Professional Documents
Culture Documents
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at .
http://www.jstor.org/action/showPublisher?publisherCode=informs. .
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science.
http://www.jstor.org
The Lot-SizingProblem
Capacitated
with Linked Lot Sizes
n this paper a new mixed integer programming (MIP) model formulation and its incorpo-
ration into a time-oriented decomposition heuristic for the capacitated lot-sizing problem
with linked lot sizes (CLSPL) is proposed. The solution approach is based on an extended
model formulation and valid inequalities to yield a tight formulation. Extensive computa-
tional tests prove the capability of this approach and show a superior solution quality with
respect to other solution algorithms published so far.
(Lot-Sizing; MIP; Valid Inequalities;CLSPL)
1. Introduction only solve the lot-sizing problem, but also the schedul-
Depending on the characteristics of the production ing problem. Different models are distinguished with
process and the planning detail required, different respect to the number of items producible on each
types of lot-sizing models are commonly used in prac- resource per period (one or two) and the extent of
tice. The main difference is between big-bucket and a bucket's capacity utilization (all-or-nothing or par-
small-bucket models. In big-bucket models a bucket tial). The discrete lot sizing and scheduling problem
size is selected, such that several products requir- (DLSP) allows at most one product per period, and
ing different setup states can be produced on each production must be for the full period or not at all (all-
resource in each bucket (period). This yields a model or-nothing) (Fleischmann 1990). The continuous setup
formulation, which charges a setup cost and/or time lot-sizing problem (CSLP) allows at most one product
in every period a specific item is produced. As the per period, but does not limit the production amount
bucket size is usually much larger than setup times per period (Karmarkar and Schrage 1985 and Salomon
(when present), the corresponding error due to the 1991), whereas the proportional lot sizing and schedul-
nonpreservation of the setup state between succes- ing problem (PLSP) allows up to two products per
sive buckets is usually small and negligible. The period which share available capacity, with no restric-
standard problem in this context is the capacitated tions on the extent to which available capacity has to
lot-sizing problem (CLSP), for which different mixed- be used up (Drexl and Haase 1995).
integer model formulations (Eppen and Martin 1987, In practice, to obtain more accurate plans, there is
Stadtler 1996) and various solution algorithms (e.g., a trend towards smaller bucket sizes. This holds true
Trigeiro et al. 1989, Tempelmeier and Derstroff 1996) especially in the multilevel case where large planned
are well known. lead times of orders are a drawback inherent to big-
In contrast, small-bucket models usually allow at bucket models. On the other hand, in small-bucket
most one setup activity per period. As the sequence of problems the planning horizon has to be divided into
products is clearly defined in this situation, they not many more buckets to yield a solution comparable to
1
Figure Characterization
ofModels
Big-bucket(e.g., CLSP)
Long
period
J products/period LL time
1 2 3 4 5
Small-bucket(e.g., PLSP)
Short period
?2 products/period time
1 2 3 4 5 6 7 8 9 10
CLSPL
Longperiod " -- . -- .
..
J products/period
Linkedlot-sizes time
1 2 3 4 5
I
• ~ of lotsof differentproducts
Capacityconsumption
that a setup state can no longer be preserved over (681 instances) to prove the solution quality of their
more than one consecutive bucket boundary. He did algorithm. Their model and test data again include
so because he observed in his test data that both prob- setup times.
lems yield the same or similar solutions most of the
time, but the newer model formulation is much easier.
Haase (1998) compares his heuristic BACLSPL with a
two-step procedure which first solves the CLSP and
3. Model Formulation
afterwards links two lots of the same product in adja- The model of the CLSPL under consideration here is
cent periods simply by a greedy heuristic. He is able based on the following assumptions:
to conclude from his extensive computational tests * The planning horizon T is fixed and divided into
that solutions obtained by BACLSPL are better than time buckets (1,..., T).
solutions of this two-step procedure (i.e., solutions * Resource consumption to produce a product j on
obtained by a model taking the possibility of linking a specific resource m is fixed, and there exists a unique
into account are better than solutions obtained by a assignment of products to resources.
standard CLSP approach and adding links in a second This assumption is in line with Tempelmeier and
step). Derstroff (1996), who argue that lot-size decisions in
Gopalakrishnan et al. (1995) introduce product a finite capacity model have to be based on "opera-
independent setup times in the model formulation, tions" while the term "item" or "product" are often
which become product dependent in Gopalakrishnan used for an aggregate of several operations. Still, we
(2000). Their MIP model formulation is characterized will use the latter two terms to ease understanding.
