concerned
with multiple products and special topics.

1.1. History

Although operations research methodology is based on mathematical de-

vel0Pments spanning centuries, rauch of the interest in the application of this
methodology had its origins in the logistic support activities which took place
during the Second World War. Interdisciplinary teams of scientists would
brainstorm problems that were not amenable to traditional analysis. The
success of these activities, along with the subsequent discovery of both the
simplex method and the electronic computer led to the birth of operations
research as a discipline. It is interesting to note that the first work on inventory
modeling pre-dates these developments by about thirty years.
Ford Harris [1913] is generally credited with the discovery of the original
E O Q (Economic Order Quantity, to be described in Section 2) model. Harris'
analysis was the basis for an inventory system that he marketed. Interest in
academic circles was sparked by the later studies of R.H. Wilson [1934]. Few
papers were published in the next fifteen years, although some interest in
mathematical inventory models resurfaced during the war. Landmark papers
were published in the early fifties ]Arrow, Harris & Marshak, 1951; Dvoretsky,
Kiefer & Wolfowitz, 1952a,b] which laid the groundwork for later develop-
ments in the mathematical theory of inventories. The monograph by Whitin
[1957] was also an important development in presenting the relationship of
inventory management issues and classical economic thinking, and was one the
first treatments of the (Q, r) model under uncertainty (to be described in
Section 5), which later became the cornerstone for many commercial inventory
systems.
A number of important scholars turned their attention to mathematical
inventory models during the 1950s. Bellman, Glicksberg & Gross [1955]
showed how the methods of dynamic programming, a term apparently coined
by the first author for describing a class of sequential decision problems, could
be used to obtain structural results for a simple version of the periodic review
stochastic demand problem. A collection of highly sophisticated mathematical
models by Arrow, Karlin & Scarf [1958] provided much of the impetus for later
work in this area. At about the same time, Wagner & Whitin [1958] discovered
the solution to the dynamic lot sizing problem subiect to time varying demand.
The number of papers published today on single product inventory problems
is weil into the thousands. These papers appear in journals published in the
United States including Management Science, Operations Research, AIIE
Transactions, Naval Research Logistics Quarterly, Mathematics of Operations
Research, and Decision Sciences, to name a few, and many more international
 Ch. 1. Single-Product, Single-Location Models 5

journals. Dozens of books have been published on the subject. The study of
inventory management models is now part of the required curriculums in both
Schools of Industrial Engineering and Schools of Business (where it is inte-
grated with production and operations management). Commercial computer
based systems for inventory control have been available for more than twenty-
five years and several inventory control systems for the personal computer are
now available.

1.2. Motivation

Arrow [1958] presented an interesting discussion of the motives of a firm for

holding inventories. H e argues that the motives for holding inventories and the
motives for holding cash are the same. He refers to Keynes who claimed that
there are essentially three motives for holding cash: the transaction, the
precautionary, and the speculative.

1.2.1. The transaction motive

The key to the transaction motive is that by producing or ordering large lots,
fixed costs may be reduced. Economies of scale can be achieved when the
number of set-ups are reduced or the number of transactions are minimized.
This fact was evidently recognized by Harris [1913], because the original
formulation of the E O Q model incorporates a set-up cost for ordering. Arrow
also points to a number of references of work published in the 1920s which
explicitly discuss this issue in the light of inventory considerations.

1.2.2. The precautionary motive

The precautionary motive is based on the fact that inventories are often
retained as a hedge against uncertainty. As we will see in this chapter, the
focus of a major portion of the research in inventory modeling is controlling
inventories in the face of stochastic demand. Although demand uncertainty is
the most obvious and significant source of uncertainty for most systems, other
uncertainties exist as well. There may be uncertainty in the supply. Such a
situation occurred in the late 1970s in the United States when there was an oil
embargo placed against the United States by members of O P E C . Uncertainty
in the order lead time is a common problem. Some costs may be difficult to
predict. Holding costs are tied to the rate of inflation. Commodities such as
precious metals are subject to considerable price variation. Safety stocks in
inventory systems are essentially the same as reserve ratlos of cash for banks.
A r r o w points out that the reserve ratio problem was studied as far back as
1888, and was rediscovered by Eisenhart in 1948 in the context of inventory
control.

1.2.3. The speculative motive

There are essentially two speculative motives for holding inventories. If the
cost of obtaining or producing the items is expected to rise in the near future, it
is clearly advantageous to hold inventories in anticipation of the price rise.
6 H.L. Lee, S. Nahmias

Companies which use commodities whose value may fluctuate could find it
advantageous to stockpile inventory. For example, silver is used in the
production of photographic film. Although the price of silver has been relative-
ly stable in recent years, it increased suddenly by about a factor of ten during
the late 1970s. Increases in labor costs could occur due to negotiated wage
increases, changes in the labor pool, changes in the need for specialized help,
or strikes.
Inventories may also be retained in advance of sales increases. If demand is
expected to increase, it may be more economical to build up large inventories
in advance rather than to increase production capacity at a future time.
Flowever, it is often true that inventory build-ups occur as a result of poor
sales. F o r this reason, aggregate inventory levels are often used as a means of
gauging the health of the economy.

1.3. Types of inventory models

Inventory models come in all shapes, sizes, colors and varieties. In this
chapter we will consider models for only a single product at a single location.
Even with these restrictions the number of possible models is enormous. In
general, the assumptions that one makes about three key variables determines
the essential structure of the model. These variables are demand, costs, and
physical aspects of the system. We will discuss each of these in turn.
(1) Demand. The assumptions that one makes about demand are usually the
most important in determining the complexity of the model.
(a) Deterrninistic and stationary. The simplest assumption is that the demand
is constant and known. These are really two different assumptions: one, that
the d e m a n d is not anticipated to change, and the other is that the demand can
be predicted in advance. The simple E O Q model is based on constant and
k n o w n demand.
(b) Deterministic and time varying. Changes in demand may be systematic or
unsystematic. Systematic changes are those that can be forecasted in advance.
Lot sizing under time varying demand patterns is a problem that arises in the
context of manufacturing final products from components and raw materials.
(c) Uncertain. We use the term uncertainty to mean that the distribution of
d e m a n d is known, but the exact values of the demand cannot be predicted in
advance. In most contexts, this means that there is a history of past observa-
tions from which to estimate the form of the demand distribution and the
values of the parameters. In some situations, such as with new products, the
d e m a n d uncertainty could be assumed but some estimate of the probability
distribution would be required.
(d) Unknown. In this case even the distribution of the demand is unknown.
The traditional approach in this case has been to assume some form of a
distribution for the demand and update the parameter estimates using Bayes
rule each time a new observation becomes available.
(2) Costs. Since the objecti'ee is to mininaize costs, the assumptions one
 Ch. 1. Single-Product, Single-Location Models 7

makes about the cost structure are also important in determining the com-
plexity of the model.
(a) Averaging versus discounting. When the time value of money is consid-
ered, costs must be discounted rather than averaged. The discount factor, «, is
given by a = (1 + r)-i where r is the interest rate. If C1, C~ . . . . represents a
stream of costs incurred over periods 1, 2 , . . (or expected costs if there is
uncertainty), then the present value of the cost stream is given by:

Ol i c z .
i=1

On the other hand, if costs are averaged, then the average cost (again,
assuming infinitely many periods in the planning horizon) is given by:

lim 1 ~ Ci "
n-~~ n i=-I

(b) Structure of the order cost. The assumptions that one makes about the
order cost function can make a substantial difference in the complexity of the
resulting model. The simplest assumption is that the cost of obtaining or
producing y items is cy for some constant c. This is known as a proportional
order cost and is often assumed when demand is uncertain. However, it is
much more realistic to assume that the order cost has both fixed and variable
components. That is, it is of the form cy + K3(y), where 6(y) is the Dirac
delta function satisfying

6(y)= ify=0.

Most deterministic inventory models assume an order cost of this form. The
analysis of stochastic models is considerably more complex when a fixed order
cost is included, however.
(c) Tirne varying costs. Most inventory models assume that costs are time
invariant. Time varying costs can orten be included without increasing the
complexity of the analysis.
(d) Penalty costs. Most stochastic, and many deterministic models, include a
specific penalty, p, for not being able to satisfy a demand when it occurs. In
many circumstances p can be difficult to estimate. For that reason, in many
systems one substitutes a service level for p. The service level is the acceptable
proportion of demands filled from stock, or the acceptable proportion of order
cycles in which all demand is satisfied.
(3) Other distinguishing physical aspects. Inventory models are also dis-
tinguished by the assumptions made about various aspects of the timing and
logistics of the model. Some of these include:
(a) Lead time assumptions. The lead time is defined as the amount of time
that elapses from the point that a replenishment order is placed until it arrives.
8 H.L. Lee, S. Nahrnias

T h e lead time is a very important quantity in inventory analysis; it is a measure

of the system response time. The simplest assumption is that the lead time is
zero. This is, of course, analytically expedient but not very realistic in practice.
It makes sense only if the time required for replenishment is short compared
with the time between reorder decisions.
The most common assumption is that the lead time is a fixed constant. The
analysis is much more complicated if the lead time is assumed to be a random
variable, lssues such as order crossing (that is, orders not arriving in the same
sequence that they were placed), and independence must be considered.
(b) Backordering assumptions. Assumptions are required about the way that
the system reacts when demand exceeds supply. The simplest and most
c o m m o n assumption is that all excess demand is backordered. Backordered
demand is represented by a negative inventory level. The other extreme is that
all excess demand is lost. This latter case, known as lost sales, is most common
in retailing environments.
Mixtures of backordering and lost sales have also been explored. Various
alternatives exist for mixture models. One is that a fixed fraction of demands is
backordered and a fixed fraction lost. A n o t h e r is that customers are willing to
wait a fixed time for their orders to be filled.
(c) The review process. Continuous review means that the level of inventory
is known at all times. This has also been referred to as transactions reporting
because it means that each demand transaction is recorded as it occurs.
M o d e r n supermarkets with scanning devices at the checkout counter are an
example of this sott of system (assuming, of course, that the devices are
connected to the computer used for stock replenishment decisions). The other
extreme is periodic review. This means that the inventory level is known only
at discrete points when physical stock taking occurs. Most systems are periodic
review, although continuous review approximations are common. In continu-
ous review systems one generally assumes that reorder decisions can be made
at any time while in periodic review they can be made only at the pre-
determined times corresponding to the start of periods.
(d) Changes which occur in the inventory during storage. Traditional inven-
tory theory assumes that the inventory items do not change character while
they are in stock. Some inventories, such as volatile liquids or radioactive
materials, experience exponential decay in which a fixed fraction of the
inventory is lost each unit of time. Fixed life inventories, such as food or
photographic film, are assumed to have constant utility until they reach their
expiration date and must be discarded. Finally, inventories m a y b e subject to
obsolescence, which means their useful lifetime may not be predictable in
advance.
This chapter will present an overview of the major developments and results
for single product inventory management. Porteus [1988] in a chapter in an
earlier volume discusses stochastic inventory models. Other chapters will be
devoted to inventory systems with multiple echelons and to production/
distribution systems. This chapter is organized as follows: 1. Introduction; 2.
 Ch. 1. Single-Product, Single-Location Mode& 9

Models with constant demand rates; 3. Models with deterministic time-varying

demand; 4. Periodic review stochastic demand models; 5. Continuous review
stochastic demand models; 6. Application of inventory theory in industry.

2. Models with constant demand rates

In this section we introduce a class of models that is based on the simplest

demand assumption: demand is deterministic and stationary. We concentrate
primarily on the case where the demand rate is constant and not anticipated to
change.
Although the assumption of deterministic and stationary demands seems
quite restrictive, models requiring that assumption are still important for the
following reasons. First, many results are quite robust with respect to the
model parameters, such as the demand rate and costs. The Economic Order
Quantity (EOQ) model is an excellent example. Second, the results obtained
from these simple models are offen good starting solutions for more complex
models. This point will be more transparent in Section 4.

2.1. The basic EOQ model

The assumptions of the basic EOQ model are:

(1) Demand is known with certainty and fixed at A units per unit time.
(2) Shortages are not permitted.
(3) Lead time for delivery is instantaneous.
(4) There is no time discounting of money. The objective is to minimize
average costs per unit time over an infinite time horizon.
(5) Costs include K per order, and h per unit held per unit time.
The unit holding cost h is composed of the cost of physical storage (e.g.,
handling, insurance, taxes, warehousing) h' and the cost of capital invested in
inventory. Let I be the capital cost per dollar of inventory investment, and c be
the unit cost of the item, then h = h' + Ic.
Since delivery lead time is instantaneous and demand is known with certain-
ty, one only orders whenever the inventory level hits zero. Suppose the order
size is Q. Define the cycle time as the time between two successive arrivals of
orders, called T. Note that T = Q/A. Since all cycles are identical, the average
cost per unit time is simply the total cost incurred in a single cycle divided by
the cycle length.
In each cycle, the order cost is K. Consider a cycle from time t -- 0 to T. The
total holding cost incurred in a cycle is

h
f0~ ( Q - A t ) dt=h(QT- ½,~T2) = ~hTQ.

Figure 1 describes the inventory level during cycles. It follows that the total
average cost per unit time is
10 H.L. Lee, S. Nahmias

Inventory
Level ~__ a cycle
(D/Q)

ä~e

Fig. 1. Inventory level during cycles.

C(Q) = (K + ½hTQ)/T= KA/Q + ~hQ .

C(Q) is convex in Q. To find the Q which minimizes C(Q) we solve
dC/dQ = 0. The optimal solution Q*, commonly known as the E O Q , is:

Q* = ~ / h . (2.1)

When there is a positive lead time of, say, r time units, then one merely
places each order r units of time before the inventory level hits zero.
Substituting Q* in the expression for C(Q), one gets C(Q*)= 2X/UK-Ah. It is
easy to show that

C(Q*) - 2 --Q- + "

If Q*/Q = 2, then C(Q)/C(Q*)= 1.25. This means that an error of 100% in

Q results in an increase in the average costs of only 25%. Hence, if there are
errors in the estimation of the cost and demand parameters, the resultant error
in Q may not result in a substantial cost penalty. The E O Q formula has thus
b e e n found to be extremely insensitive in average costs to errors in parameter
estimation [see Hadley & Whitin, 1963; Silver & Peterson, 1985].
A natural extension of the basic E O Q model is the Economic Manufacturing
Quantity ( E M Q ) model (also called E P Q for Economic Production Quantity).
In the E M Q model, an order of stock does not arrive instantaneously but
instead, stock is produced at a finite rate of 4' per unit time, where 4' > A.
H e r e , while production of a batch is underway, stock is accumulating at a rate
of 4' - A per unit time. The total time of production in a cycle is Q/4', so that
the peak level of inventory in a cycle is (41 - A)Q/t). The total inventory held in
a cyele is thus (~ - A)QT/2& Proceeding as with the E O Q model, we obtain
the optimal lot size Q*, called the E M Q , as:

Q* ~- ~/2KA4'/h(O - A). (2.2)

As q, goes to infinity, (2.2) becomes the E O Q .

