-
Seediscussions,stats,andauthorprofilesforthispublicationat:http://www.researchgate.net/publication/228312281
TheDynamicLotSizeModelwithStochasticDemands:ADecisionHorizonStudyARTICLEinINFORINFORMATIONSYSTEMSANDOPERATIONALRESEARCHJUNE2009ImpactFactor:0.41
CITATIONS4
DOWNLOADS223
VIEWS158
2AUTHORS,INCLUDING:
SureshSethiUniversityofTexasatDallas408PUBLICATIONS8,157CITATIONS
SEEPROFILE
Availablefrom:SureshSethiRetrievedon:15June2015
-
Electronic copy available at:
http://ssrn.com/abstract=1422108
THE DYNAMIC LOT SIZE MODEL WITHSTOCHASTIC DEMANDS:
A DECISION trORIZON STUDY*SITA BHASKARAN
Graduate School of Business, University of Pittsburgh,
Pittsburgh, PA 15260and
Operating Sciences Department, General Motors Research
Laboratories, Warren, MI 48090
SURESH P. SETHI
Faculty of Management Studies, University of Toronto, Toronto,
M5S 1V4
ABSTRACT
In a multiperiod decision problem it is usually the decisions in
the first or firstfew periods that are of immediate importance to
the manager. Decision/Forecasthorizon (DH/FH) research deals with
the question of whether optimal decisions inthe first or first few
periods (known as DH) can be made without regards to thedata from
some future period (known as FH) onwards.
In this paper the Dynamic Lot Size Model (DLSM) with stochastic
demands isstudied. The model has a setup cost and hence does not
have a convex cost strvictureusually assumed in the earlier DH/FH
research involving stochastic problems. It isshown that in the
stochastic DLSM, unlike in the convex cost problems, there
areexamples without DH/FH. However, randomly generated problems
seem univer-sally to have decision horizons. Sufficient conditions
for the existence of DH/FHare obtained in terms of salvage
functions, in terms of states of the system, andin terms of
decision strategies. A comparison of these approaches concludes
thepaper.
Keywords : Decision horizon, forecast horizon, stochastic
demands, concave cost,dynamic lot size model (DLSM), Wagner Whitin
algorithm.
RESUMEDans un probleme de decision multi-periode, ce sont
normalement les decisionsdans la ou les premiere(s) periodes qui
sont d'une importance immediate pour lemanager. La recherche de
l'horizon de decision / de prevision (DH/FH) a pourbut de savoir si
les decisions (connues comme DH) peuvent etre faites sans
con-sideration de donnees des periodes futures (connues comme
FH).
Le modele dynamique de la taille des lots (DLSM) en cas de
demande stochas-tique est etudie dans cet article. Le modele a un
cout de commande et de la, n'apas la structure convexe des couts
normalement supposee dans la recherche DH/FHanterieure impliquant
des problemes stochastiques. II est demontre que dans leDLSM
stochastique, contrairement aux problemes de cout convexe, il
existe des
Received Dec 1983, Revised May 1987 INFOR vol. 26, no. 3,
1988
213
-
Electronic copy available at:
http://ssrn.com/abstract=1422108
214 S. BHASKARAN AND S. SETHI
exemples sans DH/FH. Cependant, des problemes generes au hasard
semblent uni-versellement avoir des horizons de decision. On
obtient des conditions suffisantespour l'existence de DH/FH en
termes de fonction de cout de remplacement, entermes d'etats du
systeme et en termes de strategies de decision. Une comparaisonde
ces approches conclut cet article.
1. INTRODUCTIONIn a multiperiod decision problem it is usually
the decisions in the first or firstfew periods that are of
immediate importance to the manager. Decision/Forecasthorizon
research deals with decision making in these periods. To find the
optimalinitial decision, finite horizon problems of increasingly
longer horizon are solvedtill the initial decision converges. A
decision horizon procedure is a stopping rulespecifying when
further increasing of the problem horizon will have negligible or
noeffect on the initial decision. The horizon {T) when the stopping
rule is satisfied iscalled the forecast horizon (FH) and the number
of initial periods (t < T) for whichthe optimality of the
solution is unaffected by demand beyond period T is calledthe
decision horizon (DH). Bes and Sethi (1987) provide a more
elaborate definitionof DH/FH.