by an extensive use of binary variables in comparison The index m for resources is introduced here to
to the model formulations by Dillenberger et al. (1993)
anticipate the multilevel formulation in ?3.4. For
and Haase (1994). The decision problem, motivated
single-resource problems (??3.1-3.3), the index can be
by a real-world case, is then solved by a standard MIP abandoned (M = 1).
solver (LINDO). * Setups incur setup costs and consume setup time,
Sox and Gao (1999) again consider the CLSPL with-
thereby reducing capacity in periods in which setups
out setup times. They propose a new model formu-
occur. Setup costs, as well as setup times, are sequence
lation based on a shortest-route representation of the
independent.
problem yielding tight LP bounds. Their model for- * At most one setup state can be carried over on
mulation is sufficient to solve smaller test instances
each resource from one period to the next, such that
(J = 8 products, T = 8 periods, M = 1 resource) to no setup activity is necessary in the second period.
optimality within seconds, whereas they propose a * Single-item production is possible (i.e., the con-
Lagrangean decomposition heuristic to solve larger
servation of one setup state for the same product over
problems. As already stated by Haase (1998), they
also observe that restricting the preservation of setup two consecutive bucket boundaries).
* A setup state is not lost if there is no production
states, such that one setup state cannot be linked on
both bucket boundaries, is advantageous with respect on the resource within a bucket.
to the solution time of their algorithm. However, Therefore, the CLSPL contains a property of small-
buckets models, which is the preservation of setup
they derive this observation from a rather small test
set (five instances) citing a maximum deviation from states over bucket boundaries, while retaining its big-
optimality of 2.19%. bucket character.
Recently, Gopalakrishnan et al. (2001) take up To derive a mathematical model formulation, we
the CLSPL, now called capacitated lot-sizing prob- start with the standard model CLSP (e.g., Trigeiro
lem with setup carryover (CLSP-SC). They present et al. 1989). The proposed model formulation will be
a tabu-search heuristic to solve it and offer some called "Inventory & Lotsize" (I&L)-representation due
computational results on a comprehensive test set to the variables used.
MANAGEMENT
SCIENCE/Vol. 49, No. 8, August 2003 1041
SUERIE AND STADTLER
TheCapacitated
Lot-SizingProblem
3.1. Starting Base: CLSP The objective function (1) aims at the minimiza-
CLSP model formulation (I&L representation): tion of inventory holding and setup costs. Produc-
tion costs are assumed to be constant over time and
J T-1 IT
sc1 therefore irrelevant-in this decision situation. Con-
MinE• + . (1)
hj It, yE Yj straints (2) are inventory balance constraints, and
j=1 t=1 j=1 t=l
constraints (3) ensure that capacity required for pro-
Ijt,-l+ Xj=Ij+Pjt Vj= 1,... J, duction plus capacity used up by setup operations
t= 1, ..., T, (2) do not exceed available capacity. Setup constraints (4)
* +1 - force the binary setup variable Yt to "1" in case of
E X1t Y;jt C,,,t a positive production quantity X1t. Finally, variables
jR?,,a,, jR,,, st.
are restricted to be nonnegative or binary, respectively
Vm= 1,... , M, t=1,..., T, (3)
(5)-(7).
<O it Vj = 1,..., J, t= 1,... , T, (4)
Xjt Bjt =
0 (Ijo= = j 1,... , J, 3.2. Simple Plant Location (SPL) Representation
Ij IjT 0)
>_ t = 1, To obtain a tight model formulation, a standard refor-
,T-1, (5)
.... mulation, called simple plant location (SPL) represen-
Xjt >
0 Vj = 1 ...., J, t = 1, .... T, (6) tation, will be used (Rosling 1986, Stadtler 1996). In
this model formulation the production quantity vari-
YitE {0; 1} Vj= 1,...f, J, t= 1 , T. (7)
.... ables Xjt are replaced by variables Zt,, according to
Indices and index sets:
T
lines), m -=1,.../ M;
t = Periods, t = 1,..., T; D't denotes net demand for product j in period t,
whereas variables Z,,t can be interpreted as the por-
R,,, = Set of products j produced on resource m. tion of demand of product j in period t fulfilled by
Data: production in period s (see Stadtler 1996 for a proce-
dure to calculate net demands D't). New constraints
a,,,j= Capacity needed on resource m to produce one
unit of item j; (9) have to be added to ensure that each period's
= demand is met.
