0 Bewertungen0% fanden dieses Dokument nützlich (0 Abstimmungen)

1 Ansichten200 Seiten© © All Rights Reserved

0 Bewertungen0% fanden dieses Dokument nützlich (0 Abstimmungen)

1 Ansichten200 SeitenSie sind auf Seite 1von 200

Herausgegeben von

B. Fleischmann, Augsburg, Deutschland

M. Grunow, München, Deutschland

H.-O. Günther, Berlin, Deutschland

S. Helber, Hannover, Deutschland

K. Inderfurth, Magdeburg, Deutschland

H. Kopfer, Bremen, Deutschland

H. Meyr, Hohenheim, Deutschland

Th. S. Spengler, Braunschweig, Deutschland

H. Stadtler, Hamburg, Deutschland

H. Tempelmeier, Köln, Deutschland

G. Wäscher, Magdeburg, Deutschland

Diese Reihe dient der Veröffentlichung neuer Forschungsergebnisse auf den Gebieten

der Produktion und Logistik. Aufgenommen werden vor allem herausragende

quantitativ orientierte Dissertationen und Habilitationsschriften. Die Publikationen

vermitteln innovative Beiträge zur Lösung praktischer Anwendungsprobleme der

Produktion und Logistik unter Einsatz quantitativer Methoden und moderner

Informationstechnologie.

Herausgegeben von

Professor Dr. Bernhard Fleischmann Professor Dr. Herbert Meyr

Universität Augsburg Universität Hohenheim

Technische Universität München Technische Universität Braunschweig

Technische Universität Berlin Universität Hamburg

Universität Hannover Universität Köln

Universität Magdeburg Universität Magdeburg

Universität Bremen

Kontakt

Professor Dr. Hans-Otto Günther

Technische Universität Berlin

H 95, Straße des 17. Juni 135

10623 Berlin

Florian Seeanner

Multi-Stage

Simultaneous Lot-Sizing

and Scheduling

Planning of Flow Lines with Shifting

Bottlenecks

Florian Seeanner

Darmstadt, Germany

D 17

DOI 10.1007/978-3-658-02089-7

detailed bibliographic data are available in the Internet at http://dnb.d-nb.de.

Springer Gabler

© Springer Fachmedien Wiesbaden 2013

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole

or part of the material is concerned, specifically the rights of translation, reprinting, reuse of

illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical

way, and transmission or information storage and retrieval, electronic adaptation, compu-

ter software, or by similar or dissimilar methodology now known or hereafter developed.

Exempted from this legal reservation are brief excerpts in connection with reviews or schol-

arly analysis or material supplied specifically for the purpose of being entered and executed

on a computer system, for exclusive use by the purchaser of the work. Duplication of this

publication or parts thereof is permitted only under the provisions of the Copyright Law of

the Publisher’s location, in its current version, and permission for use must always be obtained

from Springer. Permissions for use may be obtained through RightsLink at the Copyright

Clearance Center. Violations are liable to prosecution under the respective Copyright Law.

The use of general descriptive names, registered names, trademarks, service marks, etc. in this

publication does not imply, even in the absence of a specific statement, that such names are

exempt from the relevant protective laws and regulations and therefore free for general use.

While the advice and information in this book are believed to be true and accurate at the date

of publication, neither the authors nor the editors nor the publisher can accept any legal re-

sponsibility for any errors or omissions that may be made. The publisher makes no warranty,

express or implied, with respect to the material contained herein.

Springer DE is part of Springer Science+Business Media.

www.springer-gabler.de

Foreword

Management concepts often repeat. Currently, another “wave of lean management” seems

to roll, preaching that inventories are waste. Inventories would blur a clear view on mistakes

which are made in production planning. Thus, inventories should be reduced. All drivers

that necessitate inventories should be eliminated by means of investments. This makes sense

in many industries, but not in all. Some industries utilize expensive, automated machines.

These cannot easily be replaced by newer ones allowing automated changeovers and thus

generating almost no setup times or setup costs. They apply a machinery, which has grown

over time. It consists of many machines which show a similar functionality, but have been

purchased successively over one or several decades. Hence, there are some older machines,

which are less eﬃcient, and a few newer ones, which are more eﬃcient.

An example is the consumer packaged goods industry. There are often only a few pro-

duction stages (like Make and Pack) with several heterogeneous, parallel production lines

per stage. Parallel means that these lines show a similar functionality, i.e., they can be

used alternatively to produce—more or less—the same products. Heterogeneous means that

they, nevertheless, do not need to be identical. This concerns production speeds, produc-

tion coeﬃcients, production costs, setup costs and setup times. Both setup costs and times

may even be sequence-dependent. The individual work stations of a production line are

usually connected by an automated transport system with a ﬁxed cycle time. Thus, each

production line can be considered as a single planning unit. Sometimes only one of these

stages constitutes a stationary bottleneck. However, sometimes the bottleneck may shift

dynamically, dependent on the mix of demand the customers ask for. This book focuses on

this kind of industries, more concretely on short- to medium-term production planning and

scheduling in this kind of industries when both lot-sizing and scheduling decisions have to be

taken simultaneously in order to consider the crucial interdependencies between the various

predecessor and successor products on the diﬀerent production stages.

The author proposes and compares diﬀerent mixed-integer programming formulations,

which base on the so-called General Lot-sizing and Scheduling Problem (GLSP), to accu-

rately model this planning situation. Furthermore, he develops and tests an innovative

solution heuristic, combining the principles of Variable Neighborhood Decomposition Search

with the Exchange heuristic, which not only behaves well for small “theoretical” problem

instances, but also proves to be applicable for a large industrial problem of plastic foil pro-

VI

duction. Despite of the high practical relevance of this type of planning problem, decision

support being oﬀered by science and by commercial software (e.g. as part of so-called “Ad-

vanced Planning Systems”) seemed to be rather poor and disappointing until now. The

author manages to make a ﬁrst—and already very impressive—step to improve this situa-

tion.

Thus, I invite the interested production planner, operations researcher and software de-

veloper to carefully read this book. Even though it deals with consumer packaged goods,

this book is no fast food. You will have to take your time to eat it. It will often taste sweet

and sometimes be hard to digest. But I promise—it is worth every bite of it.

Acknowledgments

Facing high cost pressure, industries like the consumer packaged goods industry are forced

to utilize their machine capacities as eﬃciently as possible. Hence, simultaneous lot-sizing

and scheduling is very important for these industries if they use machines with signiﬁcant,

sequence-dependent setup times or costs in a Make-To-Stock environment. Since often a

stationary bottleneck exists, to focus on the corresponding production stage is usually suf-

ﬁcient. Unfortunately, this is not the case if—due to varying product demands/mixes and

diﬀerent production speeds on the machines—the bottleneck shifts between the stages. The

thesis at hand is intented for helping solve such multi-stage lot-sizing and scheduling prob-

lems. Besides a mathematical model formulation, a new heuristic solution procedure is given

which allows ﬁnding good solutions for this model including real-world problem sizes.

The results of this thesis were made during my engagement as a research assistant at the

Chair of Production and Supply Chain Management at Darmstadt University of Technology.

So ﬁrst of all, I want to thank Prof. Herbert Meyr, the former owner of this chair and my ﬁrst

referee, who gave me this great opportunity to do a doctorate. Due to his deep knowledge

and his friendly type, it was a pleasure to work with him. The same is true for Prof. Bernardo

Almada Lobo, my second referee. His amicable and positive attitude always stimulated me

to achieve the aim. Moreover, I want to thank Prof. Rainer Quick who volunteered for acting

as the third referee.

As usual with bigger projects, some other people were also involved. Therefore, I want

to thank Prof. Wolfgang Domschke, Prof. Marco Lübbecke, and Prof. Laurence Wolsey for

their helpful suggestions. Furthermore, I would like to thank all my colleagues at university

and the members of the QBWL community for the warm and friendly atmosphere.

A special thank goes to my parents. Not only did they allow me this education but they

always supported me in everything I did. Last but not least I want to thank my wife Anne.

By her happy nature, she was a great backup and let me cope with all the challenges.

Sincere thanks are given to all of them!

Florian Seeanner

Contents

List of Figures XV

1 Introduction 1

1.1 Cost pressure in the consumer packaged goods industry . . . . . . . . . . . . 1

1.2 Motivation and goals of this thesis . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1 Supply Chain Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 Supply chain typology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.2.1 Functional attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.2.2 Structural attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.3 The supply chain of the consumer packaged goods industry . . . . . . . . . . 14

2.3.1 Topography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.3.2 Integration and coordination . . . . . . . . . . . . . . . . . . . . . . . 15

2.3.3 Functional attributes of the manufacturer . . . . . . . . . . . . . . . 15

2.3.3.1 Procurement type . . . . . . . . . . . . . . . . . . . . . . . 15

2.3.3.2 Production type . . . . . . . . . . . . . . . . . . . . . . . . 16

2.3.3.3 Distribution type . . . . . . . . . . . . . . . . . . . . . . . . 19

2.3.3.4 Sales type . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.4 Lot-sizing and scheduling in the consumer packaged goods industry . . . . . 21

2.5 Advanced Planning Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.1 General assumptions and limitations . . . . . . . . . . . . . . . . . . . . . . 25

3.1.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.1.2 Time structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

X Contents

3.2.1 The Capacitated Lot-sizing Problem (CLSP) . . . . . . . . . . . . . . 29

3.2.2 The Discrete Lot-sizing and Scheduling Problem (DLSP) . . . . . . . 30

3.2.3 The Continuous Setup Lot-sizing Problem (CSLP) . . . . . . . . . . 32

3.2.4 The Proportional Lot-sizing and Scheduling Problem (PLSP) . . . . . 33

3.2.5 The Capacitated Lot-sizing problem with Sequence-Dependent setup

costs (CLSD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3.2.6 The General Lot-sizing and Scheduling Problem (GLSP) . . . . . . . 36

3.3 Multi-level models in the literature . . . . . . . . . . . . . . . . . . . . . . . 39

3.3.1 Synchronization between the levels . . . . . . . . . . . . . . . . . . . 39

3.3.2 PLSP-based models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.3.2.1 Proportional Lot-sizing and Scheduling Problem with Paral-

lel Machines (PLSP-PM) . . . . . . . . . . . . . . . . . . . . 42

3.3.2.2 The multi-level single machine PLSP (PLSP-ML-SM) . . . . 43

3.3.3 CLSD-based models . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3.3.3.1 The Multi-level Sequence-dependent Lot-sizing and Schedul-

ing problem (MSLS) . . . . . . . . . . . . . . . . . . . . . . 47

3.3.3.2 The Multi-Level Capacitated Lotsizing Problem with Sequence-

Dependent Setup Costs using Parallel Machines at Multiple

Locations (MLCLSD-PM-ML) . . . . . . . . . . . . . . . . . 49

3.3.4 GLSP-based models . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.3.4.1 The General Lot-sizing and Scheduling Problem for Multiple

production Stages (GLSPMS) . . . . . . . . . . . . . . . . . 53

3.3.4.2 The two-stage multi-machine lot-scheduling model (P2SMM) 60

3.3.5 Other models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

4.1 Shortcomings of the basic GLSPMS . . . . . . . . . . . . . . . . . . . . . . . 67

4.2 Improved GLSPMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

4.2.1 New line synchronization . . . . . . . . . . . . . . . . . . . . . . . . . 70

4.2.2 Model formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4.3 Properties of the improved GLSPMS . . . . . . . . . . . . . . . . . . . . . . 73

4.3.1 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

4.3.2 Advantages and disadvantages . . . . . . . . . . . . . . . . . . . . . . 75

4.3.3 Generalization of other models . . . . . . . . . . . . . . . . . . . . . . 77

4.4 Computational tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.4.1 Base scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.4.1.1 Serial product structure (SER) . . . . . . . . . . . . . . . . 79

4.4.1.2 Divergent product structure (DIV) . . . . . . . . . . . . . . 80

Contents XI

4.4.2 Results with diﬀerent settings of a standard MIP-solver . . . . . . . . 82

5.1 General reformulation techniques . . . . . . . . . . . . . . . . . . . . . . . . 85

5.2 Selected approaches for lot-sizing and scheduling models . . . . . . . . . . . 87

5.2.1 Extended reformulations . . . . . . . . . . . . . . . . . . . . . . . . . 87

5.2.1.1 Simple Plant Location (SPL) . . . . . . . . . . . . . . . . . 87

5.2.1.2 Echelon Stocks . . . . . . . . . . . . . . . . . . . . . . . . . 88

5.2.1.3 Flow conservation for changeovers . . . . . . . . . . . . . . . 89

5.2.1.4 Elimination of the inventory stock variables . . . . . . . . . 90

5.2.2 Valid inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

5.3 Extended reformulations of the GLSPMS . . . . . . . . . . . . . . . . . . . . 91

5.3.1 SPL formulation (S) . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

5.3.2 Flow formulation (F) . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

5.4 Valid Inequalities for the GLSPMS . . . . . . . . . . . . . . . . . . . . . . . 95

5.4.1 Stock inequalities (K) . . . . . . . . . . . . . . . . . . . . . . . . . . 95

5.4.2 Tighter setup forcing constraints (M) . . . . . . . . . . . . . . . . . . 96

5.5 Computational results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

5.5.1 Comparison of the diﬀerent model formulations . . . . . . . . . . . . 97

5.5.2 Inﬂuence of varying problem characteristics . . . . . . . . . . . . . . 99

6.1 Selected meta-heuristic approaches . . . . . . . . . . . . . . . . . . . . . . . 103

6.1.1 Truncated MIP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6.1.2 LP&Fix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6.1.3 Relax&Fix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6.1.4 Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

6.1.4.1 General idea . . . . . . . . . . . . . . . . . . . . . . . . . . 105

6.1.4.2 Fix&Optimize as an example . . . . . . . . . . . . . . . . . 105

6.1.5 Local Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

6.1.5.1 General idea . . . . . . . . . . . . . . . . . . . . . . . . . . 109

6.1.5.2 2-opt as an example . . . . . . . . . . . . . . . . . . . . . . 109

6.1.5.3 Descent procedures . . . . . . . . . . . . . . . . . . . . . . . 110

6.1.5.4 Descent-ascent procedures . . . . . . . . . . . . . . . . . . . 110

6.1.6 Variable Neighborhood Search (VNS) . . . . . . . . . . . . . . . . . . 110

6.1.7 Variable Neighborhood Decomposition Search (VNDS) . . . . . . . . 113

6.2 Heuristics for the GLSPMS . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

6.2.1 Truncated MIP (TM) . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

6.2.2 LP&Fix (LF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

XII Contents

6.2.4 Trivial solution (TS) . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

6.2.5 Repeating setup sequence (RS) . . . . . . . . . . . . . . . . . . . . . 116

6.2.6 GLSPMS tailored Fix&Optimize (FO) . . . . . . . . . . . . . . . . . 118

6.2.7 Variable Neighborhood Decomposition Search with Exchange (VNDS+E)119

6.2.7.1 Main control procedure . . . . . . . . . . . . . . . . . . . . 121

6.2.7.2 Microperiod-based neighborhood . . . . . . . . . . . . . . . 123

6.2.7.3 Product-based neighborhood . . . . . . . . . . . . . . . . . 125

6.2.7.4 BOM-based neighborhood . . . . . . . . . . . . . . . . . . . 126

6.2.7.5 Line-based neighborhood . . . . . . . . . . . . . . . . . . . . 129

6.3 Computational tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

6.3.1 Small instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

6.3.2 Big instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

6.3.2.1 Test settings . . . . . . . . . . . . . . . . . . . . . . . . . . 132

6.3.2.2 Parameter tuning for the VNDS+E heuristic . . . . . . . . . 132

6.3.2.3 Results for VNDS+E, tailored Fix&Optimize, and Truncated

MIP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

6.3.2.4 Results for VNDS+E and Truncated MIP after 1 hour of

runtime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

6.3.3 A real-world instance . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

6.4 Brief summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

7.1 Availability of resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

7.1.1 Limited raw materials . . . . . . . . . . . . . . . . . . . . . . . . . . 143

7.1.2 Scarce common setup operators . . . . . . . . . . . . . . . . . . . . . 144

7.2 Time continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

7.2.1 Setup times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

7.2.2 Minimum lot-sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

8.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

8.2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

A.1 Serial instance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

A.2 Divergent instance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

A.3 General instance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

Contents XIII

Bibliography 173

List of Figures

1.1 Price indices of important raw materials (Food and Agriculture Organization

of the United Nations, 2011). . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.1 The Supply Chain Planning Matrix (Fleischmann et al., 2008, p. 87). . . . . 8

2.2 Trade-oﬀ between ﬁxed and variable costs per piece of a production lot (cf.

Harris, 1913). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.3 Gozinto graphs of the four product structures. . . . . . . . . . . . . . . . . . 13

2.4 Decoupling point in a Make-To-Stock environment (adapted from Fleischmann

and Meyr, 2004). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.5 A ﬂexible ﬂow line in the consumer goods industry. . . . . . . . . . . . . . . 16

2.6 An example of the soft drink industry (as in Ferreira et al. (2009)). . . . . . 17

2.7 An example of the plasters production. . . . . . . . . . . . . . . . . . . . . . 17

2.8 Small example of a shifting bottleneck. . . . . . . . . . . . . . . . . . . . . . 19

2.9 Typical demand curves (idealized). . . . . . . . . . . . . . . . . . . . . . . . 20

2.10 Example for low level coding. . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.11 Production planning the consumer goods industry (adapted from Fleischmann

et al., 2008, p. 97). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.12 Software modules covering the SCP-matrix (Meyr et al., 2008, p. 109). . . . 24

3.1 Continuous vs. discrete demand (in accordance to Meyr, 1999, p. 50). . . . . 27

3.2 Time buckets of identical length. . . . . . . . . . . . . . . . . . . . . . . . . 28

3.3 Diﬀerent lengths of time buckets. . . . . . . . . . . . . . . . . . . . . . . . . 28

3.4 Example for the two-level time structure of the GLSP. . . . . . . . . . . . . 36

3.5 PLSP with zero lead-times. . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.6 Possible structure of a microperiod. . . . . . . . . . . . . . . . . . . . . . . . 53

3.7 Example for the common time structure. . . . . . . . . . . . . . . . . . . . . 54

3.8 Tighter plans with setup and quantity splitting (in accordance with Meyr,

2004). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3.9 Impact of “slower” direct predecessor. . . . . . . . . . . . . . . . . . . . . . . 60

3.10 Line synchronization for split setup times. . . . . . . . . . . . . . . . . . . . 61

3.11 Schedules with and without synchronization (cf. Ferreira et al., 2009). . . . . 64

XVI List of Figures

4.2 Synchronization issue for more than two stages. . . . . . . . . . . . . . . . . 69

4.3 Predecessor item is transferred to several lines. . . . . . . . . . . . . . . . . . 69

4.4 New microperiod structure with two diﬀerent elements of idle time. . . . . . 70

4.5 Production of direct successor product j on line l must neither start nor end

before predecessor product i on line k in every microperiod s. . . . . . . . . 70

4.6 Travel distances between all pairs of ﬁve cities. . . . . . . . . . . . . . . . . . 74

4.7 Feasible solution for w1 = 0, w5 = 94, and w6 = 95. . . . . . . . . . . . . . . 75

4.8 High number of microperiods. . . . . . . . . . . . . . . . . . . . . . . . . . . 76

4.9 BOM and product-line assignment of the serial instance. . . . . . . . . . . . 79

4.10 BOM and product-line assignment of the divergent instance. . . . . . . . . . 80

4.11 BOM and product-line assignment of the instance with a general product

structure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

4.12 Progression of the upper and lower bound for the divergent instance with

L=3, J=12, T=5, and S=30. . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

5.2 Valid inequality for the regarded problem. . . . . . . . . . . . . . . . . . . . 86

5.3 Simple Plant Location problem. . . . . . . . . . . . . . . . . . . . . . . . . . 88

5.4 Product structure of the divergent base scenario. . . . . . . . . . . . . . . . . 89

5.5 Network ﬂow problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

6.2 Example for the product-oriented decomposition. . . . . . . . . . . . . . . . 107

6.3 Example for the resource-oriented decomposition. . . . . . . . . . . . . . . . 107

6.4 Example for the process-oriented decomposition. . . . . . . . . . . . . . . . . 108

6.5 Example for exchanging two edges as the transformation step of 2-opt. . . . 109

6.6 Neighborhood Nk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

6.7 Neighborhood Nk+1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

6.8 Random neighbor x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

6.9 Local optimum x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

6.10 A possible Relax&Fix scheme without overlap. . . . . . . . . . . . . . . . . . 115

6.11 A possible Relax&Fix scheme with overlap. . . . . . . . . . . . . . . . . . . . 115

6.12 Setup sequence obtained by the RS model. . . . . . . . . . . . . . . . . . . . 118

6.13 Repeating the setup sequence in every macroperiod t. . . . . . . . . . . . . . 118

6.14 Loop iteration 29—Neighborhood N1B . . . . . . . . . . . . . . . . . . . . . . 127

6.15 Loop iteration 38—Neighborhood N2B . . . . . . . . . . . . . . . . . . . . . . 128

6.16 Problem 24 solved with diﬀerent parameter scenarios. . . . . . . . . . . . . . 138

6.17 The percentage gap of the best VNDS+E run with the O-formulation and

scenario 2 for 204 microperiods. . . . . . . . . . . . . . . . . . . . . . . . . . 142

List of Figures XVII

7.1 The optimal solution of an artiﬁcial instance without considering any setup

operator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

7.2 Scheduling of a setup operator. . . . . . . . . . . . . . . . . . . . . . . . . . 146

7.3 The optimal solution for the same artiﬁcial instance with one setup team. . . 147

7.4 Example for a setup time overlapping more than two microperiods. . . . . . 149

7.5 Example for a continuous setup. . . . . . . . . . . . . . . . . . . . . . . . . . 152

B.2 Comparison of the solution procedures for problem 17. . . . . . . . . . . . . 172

B.3 Comparison of the solution procedures for problem 29. . . . . . . . . . . . . 172

List of Tables

increase (OECD, 2011). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2.1 Functional attributes of a supply chain typology (adapted from Meyr and

Stadtler, 2008, p. 67). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.2 Structural attributes of a supply chain typology (Meyr and Stadtler, 2008,

p. 69). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.3 Demand series causing bottlenecks on diﬀerent stages. . . . . . . . . . . . . . 19

4.2 Relaxed changeover variables smeared over the planning horizon. . . . . . . . 77

4.3 Overview of test instances. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

4.4 Percentage of test instances solved to optimality in 3600 sec. (PS) and per-

centage of instances performing best (PB) with solver settings “presolving”

(PR) and “cut generation” (CG) set on (1) and oﬀ (0). . . . . . . . . . . . . 83

5.1 Holding stocks and costs as well as echelon stocks and marginal holding costs. 89

5.2 Percentage of test instances solved to optimality in 3600 sec. (PS), the cor-

responding integrality gap (IG), and the average runtime (AR) in sec. for all

formulations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

5.3 Percentage of instances that are unﬁnished after 3600 sec. (PU) and the cor-

responding duality gap (DG) for all formulations. . . . . . . . . . . . . . . . 99

5.4 Percentage of test instances solved to optimality in 3600 sec. (PS), the cor-

responding integrality gap (IG), and the average runtime (AR) in sec. for all

formulations—now grouped by the product structure. . . . . . . . . . . . . . 100

5.5 Percentage of test instances with diﬀerent problem characteristics solved to

optimality in 3600 sec. (PS), the corresponding integrality gap (IG), and the

average runtime (AR) in sec. for the two best formulations S and F including

stock inequalities (K) as well as big-M-inequalities (M). . . . . . . . . . . . . 100

XX List of Tables

6.3 Loop iteration 12—Neighborhood N1J . . . . . . . . . . . . . . . . . . . . . . 126

6.4 Loop iteration 29—Neighborhood N1B . . . . . . . . . . . . . . . . . . . . . . 128

6.5 Loop iteration 38—Neighborhood N2B . . . . . . . . . . . . . . . . . . . . . . 128

6.6 Loop iteration 41—Neighborhood N1L . . . . . . . . . . . . . . . . . . . . . . 129

6.7 Gap (GP) and average runtime (AR) of Truncated MIP (TM), LP&Fix (LF),

and Relax&Fix (RF) for the extended formulations “ﬂow conservation” (F)

and “Simple Plant Location” (S)—both extended by stock inequalities (K) as

well as big-M-inequalities (M). . . . . . . . . . . . . . . . . . . . . . . . . . . 130

6.8 All 5 scenarios with their parameter values. . . . . . . . . . . . . . . . . . . 133

6.9 Comparison of diﬀerent sequences of the bill-of-materials based neighborhood

structures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

6.10 Number of times a new solution has been found by the neighborhood struc-

tures on average. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

6.11 Results of all 30 problem instances. . . . . . . . . . . . . . . . . . . . . . . . 136

6.12 Comparison of scenario 5 (SC5), Truncated MIP (TM), and tailored Fix&Optimize

(FO). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

6.13 Comparison of scenario 5 and Truncated MIP (TM) after 1 hour runtime. . . 139

6.14 Comparison of Truncated MIP (TM) and the new heuristic after 1 hour of

runtime. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

6.15 Comparison of scenarios 2 and 5 for formulations F+M+K and O. . . . . . . 141

7.2 Solution values of the example for a continuous setup. . . . . . . . . . . . . . 153

A.1 Number of lines, products, macroperiods, microperiods as well as set of all last

microperiods and all ﬁxed microperiods with their respective starting times. 161

A.2 Initial and maximum inventory, holding and purchasing costs as well as max-

imum purchase and sets of immediate and all successors. . . . . . . . . . . . 161

A.3 Index sets for line synchronization constraints. . . . . . . . . . . . . . . . . . 162

A.4 Overtime costs and maximum overtime. . . . . . . . . . . . . . . . . . . . . . 162

A.5 Production coeﬃcients. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

A.6 First microperiod and set of microperiods belonging to a macroperiod. . . . . 162

A.7 Demand. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

A.8 Maximum WIP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

A.9 Setup costs and times. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

A.10 Production costs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

A.11 Minimum lot-sizes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

A.12 Initial setup. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

A.13 Gozinto factors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

A.14 Allowed products and microperiods as well as standby costs. . . . . . . . . . 163

List of Tables XXI

A.15 Number of lines, products, macroperiods, microperiods as well as set of all last

microperiods and all ﬁxed microperiods with their respective starting times. 164

A.16 Initial and maximum inventory, holding and purchasing costs as well as max-

imum purchase and sets of immediate and all successors. . . . . . . . . . . . 164

A.17 Index sets for line synchronization constraints. . . . . . . . . . . . . . . . . . 164

A.18 Overtime costs and maximum overtime. . . . . . . . . . . . . . . . . . . . . . 165

A.19 Production coeﬃcients. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

A.20 First microperiod and set of microperiods belonging to a macroperiod. . . . . 165

A.21 Demand. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

A.22 Maximum WIP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

A.23 Setup costs and times. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

A.24 Production costs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

A.25 Minimum lot-sizes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

A.26 Initial setup. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

A.27 Gozinto factors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

A.28 Allowed products and microperiods as well as standby costs. . . . . . . . . . 166

A.29 Number of lines, products, macroperiods, microperiods as well as set of all last

microperiods and all ﬁxed microperiods with their respective starting times. 167

A.30 Initial and maximum inventory, holding and purchasing costs as well as max-

imum purchase and sets of immediate and all successors. . . . . . . . . . . . 167

A.31 Index sets for line synchronization constraints. . . . . . . . . . . . . . . . . . 167

A.32 Overtime costs and maximum overtime. . . . . . . . . . . . . . . . . . . . . . 168

A.33 Production coeﬃcients. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

A.34 First microperiod and set of microperiods belonging to a macroperiod. . . . . 168

A.35 Demand. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

A.36 Maximum WIP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

A.37 Setup costs and times. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

A.38 Production costs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

A.39 Minimum lot-sizes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

A.40 Initial setup. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

A.41 Gozinto factors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

A.42 Allowed products and microperiods as well as standby costs. . . . . . . . . . 170

List of Algorithms

1 GetNetDemands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3 Basic Variable Neighborhood Decomposition Search . . . . . . . . . . . . . . 113

4 Main control procedure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

5 StartExchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

6 GetMicroperiodNeighborhood . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

7 GetRandomSubSet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

8 GetProductNeighborhood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

9 GetUpwardsBomNeighborhood . . . . . . . . . . . . . . . . . . . . . . . . . . 127

10 GetDownwardsBomNeighborhood . . . . . . . . . . . . . . . . . . . . . . . . . 127

11 GetLineNeighborhood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

List of Abbreviations

Stages

LF LP&Fix

Setup Costs using Parallel Machines at Multiple Locations

MTS Make-To-Stock

XXVI List of Abbreviations

RF Relax&Fix

lem

TM Truncated MIP

WIP Work-in-process

Chapter 1

Introduction

Since the ﬁnancial crisis in 2009, the global economy has begun growing steadily again.

Especially in the emerging markets of Brazil, India, and China, there have been high growth

rates in the last few years. As illustrated in Table 1.1, the forecasts (marked with an asterisk)

still assume a positive trend in future years.

Brazil -0.7 7.5 3.4 3.2 3.9

China 9.2 10.4 9.3 8.5 9.5

Germany -5.1 3.6 3.0 0.6 1.9

India 7.0 9.9 7.7 7.2 8.2

Japan -6.3 4.1 -0.3 2 1.6

United States -3.5 3.0 1.7 2.0 2.5

World -1.2 5.0 3.8 3.4 4.3

Table 1.1: Real Gross Domestic Product growth of selected countries as a year-on-year

increase (OECD, 2011).

surprising that companies producing consumer packaged goods1 such as household products,

beverages, and food, were able increase their turnover after 2009.

try

Although the total sales of the top 50 FMCG companies have risen constantly, by up to

8% per year in 2010 and 2011, their net earnings even started to decline at the beginning

of 2011 (cf. OC&C Strategy Consultants, 2011). The main reason is that the prices for

raw materials have been increasing more and more, as shown in Figure 1.1. Furthermore,

new players from the emerging countries have entered the markets as well. Thus, consumer

1

They are also called fast moving consumer goods (FMCG) as they have a high stock turnover.

DOI 10.1007/978-3-658-02089-7_1, © Springer Fachmedien Wiesbaden 2013

2 Chapter 1. Introduction

325

300

275

250

200 Oils Price Index

Cereals Price Index

175

Dairy Price Index

150

125

100

75

2002 2003 2004 2005 2006 2007 2008 2009 2010 2011

Figure 1.1: Price indices of important raw materials (Food and Agriculture Organization of

the United Nations, 2011).

packaged goods companies are exposed to high pressure from competition, which results in

very low contribution margins.

For consumer packaged goods companies, the globalization has also led to very close

upstream and downstream linkages to suppliers and to consumers. On the one hand, raw

materials are sourced today from all over the world. On the other hand, the emerging

markets have to be satisﬁed with the products, too. Accordingly, a network of organizations

is “involved in the diﬀerent processes and activities that produce value in the form of products

and services in the hands of the ultimate consumer” (Christopher, 2005, p. 17). Such a

network represents a supply chain and as a consequence, today whole supply chains are

competing with each other. For this reason, managing such supply chains has become

increasingly important, where Supply Chain Management can be deﬁned as the “task of

integrating organizational units along a supply chain and coordinating material, information

and ﬁnancial ﬂows in order to fulﬁll (ultimate) customer demands with the aim of improving

the competitiveness of a supply chain as a whole” (Stadtler and Kilger, 2008, p. 11).

Due to the complexity of the supply chain, consumer packaged goods companies rely on IT

support. However, traditional Production Planning and Control (PPC) systems are not able

to adequately support the needed planning. Main reason is that PPC systems are typically

based on the concept of Material Requirements Planning (MRP), which was developed by

Orlicky (1975) and originally was meant for job shop manufacturing. Nevertheless, it has

been used for a wide range of diﬀerent production systems in the past, also including the

consumer packaged goods industry. But since MRP determines production quantities for

each item (of a multi-level structure) individually, and does not consider limited capacity

in a satisfactory manner, it is hardly suited for the consumer packaged goods industry.

Accordingly, it is advisable for the companies of this industry to introduce so-called Advanced

1.2. Motivation and goals of this thesis 3

Planning Systems (APS) because they are intended to provide “true planning capabilities”

(Fleischmann and Meyr, 2003) which appropriately consider the constraints of the production

system.

In the consumer packaged goods industry, the production system generally comprises two

or three production stages, such as make&pack, where in the ﬁrst stage the products are

made and then packed in the second stage, sometimes with an additional reﬁnement process

in between. At each stage, one or several parallel production lines might exist, which are all

highly utilized and thus represent potential bottlenecks. But before a certain item can be

produced, the production line has to be set up, which costs time and money. Generally, the

items can be assigned to a few so-called setup families. Changeovers between items of the

same family incur low setup costs as well as low setup times, and thus can be disregarded.

By contrast, high setup costs and setup times may result from changeovers between two

items of diﬀerent setup families (cf. the typology presented by Meyr and Stadtler (2008)).

As a result, there is some trade-oﬀ between reducing the number of changeovers (minimiz-

ing setup costs and times) and reducing stock levels (minimizing holding costs). Moreover,

setup times and costs are very often sequence-dependent, due to cleansing and other eﬀects.

Then not only the number of setups but especially their sequence has an impact on the

capacity available for production. Once several production lines are installed in parallel, not

only do decisions about the sizes of the lots have to be made, but also about the assignment

of these lots to production lines and the line-speciﬁc sequences of the lots.

Even though several APS claim to support the planning of a production system as outlined

above, some production planners in the consumer packaged goods industry are still not

satisﬁed.

One reason might be that some APS stick to MRP procedures—even for the lot-sizing

and scheduling. Therefore, they split demand into numerous jobs of a pre-deﬁned size so

that job shop planning methods can still be used. But the size of the jobs is a critical issue,

since a bigger size restricts the degree of freedom, and a smaller one drastically increases the

complexity. Either way, this approach seems to be inadequate.

A second reason could be that usually the planning problem is decomposed by stages

and the single-staged sub-problems are then solved sequentially. In general, this approach

is completely suﬃcient, as in many situations only one of the production stages actually

represents a (stationary) bottleneck (cf. Bolander and Taylor (1990) and (Meyr, 1999, p. 86)).

But because of line- and product-speciﬁc production speeds and time-varying demands, the

bottleneck might also shift dynamically between the stages. In that case, the decomposition

approach is insuﬃcient and a simultaneous consideration of multiple stages as well as parallel

4 Chapter 1. Introduction

particularly challenging.

Motivated by the dissatisfaction with such standard approaches on the part of production

planners in the consumer packaged goods industry, this thesis provides a possible component

for supporting them in making better decisions for the problems outlined above. As the

quality of decision support mainly depends on the adequacy of the models and applicable

solution methods (cf. Stadtler and Kilger, 2008, p. 4), two main goals can be identiﬁed for

this thesis.

1. A mathematical model for the simultaneous lot-sizing and scheduling of multiple stages

must be given which determines both the sizes and the sequences of the lots for all

relevant lines at all relevant stages. Therefore, the trade-oﬀ between models which are

as simple as possible and models that are as detailed as necessary has to be considered.

model in a reasonable amount of time, either exactly or near-optimally via heuristics.

Regarding the entire picture, these two elements should integrate into a decision support

system. Furthermore, in order to build a model which exhibits the relevant characteristics,

all aspects of the underlying planning problem have to be examined.

1.3 Outline

Due to the close linkages between suppliers, manufacturers, and customers, the supply chain

as a whole has to be analyzed in order to identify the relevant problem characteristics.

Therefore, Chapter 2 will introduce the basics of Supply Chain Planning in general and

of production planning in particular. Afterwards, the supply chain of a manufacturer in

the consumer packaged goods industry will be analyzed and special planning requirements

identiﬁed. Moreover, the structure and functionality of APS will be considered in order to

determine the requirements for a decision support system.

In a second step, Chapter 3 will review the current literature on multi-level lot-sizing

and scheduling models. For that purpose, some well-known single stage models will be in-

troduced ﬁrst: they will be the starting ground for several multi-level variants and help

illustrate several modeling aspects. As a result of this review, we will have to conclude that

only one existing model seems suitable for tackling the identiﬁed problem of the consumer

packaged goods industry, namely, the General Lot-sizing and Scheduling Problem for Multi-

ple Production Stages. Due to a few shortcomings, this model will be revised in Chapter 4

and the ﬁrst computational results will be given there. However, as will become clear, stan-

dard solution methods do not allow solving the model formulation at hand to optimality in

a reasonable amount of time.

1.3. Outline 5

Consequently, the following two chapters concern diﬀerent solution approaches. Chap-

ter 5 will examine diﬀerent reformulation techniques. By interpreting the underlying problem

as another well-known (combinatorial) problem, a redeﬁnition of the model variables and

constraints might allow of solving the model exactly by standard methods. However, due

to the complexity of lot-sizing and scheduling problems, heuristics will also be analyzed in

Chapter 6. In order to emphasize the quality of the revised model and the new solution

procedure, a real-world case will be explained and then solved. Finally, diﬀerent model

extensions such as for common setup operators will be presented in Chapter 7.

This thesis ends with a summary of its results, a conclusion, and an outlook on further

research opportunities in Chapter 8.

Chapter 2

packaged goods industry

Chapter 1 presented the motivation for the focus of this thesis, which is production plan-

ning in the consumer packaged goods industry—more speciﬁcally, lot-sizing and scheduling.

However, this planning cannot be considered separately, as will be shown in Section 2.1. In

fact, the whole supply chain needs to be regarded in order to adequately develop models

and methods (cf. Silver et al., 1998, p. 36). For this reason, a supply chain typology is pre-

sented in Section 2.2, which will help describe the supply chain in general and the regarded

production system in particular. Based on that typology, a detailed characterization of the

consumer goods industry is then presented and the planning requirements for lot-sizing and

scheduling will be derived. Furthermore, in Section 2.4, an operational planning concept will

be introduced which meets the identiﬁed requirements. Finally, the general structure and

functionality of an APS will be brieﬂy sketched in the last section.

In a consumer packaged goods company, a large number of decisions have to be made day

by day. Such decision making needs to be supported by planning, which is an instrument for

preparing decisions in a situation of uncertainty. The main tasks of planning are to identify

and evaluate alternative courses of action (cf. Fleischmann et al., 2008, p. 81). Accordingly,

making the right decisions is the aim of “Supply Chain Planning”.

Following the deﬁnition of Fleischmann and Meyr (2003), we understand Supply Chain

Planning (SCP) “as a generic term for the whole range of those decisions on the design

of the supply chain, on the mid–term coordination and on the short–term scheduling of

the processes in the supply chain.” As can be seen from this deﬁnition, decisions have

diﬀering importances and concern diﬀerent planning horizons. For this reason, we typically

DOI 10.1007/978-3-658-02089-7_2, © Springer Fachmedien Wiesbaden 2013

8 Chapter 2. Production planning in the consumer packaged goods industry

• supplier selection • production system structure • strategic sales

• cooperations planning

• material requirements scheduling • mid-term

• distribution planning

planning • capacity planning sales planning

• contracts

• personnel planning • short-term

• machine scheduling replenishment

• ordering materials sales planning

• shop floor control • transport planning

Figure 2.1: The Supply Chain Planning Matrix (Fleischmann et al., 2008, p. 87).

Long-term planning covers strategic decisions about the future development of the

supply chain. This includes decisions about the design of the supply chain network. The

planned and implemented measures usually aﬀect several years.

The main task of mid-term planning is to concretize regular operations based on the

strategic decisions. Therefore, rough quantities and points in time for the material ﬂow have

to be determined. The horizon generally covers 6–24 months.

Depending on the decisions of the two upper levels, short-term planning has to specify

detailed instructions for the execution and control of operations. The planning horizon ranges

from a few days to several months (cf. Fleischmann et al., 2008, p. 82).

But planning decisions not only diﬀer with respect to their horizon, they can also be

assigned to diﬀerent functional areas. Typically, the four functional processes procurement,

production, distribution, and sales are diﬀerentiated.1

Rohde et al. (2000) identify general supply chain planning tasks2 that can be assigned to

diﬀerent planning levels and functional processes. The result is the Supply Chain Planning

Matrix (see Figure 2.1) which gives an overview of the material and informational ﬂows be-

tween the diﬀerent planning tasks. As already mentioned, the long-term tasks concern the

design of the network. The SCP Matrix lists all these tasks in the same “box” to emphasize

the strong interdependency: the supply chain member has to decide which markets to enter

1

Moreover, Fleischmann et al. (2008, p. 86) state that every supply chain can be split into several internal

ones where each of them consists of these four processes.

2

“General” means that the tasks do not have to appear in every company, as they vary from industry to

industry. Nevertheless, the reader gets an impression of the typical decision problems a planner has to face.

2.1. Supply Chain Planning 9

or remain in, with existing or with new products. The chosen product program directly

inﬂuences the materials program. As today more and more components are purchased, the

suppliers have to be selected carefully in terms of product quality, service, and further prop-

erties.3 Another topic is to build cooperations with suppliers. By sharing information about

ﬁnal customer demands and production capacities (among other things), such a collabo-

ration may help to reduce, e.g., inventories, and thus costs, on both sides. Moreover, the

decisions about the product and material program are also directly connected to decisions

about the location of plants and the corresponding production systems, as they have an ef-

fect on how a customer can be reached. Since the location and capacity of a plant inﬂuence

the physical distribution structure, they are also usually planned together. Depending on

the product program and the sales forecast, the number of warehouses has to be determined

or the decision for a logistic service provider has to be taken (cf. Fleischmann et al., 2008,

pp. 88 and 89).

On the mid-term level, sales planning has to forecast the potential sales in speciﬁc regions.

Based on the given network and these sales forecasts, a rough capacity planning needs to

be executed and a master production schedule speciﬁed. This schedule, which may concern

several facilities simultaneously, determines the rough quantities which have to be produced

in certain periods, in order to deal with time-varying demands. With the help of capacity

planning, the required material as well as the required personnel has to be calculated and

the appropriate contracts need to be signed. Furthermore, distribution planning determines

the transport needs between the warehouses and the necessary stock levels. It also decides

whether an external logistic service provider has to be contracted (cf. Fleischmann et al.,

2008, p. 89).

On the short-term level, sales planning is responsible for detailed forecasts (if needed) and

demand fulﬁllment. In the latter case, sales staﬀ has to check whether a customer order can

be satisﬁed with the products already available or whether a new production order has to be

created. Distribution planning determines the detailed schedule for transport to replenish

the stocks in the warehouses or to deliver them to the customers (cf. Fleischmann et al.,

2008, p. 92). As already outlined in Section 1.1, minimizing setup costs and minimizing

inventory holding costs are competing objectives. Accordingly, one of the main tasks of

short-term production planning is to decide about the size of a “production lot”, which is

the quantity of a product that should be made at once. Figure 2.2 illustrates this trade-

oﬀ between setup costs and holding costs as a function of the lot-size. In addition to this

lot-sizing issue, another important short-term planning task is to schedule the lots on the

machines.4 Moreover, the shop ﬂoor control monitors all production activities and resources

3

Usually an ABC-classiﬁcation is executed (see Silver et al. (1998)) to focus on the most important

A-products and their suppliers.

4

Note that lot-sizing and scheduling is not necessarily a short-term planning task. Because of high setup

times (several days or even weeks) and long production runs, it might also be a mid-term task (see also

Suerie, 2005b, p. 69).

10 Chapter 2. Production planning in the consumer packaged goods industry

total costs

holding costs

setup costs

quantity

Figure 2.2: Trade-oﬀ between ﬁxed and variable costs per piece of a production lot (cf.

Harris, 1913).

time-eﬃcient shift schedule (extra shifts, overtime, capabilities of personnel, etc.) has to be

determined.

As shown with the Supply Chain Planning Matrix, there are strong interdependencies

between the diﬀerent planning levels (vertically) and even between diﬀerent tasks on the

same level (horizontally). Therefore, a simultaneous consideration of all tasks would be

preferable. But this monolithic planning approach is unrealistic, since the problem becomes

much too complex and thus is not solvable at all. On the contrary, a pure successive planning

approach might miss optimality by a great deal. A compromise is to decompose the tasks

hierarchically which has several advantages (cf. Dempster et al. (1981) and Fleischmann and

Meyr (2003)):

Since planning uncertainty increases with the length of the planning horizon, diﬀerent

planning levels should be based on diﬀerent degrees of aggregation. For instance, on a

strategic level, it might be suﬃcient to consider the yearly output for a product family

of a whole plant and the demand of an entire sales region. Besides, this aggregation is

also necessary because it is impossible to get detailed forecasts for several years at once.

Another reason for such a decomposition is that the shorter the planning horizon, the more

frequently the planning needs to be executed. Furthermore, strategic decisions are usually

made centrally by a top manager, whereas on, e.g., the operational level, the production

planners of diﬀerent plants decide in a decentralized fashion.

For all these reasons, it makes sense to build “planning modules” which integrate the

decisions which should be made simultaneously (cf. Fleischmann and Meyr, 2003). The aim

of hierarchical planning is to ﬁnd the right level of detail and the right planning horizon

2.2. Supply chain typology 11

for the modules, and to coordinate them. The modules are arranged hierarchically and

they exchange information and instructions (cf. Hax and Meal (1975) as well as Schneeweiss

(2003)).

To decide whether the planning tasks of lot-sizing and machine scheduling should build

a common planning module, the supply chain of the consumer packaged goods industry has

to be analyzed. For that purpose, we introduce a typology in the next section which helps

to characterize the supply chain.

Meyr and Stadtler (2008) present a supply chain typology 5 which allows identifying the

planning/decision problems of a supply chain. They diﬀerentiate between functional and

structural attributes.

Meyr and Stadtler follow the logic of the Supply Chain Planning Matrix to structure the

functional attributes of a supply chain (see Section 2.1). Accordingly, these attributes de-

scribe the procurement, production, distribution, and sales type of each member in the

supply chain. An overview of the categories and attributes considered is given in Table 2.1.

To begin with the inbound side of a supply chain member, the procurement type has to

be determined. At ﬁrst, the type and the number of products which are procured have to

be considered. The product type may range from rather simple standard products such as

raw materials (e.g., milk or cereals) up to very complex and specialized items (e.g., entire

consoles for car manufacturing). Since the products can be procured from diﬀerent suppliers,

the sourcing type describes their number (single-, double-, or multi-sourcing). Furthermore,

the ﬂexibility of a supplier in terms of quantities plays an important role, as well as the

lead time and reliability (cf. Meyr and Stadtler, 2008, p. 66). A further relevant topic is the

life-cycle of the externally obtained products, as a short life carries the risk that the product

gets obsolescent in stock.

The production process can be organized in diﬀerent ways, such as in a job shop or

ﬂow shop system. The repetition of operations is closely linked to the production type. It

may vary from mass production (with a continuous repetition) to one–of–a–kind produc-

tion (without any repetition). But there are also intermediate degrees between these two

extremes. Then, typically, batches or lots have to be built (see also page 9). Before the

5

This typology is supposed to be supplementing the SCOR-model which was published by the Supply

Chain Council (2011). For further reading, see Suerie and Wagner (2008), among others.

12 Chapter 2. Production planning in the consumer packaged goods industry

Functional attributes

Categories Attributes

Procurement type number and type of products procured

sourcing type

ﬂexibility of suppliers

supplier lead time and reliability

materials’ life cycle

Production type organization of the production process

repetition of operations

changeover characteristics

bottlenecks in production

working time ﬂexibility

Distribution type distribution structure

pattern of delivery

deployment of means of transportation

loading restrictions

Sales type relation to customers

availability of future demand

demand curve

number of product types

bill-of-materials (BOM)

product life cycle

degree of customization

portion of service operations

Table 2.1: Functional attributes of a supply chain typology (adapted from Meyr and Stadtler,

2008, p. 67).

production of such a batch (lot) can be started, usually a setup needs to be executed to pre-

pare the machine. Accordingly, the changeover characteristics describe the time and costs

which are necessary for this preparation. Setup times and costs can be either dependent or

independent of the sequence of the batches or lots. Furthermore, if there are bottlenecks

in production, the types of the bottlenecks have to be examined as well as the ﬂexibility of

extending or reducing the working time.

