0 visualizações

Enviado por Gerson Schaffer

LOt Sizing Pesquisa Operacional

- Set Theory
- Operation Research
- An Introduction to Optimization Chong Solution Manual PDF
- Reference Material I MEE437 or T Module 2 WS2017 18 Tutorial Problem
- MATH-162
- A Survey of Modern Data Classification Techniques
- Model Predictive Control
- Centrifugal Pump
- Vector Expansion Technique
- Robust design optimisation .pdf
- Undergraduate 14[1]
- DSP-sharif university
- Literature Survey for Face Detection under Illumination Variation
- iberamia02pp69.pdf
- Imperialist Competitive Algorithm
- Integrated Framework of Optimization Technique and Information Theory Measures for Modeling Neutrosophic Variables
- 1.pdf
- threedetailedexamplesofsystemsofequations
- KOMAT---ARRAY & LINSPACE.docx
- notes le3

Você está na página 1de 240

Lisbeth Buschkühl

Multi-Level

Capacitated Lotsizing

with Setup Carryover

Edition KWV

Die „Edition KWV“ beinhaltet hochwertige Werke aus dem Bereich der Wirtschaftswissen-

schaften. Alle Werke in der Reihe erschienen ursprünglich im Kölner Wissenschaftsverlag,

dessen Programm Springer Gabler 2018 übernommen hat.

Lisbeth Buschkühl

Multi-Level

Capacitated Lotsizing

with Setup Carryover

Lisbeth Buschkühl

Wiesbaden, Germany

Dissertation Universität zu Köln, 2008

Edition KWV

ISBN 978-3-658-24033-2 ISBN 978-3-658-24034-9 (eBook)

https://doi.org/10.1007/978-3-658-24034-9

Springer Gabler

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2008, Reprint 2019

Originally published by Kölner Wissenschaftsverlag, Köln, 2008

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, computer software, or by similar or dissimilar methodology now known or

hereafter developed.

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.

The publisher, the authors, and the editors are safe to assume that the advice and information in this book are

believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors

give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions

that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps

and institutional affiliations.

This Springer Gabler imprint is published by the registered company Springer Fachmedien Wiesbaden GmbH

part of Springer Nature

The registered company address is: Abraham-Lincoln-Str. 46, 65189 Wiesbaden, Germany

Preface

Chain Management and Production at the University of Cologne. It was ac-

cepted as a dissertation in July 2008. I would like to seize this opportunity

to thank the numerous supporters, whose backing has rendered this work pos-

sible. First of all, I am deeply indepted to my advisor Prof. Dr. Horst

Tempelmeier for his ongoing guidance and support. His profound knowledge

and high standards made working at his department a challenging and enrich-

ing experience. Also, I would like to thank Prof. Dr. Ulrich W. Thonemann

for serving as a co-referee of this dissertation.

Not less important was the support of my colleagues, Lars Fischer, Sascha

Herpers and Dr. Michael Manitz. Although I was the only member of our team

living in the ‘deterministic world’, I could always count on their help, whenever

I needed it. This also holds true for our team assistant Manuela Pioch, whose

employment at our department has proven to be a great enrichment. My

special regards go to Christian Steffens and Florian Sahling for proofreading

the manuscript. Also, they are my living proof that it is not insane to think

the world deterministic. I would also like to thank my colleague Dr. Johannes

Antweiler for making me do the dissertation in the first place. His never ending

drive to get the best out of everything for the members of our department has

been essential to creating a most comfortable atmosphere for the rest of us.

Above all, I’d like to thank my parents for seeing me through my education

and for their love, their understanding and their support, which I could always

rely on. Finally, Robert, Bernd, Angela and Stephie, together the rest of my

family and of my friends, especially for their backing over the past months.