 Ch. 1. Single-Product, Single-Location Models 11

The E O Q model assumes an average cost per unit time criterion. It ignores
the time value of money. It is easy to generalize the model to the case of
discounted costs. Let r be the continuous discount factor, and consider the
present value of total costs. The resulting E O Q value can be approximated by
the use of Maclaurin expansion [see Jesse, Mitra & Cox, 1983; Porteus, 1985b]:

Q* = V~2KA/(h' + rc). (2.3)

Recall that h' is the component of the inventory cost due to cost of capital
only.
As it turns out, (2.3) is exactly the same as (2.1) with I replaced by r.
Hadley [1964] compared the above formula with the exact solution of Q* and
found that the approximation is excellent.
When the time horizon is finite, and beginning and ending inventory levels
are constrained to be zero, the optimal order sizes within the time horizon have
been shown to be all equal to AT~n* [Schwarz, 1972], where T is the time
horizon, and n* is the smallest integer satisfying

T <~[ Ä(~__~/_~]2n(
+n 1)K 1/2

Schwarz [1977] showed that if T corresponds to a duration of 5-EOQ's worth of

supply, the cost penalty of operating under the optimal finite horizon policy
instead of the infinite horizon is only 1%.
In the remainder of this section, we consider several extensions of the basic
E O Q model under infinite time horizons.

2.2. Quantity discounts

In the basic E O Q model, the unit procurement cost of the item is assumed to
be independent of the order size, and hence has been excluded from considera-
tion in the cost function. In many situations, the supplier is willing to charge
less per unit for larger orders, thus providing an incentive to increase the lot
size. Three types of discount schedules are most commonly used: (t) a
continuous discount schedule where the price of the item is a continuous (often
linear) function of the order size [Ladany & Sternlieb, 1974; Rosenblatt & Lee,
1985], (2) an all units discount schedule in which the lower unit cost is applied
to all of the units in the order [Hadley & Whitin, 1963], and (3) incremental
discount which is applied only to the additional units beyond some breakpoint
[Hadley & Whitin, 1963; Hax & Candea, 1984]. The example in Table 1
illustrates the distinction between the all units and incremental discount
schedules.
With a quantity discount, the cost function for determining the optimal lot
size would have to include the average procurement cost. Moreover, the unit
12 H . L . L e e , S. N a h m , a s

Table 1
Order size Q All unit discount Incremental discount

Unit cost Total cost Unit cost Total cost

0 < Q < 100 1.00 Q x 1.00 1.00 Q x 1.00
100 ~ Q < 200 0.98 Q x 0.98 0.98 100 x 1.00+ ( Q - 100) x 0.98
200 ~< Q 0.95 Q × 0.95 0.95 100 x 1.00 + 100 x 0.98 + (Q - 200) x 0.95

holding cost would change as a result of changing unit cost of the item due to a
change of the lot size. With the continuous discount schedule, determining the
optimal solution may require solving a transcendental equation, so that the
optimal lot size would have to be obtained by numerical search.
For the all units quantity discounts schedule, assume several price breaks bi,
b z . . . . , b i, . . . , b .... such that if the order lot size Q is within discount interval
i, i.e., bi_a ~< Q < b,, then the unit price for each of the Q units is c~, where
c, < ce_~, and m is the number of discount intervals. Let I be the rate per dollar
of carrying inventory. The average cost per unit time if discount level i is used
is:

C~(Q) = «,A + K A / Q + ½(h' + I c , ) Q .

It can be shown that the optimal Q* must be either the E O Q for some
discount level, or one of the breakpoints of some discount interval. Hence, to
find the Q*, we can proceed by determining the individual Q,'s that optimize
C , ( Q ) subject to the constraint b,_ 1 < Q < ~ b , , respectively (which should
either be the E O Q for discount level i, b,_l, or b,). Hadley & Whitin [1963]
suggested an efficient algorithm for finding Q*.
For the incremental quantity discounts schedule, suppose the first b~ units
cost c~ each, the n e x t ( b 2 - b i ) units cost c 2 each, and, in general, the units in
the (b,_l, b,) interval cost ci each, with c, < c~_1 and b 0 defined to be 0.
Suppose Q lies in [b~_~, b,), then the purchase cost of a lot of size Q is given
by:
i-1
Pi = ~ cj(b» - bi_l) + c i ( Q - b i _ l ) .
j=1

The total cost per unit time for discount schedule i is thus:

C i ( Q ) = ( K + P i ) A / Q + ½(h' + I P , ) .

As C i ( Q ) has basically the same form as the cost function of the E O Q

model, it is straightforward to find the optimal Q, that minimizes C , ( Q ) . The
overall global optimal Q* is again obtained by searching over the various
intervals.
 Ch. 1. Single-Product, Single-Location Models t3

2.3. Learning effects

Learning effects are said to exist when, as cumulative production volume

increases, the unit production cost or time decreases as a result of the
experiences gained. The implication to the lot-sizing problem is that the
production cost is not constant over time.
When there are learning effects, three modifications for the basic E O Q
model are required. First, learning effects can result in lower manufacturing
costs [Keachie & Fontana, 1966; Muth & Spremann, 1983; Smunt & Morton,
1985]. Second, learning effccts can increase the production rate as the cumula-
tive production volume increases. Hence, the production time for a lot of Q
units is no longer linear in Q [Adler & Nanda, 1974; Sutc, 1978; Fisk & Ballou,
1982]. Third, the lower manufacturing cost as a result of learning lowers the
inventory carrying cost [Wortham & Mayyasi, 1972; Staunt & Morton, 1985].
The second and third effects can only be analyzed in a finite time horizon.
With the presence of learning, one would also have to specify how the effects
of learning are carried over from one production cycle to another. Between
one production run and another, therc will be time when items are not being
produced. During this time, it is possible that some 'forgetting' or 'unlcarning'
will occur, so that at the time of the heXt production run, the unit manufactur-
ing cost is not the same as that of the last production run. The term
'transmission' of learning is usually used in this context. When there is perfect
transmission of learning, thcn the unit manufacturing cost will decrease at a
rate independent of the lot size. Hence, if the holding cost is constant over
time, i.e., the third effect of learning does not exist, then the learning effect
should have no impact on the optimal EOQ. Most of the studies focus on
partial or no transmission of learning. Again, the effect of partial transmission
of learning can be meaningfuUy studied for the finite horizon case. H e r e we
illustrate the case with no transmission of learning.
Let y ( i ) = cost of the /th unit in a production run of Q, i~< Q. Then the
learning curve function specifies that as cumulative production increases by a
constant percentage, unit cost decreases by a constant percentage. This rela-
tionship gives rise to the following functional form of the learning curve:
y(i) = ai-b, where a is the cost of the first and b is the rate of cost decrease,
0<b<l.
The cumulative average cost per unit, for a lot size of Q, is:
Q
( l / Q ) ~'~ y ( i ) = a Q - b / ( 1 - b) ,
i=1

which is obtained by approximating E°=j y(i) by fo°i b di.

The optimal Q* minimizes

C ( Q ) = K A / Q + ½hQ + a Q - b / ( 1 - b) .
 14 H.L. Lee, S. Nahmias

2.4. Uncertainty in orders received

The basic EOQ model assumes that the amount received is the same as the
amount ordered, i.e., Q. In many inventory systems, this may not be the case.
Recognizing that production processes are orten imperfect so that production
yields may be random, there has been considerable interest in inventory
models where the amount received is uncertain.
Let Y = the amount received given lot size Q, and E(Y I Q) and V(Y ] Q) be
the expectation and variance of Y, respectively. Because the beginning of a
cycle is a regeneration point for a renewal process, and thus the expected
average costs per unit time, C(Q), is the ratio of the expected cost per cycle,
and the expected cycle length. The expected cycle length is now given by
E(Y I Q)/A. We thus have:

K + h{V(Y ] Q) + [E(Y[ Q)]2}/2A

C(Q) = E(Y] Q)/A

The optimal Q* minimizes C(Q). Explicit expressions for Q* have been

derived for specific forms of the density function of Y, given Q [see Silver,
1976; Shih, 1980].
Recently, Porteus [1986] and Rosenblatt & Lee [1986] have independently
incorporated the effect of defective items into the basic EOQ model. An order
would initiate the production process to produce Q units. The production
process is either in control and defect-free items are produced, or out of
control and defective items are produced. Defective items can be reworked
instantaneously at a cost. The production process begins in the in-control state
after a setup is performed. At some point in time, the process goes out of
control, and would remain in that state until the next setup, which corrects the
process back in control. Two essentially equivalent assumptions are used:
Porteus [1986] assumes that there is a probability q that the process would go
out of control while producing one unit of the product; whereas Rosenblatt &
Lee [1986] assume that the time between the beginning of the production run
(in the in-control state) until the process goes out of control is exponential with
mean 1//x. In Rosenblatt & Lee [1986], once the process is out of control, a
fraction o~ of the items produced is assumed to be defective. Let s be the unit
rework cost. The total cost functions now consist of the usual setup and holding
costs, plus the rework costs. Using approximations based on Maclaurin expan-
sions, Porteus [1986] and Rosenblatt & Lee [1986] obtained explicit modi-
fications to the EOQ and the EMQ formula, respectively:

Q.=I 2KA ] 1/2

h + Aqs '

F 2kA~ ]1/2
Q* = L h ,t -q; ---c-;-:
a) cqxls j .
 Ch. I. Single-Product, Single-Location Models 15

It can be seen that the above Q*'s are smaller than the E O Q or the EMQ.
Thus, the presence of defective products motivates smaller lot sizes. Numerical
results show that, even though the probability of the shift is usually very small,
the cost penalty for not incorporating the effects of defective items can be
significant. In later work, Lee & Rosenblattt [1987] considered using process
inspection during the production run so that the shift to the out-of-control stare
can be detected and restorations made earlier.

2.5. Pricing and perishability

In the E O Q model, the demand rate A is assumed to be constant and

independent of the lot size. The model can be easily extended to include the
effects of marketing policies. Ladany & Sternlieb [1974] formulated such an
extension by considering net profit as the objective. The net profit is A [ P -
c ( Q ) ] minus the ordering and holding costs, where P is the unit selling price
and c ( Q ) is the procurement or unit production cost as a function of Q. The
selling price P is modelled as a fixed mark-up of the unit cost, and hence is a
function of Q. Also the demand rate A is a function of the selling price.
Normally, for a downward sloping demand curve, A is decreasing in P. The
implication of such a formulation is that the objective function is a much more
complex function of the lot size. By making simple assumptions on the form of
c ( Q ) and the relationship of A and P, closed form expressions for the optimal
lot size can be obtained [Ladany & Sternlieb, 1974; Lee & Rosenblatt, 1986].
In general, however, search routines would have to be used to find the optimal
lot size.
P may be considered to be a decision variable, instead of a given function of
Q. This is the approach taken by Whitin [1955] and Cohen [1977]. In this case,
the E O Q is still the optimal lot size, although the optimal price can rarely be
expressed in closed form.
The E O Q extensions for perishable items usually assume exponential decay
of the items [Ghare & Schrader, 1963; Cohen, 1977]. L e t / x be the stock decay
rate and I(t) be the one hand inventory at time t. The differential equation
describing the time behavior of the inventory level is d I ( t ) / d t = - / x I ( t ) - A,
from which one obtains l(t) as:

I(t) = I(0) e x p ( - / x t ) - (A//x)[1 - e x p ( - / x t ) ] .

Using the boundary condition I(0) = Q, one can compute the inventory costs in
a cycle. The total cost functi0n should also include the costs of the decayed
items. Let c be the unit cost of an item. Using approximations based on the
Maclaurin series, the optimal Q* that minimizes expected costs is:

Q* = ~/2KA/(c/x + h ) .

It can be seen that the optimal lot size is similar to the E O Q , with a
modification of the holding costs to include the cost of item decay.
16 H.L. Lee, S. Nahmias

2.6. A n n o u n c e d price increase

The basic E O Q model assumes that all cost parameters remain unchanged
during the time horizon. Extensions exist which allow these parameters to
change after a known future point in time. Most of these studied [e.g., Lev &
Soyster, 1979; Lev, Weiss & Soyster, 1981; Aggarwal, 1981; Taylor & Bradley,
1985] focus on the unit cost of the item. In particular, it is assumed that there is
an announced price increase at some future point in time. Obviously, after the
new price is in effect, the optimal order size should be the new E O Q based on
the new price to compute the inventory holding cost. The problem is, what
should the ordering policy be between the announcement of the price increase
(at time 0) and the time at which the new price becomes effective (at time T)?
B e f o r e the price increase, the ordering policy is the E O Q based on the
current price. Depending on whether there exist regular 'scheduled' replenish-
ment orders under the current E O Q between times 0 and 7, different results
are obtained [Taylor & Bradley, 1985]. If ~-is small so that there does not exist
any current E O Q orders between times 0 and T, then we may want to place a
special order just prior to T, under the current price, to be consumed after z.
The optimal policy is that a special order should be placed if the inventory level
just prior to ~- is less than some threshold, which is a function of the E O Q
model parameters, and the current and new prices, but is independent of T.
Otherwise, no special order should be placed.
Suppose T is large enough so that there exist current E O Q orders between
times 0 and z, l e t z I be the time between the first current E O Q order time and
T. If T1 is very small or very large, then Taylor and Bradley showed that it is
optimal to modify the current E O Q so that the inventory level just prior to z
will be zero, and to place a special order at 7. If T1 is of intermediate value,
then it is optimal to continue using the current E O Q ordering policy, but place
a special order at the time when the last E O Q order is placed prior to Z.

2.7. Inflation

Instead of a one-time increase in the future, the E O Q model can also

incorporate inflationary conditions. Consider the E O Q model with discount
rate r for the time value of money. Suppose there is an inflation rate of ~,
< r, so that all costs are increasing continuously at a rate of ~"per year. Such
a formulation is actually equivalent to stating that the 'effective discount rate'
is r - ~. The resulting modified E O Q is thus the same as given by (2.1.3), with
r replaced by r - ( [see Trippi & Lewin, 1974; Jesse, Mitra & Cox, 1983].
Buzacott [1975] and Kanet & Miles [1985] obtained essentially the same result
for the total cost approach. When r > r, then the total cost function is
u n b o u n d e d for an infinite horizon, and only the finite horizon analysis is
meaningful [Bierman & Thomas, 1977].
 Ch. 1. Single-Product, Single-Location Models 17

2.8. D e m a n d with trends

When demand is deterministic but foltows some known trend, the optimal
ordering policy is much more complicated. Resh, Friedman & Barbosa [1976]
solved the problem for the linear trend case. Suppose time starts at the origin,
and the demand rate at time t is given by A(t)= At. Consider first a finite
horizon T. Let t 0, t ~ , . . , tm_ ~ be the replenishment times, and Qo . . . . , Qm-~
be the respective order quantities. Hence, there are m replenishments through-
out T. Define tù, = T. Let yi(t) be the inventory level at t, t~<~t<~ti+l,
i = 0 . . . . , r n - 1 . Then,

Qi = ft ti+l A(t) dt = A(t~+ 1 -- t~)/2A

i

and

t
yi(t) = Q~ m
f i
2 1 - ta)/2A.
A(u) du = A(ti+

Given m, the total inventory costs in T can then be computed. Resh, Friedman
& Barbosa [1976] provided an efficient älgorithm to compute the optimal
values of t~'s, from which the optimal total cost for a given m can be obtained.
Such a cost is shown to be convex in m, so that the optimal m* can be readily
found. Of interest is the infinite horizon case, i.e., when m* goes to infinity.
F o r a given m*, there is a range of T s u c h that m* remains unchanged. For this
m*, Resh, Friedman and Barbosa showed that

t* = A i ( 2 K / A h ) ~/3, i= l,...,m*-l, (2.4)

where A~'s are constants that are functions of m* and i only. It can be shown
that each t~* converges to a limit as m* tends to infinity. Moreover, (2.4)
compares rather strikingly with the E O Q cycle length, which is given by
( 2 K / A h ) 1/2 for demand rate A. For this reason, Resh, Friedman & Barbosa
t e r m e d (2.4) as the 'cubic root law' for a growing market, as opposed to the
'square root law' for a stable market.
Barbosa & Friedman [1978[ generalized the above result for the case where
demand is given by a power form: A(t)= Ata, a > - 2 . An '(a + 2) root law'
resulted for the optimal replenishment times over an infinite horizon,

ti* = L~[(a + 2)K/,~h] 1/°+2)

where L i is a constant which is a function of i and a.