The decision horizon concept has been studied mainly in problems
with de-terministic demands. Wagner and Whitin (1958) were the
first to introduce thedecision horizon concept. They used it in the
area of production planning prob-lems and solved the Dynamic Lot
Size Model. The work on this problem was latersurveyed and extended
by Lundin and Morton (1975).
Decision horizon research in problems with stochastic demands
has been re-stricted to cases with convex or proportional cost
structure. Bhaskaran and Sethi(1987) survey this literature. In
this paper the Dynamic Lot Size Model (DLSM)with stochastic demands
is studied. This study differs fundamentally from otherstochastic
decision horizon research because the DLSM has a setup cost and
hencedoes not have convex cost structure. Thus unlike all DH
research till now the twincomplexities of stochastic demand and
non-convex costs are considered simultane-ously.
In Section 2 we formulate the stochastic DLSM. A backward
algorithm is de-rived for its solution and an illustrative example
presented. Section 3 illustrates thatunlike convex cost problems,
DH/FH may not exist in non-convex cost problems.Section 4 derives
three sufficient conditions for the existence of FH in the
stochasticDLSM.
2. STOCHASTIC DYNAMIC LOT SIZE MODEL (DLSM)
2.1 Decision Horizon Research Undiscounted Problems with
Stochastic Demands.In deterministic undiscounted problems, decision
horizon results depend on the no-tions of regeneration points,
production points and regeneration sets (Lundin andMorton (1975)).
These notions do not generalize to the case of stochastic
demandsbecause the inventory levels to be visited at future periods
are not known beforehand.The notion of regeneration set was
generalized by Morton (1979) for the stochastic
-
DYNAMIC LOT SIZE MODEL 215
case in two ways the state regeneration framework and the value
function regen-eration framework. Stochastic decision horizon
research for undiscounted problemstill now has used certain
monotonicity properties of convex cost problems arisingfrom the
value function regeneration framework to get decision horizon
results andto prove, under mild conditions, that DH/FH exist. These
monotonicity propertiesdo not hold for non-convex cost problems and
hence the methods of convex costproblems will not work for
non-convex cost problems. Bes and Sethi (1987) studydiscounted
stochastic problems.
2.2 The Stochastic DLSMConsider a multiperiod undiscounted
inventory problem with Dt, Pt, It ' f = 1,2,...denoting the
nonnegative demand, production, and ending inventory
respectively,in period t. The production cost is a fixed cost and
is K, if Pt > 0, and 0 otherwise.Holding and penalty costs of h
and p per unit are charged on the average inven-tory and average
shortage, respectively, in a period. Demands A are
independentrandom variables. The problem with horizon T (or
T-period problem) is
C(T) = min'{P,}
subject toIt = It-i + Pt - Dt, lo = 0.
Pt>0, te (1, T)where
i^^^^^\K, ifx>0
and
riyx Utj ^) ^ X ^ Ut^'^"^ ^ \^ hx/2 -h p{Dt - x)/2 if X <
A-
2.3 Form of optimal policyThe optimal policy (see e.g. Hillier
and Lieberman (1980)) at period t, t e (1, T) isdetermined by two
critical numbers St and St{>St) as follows:
St L-l if I 1 < s0 otherwise.
2.4 Two point demand distributionIn numerical examples, for
computational simplicity, we assume that demand Dttakes on either a
high value Ht or a low value Lt, both being equally
probableintegers. We also assume that backlogging is not allowed
and that demands andinventory levels are bounded, i.e., there exist
H" and / " such that
-
216 S. BHASKARAN AND S. SETHI
Ht < H" and 0 < /, < /",for every t e (1, T).