To analyze the distribution type, we have to consider the product ﬂow to the customer /

supply chain neighbor. First of all, the distribution structure may consist of several stages

with diﬀerent types, such as a two- or three-stage structure with central and regional ware-

houses. Another important issue is how regularly the supply chain neighbors are delivered

(cyclic or dynamic). Besides, further loading restrictions, such as “full truck load” have to

be identiﬁed.

The sales type is mainly characterized by the relation to the customer(s), which may

range from a single direct supply chain neighbor to an anonymous market with thousands of

participants. This relation also determines whether future demands are available as orders

or forecasts. Especially for forecasts, the demand curve, which might be static, seasonal, and

so on, needs to be known. For selling the products, the degree of customization, the number

2.2. Supply chain typology 13

of diﬀerent product types, and the product life cycle play an important role, as they directly

inﬂuence sales activities like marketing. Closely related to these product properties, the

bill-of-materials (BOM) describes how raw materials and components are processed for ﬁnal

items. In general, products might consist of diﬀerent parts and components. In accordance

with the relationship between items and their components, four diﬀerent product structures

are usually distinguished. If a product is only processed in further production steps, a

serial product structure is assumed. However, if an item has exactly one component as

predecessor, there is a divergent structure. Vice versa, a convergent structure means each

item goes into exactly one product. Accordingly, a general structure shows none of these

relationships. Since the bill-of-materials is a description of the product structure, the term

BOM is often used synonymously with the term product structure in the literature. Another

way to describe the product structure is a gozinto graph (see Figure 2.3). This is a directed

graph where the nodes represent products and each arc is weighted with a “gozinto factor”

which shows how many items of the source-product go into the sink-product.

The structural attributes given in Table 2.2 help describe the linkage between the members

of a supply chain. They can be roughly grouped into the categories “topography” and

“coordination and integration” (cf. Meyr and Stadtler, 2008, p. 69). The ﬁrst category

contains all attributes which characterize the supply chain. The network structure shows

the links along the material ﬂows in the supply chain. Moreover, the whole network may span

several countries. Besides, the location of the decoupling point indicates which processes are

executed before a customer order arrives (anticipative) or after (reactive). Finally, the major

constraints arising through limitations of resources such as capacity, labor, and material have

to be considered.

The integration and coordination of the supply chain members has to be analyzed as well.

If legally separated companies build a supply chain, the legal position of each company is a

relevant attribute. Accordingly, the balance of power must be considered, too. The direction

14 Chapter 2. Production planning in the consumer packaged goods industry

Structural attributes

Categories Attributes

Topography of a supply chain network structure

degree of globalization

location of decoupling point(s)

major constraints

Integration and coordination legal position

balance of power

direction of coordination

type of information exchanged

Table 2.2: Structural attributes of a supply chain typology (Meyr and Stadtler, 2008, p. 69).

of coordination which is also very important can usually be derived from the two previous

attributes. Between diﬀerent companies, the direction is horizontal, whereas it is vertical

within a single company. Finally, the type of information which is exchanged between supply

chain partners should be speciﬁed.

industry

With the help of the typology presented in the previous section, the supply chain of the

consumer packaged goods industry can be characterized (cf. Fleischmann and Meyr (2003)

as well as Meyr and Stadtler (2008)).6 At ﬁrst, the topography of the supply chain and

the integration and coordination will be explained. Afterwards, the functional attributes of

the supply chain member “manufacturer” will be illustrated with a strong emphasis on the

production and the sales type7 categories due to the focus of this thesis.

2.3.1 Topography

Consumer packaged goods companies usually produce at several places in diﬀerent countries.

Furthermore, the raw materials are often purchased from diﬀerent parts of the world and

the ﬁnished products are distributed globally as well. Accordingly, the production and

distribution network is widely ramiﬁed.

Especially in the consumer goods industry, ﬁnal customer demands need to be met di-

rectly. If an item is not available on time, then, typically, sales are lost. Accordingly, the

6

Not all of the companies in that industry have to show exactly the same characteristics. However, for

those which can be characterized in a similar fashion, the models and solution methods presented in this

thesis should be applicable.

7

As Meyr and Stadtler (2008) point out, the ﬁnal products (consumer goods) represent the determining

factor of the supply chain.

2.3. The supply chain of the consumer packaged goods industry 15

delivery time has to be reduced to a minimum, which can be achieved by keeping ﬁnished

goods in stock. This approach is called a Make-To-Stock (MTS) production, where the whole

production process is decoupled from customer orders. Hence, the ﬁnished goods inventory

decoupling point

anticipative reactive

inventory of

finished goods

and Meyr, 2004).

represents the decoupling point by deﬁnition (see Figure 2.4). This means that the whole

production process is anticipative, and thus based on demand forecasts. Accordingly, fore-

casting accurately the demand is a very important task in an MTS environment (Vollmann

et al., 2005, pp. 21–22). Because of the location of the decoupling point, central and regional

warehouses are needed near the markets all over the world.

The major constraint is the limited production capacity. Due to the high degree of

automation of the production lines this capacity is usually very expensive.

Since consumer goods are substitutable, due to their low product diﬀerentiation, the retailers

(customers) are in a powerful position (buyer’s market). Because of the global presence and

competitive pressure, information must be exchanged horizontally as well as vertically to

coordinate the supply chain.

For producing consumer packaged goods, mainly raw materials such as dairy or cereals have

to be procured. Usually, there are several possible suppliers on the market (multi-sourcing)

with short lead-times. Moreover, long-term contracts are often signed with suppliers in

order to bind them, even though the availability and quality of natural products cannot be

guaranteed (e.g., due to bad weather and other critical factors).

16 Chapter 2. Production planning in the consumer packaged goods industry

Organization of the production process Usually, the whole production process consists

of several sub-processes, where each of these processes changes an item in the sense of adding

value. Generally speaking, a sub-process transforms input, in the form of resources and pre-

items, to output, in the form of intermediate and ﬁnal products (cf. Meredith and Shafer,

2010, p. 8). In our case, such a transformation process corresponds to a certain functionality

which is oﬀered by a machine. We deﬁne this functionality as a production stage.8

The organization is mainly described by the layout of the machines and the ﬂow of

jobs/products (Silver et al., 1998, p. 669). The production in the consumer goods industry

is usually organized as a ﬂexible ﬂow line system. But since there are diﬀerent deﬁnitions

of such a ﬂow line in the literature (cf. Quadt and Kuhn (2007, pp. 686–687) and Ruiz and

Rodríguez (2010) among others)9 , we need to specify this term for the consumer packaged

goods industry. Hence, we deﬁne a ﬂow line as illustrated in Figure 2.5: All products have

to follow the same ﬂow through the system, from production stage 1 to E. The production

stages are serially arranged. At each stage, one or several parallel machines can be used

alternatively as they provide partly the same functionality. They can be either identical in

terms of the same setup and production times or heterogeneous with diﬀerent times and

speeds. A product may skip one or more production stages. Furthermore, the stages can be

decoupled by buﬀers. The lots do not have to be transferred in whole batches in general, but

can be streamed. That is why through-put times are low. This decrease, in turn, reduces

the Work-In-Process (WIP) stocks as well.

machine 1

machine i-1

machine 2 machine 4

machine i

buffer

machine 3

(Work-In-Process)

Note that “machine” is used here as a generic term for a planning resource which is

considered as a whole. For instance, such a planning unit might be a (chemical) reactor.

But also a ﬂow shop with a pre-deﬁned material ﬂow can be aggregated to a single resource

production line. Due to this aggregate view, we can assume that there are only two or three

8

Accordingly, we get a “new” product at each stage, since value is added.

9

The deﬁnitions are usually driven by the regarded industry. For instance, Quadt (2004) considers the

case of the semi-conductor industry, where the machines at each stage have to be identical and the products

have to be processed at all stages.

2.3. The supply chain of the consumer packaged goods industry 17

stages in general. Furthermore, each machine belongs uniquely to one production stage.

In the following, two small examples will be described. They give an impression of the

possible variants of such a production system. The ﬁrst one is illustrated in Figure 2.6. It

shows a slimmed-down version of a production system from the Brazilian soft-drink industry.

In the ﬁrst stage, the diﬀerent soft-drinks are prepared in several tanks. Then these tanks

provide the ﬁlling lines at the second stage with ﬂuids. Every tank is connected to every

line. However, each line can only be fed by one tank at the same time.

liquid flavor

bottling stage

preparation stage

tank 1

filling line 1

tank 2

filling line k

tank l wash bottles fill label seal package palett

Figure 2.6: An example of the soft drink industry (as in Ferreira et al. (2009)).

Figure 2.7, the production of wound plasters10 also comprises two diﬀerent stages. Generally,

both the making and the packaging stage represent potential bottlenecks.

In the ﬁrst stage, diﬀerent components, such as silicon foil, glue, and cotton, are processed

on several production lines to plasters of diﬀerent sizes and forms. After the plasters are

made, they are packaged in variety (combination) or non-variety packs at the second stage.

The stages are decoupled by using the small spaces between the lines to temporarily store

the plasters.

10

In several countries such as the United States, Australia, and India, they are also called band-aids.

18 Chapter 2. Production planning in the consumer packaged goods industry

Repetition of operations As the job ﬂow is ﬁxed, only technologically related products

can be produced on such ﬂow lines. In the consumer goods industry, production is typically

continuous (e.g., beverages or ﬂuids in general) as well as discrete (e.g., plasters). Due to the

high degree of automation, short throughput rates are possible. Hence, discrete production

can be seen as quasi-continuous, since a huge number of items are produced. That is why a

continuous production can be assumed for the consumer goods industry.

Changeover characteristics As there is a huge number of ﬁnal items which are somehow

technologically related, the products can be assigned to a few setup “families”. For instance,

we assume that the products i, j, and v belong to diﬀerent families. Changeovers between

items of the same family are negligible. But typically setup operators have to change machine

tools or clean the machines for changeovers between items of diﬀerent families. Then, such

a changeover, e.g., between item i and j incurs signiﬁcant setup time stij and costs sij .

Furthermore, in many cases these changeovers are highly sequence-dependent (e.g., stij =

stvj ), particularly due to needed cleaning processes (among others). For instance, in the soft

drink industry the setup time to prepare the bottling line for a soft drink with sugar after

one without is less time-consuming than the other way round (cf. Ferreira et al., 2009).

Usually, setup times and costs fulﬁll the triangle inequality which means that stiv +stvj ≥

stij and siv + svj ≥ sij , respectively. But if a product v has, e.g., a cleaning eﬀect on

a production line, it might be the case that two changeovers from product i to v and v

to j are cheaper or less time-consuming than a direct changeover from i to j. Thus, the

triangle inequality is not fulﬁlled. This fact is another challenge for modeling lot-sizing and

scheduling problems, as such a product v is usually set up more than once per period (cf.

Menezes et al., 2011). A detailed explanation can be found in Chapter 3.

stage, and is known. Then this production stage can be regarded in isolation and the

related production planning problems can be solved consecutively. But as pointed out in

Chapter 1, there are shifting bottlenecks as well. In such cases, the production stages have

to be considered simultaneously. In order to clarify this matter further we consider a small

example from the consumer goods industry:

Assume a company sells two diﬀerent products, X and Y, as two diﬀerent packaged goods

A and B. In a simple make&pack environment, X and Y are produced in the ﬁrst stage

and then packaged in the second. Accordingly, the gozinto factor is 1. Figure 2.8 gives

an overview. The available capacity is 100 time units per period and the production line

in the ﬁrst stage is able to make one item of X in aM X = 5 time units, and one of Y, in

aM Y = 4 time units. Furthermore, the packaging line processes item X in aP A = 3 time

units (resulting in A) and item Y in aP B = 7 (resulting in B ) time units. If, then, the

external demand varies over time (see also Section 2.3.3.4), a shifting bottleneck as shown

2.3. The supply chain of the consumer packaged goods industry 19

gozinto factor

pXA=1 pYB=1

in Table 2.3 occurs. As can be seen, in period 3 the available capacity in the making stage

is not suﬃcient to produce the needed quantities in that period. Moreover, in period 6, the

packaging stage becomes the bottleneck.

demand for product A (qty) 14 15 16 14 13 11

demand for product B (qty) 6 5 6 7 8 10

needed production time in stage M 94 95 104 98 97 95

needed production time in stage P 84 80 90 91 95 103

So the production planner is on the horns of a dilemma: both stages have bottlenecks,

in diﬀerent periods. If the production of the packaged goods is moved to earlier periods in

order to build stocks, this can create additional capacity shortages at the ﬁrst stage. As

stocks might also be built at the making stage, the planner has to determine lot-sizes and

sequences carefully. Consequently, a successive approach is not suﬃcient. The only chance

of solving this conﬂict is to consider and optimize both stages simultaneously. Although this

is quite a small example, it gives a ﬁrst impression of the problem and its diﬃculty. In a

real-world case, hundreds of items and several hundreds of ﬁnished goods might have to be

taken into account.

Working time ﬂexibility Generally, in a ﬂow line production system, only a few but

well-trained operators are needed. As described in Section 2.1, the working times of these

operators are usually determined on a mid-term level, and thus are ﬁxed on the short-

term range. Furthermore, in many companies, production lines are already operating 24/7.

Because of this, working time ﬂexibility is quite limited.

The distribution network of the supply chain normally comprises two or three stages (cf.

Fleischmann, 1998). This structure of central and regional warehouses help realizing short

20 Chapter 2. Production planning in the consumer packaged goods industry

lead times. Accordingly, the customers (retailers) can be supplied almost daily from the

regional warehouse, whereas the regional warehouses are fed with bundled deliveries from

the central one. Therefore, central warehouses generally keep the whole range of products

in stock.

Availability of future demand and the demand curve Since production is Make-To-

Stock, forecasts play an important role. Normally, a solid data base is available, as there

have been many observations made in the past, thanks to the long life cycle of the product.

However, several factors may complicate the forecasting. For instance, there is usually a

strong seasonality in the demand curve.11 But also forward buying12 as well as product

cannibalization13 have signiﬁcant impact. Figure 2.9 shows typical demand curves.

demand demand

average

level base

level

time time

(a) Seasonality. (b) Promotion.

But these are not the only diﬃculties a demand planner has to face. Generally, the

forecast error increases the further one looks into the future. Accordingly, the quality of the

forecast has an impact on the length of the planning horizon for lot-sizing and scheduling.

the bill-of-materials (BOM) describes the product structure. Since it might be the case that

the same intermediate product/component has multiple uses, as shown in Figure 2.10a, they

are usually sorted according to their low level. This means each item is located on the level

with the longest path (or highest number of stages) to the ﬁnished good (cf. Figure 2.10b).

Then, the material requirements for each item can be easily computed level-by-level.14

11

As an example, the demand for ice-cream is signiﬁcantly higher in the summer than in the winter.

12

This means that retailers buy products ahead and keep them in stock. This eﬀect is often the conse-

quence of a (price) promotion.

13

For instance, very often a company sells items which are closely related and thus can be substituted.

Then, an increase in the demand for one product may reduce the demand for the other. This eﬀect is called

cannibalization (cf. Meredith and Maki (2001); Srinivasan et al. (2005)).

14

A similar procedure can be found with the description of Stadtler’s multi-level PLSP in Section 3.3.2.2.

2.4. Lot-sizing and scheduling in the consumer packaged goods industry 21

1 1 low level 0

2 3 2 low level 1

4 3 4 3 low level 2

Multi-level product structures normally result from the application of diﬀerent trans-

formation processes. Accordingly, these structures usually imply a production system with

multiple production stages. However, it might be the case that a “new product” is gener-

ated by reapplying the same process and thus, it is still the same production stage. Then,

the product structure has several levels, but not the production in a narrow sense. For

instance, chemical processes such as ﬁltration, extraction, etc., are often executed several

times. Nevertheless, this scenario is very unusual for the consumer packaged goods industry.

Furthermore, in the regarded case, the bill-of-materials generally has a divergent struc-

ture, since only a few pre-items are processed to several hundreds of ﬁnal items. However,

there is also a convergent element in this structure, since diﬀerent products are typically

combined and packaged as ﬁnal “goods”.

Product life cycles and the degree of customization Consumer packaged goods used

to be standard items with a life cycle of, usually, several years. However, in the recent past,

companies have tended to shorten this life cycle, by changing the packaging or by slightly

“improving” the components of their goods to make them seem more attractive to the market.

First and foremost, this makes the forecasting of future demand even more diﬃcult.

goods industry

Section 2.1 presented general planning tasks which can be somehow found in every supply

chain. But depending on the type of the supply chain under consideration, these tasks may,

in the concrete, vary signiﬁcantly.15 As illustrated in the previous section, the consumer

packaged goods supply chain has many special properties, which have a strong impact on

15

Note that this is one of the reasons why planning concepts such as MRP or MRP-II are not able to

properly support all kinds of industries and their resulting problems.

22 Chapter 2. Production planning in the consumer packaged goods industry

the planning tasks and on their coordination, especially in the mid- and short-term horizon

(cf. Drexl et al., 1994).

One possible way to consider the interdependencies between the planning tasks of the

consumer packaged goods industry appropriately is proposed by Fleischmann et al. (2008).

They present a hierarchical planning concept which is sketched in Figure 2.11.

cast

working setup seasonal Demand

time times stock Planning

net Netting, fore

Simultaneous demand Factory cast

Assignment

Lot-sizing

and

short-term Scheduling

Figure 2.11: Production planning the consumer goods industry (adapted from Fleischmann

et al., 2008, p. 97).

Starting point is the centrally executed demand planning. Due to seasonal demands

and the ﬂuctuations induced by marketing activities, the forecast horizon on the mid-term

level should comprise at least one seasonal cycle (cf. Fleischmann et al., 2008). But then,

because of the high number of ﬁnal items, it is only possible to forecast product families with

similar demand patterns. Besides, the aggregation helps to reduce the uncertainty because

up- and downward deviations can be balanced. The resulting forecasts are input to the

master planning. As the production capacity represents the major constraint, this module

must decide about pre-production and building seasonal stocks. Moreover, capacity can only

be extended on that level, as the working time of the setup operators is only ﬂexible on a

mid-term horizon, due to the legal framework. But production quantities might be shifted

between several available sites—however, this incurs higher transportation costs. To cope

with the demand uncertainty, this concept additionally considers safety stock levels in the

master planning based on the existing forecast error.16 Accordingly, the result of the master

planning is the required capacity (working time) and the product quantities that have to

be pre-produced (seasonal stocks) for each plant as well as the needed transport capacity

between the plants and the warehouses.

16

Depending on the service level that should be reached, the safety stock levels can be calculated, as

shown for instance by Nahmias (2009, p. 272) and Silver et al. (1998, p. 242).

2.5. Advanced Planning Systems 23

Based on the initial inventories and the updated forecasts—now for single ﬁnal items

on the short-term level—the net demands can be calculated. Net demand constitutes input

data for simultaneous lot-sizing and scheduling. Moreover, the working time determines the

available capacity for production.

Due to sequence-dependent changeovers, a simultaneous lot-sizing and scheduling is nec-

essary since the sequence of the lots has an impact on the available capacity, and the capacity

inﬂuences the sizes of the lots. Furthermore, if shifting bottlenecks occur, several stages must

be considered simultaneously as well.

As the lot-sizing and scheduling problem is supposed to be a very hard planning problem,

it is a good idea to use mathematical models to ﬁnd feasible plans. Generally, these kinds of

models describe the technological, physical, legal, and further constraints on a real object as

mathematical relationships (cf. Williams, 1999, p. 3). With the help of a suitable objective

function, the solutions of the model can be evaluated and the optimum determined. How-

ever, to keep such an optimization model solvable, it should be as abstract as possible and as

detailed as necessary. Moreover, if the model only includes linear constraints and a linear ob-

jective function with continuous or even discrete variables, the resulting optimization model

can be tackled with standard Mixed Integer Programming (MIP) solvers which are usually

part of Advanced Planning Systems (cf. Stadtler, 2008). To understand the functionality

and interaction of such systems, a general overview of the planning capabilities of APS is

ﬁnally given next.

In contrast to Enterprise Resource Planning (ERP) systems, Advanced Planning Systems

provide real planning functionality as postulated in Section 2.1. As Fleischmann et al. (2008,

p. 84) state, an APS represents a hierarchical planning system which allows an integral plan-

ning of the supply chain. Therefore, APS typically consist of several modules which comprise

diﬀerent planning tasks illustrated by the SCP-Matrix (see Section 2.1 and cf. Meyr et al.

(2008, p. 109)). Figure 2.12 gives a general overview. At the long-term level, a Strategic Net-

work Design module possibly helps to decide about the production and distribution network

as well as other strategic questions. The Master Planning module allows the coordination

of the material ﬂow, rough capacity, and mid-term material planning. With the help of the

Production Planning and Scheduling modules, tasks such as lot-sizing, machine scheduling,

and sequencing can be carried out. However, as outlined in Section 2.2, the requirements

for such modules depend strongly on the type of the supply chain. That is why some APS

vendors oﬀer tailored modules for the consumer goods industry, which include some kind of

simultaneous planning functionality. Moreover, APS provide “true optimization capabilities”

by exact or heuristic methods which are available for the diﬀerent modules (cf. Fleischmann

and Meyr, 2003).

24 Chapter 2. Production planning in the consumer packaged goods industry

Demand

Planning

Purchasing Production Distribution

& Planning Planning

Material

Requirements Demand

Planning Transport Fulfillment

short-term Scheduling

Planning & ATP

Figure 2.12: Software modules covering the SCP-matrix (Meyr et al., 2008, p. 109).

Today, several Advanced Planning Systems17 exist on the market. Since this supply

chain software market is extremely competitive, Advanced Planning Systems have matured

to provide an extensive level of planning capabilities (cf. Kilger and Wetterauer, 2008). But

this also leads to the point, why APS vendors do not disclose their underlying models to

secure a competitive advantage. Accordingly, the planner does not exactly know the model

and the applied solution method. Hence, it is not clear if an APS exists that actually allows

a simultaneous lot-sizing and scheduling of multiple stages at the required level.

For this reason, we have to check whether adequate models exist in the literature at all. If

this is the case, the question about adequate solution methods which are able to solve these

models naturally arises. Therefore, a literature review follows, which answers this question

for the appropriate models.

17

APS are based on the data of Enterprise Resource Planning systems. Accordingly, APS cannot replace

an ERP system but should be seamlessly integrated.

Chapter 3

As indicated in Section 2.5, today’s Advanced Planning Systems provide sophisticated plan-

ning functionality. However, since APS providers are not willing to disclose the used lot-sizing

and scheduling models nor the applied solution methods, the planner can barely know how

the results are produced. For this reason, a literature review will be given in this chapter

with a focus on multi-level lot-sizing and scheduling models. Furthermore, we restrict the

review to mixed-integer linear programming (MIP) models, for which standard solvers and

solution methods are available, and which are compliant to the planning concept presented

in Section 2.4.

Hence, Section 3.1 will brieﬂy outline the resulting assumptions and limitations for the re-

view. Due to synchronization issues, which are caused by the scheduling on the one hand and

by the multi-level structure on the other hand, the modeling is of central importance. There-

fore, Section 3.2 will present some well-known single-level models to demonstrate diﬀerent

potential approaches. In this context, relevant characteristics are derived and highlighted.

Afterwards, an overview over existing multi-level models in the literature will be given in

Section 3.3. The ﬁnal section will draw a conclusion, whether an existing model is suitable

to solve the regarded problem and how to proceed in the following.

Based on the planning concept shown in Section 2.4, several assumptions can be made to

deﬁne the scope of this review. At ﬁrst, a ﬁnite planning horizon and a deterministic dynamic

demand is assumed. Accordingly, stochastic models with random demand functions (an

overview is given by Winands et al. (2011); see also Tempelmeier and Herpers (2011) among

others) as well as static models with static demand rates and usually an inﬁnite planning

horizon (e.g., Ben-Daya et al. (2008) and Jenabi et al. (2008)) are disregarded. Moreover,

further limitations concerning the objective and the time structure are derived next.

DOI 10.1007/978-3-658-02089-7_3, © Springer Fachmedien Wiesbaden 2013

26 Chapter 3. Multi-level lot-sizing and scheduling

3.1.1 Objective

As denoted in the chapters before, the aim of lot-sizing and scheduling models is to minimize

the sum of ﬁxed and variable costs—more particular the sum of setup and holding costs:

Holdings costs typically consist of four diﬀerent elements (cf. Nahmias, 2009, p. 268).

The most obvious one is the costs for the physical storage space. But also taxes

and insurances have to be added. Furthermore, if an item is kept in stock, eﬀects like

breakage, spoilage, deterioration, or obsolescence might occur and should be quantiﬁed

in terms of costs, too. Finally, we have to consider the capital being tied up by the

items in the stock. Hence, opportunity costs of an alternative investment need to be

part of the holding costs. This latter element is usually the most important one and

is the main reason, why companies usually want to reduce their stock levels.

Setup costs normally comprise all costs associated with starting or stopping the produc-

tion of an item. This includes costs for cleaning material as well as for the wastage

which occurs during the start-up of a line. Accordingly, setting up the line from the un-

prepared/neutral status incurs start-up costs. Similarly, shutting down a line induces

shut-down costs.

But very often other cost components such as production, standby, and penalty costs

have to be included in the objective function as well:

Production costs need to be considered, if, for instance, parallel lines are available which

can be used alternatively and have diﬀerent cost rates. Then, these costs help decide,

on which line a product should be made.

Standby costs can be incurred, e.g., by the conservation of a setup state and by the reser-

vation of the production line for a certain product.

Penalty costs represent a further cost component. Generally, these costs should penalize

unfulﬁlled conditions. For instance, the aim of a customer-oriented company is to

avoid lost-sales or backorders. However, it might be useful to allow them in a planning

model, otherwise the model can become too restrictive and no solution might be found.

But then lost-sales/backorders must be penalized to be less attractive for being part of

a problem solution.1 Another application of penalty costs is the consideration of non-

monetary goals. For instance, if the level of customer satisfaction is to be optimized,

the deviation to a pre-determined goal can be penalized. In general, penalty costs

signiﬁcantly exceed the other cost components. Nevertheless, they have to be chosen

carefully to avoid distorting the actual results.

1

In this context it should be mentioned that costs for lost-sales and backorders are hard to quantify, as

the loss-of-goodwill by the customer needs to be measured. Nevertheless, Anderson et al. (2006) point out

that it is important to consider these opportunity costs for stock-outs. Furthermore, they show a possible

approach for calculating these.

3.1. General assumptions and limitations 27

goals, such as the minimization of earliness, tardiness or makespan, since the production

quantities were already determined. Correspondingly, these models will be dropped from

the following review. The same applies to models which disregard stock levels and the

corresponding holding costs. Hence, models for simultaneous batching and scheduling are

also dismissed if they neglect these costs.2

The state of a production system is steadily changing over time. These state changes result

from the external and from the internal dynamics of the system. They may occur continu-

ously or at discrete events (cf. Pressmar, 1980). External dynamics are “perturbations” from

outside which cannot be inﬂuenced by the decision maker (anymore). They ensue from the

strong connection to the system environment and lead to exogenous state changes (cf. Meyr,

1999, pp. 49). For lot-sizing and scheduling, mainly the product demand and the availability

of resources (machine capacity, personnel, raw materials)3 are these external dynamics.4 In

contrast, internal dynamics come from the decisions by the planner. These endogenous state

changes are expressed by the system variables.

demand demand

time time

(a) Continuous demand function. (b) Discrete demands.

Figure 3.1: Continuous vs. discrete demand (in accordance to Meyr, 1999, p. 50).

As mentioned above, state changes may occur at arbitrary points of time. Correspond-

ingly, there are approaches to capture, e.g., demand and production processes (thus also

stock levels) as continuous functions over time (see Figure 3.1a). Diﬀerential equation mod-

els are formulated and solved with the help of the optimal control theory (cf. Chapter 6 by

Sethi and Thompson (2005)).

2

That is why in contrast to authors like Kuik et al. (1994), batching is not used synonymously for

lot-sizing in this thesis.

3

Note that these external dynamics may be the result of a previous planning step.

4

In rare cases even production speeds may vary due to continuous improvements of the production

process.

28 Chapter 3. Multi-level lot-sizing and scheduling

But usually, lot-sizing and scheduling models are based on a discrete time structure

(see Figure 3.1b). For discretization, the horizon is divided into T time buckets/periods by

deﬁning ﬁxed points wt , t = 1, . . . , T + 1. The time span between two consecutive points wt

and wt+1 represents the capacity Ct of such a time bucket. As illustrated in Figure 3.2 these

periods can be of equidistant length (identical capacity).

planning horizon

time time e.g. day, time

bucket bucket week, bucket

t=2 month

t=1 t=T

ഥ1 ഥ2 ഥ3 ഥT time

w w w w ഥ T+1

w

C1 = C2 = CT

But sometimes they are also of diﬀerent lengths (days, weeks, or even months). For

instance, a telescopic time scale with bigger buckets towards the end of the planning horizon

can be used as shown in Figure 3.3. This might be useful because forecasts become imprecise

the more one looks ahead.

planning horizon

e.g. weeks e.g. months e.g. quarter

Since such discrete models are solvable as MIPs, the focus lies on this kind of models in

the remainder of this thesis. Due to the scheduling part and the synchronization between

the production stages, many challenging issues arise. They are illustrated in the following

by some well-known single-level lot-sizing and scheduling models which are at the same time

the starting point for the majority of multi-level models.

In the literature, lot-sizing models are mainly distinguished between large-time bucket 5

(LTB) and small-time bucket (STB) models. To show the diﬀerences between these two

types, three basic single-level models will be presented ﬁrst—one LTB- and two (correlated)

STB-models. Moreover, three further single-level models will be introduced, as most of the

multi-level models are based on these.

5

Sometimes they are also called big-time buckets (see, for instance, Suerie (2005a)).

3.2. Single-level models with diﬀerent time structures 29

To begin with LTB-models the Capacitated Lot-sizing Problem (CLSP) is explained next.

Even though the CLSP is meant for pure lot-sizing, it helps highlight the general properties

of models with large-time buckets.

Due to its few basic assumptions, the CLSP is the mostly studied multi-item lot-sizing model

(see Dixon and Silver (1981), Maes and van Wassenhove (1988), and Trigeiro et al. (1989)

among others). The following notation is used for the CLSP:6

Indices

j = 1, . . . , J products

t = 1, . . . , T periods

Data

aj capacity consumption (time) needed to produce one unit of product j

hj holding costs per unit of product j for one period

sj setup costs for product j

djt demand for product j in period t (qty)

Ij0 initial inventory of product j at the beginning of the planning horizon

(qty)

Ct available capacity in period t (time)

Mjt suﬃciently large number for product j and period t

Variables

Ijt ≥ 0 inventory of product j at the end of period t (qty)

xjt ≥ 0 total quantity of product j produced in period t (qty)

zjt ∈ {0, 1} setup changeover: zjt = 1 if a setup changeover to product j is executed

in period t (0 otherwise)

Objective function

min hj Ijt + sj zjt (3.1)

j,t j,t

6

Note that all superscripts used in this thesis do not represent indices but indicate diﬀerent variables.

30 Chapter 3. Multi-level lot-sizing and scheduling

subject to

xjt ≤ Mjt zjt ∀j, t (3.3)

aj xjt ≤ Ct ∀t (3.4)

j

The CLSP aims at minimizing the sum of holding and setup costs (3.1). To measure the

stock levels over time, inventory balance constraints (3.2) are used. They assure that the

inventory of product j at the end of period t equals the inventory of the same product at

the end of the preceding period plus the total inﬂow on stock (production) minus the total

outﬂow from stock (demand) during period t. Furthermore, setup forcing constraints (3.3)

ensure that an item j can only be produced (xjt > 0) in period t if it is set up (zjt = 1).

Hence, these constraints connect the continuous and the binary variables. Note that for the

right-hand side, a suﬃciently large value has to be chosen—this value is often called “big

M ”—which must not limit the production quantity unnecessarily. Nevertheless, this value

should be chosen as small as possible, because this generally leads to a better performance

of a standard MIP-solver. Obviously, in the case of the CLSP, a reasonable value for Mjt is

Ct

aj

. Finally, the limited capacity has to be respected by summing up the processing times of

all production quantities (3.4).

It can be seen that up to J diﬀerent production lots might be produced in a single period t.

For this reason, such a period represents a large-time bucket 7 . Besides, the sequence of the

production lots within a period is not determined by the CLSP. Correspondingly, the setup

state is lost with the time bucket boundaries8 , since it is not known, which is the last item

in period t and the ﬁrst in period t + 1. Finally, a production lot cannot overlap several

periods but is capped by the limited capacity of a period.

Even though the CLSP is not able to determine the sequences of the production lots,

it represents a basic model whose properties are more or less common for all simultaneous

lot-sizing and scheduling models with large-time buckets.

In contrast to the CLSP, the following model allows for a simultaneous lot-sizing and schedul-

ing. According to its structure, Fleischmann (1990) calls this model the Discrete Lot-sizing

and Scheduling Problem as lot-sizes can only be an integer multiple of the production quan-

7

Sometimes they are also called big-time buckets (cf. Suerie, 2005a).

8

For this reason, a variant of the CLSP with Linked Lot-Sizes has been developed and proposed by Haase

(1994, pp. 18) which explicitly determines the products at the boundaries.

3.2. Single-level models with diﬀerent time structures 31

Index

s = 1, . . . , S periods

Data

yj0 equals 1 if product j is set up at the beginning of the planning horizon (0

otherwise)

Variables

yjs ∈ {0, 1} setup state: yjs = 1 if item j is set up and produced in period s (0

otherwise)

Objective function

min hj Ijs + sj zjs (3.5)

j,s j,s

subject to

Cs

Ijs = Ij,s−1 + yjs − djs ∀j, s (3.6)

aj

yjs ≤ 1 ∀s (3.7)

j

Again, the objective is to minimize the sum of holding and setup costs (3.5). The inven-

tory balancing constraints (3.6) assure that the demands are met. Here, the “all-or-nothing”

assumption comes into play: If the setup state for product j becomes 1 in period s, automat-

ically Cajs units of this item are produced. This is possible because constraints (3.7) enforce

a unique setup state for each period. Accordingly, at most one of the products can be set

up. Finally, the start of a new production lot is detected by (3.8) to capture the setup costs

properly.

Since the DLSP enforces a unique setup state per period, it automatically determines the

sequence of the lots. But this constraint only makes sense if the bucket lengths are rather

9

In the following, only new or changed notation elements are listed.

10

Note that a diﬀerent letter is used for the period index to emphasize the diﬀerence between LTB- and

STB-models.

32 Chapter 3. Multi-level lot-sizing and scheduling

short.11 That is why the DLSP is called a Small-Time Bucket (STB) model. By deﬁnition,

all STB-models allow at most one setup changeover in a single period. Thus, this property

allows for conserving the setup state for an arbitrary number of consecutive periods. This is

reasonable because otherwise unnecessary high setup costs would result. Correspondingly,

the size of a production lot needs not to stick to the length of a time bucket. However, the

setup state is lost as soon as no production takes place.

The all-or-nothing assumption is very restrictive. That is why it is dropped by other

models like the CSLP.

The Continuous Setup Lot-sizing Problem is very similar to the DLSP (cf. Karmarkar and

Schrage (1985) and Salomon (1991) among others). The main diﬀerence is that there is now

a continuous variable xjs for the production quantity of product j in period s. Accordingly,

the binary variable yjs = 1 only indicates whether product j in period s is set up, and not

whether it is produced.

Variables

xjs ≥ 0 production quantity of item j in period s

yjs ∈ {0, 1} setup state: yjs = 1 if product j is set up in period s (0 otherwise)

Objective function

min hj Ijs + sj zjs (3.9)

j,s j,s

subject to

Cs

xjs ≤ yjs ∀j, s (3.11)

aj

yjs ≤ 1 ∀s (3.12)

j

The objective (3.9) as well as the constraints (3.12) and (3.13) are identical to the DLSP.

The inventory balancing constraints, however, consider the production quantities xjs now.

For the CSLP, the setup forcing constraints (3.11) are at the same time the capacity con-

11

Note that the actual length of a small-time bucket strongly depends on the modeled production system.

For instance, de Matta (1994) uses an STB-model, where each “small-time” bucket comprises several days.

3.2. Single-level models with diﬀerent time structures 33

straints. This is possible because the CSLP allows at most one product to be produced in

a single period. Since each period has its unique setup state it is easy to see that a setup

state can be “conserved”, even for periods with idle time (xjs = 0 for all j). Therefore, setup

costs are only calculated for a new setup.

Another very similar STB-model is called the Proportional Lot-sizing and Scheduling

Problem. This model is examined next in more detail because several multi-level models use

the same time structure.

Drexl and Haase (1995) introduced the Proportional Lot-sizing and Scheduling Problem

(PLSP). In contrast to the CSLP this model allows the production of two diﬀerent products

in one period separated by a single setup. Nevertheless, the PLSP is quite similar to the

CSLP:

Objective function

min hj Ijs + sj zjs (3.14)

j,s j,s

subject to

Cs

xjs ≤ (yj,s−1 + yjs ) ∀s (3.16)

aj

aj xjs ≤ Cs ∀s (3.17)

j

yjs ≤ 1 ∀s (3.18)

j

Once more, the objective is to minimize the sum of holding and setup costs (3.14). The

inventory balance constraints (3.15) are identical to the ones of the CSLP. Constraints

(3.16) are the special feature of the PLSP: Item j can be produced in period s if either

it is set up in this period s or in the period before (s − 1). Accordingly, two diﬀerent

production quantities are possible within a single period. Hence, the capacity limitation is

not automatically respected by the setup forcing constraints anymore and thus (3.17) have

to be included. Finally, constraints (3.18) and (3.19) equal the ones from the CSLP (see

(3.12) and (3.13)).

It is important to note that an optimal solution of the PLSP is at least as good as the

34 Chapter 3. Multi-level lot-sizing and scheduling

optimal one of the CSLP for the same problem data. Moreover, the PLSP is generally able

to ﬁnd any feasible production plan if the length of a period is suﬃciently short.12

However, the period length is generally a critical issue for STB-models like the PLSP.

If a setup needs to be executed for the production in period s, the setup time is usually

assigned to this period, and thus it has to lie in between the two corresponding ﬁx points

ws and ws+1 (cf. Section 3.1.2). In that case, the length of a period must be greater than

or equal to the maximum possible setup time. But usually, this leads directly to a loss of

modeling details and an ineﬃcient capacity utilization might result.

To cope with this limitation, diﬀerent solution approaches are possible. The easiest one

is to “attach” the setup time to the bucket boundaries and divide it into two parts—a setup

time fraction at the end of a period and a fraction at the beginning of the following period.

However, this approach can only be used for models like the DLSP or the CSLP. In contrast,

the PLSP, which separates the two production quantities within a period by the setup, needs

additional control (binary) variables. These variables are necessary to forbid production if

a setup is “distributed” over several periods.13

More complex formulations are proposed by Kallrath (1999) and Suerie (2006) who use

additional time-indexed variables to sum up the assigned setup times. Especially the formu-

lation by Suerie is of high interest, as the approach is very general and can also be applied

for sequence-dependent setups (see also Chapter 7).

setup costs (CLSD)

Even though LTB-models like the CLSP do not determine a schedule of the production lots,

a sequencing14 of the lots can be achieved by introducing auxiliary variables and additional

constraints. This approach was used by Haase (1996) who published the Capacitated Lot-

sizing Problem with Sequence-Dependent Setup Costs (CLSD).

Index

i = 1, . . . , J products

12

However, since the lot-sizes result from the optimization it is hard to determine a reasonable period

length a priori.

13

Drexl and Haase (1995) also present an extension of the PLSP with period overlapping setup times.

However, their approach is quite inﬂexible for sequence-dependent setup times.

14

Note that we are talking about sequencing because only the “sequence” of the lots is determined. The

main diﬀerence to scheduling is that the exact point in time, when a production lot is about to start, is not

directly known. This will play an important role in the next section about the stage synchronization.

3.2. Single-level models with diﬀerent time structures 35

Data

sij setup costs of a changeover from product i to j

Variables

yjt ∈ {0, 1} yjt = 1 if product j is set up at the end of period t and at the beginning

of t + 1, respectively

zijt ∈ {0, 1} equals 1 if a changeover from product i to product j takes place in period t

bjt ≥ 0 auxiliary variable to determine the position of item j in the setup sequence

of period t

Objective function

min hj Ijt + sij zijt (3.20)

j,t i,j,t

subject to

aj xjt ≤ Ct ∀t (3.22)

j

Ct

xjt ≤ zijt + yj,t−1 ∀j, t (3.23)

aj i

yjt = 1 ∀t (3.24)

j

zijt + yj,t−1 = zjit + yjt ∀j, t (3.25)

i i

The objective of the CLSD is to minimize holding costs and sequence-dependent setup

costs. The constraints for inventory balancing (3.21) and capacity restriction (3.22) are

identical to the ones of the CLSP. Constraints (3.23), however, diﬀer as they enforce either

a corresponding changeover or the transfer of the setup state from the previous period for

the production of item j in t. (3.24) ensure that the line is set up for exactly one product at

the end of a period. Constraints (3.25) are network ﬂow conditions. Starting from a unique

setup state at the bucket boundary (or the initial state yj0 for t = 1), each item in a setup

sequence of period t has a unique predecessor in the same period. Vice versa, all items have

a unique successor item while ending with unique setup state at the next bucket boundary.

Constraints (3.26) are so-called subtour elimination constraints (cf. Miller et al., 1960).

They originate from the modeling of the Traveling Salesman Problem (TSP) which can be

36 Chapter 3. Multi-level lot-sizing and scheduling

described as follows: A traveling salesman has to visit a ﬁxed number of cities exactly once.

The cities are connected to each other and the distances between them may diﬀer. The

problem is to ﬁnd the shortest round trip through all of the cities. The analogy is that each

item, which has to be produced, needs to be “visited”.

The sequence of period t is determined by the auxiliary variables bjt , because the higher

the value, the later product j is set up. For instance, if zijt equals 1, product i has to be set

up before j and the condition bjt ≥ bit + 1 must hold. Obviously, the changeover from i to j

can take place at most once in a period. This condition is in line with most of the large-time

bucket models where only J diﬀerent production lots are possible in a single period. But

this assumption might lead to bad production plans, if the triangle inequality is not fulﬁlled

(cf. Section 2.3.3.2).15

The General Lot-sizing and Scheduling Problem (GLSP) by Fleischmann and Meyr (1997)

is one of the ﬁrst models which allow small-time buckets with variable length. Therefore,

macroperiod 1 macroperiod 2

fixed fixed

micro- micro-

period 1 period 2

variable variable

time

Figure 3.4: Example for the two-level time structure of the GLSP.

a two-level time structure is deﬁned: Again, the planning horizon is divided into T time

buckets which are called macroperiods. But now the horizon is additionally divided into S

small-time buckets called microperiods (see Figure 3.4). Each microperiod s belongs exactly

to one macroperiod t, and thus a mapping set can be deﬁned as St = {ft , . . . , ft+1 − 1}

for all t = 1, . . . , T , where ft is the ﬁrst microperiod of macroperiod t, ft+1 − 1 the last in

macroperiod t, and f1 = 1. This structure should help combine the advantages of small-time

and large-time bucket models while avoiding their disadvantages at the same time.16

Meyr (1999, pp. 75–86) gives an extended version of the GLSP. As an explanatory model,

this variant is able to describe all the models presented above (CLSP, DLSP, CSLP, PLSP

15

Menezes et al. (2011) are supposed to be the ﬁrst authors who add the feature of non-triangular setups

to the CLSP. Similar to the CLSD, they include subtour-elimination constraints to forbid disconnected

subtours. But they still stick to the production variables xjt which represent the production quantity of j

for the whole time bucket t. Accordingly, the start times of the production lots are unknown and due to the

“aggregation” the exact production quantities of items with several setups are not known either.

16

Note that Meyr (1999) calls the GLSP an LTB-model as several diﬀerent production lots are allowed

between two consecutive external state changes. By contrast, authors like Suerie (2005a) describe the GLSP

as a hybrid model.

3.2. Single-level models with diﬀerent time structures 37

and CLSD) as special cases. Moreover, it allows sequence-dependent setups, parallel hetero-

geneous lines, and a conservation of the setup state as well as its lost (by introducing the

ﬁctitious product i, j = 0). The extended GLSP looks as follows:

Indices

i, j = 0, . . . , J products (=0 means neutral)

l = 1, . . . , L production lines

s = 1, . . . , S microperiods

t = 1, . . . , T macroperiods

Index set

St set of microperiods s belonging to macroperiod t

Data

hj holding costs of product j per unit and per macroperiod

djt demand for product j in macroperiod t (qty)

alj capacity consumption (time) needed to produce one unit of product j on

line l

mlj minimum lot-size of product j (qty) if produced on line l

ml0 minimum time line l has to remain shut down

cplj production costs of product j (per item) on line l

cblj standby costs for conserving setup state j on line l

slij setup costs of a changeover from product i to product j on line l

stlij setup time of a changeover from product i to product j on production

line l

Ij0 initial inventory of product j at the beginning of the planning horizon

(qty)

ylj0 equals 1 if line l is set up for product j at the beginning of the planning

horizon (0 otherwise)

Clt available capacity on line l in macroperiod t (time)

ft ﬁrst microperiod in macroperiod t

Variables

Ijt ≥ 0 inventory of product j at the end of macroperiod t (qty)

xljs ≥ 0 quantity of product j produced in microperiod s on line l (qty)

xljs ≥ 0 standby time (conservation of setup state j on line l without production)

in microperiod s

38 Chapter 3. Multi-level lot-sizing and scheduling

otherwise)

zlijs ≥ 0 equals 1 if a changeover from product i to j takes place at the beginning

of microperiod s on line l (0 otherwise)

Objective function

min hj Ijt + slij zlijs + cplj xljs + cblj xljs (3.27)

j,t l,i,j,s l,j,s l,j,s

subject to

Ijt = Ij,t−1 + xljs − djt ∀j > 0, t (3.28)

l,s ∈ St

(alj xljs + xljs ) + stlij zlijs = Clt ∀l, t (3.29)

j,s ∈ St i,j,s ∈ St

Clt

alj xljs + xljs ≤ yljs ∀l.j, s (3.30)

alj

xljs ≥ mlj (yljs − ylj,s−1 ) ∀l, j, s (3.31)

yljs = 1 ∀l, s (3.32)

j

Similar to the preceding models, the GLSP aims at minimizing the sum of holding and

setup costs (3.27). However, since parallel lines are included, it is necessary to consider

production costs. Moreover, the model captures standby/idle times which also might impose

costs. Constraints (3.28) assure that the demands are met, where the production quantities of

all microperiods belonging to the regarded macroperiod represent the inﬂow. The size of the

microperiods and thus the corresponding production quantities are limited by the capacity

of the macroperiod (3.29). Constraints (3.30) force a corresponding setup for production (or

standby) in each microperiod. Furthermore, minimum lot-sizes are introduced (3.31). They

help to avoid setup changeovers without product changes. This is necessary, as the triangle

inequality needs not to be fulﬁlled by the setup costs (siv + svj < sij ) and the “optimization”

makes use of this property: The changeover from i to j would always go via product v—

because it is cheaper—even though no item of product v is actually made. For this and

further technical reasons, minimum lot-sizes might become necessary.

The GLSP formulation has several advantages: First of all, it shows an easy way of

modeling simultaneous lot-sizing and scheduling problems. In addition, the variable period

lengths allow for long setup times (compared to the PLSP). Furthermore, if the triangle

inequality is not fulﬁlled (see page 18), it might be the case that a product is setup several

3.3. Multi-level models in the literature 39

times in a period. This cannot be modeled via TSP constraints without further ado.

The main disadvantage of the GLSP is that the optimal solution of the corresponding

LP (the integer condition is dropped/relaxed) often shows high gaps to the optimal incum-

bent solution. That is why a standard MIP-solver often has diﬃculties with the GLSP.