Bjt Large number, not limiting feasible lot sizes of
This leads to the following CLSP model formula-
product j in period t;
= tion (SPL representation) with (la)-(7a) correspond-
C,,, Available capacity of resource m in period t;
hi =Holding cost for one unit of product j per ing to (1) to (7) of the first model formulation:
period; I T-1T I T
t
Pit Primary, gross demand for item j in period
= Min E hj (t- s). +
Zy scj.-Yj; (la)
(with P)T including final inventory, if given for j=l s=l D't.Z1t
t=s j=1 t=l
the planning horizon T); T
sc1 = Setup cost for product j;
E YaI] E C,,,t
= Setup time for product j. jER,,, s=t j R,,, stj. Yj, <
D•'Zjts+
sty
Variables: Vm = 1,... , M, t = 1,... , T, (3a)
= Inventory of item j at the end of period t; = 1
Ij = Production amount of item j in period t (lot Z.ts < Vj ... , J, t= 1,..., T,
Xit Y.t s = t,... , T, (4a)
size);
Binary setup variable (=1, if a setup for item j
Yjt= Vj =1,..., J, t=1,..., T, D > 0, (9)
is performed in period t, 0 otherwise). EZs=1
s=-
Additionally, an assignment of products to resour- two of them. Constraints (19), now including vari-
ces seems necessary, which was already introduced in ables QQjt, guarantee that a setup can be carried over
the base model formulation (??3.1-3.3) by means of in period t only if either item j was set up in period
index m. t - 1 (that is, = 1) or the setup state is already
Yjt-
carried over from period t- 2 to t - 1, which implies
single-item production of product j in period t - 1
4. Extended Model Formulation (that is, QQjt-1 = 1). They replace constraints (11):
and Valid Inequalities
Now that a basic model formulation for CLSPL has _ , TJ,
<Yjt-1 + QQjt-=...
Wit
been described, we will try to strengthen it. Strength- t =2,...,T. (19)
ening in this context means to find a MIP model for-
mulation that yields tighter LP bounds. The first step Finally, constraints (20) have to be added to the model
will be to derive an extended model formulation by formulation, restricting the range of values for vari-
ables QQjt and, of course, nonnegativity constraints
adding variables; later, three groups of valid inequal-
ities will be presented. (21):
derived solely from the model data, this step is called ables QQj, hold ((23)-(24)):
preprocessing, which was also used by Maes et al. t+U-1
(1991) for the serial MLCLSP. Before a mathematical u-1 Vm= 1,...,M,je
formulation is presented, the idea will be shown by s=-t
QQjs Rm,t=2,...,
means of an example (Table 1). T-U+1,U=1,...,3, if
Net demands for items 1 to 3 of periods 1 to 3 are t-1 t-1
given, such that initial inventory is already absorbed LCms-L 1 Ds- E Stk
s=1 s=1kERm amk' kERm
in the demand figures (Ijo= 0). Taking into considera- I"-1 Dn>0
tion available capacity and that no backlog is allowed, t+U-1
one can compute "cumulative slack capacities," which - amk-D, <0. (23)
are the units of capacity available in or before period t s=t
kERM
k#j
for production of demands due in later periods. In
t+V-1
the example (Table 1) there cannot be single-item
E QQs< V-1 Vm=l ,..., M,t=2,...,
production in period 2: This is because to produce s=t jERm
of small-bucket models (Pochet and Wolsey 1995, Constraints (25) and (26) can be formulated with
Wolsey 1998) and will be transferred to CLSPL. The variables Zis, in the presence of the SPL representation
setup state conservation property of CLSPL will be (Wolsey 1989, Van Hoesel et al. 1994):
used here: If Yit = Wit = 0 for a product j in period t, r r
1046 MANAGEMENT
SCIENCE/Vol.49, No. 8, August 2003
SUERIEAND STADTLER
TheCapacitated
Lot-SizingProblem
those cuts that are violated during the MIP search are
< min ET
XQjtX~,( ami =m, St
jCI "QQjt
put into the matrix. Compared to C&B, this approach
tPj yields smaller matrices and faster solution times at
\dj = J/. . fJ, t =
21 T -1 (31) each node at the price of some separation procedure.
Variables: Here, no effort was taken to develop specialized sepa-
ration routines for the proposed cuts, but the standard
XQjt = Production quantity of item j in period t, if "model cut" feature of XPRESS-MP was used instead
this is a single-item production period.
These new inequalities (29) can be strengthened (Dash Associates 2000).
even further by aggregation over all products j pro- There are some advantages to these approaches
ducible on resource m, giving (32). This aggregation is (Miller et al. 2000). The first one is that the algor-
ithm-in principle-will terminate with an optimal
stronger than each inequality covering only a subset
of products because the term added on the left-hand solution. The second is that it will find a feasible
side (E;jERs(amj Xjt + stj Yjt)) is always solution if one exists, which is not guaranteed when
greater than
or equal to the term added on the right-hand side using a heuristic approach. On the other hand, both
might require immense amounts of memory and time,
.XQjt) due to (30).