This leads to an optimal policy based on a single critical
number St for everyperiod:
' \ 0, otherwise.Also, because it is never optimal to hold
inventory for more than {K/h) periods, thefollowing stronger bounds
on inventory are obtained:
/", V)where V = {K/h)H". Thus, /, G {0,1,..., W}.
2.5 Solution methods: Forward vs. Backward AlgorithmsThe Wagner
and Whitin forward algorithm and the Lundin and Morton results
solvethe deterministic DLSM efficiently but are not applicable to
stochastic problemsbecause the notions of regeneration set and
regeneration monotonicity (Lundin andMorton (1975)) are not
applicable to stochastic problems. Backward procedures arereadily
applicable to stochastic problems since they derive an optimal
action forevery possible inventory level. We will derive such an
algorithm. This algorithm iscomputationally more demanding than the
Wagner and Whitin algorithm as it doesnot use previously obtained
solutions.Backward algorithm. Since actions depend only on the cost
function modulo anadditive constant, the cost relative to a base
state /* is defined to always be zero.Let
Ft,T{I) ' Minimum relative cost of going from inventory level /
at the end ofperiod t to the end of period T. The cost is relative
because a largeenough constant is subtracted to make iv,r(^*) =
0-
Thus Ftj{I) satisfies these backward relations for every f e (0,
T - 1) :FtAJ) = ftAi) - ftAn FTAI) = 0
The above equations are solved iteratively to compute i^ o,
2.6 A Typical ExampleWe now present a typical example to
illustrate the nature of the solution to thestochastic DLSM: Let K
= 1, h = 0.05, I" = oo and the demand sequence bePeriod (/): 1 2 3
4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20Ht: 2 4 8 3 1 5 4 7 3 2
6 8 9 1 1 4 3 6 5 5Lt-. 1 3 4 2 1 4 4 5 1 2 5 3 8 1 1 2 3 4 2 4
-
DYNAMIC LOT SIZE MODEL 217
As increasingly longer horizon problems are solved, the nature
of the solution andthe inventory level in the various periods are
sho^wi in Figure 1. The numbersHt, St are marked as heavy dots. We
note that for 7" > 6 we have Si = 6. Thusthe first period
decision converges. As problem horizon increases, the number
ofinitial periods, for which the decision converges, also
increases. This indicates theexistence of DH/FH but there is no way
of rigorously identifying them.
10-;
i 1 -^1 2 3 4 5 6 7t
T = 8
Figure la. A Typical Example (Part 1)
-
218 S. BHASKARAN AND S. SETHI
1 2 3 4 5 6 7 8 9 10
T = 10
15-
10^
I' ^
1 1 1
= 11
1 2 3 4 5 6 7 8 9 10 Ut
9 10 11 12
Figure 1b. A Typical Example (Part 2)
3. NONEXISTENCE OF DECISION HORIZONCases where decision/forecast
horizons do not exist have not been studied. Exis-tence issues in
DH/FH research are not yet settled. Lundin and Morton (1975)remark
that decision horizons or near decision horizons existed in their
vast empir-ical study. But there are problems where DH/FH do not
exist, as we will show. Inthe DLSM with almost stationary (but not
stationary) demand it sometimes hap-pens, in both the deterministic
and stochastic cases, that as problems of increasinglylonger
horizon are solved the optimal initial production keeps oscillating
infinitely
-
DYNAMIC LOT SIZE MODEL 219
proving that decision horizons do not always exist. This does
not happen in convexcost problems. Below are such examples.
3.1 Deterministic ExampleLet K = I, h = .Qh and consider the
following three demand sequencesPeriod (0 ".Demand (A) Case
1:Demand (A) Case 2:Demand (A) Case 3:
1 2 3 4 5 6 7 8 9 10 11 12 13 14... T5 5 5 5 5 5 5 5 5 5 5 5 5
5... 55 4 5 4 5 5 5 5 5 5 5 5 5 5 . . . 54 5 5 4 4 4 5 4 5 4 5 4 5
4 ...
