Sunteți pe pagina 1din 9

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/245314982

The Multi-Product Production Cycling Problem

Article in IIE Transactions September 1980


DOI: 10.1080/05695558008974511

CITATIONS READS

60 37

1 author:

Stephen C. Graves
Massachusetts Institute of Technology
163 PUBLICATIONS 6,635 CITATIONS

SEE PROFILE

All content following this page was uploaded by Stephen C. Graves on 28 March 2017.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
The ,Multi-Product Production
Cycling Problem

STEPHEN C. GRAYES
Massachusetts Institute of Technology
Sloan School of Management
Cambridge, Massachusetts 02139

Abstract: The multi-product production cycling problem is concerned with the determination of a
production/inventory policy for a single capacitated production facility which is dedicated to producing
a family of products. This paper studies this problem assuming stochastic demand. The one-product
problem may be analyzed as a Markov decision problem and solved as such. For the multi-product
problem, a heuristic decision rule is proposed based on the analysis of the one-product problem, and a'
new notion: the composite product. This heuristic is then tested by simulation on a series of problems,
and is shown to be the most effective of the heuristics considered for the set of test problems.

The multi-product production cycling problem( MPCP) is primarily on lot-sizing; the capacity allocation decisions come
to determine production and inventory control policies for a into play only when capacity constraints are violated which
family of products, each of which requires processing on a force an adjustment in the lot-sizing decisions (e.g. see [1]).
single capacitated facility. The production of each product For stochastic demand, it is not possible to consider lot-
goes into inventory from which demand for the product is sizing and capacity sequentially. A solution procedure which
serviced. The criterion for the control policy is minimum focuses on lot-sizing loses the flexibility needed to cope
cost per unit time, where cost consists of set up costs, with the capacity allocation decisions under unknown
inventory holding costs, and backorder costs. demand. Similarly, considering capacity allocation at the
Most of the literature on the MPCP has been confined to expense of lot-sizing is too rigid to handle fluctuating
models which assume deterministic demand. Previous work demand and inventory levels.
on the MPCP for stochastic demand has been either very ex- The intent of this paper is to develop and test heuristic
ploratory in nature ([3], [15)), or hampered by extremely control policies for the stochastic demand MPCP. The
restrictive assumptions [7}., Elmaghraby [2] gives a comp- specific problem to be considered is defined by the follOWing
rehensive bibliography for the MPCP with constant known assumptions:
demand rates. Deterministic results do not seem to be
I. Demand for the product family is characterized by a
particularly helpful in solving the stochastic problem. The
known stationary distribution function.
reason seems to lie in the complexity inherent in the prob.
lem; any solution procedure for the MPCP must simultane- 2. The form of the decision policy is periodic reView,
ously consider lot-Sizing and capacity allocation decisions. where at most one product may be produced in a period.
For deterministic demand, the existing procedures focus Furthermore, the product must be produced for the entire
duration of the period at the product's finite productiOn
rate. The period length must be chosen such that these
assumptions are reasonable.
3. Production in a period is available for inventory after
Received April 1979; revised July 1979; re-revised April 1980. Paper
was handled by Scheduling, Planning and Control Department. a fixed lead time.
Work supported in part by the Office of Naval Research under 4. ~ set up cost is incurred if the product to be produced
Contract NOOOl4-7S-e-oSS6. was not prod.,ed (set up) in the immediately preceding