Furthermore, if the number of microperiods is too large, the GLSP might have many sym-

metric optimal solutions (microperiods with zero length). Therefore, additional constraints

are needed which force these periods to be at the end of a macroperiod (cf. Fleischmann and

Meyr, 1997). That is why a good choice of the number of microperiods is crucial for solving

the model (see also Section 4.3.2).

The PLSP, the CLSD as well as the GLSP are all starting ground for several multi-level lot-

sizing and scheduling models which are described in this section. But ﬁrst, the multi-level

variant of the CLSP is introduced. Even though this model shows no scheduling capability

(just like the single-level CLSP), it helps demonstrate how a feasible material ﬂow between

the levels can be achieved.

Similar to the single-level variant, the Multi-Level Capacitated Lot-sizing Problem (ML-

CLSP) is probably the most prominent multi-level lot-sizing model (cf. Maes et al. (1991)

and Stadtler (1996) among others). Therefore, the synchronization of the quantity ﬂows

between the levels is explained with its help.

sent ﬁnished products and i, j = J0 + 1, . . . , J the corresponding pre-products.17 For each

product j, the index set Nj contains all direct successor products of j. Accordingly, this

set is empty for all ﬁnished products. The corresponding gozinto factor pji describes the

quantity of item j which is necessary for producing one unit of i. Furthermore, demand is

diﬀerentiated between external (primary) demand18 and internal (secondary) demand. The

internal one arises from the production of a successor product.

Since the MLCLSP is meant for job shop planning, the products are processed on L

diﬀerent machines (not production lines). Moreover, each item can only be produced on

17

The items are sorted according to their low level code in the bill-of-materials. For a better understanding

of the low level coding see also Section 2.3.3.4 on page 20.

18

Sometimes it is also called independent demand (cf. Pochet and Wolsey, 2006, p. 47).

40 Chapter 3. Multi-level lot-sizing and scheduling

Indices

t = 1, . . . , T periods

l = 1, . . . , L machines

Index set

Nj set of all direct successor items of product j

Data

hj holding costs of product j per unit and per period

sj costs for setting up product j

stlj setup time for product j on machine l

mj minimum lot-size of product j (qty)

djt primary demand for product j in period t (qty)

Ij0 initial inventory of product j at the beginning of the planning horizon

(qty)

pij amount of item i that is needed to produce one unit of item j (gozinto-

factor)

Δ minimum lead-time oﬀset

Clt available capacity on machine l in period t (time)

Variables

Ijt ≥ 0 inventory of product j at the end of period t (qty)

xjt ≥ 0 quantity of product j produced in period t (qty)

zjt ∈ {0, 1} setup state: zjt = 1 if product j is set up in period t (0 otherwise)

Objective function

min hj Ijt + sj zjt (3.34)

j,t j,t

subject to

Ijt = Ij,t−1 + xjt − djt − pji xi,t+Δ ∀j, t (3.35)

i ∈ Nj

(alj xjt + stlj zjt ) ≤ Clt ∀l, t (3.36)

j

3.3. Multi-level models in the literature 41

The multi-level CLSP also aims at minimizing setup and holding costs (3.34). To syn-

chronize the diﬀerent levels, the “secondary” demand, which is induced by the production

of a direct successor item, is integrated to the inventory balance constraints (3.35). Corre-

spondingly, the stock level of product j at the end of period t equals the stock level in period

t − 1 plus the production quantity of period t minus the primary and the secondary demand

in that period t. The secondary demand is induced by the production of successor i in period

t + Δ (must be ≤ T ). To determine all secondary demands, the index set Nj is needed which

contains all direct successor products of j. Furthermore, Δ describes the lead-time oﬀset,

before a product is available for the next stage. Finally, the available capacities must be

respected (3.36) as well as the corresponding setups for production have to be forced (3.37).

Since the MLCLSP is used to solve job shop production problems, lot-sizing is executed

separately from scheduling on a more aggregate level. Moreover, in a job shop manufacturing

environment, setup times are rather short and lot-sizes usually very small. Accordingly,

authors like Stadtler or Maes and van Wassenhove assume that the periods are suﬃciently

long to abandon lead-times (Δ = 0). However, as soon as a more detailed view becomes

necessary and thus the sequence of the production lots becomes important, the feasibility

of the quantity ﬂows between the production stages has to be ensured. The easiest way to

handle this synchronization issue is to use the starting/ending times of the time buckets,

because it is known that the production quantity of item j in period t is available by wt+1

(cf. Figure 3.2). Accordingly, the quantity can be used on the next level in period t + 1

(Δ = 1).19

For the modeling of a ﬂow line production system (as considered in this thesis), lead-

times of one large-time bucket are unrealistically high. But even lead-times of one small-time

bucket might not be suﬃciently short. Accordingly, a zero lead-time (Δ = 0)—this means

that production quantities are transfered to the next stage within the same period—is to be

aspired.

In the following, multi-level models existing in the literature are presented and examined.

According to their time structure, the models are subdivided into PLSP-, CLSD-, and GLSP-

based models as well as other models.

Several multi-level models in the literature are based on the PLSP. One variant presented

by Haase (1994) regards a multi-level product structure and a single bottleneck machine.

This model synchronizes the production stages via the inventory balancing constraints which

19

This approach, for instance, is chosen by Tempelmeier and Derstroﬀ (1996) who consider the operations

on the resources in detail.

42 Chapter 3. Multi-level lot-sizing and scheduling

account for primary and induced secondary demands as well as product-speciﬁc lead-times

of at least one period. All other constraints are equivalent to those of the single-level PLSP.

This means that setup states and the corresponding costs are only considered for products

on that bottleneck machine. Similar to the PLSP, setup times are disregarded and setup

costs are sequence-independent.

A very similar approach is introduced by Kimms (1996) who also focus on a single

bottleneck machine. Moreover, the quantity ﬂows between the stages are synchronized with

the help of the inventory balancing constraint as well, but without considering any lead-times.

Instead, Kimms introduces additional constraints to ensure that pre-items are available on

time. This modiﬁcation is supposed to lead to a better computation of holding costs and

a reduction of end horizon eﬀects.20 The same approach is used by Kimms (1997) for the

multi-level PLSP with Parallel Machines. Since this model allows parallel heterogeneous

machines, it will be presented in a more detail next.

Another PLSP variant is introduced by Stadtler (2011) who includes setup times and

allows zero lead-times while regarding a single production line. This model will be shown in

the following as well.

(PLSP-PM)

The Proportional Lot-sizing and Scheduling Problem with Parallel Machines (PLSP-PM)

by Kimms (1997) allows for considering a multi-level product structure, but only a single

production stage with parallel heterogeneous machines. Moreover, each item is produced

on an item-speciﬁc machine without a ﬁxed machine assignment. Accordingly, a new index

set Lj is needed which contains all machines, where product j can be produced on. This

set is complemented by a second new index set Jl consisting of all products which can be

produced on machine l. This leads to the following formulation:

Index

s = 0, . . . , S periods

τ = 1, . . . , S periods

Index sets

Jl set of all products that are producible on line l

Lj set of production lines where item j can be produced on

20

For instance, ﬁnal inventories are usually optimized to zero, as they are not needed for satisfying

demand. However, this eﬀect might have an impact on the entire solution.

3.3. Multi-level models in the literature 43

Data

Δj integer lead-time of product j (at least 1 period)

Variables

xljs ≥ 0 quantity of product j produced in period s on machine l (qty)

yljs ∈ {0, 1} setup state: yljs = 1 if product j is set up in period s on machine l (0

otherwise)

zljs ∈ {0, 1} equals 1 if there is a changeover for item j in period s on machine l (0

otherwise)

Objective function

min hj Ijs + slj zljs (3.38)

j,s l j ∈ Jl s

subject to

Ijs = Ij,s−1 + xljs − pji xlis − djs ∀j, s > 0 (3.39)

l ∈ Lj i ∈ Nj l ∈ L i

min{s+Δ

j , S}

Ijs ≥ pji xliτ ∀j, s < S (3.40)

i ∈ Nj l ∈ L i τ =s+1

alj xljs ≤ Cls ∀l, s > 0 (3.41)

j ∈ Jl

Cls

xljs ≤ (ylj,s−1 + yljs ) ∀l, j ∈ Jl , s > 0 (3.42)

alj

yljs ≤ 1 ∀l, s > 0 (3.43)

j ∈ Jl

Once more the elements of the objective function are holding and setup costs (3.38).

Although the inventory balance constraints (3.39) do not include any lead-time oﬀset, con-

straints (3.40) ensure that a suﬃcient number of pre-items is in stock on time with respect

to the item-speciﬁc lead-time. With the help of (3.41) the limited production capacities are

taken into account. Constraints (3.42)–(3.44) equal the ones of the single-level PLSP (cf.

page 33), except for the fact that parallel heterogeneous lines have to be considered.

variant of the PLSP with a single production line. This line is “able” to process all items

44 Chapter 3. Multi-level lot-sizing and scheduling

unacceptably large in this practical case, Stadtler implements zero lead-times. Before the

corresponding model formulation is given, the diﬃculties with zero lead-times arising for

the PLSP will be outlined ﬁrst. Moreover, Stadtler uses “net demands” which will also be

explained beforehand.

Zero lead-times In order to demonstrate the diﬃculties of zero lead-times for the PLSP,

the example for shifting bottlenecks introduced in Section 2.3.3.2 is used again. Suppose

that only the inventory balance constraints are extended for the multi-level structure. A

possible solution is given in Figure 3.5 and Table 3.1. Although the production of item Y

starts as early as possible and the packaging to B ends as late as possible, the material

ﬂow is obviously not feasible as X is packaged before it was produced. Even though this

example makes use of diﬀerent lines, a similar result occurs for the real-world case considered

by Stadtler. Here, a successor item could be produced while its predecessor is actually not

available yet.

period s

stage A B

j=A 1 0 13

P

making j=B 0 1 8

stage Y X

j=X 0 1 13

M

j=Y 1 0 8

time

Figure 3.5: PLSP with zero lead-times. Table 3.1: PLSP solution values.

Stadtler solves the real-world problem by introducing two additional groups of variables

to diﬀerentiate between the production quantities xαjs of product j at the beginning and the

quantities xωjs at the end of period s. Applied to the example of Figure 3.5 and Table 3.1, it

means that the quantity of j = Y is represented by xαjs and of j = X by xωjs .

Demand netting As mentioned before, the big M value should be chosen carefully, be-

cause it has a signiﬁcant impact on the computational performance of a MIP-solver (see also

Chapter 5). A quite intuitive approach is to use the maximum quantity of producible units

(capacity divided by the processing time of one item) as shown in some models before. How-

ever, this value could be even “tighter” if the actual demand is considered as well. Obviously,

it makes no sense to produce more than demanded in total.

Accordingly, it is useful to execute a net demand calculation for the products as illustrated

in Algorithm 1. It is assumed that all products are sorted by their low level code (cf.

Section 2.3.3.4), thus products j = 1, . . . , J0 are ﬁnal items. Because of the level coding,

the condition max{i | i ∈ Nj } < j holds for all pre-products j. Hence, the demand can be

3.3. Multi-level models in the literature 45

determined for each product in an ascending order by summing up the primary and the

secondary demands.

Algorithm 1: GetNetDemands

input: Primary demands djs , gozinto factors pji , successor sets Nj and initial

inventories Ij0

output: Net demands dnjs

1 for j ← 1 to J do

2 aQty ← Ij0 ;

3 for s ← 1 to S do

4 dnjs ← max{0, djs + i ∈ Nj pji dnis − aQty};

5 aQty ← max{0, aQty − djs − i ∈ Nj pji dnis };

6 end

7 end

Starting with period 1, the available quantity corresponds to the initial stock level. Then,

for each period, the net demand is computed as the sum of primary and secondary demand

minus the available quantity, where this quantity must be reduced by the consumed share

afterwards.21

Model formulation Based on the idea outlined above and the computed net demands,

the following model helps to solve the real-world problem:

Data

dnjs net demand of product j in period s

Variables

xαjs production quantity of product j at the beginning of period s

xωjs production quantity of product j at the end of period s

Objective function

min hj Ijs + sj zjs (3.45)

j,s j,s

subject to

Ijs = Ij,s−1 + xjs − pji xis − djs ∀j, s (3.46)

i ∈ Nj

21

Note that the max function is needed to avoid negative values for the net demand and the available

quantity.

46 Chapter 3. Multi-level lot-sizing and scheduling

aj xjs + stj zjs ≤ Cs ∀s (3.48)

j j

Cs n

S

xαjs≤ min , d · yj,s−1 ∀j, s (3.49)

aj τ =s jτ

Cs n

S

xjs ≤ min

ω

, d · zjs ∀j, s (3.50)

aj τ =s jτ

Ij,s−1 ≥ pji xαis ∀j ≥ J0 + 1, s (3.51)

i ∈ Nj

Ijs + xαjs ≥ pji xωis ∀j ≥ J0 + 1, s (3.52)

i ∈ Nj

yjs ≤ 1 ∀s (3.54)

j

The objective (3.45) is the same again as for the other PLSP models. The inventory

balance constraints (3.46) show zero lead-times now. According to the introductory remarks,

the total production quantity xjt of item j in period t consists of the two fractions xαjs and

xωjs .

As setup times are included in the model, they also consume capacity and have to be

considered in constraints (3.48).22 With the help of the setup forcing constraints (3.49) and

(3.50), the production quantities are properly connected to the corresponding setup state.

At this place, Stadtler makes use of the calculated net demands. Constraints (3.51) assure

that for a production lot at the beginning of period t the needed pre-products are already in

stock at the end of period t − 1. Analogously, constraints (3.52) guarantee that the available

quantity of product j in period t (inventory plus produced items in that period) is suﬃcient

for the production of a successor item at the end of period t. The constraints for the setup

state (3.53), (3.54), and (3.55) are the same as for the single-level PLSP.

The current model allows solve the case for a single production line. But to generalize

this approach for an arbitrary number of production lines, further additional variables and

constraints would be necessary because the start and end times of the production lots are

not known yet by the model. Stadtler and Sahling (2013) present such a formulation with

zero lead-times for multiple stages and lines. This model, which is called Multi-Stage PLSP

without a ﬁxed lead-time oﬀset (MS-PLSP-LT0), comprises about 40 groups of constraints.

Setups are still sequence-independent.

22

Here, it becomes apparent again that the period length has to exceed the maximum setup time. There-

fore, in a second step, Stadtler extends his formulation for period overlapping setup times by using the

formulation presented by Suerie (2005b).

3.3. Multi-level models in the literature 47

In the literature, there are also multi-level variants of the CLSD. Within the scope of their

theses, Grünert (1998) as well as Sahling (2010) propose such models for multiple production

stages. Both models are presented next.

(MSLS)

Grünert (1998, pp. 47–53) developed the Multi-level Sequence-dependent Lot-sizing and

Scheduling problem (MSLS) which should be generally applicable for diﬀerent industries—

especially for the process industry. Grünert diﬀerentiates between two diﬀerent kinds of

lead-time: Δjj comprises the lead-time, which is directly caused by the production process

(e.g., by drying), whereas Δij describes the complete lead-time oﬀset, until product i can

be processed to product j (e.g., including transportation time). Moreover, he assumes that

the corresponding planning is executed in a rolling manner and that each product has a

speciﬁc planning/forecast horizon. That is why each product (group) j has its own planning

horizon which starts in a period ψj and ends in period ψj + T . The index set Ψj deﬁned as

{ψj , . . . , ψj + T } contains all these planning periods. Each product can only be processed on

a single machine (|Lj | = 1 for all j). In contrast, a machine l can produce several diﬀerent

items (|Jl | ≥ 1), thus the corresponding index set of planning periods equals Πl := j ∈ Jl Ψj .

The MSLS can be formulated as follows:23

Index

u = 1, . . . , T periods

Index sets

Πl set of periods for which line l is to be planned

Ψj set of periods for which product j is to be planned

Data

yj0 equals 1 if product j is set up at the beginning of the planning horizon (0

otherwise)

Δjj process lead-time of item j

Δij process lead-time of item i plus transportation lead-time needed to be

available for item j

πl the ﬁrst planning period of line l (= min{t | t ∈ Πl })

23

Note that the original MSLS is slightly adapted to get a consistent formulation with nonredundant

constraints.

48 Chapter 3. Multi-level lot-sizing and scheduling

cpjt production costs of item j in period t

colt costs for overtime of line l in period t

omax

lt maximum overtime possible for line l in period t

x0jt production quantity of item j in period t determined in the previous model

run

Ij0 ﬁnal stocks of item j determined in the previous model run

Clt available capacity on line l in period t (time)

Variables

yjt ∈ {0, 1} yjt = 1 if product j is set up at the end of period t and at the beginning

of t + 1, respectively

zijt ∈ {0, 1} equals 1 if a changeover from product i (yi,t−1 = 1) to product j (yjt = 1)

takes place

olt used overtime for line l in period t

bjt ≥ 0 auxiliary variable to determine the position of item j in the setup sequence

of period t

Objective function

min hjt Ijt + s0j yj,πl −1 + slij zijt

j,t j l ∈ Lj l i ∈ Jl j ∈ J l t ∈ Π l

+ cpjt xjt + colt olt (3.56)

j t ∈ Ψj l t ∈ Πl

subject to

ψi −Δij +T

−dj,t+Δjj − pij xiu ∀j, t ∈ Ψj (3.57)

i ∈ Nj u=ψi −Δij

aj xjt + stij zijt ≤ Clt + olt ∀l, t ∈ Πl (3.58)

j ∈ Jl t ∈ Ψj i ∈ Jl : i=j

olt ≤ omax

lt ∀l, t ∈ Πl (3.59)

xjt ≤ Mjt zijt + yj,t−1 ∀l, j ∈ Jl , t ∈ Ψj (3.60)

i ∈ Jl :i=j

yj,t−1 + zijt = zjit + yjt ∀l, j ∈ Jl , t ∈ Ψj (3.61)

i ∈ Jl i ∈ Jl

3.3. Multi-level models in the literature 49

yjt = 1 ∀l, t ∈ {πl − 1} ∪ Πl (3.62)

j ∈ Jl

Ij,ψj +Δjj −1 = Ij0 ∀j (3.64)

xjt = x0jt ∀l, j ∈ Jl , t ∈ Πl \Ψj (3.65)

In contrast to the previous multi-level models, not only considers the MSLS setup and

holding costs but also includes production costs. These costs are relevant, because they

are time-speciﬁc and inﬂuence the decision about when to produce. Furthermore, overtime

costs have to be considered as well. Constraints (3.57) ensure the inventory balance while

considering the lead-times of the products. Capacity constraints (3.58) must be satisﬁed

where capacity can be extended by overtime. Since overtime is not unlimitedly available,

it is capped by (3.59). Constraints (3.60)–(3.63) are more or less equal to (3.23)–(3.26) of

the single level model with the small diﬀerence of regarding several machines here. For all

planning periods of a machine, where product j was already planned in a previous model

run, the corresponding old solution values have to be set as starting inventories (3.64) and

as production quantities (3.65). The latter is necessary because these quantities might be

used as pre-products for items, whose planning horizon starts earlier.

In sum, the MSLS considers sequence-dependent setup times and costs. Moreover, it

includes product-speciﬁc lead-times of at least one period which automatically guarantee a

feasible material ﬂow between the stages.

Setup Costs using Parallel Machines at Multiple Locations (MLCLSD-

PM-ML)

Sahling (2010) solves a real-world problem of a company from the consumer goods industry.

This company produces diﬀerent packaged goods in a two-staged process with parallel pro-

duction and packaging lines at several locations. However, Sahling generalizes this practical

case. The resulting Multi-Level Capacitated Lotsizing Problem with Sequence-Dependent

Setup Costs using Parallel Machines at Multiple Locations (MLCLSD-PM-ML) allows mul-

tiple production stages with parallel heterogeneous lines. Moreover, not only considers the

model the production quantities at the diﬀerent locations but also the transportation ﬂows

between the locations.

Since rather small production quantities are planned, Sahling uses integer variables for

the lot-sizes to avoid a loss of planning accuracy potentially caused by rounding errors.

Indices

p, q = 1, . . . , P locations

50 Chapter 3. Multi-level lot-sizing and scheduling

Index set

Lpj set of lines at location p where product j can be produced on

Data

hjp holding costs per period for one unit of product j at location p

slij setup costs of a changeover from product i to product j on line l

stlij setup time of a changeover from product i to product j on line l

crpq ﬁxed transportation costs for one delivery from location p to q

Δpj production lead-time of item j (at least 1 period)

Δrjp lead-time for the transportation of product j to location p (at least 1

period)

gj volume of item j

Cpr capacity of one vehicle at location p (qty)

Ipmax storage capacity at location p (qty)

Mjt suﬃciently big number

Variables

δjpt ∈ Z+

0 auxiliary variables for the demand for product j which is met at location

p in period t

npqt ∈ Z+ 0 number of transports between locations p and q in period t

xljt ∈ Z+0 production quantity of item j on line l in period t

zlijt ∈ {0, 1} equals 1 if a changeover from item i to j on line l is executed in period t

xrjpqt ∈ Z+0 transportation quantity of item j transferred from location p to q in pe-

riod t

Objective function

min cplj xljt + hjp Ijpt + slij zlijt + colt olt + crpq npqt (3.66)

l j ∈ Jl t j,p,t l i,j ∈ Jl t l,t p,q,t

subject to

Ijpt = Ijp,t−1 + xlj,t−Δpj + xrjqp,t−Δrjp

l ∈ Lpj q : q=p

− ali xlit − xrjpqt − δjpt ∀j, p, t (3.67)

i ∈ Nj l ∈ Lpi q : q=p

δjpt = djt ∀j, t (3.68)

p

3.3. Multi-level models in the literature 51

alj xljt + stlij zlijt ≤ Clt + olt ∀l, t (3.69)

j ∈ Jl i ∈ Jl

gj xrjpqt ≤ Cpr npqt ∀p, q, t (3.70)

j

xljt ≤ Mjt · zlijt + yljt ∀l, j ∈ Jl , t (3.71)

i ∈ Jl

yljt = 1 ∀l, t (3.72)

j ∈ Jl

yljt + zlijt = zljit + ylj,t+1 ∀l, j ∈ Jl , t (3.73)

i ∈ Jl i ∈ Jl

gj Ijpt ≤ Ipmax ∀p, t (3.75)

j

The objective (3.66) is to minimize production, holding, and setup costs as well as costs

for overtime and transportation. For each product j at each location p, the inventory balance

has to be maintained (3.67). Therefore, the quantity inﬂow by the production at the current

location as well as by the transportation from the other locations have to be considered.

Moreover, product demand can be met by any location and the corresponding outﬂow has

to be captured as well. Accordingly, constraints (3.68) ensure that the external demand is

fulﬁlled centrally from the locations altogether. Constraints (3.69) forbid that production

and setup times exceed the limited capacity of a production line which is possibly extended

by overtime. Furthermore, the transportation capacity has to be respected as well (3.70).

Constraints (3.71)–(3.74) are deﬁned analogously to the single level CLSD and the MSLS,

respectively. Finally, the limited storage capacities at each location are considered by (3.75).

Both CLSD-based models in this section use lead-times of at least one period. If a zero

lead-time is implemented with the inventory balance, the arising problem is similar to the

one of the PLSP. Then, additional variables are needed which reﬂect the positions in the

sequence and the starting times of the production lots. The resulting formulation may show

strong similarities to the GLSP. Correspondingly, it is not surprising that several multi-level

models are also based on the GLSP.

The ﬁrst multi-level GLSP is proposed by Meyr (2004) who forces the start times of the

microperiods to be identical for all production lines and stages.24 The resulting General

24

Another approach to synchronize the microperiods is used by Almeder and Almada-Lobo (2011) who

consider scarce setup resources but only a single stage production. That is why this model is excluded from

this review.

52 Chapter 3. Multi-level lot-sizing and scheduling

Lot-sizing and Scheduling Problem for Multiple production Stages (GLSPMS) is tailored to

ﬂow line production systems and allows an exact synchronization of the stages. For these

reasons, this model will be examined in detail in the following.

Another GLSP-based multi-level model is presented by Fandel and Stammen–Hegener

(2006) who try to calculate inventory holding costs more accurately. This attempt leads to

a non-linear model formulation, and thus is not in the scope of this review.

Marinelli et al. (2007) regard the two-stage process of producing and packaging yoghurt.

In the ﬁrst stage, tanks are ﬁlled with diﬀerent mixtures of yoghurt. Due to the cleaning

of the tanks and the pipes, high setup times occur for changing the yoghurt. In the second

stage, the yoghurt is packaged on several lines in parallel. But since the model neglects

holding costs, it is not considered any further.

In the Brazilian soft-drink industry, the production system is similar to the one above of

the yoghurt production. In the ﬁrst stage, soft drinks with diﬀerent ﬂavors are prepared and

stored in several tanks. These tanks feed diﬀerent production lines which ﬁll the various bot-

tles. On both levels, setup times and costs are sequence-dependent (see also Section 2.3.3.2

on page 17).

To solve this problem, Toledo et al. (2007) propose a model formulation called Synchro-

nized and Integrated Two-Level Lot-sizing and Scheduling Problem (SITLSP) which is a

combination of the CSLP and the GLSP. The planning horizon is divided into macroperi-

ods. For the ﬁrst stage, each macroperiod is again divided into microperiods of a ﬁxed size

according to the CSLP. For the next level, microperiods of variable size as for the GLSP are

deﬁned. Due to the diﬀerent time structures, both stages have to be explicitly synchronized

which results in more than 30 groups of constraints. In contrast, Ferreira et al. (2009) regard

the same problem but they present a reduced model formulation of the SITLSP called the

two-stage multi-machine lot-scheduling model—P2SMM. Due to its practical relevance, the

P2SMM will be also explained at length.

Araujo et al. (2008) extend the GLSP for a second production stage to solve the produc-

tion planning problems of small foundries. These foundries transform metal into alloys in

the ﬁrst stage. In the second stage, the ﬁnal items are molded from the liquid alloy which

must be directly processed and cannot be held in stock. Since the furnace shows a limited

capacity and a new setup leads to an additional loss of capacity, sequence-dependent setups

are regarded. Furthermore, the model includes back-logging.

Lang (2009) presents an extension of GLSPMS. This extension is based on a “State-

Task-Network” approach which allows the modeling of ﬂexible production sequences and

ﬂexible BOMs. Lang calls it the The General Lot-sizing and Scheduling Problem for Multiple

production Stages with Flexible Production Sequences (GLSPMS-FPS).

Finally, Transchel et al. (2011) propose a GLSP-based model to solve a mid-term pro-

duction planning problem with two production stages in the process industry.

3.3. Multi-level models in the literature 53

3.3.4.1 The General Lot-sizing and Scheduling Problem for Multiple production

Stages (GLSPMS)

Meyr (2004) introduces the General Lot-sizing and Scheduling Problem for Multiple produc-

tion Stages (GLSPMS). The model uses a common time structure for all production lines at

all productions stages which allows short lead-times despite of realistically long setup times.

This structure is explained next followed by the model formulation.

Common time structure The idea of the common time structure is that each period s

starts at the same point in time ws for all production lines and production stages. Only

the starting times of a few a priori selected microperiods are ﬁxed. The time span between

two consecutive ﬁxed microperiods can now be interpreted as (the length of) a macroperiod.

That ﬁxing allows us to model capacity,25 demand, and holding costs based on macroperiods.

The remaining non-ﬁxed microperiods are called “free” periods because their starting times

can be determined as decision variables of the model. As a consequence, the time span

between two consecutive microperiods is of variable length.

Each microperiod can represent setup, production, and standby activities, i.e. the length

of a microperiod can contain shares of setup time, production time, and idle time. In order

to achieve a higher capacity utilization with a smaller number of microperiods, Meyr uses

“setup splitting” to allow period-overlapping setup times and “quantity splitting” to realize

zero lead-times.

Setup splitting means that setup time can overlap two consecutive periods where one

part might be assigned to the beginning of the current period and the remaining part to

the end of the directly preceding period. Quantity splitting allows a share of a production

quantity to be input for another product in the same microperiod (zero lead-time), whereas

the remaining part can be used by a successor product in the following microperiod at the

earliest.

Correspondingly, as shown in Figure 3.6, each microperiod may consist of ﬁve diﬀerent

time spans: a setup time fraction at the beginning of the microperiod, production being

available in the same microperiod, idle time, production being available in the subsequent

microperiod, and a setup time fraction at the end of the microperiod. Figure 3.7 shows a

possible solution of this approach where the entire time structure can be illustrated. For

instance, in Figure 3.7 the starting times w1 and w5 of microperiods s = 1 and s = 5 are ﬁxed.

Accordingly, the length and thus capacity C1 of the ﬁrst macroperiod equals the diﬀerence

25

Here, capacity means available working time.

54 Chapter 3. Multi-level lot-sizing and scheduling

w5 − w1 . The starting times w2 , w3 , and w4 are a result of the optimization process. Thus,

the lengths of all four microperiods belonging to the macroperiod can be chosen (almost)

freely. However, their sum has to equal the length of the macroperiod. On line l = 2 the

setup time of the ﬁrst changeover is split so that a fraction belongs to microperiod s = 1 and

the rest to s = 2. Moreover, on the same line in microperiod s = 3 the production quantity

of item 5 is split. The ﬁrst part is directly moved to the next stage, whereas the second part

is available in the subsequent microperiod at the earliest (causing work-in-process (WIP)

stock). In order to further clarify the impact of setup and quantity splitting, Figure 3.8

w1 w2 w3 w4 w5

l=2 j=4 j=5 j=5 j=6

time

s=1 s=2 s=3 s=4 s=5

starting time: fixed free free free fixed

length of period:

fixed

shows the enhancements outlined above. If splitting was not allowed, the result might look

like Figure 3.8a, where the complete changeover has to be executed within microperiod s and

the production of the required pre-product i must ﬁnish on time. However, Figures 3.8b–d

demonstrate the improvements. In Figure 3.8b setup splitting is allowed, thus, part of the

changeover can be postponed to microperiod s + 1 and idle time can be reduced. Quantity

splitting makes it possible to keep on producing item i, even if it is not needed in the same

microperiod in the next stage (Figure 3.8c). Unnecessary idle times could then be avoided.

Simultaneous setup and quantity splitting further condense the schedule leading to an even

higher utilization and a more realistic model of the real world problem (Figure 3.8d).

s-1 s s-1 s s-1 s s-1 s

i i i i i i

(a) No splitting. (b) Setup splitting. (c) Quantity splitting. (d) Quantity and

setup splitting.

Figure 3.8: Tighter plans with setup and quantity splitting (in accordance with Meyr, 2004).

Note that this common time structure allows that zero lead-times can be achieved on a

microperiod base by the inventory balance. Hence the signiﬁcant “macroperiod lead-times” of

3.3. Multi-level models in the literature 55

multi-stage LTB models can be eliminated. Besides, setup times can reach twice the length

of a macroperiod because of the setup time splitting.26 In addition, this time structure

makes it possible to model working time calendars or non-operation periods for maintenance

purposes as well.

over a ﬁnite planning horizon. Note that the assignment of production lines to production

stages will only indirectly be captured by the gozinto factors pji , denoting the quantity of

predecessor item j necessary to produce one unit of successor item i, and the production

coeﬃcients alj , denoting the production time needed to produce one unit of item j on line l

(alj = 0 if j cannot be produced on l). Setup costs and times for changing from a product

i to another product j on line l are expressed by slij and stlij , respectively. By introducing

a ﬁctitious product 0, Meyr is able to model a “neutral” setup state, i.e. shutting down a

line for costs sli0 , remaining idle, and starting up a line for cost sl0j again. In this case the

setup state gets lost during the idle periods (without production). Note that a conservation

of the last setup state during idle periods, as it was assumed in Figures 3.6 and 3.7, can

also be represented for physical products j > 0. This can be interpreted as holding line l in

a standby mode for producing the same product again without incurring additional setup

costs but possibly time-dependent standby costs cbl .

Capacity Ct , demand dmacrojt and inventory holding costs hmacro

jt are deﬁned per macrope-

riod t. Similar to the GLSP, the planning horizon is additionally subdivided into S micrope-

riods s = 1, . . . , S with variable length, whose starting times are represented by the decision

variables ws . The starting times of some of these periods, described by the set Φ, are ﬁxed in

advance. This is, for example, the case for microperiods, which mark the beginning (or end)

of a macroperiod, but can also be used to set ﬁxed downtimes of machines, e.g., during holi-

days or for maintenance. As outlined in Section 3.1, to model T macroperiods, T + 1 points

in time are needed. The same is true for the microperiods and thus an artiﬁcial macroperiod

t = T + 1 as well as an artiﬁcal microperiod s = S + 1 are introduced. Furthermore, let ft be

the ﬁrst microperiod of macroperiod t again. Then, the capacity Ct of a “real” macroperiod

t can be expressed by the diﬀerence ws − ws of the ﬁxed starting times of two subsequent

“ﬁrst” microperiods s = ft and s = ft+1 . Analogously, demand dmacro jt has to be met until

the end of the last microperiod of macroperiod t which corresponds to ft+1 − 1. For ease

of notation, in the following Λ denotes the set of all of these “last” microperiods. In order

to transform an LTB-model into this GLSPMS time structure, demand djs has to be set to

zero for all s ∈

/ Λ and to dmacro

jt otherwise. In the same way the holding costs hjs are deﬁned.

Furthermore, for each time interval shown in Figure 3.6, decision variables have to be

introduced. Accordingly, xbls denotes the setup time fraction at the beginning of microperiod s

on line l which is caused by a changeover. Analogously, xels is deﬁned as the setup time

26

Note that this is only possible if a microperiod becomes as long as the corresponding macroperiod.

56 Chapter 3. Multi-level lot-sizing and scheduling

fraction at the end of microperiod s. The production time is computed by alj · xljs where

xljs describes the quantity of item j produced on line l in microperiod s. Following the

ljs and xljs , whose production

quantity splitting idea, xljs will further be subdivided into x

times are the second and fourth element of Figure 3.6. Finally, the idle time xls represents

the remaining usage (or better non-usage) of line l in microperiod s (shutdown time for j = 0

and standby time for j > 0).

For instance, in Figure 3.7 the microperiod s = 2 starts with idle time x12 > 0 on line

l = 1. On line l = 2, item j = 5 is produced over the entire microperiod s = 3. However,

253 and x253 since the quantity x253 is not needed

this production quantity x253 is split into x

before microperiod s = 4. Besides, a splitting of the setup time occurs on line l = 2 in

microperiods s = 1 and s = 2. Therefore, xe21 > 0 and xb22 > 0.

The basic version of the GLSPMS can be summarized as follows:

Indices

i, j = 1, ..., J products, whereas = 0 means neutral

k, l = 1, ..., L production lines (multi-stage and/or parallel)

t = 1, ..., T macroperiods (e.g. weeks, months)

= T + 1 artiﬁcial macroperiod

s = 1, ..., S microperiods

= S + 1 artiﬁcial microperiod

Index sets

St set of microperiods s belonging to macroperiod t

D− set of all 4-tuples consisting of product-line combinations (k, i) and (l, j)

where product i on line k is produced at most as fast as j on l

D+ set of all 4-tuples consisting of product-line combinations (k, i) and (l, j)

where product i on line k is produced faster than j on l

Jl set of products which can be produced on line l

Φ set of all microperiods with ﬁxed starting times

Λ set of all last microperiods of macroperiods

Πl set of all microperiods that do not allow a production on line l

Data

mlj minimum lot-size of product j (qty) if produced on line l

ml0 minimum time line l has to remain shut down

hjs holding costs of product j (per unit and per macroperiod t with s =

ft+1 − 1)

cplj production costs of product j (per unit) on line l

3.3. Multi-level models in the literature 57

zljj1 = ylj0 equals 1 if line l is set up for product j at the beginning of the planning

horizon (0 otherwise)

djs demand for product j in microperiod s (qty)

Ij0 initial inventory of product j at the beginning of the planning horizon

(qty)

ws start time of ﬁxed period s ∈ Φ

pji number of (qty) units of product j required to produce one unit of the

direct or indirect successor i

Ijmax maximum stock level of product j (qty)

Wljmax maximum WIP-stock level of product j after production on line l (qty)

cej purchasing costs (per unit) of product j

emax

j maximum number of (qty) units that product j can be externally pur-

chased

co overtime costs

omax maximum overtime

Variables

ws ≥ 0 start time of microperiod s

Ijs ≥ 0 inventory of product j at the end of microperiod s (qty)

xbls ≥ 0 fractional setup time for changeover at the beginning of microperiod s

xels ≥ 0 fractional setup time for changeover at the end of microperiod s

xljs ≥ 0 total quantity of product j produced in microperiod s on line l (qty)

ljs ≥ 0

x share of xljs that can be used by successors in the same microperiod s

(qty)

xljs ≥ 0 share of xljs that can as WIP-stock ﬁrst be used by successors in the

following microperiod s + 1 (qty)

xls ≥ 0 standby time on line l in microperiod s

ejs ≥ 0 externally purchased quantity of product j in microperiod s (qty)

os ≥ 0 overtime used in microperiod s

yljs ∈ {0, 1} setup state: yljs = 1 if line l is set up for product j in microperiod s (0

otherwise)

zlijs ≥ 0 takes on 1 if a changeover from product i to product j takes place on line

l during microperiod s (0 otherwise)

58 Chapter 3. Multi-level lot-sizing and scheduling

Objective function

min hjs Ijs + slij zlijs + cplj xljs

s∈Λ,j=0 l,i,j,s l,j,s

+ cbl xls + cej ejs + co · o s + hjsxljs (3.76)

l,s j=0,s s l,s∈Λ,j=0

subject to

ws = w s ∀s ∈ Φ (3.77)

ljs + xljs = xljs

x ∀l, j = 0, s (3.78)

Ijs = Ij,s−1 + ljs

x

l

+ xlj,s−1 − pji xlis + ejs − djs ∀j = 0, s (3.79)

l l i ∈ Nj

xljs ≤ Wmax

lj ∀l, j = 0, s (3.81)

xel,s−1 + xbls = stlij zlijs ∀l, s ≥ 2 (3.82)

i,j

xbls + xls + alj xljs + xels = (ws+1 − ws ) + os ∀l, s (3.83)

j

xljs ≥ mlj zlijs ∀l, j, s (3.85)

i:i=j

yljs = 1 ∀l, s (3.86)

j

yljs = 0 ∀l, s, j ∈

/ Jl (3.87)

zlijs ≥ yljs + yli,s−1 − 1 ∀l, i, j, s ≥ 2 (3.88)

zlijs = 1 ∀l, s ≥ 2 (3.89)

i,j

xljs = 0 ∀l, s ∈

/ Πl (3.90)

j=0

ejs ≤ emax

j ∀j = 0, s (3.91)

os ≤ omax ∀s ∈ Λ (3.92)

os = 0 ∀s ∈

/Λ (3.93)

−

(ws+1 − ws ) + os ≥ xbks + aki x

kis + xels ∀s, (k, i, l, j) ∈ D (3.94)

(ws+1 − ws ) + os ≥ xbks + alj xljs + xels ∀s, (k, i, l, j) ∈ D+ (3.95)

The objective is to minimize the sum of holding costs of the lot-sizing stock, sequence-

3.3. Multi-level models in the literature 59

dependent setup costs, and production costs, as well as costs for standby, external purchase,

overtime, and holding of WIP-stock (3.76). With the help of (3.77), the starting times of all

microperiods in Φ (including the macroperiods) are ﬁxed. The quantity split is allowed by

(3.78) which divides the production quantity xljs into a part x ljs that is directly available in

the same period s and into xljs that is ﬁrst available in the next microperiod s + 1.

The inventory balancing constraints (3.79) ensure that primary as well as secondary

demand is met without back-logging. The inﬂow to the stock is composed of the total

production of j during s, the WIP-stock of j that has been built up in s − 1, and the

externally purchased quantities. Note that this WIP-stock has a lead-time of a single, usually

quite short microperiod and thus allows a more realistic modeling than commonly used LTB-

models. The outﬂow is the primary demand of j in s and the secondary demand generated

by direct successors i that are also produced in s. Constraints (3.80) restrict the overall

inventory of each product if necessary, whereas the line-speciﬁc WIP-stock can be limited

by (3.81).

Constraints (3.82) ensure that the required setup time stlij of a changeover from i to

j is actually spent while making the setup split possible. According to the time structure

outlined above, constraints (3.83) build up a microperiod s which consists of setup time

fractions, idle and production time. The length of this microperiod equals the time interval

ws+1 − ws between the two subsequent microperiods s and s + 1 which can be prolonged

if overtime os is used. Since the starting times of some selected microperiods are ﬁxed by

(3.77), all in all limited production capacity can be respected.

Because of the setup forcing constraints (3.84), production can only take place if the line

is set up accordingly. The ﬁxed end of the planning horizon wS+1 is proposed as the big-M

value. Constraints (3.85) enforce a minimum lot-size and are needed because setup costs

(or times) do not always satisfy the triangle inequality. Furthermore, (3.86) enforce that

a line can only be set up for exactly one product per microperiod. Since not all products

can be produced on every line, (3.87) forbid illegal combinations. (3.88) link the setup state

indicators y with the changeover indicators z. Together with the objective function (3.76),

they ensure that zlijs is only set to 1 if line l was set up for i in s − 1 and for j in s. To get

a tighter formulation, (3.89) are added (but not necessarily needed).

With respect to, e.g., working calendars or pre-determined maintenance activities, (3.90)

prevent production on certain lines in certain microperiods. Constraints (3.91), (3.92), and

(3.93) impose limits on external purchasing and overtime, respectively. Note that only the

capacities of the last microperiods (and thus the macroperiod) can be extended by overtime.

(3.94) and (3.95) to guarantee feasible plans. When considering a predecessor line k produc-

ing a (direct on indirect) predecessor product i and a successor line l producing a (direct on

indirect) successor product j, both equations ensure that microperiod s is long enough to

60 Chapter 3. Multi-level lot-sizing and scheduling

contain the period’s beginning setup fraction on the predecessor line k, the production on

the “slower” one of both lines, and the period’s ending setup fraction on the successor line

l. Constraints (3.94) are responsible for “slower” predecessor lines and products. Here, (ki)

represents the slower line-/product-combination which means that aki pij ≥ alj . Hence, pro-

ducing one unit of the successor product j on line l does not take more time than producing

the necessary quantity of the corresponding predecessor product k on line i. Of course, a

necessary prerequisite is that both products can be produced on the respective lines at all,

i.e. that aki > 0 and alj > 0. This is expressed by the index set D− which contains all tuples

(k, i, l, j) that fulﬁll these conditions. It is thus deﬁned by

where the index set Ni+ contains all direct and indirect successor products of i. In this

case, a potential setup time on the “faster” line l can become crucial. If both products are

produced in the same microperiod s—as shown in Figure 3.9 where all units of i are directly

consumed by j—we have to ensure that i is ﬁnished on time. Therefore, the length of this

microperiod must exceed the time for production on line k plus the setup time on line l. For

those successors the line synchronization constraints (3.94) have to be respected.

period period

s-1 s

line l j setup

line k i

time

Analogously, constraints (3.95) model the complementary case where the successor lines

and successor products are the slower ones. For this, the set D+ has to be deﬁned by

D+ = {(k, i, l, j) | aki > 0 ∧ alj > 0 ∧ aki pij < alj ∧ j ∈ Ni+ }

For ease of understandability, setup splitting has not been considered in Figure 3.9.

However, because of the setup splitting possibility, the setup fraction of the predecessor line

has to be included in the synchronization as well. As shown in Figures 3.10a and 3.10b, in

both cases—slower and faster production—the setup on line k in s must be respected, too.

ﬁlling lines, J0 diﬀerent soft-drinks are bottled. Each line l has its dedicated tank k. Ac-

3.3. Multi-level models in the literature 61

s-1 s s+1 s-1 s s+1

line l j setup j

line k i setup i

time

(a) “Slower” predecessor. (b) “Slower” successor.

these tanks, J1 diﬀerent liquid ﬂavors are prepared. In sum, J(= J0 + J1 ) products have

to be scheduled on L(= L0 + L1 ) machines (lines+tanks). Notice that inventory stocks can

only be hold for the soft-drinks. Thus, no decision variables are needed for the production

quantities of the ﬁrst stage, as these quantities can be directly obtained by the ones of the

second stage and the corresponding gozinto factors. The further notation of the P2SMM is

adapted to the one of the GLSPMS. The resulting model formulation looks as follows:

Indices

i, j products (soft-drinks 1, . . . , J0 ; liquid ﬂavors J0 + 1, . . . , J0 + J1 )

k, l machines (bottling lines 1, . . . , L0 ; tanks L0 + 1, . . . , L0 + L1 )

s = 1, . . . , S microperiods

t = 1, . . . , T macroperiods

Index sets

St set of microperiods s belonging to macroperiod t

Jl set of products which can be produced on machine l (soft-drinks on bot-

tling line l ≤ L0 ; liquid ﬂavors in tank l > L0 )

Data

djt demand for product j in macroperiod t (qty)

hj holding costs of ﬁnal item j (soft-drink) per unit and per macroperiod

bj back-logging costs of product j (soft-drink) per unit and per macroperiod

alj capacity consumption (time) needed to produce one unit of soft-drink j

on line l

pji quantity of liquid ﬂavor j necessary to produce one unit of soft-drink i

sij setup costs of a changeover from liquid ﬂavor i to liquid ﬂavor j and from

soft-drink i to soft-drink j, respectively

stij setup time of a changeover from liquid ﬂavor i to liquid ﬂavor j and from

soft-drink i to soft-drink j, respectively

62 Chapter 3. Multi-level lot-sizing and scheduling

liquid homogeneity)

+

Ij0 initial inventory of soft-drink j at the beginning of the planning horizon

(qty)

−

Ij0 initial back-log of soft-drink j at the beginning of the planning horizon

(qty)

ylj0 equals 1 if tank l is set up for liquid j at the beginning of the planning

horizon (0 otherwise)

Vl available capacity of tank l (qty)

Clt available capacity of line l in macroperiod t (time)

L0 number of ﬁlling lines

L1 number of tanks

J0 number of soft-drinks

J1 number of liquid ﬂavors

Variables

+

Ijt ≥0 inventory of product j at the end of macroperiod t (qty)

−

Ijt ≥0 back-order of product j at the end of macroperiod t (qty)

xljs ≥ 0 quantity of soft-drink j ≤ J0 produced in microperiod s (qty)

xbls ≥ 0 standby time in microperiod s on line/tank l (qty)

yljs ∈ {0, 1} setup state: yljs = 1 if product j is set up on machine l in microperiod s

(0 otherwise)

zlijs ≥ 0 takes on 1 if a changeover from product i to product j takes place on

machine l during microperiod s (0 otherwise)

Objective function

J0

+ −

min hj Ijt + bj Ijt + sij zlijs (3.96)

j=1 t l i ∈ Jl j ∈ J l s

subject to

stage 1 (tank)

pji xlis ≤ Vk ykjs ∀k > L0 , j ∈ Jk , l = k − L0 , s (3.97)

i ∈ Jl

pji xlis ≥ mkj ykjs ∀k > L0 , j ∈ Jk , l = k − L0 , s (3.98)

i ∈ Jl

ykj,s−1 ≥ ykjs k > L0 , t, s ∈ St (3.99)

j ∈ Jk j ∈ Jk

3.3. Multi-level models in the literature 63

zkijs ≥ yli,s−1 + ykjs − 1 k > L0 , j ∈ Jk , l = k − L0 , t ≥ 2, s = ft (3.101)

i ∈ Jl

zkijs ≤ 1 k > L0 , t, s ∈ St (3.102)

i,j∈Jk

stage 2 (bottling)

+ − + −

Ij,t−1 + Ijt + xljs = Ijt + Ij,t−1 + djt j ≤ J0 , t (3.103)

l ∈ Lj s ∈ S t

alj xljs + stij zlijs

j ∈ Jl s ∈ S t i ∈ Jl j ∈ J l s ∈ S t

+ xbls ≤ Clt l ≤ L0 , t (3.104)

s ∈ St

xbls ≥ stij zkijs − stij zlijs l ≤ L0 , k = l + L 0 , s (3.105)

i ∈ Jk j ∈ J k i ∈ Jl j ∈ J l

Clt

xljs ≤ yljs l ≤ L0 , j ∈ Jl , t, s ∈ St (3.106)

alj

yljs = 1 l ≤ L0 , s (3.107)

j ∈ Jl

zlijs ≤ 1 l ≤ L0 , s (3.109)

i ∈ Jl j ∈ J l

In addition to setup and holding costs, the objective function (3.96) also includes back-

logging costs. The constraints of the P2SMM are divided into two groups in accordance

with the production stage they belong to. (3.97) are the setup forcing constraints for stage 1

and represent the capacity constraints for the tanks at the same time. Accordingly, on

the left hand side the needed liquid ﬂavor has to be determined which is the result of the

bottled quantity on the directly connected ﬁlling line in the regarded period. To assure the

homogeneity of the liquid ﬂavor, a minimum lot-size is necessary (3.98). As the number of

microperiods is chosen a-priori, it might be the case that more microperiods are available

than needed. Thus, periods without production (idle periods) might occur. Constraints

(3.99) then help reduce the number of symmetric solutions by forcing idle periods to appear

at the end of a microperiod (cf. Fleischmann and Meyr, 1997). Therefore, a new setup is not

allowed in a macroperiod anymore, as soon as the setup state gets lost in that macroperiod.

