24226.lotsizing and Scheduling For Production Planning (Lecture Notes in Economics and Mathematical Systems)

Lecture Notes in Economics
and Mathematical Systems

Founding Editors:
M. Beckmann
H. P. Kiinzi
Editorial Board:
H. Albach, M. Beckmann, O. Feichtinger, W. Hildenbrand, W. Krelle
H. P. Kiinzi, K. Ritter, U. Schittko, P. Schonfeld, R. Selten
Managing Editors:
Prof. Dr. O. Fandel
Fachbereich Wirtschaftswissenschaften
Femuniversitat Hagen
Feithstr. 140/AVZ II, D-58097 Hagen, FRO
Prof. Dr. W. Trockel
Institut fUr Mathematische Wirtschaftsforschung (IMW)
Universitat Bielefeld
Universitatsstr. 25, D-33615 Bielefeld, FRO
408
KnutHaase
Lotsizing and Scheduling

for Production Planning
Springer-Verlag
Berlin Heidelberg New York
London Paris Tokyo
Hong Kong Barcelona
Budapest
Author
Dr. Knut Haase
Christian-Albrechts-UniversWit zu Kiel
Institut flir Betriebswirtschaftslehre
OlshausenstraBe 40
D-24118 Kiel, FRG
ISBN-13: 978-3-540-57833-8
001: 10.1007/978-3-642-45735-7
e-ISBN-13: 978-3-642-45735-7
CIP data applied for

This work is subject to copyright. All rights are reserved, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, re-use
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other
way, and storage in data banks. Duplication of this publication or parts thereof is
permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from
Springer-Verlag. Violations are liable for prosecution under the German Copyright
Law.
Springer-Verlag Berlin Heidelberg 1994
Typesetting: Camera ready by author
SPIN: 10083822
42/3140-543210 - Printed on acid-free paper
Acknowledgements
This work could not have been finished without the help of many people. They have
commented on portions of the manuscript and have been available for discussions on the
covered topics. These comments and discussions have been very helpful for my own
understanding of the subject and have resulted in improvements of the manuscript. I wish to
express my special gratitude to some of them.
Prof. Dr. Andreas Drexl, deserves special thanks because he provided me the helpful
guidance for realizing this work. Prof. Dr. Bernhard Fleischmann provided some codes and
data sets. The colleagues of the Institute of Business Administration at the Christian-AlbrechtsUniversity in Kiel, especially Carsten Jordan, AlfKimms, Rainer Kolisch, and Arno Sprecher,
have made many useful suggestions. The discussions in the seminar of the Institute of Statistics
and Econometrics, especially with Prof. Dr. Gerd Hansen and Prof. Dr. Helmut Liitkepohl,
have been very helpful. Dr. habil. Gerold Alsmeyer has supported me with respect to the
theory of sequential analysis. Jorg Latteier performed some computations for the examples in
Chapter 4. Daniel Lincoln III read the entire manuscript to improve the English.
Finally, I wish to thank Hilke. As always, she has been a wonderfully supportive girl
friend, despite my late evenings in the office.
Kiel,1993
,Knut Haase
Contents
Chapter 1.
Introduction
1.1 Objectives of Lotsizing and Scheduling ..................................................................... 1

1.2 Classification of Problems ......................................................................................... 3
1.3 Outline of the Following Chapters............................................................................. 7
Chapter 2.
Single-Level Capacitated Lotsizing Problems
2.1 The Capacitated Lotsizing Problem (CLSP) .............................................................. 9

2.2 Critique on the CLSP ................................................................................................ 11
2.3 The CLSP with Linked Lotsizes (CLSPL) ................................................................ 18
Chapter 3.
Single-Level Lotsizing and Scheduling Problems
22
3.1 The Discrete Lotsizing and Scheduling Problem (DLSP) ........................................... 22

3.2 The Continuous Setup Lotsizing Problem (CSLP) .................................................... 24
3.3 A New Model- The Proportional Lotsizing and Scheduling Problem (PLSP) ............ 26
3.4 Model Comparison ................................................................................................... 28
Chapter 4.
Extensions of the PLSP
31
4.1 The PLSP with Setup Times ..................................................................................... 31

4.2 The PLSP with Sequence Dependent Setup Costs ..................................................... 35
4.3 The Multi-Machine PLSP with Make-or-Buy Decisions ............................................ 38
4.4 The PLSP with Backordering and Stockouts............................................................ .41
4.5 The Multi-Level PLSP with One Bottleneck ............................................................ .45
4.6 Concluding Remarks ................................................................................................. 48
VIII
Chapter 5.
Control of Stochastic Algorithms via Sequential Analysis
50
5.1 Hypotheses Test Problem for Reducing the Parameter Space .................................... 50
5.2 Sequential Tests of Hypotheses Concerning Quantiles ............................................... 54
5.3 Monte-Carlo Study for Sequential Tests ................................................................... 59

5.4 AnExample .............................................................................................................. 66
5.5 Concluding Remarks ................................................................................................. 68
Chapter 6.
A New Class of Stochastic Heuristics for the PLSP and the CLSPL
69
6.1 Biased Random Sampling via Randomized Regrets for the PLSP .............................. 69
6.2 Regrets for PLSP Extensions .................................................................................... 76
6.3 Modifications for the CLSPL .................................................................................... 78
6.4 Integration of Parameter Control via Sequential Analysis .......................................... 79
6.5 Relations to Local Search Methods ........................................................................... 83
Chapter 7.
Computational Results
86
7.1 An Instance Generator .............................................................................................. 87

7.2 Computational Results for the PLSP ......................................................................... 92
7.3 Computational Results for the CLSPL ...................................................................... 102
Chapter 8.
Summary and Future Work ........................................................................ 107
Bibliography .................................................................................................................. 108
Appendix
Monte-Carlo Study for Sequential Tests ..................................................... 116
Chapter 1.
Introduction
Billions of dollars are tied up in the inventories of manufacturing companies which cause large
(interest) costs. A small decrease of the inventory and/or production costs without reduction of
the service level can increase the profit substantially. Especially in the case of scarce capacity,
efficient production schedules are fundamental for short delivery time and on-time delivery
which are important competitive priorities. To support decision makers by improving their
manufacturing resource planning system with appropriate methods is one of the most
interesting challenges of production planning.
The following chapters contain new models and new solution strategies which may be
helpful for decision makers and for further research in the areas of production planning and
operations research.
The main subject is on lotsizing and scheduling. The objectives and further characteristics
of such problems can be inferred from practical need. Thus, before an outline is given, we
consider the general objectives of lotsizing and scheduling and classify the most important
characteristics of such problems in the following sections.
1.1 Objectives of Lotsizing and Scheduling
Let us consider a machine (or more generally a production facility) on which a number of
different items are producible. To produce an item requires that the machine is setup for the
item. If the machine is setup for the item a setup activity is necessary. The activity to change
the setup state is called changeover, while equal items produced one after the other without
changeover form a lot. The number of items in a lot is called lotsize. Lotsizing and scheduling
is the activity to obtain simultaneously in which period, in which number (i.e. lotsize), and in
which sequence (Le. schedule) different items should be produced such that the production
plan is feasible (e.g. no shortages) and results in minimal costs. Usually, holding costs and
setup costs are the relevant costs which have to be minimized. The number of setups depends
0., the lotsize; the greater the lotsize the fewer the number of setups and therefore the smaller
the setup costs. However, this causes higher inventory costs; the greater the lotsize the greater
the inventory in storage and therefore the higher the inventory costs. Consequently regarding
to the lotsize there exist two cost components which have to balanced such that the sum is
minimized. This relation is sketched out in Figure 1.1.
costs per unit in lot
total costs
holding costs
setup costs
lotsize
Fig 1.1 Relationships between lotsize and costs
It is not a general result that the total costs are minimized when the setup costs and the
holding costs are equal. The result depends on the considered problem. In the production
planning literature different lotsizing and scheduling problems with specific solution methods
are introduced. A problem can be characterized by different criteria which we consider in the
next section.
1.2 Classification of Problems
Depending on the potential application of a lotsizing and scheduling model, the relative
importance of problem characteristics will vary. Some of the most important criteria are
described below.
1) Degree of information
In deterministic models each parameter has a fixed known value. Stochastic models
contain parameter(s) which are random variables. Typical stochastic parameters are the
external demand (time and/or quantity), lead time, the quantity of defective items, and
machine breakdowns [cf. Anderson 1989].
2) Temporal development of parameters

The parameters in static models are constant over time (time invariant). In dynamic
models some parameters vary over time. Such parameters may be demand, available
capacity (e.g. caused by holidays or extra shifts), production costs (e.g. caused by
overtime costs per hour) and setup times (learning in setups).
3) Horizon
The planning horizon may be assumed to be infinite, finite, or variable. A medium-range
(short-range) phmning problem has a time horizon that is typically months (weeks) away.
4) Time scale
Planning may be done in small (hours, shifts or days), or large (weeks, months) discrete
periods or over a continuous time scale.
5) Number of items
If there is no (relevant) interdependency between items, either because of joint utilization
of capacity or parent-component relationships, independent single-item problems occur,
and otherwise we face a multi-item problem. Often, the term product is used instead of
item.
6) Number oflevels
If there is no parent-component relationship between items a single-level problem arises,
and otherwise we have a multi-level problem. The parent-component relationships are
defined by the item structure.
Level
Series
Assembly
General
Final
Item:
2
3
4
Fig. 1.2 Item structures
In Figure 1.2 typical item structures are depicted [cf Billington et aI. 1983]
- series, i.e. the produced item proceeds through a series of production levels.
- assembly, i.e. the final item is made by a multi-level process, where each item (out of
the final one) has exactly one successor.
- general, i.e. the final item is made by a multi-level process, where either items are
used in more than one successor and items consist of more than one predecessor.
Often, the word stage is used instead oflevel.
7) Relevant costs
There are a number of cost categories that may be relevant to lotsizing and scheduling.
Following [Salomon 1991], these include:
i)
Setup related costs: These are the costs incurred each time a setup takes place for an
item on a facility (e.g. machine). In many lotsizing and scheduling models it is assumed
that for each lot separable item specific setup costs are incurred. If the setup costs
depend on the sequence in which the items are scheduled a problem with sequence
dependent setup costs arises. Sometimes startup and setup costs are distinguished.
In these cases the startup costs are incurred to setup the production facility for a
specific item type (e.g. smelt iron to produce line of cylinder) and the setup costs arise
for switching on the facility for a specific item (e.g. to found lO-inch-cylinder). A
synonym for "setup costs" is "changeover costs". Reservation costs can be charged
for keeping up the setup state whether or not the facility is used for production [cf
e.g. Karmarkar et al. 1987; Sandbothe 1991].
ii) Inventory related costs: These include holding costs which are the costs incurred for
holding an item in inventory during one or more periods, and consist mainly of the
costs of capital tied up in inventory, potential spoilage or obsolescence, taxes,
insurance and warehouse operation costs [cf Silver 1981]. A synonym for "holding
costs" is "carrying costs". Further inventory related costs are shortage costs
(backorder costs, backlog costs, loss of sales).
iii) Capacity related costs: These are the costs incurred for using regular capacity and/or
extra capacity (overtime costs, cost due to subcontracting or changing the workforce
level).
8) Resource constraints
Disregarding resource constraints (e.g. the machine capacity, the number of workers, the
storage capacity, the transportation capacity, the budget limitations) leads to an
uncapacitated problem, otherwise to a capacitated one (with single or multiple resource
constraints). A single resource constraint in the multi-level case is often called a
"bottleneck". If more than one machine is considered a multi-machine problem arises,

otherwise a single-machine problem.
9) Supply process
Sometimes batch-production is considered such that items arrive in inventory in one
batch no earlier than completion of the whole lot. Usually, it is assumed that an ordered
item is received after a lead time. Especially, when the lead time is a random variable a
complex planning problem arises. Furthermore, it may happen that only a (random) portion
(e.g. strike, defective items) ofthe number of ordered items is received.
10) Service policy

The service policy includes the decision about the service level. A service level of one (less
than one) disallows (allows) a shortage. For avoiding customer service problems and the
hidden costs of unavailable items a safety stock inventory can be used to protect against
uncertainties in demand, lead time, and supply. Subcontracting (make-or-buy decisions),
backlogging (accumulation of orders that have been promised for delivery at some future
date), backorders (filling orders with delay), and stockouts (loss-of-sales) are possibilities
to cope with shortages [cf Krajewski, Ritzman 1990, p. 415 and p. 488] .
11) Consideration of time consuming activities

These include transportation time, lead time, setup time, processing time per unit,
production speed. Infinite production speed (or equivalently zero production time per
unit) are assumed in models where capacity limits are ignored. In most cases
transportation, production and/or setup are assumed to be instantaneous. Other situations
are modelled sufficiently by assuming that production speeds are constant over time. Setup
times can be sequence dependent. Furthermore, two types of setup times can be
distinguished when items belong to different families: Minor setup times (major setup
times) are required when changing items that belong to the same (different) family(ies) [cf
Potts, Van Wassenhove 1992]. Ifleaming effects are included processing times per unit or
setup times decrease over time. Usually, the lead time is unaltered by the number of
ordered items.
12) Objectives
Usually, the total costs have to be minimized. Further possible objectives can be
maximization of the service level or smoothing the production load [cf. Domschke et al.
1993, pp. 63-69].
The focus here is on lotsizing and scheduling problems with finite capacity where
deterministic, dynamic demand for multiple items occurs in discrete time periods until the finite
planning horizon.
1.3 Outline of the Following Chapters
In the following chapters multi-item, dynamic, capacitated lotsizing and scheduling problems
will be addressed.
Chapter 2 considers the well-known capacitated lotsizing problem, denoted by CLSP.

Instances are explored where it is advisable to modify the model. Especially the time scale, the
valuation of the setup costs and difficulties associated with the multi-level extension are
drawbacks of the CLSP. To overcome these shortcomings a new model, denoted by CLSPL, is
introduced, in which the lotsizes of adjacent periods can be linked to save setup costs.
Chapter 3 considers two types of well-known lot sizing and scheduling models, the discrete
lotsizing and scheduling problem (DLSP) and the continuous setup lotsizing problem (CSLP).
They are based on different fundamental assumptions. To provide a more efficient capacity
utilization a new model type, the proportional lotsizing and scheduling problem (PLSP) is
presented. We finish this chapter with an example and a discussion of the relationship between
the models.
Chapter 4 presents some extensions of the PLSP for other potential applications. We consider
setup times, sequence dependent setup costs, the integration of make-or-buy decisions in a
multi-machine environment, the integration to plan for delivery-times, and a multi-level facility
with one bottleneck. This chapter finishes with some remarks on a general PLSP in a material
requirements planning system.
Chapter 5 provides a Monte-Carlo study of so-called sequential tests. A fictitious stochastic

algorithm is given whose quality depends on the choice of two parameter values. A sequential
test can be used to improve the quality of the stochastic algorithm.
Chapter 6 introduces a new class of stochastic heuristics applicable for the PLSP and, slightly
modified, for the CLSPL. A so-called randomized regret-measure is introduced which is used
to perform sequence and lotsize decisions. The regret-measure depends on two parameters.
The choice of parameter values can be controlled by a sequential test. The regret-measure is
modified for PLSP extensions. The relations to other local search methods are discussed.
Chapter 7 first presents an instance generator. A computational study is carried out to analyze
the performance of the heuristic for the PLSP and for the CLSPL, respectively. The effects of
the integration of a sequential test to guide the stochastic algorithm is analyzed.
Chapter 8 is dedicated to finish this manuscript with a summary and some remarks on future
research.
Chapter 2.
Single-Level Capacitated Lotsizing Problems
Whereas uncapacitated lotsizing problems can be solved efficiently to optimality [cf.

Federgruen, Tzur 1991], solving capacitated lotsizing problems is still a complex challenge in
the area of production planning. In this chapter a well-known capacitated lotsizing problem is
discussed. Some shortcomings of the corresponding model formulation force to consider a
modification.
2.1 The Capacitated Lotsizing Problem (CLSP)
We use the classification scheme of Section 1.2 to characterize the so-called capacitated
lotsizing problem, denoted by CLSP: The parameters in the CLSP are known (deterministic
information degree). The time until the finite planning horizon is divided into T (large) discrete
periods t=I, ... ,T. The different items j=I, ... ,J with a single-level item structure are to be
produced on a single machine (single resource constraint) available with Ct capacity units in
period t. The demand for item j in period t, djt, has to be satisfied without delay (shortages are
disallowed). Producing one unit of item j absorbs Pj capacity units (finite production speed).
To setup the machine for item j causes Sj setup costs. Inventory costs of hj (holding costs
coefficient) are incurred to hold one unit ofitemj one period in inventory. The objective is to
minimize the sum of setup and holding costs.
The CLSP is based on the following fundamental assumption:
AO: Setup costs occur for each lot in a period.
We define
Ijt
the inventory of item j at the end of period t,
qjt
the quantity (lotsize) of item j to be produced in period t, and
10
Xjt
a binary variable indicating whether setup occurs for itemj in period t ('1t=l) or
not (Xjt=O).
Thus, mathematical1y the CLSP can be stated as fol1ows:

I
(2.1)
mm LL(s)X)t +h)Iit)
)=l t=l
subject to
= d)t
O=I, ... ,J; t=I, ... ,T)
(2.2)
(t=I, ... ,T)
(2.3)
O=I, ... ,J; t=I, ... ,T)
(2.4)
I)t,qit ~ 0
0=1, .... ,1; t=I, ... ,T)
(2.5)
E{O,I}
0=1, ... ,1; t=I, ... ,T)
(2.6)
I i.t-l +qit - lit

I
LP)q)t ~Ct
)=l
B,x it -qit
x)t
where B
~O
(~max{CJ"",CT}/min{pl> ... ,PJ})
is a large number, and w.1.o.g. the inventory in
period t=O is Ijo=O 0=1, ... ,1).
The objective to minimize the total costs, i.e. the sum of setup and holding costs, is
expressed by (2.1) whereby only the inventory, obtained by the equations (2.2), at the end of a
period is valued. Constraints (2.3) ensure that the total production in each period does not
exceed the capacity. Constraints (2.4) force for each lot (greater than 0) in a period the binary
setup variable to be one which increases the sum of setup costs. The suitable domains of the
variables are determined by the restrictions (2.5) and (2.6). The non-negativity condition of the
inventory ensures that a shortage does not occur.
Numerous authors have considered the CLSP and its extensions. For a comprehensive
survey of problems and solution techniques we refer to [Bahl et a1. 1987; De Bodt et a1. 1984;
Domschke et a1. 1993; Maes, Van Wassenhove 1988; Salomon 1991; Tempelmeier 1992].
Recently a Lagrangean relaxation-based heuristic procedure to generate near-optimal solutions
11
to the CLSP with setup times and limited overtime is introduced in [Diaby et al. 1992]. An
optimal solution procedure for a variant of the CLSP is introduced in [Madan, Gilbert 1992].
The given algorithm has solved problems with up to 9 products and 36 periods within 30
minutes on an mM 4381 system.
2.2 Critique on the CLSP
With respect to lotsizing (and scheduling) there are some shortcomings of the CLSP model and
related solutions. We will therefore have a closer look at
- the basic assumption (AO),
- the (large) time scale (weeks or months) and
- the problems associated with extensions to the multi-level, multi-machine case.
Basic assumption
Exploring the assumption that "setup costs occur for each lot in a period" we consider the
following example.
Example 2.1: Let J=2, T=3, SI=s2=100, and h 1=h2=1. For each period t the demand djt
of the itemj and the capacity Ct are given below (missing entries are zero):
d lt
d2t
Ct
3
6
4
10
10
10
The optimal solution of the CLSP with total costs Z*=300 is determined by equalizing
lotsizes and demands, i.e. qjt=~it 0=1, ... ,J; t=l, ... ,T). Let us now consider the following
solution:
12
qlt
5u
q2t
6
4
where "u" denotes the linking of the production quantities of adjacent periods, i.e. we
start in the second period with the item j= 1. Hence, it is not necessary to change the
setup state of the machine because it was already prepared for item j=1 at the end of
the period t=1. The total costs are reduced to SI+S2+6hl=2 100+6 1=206.
The reason for the poor solution of the CLSP is, because of the assumption AO, that
sequencing aspects between adjacent periods are disregarded which causes avoidable costs of
substantial amount.
Timescale
Usually, in short-term planning problems we are interested in a schedule with small time
periods as days or shifts. The CLSP is performed for large periods as weeks or months. One
idea could be to solve a CLSP first, and afterwards compute a (disaggregate) schedule for
small time periods based on the CLSP solution. But this causes some serious problems as
shown in Example 2.2.
Example 2.2: Let J=I, T=4, s=o and h>O. Thus the "optimal" solution is determined by
producing as late as possible. For every small period t (=1, ... ,4) the demand of the item,
dt, the available capacity, Ct , and the optimal lotsizes, q;, are given below (subscript
j=1 is omitted for ease of notation):
13
10
15
10
10
10
10
10
Now, we double the period length such that we aggregate two periods to one. A
suggested result of the aggregation may be as follows
l'
5
20
20
20
20
But the solution is not feasible because the demand d3=15 cannot be satisfied (cf. the
disaggregate data) without delay. In the first half of period 1'=(3,4), i.e. in t=3, only 10
units of the item are producible so that a shortage of S units occurs. To avoid such an
infeasible solution, we have to aggregate the demand as follows:
l'
20
We have to shift d3=IS to period 1'=(1,2). The "optimal" solution based on the
aggregate data is (determined by equalizing demand and production quantities, i.e.
q;. =dr (1'=(1,2),(3,4 is more expensive than the optimal solution determined by using
disaggregate data, since avoidable holding costs of IS are incurred for the inventory at
the end of period 1'=(1,2) or t=2, respectively.
One may suggest that the CLSP can obtain a solution directly (i.e. without an aggregation
in the time scale). In this case the disadvantage of the basic assumption AO becomes more
14
important. Especially with demand exceeding the period capacity AO is not very appropriate.
Thus, the CLSP has substantial drawbacks for short term planning problems with small
periods.
Problems associated with extensions to the multi-level, multi-machine case

Extending the CLSP to the multi-level case, we have to take into account that capacity
constraints can occur at different levels. An item is producible when there is a sufficient
number of available predecessors (cf Figure l.2). How these requirements should be included
in a multi-level extension is considered in the following example:
Example 2.3: Let J=4, T=2, M(o=20 (~U~bJer of machines), Jl={1,2}, J2={3,4}, M1=
M2={ I}, M3=M4={2}, (ajJ =
~ ~ ~ ~
where J m is the set of items producible on
1 100
the machine m (=1,2), Mj the set of machines capable to produce itemj, and!lji is the
number of units of item j required to produce one unit of item i. The quantity of item j
to be produced on machine m in period t is denoted as 'ljmt. Figure 2.1 illustrates the
parent-component relationships and the machine required by an item.
r-
Fig. 2.1 General item structure with two machines
Only for the final items (independent) demand occurs. Furthermore, the processing
time, Pjm, i.e. the time to produce one unit of item j on machine m, is undefined if the
item j does not require the machine m. The demand of the item j in period t, djt, the
15
capacity of the machine m in period t, denoted by C mt, as well as the processing time of
item j on machine In, Pjm, are given below:
d
2
.5
.5
If we only consider the capacity restrictions
LPjmqjmt ::;;C mt
(t=1,2;m=1,2)
jelrn
the "solution" illustrated in Figure 2.2 would be possible:
[] production of item j
m=l
1
m=2
2
4
,
2
Fig. 2.2 Schedule ofthe machines m=l, 2
The "solution" is infeasible: Itemj=l cannot be produced at the beginning of the second
period, since item j=4 is not available. (Note item j=3 would be sufficiently available if
16
we assume that infinitesimal quantities of item j=3 can be passed from machine m=2 to
machine m=I and the transportation time hereto needed is zero).
To avoid such an infeasible solution it is necessary to take into account item specific lead
time of at least one period at each stage in the multi-level structure as in the following
inventory balance equation (remember that we do not consider the schedule within the period,
otherwise shorter lead time would be possible)
11t-1+ Lqjmt meMj
L Lajjqjm.l+1 - Ijt =d jt
O=I, ... ,J; t=I, ... ,T)
ieNjmeMj
where Nj denotes the set of immediate successors of item j, so that
L L ajjqjm.t+1
stands for
iENjmeM j
the dependent demand of item j in period t. Typically, the CLSP has large discrete time
periods; so making a shift of one period at each stage, the total lead time of a final item will be
of a substantial size. Thus models with small discrete time periods seem to be preferable for
multi-level production.
II]
m=l
production of item j
t'
m=2
Qj2t'
t'
Fig. 2.3 Feasible solution with small discrete periods
An additional advantage of a small discrete time scale is the larger set of feasible solutions;
e.g. if we take a time scale with four small periods (t'=1, ... ,4) instead of two large periods
(t=I,2) a solution as illustrated in Figure 2.3 can be provided for the problem in Example 2.3.
17
In the case with two large periods, we have to start with itemj=3 (or j=4) at the beginning of
period t=1 and not one small period l' later as depicted in Figure 2.3. Note that some units of
the item j=4 are passed to the machine m= 1 before finishing the entire lot of the item j=4 which
is called overlapping. (The situation in Figure 2.2 can be seen as a special case of overlapping
between item j= 1 and item j=3.) If overlapping of lots is not allowed the problems associated
with multi-level extensions of the CLSP will not change as the following Example 2.4 shows.
[]
qj1t
:-----l
I 1 m= 1 i
I
m=1
1
r"
qj2t
II
I
I
m=2
1
b)
a)
qj1t
qj1t'
m=1
m=1
3 4
1 2
m=2
qj2t
t'
m=2
qj2t'
2
1
c)
t'
d)
Fig 2.4 Assembly item structure and production schedules with different time scales
18
Example 2.4: Le(~
J=g, ~=J2' M=2, J1={1}, J2={2,3}, Crnt=2 (m=1,2; t=1,2), d2t=d3t=0
(t=1,2), (ajJ = 1 0 0, Pjrn=1 (m=1,2, jEJrn) where the symbols are defined as in
100
Example 2.3. Overlapping ofJots is not allowed. Only in period t=2 a demand for item
j=1 of one unit occurs, i.e. d 12=1. The item structure is depicted in Fig. 2.4 a). The
schedule of Fig. 2.4 b) is not feasible due to the overlapping of the lots Q12, Q32. In Fig.
2.4 c) (Fig. 2.4 d a feasible large (small) time scale solution is provided where for the
itemsj=2 andj=3 a specific lead time of one period is considered (to avoid an infeasible
solution). The small time scale solution will be less costly if positive holding costs of
items j=2 and j=3 are considered.
The disadvantage of a multi-level CLSP can be summarized as follows: A multi-level