V

Contents

Preface V

List of Symbols XX

B Literature Review 7

B.1 Model Formulations . . . . . . . . . . . . . . . . . . . . . . . 8

B.1.1 The Single-Level Capacitated Lotsizing Problem . . . . 8

B.1.2 The Multi-Level Capacitated Lotsizing Problem . . . . 11

B.1.3 The Single-Level Capacitated Lotsizing Problem with

Linked Lotsizes . . . . . . . . . . . . . . . . . . . . . . 12

B.2 Complexity Considerations . . . . . . . . . . . . . . . . . . . 19

VII

VIII Multi-Level Capacitated Lotsizing with Setup Carryover

B.3.1 Mathematical Programming based Approaches . . . . . 23

B.3.1.1 Branch&Bound Heuristics . . . . . . . . . . . . . . 23

B.3.1.1.1 Reformulation . . . . . . . . . . . . . . . . . 24

B.3.1.1.2 Valid Inequalities . . . . . . . . . . . . . . . . 28

B.3.1.2 Fix&Relax Heuristics . . . . . . . . . . . . . . . . 29

B.3.1.3 Rounding Heuristics . . . . . . . . . . . . . . . . . 30

B.3.1.4 Linear Programming based Approaches . . . . . . 31

B.3.1.5 Dantzig-Wolfe Decomposition and Column Gener-

ation . . . . . . . . . . . . . . . . . . . . . . . . . . 32

B.3.2 Lagrangean Heuristics . . . . . . . . . . . . . . . . . . . 33

B.3.3 Decomposition and Aggregation Approaches . . . . . . . 40

B.3.4 Metaheuristics . . . . . . . . . . . . . . . . . . . . . . . 44

B.3.4.1 Local Search . . . . . . . . . . . . . . . . . . . . . 46

B.3.4.2 Simulated Annealing . . . . . . . . . . . . . . . . . 46

B.3.4.3 Tabu Search . . . . . . . . . . . . . . . . . . . . . 47

B.3.4.4 Variable Neighborhood Search . . . . . . . . . . . 48

B.3.4.5 Genetic Algorithms . . . . . . . . . . . . . . . . . 48

B.3.4.6 Ant Colony Optimization . . . . . . . . . . . . . . 49

B.3.5 Greedy Heuristics . . . . . . . . . . . . . . . . . . . . . 50

C Test Instances 55

C.1 Problem Sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

C.2 Product and Process Structures . . . . . . . . . . . . . . . . 56

C.3 Demand Profiles . . . . . . . . . . . . . . . . . . . . . . . . . 59

C.4 Setup Time Profiles . . . . . . . . . . . . . . . . . . . . . . . 60

C.5 Setup and Holding Cost Ratios . . . . . . . . . . . . . . . . . 61

C.6 Capacity Utilization Profiles . . . . . . . . . . . . . . . . . . . 62

Contents IX

D Model Formulations 65

D.1 The Multi-level Dynamic Capacitated Lotsizing Problem with

Linked Lotsizes . . . . . . . . . . . . . . . . . . . . . . . . . 66

D.1.1 Inventory & Lotsize Formulation . . . . . . . . . . . . . 68

D.1.2 Echelon Formulation . . . . . . . . . . . . . . . . . . . . 70

D.1.3 Shortest Route Formulation . . . . . . . . . . . . . . . . 75

D.1.4 Simple Plant Location Formulation . . . . . . . . . . . . 77

D.1.5 Computational Results . . . . . . . . . . . . . . . . . . 79

D.2 Consecutive Setup Carryovers . . . . . . . . . . . . . . . . . 82

D.2.1 Standard formulation . . . . . . . . . . . . . . . . . . . 82

D.2.2 Extended formulation . . . . . . . . . . . . . . . . . . . 83

D.2.3 Computational Results . . . . . . . . . . . . . . . . . . . 84

D.3 Parallel Machines . . . . . . . . . . . . . . . . . . . . . . . . 86

D.3.1 Model Formulation . . . . . . . . . . . . . . . . . . . . . 86

D.3.2 Computational Results . . . . . . . . . . . . . . . . . . . 88

E.1 Outline of the Procedure . . . . . . . . . . . . . . . . . . . . . 93

E.2 Computation of the Lower Bound . . . . . . . . . . . . . . . . 95

E.2.1 Single Setup Carryover . . . . . . . . . . . . . . . . . . 95

E.2.2 Consecutive Setup Carryover . . . . . . . . . . . . . . . 102

E.3 Lagrangean Multiplier Update . . . . . . . . . . . . . . . . . 104

E.4 Computation of the Upper Bound . . . . . . . . . . . . . . . 109

E.4.1 Inventory Balance Constraints . . . . . . . . . . . . . . 109

E.4.2 Setup Carryover Constraints . . . . . . . . . . . . . . . 111

E.4.2.1 Single Setup Carryover . . . . . . . . . . . . . . . 112

X Multi-Level Capacitated Lotsizing with Setup Carryover

E.4.3 Capacity Constraints . . . . . . . . . . . . . . . . . . . 114

E.4.3.1 Outline of the Finite Loading Heuristic . . . . . . 115

E.4.3.2 Order of Consideration of Shifts . . . . . . . . . . 117

E.4.3.3 Shifts . . . . . . . . . . . . . . . . . . . . . . . . . 118

E.4.3.3.1 Changes to the Setup Pattern in the Target

Period . . . . . . . . . . . . . . . . . . . . . 121

E.4.3.3.2 Changes to the Setup Pattern in the Original

Period . . . . . . . . . . . . . . . . . . . . . 122

E.4.3.3.3 Single-Item Forward Shifts . . . . . . . . . . 123

E.4.3.3.4 Multi-Item Forward Shifts . . . . . . . . . . . 126

E.4.3.3.5 Single-Item Backward Shifts . . . . . . . . . 129

E.4.3.3.6 Multi-Item Backward Shifts . . . . . . . . . 132

E.4.3.3.7 Priority Values . . . . . . . . . . . . . . . . 134