Constraints (3.100) and (3.101) capture the setup changeovers. The second ones are

needed, because due to the possible setup loss for a tank, the setup state of the last micrope-

riod of a macroperiod has to be detected via the corresponding setup state of the connected

bottling line. Moreover, (3.102) limit the maximum changeovers per microperiod to 1.

64 Chapter 3. Multi-level lot-sizing and scheduling

In contrast to stage 1, stage 2 allows for building inventory stocks. As back-logs are also

allowed (negative stock levels), the inventory balance constraints look like (3.103). Con-

straints (3.104) assure that production, setup, and idle time do not exceed the available

capacity of each line in each macroperiod.

macroperiod t macroperiod t

s s+1

line l j j+1 j j+1

s s+1

time time

(a) Schedule without synchronization. (b) Schedule with synchronization.

Figure 3.11: Schedules with and without synchronization (cf. Ferreira et al., 2009).

The necessity of idle time in the bottling stage is illustrated by Figure 3.11. Even though

the tanks are “fast” enough to feed the lines, diﬀerent setup times in the stages can lead to

synchronization problems (see Figure 3.11a). Accordingly, idle time must be inserted if the

setup changeover for a tank takes longer than the changeover for the connected bottling line

(see Figure 3.11b). This is achieved by constraints (3.105). Note that idle time needs not to

be added for the ﬁrst stage, since the capacity constraints are volume- and not time-based

(and the tanks are “fast” enough). Constraints (3.106), (3.107), and (3.108) are quite the

same as for the GLSPMS. Finally, at most one changeover is allowed in each microperiod by

(3.109).

Besides the PLSP-, CLSD-, and the GLSP-based models, there are also other models which

have diﬀering structures or cannot be classiﬁed in that way. Sikora et al. (1996) solve a

lot-sizing and scheduling problem with even ﬁve production stages and a serial product

structure. Unfortunately, they do not present a mathematical formulation.

Another multi-level STB-model is suggested by Persson et al. (2004), who solve a real-

world problem with two production stages which occurs in the oil industry. They disregard

sequence-dependent setup times and deﬁne small-time-buckets which only allow a single

mode of operation on each machine (processing unit) in each microperiod. Furthermore,

each mode is dedicated to a single machine and directly linked to a ﬁxed production quantity.

This corresponds to an all-or-nothing assumption similar to the DLSP.

A further approach is presented by Mohammadi et al. (2009). They consider a multi-stage

ﬂow shop production where all machines are arranged serially. Mohammadi et al. propose

synchronization constraints that forbid the production of an item on a certain machine

3.4. Conclusions 65

unless it has been ﬁnished on the predecessor machine. For this purpose so-called “shadow

products” representing idle time are introduced.

3.4 Conclusions

In the literature, only a few multi-level models do exist. But hardly any of them are able to

cope with the requirements for several stages, parallel production lines, sequence-dependent

setup times and costs as well as zero lead-times. Main reason is that most of them are based

on the time structure of single-level models which were originally not intended to meet the

requirements, the multi-level variants “suﬀer” from these time structures.

Correspondingly, the major drawbacks of STB-models can also be found in the multi-

level PLSP models by Haase (1994), Kimms (1996) and Kimms (1997). On the one hand,

setup times (which are not part of the DLSP) are generally a critical issue. If setup times (or

minimal lot-sizes) exceed the deﬁned length of a period, complex formulations are necessary

as proposed by Kallrath (1999) and Suerie (2005a) who use additional time-indexed variables.

On the other hand, if the length of a period is chosen too long, a loss of modeling detail

and thus ineﬃcient capacity utilization might result. Only the models by Stadtler (2011) as

well as Stadtler and Sahling (2013) include setup times and allow zero lead-times. However,

setups are still sequence-independent.

On the opposite side, the CLSD-based models as proposed by Sahling (2010) and Grünert

(1998) are not able to determine the starting times of the production lots. Accordingly, they

make use of at least one large-time bucket lead-time for the synchronization of the production

stages and cannot cope with non-triangular setups. Furthermore, Grünert considers no

parallel lines and Sahling assumes small lot-sizes.

In contrast to the single-level time structure of PLSP and CLSD, the GLSP uses a two-

level time structure which combines the advantages of large-time and small-time buckets.

That is why models like the ones by Araujo et al. (2008) and Ferreira et al. (2009) are able to

meet most of the requirements. However, these models are specially tailored to the regarded

production system with two production stages. They use explicitly deﬁned constraints for

the synchronization between these two stages, and thus they cannot be easily extended for

further stages.

In sum, the GLSPMS seems best suited to model diﬀerent problem settings of the con-

sumer packaged goods industry, and thus it will be examined thoroughly in the next chapter.

Chapter 4

In the previous chapter, it became clear that the General Lot-sizing and Scheduling Problem

for Multiple Production Stages (GLSPMS) ﬁts best the requirements for solving the problem

of shifting bottlenecks in the consumer goods industry. Unfortunately, the GLSPMS shows

a few shortcomings which are pinpointed next in the ﬁrst section. Afterwards an improved

formulation resolving these shortcomings will be given in Section 4.2. Finally, the properties

of the improved GLSPMS will be discussed and ﬁrst computational results will be presented

in Sections 4.3 and 4.4, respectively.

The basic model as proposed by Meyr contains some shortcomings concerning the line syn-

chronization. First of all, constraints (3.94) and (3.95) do not consider, for which items the

corresponding lines are actually set up. This may lead to unnecessarily large microperiods.

Figure 4.1 illustrates such a case.

line l j xlse j x lse j x lse

b

line k b

xks i‘ xks i‘ x bks i‘

time

(a) Feasible solution but precluded (b) Solution potentially forced (c) Solution potentially forced

by constraints (3.94) and (3.95) by constraints (3.94) by constraints (3.95)

Figure 4.1: Unnecessarily large microperiods due to constraints (3.94) and (3.95).

In this example, item j can be produced on line l. Moreover, there are two further

products i and i which both are producible on line k. We assume now that item i is a

predecessor of j but not i . Accordingly, a production plan as shown in Figure 4.1a could be

DOI 10.1007/978-3-658-02089-7_4, © Springer Fachmedien Wiesbaden 2013

68 Chapter 4. Improvements of the GLSPMS

j are independent from each other. However, constraints (3.94) and (3.95) do not allow such

a solution. If product i is a slower predecessor, then the solution given in Figure 4.1b would

be the result from constraints (3.94), although i is not produced. If j is slower, even a larger

microperiod could be potentially enforced by constraints (3.95) as shown in Figure 4.1c.

Note that this illustrated issue results in plans with unnecessary idle time. However, these

plans are feasible. But there are two further synchronization issues which indeed can lead

to infeasible production plans.

The ﬁrst one arises if more than two production stages have to be considered. The

GLSPMS makes no diﬀerence whether idle time is needed before or after the production. Let

us consider a small example. We suppose that product j on line l is faster than its predecessor

i on k and its successor j on l . For the ease of understanding, we further assume that j on

l and i on k have the same production speed. In this case1 , constraints (3.94) enforce the

length of microperiod s to be at least ws+1 − ws ≥ xbks + aki xkis + xels for the tuple (k, i, l, j).

Accordingly, idle time occurs before production on line l as shown in Figure 4.2a. Since

product j is faster than j , constraints (3.95) allow that j might start at the same time as j

(see Figures 4.2b). Figure 4.2c illustrates these constraints (3.95) also allow the production of

j to start at the same time as of i. This means, the resulting microperiod length for the tuple

(k, i, l, j) is suﬃcient for (l, j, l , j ) and (k, i, l , j ) because ws+1 − ws ≥ xbls + al j xl j s + xel s

and ws+1 − ws ≥ xbks + aki xkis + xel s , respectively. However, the plan as given in Figure 4.2d

is not feasible in reality because the synchronization constraints do not consider idle time

and because the production of item j starts too early.

The second synchronization problem might occur for parallel production lines. Keeping

track of the relations between predecessor and successor products in the BOM is generally

necessary for the synchronization, but it is not suﬃcient. The presence of parallel lines makes

it necessary to consider the actual material ﬂow between the stages. For instance, we assume

that the successor item j is slower on line l than its predecessor i on line k (alj > pij aki ). In

addition, item j is also slower on line l than the predecessor (e.g., al j = alj ). But if item

j is produced on both lines in parallel, the “total” production rate might2 become faster

a

than the one of item i on line k ( 2lj < pij aki for al j = alj ) and a new infeasibility arises as

illustrated Figure 4.3.

In order to avoid all these shortcomings and to guarantee feasible plans, the line synchro-

nization constraints have to be revised. A possible improvement is developed and presented

in the following.

1

Note that such an example can also be created including setup times and dropping the last assumption

of equal production speeds.

2

Note that the inverse of the production coeﬃcient equals the production rate.

4.1. Shortcomings of the basic GLSPMS 69

period s period s

line l‘ j‘

line l j line l j

line k i

time time

(a) Period length suﬃcient for the synchro- (b) Period length suﬃcient for the synchro-

nization of lines k and l. nization of lines l and l .

period s period s

line l‘ j‘ line l‘ j‘

line l line l j

line k i line k i

time time

(c) Period length suﬃcient for the synchro- (d) Period length insuﬃcient for a completely

nization of lines k and l . feasible plan.

period s

line l‘ j setup

line l j setup

line k i

time

70 Chapter 4. Improvements of the GLSPMS

This section presents an improved variant of the GLSPMS. The major change concerns the

new line synchronization which is explained next. Additionally, some minor changes are

made as well. For this reason, the complete new formulation is given afterwards.

To cope with the shortcomings outlined above, the microperiod structure is extended. As

shown in Figure 4.4, each microperiod may now consist of six diﬀerent time spans. The

original ﬁve elements are complemented by a second idle time which may arise between

production and the setup time fraction at the end of the microperiod. Moreover, the ﬁrst

idle time is “moved” before the production.

Figure 4.4: New microperiod structure with two diﬀerent elements of idle time.

Correspondingly, the new structure explicitly diﬀerentiates between idle time which is

needed before and which is needed after production. This allows to enforce that the produc-

tion of a successor product j on line l must not start and must not end before the production

of a direct predecessor product i on line k. Figure 4.5 illustrates this condition which results

in two new types of constraints (see (4.21) and (4.22)).

line l, product j b

xls ഥls

x b

ෝljs

alj x alj xljs ഥls

x e e

xls

∙

¸

line k, product i b

xks ഥks

x b

ෝkis

aki x aki xkis ഥks

x e e

xks

Figure 4.5: Production of direct successor product j on line l must neither start nor end

before predecessor product i on line k in every microperiod s.

The ﬁrst one forces setup and idle time before the production on successor line l to be at

least as long as setup and idle time before the production on line k. The second one assures

that setup and idle time after production on line l are at most as long as the time needed

on line k for producing the parts which are ﬁrst available in the next microperiod, plus the

standby and setup times at the end of the current microperiod.

Even though this condition is only formulated for a pair of direct predecessor/successor

products, it is “transitive”: If the production of j must neither start nor end before j and

j must neither start nor end before i, then the condition is implicitly fulﬁlled by j and i.

Thus, an infeasible solution as shown in Figure 4.2d is not possible anymore. Moreover, the

4.2. Improved GLSPMS 71

new condition also prevents the synchronization issue illustrated in Figure 4.3, as on all lines

the production of a successor product must not end before the one of the pre-item.

The indices as well as most of the index sets stay the same for the improved model. However,

a new index set D is introduced which replaces the sets D+ and D− . Due to the new

microperiod structure and the new synchronization approach, the diﬀerentiation between

slower and faster product/line-combinations is not needed anymore and thus D contains all

tuples (k, i, l, j) where j is a direct successor of i and j is producible on l as well as i on

k. Furthermore, new variables have to be included for the second idle time. The resulting

improved formulation which corresponds to the one given by Seeanner and Meyr (2013) is

completely listed below, where the numbers of unchanged constraints are marked with an

asterisk.

Index set

D set of all (k, i, l, j)-tuples consisting of line-product combinations (k, i) and

(l, j) where product j is a direct successor of i (j ∈ Nj ) and j is producible

on line l (alj > 0) and i on k (aki > 0)

Variables

xbls ≥ 0 standby time on line l in microperiod s before production

xels ≥ 0 standby time on line l in microperiod s after production

Objective function

min hjs Ijs + slij zlijs + cplj xljs

s∈Λ,j=0 l,i,j,s l,j,s

+ cbl (xbls + xels ) + cej ejs + co · o s + hjsxljs (4.1)

l,s j=0,s s l,s∈Λ,j=0

subject to

ws = w s ∀s ∈ Φ (4.2*)

ljs + xljs = xljs

x ∀l, j = 0, s (4.3*)

Ijs = Ij,s−1 + ljs +

x xlj,s−1 − pji xlis + ejs − djs ∀j = 0, s (4.4*)

l l l i∈Nj

xljs ≤ Wmax

lj ∀l, j, s (4.6*)

72 Chapter 4. Improvements of the GLSPMS

xbl1 = stlij zlij1 ∀l (4.8)

i,j

xel,s−1 + xbls = stlij zlijs ∀l, s ≥ 2 (4.9*)

i,j

xbls + xbls + alj xljs + xels + xels = (ws+1 − ws ) + os ∀l, s (4.10)

j

xljs ≥ mlj (yljs − ylj,s−1 ) ∀l, j, s (4.12*)

yljs = 1 ∀l, s (4.13*)

j

yljs = 0 ∀l, s, j ∈

/ Jl (4.14*)

zlijs ≥ yljs + yli,s−1 − 1 ∀l, i, j, s (4.15*)

zlijs = 1 ∀l, s (4.16*)

i,j

xljs = 0 ∀l, s ∈

/ Πl (4.17*)

j=0

ejs ≤ emax

j ∀j = 0, s (4.18*)

os ≤ omax ∀s ∈ Λ (4.19*)

os = 0 ∀s ∈

/ Λ (4.20*)

xbls + xbls ≥ xbks + xbks − wS+1 (2 − yljs − ykis ) ∀s, (k, i, l, j) ∈ D (4.21)

kis + xeks + xeks ≥ xels + xels − wS+1 (2 − ykis − yljs ) ∀s, (k, i, l, j) ∈ D

aki x (4.22)

In accordance with the new microperiod structure, the objective function is extended by the

new variables in the standby costs term. Moreover, constraints (4.7) prevent zero “end-of-

horizon” inventories which may have a negative impact on later periods beyond the planning

horizon (cf. Stadtler, 2000). In contrast to Meyr’s formulation, (4.8) allow setup time at the

beginning of the ﬁrst microperiod.3 The six elements of the new structure are now included

in (4.10). As explained in Section 4.2.1, the new constraints (4.21) and (4.22) assure the

line synchronization. However, the resulting conditions only make sense, if pre-product i is

actually produced on k and successor j on l. Therefore, both types of constraints are only

“active” if products i and j are set up on lines k and l, respectively.

As next, the properties of the improved GLSPMS are discussed, before the ﬁrst com-

putational results will be presented. Note that for the remainder of this thesis, the term

GLSPMS exclusively refers to the improved model formulation given in this section.

3

Correspondingly, the condition zljj1 = ylj0 needs to be removed (see Data in Section 3.3.4.1).

4.3. Properties of the improved GLSPMS 73

At ﬁrst, the computational complexity of the GLSPMS is examined. In a second step, the

advantages and disadvantages will be elaborated. Finally, the powerfulness of the GLSPMS

will be demonstrated.

4.3.1 Complexity

problem is in the class NP and each decision problem in NP is polynomial time reducible

to DP (Garey and Johnson, 2003, p. 37). The question whether the GLSPMS has a feasible

solution, also represents a decision problem which is called GLSPMSDP in the following. It

can be easily seen that this problem belongs to the class NP, since any production plan can

be checked for feasibility in polynomial time. A corresponding algorithm for the feasibility

check may easily be constructed.

In order to show that each decision problem in NP is reducible to GLSPMSDP , it is

suﬃcient to reduce a known NP-complete problem to the GLSPMSDP . One of these NP-

complete problems is the decision version of the Traveling Salesman Problem (TSP) which

can be formulated as follows (Garey and Johnson, 2003, p. 211):

Given a set of N cities and the distance tdij between all pairs of cities i, j.

Does there exist a tour (Hamiltonian Cycle) with a length of at most C?

This decision problem TSPDP is in polynomial time reducible to the GLSPMSDP (noted

as TSP ≺p GLSPMSDP ) by interpreting the cities as products (J = N ) and the distances

as setup times (stlij = tdij ) of a single production line (L = 1). Therefore, the index sets

Nj = ∅(∀j) and D = ∅ are empty because of the single production stage. Furthermore,

on this line, production is allowed in all microperiods (Π1 = {1, . . . , S}) for all items J1 =

{1, . . . , J}.

By setting a positive demand for every product, we enforce the visit of each city (produc-

tion of each item). Note that for a round trip any product might be the starting point. Hence,

y110 can be set to 1 without loss of generality. To assure that product 1 is set up again at

the end of the planning horizon, we use two macroperiods (T = 2) and enforce the exclusive

production of item 1 in macroperiod 2. This can be easily achieved by setting the number

of microperiods in macroperiod 1 to N − 1 and in macroperiod 2 to 1 (Φ = {1, N, N + 1},

Λ = {N − 1, N }).5 Moreover, we have demand for all products except for product 1 in

macroperiod t = 1, i.e. dj,N −1 = 1 for all j = 1 and d1,N −1 = 0. In macroperiod t = 2 this

4

For a deeper introduction refer to Garey and Johnson (2003) among others.

5

Alternatively, the maximum inventory level of product 1 can be set to 0 to prevent a production in the

ﬁrst macroperiod.

74 Chapter 4. Improvements of the GLSPMS

assumption is reversed as djN = 0 for all j = 1 and d1N = 1. In sum, there are N − 1

items that have to be produced in macroperiod t = 1. Since this macroperiod consists of

N − 1 microperiods and due to the fact that at most one item can be produced in a single

microperiod, the requirement of a closed tour is fulﬁlled.

0),6 the production of one piece per item is suﬃcient and the maximum inventory Ijmax can be

set to 1. Since the setup forcing constraints only work properly if the production coeﬃcients

a1j > 0, capacity is consumed by the production. As we do not have any minimum lot-

sizes and by setting a1j = 1, an additional capacity of N time units is needed. Due to the

possibility of setup splitting we can suppose that in macroperiod t = 2 only the production of

item j = 1 needs to take place. For this reason, the capacity of the second macroperiod can

be set to 1 (wN +1 − wN = 1) and the remaining capacity is assigned to the ﬁrst macroperiod.

This results in w1 = 0, wN = C + N − 1, and wN +1 = C + N . The remaining data (omax j ,

g max , W1j

max

) is set to 0.

If the instance of GLSPMSDP at hand can be answered with yes which means that a

production plan exists, then the TSPDP has the same answer.7 In the following, a small

example should clarify the transformation steps.

Example

With the help of the following example, the transformation procedure is illustrated. We

assume that a circular tour through ﬁve cities has to be found. The travel distances are

given in Figure 4.6. The question is whether a tour with a length up to 90 units is possible.

First, the distances are interpreted as setup times (see Table 4.1).

1 21

38

4 st1ij j=1 j=2 j=3 j=4 j=5

19

15 i=1 0 25 38 21 19

25 17

i=2 25 0 33 36 14

36 5

13

i=3 38 33 0 17 13

14 3 i=4 21 36 17 0 15

33

2 i=5 19 14 13 15 0

Figure 4.6: Travel distances between all Table 4.1: Travel distances as setup times.

pairs of ﬁve cities.

6

Since only one production lot per item should be produced, the minimum lot-size can also be set to the

corresponding demand.

7

Finally, it is worth mentioning that the same reduction as by Fleischmann and Meyr (1997) could have

been used as well. They transform the 3 Partitions Problem (3PP) to the GLSP with non-zero minimum

lot-sizes. Even though the GLSPMS also includes minimum lot-sizes, constraints (4.7), however, do not

allow to reduce the GLSP to the GLSPMS since not all GLSP instances with non-zero minimum lot-sizes

can be transformed. Accordingly, the GLSP is only partly a subproblem. Nevertheless, the 3PP can be

reduced to the GLSPMS analogously.

4.3. Properties of the improved GLSPMS 75

In accordance with the transformation explained above, ﬁve microperiods are used and

the ending times w5 = 94 and w6 = 95 are ﬁxed. Then the GLSPMSDP must be answered

with yes and so is the TSPDP (see the production plan given in Figure 4.7 with the round

trip 1,4,3,5,2,1).

w1 w2 w3 w4 w5 w6

problem can be called NP-hard (Florian et al., 1980). But the proof that the TSPDP is a

special case of the GLSPMSDP allows further conclusions (cf. Chapter 8). Even though the

TSP—as the optimization problem of ﬁnding the shortest tour—is a very hard problem, the

research community has made a lot of progress in the past. Applegate et al. (2006, p. 317)

show in their historical overview over the TSP that today it is possible to solve real-world

problem instances with 1000 cities to optimality within a couple of minutes.

Hence, despite the computational complexity, real-world instances of the GLSPMS might

be solvable by standard MIP-solvers in a reasonable time. Furthermore, standard MIP-

solvers oﬀer functionality such as presolving and cut generation which may accelerate the

solution procedure depending on the data structure.

Therefore, it might be possible that the problem data for the GLSPMS allows exact

methods to solve instances with suﬃcient size. In the following it has to be tested how

the GLSPMS behaves from a computational point of view. For this purpose several test

scenarios were created which will be described at the beginning of Section 4.4.

Due to the linkage of the start times ws for all lines and stages, each state change on a certain

line implies a new microperiod for all production lines. Correspondingly, the structure carries

the risk that a huge number of microperiods is needed and thus the model gets harder to be

solved. In combination with the setup splitting, many symmetric solutions can result.

Exemplarily, we suppose that the plan shown in Figure 4.8a is the optimal solution, with

two diﬀerent production lots on each of the three production lines. The GLSPMS needs four

microperiods to ﬁnd this plan (see Figure 4.8b). The resulting redundancy can directly be

seen by Figure 4.8c.

Another potential drawback stems from the modeling of the setup times. On the one

hand, the GLSPMS allows setup time splitting and thus setup time lengths up to two

76 Chapter 4. Improvements of the GLSPMS

line l‘

line l

line k

time

(a) Two diﬀerent production lots (b) Four microperiods needed (c) Redundant solution.

on each line. for two diﬀerent lots on each

line.

macroperiods. But on the other hand, the microperiods must become correspondingly large.

This could lead to high idle times because all other production lines are automatically aﬀected

by the time structure. In Chapter 7 it is shown how to cope with that situation.

Models which are based on a “GLSP structure” as given in (4.23)–(4.25) are known to

provide bad dual bounds (also called lower bounds for minimization problems). As one can

imagine the chosen “big” M plays an important role for the tightness of the formulation be-

cause the lower this M-value, the higher the value for the setup state variable must be in the

LP relaxation (as a result of the multiplication within the setup forcing constraints). More-

over, it follows that the higher the values of the setup state variables, the higher the values of

the changeover variables. Accordingly, setup costs (and times) are captured more properly.

Therefore, M should be as small as possible and as high as necessary (cf. Section 3.3.2.2).

yjs = 1 ∀s (4.24)

j

But even then, the LP relaxation of the GLSP shows the eﬀect of “smearing” where many

setup state variables have small values which allow production but not incur costs. This

can be easily illustrated by a small example. Consider a problem instance with one line,

one macroperiod, four items j = 1, ..., 4, each with a production coeﬃcient aj = 1 as well

as demand dj > 0 while neglecting setup times and capacity limitations. The corresponding

GLSP instance must have at least four microperiods to allow the production of all items.

We assume that for each item the big M is set to the demand value as the tightest possible

value. Then a solution as shown in Table 4.2 might be the result of the LP relaxation. For

each product j (table row), the yjs sum up to 1. So each demand is producible and can

be fulﬁlled. It can be seen that small fractions got smeared over all products. Let us ﬁrst

concentrate on the changeover variables belonging to microperiod s = 1. By “taking” one

third from product j = 4 in period s = 1, product j = 3 can be produced and the condition

4.3. Properties of the improved GLSPMS 77

1 1 1

j=1 0 0 3 3 3

1 1 1

j=2 0 0 3 3 3

1 1 1

j=3 0 3 3 3 0

1 1

j=4 1 1− 3 0 0 3

Table 4.2: Relaxed changeover variables smeared over the planning horizon.

z431 ≥ 1 + 13 − 1 is fulﬁlled for z431 = 13 . This means only one third of the corresponding setup

costs is actually incurred. Variable z441 has a value of 23 which ends up in no costs (setup

state conservation). For the next periods the changeover variables z312 , z322 , and z332 can

even be minimized to zero as they only need to be greater than or equal to 13 + 13 − 1 = − 13 .

This eﬀect gets even ampliﬁed if the same product is set up in several consecutive mi-

croperiods. Nevertheless, it is important to see that the primal solution can be very good,

unfortunately without knowing it due to a possibly large gap to the dual bound.

In spite of all these potential drawbacks, the GLSPMS shows several important advan-

tages which signiﬁcantly outweigh the disadvantages. First of all, it is the only multi-level

model which includes parallel heterogeneous lines, sequence-dependent setups as well as zero

lead-times. Moreover, by penalizing idle time in an appropriate manner (cbl ), it is up to the

GLSPMS to decide for a period s with idle time (xbls + xels > 0) and without production

( j xljs = 0) on a certain line l, whether the setup state should be conserved (zljjs = 1)

or lost (zlj0s = 1). In contrast, most models for simultaneous lot-sizing and scheduling only

allow for modeling either a conservation or a loss of the setup state after idle periods. Fur-

thermore, the synchronization capabilities allow for incorporating further scheduling aspects.

In sum, the formulation can be applied for many diﬀerent problem settings from the

consumer goods industry. This characteristic is emphasized a little more in the next section.

The capability to solve diﬀerent problem settings from the consumer goods industry—even

with shifting bottlenecks—is supposed to be the most outstanding feature of the GLSPMS.

Actually, the real-world problems underlying the models shown in Section 3.3 can be tackled

in large part with the GLSPMS.

To demonstrate this, it will be exemplarily shown in the following how the GLSPMS

is able to solve the problem of the Brazilian soft-drink industry as described by Ferreira

et al. (2009). Therefore, we will take a closer look to the P2SMM again and show that each

78 Chapter 4. Improvements of the GLSPMS

solution of this model which does not include back-logs can also be found by the GLSPMS.8

But for that purpose several data transformation steps have to be executed:

First of all, production and standby costs as well as costs for external purchasing and

overtime can be set to zero as they are not included in the P2SMM. Since the capacities at the

ﬁrst production stage (3.97) are not time but volume-based (tank capacity), the GLSPMS can

model them (in the same way as the P2SMM) by the setup forcing constraints. Furthermore,

we have to assure that the time structure of the GLSPMS does not unnecessarily limit

this stage by the constraints (4.10). So the production coeﬃcients alj must be adapted

accordingly. As can be seen by (3.97), stocks are not allowed/possible at the ﬁrst stage.

Correspondingly, Ferreira et al. drop the variables for the production quantity in this stage.

For the GLSPMS this can be achieved by setting the maximum inventory Ijmax = 0 for all

liquid ﬂavors. Then, it automatically follows from the inventory balance constraints that

0 = 0 + l xljs − l i∈Nj pji xlis and l xljs = l i∈Nj pji xlis . However, as becomes

directly apparent from this last equation, the fact that a tank is dedicated to a certain ﬁlling

line is a critical issue for the GLSPMS because the production quantities are aggregated over

all tanks and bottling lines.9 An obvious workaround is to introduce “ﬁctitious” products for

each combination of tank and liquid ﬂavor as well as bottling line and soft-drink, respectively.

Then, a product is only producible on a single machine and the aggregation is annulled. But

in that case, a third, also ﬁctitious stage which comprises the “original“ soft-drinks has to be

introduced additionally. On this way, the corresponding primary demand can be fulﬁlled via

the inventory balance constraints as well. The remaining parts of the P2SMM are uncritical.

The synchronization of the material ﬂow as forced by constraints (3.105) is immediately

realized by the GLSPMS with its synchronization constraints. Besides, Ferreira et al. use

constraints (3.99), (3.100), and (3.101) to reduce the number of symmetric solutions by

forcing idle time to appear at the end of a macroperiod. Hence, they are only meant for

improving the solvability of the model and can be neglected for now. Finally, the constraints

for the minimum lot-size (3.98), the setup changeovers (3.100) and (3.108), the setup forcing

(3.106) and the unique setup states (3.107) are the same for the GLSPMS.

A very similar comparison can be made for the PLSP-ML-SM by Stadtler. The underlying

problem might be solved by the GLSPMS as well. In the following, the ﬁrst computational

results for the improved GLSPMS will be shown. For the tests, three basic problem scenarios

are used which illustrate at the same time the diﬀerent areas of application for the GLSPMS

and which will be described next.

8

Note that back-logging is quite often only used for identifying the bottleneck and is not allowed in

reality. Then relaxations like overtime or purchasing as used by the GLSPMS can also serve this purpose.

9

This assumption by the authors is only meant for simpliﬁcation. That is why the GLSPMS is even able

to ﬁnd even more realistic solutions to the original problem described by Toledo et al. (2007).

4.4. Computational tests 79

The GLSPMS is implemented and tested with the FICO Xpress Optimization Suite. Ac-

cordingly, Xpress Optimizer Version 20.00.22 is used as a standard MIP-solver. The test

instances were run on a single core (2 threads) of a Dell Precision T1500 with an Intel

Core i7-860 (2.8 GHz) processor, 8 GB RAM, and the Windows Server 2008 64bit operating

system.

Three diﬀerent base scenarios with a divergent, general, and serial product structure were

used (cf. Meyr (2004) and Seeanner and Meyr (2013)). They are motivated by practical

problems and represent realistic situations. Even though the dimensions may diﬀer from

reality, the data allows for comparing diﬀerent formulations and solution approaches. In the

following the structure of these scenarios is illustrated. The complete data can be found in

Appendix A.

The ﬁrst scenario is a three-level serial production of juices. There are two ﬁnal items, a

six-pack of apple juice (j = 1) and a six-pack of cherry juice (j = 2). The apple juice is ﬁlled

in standard 1L bottles (j = 3), whereas other, special 1L bottles—diﬀering in shape and

texture for marketing purposes—are used for cherry (j = 4). Accordingly, in a ﬁrst step two

diﬀerent types of PET-bottles (j = 5, 6) are produced on stretch blow molding machines.

Since ﬁlling of bottles is faster than molding, two identical molding lines l = 3, 4 are needed.

A schematic representation of the corresponding BOM and the assignment of products to

lines is given in Figure 4.9. In the second production stage (and level e = 1 of the BOM

l=1 j=1,2

level e=0 j=1 2

p31=6 6

l=2 3,4

level e=1 3 4

1 1

l=3 5,6

level e=2 5 6 l=4 5,6

structure, respectively), the bottles are ﬁlled with the help of a single tank l = 2. Since

this tank has a minimum ﬁll level, minimum lot-sizes are required. Setup times are highly

sequence-dependent due to cleansing processes. As mentioned, ﬁlling is quite fast, however,

80 Chapter 4. Improvements of the GLSPMS

WIP-stock of ﬁlled bottles is limited because of the restricted space between the ﬁlling and

the wrapping machine.

At the ﬁnal stage (l = 1), shrink-packaging takes place where six bottles of each sort

are wrapped. Due to the diﬀerent types of bottles the corresponding calibration is quite

time-consuming. Each macroperiod has a capacity of 80 and the demand varies between 2

and 8 units for the ﬁnal items. The production costs are the same (=1) for all lines, except

for line 4, which causes higher costs for maintenance reasons. The lines are already setup for

products 1, 3, 5, and 6, respectively. The remaining data can be found in Tables A.1–A.14

of the Appendix.

In this divergent scenario, sheet glass is produced in diﬀerent colors and lengths on two

production stages. The scenario is illustrated in Figure 4.10. At the ﬁrst stage, there is

a single line l = 3 producing light-colored (j = 5) and dark-colored (j = 6) glass. The

production of light-colored glass is more time-consuming (a35 = 4) than the production of

the dark one (a36 = 2). The setup changeover from light to dark is less complex (st356 = 2)

than vice versa (st365 = 6). Both pre-products can be cut in two lengths—short and long—at

the second stage. Thus, there are four ﬁnal items j = 1, ..., 4 possible. Long items need twice

as much of glass (p51 = p63 = 2) than short cut (p52 = p64 = 1).

l=1 j=1,2,3,4

level e=0 j=1 2 3 4 2,4

l=2

p51=2 1 2 1

l=3 5,6

level e=1 5 6

For cutting, two parallel lines are available. The newer one l = 1 needs a11 = a13 = 3 for

a long and a12 = a14 = 4 for a short item. Changeover from short to long takes 2 time units,

whereas long to short only 1. Besides, a cleansing of 3 time units is needed for a change from

dark to light and of 1 time unit for the reverse. The older line l = 2 is only able to produce

short items in 8 units of time. On that machine a setup change from dark to light requires

4 and 2 for the reverse, respectively. Stocks are not allowed for the pre-products.

The minimum lot-size is one unit. At the beginning all production lines l = 1, 2, 3 are

already set up for j = 1, 2, 6. The older one causes production costs of 2 per unit, the other

lines expense 1 per unit. Setup costs equal the corresponding setup times. The planning

horizon consists of three macroperiods t = 1, . . . , 3. Accordingly, there are four microperiods

with ﬁxed starting times of 0, 80, 160, and 240, the last one being the artiﬁcial one. There

4.4. Computational tests 81

is only primary demand for ﬁnal items varying between 2 and 6 units. All other values are

listed in Tables A.15–A.28 of the Appendix.

A third scenario with a general product structures describes the production of yogurt. Two

kinds of yogurt (j = 6 and j = 7) are ﬁlled in big and small cups (j = 5 and j = 8) resulting

in four diﬀerent ﬁnal items j = 1, . . . , 4. In this make-and-pack environment, the cups as

well as the yogurts are produced in the ﬁrst stage, each of them on a dedicated production

line which is shown in Figure 4.11. The cups on line l = 3 need a35 = 8 and a38 = 2. A

changeover st358 from the big to the small packaging format needs 8 units of time and 6 for

the reverse. Besides, after a machine breakdown (j = 0) it takes st30j = 12 to ramp up. But

standby is also possible with b3 = 1. The yogurt can be produced on line l = 2 in parallel

with a2j = 6. Both setups are sequence-independent (st267 = st276 = 4).

l=1 j=1,2,3,4

level e=0 j=1 2 3 4

0.5 0.5

p71=1 1 1 1 1 1

l=2 6,7

level e=1 5 6 7 8 5,8

l=3

Figure 4.11: BOM and product-line assignment of the instance with a general product

structure.

In the second stage, the ﬁnal items are ﬁlled and packed on line l = 1 using p63 = p74 = 0.5

units of yogurt for small cups, else 1. Making the yogurt on line l = 2 and ﬁlling it on line

l = 1 are synchronized. Thus, building stocks for items 6 and 7 is not allowed. However,

producing the cups is decoupled by an unlimited buﬀer. The corresponding production

coeﬃcients are a11 = a12 = 6 and a13 = a14 = 3. A changeover from big to small cups needs

8 and from small to big 6 time units, whereas changing only the yogurt ﬂavor needs 4. All

further data is presented in the Appendix (Tables A.29–A.42).

These three base scenarios are varied by copying the bill-of-material structures as well as

the production lines and by increasing the number of (macro)periods of demand. As shown

in Table 4.3, (2 · 3 · 5 + 3 · 3 · 4 + 2 · 2 · 4 =) 82 problem instances result with 6 to 12 products,

3 to 9 lines, and 3 to 8 macroperiods.

L J T

SER 4,8 6,9,12 4,5,6,7,8

DIV 3,6,9 6,9,12 3,4,5,6

CONV 3,6 8,11 3,4,5,6

82 Chapter 4. Improvements of the GLSPMS

At ﬁrst, the aim is to ﬁnd out the maximum size of problem instances that can be solved to

optimality. Since the regarded planning problem is typically a short-term task, the compu-

tational runtime is limited to 3600 sec.

To determine the maximum solvable problem size, the number of microperiods per

macroperiod10 is systematically increased until a solution keeps unchanged or none of the

formulations can be solved to optimality any more.11 Besides, the impact of automatic cut

generation (CG) and presolving (PR) of the standard MIP-solver is analyzed as well.

Table 4.4 shows the results of diﬀerent solver settings grouped by the structure of the

products (serial (SER), divergent (DIV) and general (GEN)) and the problem instances’

sizes. Thereby, the size of a problem instance is deﬁned as the number of products multiplied

by the number of lines and the total number of microperiods (J × L × S). Each column

presents a diﬀerent combination of settings, where presolving is switched on (PR=1) and

oﬀ (PR=0) and cut generation is switched on (CG=1) and oﬀ (CG=0), respectively. The

ﬁrst entry represents the percentage of instances solved to optimality within 3600 sec. (PS).

Note that due to the external purchase-relaxation all instances are solvable. Since some

solver settings perform better than others, the second entry is the percentage of performing

best (PB). That means, for example, 39% of all instances with a serial product structure

and a size greater than 500 can be solved to optimality when presolving and cut generation

is enabled. This solver setting combination performs best in 44% of all instances. If cut

generation is switched oﬀ, the MIP-solver performs best for 31% of this group of problems.

As there are instances that are solved in the same amount of time (seconds with an accuracy

of one digit after the decimal point) but with diﬀerent settings, the percentages do not need

to sum up to 100.

Across all three product structures, the optimizer performs best when the default set-

tings for presolving and cut generation (1+1) are used. Accordingly, the highest number

of instances can be solved with them (64% on average). Only for small problem variants

(J × L × S ≤ 250) cut generation seems to reduce the performance. “Big” problems (>

500; for instance, J = 12, L = 3, T = 4 with six microperiods each) cannot be solved at all

except for the serial structure. Their share is relatively high with 32% on average. So there

seems to be some signiﬁcant diﬀerence depending on the product structure.

In summary, we have to state that the capabilities of standard MIP-solvers like Xpress are

not suﬃcient to solve the GLSPMS exactly—at least with the current formulation. However,

as shown in Figure 4.12 feasible solutions can be found quite fast, even for “big” instances,

and most of the computational time is spent for improving the lower bound. That is why

10

Note that we use the same number of microperiods for each macroperiod and thus the total number of

microperiods (S) is calculated as number of macroperiods (T) multiplied with “microperiods per macrope-

riod”.

11

Due to the external purchase-relaxation all instances are solvable.

4.4. Computational tests 83

J×L×S PS PB PS PB PS PB PS PB PS

SER 1- 250 100 7 100 50 100 0 100 43 100

251 - 500 93 50 83 27 87 17 70 7 83

> 500 39 44 31 31 33 22 25 3 32

DIV 1 - 250 100 24 100 62 100 3 97 12 99

251 - 500 56 40 52 19 42 27 29 15 45

> 500 0 33 0 33 5 19 0 24 1

GEN 1 - 250 100 12 100 71 100 12 100 6 100

251 - 500 54 23 62 31 46 27 31 19 48

> 500 0 43 0 0 0 29 0 29 0

average PS 64 61 58 50

Table 4.4: Percentage of test instances solved to optimality in 3600 sec. (PS) and percentage

of instances performing best (PB) with solver settings “presolving” (PR) and “cut generation”

(CG) set on (1) and oﬀ (0).

Figure 4.12: Progression of the upper and lower bound for the divergent instance with L=3,

J=12, T=5, and S=30.

84 Chapter 4. Improvements of the GLSPMS

it seems worth checking whether reformulation techniques can strengthen the above model

formulation of the GLSPMS and improve the solution performance. Therefore, in the next

chapter a couple of well-known reformulation techniques, which have been proven successful

to tighten other dynamic lot-sizing problems, are adapted to the GLSPMS.

Chapter 5

GLSPMS

Section 4.2 showed that a standard MIP-solver is not able to solve mid-sized instances of

the improved GLSPMS to optimality within 1h of runtime. But as mentioned in Chapter 3,

the way of formulating a model has impact on the performance of a solver. Accordingly,

this chapter which is based on the results presented by Seeanner and Meyr (2013) will

examine, whether it is possible to solve the problem instances exactly by “reformulating” the

GLSPMS. Hence, Section 5.1 will ﬁrst explain some general reformulation techniques, before,

in particular, selected approaches for lot-sizing and scheduling models will be introduced in

Section 5.2. Afterwards, these approaches will be adapted for the GLSPMS in Sections 5.3

and 5.4, respectively. To end up this chapter, Section 5.5 will give the computational results

of the developed reformulations.

The improved GLSPMS model is a formulation for the set of all feasible solutions to the

regarded multi-level lot-sizing and scheduling problem (cf. Pochet and Wolsey, 2006, p. 79).

But there are also other formulations, which describe the same set of solutions as well. This

fact can be illustrated by the following small example:

We suppose a pure-integer-program with two variables x and y. The feasible solutions

which build the feasible set X are highlighted as solid points in Figure 5.1. Obviously, the

convex sets P1 and P2 both are valid formulations as they contain all feasible solutions but

no infeasible ones. However, they are of diﬀerent quality because P1 is a better formulation

than P2, as P1 ⊂ P2. Since the performance of a standard MIP-solver strongly depends

on the quality of the formulation used, “reformulations” of the GLSPMS need to be checked

whether they allow reducing the computational burden (cf. Pochet and Wolsey, 2006, p. 93).

As also shown in Figure 5.1, the best formulation is the convex hull formulation conv(X).

DOI 10.1007/978-3-658-02089-7_5, © Springer Fachmedien Wiesbaden 2013

86 Chapter 5. Reformulations of the improved GLSPMS

It even allows to solve the integer problem as an LP, since all extreme points of conv(X)

belong to X. But ﬁnding the convex hull is a very hard problem as well (cf. Pochet and

Wolsey, 2006, pp. 93–94).

Nevertheless, it is generally a good idea to use a formulation as tight as possible. There are

two general ways to obtain a tighter formulation. On the one hand, extended reformulations

make use of redeﬁned and/or additional variables (plus adapted constraints) which give a

tighter description of the problem. A common way to get an extended formulation for an

optimization problem is to “reformulate” it as another (combinatorial) one. This approach is

also widespread for dynamic lot-sizing problems and will be illustrated in the next section.

On the other hand, a formulation can also be improved by using additional linear con-

straints (valid inequalities). As illustrated in Figure 5.2 a valid inequality/cut is a constraint

which eliminates part of the (initial) solution space and which is fulﬁlled by all solutions in

X (cf. Pochet and Wolsey, 2006, p. 95).

In the following, a few selected reformulation techniques will be explained which have

been successfully applied for lot-sizing and scheduling models in past.

5.2. Selected approaches for lot-sizing and scheduling models 87

els

To begin with extended reformulations, this section will explain the Simple Plant Loca-

tion, the echelon stocks and the network ﬂow formulation. Afterwards, the so-called (l,S)-

inequalities will be described.

The ﬁrst extended reformulation which will be presented in this section is the Simple Plant

Location (SPL) reformulation (cf. Krarup and Bilde, 1977; Rosling, 1986). It belongs to the

mostly used reformulations for lot-sizing (and scheduling) problems (cf. Jans and Degraeve,

2007).1 Afterwards, the basic concept of echelon stocks will be explained. Moreover, the

idea of reformulating a lot-sizing problem as a network ﬂow problem will be illustrated.

Given a set of potential plant locations as well as ﬁxed customer locations, the Simple Plant

Location problem2 is about how to satisfy customer demands while minimizing costs. These

costs consist of ﬁx costs for opening such facilities on the one hand, and transportation

costs for delivering a certain customer from a certain facility on the other hand (cf. Daskin,

1995, Chapter 7). The basic idea of the SPL reformulation is to interpret an item/period

combination as a plant location. Then, the opening of a facility is similar to setting up a

machine for production in a certain period. Furthermore, the demand of subsequent periods

can be interpreted as customer demand. Hence, the transportation costs correspond to

holding costs. Accordingly, we obtain a Simple Plant Location (SPL) problem as illustrated

in Figure 5.3. In this example, variable xjsτ describes the quantity of item j which is

produced in period s and meets the demand of period τ ≥ s. The transportation costs for

the relation between facility location s and customer τ equal the holding costs for keeping

the items τ − s periods in stock.

Obviously, the SPL approach needs no variables for the inventory stocks, as xjsτ already

carries the information, how many items of product j are kept in stock and how long.

1

Besides, the Shortest Route (cf. Eppen and Martin, 1987) and the Multi-Commodity Flow reformu-

lation (cf. Pochet and Wolsey, 1995) are often applied for lot-sizing problems as well. But as Jans and

Degraeve (2007) state these formulations produce the same lower bound for the single-level CLSP and can

be transformed to each other, respectively. Denizel et al. (2008) even prove that Simple Plant Location and

Shortest Route give the same LP bound for the CLSP with setup times. Since the same observation is made

by Stadtler (1996) for the multi-level CLSP, we only concentrate on the SPL formulation in this thesis.

2

This problem is also called Warehouse Location Problem or (Uncapacitated) Facility Location Problem.

88 Chapter 5. Reformulations of the improved GLSPMS

The concept of echelon stocks is very important for multi-level modeling as it allows de-

composing the single levels (cf. Clark and Scarf (1960) and Afentakis et al. (1984) among

others). The echelon stock Ejs is deﬁned as the total quantity of product j in the whole

system (as item or as component of a successor product) in period s (cf. Pochet and Wolsey,

2006, p. 396). The corresponding mathematical deﬁnition is given in (5.1), where again the

index set Nj+ contains all direct and indirect successor products of j. From this deﬁnition

follows that the inventory stock variables Ijs can be substituted according to (5.2).

Ejs :=Ijs + pji Iis = Ijs + pji Eis (5.1)

i∈Nj+ i∈Nj

Ijs =Ejs − pji Eis (5.2)

i∈Nj

This substitution has also impact on the objective function. However, as can be seen

in (5.3), the resulting expression can be simpliﬁed by adapting the cost values. There-

fore, the successor-oriented expression j i∈Nj pji Eis hjs only needs to be transformed to a

predecessor-oriented j i∈Vj pij Ejs his . This transformation becomes apparent by looking

at the arcs of a gozinto graph.

⎛ ⎞ ⎛ ⎞

Ijs hjs = ⎝Ejs − pji Eis ⎠ hjs = Ejs ⎝hjs − pij his ⎠ =: Ejs hm

js (5.3)

j,s j,s i∈Nj j,s i∈Vj j,s

These new values are marginal holding costs hmjs which describe the value added at the

regarded production stage by processing the needed components to one piece of product j.

5.2. Selected approaches for lot-sizing and scheduling models 89

It is intuitive to “value” the echelon stocks in this way, since it cannot be seen from Ejs

j=1 2

Ijs hjs Ejs hm

js

p51=2 1

j=1 12 3 12 1

5 j=2 2 3 2 2

j=5 5 1 31 1

Figure 5.4: Product structure of

the divergent base scenario (see Sec- Table 5.1: Holding stocks and costs as well as

tion 4.4.1.2). echelon stocks and marginal holding costs.

whether j has already been processed on further production stages. To illustrate now the

new calculation of holding costs, we assume the product structure given in Figure 5.4 as well

as inventories stocks and holding costs given in Tab. 5.1.