CLSP either has an "optimal" solution with long lead time or, if a shift at each stage is not
considered; a "solution" with short lead time which is possibly infeasible.
2.3 The CLSP with Linked Lotsizes (CLSPL)
One drawback of the (single-level) CLSP is the basic assumption (AO) ignoring the setup state
at the beginning of a period. With regard to this aspect a modified CLSP with linked lotsizes,
denoted by CLSPL, can mathematically be stated as follows:
J
(2.7)
min LL[Sj(X jt -zJt)+hjI Jt ]

j=1 t=1
subject to
I~'.1 +qj' -Ijt =d j,
0=1, ... ,1; t=1, ... ,T)
(2.8)
(t=I, ... ,T)
(2.9)
0=1, ... ,1; t=I, ... ,T)
(2.10)
LPjQj':::::C,
j=1
19
J
~>jt~1
(t=I, ... ,T)
(2.11)
Zjt - Xj.t_1 ~ 0
O=I, ... ,J; t=I, ... ,T)
(2.12)
Zit-Xjt~O
0=1, ... ,1; t=I, ... ,T)
(2.13)
l-LXjt+J'Vt~O
i=\
(t=I, ... ,T)
(2.14)
Zjt +Zj.t-1 +V t_\ ~ 2
0=1, ... ,1; t=I, ... ,T)
(2.15)
v t ~O
(t=I, ... ,T)
(2.16)
Ijl'qjt ~ 0
0=1, ... ,1; t=I, ... ,T)
(2.17)
xjt>zJt E{O,I}
O=I, ... ,J; t=l, ... ,T)
(2.18)
i=\
where
Vt =
a variable indicating whether more than one item
IS
produced in period t
(v t >0) ornot{v t =0),

Xjt =
a binary variable indicating whether a quantity is produced for item j in period t (xjt=l)
or not (Xjt=O),
Zjt =
a binary variable indicating whether the quantities of item j in period t-l and period t
are linked (zjt=l) or not (Zjt=O),
and Ijo=O, vo=O, Xjo=O, ZjO=Zjl=O O=I, ... ,J), and the other parameters and variables are defined
as in the CLSP.
Now, the objective function (2.7) calculates only setup costs for an item j in a period t
when not linking with qj,t-l (i.e. the production quantity of the preceding period of the item) is
performed. The restrictions (2.11) to (2.13) secure that only one product can be produced at
the end of a considered period and produced further (linking) in the following period. If more
than one item is produced in a period t (Vt>O) it is necessary to avoid that the same quantity
will be considered as produced at the beginning and at the end of a period t; i.e. the quantity
can either be linked with a quantity of the preceding period t-l or with a quantity of the
20
succeeding period t+1 (see (2.14) and (2.15, only. If nothing is produced in a period, i.e. the
quantity is zero, it is possible to keep up the setup state.
Example 2.5: Let J=3, T=5, Ct=100 (t=1, ... ,5) and the other data as provided in Table
2.1.
Table 2.1 Data ofExam)2le 2.5

t
d lt
2
30
Pj
80
40
d2t
30
70
d3t
40
60
hj
Sj
400
150
100
Optimal solutions for the CLSP as well as the CLSPL are entered in Table 2.2.
Table 2.2 O)2timal solutions of the CLSP and the CLSPL

t
qlt
CLSP
CLSPL
q2t
30
931
40
30u
q2t
30
931
40
90
30
qlt
Z*
30
70
2070
70
1460
60
80u
40
30u
30
We see that the optimal objective function value Z~sp is 41.78% larger than Z~SPL .
The sets of feasible solutions of both models are identical whereas an optimal solution of
the CLSP can be more costly, so
Z~sp ~ Z;LSPL .
21
Since the CLSPL contains more binary variables than the CLSP it may be more difficult to
determine an optimal solution in reasonable time. However, in practical applications fast
heuristics are preferred which are based on the CLSP (e.g. the heuristic in [Dixon, Silver 1981]
or in [GUnther 1987]). Thus it is an interesting task to extend existing heuristics such that they
are applicable to the CLSPL.
Recently a similar model for the multi-machine case with setup times has been presented in
[Dillenberger et al. 1992]. Linking of production quantities of adjacent periods is considered in
[Haase 1993]; a heuristic is presented which is compared with the Dixon-Silver heuristic.
Note, due to the "correct" calculation of the setup costs by linking lots of adjacent periods
the CLSPL is applicable for problems with large and small discrete periods. Moreover, it is
appropriate for multi-level extensions.
Chapter 3.
Single-Level Lotsizing and Scheduling Problems
A solution of the problems considered in the preceding chapter does not yield the sequence of
the lots. Especiai1y in the short term planning the sequence of the production quantities is
required. This chapter deals with models where the size and the sequence of lots are
determined simultaneously. These include the following (well-known) models for single-level,
multi-item, dynamic, capacitated lotsizing and scheduling:
- the discrete lotsizing and scheduling problem (DLSP)
- the continuous setup lotsizing problem (CSLP)
which are considered in Sections 3.1 and 3.2, respectively. Furthermore, in Section 3.3 we
introduce a new type of model. Although these types of models are based on different
fundamental assumptions, each of them is applicable for short-term scheduling with periods as
days or shifts. In the last section of this chapter we compare the three models and stress the
relevance of the new model.
3.1 The Discrete Lotsizing and Scheduling Problem (DLSP)
The discrete lotsizing and scheduling problem (DLSP) has been suggested in [Fleischmann
1990] and is based on the following assumption:
AI: The production process always runs full periods without changeover.
This "all or nothing production" implies that at most one item can be produced in a fixed
number per period using all capacity. The capacity is assumed to be constant over time. Setup
costs have to be paid at the beginning of the periods only. A lot is produced in a batch of
immediately adjacent production periods. Thus the lotsizes are multiples of a full period
production.
23
Mathematically the DLSP can be stated as follows:
J
(3.1)
min LL(sjXj, +hh)

j=1 1=1
subject to
_
J. LJr; ~l:
Ij.,.1 +PjYjl -I j, =d j,
~f"\
J J)
-1\t -.O\l
._.
0-1, ... ,1, t-l, ... ,T)
(3.2)
(t=l, ... ,T)
(3.3)
LYj,:S; 1
j=1
Xjl -Yi+Yj.,-1 ~o
0=1, ... ,!; t=I, ... ,T)
(3.4)
Ij"x j, ~ 0
O=I, ... ,J; t=I, ... ,T)
(3.5)
0=1, ... ,1; t=I, ... ,T)
(3.6)
where
Pj
quantity of item j which is producible in one period (=pFt),
Yjt.
a binary variable indicating whether item j is produced in period t (Yjt= 1) or not (Yjt=O),
Xjt
a variable indicating whether setup occurs for itemj in period t (xjt~l) or not (Xjt=O),
and w.1.o.g. Ijo=O, YjO=O 0=1, ... ,1), and the other symbols are the same as in the CLSP. Note,
due to the minimization objective Xjt will be from {O, 1}.
There are some differences between the model formulation of the CLSP and the DLSP.
The inventory input of an item is either zero or the fixed quantity PJ' That one item per period
is produced only is guaranteed by (3.3). Setup costs occur at the beginning of a batch due to
inequalities (3.4). This implies, if the machine is idle between two batches of the same item
setup costs are incurred twice. This is to say, that the DLSP does not "look back" for more
than one period. (It is possible to keep up the setup state by additional (unnecessary)
production such that no idle periods occur between two batches of the same item. Note,
"unnecessary production" is not an inconsistency, because there exists, if h/O, always a
solution without unnecessary production which is less costly.)
24
To solve a CLSP as a DLSP the periods of the CLSP are be considered as macro-periods
which are divided in sub-periods, i.e. in "DLSP periods"; e.g. a week can be divided in ten
shifts. Furthermore, variability in the capacity has been taken into account by varying the
number of sub-periods in a macro-period [cf. Fleischmann 1990].
We will now discuss briefly some references related to the DLSP: The first mixed binary
programming models based on assumption Al have been developed some decades ago [cf.
Dinkelbach 1964; Haehling von Lanzenauer 1970], but there was no method for solving
medium- to large-size problem instances. The DLSP is a special case of the multi-machine
problem formulated in [Lasdon, Terjung 1971; Eppen, Martin 1987]. They consider solution
methods which are practicable only for problems where the number of periods is substantially
smaller than the number of products. A dynamic programming algorithm for a special case of
the DLSP is presented in [Gascon, Leachman 1988]. In [Salomon et al. 1991] a problem
classification for the DLSP is introduced and results on computational complexity are derived
for a number of single and parallel (more than one machine is capable to produce a certain
item) machine variants.
An exact branch-and-bound approach based on Lagrangean relaxation of the capacity
constraints has been presented in [Fleischmann 1990]. The resulting subproblems (which are
separately studied in [Wolsey 1989]) have been solved by dynamic programming. This
approach is capable of solving problems with (e.g.) J=12, T=122 or J=3, T=250 to optimality
on a (slow) personal computer within a reasonable amount of time.
3.2 The Continuous Setup Lotsizing Problem (CSLP)
The CSLP is based on the following assumption:
A2: At most one item can be produced per period.
25
This implies that there is either a setup at the beginning of a period or a setup occurs not
at all (within the period). Moreover, lotsizes are continuous quantities between zero and the
maximum possible production of item j in period t.
Mathematically the CSLP can be stated as follows:

J
(3.7)
min LL(sjXjt +hh)

J=! t=!
subject to
I jt _! +qjt -Ijt
J
=d jt
LYjt~1
0=1, ... ,1; t=I, ... ,T)
(3.8)
(t=I, ... ,T)
(3.9)
O=I, ... ,J; t=I, ... ,T)
(3.10)
O=I, ... ,J; t=I, ... ,T)
(3.11)
0=1, ... ,1; t=I, ... ,T)
(3.12)
0=1, ... ,1; t=I, ... ,T)
(3.13)
j=!
where
Yjt
= a binary variable indicating whether the machine is setup for item j in period t (Yjt= I) or
not (Yjt=O),
and w.l.o.g. IjO=O, YjO=O 0=1, ... ,1), and the other symbols are the same as in the CLSP.
The CSLP differs from the DLSP in the restrictions (3.8) and (3.11). (3.8) is the
inventory balance equation (cf (2.2. (3.11) defines an upper bound for the production
quantity of an item. The setup state, expressed by Yjt, can be kept up (memorized) if the
machine is idle. Thus no setup occurs between two batches of the same item if no other item
has been produced meanwhile.
26
The single-item CSLP has been studied in [Bitran, Matsuo 1986; Karmarkar et at. 1987].
In [Karmarkar, Schrage 1985] a branch-and-bound approach based on Lagrangean relaxation
of the capacity constraints (3. 11 ) has been presented, along with rather discouraging
computational experiences even for small-sized problems. MPSARX, an MPS-system with
automatic cut generation capabilities [cf Pochet, Wolsey 1991; Van Roy, Wolsey 1987],
allows to solve medium-sized problems to optimality.
Recently, a parallel machine CSLP has been studied in [de Matta, Guignard 1989]. In
IMagnanti, Vachani 1990] a cutting plane algorithm is presented for solving the Changeover
Cost Scheduling Problem (CSP), which is closely related to the CSLP, too.
3.3 A New Model - The Proportional Lotsizing and Scheduling Problem (PLSP)
The DLSP does not track of the setup state over idle periods. The CSLP wastes capacity by
not making fully use of the capacity while producing lots of items. Now we introduce a new
type of model which produces continuous lot sizes over one or several, adjacent or nonadjacent periods where the setup costs are "correctly" calculated by keeping track of the setup
state during idle time. Moreover, we allow one changeover within each period. Thus two
items (the first and second item) can be produced per period. Hence, if the first item does not
fully use the capacity of a period, the remaining capacity can be used by the second item. The
PLSP is based on the following assumption:
A3: At most one changeover is allowed within each period.
The four cases that may occur are: No setup in a period, setup at the beginning of a
period, setup during a period or setup at the end of a period. (Note, the case that more than
one changeover may be allowed within a period of a certain length can be reached by shorten
the period length.)
27
Mathematically the PLSP can be stated as follows:

J
(3.14)
min LL(sjX jt +hJl jt )

J~J
t~J
subject to
Ijt-l + qjt - ~t = djt
O=I, ... ,J; t=I, ... ,T)
(3.15)
LYjt:O;1
(t=I, ... ,T)
(3.16)
0=1, ... ); t=I, ... ,T)
(3.17)
0=1, ... ); t=I, ... ,T)
(3.18)
(t=I, ... ,T)
(3.19)
Ijt, qjt, Xjt ~ 0
0=1, ... ); t=I, ... ,T)
(3.20)
Yjt E{O,I}
O=I, ... ,J; t=I, ... ,T)
(3.21)
'f
j~J
Xjt - Yjt + Yj,t-l ~ 0

~Yjt
+ BYj,t-l - qjt ~ 0
LP8jt sCt
j~J
where
Yjt
= a binary variable indicating whether the machine is setup for item j at the end of period t
(Yjt=l) or not (Yjt=O),
and w.1.o.g. Ijo=O, YjO=O
0=1, ... ), and the other symbols are the same as in the CLSP.
Due to the term "BYj,t_l" in (3.18) it is possible to produce two items per period; except
for the first period if the machine is not setup for a specific item. If two items are produced per
period t then the first item is that item which is also produced (at least) at the end of the
preceding period t-l or for which the machine is setup at the end of the preceding period t-1.
The possible splitting of the machine capacity for the production of two items within one
period proportional to the quantities (Pjqjt) needed motivates the name of the model.
28
3.4 Model Comparison
Some important characteristics of the three lotsizing and scheduling models are summarized in
Table 3.1.
Table 3.1 Characteristics of the lotsizing and scheduling models

type of
max. no. of
model
different items setup state during period

per period
keeping up the
setup in a
capacity
constant
idle periods
DLSP
no
CSLP
yes
PLSP
lotsize
yes
at the
multiples of a full
beginning
period production
at the
continuous
time-
beginning
quantities
varying
at any point
continuous
time-
of time
guantities
varying
As an example we take the same data as in Example 2.5 (cf. Table 2.1) where we
duplicate the number of periods and cut in halves the capacities and the holding cost
coefficients.
Example 3.1: Let J=3, T=lO, C t=50 for all t, and the demand, the holding and the
setup costs as provided in Table 3.2.
Table 3.2 Data of example 3.1
2
30
80
40
400
30
70
1.5
150
40
60
Table 3.3 provides the optimal solutions of the models.
100
29
Table 3.3 DLSP, CSLP, and PLSP solutions of the examEle 3.1
gil
DLSP
50
50
50
30
30
50
40
30
20
50
1910
50
50
931
30
gil
50
50
20
30
g2t
20
z*
50 2140
50
g2t
931
10
50
gil
PLSP
50
g2t
931
CSLP
40
30
1710
10
20
50
Figure 3.1 illustrates the schedules and the capacity utilization resulting from the different
basic assumptions.
C fp.
t
DLSP
nl
C fp .
I
A
50
:\
10
\0
(i
I
7
10
>
>
CSLP
:'\
/ Pj
A
50
>1
Fig. 3.1 Gantt charts for the machine representing the schedules in Example 3.1
30
The solutions of the DLSP and CSLP are rather poor compared with the results obtained
by the PLSP. There are two reasons to explain the poor results of the DLSP: The DLSP does
not keep track of the setup state over idle periods; furthermore batches are too large in some
periods thus increasing holding costs. The interchange of the quantities of the periods 9 and 10
and the splitting of q26=50 (such that q26=30 and q2s=20) would decrease the optimal
objective function value to 1955. The reason for the poor solution of the CSLP lies in wasting
capacity by not making full use of the capacities while producing lots of items.
Let us now discuss the relations between the three types of lotsizing and scheduling
models in more detail:
The "all or nothing production" assumption of the DLSP and the feasibility of continuous
lotsizes for the CSLP implies that the set of feasible solutions of the DLSP is a subset of
the set offeasible solutions of the CSLP.
The PLSP differs from the CSLP only in the restriction (3.18). Ifwe drop the term "BYj.t_
I" the models are identical. Thus the set of feasible solutions of the CSLP is a subset ofthe
set of feasible solutions of the PLSP.
These facts about the sets of feasible solutions lead to the following chain of inequalities:
Z~LSP
;::
Z~SLP
;:: Z;LSP
Thus, disregarding the effort which has to be spent for solving the different models, the
PLSP is superior to the DLSP and the CSLP with respect to the quality of solutions.
Note, that the solution space of the CLSPL (cf. Section 2.3) is greater than the solution
space of the PLSP because more than two items are producible in a period. The CLSPL can
also connect lotsizes of adjacent periods. Thus the following inequality is valid:
Z;LSP ;::
Z~SPl.
Chapter 4.
Extensions of the PLSP
This chapter is concerned with extensions of the PLSP for other potential applications. We
consider setup times and present how sequence dependent setup costs and times can be
included. Normally more than one machine exists which is able to produce certain items.
Hence, we are going to model the multi-machine case. A make-or-buy decision is incorporated
by defining one machine as a "buy-machine". Using the buy-machine, i.e. purchasing items, is
one possibility to overcome capacity shortages or to obtain a more efficient capacity
utilization; filling orders with delay or planned stockouts will also be treated. To take care of
parent-component relationships between items, a multi-level PLSP is introduced modelling a
production facility with one bottleneck. We conclude this chapter with remarks on the
presented extensions, and give an overview over further modifications as well as over aspects
of an integration into a material requirements planning (MRP) system.
4.1 The PLSP with Setup Times
Setup costs include opportunity costs for setup times. Unfortunately, the correct estimation of
these costs is not easy. Setup times consume capacity which is not directly taken in
consideration in generic lotsizing models. Thus disregarding setup times may be very crucial in
practice; scarce machine times as well as substantial setup times may result in infeasible
production schedules. We therefore introduce the PLSPST, i.e. the PLSP with setup times.
Mathematically the PLSPST can be stated as follows:

J
(4.1)
min LL(sJxJt +h)JJ

J=1 t=1
subject to
(j=I, ... ,1; t=I, ... ,T)
(4.2)
32
t
O=I, ... ,J; t=I, ... T; 8=1, ... ,SPj)
(4.3)
(t=I, ... ,T)
(4.4)
B'Yjt + S.Yj,t-1 - Wjt - Pjqjt :?: 0
0=1, ... ,1; t=1, ... ,T)
(4.5)
B(I-zi,)-qit:?: 0
0=1, ... ,1; t=I, ... ,T)
(4.6)
Xjt - Yjt + Yj,t-l :?: 0
O=I, ... ,J; t=I, ... ,T)
(4.7)
2:[P8it +W it ] ~ Ct
(t=I, ... ,T)
(4.8)
Ijt, qjt, Xjt, Wjt :?: 0
0=1, ... ,1; t=I, ... ,T)
(4.9)
YiP Zit E{O,I}
0=1, ... ,1; t=l, ... ,T)
(4.10)
BOzjt - stj(Yjt - Yj,t-8)+ 2: Wit:?: 0

t=t-8+1
i=l
where
B
large number,
Ct
the capacity (hours) of the machine in period t,
hj
holding costs per unit and period of item j (holding costs coefficient),
Ijt
the inventory of item j at the end of period t,
number of items,
Pj
number of hours necessary to produce one unit of item j (processing time),
qjt
the quantity of item j to be produced in period t,
Sj
setup costs of item j,
SPj
maximum number of setup periods necessary for item j (sPj may be set equal to
1 + st J min {C t > Olt = I, ... ,
Tn
Il
with a
being the smallest integer greater than or
equal to a),
stj
the number of hours to setup the machine for itemj (setup time),
planning horizon,
Wjt
the setup time for item j required in period t,
Xjt
a variable indicating whether setup occurs for itemj at the end of period t (Xjt>O) or not
(Xjt=O),
~it
a binary variable indicating whether the machine is setup for item j at the end of period t
(Yjt=l) or not (Yjt=O),
33
Zjt
a binary variable indicating whether setup for item j has been finished in period t (Zjt=O)
or not (Zjt=I).
W.l.o.g. let ljO
=0, Yj.-B =... =YjO =0, w j._ =... =w jO =

B
for j=I, ... ,J. Note that Cb s~
and Wjt must have appropriate dimensions. That means, if e.g. Ct is measured in hours per
period, then s~ and Wjt must be measured in hours, too.
The objective function (4.1) counts, in conjunction with (4.7), for setup and holding costs.
(4.2) are ordinary inventory balance equations. (4.3) "distributes" the setup times over some
periods. The "distribution" of the setup time for item j must take place completely before
producing item j and the setup time of a preceding lot of the item j must not be taken into
account. (4.4) ensures that the machine can be reserved for only one item at a time. (4.5)
guarantees that a production or a setup activity for an item can be performed in a period t only
if the machine is "reserved" for the considered item. Thereby the reservation can be done in the
period t or at the end of the preceding period t-l. Due to (4.6) nothing is produced before the
setup activity is finished completely. (4.8) are the capacity constraints.
Example 4.1: Let J=3, T=8, and Ct=10 (t=I, ... ,T). The processing time Pj, the setup
costs Sj, the holding costs hj , and setup time stj ofitemj (=1, ... ,1) are as follows:
Sj
hj
s~
j=1
50
j=2
SO
j=3
SO
Pj
34
The demand of the item j in period t, djt, is (missing entries are zero) :
12
14
10
The optimal objective function value is Z;LSPST = 304 . The lotsizes gjt and,
In
parenthesis, the setup times Wjt of the optimal PLSPST solution are given below:
glt (wit) 6 (4)
(2)
q2t (w2t)
2(4)
10
10
6 (4)
931 (W31)
8 (2)
This solution is illustrated in Figure 4.1 :
sr\llp activil
[jJ
10
I-
~l
prodllrtlon of Ilem
:J [
2
6
Fig. 4.1 Gantt chart of the machine schedule with setup times
Note that the setup activity for item j=2 is distributed over two periods (cf. t=3 and t=4
in Figure 4.1).
35
Until now only very little research has been published on lotsizing (and scheduling) with
setup times. Setup times have been considered for the CLSP, a CLSPL like model, and the
DLSP, in [Cattrysse et al. 1990], in [Dillenberger et al. 1992], and in [Trigeiro et al. 1989]
respectively.
Note that in the CLSP and CLSPL context the setup times are restricted to be less than
one period length such that the setup activity can take place in a period completely, i.e. it is not
possible to "distribute" setup times over more than one period. The extended DLSP requires
setup times which are multiples of period lengths. Thus, regarding existing models with setup
times, the important features of the PLSPST are:
(i) the length of setup times is not restricted and
(ii) the setup activity can start at any (continuous) point of time within a period.
4.2 The PLSP with Sequence Dependent Setup Costs
In many real life problems setup costs (and times) are not sequence independent. We are now
going to present the PLSPSDSC, i.e. the PLSP with sequence dependent setup costs.
Mathematically the PLSPSDSC can be stated as follows:

J
(4.11 )
min LL(SC jt +hjl jt )

j=1
t~1
subject to
Ij t-1 + qjt - Ijt = djt
0=1, ... ,1; t=I, ... ,T)
(4.12)
LYj, ~1
(t=I, ... ,T)
(4.13)
0=1, ... ,1; t=I, ... ,T)
(4.14)
0=1, ... ,1; t=I, ... ,T)
(4.15)
pI
36
1
LPjqjt::;;C t
j=1
(t=I, ... ,T)
(4.16)
qjt,SCjt :<:0
G =1, ... ,1; t=I, ... ,T)
(4.17)
Yjt E{O,I}
G =1, ... ,1; t=1, ... ,T)
(4.18)
where
SCjt= setup costs for itemj to be paid in period t,
sC ij = setup costs which occur when the setup state of the machine changes from item i to
itemj,
sc
max {sca.~ Ia.,
f3 = I, ... ,J},
SCij = sc-sc ij ,
and the other symbols are the same as in the PLSPST.
The costs of the first setup are calculated correctly in (4.15) only ifYjo is given, otherwise
the first setup activity is valued with sc. These maximal costs are also incurred by a
changeover from a non-setup in a setup state; but this is not a crucial aspect since no additional
costs are incurred for keeping up a setup, i.e. SCjj=O G=I, ... ,J), so it cannot be less costly to
change in a non-setup state. However, if a setup changes from item i to item j in period t, that
means Yi.t-l=1 and Yjt=I. it follows: SC jt - SC+SCij :<: 0 e> SC jl :<: sc ij ; thus according to the
objective function SCjt will take the correct setup costs SCij-
Example 4.2: Let YlO=1. Despite of setup costs and times we take the same data as in
Example 4.1. We assume that item j=1 is member of one family of items and the items
j=2 and j=3 belong to another family. A changeover between items of different families
causes (major) setup costs of a relative high amount as may be seen in the following
table of sequence dependent setup costs:
37
2
SClj
200
200
SC2j
200
50
SC3j
200
50
We get
Z;LSPSDSC
= 396 as optimal objective function value. The lotsizes of the optimal
PLSPSDSC solution are given below.
10
8
8
2
10
10
This solution is illustrated in Figure 4.2:
C/Pj
[] production of item j
lOr
1
I
[31
2
I
5
I
6
I
7
Fig. 4.2 Gantt chart of the machine schedule with sequence dependent setup costs
Thus in period t=3 a changeover between the families is necessary only.
Note that sequence dependent setup times can be included into the PLSPST by replacing
(4.3) by:
J _
Bz jt -stYj.t_/i + LStijYi.H + LW j< 20 (j=l, ... ,J; t= 1, ... ,T; 8=1, ... ,sPj) (4.3')
i~l
't=t-8+1
38
where st (Stii) is defined analogous to sc (SCii) and the other symbols as well as the necessary
other required restrictions are the same as in the PLSPST.
Until now only little research has been published on problems with sequence dependent
setup costs. In [Schrage 1982] a DLSP-Iike model with sequence dependent setup costs is
considered. In [Fleischmann, Popp 1989] the approach of [Fleischmann 1990] is extended to
the DLSP with sequence dependent setup costs. In [Fleischmann 1992] the DLSP with
sequence dependent setup costs is formulated as a travelling salesman problem with time
windows. A model which requires a lot of additional variables is presented in [Dilts, Ramsing
1989]. A heuristic for a static (i.e. constant demand per period) lot scheduling problem with
sequence dependent setup costs and times is introduced in [Dobson 1992].
4.3 The Multi-Machine PLSP with Make-or-Buy Decisions
Often, in practice lotsizing and scheduling takes place on more than one scarce machine.
Hence, we will now consider the PLSPMM, i.e. the multi-machine PLSP. We suppose
m=1, .. ,M machines whereby the machine m=M is defined as a "buy-machine" in order to
include make-or-buy decisions.
Mathematically the PLSPMM can be stated as follows:
min
t~L~j(SimXimt +kimqimt)+hJIit)
(4.19)
subject to
0=1, ... ,1; t=1, ... ,T)
(4.20)
(m=1, ... ,M-1; t=1, ... ,T)
(4.21)
B'Yjmt + B'Yjm,t-l - qjmt ~ 0
(m=1, ... ,M-l;j EJm; t=1, ... ,T)
(4.22)
B'XjMt - qjMt ~ 0
(jEJ M; t=1, ... ,T)
(4.23)
Ij,t-l
Lqiml meM j
~t = djt
39
Xjmt - Yjmt + Yjm,t-l ~ 0
(m=I, ... ,M-I;j EJm; t=I, ... ,T)
(4.24)
LPjmqjmt :;;;C mt
(m=I, ... ,M-I; t=I, ... ,T)
(4.25)
(m=I, ... ,M-I;j EJm; t=I, ... ,T)
(4.26)
G EJM; t=I, ... ,T)
(4.27)
jeJ m
C\jmb Xjmt ~ 0; Yjmt E {O, 1}

qjMt~O; XjMt E
to, I}
where
Cmt = the capacity of the machine m < M in period t
J m = the set of items producible on machine m
kjm = the costs to produce (purchase) one unit of item j on (with) the machine m<M (m=M)
(Usually, it is supposed that it is substantial less costly to produce the itemj than to buy
the itemj, i.e. kjm (m<M) kjM GEJM
Mj = the set of machines capable of producing item j,
qjmt= the quantity ofitemj to be produced (purchased) on machine m<M (m=M) in period t,
Sjm = setup (order) costs for itemj on machine m<M (m=M),
Xjmt = a variable indicating whether setup (purchase) occurs on machine m<M (m=M) for item
j in period t (xjmt=l) or not (Xjmt=O),
Yjmt = a binary variable indicating whether machine m<M is setup for item j at the end of
period t (Yjmt=l) or not (Yjmt=O),
The buy-machine (the purchase) is unconstrained (cf (4.25. Since, it is not sensible to
define a setup state of the buy-machine the "setup costs", i.e. order costs, occur for each
purchase ofa lot of an item (cf (4.23.
Example 4.3: Let M=2, the demand as in Example 4.1, C 1t=5 (t=I, ... ,T),
JI
={I, 2, 3},J 2 = {I, 2, 3} and the other data as follows:
40
Pil
j=1
j=2
j=3
~l
~2
hi
kjl
kj2
50
25
10
50
25
10
50
25
10
The optimal objective function value is Z~PMM = 424. The lotsizes of the optimal
PLSPMM solution are given below.
qllt
5
4
q2lt
q3lt
ql2t
q22t
10
q32t
The solution is illustrated in Figure 4.3.
c1t /pIj
51
[}
qj2t
production of item j on machine m=l
[]
l:f
>
purchase of item j (buy-machine m=2)
[J]
4
~
8
Fig. 4.3 Gantt charts of the two machine schedules
>
41
In this example, it is too costly and/or there is no capacity to produce item j=3 so a
buy-decision is recommended.
Note, the integration of a buy-machine is "realistic" and provides a way to overcome

feasibility problems (e.g. if J~{I, ... ,J} one feasible solution can be determined by using the
buy-machine only).
The CLSP with multiple machines has been dealt with in, among others, [Maes et al.
1991]. The DLSP with multiple machines is addressed in [Eppen, Martin 1987] and analyzed in
[Salomon et al. 1991]. Optimality properties for a CLSP-like single item model with make-orbuy decisions are presented in [Lee and Zipkin 1989].
4.4 The PLSP' with Backordering and Stockouts
Instead of using a "buy-machine", alternatives to cope with a given demand or scarce

capacities are to plan for backorders or stockouts. Recall, a backorder is an order not ready
when promised and a stockout is a matter of not satisfying the order. In the former case, the
order is not cancelled, but the demand is pushed in later periods. In the latter case, the
customer may go to a competitor, resulting in a lost sale. However, costs are incurred for
expediting the past-due order, losing the customer in the future (loss of good will) as well as
losing profit from the sale.
Now we are going to present the PLSPBOSO, i.e. the PLSP with backordering and
stockouts.
Mathematically the PLSPBOSO can be stated as follows:
42
1
LDr+1
min LL(sjXjt + hjIjJ +

Lgnun
j=1 t=1
f=1 t=EDr
(4.28)
subject to
F
Lf=1 d jf r ft uft=O
O=I, ... ,J; t=I, ... ,T)
(4.29)
(f=I, ... ,F)
(4.30)
LY1t $; 1
1=1
(t=I, ... ,T)
(4.31)
B'Yjt + B'Yj,t-1 - %t ~ 0
0=1, ... ,1; t=l, ... ,T)
(4.32)
Xjt - Yjt + Yj.t-l ~ 0
O=I, ... ,J; t=I, ... ,T)
(4.33)
LPlqJt $;C t
1=1
(t=I, ... ,T)
(4.34)
Xjt, qjt~ 0
0=1, ... ,1; t=I, ... ,T)
(4.35)
Y1t E{O,I}
0=1, ... ,1; t=I, ... ,T)
(4.36)
un E{O,I}
(f = 1, ... ,F; t E[EDf,LD f + 1])
(4.37)
Ij.t-l + qjt - Ijt LDr+1
LU n =1
t=EDr
1
where
r ft = {
ifEDf $; t $; LD f
.
, i.e. a "binary parameter" which indicates whether the demand of
otheTWIse
orderf(=l, ... ,F) can occur in period t (rft=l) or not (rft=O).

gft
costs for delivery or cancellation of order f in period t
djf
demand for item j in order f
EDf= earliest delivery date of order f ($;LDf)

F
number of orders
LDf = latest delivery date of order f (LD(=T)

uft = binary variable indicating whether order fis delivered or cancelled in period t (uft=l) or
An order may consist of the demand for more than one item which must be delivered
jointly (cf. (4.29)). Furthermore, an order f can be delivered in the period t=ED f at the earliest
43
and in t=LDf at the latest. Is that impossible, a lost sale of order f occurs in the period f=LDf
+1 (i.e. uft'=1 cf. (4.30 with costs gft' (cf. (4.29.
Example 4.4: Let J=3, T=9, F=4, and Ct=6 (t=I, ... ,T). The processing times, holding
and setup costs are the same as in Example 4.l. For each order fthe earliest and latest
delivery times, and the amounts of the demand for each item are given below:
EDf LDf dlf
d2f
d3f
f=1
f=2
f=3
10
f=4
4.
0 0
0 1
.
0 0 1 1 0 0 o
000
Thus (fft) = ~ 0 0 0 0 0 0 1 0
0 0 0 0 0 010
0 0]
[0
The costs which are incurred for delay or cancellation of an order are as follows:
glt
10
200
g2t
10
200
g3t
480
~t
240
We get Z~SPSBOSO
=524
as optimal objective function value. The following table
provides the optimal production quantities and, for each order, the date of delivery or
cancellation.
44
qlt
q2t
q3t
ult
3
1
U2t
U3t
U4t
The orders f-=3 and f-=4 will be delivered without delay, the order f-=1 will be cancelled
(stockout) and order f-=2 will be delayed (backorder) one period due to capacity
restrictions.
Note that backlogging (accumulation of orders that have been promised for delivery at
some future date) can be modelled equivalently to backordering. In the backlogging case, EDf
is the day we have promised to satisfy the order f in the future and gft are the shortage costs of
order f in period t, respectively.
Thus, in the following we refer to the literature for backordering as well as for
backlogging. In [Zangwill 1969] a dynamic economic lotsize production system with
backlogging is analyzed. [Haehling von Lanzenauer 1970] presents a multi-stage DLSP-like
model where the possibility of unsatisfied demand is included by introducing of shortage costs.
A backorder version of the Wagner-Whithin [cf. Wagner, Whitin 1958] algorithm is presented
in [Webster 1989]. In [Choo, Chan 1990] a so-called two-way eyeballing heuristic for the
dynamic lotsizing problem with backlogging is introduced. This heuristic is compared with the
Silver-Meal and Groffheuristic, which have been modified for the backlogging case in [Hsieh,
Lam Choo 1992]. The results of a simulation study with respect to the performance of
traditional lotsizing algorithms, an algorithm for lotsizing with backorder as well as the
backorder version of the EOQ algorithm are presented in [Gupta, Brennan 1992].
45
4.5 The Multi-Level PLSP with One Bottleneck
In most real-life manufacturing situations the production process covers several levels. In a
multi-level process bills of materials have to be regarded, i.e. an item is producible only if its
components are available. To ensure that a component, transformed by one or more
operations, is available on time an item specific lead time has to be considered (cf. Chapter
2.2). We assume that such a lead time may not depend on the ordered quantity.
A machine that limits the output of the multi-level process is called bottleneck or critical
resource. We assume that the bottleneck is always the same resource, i.e. the bottleneck does
not shift from one machine to another due to demand variation. We formulate the
PLSPMSOB, i.e. the multi-stage PLSP with one bottleneck.
Mathematically the PLSPMSOB can be stated as follows:

J
(4.38)
min L:L:(sJXjt +hh)

FI H
subject to
Ij.H +qj.t-Vj - L:a'Jq;t -I jl =d jl
0=1, ... ); t=I, ... ,T)
(4.39)
(t=I, ... ,T)
(4.40)
OEJ BN; t=l, ... ,T)
(4.41 )
t=I, ... ,T)
(4.42)
iEN j
O~JBN;
Xjt - Yjt
(4.43)
+ Yj-t-l ~ 0
(t=I, ... ,T)
(4.44)
qjt, Xjt ~ 0
0=1, ... ); t=I, ... ,T)
(4.45)
do, I}
O=I, ... ,J; t=I, ... ,T)
(4.46)
Yjl
where
46
Ilij = the number of units ofitemj required to produce one unit of item i
JBN = set of items which run through the bottleneck
Nj
set of immediate successor items of item j,
Vj
lead time of item j (in periods) without the time for passing the bottleneck,
Due to constraints (4.39) inventory is balanced in the multi-level system. A lead time of at
least one period for components ensures feasibility (cf Chapter 2.2). Note that the lead time of
an item cannot consist of the time to pass the bottleneck since it is not known before finishing
the planning process, i.e. the total lead time is a result of planning.
Example 4.5: Let J=4, T=8, and Ct=8 (t=I, ... ,T). The final items j=1 and j=2 are
produced by a multi-level process as illustrated in Figure 4.4.
bottleneck
Fig. 4.4 Multi-level process with one bottleneck
The items j=2 and j=3 are produced on the bottleneck machine (cf dotted box in Figure
4.4) whereby P2=P3=1. The quantity of item i necessary to produce one unit of item j,
Ilij, is depicted at the arc from node i to node j. It is assumed that external demand
occurs for the final items only (i.e. d 3t=d4t=O (t=I, ... ,8. The demand for the final
items, the setup costs, the holding costs coefficients, and the lead times are given
below:
47
2
Sj
hj
Vj
j=1
25
j=2
50
j=3
50
j=4
25
12
14
We get Z~SPMSOB = 328 as optimal objective function value. The following table
provides the optimal production quantities:
12
2
6
q4t
16
4
4
24
In the optimal solution of this multi-level problem the quantity of item j=4 in period
t=l, i.e. q41 satisfies the demand for its "parents" j=2 and j=3 from peri04 t=3 (v4=2!)
up to t=4. Thus, for producing 6 units in period t=4 of item j= 1 6 units of item 3 must
arrive in inventory not later than in period t=3.
One of the first multi-stage models is due to [Haehling von Lanzenauer 1970] which is
based on the "all or nothing production" assumption as in the DLSP. A simulated annealing
approach for a 2-stage DLSP with multi-item batch production is suggested in [BrOggemann,
Jahnke 1992].
48
Multi-level lotsizing problems are modelled as constrained generalized networks in
[Steinberg, Napier 1980]. Uncapacitated multi-stage lotsizing problems are solved via shortest
path relaxations in [Afentakis, Gavish 1986]. The efficiency of single-stage heuristics in a
multi-stage setting is analyzed in [Blackburn, Millen 1985].
A lot of papers focus on multi-level CLSP extensions. Heuristics for serial multi-stage
systems are provided in [Billington et al. 1988]. A heuristic method, based on Largrangian
relaxation, for multi-levellotsizing with a single bottleneck facility is presented in [Billington et
aI. 1986]. LP-based heuristics can be found in [Maes et al. 1991]. Simulated annealing
approaches are evaluated in [Kuik, Salomon 1990]. Further special heuristics may be found in
[Afentakis 1987; Heinrich 1987; Roll, Karni 1991; Tempelmeier, Helber 1992].
4.6 Concluding. Remarks
It is not difficult to incorporate the different PLSP extensions treated above or additional
aspects (e.g. multi-resource constraints, safety stocks) in a general model. There are several
important aspects with respect to the presented extensions:
(i) By variations in item mix, even low-usage machines can become bottlenecks from time to
time. Thus, especially the multi-machine case in a multi-level environment is rather important.
(ii) Even the single-stage single-machine DLSP with non zero setup times, the single-stage
multiple machine DLSP without setup times as well as (beside others) the multi-stage CLSP
with non-zero setup times belong to the class of the NP-complete feasibility problems [cf
Salomon et al. 1991; Maes et al. 1991]. Thus, to develop an algorithm that computes a optimal
solution for a more realistic model in reasonable time seems to be a very difficult task.
However, including make-or-buy decisions or backordering and stockouts in our model
provides a way to overcome the feasibility problems associated with non-zero setup times or
multiple machines.
49
(iii) The incorporation of backordering (backlogging) facilitates the computation of optimal
delivery dates in a sales department. This might be more realistic than the typical assumption
that each demand has to be satisfied without any delay.
(vi) The high complexity of capacitated multi-level problems has forced the development of
"rough" solutions approaches for existing manufacturing resource planning systems (MRP II).
They are mostly based on a successive concept which has a lot of shortcomings (ineffective
capacity usage, past due-dates and large lead times). Thus there is still a potential space for
improvements. In the basic successive planning concept capacity is disregarded in the first
phase (and others). This may be one reason for the poor performance ofMRP II systems. To
overcome this problem it seems to be more effective to follow a hierarchical planning concept
where in each step the capacity is considered [Drexl et al. 1993]. For example, in a first step
one may solve a capacitated single-level multi-item problem (CLSP) with high aggregated data
(capacity, level). Based on that solution, the aggregation level can be reduced and resulting
(sub-) problems may be solved by approaches which are based on lotsizing and scheduling
models (e.g. PLSP).
Chapter 5.
Control of Stochastic Algorithms via Sequential

Analysis
This chapter deals with the control of stochastic algorithms which depend on some parameters.
Stochastic algorithms generate a (large) number of solutions. The information in the objective
function values of already generated solutions can be used to "tune" a stochastic algorithm for
further computations. This is to say, that certain "good" parameter values are preferred in hope
for a solution which is better than any solution already known. The question of which
parameter values are to be preferred will be answered by a statistical test.
The expected sample size of so-called sequential tests is smaller than the sample size of
tests with fixed sample size and same power. The application of a sequential test to identifY
non-preferable parameter values saves computation time which can be spent for a search with
more interesting parameter values.
The' test problem of identifYing non-preferable parameter values is introduced in detail
in Section 5.1. To solve this test problem, in Section 5.2 some sequential tests are presented.
They are compared by a Monte-Carlo study in Section 5.3. In Section 5.4 a simple example
shows how substantial improvements can be achieved using a (sequential) test. A summary and
some remarks on potential extensions to the method are given in Section 5.5.
5.1
Hypotheses Test Problem for Reducing the Parameter Space
Let SA(8) be a stochastic algorithm for a minimization problem (e.g. the PLSP) the
performance of which heavily depends on 8. For the sake of simplicity let us assume that 8 is a
two dimensional vector consisting of the non-negative real-valued parameters y and 0, I.e.
8=(y,0).
51
To support the choice of the parameter values we partition the parameter space as
follows:
Consider
[Yo,YI),[Yl>Y2), .,[YhYi+ I)' .,[YI-l>yd and

[oO,ol),[0I,02),,[~,ok+ I)' .. ,[OK-I ,OK].
Then construct the following H=IK cross products (see the instance partitioning
provided in Figure 5.1 for I=K=2):
8 1 = [YO,Yl)x[OO,OI)
8 2 = [Yo,Yl)x[l>2)
Fig. 5.1 Partition of the parameter space for I=K=2
Thus {8 1, ... ,8H} corresponds to a partition of the parameter space 0, so that

H
0:== U8h and 8 j (l8 j == 0 for i 1= j.

h=1
Spaces of discrete parameter sets (cf Section 5.4) or of higher dimensions may be
partitioned analogously way.
52
In the following we need some additional symbols: x denotes a solution of the
considered minimization problem, Z(x) the objective function value associated with x. SA(0),
operating on the full parameter space 0, tailors to SA(9tJ in the case where it only operates on
the subspace 9 h. In addition the application of SA(9tJ reveals Z(x(9tJ) to become a random
variable and Zl(x(9tJ), Z2(x(9tJ), ... to be a stochastic process. We assume that the parameter
values are chosen from a subspace 9 h independently before executing SA(9tJ thus Zl(x(9tJ),
Z2(x(9h, ... are independent.
------- mean
----O.2th quantile of
o
eh
obj. funct. value
repetitions of SA(e h)
Fig. 5.2 Objective function values depending on subspaces 9 h
Now the fundamental question arises, how (often) to perform SA(9tJ, h=l, ... ,H. Figure
5.2 provides a fictitious example demonstrating potential relationships between the mean and
the q-th quantile, ~(9h)q, associated with the random variables Z(x(9tJ), h=1, ... ,4. It seems to
be meaningless to take (an estimate of) the means as the only criterion. Rather, an estimate of a
combination of the mean and the standard deviation could be a good suggestion for choosing
subspaces (in the case that all solutions are feasible and distributed with symmetry). Moreover,
as Figure 5.2 indicates, as a single criterion the q-th quantile
~(9tJq
(e.g. q=.2) seems to be a
53
good indicator for choosing subspaces; i.e. in subspaces 8 h for which the q-th quantile is
relatively high, the probability of determining low costs objective function values is relatively
small. Thus in the following we try to identify those subspaces 8 h of E> which have a small
quantile (e.g. 1;(8hb
::; 1;); and the subspaces 8 h ofE> which have a large quantile (e.g. 1;(8hb
> 1;) will be "eliminated" from E>. Thus SA(E will operate on subspaces 8 h ofE> in such a way
that the probability to determine solutions with low costs is increased.
Whether 1;(8h)q::;1; or 1;(8 h)q>1; holds may not be known before performing SA(8 h)
(h=l, ... ,H). But a sample of objective function values can be determined by SA(8h) to estimate
1;(8 h)q which can be used to decide the following test problem
where H 0 is called the null-hypothesis, being tested, and HI the alternative hypothesis.
Let us assume that for our instance partitioning, I=K=2, the hypotheses regarding the
quantiles of 8 1 and 8 3 are rejected. Then the parameter space is reduced by eliminating the
subspaces 8 1 and 8 3 as illustrated in Figure 5.3. Thus, SA(8)) and SA(8 3) will not be
performed any more after rejecting the hypotheses regarding to 8 1 and 8 3 ,
Fig. 5.3 Reducing the parameter space via test of hypotheses
Suppose that we have planned to perform, say, a total of 1000 further executions of
SA(E. After rejecting the hypotheses regarding 8 1 and 8 3, then SA(8 z ) and SA(8 4 ) could be
performed 1000/2=500 times each instead of 1000/4=250 times only which may increase the
54
probability to obtain a solution which is better than an already known solution. Another
strategy is to perform 250 executions of each of SA(8 2 ), and SA(8 4 ) as planned, and save
computation time due to non executing SA(8]) and SA(8 3 ).
How such test problems can be solved with a so-called sequential test is considered in
the following section.
5.2 Sequential Tests of Hypotheses Concerning Quantiles
Let P[ro] denote the probability that an event ro occurs and fz(z) be the (unspecified) density
function of the continuous random variable Z. The q-th quantile ofZ, denoted by~, satisfies
q=P[Z~q] =
Jfz(z)dz.
~q
Let ZI, ... ,Zn denote a sample from fz(z). Suppose that it is desired to test whether 1;q
offz(z) is less than or equal a specified value, say 1;, i.e. we would like to test
Let Un denote the number OfZi'S that don't exceed 1;, i.e.
n
Un=L~'
i=l
where
ifZj ~1;
V ={l,
,
0, otherwise
Due to the independence ofZi Vi has a Bernoulli density function,

