ARTICLE IN PRESS

Int. J. Production Economics 97 (2005) 18–30

www.elsevier.com/locate/dsw

A model for optimizing multi-product inventory systems with

multiple constraints
Cengiz Haksever, John Moussourakis
Department of Management Sciences, College of Business Administration, Rider University, Lawrenceville, NJ 08648, USA
Received 14 March 2003; accepted 12 April 2003
Available online 21 August 2004

Abstract

‘‘How much to order’’ and ‘‘when to order’’ are the two fundamental issues managers have to resolve in an inventory
system. Making these decisions in inventory systems with multiple products is a challenging task for managers because
these decisions are often subject to several constraints due to limited resources such as budget, space, and the maximum
weight of goods that can be stored. Most approaches in the literature for optimizing decisions in such an environment
consider only a single budgetary constraint. This paper presents a mixed-integer programming model to optimize the
two fundamental decisions of inventory management for ordering multiple inventory items subject to multiple resource
constraints. It also determines whether a fixed cycle for all products or an independent cycle for each should be used for
a lower total cost. The solution of the model does not seem to require excessive central processing unit (cpu) time as
indicated by the computational experience reported in this paper; solution of the largest test problem, with 30 products
and five resource constraints, required less than 20 cpu seconds on a personal computer.
r 2004 Elsevier B.V. All rights reserved.

Keywords: Optimization; Multi-product; Multi-constraint; Inventory model

1. Introduction approach, many manufacturers still have to order

and keep inventories of raw materials, parts,
It is well known that costs of ordering and finished goods, and supplies. This is mainly
holding inventories of goods may be significant for attributable to the types of products and/or supply
any company. For this reason some manufacturers systems.
have been implementing just-in-time (JIT) systems. It is clear that deciding ‘‘how much to order’’
However, despite the popularity of the JIT and ‘‘when to order’’ is still a challenge to
managers in both manufacturing and service
Corresponding author. Tel.: +609-895-5476; fax: +609- industries. Intensifying competition and desire to
896-5304. achieve ever-increasing levels of customer satisfac-
E-mail address: [email protected] (C. Haksever). tion have increased the complexity and importance

0925-5273/$ - see front matter r 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.ijpe.2004.05.004
 ARTICLE IN PRESS

C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30 19

of this task. One of the most challenging aspects of rated problems. In Section 7 we offer some
most inventory decisions is the multitude of conclusions.
products involved. Another is the set of con-
straints within which managers have to operate;
very often these two challenges exist at the same 2. Review of literature
time. This is the decision-making situation we
address in this paper. Published research on the multi-product inven-
Operations management and operations re- tory system under a single constraint may be
search literature is rich with articles focused on organized into three groups. The first group can be
meeting these challenges. However, the great characterized by its ‘‘independent cycle’’ assump-
majority of these papers consider only one tion. It is assumed that the quantity of each
constraint; two exceptions are approaches pro- product ordered, and therefore its order cycle, is
posed by Maloney and Klein (1997) that can determined independently of other products. In
handle two resource constraints and Güder and this system, at some point, all product inventories
Zydiak (1999) that can handle multiple con- will reach their peak at the same time.
straints. The approach presented in this paper The critical issue then, is to make sure that a
goes beyond two resource constraints and deals resource constraint (for example, budget, ware-
with multi-product multi-constraint (MPMC) in- house space, etc.) is not violated when this
ventory systems with stationary ordering policies. happens. The usual approach is to first determine
More specifically, this paper extends Rosenblatt the order quantity for each item from the classical
(1981) algorithm for determining whether an economic order quantity (EOQ) formula and
independent or fixed cycle approach leads to lower check if the resource constraint will be violated
cost under a single budget constraint to multiple when all inventories reach their peak at the same
(two or more) constraints. We formulate the time for these order quantities. In case the
MPMC inventory decision as a mixed-integer constraint is violated, order quantities must be
programming model which is based on a piecewise reduced. The most frequently suggested approach
linear approximation of the number of orders is to use Lagrange multipliers to find order
function, N ¼ f ðX Þ ¼ D=X , where D is annual quantities that do not violate the resource con-
known demand and X is order quantity. Most straint. This is the method that was suggested for
discussions of inventory models focus on the the solution of the problem in many textbooks
relationship between total cost and order quantity. including Holt et al. (1960), Hadley and Whitin
In this paper, we rely on the relationship between (1963), Buchan and Koenigsberg (1963), Johnson
the number of orders and order quantity which and Montgomery (1974), Tersine (1976), and
enables us to handle multiple constraints through Nahmias (1993). In other publications, Parsons
a linear model. (1966) considered the problem for both the classic
The rest of the paper is organized as follows. EOQ model and production lot size model. Using
Section 2 presents a brief review of the relevant Lagrange multipliers, Parsons developed formulas
literature. Section 3 summarizes the Rosenblatt for optimum order quantity for specific types of
(1981) algorithm for choosing between indepen- constraints, assuming only one constraint active at
dent vs. fixed cycle approaches. Section 4 sum- a time. Ziegler (1982) proposed an approxima-
marizes the procedure we use for the approxi- tion algorithm, which determines upper and
mation of the number of orders function; details of lower bounds on Lagrangian multipliers. Maloney
the procedure are given in the Appendix. Section 4 and Klein (1993) developed an algorithm which
also presents the mathematical model for opti- provides bounds on the optimal value of a
mizing multi-product inventory systems with Lagrange multiplier and converges to the value
multiple constraints. Section 5 discusses imple- needed to optimize the multi-item inventory
mentation issues and Section 6 presents the results model. In a later paper Maloney and Klein
of computational testing on randomly gene- (1997) developed an algorithm for the problem
 ARTICLE IN PRESS