Since product 5 is part of product 1 and of product 2, the echelon stock of 5 in the regarded

period s is 5 + 12 · 2 + 2 · 1 = 31. The echelon stocks of the ﬁnal products equal the inventory

stocks. However, the holding cost values for these products diﬀer. Since, e.g., the inventory

holding costs for product 1 are partly covered by the holding costs of the pre-product 5, the

additional value of product 1—and thus the marginal holding costs—is 3 − 2 · 1 = 1. As can

be seen now, j Ijs hjs = 12 · 3 + 2 · 3 + 5 · 1 = 47 and j Ejs hm js = 12 · 1 + 2 · 2 + 31 · 1 = 47

are the same.

The above mentioned reformulations refer to the “inventory holding” and “production” vari-

ables. In contrast, Karmarkar and Schrage (Karmarkar and Schrage, 1985, p. 330) propose

to reformulate the “setup” variables (cf. Wolsey, 1997, p. 158). This approach can be illus-

trated by the following example: We assume a single production line, four diﬀerent products

y10=1

j=1 j=1

z141=1

y20=0

j=2 j=2

j=3 j=3

j=4 j=4

s=0 s=1 s=3 s=0 s=1 s=3

(a) Unique setup state. (b) Unique changeover.

j = 1, ..., 4, and four periods s = 1, ..., 4. In scheduling models, each setup state yjs is unique

as illustrated by the solid nodes of a possible model solution in Figure 5.5a. Due to that,

the changeover zijs must be unique as well (a single arc between two consecutive periods as

90 Chapter 5. Reformulations of the improved GLSPMS

shown in Figure 5.5b). Karmarkar and Schrage make use of this property. If a changeover

from product i to product j is executed in period s, a changeover in period s + 1 has to

start from product j consequently. This relationship can be expressed by the network ﬂow

condition (5.4).

zij,s−1 = zjis ∀j, s > 0 (5.4)

i i

With the help of this ﬂow conservation, the setup state variables yjs can be substituted by

i zijs .

Due to the direct link between inventory stock, production quantity, and demand, it is

possible to drop the inventory stock variables. Since production quantities and demand

values suﬃce to calculate the stock levels, the recursion of the inventory balance constraints

can be eliminated as shown in equation (5.5).

s

s

s

Ijs = Ij0 + xjτ − djτ − pji xiτ (5.5)

τ =1 τ =1 τ =1 i∈Nj

Then, each occurrence of Ijs can be substituted accordingly. However, the non-negativity

constraints (5.6) still have to be considered. From these constraints and from the substitu-

tion, it follows that constraints (5.7) must hold.

s

s

s

Ij0 + xjτ − djτ − pji xiτ ≥ 0 ∀j, s (5.7)

τ =1 τ =1 τ =1 i∈Nj

s

s

s

Ij0 + xjτ ≥ djτ + pji xiτ ∀j, s (5.8)

τ =1 τ =1 τ =1 i∈Nj

By rearranging the terms of (5.7), the new constraints (5.8) result. As can be seen, these

new constraints force the cumulated production quantities to be always suﬃcient for meeting

the cumulated primary and secondary demands.

In the literature, diﬀerent valid inequalities (cuts) for a variety of lot-sizing problems can be

found (cf. Pochet and Wolsey, 2006). One well-known and successfully applied class of those

cuts are the (l,S)-inequalities. They have been proposed by Barany et al. (1984) who regard

an uncapacitated single-item problem with s, τ = 1, ..., S periods where production enforces

5.3. Extended reformulations of the GLSPMS 91

Moreover, dynamic demand ds is considered as part of cumulated demands Dsτ = τσ=s dσ .

Barany et al. prove that constraints (5.9) are valid inequalities for the problem at hand.

xs + Dsτ ys ≥ D1τ ∀τ, T = {1, . . . , τ }, S ⊆ T (5.9)

s∈S s∈T \S

These inequalities enforce that in order to fulﬁll the cumulated demand D1τ of periods 1 to

τ , a suﬃcient quantity xs must be produced in a subset of these periods (S) if the machine

is not set up in the complementing set of periods (T \ S). But if there is a setup in a

period s belonging to T \ S, then it must be assumed that a quantity is produced which

might fulﬁll the demand of all consecutive periods up to period τ . Hence, Dsτ is used as the

corresponding coeﬃcient.3

Since the regarded problem is part of most of the lot-sizing problems, these (l,S)-cuts can

be often directly applied (or with slight modiﬁcations). This is also true for the GLSPMS.

Hence, Section 5.4 will introduce an adapted version of the (l,S)-inequalities.

As mentioned before, the reformulation approaches presented in Section 5.2 are now to be

applied to the GLSPMS. To begin with the extended reformulations, an SPL as well as a

network ﬂow formulation will be introduced next (q.v. Seeanner and Meyr, 2013).

The following SPL formulation is based on the reformulation of the MLCLSP presented

by Stadtler (1996). As Section 5.2.1.1 showed, SPL allows dropping the inventory stock

variables. Hence, the inventory balance constraints can be eliminated as well. However, the

additional constraints (5.8) shown in Section 5.2.1.4 still need to be considered for pre-items

to ensure the fulﬁllment of secondary demand in time.

Another specialty of the MLCLSP reformulation by Stadtler is that the production vari-

ables do not describe the total production quantities which are needed for “target” demands

but fractions of these demands. Hence, secondary demand needs to be a fraction as well.

That is why Stadtler makes use of total (net) demands (cf. Section 3.3.2.2) and the concept

of echelon stocks. This approach will be applied to the GLSPMS, too.

While using the same data, indices, index sets, and variables as in Section 4.2, the

resulting reformulation of the GLSPMS (q.v. Seeanner and Meyr, 2013) looks as follows:4

3

Note that these inequalities are called (l,S) because in the notation of the authors l = τ and S = S.

4

The numbers of the unchanged constraints compared to the improved formulation in Section 4.2.2 are

92 Chapter 5. Reformulations of the improved GLSPMS

Indices

σ, τ = 1, ..., S microperiods

u, ρ = 1, ..., T macroperiods

Index set

Nj+ set of all direct and indirect successors of product j

Data

hmjs marginal holding costs of product j in period s

dnjs net demand of product j in microperiod s

ft ﬁrst microperiod of macroperiod t

Variables

xljsτ ≥ 0 fraction of dnjτ which is produced on line l during microperiod s(≤ τ )

ljsτ ≥ 0

x share of xljsτ that can be used by successors in the same microperiod s

xljsτ ≥ 0 share of xljsτ that can (as WIP-stock) ﬁrst be used by successors in the

following microperiod s + 1 so that (5.12) and (5.13) hold

ejsτ ≥ 0 fraction of dnjτ which is externally purchased in microperiod s

Objective function

T −1

T

u−1

Min hm

j,fρ+1 −1 dnjτ xljsτ

j>0 t=1 u=t+1 ρ=t l s∈St τ ∈Su

s

T −1

T

u−1

+ hjs · max 0, Ij0 − djτ + hj,fρ+1 −1 dnjτ ejsτ

j>0 s∈Λ τ =1 j t=1 u=t+1 ρ=t s∈St τ ∈Su

S

S

+ slij zlijs + cplj dnjτ xljsτ + cl xbls + xls +

b e

cej dnjτ ejsτ + co · o s

l,i,j,s l,j,s τ =s l,s j,s τ =s s

(5.10)

subject to

ws = w s ∀s ∈ Φ (5.11*)

ljsτ + xljsτ = xljsτ

x ∀l, j, s, τ ≥ s (5.12)

xljss = 0 ∀l, j, s (5.13)

xljsτ ≤ yljs ∀l, j, s, τ ≥ s (5.14)

5.3. Extended reformulations of the GLSPMS 93

τ

τ τ

dniτ

xljsτ + ejsτ + pji e =1

n isτ

∀j, τ : dnjτ > 0 (5.15)

d

s=1 jτ

l s=1 s=1 i∈Nj+

σ

S

Ij0 + dnjτ (

xljsτ + xlj,s−1,τ )

l s=1 τ =s

σ

S

σ

σ

S

+ ejsτ dnjτ ≥ djτ + pji dniτ xlisτ ∀j > J0 , σ (5.16)

s=1 τ =s τ =1 l,i∈Nj s=1 τ =s

σ

S

S

Ij0 + dnjτ xljsτ + dnjτ ejsτ

l s=1 τ =σ+1 τ =σ+1

σ

S

σ

− dniτ xlisτ pji − djτ ≤ Ijmax ∀j, σ (5.17)

l i∈Nj+ s=1 τ =σ+1 τ =1

xbl1 = stlij zlij1 ∀l (5.18*)

i,j

xel,s−1 + xbls = stlij zlijs ∀l, s ≥ 2 (5.19*)

i,j

S

xbls + xls + alj dnjv xljsv + xels + xels = (ws+1 − ws ) + os ∀l, s (5.20)

j v=s

S

dnjτ xljsτ ≥ mlj (yljs − ylj,s−1 ) ∀l, j, s (5.21)

τ =s

yljs = 1 ∀l, s (5.22*)

j

yljs = 0 ∀l, j ∈

/ Jl , s (5.23*)

zlijs ≥ yljs + yli,s−1 − 1 ∀l, i, j, s (5.24*)

zlijs = 1 ∀l, s (5.25*)

i,j

S

xljsτ = 0 ∀l, s ∈

/ Πl (5.26)

j=0 τ =s

S

dnjτ ejsτ ≤ emax

j ∀j, s (5.27)

τ =s

os ≤ omax ∀s ∈ Λ (5.28*)

os = 0 ∀s ∈

/ Λ (5.29*)

xbls + xbls ≥ xbks + xbks − wS+1 (2 − yljs − ykis ) ∀s, (k, i, l, j) ∈ D (5.30*)

S

kisτ + xeks + xeks ≥ xels + xels − wS+1 (2 − ykis − yljs ) ∀s, (k, i, l, j) ∈ D

aki dniτ x (5.31)

τ =s+1

In the ﬁrst two rows of the new objective function (5.10) the inventory holding and WIP-stock

94 Chapter 5. Reformulations of the improved GLSPMS

costs are calculated. Therefore, we diﬀerentiate between three diﬀerent kinds of “inﬂow” to

the stock. The ﬁrst term refers to the production quantities which are determined by mul-

tiplying the production ratios by the net demand. In contrast to the GLSPMS formulation

in Section 4.2.2, the WIP costs are directly included here as the total shares xljsτ are con-

sidered. Since items which are in stock from macroperiod t to u incur (marginal) holding

costs in macroperiods t, ..., u − 1, the sum over all these macroperiods has to be taken. The

second term is necessary, because for the starting inventories the full holding costs have to

be charged as long as they are available. And ﬁnally, there might be items in the stock

which are externally purchased. These may also incur holding costs. The other terms on

the third line of the objective function (5.10)—setup, production, and standby costs as well

as costs for purchasing and overtime—are analogous to the original objective (4.1), but use

the demand fractions.

Since production quantities now carry the information how long they are stored, neither

are the inventory variables Ijs nor the inventory balancing constraints (3.79) needed any

longer. Holding costs can be directly calculated using xljsτ and the marginal holding cost

values. Fractions ejsτ for external purchasing also need two time indices s and τ because it

might be necessary to buy and process a pre-product earlier than its corresponding ﬁnal item

is needed. Thus, costs for external purchasing and constraints (5.27) have to be adapted

analogously. Constraints (5.12) allow the quantity splitting, whereas constraints (5.13) forbid

that the WIP-stock is used in the same period. Moreover, constraints (5.14) enforce the

needed setups.

As shown at the beginning of this section, it has to be ensured that primary demands are

fulﬁlled and pre-products are produced on time. Therefore, constraints (5.15) enforce that

the production fractions sum up to 100% where the external purchase has to be considered

as well—also of successor products because purchasing them reduces secondary demand.

However, we can restrict ourselves to products and periods with a positive net demand.

Additionally, constraints (5.16) replace Ijs ≥ 0 as illustrated above. They guarantee that

cumulated supply of pre-products equals or exceeds their cumulated demand for each mi-

croperiod σ, i.e. initial inventory plus the total production or external purchase for net

demand up to σ is not lower than the—up to the same period—total original gross demand

and secondary net demand, which has been induced by production.

It is interesting to note that the one-period lead-time of the variables x has to be taken

into account (xlj0τ = 0 ∀l, j, τ if Ij0 also includes the WIP-stock) when calculating the over-

all production quantity of pre-product j that is available to fulﬁll demand of microperiod τ

on the left-hand side of (5.16). Since the inventory stock variables are dropped, constraints

(5.17) have to be adapted analogously to (5.16). Note that variables xljsτ could be elimi-

nated by replacing them with (5.12) in the respective constraints. They are only introduced

for improving readability. The remaining constraints for minimum lot-sizes (5.26) and for

periods without production (5.27) as well as the second group of synchronization constraints

5.4. Valid Inequalities for the GLSPMS 95

As shown in Section 5.2.1.4, the ﬂow conservation constraints can be used for all production

lines. Hence, by introducing (5.32), constraints (3.88) (or (5.24*)) can be eliminated:

zlij,s−1 = zljis ∀l, j, s > 0 (5.32)

i i

Furthermore, the setup state indicators yljs in the remaining constraints are substitutable

by i zlijs . Since both variants—the Simple Plant Location and the network ﬂow for the

changeovers—are applicable, four diﬀerent extended reformulations are considered in the fol-

lowing: original (O), Simple Plant Location (S), ﬂow for changeover (F), and its combination

(SF).5

This section shows two diﬀerent groups of valid inequalities for the GLSPMS. First, an

adapted subset of the (l,S)-inequalities will be presented. And second, the setup forcing

constraints will be tightened up.

In Section 5.2.2 the (l,S)-inequalities were introduced. Constraints (5.33) represent a subset

of these inequalities which are adapted for the GLSPMS. If there is no setup for an item for

several periods (s+1, ..., τ ), the stock level in period s must equal or exceed the corresponding

cumulative demand in these periods. For the GLSPMS, it means that also the WIP-stock in

microperiod s and the purchased quantity of microperiods s + 1, ..., τ have to be taken into

account:

τ

τ

σ

Ijs + xljs + ejσ ≥ djσ · 1− zlijϑ ∀j ≤ J0 , s < S, τ > s : djτ > 0

l σ=s+1 σ=s+1 l,i ϑ=s+1

(5.33)

These inequalities can be adapted and limited to (inventories at the end of) macroperiods t

because positive demand can only occur during the last microperiod of a macroperiod, i.e.

5

Original here means the improved model as it was introduced in Section 4.2.2.

96 Chapter 5. Reformulations of the improved GLSPMS

djs = 0 for s ∈ / Λ (see Section 3.3.4.1). However, since the SPL formulation is based on

echelon stocks, the (l,S)-inequalities can be applied for all products at all stages. Accord-

ingly, constraints (5.34) enforce that if there is no setup, all fractions that are produced or

externally purchased (directly or already built in externally purchased successor items) must

sum up to 1.

⎛ ⎞

s

τ

⎜ dniσ ⎟

⎝ xljϑσ + ejϑσ + pji n eiϑσ ⎠ ≥

ϑ=1 σ=s+1:dn >0 l +

d jσ

jσ i∈Nj

τ

σ

1− zlijϑ ∀j, s < S, τ > s : dnjτ > 0 (5.34)

σ=s+1 l,i ϑ=s+1

binary and continuous variables. Often, a good choice of the “big-M” increases notably the

lower bound of the LP relaxation and thus reduces computational runtime. Since this value

represents an upper bound for the production quantity, a widely spread approach is to deﬁne

M as the minimum of accumulated remaining demand and available capacity in the current

period divided by the production coeﬃcient (see also Section 3.3.2.2):

Kt n

S

xljs ≤ min , d · zlijs ∀l, j, t, s ∈ St . (5.35)

alj σ=s jσ i

In the following, (K) denotes the stock-inequalities (5.33)–(5.34) and (M) the valid in-

equalities (5.35) for the big-M variant, respectively.

The previous chapter analyzed the maximum solvable problem size of the artiﬁcial instances

as well as the impact of diﬀerent solver settings like “automatic cut generation” and “presolv-

ing” of the standard MIP-solver. The corresponding tests have shown that the default solver

settings perform best on average. Thus, they are used for all experiments in the remainder

of this thesis.

5.5. Computational results 97

In the current section, the extended formulations “Flow Conservation” (F), “Simple Plant

Location” (S) and their combination (SF) are compared with the original formulation (O)

of Section 4.2.2, each with the stock inequalities (K) and valid inequalities for the big-

M variant (M) of Section 5.4 potentially being added. According to the observations of

Section 4.4.2, the test instances are diﬀerentiated by the three groups of size. Group G1

contains all instances with J × L × S ≤ 250, G2 with 250 < J × L × S ≤ 500, and G3 with

J × L × S > 500. The results are shown in Table 5.2.

- M K M+K average

PS IG AR PS IG AR PS IG AR PS IG AR PS IG AR

G1 O 100 126 13 100 73 13 100 69 11 100 54 12 100 80 12

F 100 127 19 100 65 11 100 67 18 100 50 25 100 77 18

S 100 41 13 100 50 9 100 41 11 100 46 12 100 44 11

SF 100 71 15 100 65 14 100 58 14 100 51 14 100 61 14

G2 O 72 54 477 72 49 417 67 41 545 74 40 558 71 46 499

F 71 49 424 75 42 430 68 36 428 73 34 468 72 40 438

S 72 34 415 73 38 371 70 33 448 75 37 506 73 36 435

SF 70 44 480 74 45 463 68 40 412 75 39 506 72 42 466

G3 O 22 4 349 27 7 634 25 4 603 27 7 568 25 6 546

F 30 4 451 37 4 732 38 6 927 44 6 860 37 5 763

S 27 3 976 29 4 513 25 3 709 25 3 487 27 3 671

SF 30 4 414 33 6 565 35 6 678 35 4 597 33 5 569

average O 66 80 269 68 54 273 65 49 322 69 42 336 67 56 300

F 68 76 261 72 46 308 69 44 336 73 35 363 70 50 317

S 68 34 309 69 39 239 66 33 291 69 37 302 68 36 285

SF 68 50 280 70 48 297 68 43 285 71 39 324 69 45 297

Table 5.2: Percentage of test instances solved to optimality in 3600 sec. (PS), the corre-

sponding integrality gap (IG), and the average runtime (AR) in sec. for the original formu-

lation (O), the extended formulations “ﬂow conservation” (F), “Simple Plant Location” (S),

and their combination (SF)—possibly complemented by stock inequalities (K) and big-M-

inequalities (M).

The ﬁrst column (PS) describes the percentage of instances that are solved in 3600 sec. to

optimality. For those instances, the percentage integrality gap = optimal solution - LP relaxation

LP relaxation

×

100 can be measured, which is given in the second column (IG). The quality of the diﬀerent

formulations is evaluated by inspecting the IG, since the LP relaxation describes the tightness

of the formulation: the smaller the gap within the same group of instances, the better

the formulation. In the third column the average runtime (AR) of the solved instances is

presented (sec.).

Interestingly, the Simple Plant Location formulations (S, SF) produce better IG values (36

and 45% on average) than are possible when using the original formulation or F standalone

(56 and 50% on average). These observations coincide with the results of Denizel and Süral

(2006) who regarded the CLSP with setup times. If S is not applied, the IG can at least

be improved by adding the big-M inequalities M (e.g., for O and F the IG can be reduced

98 Chapter 5. Reformulations of the improved GLSPMS

from 80 to 54% and from 76 to 46% when not using K, and from 49/44% to 42/35% when

using K). Unfortunately however, they do no make it possible to solve more instances. The

percentage of instances PS that can be solved to optimality within the time limit almost

stays the same regardless of the type of formulation used (67–70% on average).

What can also be seen is that for the Simple Plant Location formulation (S) the big-M

inequalities produce higher integrality gaps (raise from 34 to 39%). The plain explanation

is that constraints (5.14) are tighter than the corresponding adapted constraints (5.35).

As expected, less instances can be solved to optimality when the problem sizes increase.

Within group G1 all problem instances can be solved within the one hour time limit (PS =

100). This share reduces to an average of 25–37% within group G3. Within the groups G1

and G2 the values of PS are quite independent from the chosen formulation. However, for

the big instances within G3, the ﬂow formulation F performs signiﬁcantly better than the

others (e.g., 37 compared to 25% of the original formulation). But obviously, only those G3

instances which have a small IG can be solved to optimality at all.

Concerning the valid inequalities, it can be stated that it is always advantageous to use

the big-M-inequalities. The PS-values of column “M” are never worse than the corresponding

ones of column “-”. The same holds true when comparing columns “M+K” and “K”. All in

all the combination M+K performs best with respect to the number of optimally solvable

instances PS.

To complete the picture, Table 5.3 also shows diﬀerent performance indicators for the

problem instances that have proven to be too diﬃcult to be solved to optimality within the

time limit. The ﬁrst column PU—showing the percentage number of unﬁnished problems—

complements the PS column of Table 5.2. Correspondingly, group G1 is dropped since

all instances of this group have been solved to optimality. For all other instances, the

percentage duality gap (DG) after 3600 sec. is shown, where the duality gap is measured

as best solution - best lower bound

best lower bound

× 100. It gives an idea about the worst case quality of the best

(feasible) solution found after one hour.

Interestingly, the duality gaps DG decrease while the problem sizes are increasing from

G2 (21–28%) to G3 (14–18%). This appears counterintuitive at the ﬁrst sight. But one

has to keep in mind that—because of their smaller size— within the one hour time limit

more than twice the number of G2-instances have been solved to optimality as compared

to G3-instances. Thus, after one hour only the 27–29% “hard” instances of the middle-size

class G2 remain, which have a quite high DG. The larger instances of class G3 need longer

computation times anyway. Accordingly, after one hour also some “easier” problem instances

are left which already show a rather low DG.

Moreover, the S formulations oﬀer preferable duality gaps (the DGs of S are never higher

than the ones of O; the same holds true for SF and F), even though the percentages of

unﬁnished instances PU might be higher (F and SF for G3).

5.5. Computational results 99

- M K M+K average

PU DG PU DG PU DG PU DG PU DG

G2 O 28 31 28 27 33 25 26 28 29 28

F 29 30 25 24 32 23 27 20 28 24

S 28 26 27 24 30 25 25 25 27 25

SF 30 24 26 18 32 21 25 20 28 21

G3 O 78 15 73 17 75 17 73 14 75 16

F 70 18 63 18 62 18 56 18 63 18

S 73 14 71 13 75 14 75 13 73 14

SF 70 16 67 15 65 17 65 15 67 16

average O 34 21 32 21 35 20 31 19 33 20

F 32 22 28 20 31 20 27 19 30 21

S 32 19 31 18 34 18 31 18 32 18

SF 32 19 30 16 32 18 29 17 31 18

Table 5.3: Percentage of instances that are unﬁnished after 3600 sec. (PU) and the cor-

responding duality gap (DG) for the original formulation (O), the extended formulations

“ﬂow conservation” (F), “Simple Plant Location” (S), and their combination (SF)—possibly

extended by stock inequalities (K) and big-M-inequalities (M).

So far the instances have only been examined in terms of their size. But it might also be

important to learn about the impact of the underlying product structure or other character-

istics of the test instances like the utilization rate, the time-between-orders, and the relation

between setup costs and setup times.

For this reason, Table 5.4 shows the same results as Table 5.2, but now grouped by the

product structure. Indeed, three main eﬀects can be seen: ﬁrst, the big-M variants M and

M+K are still preferable (best value per model formulation is marked in italics). Usually,

M+K performs best, but especially for divergent product structures, M also shows very good

results. Second, the integrality gap of the serial instances is very low compared to the other

product structures (between 5 and 9% on average). In fact, almost all instances of group

G3, which have been solved to optimality according to Table 5.2, actually show a serial

product structure. Thus, divergent and general problem structures seem to be signiﬁcantly

more diﬃcult than serial ones. And third, the ﬂow formulation is beneﬁcial for serial and

general product structures. The PS values of F outperform O and the PS values of SF

outperform S for these kinds of problems. In most cases F again seems preferable to SF.

However, the situation is diﬀerent for the divergent instances. Here, F seems to be rather

counterproductive and S performs best (followed by O).

Summing up, the ﬂow formulation F in combination with both valid inequalities M+K

performs best for the serial instances (88%) and for the general instances (78%). On the

contrary, F is rather disadvantageous for divergent product structures. Here, the Simple

Plant Location formulation S—again using the extension of M+K—is superior (66%).

In accordance with these results, the potential inﬂuence of the other problem charac-

100 Chapter 5. Reformulations of the improved GLSPMS

- M K M+K average

PS IG AR PS IG AR PS IG AR PS IG AR PS IG AR

SER O 70 11 176 73 9 222 73 7 286 74 6 205 72 9 223

F 78 11 192 84 8 291 83 7 352 88 6 335 83 8 295

S 74 5 373 75 6 165 73 5 324 74 6 258 74 5 279

SF 78 9 250 79 8 187 79 7 275 83 7 306 79 8 255

DIV O 63 124 375 64 95 352 59 92 388 64 75 480 62 96 399

F 60 121 409 62 85 385 56 87 307 59 69 348 59 91 363

S 64 58 342 65 66 345 62 57 339 66 64 398 64 61 356

SF 58 92 327 63 87 439 58 79 325 61 73 337 60 83 358

GEN O 68 110 223 68 54 208 66 44 263 70 41 289 68 62 246

F 70 111 123 72 48 202 74 46 352 78 37 436 74 59 283

S 66 37 131 66 46 163 66 37 140 66 39 185 66 40 154

SF 72 54 253 72 48 237 72 45 239 74 41 335 73 47 267

average O 66 80 269 68 54 273 65 49 322 69 42 336 67 56 300

F 68 76 261 72 46 308 69 44 336 73 35 363 70 50 317

S 68 34 309 69 39 239 66 33 291 69 37 302 68 36 285

SF 68 50 280 70 48 297 68 43 285 71 39 324 69 45 297

Table 5.4: Percentage of test instances solved to optimality in 3600 sec. (PS), the correspond-

ing integrality gap (IG), and the average runtime (AR) in sec. for all formulations—now

grouped by the product structure.

teristics is examined with the help of the two “best” formulations F+M+K and S+M+K.

To see the impact of the utilization rate, the capacity (cap) of the instances is reduced

by 20% (w̄scap := 0.8w̄s ∀s ∈ Φ). The time-between-orders (tbo) is supposed to be a

second factor. Therefore, the setup costs are doubled (stbo lij := 2slij ∀l, i, j) so that hold-

ing inventory becomes more attractive and larger lot-sizes are to be expected. Finally,

since setup costs and times are the same for the current instances, the setup costs are

“inverted” (setup). This means that for each line the setup cost for the most expensive

changeover is the cheapest, the second most expensive is now the second cheapest, and so on

(ssetup

lij := maxi,j {slij } + mini,j {slij } − slij ∀l, i, j = i with ali > 0 and alj > 0). Thus, setup

costs and setup times are not proportional any longer.

PS IG AR PS IG AR PS IG AR PS IG AR

SER F+M+K 88 6 335 89 5 129 89 7 195 88 6 302

S+M+K 74 6 258 73 5 125 76 7 326 75 6 340

DIV F+M+K 59 69 348 57 39 461 66 68 416 63 70 347

S+M+K 66 64 398 55 40 226 65 70 326 64 69 365

GEN F+M+K 78 37 436 76 70 320 78 50 405 82 39 464

S+M+K 66 39 185 72 79 446 66 56 280 66 39 124

average F+M+K 73 35 363 72 32 288 76 40 325 75 37 356

S+M+K 69 37 302 65 36 240 69 43 317 68 38 305

Table 5.5: Percentage of test instances with diﬀerent problem characteristics solved to op-

timality in 3600 sec. (PS), the corresponding integrality gap (IG), and the average runtime

(AR) in sec. for the two best formulations S and F including stock inequalities (K) as well

as big-M-inequalities (M).

5.5. Computational results 101

In Table 5.5 the corresponding results are shown. In column “normal” the values of Ta-

ble 5.4 are repeated for comparison. Obviously, general conclusions in terms of solvability

can hardly be derived. For instance, by reducing the capacity the IGs for the divergent prod-

uct structures become smaller (39%/40%), but less instances could be solved (57%/55%).

In contrast, PS is nearly the same for the general structure with the ﬂow formulation and

even higher for the Simple Plant Location formulation, although the IGs are substantially

larger now (70%/79%).

The impact of a diﬀerent time-between-orders is almost negligible in terms of PS. Only

the IGs increase slightly for most of the instances, again to the highest degree for the general

product structures. The reason is that—compared to the other cost terms of the objective

function—mainly the setup costs are “aﬀected” by the LP relaxation. By allowing fractional

setup costs instead of binary ones, the inﬂuence of an increase gets mitigated. For the “setup”

instances no real diﬀerence could be observed at all. Accordingly, it seems that instances

with proportional setup times and costs are not easier to solve.

As shown in Section 5.5.2, the “Simple Plant Location” based reformulation appears to

perform slightly better for instances with a divergent product structure, whereas the ﬂow

formulation shows advantages for all other instances. Furthermore, the usage of the stock

and the big-M inequalities are always recommendable. Nevertheless, even for instances of

moderate size the percentage of unﬁnished problems (that cannot be solved to optimality

within a reasonable amount of time) as well as the corresponding duality gaps are quite

high. Thus, standard MIP-solvers do not appear to be promising for solving this problem

satisfactorily. For this reason, heuristics will be designed and tested in the next chapter.

Chapter 6

Even though the reformulations presented in Chapter 5 provide slightly improved (dual)

lower bounds and thus reduce the computational eﬀort of a standard MIP-solver, it is still

not possible to solve mid- and large-sized problem instances exactly.

Accordingly, a heuristic approach is needed which is able to ﬁnd suﬃciently good (primal)

feasible solutions in a reasonable amount of time. Therefore, Section 6.1 will ﬁrst present

selected meta-heuristic approaches which can be applied to diﬀerent kinds of problems.

Afterwards, several new heuristics for the GLSPMS will be introduced in Section 6.2 which

are based on the approaches before. Section 6.3 will show the computational results of

the new heuristics and complement the ﬁndings of Seeanner et al. (2013). Finally, a brief

summary will be given in Section 6.4.

Typically, heuristics are grouped into construction and improvement heuristics. The former

try to generate a ﬁrst initial solution from scratch, whereas the latter try to improve a given

feasible solution (cf. Pochet and Wolsey (2006) and Zäpfel et al. (2010) among others). How-

ever, there are procedures which are able to construct and improve solutions. For instance,

exact algorithms, which are truncated during the search, belong to both groups. One im-

portant representative of these algorithms is Truncated MIP which will be presented next.

After that, the construction heuristics LP&Fix and Relax&Fix will be introduced. They

have the advantage of being easy to adapt and implement. Furthermore, the improvement

heuristics Exchange, Variable Neighborhood Search (VNS), and Variable Neighborhood De-

composition Search (VNDS) will be explained. As the latter two are both based on Local

Search, this principle will be illustrated as well.

DOI 10.1007/978-3-658-02089-7_6, © Springer Fachmedien Wiesbaden 2013

104 Chapter 6. Heuristics for the improved GLSPMS

a problem instance exactly (cf. Pochet and Wolsey, 2006, pp. 87–113). An example for a

straightforward heuristic is simply to truncate the exact search procedure of a MIP-solver.

The best solution found so far will be accepted as the solution output. For that purpose,

several diﬀerent stopping conditions can be applied. For instance, as soon as a ﬁxed amount

of time has elapsed, the i -th incumbent solution is found, or a certain duality gap is reached,

the solver run might be terminated. Obviously, this heuristic is both at the same time—

construction as well as improvement heuristic.

As next, the two well-known construction heuristics LP&Fix and Relax&Fix will be

presented. These heuristics also belong to the class of MIP-based heuristics1 , because they

generate subproblems of the regarded MIP and pass them to a MIP-solver.

6.1.2 LP&Fix

The LP&Fix heuristic approach consists of two simple steps. First, all integer variables are

relaxed (set to continuous) and the resulting LP is solved. Those integer variables which

already have an integer value in the LP solution get ﬁxed to that value. Finally, the remaining

MIP is solved regularly (cf. Pochet and Wolsey, 2006, pp. 108–109).2

6.1.3 Relax&Fix

The Relax&Fix3 heuristic tries to reduce the complexity of the original problem by solving

iteratively R diﬀerent MIP-subproblems: In each run r = 1, ..., R the integer variables are

decomposed into three pairwise disjoint sets Fr , Rr , and Qr . The set Fr contains all variables

that are ﬁxed to a certain value, whereas the variables in Rr and Qr are optimized. However,

only the elements of Rr have to fulﬁll the integrality constraints. The variables belonging

to Qr are relaxed.

To ﬁnd a feasible solution to the original problem instance, the idea of Relax&Fix is to

use a partition F1Δ , ..., FRΔ of all integer variables, where FrΔ contains all variables which

have to be ﬁxed after run r. Accordingly, the relation between the sets is Fr = Fr−1 ∪ FrΔ

for all r, where F0 = ∅. In order to ﬁx variables after run r, they should have been optimized

in the run r − 1 before. Accordingly, FrΔ ⊆ Rr−1 must be fulﬁlled (cf. Pochet and Wolsey,

2006, pp. 109–111).

1

Some authors also classify these procedures as combinations of exact and heuristic algorithms (cf., e.g.,

Puchinger and Raidl (2005) and Lang (2009)).

2

Note that after the ﬁrst solution is found, the LP&Fix becomes an improvement heuristic.

3

This heuristic is also known as Fix&Relax.

6.1. Selected meta-heuristic approaches 105

The ﬁrst improvement heuristic which will be described next is the MIP-based procedure

Exchange.

6.1.4 Exchange

This heuristic follows a similar idea as Relax&Fix to reduce the size of the solution space

which is presented next. Afterwards, an implementation of the Exchange for the MLCLSP

will be outlined.

For Exchange, a decomposition method needs to be deﬁned which partitions all integer

variables but now into two subsets F and R. Given an initial solution, the variables in F

are ﬁxed to the values of this solution, the other variables are “released”. Then, the resulting

“sub-MIP” is solved to optimality. Since at least the initial solution can be found by the

MIP-solver, the new solution is always better than or equal to the old one. Due to this and

the fact that Exchange does not use any relaxation (in contrast to the Relax&Fix), at all

times a feasible solution is available.

Obviously, this method can also be applied in an iterative manner. Then, the newly

computed solution becomes the “initial” one and the process of ﬁxing and optimizing is

repeated (cf. Pochet and Wolsey, 2006, p. 113).

A problem-speciﬁc implementation of this procedure is Fix&Optimize which solves the

MLCLSP.

Similar to Relax&Fix the decompositions for the Exchange heuristic can be built in many

diﬀerent ways. In order to solve the MLCLSP, Helber and Sahling (2010) propose a sys-

tematic approach for decomposing the setup variables zjt (cf. Section 3.3.1) into the disjoint

sets F and R. They embed product-, resource-, and process-oriented decompositions in an

iterative procedure and call this Fix&Optimize.

Iterative procedure Helber and Sahling combine these three types of decomposition in

four diﬀerent ways and embed them into an iterative procedure which applies them one

or several times (in a loop). This procedure starts with an initial solution as the current

best. Then, in each iteration, a new subproblem is created by using the Exchange procedure

with the corresponding decomposition which ﬁxes the variables zjt ∈ F to the values of the

current best solution. The resulting problem is solved to optimality. If a better solution

106 Chapter 6. Heuristics for the improved GLSPMS

can be found, it becomes the new current best.4 Fix&Optimize terminates if the maximum

number of iterations is reached (an external parameter to be set) or all decompositions are

not able to ﬁnd a better solution.

The best results for the MLCLSP are obtained with the sequence “product-oriented”,

“resource-oriented” and ﬁnally “process-oriented” decomposition. In the following, the three

diﬀerent decompositions are explained in more detail. For that purpose, a small example

j=1 2 resource 1

3 4 resource 2

with four products and two resources is used (Figure 6.1), where products j = 1 and j = 2

are producible on resource l = 1 as well as products j = 3 and 4 on resource l = 2. The

corresponding planning horizon comprises T = 6 periods.5

a sub-MIP is built by releasing the setup variables of item j for all periods t. This means

that in each iteration the procedure decides whether the regarded product should be set

up. Accordingly, there are J diﬀerent product-oriented decompositions. Helber and Sahling

point out that the sequence of releasing products plays an important role for the quality

of the results. They suppose that the higher the costs incurred by a speciﬁc item the

more important this item is. To build a “good sequence” they propose a cost function for

evaluation. This function is based on the LP relaxation values of the model and returns a

cost value consisting of setup, holding, and weighted overtime costs for each product. The

products are then ordered according to decreasing values.

To demonstrate this procedure, Figure 6.2 shows the decompositions of product 1 and 2

of the regarded example. For instance, the former one builds the subsets R = {z11 , . . . , z16 }

and F = {z21 , . . . , z46 }.

product but all products which are producible on a certain resource. Starting with l = 1, all

resources are considered in an ascending order.

However, since all sub-MIPs should be solved to optimality in a reasonable amount of

time, the setup variables are not released for all periods at once. Helber and Sahling apply a

4

Helber and Sahling use overtime as a relaxation for the MLCLSP. Therefore, Fix&Optimize does not

accept a better solution which includes overtime if the current best is feasible with the original capacity.

5

Note that each product can only be processed on a single machine (see also Section 3.3.1).

6.1. Selected meta-heuristic approaches 107

periods periods

t=1 t=2 t=3 t=4 t=5 t=6 t=1 t=2 t=3 t=4 t=5 t=6

products

products

j=2 fixed j=2 binary

periods periods

t=1 t=2 t=3 t=4 t=5 t=6 t=1 t=2 t=3 t=4 t=5 t=6

products

products

108 Chapter 6. Heuristics for the improved GLSPMS

“rolling time window“ with a size of four periods (see also Section 6.2.3). They start at the

beginning of the planning horizon with an overlap of two periods in each iteration. Hence,

periods t = 1, ..., 4 are regarded ﬁrst, then periods t = 3, ..., 6 and so on. Note that for the

last decomposition the time window might contain less than four microperiods. Figure 6.3

illustrates this decomposition for resource l = 1 of the above example.

product, the setup variables are released. Helber and Sahling also consider these product-

couples in an ascending order. To get sub-MIPs with a reasonable size, the planning horizon

is divided into two parts (as two products are considered). From the product structure of

periods periods

t=1 t=2 t=3 t=4 t=5 t=6 t=1 t=2 t=3 t=4 t=5 t=6

products

products

(a) 1st process decomposition of product-couple 1 (b) 2nd process decomposition of product-couple 1

and 4. and 4.

the example in Figure 6.1 three diﬀerent couples result. Therefore, six sub-MIPs are created.

The two subproblems of the couple (1,4) are illustrated in Figure 6.4.

In addition to this MIP-based heuristic, there are also several other types of improvement

heuristics like Population-, Swarm-, and Local Search-based procedures among others (cf.

Zäpfel et al., 2010).6 Especially the last ones are wide-spread in the literature and play an

important role in the following. Hence, the principle of Local Search is introduced next.

In this section, the general idea of Local Search will be explained ﬁrst. Afterwards, a well-

known representative will be introduced as an example to illustrate this procedure.

6

Note that this diﬀerentiation is not clear-cut. Indeed, there are many heuristics which belong to several

classes or even represent hybrid procedures.

6.1. Selected meta-heuristic approaches 109

Starting from a current feasible solution x, Local Search does not examine the whole solution

space but the (local) neighborhood N (x) of x. The neighborhood is deﬁned as the set of all

solutions x (also termed candidates) which can be obtained by an atomic transformation

step from x. A very useful assumption for this transformation is that the whole solution

space should be reachable by a repeated application. To illustrate this procedure, 2-opt is

introduced next as an example.

A well-known Local Search procedure is the 2-opt heuristic which has been developed for the

Traveling Salesman Problem (cf. Croes (1958) and Applegate et al. (2006) among others).

The transformation step of 2-opt is exchanging two edges of a feasible solution by two

other edges to build a new feasible tour. Accordingly, the neighborhood of x is deﬁned as

the set of all tours which can be reached by this transformation step. To illustrate this,

Figure 6.5a shows a feasible solution for the instance of the Traveling Salesman Problem

given in Section 4.3.1. This round trip x has a length of 109 units. By exchanging the edges

(1, 2) and (4, 5) as shown in Figures 6.5b and 6.5c, a new neighbor x is reached with length

104.

1 1 1 21

4 4 4

19 19 19

15 15

25 17 25 17 17

5 5 5

3 3 14 3

2 33 2 33 2 33

(a) Feasible solution. (b) Removing two edges. (c) Adding two new edges.

Figure 6.5: Example for exchanging two edges as the transformation step of 2-opt.

The complete neighborhood N (x) consists of ﬁve neighbors where the other ones can be

determined in the same way (exchanging the pairs of edges {(1, 2), (3, 4)}, {(2, 3), (1, 5)},

{(2, 3), (4, 5)}, and {(3, 4), (1, 5)}). The original 2-opt only allows improvements and thus

belongs to the class of descent procedures.7 But the question which arises now is how to

explore the resulting neighborhood and ﬁnd a better solution.

7

Note that this term refers to minimization problems. Correspondingly, these procedures are called

“ascent” for maximization problems.

110 Chapter 6. Heuristics for the improved GLSPMS

Descent procedures only allow a candidate to become a new solution if it is better than the

current solution. There are two possible ways to get a better solution: Either the Local

Search procedure examines successively one candidate after another. As soon as a better

solution is found, it becomes the current best (ﬁrst ﬁt). Here, the sequence of the candidates,

which can be ﬁxed or randomly generated, plays an important role. Alternatively, a subset

of the neighborhood can be created (this can also be the complete neighborhood) and the

best candidate is chosen out of it (best ﬁt). For instance, a ﬁrst ﬁt method would directly

accept the candidate shown in Figure 6.5c because it is better than the one in Figure 6.5a

(104 < 109). By contrast, a best ﬁt method would examine, e.g., all ﬁve neighbors to get

the best (which is actually better than 104).

Descent procedures terminate if no improvement can be achieved anymore. In that

case, the current solution is called a local optimum as it is better than all of its neighbors.

Unfortunately, such local optima are usually reached quite fast by descent Local Search

methods and may have big gaps to the “global” optimum. Correspondingly, the major

challenge of this kind of heuristics is to escape from these local optima. That is why there

are also descent-ascent Local Search heuristics which accept worse solutions temporarily as

illustrated next.

By also allowing a worsening within the search, a descent-ascent procedure does not check

the candidate for improvement but whether it can be accepted. Hence, an acceptance rule

has to be deﬁned. Note that this is a generalization since descent methods use a “special”

acceptance rule which only allows improvements.

As one can imagine, Local Search procedures can be implemented in many diﬀerent ways

(diﬀerent neighborhoods, diﬀerent acceptance rules, etc.).8 One popular representative is

the Variable Neighborhood Search.

Mladenović and Hansen (1997) introduce the principle of Variable Neighborhood Search

(VNS). Generally speaking, to avoid being entrapped in a local optimum which is not the

global one, VNS uses a series of K diﬀerent neighborhoods Nk which are usually nested as

Nk ⊆ Nk+1 . However, the relation between the neighborhoods of VNS and Local Search is

not uniquely deﬁned (cf. Hansen and Mladenović, 2001). In this thesis, it is assumed that

these neighborhoods Nk are independent from the neighborhood of the underlying Local

8

To name only a few, there are Tabu Search, Simulated Annealing or Threshold Accepting (Meyr, 2000).

6.1. Selected meta-heuristic approaches 111

Search method.9 Hence, existing Local Search-based heuristics need not to be adapted but

can be directly embedded into VNS.10 Algorithm 2 illustrates the underlying procedure.

initialization: ﬁnd an initial solution x with any construction heuristic

1 k ← 1;

2 repeat

3 Shaking: choose solution x’ randomly from neighborhood Nk (x);

4 Local Search: ﬁnd local optimum x” starting from x’;

5 if x” better than x then

6 x ← x”;

7 k ← 1;

8 else

9 k ← k + 1;

10 end

11 until k = K;

the loop is entered. Each loop run consists of two steps. In the ﬁrst Shaking step, a

new solution x is chosen randomly from the current neighborhood Nk of the current best

solution x as illustrated in Figure 6.6a. Then, in a second step, a Local Search procedure is

started. If the local optimum x is found (Figure 6.6b), the quality of this new solution is

compared to the previous solution x.

Local Search

x“

x‘ x‘

Shaking

x x

Nk

(a) Shaking phase. (b) Local Search phase.

If x is better than x, the new solution becomes the current best and the ﬁrst neigh-

borhood is searched again. Otherwise the next neighborhood Nk+1 is used as shown in

Figure 6.7a and a new neighbor x is randomly chosen from that. This neighbor might pos-

sibly lie outside the old neighborhood (but not has to). From x , the independent Local

Search heuristic is started again (Figure 6.7b).

9

Mladenović and Hansen also propose a VNS-variant which uses the diﬀerent neighborhoods Nk for the

Local Search. They call this a Variable Neighborhood Descent (VND) heuristic.

10

Note that the existing heuristic becomes a “multi-start” procedure.

112 Chapter 6. Heuristics for the improved GLSPMS

Nk+1

x x

Shaking

Local Search

x‘ x‘

Example The principle of 2-opt and its neighborhood structure can be generalized to p-

opt where the neighborhoods are based on the exchange of p edges. We assume now that

the neighborhoods of VNS are N1 = 2-opt, N2 = 3-opt, Nk = (k+1)-opt. Starting with the

solution given in Figure 6.5c as x, the Shaking might result in solution of Figure 6.8 which

is a 2-opt neighbor of x. From this solution with length 121, the Local Search is started. In

1 21 1 21

4 4

38

15 15

25

5 5

13

14 3 3

2 33 2 33

this example, 2-opt is also used as Local Search procedure for simpliﬁcation.11 It terminates

with the local optimum x that has a solution value of 107 which is worse than 104 (cf.

Figure 6.9). Correspondingly, VNS might move to 3-opt as the succeeding neighborhood. In

the next step, a random neighbor x is chosen—but now from the 3-opt neighborhood of x.

Subsequently, the 2-opt Local Search is executed again starting from x .

In recent years, many variants of the VNS have been successfully applied to a wide range

of optimization problems. Examples include problems of graph theory (e.g., the traveling

salesman problem (Felipe et al., 2009; Hansen and Mladenović, 2001), minimum spanning

trees (Naji-Azimi et al., 2010; Ribeiro and Souza, 2002)), supply chain planning problems

(e.g., car sequencing (Prandtstetter and Raidl, 2008)), continuous optimization (Mladenović

et al., 2008) or lot-sizing and scheduling problems (Almada-Lobo and James, 2010; Almada-

11

The neighborhood for the Local Search should be smaller than the ones of VNS, otherwise VNS would

be useless.

6.1. Selected meta-heuristic approaches 113

Lobo et al., 2008). One important variant of the VNS is called the Variable Neighborhood

Decomposition Search which is shown next.

The Variable Neighborhood Decomposition Search (VNDS) aims at solving large problem

instances. Therefore, the search space for the Local Search is reduced by ﬁxing parts of

the solution (space). The parts to ﬁx are determined by decomposition methods which

are controlled by the diﬀerent neighborhoods of VNDS. Hansen et al. (2001) deﬁne such

a neighborhood Nk (x) as the set of every possible ﬁxation where “all but k attributes (or

variables or subset of variables) are ﬁxed” of solution x. Unfortunately, it does not become

clear from this deﬁnition if the neighborhood contains all possible solutions which can be

obtained from x when all but k out of n variables/attributes are ﬁxed, or if it contains all

n

k

possible combinations of ﬁxations (cf. Lazić et al., 2010).

initialization: ﬁnd an initial solution x with any construction heuristic

1 k ← 1;

2 repeat

3 Shaking: choose randomly f from neighborhood Nk (x);

4 Local Search: ﬁnd local optimum x’ in subspace Dk deﬁned by the underlying

ﬁxation f;

5 if x’ better than x then

6 x ← x’;

7 k ← 1;

8 else

9 k ← k + 1;

10 end

11 until k = K;

Nevertheless, the general idea is that the Shaking implies a ﬁxation which results in a

subset Dk of the solution space. Hence, we deﬁne now that the neighborhood Nk (x) only

comprises the ﬁxations as illustrated in Algorithm 3. Then, the Local Search determines

the local optimum in the resulting subspace Dk .12 Accordingly, the major diﬀerence of the

VNDS compared to the VNS is in the Local Search phase. After the local optimum is found,

the steps are repeated in the same way as for the VNS. The following example demonstrates

this procedure.