(E•,Rs am,, but this drawback diminishes somewhat with the
E (amjXjt+ stj . iYt) progress in information technology. Nevertheless, to
jR,tl " solve larger instances, some heuristic approach seems
appropriate.
~ QQjt/ + Therefore, a time-oriented decomposition heuris-
\ jR,,, EajXQ5t
jER,,
tic proposed by Stadtler (2003) for the multilevel
Cmt'1-
Vm=1,...,M, t=2,...,T-1. (32) capacitated lot-sizing problem (MLCLSP) will be
The reasoning for (32) is: If there is single-item pro- adapted to fit CLSPL needs. The algorithm can
be characterized as a time-oriented decomposition
duction on resource m in period t, capacity consump-
tion on resource m is restricted to production of XQj,. approach. The planning horizon is divided into three
On the other hand, if there is no single-item produc- parts: the lot-sizing window and the time intervals
tion on resource m in period t, (32) simply reduces to preceding and following the lot-sizing window (see
the capacity constraints (3). Again, multilevel lot siz- Figure 2). In successive planning steps, a lot-sizing
window is moved through the planning horizon. Lot-
ing requires no modifications. Consequently, only (32)
is considered in the model formulations tested. sizing decisions, in terms of the model formulation
that are those constraints which include binary vari-
ables, are only considered in the lot-sizing window.
5. Solution Approaches In the time interval preceding the lot-sizing window,
Various solution methods (MIP, Lagrangean decom- setup decisions (binary variables) are fixed. In the
time interval following the lot-sizing window, only
position, tabu search) have been proposed for the
CLSPL (see ?2). Here, the CLSPL is tackled by a stan- inventory balance constraints and capacity constraints
dard MIP solver (XPRESS-MP,Release 12). (without the inclusion of setup times) are included in
One variant, called "Cut and Branch" (C&B), adds the model formulation to anticipate future capacity
all valid inequalities proposed in ?4 to the initial bottlenecks. The objective in each planning step is to
model formulation and then solves the resulting LP, minimize setup and inventory holding costs incurred
before entering the branch and bound process. up to the end of the current lot-sizing window.
A second variant, in the following called "Branch The reasoning for this approach is that a tight
and Cut" (B&C), is also tested. In this case, the valid model formulation can be used inside the lot-sizing
inequalities are not added to the initial model for- window, gathering their benefits without accepting
mulation, but stored separately in a cut pool. Only the drawback of an inflated matrix, if such a model
2
Figure Internally RollingScheduleswith Lot-Sizing Windows more planning steps will sometimes emerge, result-
(A/4r/( -, 4/1/2, Stadtler 2003) ing in no solution for the complete problem. In real-
periods t =T ity, actual capacity consumption by setups will lie
5
somewhere in between these two extremes. This is
0 00 & why a third variant (STE)is introduced. Here, capacity
1. Planning step:
--0&-0 O losses due to setups are estimated by their mean real-
2. Planning ized in the periods preceding the lot-sizing window
step: &00o
000
plus some safety margin. Furthermore, the planning
3. Planning step: 0 0 00 0
horizon effect can be reduced by the introduction
of bonuses in the objective function. Because the
4. Planning
step:-
0 00 bonus concept has originally been developed for the
single-level uncapacitated case (Stadtler 2000), it only
Legend: approximates the true costs of lot sizing within the
I I Lot-sizingwindow lot-sizing window in an LP relaxation of a capacitated
Intervalof timewithrelaxedintegrality constraints model. Still, bonuses have improved the solution
--------. Intervalof timefollowingthe lot-sizingwindow quality significantly for the CLSP (Stadtler 2003) as
well as for the CLSPL (this paper). Here, bonuses are
Intervalof timeprecedingthe lot-sizingwindow
calculated along the lines of Stadtler (2003) for setup
costs but not for setup times sty. Of course, the
formulation is used for the whole planning hori- scjof
overlap lot-sizing windows also reduces the plan-
zon. The heuristic is controlled by three parameters horizon effect (see Stadtler 2003 for more details
ning
A/T/I( and a time limit. A indicates the length of the regarding the time-oriented decomposition heuristic).
lot-sizing window, whereas (D represents the overlap
(in periods) of two consecutive lot-sizing windows.