18 H.L. Lee, S. Nahmias

3. Models with deterministic time-varying demand

As noted above in Section 1.3 it is generally the assumption about demand

that is most important in determining the complexity and structure of an
inventory model. In this section we generalize the premise of the previous
section to allow demand to be time-varying. However, we will retain the
assumption that demand is deterministic.
A continuous review model which allowed for time varying demand was
considered in Section 2.8. However, since the vast majority of the research on
time-varying demand problem assumes periodic review, we will henceforth
restrict attention to that case. Let (rl, r 2 , . . , r n) be known requirements for a
single product over the next n periods. Periods are defined as times of
reptenishment opportunity. Let (Yl, Y 2 , • - , Yn) be the order sizes placed in
these periods. For convenience we add the assumption that the order lead time
is zero. As with the E O Q type models discussed above, the extension to a
positive order lead time is not difficult. One merely translates the order
backwards by the n u m b e r of periods in the lead time. Furthermore, since the
models described below are offen used in M R P (Material Requirements
Planning) settings for production, we will use the terms order sizes and
production quantities interchangeably.

3.1. The Wagner-Whitin model

The original formulation of this problem is due to Wagner & Whitin [1958]
(henceforth referred to as WW). They assume:
(1) Shortages are not permitted.
(2) Starting inventory is zero. This assumption may be relaxed by netting
out starting inventory from the first period's (or additional periods', if neces-
sary) demand.
(3) Only linear costs of holding and fixed order costs are present. Note that
(3) means that the total cost function is concave. WW's results hold under the
more general assumption:
(3') The holding cost is a concave function of the ending inventory in each
period 1 and the ordering cost is a concave function of the order quantity in
each period.
The ending inventory in period t is given by:

xt = ~ (yj - rj).
y=l

Let ht(xt) be the holding cost incurred in period t and c,(yt) the order cost.

lWW assume holding cost is based on starting rather than ending inventory.
 Ch. 1. Single-Product, Single-Location Models 19

The objective is to find non-negative order or production quantities

(Yl, Y2, . . , Yù) to

minimize B [«,(y,) + h,(x,)]

t=l

subject to x, = ~ ( y j - rj),
j=l

xt>~O, t=l,...,n,

x0=0.

The most common form of the cost functions is:

c,(y,) = K6(yt) + cyt

and

h,(x,) = hx, ,

where 6(y,) equals 1 if y, > O, and 0 otherwise.

Wagner and Whitin developed an efficient algorithm for solving this problem
which was based on the following observation.

Result. An optimal ordering policy has the property

YtX,_l = 0 , t = 1, 2 , . . . , n .

This means y, > 0 only if x,_ 1 = 0 or in words, ordering takes place only in
periods when starting inventory is zero. A little reflection will show that this
result means that Yt = 0 , r t , r t "+ F r + l , . . . , r t + r t + 1 -4- - - - -b F n , are the only
permissible values for y« If not, it could happen that for some t, 0 < xy_ 1 < r t
and ordering would have to occur to avoid shortages.
The optimal policy is then an 'exact requirements' policy and is completely
specified by the periods in which ordering is to occur. The optimal policy may
be found by finding the shortest path through an acyclic network with nodes
labelled 1, 2 . . . . . n + 1 and arcs (i, j) connecting all pairs of nodes with i < j.
The length of the arc connecting nodes i and j, cij, is the cost of ordering in
period i to satisfy requirements through period j - 1, where i < j ~< n + 1, and
j-] j-2 /-1
cq=ci(~r~)+ ~,hg( ~ r,).
"k=i k=i ~l=k+l

Given the values of cij for 1 ~< i < j ~< n + 1, one may find the optimal order
20 H.L. Lee, S. Nahmias

policy by solving either a backwards or forwards dynamic program. The

backward formulation is:

B=min(cij+fj) fori=l .... n

j>i »

fù+l = 0 .

Interpret f,. as the cost of following an optimal policy when i periods remain.
The optimal solution is found by re-tracing the optimal values of j. The
procedure is best illustrated by example. (This example appears in more detail
in Nahmias [1989]).

Example. Suppose r = (52,87,23,56), c,(y)= K•(y), where K = 7 5 for t =

1,..,4, a n d h , ( x ) = h x , w h e r e h = l for t = l . . . . ,4.
Then the dynamic programming equations yield:

L=O~

f4 = 75 at j = 5 ,

B -- 131 at j = 5 ,

f2 = 173 at j = 4 ,

fl --- 241 at j = 2 .
The optimal solution is determined in the following manner. Because the
minimizing value of fl occurs at j = 2, one orders in period 1 and again in
period 2. Since the minimizing value of f2 occurs at j = 4, one orders in period 2
and not again until period 4. Finally one must order in period 4 to satisfy the
requirement in that period. Hence the optimal policy is

Yl = rl = 52,

Y2 = r2 + r3 = 110,

Y3=0,

Y4 = r4 = 56.

3.2. Planning horizons

Wagner and Whitin's work sparked considerable interest in the concave cost
dynamic lot size problem. One area of interest is planning horizons. Wagner
and Whitin's original observation was the following.
 Ch. 1. Single-Product, Single-Location Models 21

W W Planning Horizon Theorem. Let l(t) be the last period a set-up occurs for
the optimal order policy associated with a t period problem. Then for any
problem of length t* > t it is necessary to consider only periods l(t) <~j <~t* as
candidate periods for the last set-up. Furthermore, if l(t)= t, the optimal
solution to a t* period problem has y,>O. In the latter case, periods 1,
2 , . . . , t - 1 constitute a planning horizon.

What is the advantage of locating a planning horizon? Its identification

means that the optimal policy for periods 1, 2 . . . . . t - 1 does not depend upon
the demand forecasts beyond period t. When a forward algorithm is used to
compute the optimal policy, the computations can go forward from the end of
the last planning horizon rather than from period 1.
Additional results on planning horizons for the no backlog case were
obtained by Zabel [1969], Eppen, Gould & Pashigian [1969], and Lundin &
Morton [1975], and for the backlog case by Blackburn & Kunreuther [1974].
Planning horizons are, we feel, of more theoretical than practical interest. The
W W algorithm (forward or backward) is extremety efficient. As one extends
the length of the horizon the salient issue is forecast accuracy and not
computational efficiency.

3.3. Backordering

Zangwill [1966a] generalizes the work of WW to allow for backordering of

demand. When backordering is permitted the fundamental structural result of
W W that Yt ~ 0 only if x, = 0, where x t is the starting inventory of period t,
generalizes to y, > 0 only if x t ~< 0. Furthermore an optimal policy still orders
exact requirements. That is, there exist integers 0 = S o ~< S 1 ~ • • • <~ Sù = n such
that:

y,=0 ifS,_l=St,
and

St

y,= ~ rj for l ~ < t ~ < n .

j=l+St_ 1

Backorders occur when S, < t.

Zangwill's analysis permitted general concave holding and shortage costs and
a limit on the n u m b e r of periods a backorder stays on the books. He gave a
forward dynamic programming algorithm for computing the optimal policy. An
extension of these results to allow for multiple products and echelons appears
in Zangwill [1966b].
22 H.L. Lee, S. Nahmias

3.4. Heuristic methods

T h e r e has been considerable interest in approximate solutions to the basic

WW lot sizing modet. In some sense this is surprising since dynamic program-
ming efficiently finds the optimal solution. Production lot-sizing plays an
important role in M R P and practitioners have shied away from incorporating
dynamic programming into their planning systems. Hence there is strong
interest in more intuitively appealing heuristic methods. Also, the heuristics
often provide more stable solutions in a rolling horizon environment in which
demand forecasts are revised periodically. We discuss a number of better
known methods.

3.4.1. The Silver-Meal heuristic

This method, due to Silver & Meal [1973], is a forward algorithm that
requires computing the average cost per period as a function of the n u m b e r of
periods in the current order horizon.
Define C ( T ) = t h e average cost of holding and setup per period if the
current order spans the next T periods.
If (r 1. . . . , r n) are the requirements over the next n periods and all costs are
constant, it follows that

C ( T ) = [K + hr I + 2hr 2 + . . . + ( T - 1 ) h r r ] / T .

The m e t h o d is to compute C ( T ) for T = 1, 2 . . . . , stop the first time that

C ( T ) > C ( T - 1), and set Yl = E jr=-~ rj. The process is started again at period T
and continues until the end of the horizon is reached.

3.4.2. Part period balancing

Part period balancing is generally credited to DeMatteis [1968]. Like the
Silver-Meal heuristic, the method finds the length of the horizon to be spanned
by the current order. The horizon length is that whose total holding cost most
closely matches the set-up cost.

Example. Consider the previous example with

r = (52, 87, 23, 56), K = 75, h -- 1.

(a) Solving by the Silver-Meal heuristic:

Starting in period 1:
C(1) = 75
C(2) = (75 + 87)/2 = 81.
Stop and set Yl -- rl = 52.
Starting in period 2:
C(1) = 75
C(2) = (75 + 23)/2 = 49
 Ch. 1. Single-Product, Single-Location Models 23

C(3) = [75 + 23 + (2)(56)]/3 = 70.

Stop and set Y2 = r2 + rB = 110.
Finally Y4 = 56.
The Silver-Meal solution is y = (52,110, 0, 56).
(b) Solving by part period balancing:
Starting in period 1:

Order horizon Total holding cost

1 0
2 87

Since the set-up cost of 75 is closer to 87 than 0, set y~ = r I + r 2 = 139 and

commence the process again in period 3.
Starting in period 3:

Order horizon Total holding cost

1 0
2 56

Since the set-up cost of 75 has not yet been reached, set Y3 r3 + r4 79. The
= =

part period balancing solution is thus: y = (139, 0, 79, 0).

For this example, Silver-Meal gives the optimal solution while part period
balancing does not. The cost of the optimal solution is (75)(3) + (1)(23) = 248.
The cost of the part period balancing solution is (75)(2) + (1)(87 + 56) = 293.
It has been demonstrated that as the number of periods grows without
bound, the relative error of Silver-Meal can be arbitrarily large [Axsäter,
1982] but the relative error of part period balancing is bounded by 3 [Bitran,
Magnanti & Yanasse, 1984]. This might suggest that part period balancing is a
better heuristic. However, worst-case performance is not necessarily an accur-
ate indicator for the effectiveness of a heuristic. In fact, simulation studies
[Silver & Peterson, 1985] indicate that on average Silver-Meal is a better
performer. (A more extensive discussion of heuristics is given by Baker in
Chapter 11 of this volume.)

3.5. Capacitated lot sizing

Practically speaking, an important generalization of W W is the case where

there are upper bounds on the production quantities each period. That is, there
are capacities U1, U2 . . . . . U n, such that Yi ~< bi, for 1 ~i~< n. Limiting
production capacity each period significantly alters the properties of the
optimal solution, and the computational complexity of the optimal solution
procedure.
Florian & Klein [1971] have discovered a solution procedure based on the
24 H.L. Lee, S. Nahmias

following property of an optimal policy: between any two periods in which

starting inventory is zero there is at most one period in which production is
neither zero nor at capacity. They constructed a solution algorithm based on
solving a collection of acyclic network problems. In general, however, the
method is quite tedious. Love [1973] also obtained a characterization of the
optimal policy under a more general model structure, but his results do not
lead to an efficient algorithm in general.
The most computationally efficient method of solving the problem optimally
seems to be the one proposed by Baker, Dixon, Magazine & Silver [1978].
They developed a tree search algorithm based on properties of an optimal
solution. Still the authors note that, in the worst case, their algorithm is not
computationally effective. For example, a typical 24 period problem required
the solution of 18 000 different subproblems.
Because of the computational complexity of optimal solution algorithms,
approximate methods are of interest. Heuristic methods for solving the capaci-
tated problem have been explored by Dixon & Silver [1981] and Karni [1981].
These heuristics involve determining a feasible solution and subsequently
combining lots to reduce setups. Nahmias [1989] considers such a method in
the context of material requirements planning.
A number of researchers have considered multi-item lot-sizing subject to
capacity constraints. We will not review these methods here but refer the
interested reader to Chapter 8 of this volume.

3.6. Convex costs

In the production planning context, concave costs arise much more frequent-
ly than do convex costs. The primary reason is that if a fixed set-up cost is
present and all other costs are linear (a common situation), then the production
cost function is concave. Furthermore, concavity oecurs when there are
declining marginal costs of production or other scale eeonomies present. There
are some cases, however, where costs are convex rather than concave func-
tions. An example is where fixed costs are negligible, and there is a capacity
constraint on production. Another example is where there are several sources
of limited production which are used in order to ascending costs. This occurs in
the generation of electric power. As with the WW model above, the problem is
to find non-negative production quantities (Yl, Y2, • •, Y,,) to:

minimize ~ Ct(y,) + h,(xt)

t-1

subjectto xt=2 (yj-rj),

j=l

where x t can be positive or negative. Interpret C,(yt) as the cost of producing

y, in period t, ht(xt) as the cost of holding x t in period t ifx, > O, and the cost of
 Ch. 1. Single-Product, Single-Location Models 25

backordering x t in period t if x, < 0. The functions C t and h« are assumed to be

convex functions of their arguments for 1 ~< t ~< n.
Veinott [1964] has developed a parametric algorithm for solving this prob-
lem. It is based on the observation that if costs are convex, the optimal
production quantities are non-decreasing functions of the requirements. Includ-
ing the additional assumption that the cost functions are piecewise linear with
endpoints on the integers assures that the optimal production amounts are also
integral.
Assume that backorders are not permitted and suppose (Yl . . . . . y,,) is an
optimal production plan for (ri . . . . . r,,). Then Veinott proved that
( y ~ , . . , yj + 1 . . . . . Yn) is optimal for the requirements schedule
(r~ . . . . , r i + 1 . . . . , rn) where 1~< i~< n and 1 <~j<~ n. Since Yi = 0 is almost
always optimal for r i = 0, 1 ~ i ~< n, a solution procedure is to start with a zero
requirements vector and increase a component of the requirements vector by
one unit and find the component of production that increases by simply
comparing the costs of the n possibilities. This process is continued until the
original requirements vector is obtained. A more complex procedure is re-
quired when backordering is permitted.