(for ViE {O, I}),
and Un has a binomial density function,
f(u,;n,p) =(:Jpu, (1- p)n-u;
(for uiE {O, 1, ... ,n}),
55
where p (O:::;p:::;l) denotes the probability that the Zi' s don't exceed
~.
(Note that Un, Vi as well
as Zi are random variables and un' vi, and Zi are realizations of the random variables,
respectively. )
The smaller the quantile
~q,
the higher the probability p; e.g. if ~q :::; ~ it is
Ho is true Vi is f(vi;P) and Un is f(un;n,p) distributed satisfying
p~q.
p~q.
Thus if
So if Ho is true, one
would expect Un to be near or greater than n'q, and Ho will be rejected if and only if Un is
small [cf. Mood et al. 1982, p. 515]. That means the test problem concerning a quantile can be
transformed in a test problem concerning the parameter of a Bernoulli or binomial density
function, i.e.
Ho: P ~ q versus H{ p < q
The power function of a test, denoted by 1t(\v), is defined to be the probability that H 0
is rejected when the sample distribution is parameterized by \11 (=p in our problem), i.e.
1t(\I/) = P[H0 is rejected; f(.;\V)].
The power function is used to assess the goodness of a test. To find a powerful test we
consider so-called sequential tests which refer to techniques for testing hypotheses when the
sample size is not fixed in advance but is determined by criteria which depend on the
\
observations. A test problem can be solved either by a test with fixed sample size or by a
sequential test with same power. Generally it is known, that the expected sample size of a
sequential test is smaller than the sample size of a test with fixed sample size. Thus, we will
prefer sequential tests. There are a number of different sequential tests [cf. e.g. Siegmund
1985, Irle 1990]. Among them are sequential probability ratio tests as well as repeated
significance tests which will be discussed in the sequel.
56
Sequential probability ratio tests
The sequential probability ratio test, denoted by SPRT, is derived as a test of a simple
hypothesis against a simple alternative. For the composite hypotheses from' above, it is
necessary to choose surrogate simple hypotheses pO;e:q and pI<q, and use a SPRT of pO against
pI [cf. Sigmund 1985, p. 14]. The SPRT is based on the likelihood-ratios
(for n=I,2, ... ) where f(.;pi) denotes the Bernoulli density function with the parameter pi, je
{O,I} [cf. Mood et al. 1982, p. 466]. Using the conventional symbols, i.e. ex. (:;; .1) denotes the
type I error (rejection of H 0 when it is true; significance level) and
(:;; .1) the type II error
(acceptance of Ho when it is false), we get the following stopping and decision rule for the
SPRT:
Stopping nile: Stop at N= min {nlLRn (l(A,B)}, where A=a.I(I-~) and B=(1-ex.)/~.
Decision nile: Reject Ho. ifLRN ;e:B; accept Ho. ifLRN:;;A.
Besides the observations the SPRT depends on the type I and II errors,
(ex.,~),
and the
"surrogates" po and pI, which will be denoted by SPRT(po,pI,ex.,~).
Difficulties in application caused by the absence of a definite upper bound, say r, on the
stopping rule of the SPRT have forced to consider the truncated sequential probability ratio
test, denoted by SPRT-t. The SPRT-t stops at the latest at the upper bound r. If LRn e(A,B)
for n=I, ... ,r one possibility to make a decision is to reject Ho. ifLRn is more closely to B than
to A, otherwise accept Ho. Thus, the stopping and decision rule for the SPRT-t(pO,pI,ex.,~,r)
can be stated as follows:
Stopping nile: Stop at N= min {nlLRn ~(A,B)vn;e: r}.
Decision rule: Reject Ho. ifLRN>.5 (B-A); otherwise accept Ho.
57
Repeated significance tests

Let r be the arbitrary, but fixed (maximum) sample size, and c' a constant satisfYing O<c' ~r. As
mentioned above Ho will be rejected ifand only ifU r is small, i.e. ifU r < qr or
Ur
qr-c'.
In a standard fixed sample .05 level significance test c' will be determined by:
P[Ur ~ qr-c']=.95
(I.e. the probability of type I error is .05). Now
P[Ur ~ qr-c'] =
L:::
'
f(u;r,q)
thus c' can be determined from a binomial table. To terminate as soon as possible, in a repeated
significance test (RST) the sample is taken sequentially and after each drawn observation a
decision rule is used, based on the standard fixed sample significance test. The RST stops as
soon as the null-hypothesis can be rejected or the maximum sample size r is reached.
Whenever the decision rule is performed (after each drawn observation) in the RST, it
is possible to reject a true null-hypothesis, i.e. the probability to make an error of type I will be
greater than .05 if c' is used. Thus in order to guarantee that the significance level does not
decrease (i.e. a> .05) the decision rule has to be defined more carefully as in a standard fixed
sample significance test, i.e. we have to choose a constant c which is larger than c' in our
decision rule. The rejection region of such a test is bounded by a linear function fjb(n)=q'n-c;
so a RST-Ib(q,c,r) denotes an RST with linear boundary qn-c, and maximum sample size r.
From the central-limit theorem about the distribution of a sum of independent and
identically distributed random variables it follows that
D=(U n -n.q)/(n"q(1-q)/n)
is for large n asymptotically standard normal distributed, denoted by <1>(0) [cf. e.g. Mood et aI.
1982, p. 120]. Using such an approximation the following decision rule can be derived
intuitively: reject Ho if and only if D<-b where b must be larger than 1.645 in order that the
significance level is .05 [cf. Siegmund 1985, p. 4]; note <1>(-1.645)=.05. The rejection region of
58
such a test is bounded by a curved boundary fcb(n)=nq-nb~q(1- q) / n; so in the following

such a test is denoted by RST-cb(q,b,r).
Formally the RST -lb(q,C,f} and RST -cb( q,b,r) can be stated as follows:
Stopping rule: Stop at N= min
{N, r}.
Decision rule: Reject H0, ifNS;r; otherwise accept Ho.

where
N = min{njU o < nq -c} or

N =min{nl Un <n. q -n . b ~rq-:-:(1--<!-:)-:/-n},
respectively.
(Note, q is defined by the test problem thus q has not to be determined before
performing a RST -lb( q,c,r) or RST -cb( q,b,r), i.e. q is not a free parameter. If a SPRT is used
to derive a decision we will always set pO = q which wiII be motivated later.)
To illustrate how such tests are performed we consider the Example 5.1.
Example 5.1: Let H 0: p>.s (i.e. q=.s), H]:pS;.s, /;=9, a=.Os, 13=.05; so A=a/(l-f3)
=0.053 and B=(l-a)/f3=19. To solve this test problem a sample, 21>z2, ... will be
obtained sequentially. We consider a SPRT(.2,.s,.Os,.OS), a SPRT-t(.2,.s,.Os,.OS, 10), a
RST-lb(S,4,1O), and a RST-cb(.S,2,10), i.e. it is pO=.s, pl=.2, c=4, b=2, and r=1O.
Thus it is ~b(n)=nl2-4, fcb (n)=nl2--{ri, and LRn=.4u, 0.62S u,o. The following table
provides the decisions of the tests for a fictitious sample.
59
Zn
vn
un
LRn
qb(n)
fcl,(n)
11
1.6
-3.5
-.5
15
2.56
-3
-.41
12
4.1
-2.5
-.23
13
6.6
-2
14
10.5
-1.5
.26
15
16.8
-1
16
26.
-.5
11
10
18
0
.5
1
We see that the required sample size and the recommended decision of the tests can be
different. The SPRT's stop at the 7th, the RST-Ib and the RST-cb at the 10th and 5th
observation, respectively. Since, U lO is not less than qb(lO) the RST-Ib does not reject the
null-hypotheses. By the other tests the alternative hypothesis is accepted.
Which decision is correct depends on the specific distribution of Z which is generally
unknown. To analyze the error probability and to derive recommendations for the test
parameter values (P0,pl,b,c,r) we have performed a Monte-Carlo study which we present in
the following section.
5.3 Monte-Carlo Study for Sequential Tests
Let us assume that we are only interested

to recognize as soon as possible that H0 is not true
,(i.e. p<q); and if H 0 is true, it does not matter whether the sample size is large. Thus, for our
purpose we are interested in the expected sample size of a test rejecting H(), denoted by E[N
IHo
is rejected]p where p denotes the true percent point. Furthermore, let SR(k) denote a
sequential test procedure which draws at maximum k
(~r)
observations and stops before the
60
k-th observation if and only if HO is rejected. Thus E[Ni SR(k)]p denotes the expected sample
size under the condition that the procedure SR(k) is used and the true percent point is p.
Remark: SR(k) refers to the case that a stochastic heuristic is executed at maximum k times.
Thus, only ifE[NISR(k)]p is less than k, computation time is saved.
Let a 10000 replication Monte Carlo experiment denote that 10000 times a sample is
generated sequentially (by a probability function with a known q"th quantile) as long as a
considered test rejects or accepts a considered null-hypothesis. The average sample size as well
as the number of rejections are used as approximations for the expected sample size and the
expected power function value, respectively.
For gaining a better assessment of expected sample sizes and the power function we
have analyzed the sequential tests described above by a Monte-Carlo experiment (cf
Appendix). The performance of a test may depend on the test problem, especially on q, and on
the values of the test parameters.
The error of type I depends on the respective parameter values of the SPRTs and
RSTs. Thus, if a .05 significance level is desired we have to determine the corresponding
parameter values. For our test problem the condition is satisfied for a SPRT if 0.=.05 and pO=q.
More difficult is the determination of the respective parameter values for a RSTIb(q,c,r) and a RST -cb(q,b,r), since there exits, in the authors opinion, no appropriate approach
of doing so. Thus for a certain hypotheses test problem, (i.e. for qE[O, 1]), a simulation study is
required to obtain suitable test parameter values (b, c, r).
Such a simulation study requires a comprehensive design. For instance, for a RSTIb(q,c,r) many combinations of c and r must be compared in order to have a test which has a
good performance and a .05 significance level. This is to say, that for each combination of c
and r the expected sample size and the significance level must be approximated by simulation.
61
Let q=po=.2, pl=.05, 0.=.05, ~=.05, c=5.7, b=2.12, and r=60 (determined by a preexperiment). Table 5.1 gives the (final) results of a 10000 replication Monte-Carlo experiment
for the test problem
Ho: P ~ .2 versus HI : p < .2.

I size
. approximatIOns fior Hrn : p~. 2
T a ble 51 Power and sample
SPRT
SPRT-t
.2,.05.05,.05)
truep
RST-Ib
.2,.05,.05,.05,60) .2,5.7,60)
RST-cb
'.2,2.12,60)
.2
.05
.2
.05
.2
.05
.2
.05
P[rej.HO]p
.040
.967
.044
.934
.047
.976
.044
.945
E[N]p
22.6
30.5
21.8
29.5
59.5
38.5
58.8
30.4
E[NIHO is rej.]p
31.8
30.8
28.5
28.7
48.8
38.0
32.0
28.6
E[NISR(60}]p
58.8
30.7
58.6
30.8
59.5
38.5
58.8
30.4
E[NISR(80}]p
78.1
32.1
77.8
32.1
78.5
39.0
77.9
31.5
E[NISR(100}]p
97.3
33.0
96.9
33.5
97.5
39.5
97.0
32.5
Table 5.1 indicates that the SPRT(.2,.05,.05,.05) is more powerful than the SPRTt(.2,.05,.05,.05,60} and RST -cb(.2,2.12,60) due to the low (approximated) probability to make
an error of type I (=.04 ) or II (=1-.967). The approximated probability to make an error of
type II of the RST-Ib(.2,5.7,60} (=1-.976) is less than in the other tests. However, the
approximation for the expected sample sizes of the RST-lb(.2,5.7,60} is poor compared to the
respective values in the other tests. Especially, in the interesting case of a false null-hypothesis
(i.e. p=.05) the sample size is relative large.; e.g.
E[Nlos=38.5
of the RST-
Ib(.2,5.7,60} is more than 20 percent larger than the corresponding value of the
SPRT(.2,.05,.05,.05}. (Note, in simulations with a 5.3 :::; c < 5.7 and P[rej.Holz >.05 it turned
out that E[Nlos > 30.) Comparing the other tests no substantial differences can be found
between the expected sample sizes. Let us consider the approximated power functions in
Figure 5.4 and expected sample sizes. using SR(80). in Figure 5.5.
62
w(P) 100
-+- SPKI'(.2,.OS,.OS,.OS)
100
_ _ _ SPRrt(.2,.OS,.OS,.OS,60)
90
80
70
80
50
40
30
20
10
- . - RST-Ib(.2,5.7,60)
--*- RST-cb(.2,2.12,60)
O. .====~~----~------+-----~
.25
.2
.15
.1
.05 P
Fig. 5.4 Power functions for Ho: p~.2
E[NISR(80)]p
-+- SPKI'(.2,.OS,.OS,.OS)
80
_ _ _ SPRrt(.2,.OS,.OS,.OS,60)
70
- . - RST-lb(.2,5.7,60)
--*- RST-cb(.2,2.12,60)
60
50
40
30
20
10
0
.25
.2
.15
.1
Fig. 5.5 Expected sample sizes for Ho:
.05
p~.2
using SR(80)
The figures strengthen the interpretation of the results in Table 5.1. (The specific values
which are depicted in the figures are provided in the Appendix.)
To check whether these results are valid for other hypotheses with an underlying
Bernoulli or binomial distribution let us now consider q=.125. Let pO=.125, pi=.025, a.=.05,
~=.05,
c=4.5, b=2.12, and r=60. Table 5.2 provides the results of a 10000 replication Monte-
Carlo experiment for the new test problem:
Ho: P ~ .l25 versusHJ : p < .125
63
The power functions as wen as the expected sample sizes under SR(80) are illustrated
in Figure 5.6 and 5.7, respectively.
I sIze
. a 'proxImafIons t1or Hrn: )~.
T a hie 52
Power and sample
> 125
SPRT
SPRT-t
RST-Ib
.125,.025,.05,.05 ,(.125,.025,.05,.05,60J (.125,4.5,60l
RST-cb
(.125,1.98,60)
true p_
.125
.025
.125
.025
.125
.025
.125
.025
P[rej.HO]p
.048
.973
.043
.850
.063
.936
.051
.867
E[N]p
30.8
43.8
28.2
39.4
59.4
44.2
58.9
39.2
E[NIHo is rej.]p
44.6
44.2
37.4
36.9
51.0
43.1
37.7
36.0
E[NISR(60)]p
59.1
40.5
59.0
40.4
59.4
44.2
58.9
39.0
E[NISR(80)]p
78.2
43.2
78.2
43.4
78.2
45.5
77.8
41.9
E[NISR(100)]p
97.3
44.8
97.3
46.3
96.9
46.8
96.8
44.5
\fI(p)-lOO
---+- SPRT(.125 ..025 .05 ..05)
100
90
SPKft(.125 .025 .05 .05.60)
~ R<;Tlb(.I25.45.60)
80
70
-*- R<;Tcb(.125.1.98.60)
60
50
40
30
20
10
0
.15
.125
.1
.75
.05
.025
Fig. 5.6 Power functions for H 0:
P
p~.125
64
E[NISR(80)]p
8O-r------.....____
-+- SPKf(.125.025 .05.05)

_
10
SPRf-t(.125 .025 .05,.05.60)
- - . - RSf-lb(.125,4,s.60)
60
~ RSf-cb(.125.1.98.60)
50
40
30
20
10
O~------~----~-------+------~----~
.15
.125
.1
.075
.05
Fig. 5.7 Expected sample sizes for H 0:
.025
p~.125
using SR(80)
According to the illustrations in Figure 5.6 the SPRT seems to be more powerful than
the other tests while the conditionally expected sample sizes of the tests not substantially
different (cf Figure 5.7).
However, there is no test which outperforms the others regarding all chosen criteria.
The parameter values of the tests are chosen such that the type I error is less than or equal .05.
Obviously this is not difficult for the SPRTs but it is more difficult for the RSTs; where a
simulation is necessary to obtain band c. Thus we recommend the SPRT because:
It seems to be more powerful than the others.
The expected sample size is as well as in the other tests.
The determination of the test parameter values is easier than in the others
In our application, the absence of a definite upper bound in the stopping rule makes no
difficulty.
In order to analyze the SPRT in more detail we have investigated the effects of
variations in pO,pl,a,J3. Again a 10000 replication Monte-Carlo experiment has been performed
for the test problem:
Ho: P ~ .125 versus HI: p <.125
65
(Note, this test problem will be used to identify "bad" parameter subspaces in the solution
approach presented in the next chapter).
For the test problem the SPRT(po,pI,a.,I3) with pO=.125, pI=.025, and a.=13=.05 is
already analyzed. (The results are taken over from Table 5.2 in Table 5.3.) Now, we will
consider alternative surrogates, that is, pO=.15, and pI=.025. Using such surrogates for our test
problem we have to reduce a. (e.g. a.=.02), to avoid that the error of type I increases.
Furthermore, we will examine more narrow surrogates, that is, pO=.125, and pI=.075. We are
also interested in the effect of a larger 13. The results of a 10000 replication Monte-Carlo
experiment are given in Table 5.3.
Table 5.3 Power and sam21e size a22roximations for the SPRT{2,2 I ,a.,~};Ha: 2 ~ .125
21
a.
true 2
P[HO is rej.]p
E[N]p
E[NISR(80)]p
.125
.025
.05
.05
.125
.048
30.8
78.2
.125
.025
.05
.05
.025
.973
43.8
43.2
.15
.025
.02
.05
.125
.056
31.3
78.3
.15
.025
.02
.05
.025
.975
42.4
42.9
.125
.075
.05
.05
.125
.052
185.9
79.9
.125
.075
.05
.05
.025
71.5
67.4
.125
.025
.05
.10
.125
.056
24.3
77.7
.125
.025
.05
.10
.025
.945
40.2
42.0
no.
The increase of pO (=.15) has no substantial (positive) effect. If the null-hypothesis is

false (i.e. p=.025) there is a slight improvement ofE[NISR(80)1025 (from 43.2 to 42.9). But
the (approximated) probability to make on error of type I is increased to 0.56. (cf. no. 1 and
no. 3 in Table 5.3). Thus the results do not force to increase po.
If the difference between po and pI is relatively small, the expected sample size
increases drastically whereas the power becomes better (cf. no. 5 and 6 in Table 5.3). Such a
test may be less effective because a large number of executions of a stochastic algorithm is
necessary before the test performs a decision, i.e. reduces the parameter space.
66
An increase of the error type IT parameter value does not lead to a substantial decrease
in the power though the sample size becomes slightly smaller.
Thus there is no result which indicates to change the SPRT(.125,.025,.05,.05). We will
therefore use it for the parameter space reduction of stochastic algorithms. An example is
given in the following section.
5.4 An Example
In the preceding sections we have seen how a parameter space of a stochastic algorithm can be
reduced by a sequential test. A Monte-Carlo study resulted in a recommendation for the SPRT.
To motivate such a control of a stochastic algorithm we consider the Example 5.2 (where a
discrete parameter space is considered).
Example 5.2: Let Y E {I, 2}, 0
{I, 2}. So the following parameter space partition can
be chosen 0=LJ:=19b =LJ:=I(y,O)b ={(I,I),(l,2),(2,1),(2,2)}. We assume that a

stochastic algorithm SA(0) determines solutions for a minimization problem such that
the objective function value (cost function) is given by Zy ,O)b) = y
I:::l X~
where Xi
is uniformly distributed over the interval [0,1]. Furthermore, we assume that the
relationship between the objective function value and the parameters is unknown, so it
is necessary to start with all parameter subspaces.
Fig. 5.8 Sequence of parameter subspaces
67
The computation is restricted to #exc executions of SA(0). We consider two

algorithms, SAl and SA2 which are based on SA(0). In SAl each 9 h is chosen #exc/4
times whereas in SA2 the choice of parameter values is controlled by the
SPRT(.125,.025,.05,.05) (Ha: p
.125 versus HI : p <.125). Before applying the test
each parameter subspace is chosen 10 times, the resulting 40 objective function values
are sorted in ascending order and the 5-th value is taken as . 125-th quantile estimate.
The sequence in which the parameter subspaces are then considered is illustrated in
Figure 5.8. We start with 91> then we use 9 2, 9 3, 9 4 , 91> and so forth with respect to
the decision of the test; i.e. if the test rejects the null-hypothesis regarding 8 h then 8 h
will be eliminated from the sequence.
Let Z:~~c and Z;~x; denote the smallest objective function value of the solutions
which are determined within #exc executions of SAl and SA2, respectively. We repeat
this test 100 times. Thus Z:~xlc and Z;~~ are computed 100 times, denoted with Z:~'lc (i)
and
Z:~:; (i)
(i=1, ... ,100).
The
average
objective
function
values,
0z#exc = _l_"'\'loo z #exc(j) and 0Z#exc = _l_"'\'loozHexc(i) and the deviations in

SA2 100 "-"1=1 SA2 ,
SAl 100 "-"1=1 SAl
'
percent are given in Table 5.4.
Table 5.4 Average costs of SAl and SA2

0Z;~~ - 0Z:~'lc
0Z;~lc
0Z;:~
100
1.585
1.606
1.3%
200
1.429
1.253
-12.3%
400
1.266
.976
-22.9%
800
1.079
.838
-22.3%
1600
.912
.717
-21.4%
3200
.768
.607
-20.9%
#exc
HCXC
0ZSAl
For a small number (#exc=lOO) of executions the (expected) costs computed by

SAl and SA2 do not differ (substantially). Recall the expected sample size of the
68
SPRT(.125,.25,.05,.05) is about 30 but the sample size is 25 (per 8 h) if #exc=IOO.
Thus
the
parameter
space
will
not
be
reduced.
In
the
other
cases
(#exc=200, ... ,#exc=3200) the resulting values of SA2 are substantially smaller than
those of SAl. So the results of SA2 are much more better due to the integration of
parameter control (by a sequential test).
Thus, the example forces to incorporate a (sequential test based) parameter control if a
stochastic algorithm depends on the choice of parameter values.
5.5 Concluding Remarks
We introduced a method to control the choice of parameter values which turned out to
work very well if the stochastic process depends heavily on the parameter values. Let us now
give some concluding remarks concerning technical and other aspects:
I. If one or more hypotheses are rejected or accepted a new parameter space reduction could
be started on the remaining parameter space using a new partition. (In that case the SPRT-t
seems preferable to avoid problems regarding the indefinite upper bound of the sample size
of the SPRT.)
2. The method can be combined with others. For example, one can consider a parameter
subspace 8 h only with a certain probability which depends on the difference between the
minimum,
Z(8 h ), which has been found using 8 h and the global minimum
Z(0) =min {Z(8 h )Ih =1, ... , H} so that the higher the difference the smaller the probability
to consider 8 h.
3. Despite the independence of the objective function values, there is no other restrictive
assumption made concerning the problem type, the distribution of the objective function
value or the type of stochastic algorithm.
4. The method is not restricted to the choice of parameter values. The choice of decision rules
could also be performed by a sequential test.
Chapter 6.
A New Class of Stochastic Heuristics for the PLSP

and the CLSPL
In this chapter a stochastic backward add-method is introduced which is applicable for the
PLSP and, slightly modified, for the CLSPL. The heuristic starts at the planning horizon T and
steps backwards to the first period t=1. In each period (add-step) setup and lotsize decisions
are performed based on a so-called randomized regret-measure. The performance of the
heuristic depends on the choice of two real-valued parameters from a (partitioned) parameter
space. Therefore a sequential test controls the choice of the parameter values (cf Chapter 5).
In Section 6.1 we present a formal description of the backward add-method for the
PLSP. Section 6.2 gives recommendations how to modify the method for PLSP extensions.
This stochastic heuristic is modified for the CLSPL in Section 6.3. Section 6.4 treats the
integration of a sequential test based parameter control in a stochastic algorithm. Relations to
other local search methods are discussed in Section 6.5.
6.1
Biased Random Sampling via Randomized Regrets for the PLSP
In the following we describe a simple biased sampling (i.e. no pure random sampling) method
which essentially relies on randomized regrets (cf Drexl1991).
The symbols used in the following are defined in Chapters 2 to 4.
Demand is given in capacity units, i.e. we assume that pr1 0=1, ... ,1). The cumulative
demand (required capacity) ofitemj from period t to the horizon T which has to be satisfied in
the periods t, ... , 1 is denoted by
Djt :=max{o,t,(djt-qjt)}
For initialization purposes we define DjO =
0=1, ... ,J; t=1, ... ,T)
0=1,...,1) (cf below).
The total demand (total required capacity) which has (still) to be satisfied is specified by
70
J
TD:=LDj\.
j=\
For each itemj the next but last period less than or equal to t with positive demand is
nd jt := {
Ot
nd j.t-\
if t=O
ifd jt >0 O=1...,J; t=l, ... ,T)
otherwise
The cumulative capacity from period 't=1 to period 't=t is denoted by
(t=l, ... ,T).
Scheduling will be done backward oriented. If the machine is setup for the item i in
period t, item i will be scheduled until Dh =
('t
t). Hence, if the item i is scheduled over
more than one period, say from period t to period 't t), then Dit, and the demands di,t1>"', di;t will be scheduled. Then f= ndi;t-l is the next period with positive demand of item i
and we have to decide whether to preserve the machine state and "jump-back" from period 't to
f (as target period) in order to save setup costs or not. A jump-back is infeasible if TD > Ct"
However, if a new item j~ should be scheduled in a period t we consider only those items
where Djt>O because changing the setup state in period t to the item j with Djt=O, qjt will be
zero too, which may be a non-"optimal" decision (in most cases).
Ifwe assume that Ct is constant, then scheduling an itemj (as second item) in period t
(or f=ndj,t_l) and not in one of the earlier periods 't=t-1, ... , 1 (or f,f -1, ... ,1) saves at least hpjt
(or hpjr) holding costs for itemj. If the machine is not setup in period t for itemj, setup costs
ofSj are incurred. In such a case, the total "savings" ofitemj are at least hjDjt - sjBased on the setup state i of the machine in period t we define "savings" of scheduling
item j in period t as follows:
if (j:;t:i)!\(Djt >0)
case a)
if(j=i)!\(TD~Cnd.jl-l )
case b)
case c)
otherwise
O=l, ... ,J; 0~'Y~1)
71
Case a) occurs if the machine is not setup for the item j and there exists in period t
cumulated demand of the periods 't (=t, ... ,T) which has to be satisfied in the (earlier) periods
R,t-I, ... ,l. Case b) occurs if the machine is setup for the item j (i.e. j=i) in period t and a
jump-back to the target period ndj,t_l is feasible. Case c) occurs for the item j in period t if
Djt=O and the machine is not setup for item j in period t or a jump-back to the target period
ndj,t_Ileads to infeasibility. If the parameter y is near to one (to zero) then we expect that a
schedule will be computed where the lotsizes are relative large (small) and (no) "jump-back"
takes place.
As decision criteria we define
Pjt(i):= (rjt (i) - min {rkt (i )Ik = 1, ... , J 1\ rkt (i) > -00} +s
where 0 ~ 0,
I>
if rjt (i) =-00

otherwise
> 0,
which corresponds to the regret for producing item j not in period t or ndj,t_l> respectively.
Items which are "feasible" contain a positive regret (since
I>
> 0) and are "compared" with the
worst feasible alternative (i.e. the item with smallest finite savings). The larger the regret of the
itemj the more preferable it is to produce itemj (in the period t or ndj,t_l) where 0> 1 (0
<1) increases (decreases) the preferences. The regret of an "infeasible" item is zero.
The method for solving the PLSP incorporates the following three main ideas:
(i)
Lotsizing and scheduling is done backward oriented starting in period T.
(ii) In each period we add at most two items.