E.4.3.3.8 Tabu Lists . . . . . . . . . . . . . . . . . . . 136

E.4.4 Postoptimization . . . . . . . . . . . . . . . . . . . . . . 136

E.5 Adaptation for Parallel Machines . . . . . . . . . . . . . . . 138

E.5.1 Aggregate Model Formulation for Resource Types . . . 140

E.5.2 Computation of a Feasible Solution for Parallel Machines 145

E.5.2.1 Assignment of Production to Individual Resources 145

E.5.2.2 Feasibility Routine . . . . . . . . . . . . . . . . . . 149

E.5.2.2.1 Lot Splitting . . . . . . . . . . . . . . . . . . 149

E.5.2.2.2 Shifting of Setup Activities . . . . . . . . . . 151

E.5.2.2.3 Modified Finite Loading Heuristic . . . . . . 152

E.5.3 Postoptimization . . . . . . . . . . . . . . . . . . . . . . 154

F.1 Comparison of the Performance for Various Parameter Com-

binations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156

Contents XI

F.1.1.1 Parameters for the Smoothing of the Subgradients 157

F.1.1.2 Parameters for the Update of the Lagrangean Mul-

tipliers . . . . . . . . . . . . . . . . . . . . . . . . . 158

F.1.1.2.1 Parameters for Factor δ l . . . . . . . . . . . . 159

F.1.1.2.2 Parameters for the Estimate of the Optimal

Solution . . . . . . . . . . . . . . . . . . . . . 160

F.1.1.2.3 Parameters for the Updating Modi . . . . . . 161

F.1.2 Parameter for the Termination of the Finite Loading

Heuristic . . . . . . . . . . . . . . . . . . . . . . . . . . 162

F.1.3 Parameters for the Sorting of Resources . . . . . . . . . 162

F.1.4 Parameters for the Sorting of Resources and Periods . . 165

F.1.5 Parameters for the Shifts . . . . . . . . . . . . . . . . . 166

F.1.5.1 Parameters for the Backward Shifts . . . . . . . . 166

F.1.5.2 Parameters for the Multi-item Shifts . . . . . . . . 167

F.1.5.3 Parameters for Careful Shifts . . . . . . . . . . . . 168

F.1.6 Parameters for the Computation of Priority Values . . 169

F.1.7 Parameters for the Application of Tabu Lists . . . . . . 171

F.1.7.1 Parameters for the Tabu Status . . . . . . . . . . . 171

F.1.7.2 Parameters for the Aspiration Criterion . . . . . . 172

F.2 Dominant Parameter Configurations and Further Tests for Se-

lected Parameters Parameters . . . . . . . . . . . . . . . . . 173

F.2.1 Dominant Parameter Configurations . . . . . . . . . . . 173

F.2.2 Further tested Parameter Combinations . . . . . . . . . 175

F.2.3 Parameters for Parallel Machines . . . . . . . . . . . . . 177

F.3 Evaluation of the Overall Performance of the Heuristic . . . 180

F.3.1 Results for the MLCLSPL with Single Setup Carryover 183

F.3.2 Results for the MLCLSPL with Consecutive Setup Car-

ryover . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

XII Multi-Level Capacitated Lotsizing with Setup Carryover

F.3.3.1 Single Setup Carryover . . . . . . . . . . . . . . . 194

F.3.3.2 Consecutive Setup Carryover . . . . . . . . . . . . 198

Bibliography 205

List of Figures

C.2 Product and process structures for 20 items and 6 resources . 58

C.3 Product structures and process structures for 40 items and 6

resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

violation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

F.1 Development of the upper and the lower bound for a given

test instances . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

XIII

List of Tables

B.2 MP-based heuristics . . . . . . . . . . . . . . . . . . . . . . . 22

B.3 Lagrangean Heuristics . . . . . . . . . . . . . . . . . . . . . . 36

B.4 Decomposition and Aggregation Heuristics . . . . . . . . . . . 41

B.5 Metaheuristics . . . . . . . . . . . . . . . . . . . . . . . . . . 45

B.6 Greedy Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . 51

C.2 Mean demand . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

C.3 Setup time profiles . . . . . . . . . . . . . . . . . . . . . . . . 60

C.4 TBO profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

C.5 Capacity profiles . . . . . . . . . . . . . . . . . . . . . . . . . 64

C.6 Number of parallel machines . . . . . . . . . . . . . . . . . . 64

number of optimal solutions . . . . . . . . . . . . . . . . . . . 80

D.2 Comparison of the model formulations for single carryovers:

mipgap and computation time . . . . . . . . . . . . . . . . . . 81

D.3 Comparison of the model formulations for consecutive carry-

overs: number of optimal solutions . . . . . . . . . . . . . . . 85

XV

XVI Multi-Level Capacitated Lotsizing with Setup Carryover

overs: mipgap and computation time . . . . . . . . . . . . . . 86

D.5 Comparison of the model formulations for parallel machines:

number of optimal solutions . . . . . . . . . . . . . . . . . . . 89

D.6 Comparison of the model formulations for parallel machines:

mipgap and computation time . . . . . . . . . . . . . . . . . . 89

F.1 Tested values for the smoothing parameters of the subgradients 157

F.2 Results for the smoothing parameters of the subgradients . . 158

F.3 Tested Combinations for the Parameters for the Lagrangean

multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

F.4 Tested values for the stepsize update factor δ l . . . . . . . . . 159

l

F.5 Results for the stepsize update factor δ . . . . . . . . . . . . 160

F.6 Tested values for the estimate of the upper bound . . . . . . 160

F.7 Results for the estimate of the optimal solution . . . . . . . . 160

F.8 Tested values for the stepsize update: upper and lower bounds 161

F.9 Results for the stepsize update . . . . . . . . . . . . . . . . . 162

F.10 Tested values for the sorting of resources . . . . . . . . . . . . 163

F.11 Results for the sorting of resources . . . . . . . . . . . . . . . 164

F.12 Tested values for the order of consideration of resources and

periods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

F.13 Results for the order of consideration of resources and periods 165

F.14 Tested values for backward shifts . . . . . . . . . . . . . . . . 166

F.15 Results for backward shifts . . . . . . . . . . . . . . . . . . . 167

F.16 Tested values for multi-item shifts . . . . . . . . . . . . . . . 168

F.17 Results for multi-item shifts . . . . . . . . . . . . . . . . . . . 168

F.18 Tested values for the first iteration of careful shifts . . . . . . 169

F.19 Results for the first iteration of careful shifts . . . . . . . . . 169

F.20 Tested values for the computation of the priority values . . . 170

List of Tables XVII

F.22 Tested values for the tabu lists . . . . . . . . . . . . . . . . . 171

F.23 Results for the tabu lists . . . . . . . . . . . . . . . . . . . . . 172

F.24 Tested values for the aspiration criterion . . . . . . . . . . . . 172

F.25 Dominant configuration for the parameters for the Lagrangean

multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

F.26 Dominant Configuration for the Parameter for the termination

of the finite loading heuristic . . . . . . . . . . . . . . . . . . 174

F.27 Dominant configuration for the parameters for the ordering of

shifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

F.28 Dominant configuration for the parameters for the shifts . . . 174

F.29 Dominant configuration for the parameters for the application

of tabu lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

F.30 Tested values for the computation of the priority values . . . 176

F.31 Results for the maximum number of per TBO profile . . . . . 176

F.32 Dominant configuration for the parameters for the computa-

tion of priority values . . . . . . . . . . . . . . . . . . . . . . 177

F.33 Tested values for the setup cost modification - test 1 . . . . . 178

F.34 Results for the setup cost modification for the MLCLSPLP M

- test 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178

F.35 Tested values for the setup cost modification for the

MLCLSPLP M - test 2 . . . . . . . . . . . . . . . . . . . . . . 178

F.36 Results for the setup cost modification for the MLCLSPLP M

- test 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

F.37 Results for the setup cost modification for the

MLCLSPLP M CC - test 1 . . . . . . . . . . . . . . . . . . . . . 179

F.38 Tested values for the setup cost modification for the

MLCLSPLP M CC - test 2 . . . . . . . . . . . . . . . . . . . . . 180

F.39 Results for the setup cost modification for the

MLCLSPLP M CC - test 2 . . . . . . . . . . . . . . . . . . . . . 180

max

F.40 Tested values for the maximum number of iterations l for

the MLCLSPL without parallel machines . . . . . . . . . . . 182

XVIII Multi-Level Capacitated Lotsizing with Setup Carryover

F.41 Tested values for the maximum number of iterations lmax for

the MLCLSPL with parallel machines . . . . . . . . . . . . . 182

F.42 Number of feasible solutions for the MLCLSPLe . . . . . . . 183

F.43 Results for the MLCLSPLe . . . . . . . . . . . . . . . . . . . 184

F.44 Results for the MLCLSPLe per class . . . . . . . . . . . . . . 185

F.45 Results for the MLCLSPLe per capacity and setup time profile 186

F.46 Results for the MLCLSPLe per setup and holding cost ratio . 187

F.47 Results for the MLCLSPLe per product and process structure

and demand profile . . . . . . . . . . . . . . . . . . . . . . . . 188

F.48 Number of feasible solutions for the MLCLSPLCC . . . . . . 189

F.49 Results for the MLCLSPLCC . . . . . . . . . . . . . . . . . . 189

F.50 Results for the MLCLSPLCC per class . . . . . . . . . . . . . 190

F.51 Results for the MLCLSPLCC per capacity and setup time profile 191

F.52 Results for the MLCLSPLCC per setup and holding cost ratio 192

F.53 Results for the MLCLSPLCC per product and process struc-

ture and demand profile . . . . . . . . . . . . . . . . . . . . . 193

F.54 Number of feasible solutions for the MLCLSPLP M . . . . . . 194

F.55 Results for the MLCLSPLP M . . . . . . . . . . . . . . . . . 194

F.56 Results for the MLCLSPLP M per class . . . . . . . . . . . . . 195

F.57 Results for the MLCLSPLP M per capacity profile . . . . . . . 196

F.58 Results for the MLCLSPLP M per product and process struc-

ture and demand profile . . . . . . . . . . . . . . . . . . . . . 197

F.59 Number of feasible solutions for the MLCLSPLP M CC . . . . 198

F.60 Results for the MLCLSPLP M CC . . . . . . . . . . . . . . . . 198

F.61 Results for the MLCLSPLP M CC per class . . . . . . . . . . . 199

F.62 Results for the MLCLSPLP M CC per capacity profile . . . . . 200

F.63 Results for the MLCLSPLP M CC per product and process

structure and demand profile . . . . . . . . . . . . . . . . . . 201

List of Algorithms

2 Lagrangean heuristic . . . . . . . . . . . . . . . . . . . . . . . . 94

3 Computation of the lower bound . . . . . . . . . . . . . . . . . 102

4 Upper bound - inventory balance constraints . . . . . . . . . . 110

5 Upper bound - setup carryover constraints . . . . . . . . . . . 113

6 Upper bound - consecutive setup carryover constraints . . . . . 114

7 Upper bound - capacity constraints . . . . . . . . . . . . . . . 116

8 Upper bound - single-item forward shifts . . . . . . . . . . . . 124

9 Upper bound - multi-item forward shifts . . . . . . . . . . . . 127

10 Upper bound - single-item backward shifts . . . . . . . . . . . 131

11 Upper bound - multi-item backward shifts . . . . . . . . . . . 133

12 Iterative solution procedure for parallel machines . . . . . . . . 140

13 Assignment of production to parallel machines . . . . . . . . . 146

14 Feasibility routine . . . . . . . . . . . . . . . . . . . . . . . . . 150

XIX

List of Symbols

tipliers

Dn

δkts production quantity of item k on machine n in period t to satisfy

demand in period s

satisfied by production in period t

δkts fraction of demand of item k in period s, which is satisfied by produc-

tion in period t

D D

δkts demand of item k in period s, which is produced in period t (δkts =

δkts · Dks )

γkt binary setup variable

n

γkt binary setup variable on machine n

λlu stepsize for the update of the Lagrangean multiplier of the inventory

balance constraint for item k in period t at iteration l

λlv stepsize for the update of the Lagrangean multiplier of the capacity

constraint for resource m in period t at iteration l

λlw stepsize for the update of the Lagrangean multiplier of the setup car-

ryover constraint for resource m in period t at iteration l

Ω great number

ωkt binary setup state variable for item k at the beginning of period t

XXI

XXII Multi-Level Capacitated Lotsizing with Setup Carryover

n

ωkt binary setup state variable for item k on machine n at the beginning

of period t

Dk average demand of item k

Dmt̂t minimum average capacity consumption through production times on

resource m in periods t̂ to t

ςtml subgradient of the setup carryover constraint for resource m in period

t at iteration l

ϑm

t dummy variable to prohibit a changeover, when the setup state of the

same item k is carried over from period t − 1 to period t and from

period t to period t + 1

ϑkt dummy variable to indicate a consecutive carryover for item k from

period t − 1 to t and to t + 1

bbn available capacity of an individual resource n in the case of parallel

t

machines in period t

qbkt quantity to be assigned on parallel machines for item k in period t

ebm modified available capacity of resource type m in period t

t

ξtml subgradient of the capacity constraint for resource m in period t at

iteration l

l

ζkt subgradient of the inventory balance constraint for item k in period t

at iteration l

akj production coefficient (quantity of item k, which is directly needed to

produce one item j)

bm

t total capacity of resource m in period t in time units

dkts cumulated demand of item k in periods t to s

dkt external demand of item k in period t

ek echelon holding cost coefficient of item k

List of Symbols XXIII

K total number of items

k items, k = 1, . . . , K

Km set of items k that are produced on the same resource m

Lm cumulated lead time of resource m

M total number of non-identical resources

m non-identical resources, m = 1, . . . , M

n parallel machines, n = 1 . . . Pm

Nk set of direct successors of item k

Pm number of parallel machines of resource type m

s periods, s = 1, . . . , T

sk setup cost coefficient for item k

t production periods, t = 1, . . . , T

trk time needed to set up a resource for the production of item k

Vk set of direct predecessors of item k

wtm Lagrangean multiplier for the constraint of total number of setup car-

ryovers

List of Abbreviations

B&B Branch&Bound

B&C Branch&Cut

BO Backorders

C&B Cut&Branch

CC Consecutive setup carryover (over more than one adjacent pe-

riod transition)