20 C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30

with two constraints. Bretthauer et al. (1994) Chen and Min (1994) also considered a two-
focused on a general integer programming model constraint, fixed cycle problem and proposed a
for the single constraint problem and developed a closed-form solution and identified the conditions
heuristic and a branch and bound algorithm under which the optimal closed-form solution is
solution method. The authors identified several valid. Güder and Zydiak (1999) proposed two
other problems from operations management and algorithms for the solution of the multiple
statistics with the same structure that may be constraint problem; one for the fixed cycle case
solved by the heuristic. The major disadvantage of and another for non-stationary.
the Lagrangian method is that in many cases Page and Paul (1976) suggested a method, which
Lagrange multipliers cannot be obtained directly first determines the joint cycle time for product
but must be estimated by trial and error (Johnson groups without considering any resource con-
and Montgomery, 1974); applying the Lagrangian straint, and then calculates order arrival times
method to problems with more than one con- for each group. If the resource constraint (either
straint is even more difficult and may not give the maximum inventory investment or maximum
correct result (Taha, 1987, p. 517). However, a inventory space) is violated, then they propose to
special case of the problem can be solved without modify the order interval by using the Lagrange
resorting to Lagrange multipliers. It has been multiplier method. They also offered a heuristic
shown that (Rosenblatt, 1981) when inventory for grouping similar products to achieve better
carrying cost is a fixed percentage of the cost of an results. Goyal (1978) showed that it is possible to
item, and the single constraint is a budget-type reduce the total resource requirement by consider-
restriction, where resource usage is proportional to ing products belonging to more than one group. In
the cost of the item, the optimal solution can be addition, Rosenblatt and Rothblum (1990) also
obtained without computing the Lagrange multi- considered the fixed cycle problem with phasing of
pliers. Optimal order quantities can be determined orders but treated available capacity (for example,
by simply reducing EOQs by the ratio of the storage space, budget, etc.) as a decision variable
resource to the total amount of that resource rather than a constraint in a nonlinear program-
needed if EOQs are used as order quantities. ming model; they developed two solution ap-
Unfortunately, most constraints do not meet this proaches which lead to the same result.
requirement. A third group of research papers, addressing
A second strand of research assumes a joint mostly unconstrained multi-item inventory sys-
fixed order cycle time for all products; orders for tems, can be characterized as basic cycle ap-
individual products are phased within the cycle. In proaches. This approach computes individual
this approach, the optimization problem can be cycle times that are integer multiples, or powers
stated as ‘‘determine a joint order cycle time and of two multiples, of a basic cycle time. Papers by
phasing of individual orders within it so that the Goyal (1973), Silver (1976), Goyal and Belton
total cost is minimized subject to a single resource (1979), Kaspi and Rosenblatt (1983), and Roundy
constraint.’’ In other words, the time between (1985) are in this category.
replenishments for each product is the same, but In addition to these works, several researchers
order arrivals are staggered so that individual dealt with variants of the problem. Among these is
inventory levels do not peak at the same time. an approach developed by Güder et al. (1995)
Researchers who followed this approach include which calculates non-stationary order quantities
Homer (1966), Page and Paul (1976), Goyal for a multi-item inventory system with a single
(1978), Zoller (1977), Hartley and Thomas (1982) resource constraint. This approach generates an
and Thomas and Hartley (1983). Lee (1994) order quantity for each item that is between the
considered a two-constraint, fixed cycle problem order quantities obtained from the Lagrangian
and incorporated quantity discounts, and pro- approach and classical EOQ formula and that is
posed a global optimization approach based on allowed to vary over time. Gallego et al. (1996)
Lagrangian duality and geometric programming. developed two models, which can be considered as
 ARTICLE IN PRESS

C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30 21

variants of the multi-item single constraint pro- X j X0; j ¼ 1; 2; . . . ; k;