Case 1. The T-period solution is as follows: If, for some
positive integer n (n > 2) T = Zn, the optimal solution is
unique and the initial production is 15, r = 3n -I-1, the optimal
solution is not unique and initial productions of 15
and 20 are both optimal, T = Zrt -\- 2, the optimal solution is
not unique and initial productions of 10
and 15 are both optimal.
Thus, in this case the initial decision does converge (to 15)
and there is a DH(= 3) even though there may be other optimal
initial decisions as well. By slightlychanging the demand structure
in the first few periods it is possible to constructexamples where
the optimal initial solution is unique and the initial
productionoscillates. This is shown in Case 2.
Case 2. The solution, in this case, is always unique and the
optimal initial productionis 14, 18 or 9 according asT = 3, 3n
-I-1, or 3 -f- 2 respectively. Hence no DH. Thecost penalty (i.e.,
excess over minimum cost) on using the three initial productionsPi
= 14, 18 and 9 is given in Table 1.
Table 1.Cost Penalty Table: Deterministic Example (Case 2)
Pi14189
T3n
0.1.2
3/7-1-1
.150.1
3n + 2.05.15
Thus, for e < .15, there is no production Pi with cost
guaranteed to be within eof optimal. Oscillation of the initial
decision can also occur when the demand hasa repetitive pattern as
shown in Case 3.
Case 3. Here, except for the initial six periods, the demand is
alternately 5 and 4.The optimal initial decision is again unique
and is 14, 18 or 9 according as T ~ 3n,3 -h 1 or 3 -f- 2
respectively. Again, no DH.
-
220 S. BHASKARAN AND S. SETHI
3.2 Stochastic ExampleConsider the previous example with the
following stochastic demand sequence (hav-ing a two point
distribution):Period (t) : 1 2 3 4 5 6 7 8 9
7 7 7 7 7 7 7 7 77 7 6 6 6 6 6 6 6
T16
For every T > 2, the solution of the T-period problem is
unique. The optimalinitial production (Pi) is 14 or 21 according as
T is even or odd respectively. Thecost penalty (i.e., excess over
minimum cost) on using the two initial productionsPi = 14 and 21
are given in Table 2:
Table 2.Cost Penalty Table: Stochastic Example
Pi1421
TEven
0.05
Odd.075
0
For e < .05, there is no production Pi with cost guaranteed
to be within e ofoptimal. It should be noted that the demand is
nearly stationary and the dispersionof demand is small (i.e.,
difference between high and low demand is small) thus the demand is
only marginally stochastic. Only in such cases, to the
authors'knowledge and experience, DH/FH does not exist.
4. THREE APPROACHES FOR DETERMINING DH/FHSufficient conditions
for the existence of DH/FH in the stochastic DLSM are ob-tained
using three different approaches: States, Decision strategies, and
Salvagefunctions.
4.1 StatesIn problems with stochastic demands, decisions only
determine the expected prob-ability distribution of inventory
levels. Thus, it helps to generalize the notion of"inventory level"
to that of "state" of the system, where a "state" is defined as
aprobability distribution of inventory levels. This done, it is
possible to formulatea forward dynamic programming algorithm to
solve the stochastic problem in thefollowing manner:
Step 1. Set^ = O.Step 2. Enumerate all "states" that can be
reached at the end of period t-i-1
from the end of period t. For every such "state" consider only
thecheapest path (from t = O)hy which it can be reached.
-
DYNAMIC LOT SIZE MODEL 221
Step 3. If the optimal initial decision corresponding to all
"states" of thesystem at period / + 1 is the same, stop, f + 1 is a
FH. If not, setf = r + 1 and go to Step 2.