(iii) The regrets Pjt(i) are used in a randomized way to decide which item(s) should be
produced in a specific period.
A formal description of the stochastic backward add-method for the PLSP, denoted by
BAPLSP, may be given as follows:
72
IBAPLSP I
Initialization:
Yjt:=O 0=1, .. .1; t=O,I, ... ,T);
{in all periods the machine is not setup}
i:=O;
{the machine is not setup for any item}
t:=T;
{start at the planning horizon}
Qjt:=O O=I, ... ,J; t=I, ... ,T);

T
Djt := Ldj, O=I, ... ,J; t=I, ... ,T);
{all quantities of demand are unsatisfied}
't=t
]
{the total demand is unsatisfied}
TO:=LD jl ;
j=1
{start ofloop I}
while TD>O and t>O do

begin
determine Pjt(i) 0=1, ... ,1);
]
if LPjt (i)=O then t:=t-I

j=1
{backward step if no scheduling is required in period t}
else begin
{select and schedule an item}
choose j* at random proportional to Pjt(i);

i:=j*;
{the setup state changes to itemj*}
if Yit=1 then
{check whether jump-back to target period is required}
begin
Yi,t-I"'" Yi,ndu_I+I: =
{machine state preserving}
t:=ndi,t_l;
end;
while Dit > 0 do
begin
{start of loop 2; schedule selected item}
if LY jt = 0 then Yi(= 1 {determine setup state for (second) item in period t}

j=1
else Yi,t-l :=1;
{determine setup state for first item of period t}
qit:=min {Ct,Dit}
{determine production quantity}
Ct:= Ct - qit;
{update capacity}
Dj,;:=Dj,; - qit ('t=I, ... ,t);{update demand which has still to be produced}
TO := TO - %;
{update total demand which has still to be produced}
ifCt = 0 then t := t - 1; {backward step if no more capacity is available}

end;
{ end ofloop 2 }
ifLYj,t-1 = 1 then t:= t - 1; {backward step if two items are scheduled}

j=1
end;
{of scheduling the selected item}
73
end;
{end ofloop I}
lfTD > 0 then Z := 00

else evaluate solution;
If a jump-back for the item j occurs from period t to target period f=ndj,t_lt), we
preserve the machine state (i.e. Yj,t-l=...=Yjr=I). In the case where two items are scheduled in
a period t a step in the preceding period t-l is enforced . The capacity left over in period t
cannot be used which may lead to an infeasible solution (indicated by TD > 0 at the end of
BAPLSP). The objective function value of an infeasible solution will be set to infinity.
BAPLSP determines one solution of the PLSP per execution. #exc executions of
BAPLSP in general yield several (::;#exc) different feasible solutions. To see how BAPLSP
works we consider the following example.
Example 6.1: Let J=3, T=9, Ct=50 (t=I, ... ,9) and the other PLSP data as provided in
Table 6.1:
Table 6.1 Data ofExamEle 6.1

t
hj
Sj
d lt
20
80
450
d2t
40
50
10 3
300
d3t
30
30
90
150
Djt (as initialized in the beginning) and ndjt are given in Table 6.2.
74
Table 6.2 Initialised data

4
100 100 100
80
80
80
100 100 100
60
60
60
10
10
10
D3t
150 150 150 120 120 120
90
90
90
ndIt
nd2t
\ 6
nd3t
DIt
D2t
Let y=.5, /)=1 and &=1. The regret-measures for the items j=I, ... ,3 at the planning
horizon T=9 are determined as follows:
r19(0) = -00, since D19=O
=>P19(0)=O
r29(0) =0.5310 - 0.5300 = -135
=>P29(0)= -135 - min {-135,-30} + 1 =1
r39(0) =0.5190 - 0.5150 = -30
=>P39(0)= -30 - min {-135,-30} + 1 =106
Let P[G,t)] denote the probability to select the itemj in period t. Then
P[(1,9)] = 0, P[(2,9)] = 11107, and P[(3,9)] = 106/107.
So the probability that BAPLSP starts with item j=3 at the planning horizon is very
high.
In Table 6.3 a record of one execution ofBAPLSP is given:
75
Table 6.3 Record of one execution ofBAPLSP {y= .5, B=1 and 8=1}
t
ru(i)
r2t(i)
r3t(i) Pu(i) P2t(i) P3t(i)
-00
-135
-30
106
qit
Dil
Dit
TO
50
100
40
300
40
60
260
10
90
250
90
250
-00
-135
15
151
-00
75
-00
6
5
-145
-00
-60
-145
-00
-00
50
40
40
200
86
30
60
170
20
80
60
150
50
30
10
100
10
20
20
90
20
70
20
20
20
50
20
30
30
3
3
-00
-90
-60
31
2
2
-00
-00
-60
As expected, item j=3 is selected at first. However, in period t=8 item j=2 is
selected in spite of a very small choice probability, i.e. P[(2,8)]=1I152 and is scheduled
until the period t=7. In period t=7 item j=2 is selected again, which forces a jump-back
to period n~6=6; and so on (cf Table 6.3).
The solution in Table 6.3 of Example 6.1 is the best solution determined within 1000
executions ofBAPLSP. In each execution we have used the same parameter values as in the
Example 6.1 (i.e. y= .5, B=1 and 8=1). The corresponding objective function value is Z=1990
which has been obtained two times during the 1000 executions, only. Trying to get a better
solution, we have used other parameter values. This was done by choosing y from the interval
[0,9] and B form the interval [0,1] at random with uniform distribution before starting a new
execution ofBAPLSP. Within further 1000 executions ofBAPLSP the optimal solution (with
Z=1840) was reached 11 times, while the second best solution (with Z=1990) was computed 8
76
times. This small investigation indicates that the performance of BAPLSP heavily depends on
the choice of the parameters 'Y and B, i.e. on the weight of the setuplholding cost part as well as
on the randomization scheme. Therefore the choice of the parameter values ought to be
controlled by a sequential test (cf. Section 6.4).
The presented backward add-method can be modified by relaxing (ii) (cf. Section 6.3)
for the CLSPL. The next section provides some suggestions to modifY the method for PLSP
extensions.
6.2 Regrets for PLSP Extensions
The regret-measure in the BAPLSP is based on savings. In other potential applications these
savings may be different. Thus, to develop a backward add-method for a PLSP extension we
"simply" have to define an appropriate regret-measure.
The regret-measure rii) will be also used in an algorithm for the CLSPL (cf. Chapter
6.3). In the PLSP is Yit=l if i:;t:O. Thus, for the PLSP extensions we define specific regretmeasures where we use Yjt (or Yjrnt in the multi-machine case).
If a PLSPSDSC is considered SCji setup cost are incurred in period t if the setup state
changes from item j to item i, i.e. if Yit=l and Yj,t-l=l (cf. Chapter 4.2). Now let w.l.o.g.
Yj,T+l=O
0=1, ... ,1). So with respect to a backward oriented view a measure for the savings for
item j in period t is
J
ri~DSC:= (1- 'Y)h P it - 'Y ~)ciiY ~t+l

i=l
which is easy to integrate in BAPLSP, in that rjt(i) is replaced by rj~DSC.
More comprehensive. modifications are needed when integrating setup times and
multiples machines into BAPLSP.
77
To consider setup times we suggest as savings for itemj in period t
rj~:= (l-y l)hPjl -y 1((1-Y 2)SP -
Yj.I+I)+Y 2st j(1- Yj,1+1)
where 0 :;;; Yh Y2:;;; 1, and s~ is defined as in the PLSPST (cf Chapter 4.1).
In the multi-machine case the (machine dependent) production costs of item j on
machine m are kjrn. To produce item j on machine m and not on the most expensive alternative
machine m=M (buy-machine) saves Kjt=kjM - kjrn per unit. Thus the total savings to schedule
item j on machine m in period t are
rjrnt:=(1-Y)(Kjt+ hj)Djt - Y Sjrn (1-Yjrn,t+l)
ifm<M, and
rjMt:=(1-y)hPjt - YSjM
ifm=M
where Yjrnt and Sjrn are defined as in the PLSPMM (cf Chapter 4.3).
Capacity aspects enforce to substantial changes of BAPLSP if setup times are

significant or multiple machines are available. For instance, the capacity check is more difficult
if setup times are considered because the required capacity depends on the unknown number of
setups. The machine selection for the items makes the decision problem more complex in the
multi-machine case. However, the basic ideas ofBAPLSP will not change. So it must not be a
difficult task to develop an algorithm for such extensions. An extension for the multi-level case
is introduced in [Kimms 1993].
In the next section we will see how these ideas are integrated in an approach for the
CLSPL.
78
6.3 Modifications for the CLSPL
Since in a CLSPL the number of items producible per period is not restricted we have to
modify BAPLSP for the CLSPL by relaxing the assumption that "In each period we add at
most two items" (cf. Chapter 6.1).
A formal description of such a modification, denoted by BACLSPL; is given in the
following:
IBACLSPLI
Initialization:
Xjt:=O (j=I, ... ,J; t=O,I, ... ,T);
{in all periods the machine is not setup}
,t:=O (j=I, ... ,J.; t=O,I, ... ,T);

i:=O;
{in all periods no linking of quantities is performed}

{the machine is not setup for any item}
t:=T;
{start at the planning horizon}
qjt:=O (j=I, ... ,J; t=I, ... ,T);

T
Dj\ := Ldj't (j=I, ... ,J; t=I, ... ,T);
{all quantities of demand are unsatisfied}
't=t
{the total demand is unsatisfied}
TD:=LD jl ;
pi
while TD > 0 do
{start of loop I}
begin
determine Pjt(i) (j=I, ... ,J);
J
{backward step if no scheduling is possible in period t}
ifLPj\ =0 then t := t - 1
j=1
else begin
{select and schedule an item}
choose j* at random proportional to Pjt(i);

i:=j*
ifXit=1 then
begin
Zjt = Zj,t-l = ... =
{check whether jump-back to target periodis required}

{machine state preserving}
Zi,ndu_,+1
Xi,t-l = Xi,t-2="'=
t
= 1;
xi,nd,t_,+1
=1;
:= ndi ,t-l;
end;
while Dit > 0 do
{start of loop 2; schedule selected item}
79
begin
Xit = 1;
{fix setup state for item in period t}
qit := mint Ct,Dit }
{determine production quantity}
{update capacity}
Ct := Ct - %;
Di't=Dh - qit ('t=1, ... ,t); {update unsatisfied demand of selected item}
TD:=TD - %;
{update total unsatisfied demand}
ijCt = 0 then
begin
t := t - 1;
{step backward if no more capacity is available}
ijDit >0 then Zit:=1; {fix "linking" variable}

end;
end;
end;
{end of loop 2}
{of scheduling the selected item}
. {end ofloop I}
end;
IfTD > 0 then Z := 00

else evaluate solution;
The linking variable Zjt for item j in period t is 1 if a lot size is scheduled over more than
one period or a jump-back occurs. Note, if a feasible solution exists, in each execution of
BACLSPL a feasible solution will be computed.
It should be noted here that backward lotsizing and scheduling methods in the
capacitated case are from a conceptual point of view superior to forward-oriented ones (e.g.
the famous Dixon, Silver 1981 heuristic): There is no need for complicat~d and timeconsuming look ahead procedures in order to secure resource feasibility.
6.4 Integration of Parameter Control via Sequential Analysis
The Example 6.1 in Section 6.1 indicates that the regret based backward add-method heavily
depends on the choice of the parameter values for y, o. (Note the parameter
will be fixed to
one in the following). To integrate a sequential test we define a partitioned parameter space 0.
80
The parameter 'Y will be taken from the interval [0,1]. The smaller a positive 5 is chosen (less
than one) the more similar the regrets of the items are; if we choose 5=0 all regrets equal one.
On the other hand the larger 5 is taken, the more larger the probability is to select the item(s)
with maximum regret. To determine a manageable parameter space, we restrict 5 to be less
than 9, i.e. 5 e[0,9]
Before performing an execution ofBAPLSP (or BACLSPL) we have to decide from
which subspace 9 j (i=I, ... ,H) the parameter values for 'Y and 5 are randomly chosen. We start
with 9 1, Then we consider the subspaces 9 2, 9 3, ... ,9H, and start again with 9 1 (and so on).
During the executions the sequential test eliminates one or more subspaces from the sequence
(parameter space reduction). For example, if the subspace 9 5 is eliminated the new sequence
will be 9}. ... ,94,96, ...
Let ii be the pre-sample size per subspace 9 h (h=I, ... ,H) for estimating the .125-th
quantile. We use t~e SPRT(.125, .025, .05, .05) (cf. Chapter 5), to decide whether a subspace
is to eliminate or not. The .125-th quantile will be estimated by the r.125. ii Hl-th smallest
value from the objective function values of the first ii H solutions determined with BAPLSP
(orBACLSPL).
Executing BAPLSP (or BACLSPL), using 9 h, n times reveals n solutions. Let uh
denote the number of objective function values of the corresponding solutions which are
smaller than or equal the estimated quantile
~.
The SPRT(.125, .025, .05, .05) regarding to 9 h
(h=l, ... ,H) stops at

N h= min {nlLRhn E(A,B)}=min{ nlln(LRhn ) E(ln(A),ln(B)}, where .
A=aI(l-J3)=.053, and
B=(I-a.)/J3= .95/.05 = 19,
LRhn = (pI/pot (1- pO)/(l- pl)t-n =(. 025f. 125th ((l-.125)/(I-.025)th - n
=> Nh=min{ nl-1.61u h +.108(n-uh ) E(-2.94,2.94)},
Thus the subspace 9h will not be considered (= rejected) anymore if In(LRmJ
;?!
In(B)
where the transformation by the monotonous increasing logarithmic function lnO simplifies the
computation ofNh' (Note, each 9h will be considered at least 2.94/.1081 = 28 times).
81
A formal description of the stochastic backward add-method with integrated parameter

control, denoted by MAIN, may be given as follows:
MAIN
Input:
9 1,0 .. ,9H
{subspaces}
{noo of executions}
{pre-sample size per subspace}
{stochastic algorithm using y, o}
#eKc
Ii
SA(y,o)
Quantile estimation:
{sequence of subspaces}
SEQ := 9 1 9H, 9 1
Z:=oo;
{best objective function value}
h:=I;
for i:=1 to Ii H do
begin
{select parameter values at random with uniform distribution}
y,o E 9h
{execute BAPLSPIBACLSPL using y, o}
Zf= SA(y,o)
ijZi < Z then
begin
Z:=Zi;
memorize solution;
end;
h:= index of the next parameter subspace in SEQ;
end;
sort in nondecreasing order the objective function values Z\' ZiioH;
~:= Zr.125.fi Hl;
,000,
,000
0 0 0'
Uh:=O (h=l,ooo,H);
Determine test data:
{the sample size of each subspace}
n:=Ii;
h:=I;
{determine noo of objective function values}
for i:=1 to Ii H do
{less than or equal the quantile}
begin
ifZi S; ~ then uh:= uh + 1;
h:= index of the next parameter subspace in SEQ;
end;
Sequential test based parameter control:
0
82
for i:=ii H +1 to #exc do

begin
y,6 e 9h
Zj:= SA(y,6)
ifZj < Z then

begin
Z:=Zj;
memorize solution;
end;
ifZj:;;; I; then uh:= uh + I;
if -1.61uh + .108(n-u~:?!: 2.94 then eliminate 9 h from SEQ
else if -1.61uh + .108(n-u~:;;; -2.94 then uh:= 00;
{reject null-hypotheses}
{accept null-hypotheses}
ifSEQ={} then
begin
h :=1;
SEQ := 91> ... , 9R, 9 1, ... ;
end else h := index of the next parameter subspace in SEQ;
if~h=first subspace in SEQ
then n:=n+1;
{update sample size}
end;
output: memorized solution;

SA(y,6)
(=SA(9~)
denotes that the stochastic algorithm, i.e. BAPLSP or BACLSPL,
computes one solution using y,6 (=(9~); the objective function value of the solution of the ith execution is Zj. The objective function values ZI, ... ,Zfi.H determined during the quantile
estimation have to be memorized in order to use their information for the sequential test (cf.
Determine test data). The parameter n (:?!: ii) denotes the sample size which is identical for all
remaining 9 h (h eH). If the null-hypothesis corresponding to a subspace 9 h is accepted after
the n-th execution of SA(9~, i.e. LRhn:;;;A, the counter uh is set to infinity to avoid that for an
i>n LRhj=B. Thus the decision to accept the null-hypothesis will not change in later executions
(i>n) and therefore the type I error probability does not increase.
83
6.5 Relations to Local Search Methods
Local search provides extremely powerful methods for solving hard (combinatorial)
optimization problems to sub-optimality (for the complexity oflocal search see [Johnson et al.
1988] as well as [Yannakakis 1990]. In the following we will relate the methods presented here
especially to simulated annealing, tabu search as well as genetic algorithms.
In general, local search procedures consist out of an exchange and an acceptance

component. Starting from a current solution, the exchange (or generation) component
provides neighbouring solutions. The acceptance component decides whether to accept the
new solution(s) or not. From a conceptual point of view an exchange (or interchange) method
consists out ofa sequence of add and/or drop operations. Omitting one of two reveals add and
drop to be special cases of local search. Thus BAPLSP (or BACLSPL) essentially is a local
search method.
simple random
guided random
det. acceptance
prob. acceptance
stochastic process
Fig. 6.2 Relations to other local search methods
In Figure 6.2 we first distinguish local search methods to be either deterministic or
randomized exchange procedures. Second, both types of exchange methods may be either
combined with a deterministic or with a probabilistic acceptance method. Exchange and/or
acceptance probabilities may either be modified or used without modification. Randomized
exchange and/or acceptance produces a sequence of objective function values corresponding to
a stochastic process. Note that deterministic exchange in combination with deterministic
84
acceptance is likely to be trapped in local optima. With respect to threshold accepting, in
Figure 6.2 an additional arc between "rand. exchange" and "det. acceptance" could be
introduced [cf. Dueck, Scheuer 1988].
Randomized exchange procedures may either be of a simple or guided random type.
While in the former case each neighbour has the same probability of being chosen, in the latter
neighbours are chosen with a probability being controlled by some method. Thus BAPLSP (or
BACLSPL) in conjunction with sequential analysis (for controlling parameter subspaces)
corresponds to guided random exchange methods. (Up to our knowledge the term "guided
random search" has been introduced in [Hartmann, van Hee 1992] in the context of Markov
decision process-based search).
In simulated annealing methods (cf. e.g. [Johnson et al. 1989, 1991], [van Laarhoven
et al. 1992], [Martin et al. 1992]) the sequence of solutions does not converge monotonically
towards a local optimum. Rather, solutions walk up and down through the feasible set.
Starting from a current solution, first a neighbouring solution is selected, either in a
deterministic or in a random way. Then, based on the induced change of the objective function
value, this transition is either accepted or rejected with a probability being controlled by some
parameters. Clearly, the selection of neighbouring solutions may be based on randomized
regrets (related to the respective changes of the objective function values) by making properly
use of the sequential analysis motivated controlled search in the parameter space as described
above.
In tabu search (cf. e.g. [Glover 1989, 1990], [Hertz, de Werra 1990] one of the main
ideas is to guide the local search process deterministically out of local optima. This can be
done using several criteria, which have to ensure the increase in the objective function value
not being too high or being compensated later on. One of the essential ingredients of tabu
search is to maintain throughout the search a tabu list (of forbidden transitions between
neighbouring solutions) in order to prevent from cycling. Thus basically tabu search is some
kind of deterministic exchange as well as deterministic acceptance method. In addition,
randomized exchange and/or acceptance may be incorporated into tabu search as well, see e.g.
85
[Faigle, Kern 1992] as a recent reference to probabilistic tabu search. To be more specific, the
selection of neighbouring solutions may be based on randomized regrets once more - thus
keeping track of the ideas presented here for tabu search as well.
Genetic algorithms (see e.g. [Miihlenbein et al. 1988], [Goldberg 1989], [Liepins,
Hilliard 1989], [Dorndort: Pesch 1992] as well as [Pesch 1993]) are mo~vated by the analogy
of evolution. They have been designed as general search methods working on populations of
feasible solutions in order to make use of properties which "good" solutions have in common.
The three basic operators used by genetic algorithms when constructing a new population are
reproduction, crossover as well as mutation. For reproduction an individual (coded as a
string) is copied with probability proportional to its fitness value. Then crossover is applied to
randomly partitioned pairs of the new population by choosing randomly a position of the string
and exchanging the tails. Random changes to single elements of an individual, i.e. mutation,
comprises the last operating within genetic algorithms. Thus genetic algorithms may be seen to
be randomized exchange methods, where the exchange process is guided via fitness values, in
combination with deterministic or probabilistic acceptance. Clearly, fitness based reproduction
of individuals may be replaced by taking randomized fitness regrets, which is closely related to
scaling and windowing. In addition, the selection of pairs of the new population once more
may be based on randomized fitness regrets. Last but not least, the choice of the position of the
string may be based on randomized regrets. Thus genetic algorithms bear a wide range of
potential applications of the ideas presented here.
Note that our concept of guided random exchange is closely related to the
"modification of generation probabilities" [Faigle, Kern 1992] as follows: Let us assume that
the parameter space defines in advance the whole set of probabilities for exchanging
neighbouring solutions. Then, ceteris paribus, w.r.t. the subspace h exchange takes place with
some given probabilities (defined by
a~.
Changing from subspace
ah
to subspace
ak
conceptually is a method of "modification of generation probabilities". Thus the controlled