blem. In the first model (tactical), they determine where TC is the total cost of holding and ordering
and coordinate order quantities to minimize total inventory, Xj the order quantity for product j, Dj
ordering and carrying costs subject to a capacity the annual demand for product j, Chj the holding
constraint. In the second model (strategic), which cost for product j, Coj the ordering cost for
was also considered by Hall (1988) earlier, they product j, wrj the amount of resource r (for
determine and coordinate order quantities to example, budget, space) consumed by one unit of
minimize the average cost which includes, in product j, and Br the maximum amount of
addition to ordering and carrying costs, a cost resource r available.
proportional to the peak usage of the resource. The Rosenblatt (1981) considered the independent
authors developed a heuristic for staggering orders cycle (Lagrangian) and the fixed cycle approaches
and employ it for the solution of both models. for a multi-item inventory system with a budget
Hariga and Jackson (1996) developed a third constraint. His algorithm determines not only the
variant of the multi-product single constraint lower cost approach, but also calculates order
inventory problem as a mixed-integer program- quantities and cycle times. Rosenblatt’s algorithm
ming model. These researchers developed a heur- is straightforward and relatively easy to imple-
istic for staggering orders and combined it with a ment. However, it is limited to a single constraint
generalized version of Roundy (1985) the powers- and is based on the assumption that carrying cost
of-two techniques for generating order frequencies. is a percent of the cost of the item and therefore is
limited to a single constraint of the budgetary type.
In other words, the constraint must be such that
3. Rosenblatt’s algorithm: Independent vs. fixed resource usage of each product is a multiple of the
cycle approach cost of the item. This requirement severely reduces
the applicability of this useful algorithm.
As indicated by Rosenblatt (1981), Hall (1988) In this paper we propose an alternative to
and Güder and Zydiak (1999), neither the in- Rosenblatt (1981) algorithm, which may be
dependent (Lagrangian) nor fixed cycle approach computationally less efficient, but is applicable to
dominates the other for all problem instances. a wide variety of inventory systems. Specifically,
Rosenblatt (1981) developed an algorithm for a while Rosenblatt (1981) algorithm can handle only
single resource constraint case to test which one budget-type constraint, we propose a mixed-
solution approach would result in lower cost when integer programming model that can handle
applied to a data set. multiple linear constraints of any type. The model
The common assumptions made for the multi- is based on a piecewise linear approximation of the
constraint multi-product inventory model are that number of orders function.
product demand is known and evenly distributed,
lead time is zero, the entire order quantity is received
at one point in time, no excess stock is carried, and 4. The model
stock outs are not allowed. We also assume linearity
in the amounts of resource consumption. The The number of orders function is strictly convex
problem considered can be formulated as and can be approximated with a series of linear
Xk   functions (see Appendix, Fig. 1). Thus, any line
Dj C oj X j C hj
Min TCðX Þ ¼ þ segment such as Lij ¼ aij  bij X ij , passing through
j¼1
Xj 2
the end points of interval i for product j will always
subject to be above the curve and Lij will always overestimate
the true number of orders in that interval. The
X
k
error of estimation, Eij, will be given by
wrj X j pBr ; r ¼ 1; 2; . . . ; v;
j¼1 E ij ¼ Lij  N j ¼ aij  bij X ij  ðDj =X ij Þ:
 ARTICLE IN PRESS

22 C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30

The maximum error can be reduced to any finite Yij auxiliary variables for product j: Y ij ¼ 1,
number by increasing the number of line segments. if X j 2 ½nij ; mij
; Y ij ¼ 0, otherwise,
Once a decision-maker determines the maximum Z1, Z2 auxiliary variables: Z 1 ¼ 1 and Z 2 ¼ 0,
tolerable error (TE), the range of possible order sizes fixed cycle approach to be used; Z 1 ¼ 0
(X) is split into as many intervals as necessary so and Z 2 ¼ 1, independent cycle approach
that no line segment overestimates N by more than to be used.
TE. We follow a procedure that splits an interval at
the point (Xio) where the error Eij is maximum, The resulting model follows:
thereby reducing overestimation by the greatest Xk   
amount. The details of the linearization of the Xj
Min TCðN j ; X j Þ ¼ C oj N j þ C hj ð1Þ
number of orders function are given in Appendix A. j¼1
2

subject to
Model parameters:
Lij ¼ aij Y ij  bij X ij ; j ¼ 1; 2; . . . ; k; ð2Þ
TC total annual inventory cost,
k number of products, X
pj
Coj ordering cost for product j, Y ij ¼ 1; j ¼ 1; 2; . . . ; k; ð3Þ
Chj holding cost per unit per year for pro- i¼1

duct j,
X
pj
Dj annual demand for product j, Nj ¼ Lij ; j ¼ 1; 2; . . . ; k; ð4Þ
aij y-intercept of the line for product j i¼1
passing through the end points of
interval i, X ij Xnij Y ij ; j ¼ 1; 2; . . . ; k; ð5Þ
bij slope of the line for product j passing
X ij pmij Y ij ; j ¼ 1; 2; . . . ; k; ð6Þ
through the end points of interval i,
nij, mij lower and upper end points of interval i of
X
pj
Nj values, X ij ¼ X j ; j ¼ 1; 2; . . . ; k; ð7Þ
pj number of intervals into which the num- i¼1
ber of orders (Nj) function for product j
has been divided, X
k
wrj X j pMZ1 þ Br Z 2 ; r ¼ 1; 2; . . . ; v; ð8Þ
wrj amount of resource r consumed by one
j¼1
unit of product j,
v number of constrained resources, T j XRZ 1 ; j ¼ 1; 2; . . . ; k; ð9Þ
Br availability of resource r,
R Minft; tr ; r ¼ 1; 2; . . . ; vg, T j pRZ 1 þ Z 2 ; j ¼ 1; 2; . . . ; k; ð10Þ
t the optimal cycle length for the uncon-
strained problem, T j Dj  X j ¼ 0; j ¼ 1; 2; . . . ; k; ð11Þ
tr the maximum cycle length allowed by the
rth resource constraint, Z1 þ Z 2 ¼ 1; ð12Þ
M a very large positive constant. T j ; X ij ; X j ; N j ; Lij X0 8i and j;
Decision variables: Y ij ¼ 0; 1 integer;
Nj number of orders of product j,
Xj order size for product j, Z1 ; Z 2 ¼ 0; 1 integer:
Lij number of orders for product j if the order Constraint (2) represents the linear segments
size is in interval i, approximating the number of orders function for
Xij order size for product j in interval i, each product. Constraints (3) and (4) make sure
Tj cycle time for product j, that only one of these segments is selected and
 ARTICLE IN PRESS