Finally, T stands for the number of periods at the 6. Computational Results
end of the lot-sizing window with relaxed integral- In the remainder, some computational results re-
ity constraints. Thus, A > T holds. Within these garding the model formulation and the solution ap-
settings, a so-called >_
planning horizon effect or end proaches proposed will be given. Additionally, some
effect occurs (Fisher et al. 2001): Due to the objec- comparisons to solutions reported in the literature are
tive function, inventories are minimized at the end drawn. All computations are carried out on a PC
of each lot-sizing window; e.g., they will be zero (Windows NT 4.0) with Pentium IV, 1.7 GHz micro-
if no capacity bottleneck is anticipated. Therefore, a processor, and 256 MB RAM. As a MIP solver,
time-between-order cycle is imposed by the choice of XPRESS-MP release 12 with standard settings is used.
the length of the lot-sizing window. To overcome the
planning horizon effect, several options are proposed 6.1. Single-Level Test Instances Taken from
by Stadtler (2003) and will be used here. the Literature
To anticipate future capacity bottlenecks, three A starting point established four test instances by
variants were tested. The first one (STMIN) neglects Thizy and van Wassenhove (1985), which were also
capacity consumption due to setup times in periods used by Haase (1994) and Sox and Gao (1999). These
following the lot-sizing window. The second variant test instances are solved to optimality by the MIP
(STMAx)depicts the worst case, that is, if all items formulation of Sox and Gao (1999) as well as by
will have to be produced every period. In this case, the model formulation proposed here by a stan-
available capacity per period is reduced by the sum dard MIP solver, within seconds. Hence, it is possi-
of setup times of items producible on the specific ble to evaluate the solutions of heuristic BACLSPL
machine (less the smallest setup time). Consequently, of Haase (1994). These deviate from optimal solu-
if capacities are tight, infeasible problems for one or tions by 5.0% on average (TV11-TV14; J = 8 products,
T = 8 periods, M = 1 resource) on Haase's test * SPL/ext/vi- B&C: the extended formulation and
instances, which differ from the original in the setup all valid inequalities, Branch and Cut approach (??3.2
cost of product 5 due to a typing error. and 4);
As these test instances are rather small, a test set * Sox/Gao: Model formulation proposed by Sox
introduced by Trigeiro et al. (1989) for the CLSP, and Gao (1999) extended to work with setup times
which was also used by Gopalakrishnan et al. (2001) (see Appendix B for details);
* Gopalakrishnan: Model formulation proposed by
for the CLSPL, will be used in the following (named
TTM-test set). The test set is divided into three Gopalakrishnan (2000).
In the columns, the gaps to lower bound (LB)
phases (Phase I-III). No results for Phase I test
instances are reported in the literature, as these test [=(objective function value - LB)/LB] are calculated
based on the following concept: The solution consid-
instances are regularly used for parameter tuning.
ered is the best solution found after 10 seconds. In
Here, they are used to set the parameters for the case no solution was found after 10 seconds (num-
time-oriented decomposition heuristic (i.e., parame-
ber in brackets), the first solution was taken instead
ters of the lot-sizing window). From Phase I problems,
(which was found after maximal 17 seconds). The
we can conclude that the incorporation of bonuses is lower bound (LB) is either the model's own LP relax-
worthwhile, whereas we discard version STMAX.The ation (LBLP), the LP relaxation after automatic cut gen-
Phase II and III problems are clustered into Classes eration by XPRESS-MP (LBXLP), or the LP relaxation
1-9 according to Table 2. after automatic cut generation of model SPL/ext/vi
In Table 3, gaps to different lower bounds are (LBXLP(ext/vi)),which includes all valid inequalities
reported for Class 1 test instances. The rows show defined in ?4.
the different model formulations that have been The MIP model formulations I&L/basis/-, Sox/
tested: Gao, and Gopalakrishnan are intended to be used as
* I&L/basis/-: the initial formulation (??3.1 and a reference. From Table 3 it can be observed that the
3.3); proposed model formulation with valid inequalities
* SPL/ext/-: the extended formulation (??3.2 and not only yields better solutions (Column 4), but also
4.1); better bounds (Columns 2 and 3) than the model for-
* SPL/ext/vi- C&B: the extended formulation and mulations reported in the literature. Solutions of ver-
all valid inequalities, Cut and Branch approach (??3.2 sions SPL/ext/vi are-independent of the approach
and 4); chosen (C&B vs. B&C)-significantly better than solu-
tions of the three reference model formulations (based
on the one-sided Wilcoxon matched-pairs signed rank
Table2 ofTTM-Test
Classification Set test, which will be used throughout this paper at con-
fidence limits of 99%). Moreover, they both provide
PhaseII
significantly better solutions than SPL/ext/-.