search within the parameter space via sequential analysis as described above essentially is a
general method of modifying exchange (or generation) probabilities.
Chapter 7.
Computational Results
In Chapter 2 and Chapter 3 we have introduced the CLSPL and the PLSP, respectively. A
stochastic heuristic applicable for the PLSP (BAPLSP) and, slightly modified, for the CLSPL
(BACLSPL) is presented in the preceding chapter. Since it is supposed that the approach
depends (heavily) on two parameter values, a sequential test based parameter control is
provided.
This chapter investigates the performance of the algorithm by a computational study.
Since no tailor-made optimal solution approach for the PLSP (as well as for the CLSPL) exists
we compute optimal benchmark solutions with the standard mixed integer programming (MIP)
solver LINDO [cf Schrage 1986]. The optimal objective function value of such a reference
solution is compared with the objective function value of the best solution which is computed
with the heuristic. Only small instances of the PLSP with up to 50 binary variables are solvable
in reasonable time with LINDO. For the CLSPL some CLSP (medium size) instances exists in
the literature which are solved to optimality as CLSP (and as DLSP) and transformed into a
(non-optimal) CLSPL solution.
Another inquiry is on the effect of the integration of the sequential test based parameter
control.
The performance of a heuristic may depend on certain problem characteristics. Hence,
we describe an algorithm for the generation of PLSP's as well as CLSPL's with specific
properties.
The instance generator is described in the following section. The computational results
of the PLSP are given in Section 7.2 where especially the parameter control is analyzed. A
brief performance investigation of the backward add-method for the CLSPL is presented in
Section 7.3.
We will see that backward add-methods with sequential test based parameter control
are very efficient for lotsizing and scheduling.
87
7.1
An Instance Generator
The difficulty of a lotsizing and (semi-) scheduling problem may depend on several criteria. So
it is interesting to explore the performance of an approach applied to problems with different
properties. In order to perform such a study we require problems which possess certain
criteria. As relevant criteria (parameters) we will consider
number of items (J)
number of periods (T)
holding costs (hc)
time between order (TBO)
demand pattern (dp)
capacity pattern (cp)
capacity utilization (U)
which are explained in the following.
Number of items
Usually, the more items we consider the more complex it is to solve a PLSP or CLSPL to
optimality due to the increase of the number ~f binary variables. Hence, we will analyse the
sensitivity of the performance of the backward add-method regarding to the numbers of items
1.
Number of periods
We suppose that the number of periods T has an effect on the performance of the backward
add-method. In [Fleischmann 1992] it is distinguished between macro-periods t m (=l, ... ,Tm)
and sub-periods t S(=l, ... ,T). A macro-period is a "large" period (e.g. week) which is split up in
"small" sub-periods (e.g. days or shifts) (cf. Chapter I). We assume that the number of subperiods per macro-period TS is constant. Thus the number of periods is T=Tm"Ts In a PLSP
the number of different items producible in the first macro-period is less than or equal TS due
88
to the assumption A3 (cf Chapter 3). Therefore, 'fS
J must hold (if the items j =I, ... ,J are
demanded in tlll::l).
Holding costs
Let Pj be the capacity used to produce one unit of the item j. To keep one unit of item j in
inventory causes ii j holding costs per period. Since we assume that the demand djt is given in
capacity units the holding costs per capacity unit of the item j are hriij 1Pj. So the holding
costs coefficients hj 0=1, .. ,1) are usually unequal if the production speeds or the capital tied up
in itemj per unit are varying. Therefore for our instances h:i O=I, ... ,J) will be chosen at random
from the set {l,2, ... ,hc} with uniform distribution where hc is a positive integer. Ifhc = 1 then
hrl for all j (=I, ... ,J).
Time between order

Let TBO (>0), and i\TBO (<TBO) be positive integers. Define A as a set of integers as
follows A={TBO-i\TBO, ... ,TBO+i\TBO}. We determine the (approximated) time between
order of item j, denoted by TB0j. at random from A with uniform distribution. The average
demand ofitemj per period is d j = IITL~=ldjt. The holding costs ofa lot size with a length of
TBOj periods can be approximated by djhjL:~j("t-I). Now, let us assume that (on the
average) the approximate holding costs are equal to the setup costs. Then the rounded setup
costs of the item j are
sJ = rdJ h.TBO.(TBO
J
J
J -1)/21
0=1, ... ,1).
TBO influences the size of lots which are profitable. i\TBO determines the (expected)
differences between the time between order of the items TBOj 0=1, ... ,1). Therefore the larger
i\TBO is chosen the larger the (expected) differences between profitable lotsizes of the items
are.
89
(Note in [Salomon 1990] Sj = djhjTBO~ /2 is proposed which is derived from the so
called economic order quantity (EOQ-) formula for single level problems with constant
average demand and contimlOstime horizon [cf. Harris 1913]).
Demand pattern
The (cumulated) capacity of a macro-period t m is Cn;,
= ~T'T"lm
m C I . We assume that in a
I
"-'1= {I -1)+1
macro-period t m no more than one positive integer demand of the item j occurs, denoted by
dm
jlm where the superscript m stands for macro. The demand pattern may have an effect on the
performance of a heuristic. Therefore, we will consider the following types of demand patterns
td e{ud,pt,ds}:
ud
In the macro-period t m the one demand d m

jl m of the itemj will be generated at random
with uniform distribution so that d;m e{o, ... ,r 2C~ / J l} (j=I, ... ,J;tm=I, .... Tm).
pt
The pattern of the randomly generated demand of the itemj will have a positive trend.
i.e. in the macro period t m the demand of the item j will be generated with uniform
distribution so that d;m e{ o, .... ,r 4 t mC~ / (1 . Tm) l} (j=l, ... ,J;tm=I, ... ,Tm). For
example. if C~ is constant for all t m then the expected increase per macro-period is
E[d m
m 1 -dmm]
= r 2 .Cn;,
/(1. Tm)l
.I.t +
Jt
t
ds
The demand will be generated at random with uniform distribution whereby the
expected values of the demand of the items are of different seize; that is. in the macro
period tmthe demand of the itemj will be generated with uniform distribution so that
d;m e{o, .... ,rj-4-C~ / (J.(1 +
1))1} (j=I, ... ,J;tm=I .....Tm). Thus the differ~nce
between the expected demand of item j and item j+ 1 is in the macro-period t m

E[d ml m-dmm] = r 2 .Cn;, / (1.(1 +
J+ ,I
Jt
11
Let tS(tm.j) be the (sub-) period in which a positive demand of item j in the macro-period t m
d m ift=t"(t m .)
. 'J
occurs. Then. it is d jl = { JIm
o
otherwise
where tS(tm.j) depends on the parameter d m e {e.i} as follows:
e
Demand occurs at the end of the macro-period tm. i.e.
t~
= tm.p.
90
If tm<Tm the demand occurs at the end of a macro-period or TS-l (sub-) periods later
which will be determined at random with uniform distribution. In the last macroperiod tltl::Tm the demand will always occur in the last (sub-) period T=TmTs. Thus,
ts(tm,j)
{tmTS,tmTS+ 1, ... ,min{ (tm+ I)TS-l,T}}.
Hence, no demand occurs before the end of the first macro-period. This avoids that
more items are demanded than being producible in the (sub-) periods t=I, ... ,TS-l.
Remark: The demand djt will be generated according to the fact that the cumulated demand
must not be greater than the cumulated capacity, i.e. the following inequality holds:
T
L~:Ait::;LCt
t~l
rl
t~l
In the following a specific demand pattern is denoted by dp = (td,dm); e.g. dp=(pt,e)

means that a demand with positive trend is randomly generated whereby a demand occurs at
the end of a macro-period.
Capacity pattern
We suppose that the capacity pattern may have an effect on the performance of the scheduling
methods. Two types of capacity pattern cp will be considered:
co
The capacity stays the same in each period (constant pattern): Ct= 100 (t=I, ... ,T).
ic
The capacity is increasing:
100
Ct
ift::;fT/31
150 iffT/31<t::;f2.T/31
200 iff2. T 131 <t::; T
91
Capacity utilization
The scarcity of capacity may have a substantial effect on the performance of a heuristic.
Therefore we will consider instances with different capacity utilization U which is defined as
follows:
U=(tt.djt)/(tCt)-B (O<U~I; 0~B ~y~Ct)
Definition 7.1: INST(J,Tffi,Ts,hc,TBO,dTBO,dp,cp,U;n) is an algorithm which generates n

(>0) lotsizing and scheduling problems which possess the following criteria
the number of items is J,
the number of periods is T=Tffi. TS,
the randomly generated holding costs coefficient ofitemj is hj E{I,2, ... ,hc} (j=I, ... ,J),
the randomly generated approximated time between order of the item j is
TBOj
E{TBO-dTBO, ... ,TBO+dTBO} and the setup costs are

Sj
=fdjhjTBOj(TBO
-1)/21 (j=I, ... ,J),
the randomly generated demand has a pattern dp Etdxdffi = {{ud,pt,ds} x {e,i} }
the capacity pattern is CpE {co,ic},
the capacity utilization is U.
Note,
INST(J,Tffi,Ts,hc,TBO,dTBO,dp,cp,U;n)
generates
(usually)
different
problems.
Exampel 7.1: INST(3,2,4,1,6,0,(ud,e),co,0.9,1) will generate one lotsizing and

scheduling problem with J=3 items, T=8 periods, a constant capacity per period Ct= 100
(t=I, ... ,8), and (as one example) the following demand djt, the holding costs
coefficients hj and the setup costs Sf
92
dj1
dj2
~3
dj4
djS
dj6
dj8
dj7
j=1
23
278
j=2
58
46
j=3
149
166
The average demand of item j=1 is (23+278)/8=37.625. Since
hj
~TBO=O
Sj
565
1
195
591
it is TB0 1=6.
So the setup costs ofitemj=1 are

~ = f37.625.16(6-1)/21
=f564.3751 =565
as provided in the table above. Furthermore, the capacity utilization is

U = (23+58+149+278+46+ 166) 1 (8100) = 720 1 800 = 0.9
as required.
In the next section the instance generator is used to analyze the performance of
BAPLSP.
7.2
Computational Results for the PLSP
In this section the performance of BAPLSP with and without parameter control is analyzed.
The computation time as well as the deviation of the best objective function value from the
objective function value of the optimal solution are valid performance characteristics.
Let BAPLSP_SEQ denote that MAIN is performed using BAPLSP where Ii = 8, H=9,
and
9 1 = [ 0, .33) x [0,3),
92 = [ 0, .33) x [3,6),
9 3 = [ 0, .33) x [6,9]
94 = [.33, .66) x [0,3),
9 s = [.33, .66) x [3,6),
96 = [.33, .66) x [6,9]
9 7 = [.66, 1] x [0,3),
98 = [.66,1] x [3,6),
99 = [.66, I]
x [6,9]
93
a), ... a9
Thus
are of same size and in the first execution ofBAPLSP in MAIN is
y e[0.33) and 0 e[0.3).
Let Z~sP denote the optimal objective function value ofa PLSP (as in Chapter 3).
Definition 7.2: Z~sP and Z:~~sP _SEQ denote the objective function value of the best solution
which has been found within #exc executions ofBAPLSP without and with parameter control,
respectively.
Definition 7.3: The percentage deviation ofZ~spand Z:;LSP _SEQ from Z~LSP ofa PLSP is
&:;LSP: = 100 (Z;~~sP - Z~sP) I Z;LSP

and
A7#/e,e _SEQ'. = 100 . (-z#exe
ULBAPLSP
BAPLSP _SEQ - zPLSP )/Z'PLSP
respectively.
Definition 7.4: The percentage improvement by using sequential test based parameter control
is denoted by
exc
A7#1exe. 100 (-z#<xe
ULSEQ
.=
. BAPLSP - -z.,e
BAPLSP _SEQ )/-ZBAPLSP
Definition 7.5: The percentage fraction of infeasible solutions determined in #e?<-c executions
ofBAPLSP without and with parameter control is

INF~;;LSP:= 100(no. of infeasible solutions using BAPLSP)/#exc
and
INF::;LSP _SEQ: = 100(no. of infeasible solutions using BAPLSP_SEQ) I #exc
respectively.
For 16 parameterizations ofINST(J,Tm,TS,hc,TBO,LlTBO,dp,cp,U;S) we generate n=5

instances
each
(cf
Table
7.1).
We
start
with
the
basic
parameterization
94
INST(3,5,3,1,8,0,(ud,e),co,.8;5) (cf. SOOI to S005 in Table 7.1). Then we consider a larger

and a smaller capacity utilization, i.e. U=.95 and U=.65 whereby the other parameters are kept
as in the basic parameterization (cf. SOIl to SOlS and S021 to S025 in Table 7.1). The
instances S031 to S 105 also differ in only one parameter position from the basic
parameterization. In S III to S 115 the number of sub-periods per macro-period is increased
from 3 to 4 and the number of macro-periods is decreased from 5 to 4. Thus the problems
contain less than 50 binary variables (=JT) which is necessary to determine an optimal PLSP
solution within reasonable time
(~
2 hours on a Personal Systeml2 Model 70 486 using
LINDO). In S121 to S125 the number of items is decreased from 3 to 2 and the number of
macro-periods is increased from 5 to 12. (So the number of binary variables of each of these
problems is less than 50). In S 131 to S 155 more than one parameter "values" of the basic
parameterization are changed. Such problems are supposed to be more complex than the
preceding ones.
95
Table 7.1 8mall instances by IN8Tp,Tm,TS,hc,TBO,8TBO,d~,c~,U;n}
J
Tm
TS
8001 to 8005
8011 to 8015
8021 to 8025
8031 to 8035
8041 to 8045
d~
c~
(ud,e)
co
.8
(ud,e)
co
.95
(ud,e)
co
.65
(ud,e)
pt
.8
(ud,i)
co
.8
8051 to 8055
(pt,e)
co
.8
8061 to 8065
(ds,e)
co
.8
8071 to 8075
(ud,e)
co
.8
8081 to 8085
12
(ud, e)
co
.8
8091 to 8095
(ud,e)
co
.8
8101 to 8105
(ud,e)
co
.8
8111 to 8115
(ud,e)
co
.8
8121 to 8125
12
12
(ud,e)
co
.8
8131 to 8135
(ud,e)
co
.8
8141 to 8145
(ds,i)
pt
.8
8151 to 8155
(~t,Q
co
.8
he
TBO 8TBO
Each (small) problem (cf. Table 7.1) has been solved by BAPL8P and BAPL8P_SEQ
using #exc=1000 each. Table 7.2 provides the average deviation between the optimal objective
function value and the objective function value of the best solution of BAPLSP and
BAPL8P_8EQ, respectively (average of 5 instances per entry). Furthermore the average
fraction of infeasible solutions is given in Table 7.2.
96
Table 7.2 Average deviation from optimum and average fraction of infeasible solutions
-]1)00
0LiZ~~p
0LiZ BAPLsP _SEQ
0INF~~
0INF~~LSP _SEQ
SOOI to S005
18
14
SOlI to SOlS
.76
.76
50
41
S021 to S025
12
10
S031 to S035
.25
.25
29
22
S041 to S045
.06
S051 to S055
.73
.05
19
S061 to S065
13
S071 to S075
10
S081 to S085
46
34
S091 to S095
.224
.98
12
S101 to S105
2.4
.75
28
23
SIll to S115
S121 to S 125
.08
.08
S131 to S135
.03
24
21
S141 to S145
1.06
.057
S151 to S155
.01
.01
In all 16 parameterizations with 5 instances the average deviation from optimum is less
than one percent for the BAPLSPL_SEQ. The objective function value of the solution
determined with BAPLSP (BAPLSP_SEQ) for the problem S 105 deviates from the optimal
objective functions value by 11.9 (3.6) percent. Such a poor result occurred only once in our
test. Thus the solution quality for small instances is very high. In the total average (all 165
instances) Z~~I.SP is .12% larger than Z~':.sP SEQ' i.e. the integration of parameter control has
improved (slightly) the solution quality. Moreover, the fraction of infeasible solutions is
substantially decreased due to the integration of parameter control (cf. Table 7.2).
In order to analyze the effect of the sequential test we have generated 2010 large
lotsizing and scheduling problems with specific criteria as provided in Table 7.3. Again we
97
start with a basic parameterization, this time we use INST(6, 10,8, 1,30,0,(ud,e),co,.8;10). Then
we change successively the capacity utilization, the capacity pattern, and so on. For the
problems LI70 to LI99 more than one parameter "value" differs from the basic
parameterization.
Table 7.3 Large instances by INST{ J, Tm, TS, hc, TBO,
J
Tm
TS
LOOO to L009
10
LOIO to LOl9
L020 to L029
~TBO,d2,c2,U;
n}
TBO
~TBO
d2
c2
30
(ud,e)
co
.8
10
10
30
(ud,e)
co
.95
10
10
30
(ud,e)
co
.65
10
L030 to L039
10
30
(ud,e)
ic
.8
10
L040 to L049
10
30
(ud,i)
co
.8
10
L050 to L059
10
30
(pt,e)
co
.8
10
L060 to L069
10
30
(ds,e)
co
.8
10
L070 to L079
10
30
15
(ud,e)
co
.8
10
L080 to L089
10
60
(ud,e)
co
.8
10
L090 to L099
10
15
(ud,e)
co
.8
10
LI00 to LI09
10
30
(ud,e)
co
.8
10
LllO to LI19
10
12
30
(ud,e)
co
.8
10
LI20 to LI29
10
30
(ud,e)
co
.8
10
LI30 to LI39
30
(ud,e)
co
.8
10
LI40 to LI49
15
30
(ud,e)
co
.8
10
LI50 to LI59
10
30
(ud,e)
co
.8
10
LI60 to LI69
10
10
12
30
(ud,e)
co
.8
10
LI70 to LI79
25
50
20
(pt,i)
co
.9
10
LI80 to LI89
10
11
50
20
(ds,i)
pt
.75
10
LI90 to LI99
10
30
{ds,Q
co
.8
10
hc
98
Each instance of Table 7.3 has been solved with BAPLSP and BAPLSP_SEQ. The average
-1000
-1000
1000
percentage deviations between ZBA)'LSP and ZBAPLSP _SEQ' denoted 0.1ZSEQ ' and the average of
INF~~sp and INF~~sp _SEQ are provided in Table 7.4 (10 instances per entry).
Table 7.4 Large instances by INST( J, Tm, TS, hc, TBO, .1TBO,dp,cp,U; n)
0.1Z~:
0INF~:LSP
0INF~~sp SEQ
LOOO to L009
1.16
24.6
16.3
LOIO to L019
.67
33.2
26
L020 to L029
0.94
44.3
14.6
L030 to L039
1.81
31.7
14.7
L040 to L049
.34
10.3
5.7
L050to L059
1.61
46.1
31.3
L060 to L069
.86
15.4
13.5
L070 to L079
.51
18.8
13.7
L080to L089
1.07
17.9
16.8
L090 to L099
1.14
12.4
4.8
L100 to Ll09
1.04
21.4
17
LlI0 to Ll19
1.27
13.1
7.1
Ll20 to Ll29
.88
21.4
16.7
LBO to L139.
.98
3.9
3.9
Ll40to Ll49
.78
16.6
10
Ll50 to Ll59
.72
Ll60 to Ll69
.55
44.1
26
Ll70 to Ll79
.12
41
39.8
Ll80 to Ll89
.76
Ll90 to Ll99
1.02
.9
10.6
9.3
Independently of the chosen parameterization the integration of parameter control has

increased the solution quality (on the average). The improvement is not as large as in the
99
fictitious Example 5.1 (cf. Chapter 5). This may be reasoned by the efficiency of the heuristic:
The results for the small instances (cf Table 7.2) indicate that the best objective function value
of the BAPLSP is near the optimal solution. It is thus difficult to determine a less costly
solution. Whereas a heuristic which generates a solution with an objective function value which
is much larger than the optimal objective function value may easily be improved. Let us assume
that the parameter control improves a heuristic. Then a small improvement indicates that the
basic heuristic is very efficient. Thus BAPLSP seems to be efficient for large problems as well,
and does not depend on the problem criteria.
Let us consider the problem LOOO in more detail. In Figure 7.1 the sorted objective
function values of 1000 executions ofBAPLSP and BAPLSP_SEQ are plotted.
Z
300000
280000
260000
240000
--BAPLSP
- - - BAPLSP_SEQ
--- ---
220000
200000
180000
160000
200
400
600
800
1000
Fig. 7.1 Sorted objective function values computed by BAPLSP

and BAPLSP_SEQ for LOOO
Figure 7.1 indicates that the number of infeasible solutions computed by BAPLSP as
well as by BAPLSP_SEQ is about 100. BAPLSP_SEQ has found some solutions which
corresponding objective function values are substantially larger (smaller) than the largest
(smallest) objective function value computed by BAPLSP. In the most cases, the objective
function values determined by BAPLSP_SEQ are smaller than the objective values of
100
BAPLSP. Thus the quality of solution is increased due to the integration of the sequential test.
Such a result was observable in most of the instances which we have analyzed.
In Figure 7.2 the smallest objective function value of the best solution found until the ith execution (i=I, ... ,IOOO) ofBAPLSP and BAPLSP_SEQ are depicted.
220000
--BAPLSP
- - - BAPLSP_SEQ
200000
180000
-------------------
180000+---~---+--~~--+---~---+--~~--+---~--~
100
200
300
400
500
600
700
600
900
1000
Fig. 7.2 Evolution of the objective function values of Z~':LSP and Z~':LSP _SEQ for LOOO
Recall that we have considered a partition of the parameter space with 9 subspaces.
The expected sample size to reject or accept the null-hypothesis is around 30. So after 9
30= 270 executions parameter space reduction may take place. Thus the improvement of the
minimum around the 400-th executions of BAPLSP_SEQ may be caused by reducing the
parameter space, because the probability to find a better solution has increased. Such a result
was observable in most of the instances which we have analyzed.
We consider the instances provided in Table 7.5 to analyze the influence on

computation time of the BAPLSP and BAPLSP_SEQ by increasing the number of items
[CPUl to CPU3], the number of macro-periods [CPU4, CPUI,CPU5], or the number of subperiods per macro-period [CPU6, CPU!, CPU7].
101
Table 7.5 ComEutation time!) {in seconds} for 1000 executions
instance
Tm
TS
BAPLSP
BAPLSP SEQ
CPUl
10
10
8.4
8.68
CPU2
10
10
23.99
23.18
CPU3
10
10
46.36
42.95
CPU4
10
5.5
5.71
CPUS
IS
10
11.48
11.2
CPU6
10
9.06
9.23
CPU7
10
IS
9.17
9.39
I) On a Personal Systeml2 Model 70 486 using the computer language
Turbo Pascal 5.0
(The other input parameters of

instances as follows: hc=I, TBO=25,
INST(J,Tm,Ts,hc,TBO,~TBO,dp,cp,U;I)
~TBO=O,
are in all
dp=(ud,i), U=.8, cp=co).
The computation time increases substantially (but almost linearly) with the number of
items (cf. [CPU! to CPU3] in Table 7.5 and Figure 7.3 a, less strongly with the number of
macro-periods (cf. [CPU4, CPUl, CPU5] in Table 7.5 and Figure 7.3 b, and insignificant
with the number of sub-periods (cf. [CPU6, CPUI, CPU7] in Table 7.5 and Figure 7.3 c. The
integration of the sequential test requires insignificant computation time. Determining a
changeover, i.e. computing the regret measures of the items, requires most of computation
time. Thus if there are parameter values which lead to a solution with a lot of changeovers, the
computation of such a solution requires a lot of time due to the high number of regretcalculations. If such parameter values are eliminated by the sequential test the computation
time per schedule decreases. This may be the reason for a shorter computation time in CPU2
and CPU5 comparing BAPLSP_SEQ and BAPLSP.
102
b)
a)
c)
12
10
8
6
850
840
Coo
n20
d 10
s 0
12
10
8
6
nO.ofi1ems
2
0
10
15
no. of macro-periods
10
15
no. of sub-periods
Fig. 7.3 Computation time ofBAPLSP
7.3
Computational Results for the CLSPL
Let BACLSPL_SEQ denote that MAIN is performed using BACLSPL where the pre-sample
size and parameter subspaces are chosen as for BAPLSP_SEQ (cf Chapter 7.2).
We have tested BACLSPL and BACLSPL_SEQ on the problems 11 to 14 from [Thizy