C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30 23

determine the number of orders. Constraints (5), If the fixed cycle approach is chosen, the next
(6), and (7) ensure that only one value for order step is to determine the phasing of the orders. The
size is selected. Resource constraints are repre- following formula can be used to determine the
sented by (8), as mentioned earlier, resource time gaps between the replenishment of products:
consumption is assumed to be linear. Z1 and Z2
wsj Dj
are binary variables that help decide whether the tj ¼ R Pk ; j ¼ 1; 2; . . . ; k; ð16Þ
fixed cycle or independent cycle approach will be h¼1 wsj Dh
used. Specifically, constraint (12) assures that only where t1 is defined as the time interval from the
one approach is used. Constraints (9) and (10) replenishment of product k to the replenishment of
guarantee that if a fixed cycle solution is chosen, product 1, and in general tj is the time interval
the cycle times, stipulated by either the uncon- between the replenishment of product ðj  1Þ to
strained or the constrained problem, will be met. the replenishment of product j. Subscript s is equal
Constraints (11) compute the order quantities for to the subscript of the Min of tr that determines R,
all products. R is a constant supplied by a user of and wsj is the consumption of resource s, whose
the model and is the minimum of t and tr . t is the availability restricts the cycle time most, by
common (fixed) cycle length for all products for product j. It is interesting to note that a concept
the unconstrained problem and is calculated similar to the definition of R has been suggested by
according to the following formula and corre- Haji and Mansuri (1995) for the solution of
sponds to Eq. (14) of Rosenblatt (1981): economic lot scheduling problem with multiple
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi constraints. Eq. (16) is equivalent to Eq. (7) of
u Pk
u 2 j¼1 C oj Güder and Zydiak (1999) and is an extension of
t ¼ tPk ; ð13Þ
Eq. (19) of Rosenblatt (1981) and Eq. (10) of Page
j¼1 C hj Dj
and Paul (1976). If the fixed cycle is chosen, tj’s can
tr is the cycle time imposed by constraint r and is be calculated outside the model.
calculated as follows:
2Br
tr ¼ ; ð14Þ 5. Implementation
Qr
where When implemented, the model presented in this
" Pk 2
# paper leads to a mixed-integer programming
X
k
j¼1 ðwrj Dj Þ problem that may be quite large depending on
Qr ¼ wrj Dj þ Pk : ð15Þ
j¼1 j¼1 wrj Dj
various factors. It is well known that the solution
of mixed-integer programming problems may
Eq. (14) is the multi-constraint version of Eq. require considerable computation time. Therefore,
(18) of Rosenblatt (1981) (see also Page and Paul, for the proposed model to be a viable tool for
1976). If Z 1 ¼ 1, the model will choose the fixed managers, it must meet two criteria: It must be
cycle solution, and all the resource constraints will practical and the time requirement for its solution
be satisfied. However, if Z 2 ¼ 1, then the model must not be excessive. The criteria imply that an
will pick the independent cycle approach and automated procedure must exist for the linear
determine the optimum cycle time for each approximation of the number of orders function,
product. Although we represented only resource and once the problem is set up it must be solved in
constraints (8), other linear constraints can be a reasonable amount of time.
added to the model. It also should be noted that We developed a computer code (SplitV2) in
when independent cycle approach is chosen, the Visual Basic for Applications (VBA) for Excel,
model does not use the Lagrangian method which splits the number of orders function into
and thereby avoids all the problems associated intervals that are approximated by linear equa-
with it. tions. A user of this code must supply the number
 ARTICLE IN PRESS