Class # Items # Periods # Instances The LP bounds are rather weak if the valid inequal-
1 6 15 116 ities are not included in the model formulation right
2 6 30 5 from the beginning. They are considerably improved
3 12 15 5
4 12 30 5
by the automatic cut generation for I&L/basis/-
and Gopalakrishnan as well as for the B&C version,
5 24 15 5
6 24 30 5 because the valid inequalities can be drawn from
the cutpool here. The best solutions are obtained by
PhaseIII
the B&C approach, which also obtains a first solu-
7 10 20 180 tion faster.
8 20 20 180 It should be added that the model formulation
9 30 20 180
Sox/Gao, which is rather tight, is not as compact as
this model formulation. Although it requires fewer generation of model SPL/ext/vi (LBXLP(ext/vi)), which
constraints (1,308 vs. 2,118), the matrix is blown up includes all valid inequalities defined in ?4, is cho-
due to the use of variables (3,150 vs. 1,156) and sen here and in the remainder-with the exception
nonzero coefficients (57,150 vs. 16,071), which slows of the results of Gopalakrishnan et al. (2001), where
down the solution process in each node (in compar- gaps to LB are taken from their paper (Table 5). For
ison to SPL/ext/vi on Class 1 test instances). Fur- some test instances, to find even a first solution takes
thermore, in some rare cases their model formulation a lot of computational time (>20 minutes). Mean com-
will exclude the optimal solution (see Appendix B putational effort for Classes 7-9 is 61.2 seconds for
for details). SPL/ext/- (standard deviation 51.0), 145.3 seconds
Additionally, to judge the solution quality of the for the C&B approach (standard deviation 339.7), and
B&C approach, the MIP solver was given 600 seconds 142.9 seconds for the B&C approach (standard devi-
for each of the 116 Class 1 test instances. In 91 cases ation 283.8). In contrast, the time-oriented decom-
optimality could be proven within this time limit. position heuristic provides excellent solutions in a
Table 4 shows the capabilities of the time-oriented very short time, which shows the effectiveness of the
decomposition heuristic (?5) compared to the MIP model formulation.
model formulations. For the MIP model formulations, Table 5 compares the proposed approaches to the
the solution after 30 seconds is taken for Classes 1-3 tabu-search heuristic of Gopalakrishnan et al. (2001)-
and 5, whereas 60 seconds of computational time are the most recent results reported for the CLSPL.
allowed for Classes 4 and 6-9. Whenever no solu- The B&C approach (SPL/ext/vi- B&C) yields mean
tion is reached within these time limits (number in gaps to LB of about 1/10th of those reported there,
brackets), the first solution is taken instead. As a but computational efforts are much bigger, espe-
lower bound, the LP relaxation after automatic cut cially in Classes 7 to 9. However, when considering
1050 MANAGEMENT
SCIENCE/Vol. 49, No. 8, August 2003
SUERIE AND STADTLER
Lot-SizingProblem
TheCapacitated
Table5 TimesforTTM-Test
Gapsto LBandSolution ofClasses)
Sets(Aggregation
et al. (2001)
Gopalakrishnan
B&C
SPL/ext/vi- Heuristic
(6/2/2,B,STMIN) 550
[PentiumIII, MHz]
0 time 0 time
Classes Gapto LB [max:30/60s] Gapto LB [max:15s] Gapto LB 0 time[s]
the time-oriented decomposition heuristic, both the taken as the demand (setup cost, setup time) of the
solution quality as well as the computational efforts new item, whereas the average is taken for production
outperform those reported by Gopalakrishnan et al. coefficients and holding-cost coefficients. This setting
(2001). could also be the result of some aggregation of items
to product families. In such a case, the model formu-
6.2. Modified Single-Level Test Instances lation could be streamlined further by restricting the
Considering Classes 7 to 9, the CLSPL does not really link option to the long runners, which has not been
seem to be an appropriate model. What can be gained done here.
if one out of 30 setup states is carried over every Table 6 shows computational results for the new
period, especially if the feature to carry over one test instances. Again, 60 seconds of computational
setup state over two consecutive bucket boundaries time are allowed for each test instance. It can be
is never used here? Instead, the CLSPL should be
observed that the option to carry over one setup state
applied if only a few items require one resource over two consecutive periods is now used frequently.
and/or some of them are long runners, whereas
There are 3.9 single-item production periods on aver-
demand for the other items is rather low. Therefore, a
new data set is generated from Classes 7 to 9. In Class age for Class 7a in the solutions of SPL/ext/vi- B&C
7, Items 1-4 and 5-8 are aggregated to form two new (Class 8a: 3.7; Class 9a: 2.4). Compared to the origi-
nal data, the newly created test classes are more dif-
items, such that together with the unchanged Items
9 and 10, Class 7a now has four items. In Class 8a, ficult to solve with respect to observed gaps to LB.