3.7. Lot sizing and materials requirements planning

As noted earlier in this section, one of the most important applications of

dynamic lot sizing algorithms is Materials Requirements Planning (MRP)
systems [see, e.g., Orlicky, 1975, and Chapter 11 of this volume]. Briefly, M R P
is the means by which a set of forecasts for an end item (known as the master
production schedule) are converted to requirements for raw materials and
subassemblies. At each stage of the system production lot sizes must be
determined over the appropriate planning horizon.
Several problems arise when considering multi-level lot-sizing problems
which are not present in the single-level problem. Capacity constraints may
render independent single-level solutions infeasible, thus making multi-level
optimization desirable problems related to rolling horizons [Carlson, Jucker &
Kropp, 1982] and nervousness [Carlson, Beckman & Kropp, 1979; Kropp &
Carlson, 1984] also affect the choice of a lot sizing algorithm. We merely note
these issues here and refer the interested reader to Baker's more in depth
treatment in Chapter 11 of this volume.

4. Periodic review stochastic demand models

The majority of research in inventory has been focused on stochastic demand

models. These models are more sophisticated and usually provide a better
description of reality. Because of their importance, we will consider separately
discrete time and continuous time cases.
26 H.L. Lee, S. Nahmias

4.1. The newsboy model

The newsboy model is the basis for most discrete time stochastic inventory
models. It applies when the product's useful life is only one planning period.
This would be the case when a product perishes quickly such as fresh produce,
certain short-lived style goods, or newspapers (hence the name 'newsboy'
model). It is more interesting for its structural importance than its ap-
plicability.
Suppose that the demand during a period is a random variable D which has
known Cumulative Distribution Function ( C D F ) F(t). For convenience we
assume that D is purely continuous. Similar results apply when D is discrete or
partially continuous and partially discrete. Let f(t) be the density of one
period's demand.
T h e r e are several ways to define costs. We adopt the traditional one used by
Arrow, Karlin & Scarf [1958] and others. Assume it costs c to purchase one
unit, inventory remaining at the end of the period costs h per unit and there is
a penalty cost of p per unit of unsatisfied demand. Assume p > c.
Suppose the initial inventory on hand at the start of the period is x ~> 0. The
decision variable, y, is the inventory on hand after ordering. Hence y may be
called the order-up-to-point and satisfies y I> x. We assume that the delivery
time is zero or small enough so that the order can be used to satisfy the
d e m a n d during the period.
T h e number of units remaining on hand at the end of the period is the
r a n d o m variable y - D , which may be positive or negative. Denote X + =
max{X, 0}. It follows that the cost incurred in the period for a given y is

c(y-x) + h ( y - D) + + p ( D - y)+ ,

which is also a random variable. The usual approach in analyzing such

problems is to determine the expectation and choose the decision variable to
minimize the expected value. This is generally justified by the Law of Large
Numbers. If the decision process is repeated many times, then the average cost
per period will be minimized.
The expected cost, G(y, x), is

G(y,x)=c(y-x)+h
foy ( y - t ) f ( t ) d t + p f; ( t - y ) f ( t ) d t
= G(y) - cx,

where

G(y) = cy + h ( y - t)f(t) dt + p ( t - y)f(t) dt.

 Ch. 1. Single-Product, Single-Location Models 27

It is easy to show that G ( y ) is convex in y so that the minimizing value of y,

say y*, occurs where G'(y*) = 0. Taking the derivative and solving gives
p--c
F(y*)=pTh "

The optimal solution is: if x < y* order y* - x, otherwise do not order. This
is known as a critical number policy. Since F(y*) is the probability that
demand does not exceed y*, the optimal solution occurs where this probability
is equal to the critical ratio. Note that the critical ratio is often expressed in the
form Cu/(Cu + Co) where c u is the underage cost (i.e., p - c), and c o is the
overage cost (i.e., h + c). Also, the assumption p > c guarantees that the
critical ratio is a probability.
The newsboy model may be formulated and solved assuming discrete
demand and discrete variables, or with the objective of maximizing profit
rather than minimizing costs [see Hadley & Whitin, 1963]. One must be careful
to properly interpret costs when applying the model in practice, however. (This
issue is discussed in Hillier & Lieberman [1987] and Nahmias [1989].)

4.2. Dynamic models with no set-up cost

4.2.1. The basic model

Most discrete stochastic inventory problems have been formulated as dy-
namic programs. Beliman, Glicksberg & Gross [1955] appear to be the first to
formulate and analyze dynamic inventory models in this fashion. Assume the
same cost structure as in the previous section, but suppose demands in
successive periods are IID (Independent Identically Distributed) random
variables Dl, D 2 , . . with common CDF F(t) and density f(t). Define Cn(x ) as
the minimum expected discounted cost when n periods remain and x is the
current inventory level. Assume the discount factor, 0 ~< o~ ~< 1 where a = 1
only if n < ~. Furthermore assume zero order lead time. Then the functional
equations defining an optimal policy are

C,,(x) = min G(y) - cx + «

y~x w. f0 C~_l[s(y, t)]f(t) dt
/ for n I> 1 ,

where G ( y ) is defined in Section 4.1.

The function s(y, t) is known as the transfer function. If excess demand is
backordered then s ( y , t ) = y - t while if excess demand is lost s(y, t)=
( y - t) +. One generally assumes the initial condition Co(x ) --0, although the
assumption Co(x ) = - c x (meaning that inventory left at the end of the horizon
is returned for the purchase price) leads to a stationary optimal policy (see
Section 4.2.4 below).
If Co(x ) = 0, then the optimal policy has the form
28 H . L . Lee, S. Nahmias

if x < y,, order up to Yn,

if x I> Yn do not o r d e r ,

where yù minimizes the bracketed sum above. It is relatively easy to show that

ya~y2>-...>~y=.

If demand is backordered Yl solves the newsboy model described in the

previous section and y= solves

p - (1 - c 0 c
F(y=) =
p+h

Similar results apply in the lost sales case.

4.2.2. E x t e n s i o n to include positive order lead time

W h e n a positive order lead time is included, the complexity of the resulting
model is much more sensitive to backlogging assumptions. The dynamic
programming formulation now requires a multi-dimensional state variable
indicating sizes of all outstanding orders. As Karlin & Scarf [1958] proved,
however, as long as excess demand is backordered, the optimal policy depends
only on the sum of the on-hand and on-order stock. Furthermore, the optimal
policy has the same structure as the zero lead time case although the critical
numbers are computed differently.
Unfortunately, when there is a positive order lead time and excess demand is
lost rather than backordered, the optimal policy is extremely complex. It is a
non-linear function of the vector of outstanding orders. The functional equa-
tions defining an optimal policy in the lead time lost sales case are

Cù(x0, xl ..... xT-l) =

y~O
t
min cy + L ( x o ) + a f;0 Cn_l(x o + x I - t, x 2 . . . . , x t _ l , y ) f ( t ) dt

+ «Cn_l(xl, x2,.., y)(1 - F(xo))},

where T = order lead time in periods, x 0 = inventory on hand, x i = outstanding

o r d e r scheduled to arrive in i periods for 1 ~< i ~< T - 1, y = size of the current
order, to be delivered in T periods.
T h e function L ( x o ) is the expected holding and shortage cost in one period if
inventory at the start of the period is x 0. When all costs are linear

L(xo) = h
f)° (x o - O f ( t ) dt + p
F 0
(t - x o ) f ( t ) d t .
 Ch. 1. Single-Product, Single-Location Models 29

Because of the complexity of the optimal policy, approximations are im-

portant. Morton [1969] suggests the following approximation for the optimal
order quantity assuming infinitely many periods remain in the planning horizon

= min{y1 , Y2 . . . . . YT+I} ,

where y* solves
-T
Fi(y * ) = p-a c
p + h- a-T+1c

where F i is the CDF of the i-fold convolution of the one period CDF F(. ).
Nahmias [1979] using different methods develops approximations for several
lead time lost sales inventory models including a positive set-up cost for
ordering, uncertainty in the lead time, and partial backordering of demand.

4.2.3. L e a d time uncertainty

Kaplan [1970] analyzed an extension of the dynamic lead time model to
incorporate lead time uncertainty. What makes the random lead time problem
difficult is the way one treats order crossing. If orders are placed with one
supplier it is unlikely that they would cross in time; that is, an order placed on
Monday should arrive before one placed on Tuesday even though the exact
arrival times may not be certain. The difficulty is that if orders are not
permitted to cross, successive lead times are dependent random variables.
Hadley & Whitin [1963] discuss the order crossing problem.
Kaplan's formulation of the problem was quite ingenious. Let

pi = Pr{all orders placed i or more periods ago

arrive in the current period}.

This formulation guarantees that orders do not cross since the arrival of an
order i periods ago forces the arrival of orders placed i + 1, i + 2 , . . , m
periods ago as weil. The likelihood that the lead time is i periods, qi, may be
computed from Pi. It is the values of qi one would observe in a real system.
Kaplan showed that if there is no order set-up cost, the optimal policy is a
critical number policy in every period.

4.2.4. Batch ordering

Veinott [1965] discovered a forward formulation of the dynamic problem that

leads to several interesting results. In particular he showed that under reason-
able assumptions about the salvage value and the transfer function, an n period
multi-period dynamic inventory problem can be decomposed into n single-
period inventory problems.
30 H.L. Lee, S. Nahmias

As above, let L ( y ) be the expected one-period holding and shortage cost

function. For convenience assume zero order lead time.
Ler x n = starting inventory in period n and yn be the order-up-to point in
period n. T h e n the expected discounted cost for an N period problem may be
written

n-1 N
fN = E ~ I c ( y ù - x,,) + L ( y ù ) - « «xN+t.

This formulation assumes that the stock remaining at the end of the horizon,
XN+1, can be returned at the original purchase price of c. If excess demand is
b a c k o r d e r e d , xù+l = y ù - Dù for 1 ~< n ~< N. Making this substitution and re-
arranging terms gives
N N
f. = E a ù-1 E{cy,,(1 - oe) + L ( y ù ) } - cx I ä~ a~«a.
n=l n=l

N N

= Z «~-'w(yù)-cx~- E ~~ca,
n=l n=I

wherea=E(Dù) forn=l, 2,..,N.

If there are no constraints on y, then the optimal policy is the value of y*
minimizing W(y). Because fu is the sum of N one period expected cost
functions, one can find the form of the optimal policy for the multi-period
p r o b l e m as easily as for the one period. Similar results apply when there is a
positive order lead time and excess d e m a n d is backordered.
This m e t h o d can often be used to characterize the form of the optimal policy
when there are constraints on y. As an example, suppose the minimum order
size is M. This problem is c o m m o n in dry goods inventory control where items
are packaged in groups of six or twelve. T h e n it is easy to find the optimal
policy. Consider Figure 2. If W(y) is a strictly convex function of y there is a

w(y)

f I

I I
I i
[ I
1 [
k k+M Y

Fig. 2. Optimal policy for batch ordering.

 Ch. 1. Single-Product, Single-Location Models 31

unique value k satisfying W(k)= W(k + M ) . 2 Suppose in any period that

k < x < k + M. An order size of one batch brings the inventory level to
x + M > k + M and to a necessarily higher cost. Hence it is optimal not to
order in this case. If x < k then it is clear that one should order that number of
batches that brings the inventory into the interval [k, k + M].
This forward formulation is also useful for characterizing the optimal policy
in multiproduct systems with no set-up cost.

4.3. Dynamic models with positive set-up cost

4.3.1. Structure of the optimal policy

A set-up cost for ordering means that the cost of ordering u units is
K6(u) + cu where
(~ if u = 0 ,
6(u) =- if u > 0.

What makes the analysis difficult is that ö(u) is discontinuous at zero. The
form of the optimal policy is easy to derive for a one-period problem, but
proving that form is optimal for the multi-period problem is more difficult. The
function to be minimized in the newsboy model, G ( y ) , is convex. It appears in
Figure 3.
Define S as the value of y at which G is minimized, and s to satisfy s < S,
G(s) = G(S)+ K. As above, let x be the starting inventory in a period. If
s ~< x < S then G(x) <~G(S) + K and it is clearly advantageous not to order. If

C(y)

I
!
I
I
I
I y

order % do not o r d e r >

Fig. 3. Optimal policy for the one-period problem with set-up cost.

2This holds under somewhat weaker conditions as weil. For example if W is strictly quasiconvex.
32 H . L . Lee, S. Nahmias

x < s then G ( x ) > G ( S ) + K and it is optimal to order up to S. Hence, the

optimal policy is an (s, S) policy:

i f x < s order to S (i.e., order S - x ) ,

if x 1>s do not order.

The functional equations defining an optimal policy for the multi-period

version of the problem are

Cù(x) = min
y~x / K6(y - x) - cx + G ( y ) + «
f0~ C , , _ l ( s ( y , t ) ) f ( t ) dt / .
If we could show that the bracketed term is convex in y, then the arguments
above establish that an (s, S) policy is optimal in every period. However, the
bracketed term is n o t convex as a function of y. The goal of the analysis is to
determine if there is some property possessed by the bracketed term as a
function of y for which (s, S) policies remain optimal. This property was
discovered by Scarf [1960] and is known as K-convexity. A differentiable
function g ( x ) is K-convex if K + g ( x + y ) - g ( x ) -- y g ' ( x ) >t O. Scarf showed
that an (s, S) policy is optimal for a K-convex function and established
properties of K-convex functions that allowed hirn to prove inductively that the
bracketed term is a K-convex function of y for the case S(y, t) = y - t. This
property allowed Scarf to show that the form of the optimal policy was (s, S) in
every period of a finite horizon problem, although the values of s and S could
change from one period to the next. That a stationary (s, S) policy is optimal
for the infinite horizon problem was proven by Iglehart [1963].

4.3.2. G e n e r a l i z e d (s, S ) policies

A n order cost function consisting of both proportional and fixed components
is a special type of concave order cost. More general concave order costs are of
interest, since they can be used to describe quantity discounts. Porteus [1971]
was able to characterize the form of an optimal policy when the order cost
function is concave. For his proof he required the demand density to be a
one-sided Polya frequency function. (The definition and the properties of
Polya frequency functions are diseussed in detail by Karlin [1968].) Porteus
calls the optimal policy a generalized (s, S) policy. If the order cost function is
piecewise linear concave then the generalized (s, S) policy has the following
form. If x is the starting inventory in any period and y ( x ) the inventory after
ordering then there exist 2r numbers satisfying s r < St_ 1 < " " " < S 1 ~ S 1 <
S2<...<S r s u c h that

y(x) = {!
r for x < S r ,
i forsi+l«-x<s»
for x > s 1 .
l<~i«-r,

The optimal order-up-to quantity, y ( x ) , is a more complex non-linear

 Ch. 1. Single-Product, Single-Location Models 33

function of the starting inventory x when the order cost function is an arbitrary
concave function.