September 1980, AIlE TRANSACTIONS, 0569-SS54/80/0900-o233/$2.00/0 1980 AIlE 233


period. Whenever the facility is shut down, any prior set up Thus, the expected cost from following the optimal decision
is lost and is not carried over. Set up times are assumed to policy for the remaining t periods is the minimum of:
be zero. (1) the expected cost if we do not produce in the current
5. Demand exceeding on-hand inventory is backordered. period; (2) the expected cost if we do produce in the
current period. In each case, the expected cost is the
6. The objective is to minimize the expected cost per
expected backorder and inventory holding cost for the
unit time over an infmite horizon over all products. The
current period, plus a set up charge, if applicable, plus the'
components of the total cost are set up, inventory holding,
expected cost for the remaining t-l periods. Note that the,
and backorder costs.
machine status (k) in the next period depends only on the
The remainder of the paper is divided into three sections.
current decision. Also, if the current decision is to prt)duce,
In the next section, heuristics are developed for the MPCp
inventory is increased by P, and a sei up charge is incurred
based on the analysis of the one-product problem. Then
only if the machine was not previously set up.
these heuristics are tested by simulation over a series of test
This formulation may be easily modifled to allow for a
problems. Finally, we give a brief summary of the results of
the paper. lead time equal to an integral number of periods; this exten-
sion is analogous to that given in (14] for a pure inventory
model.
Development of Heuristics
We conjecture that the stationary steady-state optimal
The One-Product Problem policy is a two-c.itical-number policy (/*,I**), such that the
facility is turned on when the inventory frrst drops below
The simplest version of the MPCP is the one-product problem. 1*, and is not shut down until the inventory fIrst exceeds
This problem may be formulated as' a Markov decision 1** (1** > 1*). This conjectu,re is very reasonable; indeed it
problem over a two-dimensional state space consisting of is difficult to fathom any other poli<:y form being optimal.
the product's inventory level, and the machine's status (set Unfortunately, no proof or counterexample has been found
up or shut down). Each period the decision is either to shut for this conjecture. However, in [8] strong evidence is given
down or to produce for the entire period. The cost expres- in support of the conjecture: two-critical-number policies
sion for the one-product problem for a fInite time horizon are proved' optimal for several special cases, and are found
with zero lead time can be defIned by backward recursior. , to be optimal for several numerical examples.
as follows: To determine optimal 'policies for the one-product prob-
lem, one can use (I) to recursively compute the optimal
C(I,k,t) = min {G(l) + EA[C(I-A, 0, t-l)]; (1) production decisions for all values of (/,k,t). Altermi.tively,
to fInd the stationary steady-state optimal policy as t..,. 00 ,
G(I+P) + (l-k)K + EA[C(I+P-A,l, t-l)] } the policy iteration method for sequential decision problems
where [13] can be used. Since this method requires a fmite state
space; it is necessary to restrict the inventory level to range
C(I,k, t) = expected cost for remaining t periods assuming between some lower and upper inventory levels. By choos-
an optimal decision policy, given inventory level ing these boundaries intelligently, the truncated problem
I and machine status k. will be an accurate representation of the original problem. '
To gain insight into the behavior of the optimal policy
C(I, k, /=0) =0 for all I, k for the one-product problem, we defIne a value function
V(I, t) as
o if facility is not set up for production
k ={ 1 if facility is set up for production =
V(I,t) G(l) -G(I+P)+ E A[C(I-A,O,t-l)] ,
;
(2)
; .,,'.

G(l) = expected backorder and inventory holding cost for -EA [C(I+P-A,I,t-1)).
'H~n!q :;,:;;i.:~).;.;;
current period, given net inventory levell available 0":

for servicing demand. , The optimal production policy is determined'by this (unc-
tion: if the machine is set up (k=l); a decision to produce is
EA[C(I-~k,t)] =expectation of the cost function taken ,made if and only if V(I it) is positive; if th~ machine is not
, over A, the demimd random variable. 'set up (k=O), a decision to produce is made if and only if
, V(l,t) is greaterthari K, the set up cost. Note thatif V(I,t)
, , is non-increasing in I, the optimal policy is a two-critical-
= ~ p(i) C(I-i, k, t) for P(i) = probability number policy.
1=0 V(/,t) may be interpreted as the value of a producing
that demand (A) equals i. versus nonproducing production facility given an inventory
level of I and t periods remaining; that is, a rational product
=
P production rate measured in units per period. manager would be willing to pay as much as V(I,t) to turn
on the machine in the current period, given that in subse
K =set up cost. quent periods it will cost K to set up the machine. Hence,