24 C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30

of products, annual demand for each product, from user-provided parameter intervals. The
holding cost, ordering cost, and resource con- program also computes t and tr ’s, cycle time for
straints. The code creates an Excel sheet, which is a the unconstrained problem and maximum cycle
complete representation of the mixed-integer lengths allowed by the resource constraints,
model ready for solution with Solver. Solver is respectively.
developed by Frontline Systems and its standard We generated and solved 24 sets of problems,
version is usually a part of Microsoft’s Excel each set consisting of five different problems by
program. using the same parameters. Product demands in
the first 12 sets of problems were randomly
generated from the interval 200–1000 and we used
a TE of 0.25, which corresponds to an error
6. Computational experience
tolerance of 25% of an order. As an example,
Table 2 shows the data and solution of a problem
To test if the proposed algorithm requires a
with 10 products and five resource constraints
reasonable amount of time to find a solution, we
from this group of test problems. In the second set
developed another VBA code (MPMC Problem
of 12 problems, demands were generated from
Generator—MPMC-GEN), which randomly gen-
200–2000 interval and we used a TE of 0.35.
erates model parameters as well as resource
All the test problems were generated and solved
constraints. A user of this program must supply
on a Dell OptiPlex GX260 personal computer
the parameters shown in Table 1, as well as a TE
(PC) with Pentium 4, 2.33 GHz processor, using
for piecewise linear approximation of the number
Microsoft Windows XP Professional operating
of orders function, and an extremely large positive
system. Optimization software used in these tests
number for M. Numbers in the table represent the
was Premium Solver Platform Version 3.5. The
values we used in generating the test problems.
version of the software we used is limited to linear
MPMC-GEN uses VBA’s built-in function Rnd
and quadratic programming problems with a
to generate uniformly distributed random numbers
maximum of 2000 variables and 8192 constraints.
that are used to create the coefficients of the model
Important statistics from computational testing
are given in Table 3. The central processing unit
Table 1
(cpu) times include problem set-up phase. Twenty
Parameters used for generating test problems five percent of the test problems were of the fixed
cycle type; the solution of these problems required
Parameters Definition Values used more cpu time (average 9.93 seconds) than
NPROD Number of products 5, 10, 20, 30 independent-cycle-type problems (average 6.82
MINDEM Minimum demand 200 seconds), but the difference was much more
MAXDEM Maximum demand 1000, 2000 significant for problems with 20 or 30 products.
MINHOLDC Minimum holding cost $10 The large standard deviations (see Table 3) for
MAXHOLDC Maximum holding cost $100
some of the problem sets were due to a few fixed-
MINORDC Minimum ordering cost $100
MAXORDC Maximum ordering $400 cycle-type problems.
cost Several factors influence problem size; they
NCONSTR Number of resource 3, 4, 5 are, demand, length of the number of orders
constraints interval considered, TE, and the number of
MINRCOEF Minimum value for 10
products. The first three have a direct impact on
constraint coefficients
MAXRCOEF Maximum value for 100 the number of linear equations that approximate
constraint coefficients the number of orders function. A large demand
TOLERABLE Tolerance for linear .25, .35 implies a relatively large interval of possible
ERROR approximation values for the number of orders. We used 1000
BIG M An extremely large 1,000,000
as the maximum annual demand in the first 12
positive value
problem sets, and 2000 in the second 12 to
 ARTICLE IN PRESS

C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30 25

Table 2
Output from a test problem with 10 products and five constraints

Product 1 2 3 4 5 6 7 8 9 10

Demand 354 743 563 486 320 764 943 624 272 806
Coj 221 239 248 162 199 129 277 151 378 129
(1/2)Chj 25 17 44 39 17 35 17 9 6 20
EOQ 56 101 56 45 61 53 126 102 127 73
w1j 37 53 41 53 88 78 40 17 47 20
B1 42,800
w2j 66 23 30 22 41 93 47 84 75 41
B2 34,260
w3j 47 26 59 59 56 66 71 43 36 58
B3 30,000
w4j 63 89 27 77 80 83 96 16 44 21
B4 27,864
w5j 26 74 60 52 78 91 18 74 49 35
B5 50,000

BIG M 1,000,000
R value 0.124875
Tolerance 0.25

Xj 44.2057 92.7819 70.3045 60.6891 39.9599 95.4043 117.7569 77.9218 33.9659 100.6490
Nj 8.0348 8.0359 8.0630 8.0479 8.1103 8.0701 8.0552 8.0264 8.0575 8.1230
Tj 0.1249 0.1249 0.1249 0.1249 0.1249 0.1249 0.1249 0.1249 0.1249 0.1249

Z1 1
Z2 0
Z1+Z2=1 1

OBJFUN= $34,162.23
Time for this run 2.56 Seconds

keep the test problems within the limitations of smaller the TE, the larger the number of equa-
the software. tions, and hence the larger the number of variables
The largest interval that can be considered for and constraints. A maximum TE of 0.25 of an
the number of orders is, 1  D, one to annual order size appeared reasonable for test purposes.
demand. However, this is an unnecessarily wide In the second set of 12 problems, a TE of 0.35 was
interval that will require a very large number of used to keep problem size within the limits of the
approximating equations, and hence make the software.
problem unwieldy. On the other hand, too narrow It is also clear that the larger the number of
an interval may miss the optimum. A compromise products the larger the number of variables and
approach of using an interval around the uncon- constraints. Our largest test problem involved
strained EOQ gave completely satisfactory results 1910 variables and 2004 constraints for a 30-
in preliminary tests; therefore, the number of product problem. It was felt that solving test
orders interval considered for each product was: problems involving up to 30 products would
EOQ70.95EOQ. provide sufficient data for making reasonable
The third factor that has a direct impact on the extrapolations about solution times for problems
number of approximating equations is TE; the with a larger number of products.
 ARTICLE IN PRESS

26 C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30

Table 3
Multi-product multi-constraint inventory model computational test results

Problem set No. of products No. of resource Avg. cpu time (s) Std. of cpu times
constraints

1 5 3 1.87 0.52
2 5 4 1.72 0.17
3 5 5 1.75 0.32
4 10 3 5.13 3.26
5 10 4 3.33 1.66
6 10 5 2.82 1.42
7 20 3 13.53 7.25
8 20 4 14.22 9.31
9 20 5 18.13 17.15
10 30 3 8.84 1.05
11 30 4 9.48 1.05
12 30 5 11.22 1.80
13 5 3 1.68 0.08
14 5 4 2.46 1.07
15 5 5 3.15 1.79
16 10 3 2.63 1.85
17 10 4 2.88 2.26
18 10 5 3.30 1.25
19 20 3 14.62 11.63
20 20 4 19.17 21.27
21 20 5 10.00 0.61
22 30 3 9.92 1.26
23 30 4 10.66 2.09
24 30 5 9.91 0.81