4.3.3. Approximations
Optimal (s, S) policies may be computed by either successive approxi-
mations using the functional equations [Wagner, O'Hagen & Lund, 1965], by
policy iteration methods [Federgruen & Zipkin, 1984], or by Markovian
methods [Karlin, 1958]. However, these calculations are generally complex and
time consuming. For that reason approximate (s, S) policies are of interest.
Several researchers have suggested methods for computing approximate
(s, S) policies. One is based on a continuous review analogue (We discuss
continuous review models in the next section.) We will not review all of the
methods here but note that Porteus [1985a] has performed a computational
comparison of 17 methods. He found that several methods (in particular, two
based on his own prior work with Freeland) seemed to perform best. As he
notes in his conclusion, stationary (s, S) policies are of only limited interest in
practice since the distribution of demand is time varying in most real environ-
ments.
From our experience, practitioners seem to favor the continuous review
model discussed in the next seetion and simply recompute the lot size and
reorder point on a periodic basis as new estimates of the mean and standard
deviation of demand are made. These estimates are generally obtained using a
forecasting tool such as simple exponential smoothing.

4.4. Service constraints in periodic review systems

It is orten difficult for management to accurately estimate shortage costs.

Shortage costs should include both direct costs and indirect costs of shortage.
Direct costs include clerical expenses of keeping track of unfilled orders and
lost or deferred profit from sales. Indirect costs include disruptions that might
result elsewhere in the firm due to shortages of key equipment and loss of
customer goodwill. The consequences of loss of customer goodwill are very
difficult to gauge. In particular, customers may choose to go elsewhere which
means that the pattern of future demand is altered.
Primarily for this reason service levels are very popular in practice. Service
levels may be defined in several ways, but two seem to be the most common:
(1) the proportion of periods in which all demand is met;
(2) the proportion of demand satisfied immediately from inventory.
Refer to these as Type 1 and Type 2 measures of service, respectively. One
offen sees policies determined assuming a Type 1 service level because of its
simplicity, but Type 2 service is what one commonly means by service. In order
to understand the difference between these two measures consider the follow-
ing simple example from Nahmias [1989]. Suppose that over ten order cycles
the following has occurred:
34 H . L . Lee, S. Nahmias

Order cycle Demand Stockouts

1 180 0
2 75 0
3 235 45
4 140 0
5 180 0
6 200 10
7 150 0
8 90 0
9 160 0
10 40 0

Based on a Type 1 service, there were 8/10 -- 80% of the periods in which no
stockout occurred. However, the Type 2 service measure is rauch higher. The
total n u m b e r of demands over the ten periods is 1450 (the sum of the entries in
column 2) while the total number of demands which could not be met
immediately was 55. The proportion of satisfied demands was 1395/1450=
0.9621 or slightly more than 96%.
Clearly these measures of service are quite different. As a result, they yield
very different operating policies. Other service measures are possible as well.
For example one might wish to minimize time weighted backorders or the
probability that stockouts do not exceed some critical level.

4.4.1. Service levels when set-up cost is zero

Optimal policies subject to service level constraints for periodic review
systems with no order set-up cost are easy to find. Let/31 be the desired level of
T y p e 1 service and /~2 the desired level of Type 2 service (assume both are
expressed as ffactions). Because F(t) is the probability distribution of one
period's demand and is the probability that demand is less than or equa! to t, it
follows that the order up to level, say S, should satisfy

F(S) = ~,

for a Type 1 service measure.

For Type 2 service, define n(S) as the expected number of unsatisfied
demands at the end of a period when ordering to S. Then

n(S) = E[(D - S)] + = f f (x - S ) f ( x ) d x ,

where D is the demand during one period.

If the expected demand per period is A, then n(S)/A is the expected ffaction
of demands not satisfied in one period. Since J~2 is the average ffaction of
demands met in one period, it follows that

n(s)/a = 1 - t~~.
 Ch. 1. Single-Product, Single-Location Models 35

When lead time demand is normal with mean I r and standard deviation o-r,
n(S) = o-rL((S - At)/~rr) where L ( z ) is the standardized loss integral for which
both tables and approximations are available. [See Brown, 1967 for example.]

4.4.2. Service levels when set-up cost is positive

Since set-up costs are common, it is more important to be able to find
optimal and suboptimal policies in this case. Unfortunately, computations are
just as complex for service levels as they are when penalty costs are used. For
this reason approximations are of interest hefe as well.
One approach for finding approximate (s, S) policies satisfying a service
level constraint is to use the (Q, R) approximation assuming continuous review
and set s = R and S = R + Q. Continuous review policies subject to service
level constraints are discussed in the next section.
While the results obtained from this approach are reasonable in many
circumstances, more accurate approximations may be required. Several re-
searchers have suggested (s, S) approximations subject to service levels. Their
methods are based on the results of Roberts [1962].
Roberts considered an asymptotic analysis of the underlying renewal pro-
cess. H e showed that for the cost model when both K and p were large one
could approximate (s, S) by

S - s ~ ~ + B T ,
and

p ( t - s)f(t) dt = ~ ,

assuming no discounting and an order lead time of T periods. Interpret a as the

expected value of one period's demand and BT a constant depending only on
the length of the lead time. Similar, but somewhat more complex, formulas
were obtained for the case where costs are discounted.
Roberts' results were first extended to consider service levels by Schneider
[1978] and later refined by Tijms & Groenevelt [1984]. They suggest for a Type
2 service criterion that the reorder level, s, solve

£~ (t - s)2fr+l(t) dt = "tI ,

where

= (1 -- /32)2A[S - s + A2/2A ]

and

Az = t2f(t) d t .
f0+
36 H.L. Lee, S. Nahmias

Interpret /32 as the desired service level and f r + l ( t ) as the density of the
T + 1 fold convolution of one period's demand. Tijms and Groenevelt use
these results to obtain approximations for s based on the first two moments of
the distribution of one period's demand. The authors present extensive compu-
tations and discuss under which circumstances their approach is accurate.
Recently Cohen, Kleindorfer & Lee [1988] considered an extension of these
results to lost sales and priority classes of demand.

5. Continuous time stochastic demand models

Continuous review has also been referred to as transactions reporting [see,

e.g., Hadley & Whitin, 1963], because all demand transactions are recorded as
they occur. A supermarket with an optical scanner connected to a centralized
inventory eontrol system is an example of such a system. In this system, all
transactions are monitored continuously and thus inventory ordering deeisions
can be made as soon as these transactions occurred. Hence, such a system will
be more responsive than the periodic review system, where inventory ordering
actions can only take place at the specific review times. Such responsiveness is
achieved, however, at the expense of a continuous-time monitoring system.
When the inventory system is monitored on a continuous time basis, the
analysis of such system requires the specification of the demand process to the
system. Consider a demand process where the interarrival times between
demands form a sequence of IID random variables with a finite mean l/A, and
with density q,(.) and CDF ~ ( . ) . Any sequence of IID random variables
forms an ordinary renewal process. Assume that the quantity demanded at
each occurrence of demand is a random variable. By treating stationary
continuous time demand with stationary costs as a limiting case of the periodic
review models discussed above, we can argue, heuristically, that an (s, S)
policy is optimal. This result ean be proved rigorously [see Beckmann, 1961;
Hadley & Whitin, 1963.] Hence, the management problem of interest is the
determination of the optimal (s, S) values that would minimize some long run
average cost measures.
In this section, we will first present the cost model based upon which the
(s, S) values are to be found. A heuristic solution procedure to find near-
optimal (s, S) values is then described. To compute the exact value of the
expected cost components in this cost model, however, we need to find the
steady state behavior of several key operating characteristics of the system.
One such key characteristic is the inventory position (on hand plus on order).
We first show how this operating characteristic can be found in the case of
general renewal process demands. We then show how the model can be
considerably simplified for the cases of Poisson demands, and the one-for-one
ordering policy (where s = S - 1). Besides useful to compute the cost function,
knowledge of the distribution of the inventory position is orten sufficient to
determine other characteristics of interest as weil. One such characteristic, the
waiting time distribution, will be examined.
 Ch. 1. Single-Product, Single-Location Models 37

5.1. The cost f u n c t i o n and heuristic m o d e l s

The long run average cost is orten used as the performance measure for a
continuous review inventory system, Similar to periodic review models, this
cost consists of order, inventory holding, and shortage costs. Let K be the
order cost, h be the unit holding cost per unit time, p be the unit shortage eost
(for each unit that is backordered or lost as a result of shortage), and p' the
cost per unit time that a unit is backordered (for the case of backordering of
unmet demands). Then, the average cost C equals:

C = KN +pM + hH +p'B, (5.1)

where N is the expected number of orders placed per unit time, M is the
expected number of units short per unit time, H is the expected on-hand
inventory level, and B is the expected backorder level. The problem is then to
find N, U, H and B.
Instead of finding the exact values of N, M, H and B, an approach is to find
an approximate expression for C in (5.1), and use heuristics to find the optimal
values of (s, S). Here we describe simple, yet effective ways to find the optimal
(s, S) values in the backorder case where p' = 0, and the lost sales case [see
Hadley & Whitin, 1963].
Let Q -- S - s. The (s, S) inventory control system is sometimes referred to
as the (Q, r) inventory control system, where the 'r' corresponds to the 's'.
The problem is to find optimal values of Q and s. Define a replenishment cycle
as the length of time between two successive points in time where orders are
placed. The approximate cost function is based on assuming that there is at
most one order outstanding, and demands arrive one at a time. It is also based
on treating replenishment cycles as renewal cycles so that the order and
shortage components of the average cost can be computed as the ratio of the
expected order and shortage tost per replenishment cycle and the expected
length of a replenishment cycle. Such a treatment, of course, is only an
approximation.
Let )~ be the average demand per unit time, ~" be the random variable
denoting the lead time demand. Suppose unmet demands are backordered.
The expeeted number of replenishment cyeles per unit time is 2t/Q. In each
cycle, the expected number of units backordered is E ( f f - s ) +, and so the
expected units backordered per unit time is ( ) t / Q ) E ( f f - s) +. The expected
ending inventory at the time when an order arrives is E ( s - ~'), and the
expected inventory level after an order arrival is Q + E ( s - ~). The expected
on-hand inventory level is then approxirnated by the average of these two
inventory levels:

C = KA/Q + p(h/Q)E(« - s) + + h[s - E(~r) + ½QI.

Setting the first derivatives of C with respect to s and Q, respectively, to O,
we get:
38 H.L. Lee, S. Nahmias

Q = ~/2A[K + pE(~" - s)+]/h, (5.2)

Pr{~"/> s} = Qh/pA. (5.3)

An iterative procedure can then be used to find the optimal values of s and
Q. The procedure starts with using an initial value of Q (e.g., the EOQ) to
obtain s from (5.3). From this value of s, a new Q can be computed from (5.2),
which can then be used in (5.3) to update s. The procedure ends when the
values of s and Q stabilize.
A similar approximate cost function can also be developed for the lost sales
case. Here, with at most one order outstanding, the expected ending inventory
at the time just before a replenishment order arrives is given by E(s - ()+. The
expected inventory level right after the order arrives is Q + E ( s - ~)+. The
average inventory level is obtained by taking the average of the two, i.e.,
( Q / 2 ) + E(s - ~)+. The expected number of shortages in a cycle is E(~" - s) +.
Hence, the expected quantity demanded in a replenishment cycle is Q +
E ( ~ ' - s ) +. The expected number of orders placed per unit time is thus
A/[Q+E(~-s)+]. As E ( ~ r - s ) + is usually much smaller than Q, one
can approximate the expected number of orders placed per unit time by A/Q.
The approximate cost function is thus given by:

C= A[K + p E ( « - s)+]/Q + h[½Q + E ( s - •)+].

Setting the derivatives of C with respect to Q and s to zero yield the same
expression for Q as in (5.2), and:

Pr(« >~s) = Qh/(Qh + Ap). (5.4)

A similar iterative procedure can then be used to obtain the optimal values
of s and Q.
There are some interesting observations that one can draw from the heuristic
models. First, from (5.2), the optimal value of Q must be larger than or equal
to the E O Q , i.e., a larger order quantity is needed as a result of uncertainty in
demand and the penalty cost of shortages. The reason for this is as follows.
The inventory system will be vulnerable to shortages when we have placed an
order and it has not arrived. A larger order quantity would mean that the
above situation will happen less frequently. Second, from both (5.3) and (5.4),
the optimal value of s is non-increasing in Q. Hefe the interpretation is that,
when the lot size is larger, it is less frequent that the inventory system will run
into the position in which it is vulnerable to shortages. Thus, the need to have
a higher reorder point for safety purposes is reduced, and so the optimal value
of s is smaller.
When there is a service criterion, such as the two service measures of Section
4, the usual way is to add the service requirement and adjust the iterative
procedure above. A description of these heuristic methods can be found in
Nahmias [1976] and Yano [1985].
 Ch. 1. Single-Product, Single-Location Models 39

5.2. Renewal process demand models

In this section, we focus on the backorder case with p = 0. The major

concern is then to find the exaet values of H and B. It turns out that, in the
case of backordered demands and constant lead times, the on-hand inventory
level and the backorder level are related to the inventory position. Hence, the
steady state distribution of the inventory position in continuous review inven-
tory systems has been an important topic of research.
To see the relationship between the on-hand inventory level and the
backorder level with the inventory position, let V(t) and l(t) be the installation
inventory level and the inventory position at time t respectively. Let T be the
replenishment lead time, which is a constant. Let also D ( t - T, t) denote the
demand during the interval ( t - T, T]. Thus, the on-hand inventory level and
the backorder level at time t are given by max{V(t), 0} and max{-V(t), 0},
respectively. Consider l ( t - T). Note first that all the orders outstanding that
made up I ( t - T ) would have arrived by time t, and thus all the items in
I ( t - T) would contribute to the inventory level V(t). The demands in ( t -
T, T], however, will deplete this inventory level. Hence, we have the following
relationship:

V(«) = I ( t - T ) - D ( t - T, t) . (5.5)
The relationship above shows that V(t) can be obtained from l ( t - T) and
D ( t - T, t). Under the (s, S) policy, when the inventory position reaches or
drops below the reorder level s, an order is immediately placed so that the
inventory position is brought back to S. Hence, the inventory position is always
between s + 1 and S. There are several methods from which the steady state
distribution of the inventory position can be found.
Sivazlian [1974] traced the transient behavior of the inventory position, and
obtained the following result for the unit demand case. Without loss of
generality, suppose we are now at time 0, where an order has just been made,
bringing the inventory position up to S. Let Q = S - s. Denote 0Q( • ) and
~ Q ( . ) as the density and CDF of the Q-fold convolution of 0(" ), respectively.
The convention 0°(0) -- 1, i.e., qr0(. ) = 1 will be used. Sivazlian shows that,
for n= l , . . . , Q,

Pr{l(t) = s + n} = ~ [~ke+e n(t) _ q~ke+O-n+l(t)]

k=O

Using the Laplace transform of the probability function of I(t) and taking the
limit as t---~ % it can be shown that limt~~ Pr{I(t) = s + n} = 1/Q. The remark-
able result is that, in the steady stare, the inventory position is uniformly
distributed in Is + 1, s + 2 , . . . , s + Q], and is independent of the distribution
of the interarrival times between demands. This simplifies tremendously the
derivation of the distribution of on-hand inventory and backorder level from
(5.5).
40 H . L . Lee, S. Nahmias

Instead of unit demand, suppose now that the quantity demanded in each
demand event is an liD random variable greater than or equal to one, and
which has a finite mean. Ler the probability function and cumulative dis-
tribution of each demand quantity be given by ~b(. ) and q)(. ) respectively. Let
also r(. ) and R(. ) denote its renewal function and the associated distribution
function respectively. Under the (s, S) policy, the inventory position at the
time when an order is triggered may be below s. The quantity of the difference
between s and the inventory position at this time is usually called the
undershoot. It turns out that the uniformity of the distribution of the inventory
position may no longer hold.
Let the steady state distribution of the inventory position be given by fr(x),
x E Is, S]. Richards [1975] used a simple approach that focused on the expec-
ted holding time of the inventory position within each replenishment cycle to
obtain:

fr(x) = r ( S - x)/[1 + R ( S - s)].