Example For the small TSP example from above, a neighborhood decomposition could be

deﬁned as the ﬁxation of 1 edge. Accordingly, the VNDS neighborhood of the solution given

12

Note that the idea of ﬁxing parts and solving the remaining is identical to Exchange.

114 Chapter 6. Heuristics for the improved GLSPMS

in Figure 6.5a consists of the ﬁve edges (1, 2), (1, 5), (2, 3), (3, 4), and (4, 5). By selecting one

of these in the Shaking phase, e.g., edge (2, 3), the subspace Dk comprises all solutions which

include (2, 3). Hence, at the beginning of the Local Search phase, 2-opt is only allowed to

examine the three neighbors {{(1, 2), (3, 4)}, {(1, 2), (4, 5)}, {(1, 5), (3, 4)}}. However, such a

decomposition only becomes useful for larger problems. If, for instance, a TSP has to be

solved with a large number of cites/nodes, then even the 2-opt neighborhood of a solution x

would be large as well.

Based on the meta-heuristics illustrated in Section 6.1, several new procedures for the

GLSPMS are developed (q.v. Seeanner and Meyr (2013) and Seeanner et al. (2013)). They

will be presented now.

As mentioned in Section 4.4, Xpress is used as MIP-solver throughout this thesis. Hence, for

the Truncated MIP heuristic, the corresponding GLSPMS model formulation is loaded and

started in Xpress. Due to the high gaps of the GLSPMS as a GLSP-based formulation (cf.

explanations of Section 4.3.2 and results of Section 5.5), the runtime is used as the stopping

condition.

The integer variables of the improved GLSPMS and of its reformulations are either the binary

setup state variables yljs or the changeover variables zlijs . Hence, for the LP&Fix heuristic,

these binary variables are allowed to take on continuous values in the interval [0; 1]. The

resulting LP is solved and afterwards the variables which are either 0 or 1 are ﬁxed to the

corresponding value. Finally, the remaining MIP has to be computed.

Since the Relax&Fix procedure as presented in Section 6.1.3 is a general approach, there

are, of course, diﬀerent ways to build these decompositions. However, for lot-sizing (and

scheduling) problems, it is common to use a time decomposition with a rolling scheme as

illustrated in Figure 6.10 (cf. Stadtler (2003) or Ferreira et al. (2009) among others). In

that example, a time window with a ﬁxed size of three periods is used. All setup variables

belonging to the time window are solved as binaries. In the ﬁrst run r = 1, the time window

6.2. Heuristics for the GLSPMS 115

periods

t=1 t=2 t=3 t=4 t=5 t=6 t=7 t=8 t=9 t=10 t=11 t=12

…

starts at the beginning of the planning horizon. Accordingly, the setup variables of the

periods t = 1, ..., 3 are set to binary, all others to continuous and the optimization process is

started. In the second run r = 2, the time window is moved three periods forward (rolling

scheme). The optimized binary variables of the ﬁrst three periods are ﬁxed. After the

ﬁxation, the optimization is started again with the new time window (periods t = 4, ..., 6).

This procedure is repeated until the time window reaches the end of the planning horizon.

For models with a two-level time structure like the GLSPMS, such a time window should

always comprise a whole macroperiod. Otherwise a decision might become too myopic, since

inventory holding costs are not regarded properly and a good solution cannot be found.

Regarding the above example, the sets of binary variables of all runs are disjoint.13 However,

it is also possible to let the time windows overlap. Such a scheme is presented in Figure 6.11.

Here, the time window comprises four periods with an overlap of two. Since the relaxation

periods

t=1 t=2 t=3 t=4 t=5 t=6 t=7 t=8 t=9 t=10 t=11 t=12

runs

…

of the binary variables may have a strong impact on the variables of the time window, the

overlap helps avoid ﬁxing variables to disadvantageous values by looking more ahead. But

it also leads to subproblems which are harder to solve (more binary variables).

As extensively explained in Chapter 4 the GLSPMS oﬀers the two diﬀerent model relax-

ations of external purchasing and overtime. These two allow to construct easily and fast a

simple initial solution as shown next.

13

Some authors deﬁne Relax&Fix only for disjoint sets (cf. Pochet and Wolsey, 2006, p. 109).

116 Chapter 6. Heuristics for the improved GLSPMS

The ﬁrst approach is to preserve the initial setup state over the whole planning horizon, i.e.

no changeover is allowed. This is achieved by adding constraints (6.1) to the GLSPMS.

Then, most of the products cannot be produced but have to be externally purchased. If

a product j cannot be purchased in reality, purchasing costs become penalty costs for not

producing. These costs are deﬁned for each product as ej = maxl,i {sclij + hj }. The upper

bounds emax

j are set to the net demand values (see Section 3.3.2.2) to prevent them from

becoming a limiting factor.

Another possibility to get a ﬁrst initial for the GLSPMS solution is to apply a setup sequence

where each product is set up at least once per macroperiod. Hence, the limited capacity is

disregarded and the overtime relaxation comes into play. The easiest way to generate this

setup sequence is to determine a sequence for a single macroperiod which is then repeated for

the whole planning horizon. For that purpose, the following model formulation was devel-

oped. This model uses the average net demand of the products and divides the macroperiod

into V − 1 microperiods as follows:14

RS model

Index

v = 1, ..., V microperiods

Data

drs

j average net primary demand for product j (= [ s djs − Ij0 ]+ /T )

rs

Ij0 =0 initial inventory is set to 0

Mlj big-M which is needed for the setup forcing constraints. This value

should be as tight as possible and therefore it is computed recur-

sively for each line l and item j. Assuming that the items are sorted

by the lowlevel code, we iterate over j = 1, ..., J and l: Mlj =

max mlj , i∈N + pji k max{mki , Mki } + drs j

j

14

Note that all data without the superscript ’rs’ is derived from the model presented in Section 4.2.2.

6.2. Heuristics for the GLSPMS 117

Variables

rs

Ijv ≥0 inventory of product j at the end of microperiod v (units)

xljv ≥ 0

rs

total quantity of product j produced in microperiod v on line l (units)

rs

yljv ∈ {0, 1} setup state: yljv = 1 if line l is set up for product j in microperiod v (0

otherwise)

rs

zlijv ≥0 takes on 1 if a changeover from product i to product j takes place on line

l during microperiod v (0 otherwise)

Objective function

min rs

slij zlijv + cplj xrs

ljv (6.2)

l,i,j,v l,j,v

subject to

rs

Ijv = Ij,v−1

rs

+ ljv −

xrs pji xrs

liv ∀j, v = (V − 1) (6.3)

l l i∈Nj

−1 = Ij,V −2 + lj,V −1 − li,V −1 − dj ∀j

rs rs

Ij,V xrs pji xrs rs

(6.4)

l l i∈Nj

rs

Ijv ≤ Ijmax ∀j, v (6.5)

xrs

ljv ≤ rs

Mlj yljv ∀l, j, v (6.6)

rs

yli,v−1 + yljv

rs

− 1 ≤ zlijv

rs

∀l, i, j, v (6.7)

rs

ylj,V = rs

ylj1 ∀l, j (6.8)

ljv ≥ mlj (yljv − ylj,v−1 )

xrs ∀l, j, v

rs rs

(6.9)

rs

yljv =1 ∀l, v (6.10)

j

The objective (6.2) is to minimize the setup and production costs. Constraints (6.3)

and (6.4) represent the inventory balance. The former assure that the secondary demand is

fulﬁlled. The latter enforce that the primary demand is met which only occurs in microperiod

V −1. Constraints (6.5) limit the stock levels and (6.6) are the setup forcing constraints. Note

that “big-M” is reduced to the maximum of the minimum lot-size and the sum of primary and

secondary demand. With the help of (6.7) the change of diﬀerent setup states is captured.

As the resulting setup pattern of this model has to be applied repeatedly to the GLSPMS,

constraints (6.8) are needed to assure that the setup sequence ﬁts together. Note that there

is only positive demand in microperiod V − 1 per product. However, one further microperiod

must be considered because the changeover between the two microperiods V −1 and V might

diﬀer from the changeover between microperiods 0 and 1. Otherwise, minimum lotsizes in

combination with limited stock levels potentially lead to infeasible plans. Furthermore, (6.8)

enforce the equality with v = 1 because for some test instances the production lines are

118 Chapter 6. Heuristics for the improved GLSPMS

supposed to be in an idle state at the beginning (yl00 = 1). Again, a minimum lot-size has

to be produced after a changeover (6.9) and a unique setup state is enforced for each line

and microperiod (6.10).

To ﬁnd a feasible solution for the model above, a certain, unknown minimum number of

microperiods V is needed. Starting with V = 1, the number of microperiods V is incremented

stepwise by 1. As soon as the ﬁrst feasible solution for the RS model has been found for a

value V = V , the execution is stopped. The computed setup sequence will then be ﬁxed for

every macroperiod. This is achieved by adding constraints (6.11) to the GLSPMS model.

yljs = yljv

rs

∀l, j, s, v with v = (s − 1) modulo (V − 1) + 1 (6.11)

For instance, if the setup sequence of Figure 6.12 represents a feasible solution of the RS

model, Figure 6.13 illustrates the ﬁrst GLSPMS solution.

1 4 2 1 3 4

s=0 s=1 … s=V-1

1 4 2 1 3 4 2 1 3 4 2 1 3

not allowed in reality. Hence, ﬁctitious overtime costs can be deﬁned as g = l (maxi,j {slij }+

maxj,s {hjs }/ minj {alj }). Moreover, g max = wS+1 /T is an appropriate upper bound.15

Since the Fix&Optimize yields very good results for the MLCLSP, it seems reasonable to

adapt it to the GLSPMS. Due to the two-level time structure of the GLSPMS, the decompo-

sitions of the Fix&Optimize cannot be directly transferred to the setup variables yljs . That

is why several adaptations are needed.

can only be included in a new setup sequence if the item being set up in the recent

solution is released in the current sequence as well. Otherwise no new sequence can be

generated.

15

This upper bound is suﬃcient for the used test instances, however it has to be adapted if another

problem is considered.

6.2. Heuristics for the GLSPMS 119

a resource. Thus, all products that can be processed on this line l are considered

simultaneously. To limit the size of subset R, the same rolling time window scheme

with the same number of macroperiods is applied to the GLSPMS as to the MLCLSP.

GLSPMS. For each pair of predecessor / successor product, all concerned production

lines and all microperiods are regarded.

Unfortunately, the cost function which is needed to determine the sequence of the product-

oriented decompositions cannot be easily adapted to the GLSPMS. The reason for this is

two-fold: ﬁrst of all, the GLSPMS considers sequence-dependent setups and the correspond-

ing costs cannot be clearly assigned to one product. And second, the used formulation itself

and the LP relaxation have crucial impact on the results of the cost function. This becomes

particularly clear through the observations by Helber and Sahling. They found out that a

Simple Plant Location based reformulation of the MLCLSP leads to a cost function that

ends up in a worse sequence of products, although this formulation produces better lower

bounds.

As the GLSPMS shows a totally diﬀerent model formulation and due to the other reason,

the natural sequence of the product index is used.

The best results for the MLCLSP can be obtained with the sequence “product-oriented”,

“resource-oriented” and ﬁnally “process-oriented” decomposition. This variant is also used

for the GLSPMS tailored version.

(VNDS+E)

As the basic concepts of VNDS and Exchange are equivalent, the idea now is to bring together

these two solution procedures.16 Therefore, similar to the approach of Helber and Sahling a

microperiod, a product, a bill-of-materials, and a production line decomposition are deﬁned.

All the subsets R and F resulting from a certain decomposition deﬁne the neighborhood

structures for the VNDS. Accordingly, the following neighborhood structures can be derived

(in the sequence above):

NσS These neighborhoods contain all setup sequences that can be found by releasing σ

microperiods.

16

A similar but more general approach is presented by Lazić et al. (2010) who use VNDS for binary mixed

integer programs.

120 Chapter 6. Heuristics for the improved GLSPMS

NιJ This group of neighborhoods consists of all solutions that can be reached by releasing

the current setup sequence and ι products. This is equivalent to allowing ι items to be

included in a new setup sequence.

NβB Two diﬀerent neighborhoods β = 1 and β = 2 similar to NιJ are deﬁned here. They do

not only regard products separately but their dependencies expressed by the product

structure. That means whole parts of this structure are chosen—from the ﬁrst till the

last level. Details of the implementation are given in the following.

NλL Finally, these neighborhoods refer to all possible solutions that can be found by releas-

ing λ production lines.

These decompositions should be controlled by the VNDS which is intended to increase the

solution space by ﬁxing a decreasing number of attributes. Following this idea, the neighbor-

hood structures presented above are systematically arranged to get subsets R with increasing

cardinality. To achieve this, we take advantage of the following relation: For problem in-

stances of the consumer goods industry, the sequence L < J < S holds in general. This

leads directly to L · J < L · S < J · S. Now a connection to the neighborhood structures

can be established, since the elements of this derived sequence are equivalent to the number

of released variables (e.g. releasing a single microperiod aﬀects L · J variables of yljs and so

on). That is why the sequence NσS , NιJ , NβB , NλL is applied in the following. Note that NβB

is ranked before NλL because at least for λ = L it follows that NβB ⊂ NLL . To adhere to the

VNDS principle only certain ranges should be allowed for σ, ι, and λ:

S

• The restriction σ · L · J ≤ L · S implies directly σ ≤ L·S

L·J

⇒σ≤ J

.

• From ι · L · S ≤ J · S it follows that ι ≤ L·S

J·S

⇒ ι ≤ LJ .

In the remainder of this chapter, k is an identiﬁer for each neighborhood structure and

describes indirectly the number of ﬁxed attributes. According to the random ﬁxation of

attributes, the new algorithm also ﬁxes/releases diﬀerent parts of a neighborhood at random,

e.g., for neighborhood N1S the “released” microperiod is randomly selected. However, it still

might be impossible to run the MIP-solver for the Exchange until the optimum is found.17

Hence, for a larger k, it is more likely that no improvement can be achieved. Accordingly, it

is important to test diﬀerent parts of a certain neighborhood structure and to execute several

runs to test it more intensively. Otherwise, the solution procedure might terminate too early.

This can be achieved by releasing diﬀerent variables. For this purpose, a random function

f (k) is implemented which returns the subset R according to the current neighborhood

structure k. Note that only one of the subsets is necessary as they complement each other.

17

This is at least true for NLL , otherwise the original problem should be solvable to optimality.

6.2. Heuristics for the GLSPMS 121

If f (k) is called sequentially with the same k, it returns diﬀerent sets of variables. This

behavior is equivalent to “sampling without replacement”.

Example To demonstrate the workﬂow of the new VNDS with Exchange (VNDS+E)

heuristic, the base divergent scenario of Section 4.4.1.2 with (J=)6 items, (L=)3 diﬀerent

lines, and 3 macroperiods with 4 microperiods each (S=12) will be examplarily solved in the

following. Since 12 6

= 2 and 63 = 2, the developed algorithm will use the neighborhoods

N1S , N2S , N1J , N2J , N1B , N2B , N1L , N2L , and N3L in the given sequence with k = 1, . . . , K = 9.

As next, the main procedure and the implementation of the particular function f (k) for

each neighborhood structure are sketched.

The main control procedure is illustrated in Algorithm 4. First, the algorithm is initialized

(lines 1–3). As mentioned before, several consecutive iterations with the same neighbor-

hood k are allowed. For that purpose, a neighborhood iteration counter i is introduced and

set to 0. Since the algorithm starts with the ﬁrst neighborhood, k is set to 1.

The next step is to retrieve an initial solution and the corresponding objective function

value. After the initialization, the procedure calls the StartExchange() function with the

ﬁrst structure. If a new better incumbent solution is obtained, the algorithm restarts with the

ﬁrst structure k = 1 and the neighborhood iteration counter is reset. Otherwise, counter i is

increased till the maximum I (external parameter) is reached. In that case, again the counter

i is reset and the procedure changes to the next neighborhood structure k + 1. These steps

are repeated until all K structures have been visited.

The StartExchange() function is shown in Algorithm 5. According to the parameter k,

this function retrieves the corresponding subset R. After that SolveSubMIP(sol, R) ﬁxes the

setup variables and starts the MIP-solver. If a new solution can be found, it will be returned.

Otherwise this function returns the old solution sol. Algorithms 6–11 brieﬂy describe how

the subset R (and thus F) is chosen.

Example The heuristic starts with the initialization. Therefore, the external parameter

initial solution (IS) must be given which determines the construction heuristic to be executed.

In the regarded example, the trivial solution (TS) heuristic is chosen (see Section 6.2.4).

Accordingly, GetInitSol obtains the trivial solution by the MIP-solver in 0.7 sec. For each

setup variable yljs the corresponding value is given in Table 6.1. For this solution where

product j = 1 is set up on line l = 1, j = 4 on l = 2, and j = 6 on l = 3 the objective

function value is 5300.0 (including penalty costs for purchasing).

According to Algorithm 4, in the ﬁrst while loop iteration, neighborhood N1S is used

which will be described next.

122 Chapter 6. Heuristics for the improved GLSPMS

1 i ← 0;

2 k ← 1;

3 bestSol ← GetInitSol();

4 while true do

5 newSol ← StartExchange(bestSol, k, i);

6 if newSol better than bestSol then

7 bestSol ← newSol;

8 i ← 0;

9 k ← 1;

10 else

11 i ← i + 1;

12 end

13 if i = I then

14 i ← 0;

15 clear the history of subsets belonging to neighborhood k;

16 k ← k + 1;

17 end

18 if k = K + 1 then

19 if runtime is stopping criterion and runtime < maxRuntime then

20 k ← 1;

21 else

22 break;

23 end

24 end

25 end

Algorithm 5: StartExchange

input: A feasible incumbent solution sol, neighborhood k, neighborhood iteration

counter i

output: if a solution was found, newSol is returned or sol otherwise

1 switch k do

2 case NσS

3 R ← GetMicroPeriodNeighborhood(σ,i);

4 case NιJ

5 R ← GetProductNeighborhood(ι,i);

6 case N1B

7 R ← GetUpwardsBomNeighborhood(i);

8 case N2B

9 R ← GetDownwardsBomNeighborhood(i);

10 case NλL

11 R ← GetLineNeighborhood(λ,i);

12 endsw

13 newSol ← SolveSubMIP(sol, R);

6.2. Heuristics for the GLSPMS 123

yljs s=0 1 2 3 4 5 6 7 8 9 10 11 12

l=1 j=1 1 1 1 1 1 1 1 1 1 1 1 1 1

j=2 0 0 0 0 0 0 0 0 0 0 0 0 0

j=3 0 0 0 0 0 0 0 0 0 0 0 0 0

j=4 0 0 0 0 0 0 0 0 0 0 0 0 0

l=2 j=2 1 1 1 1 1 1 1 1 1 1 1 1 1

j=4 0 0 0 0 0 0 0 0 0 0 0 0 0

l=3 j=5 0 0 0 0 0 0 0 0 0 0 0 0 0

j=6 1 1 1 1 1 1 1 1 1 1 1 1 1

parameter σ, a subset Si of all S microperiods with |Si | = σ is randomly generated in

neighborhood iteration i.

Algorithm 6: GetMicroperiodNeighborhood

input: Number of microperiods to release σ, neighborhood iteration counter i

output: subset R with all setup variables to be released

1 R ← ∅;

2 Si ← GetRandomSubset(S, σ, history of subsets belonging to this neighborhood );

3 add Si to the history;

4 forall the production lines l, products j, microperiods s ∈ Si do

5 add yljs to R

6 end

need to be diﬀerent (pairwise), otherwise a useless (multiple) iteration is executed. Therefore,

GetRandomSubset() requires the former generated subsets as a parameter. This function is

generalized as shown in Algorithm 7 and can be applied to the next neighborhood functions

as well. By initializing an auxiliary set with all elements and randomly removing elements

step-by-step, the subset with the necessary size is created.18 For each s ∈ Si and for all l, j

the setup variables yljs are added to R.

Example In the ﬁrst while loop iteration, a single microperiod is randomly chosen. As

shown in Table 6.2a, microperiod 12 is released (highlighted gray).

18

Note that GetRandomSubset() is only useful, if the binomial coeﬃcient nSize oSize is suﬃciently large.

Otherwise, it would be very likely that the while-loop will run several times. In this case, this problem

should only occur while determining subsets of the production lines (assuming that L < J < S). Then, it

might be advantageous to compute all possible subsets in a preprocessing step and to randomly choose one

of the possibilities.

124 Chapter 6. Heuristics for the improved GLSPMS

Algorithm 7: GetRandomSubSet

input: Original size of the set oSize, size of the new subset nSize, history of subsets H

output: random subset Q

1 while true do

2 Q ← {1, 2, . . . , oSize};

3 for 1 to (oSize − nSize) do

4 remove random element from Q;

5 end

6 if Q ∈/ H then

7 break;

8 end

9 end

j=2 0 0 0 0 0 0 0 0 0 0 0 0 0 j=2 0 0 0 0 0 0 0 0 0 0 0 0 0

j=3 0 0 0 0 0 0 0 0 0 0 0 0 0 j=3 0 0 0 0 0 0 0 0 0 0 0 0 1

j=4 0 0 0 0 0 0 0 0 0 0 0 0 0 j=4 0 0 0 0 0 0 0 0 0 0 0 0 0

j=4 0 0 0 0 0 0 0 0 0 0 0 0 0 j=4 0 0 0 0 0 0 0 0 0 0 0 0 1

j=6 1 1 1 1 1 1 1 1 1 1 1 1 1 j=6 1 1 1 1 1 1 1 1 1 1 1 1 1

0.8 sec.

6.2. Heuristics for the GLSPMS 125

Then, the MIP-solver is started with the resulting subproblem. An important question

which arises here (and especially for larger instances/neighborhoods) is how much runtime

should be spent on the optimization. Accordingly, the maximum runtime for each sub-MIP

(MR) represents another parameter of this new heuristic. Fortunately, the instance at hand

is easy to solve and thus after 0.1 sec. (0.8 in total) the solution given in Table 6.2b is found.

The changed values are highlighted with white font color.

In the second loop iteration, again the neighborhood N1S is used, because a new solution

was found (see line 9 in Algorithm 4). But now, microperiod 12 is blocked as it belongs to the

“history” and microperiod 10 is randomly chosen. However, the new subproblem not allows

ﬁnding a better solution. The next question is, how many further single microperiods should

be tested. Hence, the maximum number of neighbors within a single neighborhood (MN) is

a further parameter for VNDS+E which corresponds to I in Algorithm 4 (line 13). For the

current example, three diﬀerent neighbors (I = 3) of neighborhood k must be unsuccessfully

tested sequently, before the neighborhood k + 1 is used.

Afterwards, the product neighborhood (k = 3) is used in loop iteration 12 for the ﬁrst time.

a set of products Ji with cardinality of ι is generated. After that for all j ∈ Ji , l, and s the

variables are added to R. However, this is not suﬃcient yet to ﬁnd a new setup sequence.

Due to constraints (4.13*), for each line and microperiod the setup variables of at least two

products (the current set up and another one) need to be released to allow a new setup state.

Since the product-based neighborhood aims at including a random chosen product in the

setup sequence, all yljs which are 1 in the current solution are added to R additionally.

Algorithm 8: GetProductNeighborhood

input: Number of products to release ι, neighborhood iteration counter i

output: subset R with all setup variables to be released

1 R ← ∅;

2 Ji ← GetRandomSubset(J, ι, history of subsets belonging to this neighborhood );

3 add Ji to the history;

4 forall the production lines l, products j ∈ Ji , microperiods s do

5 add yljs to R

6 end

7 forall the yljs which are set to 1 in the current solution do

8 add yljs to R

9 end

126 Chapter 6. Heuristics for the improved GLSPMS

Example As mentioned above, in loop iteration 12, the product neighborhood N1J is ap-

plied for the ﬁrst time. The current solution is shown in Table 6.3a.

j=2 0 0 0 0 0 0 1 0 0 0 0 0 0 j=2 0 0 0 0 0 0 1 0 0 0 0 0 0

j=3 0 1 0 0 0 0 0 0 0 0 0 0 1 j=3 0 1 0 0 0 0 0 0 0 0 1 1 1

j=4 0 0 1 0 0 1 0 0 0 0 1 1 0 j=4 0 0 1 0 0 1 0 0 0 0 0 0 0

j=4 0 1 1 1 1 1 1 1 1 1 1 1 1 j=4 0 1 1 1 1 1 1 1 1 1 1 1 1

j=6 1 1 1 0 0 1 0 0 0 0 1 1 1 j=6 1 1 1 0 0 1 0 0 0 0 1 1 1

4.0 sec.

In the current iteration, product 3 is chosen at random. But as already mentioned before,

it is not suﬃcient to release only the variables belonging to product 3. In order to allow new

setup states, the old sequences have to be released as well (see also Table 6.3a). The recent

solution is given in Table 6.3b. In microperiods 10 and 11 the setup state of production line

l = 1 has changed from product 4 to 3. After 16 further loop iterations with one improvement

by the microperiod and one by the product decomposition, in iteration 29, the BOM-based

neighborhood comes into play.

Similar to the process-oriented decomposition of the Fix&Optimize heuristic (cf. Section 6.2.6),

the following two BOM-based neighborhoods consider the product structure. However, they

are not restricted to two products. Since the GLSPMS allows diﬀerent types of bills-of-

materials, an “upwards” (β = 1) and a “downwards” (β = 2) neighborhood structure is

implemented. Algorithm 9 illustrates the “upwards” approach. At ﬁrst, a pre-item j is

randomly picked. For this product all direct and indirect successors are recursively added

to R. Finally, all yljs which are 1 in the current solution have to be included again for the

same reason as for GetProductNeighborhood().

As Algorithm 10 shows the function GetDownwardsBomNeighborhood() works similarly.

Now, a ﬁnal item is selected by chance and then recursively all its predecessors as well.

Example In loop iteration 29, the BOM-based neighborhood N1B is used for the ﬁrst time.

Therefore, a pre-product has to be selected ﬁrst. Then, all corresponding successor items

are chosen. In the regarded case products 5 (see Figure 6.14a) and products 1 and 2 are

picked (Figure 6.14b).

6.2. Heuristics for the GLSPMS 127

Algorithm 9: GetUpwardsBomNeighborhood

input: neighborhood iteration counter i

output: subset R with all setup variables to be released

1 R ← ∅;

2 u ← GetRandomProductOnLastLevel(history of subsets/random numbers belonging to

this neighborhood );

3 add u to the history;

4 for j = u and j = direct and indirect successors of u, production lines l and

microperiods s do

5 add yljs to R

6 end

7 forall the yljs which are set to 1 in the current solution do

8 add yljs to R

9 end

input: neighborhood iteration counter i

output: subset R with all setup variables to be released

1 R ← ∅;

2 u ← GetRandomProductOnFirstLevel(history of subsets/random numbers belonging

to this neighborhood );

3 add u to the history;

4 for j = u and j = direct and indirect predecessors of u, production lines l and

microperiods s do

5 add yljs to R

6 end

7 forall the yljs which are set to 1 in the current solution do

8 add yljs to R

9 end

1 2 3 4 1 2 3 4

5 6 5 6

(a) Pre-item selection. (b) Selection of all successor products.

128 Chapter 6. Heuristics for the improved GLSPMS

Starting with the setup sequence from Table 6.4a and the corresponding released vari-

ables, a new solution (Table 6.4b) is obtained. The new objective function value is 226.625

and 7.2 sec. have passed in total.

j=2 0 0 0 0 0 0 1 0 0 0 0 0 0 j=2 0 0 0 0 0 0 0 0 1 1 0 0 0

j=3 0 1 0 0 0 0 0 0 0 0 1 1 1 j=3 0 1 0 0 0 0 0 0 0 0 0 1 1

j=4 0 0 1 0 0 1 0 0 0 0 0 0 0 j=4 0 0 1 0 0 0 0 0 0 0 0 0 0

j=4 0 1 1 1 1 1 1 1 1 1 1 1 1 j=4 0 1 1 1 1 1 1 1 1 1 1 1 1

j=6 1 1 1 0 0 1 0 0 0 0 1 1 1 j=6 1 1 1 0 0 0 0 0 0 0 0 1 1

runtime 7.2 sec.

hood N2B is used for the ﬁrst time. Accordingly, the current solution in Table 6.5 is the same

as in Table 6.4b. For the new neighborhood the ﬁnished product 4 is initially chosen (see

Figure 6.15a) and afterwards the only predecessor product 6 is selected (see Figure 6.15b).

1 2 3 4 1 2 3 4

5 6 5 6

(a) Finished product selection. (b) Selection of all predecessors.

yljs s=0 1 2 3 4 5 6 7 8 9 10 11 12

l=1 j=1 1 0 0 1 1 1 1 1 0 0 1 0 0

j=2 0 0 0 0 0 0 0 0 1 1 0 0 0

j=3 0 1 0 0 0 0 0 0 0 0 0 1 1

j=4 0 0 1 0 0 0 0 0 0 0 0 0 0

l=2 j=2 1 0 0 0 0 0 0 0 0 0 0 0 0

j=4 0 1 1 1 1 1 1 1 1 1 1 1 1

l=3 j=5 0 0 0 1 1 1 1 1 1 1 1 0 0

j=6 1 1 1 0 0 0 0 0 0 0 0 1 1

As becomes apparent by Table 6.5 no improved solution has been found. Since the BOM-

based neighborhoods cannot ﬁnd any new solution, three iterations later the ﬁrst line-based

neighborhood structure N1L is used (see Table 6.6a).

6.2. Heuristics for the GLSPMS 129

The last function GetLineNeighborhood() proceeds in the same manner (see Algorithm 11).

A subset of lines Li is build with |Li | = λ. For all l ∈ Li , j, and s the setup state variables

are included in R.

input: Number of lines to release λ, neighborhood iteration counter i

output: subset R with all setup variables to be released

1 R ← ∅;

2 Li ← GetRandomSubset(L, λ, history of subsets belonging to this neighborhood );

3 add Li to the history;

4 forall the production lines l ∈ Li , products j, microperiods s do

5 add yljs to R

6 end

Example In iteration 41, once more a better solution can be found by releasing randomly

production line l = 1. After less than 10 sec. in total, an objective value of 223.25 is reached.

This solution which is illustrated in Table 6.6b is at the same time the optimal one. Hence,

j=2 0 0 0 0 0 0 0 0 1 1 0 0 0 j=2 0 0 0 0 0 1 1 0 0 0 1 0 0

j=3 0 1 0 0 0 0 0 0 0 0 0 1 1 j=3 0 1 0 0 0 0 0 0 0 0 0 0 1

j=4 0 0 1 0 0 0 0 0 0 0 0 0 0 j=4 0 0 1 0 0 0 0 0 0 0 0 1 0

j=4 0 1 1 1 1 1 1 1 1 1 1 1 1 j=4 0 1 1 1 1 1 1 1 1 1 1 1 1

j=6 1 1 1 0 0 0 0 0 0 0 0 1 1 j=6 1 1 1 0 0 0 0 0 0 0 0 1 1

9.5 sec.

it is not surprising that the new heuristic probes all (K =) 9 neighborhoods again, before it

terminates in iteration 68(= 41 + 9 · 3). To ﬁnd this solution, the new heuristic has spent

13.6 sec. in total whereas pure Xpress needed 25.4 sec.

In the following section, the new heuristics will be examined. Note that the terms Trun-

cated MIP (TM), LP&Fix (LF), Relax&Fix (RF), Fix&Optimize (FO) always refer to the

GLSPMS speciﬁc implementations shown in this section—without exception.

130 Chapter 6. Heuristics for the improved GLSPMS

In a ﬁrst step, the solution quality of the simple construction heuristics Truncated MIP (TM),

the LP&Fix (LF) and the Relax&Fix (RF) will be evaluated. Therefore, these heuristics

will be tested for each problem instance used in Chapter 4 and the two best formulations

(F+M+K and S+M+K) presented Chapter 5. Unfortunately, it will become apparent that

the results are dissatisfying. This is where the tailored Fix&Optimize and the VNDS+E

heuristic come into play. As these heuristics seem quite promising, Section 6.3.2 will compare

them to Truncated MIP using bigger test instances which are closer to real-world problem

sizes. Since VNDS+E will provide good results, an actual real-world problem will then be

tackled in Section 6.3.3.

The following Table 6.7 contains all results for the three construction heuristics with a time

limit of 300 sec. In the columns denoted by “GP” the averaged percentage diﬀerence between

M+K

TM LF RF

GP AR GP AR GP AR

SER F 1 13 21 3 1 6

S 1 16 1 109 1 14

DIV F 12 14 519 2 26 19

S 6 19 36 108 19 16

GEN F 8 15 1026 1 29 6

S 20 17 20 109 35 9

Table 6.7: Gap (GP) and average runtime (AR) of Truncated MIP (TM), LP&Fix (LF), and

Relax&Fix (RF) for the extended formulations “ﬂow conservation” (F) and “Simple Plant

Location” (S)—both extended by stock inequalities (K) as well as big-M-inequalities (M).

the best feasible solution found within 300 sec. and the overall best lower bound found by

any formulation within 3600 sec. (results of Section 5.5.1) is given.19 Moreover, the average

runtime (AR) is listed as well (sec.).

Once more, it can be seen that the serial instances are easier to solve. TM as well as RF

are both able to solve most of the instances to optimality (all GPs ≤ 1%)—independently

of the used model formulation. Only LF shows some weaknesses when using the ﬂow for-

mulation (GP=21%). But with the S formulation, LF produces competitive results as well,

even though the average runtime is signiﬁcantly higher than the one of TM and RF. Note

that RF also runs considerably faster than TM.

19

Thus, diﬀerences stem from the formulations’ potentials to ﬁnd feasible solutions.

6.3. Computational tests 131

For the divergent instances all three heuristics perform diﬀerently. TM performs best

(GP between 6 and 12%) and in particular the S formulation appears to be superior as in

the chapter before. In contrast, RF is not that good anymore. The gaps (GP) lie between

19 and 26%, but with an average runtime between 16 and 19 secs. as compared to the 14–19

secs. of TM. Finally, the LF produces the worst results. The lower gap is 36%, the higher

even 519%. Again a systematic diﬀerence related to the formulations can be observed. It

seems that for the F formulation more variables can be ﬁxated. On the one hand, this results

in a lower runtime. But on the other hand, the GP values are quite high.

This contrast between F and S for LF becomes dramatically obvious for the instances

with a general product structure. Here, LF produces a gap of 1026% in 1 sec. compared to a

gap of 20% after an average runtime of 109 secs. The results of TM are again the best with

gaps between 8 and 20%. RF performs bad (GP=35%).

In sum, TM performs best in terms of solution quality, in particular the S formulation

for the DIV instances and the F formulation for the GEN instances. The main reason is

probably that these formulations beneﬁt from their similar behavior in Table 5.4. LF is

fast, but of very bad quality when applied in a ﬂow formulation. For the S formulation,

LF behaves substantially better with respect to solution quality, however, at costs of higher

computation times. RF is also worse than TM in terms of solution quality. However, the

computational eﬀort for RF is signiﬁcantly lower.

Note that increasing the permitted computation time from 300 sec. to 3600 sec. does not

really change the picture. For TM the GP stays almost the same when the SER instances

are considered. The reductions (GP 300sec − GP 3600sec ) range between 1 and 5 for the DIV

instances with F+M+K being the only exception since a reduction by 11 is possible. The

GEN instances are reduced by 1–6 as compared to Table 6.7. The results of LF and RF

do not change substantially when longer running times are allowed. Within 300 secs., all

remaining MIPs of the (bad performing) ﬂow formulations of LF and around 69% of the

other formulations have already been solved to optimality. Thus, there is not much space

left for improvement. For RF more than 99% of the “sub-MIPs” of RF can be solved to

optimality within 300 secs.

As already shown at the end of Chapter 5, most of the improvements are made by the

TM in the ﬁrst minutes. Accordingly, this is very promising that more sophisticated MIP-

based heuristics like the tailored Fix&Optimize (FO) and the new VNDS+E heuristic allow

for solving the GLSPMS. Moreover, the example in Section 6.2.7 already indicated that

VNDS+E has the potential to outperform pure standard MIP-solvers.

Due to the dissatisfying results of the construction heuristics for small instances, the focus

is shifted to the improvement heuristics. Hence, in this section the tailored Fix&Optimize

132 Chapter 6. Heuristics for the improved GLSPMS

as well as the VNDS+E will be compared to each other. The following settings are used for

comparison.

Section 5.5.1 showed that the majority of the bigger instances cannot be solved to optimality.

From these “unﬁnished” instances 30 are selected now—10 of each product structure. But

since the chosen instances are still away from real-world problem sizes, they are enlarged by

copying the number of periods (demand and capacity vector). As a result, the tests cover a

range of 6-12 products, 3-9 lines, and 8-16 macroperiods with 24-72 microperiods.

Since the network ﬂow formulation in combination with the tighter setup forcing con-

straints and the stock inequalities (F+M+K) has proven to be advantageous, this formula-

tion is used in the following.20 Correspondingly, not have the setup state variables yljs to be

released or ﬁxed but the changeover variables zlijs .

Before the tailored Fix&Optimize and VNDS+E can be compared to each other, the

parameters of the latter have to be tuned.

In Section 6.2.7, the following three diﬀerent parameters were already introduced:

Initial solution (IS) Due to the bad performance of RF and LF, the new heuristic will only

use the trivial solution (TS) and the repeating setup sequence (RS) as construction

heuristics. Since they are able to ﬁnd an initial solution very quickly, more time can

be spent on improving this solution by VNDS+E.21

be very diﬃcult to ﬁnd a new solution with limited runtime and due to the high num-

ber of neighbors that constitute a neighborhood, several ﬁxations should be tested.

Therefore parameter MF is introduced which describes the number of neighbors ex-

plored. The value of MF is equivalent to the date I in Algorithm 2. Note that the

value for MF might be automatically decreased for a certain neighborhood structure

if the number of possible subsets is less than MF.

Maximum runtime for each MIP (MR) The more variables are released the more com-

putational eﬀort is necessary. Hence, the MIP run might be aborted before proving

20

Note that Seeanner et al. (2013) use a diﬀerent formulation based on the GLSPMS by Meyr (2004).

This leads to other results because the improved formulation is closer to reality and the new synchronization

constraints cause higher objective function values. Moreover, the improved GLSPMS of Section 4 needs

more random access memory (RAM) and seems harder to solve.

21

Note that for all artiﬁcial problems a ﬁrst incumbent solution can be found by the RS Model in less

than 10 sec.

6.3. Computational tests 133

optimality. A question that naturally arises is how much time should be spent on each

run of the Exchange. Therefore, two diﬀerent values (20 sec., 30 sec.) will be checked.

Another question is if the maximum gap should be considered as a criterion for the

termination as well. In the regarded case, the stopping criterion is only based on the

runtime because of the bad quality of lower bounds provided by the GLSPMS in gen-

eral. Depending on the number of variables that are released the gap might strongly

vary (typically between 10% and 50%). Consequently, it is hard to determine a “good“

bound on the gap. This, indeed, is really crucial because otherwise too much time is

spent for a single subproblem without improving the solution.

Since at most 30 sec. should be spent for a single sub-MIP, a fourth parameter is used which

might have an impact on the performance of the solver:

at the beginning of the Branch&Bound procedure to ﬁnd a ﬁrst high quality incumbent

solution. This step usually reduces the overall runtime signiﬁcantly. Since less runtime

is provided for each sub problem, those heuristics represent a potential drawback as

they can become very time-consuming. For this reason, the impact of this start heuris-

tics will be analyzed. Therefore, the boolean parameter SH allows to include the use

of the start heuristic.

To ﬁnd the best parameter settings for the VNDS+E heuristic, several scenarios are

deﬁned in the following. Since the number of possible combinations of the parameter settings

for the new heuristic is large, ﬁve diﬀerent scenarios are created. The ﬁrst one represents a

base scenario. All other scenarios only diﬀer in a single parameter value. Table 6.8 gives an

overview of the scenarios with the corresponding parameter values.

Scenario IS MR MN SH

1 TS 20 3 ON

2 RS 20 3 ON

3 TS 30 3 ON

4 TS 20 5 ON

5 TS 20 3 OFF

But before the scenarios will be compared to each other, it is important to assess the

eﬀectiveness of the sequence of the BOM-based neighborhoods ﬁrst. Therefore, only the

base scenario is considered and runtime issues are neglected.

As there is no “logical” sequence for the BOM-based neighborhood structure (concerning the

size of the neighborhood), in a ﬁrst step three diﬀerent kinds of sequences are tested:

134 Chapter 6. Heuristics for the improved GLSPMS

3. Each time the search was successful and the algorithm starts again with the ﬁrst

structure (k=1, on line 9 of Algorithm 4), the sequence is randomly chosen (RND).

Since both BOM strategies are identical for the serial problem instances, only the other two

product structures are regarded. The corresponding results are shown in Table 6.9. Each

problem (Prob) instance has a unique number (ID). The random strategy (RND) serves

as a benchmark. Therefore, the average objective values of 30 test runs (Avg) as well as

the corresponding coeﬃcient of variation (CV) are given in the second and third column,

respectively. For both other combinations (N1B , N2B and N2B , N1B ) the relative percentage gap

between the average objective value and the reference (AG) and the coeﬃcient of variation

(CV) are listed.

Prob RND N1B , N2B N2B , N1B Prob RND N1B , N2B N2B , N1B

ID Avg CV AG CV AG CV ID Avg CV AG CV AG CV

1 935 5 -0.9 3 -0.4 4 11 788 2 -0.2 2 -0.4 2

2 1820 5 -0.4 5 -0.2 4 12 997 2 1.0 2 0.3 2

3 913 3 0.2 4 -0.2 5 13 1582 2 0.4 2 0.8 2

4 1733 4 1.4 4 1.0 4 14 773 2 0.5 5 0.3 4

5 1668 5 1.0 10 0.2 6 15 1002 4 -1.2 4 -0.7 6

6 479 2 0.1 1 0.2 1 16 1542 2 -0.4 3 0.3 3

7 879 3 -0.8 3 0.8 4 17 692 2 0.1 2 0.1 2

8 775 0 0.0 0 0.1 0 18 1086 3 -0.4 2 0.0 2

9 1322 2 -0.2 2 -0.1 2 19 654 2 -0.1 2 -0.2 2

10 1209 4 -2.4 3 -1.2 4 20 1015 2 -0.2 2 -0.2 2

Avg 3 -0.2 4 0.0 3 Avg 2 0.0 3 0.0 3

structures.

Regarding to the average values in the last rows, no real diﬀerence between the structures

can be found. In fact, the CV is almost the same per product structure and the AG equals or

is close to 0.0%. That is why the hypothesis that a diﬀerent sequence gives diﬀerent results

has to be rejected as the diﬀerence is not signiﬁcantly statistically proven by the analysis of

variance (ANOVA, p > 0.1). Although no real link between the product structures and the

neighborhood structures can be derived from this, the BOM-based neighborhoods are useful

parts of the solution procedure anyway.

As illustrated in Tables 6.10a and 6.10b, both neighborhood structures help to ﬁnd new

solutions.

Columns AN1 and AN2 show how often a new solution is found on average by N1B and

N2B ,respectively. For example, on average structure N1B ﬁnds a new solution 7.5 times for

problem 1 when this structure is arranged before N2B . Interestingly, the neighborhood which

6.3. Computational tests 135

Prob RND N1B , N2B N2B , N1B Prob RND N1B , N2B N2B , N1B

ID AN1 AN2 AN1 AN2 AN1 AN2 ID AN1 AN2 AN1 AN2 AN1 AN2

1 4.7 7.6 7.5 2.7 5.1 8.6 11 3.2 4.5 4.9 3.9 3.7 5.3

2 0.9 3.7 0.7 2.9 0.7 3.2 12 2.5 3.4 2.4 1.6 2.3 2.6

3 5.2 12.3 10.0 2.3 6.2 13.9 13 1.6 2.4 2.1 1.0 2.1 3.5

4 0.3 4.7 1.1 3.8 0.6 4.3 14 9.8 19.5 17.0 8.0 11.5 20.8

5 3.1 8.4 6.9 4.0 3.2 9.3 15 9.5 20.1 18.6 7.3 8.6 17.1

6 2.2 3.1 2.3 1.9 1.7 2.7 16 3.8 10.6 7.9 7.0 3.6 9.7

7 2.2 3.2 2.2 2.2 1.4 2.5 17 2.8 2.5 2.0 2.1 2.3 2.3

8 0.6 1.1 0.3 0.3 0.3 0.5 18 1.7 2.4 2.0 1.1 1.7 2.5

9 3.1 3.4 3.5 2.6 3.2 3.7 19 8.1 13.8 13.3 9.5 8.1 15.6

10 4.8 9.3 7.9 2.1 5.1 8.7 20 6.3 9.7 7.3 5.3 6.5 9.5

Avg 2.7 5.7 4.2 2.5 2.7 5.7 Avg 4.9 8.9 7.7 4.7 5.0 8.9

Table 6.10: Number of times a new solution has been found by the neighborhood structures

on average.

comes ﬁrst ﬁnds almost always more new solutions. Consequently, the random sequence is

included in the procedure.

Results of the diﬀerent parameter scenarios As a next step, the diﬀerent scenarios

presented above are tested. In Table 6.11, the results of all ﬁve scenarios for the 30 problem

instances are shown. The problem ID is given in the ﬁrst column. In columns 2–7, problem

details like product structure (PS), the number of production stages (E), the number of

lines (L), the number of products (J), the number of macroperiods (T) and the number of

microperiods (S) are given. The number of microperiods is chosen in a way that the problem

is not solvable to optimality within 1 hour of runtime and a “repeating setup sequence”

can be found for it (see Section 6.2.5). In order to make the results comparable, for each

problem the minimum runtime (secs.) over all test instances and scenarios is determined

(column RT). The objective function values of all of these test instances are then measured

at this point in time. Column Min shows the best objective value found by any scenario.

Since every problem instance has been solved 30 times for every scenario parameter, the

percentage gap between the best value (Min) over all scenarios and the minimum (MG) or

the average value (AG) of each scenario can be calculated. Additionally, the coeﬃcient of

variation (CV) is given to show the robustness of the solution procedure.

For instance, the fastest test run for problem 7 terminates after 387 sec. In that time

span the best solution found by one of the 150 runs has an objective value of 843. The best

solution for scenario 1 is 0.2% worse than 843 and on average scenario 1 is 6.2% away from

this overall best. When considering all divergent problems, scenario 5 performs best with

only a 0.3% gap between its best results and the overall best results. Moreover, scenario 5

shows the lowest average gap (AG) for these instances with 6.1% as well. However, all other

scenarios produce a lower coeﬃcient of variation than scenario 5.