Example 4.1. Let AT = 1, A = .1, / " = 10 and as in previous
examples let A have atwo point distribution:
Period (0 : 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
2122Ht\ 5 4 5 5 3 5 5 4 5 5 5 4 5 5 3 5 4 5 4 5 5 5Lt: 2 3 3 5 2 4
2 4 3 4 4 2 5 4 2 5 4 3 4 4 3 4
On applying the above forward algorithm FH = 22, 5i = 6 and DH ~
1. Let IX^ I :be the number of states at end of period t. Then \Xt\
is shown below:
t: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22\Xt\:
5 13 11 13 26 83 29 55 85 184 148 109 156 97 113 97 97 65 161 129
209 161
In the backward algorithm of Section 2 there are only {W + 1)
inventor}^ levels ateach period. In the forward algorithm the
number of states at end of period t canincrease rapidly with t and
make the algorithm computationally complex. However,if as in
Example 4.1, the number of possible decisions at any period can be
reducedto a few possibilities (in this example the number varies
between 2 and 5) then theforward algorithm is applicable.
4.2 Decision StrategiesIt is a common property for the product
of the transition matrices of a Markovchain to tend to a constant
matrix (i.e., a matrix with identical rows). This property,known as
"loss of memory" implies that the effect of the initial decision
wears outafter sufficient time. This leads to the following
result:
Theorem 4.1 : If there exists a T > 2, such that for all
possible decisions betweenperiods 2 and T, the "state" at the end
of period T is independent of the "state" at theend of period 1,
then T is a forecast horizon.
The above theorem provides conditions for the existence of FH
which are indepen-dent of K and h. They are strong conditions which
are sufficient but not necessary.In particular, these conditions
will never be satisfied by deterministic problems.
Example 4.2. Let /" = 10. Consider the two point demand
sequence
Period {t)\ 1 2 3 4 5 6 7 8 9 10Ht: 5 4 6 5 7 3 5 6 4 6Lt: 1 2 2
2 1 2 2 2 2 2
On applying Theorem 4.1, FH = 10. DH depends on K and h. Morton
and Wecker(1977) deal with infinite horizon optimality for Markov
Decision Processes.
-
222 S. BHASKARAN AND S. SETHI
4.3 Salvage FunctionsIncorporating a salvage function for
terminal inventory leads to the following obser-vation: An A'^
-period problem is equivalent for any T T is equivalent to a
problem with horizon lengthT and appropriate salvage function G{I)
for terminal inventory. G{I) is subject tothe bounds (proof
omitted)
If I* = 0 the above inequality reduces to
-K < G{I) 1, the critical number Si corresponding to the
solutionto the T-period problem with salvage function G, is the
same for every function Gwithin the bounds given above, then T is
an FH.
The number of possible salvage functions between the given
bounds is very large.However, one could try out an arbitrary number
of salvage functions and if they allgive the same first period
decision, conclude that T is an apparent FH.
Example 4.3 In the typical example of Section 2, let disposal be
free. On solvingincreasingly longer horizon problems and randomly
generating 50 possible salvagefunctions for every horizon length,
an apparent FH of 6 was obtained with a DHof 2.
-
DYNAMIC LOT SIZE MODEL 223
4.5 Comparison of the three approachesNo disposal vs. Disposal.
In most of the paper the model considered is the stan-dard DLSM in
which disposal of inventory is not allowed. This keeps the form
ofpolicy simple. This is suitable for the state and decision
strategy approaches whichare based on enumeration of possible
decision strategies. In the salvage functionapproach, the model is
altered to allow disposal as this bounds salvage functions.Disposal
complicates the form of policy but a complicated policy does not
affect theefficiency of the salvage function approach.
While in the state and salvage function approaches, the T-period
problem issolved for increasing values of T, this is not so in the
decision strategy approach.
The first two approaches are feasible only if a few alternative
decisions are underconsideration at every period. Otherwise, one
runs into computer storage problemsin the state approach and into
CPU time problems in the decision strategy approach.The
combinatorial complexity of the salvage function approach exceeds
that of thefirst two approaches. However, trying out a number of
possible salvage functions isone way of determining apparent
forecast horizons.
ACKNOVt'LEDGEMENTThis research is supported in part by grant
A4619 from NSERC of Canada.