He also showed that this distribution is uniform if and only if demands are of
unit size.
A similar result for the steady state distribution of the inventory position has
also been developed by Tijms [1972] and Sahin [1979], using a different
approach. First they obtained the probability function of the inventory position
at a given point in time, t. The probability distribution of a replenishment cycle
was shown to be

1 - ~ [gtk( • ) - 't//k+l( " )] (J~k(s -- S),

k=O

with expected length given by [1 + R(S - s)]/A. Let m(. ) denote the renewal
density of the cycle. Consequently, limx__,= m(x) = A/[1 + R(S - s)]. Next,

Pr{x ~< I(t) ~ x + dx}

c~

=- Z [ a / t k ( / ) -- ~[tk+l(t)]~)k( S -- X)
k=l

+~=lfom(t-u)[gtk(u)-qtk÷l(u)]4)k(S-x)du.

Taking the limit of the above expression as t-+ 0% one obtains the steady state
distribution of the inventory position, which is the same as the one obtained by
Richards. Stidham [1974] has also noted that the (s, S) system is a special case
of a 'stochastic clearing system' from which the steady state distributions can be
obtained.
The uniform distribution of the inventory position can, however, be re-
captured by modifying the procurement policy slightly. Consider the following
 Ch. 1. Single-Product, Single-Location Models 41

policy, termed the (s, nQ) policy, that orders an amount of nQ, n = 1, 2 , . . . ,
when the inventory position falls to or below s. The value of n is determined so
that the inventory position after ordering is between s + 1 and S = s + Q.
Here, the inventory position after ordering is itself a random variable, and
depends on the magnitude of the undershoot at the time when an order is
placed. The (s, nQ) policy and the (s, S) policy are the same when demands
are all of unit size. Consider now the embedded Markov chain of the inventory
position at the times of customer arrival. Specifically, let Wn denote the time of
arrival of the nth customer, n = 1, 2 , . . . Define the stochastic process
{Yn, n = 0 , 1 , 2 , . . } , where Yo = I(0), and Yn=I(Wn), n = l , 2 . . . . . It fol-
lows that { Yn} is a finite state discrete Markov chain. The one step transition
probability matrix is given by:

(a(mQ + k - j ) if j < k ,
pik = Pr{Y,,+~ = k [ Y ~ = j } = ~ ~

L~o~b(mQ + k - j ) if j>~k.

This Markov chain is irreducible if ~b(1)> 0. Noting that the matrix {Pik} is
doubly stochastic, Simon [1968] shows that the stationary distribution of the
inventory position immediately following a customer arrival, and consequently,
the steady state distribution of the inventory position, is uniform.
Given the steady state distribution of the inventory position, we now turn to
the derivation of the steady state distribution of the inventory level from (5.5).
Note first that I(t - T) and D ( t - T, t) will be independent only if t - T is a
customer arrival point or if the customer arrival process is Poisson. Sahin
[1979] derived expressions for the transient probability of V(t). Let V(oo), I(oo)
and D(~) denote random variables having the limiting distributions of the
inventory level, inventory position, and demand in lead time, respectively.
Sahin [1979] proved that I(~) and D(o0) are independent. Therefore, from
(5.5), if I(~) and D(~) exist, we have:

V(~) = I(c¢) - D ( ~ ) . (5.6)

Thus, the distribution of V(oo) can be obtained ffom those of l(o0) and D(~)
by simple convolution. The distribution of I(oo) is 7r(. ). D(w) is related to an
equilibrium cumulative renewal process, and has density g(y) given by:

g(y) = A ~(U)][~k-~(T -- U) -- q/k(T -- U)] d u ,

~~ thk(y) ~;0~[1 -
y>O,

g(0) = A , /Io = [ 1 - ~ ( u + T)] d u .

42 H.L. Lee, S. Nahmias

Zipkin [1986] gave a rigorous treatment of when the relationship given by

(5.6) and the independence property would hold. In the case of stochastic lead
times, we need to construct a similar relationship as (5.6). Suppose lead time ~-
is a random variable, and that D~(oo) denote the lead time demand in the limit,
given that the lead time is ~-. The standard approacb [see Hadley & Whitin,
1963] is then to find the marginal distribution of D,(oo) over the range of T. Let
D(o~) denote the variable with such a marginal distribution. We then obtain
V(~) from (5.6).
The assumptions that justify (5.6) are that orders do not cross and their lead
times are independent. One way to ensure this, due to Zipkin [1986], is to
focus on the process by which orders arrive. Consider a real-valued, stationary
stochastic process {U(t): t E ~} satisfying the following assumptions:
(1) U(t) >10 and E [ U ( t ) ] < c o ;
(2) t - U(t) is nondecreasing;
(3) sample paths of {U(t)) are continuous to the right;
(4) {U(t)) is independent of the placement and size of orders and of
d e m a n d process.
Suppose an order is placed at time u and that it arrives at time z =
min{t: t - U ( t ) > ~ u}. This way, it is clear that lead times are nonnegative.
M o r e o v e r , orders never cross, since if u 1 < u2, then, whenever { t - U ( t ) )
passes the value u 1, the process forces the order to arrive at time t, where
t = U(t) + ul, but t - U(t) = ul < u 2, so that the second order cannot arrive at
t.

5.3. Poisson d e m a n d s

In the last section, we have seen how the steady state distribution of the
inventory position can be used to compute the steady state distribution of the
on-hand inventory and backorder level. An important renewal demand process
in inventory modelling is the Poisson process. Continuous time inventory
models are relatively straightforward when demand is Poisson. Models with
more general lead time assumptions can be studied. Expressions for the cost
function can be derived explicitly when demand is Poisson, so that algorithms
to find the optimal (s, S) values can be readily found [see Beckmann, 1961;
Archibald & Silver, 1978].
As with periodic review, the lost sales case is much more difficult to analyze.
T h e renewal approach of the previous section fails here. The major problem
arises in the transition of the inventory position when the o n - h a n d stock
becomes zero. When demand occurs and the system is out of stock, the
inventory position remains unchanged. Nevertheless, it should be easy to see
that the maximum number of outstanding orders at any point in time in lost
sales systems is [ S / ( S - s)J, where [xJ denotes the largest integer smaller than
or equal to x. Thus, when S - s > s and unmet demands are lost, there can be
 Ch. 1. Single-Product, Single-Location Models 43

at most one order outstanding at any point in time. In that case, the inventory
position and the inventory level are the same just before a replenishment order
is placed. As a result, with Poisson demands, constant lead time, and the
values of (s, S) considered satisfying Q = S - s > s, exact results for some
operating characteristics of the system can be obtained [see Hadley & Whitin,
1963]. When demand arrivals are Poisson and the order quantity from each
customer arrival is a random variable (known as compound Poisson, discussed
also in Section 5.3), one can obtain exact results for the lost sales case with at
most one order outstanding [Archibald, 1981].
When replenishment lead times have the negative exponential distribution,
exact results can also be obtained. Order crossing is not a problem in this case.
One utilizes the C h a p m a n - K o l m o g o r o v equations for the steady stare prob-
abilities as in standard queueing systems. For the simple case where the arrival
rates and the replenishment times are independent of the state of the system,
Galliher, Morse & Simond [1959] defined the system state as the number of
orders outstanding at the instant in question, k, and the number of demands
which have arrived since the last order went out, n. Let Pkn be the associated
steady state probability and 1//x the mean replenishment lead time. Let
Q = S - s. Then Pkn is the solution to:

0 = Apk,n_ ~ + (k + 1)/xp~.~,n - (A + klx)pk, , , 0 < n < Q,

0 =/xp~ 0 -- Ap00 ,

0 = Apk_l,Q_ ~ + (k + 1)/xpk+l,0- (A + k t x ) p e o , k>0.

The operating characteristics of the inventory system are then obtained, as a

function of Pkn"
The problem of state dependent lead times can be solved as well when
demands are Poisson. Gross & Harris [1971] first analyzed the problem when
S - s = 1, and later extended their analysis to the general (s, S) system in
Gross & Harris [1973]. Each order placed by the inventory system is assumed
to be processed by a single server queueing system, where the service rate is
state dependent. Two types of dependencies of lead times on system states
were considered: (1) the instantaneous probability that an order will finish
service in an infinitesimal interval dt is/x(k) dt + o(dt), where k is the number
of outstanding orders, and (2) the service time for an order is exponentially
distributed with mean 1//x(rn), where m is the number of outstanding orders
just after the previous order has arrived. In both cases,/x(m) is specified to be
/x1 if m = 1, and /x2 otherwise. The first case is solved by using standard
C h a p m a n - K o l m o g o r o v approach as above, whereas the second case is solved
by deriving the steady state distribution from the limiting distribution of the
imbedded Markov chain generated by the departure points of the service
system.
44 H.L. Lee, S. Nahmias

Wijngaard & Van Winkel [1979] considered a more general problem where
the Poisson arrival rates and exponentially distributed lead time rates are both
dependent of the net inventory level of the system. Their analysis is based on
the average behavior of the inventory level in a cycle, which is defined to be
the successive points in time when the inventory level hits zero.

5.4. Order-for-order inventory policies

An important class of ordering policies is the order-for-order policy, known

also as the one-for-one replenishment policy, or the (S - 1, S) policy. This is a
special case of the (s, S) policy with s = S - 1. Here, any time when a customer
demand occurs, an order is made to return the inventory position to S. Such
policies are usually used for low-demand items where the order set-up cost is
small compared to the shortage and holding costs. When demands are Poisson,
explicit results can be obtained for both the stochastic lead time case where
orders can cross and for the lost sales case. These results are possible because
of the strong similarities between these inventory systems and conventional
queueing systems.
For an (S - 1, S) inventory system, the on-hand inventory at time t is given
by [S - W(t)] +, where W(t) is the number of orders in resupply at time t. The
number of backorders at time t is [ W ( t ) - S] +. Consider the steady state
distribution of W(t). Note that the number of customers in an M / G / ~
queueing system is the same as the number of outstanding orders in an
(S - 1, S) inventory system with Poisson demands, complete backlog of un-
satisfied demands, and arbitrary but independent lead times. Let A be the
Poisson arrival rate, and T be the mean replenishment lead time. By a result
due to Palm [1938], the number of customers in an M / G / ~ queueing system,
and consequently, the number of outstanding orders in the inventory system, is
Poisson with mean AT. Hence, this distribution only depends upon AT, and
does not depend on the distribution of the replenishment lead time! Given the
distribution of the number of orders in resupply, it is straightforward to
develop the performance measures for the inventory system. In the steady
state, the on-hand inventory is distributed as ( S - W) +, where W is Poisson
with mean AT; whereas the number of backorders is truncated Poisson, i.e.,
( w - s) +.
The number of customers in an M / G / S / S queueing system is the same as the
number of outstanding orders in an (S - 1, S) inventory system with Poisson
demands, arbitrary but independent resupply lead times, and lost sales for
shortages. From standard queueing theory, the steady state probability that the
number of customers in the queueing system is k, is

[(AT)k/k!
,/~~0 (AT)i/i!
] for0~<k~S.

Let W be a random variable with this probability distribution. The probability

 Ch. 1. Single-Product,Single-LocationModels 45

that the inventory system will have no on-hand stock is the same as the
probability that all servers in the queueing system are busy. Hence, the
expected order cost per unit time is K A I 1 - P r { W = S}]. The expected shor-
tages per unit time is APr{W = S}, and the expected on-hand stock is again
given by E ( S - W ) ÷ = E ( S - W ) . Smith [1977] described ways to obtain
optimal values of S for such inventory systems.
Suppose arrivals of customers are Poisson with rate A. The amount de-
manded by each customer is an IID random variable with ~b(j) denoting the
probability that the demand is j, j > 0 . Such a demand process is called
C o m p o u n d Poisson, and ~b(j) is sometimes referred to as the compounding
distribution. If ~b(0)>0, then we can redefine A ' = A [ 1 - ~ b ( 0 ) ] , and
4~'(j) = ~b(j)/[1 - ~b(0)], j > 0. This way, we ensure that the probability is zero
that the amount demanded ffom a customer is zero. When 4~(j) = (1 - p)pJ-~,
j > 0, i.e., a geometric compounding distribution, then the demand distribution
is called a stuttering Poisson. D e n o t e p(k; T) as the probability that the
demand in an interval of length T is k. Then:
k
p(k; r)= E Pr{k demanded Irom y customers}Pr{y arrivals in T )
y=l
k
=E
y=l
fóY(k)( A T ) y e x p ( - A T ) / y ! . (5.7)

F e e n e y & Sherbrooke [1966] extended Palm's T h e o r e m to the compound

Poisson demand case. T h e y assumed that, each time an order is issued as a
result of a customer arrival, all the items in that order will be resupplied as a
batch, i.e., no order splitting by the supplier is allowed. The order lead times
are all IID. Let h(k) be the steady state probability that there are k items in
resupply. For the complete backorder case, it can be proven that h(k)=
p(k; T). With lost sales two cases are considered. The no partial fill case is one
such that, if a customer arrives and finds the number of stock on hand is less
than what h e / s h e demands, then all the demands are lost. For this case,

h(k)=p(k; T) p(i; T), O<~i<~S.

The partial fill case is one such that only the portion not satisfied in a
customer's demands would be lost. For this case,

h(k)=p(k;T)/A(S), O<~k<S,

h(S) =
I~o
-
[(/~ T ) y e x p ( - A T ) / y ! ]
~ 1/
i=s
~)Y(i) A(S),

where A(S) is a normalizing constant.