CG Column Generation

CLSP Capacitated Lotsizing Problem

CLSPL Capacitated Lotsizing Problem with Linked Lotsizes

CSLP Continuous Setup Lotsizing Problem

DLSP Discrete Lotsizing and Scheduling Problem

DW Dantzig-Wolfe Decomposition

EA Evolutionary Algorithms

ELSP Economic Lot Scheduling Problem

EOQ Economic Order Quanitity

EP Evolutionary Programming

ES Evolutionary Strategies

F&R Fix&Relax

XXV

XXVI Multi-Level Capacitated Lotsizing with Setup Carryover

GA Genetic Algorithms

GH Greedy Heuristics

GLSP General Lotsizing and Scheduling Problem

I&L Inventory&Lotsize

ILS Iterated Local Search

LD Lagrangean Decomposition

LP Linear Programming

LR Lagrangean Relaxation

LR-MLCLSPL Lagrangean Relaxation of MLCLSPL

MA Memetic Algorithms

MIP Mixed Integer Program

MLCLSP Multi-Level Capacitated Lotsizing Problem

MLCLSPSP L Simple Plant Location Formulation of the Multi-Level Capaci-

tated Lotsizing Problem

MLCLSPSR Shortest Route Formulation of the Multi-Level Capacitated Lot-

sizing Problem

MLCLSPL Multi-Level Capacitated Lotsizing Problem with Linked

Lotsizes

MLCLSPLe Multi-Level Capacitated Lotsizing Problem with Linked Lot-

sizes with echelon holding costs

MLCLSPLCC Multi-Level Capacitated Lotsizing Problem with Linked Lot-

sizes and Consecutive Carryover

MLCLSPLPM Multi-Level Capacitated Lotsizing Problem with Linked Lot-

sizes and consecutive carryover on Parallel Resources

MLULSPL Multi-Level Uncapacitated Lotsizing Problem with Linked

Lotsizes

MRP Material Requirements Planning

MP Mathematical Programming

NFA Network Flow Algorithm

PLSP Proportional Lotsizing and Scheduling Problem

PM Parallel (identical) resources

List of Abbreviations XXVII

RH Rounding Heuristic

SA Simulated Annealing

SLULSP Single-Level Uncapacitated Lotsizing Problem (also referred to

as Wagner-Within ProblemWagner and Whitin (1958))

SLULSPL Single-Level Uncapacitated Lotsizing Problem with Linked

lotsizes

SP Set Partitioning

SPL Simple Plant Location

SPP Set Partitioning Problem

SR Shortest Route

TA Threshold Accepting

TBO Time between orders

TS Tabu Search

VI Valid Inequalities

VNS Variable Neighborhood Search

Chapter A

Lotsizing problems are found in every company, which produces various items

on a given set of resources with scarce capacities. In general the objective

is to find a production plan, which meets demand with the resources’ given

capacities and at minimal costs. However, the specific lotsizing problem is

highly dependent on the characteristics of the underlying production process.

Batch restrictions e. g. in chemical processes, multi-level product structures

and sequences-dependent setup times are only a few examples, which make it

necessary to tailor a unique solution approach to the specific problem.

are the two main reasons for the fact, that neither standard MRP-systems nor

modern Advanced Planning Systems provide adequate solutions. Hence, these

systems are not capable of generating feasible production plans.

Early theoretical work on lotsizing problems was done by Wagner and Whitin

(1958) and Manne (1958). Since then lotsizing has become a widely studied

field of research for the same reasons indicated above. First, the high level of

complexity makes it impossible to solve any lotsizing problem of realistic di-

mensions with an exact solution method. Furthermore, no dominant heuristic

L. Buschkühl, Multi-Level Capacitated Lotsizing with Setup Carryover,

Edition KWV, https://doi.org/10.1007/978-3-658-24034-9_1

2 Multi-Level Capacitated Lotsizing with Setup Carryover

solution approach to this type of problem has yet been found. This in turn has

lead to the development of a great variety of heuristics. Second, the particu-

lar production processes, demand structures and further restrictions faced by

different companies and industries have triggered the need for models, which

are respectively applicable to the specific production process.

The spectrum of models that have been developed ranges from those with a

continuous time scale, constant demand and an infinite time horizon to those

with a discrete time scale, dynamic demand and a finite time horizon. Famous

members of the former group are the Economic Order Quantity model (EOQ)

and the Economic Lot Scheduling Problem (ELSP).1

distinction can be made between small- and big-bucket formulations. The

names originate from the difference in period (or time-bucket) length. The

length of a period is not measured in time units but rather in the number of lots

that can be produced. For small-bucket models the maximum number of lots

per period is two, while there is no corresponding limit for big-bucket models.

Hence, while the sequence of lots within the period remains undetermined in

big-bucket models, it is clearly defined in small-bucket ones.2

(CLSP). The Multi-Level Capacitated Lotsizing Problem (MLCLSP) is the

multi-level extension, which accounts not only for external demand, but also

for the derived demand of components. The disadvantage of the CLSP is

that the setup state is not preserved from one period to the next. Even if

the same item is produced in two adjacent periods and production could be

continued in the second period without any further changeover, setups are

incurred twice. This can result in a false calculation of the overall production

costs and resource requirements and thus in a bad scheduling decision. Also,

if capacity is scarce, an existing feasible schedule may not be found.