Items 1-8 and 9-16 are aggregated to form a new The B&C approach outperforms the version with-
class with six items, whereas in Class 9a, Items 1-10, out valid inequalities significantly for all test classes.
11-20, 21-23, and 24-26 are aggregated, resulting in a Here again, the time-oriented decomposition heuris-
total of eight items. The aggregation is defined such tic (6/2/2, B, STMIN) shows its strength by providing
that the sum of demands (setup costs, setup times) is even better solutions than the B&C approach in much
shorter time. Results for version STEare also compet- the solution quality has improved considerably (sec-
itive, but as argued in ?5, sometimes infeasible solu- ond column, Table 8), it does not outperform the
tions (Class 7a: 20; Class 8a: 32; Class 9a: 32) occur. best decomposition heuristic with a time limit of
60 seconds.
6.3. Multilevel Test Instances
Last, to apply the model formulation in the multilevel 7. Conclusions
case, test instances from Stadtler (2003) for MLCLSP In this paper, a new model formulation for the
are used (called test set B+). The 60 test instances
CLSPL has been presented. Extensive computational
used here are those with general operations struc- tests show that this MIP model formulation yields
ture, and all comprise 10 products on 3 resources over
tight lower bounds as well as good primal solutions.
24 periods. Due to the problem size, the gap to LB
Herein, a branch-and-cut approach within a standard
of the complete model is rather large (Table 7). The MIP solver has provided the best results.
time limit was set to 600 seconds or the first solution As long as lot-sizing problems comprise only a
was taken (number in brackets). The time-oriented
single level with one resource, the proposed model
decomposition heuristic is able to close a considerable formulation, together with the time-oriented decom-
part of the gap to LB in less computational time. As position heuristic, has shown excellent solutions on
each lot-sizing window contains a more complex MIP a standard PC to problems of the size that arise in
problem than before, it is wise to reduce the size of the industrial practice. Moreover, it has been argued that
lot-sizing window or allow more computational time. the CLSPL should only be applied in the case where
Still, the time-oriented decomposition heuristic gener- one can expect that a single-item production (and a
ates presumably good solutions in reasonable time.
corresponding setup state) carryover will occur in an
For a subset of 10 arbitrarily chosen test instances, optimal solution. Otherwise, simpler models might
we computed better bounds by setting the compu- be more appropriate to solve the underlying decision
tational time to 12 hours and changing the stan- problem.
dard search strategy of XPRESS-MP to breadth search. The use of MIP models inside the lot-sizing win-
By applying this lower bound, gaps for the heuris- dows of the time-oriented decomposition heuristic
tic (4/2/2, B, STMIN) are reduced by approximately makes it easy to adjust the solution algorithm (i.e., the
one-quarter (from 31.6% to 22.6%) (Table 8). Further- MIP model formulation inside the lot-sizing window),
more, we have enlarged the time limit to 12 hours in case additional constraints apply.
for solving the complete model SPL/ext/vi- B&C by Some future research will be necessary to prove
the standard search strategy of XPRESS-MP.Although the solution quality of the proposed approach for
B&C
SPL/ext/vi- 0 time 0 time 0 time
Testset [max:12h] Gapto LB [max:60s] Gapto LB [max:180s] Gapto LB [max:60s]
B+ (10) 25.0% 26.0% 53.8 23.4% 136.8 22.6% 40.3
E
a=O•E
E Yias,
= 1 Vi= 1..., N. (A3)
s=bi +1 3=0
Optimal solution (sum of setup costs): 800
t=1
and
1 10O A 10 bi
E E
T+1 1
1 2 3 4 is added.
The model formulation by Sox and Gao (1999) will not find the and Methodsfor Production Planning. Springer, Berlin, Germany,
optimal solution (even for test instances without setup times), if in 127-146.
the optimal solution Karmarkar, U., L. Schrage. 1985. The deterministic dynamic prod-
(a) an arc (product j) with value "1" is required starting in uct cycling problem. Oper. Res. 33 326-345.
period t (index 1), Maes, J., J. O. McClain, L. N. Van Wassenhove. 1991. Multilevel
(b) an arc (product j) with value lower than "1" is required capacitated lotsizing complexity and LP-based heuristics. Eur.
starting in period t - 1 (index 0) ending in period t (index 1), and J. Oper. Res. 53 131-148.