7. Conclusion software, can be solved on a PC in a matter of

seconds or minutes.
While many manufacturers are moving toward Test results shown in Table 3 indicate that the
JIT systems, some are prevented from such a model we presented can be a viable tool for
move because of the nature of their supply managers who operate inventory systems with
systems; consequently, they still have to manage multiple products and under multiple constraints.
inventories of a large number of items. Further- Solution times for test problems of up to 30
more, most service organizations such as hos- products and five resource constraints were in
pitals, and retailers do not have much choice seconds, not hours or even minutes. In our
but to keep inventories of finished products and experiments no test problem required more than
supplies. Managers of these organizations face 20 cpu seconds, which seems to be a reasonable
an important challenge: ‘‘How much (or many) solution time on PCs. It can be concluded that
to order’’ and ‘‘when to order.’’ These decisions with commercially available modern optimizers
are often subject to limited resources, such as computation time even for much larger problems
budget, space, weight, number of times a product may not be significant. For very large problems
can be ordered in a year, to name a few. In this another version of the Solver, Large-Scale LP
paper, we presented a viable model that may aid Solver, can be used. Furthermore, computational
managers in meeting these challenges. This model experiments have shown that one does not need a
not only can handle any number of linear mainframe computer to solve even relatively large
constraints but also, with the right optimization mixed-integer programming problems to optimize
 ARTICLE IN PRESS

C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30 27

multi-product multi-constraint inventory systems; The maximum error can be reduced to any finite
they can be solved on PCs. The two VBA number by increasing the number of line segments.
programs (SplitV2 and MPMC-GEN) mentioned Once the maximum error an analyst is willing to
in this paper are available free of charge from the tolerate (TE) is determined, the next task is to split
corresponding author for research purposes. the range of possible order sizes (X) into as many
intervals as necessary so that no line segment
overestimates N by more than TE. An efficient
Acknowledgements way is to split the intervals at the point where the
error Eij is at maximum, thereby reducing over-
We would like to thank two anonymous estimation by the greatest amount.
reviewers for their useful comments. Suppose we start with the first interval as
interval 1, which is approximated by the line
segment LðpÞ
1j ¼ a1j  b1j X 1j , where the superscript
p represents the iteration number and is set equal
Appendix A. Linear approximation of the number of
to 0 at the beginning of the process (Fig. 2). Since
orders function
the process can be applied to only one product at a
time, the subscript j will be dropped in the
The number of orders function (Fig. 1) is strictly
discussion. Also, some values will be identified
convex and can be approximated with a series of by the iteration at which they are calculated. For
linear functions. The error of estimation, Eij, is example, bð2Þ
1 represents the slope of the line that
given by
approximates the curve in interval 1 at the second
E ij ¼ Lij  N j ¼ aij  bij X ij  ðDj =X ij Þ: ðA:1Þ iteration. The range of N values an analyst wants

Nj

L(p)1=a(p)1 − b(p)1X1j

(p)
a 2

L(p)2=a(p)2 − b(p)2X2j
a(p)3
L(p)3=a(p)3 − b(p)3X3j

L(p)4=a(p)4 − b(p)4X4j
(p)
a 4

Nj = Dj /Xj

Interval 1 Interval 2 Interval 3 Interval 4 Xj

Fig. 1. Piecewise linear approximation of the number of orders function.

 ARTICLE IN PRESS

28 C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30

Nj

N1L

L(1)1=a(1)1 − b(1)1X1

L(0)1=a(0)1 − b(0)1X1

E max

L(1)2=a(1)2 − b(1)2X2
N1o =
Db1(0)

N1R
Xj
n1 (0)
X1o = / (0)
Db 1 m 1
(0)

Fig. 2. Splitting the first interval.

to consider will determine the end points of the the end points (ni, NiL) and (mi, NiR) of any
first interval at iteration p ¼ 0. One way to choose interval i, the equation of a line segment
the end points of the first interval is to set the
ordinate of the left end n(0) Li ¼ ai  bi X i ðA:2Þ
1 equal to the minimum
order quantity acceptable by a supplier, and passing through these end points can be con-
for a problem with one or more constraints, set structed:
the ordinate of the right end m(0) 1 equal to  
Min ðDj ; Br =wrj Þ. Then, abscissas can be deter- N iL  N iR 
bi ¼  ; ðA:3Þ
mined as N 1L ¼ D=nð0Þ ð0Þ
1 , and N 1R ¼ D=m1 for the ni  m i 
left and right end points, respectively. This will
 
create the widest interval that needs to be N iL  N iR 

ai ¼  X i þ L i : ðA:4Þ
considered. ni  m i 
From Eq. (A.1), and by ordinary differentiation,
the error for any interval i, Ei, is maximum
pffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiatffi The value of ai can be calculated by substituting
X io ¼ D=bi with the corresponding N io ¼ Dbi . the coordinates of one of the end points of the
Where the subscript o refers to the point at which interval in (A.5) as explained shortly.
the error of estimation is maximum (Emax). For any product, the maximum error, EiMax, for
For any interval i, let NiL=D/ni and NiR=D/mi any interval i can be computed as
represent the abscissas of the left and right end pffiffiffiffiffiffiffiffi
points, respectively. Then, given the coordinates of E iMax ¼ ai  2 Dbi ðA:5Þ
 ARTICLE IN PRESS