2 Compare Sürie and Stadtler (2003).

Chapter A. Introduction and Outline 3

These include the Discrete Lotsizing and Scheduling Problem (DLSP) 3 , the

Continuous Setup Lotsizing Problem (CSLP)4 and the Proportional Lotsizing

and Scheduling Problem (PLSP)5 .

The PLSP is the most advanced small-bucket model, as it allows for up to two

items to be produced in one period. It is appropriate for the representation

of a continuous production, when setup times can be split across periods and

periods are chosen appropriately.6 However, holding costs and demand are

given for each time bucket. Thus, the bucket length influences the time and

demand structure and hence the optimal solution.

To overcome this drawback, Meyr (1999) extends the PLSP to the General Lot-

sizing and Scheduling Problem (GLSP). The GLSP works with two different

time schemes. First, the planning horizon is divided into so-called macro-

periods to compute demand, inventory and holding costs. Each macro-period

is in turn divided into micro-periods of variable length. They account for se-

tups and production. For more extensive reviews on small-bucket models see

e. g. Drexl and Kimms (1997) and Meyr (1999).

high level of complexity and less flexibility for the subsequent planning pro-

cesses.7 In big-bucket models lots can be moved within a period without

changing the overall schedule. Thus, they are more robust in the case of ran-

dom events than small-bucket ones.

A model which retains the advantages of the CLSP and incorporates aspects

of continuous production is the Capacitated Lotsizing Problem with Linked

Lotsizes (CLSPL). This big-bucket model was first tackled by Dillenberger

et al. (1993). It includes partial sequencing, i. e. the first and the last lot of

a period are determined. Therefore, a unique setup state can be identified

on each resource at the period transitions. This in turn implies, that at the

4 See Karmarkar and Schrage (1985).

5 See Drexl and Haase (1995).

6 See also Sürie (2005a).

7 Note however, that models integrating sequencing decisions are more easily tractable,

when setups are sequence-dependent.

4 Multi-Level Capacitated Lotsizing with Setup Carryover

beginning of a specific period, production for the respective item can take place

without incurring setup time or costs. Like the PLSP, the CLSPL is suitable

for the representation of continuous production processes, when the according

adaptations are made.

The advantage of the CLSPL over small-bucket formulations is that its com-

plexity, though higher than for the standard CLSP, still is rather low. Also, as

the sequence of lots is determined only at the period transitions, again greater

flexibility is left to subsequent planning phases, thus reducing nervousness of

the plan in case of unforeseen events. The advantage over the standard CLSP,

on the other hand, is that superfluous setups are avoided. As this saves setup

costs as well as setup times, the CLSPL generates different optimal production

plans. When capacity is scarce and setup times are considered, linking lotsizes

over adjacent periods may therefore render a problem feasible, which is not

feasible in the corresponding CLSP formulation.

The model studied in this work is the Multi-Level Capacitated Lotsizing Prob-

lem with Linked Lotsizes (MLCLSPL). It can be seen as the multi-level exten-

sion of the CLSPL or as the integration of linked lotsizes into the MLCLSP.

A set of the rather complex interdependencies that arise in practical lotsizing

problems are considered, such as scarce capacities, setup times, derived de-

mand and positive lead times. The cost of introducing these aspects into the

model formulation is that complexity increases.

formulations from the literature, some complexity considerations and an

overview on the various heuristics that have been developed.

Chapter C describes the test sample generated to test the model formula-

tions presented in chapter D and the overall performance of the heuristic, as

presented in chapter F.

In chapter D, we will first give the model formulation for the MLCLSPL with

single setup carryovers (MLCLSPLe ). Then the formulation for consecutive

setup carryovers (MLCLSPLCC ) is presented. Finally, we introduce a model

formulation for the MLCLSPL with parallel machines (MLCLSPLP M ). Fur-

Chapter A. Introduction and Outline 5

results using CPLEX.

laxation based. The performance of the heuristic for the different problems is

evaluated in chapter F. Finally, in chapter G, we give an outlook on future

work.

Chapter B

Literature Review

the class of models, to which the MLCLSPL introduced later belongs, i. e.

to dynamic big-bucket capacitated lotsizing problems of the CLSP type. For

further research on lotsizing we refer to the numerous existing reviews. These

include Bahl et al. (1987), who introduce and cover subsets of unconstrained

and constrained single- and multi-level lotsizing. Maes and van Wassenhove

(1988) give a structured overview of heuristics to the CLSP and carry out

computational studies. Gupta and Keung (1990) give a review on uncapaci-

tated single- and multi-level lotsizing models. Salomon et al. (1991) give an

overview on different dynamic capacitated lotsizing models and some solution

algorithms. Kuik et al. (1994) study the impact of lotsizing and respond to

some criticism. Wolsey (1995) and Brahimi et al. (2006) review solutions to

the single-item lotsizing problem. Drexl and Kimms (1997) analyze simulta-

neous lotsizing and scheduling models. Staggemeier and Clark (2001) review

the use of metaheuristics in lotsizing and scheduling. Karimi et al. (2003) give

a rather up to date review on solution approaches to single-level capacitated

lotsizing problems. Jans and Degraeve (2007) give a review on metaheuristics

for dynamic lotsizing.

L. Buschkühl, Multi-Level Capacitated Lotsizing with Setup Carryover,

Edition KWV, https://doi.org/10.1007/978-3-658-24034-9_2

8 Multi-Level Capacitated Lotsizing with Setup Carryover

some complexity considerations are presented in section B.2. Due to the mod-

els’ complexity, a great variety of heuristics has been developed. They are

presented thereafter in section B.3.