46 H.L. Lee, S, Nahmias

5.5. Waiting time distributions

Besides average cost, a performance measure of interest is the waiting time

distribution of customer orders. The waiting time is the amount of time a
customer must wait for his/her order to be fulfilled. For the general renewal
demand models, the usual approach to deriving the waiting time distribution is
to establish the equivalence of the event that a customer has to wait a certain
length of time, to some other event involving demands, the inventory position
or the n u m b e r of orders outstanding, and the waiting time. Because the steady
state distribution of the inventory position and the number of orders outstand-
ing can be obtained as in previous subsections, the waiting time distribution
can be easily derived.
Suppose s/> - 1 so that no customer will ever wait more than the lead time
T, assumed to be constant. Consider the probability that a customer arriving at
time t would have to wait for more than x time units, i.e., the customer is still
waiting at time t + x. Note that all the outstanding orders that are part of the
inventory position at time t + x - T would have arrived by time t + x, and can
thus be used to satisfy the customers arriving prior to that time. H e n c e , the
event that a customer arriving at time t will wait for at least x time units is
equivalent to the event that demand in (t + x - T , t] is greater than the
inventory position at time t + x - T. Because the steady state distribution of
the inventory position is known (see Section 5.1), one can obtain the dis-
tribution of waiting time for the general (s, S) inventory system [Kruse, 1981].
In the case of ( S - 1, S) inventory policy with Poisson demands, stronger
results exist. Higa, Feyerherm & Machado [1975] first studied this problem
with stuttering Poisson demands and negative exponential lead times. Unlike
other studies assuming compound Poisson demands in this case, it is assumed
that each individual item within the same batch has an independent lead time.
Waiting time is defined to be the length of time for all demands within a
customer order to be met. Let W(t) be the number of items in resupply at time
t. The approach is based on observing that the waiting time of a batch of k
units arriving at time t is zero if W ( t ) + k <~S, and is the (W(t)+ k - S)th
order statistic in the set of resupply times for the W(t) + k items. An explicit
expression was then developed for this probability distribution by noting that
the steady state distribution of W(t) is also compound Poisson. For the
constant lead time and Poisson demand case assuming an (S - 1, S) policy, the
expression can be quite simple. Sherbrooke [1975] showed that the probability
that the waiting time is less than or equal to w time units is simply a sum of
Poisson probabilities from 0 to S - 1 , where the mean of this Poisson prob-
ability is a function of w.
Higa, Feyerherm & Machado's result is not entirely correct when a First-
Come-First-Serve (FCFS) discipline is used to satisfy customer demands.
Suppose, as before, a customer arrives at time t. While this customer is waiting
for his/her order, there could be arrivals after time t. Because of stochastic
lead time, the orders generated by these arrivals could have arrived during the
 Ch. 1. Single-Product, Single-Location Models 47

waiting time of the customer arriving at t, which could be used to satisfy the
customer arriving at time t, under FCFS. Hence, the equivalence of the two
events used by Higa, Feyerherm & Machado described above does not hold
exactly. Kruse [1980] corrected for this difference with Poisson demands and
( S - 1, S) policies, and extended it to general IID lead times. A new event
equivalence has to be established: the event that a customer arriving at time t
would have to wait for more than w time units is equivalent to

W ( t ) - rl(t, t + w) - r2(t, « + w) - 6(w) >t S ,

where W ( t ) is the number of outstanding orders at time t, rl(t, t + w) is the

amount of outstanding orders generated in (0, t] which are filled by the
supplier in (t, t + w], r2(t , t + w) is the amount of outstanding orders generated
in (t, t + w] which are filled by the supplier in (t, t + w], and ~(w) is 1 if the
order triggered by the customer arrives by w, and 0 otherwise. Hefe,
Pr{6(w) = 1} is the probability that the lead time is less than or equal to w.
The waiting time distribution is obtained by using results of infinite-server
queueing systems.
The effect of waiting time on customer balking has been studied by Das
[1977]. The model assumes Poisson demands, ( S - 1 , S) policies, negative
exponential lead times, and that a customer would cancel its order after waiting
for a fixed time period. The steady stare probabilities of the number of
outstanding orders are solved by using Chapman-Kolmogorov equations.

6. Application of inventory theory in industry

Some of the fundamental concepts imbedded in the inventory models

described in this volume have been used extensively in practice. With the
advance of computer technology, many software systems for the control of
inventory have been developed, which have incorporated these concepts.
However, it is difficult to assess accurately the extent to which inventory
models have been used in industry, because documentation of its use in
publicly available literature is rare. In this section, we describe how the
inventory models described in previous sections are used in practice, as
described in published studies in the literature.

6.1. Use o f E O Q

The EOQ model is perhaps the most widely used in practice. In 1961,
APICS (American Production and Inventory Control Society), in conjunction
with Factory Magazine, surveyed the Society's members, and found that the
use of 'modern' analytical techniques such as EOQ was extremely rare [see
Factory Magazine, 1961]. A majority of the respondents cited 'pure judgement'
as the method used for determining inventory ordering and safety stock
48 H.L. Lee, S. Nahmias

decisions. However, the use of analytical methods steadily increased [Factory

Magazine, 1966]. In 1973, a similar survey found that 56% of the respondents
were using EOQ, and 32% were using some form of statistical analysis for the
calculation of the re-order point [Davis, 1975]. In a 1978 survey of manufactur-
ing firms in Arizona, California, Illinois, New York and Pennsylvania, 85% of
the respondents reported using EOQ [Reuter, 1978a,b]. Such a finding is
reconfirmed by a recent study by Osteryoung, MacCarty & Reinhart [1986], in
which 84% of the firms surveyed, which included manufacturing, retail,
services and wholesale/distributers, reported using EOQ. Of these firms, 48%
reported that the E O Q usage was for finished goods inventory control.
On the other hand, the E O Q has often been criticized. In previous sections,
we noted that most of the assumptions required in the derivation of the EOQ
formula do not hold. Similar objections have been made about the other
inventory models as weil. A major problem reported by practitioners is the
inability of most firms to accurately assess ordering and carrying costs ]Zeit-
zew, 1979]. Gardner [1980] found that the fundamental problem of inventory
practice is accurately estimating the cost parameters. Most of the cost parame-
ters are estimated by standard accounting procedures that generate at best
average costs, yet it is marginal costs that matter in inventory control.
Schonberger & Schniederjans [1984] found two major flaws with inventory
models. First, most costs are difficult to estimate. It is also difficult to
incorporate the hidden costs such as the quality costs related to large batches,
i.e., the longer response time to detect quality problems when batch sizes are
large. Second, many inventory models are static. For example, the EOQ
analysis assumes costs are stationary. However, many of these costs are
dynamic; that is, time varying. Moreover, by proper investments, some of
those costs (e.g., set-up costs) can be reduced over time, resulting in smaller lot
sizes.
Despite its shortcomings, the basic E O Q model remains the most widely
used. It is also the cornerstone of many currently commercially available
software packages for inventory control.

6.2. I B M ' s inventory package I M P A C T

In the early sixties, IBM introduced an inventory package called IMPACT

(Inventory Management Program and Control Techniques) [IBM, 1967]. The
package has steadily gained popularity among wholesalers and retailers, many
of which have adapted it to their particular needs. It contains the logic for
inventory control, as weil as complete inventory file structure and forecasting
systems. In the seventies, an enhancement of IMPACT, called INFOREM
(Inventory Forecasting and Replenishment Modules) was developed, also by
IBM [IBM, 1978]. This new system employs an improved forecasting tech-
nique, and includes simulation for forecasting and strategic planning of inven-
tory policies. IMPACT was an important step in the advancement of computer-
ized inventory control systems in industry. Many of the commercially available
 Ch. 1. Single-Product, Single-Location Models 49

inventory control software systems actually employ logics similar to those of

IMPACT. We describe the extent to which inventory models reviewed in this
chapter are used in IMPACT.
IMPACT assumes a periodic review, reorder point, fixed lot size inventory
control systems similar to the orte of Section 4. The lot size is assumed to be
the EOQ. The reorder point is the sum of mean lead time demand plus one
review period's demand plus safety stock. Such an operating system is exactly
the approximation of (s, S) policies described in Section 4.4.2. The safety stock
equals a safety factor multiplied by the standard deviation of the demand in
lead time plus a review period. Determination of the safety factor is based on
the following assumptions:
(1) demand per period is normally distributed;
(2) a target service level based on the expected amount of demands filled
over the expected demand, i.e., the Type 2 service measure used in Section
4.4.1, can be specified;
(3) unmet demands are backordered;
(4) at most one order is outstanding at all times.
The safety factor is chosen to most nearly equate the percentage of demands
filled immediately in an order cycle and the prespecified target service level.
Kleijnen & Rens [1978] conducted simulation studies and found that IM-
PACT often provides overservice; that is, the simulated services using the
reorder points of IMPACT are greater than the prespecified targets. Kleijnen
and Rens attributed this difference to: (1) unmet demands can be lost, and (2)
the approximation of a periodic review system by a continuous review (Q, r)
system requires proper adjustment of the reorder point. Actually, the criticism
of IMPACT that it does not give good reorder points when unmet demands are
lost is unreasonable, because it is a backorder model. It is easy to modify the
formulas used in IMPACT to account for lost sales. The adjustment of the
reorder point to account for the approximation of periodic review by continu-
ous review is also straightforward. Silver & Peterson [1985] suggest several
ways to make such an adjustment.

6.3. Success stories of application

There have been several successful applications of inventory models reported

in the literature. Most of these reported a significant reduction in inventory
investment while maintaining the same level of service to customers. It is
interesting to note that in many of these applications simple models were
employed, and approximations used extensively. Again, EOQ analysis pre-
vailed. We consider several such applications below.
Many applications were in the manufacturing sector. Liberatore [1979]
reported one such application for FMC's Industrial Chemical Group plants.
Here, the problem was to find ordering policies for bulk raw materials for a
chemical production process. Since the production process is continuous and
the product is produced at fairly uniform rate, the environment was ideal for
 50 H.L. Lee, S. Nahmias

E O Q analysis. The E O Q was used as the order size, with an appropriate

adjustment for full truck loads. A reorder point was used to account for the
variability of demand, even though that variability was small. A Type 1 service
measure described in Section 4.4.1, giving the probability of stockout in an
ordering cycle, was used. This criterion was appropriate because the main cost
of shortages in this environment was the cost of a shutdown of the production
process, regardless of the amount short.
An (s, S) inventory model was used for managing finished goods for Planters
Division of Standard Brands Inc. [Brout, 1981]. Again, the E O Q was used to
determine the lot size, whereas the reorder point was computed based on the
Type 2 service measure of Section 4.4.1. A similar model has been used for
managing materials for aircraft production [Lingaraj & Balasubramanian,
1983]. In this case an E O Q with an adjustment for minimum order size was
r e c o m m e n d e d , coupled with a reorder point based on Type 1 service. Similar
(s, S) systems have been employed for managing finished goods at Pfizer
Pharmaceuticals [Kleutghen & McGee, 1985].
The military has been a major user of m o d e r n inventory models. E O Q was
first implemented by the U.S. Air Force in 1952 for the purchase of expendable
weapon's system spares [Austin, 1977]. In 1958, the D e p a r t m e n t of Defense
directed all defense procurement and logistics agencies to use the basic E O Q
formula. The use of E O Q and simple approximate continuous review models
has also resulted in significant tost savings at the U.S. Navy [Gardner, 1987].
Finally, we note that there has been a significant impact of modern inventory
models on the management of blood inventories at major blood banks
[Brodheim & Prastacos, 1979].

References

Adler, G.L., and R. Nanda (1974). The effects of learning on optimal lot size determination-
Single product case. A l l e Trans. 6, 14-20.
Aggarwal, S.C. (1981). Purchase-inventory decision models for inflationary conditions. Interfaces
11, 18-23.
Archibald, B.C. (1981). Continuous review (s, S) policies with lost sales. Management Sci. 27,
1171-1177.
Archibald, B.C., and E.A. Silver (1978). (s, S) policies under continuous review and discrete
compound Poisson demand. Management Sci. 24, 899-909.
Arrow, K.A. (1958). Historical background, in: K.A. Arrow, S. Karlin and H.E. Scarf (eds.),
Studies in the Mathematical Theory of lnventory and Production, Stanford University Press,
Stanford, CA.
Arrow, K.A., T.E. Harris and J. Marschak (1951). Optimal inventory policy. Econometrica 19,
250-272.
Arrow, K.A., S. Karlin and H.E. Scarf (eds.) (1958). Studies in the Mathematical Theory of
Inventory and Production, Stanford University Press, Stanford, CA.
Austin, L.M. (1977). Project EOQ: A success story in implementingacademic research. Interfaees
7(4), 1-13.
Axsäter, S. (1982). Worst case performance for lot sizing heuristics. European J. Oper. Res. 9,
339-343.
 Ch. 1. Single-Product, Single-Location Models 51

Baker, K.R., P.R. Dixon, M.J. Magazine and E.A. Silver (1978). Algorithm for the dynamic
lot-sizc problems with time varying production capacity constraints. Management Sci. 16,
1710-1720.
Barbosa, L.C., and M. Friedman (1978). Deterministic inventory lot size models- A general root
law. Management Sci. 24, 819-826.
Beckmann, M. (1961). An inventory model for arbitrary interval and quantity distributions of
demand. Management Sci. 8, 35-57.
Bellman, R.E., I. Glicksberg and O. Gross (1955). On the optimal inventory equation. Manage-
ment Sci. 2, 83-104.
Bierman, H., and L.J. Thomas (1977). Inventory decisions under inflationary conditions. Decision
Sciences 8, 151-155.
Bitran, G.R., T.L. Magnanti and H.H. Yanasse (1984). Approximation methods for the uncapaci-
tated dynamic lot size problem. Management Sci. 30, 1121-1140.
Blackburn, J., and H. Kunreuther (1974). Planning and forecast horizons for the dynamic lot size
model with backlogging. Management Sci. 21, 251-255.
Brodheim, E., and G.P. Prastacos (1979). The long island blood distribution system as a prototype
for regional blood management. Interfaces 9(5), 3-20.
Brout, D.B. (1981). Scientific management of inventory on a hand-held calculator, lnterfaces
11(6), 57-69.
Brown, R.G. (1967). Decision Rules for Inventory Management, Dryden Press, Hinsdale, IL.
Buzacott, J.A. (1975). Economic order quantities with inflation. Oper. Res. Quart. 26, 553-558.
Carlson, R.C., J.V. Jucker and D.H. Kropp (1979). Less nervous MRP systems: A dynamic
economic lot-sizing approach. Management Sci. 25 754-761.
Carlson, R.C., S.L. Beckman and D.H. Kropp (1982). The effectiveness of extending the horizon
in rolling production scheduling. Decision Sciences 13, 129-146.
Cohen, M.A. (1977). Joint pricing and ordering policy for exponentially decaying inventory with
known demand. Naval. Res. Logist. Quart. 24, 257-268.
Cohen, M.A., P. Kleindorfer and H. Lee (1988). Service constrained (s, S) inventory systems with
priority demand classes and lose sales. Management Sci. 34, 482-499.
Das, C. (1977). The (S - 1, S) inventory model under time limit on backorders. Oper. Res. 25,
835 -850.
Davis, E.W. (1975). A look at the use of production-inventory techniques: Past and present. Prod.
Invent. Management J., December, 1-19.
DeMatteis, J.J. (1968). An economic lot sizing technique: The part-period algorithms. IBM Syst.
J. 7, 30-38.
Dixon, P.S., and E.A. Silver (1981). A heuristic solution procedure for the multi-item, single-
level, limited capacity, lot sizing problem. J. Oper. Management 2, 23-39.
Dvoretzky, A., J. Kiefer and J. Wolfowitz (1952a). The inventory problem: I. Case of known
distributions of demand. Econometrica 20, 187-222.
Dvoretzky, A., J. Kiefer and J. Wolfowitz (1952b). The inventory problem: II. Case of unknown
distributions of demand. Econometrica 20, 450-466.
Eppen, G.D., F.J. Gould and B.P. Pashigian (1969). Extensions of the planning horizon theorem
in the dynamic lot size model. Management Sci. 15,268-277.
Factory Magazine (1961). Exclusive survey of production and inventory control. Factory
Magazine, April.
Factory Magazine (1966). Exclusive survey shows new directions in production and inventory
control. Factory Magazine, October.
Federgruen, A., and P. Zipkin (1984). An efficient algorithm for computing optimal (s, S) policies.
Oper. Res. 32, 1268-1285.
Feeney, G.J., and C.C. Sherbrooke (1966). The ( S - 1 , S) inventory policy under compound
Poisson demand. Management Sci. 12, 391-411.
Fisk, J.C., and D.P. Ballou (1982). Production lot sizing under a learning effect. HE Trans. 14,
257 -264.
52 H . L . Lee, S. Nahmias