C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30 29

pffiffiffiffiffiffiffiffiffiffi
by substitution of X io ¼ D=bi and Nio=D/Xio in Substituting these values in (A.9), a(1)1 will be
Eq. (A.1). determined:
If the maximum error is above the TE, the  qffiffiffiffiffiffiffiffiffiffiffi

interval will be split into two at Xio. Let us assume D=nð1Þ  Dbð0Þ qffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffi
 1 1 
that EiMax>TE for the first interval, hence, we að1Þ
1 ¼  qffiffiffiffiffiffiffiffiffiffiffiffiffi  D=bð0Þ þ Dbð0Þ 1 :
qffiffiffiffiffiffiffiffiffiffiffiffiffi ð1Þ ð0Þ 
1
 n1  D=b1 
split interval 1 into two at X 1o ¼ D=bð0Þ
1 . The
coordinates of the end points of the two newly ðA:10Þ
constructed intervals are (Fig. 2):
Similarly, the slope and y-intercept for the line
Interval 1: segment approximating the second interval can be
calculated, for example, using the coordinates of
ðnð1Þ ð1Þ ð1Þ ð1Þ
1 ; N 1L Þ and ðm1 ; N 1R Þ or its left end, as
qffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffi
 
ðnð1Þ ; D=n ð1Þ
Þ and D=bð0Þ1 ; Dbð0Þ ; ðA:6Þ   qffiffiffiffiffiffiffiffiffiffiffi 
1 1 1 N ð1Þ  N ð1Þ   Dbð0Þ  D=mð1Þ 
ð1Þ  2L 2R   1 2 
b2 ¼  ð1Þ  ¼  qffiffiffiffiffiffiffiffiffiffiffiffiffi ; ðA:11Þ
Interval 2:  n  mð1Þ   ð1Þ 
2 2  D=bð0Þ 1  m 2 
ðnð1Þ ð1Þ ð1Þ
2 ; N 2L Þ and ðm2 ; N 2R Þ or
ð1Þ
qffiffiffiffiffiffiffiffiffiffiffi 
qffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffi  
 Dbð0Þ  D=mð1Þ qffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffi
D=bð0Þ1 ; Dbð0Þ and ðmð1Þ ð1Þ
2 ; D=m2 Þ ðA:7Þ  1 2 
1 að1Þ
2 ¼  qffiffiffiffiffiffiffiffiffiffiffiffiffi  D=bð0Þ
1 þ Dbð0Þ 1 :
 ð0Þ ð1Þ 
 D=b1  m2 
ð1Þ ð1Þ ð0Þ ð1Þ ð1Þ
where qm 1 ¼ ffi n2 ¼ X 1o , and N 1R ¼ N 2L ¼
ffiffiffiffiffiffiffiffiffiffi ðA:12Þ
N ð0Þ
1o ¼ Dbð0Þ 1 . Also, note that the left end of the
first interval and the right end of the last interval Next, we calculate the maximum error and
will always remain the same regardless of the check if it is below TE. For example, the maximum
number of intervals created. These are the end error for the two intervals created at the end of
points that defined the first interval at iteration 0. step 1 would be
qffiffiffiffiffiffiffiffiffiffiffi
In other words, set nð1Þ ð0Þ ð1Þ
1 ¼ n1 and m2 ¼ m1 . The
ð0Þ

next step would be to construct the equations of E 1 Max ¼ a1  2 Dbð1Þ

ð1Þ
1
the lines approximating the newly created inter- and
vals. The slope and the y-intercept can be qffiffiffiffiffiffiffiffiffiffiffi
calculated according to Eqs. (A.3) and (A.4) and E 2 Max ¼ að1Þ
2 2 Dbð1Þ 2 :
by substitution from (A.6):
 This process of splitting intervals can be
   qffiffiffiffiffiffiffiffiffiffiffi
N ð1Þ  N ð1Þ  D=nð1Þ  Dbð0Þ  repeated until the maximum error in every interval
 1L 1R   1 1  is below TE.
bð1Þ
1 ¼  ð1Þ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffi ; ðA:8Þ
n1  mð1Þ   ð1Þ ð0Þ 
1  n1  D=b1 
 qffiffiffiffiffiffiffiffiffiffiffi References