In this section, we present the standard formulations for the CLSP, the ML-

CLSP and the CLSPL. The focus is on the difference in the modeling assump-

tions. Further extensions such as backorders and overtime are not considered

here. Reformulations will be presented in section B.3.1, as their purpose is to

increase quality and solution speed of the solution approaches presented there.

of each item to a specific resource. All items assigned to the same re-

source are interdependent in that they compete on its scarce capacity.

However, there is no input-output relationship between any of the items

(single-level production).

• The planning horizon is finite and divided into discrete periods (1,. . . ,T ).

the total number of items K.

• The objective of the model is to minimize the sum of inventory and setup

costs.

demand are exogenous to the model and assumed to be deterministic.

Production costs are assumed to be constant over time.

Chapter B. Literature Review 9

ages.

• Setup costs are considered as the out of pocket costs arising from a

changeover. Opportunity costs for lost production time during a setup

will not be regarded, as setup times are directly included in the model.

Inventory costs consist of all quantity-dependent cost factors.

time option. Setup and production times are considered. They reduce

capacity in the periods they occur. Other time consuming activities,

such as transportation times and lead times, are ignored.

Model CLSP

K X

X T

min Z = (sk · γkt + hk · ykt ) (B.1)

k=1 t=1

subject to

X

(tbk · qkt + trk · γkt ) ≤ bm

t ∀m, t = 1, . . . , T (B.3)

k∈Km

10 Multi-Level Capacitated Lotsizing with Setup Carryover

Indices:

k items, k = 1, . . . , K

K total number of items

Km set of items k that are produced on the same resource m

m non-identical resources, m = 1, . . . , M

M total number of non-identical resources

t production periods, t = 1, . . . , T

T length of the planning horizon

Data:

bmt total capacity of resource m in period t in time units

dkt external demand of item k in period t

hk full holding cost coefficient for item k

M total number of non-identical resources

Ω great number

sk setup cost coefficient for item k

tbk time needed to produce one unit of item k

trk time needed to set up a resource for the production of item k

ybk initial inventory of item k at the beginning of the planning hori-

zon

Variables:

γkt binary setup variable

qkt production quantity of item k in period t

ykt inventory of item k available at the end of period t

The objective function (B.1) minimizes setup costs and inventory holding costs.

Constraints (B.2) are the inventory balance constraints. External demand dkt

arises at the end of period t. Hence, the inventory available at the end of the

prior period (yk,t−1 ) and the quantity qkt produced within t can be used to

fulfill it. Excess production is kept on hand (ykt ). The first inventory variable

yk0 is set to some exogenous initial inventory by constraints (B.5).

Constraints (B.3) are the capacity constraints, which are defined for each re-

source considered. Note here that qkt and γkt do not need any resource index,

due to the unique assignment of each item k to one resource m.

only if the resource is set up for this item. Ω is a number great enough not to

Chapter B. Literature Review 11

½ m¾

bt

Ω= max (B.8)

m∈M,k∈Km ,t∈T tbk

Finally, qkt and ykt are defined as non-negative in constraints (B.6) and γkt as

binary in constraints (B.7).

duction systems, end items are produced as well as their components. For

the latter, demand arises from the production quantities of successor items.

Therefore, the lotsizing decision for a given production stage item also influ-

ences the optimal production plan on previous stages. Therefore, an overall

optimal plan can only be generated, when the complete product structure is

considered simultaneously. Hence, the assumptions of the CLSP are adapted

as follows:

level production).

omitted here.

Model MLCLSP

K X

X T

min Z = (sk · γkt + hk · ykt ) (B.9)

k=1 t=1

subject to

X

yk,t−1 + qkt − ykt = dkt + akj · qjt ∀k, t = 1, . . . , T (B.10)

j∈Nk

12 Multi-Level Capacitated Lotsizing with Setup Carryover

X

(tbk · qkt + trk · γkt ) ≤ bm

t ∀m, t = 1, . . . , T (B.11)

k∈Km

Indices:

Nk set of direct successors of item k

Data:

akj production coefficient (quantity of item k, which is directly

needed to produce one item j)

The objective function and the restrictions of the MLCLSP equal those of

the CLSP with the exception that the inventory balance constraints (B.10)

account for derived demand.

with Linked Lotsizes

are changed as follows:

• A setup state can be carried over from one period, if there is a corre-

sponding setup operation in the same period or the setup state has been

carried over to that period. In the latter case, the setup state is preserved

Chapter B. Literature Review 13

two consecutive periods, no other item can be produced in the middle

period.

Model CLSPL

K X

X T

min Z = (sk · γkt + hk · ykt ) (B.16)

k=1 t=1

subject to

X

(tbk · qkt + trk · γkt ) ≤ bm

t ∀m, t = 1, . . . , T (B.18)

k∈Km

X

ωkt ≤ 1 ∀m, t = 1, . . . , T (B.20)

k∈Km

ωk,t−1 + ωkt ≤ 1 + ϑm

t ∀m, k ∈ Km , t = 1, . . . , T (B.22)

γkt + ϑm

t ≤1 ∀m, k ∈ Km , t = 1, . . . , T (B.23)

ωk1 = 0 ∀k (B.24)

14 Multi-Level Capacitated Lotsizing with Setup Carryover

ϑm

t ≥0 ∀m, t = 1, . . . , T (B.27)

Variables:

ωkt binary setup state variable for item k at the beginning of period

t

ϑm

t dummy variable to prohibit a changeover, when the setup state

of the same item k is carried over from period t − 1 to period t

and from period t to period t + 1

Again, most restrictions are the same as for the CLSP. However, restrictions

(B.19) differ from restrictions (B.4) in that the latter enforce a setup for every

positive production quantity. In the CLSPL, however, a particular setup state

can be preserved from the precedent period. In this case, the resource is in the

respective setup state already (ωkt = 1) and thus production can take place

without any additional changeover.