0234 AIlE TRANSACTIONS, VolUme 12, No.3


if this value is greater than the actual cost of turning on the tion is still desirable. A weakness, however, is that the
maahine in the current period (K or 0, depending on the heuristic does not try to anticipate potential production
machine's status), a decision to produce is made. conflicts. As an example, consider a two-product problem
For the steady-state problem we define V(l) to be the in which the machine is shut down, and the inventory level
limit of V(l,t) as t goes to infinity. V(l) is readily computed for each product is slightly above its critical level. One.
from the results of the policy iteration method. product analysis would indicate no production in the
Two continuous-review versions of the one-product prob- current period; however, in the subsequent period hath
lem have been treated analytically. In [6] , Gavish and Graves inventory levels will likely fall below their critical levels, and
consider the problem with unit Poisson demand. Relying both products will want to be produced. Since only one
largely on results from 'queueing theory, they develop re- product can be produced per period, the inventory for one
cursive relationships for expressing expected cost as a func- of the products will continue to fall, incurring increasing
don of the inventory control policy; these cost functions ,costs, until it is able to wrest the production facility away
can therfbe: 'searched to find an optimal policy. In [9], from the other product. '
Graves and Keilson examine a similar problem with Poisson To help anticipate possible conflict situations, the com.
demand arrivals but where the size of the demand requests posite product is introduced. In the context of a two
is an exponential random variable; here an analytical expres- product example, suppose the product units are redefined
,sion is fOWld for the expected cost as a function of the so that both products have eqUivalent production rates.
control policy. Define the composite product to be the aggregation of the
two products by setting the demand for the composite
product to be the sum of the demands for the two products,
One-Product Analysis Applied to MPCP and the cost structure for the composite product to be
some composite representation of the individual product
The multi-product problem can also be formulated as a cost structures. The production rate of the composite
Markov decision problem, and in theory can be solved by product is just the production rate for either product.
the policy iteration method. However, the computation This composite product can now be used to recognize
required by' this method increases exponentially with the potential production conflicts. The composite' product can:
number of products, and is computationally infeasible for be analyzed as if it were one product to find a production
reasonably-sized problems. Consequently, we have concen- policy (1:,/'1*). A decision to "produce" the composite
trated on finding good heuristic policies Which are easy to ,product is made by comparing the composite inventory
compute and implement. Ie =II +/2 with either 1'1* or 1'1, depending upon whether
A heUristic solution for this problem can be developed the composite product is or is not "set up"; a decision to ,
on the basis of the one-product analysis: For each product, , "produce" would indicate that the composite inventory is .
compute the optimal policy assuming the facility is dedicated not adequate for the composite demand. The composite
to that product. Each product's .policy may be executed as product should be helpful in anticipating potential conflicts;
long as this does not conflict with the execution of another when bOtll individual product inventories are just adequate
product's policy. A conflict occurs whenever more than one when considered separately, in composite the composite
product is to be produced in a given period, in which case inventory should indicate the need for current production.
the heuristic must resolve the conflict. For example, consider The notion of the composite, product is immediately
the case where the machine is set up for product I, product extendable to the n-productproblem. For n products, there
1 has inventory(ll) less than its second critical number If*.
and product 2 has inventory (12 ) less than its first critical are (j) possible compOSite products consisting of; products
number If; here both products have triggered and the each, and hence, up to T.';..2(j) = 2n -n-l composite products
heuristic must choose one to produce in the immediate to defme and analyze.
period. One procedure is to use the value function derived
from the one-product analysis, and produce the product The defmition of composite product is very similar to
with highest "value" for the machine. For the stated that of the aggregate product used in aggregate production
'example, VI(lt} is to be compared against V2 (I2 ) -K, planning (e.g., [10], [12]). However there are important
. where Vie ) is the value function from the one-product differences in the intended function of these concepts. The
analysis for product i; note that if the product is not set up, need for an aggregate product is primarily due to data and
the set up charge must be subtracted' from thevaiue computational considerations; without some aggregation,
function to give the effective value of the machine. the planning problem becomes computationally unwieldy.
In comparison, the composite product is a tool for detailed
scheduling decisions where capacity planning decisions have
The Composite Product previously been made; its use is based on the need to
account for product interdependency due to the sharing of
The procedure based on the one-product analysis is reason production capacity. The composite product is intended to
able; clearly if the one-product analysis indicates production help compensate for the lack of scheduling flexibility due
'for a 'product, then in the multi-product problem, produc- to the fIxed capacity levels.
~ ..... ..