(c) a setup for some other product k is required in period t - 1 Miller, A. J., G. L. Nemhauser, M. W. P. Savelsbergh. 2000. Solving
or earlier, which prohibits other arcs of product j ending in period multi-item capacitated lot-sizing problems with setup times by
t (Index 1), otherwise violating (21) of Sox and Gao (1999). branch-and-cut. CORE discussion paper, No. 00/39, Louvain-
Figure B shows an example. la-Neuve, Belgium.
Pochet, Y., L. A. Wolsey. 1995. Algorithms and reformulations for
lot sizing problems. W. J. Cook, L. Lovaisz, P. Seymour eds.
CombinatorialOptimization. DIMACS Series in Discrete Math-
References ematics and Theoretical Computer Science, Vol. 20. American
Dash Associates. 2000. XPRESS-MP. Release 12 Supplement. Blis- Mathematical Society, Providence, RI, 245-294.
worth, Northants, U.K. Rosling, K. 1986. Optimal lot-sizing for dynamic assembly systems.
Dillenberger, C., L. F. Escudero, A. Wollensak, W. Zhang. 1993. On S. Axsater, C. SchneeweiSg, E. Silver, eds. Multi-Stage Produc-
solving a large-scale resource allocation problem in produc- tion Planning and Inventory Control. Springer, Berlin, Germany,
tion planning. G. Fandel, T. Gulledge, A. Jones, eds. Operations 119-131.
Research in Production Planning and Control. Springer, Berlin, Salomon, M. 1991. Deterministic Lotsizing Modelsfor Production Plan-
Germany, 105-119. ning. Springer, Berlin, Germany.
Drexl, A., K. Haase. 1995. Proportional lotsizing and scheduling. Sox, C. R., Y Gao. 1999. The capacitated lot sizing problem with
Internat. J. Production Econom. 40 73-87. setup carry-over. lIE Trans. 31 173-181.
Eppen, G. D., R. K. Martin. 1987. Solving multi-item capacitated Stadtler, H. 1996. Mixed integer model formulations for dynamic
lot-sizing problems using variable redefinition. Oper. Res. 35 multi-item multi-level capacitated lotsizing. Eur. J. Oper. Res.
832-848. 94 561-581.
Fisher, M., K. Ramdas, Y.-S. Zheng. 2001. Ending inventory valua- _. 2000. Improved rolling schedules for the dynamic single-level
tion in multiperiod production scheduling. Management Sci. 47 lot-sizing problem. Management Sci. 46 318-326.
679-692. _. 2003. Multi-level lot-sizing with setup times and multiple
Fleischmann, B. 1990. The discrete lot-sizing and scheduling prob- constrained resources: Internally rolling schedules with lot-
lem. Eur. J. Oper. Res. 44 337-348. sizing windows. Oper. Res. 51 487-502.
Gopalakrishnan, M. 2000. A modified framework for modelling Tempelmeier, H., M. Derstroff. 1996. A Lagrangean-based heuristic
set-up carryover in the capacitated lotsizing problem. Internat. for dynamic multilevel multiitem constrained lotsizing with
J. Production Res. 38 3421-3424. setup times. Management Sci. 42 738-757.
-, D. M. Miller, C. P. Schmidt. 1995. A framework for modelling Thizy, J. M, L. N. Van Wassenhove. 1985. Lagrangean relaxation
setup carryover in the capacitated lotsizing problem. Internat. for the multi-item capacitated lot-sizing problem: A heuristic
J. Production Res. 33 1973-1988. implementation. IIE Trans. 17 308-313.
-, K. Ding, J.-M. Bourjolly, S. Mohan. 2001. A tabu-search heuris- Trigeiro, W. W., L. J. Thomas, J. O. McClain. 1989. Capacitated lot-
tic for the capacitated lot-sizing problem with set-up carryover. sizing with setup times. Management Sci. 35 353-366.
Management Sci. 47 851-863. Van Hoesel, C. P. M., A. P. M. Wagelmans, L. A. Wolsey. 1994.
Haase, K. 1994. Lotsizing and Scheduling for Production Planning. Polyhedral characterization of the economic lot-sizing problem
Springer, Berlin, Germany. with start-up costs. Siam I. Discrete Math. 7 141-151.
- . 1998. Capacitated lot-sizing with linked production quan- Wolsey, L. A. 1989. Uncapacitated lot-sizing problems with start-up
tities of adjacent periods. A. Drexl, A. Kimms., eds. Beyond costs. Oper. Res. 37 741-747.
Manufacturing Resource Planning (MRP II). Advanced Models _. 1998. Integer Programming. Wiley, New York.
Accepted by Thomas Liebling,former departmenteditor; receivedApril 30, 2002. This paper was with the authorsfor 1 revision.