D=nð1Þ  Dbð0Þ 
 1 1 
að1Þ ð1Þ ð1Þ
1 ¼ b1 X 1 þ L1 ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffi X 1 þ Lð1Þ
1 :
Bretthauer, K., Shetty, B.S.S., White, S., 1994. A model for
 ð1Þ ð0Þ  resource constrained production and inventory manage-
 n1  D=b1 
ment. Decision Sciences 25 (4), 561–580.
ðA:9Þ Buchan, J., Koenigsberg, E., 1963. Scientific Inventory Man-
agement. Prentice-Hall, Englewood Cliffs, NJ.
To calculate the value of a(1)
we need to evaluate
1 Chen, C-K., Min, K.J., 1994. A multi-product EOQ model with
L1(1) at either end point of the interval it pricing consideration—T.C.E. Cheng’s model revisited.
approximates. For example, if we evaluate L(1) 1 at
Computers and Industrial Engineering 26 (4), 787–794.
Gallego, G., Queyranne, M., Simchi-Levi, D., 1996. Single
the right end of theqfirst interval,
ffiffiffiffiffiffiffiffiffiffiffi we know from
qffiffiffiffiffiffiffiffiffiffiffiffiffi resource multi-item inventory systems. Operations Research
(A.7) that Lð1Þ
1 ¼ Db ð0Þ
1 and X 1o ¼ D=bð0Þ 1 . 44 (4), 580–595.
 ARTICLE IN PRESS

30 C. Haksever, J. Moussourakis / Int. J. Production Economics 97 (2005) 18–30

Goyal, S.K., 1973. Determination of economic packaging Lee, W.J., 1994. Optimal order quantities and prices with
frequency for items jointly replenished. Management storage space and inventory investment limitations. Com-
Science 20 (2), 232–235. puters and Industrial Engineering 26 (3), 481–488.
Goyal, S.K., 1978. A note on ‘Multi-product inventory Maloney, B.M., Klein, C.M., 1993. Constrained multi-item
situations with one restriction’. Journal of Operational inventory systems: An implicit approach. Computers and
Research Society 29 (3), 269–271. Operations Research 20 (6), 639–649.
Goyal, S.K., Belton, A.S., 1979. On ‘A simple method of Maloney, B.M., Klein, C.M., 1997. A multi-linear constraint
determining order quantities in joint replenishments under inventory system. International Journal of Operations and
deterministic demand’. Management Science 25 (6), 604. Quantitative Management 3 (1), 27–40.
Güder, F., Zydiak, J.L., 1999. Ordering policies for multi-item Nahmias, S., 1993. Production and Operations Analysis. Irwin,
inventory systems subject to multiple resource constraints. Homewood, IL.
Computers and Operations Research 26, 583–597. Page, E., Paul, R.J., 1976. Multi-product inventory situations
Güder, F., Zydiak, J., Chaudhry, S., 1995. Non-stationary with one restriction. Operational Research Quarterly 27 (4),
ordering policies for multi-item inventory systems subject to 815–834.
a single resource constraint. Journal of the Operational Parsons, J.A., 1966. Multi-product lot size determination when
Research Society 46, 1145–1152. certain restrictions are active. Journal of Industrial En-
Hadley, G., Whitin, T.M., 1963. Analysis of Inventory Systems. gineering 17 (7), 360–365.
Prentice-Hall, Englewood Cliffs, NJ. Rosenblatt, M.J., 1981. Multi-item inventory system with
Haji, R., Mansuri, M., 1995. Optimum common cycle for budgetary constraint: A comparison between the Lagran-
scheduling a single-machine multiproduct system with a gian and the fixed cycle approach. International Journal of
budgetary constraint. Production Planning and Control 6 Production Research 19 (4), 331–339.
(2), 151–156. Rosenblatt, M.J., Rothblum, U.G., 1990. On the single
Hall, N.G., 1988. A multi-item EOQ model with inventory cycle resource capacity problem for multi-item inventory systems.
balancing. Naval Research Logistics Quarterly 35, 319–325. Operations Research 38 (4), 686–693.
Hariga, M., Jackson, P.L., 1996. The warehouse scheduling Roundy, R., 1985. 98% effective integer-ratio lot-sizing for one-
problem: Formulation and algorithms. IIE Transactions 28, warehouse multi-retailer systems. Management Science 31,
115–127. 1416–1430.
Hartley, R., Thomas, L.C., 1982. The deterministic, two- Silver, E.A., 1976. A simple method of determining order
product inventory system with capacity constraint. Journal quantities in joint replenishments under deterministic
of Operational Research Society 33, 1013–1020. demand. Management Science 22 (12), 1351–1361.
Holt, C.C., Modigliani, F., Muth, J.F., Simon, H.A., 1960. Taha, H.A., 1987. Operations Research. Macmillan, New
Planning Production, Inventories, and Work Force. Pren- York.
tice-Hall, Englewood Cliffs, NJ. Tersine, R.J., 1976. Material Management and Inventory
Homer, E.D., 1966. Space-limited aggregate inventories with Systems. Elsevier, New York.
phased deliveries. Journal of Industrial Engineering 17 (6), Thomas, L.C., Hartley, R., 1983. An algorithm for limited
327–333. capacity inventory problem with staggering. Journal of
Johnson, L.A., Montgomery, D.C., 1974. Operations Research Operational Research Society 34, 81–85.
in Production Planning, Scheduling, and Inventory Control. Ziegler, H., 1982. Solving certain singly constrained convex
Wiley, New York. optimization problems in production planning. Operations
Kaspi, M., Rosenblatt, M.J., 1983. An improvement of Silver’s Research Letters 1 (6), 246–252.
algorithm for the joint replenishment problem. IIE Transac- Zoller, K., 1977. Deterministic multi-item inventory systems
tions 15 (3), 264–267. with limited capacity. Management Science 24 (4), 451–455.