,}. September 1980, AIlE TRANSACTIONS 235


Statement of the Heuristic Testing of Composite-Product Heuristic

We now state a general heuristic procedure for the MPCP. The Identical Cost Problem
For convenience, suppose the products are arranged so that
Idd l ~Iz/dz ~ ... ~In/dn, where Ij is the inventory level The proposed heuristic relies heavily upon the use of
and dj is the expected demand rate of product j. That is, composite products. Since there may be up to 2n -n-l dis-
the products are ordered according to periods of supply. Let tinct composite products, it is important to be able to easily
k denote the status of the machine where k=j for the defme and analyze the possible composite products. A
machine set up for productj, j=l, ... ,n, and k=n+ 1 for the special case of the general MPCP for which composite
machine shut down. The notation [;] is used to represent products are immediately applicable is the identical-cost
the composite product consisting of products 1,2, .. . ,J, problem. Here, inventory holding and backorder costs for
while [;-I,k1is the composite product consisting of products all products are linear, and proportional to production
1, 2,.. ::: 1-'1 and product k. By convention, for k ~ j rates; the set up costs for all products are the same. That is,
[;-I,k] =[;], for k =n+ 1, [;-1 ,k] is not defined, and forj=l, for i=l, 2,... ,n
1I-1,k] is just product k.Note that the composite product
[i] is not currently set up for j < k, while [i-l,k] is always
considered to be set up. The heuristic considers possible bi = blPi
production of at most two products: product I, which has
least inventory, and product k, which is currently set up. A Ki = K
decision to "produce" [i] corresponds to initiating produc- where
. tion of product 1; a decision to "produce" Ii-l,k] is a con-
tinuation of production of product k. 111e general heuristic hi = inventory holding cost per unit per period for
is as follows. producti,
bi = i)ack; order cost per unit per period for product i,
The Composite-Product Heuristic
1. Letj =1. . Pi = production rate in units perp~riod for product i.
2. Is production indicated for 1I-1,k]?'
If yes, go to step 3.
Ki = set up cost for product i,
If no, go to step 4.
h,b, K= positive constants.
3. Is production indicated for II]?
If yes, choose to produce either product 1 or product The demand distributions may vary across produ'cts. Now,
k; this choice is based on comparing value functions, if a unit of each product is redefined so that all products
i.e. VI (II) vs.Vz(Ij ) -K. have the same. production rates(~.g., rescale Pi = 1 for all
. If no, produce product k. products), then the products have identical cost parameters
This problem and its assumptions are very realistic and
4. Is production indicated for Ii]? have been observed in practice III] . Consider the invento.iy
Ifyes, produce product 1. .holding and backorder costs; these costs are frequently
If no, go to step 5. taken to be proportional to the value of a product. In a
.' single machine environment the value of a product consists
5. Letj =j+1. of the cost of the input plus the value added to this input
If! ~n, go to step 2. dUring processing. Given that all products require the same
If j =n+ I, shut down the facility. or similar inputs, it is reasonable to suppose that the machine
processes the inputs at a relatively steady dollar rate,
The heuristic is an iterative procedure over j. For eachj, independent of product. Furthermore, if the rate of value
the two composite products, Ii-l,k] and II], are considered added by the machine were not nearly constant across
for possible production; these decisions are based upon the products, it could be. argued that the machine capacity is
. one-product analysis of the respective composite products. .' not being used efficiently; Hence, the products should have
A decision to "prOduce" either Ii] or Ii-l,k] terminates the . similar. value,and hence similar inventory holding and back-
process (steps 2, 4); a decision to "produce" both is contra- order costs per unit of production. Set up costs generally
dictory and is resolved in step 3. If no composite product reflect the labor cost for altering the set up of the machine.
triggers, the facility is shut down (step 5). If all products require a similar amount of work for their
In the next section, the implementation of this general set ups, it is reasonable to assume they have the same set
heuristic is discussed; in particular, when the products have up costs.
identical cost structures, composite products are easily de- Given identical-cost products, composite products are
fined and the heuristic is readily applied. This composite- easily defined. All composite products. have the cost para-
product heuristic is then compared with several alternative meters of the individual products, and the common produc-
heuristics by simulation on a series of test problems. tion rate; the demand distribution for the composite product