Nevertheless, in sum scenario 5 performs best for all kinds of problems as the total average

136

ID PS E L J T S RT Min MG AG CV MG AG CV MG AG CV MG AG CV MG AG CV

1 DIV 2 3 12 8 40 141 881 0.2 8.0 5.0 0.5 7.5 2.4 1.5 7.0 3.1 1.5 10.9 5.4 0.0 6.5 3.9

2 DIV 2 3 12 12 60 423 1716 1.9 11.6 5.2 1.3 10.8 5.0 0.1 11.1 5.0 1.0 8.6 4.0 0.0 9.3 5.3

3 DIV 2 6 12 8 24 299 861 2.7 6.6 2.5 4.0 9.9 5.4 2.5 8.0 4.1 0.8 6.2 5.1 0.0 5.5 3.3

4 DIV 2 6 12 12 36 652 1642 0.0 7.6 3.9 1.2 9.1 4.2 0.2 8.6 4.2 0.3 8.1 5.0 0.3 6.2 5.9

5 DIV 2 9 12 12 24 634 1501 3.5 13.2 6.0 4.3 17.9 8.5 0.7 13.0 6.9 0.0 11.3 5.9 0.4 11.6 7.4

6 DIV 2 6 6 8 24 284 472 0.4 3.5 3.5 0.2 3.2 2.4 0.0 3.3 2.5 0.6 2.1 1.6 0.0 2.4 1.7

7 DIV 2 6 6 12 24 387 843 0.2 6.2 3.9 0.0 6.7 4.3 0.4 7.4 4.0 1.0 5.5 2.9 0.7 6.3 4.4

8 DIV 2 9 6 12 24 610 775 0.0 1.5 3.8 0.0 0.2 0.4 0.0 1.1 1.3 0.0 0.6 0.9 0.0 1.7 3.0

9 DIV 2 3 9 12 36 150 1281 0.5 4.1 2.5 2.7 6.6 2.9 0.0 3.8 2.1 1.7 4.3 1.4 1.3 4.7 2.3

10 DIV 2 6 9 12 24 326 1128 2.2 7.8 4.0 0.0 4.7 4.0 0.8 6.4 2.9 0.6 8.1 4.9 0.3 7.2 4.2

Avg 1.2 7.0 4.0 1.4 7.7 3.9 0.6 7.0 3.6 0.8 6.6 3.7 0.3 6.1 4.1

11 GEN 2 3 11 8 48 155 755 0.0 6.4 3.2 0.4 5.6 3.3 3.4 7.3 2.8 2.4 8.7 3.0 2.5 6.2 2.7

12 GEN 2 3 11 10 60 199 977 1.0 4.8 2.2 1.2 5.5 2.4 1.9 5.1 2.1 2.1 5.9 2.6 0.0 4.4 2.0

13 GEN 2 3 11 12 72 267 1572 0.0 2.7 2.2 1.3 4.9 2.1 1.3 5.1 2.6 0.4 4.4 1.6 0.2 3.3 2.0

14 GEN 2 6 11 8 24 266 735 1.1 7.1 3.1 2.5 7.9 3.6 1.4 7.0 2.8 0.0 10.0 4.7 0.8 6.3 3.0

15 GEN 2 6 11 10 30 263 958 1.5 9.0 5.0 2.8 8.7 3.4 0.2 9.1 6.2 0.0 9.6 4.2 0.7 8.4 6.0

16 GEN 2 6 11 12 36 555 1506 0.9 5.5 3.0 1.7 6.7 2.3 1.3 5.9 2.7 0.9 5.1 2.8 0.0 5.0 3.0

17 GEN 2 3 8 10 40 168 671 1.0 5.2 2.6 1.2 4.8 2.0 2.2 5.8 2.2 2.1 5.8 2.2 0.0 4.6 2.5

18 GEN 2 3 8 12 48 332 1044 1.2 5.6 2.5 2.7 5.3 1.9 1.3 5.3 2.1 0.3 4.8 2.5 0.0 5.0 2.0

19 GEN 2 6 8 10 20 226 621 0.0 6.1 2.9 1.3 4.6 1.8 1.2 6.0 2.3 2.7 6.4 2.2 1.3 4.7 2.7

20 GEN 2 6 8 12 24 353 974 2.6 6.2 2.7 1.9 4.5 1.6 0.0 6.0 2.5 2.2 5.5 2.2 1.4 5.6 2.2

Avg 0.9 5.8 2.9 1.7 5.9 2.5 1.4 6.3 2.8 1.3 6.6 2.8 0.7 5.4 2.8

21 SER 3 4 12 8 32 187 6902 0.0 2.3 2.2 1.3 3.0 1.1 0.1 2.3 2.1 0.1 1.9 0.9 0.4 1.8 1.1

22 SER 3 4 12 12 48 293 9313 1.0 2.3 0.7 0.0 0.2 0.2 0.3 2.2 0.9 0.3 1.8 0.6 0.4 1.9 0.7

23 SER 3 4 12 16 64 505 13916 0.0 1.9 1.3 1.7 2.3 0.3 0.3 3.9 10.1 0.6 1.9 1.7 0.6 2.0 1.9

24 SER 3 8 12 8 16 359 6762 0.6 1.6 0.6 1.7 2.8 0.3 0.0 1.2 0.6 0.3 1.2 0.5 0.2 1.5 0.7

25 SER 3 8 12 12 24 459 9117 0.2 1.6 0.8 1.4 2.6 0.3 0.0 1.4 0.9 0.6 1.5 0.5 0.6 1.4 0.5

26 SER 3 8 12 16 32 295 13760 0.6 3.8 3.5 1.1 1.5 0.2 0.6 5.1 3.0 0.0 4.4 7.2 0.5 9.2 12.6

27 SER 3 4 6 12 48 341 4500 0.0 0.9 0.7 0.0 0.7 0.4 0.0 0.7 0.5 0.0 0.8 0.5 0.0 0.8 0.6

28 SER 3 4 6 16 64 387 6702 0.0 0.8 0.4 0.0 0.9 0.5 0.0 0.9 0.4 0.0 0.8 0.4 0.0 0.9 0.4

29 SER 3 4 9 12 36 153 7982 0.1 1.2 0.7 0.5 0.9 0.2 0.0 0.9 0.5 0.2 1.0 0.6 0.0 0.9 0.6

30 SER 3 4 9 16 48 358 12451 0.4 1.3 0.5 0.1 1.1 0.6 0.2 1.2 0.6 0.0 1.0 0.7 0.2 1.1 0.5

Avg 0.3 1.8 1.1 0.8 1.6 0.4 0.2 2.0 2.0 0.2 1.6 1.4 0.3 2.2 2.0

Total 0.8 4.9 2.7 1.3 5.0 2.3 0.7 5.1 2.8 0.8 4.9 2.6 0.4 4.6 3.0

Chapter 6. Heuristics for the improved GLSPMS

6.3. Computational tests 137

MG- and AG-values are the lowest (0.4 and 4.6%, respectively). One reason for this might

be that the starting heuristic is not able to ﬁnd a better upper bound. Thus, useful runtime

is “wasted” and the Branch&Bound procedure is started too late.22

Besides, scenario 2 gives the stablest results in total. However, comparing the results

of all scenarios, the new heuristic seems to be quite a robust procedure as the diﬀerences

between the scenarios are very small in terms of the best (MG) and the average solutions

found (AG).

To illustrate the diﬀerent behavior of scenario 2 and 5, the development of the objective

values of problem 24 is shown in Figure 6.16. For each point in time the best, the worst

and the mean value is determined for the heuristic. Obviously, the initial solution provided

by the RS model has a signiﬁcant lower objective function value (Figure 6.16a). Moreover,

the heuristic converges quite fast with that solution and after 155 sec. no real improvement

can be achieved. In contrast, scenario 5 starts with a signiﬁcantly worse initial solution

(Figure 6.16b). Thus, it takes a while until all 30 runs produce good results. However, after

359 sec., scenario 5 is slightly better than scenario 2—on average and in total.

The aim is now to compare VNDS+E (with its best parameter setting scenario 5), tailored

Fix&Optimize (FO), and Truncated MIP (TM) which is used as a benchmark. The results

are summarized in Table 6.12. Once more the average percentage gaps of scenario 5 (SC5)

are given (AG). Column TM shows the gap (Gap) of Truncated MIP to the best value of

all scenarios (column Min in Table 6.11) after the same minimum runtime described before

(column RT in Table 6.11). For the tailored Fix&Optimize, the gap (Gap) to this overall

best value is given as well. However, since FO often terminated earlier, the runtime (RT) is

listed additionally. For problem 7 once again it means that the execution of the MIP-solver

is stopped after 387 sec. with a gap of 34.1% to 843. What can be realized is that the new

VNDS+E beats Truncated MIP almost always by far. However, the results of problem 8

and the serial problems are comparable, as they seem easier to solve.23 Unfortunately, the

tailored Fix&Optimize performs quite bad except for problem 8. That is why FO is dropped

from further consideration.

6.3.2.4 Results for VNDS+E and Truncated MIP after 1 hour of runtime

In order to see the convergence behavior of the new heuristic, Truncated MIP as well as

VNDS+E are now run for one hour. For that purpose three problems with the highest

CV-value are chosen out of each group of product structures (see bold values in Table 6.11).

22

Note that this is the major diﬀerence to the results by Seeanner et al. (2013).

23

In order to get an impression of the heuristics’ behavior, three diﬀerent instances are graphically repre-

sented in the Appendix B (see Figures B.1, B.2, and B.3).

138 Chapter 6. Heuristics for the improved GLSPMS

!"#$

!"#$

(a) Parameter scenario 2.

!"#$

!"#$

(b) Parameter scenario 5.

6.3. Computational tests 139

ID AG Gap Gap RT ID AG Gap Gap RT ID AG Gap Gap RT

1 6.5 299.7 1255.2 27 11 6.2 32.8 1313.4 11 21 1.8 12.6 79.8 7

2 9.3 317.9 1172.2 158 12 4.4 37.2 1369.4 18 22 1.9 11.7 79.9 11

3 5.5 181.4 1284.8 26 13 3.3 70.0 1256.3 23 23 2.0 17.2 78.3 15

4 6.2 314.7 1226.4 114 14 6.3 40.7 1371.7 28 24 1.5 9.5 83.5 8

5 11.6 236.6 1349.5 25 15 8.4 41.2 1418.3 24 25 1.4 9.0 83.7 14

6 2.4 13.9 22.1 173 16 5.0 82.3 1329.8 41 26 9.2 38.1 80.3 23

7 6.3 34.1 32.3 51 17 4.6 20.8 1663.2 7 27 0.8 2.4 52.9 5

8 1.7 1.1 0.0 69 18 5.0 28.2 1465.3 10 28 0.9 2.3 54.9 7

9 4.7 83.0 861.0 23 19 4.7 28.9 1804.5 6 29 0.9 1.6 51.4 7

10 7.2 26.8 987.7 37 20 5.6 38.2 1577.2 13 30 1.1 3.1 47.6 11

Table 6.12: Comparison of scenario 5 (SC5), Truncated MIP (TM), and tailored

Fix&Optimize (FO).

To ensure a runtime of one hour, it might become necessary to “restart” the new heuristic

if all neighborhoods have been visited (cf. lines 18–24 in Algorithm 4). The results of the

procedure can be seen in Table 6.13.

ID Gap MG AG CV ID Gap MG AG CV ID Gap MG AG CV

2 77.8 -2.9 2.3 2.7 14 18.1 -2.5 2.9 3.2 21 7.8 -0.1 1.0 0.7

4 121.2 -2.6 2.6 3.3 15 21.6 -4.6 1.1 2.3 23 9.7 0.1 0.9 0.5

5 90.0 0.0 6.2 3.5 16 78.7 -4.4 0.8 2.4 26 9.1 -0.7 0.4 0.6

Table 6.13: Comparison of scenario 5 and Truncated MIP (TM) after 1 hour runtime.

Again, benchmark is the best value of all scenarios after the minimum runtime as pre-

sented in the columns Min and RT in Table 6.11. Column TM gives the percentage gap of

the solution found by Truncated MIP to this benchmark. Once more the gap of the best

objective value of 30 runs (MG) and the gap of the corresponding average value (AG) are

given as well as the coeﬃcient of variation (CV). Note that since the heuristics are run for

1 hour, the benchmark values can even be improved which results in negative MG-values.

Even though the CVs can only be slightly reduced for some of the instances, especially

for the general and the divergent product structures the supremacy of the new procedure is

shown. Since the nine chosen problem instances are the largest ones in terms of number of

setup variables, it can be presumed that the bigger the problems the more eﬀective VNDS+E

behaves compared to Truncated MIP. In the following, this supposition should be checked

with the help of a real-world problem instance.

140 Chapter 6. Heuristics for the improved GLSPMS

A European company produces in a two-stage environment plastic foils with varying com-

positions and in diﬀerent types of breadth, thickness, color, and adhesion. At the ﬁrst stage,

seven kinds of resins are produced on two heterogeneous production lines. Another three

resins can only be externally purchased. These ten types of resins are then processed to sev-

eral hundred ﬁnal items. These items can be grouped into 41 setup families. These families

are produced on the three heterogeneous lines of the second stage. Furthermore, both stages

are decoupled by tanks with limited capacity.

The planning horizon encompasses 12 weeks, production takes place 24/7, i.e. 168 hours

per week. Demand forecasts are available on a weekly base and stock holding costs are

calculated on a weekly base as well. Accordingly, a macroperiod represents one week of

planning.

The problem is solved with Truncated MIP and again 30 times with VNDS+E, each

limited to 1 hour of runtime. Once more, the ﬂow formulation with the tighter setup forcing

constraints and the stock inequalities (F+M+K) is tested using parameter scenario 5. The

results can be found in Table 6.14. The shown values refer to the instance with 17 (= V )

microperiods per macro (204 microperiods in total). Again, the column AG describes the

average gap to the best solution found in these 30 runs and CV is the coeﬃcient of variation.

Unfortunately, these values (124.1 and 38.2, respectively) are very high. But as can also be

seen, Truncated MIP (column TM) is not able to ﬁnd any feasible solution.24 One reason

might be that due to the stock inequalities (see Section 5.4), Xpress needs the 8 GB of RAM

just for reading the model. Accordingly, it is not surprising that the LP relaxation of this

instance takes more than half an hour and the initial LP with the ﬁxed setup pattern of the

trivial solution about 12 minutes.

SC5 TM

Micros AG CV Gap

204 124.1 38.2 no incumb.

Table 6.14: Comparison of Truncated MIP (TM) and the new heuristic after 1 hour of

runtime.

Since the range of solution values is quite high, the impact of the used model formulation

and the chosen parameter scenario must be clariﬁed. Hence, further comparisons are made

next.

At ﬁrst, parameter scenario 2 is applied to the ﬂow formulation again as this scenario

showed more stable results in Section 6.3.2.2. However, since the memory problem is still

present, the VNDS+E heuristic is also applied for the original improved formulation (O)

24

This is also the case for 16, 15, and 14 microperiods per macro.

6.3. Computational tests 141

which is the smallest one in terms of the number of variables and constraints.25 Both

parameter scenarios 2 and 5 are used for this second formulation. The results are given in

Table 6.15 where the best solution found for F+M+K and scenario 2 with 204 microperiods

serves as the benchmark (the corresponding MG-value is therefore 0).

F+M+K O

SC5 SC2 SC5 SC2 TM

Micros MG AG CV MG AG CV MG AG CV MG AG CV Gap

204 0.0 124.1 38.2 0.0 181.0 48.6 -17.7 -9.7 5.1 -31.7 -26.6 4.5 no incumb.

216 – – – – – – – – – -30.5 -24.0 3.9 –

228 – – – – – – – – – -27.2 -22.3 4.1 –

First of all, the new results conﬁrm the previous ones because scenario 2 performs worse

than scenario 5 for the ﬂow formulation. Moreover, not only is the average gap larger but

also the CV value is signiﬁcantly higher. In contrast, the results of the original formulation

are much better than the ones of the ﬂow formulation. For instance, the best solution found

by scenario 5 is 17.7% better than the best result of the ﬂow formulation for the same

runtime. This might be an indication that the availability of RAM is a crucial factor (about

2.2 GB are necessary for model loading). However, it seems that a large-sized instance of

the GLSPMS is hard to solve anyway because Xpress is still not able to ﬁnd an incumbent

solution.26 Probably, this is also the reason why scenario 5 performs better than scenario 2.

Surprisingly, scenario 2 beats scenario 5 for the original formulation to a large extent. The

MG-value is almost 32% better and on average the objective function values are 26.6% lower

than for the ﬂow formulation.27

Since the results of the original formulation with scenario 2 are quite stable (CV < 5%),

these settings are used to check the impact of increasing the number of microperiods.28 As

also illustrated in Table 6.15, once more VNDS+E provides stable results for 216 and 228

microperiods. However, these solutions are slightly worse than the ones with 204 micrope-

riods. Correspondingly, 17 microperiods per macro seem to be a reasonable choice as the

convergency is good as well (cf. Figure 6.17).

25

Moreover, this formulation allows to compare the results of Seeanner et al. (2013) who also use the

original formulation, but of a further GLSPMS variant.

26

The GLSPMS as presented by Seeanner et al. (2013) was solvable by Xpress. However, it did not

contain the synchronization constraints yet. Hence, it appears that the new constraints make the improved

GLSPMS much harder to solve.

27

These results are closely related to the ones by Seeanner et al. (2013) who make the same observation.

28

The “–” symbolizes that these settings are dropped.

142 Chapter 6. Heuristics for the improved GLSPMS

# !"

Figure 6.17: The percentage gap of the best VNDS+E run with the O-formulation and

scenario 2 for 204 microperiods.

In this chapter, several new heuristics for the GLSPMS have been developed and tested. To

start with simple construction heuristics, LP&Fix and Relax&Fix were examined ﬁrst. Even

for small instances, both heuristics showed bad performance. Hence, the focus was shifted to

improvement heuristics, especially to the new VNDS+E heuristic based on (micro)period,

product, bill-of-materials and production line decompositions. After the parameters were

tuned for this heuristic, VNDS+E was then compared to a GLSPMS tailored version of

Fix&Optimize and Truncated MIP which served as a benchmark. For this comparison,

bigger problem instances were created and the two very fast construction heuristics TS and

RS which make use of the GLSPMS relaxations were applied to generate initial solutions. In

sum, the results of VNDS+E were signiﬁcantly better for most of the instances. Moreover,

an indication for convergency could also be found.

Because of the good results, VNDS+E was also used to solve a large real-world problem

instance. Unfortunately, the model formulation F+M+K exceeded the available PC-memory

which seemed to have an impact on the solution quality. However, by using the original

GLSPMS formulation (improved version), VNDS+E was able to ﬁnd good solutions for this

large instance as well. Thus, a powerful solution method is available now. For this reason,

extensions of the GLSPMS can be examined in the following.

Chapter 7

Chapter 2 illustrated the attributes of a ﬂow line production system which might need to be

considered by a simultaneous lot-sizing and scheduling. Indeed, the GLSPMS regards most

of these attributes and allows hereby solving diﬀerent problem settings of the consumer pack-

aged goods industry. Despite, there are some important aspects left which are not included in

this model so far. But as already outlined before, a model should be as detailed as necessary

and as simple as possible for solvability purposes. However, as the new VNDS+E heuristic

is able to solve the GLSPMS satisfactorily, it now makes sense to develop further extensions

of the model. Especially the structural properties of the GLSPMS help integrate further

(scheduling) aspects. Hence, the ﬁrst section will present several extensions concerning the

availability of scarce resources. Since these extensions might potentially need a high number

of microperiods, the drawback of short period lengths as shown in Section 4.3.2 becomes a

severe issue. Therefore, time continuity will be the topic in the second part of this chapter.

Most of the lot-sizing and scheduling models given in Chapter 3 only regard the limited

capacity of machines and production lines, respectively. However, as outlined in Section 2.2.2,

other resources like raw materials and labor force might also be limited and thus have to be

taken into consideration.

Normally, a manufacturing company tries to ﬁnd reliable suppliers with a reasonable service

level and reasonable lead times. Nevertheless, it might happen for diﬀerent reasons1 that not

1

For example, after the earthquake, the tsunami, and the nuclear disaster in Japan, many supply chains

got disrupted in 2011.

DOI 10.1007/978-3-658-02089-7_7, © Springer Fachmedien Wiesbaden 2013

144 Chapter 7. Extensions of the improved GLSPMS

enough raw materials are available. Then, the materials have to be considered by the model

as well. In the following, we assume that M diﬀerent materials are scarce and that Rmt

describes the available quantity of the material m in macroperiod t. Furthermore, gozinto

factor armj gives the quantity of material m which is needed for producing one unit of item j.

In sum, this leads to the ﬁrst extension as follows:

Index

m = 1, ..., M raw materials

Data

armj consumption (unit) of material m needed to produce one unit of product

j

Rmt available quantity (unit) of material m in macroperiod t

armj xljs ≤ Rmt ∀m, t (7.1)

l,j s∈St

By adding the constraints (7.1) to the improved GLSPMS, it can be assured that in

each macroperiod t only the available quantity of material m is used for production. In the

regarded case, the demand for the raw material is triggered by the production activity. But

certain setup states2 or changeovers might need further resources, too. With the help of

consumption coeﬃcients for a certain resource and its available capacity, such limitations

can be modeled in the same way as for the raw materials above. Then, the resource demand

always refers to an entire micro- or macroperiod. However, this approach is quite “rough”

even for microperiods. But especially for scarce resources like setup operators, the limitation

may have a strong impact on the resulting production plan. Accordingly, a more complex

formulation must be used to include the timing aspect in a more appropriate way as shown

next.

Flow lines are sometimes operated and maintained by highly specialized personnel. This

staﬀ is usually neither working 24/7 nor available in an unlimited number. Accordingly,

it might be necessary to model the limited availability of setup operators who execute the

setup changeover. In contrast to the area of scheduling (cf. Cheng et al., 1999), researchers

2

A typical real-world case is, for instance, the availability of certain molds which are installed to the

machine and are needed for production.

7.1. Availability of resources 145

in the ﬁeld of lot-sizing (and scheduling) have paid less attention to this case so far. That

is why only a few articles can be found in the literature.

One recent approach is proposed by Tempelmeier and Buschkühl (2008). They extend

the PLSP for a common setup operator. An important issue for the deployment of the

operators is the time synchronization again because it is necessary to know the starting

and ending time of the setup operations. Therefore, Tempelmeier and Buschkühl introduce

additional variables in order to determine these points in time. Since these points can be

directly obtained within the GLSPMS, it is straightforward to include setup operators in this

model. Figure 7.1 shows the optimal solution of the basic instance with a general product

structure (see Section 4.4.1.3).

l=1

l=2

l=3

0 240

Figure 7.1: The optimal solution of an artiﬁcial instance without considering any setup

operator.

As can be seen, on all three machines, several setups (black boxes) have to be executed

in the planning horizon, many of them in parallel. However, if only a single setup team is

available in reality for operating the production lines, this shown solution would be infeasible.

Hence, a corresponding extension of the GLSPMS, which allows modeling this limitation of

setup resources, is presented next.

Therefore, we assume that o = 1, ..., O diﬀerent setup operators / setup teams are avail-

able. Each operator is able to execute diﬀerent changeovers. Moreover, a setup must be

completely executed by the same operator. The idea is now to connect a suited operator

to a changeover with the help of additional binary variables. Such a binary variable ỹlos

becomes 1 if the setup time needed for a changeover on line l in microperiod s (xel,s−1 + xbls )

has to be spent by operator o. In sum, the following index, index sets, and variables are

used:

Index

o = 1, . . . , O setup operators

Index sets

Olij set of operators which are able to execute the changeover from product i

to j on line l

146 Chapter 7. Extensions of the improved GLSPMS

Variables

ỹlos ∈ {0, 1} equals 1 if a setup changeover on production line l in microperiod s indi-

cated by i,j : i=j zlijs = 1 is assigned to setup operator o (0 otherwise)

z̃okls ∈ {0, 1} equals 1 if setup operator o is assigned to line k in microperiod s − 1 and

to l (diﬀerent from k) in microperiod s (0 otherwise)

Based on these variables, the constraints which are introduced next can be added to the

improved model formulation of Section 4.2.2.

ỹlos ≥ zlijs ∀l, i, j = i, s (7.2)

o∈Olij

ỹlos ≤ 1 ∀o, s (7.3)

l

xbks + xels − stmax (1 − z̃okl,s+1 ) ≤ (ws+1 − ws ) + os ∀o, k, l = k, s (7.5)

First of all, constraints (7.2) enforce that a suitable operator has to execute the changeover.

If a setup change from product i to j takes place on line l in microperiod s (zlijs = 1), an

operator o needs to be chosen from Olij . Moreover, constraints (7.3) ensure that an operator

can only be assigned to only one machine at the same time.

If a certain setup operator o is responsible for two consecutive changeovers on the same

line, no time synchronization problem occurs. In contrast, if the operator o needs to execute

changeovers on diﬀerent lines, an infeasible plan may result due to the setup splitting op-

portunity of the GLSPMS. As shown in Figure 7.2a, it might happen that the operator has

not ﬁnished the operation on line k for a changeover (zkii s = 1) before the setup changeover

on another line (zljj ,s+1 = 1) starts. Thus, the setup times may overlap.

s-1 s s+1 s-1 s s+1

time

(a) No time synchronization of setups. (b) Synchronization of setup operations.

7.1. Availability of resources 147

To avoid this problem, the corresponding line change must be detected and suﬃcient

time for both setups has to be made available. The former is achieved by constraints (7.4)

and the binary variable z̃okls which becomes 1, if operator o is ﬁrst assigned to line k in

microperiod s − 1 and then to l in microperiod s. The latter follows from constraints (7.5):

If the setup operator has to change from line k to l in microperiod s, the term in brackets

on the left-hand side becomes 0 and the length of the microperiod must be at least as long

as both setup time fractions xbks and xels . Otherwise, the microperiod length must not be

inﬂuenced by this constraint. For that purpose, −stmax is chosen as a coeﬃcient of suﬃcient

size because xbks ≤ stmax and xels ≤ stmax and the microperiod is per se at least as long as

each of these fractions. Correspondingly, the result of these new constraints may look like

Figure 7.2b. Note that it is allowed to assign two or more operators to a certain line in the

same microperiod because such a solution is, of course, feasible in reality. However, if setup

operators are actually scarce, a multiple assignment will not occur. To illustrate the impact

of constraints (7.2)–(7.5), the test instance considered at the beginning of this section is

extended by the following data: We assume now that only a single setup team (O = 1) is

available which is able to execute the changeovers on all lines. Accordingly, the index set is

Olij = {1} for all l and all i, j ∈ Jl . Figure 7.3 illustrates the result and Table 7.1 shows the

corresponding values of the new variables.

l=1

l=2

l=3

0 240

Figure 7.3: The optimal solution for the same artiﬁcial instance with one setup team.

yl,1,s

o s=1 s=2 s=3 s=4 s=5 s=6 s=7 s=8 s=9

l=1 0 1 1 0 0 1 0 1 0

l=2 0 0 0 1 0 0 1 0 1

l=3 1 0 0 0 1 0 0 0 0

z1(l,k)s

o =1 (3,1) (1,2) (2,3) (3,1) (1,2) (2,1) (1,2)

In this solution, the operator has to change, for instance, from line k = 3 to l = 1

in microperiod s = 5. Therefore, this period is long enough (14 time units) to allow the

operator executing the setups (xb35 = 6 and xe15 = 6).

Interestingly, the computational eﬀort for Xpress to solve the current test instance to

optimality is much lower than without regarding the setup operators. The optimal solution

148 Chapter 7. Extensions of the improved GLSPMS

as shown in Figure 7.3 could be found after 250 sec. of runtime. On the contrary, the solution

in Figure 7.3 was found in 45 sec. Obviously, the MIP solver beneﬁts in this case from the

reduction of the solution space. Instead of 12 changeovers, the single setup team is now

only able to execute eight changeovers. This leads to higher stock holding costs (about 70%

increase) and makes it necessary to purchase items (total costs raise by factor 2.5). But this

can be avoided as the production lines and the operator have idle time. One possibility would

be to increase the number of microperiods to allow more setups (e.g., using four microperiods

per macroperiod reduces the total costs by 28% for the regarded instance). Unfortunately,

this might cause another problem as will be illustrated next in Section 7.2.

Another limitation may be the availability of setup tools which are needed for executing

the changeovers. By using the same variables and constraints as for the setup operators,

these limited setup tools can be modeled exactly the same way.3

As outlined in the section before, a high number of microperiods might become necessary.

But this leads to short microperiod lengths which are a severe issue (see also Section 4.3.2).

Despite the setup splitting capabilities of the GLSPMS, high setup times can lead to unnec-

essary idle times or even make the problem infeasible.

Correspondingly, an approach which is able to “distribute” setup times over more than

two consecutive microperiods is preferable in that case. Suerie (2005b) examines how STB-

and LTB-models can be extended in a way that lot-sizes as well as setup operations do not

have to stick to any time bucket boundaries. Suerie presents diﬀerent model formulations

which he embraces under the term “time continuity”. The underlying modeling approach for

period-overlapping setup times will be applied to the improved GLSPMS in the following

(q.v. Suerie, 2006).

Actually, there are several scenarios where the setup time might become a critical factor.

For instance, if plenty more items have to be produced on a certain line than on another line

and the latter one shows signiﬁcantly longer setup times, the GLSPMS might not be able to

consider this appropriately. Figure 7.4 shows an example of a production plan which cannot

be obtained by the model so far.

3

Note that the new VNDS+E heuristic can be applied to the improved GLSPMS with the operator

extension. However, the RS model (Section 6.2.5) cannot be used to ﬁnd a ﬁrst initial solution, since

overtime might not be suﬃcient for generating a feasible solution. The reason is that more setups have to

be executed consecutively which leads to a higher number of needed microperiods. Either the RS model has

to be adapted accordingly or the TS (Section 6.2.4) solution can be used alternatively.

7.2. Time continuity 149

microperiods

s=1 s=2 s=3 s=4 s=5

time

Figure 7.4: Example for a setup time overlapping more than two microperiods.

processed in the same time as it takes to set up line l = 2 for product j = 7. Hence, the

GLSPMS should allow the setup time on this second line to overlap the microperiods s = 2,

3, and 4 to get such a compact plan.

If a setup has to span more than two microperiods, there must be at least one microperiod

in between which exclusively consists of setup time. To model this, new setup state variables

yljs are inserted into the GLSPMS in analogy to Suerie. They indicate such a “continuous”

setup on line l in microperiod s and forbid any other activity (standby or production) on the

corresponding line. Regarding the above example, it means that the continuous setup state

must be active (set to 1) for line l = 2 in microperiods s = 2, 3, and 4. To be speciﬁc, these

continuous setup states need to be activated for product j = 6 (y262 = y263 = y264 = 1).

This is necessary as in microperiod 5 we need to know the product that was set up before

the changeover has been started. Otherwise, we cannot determine which changeover is

completed.

Since the lengths of the microperiods are not ﬁxed in the improved GLSPMS, additional

variables τls and σls are needed. They cumulate the times which has been spent for setup on

line l until microperiod s. However, they should only cumulate as long as these times belong

to the same setup changeover process. Therefore, a reset of these variables is necessary: If

a new production lot begins, the corresponding τls and σls have to be set to zero. If a new

changeover starts, the variables must take the value of the time already spent for this setup.

The diﬀerences between these two types of variables, why they are needed, and how this

reset works will be explained right after the model formulation is given.

In sum, the improved GLSPMS has to be extended by three new groups of variables:

Variables

yljs ∈ {0, 1} continuous setup state yljs equals 1 if microperiod s on line l exclusively

consists of setup time and product j is the last item which was produced

in a preceding microperiod v < s

τls lower bound for the cumulated setup time until microperiod s on line l

σls upper bound for the cumulated setup time until microperiod s on line l

150 Chapter 7. Extensions of the improved GLSPMS

Even though only three new groups of variables are needed, several new constraints

have to be included to realize the period overlapping setup times. For this reason, a com-

plete formulation of this extension is given next (once more the numbers of the unchanged

elements—compared to the improved GLSPMS—are marked with an asterisk):

Objective function

min hjs Ijs + slij zlijs + cplj xljs

s∈Λ,j=0 l,i,j,s l,j,s

+ cl (xbls + xels ) + cej ejs + co · o s + hjsxljs (7.6*)

l,s j=0,s s l,s∈Λ,j=0

subject to

ws = w s ∀s ∈ Φ (7.7*)

ljs + xljs = xljs

x ∀l, j = 0, s (7.8*)

Ijs = Ij,s−1 + ljs +

x xlj,s−1 − pji xlis + ejs − djs ∀j = 0, s (7.9*)

l l l i∈Nj

Ijs ≤ Ijmax ∀j = 0, s (7.11*)

xljs ≤ Wmax

lj ∀l, j, s (7.12*)

xbls + xbls + alj xljs + xels + xels = (ws+1 − ws ) + os ∀l, s (7.13*)

j

xljs ≥ mlj (yljs − ylj,s−1 ) ∀l, j, s (7.15*)

(yljs + yljs ) = 1 ∀l, s (7.16)

j

yljs = 0 ∀l, s, j ∈

/ Jl (7.18*)

yljs = 0 ∀l, s, j ∈

/ Jl (7.19)

xbls + xels + xels ≤ wS+1 1− yljs ∀l, s (7.20)

j

τls ≥ stlij zlijs ∀l, s (7.22)

i,j

τls ≤ xel,s−1 + xbls + stmax 1− ylj,s−1 ∀l, s (7.23)

j

7.2. Time continuity 151

σls ≥ σl,s−1 + xel,s−1 + xbls − stlij zlij,s−1 ∀l, s (7.25)

i,j

σls ≤ stlij zlijs + stmax yljs ∀l, s (7.26)

i,j j

zlijs = 1 ∀l, s (7.27*)

i,j

xljs = 0 ∀l, s ∈

/ Πl (7.28*)

j=0

ejs ≤ emax

j ∀j = 0, s (7.29*)

os ≤ omax ∀s ∈ Λ (7.30*)

os = 0 ∀s ∈

/ Λ (7.31*)

xbls + xbls ≥ xbks + xbks − wS+1 (2 − yljs − ykis ) ∀s, (k, i, l, j) ∈ D (7.32*)

kis + xeks + xeks ≥ xels + xels − wS+1 (2 − ykis − yljs ) ∀s, (k, i, l, j) ∈ D (7.33*)

aki x

As described above, a new setup state is introduced. Correspondingly, the variables yljs

enter the unique setup constraints (7.16). Hence, the changeover constraints (7.17) have

to consider the new setup states as well. Moreover, constraints (7.19) ﬁx the continuous

setup state variables to zero for items which are not producible. If yljs is active (=1),

constraints (7.20) force all elements of microperiod s to be zero—except for xbls .4

Since the setup splitting constraints (4.8) and (4.9*) of the improved formulation pre-

vent setup times to overlap more than two consecutive microperiods, these constraints are

dropped. But of course, it must be still assured that the whole setup time is “spent”. Hence,

the variables for the cumulated setup time come into play. To achieve a cumulation, con-

straints (7.21) allow τls to reach a value up to the sum of the cumulated setup time in the

previous microperiod and the time actually spent for the setup between the last and the

current microperiod. Correspondingly, τls is a lower bound which must be at least as big as

the necessary setup time (7.22) if the changeover process is ﬁnished in microperiod s and

the changeover variable zlijs becomes 1 for i = j, respectively.5 Note that for these con-

straints (7.22) an equality can not be postulated as zljjs = 1 forbids a cumulation. But as

soon as a changeover is completed, τls must be reset. Otherwise, the cumulated value could

be carried to the next changeover and the corresponding setup time needs not to be spent

completely anymore. Therefore, constraints (7.23) are necessary: If there was a “normal”

setup state in the previous microperiod s − 1, τls is limited to the sum of the setup split

variables. Otherwise, τls must not be restricted which is possible by using the maximum

setup time as an upper bound. Since this reset is controlled by the “normal” setup state,

4

Note that the production variable xljs is already zero because the “normal” setup state variable yljs

has to be zero as well due to (7.16).

5

A lower bound is usually suﬃcient because a limited capacity forces xbls and xels (thus τls ) to take the

smallest possible value.

152 Chapter 7. Extensions of the improved GLSPMS

a continuous setup state yljs may only become 1 if already the continuous setup state or a

normal setup state for the same item precedes in microperiod s − 1 (7.24).

Usually, for most of the test instances in this thesis, the constraints so far will suﬃce.

However, since the GLSPMS also considers standby costs, it might be possible for τls —to be

more precise, for the corresponding xbls and xels —to take a value bigger than the corresponding

setup time. The new upper bound variables σls now help to prevent this. Constraints (7.25)

force σls to cumulate and to be an upper bound for the setup time spent. At the same time,

these constraints reset the value if a changeover is completed in the previous period s − 1.

Then, σls only has to be non-negative. Due to (7.26), σls must not exceed the needed setup

time. Accordingly, xel,s−1 and xbls are restricted and thus τls as well. Note that the second

term on the right-hand side of (7.26) is necessary because a continuous setup state with

zljjs = 1 forces σls to be zero.

To understand the interaction of the new variables and constraints, Figure 7.5 illustrates

the result of a small instance with one production line, four products, and ﬁve macroperiods

(each with a single microperiod). To show the diﬀerent eﬀects, this example includes a short

setup time at the beginning of the planning horizon as well as two longer setups which need

to overlap three and two periods, respectively.

0 50

Table 7.2 gives the solution values of the new variables as well as of the normal setup

state and changeover variables. In the ﬁrst microperiod s = 1, τ11 and σ11 take the values

of the setup time st112 = 4 since the corresponding changeover is completed. In the next

period, τ12 and σ12 are set to zero. This is important because otherwise the old value (=4)

could be taken into account for the next setup which starts in the current microperiod. This

reset is achieved by (7.23) as j y1j1 = 1 and by (7.25) as z1121 = 1.

In microperiod s = 3, the continuous setup state y123 = 1 becomes active. Hence, xb13

uses the entire microperiod. Because of z1234 = 1, τ14 = st123 , and σ14 = st123 , the setup

is ﬁnished in microperiod s = 4. Finally, the last setup is executed regularly without a

continuous setup.

A problem quite similar to this also arises for minimum lot-sizes. This will be considered

next.

7.2. Time continuity 153

y1js = 1 2 2 - 3 4

yljs = 1 - - 2 - -

xb1s 4 0 10 4 6

xe1s 0 5 0 1 0

τ1s 4 0 15 19 7

σ1s 4 0 15 19 7

z1(i,j)s = 1 (1,2) (2,2) (2,2) (2,3) (3,4)

Although production lots may overlap an arbitrary number of microperiods, the minimum

lot-size condition must be fulﬁlled within a single microperiod. But if large minimum lot-

sizes have to be assured (cf. Transchel et al. (2011)), the GLSPMS might suﬀer from the

same problem as with the setup times. Again, if the number of microperiods is very high, the

minimum lot-sizes lead to unnecessary big microperiods and potential idle times. Or even

worse, if the remaining capacity is not suﬃcient for such a lot, it may lead to infeasibility.

The minimum lot-size issue can be solved in a similar way as for the setup time issue by

using variables τ̂ljs as the cumulated lot-sizes and by adding the following constraints to the

formulation of Section 4.2.2:

τ̂ljS ≥ mlj yljS ∀l, j (7.35)

τ̂ljs ≤ xljs + mlj 1− yli,s−1 ∀l, j, s (7.36)

i=j

In contrast to the formulation for the period overlapping setup time, no additional state

variables are necessary because the changeover variables already indicate a continuous setup

state (zljjs = 1). However, the variables τ̂ljs which cumulate the production quantities still

need to be introduced. Similar as before, these new variables have to “cover” the minimum

lot-size as soon as a new setup state occurs in the next microperiod. This is achieved by

(7.34) and by (7.35) for the last microperiod S. Since the term (yljs − ylj,s+1 ) only becomes 1

if the setup of product j is lost from microperiod s to s + 1, the minimum lot-size constraints

need not to be fulﬁlled as long as the setup for product j is conserved. Accordingly, variables

τ̂ljs must be reset as soon as a new lot starts which is expressed by (7.36). Vice versa, it

154 Chapter 7. Extensions of the improved GLSPMS

means that as long as the setup state is the same, the production quantity can be added

(7.37). Here, the cumulated quantity can only be extended by the quantity of the current

microperiod.6

Summing up, this chapter showed several extensions of the improved GLSPMS. First,

scarce resources like raw materials or setup teams were considered. It became apparent

that the time structure of the GLSPMS allows integrating easily those aspects into the

model. Second, it was shown how the improved GLSPMS has to be extended to make

period-overlapping setup times and minimum lot-sizes possible.

6

Since the setup state variables are still the same as in the improved formulation, the VNDS+E can be

used without any changes.

Chapter 8

8.1 Summary

In the consumer packaged goods industry, usually a large number of ﬁnal items is produced

in a two- or three-stage ﬂow line production system with heterogeneous, parallel production

lines at each stage. Due to sequence-dependent setup times and costs a simultaneous lot-

sizing and scheduling of the bottleneck stage is necessary. But as Chapter 2 illustrated, line-

and product-speciﬁc production speeds in combination with time-varying demand can easily

lead to dynamically shifting bottlenecks. In this case, even multiple production stages need

to be considered at the same time. Unfortunately, this problem seems not to be adequately

addressed by current Advanced Planning Systems what might explain the dissatisfaction

of production planners in the regarded industry. Motivated by this situation, the thesis at

hand focused on MIP-models for the simultaneous lot-sizing and scheduling of multi-stage

ﬂow line production systems.

As became apparent in Chapter 3, the scientiﬁc community has also paid less attention

to this problem so far, since only a few multi-level models actually exist in the literature.

Besides, most of these models suﬀer from the underlying time structure. Unrealistically high

lead-time oﬀsets between the stages and setup time issues belong to the major drawbacks. In

fact, just a single model seemed suitable for tackling the considered problem: the “General

Lot-sizing and Scheduling Problem for Multiple production Stages” (GLSPMS) as proposed

by Meyr (2004). The main idea of the GLSPMS is to use a common time grid for all

production lines at the diﬀerent stages. This allows zero lead-times and gives the opportunity

of including further scheduling aspects.

But as Chapter 4 revealed, the GLSPMS in its basic form has some shortcomings con-

cerning the synchronization of the diﬀerent stages. Hence, a revised variant of the GLSPMS

was proposed and analyzed. Even though this improved model was proved in this thesis to

be NP-complete, it had to be checked whether problem instances of a suﬃcient size can be

solved to optimality anyway. Accordingly, artiﬁcial test instances with three diﬀerent types

DOI 10.1007/978-3-658-02089-7_8, © Springer Fachmedien Wiesbaden 2013

156 Chapter 8. Summary and Outlook

of product structures (serial, divergent, general) were presented. With the help of these in-

stances, the computational performance of a standard MIP-solver was examined. It turned

out that only small instances could be solved to optimality in a reasonable amount of time

(1h).

However, since the model formulation itself has a signiﬁcant impact on the computational

eﬀort of a MIP-solver, Chapter 5 probed diﬀerent reformulations of the improved GLSPMS.

For this purpose, several reformulation techniques which had been successfully applied to

other types of lot-sizing problems were adapted for the GLSPMS (see also Seeanner and

Meyr, 2013). In addition to a Simple Plant Location (SPL) formulation based on echelon

stocks (cf. Stadtler, 1997), a network ﬂow formulation (cf. Karmarkar and Schrage, 1985)

was presented. Furthermore, a subset of the (l,S) valid inequalities (cf. Barany et al., 1984)

as well as tightened setup forcing constraints were introduced. Each possible combination of

these formulations and valid inequalities was tested afterwards, again with the artiﬁcial in-

stances of the previous chapter. As a result, the SPL based reformulation performed slightly

better for instances with a divergent product structure, whereas the ﬂow formulation showed

advantages for all other instances. Moreover, both groups of valid inequalities were always

recommendable for all product structures. Nevertheless, even for instances of moderate size

the percentage of unﬁnished problems which cannot be solved to optimality within one hour

of runtime was still high.

For this reason, new heuristics for the GLSPMS were developed and tested in Chapter 6.

In a ﬁrst step, the simple construction heuristics Truncated MIP, LP&Fix, and Relax&Fix

were compared to each other. But the results were not satisfying. Correspondingly, the

focus was shifted towards two improvement heuristics: On the one hand, a GLSPMS tailored

Fix&Optimize was introduced as it is known to produce good results for another lot-sizing

model (cf. Helber and Sahling, 2010). On the other hand, a new heuristic combining the

principles of the Variable Neighborhood Decomposition Search (VNDS) heuristic and the

MIP-based Exchange was developed (cf. Seeanner et al., 2013). Moreover, for this VNDS+E

heuristic, diﬀerent neighborhood structures were deﬁned based on (micro)period, product,

bill-of-materials and production line decompositions. With the help of these decompositions,

subproblems of the GLSPMS were built and solved.

Since both improvement heuristics need an initial solution and LP&Fix as well as Re-

lax&Fix showed bad performance before, two further construction heuristics were presented.

These two take advantage of the GLSPMS structure by using penalty costs. Though they

generate bad solutions yet, these solutions are acceptable for this purpose as the heuristics

are very fast and allow to spend more time for the improvement procedures. The ﬁrst con-

struction heuristic creates a setup sequence which conserves the initial setup state and forbids

any changeover. The second one builds a repeating setup sequence where each product with

a positive net demand is set up at least once in every macroperiod.

But before a comparison between the improvement heuristics could be made, the pa-

8.2. Outlook 157

rameters of the VNDS+E had to be tuned. Therefore, parameter scenarios with diﬀerent

settings which also included the two new construction heuristics were deﬁned. In sum, the

scenario that deactivates the starting heuristic of the MIP-solver performed best. Hence, it

was used for the following tests. However, it could be observed that the other parameter

scenarios also performed quite well. Accordingly, VNDS+E is quite robust in terms of the

parameters used and the quality of the initial solution. Both heuristics were then examined.

Moreover, the artiﬁcial test instances of Chapter 4 were increased to get more realistic prob-

lem sizes. In accordance with the previous results, the ﬂow formulation with both groups of

valid inequalities was used. In addition, Truncated MIP served as a benchmark because it

is construction and improvement heuristic at the same time. As a result, the new VNDS+E

heuristic beat the other two by far.

Up to then, only (small- and big-sized) artiﬁcial test instances have been examined. Since

the VNDS+E produced good results, a large-sized real-world problem from a chemical com-

pany was examined as well. Unfortunately, the resulting MIP caused memory issues for the

PC due to the size of the model formulation with the subset of (l,S)-inequalities. Accord-

ingly, VNDS+E was tested once more, now with the original formulation of the improved

GLSPMS. Hereby, reliably good solutions (as conﬁrmed by the company) could be obtained

within 1-2 hours of runtime.

Finally, several extensions of the GLSPMS were proposed in Chapter 7. Due to its time

structure, scarce setup resources such as operators and tools which have to be synchro-

nized between the stages could be integrated in the model. Moreover, one weakness of the

GLSPMS was ﬁnally tackled by introducing period overlapping setup times as well as period

overlapping minimum lot-sizes.

8.2 Outlook

To complete this thesis, some topics will be presented next which might be interesting for on-

going research (cf. Clark et al., 2011). This includes some deepening aspects concerning the

application of the improved GLSPMS and the VNDS+E heuristic as well as the broadening

view of neighboring ﬁelds with existing interdependencies:

for solving diﬀerent problem settings from the consumer packaged goods industry.

Accordingly, it is possible to tackle some of the problems modeled by the other multi-

level formulations. Then, not only could the quality of the GLSPMS but also of the new

solution method be compared. Furthermore, the extensions presented in Chapter 7 are

only tested with small instances for feasibility. That is why these could be tested for

real-world problems, too. In that context, the question also arises whether the newly

developed solution procedure has to be adapted or still performs very good.

158 Chapter 8. Summary and Outlook

Adapting VNDS+E for other model formulations The new VNDS+E heuristic pro-

duced good results for the GLSPMS. However, its basic structure can also be applied

for other lot-sizing (and scheduling) problems. For that purpose, the neighborhoods

probably need to be adapted. Especially, an adaption for the MLCLSP would be of

high interest, since a great variety of heuristics has been developed for this model until

today (cf. Almeder (2010), Helber and Sahling (2010), Stadtler (2003) among others).

Then, a thorough comparison between all these heuristics could be made.

Including open-source based LP-/MIP-solvers In the past years, many high quality

open-source software solutions have been developed which more and more get embed-

ded in business applications. The OR community is also very active and published

many freely available open-source LP-/MIP-solvers like GLPK (2013) or CBC (2013),

to name only a few. Even though these solvers cannot compete with commercial ones

in every ﬁeld of application, they show diﬀerent strengths which might be useful (cf.