Florian, M., and M. Klein (1971). Deterministic production planning with concave costs and
capacity constraints. Management Sci. 18, 12-20.
Galliher, H.P., P.M. Morse and M. Simond (1957). Dynamics of two classes of continuous-review
inventory systems. Oper. Res. 7, 362-384.
Gardner, E.S. (1980). Inventory theory and the Gods of Olympus. Inter]aces 10(4), 42-45.
Gardner, E.S. (1987). A top-down approach to modelling US Navy inventories, lnterfaces 17(4),
1-7.
Ghare, P.M,, and G.F. Schrader (1963). A model for an exponentiat decaying inventory. J. lndust.
Engrg. 14, 238-243.
Gross, D., and C.M. Harris (1971). On one-for-one-ordering inventory policies with state-
dependent leadtimes. Oper. Res. 19, 735-760.
Gross, D., and C.M. Harris (1973). Continuous review (s, S) inventory policies with state-
dependent leadtimes. Management Sci. 19, 567-57&
Hadley, G. (1964). A comparison of order quantities computed using average annual cost and the
discounted cost. Management Sci. 10, 472-476.
Hadley, G., and T.M. Whitin (1963). Analysis of lnventory Systems, Prentice-Hall, Englewood
Cliffs, NJ.
Harris, F.W. (1913). How many parts to make at once. Factory, The Magazine of Management
10(2), 135-136, 152.
Hax, A.C., and D. Candea (1984). Production and lnventory Management, Prentice-Hall,
Englewood Cliffs, NJ.
Higa, I., A.M. Feyerherm and A.L. Machado (i975). Waiting time in an ( S - 1, S) inventory
system. Oper. Res. 23, 674-680.
Hillier, F.S., and G.T. Lieberman (1987). lntroduction to Operations Researeh, 4th edition,
Holden Day, Oakland, CA.
Iglehart, D.L. (1963). Optimality of (s, S) policies in the infinite horizon dynamic inventory
problem. Management Sci. 9, 259-267.
IBM (1967). Basis Principles of Wholesale IMPACT Inventory Management Program and
Control Techniques, E20-8105-1, IBM, White Plains, New York, revised edition.
IBM (1978). INFOREM: Principles of Inventory Management-Application Description,
GE20-0571-1, IBM, White Plains, New York, rev&ed edition.
Jesse, R.R., A. Mitra and J.F. Cox (1983). EOQ formula: Is it valid under inflationary conditions?
Deeision Sciences 14, 370-374.
Kanet, J.J., and J.A. Miles (1985). Economic order quantities and inflation. Internat. J. Produc-
tion Res. 23, 597-608.
Kaplan, R. (1970). A dynamic inventory model with stochastic lead times. Management Sci. 16,
491-507.
Karlin, S. (1958). Steady state solutions, in: Arrow, Karlin and Scarf (eds.), Studies in the
Mathematical Theory of lnventory and Production, Stanford University Press, Stanford, CA,
Chapter 14.
Karlin, S. (1968). Total Positivity, Stanford University Press, Stanford, CA.
Karlin, S., and H.E. Scarf (1958). Optimal inventory policy for the Arrow Harris Marschak
dynamic model with a time lag, in: Arrow, Karlin and Scarf (eds.), Studies in the Mathematical
Theory of lnventory and Production, Stanford University Press, Stanford, CA, Chapter 10.
Karni, R. (1981). Maximum Part Period Gain (MPG) A lot sizing proeedure for uneonstrained
and constrained requirements planning systems. Prod. lnvent. Management 22, 91-98.
Keachie, E.C., and R.J. Fontana (1966). Effects of learning on optimal lot size. Management Sci.
13, B102-108.
Kleijnen, J.P.C., and P.J. Rens (1978). IMPACT revisited: A critical analysis of IBM's inventory
paekage 'IMPACT'. Prod. lnvent. Management 19(1), 71-90.
Kleutghen, P.P., and J.C. McGee (1985). Development and implementation of an integrated
inventory management program at Pfizer Pharmaceuticals. Interfaces 15(1), 69-87.
Kropp, D.H., and R.C. Carlson (1984). A lot sizing algorithm for reducing nervousness in MRP
systems. Management Sci. 30, 240-244.
 Ch. 1. Single-Product, Single-Location Models 53

Kruse, W.K. (1980). Waiting time in an S 1, S inventory system with arbitrarily distributed lead
times. Oper. Res. 28, 348-352.
Kruse, W.K. (t981). Waiting time in a eontinuous review (s, S) inventory system with constant
lead times. Oper. Res. 29, 202-207.
Ladany, S., and A. Sternlieb (1974). The interaction of economic ordering quantities and
marketing policies. AHE Trans. 6, 35-40.
Lee, H.L., and M.J. Rosenblatt (1986). The effects of varying marketing policies and conditions
on the economic ordering quantity. Internat. J. Prod. Res. 24, 593-598.
Lee, H.L., and M.J. Rosenblatt (1987). Simultaneous determination of production cycle and
inspection schedules in a production system. Management Sci. 33.
Lev, B., and A.L. Soyster (1979). An inventory model with finite horizon and pricc changes. J.
Oper. Res. Soc. 30, 43-53.
Lev, B., H.J. Weiss and A.L. Soyster (1981). Optimal ordering policies when anticipating
parameter changes in EOQ systems. Naval. Res. Logist. Quart. 28,267-279.
Liberatore, M.J. (1979). Using MRP and EOQ/safety stock for raw materials inventory control:
Discussion and case study, lnterfaces 9(2) Part 1, l-7.
Lingaraj, B.P., and R. Ba|asubramanian (1983). An inventory management and materials informa-
tion system for aircraft production. Interfaces 13(5), 65-70.
Love, S.F. (1973). Bounded production and inventory models with piecewise concave costs.
Management Sci. 20, 313-318.
Lundin, R.A., and T.E. Morton (1975). Planning horizons for the dynamic lot size model: Zabel
vs. protective procedures and computational resutts. Oper. Res. 23; 711-734.
Morton, T.E. (t969). Bounds on the solution of the lagged optimal inventory equation with no
demand backlogging and proportional costs. SIAM Rer. 11, 527-596.
Muth, E.J., and K. Spremann (1983). Learning effects in economic lot sizing. Management Sci. 29,
264-269.
Nahmias, S. (1976). On the equivalence of three approximate continuous review inventory models.
Naval. Res. Logist. Quart. 23, 31-36.
Nahmias, S. (1979). Simple approximations for a variety of dynamic leadtime lost-sales inventory
models. Oper. Res. 27, 904-924.
Nahmias, S. (1989). Production and Operations Analysis, Richard D. Irwin, Homewood, IL.
Orlicky, J. (1975). Material Requirements Planning, McGraw-Hill, New York.
Osteryoung, J.A., D.E. McCarty and W.J. Reinhart (1986). Use of the EOQ model for inventory
analysis. Prod. Invent. Management 27(3), 39-45.
Palm, C. (1938). Analysis of the Erlang traffic formula for busy signal arrangements. Ericsson
Technics 5, 39-58.
Porteus, E.L. (1971). On the optimality of generalized (s, S) policies. Management Sci. 17,
411-427.
Porteus, E.L. (1985a). Numerical comparisons of inventory policies for periodic review systems.
Oper. Res. 33, 134-152.
Porteus, E.L. (1985b). Undiscounted approximations of discounted regenerative models, oper.
Res. Lett. 3, 293-300.
Porteus, E.L. (1986). Optimal lot sizing, process quality improvement and setup cost reduction.
Oper. Res. 34, 137-144.
Porteus, E.L. (1988). Stochastic inventory theory, in: D. Heyman and M. Sobel (eds.), Handbooks
in Operations Research and Management Science, Vol. II: Stochastic Models.
Resh, M., M. Friedman and L.C. Barbosa (1976). On a general solution of the deterministic lot
size problem with time-proportional demand. Oper. Res. 24, 718-725.
Reuter, V.G. (1978a). The big gap in inventory management. J. Purchasing Mater. Management,
Fall, 17-22.
Reuter, V.G. (1978b). Are we using the inventory management tools, lndust. Management,
May/June, 26-30.
Richards, F.R. (1975). Comments on the distribution of inventory position in a continuous-review
(s, S) inventory system. Oper. Res. 23, 366-371.
54 H.L. Lee, S. Nahmias

Roberts, D.M. (1962). Approximations to optimal policies in a dynamic inventory model, in: K.A.
Arrow, S. Karlin and H.E. Scarf (eds.), Studies in Applied Probability and Management Science,
Stanford University Press, Stanford, CA, Chapter 13, pp. 207-299.
Rosenblatt, M.J,, and H.L. Lee (1985). Improving profitability with quantity discounts under fixed
demands. HE Trans. 17, 388-395.
Rosenblatt, M.J., and H.L. Lee (1986). Economic production cycles with imperfect production
processes. HE Trans. 18, 48-55.
Sahin, I. (1979). On the stationary analysis of continuous review (s, S) inventory systems with
constant lead times. Oper. Res. 27, 717-729.
Scarf, H.E. (1960). The optimality of (s, S) policies in the dynamic inventory problem, in: K.A.
Arrow, S. Karlin and P. Suppes (eds.), Mathematical Methods in the Social Sciences, Stanford
University Press, Stanford, CA.
Schneider, H. (1978). Methods for determining the re-order point of an (s, S) ordering policy
when a service level is specified. J. Oper. Res. Soc. 29, 1181-1193.
Schonberger, R.J., and M.J. Schniederjans (1984). Reinventing inventory control, lnterfaces
14(3), 76-83.
Schwarz, L.B. (1972). Economic order quantities for produets with finite demand horizons. AllE
Trans. 4, 234-237.
Schwarz, L.B. (1977). A note on the near optimality of '5-EOQ's worth' forecast horizons. Oper.
Res. 25, 533-536.
Sherbrooke, C.C. (1975). Waiting time in an (S - 1, S) inventory system - Constant service time
case. Oper. Res. 23, 819-820.
Shih, W. (1980). Optimal inventory policies when stockouts result from defective products.
Internat. J. Prod. Res. 18, 677-686.
Silver, E.A. (1976). Establishing the order quantity when the amount received is uncertain.
INFOR 14, 32-39.
Silver, E.A., and H.C. Meal (1973). A heuristic for selecting lot size quantities for the case of a
deterministic time-varying demand rate and discrete opportunities for replenishment. Prod.
Invent. Management 14, 64-74.
Silver, E.A., and R. Peterson (1985). Decision Systems for Inventory Management and Production
Planning, 2nd edition, Wiley, New York.
Simon, R.M. (1968). The uniform distribution of inventory position for continuous review (s, Q)
policies, Rand Corporation Paper P3938, Santa Monica, CA.
Sivazlian, B.D. (1974). A continuous-review (s, S) inventory system with arbitrary interarrival
distribution between unit demand. Oper. Res. 22, 65-71.
Smith, S.A. (1977). Optimal inventories for an (S - 1, S) system with no backorders. Management
Sci. 23 522-528.
Smunt, T.L., and T.E. Morton (1985). The effects of learning on optimal lot sizes: Further
developments on the single product ease. I1E Trans. 17, 33-37.
Stidharn, S. (1974). Stochastic clearing systems. Stochastic Process. Appl. 2, 85-113.
Sule, D.R. (1978). The effect of alternative periods of learning and forgetting on economic
manufacturing quantity. A l l e Trans. 10, 338-343.
Taylor, S.G., and C.E. Bradley (1985). Optimal ordering policies for announced price increases.
Oper. Res. 33,312-325.
Tijms, H.C. (1972). Analysis of (s, S) lnventory Models, Mathematisch Centrum, Amsterdam.
Tijms, H.C., and H. Groenevelt (1984). Simple approximations for the reorder point in periodic
and continuous review (s, S) inventory systems with service level constraints. European J. Oper.
Res. 17, 175-190.
Trippi, R.R., and D.E. Lewin (1974). A present-value formulation of the classical EOQ problem.
Decision Sciences 5, 30-35.
Veinott, A.F. (1964). Production planning with convex costs: A parametric study. Management
Sei. 10, 441-460.
Veinott, A.F. Jr. (1965). The optimal inventory policy for batch ordering. Oper. Res. 13,424-432.
Wagner, H.M., and T.M. Whitin (1958). Dynamic version of the economic lot size model.
Management Sci. 23, 89-96.
 Ch. 1. Single-Product, Single-Location Models 55

Wagner, H.M., M. O'Hagen and B. Lund (1965). An empirical study of exactly and approximately
optimal inventory policies. Management Sci. 11, 690-723.
Whitin, T.M. (1955). Inventory control and price theory. Management Sci. 2, 61-68.
Whitin, T.M. (1957). The Theory of lnventory Management, Revised edition, Princeton University
Press, Princeton. NJ.
Wijngaard, J., and E.G.F. Van Winkel (1979). Average costs in a continuous review (s, S)
inventory system with exponentially distributed lead time. Oper. Res. 27, 396-401.
Wilson, R.H. (1934). A scientific routine for stock control. Harvard Business Rer. 13, 116-128.
Wortharn, A.W., and A.M. Mayyasi (1972). Learning considerations with economic order quanti-
ty. AHE Trans. 4, 69-71.
Yano, C.A. (1985). New algorithms for (Q, r) systems with complete backordering using a fill-rate
criterion. Naval. Res', Logist. Quart. 32, 675-688.
Zabel, E. (1964). Some generalizations of an inventory planning horizon theorem. Management
Sci. 10, 465-471.
Zangwill, W.I. (1966a), A determinsitic multi-period production scheduling model with backlog~
ging. Management Sci. 13, 105-119.
Zangwill, W.I. (1966b). A deterministic multi-product multi-facility production and inventory
model. Oper. Res. 14, 486-507.
Zeitzew, H. (1979). E O Q - Is there a better way? Purchasing Magazine, 29 33.
Zipkin, P. (1986). Stochastic leadtimes in continuous-time inventory models. Naval. Res. Logist.
Quart. 33,763-774.