236 AIlE TRANSACTIONS, Volume 12, No.3


is just the convolution of the demand distributions for the tively. This heuristic will switch to the most-critical product
indiVidual products. Consequently, the analysis of the whenever 1*- /2/d2 ~/** -1t/d 1
composite-products entails merely solving the one-product
problem with the common cost and production parameters These three heuristics are based on standard notions of
for a range of demand distributions inventory theory. The first heuristic H3 corresponds to a
(Q,R) policy, while the second heuristic H4 represents an
(s,S) policy. Heuristic H5 is a modification of 84 which
Description of Heuristics allows for switching of products prior to the set up prOduct
reaching the order-up-to level (1**).
Five heuristic scheduling procedures, including the com- No theory exists for determining the parameters for
posite-product heuristic, are proposed for testing on the these three -heuristics. The inventory theory developed to
identical-Qst~l?roblem. These heuristics are described below. fmd (Q,R) or (s,S) policies is not useful since it is restricted
.::- . .
to single-product uncapacitated problems. Therefore, for
each of the three heuristics, a grid search is done over a
HI : This heuristic is the naive procedure based solely restricted policy space to estimate the best set ofparameters.
on the one-product analyses of the individual products. To determine the expected cost for each policy in the grid,
Each product is governed by its optimal policy for the one-
the system is simulated for 600 periods where the first 100
product problem. If no products trigger, the machine is periods are for the initialization, while costs are recorded
shut down; if exactly one product triggers, the product is only over the last 500 periods. The policy with least cost is
produced. If more than one product triggers, the- product to then chosen to be the best choice of parameters. 1
produce is found by comparing value functions.

H2: This heuristic is the composite-product heuristic; Description of Simulation


it extends HI by using composite products in the decision
procedure as outlined in the previous section. Again, value A computer program has been developed to simulate the .
functions are used to resolve conflicts when more than one MPCP. This simulation program is characterized by the \.
composite product triggers in a period. following set of specifications: -

.These two heuristics have been selected so that the 1. Demand for each product is an independent Poisson,
effectiveness of the composite product can be studied. To process where At is the demand rate for product i, i= 1,2, , n.
test the .effectiveness of these two heuristics compared with
possible alternative heuristic policies, three additional heuris- 2. The production rate for each product is one unit per
tics have been defmed. period. This implies that expected machine utilization is
U = 1;~=1 At (U < I).
H3: The heuristic H3 is characterized by two parameters
(/*,N*). When a product is set up, it is produced for N* 3. Inventory holding and backorder costs are linear with '
consecutive periods. After N* periods of production, either respect to the net inventory level at the end of the period.
the facility is shut down or the most-critical product is set
up. The most-critical product is that product with the least 4. All products have identical cost structures. In particular, .
number of periods~of-supplyof inventory; that is, product j the inventory holding cost rate is I cost unit per unit of
is most critical if ~/dj =mini =l ... n Y,fdf } where/f is the inventory per period, while the backorder cost rate is 5 cost
inventory level and d j is the expected demand rl!,te for units per bac.korder per period. The set up costK is,an input
product i. The decision on whether to produce the most- parameter.
critical product depends on whether its inventory measured
in periods-of-supply is less than /*. 5. The production lead time is ten periods.