Mittelmann, 2012). Accordingly, it might be interesting to develop solution methods

which make use of these good properties and thus become competitive. For instance,

the CLP (2013) solver is supposed to be quite fast for LPs. Therefore, the well-known

method of generating setup sequences by a meta-heuristic and solving the remaining

LP could be a reasonable approach (cf. Kuik et al., 1993; Meyr, 2000).1

Extending the GLSPMS for distribution planning As Chandra and Fisher (1994) point

out a decoupled production and distribution planning is only acceptable if “there is

suﬃcient ﬁnished goods inventory to buﬀer the production and distribution operations

from each other”. But especially the cost pressure forces the consumer goods industry

to reduce the inventories. Furthermore, the perishability of goods can also inhibit high

stock levels, as the time between the date of production and the date of delivery is

limited. Accordingly, a simultaneous production and distribution planning might be

necessary (cf. Amorim et al., 2011). Therefore, another potential research topic is to

examine, whether the GLSPMS can be extended for a simultaneous distribution plan-

ning. As shown by the complexity proof in Section 4.3.1, the optimization version of

the Traveling Salesman Problem (TSP) is a special case of the GLSPMS. Accordingly,

it might be possible to integrate the Vehicle Routing Problem which is closely related

to the TSP into the GLSPMS.

Considering scarce setup resources In Chapter 7 several extensions were proposed. Es-

pecially in the section about scarce setup resources, only ﬁrst basic approaches were

shown which assumed that, for instance, the setup operators are available over the

whole planning horizon. However, it might be important to examine how the model

has to be adapted if the working time of machines and operators diﬀer. Another issue

1

Note that for the real-world problem instance presented in Chapter 6 the LP takes about 20 minutes if

the binary variables are relaxed and not ﬁxed.

8.2. Outlook 159

could be that setup operators do not only execute the changeover by “replacing” tools

but also have to operate the machines during production time (and not during idle

time). The same is true for scarce setup tools. On the one hand, these tools might be

necessary for the changeover as a tool for the operators (e.g., a special wrench). On

the other hand, tools might be needed for production (e.g., a die cutting tool or a top

frame). But then, it could be useful to diﬀerentiate the changeover between a set-up

and a tear-down process, especially for time-consuming changeovers.

Appendix A

In Section 4.4.1 three diﬀerent base scenarios were described. In the following, the complete

data of the resulting serial, convergent, and general instance will be given (q.v. Seeanner and

Meyr, 2013).

L 4

J 6

T 4

S 12 (3 micros per macro)

Λ {3,6,9,12}

Φ {1,4,7,10,13}

ws {0,-,-,80,-,-,160,-,-,240,-,-,320}

Table A.1: Number of lines, products, macroperiods, microperiods as well as set of all last

microperiods and all ﬁxed microperiods with their respective starting times.

Ij0 Imax

j hjs (s ∈ Λ) cej emax

j Nj

j=1 0.0 15.0 8.0 100.0 100.0 ∅

j=2 0.0 10.0 8.0 100.0 100.0 ∅

j=3 0.0 10.0 2.0 0.0 0.0 {1}

j=4 0.0 10.0 2.0 0.0 0.0 {2}

j=5 0.0 100.0 1.0 4.0 100.0 {3}

j=6 0.0 100.0 1.0 4.0 100.0 {4}

Table A.2: Initial and maximum inventory, holding and purchasing costs as well as maximum

purchase and sets of immediate and all successors.

DOI 10.1007/978-3-658-02089-7, © Springer Fachmedien Wiesbaden 2013

162 Appendix A. Artiﬁcial test instances

co 200

max

o 80

l=1 3.0 3.0 - - - -

l=2 - - 1.0 1.0 - -

l = 3, 4 - - - - 4.0 3.0

ft St

t=1 1 {1,2,3}

t=2 4 {4,5,6}

t=3 7 {7,8,9}

t=4 10 {10,11,12}

j=1 3.0 5.0 5.0 5.0

j=2 2.0 4.0 6.0 8.0

WIPmax

lj j=1 j=2 j=3 j=4 j=5 j=6

l=1 100.0 100.0 - - - -

l=2 - - 10.0 10.0 - -

l = 3, 4 - - - - 0.0 0.0

l=1 i=1 0.0 6.0 - - - -

i=2 8.0 0.0 - - - -

l=2 i=3 - - 0.0 2.0 - -

i=4 - - 12.0 0.0 - -

l = 3, 4 i=5 - - - - 0.0 8.0

i=6 - - - - 8.0 0.0

A.1. Serial instance 163

cp

lj j=1 j=2 j=3 j=4 j=5 j=6

l=1 1.0 1.0 - - - -

l=2 - - 1.0 1.0 - -

l=3 - - - - 1.0 1.0

l=4 - - - - 2.0 2.0

l=1 1.0 1.0 - - - -

l=2 - - 10.0 10.0 - -

l = 3, 4 - - - - 1.0 1.0

l=1 1 0 - - - -

l=2 - - 1 0 - -

l=3 - - - - 1 0

l=4 - - - - 0 1

i=3 6.0 - -

i=4 - 6.0 -

i=5 6.0 - 1.0 -

i=6 - 6.0 - 1.0

Jl Πl bl

l=1 {1,2} {1,2,...,12} 0.0

l=2 {3,4} {1,2,...,12} 0.0

l = 3, 4 {5,6} {1,2,...,12} 0.0

164 Appendix A. Artiﬁcial test instances

L 3

J 6

T 3

S 12 (4 micros per macro)

Λ {4,8,12}

Φ {1,5,9,13}

ws {0,-,-,-,80,-,-,-,160,-,-,-,240}

Table A.15: Number of lines, products, macroperiods, microperiods as well as set of all last

microperiods and all ﬁxed microperiods with their respective starting times.

Ij0 Imax

j hjs (s ∈ Λ) cej emax

j Nj

j=1 0.0 100.0 3.0 100.0 100.0 ∅

j=2 0.0 100.0 3.0 100.0 100.0 ∅

j=3 0.0 100.0 3.0 100.0 100.0 ∅

j=4 0.0 100.0 3.0 100.0 100.0 ∅

j=5 0.0 0.0 - 100.0 100.0 {1,2}

j=6 0.0 0.0 - 100.0 100.0 {3,4}

Table A.16: Initial and maximum inventory, holding and purchasing costs as well as maxi-

mum purchase and sets of immediate and all successors.

A.2. Divergent instance 165

co 200

omax 80

l=1 3.0 4.0 3.0 4.0 - -

l=2 - 8.0 - 8.0 - -

l=3 - - - - 4.0 2.0

ft St

t=1 1 {1,2,3,4}

t=2 5 {5,6,7,8}

t=3 9 {9,10,11,12}

j=1 0.0 6.0 6.0

j=2 0.0 6.0 6.0

j=3 2.0 6.0 6.0

j=4 3.0 6.0 6.0

WIPmax

lj j=1 j=2 j=3 j=4 j=5 j=6

l=1 100.0 100.0 100.0 100.0 - -

l=2 - 100.0 - 100.0 - -

l=3 - - - - 0.0 0.0

l=1 i=1 0.0 1.0 1.0 2.0 - -

i=2 2.0 0.0 3.0 1.0 - -

i=3 3.0 4.0 0.0 1.0 - -

i=4 5.0 3.0 2.0 0.0 - -

l=2 i=2 - 0.0 - 2.0 - -

i=4 - 4.0 - 0.0 - -

l=3 i=5 - - - - 0.0 2.0

i=6 - - - - 6.0 0.0

166 Appendix A. Artiﬁcial test instances

cp

lj j=1 j=2 j=3 j=4 j=5 j=6

l=1 1.0 1.0 1.0 1.0 - -

l=2 - 2.0 - 2.0 - -

l=3 - - - - 1.0 1.0

l=1 1.0 1.0 1.0 1.0 - -

l=2 - 1.0 - 1.0 - -

l=3 - - - - 1.0 1.0

l=1 1 0 0 0 - -

l=2 - 1 - 0 - -

l=3 - - - - 0 1

i=5 2.0 1.0 - -

i=6 - - 2.0 1.0

Jl Πl bl

l=1 {1,2,3,4} {1,2,...,12} 0.0

l=2 {2,4} {1,2,...,12} 0.0

l=3 {5,6} {1,2,...,12} 0.0

A.3. General instance 167

L 3

J 8

T 3

S 12 (4 micros per macro)

Λ {4,8,12}

Φ {1,5,9,13}

ws {0,-,-,-,80,-,-,-,160,-,-,-,240}

Table A.29: Number of lines, products, macroperiods, microperiods as well as set of all last

microperiods and all ﬁxed microperiods with their respective starting times.

Ij0 Imax

j hjs (s ∈ Λ) cej emax

j Nj

j=1 0.0 100.0 8.0 200.0 100.0 ∅

j=2 0.0 100.0 8.0 200.0 100.0 ∅

j=3 0.0 100.0 7.0 200.0 100.0 ∅

j=4 0.0 100.0 12.0 200.0 100.0 ∅

j=5 0.0 100.0 1.0 200.0 100.0 {1,2}

j=6 0.0 0.0 - 200.0 100.0 {1,3}

j=7 0.0 0.0 - 200.0 100.0 {2,4}

j=8 0.0 100.0 2.0 200.0 100.0 {3,4}

Table A.30: Initial and maximum inventory, holding and purchasing costs as well as maxi-

mum purchase and sets of immediate and all successors.

168 Appendix A. Artiﬁcial test instances

co 0

omax 0

alj j=0 j=1 j=2 j=3 j=4 j=5 j=6 j=7 j=8

l=1 1.0 6.0 6.0 3.0 3.0 - - - -

l=2 1.0 - - - - - 6.0 6.0 -

l=3 1.0 - - - - 8.0 - - 2.0

ft St

t=1 1 {1,2,3,4}

t=2 5 {5,6,7,8}

t=3 9 {9,10,11,12}

j=1 0.0 0.0 7.0

j=2 0.0 5.0 5.0

j=3 0.0 5.0 6.0

j=4 0.0 4.0 5.0

WIPmax

lj j=1 j=2 j=3 j=4 j=5 j=6 j=7 j=8

l=1 200.0 200.0 200.0 200.0 - - - -

l=2 - - - - - 200.0 200.0 -

l=3 - - - - 200.0 - - 200.0

A.3. General instance 169

slij /stlij j=0 j=1 j=2 j=3 j=4 j=5 j=6 j=7 j=8

l=1 i=0 0.0 999.0 999.0 999.0 999.0 - - - -

i=1 999.0 0.0 4.0 8.0 8.0 - - - -

i=2 999.0 4.0 0.0 8.0 8.0 - - - -

i=3 999.0 6.0 6.0 0.0 4.0 - - - -

i=4 999.0 6.0 6.0 4.0 0.0 - - - -

l=2 i=0 0.0 - - - - - 999.0 999.0 -

i=6 999.0 - - - - - 0.0 4.0 -

i=7 999.0 - - - - - 4.0 0.0 -

l=3 i=0 0.0 - - - - 12.0 - - 12.0

i=5 0.0 - - - - 0.0 - - 8.0

i=8 0.0 - - - - 6.0 - - 0.0

cp

lj j=0 j=1 j=2 j=3 j=4 j=5 j=6 j=7 j=8

l=1 1.0 1.0 1.0 1.0 1.0 - - - -

l=2 1.0 - - - - - 1.0 1.0 -

l=3 1.0 - - - - 1.0 - - 1.0

mlj j=0 j=1 j=2 j=3 j=4 j=5 j=6 j=7 j=8

l=1 1.0 1.0 1.0 1.0 1.0 - - - -

l=2 1.0 - - - - - 1.0 1.0 -

l=3 1.0 - - - - 1.0 - - 1.0

ylj0 j=0 j=1 j=2 j=3 j=4 j=5 j=6 j=7 j=8

l=1 0 1 0 0 0 - - - -

l=2 0 - - - - - 1 0 -

l=3 0 - - - - 1 - - 0

i=5 1.0 1.0 - -

i=6 1.0 - 0.5 -

i=7 - 1.0 - 0.5

i=8 - - 1.0 1.0

170 Appendix A. Artiﬁcial test instances

Jl Πl bl

l=1 {0,1,2,3,4} {1,2,...,12} 0.0

l=2 {0,6,7} {1,2,...,12} 0.0

l=3 {0,5,8} {1,2,...,12} 1.0

Appendix B

The following three selected ﬁgures show the development of the objective values for Trun-

cated MIP, the tailored Fix&Optimize, and the new VNDS+E heuristic with parameter

scenario 5. For each point in time the best, the worst, and the mean value is determined

for the VNDS+E. These representatives are chosen in a way that for each product structure

either the GLSPMS tailored Fix&Optimize or Truncated MIP performs well. Note that the

tailored Fix&Optimize performed very bad for problem 17. That is why the corresponding

curve cannot be seen in Figure B.2.

!"#$

!"#$

%&'()

&*+,-.

DOI 10.1007/978-3-658-02089-7, © Springer Fachmedien Wiesbaden 2013

172 Appendix B. Computational results of the heuristics

!"

# $%&'

$()*+,

!"#

!"#

$%&'(

%)*+,-

Bibliography

Afentakis, P.; Gavish, B.; Karmarkar, U. (1984) Computationally eﬃcient optimal solutions

to the lot-sizing problem in multistage assembly systems, Management Science, vol. 30,

no. 2, 222–239.

lot-sizing with sequence-dependent setups, International Journal of Production Research,

vol. 48, no. 3, 861–878.

Almada-Lobo, B.; Oliveira, J.; Carravilla, M. (2008) Production planning and scheduling

in the glass container industry: A VNS approach, International Journal of Production

Economics, vol. 114, no. 1, 363–375.

Almeder, C. (2010) A hybrid optimization approach for multi-level capacitated lot-sizing prob-

lems, European Journal of Operational Research, vol. 200, no. 2, 599–606.

machine lotsizing problem, International Journal of Production Research, vol. 49, no. 24,

7315–7335.

Amorim, P.; Meyr, H.; Almeder, C.; Almada-Lobo, B. (2011) Managing perishability in

production-distribution planning: a discussion and review, Flexible Services and Manufac-

turing Journal, 1–25, http://dx.doi.org/10.1007/s10696-011-9122-3.

Anderson, E.; Fitzsimons, G.; Simester, D. (2006) Measuring and mitigating the costs of

stockouts, Management Science, vol. 52, no. 11, 1751–1763.

Anthony, R. (1965) Planning and control systems: A framework for analysis, Harvard Uni-

versity, Boston.

Applegate, D.; Bixby, R.; Chvátal, V.; Cook, W. (2006) The Traveling Salesman Problem -

A Computational Study, Princeton University Press.

Araujo, S.; Arenales, M.; Clark, A. (2008) Lot sizing and furnace scheduling in small

foundries, Computers & Operations Research, vol. 35, no. 3, 916–932.

DOI 10.1007/978-3-658-02089-7, © Springer Fachmedien Wiesbaden 2013

174 Bibliography

Barany, I.; van Roy, T.; Wolsey, L. (1984) Strong formulations for multi-item capacitated lot

sizing, Management Science, vol. 30, no. 10, 1255–1261.

Ben-Daya, M.; Darwish, M.; Ertogral, K. (2008) The joint economic lot sizing problem:

Review and extensions, European Journal of Operational Research, vol. 185, no. 2, 726–

742.

Bolander, S.; Taylor, S. (1990) Process ﬂow scheduling: mixed ﬂow cases, Production and

Inventory Management Journal, vol. 31, no. 4, 1–6.

23.02.2013.

Chandra, P.; Fisher, M. (1994) Coordination of production and distribution planning, Euro-

pean Journal of Operational Research, vol. 72, no. 3, 503 – 517.

ing with setup and dismounting times, Computers & Operations Research, vol. 26, 715–730.

Christopher, M. (2005) Logistics and supply chain management, creating value-adding net-

works, Financial Times Prentice Hall, Harlow, 3rd ed.

Clark, A.; Almada-Lobo, B.; Almeder, C. (2011) Lot sizing and scheduling: industrial exten-

sions and research opportunities, International Journal of Production Research, vol. 49,

no. 9, 2457–2461.

Clark, A.; Scarf, H. (1960) Optimal policies for a multi-echelon inventory problem, Manage-

ment Science, vol. 6, no. 4, 475–490.

23.02.2013.

vol. 6, no. 6, 791–812.

Daskin, M. (1995) Network and Discrete Location: Models, Wiley & Sons.

scheduling problem, European Journal of Operational Research, vol. 79, no. 1, 25–37.

Dempster, M.; Fisher, M.; Jansen, L.; Lageweg, B.; Lenstra, J.; Rinnooy Kan, A. (1981)

Analytical Evaluation of Hierarchical Planning Systems, Operations Research, vol. 29,

no. 4, 707–716.

Denizel, M.; Altekin, F.; Süral, H.; Stadtler, H. (2008) Equivalence of the LP relaxations

of two strong formulations for the capacitated lot-sizing problem with setup times, OR

Spectrum, vol. 30, no. 4, 773–785.

Bibliography 175

Denizel, M.; Süral, H. (2006) On alternative mixed integer programming formulations and

LP-based heuristics for lot-sizing with setup times, Journal of the Operational Research

Society, vol. 57, no. 4, 389–399.

Dixon, P.; Silver, E. (1981) A Heuristic Solution Procedure for the Multi-Item, Single-Level,

Limited Capacity, Lot-Sizing Problem, Journal of Operations Management, vol. 2, no. 1,

23–39.

Drexl, A.; Fleischmann, B.; Günther, H.-O.; Stadtler, H.; Tempelmeier, H. (1994)

Konzeptionelle Grundlagen kapazitätsorientierter PPS-Systeme, Zeitschrift für betrieb-

swirtschaftliche Forschung, vol. 46, no. 12, 1022–1045.

Drexl, A.; Haase, K. (1995) Proportional lotsizing and scheduling, International Journal of

Production Economics, vol. 40, no. 1, 73–87.

Eppen, G.; Martin, R. (1987) Solving multi-item capacitated lot-sizing problems using vari-

able redeﬁnition, Operations Research, vol. 35, no. 6, 832–848.

Fandel, G.; Stammen–Hegener, C. (2006) Simultaneous lot sizing and scheduling for multi-

product multi-level production, International Journal of Production Economics, vol. 104,

no. 2, 308–316.

Felipe, .; Ortuño, M. T.; Tirado, G. (2009) The double traveling salesman problem with mul-

tiple stacks: A variable neighborhood search approach, Computers & Operations Research,

vol. 36, no. 11, 2983–2993.

Ferreira, D.; Morabito, R.; Rangel, S. (2009) Solution approaches for the soft drink integrated

production lot sizing and scheduling problem, European Journal of Operational Research,

vol. 196, no. 2, 697–706.

Fleischmann, B.; Meyr, H.; Wagner, M. (2008) Advanced Planning, in: H. Stadtler; C. Kilger

(Eds.) Supply Chain Management and Advanced Planning, Springer Berlin, 4th ed., 81–

106.

Fleischmann, B. (1990) The discrete lot–sizing and scheduling problem, European Journal of

Operational Research, vol. 44, no. 3, 337–348.

M. Speranza; P. Stähly (Eds.) Advances in Distribution Logistics, Lecture Notes in Eco-

nomics and Mathematical Systems, vol. 460, Springer, Berlin, 55–81.

Fleischmann, B.; Meyr, H. (1997) The general lotsizing and scheduling problem, OR Spec-

trum, vol. 19, no. 1, 11–21.

176 Bibliography

Fleischmann, B.; Meyr, H. (2003) Planning Hierarchy, Modeling and Advanced Planning

Systems, in: S. C. Graves; A. G. de Kok (Eds.) Supply Chain Management: Design,

Coordination, Operation, Handbooks in Operations Research and Management Science,

vol. 11, Elsevier, Amsterdam et al., 457–523.

Fleischmann, B.; Meyr, H. (2004) Customer Orientation in Advanced Planning Systems, in:

H. Dyckhoﬀ; R. Lackes; J. Reese (Eds.) Supply Chain Management and Reverse Logistics,

Springer Berlin, 297–321.

Florian, M.; Lenstra, J.; Kan, A. R. (1980) Deterministic Production Planning: Algorithms

and Complexity, Management Science, vol. 26, no. 7, 669–679.

Food and Agriculture Organization of the United Nations (2011) Annual Real Food Price

Indices.

Garey, M. R.; Johnson, D. S. (2003) Computers and Intractability, W.H. Freeman and Com-

pany, 24th ed.

23.02.2013.

Verlag, Aachen.

Haase, K. (1994) Lotsizing and scheduling for production planning, Springer, Berlin.

Haase, K. (1996) Capacitated lot–sizing with sequence dependent setup costs, OR Spectrum,

vol. 18, no. 1, 51–59.

Hansen, P.; Mladenović, N. (2001) Variable neighborhood search: Principles and applications,

European Journal of Operational Research, vol. 130, no. 3, 449–467.

Search, Journal of Heuristics, vol. 7, no. 4, 335–350.

Harris, F. (1913) How many parts to make at once, Factory, The Magazine of Management,

vol. 10, no. 2, 135–136,152.

Hax, A.; Meal, H. (1975) Hierarchical integration of production planning and scheduling, in:

M. A. Geisler (Ed.) Logistics: TIMS Studies in Management Sciences Vol. 1, Logistics:

TIMS Studies in Management Sciences, vol. 1, North-Holland, Amsterdam, 53–69.

Helber, S.; Sahling, F. (2010) A ﬁx-and-optimize approach for the multi-level capacitated lot

sizing problem, International Journal of Production Economics, vol. 123, no. 2, 247–256.

Bibliography 177

Jans, R.; Degraeve, Z. (2007) Meta-heuristics for dynamic lot sizing: A review and compar-

ison of solution approaches, European Journal of Operational Research, vol. 177, no. 3,

1855–1875.

Jenabi, M.; Ghomi, S. F.; Torabi, S. (2008) Finite horizon economic lot and delivery schedul-

ing problem: Flexible ﬂow lines with unrelated parallel machines and sequence dependent

setups., International Journal of Engineering, vol. 21, 143–158.

in: F. Keil; W. Mackens; H. Voß; J. Werther (Eds.) Scientiﬁc Computing in Chemical

Engineering, Springer, Berlin, 330–337.

Karmarkar, U.; Schrage, L. (1985) The Deterministic Dynamic Product Cycling Problem,

Operations Research, vol. 33, no. 2, 326–345.

Kilger, C.; Wetterauer, U. (2008) The Selection Process, in: H. Stadtler; C. Kilger (Eds.)

Supply Chain Management and Advanced Planning, Springer Berlin, 4th ed., 309–323.

Kimms, A. (1996) Multi-level, single-machine lot sizing and scheduling (with initial inven-

tory), European Journal of Operational Research, vol. 89, no. 1, 86–99.

Kimms, A. (1997) Multi–Level Lot Sizing and Scheduling, Physica Verlag, Heidelberg.

Krarup, J.; Bilde, O. (1977) Plant Location Set Covering and Economic Lot Sizing: An

O(mn) algorithm for structured Problems, in: L. Collatz (Ed.) Numerische Methoden

bei Optimierungsaufgaben, Bd. 3, Optimierung bei graphentheoretischen und ganzzahligen

Problemen, Birkhäuser, Basel, 155–180.

Kuik, R.; Salomon, M.; Wassenhove, L. N. V. (1994) Batch decisions: structure and models,

European Journal of Operational Research, vol. 75, no. 2, 243–263.

Kuik, R.; Salomon, M.; Wassenhove, L. V.; Maes, J. (1993) Linear programming, simu-

lated annealing and tabu search heuristics for lotsizing in bottleneck assembly systems, IIE

Transactions, vol. 25, no. 1, 62–72.

Lang, J. (2009) Production and inventory management with substitutions, Springer Verlag,

Heidelberg.

Lazić, J.; Hanaﬁ, S.; Mladenović, N.; Urošević, D. (2010) Variable neighbourhood decompo-

sition search for 0-1 mixed integer programs, Computers & Operations Research, vol. 37,

no. 6, 1055–1067.

Maes, J.; McClain, J.; van Wassenhove, L. (1991) Multilevel capacitated lotsizing complexity

and LP-based heuristics, European Journal of the Operational Research, vol. 53, no. 2,

131–148.

178 Bibliography

Maes, J.; van Wassenhove, L. (1988) Multi-item single-level capacitated dynamic lot-sizing

heuristics: a general review, Journal of the Operational Research Society, vol. 39, no. 11,

991–1004.

Marinelli, F.; Nenni, M.; Sforza, A. (2007) Capacitated lot sizing and scheduling with parallel

machines and shared buﬀers: A case study in a packaging company, Annals of Operations

Research, vol. 150, no. 1, 177–192.

Menezes, A.; Clark, A.; Almada-Lobo, B. (2011) Capacitated lot-sizing and scheduling with

sequence-dependent, period-overlapping and non-triangular setups, Journal of Scheduling,

vol. 14, no. 2, 209–219.

Meredith, J.; Shafer, S. (2010) Operations Management, International Student Version, Wi-

ley & Sons, 4th ed.

Meredith, L.; Maki, D. (2001) Product cannibalization and the role of prices, Applied Eco-

nomics, vol. 33, no. 14, 1785–1793.

tionslinien, Deutscher Universitätsverlag, Wiesbaden.

Meyr, H. (2000) Simultaneous lotsizing and scheduling by combining local search with dual

reoptimization, European Journal of Operational Research, vol. 120, no. 2, 311–326.

licher Fertigung, Zeitschrift für Betriebswirtschaft, vol. 74, no. 6, 585–610.

Meyr, H.; Stadtler, H. (2008) Types of Supply Chains, in: H. Stadtler; C. Kilger (Eds.)

Supply Chain Management and Advanced Planning, Springer Berlin, 4th ed., 65–80.

Meyr, H.; Wagner, M.; Rohde, J. (2008) Structure of Advanced Planning Systems, in:

H. Stadtler; C. Kilger (Eds.) Supply Chain Management and Advanced Planning, Springer

Berlin, 4th ed., 109–115.

Miller, C.; Tucker, A.; Zemlin, R. (1960) Integer Programming Formulation of Traveling

Salesman Problems, Journal of the ACM, vol. 7, no. 4, 326–329.

bench.html, 24.09.2012.

Mladenović, N.; Dražić, M.; Kovačevic-Vujčić, V.; Čangalović, M. (2008) General variable

neighborhood search for the continuous optimization, European Journal of Operational

Research, vol. 191, no. 3, 753–770.

Mladenović, N.; Hansen, P. (1997) Variable Neighborhood Search, Computers & Operations

Research, vol. 24, no. 11, 1097–1100.

Bibliography 179

Mohammadi, M.; Ghomi, S. F.; Karimi, B.; Torabi, S. (2009) Development of Heuristics for

Multi-Product Multi-Level Capacitated Lotsizing Problem with Sequence-Dependent Setups,

Journal of Applied Sciences, vol. 9, no. 2, 296–303.

Naji-Azimi, Z.; Salari, M.; Golden, B.; Raghavan, S.; Toth, P. (2010) Variable neighborhood

search for the cost constrained minimum label spanning tree and label constrained minimum

spanning tree problems, Computers & Operations Research, vol. 37, no. 11, 1952–1964.

Persson, J.; Göthe-Lundgren, M.; Lundgren, J.; Gendron, B. (2004) A tabu search heuristic

for scheduling the production processes at an oil reﬁnery, International Journal of Produc-

tion Research, vol. 42, no. 3, 445–471.

Pochet, Y.; Wolsey, L. (1995) Algorithms and reformulations for lot sizing problems, in:

W. Cook; L. Lovasz; P. Seymour (Eds.) Combinatorial Optimization, DIMACS Series in

Discrete Mathematics and Theoretical Computer Science, vol. 20, American Mathematical

Society, Providence, 245–293.

Pochet, Y.; Wolsey, L. (2006) Production Planning by Mixed Integer Programming, Springer

Series in Operations Research and Financial Engineering, Springer New York, ISBN 978-

0-387-29959-4 (Print) 978-0-387-33477-6 (Online).

Prandtstetter, M.; Raidl, G. (2008) An integer linear programming approach and a hybrid

variable neighborhood search for the car sequencing problem, European Journal of Opera-

tional Research, vol. 191, no. 3, 1004–1022.

D. Hahn (Ed.) Führungsprobleme industrieller Unternehmungen, de Gruyter, Berlin, 453–

463.

Puchinger, J.; Raidl, G. (2005) Combining Metaheuristics and Exact Algorithms in Combina-

torial Optimization: A Survey and Classiﬁcation, in: J. Mira; J. Álvarez (Eds.) Artiﬁcial

Intelligence and Knowledge Engineering Applications: A Bioinspired Approach, Lecture

Notes in Computer Science, vol. 3562, Springer Berlin / Heidelberg, 113–124.

Quadt, D. (2004) Lot-Sizing and Scheduling for Flexible Flow Lines, Springer Verlag, Hei-

delberg.

180 Bibliography

Quadt, D.; Kuhn, H. (2007) A taxonomy of ﬂexible ﬂow line scheduling procedures, European

Journal of Operational Research, vol. 178, no. 3, 686–698.

Ribeiro, C.; Souza, M. (2002) Variable neighborhood search for the degree-constrained mini-

mum spannign tree problem, Discrete Applied Mathematics, vol. 118, no. 1, 43–54.

Rohde, J.; Meyr, H.; Wagner, M. (2000) Die Supply Chain Planning Matrix, PPS Manage-

ment, vol. 5, no. 1, 10–15.

Rosling, K. (1986) Optimal lot-sizing for dynamic assembly systems, in: S. Axsäter;

C. Schneeweiß; E. Silver (Eds.) Multi-stage Production Planning and Inventory Control,

Springer, Berlin, 119–131.

Ruiz, R.; Rodríguez, J. (2010) The hybrid ﬂow shop scheduling problem, European Journal

of Operational Research, vol. 205, no. 1, 1–18.

Salomon, M. (1991) Deterministic Lotsizing Models for Production Planning, Springer Ver-

lag, Berlin.

Schneeweiss, C. (2003) Distributed Decision Making, Springer, Berlin Heidelberg New York,

2nd ed.

Seeanner, F.; Almada-Lobo, B.; Meyr, H. (2013) Combining the principles of Variable

Neighborhood Decomposition Search and the Fix&Optimize heuristic to solve multi-level

lot-sizing and scheduling problems, Computers & Operations Research, vol. 40, no. 1,

303–317.

Seeanner, F.; Meyr, H. (2013) Multi-stage simultaneous lot-sizing and scheduling for ﬂow

line production, OR Spectrum, vol. 35, no. 1, 33–73.

Sethi, S.; Thompson, G. (2005) Optimal Control Theory: Applications to Management Sci-

ence and Economics, Springer US.

Sikora, R.; Chhajed, D.; Shaw, M. (1996) Integrating the lot–sizing and sequencing decisions

for scheduling a capacitated ﬂow line, Computers and Industrial Engineering, vol. 30, no. 4,

659–679.

Silver, E.; Pyke, D.; Peterson, R. (1998) Inventory management and production planning

and scheduling, Wiley & Sons, New York et al., 3rd ed.

into demand forecasting, Marketing Intelligence & Planning, vol. 23, no. 5, 470–485.

Bibliography 181

Stadtler, H. (1996) Mixed integer programming model formulations for dynamic multi-item

multi-level capacitated lotsizing, European Journal of Operational Research, vol. 94, no. 3,

561–581.

Stadtler, H. (1997) Reformulations of the shortest route model for dynamic multi-item multi-

level capacitated lotsizing, OR Spectrum, vol. 19, no. 2, 87–96.

Stadtler, H. (2000) Improved Rolling Schedules for the Dynamic Single-Level Lot-Sizing Prob-

lem, Management Science, vol. 46, no. 2, 318–326.

Stadtler, H. (2003) Multilevel Lot Sizing with Setup Times and Multiple Constrained Re-

sources: Internally Rolling Schedules with Lot-Sizing Windows, Operations Research,

vol. 51, no. 3, 487–502.

Stadtler, H. (2008) Linear and Mixed Integer Programming, in: H. Stadtler; C. Kilger (Eds.)

Supply Chain Management and Advanced Planning, Springer Berlin, 4th ed., 518–527.

Stadtler, H. (2011) Multi-Level Single Machine Lot-Sizing and Scheduling with Zero Lead

Times, European Journal of Operational Research, vol. 209, no. 3, 241–252.

Stadtler, H.; Kilger, C. (Eds.) (2008) Supply Chain Management and Advanced Planning,

Springer Berlin, 4th ed.

Stadtler, H.; Sahling, F. (2013) A lot-sizing and scheduling model for multi-stage ﬂow lines

with zero lead times, European Journal of Operational Research, vol. 225, no. 3, 404–419.

Journal of Production Research, vol. 43, no. 1, 49–66.

Suerie, C. (2005b) Time Continuity in Discrete Time Models: New Approaches for Produc-

tion Planning in Process Industries, Springer Verlag, Heidelberg.

Suerie, C. (2006) Modeling of period overlapping setup times, European Journal of Opera-

tional Research, vol. 174, no. 2, 874–886.

Suerie, C.; Wagner, M. (2008) Supply Chain Analysis, in: H. Stadtler; C. Kilger (Eds.)

Supply Chain Management and Advanced Planning, Springer Berlin, 4th ed., 37–63.

10.0.

Tempelmeier, H.; Buschkühl, L. (2008) Dynamic multi-machine lotsizing and sequencing with

simultaneous scheduling of a common setup resource, International Journal of Production

Economics, vol. 113, 401–412.

182 Bibliography

multiitem constrained lotsizing with setup times, Management Science, vol. 42, no. 5, 738–

757.

Tempelmeier, H.; Herpers, S. (2011) Dynamic uncapacitated lot sizing with random demand

under a ﬁllrate constraint, European Journal of Operational Research, vol. 212, no. 3,

497–507.

Toledo, C.; França, P.; Morabito, R.; Kimms, A. (2007) Um modelo de otimização para o

problema integrado de dimensionamento de lotes e programação da produção em fábricas

de refrigerantes, Pesquisa Operacional, vol. 27, no. 1, 155–186.

Transchel, S.; Minner, S.; Kallrath, J.; Löhndorf, N.; Eberhard, U. (2011) A hybrid general

lot-sizing and scheduling formulation for a production process with a two-stage product

structure, International Journal of Production Research, vol. 49, no. 9, 2463–2480.

Trigeiro, W.; Thomas, L.; McClain, J. (1989) Capacitated lot sizing with setup times, Man-

agement Science, vol. 35, no. 3, 353–366.

Vollmann, T.; Berry, W.; Whybark, D.; Jacobs, F. (2005) Manufacturing Planning and

Control for Supply Chain Management, Irwin/McGraw-Hill, New York, 5th ed.

Williams, H. (Ed.) (1999) Model Building in Mathematical Programming, Wiley & Sons, 4th

ed.

Winands, E.; Adan, I.; van Houtum, G. (2011) The stochastic economic lot scheduling prob-

lem: A survey, European Journal of Operational Research, vol. 210, no. 1, 1–9.

Wolsey, L. (1997) MIP modelling of changeovers in production planning and scheduling prob-

lems, European Journal of Operational Research, vol. 99, no. 1, 154–165.

Zäpfel, G.; Braune, R.; Blögl, M. (Eds.) (2010) Metaheuristic Search Concepts: A Tutorial

with Applications to Production and Logistics, Springer Berlin Heidelberg.

- Koordination_im_reverse_logistic_ Eberhard_SchmidKoordination_im_reverse_logistic_ Eberhard_SchmidYuda
- (Produktion Und Logistik) Xin Wang (Auth.) - Operational Transportation Planning of Modern Freight Forwarding Companies_ Vehicle Routing Under Consideration of Subcontracting and Request Exchange-Gabl(Produktion Und Logistik) Xin Wang (Auth.) - Operational Transportation Planning of Modern Freight Forwarding Companies_ Vehicle Routing Under Consideration of Subcontracting and Request Exchange-GablSTEFANUS GUNAWAN
- Development and Evaluation of Setup Strategies in Printed Circuit Board AssemblyDevelopment and Evaluation of Setup Strategies in Printed Circuit Board AssemblyOmar Gaxiola
- Gasmesstechnik in Theorie und Praxis: Messgeräte, Sensoren, AnwendungenGasmesstechnik in Theorie und Praxis: Messgeräte, Sensoren, AnwendungenGerhard Wiegleb
- Dr. Bodo Runzheimer, Dr. Thomas Cleff, Dr. Wolfgang Schäfer (auth.) - Operations Research 1_ Lineare Planungsrechnung und Netzplantechnik-Gabler Verlag (2005).pdfDr. Bodo Runzheimer, Dr. Thomas Cleff, Dr. Wolfgang Schäfer (auth.) - Operations Research 1_ Lineare Planungsrechnung und Netzplantechnik-Gabler Verlag (2005).pdfPaulius Volonsevičius
- Gateway Management: Qualitätscontrolling in der Produktentwicklung der Zulieferindustrie: Assessment-Leitfaden für Gateway SystemeGateway Management: Qualitätscontrolling in der Produktentwicklung der Zulieferindustrie: Assessment-Leitfaden für Gateway SystemePhilipp Peters
- Produktlebenszyklus und Bestandsmanagement in der Halbleiterindustrie: Eine Analyse spezifischer HalbleiterprodukteProduktlebenszyklus und Bestandsmanagement in der Halbleiterindustrie: Eine Analyse spezifischer HalbleiterprodukteMarkus Walluschnig
- 1marz l Krug w Rose o Weigert g Hrsg Simulation Und Optimieru1marz l Krug w Rose o Weigert g Hrsg Simulation Und OptimieruLana
- Der Produktionsbetrieb 2_[Springer-Lehrbuch] Prof. Dr. h.c. mult. Dr.-Ing. Hans-Jürgen Warnecke (auth.) -Produktion, ProduDer Produktionsbetrieb 2_[Springer-Lehrbuch] Prof. Dr. h.c. mult. Dr.-Ing. Hans-Jürgen Warnecke (auth.) -Produktion, ProduAttila Gündoğan
- Curele dintate Handbuch.Zahnriementechnik.pdfCurele dintate Handbuch.Zahnriementechnik.pdfVasile Ermolai
- Aufbau und Ablauf einer IT-Integration unter Berucksichtigung spezifischer rechtlicher Aspekte PhasenmodellAufbau und Ablauf einer IT-Integration unter Berucksichtigung spezifischer rechtlicher Aspekte Phasenmodellaggrkrov
- 978-3-7315-0202-9978-3-7315-0202-9love
- Integration Des Step-Produktmodelles in Den GetriebeentwicklungIntegration Des Step-Produktmodelles in Den GetriebeentwicklungAdrian García Marcos
- Brennstoffzellen und Virtuelle Kraftwerke Energie-, umwelt- und technologiepolitische Aspekte einer effizienten... (Bert Droste-Franke, Holger Berg, Annette Kötter etc.) (z-lib.org)Brennstoffzellen und Virtuelle Kraftwerke Energie-, umwelt- und technologiepolitische Aspekte einer effizienten... (Bert Droste-Franke, Holger Berg, Annette Kötter etc.) (z-lib.org)Emre İnan
- Generisches Prozessebenenmodell (PrEMo) zwischen Strategie und technologiegestützten Maßnahmen: Systematische Darstellung für die praktische AnwendungGenerisches Prozessebenenmodell (PrEMo) zwischen Strategie und technologiegestützten Maßnahmen: Systematische Darstellung für die praktische AnwendungFranz Ringswirth
- ropers2017ropers2017Yuzar Stuff
- Employer Branding Als Bestandteil Einer Ganzheitlichen Markenfuhrung by Birgit Sponheuer (Z-lib.org)Employer Branding Als Bestandteil Einer Ganzheitlichen Markenfuhrung by Birgit Sponheuer (Z-lib.org)naomi
- Ra KitschRa Kitschvprochocky
- r18-018_obsoleszenzr18-018_obsoleszenzralf.stein9265
- Logistische Kennlinien: Grundlagen, Werkzeuge und AnwendungenLogistische Kennlinien: Grundlagen, Werkzeuge und AnwendungenPeter Nyhuis
- 1975.pdf1975.pdfcostea0028
- Dubbel Taschenbuch für den Maschinenbau 3: Maschinen und SystemeDubbel Taschenbuch für den Maschinenbau 3: Maschinen und SystemeCSPacademic
- Dynamic Fleet Management for International Truck Transportation Focusing on Occasional Transportation Tasks by Steffen Schorpp (auth.) (z-lib.org)Dynamic Fleet Management for International Truck Transportation Focusing on Occasional Transportation Tasks by Steffen Schorpp (auth.) (z-lib.org)Tahir
- Entwicklung einer Architektur für das Accounting in dynamischen Virtuellen OrganisationenEntwicklung einer Architektur für das Accounting in dynamischen Virtuellen OrganisationenMario Golling
- Betriebsfestigkeit VDI BuchBetriebsfestigkeit VDI Buchdoan quang
- Prof. Dr. Jürgen Bloech (Auth.), Prof. Dr. Jörg Biethahn, Prof. Dr. Jürgen Bloech, Dr. Ronald Bogaschewsky, Dipl. -Kfm. Uwe Hoppe (Eds.) - Wissensbasierte Systeme in Der Wirtschaft 1991_ Anwendungen uProf. Dr. Jürgen Bloech (Auth.), Prof. Dr. Jörg Biethahn, Prof. Dr. Jürgen Bloech, Dr. Ronald Bogaschewsky, Dipl. -Kfm. Uwe Hoppe (Eds.) - Wissensbasierte Systeme in Der Wirtschaft 1991_ Anwendungen umustaphar zohayr
- LTE OptimizationLTE OptimizationH_H2014
- Fabrikplanung und Fabrikbetrieb: Methoden für die wandlungsfähige, vernetzte und ressourceneffiziente FabrikFabrikplanung und Fabrikbetrieb: Methoden für die wandlungsfähige, vernetzte und ressourceneffiziente FabrikMichael Schenk
- Kostenorientiertes QualitätscontrollingKostenorientiertes QualitätscontrollingJamal Ismail
- Leonid Jasvoin (Auth.) - Integration Der Unsicherheitsaspekte in Die Schedule-Optimierung_ Empirische Modellierung Unter Anwendung Der Fuzzy-Theorie Am Beispiel Des Luftverkehrs-Deutscher UniversitätsLeonid Jasvoin (Auth.) - Integration Der Unsicherheitsaspekte in Die Schedule-Optimierung_ Empirische Modellierung Unter Anwendung Der Fuzzy-Theorie Am Beispiel Des Luftverkehrs-Deutscher Universitätsinappi loris
- Produktdatenmanagement – Anforderungen und Lösungen: Konzeption, Auswahl, Installation und Administration von PDM-SystemenProduktdatenmanagement – Anforderungen und Lösungen: Konzeption, Auswahl, Installation und Administration von PDM-SystemenThomas Mechlinski
- Dubbel Taschenbuch für den Maschinenbau 2: AnwendungenDubbel Taschenbuch für den Maschinenbau 2: AnwendungenCSPacademic
- Elektronische Rechnungseingangsbearbeitung: Eine Make-or-Buy-StudieElektronische Rechnungseingangsbearbeitung: Eine Make-or-Buy-StudieTim Wilczek
- Entwicklung und empirische Bestätigung eines Selbstbewertungsmodells für das Quality Gates ManagementEntwicklung und empirische Bestätigung eines Selbstbewertungsmodells für das Quality Gates ManagementPhilipp Peters
- Cloud Logistics by Falco Jaekel (z-lib.org)Cloud Logistics by Falco Jaekel (z-lib.org)Avelino Taipe
- Miniplant-Technik: in der ProzessindustrieMiniplant-Technik: in der ProzessindustrieWiley
- Arbeitsmethoden in der organischen Chemie: Mit EinführungsprogrammArbeitsmethoden in der organischen Chemie: Mit EinführungsprogrammSiegfried Hünig
- Geschäftsprozesse im Projektmanagement: Best Practices der ImplementierungGeschäftsprozesse im Projektmanagement: Best Practices der ImplementierungGerrit Buchenau
- Rudolf Large - Strategisches-BeschaffungsmanagementRudolf Large - Strategisches-BeschaffungsmanagementToca Disco
- Patentbasierte Methoden und lnstrumente: Einsatzmöglichkeiten im Technologie- und lnnovationsmanagementPatentbasierte Methoden und lnstrumente: Einsatzmöglichkeiten im Technologie- und lnnovationsmanagementAndreas Fraas
- Strategisches Informationsmanagement in Großprojekten der IndustrieStrategisches Informationsmanagement in Großprojekten der IndustrieGunter Gemmel
- Business Process Management: Die Bedeutung einer End-to-End Prozesssicht für die ökologische NachhaltigkeitBusiness Process Management: Die Bedeutung einer End-to-End Prozesssicht für die ökologische NachhaltigkeitKai Severini
- MPSCAI-2010 Thomas Bindel.pdfMPSCAI-2010 Thomas Bindel.pdfiuliaaaa24
- Zirkler .Führungsorientiertes US-Amerikanisches Management AccountingZirkler .Führungsorientiertes US-Amerikanisches Management AccountingJamal Ismail
- Ochoa SilviaOchoa Silviashoemakerlevy9
- Untersuchungen zur Fresstragfähigkeit von Groß-, Schräg- und Hochverzahnungen - HepermannUntersuchungen zur Fresstragfähigkeit von Groß-, Schräg- und Hochverzahnungen - Hepermannjkas
- Mapping Interface Management to Project Management (OCR)Mapping Interface Management to Project Management (OCR)nghiasup
- Lieferantenmanagement in Der Automobilindustrie Struktur Und Entwicklung Der Lieferantenbeziehungen Von Automobilherstellern by Johannes E. DölleLieferantenmanagement in Der Automobilindustrie Struktur Und Entwicklung Der Lieferantenbeziehungen Von Automobilherstellern by Johannes E. DölleJordan Rick Processes
- Akademisierung in der Pflege: Aktueller Stand und ZukunftsperspektivenAkademisierung in der Pflege: Aktueller Stand und ZukunftsperspektivenNicola Hundt
- Handbuch_Verkehrsunfallrekonstruktion.pdfHandbuch_Verkehrsunfallrekonstruktion.pdfVasicSasa
- [Wolfgang_Siebenpfeiffer_(eds.)]_Heavy-Duty-,_On-_(z-lib.org).pdf[Wolfgang_Siebenpfeiffer_(eds.)]_Heavy-Duty-,_On-_(z-lib.org).pdframzi
- Hartmut Brinkmeyer Auth. Drivers of Bank Lending New Evidence From the CrisisHartmut Brinkmeyer Auth. Drivers of Bank Lending New Evidence From the Crisisvadavada
- Einkauf und Supply Chain ManagementEinkauf und Supply Chain ManagementCSPacademic
- Empirische Analyse von preis- und verfügbarkeitsbeeinflussenden Indikatoren unter Berücksichtigung der Kritikalität von RohstoffenEmpirische Analyse von preis- und verfügbarkeitsbeeinflussenden Indikatoren unter Berücksichtigung der Kritikalität von RohstoffenBenjamin Achzet
- (Energiepolitik und Klimaschutz. Energy Policy and Climate Protection) Sevinj Amirova‐Mammadova (auth.) - Pipeline Politics and Natural Gas Supply from Azerbaijan to Europe_ Challenges and Perspectiv.pdf(Energiepolitik und Klimaschutz. Energy Policy and Climate Protection) Sevinj Amirova‐Mammadova (auth.) - Pipeline Politics and Natural Gas Supply from Azerbaijan to Europe_ Challenges and Perspectiv.pdfaramis_crosnier
- Determination of the Transfer Function of Power Transformer_IEH_Jahresbericht_2009Determination of the Transfer Function of Power Transformer_IEH_Jahresbericht_2009Parantap Raha
- Six Sigma (Aleman)Six Sigma (Aleman)César De La Cruz Oliva
- 10.1007%2F978-3-663-08987-210.1007%2F978-3-663-08987-2Nam Nhito
- torpoastorpoasKaterina Serenici
- Smart Clothes - Textilien mit Elektronik: Was bietet der Markt der Intelligenten Bekleidung?Smart Clothes - Textilien mit Elektronik: Was bietet der Markt der Intelligenten Bekleidung?Sina Carvajal Vargas