5. REFERENCES
1. Bes, C. and S.P. Sethi (1987) "Concepts of Forecast and
Decision Horizons: Applicationsto Dynamic Stochastic Optimization
Problems" Math, of O.R. forthcoming.
2. Bhaskaran, S. and S.P. Sethi (1987) "Decision and Forecast
Horizons in a StochasticEnvironment: A Survey" Optimal Control
Applications and Methods Vol.8, pp 201-217.
3. Chand, S. (1982) "A note on Dynamic Lot Sizing in a Rolling
Horizon Environment"Decision Sciences No\. 13, No. l ,pp
113-118.
4. Chand, S. and T.E. Morton (1986) "Minimal Forecast Horizon
Procedures for DynamicLot Size Models" Naval Research Logistics
Quarterly Vol. 33, No. 1, pp 111-122.
5. Hillier, F.S. and G.J. Lieberman (1980) Introduction to
Operations Research Holden DayInc.
6. Kleindorfer, P. and H. Kunreuther (1978) "Stochastic Horizons
for the Aggregate Prob-lem" Management Science Vol. 24, No. 5, pp
485- 497.
1. Lundin, R.A. and T.E. Morton (1975) "Planning Horizons for
the Dynamic Lot SizeModel: Zabel vs. Protective Procedures and
Computational Results" Operational Re-search Vol. 23, No. 4, pp
711-734.
8. Morton, T.E. (1978a) "Universal Planning Horizons for
Generalized Convex ProductionScheduling" Operations Research Vol.
26, No. 6, pp 1046-1057.
9. Morton, T.E. (1978b) "The Nonstationary Infinite Horizon
Inventory Problem" Man-agement Science Vol 24, No. 14, pp
1474-1482.
10. Morton, T.E. (1979) "Infinite Horizon Dynamic Programming
Models A PlanningHorizon Formulation" Operations Research Vol. 27,
No. 4, pp 730-742.
-
224 S. BHASKARAN AND S. SETHI
11. Morton, T.E. and W.E. Wecker (1977) "Discounting, Ergodicity
and Convergence forMarkov Decision Processes" Management Science
Vol 23, No. 8, pp 890-900.
12. Wagner, H.M. and T.M. Whitin (1958) "Dynamic Version of the
Economic Lot SizeModel" Management Science Vol 5, pp 89-96.
13. Zabel, E. (1964) "Some Generalizations of an Inventory
Planning Horizon Theorem"Management Science Vol. 10, pp
465-471.
Sita Bhaskaran is a Senior Research Scientist at the
GeneralMotors Research Laboratories. She works in the Production
andLogistics Group of the Operating Sciences Departemnt. Her
re-search interests are in applied Operations Research and her
cur-rent work is in the areas of facility location and supplier
ratio-nalization. She received her Ph.D. at the University of
Adelaide,Australia and prior to joining GM taught at the Graduate
Schoolof Business, University of Pittsburgh.
Suresh P, Sethi is General Motors Research Professor of
Opera-tions Management at the University of Toronto. He has an
MBAfrom Washington University and an M.S. and a Ph.D. in
Indus-trial Administration from Carnegie-Mellon University. He wasa
Connaught Senior Research Fellow in 1984-85. He is currently
aPrincipal Investigator for the Manufacturing Research Corpora-tion
of Ontario, a Center of Excellence funded by the
ProvincialGovernment. His research interests are in production
planning
problems under uncertainty, problems of scheduling in
manufacturing, and stochas-tic dynamic optimization problems. His
articles on these and other topics haveappeared in a variety of
journals including Management Science, Operations
Research,Mathematics of Operations Research, SIAMJ. of Control and
Optimization, SIAM Review, Ad-vances in Applied Probability, J.
ofEcon. Theory, Naval Research Logistics Quarterly, Journalof
Optimization Theory and Application, IEEE Transactions on Automatic
Control, EuropeanJournal of Operational Research, International
Journal of Production Research, and The FMSMagazine. He has also
co-authored a book and co-edited a special issue for INFORon the
subject of optimal control theory and applications.