H4: The heuristic H4 is parameterized by (1*,/**). The 6. All simulation runs are for 20,200 periods. The first 200
product currently set up continues to be produced until its periods are used to initialize the system. System costs are .
inventory, in periods-of-supply. reaches or exceeds /**. At accumulated over the remaining 20,000 periods. .
that time, the machine is shut down only if the most-
critical product has inventory, in periods-of-supply, greater Given. this description of the simulation, a t~t problem
than /*; otherwise the most-critical product is set up. is specified by setting the number of products n, the demand
rates for the products (At), and the set up c9st K. To run
H5: The heuristic H5 is identical to H4 except that it
allows for the machine to be switcl-ed from one product to
1An examination was made of the sensitivity of this search procedure
another product prior to the first produci'siriventoryreach- . to the length of the dm~lation run; a run size of 600 periods was
ing /**;, let II and 12 be the inventory levels of the product found to be sufficient for f"mding an optimal or near-optimal para-
being produced, and of the most-critical product, respec- meter choice.

September 1980, AIIETRANSA<;:TIONS 237


the spnulation.a heuristic policy must be given for making smaller as the size of the blocks increases. A block size of
production decisions. For each test problem, the generated 200 periods seemed to be sufficient to reduce the depen-
demand realization is identical for every heuristic that is dency so that each sample point could be treated as an
considered for that test problem. Hence, any variation across independent observation. A statistical test for autocorrela-
the performances of the heuristics on the test problem is tion in the time series could have been performed to verify
due to the heuristics, and is not a consequence of differences this observation.
in the generated demand. To test the null hypothesis .that two heuristics are equiva-
lent for a specific test problem, a student's t-statistic can be
formed from the differences between the two matched cost
Results of Simulation samples. Table 2 reports these t-statistics when the four
heuristics. HI, H3, H4, and H5 are compared individually
The five heuriStics are tested on four identical-cost problems. against H2. the composite-product heuristic. The null hypo-
Each test problem has five products, the demand for each thesis that H2 is equivalent to HI, H3, or H4 on any test
product i is Poisson with rate At, and the machine utilization problem can be easily rejected. In comparing H2 with HS,
is 90% for each problem. The set up cost is set to obtain the H2 clearly outperforms HS only for problem a. For the
given economic order quantity (EOQ) for the product with other three test problems, the null hypothesis cannot be
the smallest demand rate (i.e., ~1 ::: .1). The four test rejected with 95% confidence. Despite this lack of statistical
problems are as follows: significance, H2 seems to perform better than H5. The
differences between H2 and H5 on the test problems for
a) ~1 = ~2= ~3= ~= .1'~5 ::: .5; EOQ::: 2. which H2 is best are greater and more significant than the
difference between the two heuristics for problem b. Hence,
b) ~1 = ~2= ~3= ~= .1'~5 = .5: EOQ = 5. based on this limited evidence, the composite product
heuristic $eems to be the most effective procedure considered
c) ~1 = ~2= ~3= .1, ~= ~5= .3; EOQ::: 2. for the identical-cost problem~

d) ~1 = ~2= ~3=.1, ~= ~= .3; EOQ ::: 5. Table 2: HestS for heuristics compared against H2
for identical-eost problem.
The results from the simulation are given in Table 1
heuristic
where the entries in the table are the average total cost per
period. In comparing HI with H2, it is clear that the intro- test problem H1 H3 H4 H5
duction of the composite product in H2' provides a substan- a 6049 6.44 5.82 4.14
.tial improvement.' Heuristics H3 and H4 have performances b 4.03 6.08 7.08 -a.35
comparable to HI b\lt are outperformed by H2. The heuris- c 6.27 8.50 5,42 1.18
tic H5. however, compares well with heuriStic H2. Heuris- d 2.18 2.65 5.04 1.47
tic H2 performs best in three of the test problems while All t-values are with 99 degrees of freedom [6].
, H5 is slightly better in test problem b.

Table 1: Simulation results for Identical-eost problem.