and Van Wassenhove 1985] with T=8 (macro-) periods and J=8 items differing only in the
capacities; as in [Fleischmann 1990] the problems are denoted with TVll to TVI4.
The demand djt O=l,oo.,J, t=l,oo.,T), the setup costs Sj 0=1,00.,J) and the holding costs
coefficients hj O=I,oo.,J) in TVI1 to TV14 are given in Table 7.6.
103
Table 7.6 Demand, setup costs, and holding costs for TVll to TV14
dj!
j=1
dj2
dj3
d j4
djs
d j6
70
50
100
20
80
50
10
30
100
40
100
150
20
40
20
j=2
20
40
j=3
40
50
j=4
50
j=6
70
j=7
j=8
10
hj
djS
Sj
100
100
40
50
200
80
90
160
200
160
90
100
100
300
40
10
10
20
10
450
40
40
100
20
40
50
250
50
10
20
60
40
40
500
10
10
20
30
300
100
j=5
dj7
20
The capacities ofTVll to TV14 are provided in the Table 7.7.
Table 7.7 Capacities ofTVll to TV14
C1
C2
C3
C4
Cs
C6
C7
Cs
TVll
350
350
350
400
400
400
400
500
TV12
400
400
400
400
400
400
400
400
TV13
500
500
500
500
500
500
500
500
TV14
600
600
600
600
600
600
600
600
In [Fleischmann 1990] TVll to TV14 are solved as DLSP to optimality whereby the
(macro-) period t (=1, ... ,8) is divided in C.
Ip
sub-periods where p=50 has been chosen, e.g.
in TVl1 period t=1 is divided in 350/50=7 sub-periods. The total number of sub-periods is
t = (L~~, C.) I p, e.g. t = 3150 150 =63 in TV11. Furthermore the lotsizes are multiples of p
(cf Chapter 3.1). (Recall, that the demand is given in capacity units). So the total quantity of
an item j to be produced from period t=1 to period t=T is
OJ =r(:L~~' d
j .)
Ipl p, e.g.
0, =r420/501- 50 =450 in TV11. Hence, the discrete lotsizes may force the final inventory to
be greater than zero, i.e.
IjT~O,
e.g. I JT=450-420=30 in TV11 which causes avoidable holding
costs (at least) in period t=T. By an appropriate reduction of the last production quantities of
104
every item 0=1, ... ,1) the final inventories can be reduced to zero. For instance, if the last
production quantity ofitemj=l in TVII will be reduced to 20 in period t="t then the objective
function value will be reduced by h 120(T-"t+l) (note, that ql?"20 holds). Now, let reduced
costs of DLSP denote the objective function value of an optimal DLSP schedule where the last
production quantity of every item is reduced so that the final inventory is zero, i.e. IjrO
0=1, ... ,1).
TVII to TV14 have also been solved to optimality as CLSP by Branch&Bound [cf.
Gelders et al. 1986] and mixed integer programming [cf. Baranyet al. 1984]. For a given
CLSP solution setups can be reduced a posterior; by linking two lots of adjacent periods (for
the item with the highest setup costs, if there is a choice). This has been performed for TVII
to TV14 in [Fleischmann 1990] and the corresponding costs of such a modified solution are
denoted by reduced costs of CUiP.
We have applied BACLSPL and BACLSPL_SEQ for TV11 to TV14. The results are
given in Table 7.8.
Table 7.8 ComQarison of the CLSPL with the reduced costs of the DLSP and CLSP
reduced
reduced
BACLSPL BACLSPL BACLSPL BACLSPL
costs of
costs of
1000 exe.
DLSP
CLSP
_SEQ
2500 exe.
_SEQ
2500 exe.
1000 exe.
TV11
8080
7810
7470
7470
7470
7470
TV12
7090
7520
6980
6980
6980
6980
TV 13
6840
7360
6800
6900
6800
6790
TV14
6840
7520
6660
6660
6660
6660
In all instances BACLSPL and BACLSPL_SEQ have found substantial better solutions
than the other approaches. The integration of a sequential test based parameter control has not
improved the solution quality. For the problem TV13 BACLSPL_SEQ has only found a worse
solution than BACLSPL within 1000 executions. There are two possibilities for this result: It
may be an insignificant stochastic event, or a "good" subspace is "wrongly" eliminated which
105
again would be caused by the stochastic nature of our algorithm. Hence, we have increased the
number of executions from 1000 to 2500. In that case BACLSPL_SEQ has determined a
solution for the problem TV13 which is slightly better than the solution obtained by BACLSPL
(cf Table 7.6).
To improve the results we have tested other regret measures. For example we have
used
(l-Y)(h P jt ) if j*iandTD-TQ>C ndjt
r/i):= yes)
not feasible
if j = i and TD - TQ > C nd ".,

otherwise
instead of rjt(i) in BAPLSP which can be interpreted as "costs for producing item j not in
period til. Using rjt(i), BAPLSP has computed an objective function value of 7270 for the
TVl1 instance which is an improvement. For the other instances no better solution was found.
Moreover, the solutions were poor compared with the solutions provided in Table 7.8. The
TVl1 solution with objective function value Z=7270 is illustrated in Figure 7.4
qjt
linked production quantities
Qj,t+l]
j=l
j=2
j=3
j=4
j=5
j=6
F~-l
j=7
j=8
120
--4----11
--+I---II.~
3
4
5
:> t
Fig. 7.4 Solution ofTVll
We see that in this solution (as in the solution ofTVl2 to TVI4) the case of preserving
the setup state over several periods never occurs. This is important to notice in order to make
106
a fair comparison with the reduced costs of the DLSP and the CLSP since these models do not
keep track of the setup state for an item during idle periods.
The computation times for solving TVll to TV14, are given in the Table 7.9.
Table 7.9 Computation timel) (in seconds) for 1000 executions

BACLSPL
BACLSPL SEQ
TV 11
29.54
28.18
TV12
26.97
26.36
TV 13
22.91
23.02
TV14
21.91
22.78
1) On a Personal Systeml2 Model 70 486 using the computer language

Turbo Pascal 5.0
Again (cf Chapter 7.2) it can be seen the integration of the sequential tests does not
significantly increase the computation time. It is remarkable, that the computation time per
solution is less than .03 seconds (on average), which is very efficient.
Thus it can be stated that the randomized backward add-method is a very efficient
solution approach.
Note, in [Haase 1993] the Dixon-Silver heuristic and a deterministic heuristic which is
also backward oriented are compared. A computational study shows that the integration of
sequencing aspects improves the solution quality substantial.
Chapter 8.
Summary and Future Work
This manuscript deals mainly with dynamic, capacitated and multi-item lotsizing and scheduling
problems. We discusses the assumptions of the CLSP which forces to consider the CLSPL,
where the lots of adjacent periods can be linked. The CLSPL is superior to the CLSP with
respect to the quality of solutions. The CLSP and the CLSPL do not consider the sequence of
lots within a period.
Sequencing is integrated in the well-known DLSP and CSLP. A further new model, the
proportional lotsizing and scheduling problem (PLSP), is introduced which is superior to the
DLSP and the CSLP with respect to the quality of solution. Furthermore some extensions of
the PLSP are presented. Especially, if setup times and sequence dependent setup costs (and/or
times) have to be considered the PLSP extensions are more attractive than other existing
scheduling and lot sizing models.
A new .type of stochastic heuristic for lotsizing and scheduling is introduced where a
sequential test is integrated to control the choice of parameter values. The heuristic is used to
solve the PLSP, and with slight modifications, the CLSPL. Numerical results show that the
heuristic, especially the parameter control, is very efficient.
It is shown how the heuristic can easily be modified for some PLSP extensions. To
perform and evaluate such modifications is worthwhile and should be done in the next future.
Due to the choice decision which is based on randomized regrets (with parameter
control) there are relationships to other existing stochastic algorithms as genetic search,
simulated annealing as well as tabu search. Thus the integration of regret-based choice
selection as well the integration of parameter control may help to increase the performance of
these approaches significantly.
Bibliography
Afentakis, P., "A parallel heuristic algorithm for lot-sizing in multistage production systems",
lIE Transactions, Vol. 19 (1987), pp. 34-42.
Afentakis, P., Gavish, B., "Optimal lot-sizing algorithms for complex product structures",
Operations Research, Vol. 34 (1986), pp. 237-249.
Anderson, E. J., "A note on the dynamic lot-size model with uncertainty in demand and
supply processes", Management Science, Vol 35 (1989), pp. 635-640.
Bahl, H.C., Ritzman, L.P., Gupta, J.N.D., "Determining lot sizes and resource
requirements: A review", Operations Research, Vol. 35 (1987), pp. 329-345.
Barany, L, Van Roy, T., Wolsey, L.A., "Strong formulation for multi-item capacitated lot
sizing", Management Science, Vol. 30 (1984), pp. 1255-126l.
Bertocchi, M., Brandolini, L., Slominski, L., Sobczynska, J., "A Monte-Carlo approach for
0-1 programming problems", Computing, Vol. 48 (1992), pp.259-274.
Billington, P. J., McClain, J. 0., Thomas, L.J., "Mathematical programming approaches to
capacity-constrained MRP systems: Review, formulation and problem reduction",
Management Science, Vol. 29 (1983), pp. 1126-114l.
Billington, P. J., McClain, J. 0., Thomas, L.J., "Heuristics for multilevel lot sizing with a
bottleneck", Management Science, Vol. 32 (1986), pp. 989-1006.
Billington, P. J., Blackburn, J.D., Maes; J., Millen, R.A., Van Wassenhove, L.N., "Multiproduct scheduling in multi-stage serial systems", in: Chikan, A., Lovell, M.C. (eds.),
The Economics ofInventory Management, Elsevier Science Publishers B.V., Amsterdam
1988, pp. 345-355.
Bitran, G.R., Matsuo, H., " The multi-item capacitated lot size problem: Error bounds on
Manne's formulation", Management Science, Vol. 32 (1986), pp. 350-359.
Blackburn, J.D., Millen, R.A., "An evaluation of heuristic performance in multi-stage lotsizing systems", International J. of Production Research, Vo!' 23 (1985), pp. 857-866.
De Bodt, M., Gelders, L.F., Van Wassenhove, L.N., "Lot sizing under dynamic demand
conditions:a review", Engineering Costs and Production Economics (1984), pp. 165187.
Briiggemann, W., Jahnke B., "DLSP for 2-Stage Multi-Item Batch Production",
Arbeitspapier Nr. 31, Seminar fur allgemeine Betriebswirtschaftslehre Institut fur
Logistik und Transport der Universitat Hamburg (1992).
Campbell, G.M., Mabert, V.A., "Cyclical schedules for capacitated lot sizing with dynamic
demands", Management Science, Vo!' 37 (1991), pp. 409-427.
109
Carreno, J.J., "Economic lot scheduling for multiple products on parallel identical
processors", Management Science, Vol. 36 (1990), pp. 348-358.
Cattrysse, D., Maes, J., Van Wassenhove, L.N., "Set partitioning and column generation
heuristics for capacitated dynamic lotsizing", European J. of Operational Research, Vol.
46 (1990), pp. 38-47.
Cattrysse, D., Salomon, M., Kuik, R., Van Wassenhove, L.N., "Heuristics for the discrete
lotsizing and scheduling problem with setup times", Discussion Paper, ErasmusUniversiteit Rotterdam 1990.
Choo, E.U., Chan, G.H., "Two way eyeballing heuristics in dynamic lot sizing with
backlogging", Computers Operations Research, Vol. 17 (1990), pp. 359-363.
Chen, W.-H., Thizy, J.-M., "Analysis of relaxations for the multi-item capacitated lot-sizing
problem", Annals of Operations Research, Vol 26 (1990), pp. 29-72.
Diaby, M., Bahl, H.C., Karwan, M.H., Zionts, S., "A Lagrangean relaxation approach for
very-large-scale capacitated lot-sizing", Management Science, Vol 38 (1992), pp. 13291340.
DilIenberger, c., Escudero, L.F., Woolensak, A., Zhang W., "On practical resource
allocation for production planning and scheduling with different setup types", Working
paper: mM (GMTC) Sindelfingen 1992.
Dilts, D.M., Ramsing, K.D., "Joint lot sizing and scheduling of multiple items with sequencedependent setup costs", Decision Sciences, Vol 20 (1989), pp. 120-133.
Dinkelbach, W., "Zum Problem der Produktionsplanung in Ein- und Mehrproduktuntemehmen", Wurzburg-Wien 1964.
Dixon, P.S., Silver, E.A., "A heuristic solution procedure for the multi-item, single-level
limited capacity, lot-sizing problem", J. of Operations Management, Vol. 2 (1981), pp.
23-39.
Dobson, G., "The economic lot-scheduling problem: Achieving feasibility using time-varying
lot sizes", Operations Reasearch, Vol. 35 (1987), pp. 764-77l.
Dobson, G., "The cyclic lot scheduling problem with sequence-dependent setups", Operations
Research, Vol. 40 (1992), pp. 736-749.
Domschke, W., Scholl, A., VoO S., Produktionsplanung, Berlin 1993.
Dorndorf, U., Pesch, E., "Evolution based learning in a job shop scheduling environment",
Research Memorandum, Limburg University, Maastricht 1992 (Computers&Operations
Research, to appear).
Drexl, A., "Scheduling of project networks by job assignment", Management Science, Vol. 37
(1991), pp. 1590-1602.
110
Drexi, A., Fleischmann, B., Gunther, H.-O., Stadtler, H., Tempelmeier H.

"Konzeptionelle Grundlagen kapazitatsorientierter PPS-Systeme", Manuskripte aus den
Instituten fUr Betriebswirtschaftslehre der Universitlit Kie~ No. 315 (1993).
Drexi, A., Haase K., "A new type of model for multi-item capacitated dynamic lotsizing and
scheduling", Manuskripte aus den Instituten fUr Betriebswirtschaftslehre der Universitlit
Kiel No. 286 (1992).
Drexi, A., Haase K., "Sequential-analysis-based randomized-regret-methods for lotsizing and
scheduling", Manuskripte aus den Instituten fUr Betriebswirtschaftslehre der Universitlit
Kiel No. 323 (1993).
Dueck, G., Scheuer, T., "Treshold accepting", IBM Scientific Center, (Germany 1988)
Heidelberg.
Elmaghraby, S.E., "The economic lot scheduling problem (ELSP): Review and extensions",
Management Science, Vol. 24 (1978), pp. 587-59S.
Eppen, G. D., Martin, R.K., "Solving multi-item capacitated lot-sizing problems using
variable redefinition", Operations Research, Vol. 35 (1987), pp. 587-598.
Faigle, U. Kern, W., "Some convergence results for probabilistic tabu search", ORSA J. on
Computing, Vol. 4 (1992), pp. 32-37.
Federgruen, A., Tzur, M., "A simple forward algorithm to solve general dynamic lot sizing
models with n periods in O(n log n) or O(n) time", Management Science, Vol. 37 (1991),
pp. 909-925.
Fleischmann, B., "The discrete lot-sizing and scheduling problem", European J. of
Operational Research, Vol. 44 (1990), pp. 337-348.
Fleischmann, B., "The discrete lot-sizing and scheduling problem with sequence-dependent
setup costs", Discussion Paper, University Augsburg 1992 (European J. of Operational
Research", to appear).
Fleischmann, B., Popp, Th., "Das dynamische LosgroBenproblem mit reihenfolgeabhlingigen
Riistkosten", in: Pressmar, D. et al. (Eds.): "Operations Research Proceedings 1988",
Berlin 1989, pp. 510-515.
Florian, M., Lenstra, J.K., Rinnooy, A.H.G., "Deterministic production planning:
Algorithms and complexity", Management Science, Vol. 26 (1980), pp. 669-679.
Gascon, A., Leachman, R.c., "A dynamic programming solution to the dynamic multi-item,
single-machine scheduling problem", Operations Research, Vol. 36 (1988), pp. 50-56.
Gelders, L.F., Maes, J. , Van Wassenhove, L.N, "A branch and bound algorithm for the
multi-item, single-level capacitated dynamic lot-sizing problem", in: S. Axsliter et al.
(eds.), Multi-stage Production Planning and Inventory Control, Springer Verlag, Berlin
1986, pp. 92-108.
Glover, F., "Tabu search-Part I", ORSA J. on Computing, Vol. 1 (1989), pp.190-206.
111
Glover, F., "Tabu search-Part II", ORSA J. on Computing, Vol. 2 (1990), pp.4-32.
Goldberg, D.E., "Genetic algorithms in search, optimization, and machine learning",
ReadinglMass. 1989.
Giinther, H.-O., "Planning lot sizes and capacity requirements in a single stage production
system", European J. of Operational Research, Vol. 31 (1987), pp. 223-231.
Gupta, M.S., Brennan, L., "Lot sizing and backordering in multi-level product structures",
Production and Inventory Management Journal, No.1 (1992), pp. 27-34.
Haase, K., "Capacitated Lot-sizing with linked production quantities of adjacent periods",
Manuskripte aus den Instituten fur Betriebswirtschaftslehre der Universitat Kiel No. 334
(1993).
Haehling von Lanzenauer, c., "A production scheduling model by bivalent linear
programming", Management Science, Vol. 17 (1970), pp. 105-111.
Harris, F.W., "How many parts to make at onnce", Factory, The Magazine of Management,
Vol. 10 (1913), pp.135-136.
Hart, J.P., Shogan, A.W., "Semi-greedy heuristics: An empirical study", Operations Research
Letters, Vol. 6 (1987), pp. 107-114.
Hartmann, L.B., van Hee, K.M., "Application of Markov decision processes to search
problems", Paper presented at the 2nd ISDSS Conference, FAWIUlm, June 1992.
Hax, A.C., Candea, D., "Production and inventory management", Englewood Cliffs (New
Jersey) 1984.
Heinrich, C.E., "Mehrstufige Losgr6Benpianung
Produktionsplanungsystemen", Berlin 1987.
in
hierarchisch
strukturierten
Hertz, A., de Werra, D., "The tabu search metaheuristic: How we used it", Annals of
Mathematics and Artificial Intelligence, Vol. 1 (1990), pp. 111-121.
Hsieh, H., Lam, K.F., Choo, E.U., "Comparative study of dynamic lot sizing heuristics with
backlogging", Computers Operations Research, Vol. 19 (1992), pp. 393-407.
Hsu, W.-L., "On the general feasibility test of scheduling lot sizes for several products on one
machine", Management Science, Vol. 29 (1983), pp. 93-105.
Irle, A., "Sequentialanalyse: Optimale sequentielle Tests", Stuttgart 1990.
Johnson, D.S., Aragon, c.R., McGeoch, L.A., Schevon, C., "Optimization by simulated
annealing: An experimental evaluation: Part I, graph partitioning", Operations Research,
Vol. 37 (1989), pp. 865-892.
112
Johnson, D.S., Aragon, C.R., McGeoch, L.A., Schevon, C., "Optimization by simulated
annealing: An experimental evaluation: Part II, graph coloring and number partitioning",
Johnson, D.S., Papadimitriou, C.H., Yannakakis,M., "How easy is local search ?, J. of
Computer and System Sciences, Vol. 37 (1988), pp. 79-100.
Jones, P. C., Inman, R.R., "When is the economic lot scheduling problem easy?", lIE
Transactions, Vol. 21 (1989), pp. 11-20.
Karmarkar, U.S., Kekre, S., Kekre, S., "The dynamic lot-sizing problem with startup and
reservation costs", Operations Research, Vol. 35 (1987), pp. 389-398.
Karmarkar, U.S., Schrage, L., "The deterministic dynamic product cycling problem",
Kimms, A., "Multi-level, single-machine lot sizing and scheduling (with initial inventory)",
Manuskripte aus den Instituten fur Betriebswirtschaftslehre der Universitat Kiel No. 329
(1993).
Krajewski, L.J., Ritzman, L.P., Operations management, 2nd Edition, Addison Wesly
Massachusetts 1990.
Kuik, R., Salomon, M., "Multi-level lot sizing problem: Evaluation of a simulated-annealing
heuristic", European J. of Operational Research, Vol. 45 (1990) pp. 25-37.
Kuik, R., Salomon, M., "Linear programming, simulated annealing and tabu search for
lotsizing in bottleneck assembly systems", lIE Transactions, Vol. 25 (1993), pp. 62-72.
van Laarhoven, P.J.M., Aarts, E.H.L., Lenstra, J.K., "Job shop scheduling by simulated
annealing", Operations Research, Vol. 40 (1992), pp. 113-125.
Lambrecht, M.R., Vanderveken, H., "Heuristic procedures for the single operation multiitem loading problem", AIlE Transaction, Vol. 11 (1979), pp. 319-326.
Lasdon, L.S., Terjung, R.c., "An efficient algorithm for multi-item scheduling", Operations
Research, Vol. 19 (1971), pp. 946-969.
Lee, S.-B., Zipkin, P.H., "A dynamic lot-size model with make-or-buy decisions",
Management Science, Vol. 35 (1989), pp. 447-458.
Liepins, G.E. Hilliard, M.R., "Genetic algorithms: Foundations and applications", Annals of
Lozano, S., Larraneta, J., Onieva, L., "Primal-dual approach to the single level capacitated
lot-sizing problem", European J. of Operational Research, Vol. 51 (1991), pp. 354-366.
Madan, M. S., Gilbert, K.c., "An exact solution algorithm for a class of production planning
and scheduling problems", J. of the Operational Research Society, Vol. 43 (1992), pp.
961-970.
113
Maes, J., McClain, J.O., Van Wassenhove, L.N., "Multilevel capacitated lotsizing
complexity and LP-based heuristic", European J. of Operational Research, Vol. 53
(1991), pp. 131-148.
Maes, J., Van Wassenhove, L.N., "Multi-item single-level capacitated dynamic lot-sizing
heuristics: A general review", J. of the Operational Research Society, Vol. 39 (1988), pp.
991-1004.
Magnanti, T.L., Vachani, R., "A strong cutting plane algorithm for production scheduling
with changeover costs", Operations Research, Vol. 38 (1990), pp. 456-473.
Martin, 0., Otto, S.W., Felten, E.W., "Large-step Markov chains for the TSP incorporating
local search heuristics", Operations Research Letters, Vol. 11 (1992), pp. 219-224.
de Matta, R., Guignard, M., "Production scheduling with sequence independent changeover
cost", Working Paper, The Wharton School, University of Pennsylvania 1989.
McClain, J.O., Thomas, L.J., Weiss, E.N., "Efficient solutions to a linear programming
model for production scheduling with capacity constraints and no initial stock", lIE
Transactions, Vol. 21 (1989), pp. 144-152.
Miihlenbein, B., Gorges-Schleuter, M., Krimer, 0., "Evolution algorithms in combinatorial
optimization", Parallel Computing, Vol. 7 (1988), pp. 65-85.
Mood, M. A, Graybill F.A., Boes, D.C., "Introduction to the theory of statistics", 3. edition,
McGraw-Hill, Auckland 1982.
Pesch, E., "Local search based learning with applications in automated manufacturing", 1993
(in preparation).
Potts, C.N., Van Wassenhove, L.N., "Integrating scheduling with batching and lot-sizing: a
review of algorithms and complexity", Journal of the Operational Society, Vol. 43
(1992), pp. 395-406.
Pochet, Y., Wolsey, L.A., "Solving multi-item lot-sizing problems using strong cutting
planes", Management Science, Vol. 37 (1991), pp. 53-67.
Roll, Y., Karni, R., "Multi-item, multi-level lot sizing with an aggregate capacitY constraint",
European J. of Operational Research, Vol. 51 (1991), pp. 73-87.
Salomon, M., "Deterministic lotsizing models for production planning", Berlin 1991.
Salomon, M., Kroon, L.G., Kuik, R., Van Wassenhove, L.N., "Some extensions of the
discrete lotszing and scheduling problem", Management Science, Vol. 37 (1991), pp.
801-812.
Sandbothe, R.A., "The capacitated dynamic lotsizing problem with startup and reservation
costs: A forward algorithm solution", J. of Operations Mangement (1991), pp.255-266.
Schrage, L., "The multiproduct 10 scheduling problem", in: Dempster, M.A.H., Lenstra, J.K.
Rinnoy Kan, A.H.G. (Eds.), "Deterministic and stochastic scheduling",
DordrechtIHoliand, 1982, pp. 233-244.
114
Schrage, L., Linear, integer and quadratic programming with LINDO, third edition, Redwood
City 1986.
Siegmund, D., "Sequential analysis, tests and confidence intervals", New York 1985.
Silver, E.A., "Operations research in inventory management: a review and critique",
Operations Research (1981).
Steinberg, E., Napier, H. A., "Optimal multi-level lot sizing for requirements planning
systems", Management Science, Vol. 26 (1980), pp. 1258-1271.
Tempelmeier, H., Material-Logistik, Springer Berlinl992.
Tempelmeier, H., Helber, S., "A heuristic for dynamic multi-item multi-level capacitated
lotsizing for general product structures", Discussion Paper, TH Branschweig 1992.
Thizy, J.-M., "Analysis of Lagrangian decomposition for the multi-item capacitated lot-sizing
problem", lNFOR, Vol. 29 (1991), pp. 271-283.
Thizy, J.-M., Van Wassenhove, L.N., "Lagrangean relaxation for the multi-item capacitated
lot-sizing problem: A heuristic implementation", lIE Transactions, Vol. 17 (1985), pp.
308-313.
Trigeiro, W.W., Thomas, L.J., McClain, J.O., "Capacitated lot sizing with setup times",
Van Roy, T.J., Wolsey, L.A., "Solving mixed integer programming problems using automatic
reformulation", Operations Research, Vol. 35 (1987), pp. 45-57.
Wagelmans, A., van Hoesel, S., Kolen, A., "Economic lotsizing: An O(n log n) algorithm
that runs in linear time in the Wagner-Whitin case", Operations Research, Vol. 40
(1992), Suppl. No.1, pp. SI45-S156.
Wagner, H.M., Whitin, T.H., "Dynamic version of the economic lot size model",
Wolsey, L.A., "Uncapacitated lot-sizing problems with start-up costs", Operations Research,
Vol. 37 (1989), pp. 741-747.
Yannakakis, M., "The analysis of local search problems and their heuristics", in: Proc. ']th
annual symposium on theoretical aspects of computer science. Berling (Springer) 1990,
pp. 298-311.
Webster, F.M., "A back-order version of the Wagner-Whithin Discrete Demand EQQ
algorithm by dynamic programming", Production and Inventory Management Journal,
Vol. 30 (1989), pp. 1-5.
115
Zangwill, W.I., "A backlogging model and a multi-echolon model of a dynamic economic lot
size production system-a network approach", Management Science, Vol. 15 (1969), pp.
506-527.
APPENDIX
Monte-Carlo Study for Sequential Tests
The power and the sample size are criteria to assess different tests. We use the same symbols
as introduced in Chapter 5. The following tables (Table Al to A8) provide the results of a
10000 replication Monte-Carlo study for approximating the power function and, under
different circumstances, the expected sample size of the SPRT (cf Table AI, A2), SPRT-t (cf
Table A3, A4), RST-Ib (cf Table A5, A6), and RST-cb (cf Table A7, A8) using the test
problems
HO: p<':.2 versus Hj p<.2 (cf Table AI, A3, A5, A7)
HO: p<':.125 versusHj p<.125 (cf TableA2, A4, A6, A8)
where p is the parameter of the Bernoulli distribution. The tables provide the values which are
illustrated in the Figures 5.4 to 5.7 (cf Chapter 5).
Table At power and samEle size aEEroximation for the SPRT{.20,.05,.05,.05}