The simulation experiments were repeated on a series of
six test problems for which demand across products is
allowed to be correlated. HeuriStic HI. H3, H4, and H5 are
heuristic
test problem H1 H2
not altered when applied to these conelated-demand prob
H3 H4 H5
lems. For heuristic H2, however. the specification of the
a 22.50 17.41' 20.72 20.58 18.26 demand distribution for all composite products must reflect
b 28.91 26.31 28.92 30.18 26.14' the correlation between individual products. Each test'
c 22.93 18.58' 20.84 21.51 19.32 problem assumes two identical' products with machine
d 29.54 28.51 29.54 31.59 29.16 utilization of 90%. that is, )., ::: .45 for i:::l,2. Demand is a.
Indicates best performing heuristic for test problem. bivariate Poisson process with a correlation coefficient of p
where I pi'" 1 (see [4]). A test problem is characterized by
To verifY these observations. a statistical test has been specifYing a value for p=O, .33, or .66, and for E0Q=2 or
constructed from the simulation runs by breaking the 5 units.
simulation run of 20,000 periods into 100 blocks of 200 The simulation results for the test problems are presented
periods. A sample of 100 cost observations is obtained for in Table 3, where the table entries are the average total cost '
each heuristic on each test problem by observing the average per period. Heuristic H2, utilizing the composite product, is
cost over each block. The sample points for a heuristic are a dramatic improvement over HI. Indeed, the magnitude of
not totally independent since the costs incurred in one the improvement from the composite product increases
period are correlated with the. costs incurred in the previous as the demand correlation increases. Alternative heuristics
period. This dependency between sample points grows H3 and H4. again, do not perform well with respect to H2.