true p P[rej.HO]E E[N]E E[NIHO is rej']E E[NISR(60)]E E[NISR(80)]E E[NISR(100)]E
.25
.010
15.8
24.9
.20
.040
22.6
31.8
.15
.10
.192
.639
34.7
40.9
44.2
.05
.967
30.5
59.7
58.8
79.5
78.1
99.3
97.3
47.0
55.4
46.3
72.1
55.9
88.5
64.5
30.8
30.7
32.1
33.0
Table A2 Power and samEle size aEEroximation for the SPRT{.125,.025,.05,.05}

true p P[rej.HO]E E[N]E E[NIHO is rej']E E[NISR(60)]E E[NISR(80)]E .E[NISR(100)]E
.150
.125
.016
23.1
30.8
40.5
44.6
41.5
52.2
59.6
99.0
59.1
57.9
79.3
78.2
97.3
75.8
93.4
.100
.048
.130
.075
.365
55.9
64.1
55.1
70.0
84.2
.050
.737
59.1
61.8
49.8
67.2
.025
.973
43.8
44.2
40.5
59.3
43.2
44.8
117
Table A3 Power and samEle size aEEroximation for the SPRT-tPO,.05!.05 2.05 260}
true p P[rej.HO]E E[N]E E[NIHO is rej']E E[NISR(60)]E E[NISR(80)]E E[NISR(lOO)]E
.25
.on
15.6
24.4
59.6
79.4
99.2
.20
.044
21.8
28.5
58.6
77.3
96.9
.15
.171
3.8
33.1
55.4
72.0
88.6
.10
.529
36.6
34.5
46.5
56.0
65.4
.05
.934
29.5
28.7
30.8
32.1
33.5
Table A4 Power and samEle size aEEroximation for the SPRT-t{.125,.025,.05,.05,60}

true p P[rej.HO]E E[N]E E[NIHO is rej']E E[NISR(60)]E E[NISR(80)]E E[NISR(lOO)]E
.150
.016
22.4
34.8
59.6
.125
.043
.100
.107
79.3
98.9
28.2
37.4
34.7
39.8
59.0
78.2
97.3
57.8
75.7
93.6
.075
.242
41.1
40.5
55.3
7.4
85.6
.050
.504
44.1
39.8
49.8
59.8
69.7
.025
.850
39.4
36.9
4.4
43.4
46.3
Table AS Power and samEle size aEEroximation for the RST-lb{.20,5. 7,60}
true p P[rej.HO]E E[N]E E[NIHO is rej.]E E[NISR(60)]E E[NISR(80)]E E[NISR(lOO)]E
.25
.008
59.9
48.3
59.9
79.7
99.6
.20
.047
59.5
48.8
59.5
78.5
97.6
.15
.236
57.1
47.6
57.1
72.4
87.7
.10
.658
5.2
45.0
50.2
57.0
63.8
.05
.976
38.5
38.0
38.5
39.0
39.5
Table A6 Power and samEle size aEEroximation for the RST -lb{.125,4.5,60}
true p P[rej.HO]E E[N]E E[NIHO is rej']E E[NISR(60)]E E[NISR(80)]E E[NISR(100)]E
.150
.020
.125
.063
.100
.161
50.9
59.8
79.4
99.0
59.4
51.0
59.4
78.2
96.9
58.4
5.1
58.4
75.2
92.0
59.8
.075
.361
56.2
49.4
56.2
68.9
81.7
.050
.668
51.7
47.5
51.7
58.3
64.9
.025
.936
44.2
43.1
44.2
45.5
46.8
118
Table A7 Power and sam~le size a~~roximation for the RST-cb{.20,2.12,60}

true p P[rej.HO]~ E[N]~ E[NIHO is rej.]~ E[NISR(60)]~ E[NISR(80)]~ E[NISR(100)]~
.25
.012
51.6
27.6
59.6
79.4
99.2
.20
.15
.044
58.8
32.0
97.0
34.6
.10
55.1
45.7
58.8
55.1
77.9
.192
.571
45.7
71.3
54.3
87.5
62.9
.05
.945
3.4
30.4
31.5
32.5
34.9
28.6
Table AS Power and sam~le size a~~roximation for the RST-cb{.125,1.98,60}

true p P[rej.HO]~ E[N]~ E[NIHO is rej.]~ E[NISR(60)]~ E[NISR(80)]~ E[NISR(100)]~
.150
59.6
58.9
37.4
59.6
79.2
.125
.019
.051
37.7
58.9
.100
.118
57.5
38.6
57.5
77.8
75.1
.075
.267
.050
.535
54.5
48.7
39.4
38.9
48.7
69.2
58.0
.025
.867
39.2
36.0
39.0
41.9
54.5
98.8
96.8
92.8
83.8
67.3
44.5
The SPRT can be recommended. A summary and an interpretation of the results are
given in Chapter 5.
Vol. 314: R. Tietz, W. Albers, R. Selten (Eds.), Bounded

Rational Behavior in Experimental Games and Markets.
VI, 368 pages. 1988.
Vol. 338: J.K. Ho, R.P. Sundarraj, DECOMP: an

Implementation of Dantzig-Wolfe Decomposition for
Linear Programming. VI, 206 pages.
Vol. 315: I. Orishimo, G.J.D. Hewings, P. Nijkamp (Eds),

Information Technology: Social and Spatial Perspectives.
Proceedings 1986. VI, 268 pages. 1988.
Vol. 339: J. Terceiro Lomba, Estimation of Dynamic

Econometric Models with Errors in Variables. VIII, 116
pages. 1990.
Vol. 316: R.L. Basmann, D.J. Slottje, K. Hayes, J.D.

Johnson, OJ. Molina, The Generalized Fechner-Thurstone
Direct Utility Function and Some of its Uses. VIII, 159
pages. 1988.
Vol. 340: T. Vasko, R. Ayres, L. Fontvieille (Eds.), Life

Cycles and Long Waves. XIV, 293 pages. 1990.
Vol. 317: L. Bianco, A. La Bella (Eds.), Freight Transport

Planning and Logistics. Proceedings, 1987. X, 568 pages.
1988.
Vol. 342: K. Okuguchi, F. Szidarovszky, The Theory of

Oligopoly with Multi-Product Firms. V, 167 pages. 1990.
Vol. 318: T. Doup, Simplicial Algorithms on the

Simplotope. VIII, 262 pages. 1988.
Vol. 319: D.T. Luc, Theory of Vector Optimization. VIII,
173 pages. 1989.
Vol. 320: D. van der Wijst, Financial Structure in Small
Business. VII, 181 pages. 1989.
Vol. 321: M. Di Matteo, R.M. Goodwin, A. Vercelli (Eds.),
Technological and Social Factors in Long Term
Fluctuations. Proceedings. IX, 442 pages. 1989.
Vol. 322: T. Kollintzas (Ed.), The Rational Expectations
Equilibrium Inventory Model. XI, 269 pages. 1989.
Vol. 323: M.B.M. de Koster, Capacity Oriented Analysis
and Design of Production Systems. XII, 245 pages. 1989.
Vol. 324: I.M. Bomze, B.M. Potscher, Game Theoretical
Foundations of Evolutionary Stability. VI, 145 pages. 1989.
Vol. 325: P. Ferri, E. Greenberg, The Labor Market and
Business Cycle Theories. X, 183 pages. 1989.
Vol. 326: Ch. Sauer, Alternative Theories of Output, Unemployment, and Inflation in Germany: 1960-1985. XIII,
206 pages. 1989.
Vol. 327: M. Tawada, Production Structure and International Trade. V, 132 pages. 1989.
Vol. 328: W. Giith, B. Kalkofen, Unique Solutions for
Strategic Games. VII, 200 pages. 1989.
Vol. 329: G. Tillmann, Equity, Incentives, and Taxation.
VI, 132 pages. 1989.
Vol. 330: P.M. Kort, Optimal Dynamic Investment Policies
of a Value Maximizing Firm. VII, 185 pages. 1989.
Vol. 331: A. Lewandowski, A.P. Wierzbicki (Eds.),
Aspiration Based Decision Support Systems. X, 400 pages.
1989.
Vol. 341: G.R. Uhlich, Descriptive Theories of Bargaining.

IX, 165 pages. 1990.
Vol. 343: C. Chiarella, The Elements of a Nonlinear Theory

of Economic Dynamics. IX, 149 pages. 1990.
Vol. 344: K. Neumann, Stochastic Project Networks. XI,
237 pages. 1990.
Vol. 345: A. Cambini, E. Castagnoli, L. Martein, P
Mazzoleni, S. Schaible (Eds.), Generalized Convexity and
Fractional Programming with Economic Applications.
Proceedings, 1988. VII, 361 pages. 1990.
Vol. 346: R. von Randow (Ed.), Integer Programming and
Related Areas. A Classified Bibliography 1984-1987. XIII,
514 pages. 1990.
Vol. 347: D. Rfos Insua, Sensitivity Analysis in Multiobjective Decision Making. XI, 193 pages. 1990.
Vol. 348: H. Stormer, Binary Functions and their
Applications. VIII, 151 pages. 1990.
Vol. 349: G.A. pfann, Dynamic Modelling of Stochastic
Demand for Manufacturing Employment. VI, 158 pages.
1990.
Vol. 350: W.-B. Zhang, Economic Dynamics. X, 232 pages.
1990.
Vol. 351: A. Lewandowski, V. Volkovich (Eds.),
Multiobjective Problems of Mathematical Programming.
Vol. 352: O. van Hilten, Optimal Firm Behaviour in the
Context of Technological Progress and a Business Cycle.
XII, 229 pages. 1991.
Vol. 353: G. Ricci (Ed.), Decision Processes in Economics.
Proceedings, 1989. III, 209 pages 1991.
Vol. 354: M. Ivaldi, A Structural Analysis of Expectation
Formation. XII, 230 pages. 1991.
Vol. 355: M. Salomon. Deterministic Lotsizing Models for
Production Planning. VII, 158 pages. 1991.
Vol. 332: T.R. Gulledge, Jr., L.A. Litteral (Eds.), Cost

Analysis Applications of Economics and Operations
Research. Proceedings. VII, 422 pages. 1989.
Vol. 356: P. Korhonen, A. Lewandowski, J . Wallenius

(Eds.), Multiple Criteria Decision Support. Proceedings,
1989. XII, 393 pages. 1991.
Vol. 333: N. Dellaert, Production to Order. VII, 158 pages.

1989.
Vol. 357: P. Zornig, Degeneracy Graphs and Simplex

Cycling. XV, 194 pages. 1991.
Vol. 334: H.-W. Lorenz, Nonlinear Dynamical Economics

and Chaotic Motion. XI, 248 pages. 1989.
Vol. 358: P. Knottnerus, Linear Models with Correlated

Disturbances. VIII, 196 pages. 1991.
Vol. 335: A.G. Lockett, G. Islei (Eds.), Improving Decision

Making in Organisations. Proceedings. IX, 606 pages. 1989.
Vol. 359: E. de Jong, Exchange Rate Determination and

Optimal Economic Policy Under Various Exchange Rate
Regimes. VII, 270 pages. 1991.
Vol. 336: T. Puu, Nonlinear Economic Dynamics. VII, 119

pages. 1989.
Vol. 337: A. Lewandowski, I. Stanchev (Eds.),
Methodology and Software for Interactive Decision
Support. VIII, 309 pages. 1989.
Vol. 360: P. Stalder, Regime Translations, Spillovers and

Buffer Stocks. VI, 193 pages. 1991.
Vol. 361: C. F. Daganzo, Logistics Systems Analysis. X,
321 pages. 1991.
Vol. 362: F. Gehrels. Essays In Macroeconomics of an Open

Economy. VII. 183 pages. 199 I.
Vol. 363: C. Puppe. Distorted Probabilities and Choice
under Risk. VIII. 100 pages . 1991
Vol. 364: B. Horvath. Are Policy Variables Exogenous?
XII. 162 pages. 199 I.
Vol. 387: W. Gaertner, M. Klemisch-Ahlert, Social Choice

and Bargaining Perspectives on Distributive Justice. VIII,
131 pages. 1992.
Vol. 388: D. Bartrnann, M. J. Beckmann, Inventory Control.
XV, 252 pages. 1992.
Vol. 365: G. A. Heuer. U. Leopold-Wildburger. Balanced

Silverman Games on General Discrete Sets. V. 140 pages.
1991.
Vol. 389: B. Dulla, D. Mookherjee, T. Parthasarathy, T.

Raghavan, D. Ray. S. Tijs (Eds.), Game Theory and
Economic Applications. Proceedings, 1990. ??, ?? pages.
1992.
Vol. 366: J. Gruber (Ed.). Econometric Decision Models.

Proceedings. 1989. VIII. 636 pages. 1991.
Vol. 390: G. Sorger, Minimum Impatience Theorem for

Recursive Economic Models. X, 162 pages. 1992.
Vol. 367: M. Grauer. D. B. Pressmar (Eds.). Parallel

Computing and Mathematical Optimization. Proceedings.
V. 208 pages. 1991.
Vol. 391: C. Keser, Experimental Duopoly Markets with

Demand Inertia. X, 150 pages. 1992.
Vol. 368: M. Fedrizzi. J. Kacprzyk. M. Roubens (Eds.).

Interactive Fuzzy Optimization. VII. 216 pages. 1991.
Vol. 369: R. Koblo. The Visible Hand. VIII. 131
pages. 199 I.
Vol. 370: M. J. Beckmann. M. N. Gopalan. R. Subramanian
(Eds.). Stochastic Processes and their Applications.
Proceedings. 1990. XLI. 292 pages. 199 I.
Vol. 371: A. Schmutzler. Flexibility and Adjustment to Information in Sequential Decision Problems. VIII. 198
pages. 199 I.
Vol. 392: K. Frauendorfer, Stochastic Two-Stage

Programming. VIII, 228 pages. 1992.
Vol. 393: B. Lucke. Price Stabilization on World
Agricultural Markets. XI, 274 pages. 1992.
Vol. 394: Y.-J. Lai. C.-L. Hwang. Fuzzy Mathematical
Programming. XIII. 301 pages. 1992.
Vol. 395: G. Haag, U. Mueller, K. G. Troitzsch (Eds.).
Economic Evolution and Demographic Change. XVI. 409
pages. 1992.
Vol. 396: R. V. V. Vidal (Ed.). Applied Simulated
Annealing. VIII. 358 pages. 1992.
Vol. 373: A. Billot. Economic Theory of Fuzzy Equilibria.

XIII. 164 pages. 1992.
Vol. 397: J. Wessels. A. P. Wierzbicki (Eds.). UserOriented Methodology and Techniques of Decision
Analysis and Support. Proceedings. 199 I. XII. 295 pages.
1993.
Vol. 374: G. Pflug. U. Dieter (Eds.). Simulation and Optimization. Proceedings. 1990. X. 162 pages. 1992.
Vol. 398: J.-P. Urbain. Exogeneity in Error Correction

Models. XI. 189 pages. 1993.
Vol. 375: S.-J. Chen, Ch.-L. Hwang, Fuzzy Multiple

Attribute Decision Making. XII, 536 pages. 1992.
Vol. 399: F. Gori. L. Geronazzo. M. Galeotti (Eds.).

Nonlinear Dynamics in Economics and Social Sciences.
Proceedings. 1991. VIII. 367 pages. 1993.
Vol. 372: J. Esteban. The Social Viability of Money. X.

202 pages. 199 I.
Vol. 376: K.-H. Jockel, G. Rothe, W. Sendler (Eds.),

Bootstrapping and Related Techniques. Proceedings, 1990.
VIII, 247 pages. 1992.
Vol. 377: A. Villar, Operator Theorems with Applications
to Distributive Problems and Equilibrium Models. XVI,
160 pages. 1992.
Vol. 400: H. Tanizaki. Nonlinear Filters. XII. 203 pages.

1993.
Vol. 401: K. Mosler. M. Scarsini. Stochastic Orders and
Applications. V, 379 pages. 1993.
Vol. 378: W. Krabs. J. Zowe (Eds.), Modem Methods of

Optimization. Proceedings, 1990. VIII. 348 pages. 1992.
Vol. 402: A. van den Elzen, Adjustment Processes for

Exchange Economies and Noncooperative Games. VII, 146
pages. 1993.
Vol. 379: K. Marti (Ed.), Stochastic Optimization.

Vol. 403: G. Brennscheidt. Predictive Behavior. VI. 227

pages. 1993.
Vol. 380: J. Odelstad. Invariance and Structural

Dependence. XII, 245 pages. 1992.
Vol. 404: Y.-J. Lai. Ch.-L. Hwang. Fuzzy Multiple

Objective Decision Making. XIV. 475 pages. 1994.
Vol. 381: C. Giannini, Topics in Structural VAR

Econometrics. XI. 131 pages. 1992.
Vol. 405: S. Koml6si. T. Rapcslik. S. Schaible (Eds.),

Generalized Convexity. Proceedings, 1992. VIII, 404 pages.
1994.
Vol. 382: W. Oettli, D. Pallaschke (Eds.), Advances in

Optimization. Proceedings, 1991. X, 527 pages. 1992.
Vol. 383: J. Vartiainen, Capital Accumulation in a
Corporatist Economy. VII, 177 pages. 1992.
Vol. 384: A. Martina, Lectures on the Economic Theory of
Taxation. XII, 313 pages. 1992.
Vol. 385: J. Gardeazabal, M. Regulez, The Monetary Model
of Exchange Rates and Cointegration. X, 194 pages. 1992.
Vol. 386: M. Desrochers, J.-M. Rousseau (Eds.), Computer-Aided Transit Scheduling. Proceedings, 1990. XIII. 432
pages. 1992.
Vol. 406: N. M. Hung. N. V. Quyen. Dynamic Timing

Decisions Under Uncertainty. X. 194 pages. 1994.
Vol. 407: M. Ooms, Empirical Vector Autoregressive
Modeling. XIII. 380 pages. 1994.
Vol. 408: K. Haase. Lotsizing and Scheduling for
Production Planning. VIII. 118 pages. 1994.
Vol. 409: A. Sprecher. Resource-Constrained Project
Scheduling. XII, 142 pages. 1994.
Vol. 410: R. Winkelmann, Count Data Models. XI, 213
pages. 1994.

24226.lotsizing and Scheduling For Production Planning (Lecture Notes in Economics and Mathematical Systems)

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

24226.lotsizing and Scheduling For Production Planning (Lecture Notes in Economics and Mathematical Systems)

Caricato da

Copyright:

Formati disponibili

Lecture Notes in Economics

and Mathematical Systems

Lotsizing and Scheduling

CIP data applied for

1.1 Objectives of Lotsizing and Scheduling ..................................................................... 1

1.3 Outline of the Following Chapters............................................................................. 7

Single-Level Capacitated Lotsizing Problems

2.1 The Capacitated Lotsizing Problem (CLSP) .............................................................. 9

Single-Level Lotsizing and Scheduling Problems

3.1 The Discrete Lotsizing and Scheduling Problem (DLSP) ........................................... 22

Extensions of the PLSP

4.1 The PLSP with Setup Times ..................................................................................... 31

Control of Stochastic Algorithms via Sequential Analysis

5.3 Monte-Carlo Study for Sequential Tests ................................................................... 59

7.1 An Instance Generator .............................................................................................. 87

Summary and Future Work ........................................................................ 107

Bibliography .................................................................................................................. 108

Monte-Carlo Study for Sequential Tests ..................................................... 116

1.1 Objectives of Lotsizing and Scheduling

costs per unit in lot

1.2 Classification of Problems

2) Temporal development of parameters

Fig. 1.2 Item structures

"bottleneck". If more than one machine is considered a multi-machine problem arises,

10) Service policy

11) Consideration of time consuming activities

1.3 Outline of the Following Chapters

Chapter 2 considers the well-known capacitated lotsizing problem, denoted by CLSP.

Chapter 5 provides a Monte-Carlo study of so-called sequential tests. A fictitious stochastic

Single-Level Capacitated Lotsizing Problems

Whereas uncapacitated lotsizing problems can be solved efficiently to optimality [cf.

2.1 The Capacitated Lotsizing Problem (CLSP)

AO: Setup costs occur for each lot in a period.

the inventory of item j at the end of period t,

the quantity (lotsize) of item j to be produced in period t, and

Thus, mathematical1y the CLSP can be stated as fol1ows:

O=I, ... ,J; t=I, ... ,T)

(t=I, ... ,T)

O=I, ... ,J; t=I, ... ,T)

0=1, .... ,1; t=I, ... ,T)

0=1, ... ,1; t=I, ... ,T)

I i.t-l +qit - lit

(~max{CJ"",CT}/min{pl> ... ,PJ})

is a large number, and w.1.o.g. the inventory in

period t=O is Ijo=O 0=1, ... ,1).

2.2 Critique on the CLSP

Problems associated with extensions to the multi-level, multi-machine case

where J m is the set of items producible on

If we only consider the capacity restrictions

the "solution" illustrated in Figure 2.2 would be possible:

O=I, ... ,J; t=I, ... ,T)

where Nj denotes the set of immediate successors of item j, so that

Fig. 2.3 Feasible solution with small discrete periods

Example 2.4: Le(~

J=g, ~=J2' M=2, J1={1}, J2={2,3}, Crnt=2 (m=1,2; t=1,2), d2t=d3t=0

The disadvantage of a multi-level CLSP can be summarized as follows: A multi-level

2.3 The CLSP with Linked Lotsizes (CLSPL)

min LL[Sj(X jt -zJt)+hjI Jt ]

0=1, ... ,1; t=1, ... ,T)

(t=I, ... ,T)

0=1, ... ,1; t=I, ... ,T)

(t=I, ... ,T)

O=I, ... ,J; t=I, ... ,T)