'.238 AIlE TRANSACTIONS, Volume 12, No~ 3


,....Table 3: Simulation results for correlated-demand problem. the composite-product heuristic outperformed HS in only
seven of ten test problems. We should note, however, that
heuristic
whereas the parameters necessary to implement the compo-
test problem H1 H2 H3 H4 H6 site-product heuristic are found by an analysis of a series of
p=O/EOQ=2 13.52 11.21 15.69 15.54
simple one-product problems, there is no analytical proce-
13.49
17.37 12.77' 16.07
dure for fmding the parameters for H5; rather, the para-
p"'.33/EOQ=2 14.37 13.74
20.91 14.18 15.50 17.25 14.07'
meters are determined from a grid search of the parameter
p=.66/EOQ=2
p=O/EOQ=5 16.37 14.17 21.57 17.44 13.77
space, where a simulation is needed for the evaluation of
p",.33/EOQ=6 20.17. 15.10 19.75
each grid point.
19.25 15.77
p=.66/EOQ=5 23.64 16.21' 19.15 20.29 16.68
"Best performance for test problem.
:-:;: . ~
Acknowledgment
However, H5 does better than H2 in two of the six test
problems. The author wishes to thank Professors Warren H. Hausman
To test the dominance of H2, a statistical analysis and Leroy B. Schwarz for their contributions, advice, and
identical to that for the identical-cost problem has been encouragement during this research, and to acknowledge
done, and is given in Table 4. In comparifig H2 with either the helpful comments of two anonymous referees.
HI, H3, or H4, the null hypothesis that the pair ofheuristics
are identical is easlly rejected for all test problems. However,
when comparing H2 with H5, there is no simple conclusion References
on the relative performance of the two heuristics. The null
hypothesis that the two heuristics are identical can be [I} Doll, C. L., and D. C. Whybark, "An Iterative Procedure for
rejected with 90% confidence for three of the test problems the Single-Machine Multi-Product Lot Scheduling Problem,"
Mfl1IIlgement Science, 20 1;50-55 (September1973).
and with 85% confidence for four of the teSt problems. [2) Elmaghraby, S., "The Economic Lot Scheduling Problem
However, for one problem (p=O/EOQ=S) H5 is significantly (ELSP): Review and Extensions," Management Science, 24'6,
better than H2. Hence, the evidence presented is not conclu- 587-598 (February 1978).
sive, but suggests the need for additional experimentation. (3) Evans, R. V., "Inventory Control of a Multiproduct System
with a Limited Production Resource," Naval Rereorch LogiJ
tier (Jumterly, 142, 173-184 (June 1967).
(4) Fener, W., An Introduction to Probability Theory and In
Table 4: Nests.... for heuristics compared against H2 Applicationr, Volume I, Third Edition, New York: John
for correlated-demand problem.
Wiley and Sons, Inc., 279 (1968).
heuristic [5} Fishman, G. S., andP. J. Kiviat, ''The Analysis of Simulation-
Generated Time Series," Management Science, 137, 52S-557
test problem H1 H3 H4 H5 (March 1967).
[6} Gavish, B., and S. C. Graves, "A One-Product Production!
p=O/EOQ=2 4.67 6.13 6.30 4.88 Inventory Problem Under a Continuous Review Policy,"
p=.33/E00=2 4.58 3.20 4.79 2.18 Working Paper Series No. 7735, Graduate School of Manage-
p"'.66/EOQ=2 5.06 2.84 5.86 -0.75 ment, University of Rochester, Rocllester, NY, November
p-o/EOQ"'5 4.46 1977, to appear in Operationr Rereorch.
7.32 5.26 -2.24
p=.33/EOQ"5 4.97 5.71 5.10 1.58
[7) a
Goyal, S. K., "Lot Size Scheduling on Single Machine for
Stochastic Demand," Management Science, 19 11, 1322-25
p-.66/EOQ=5 5.34 6.62 6.31 1.69 (July 1973).
.. " All t-values are with 99 degrees of freedom. [8} Graves, S. C., "The Multi-Product Production Cycling Prob-
lem," unpublished dissertation, University of Rochester
(October 1977).
[9} Graves, S. C., and J. Keilson, "The Compensation Method
Applied to a One-Product Production/Inventory Pro'blem,"
Working Paper Series No. 7913, Graduate School of Manage-
,'L ".:"' ment, University of Rochester, Rochester, NY (July 1979),
this paper the MPCP has been defined and analyzed. In to appear in Mathematicr ofOperationr Rereorch.
particular a composite-product heuristic has been proposed [10} Hax, A. C., "Aggregate Production Planning," in Handbook
(If Operationr Rereorch, j. Moders and S. Elmaghraby (editors)
and tested on a series of identical-cost problems. This Van Nostrad Reinhold (1978).
. heuristic relies on the analysis of the one-product problem, [11} Hax, A. C., and H. C. Meal, "Hierarchical Integration of Pro-
which is well understood and computationally manageable, duction Planning and Scheduling," North Holland/TIMS,
and on the notion of a composite product. This composite- Studies in Management Science, Vol. 1, Logirticr, North Hol-
product heuristic was compared by a simulation experiment land-American Elsevier, 53-69 (1975).
l12} Holt, C. C., F. Modigliani, J. F. Muth, and H. A. Simon,
against four alternative heuristics. The composite-product
Planning Production, Inventorier and Work Force, Prentice
heuristic clearly dominated three of these heuristics on the Han (1960).
test problems considered, while the fourth heuristic H5 was [13] Howard, R. A., Dynamic Programming and Markov ProcerseB,
slightly inferior to the composite-product heuristic; indeed, New York: John Wiley and Sons (1960). .

September 1980, AilE TRANSACTIONS


.:' .~ ': ..: ....
't.''' .. ;.",,1 . .
[14]-"Karlin, S., and H. Scarf, "Inventory Models of the Arrow- Stephen C. Graves is an Assistant Professor of Management at .the
t' Harris-Marschak Type with Time Lag," Chapter 10 in K. J. Alfred P. Sloan School of Management, Massachusetts Institute of
Arrow," S. Karlin, and H. Scarf (eds.), Studies in the Mathe- Technology. He received an AD and MBA from Dartmouth College,
matical Theory of Inventory and Production, Stanford, CA: and a PhD in business administration from the University of
Stanford University Press (1958). Rochester. He has publiShed several articles in the area of production
{IS] Vergin, R. C., and T. N. Lee, "Scheduling Rules for the and operations management. His current research interests include
Multiple Product Single Machine System with Stochastic planning for multistage logistics systems, production planning and
Demand,"INFOR, 16 1,64-73 (February 1978). scheduling, and the design of automated production systems.

240 AIlE TRANS.l\CTlON~, Volume 12, No.3

View publication stats

S-ar putea să vă placă și