There are four possible reasons for a resource to be set up for a particular item

at the beginning of a period:1

1. Production for the item was started at the end of the previous period and

continued in this period.

Chapter B. Literature Review 15

2. Free capacity at the end of the previous period was used to set up the

resource for the item that is produced first in this period.

3. The resource was idle in the previous period and had been set up for the

relevant item before. Thus, the setup state is carried over for at least two

consecutive periods.

4. The previous period was idle and free capacity was used to set up the

resource for the considered item.

Obviously, on each resource, this can only be the case for at most one product

per period (constraints (B.20)).

corresponding changeover has been performed in the same period or it had been

carried over from the previous period (see constraints (B.21)). For the latter

case constraints (B.22) and (B.23) ensure that a consecutive setup carryover

for one item is installed, only if no other item is produced in between. ϑm

t is a

dummy variable, which prevents a consecutive carryover, if a setup operation

is undertaken in between (ϑm m

t = 0) and vice versa (ϑt = 1). Note here that

the case, where the setup state is carried over to a period, another item is

produced (with the corresponding setup state), the machine is changed back

to the first item and the setup state carried over to the next period, need not

be considered. The reason is that the first carryover is superfluous, when the

production can take place in the same period after an additional changeover.

Dillenberger et al. (1993) were the first to tackle the CLSPL. Unlike most

other lotsizing models, neither holding nor setup costs are considered. Rather,

the objective is to minimize the weighted total backlog at the end of the

planning horizon and the weighted (per period) processing time along the

planning horizon. Products are aggregated to part types, for which a maximum

backlog must not be exceeded. Part types are again aggregated to part type

families, resulting in partially sequence-dependent (major inter- and minor

intra-family) setups. Also several resource groups, respectively consisting of

16 Multi-Level Capacitated Lotsizing with Setup Carryover

resource groups with different processing and setup times. Within each group

of resources processing and setup times are identical for any given item. The

workload assigned to a specific resource group in any period of the planning

horizon is not to exceed its corresponding available capacity. In this model

formulation, setups can be carried over several consecutive periods, unless

another part type is produced in between.

Reference PC ST S SD PM MM BO ML CC

Dillenberger et al. (1993) • • • • • • •

Dillenberger et al. (1994) • • • • • • •

Haase (1994) • •

Haase (1998) •

Gopalakrishnan et al. (1995) a • • • •

Gopalakrishnan et al. (1995) b • • • • •

Gopalakrishnan (2000) • • • •

Gopalakrishnan et al. (2001) • • • •

Sox and Gao (1999) • • •

Sürie and Stadtler (2003) a • • •

Sürie and Stadtler (2003) b • • • •

Quadt (2004) • • • • •

Gupta and Magnusson (2005) • • • •

Abbreviations:

BO Backorders

CC Consecutive setup carryover (over more than one adjacent period transition)

ML Multi-level

MM Non-identical resources

PC Variable production costs

PM Parallel (identical) resources

S Setup costs

SD Sequence dependency of setups

ST Setup time

Dillenberger et al. (1993). The result is a very large-scale mixed integer pro-

gram, which cannot be solved in reasonable time. However, the model can be

transferred into a wide range of optimization models associated with different

objectives and restrictions. Alternative objectives are shipment value opti-

mization, serviceability maximization, resource cost minimization and mini-

mization of the weighted number of setups. The additional constraints com-

Chapter B. Literature Review 17

prise upper bounds for total backlog, minimum and maximum cumulative and

per period production levels, capacity restrictions for different kinds of re-

sources, maximum number of resources (and periods) per part type and of

part types per resource and major inter-family setups.

Haase (1994) introduces linked lotsizes into the standard CLSP. The objective

of the problem is the minimization of setup and holding costs. Setup times

are not considered. Also, production takes place on a single resource. He

replaces the binary setup variable by the difference between a binary setup

state variable and a binary setup carryover variable. The logic is that, if a

setup state is needed and it is not carried over from the precedent period, it

must be installed by a changeover in the same period. The restrictions for

linking equal those of Dillenberger et al. (1993). A linking can be installed

over consecutive periods, unless another item is produced in between. Haase

(1998) considers the same problem without consecutive setup carryovers.

Gopalakrishnan et al. (1995) present two versions of the CLSPL. The first one

is a single-resource CLSPL with the objective of minimizing holding, setup

and time-variant production costs. Production costs here are not incurred per

unit, but once, if the item is produced in the corresponding period. Also,

they consider (real) setup costs, as well as setup times. A new mixed integer

formulation is presented to model the setup carryovers. Unlike the previous

formulations, setup costs are constant per setup and thus independent of the

product, for which the setup is undertaken. This yields the possibility to model

the total number of setups as an integer instead of a binary variable for each

individual setup. However, three kinds of binary variables are needed for any

given product-period combination: αkt = 1 for the first item produced in t,

βkt = 1 for the last item produced in t and γkt = 1 for the item corresponding

to the setup state at the end of period t. These binary variables are used to

set a number of continuous variables such that the total number of setups in

period t is calculated as the number of products produced in t, less a setup

carryover from t − 1, if the end of period setup state in t − 1 corresponds

to the first item produced in t, plus an additional setup, if the setup state

is changed after production of the last item in t, plus one setup in case of t

18 Multi-Level Capacitated Lotsizing with Setup Carryover

comparison to the previous model formulation the use of binary variables here

is extensive. The second model, which the authors refer to as the integrated

model, incorporates parallel resources, resource assignment and tool selection.

Products are aggregated to form product families. Intra-family setups are

neglected. The objective is to minimize setup costs, holding costs and tool

carrying costs.

Gopalakrishnan et al. (1995). He relaxes the rather restrictive assumption of

product-independent setups and incorporates product-specific setup times and

costs. For this reason the setup number is now calculated for every item and

period individually. Setups during idle periods are accounted for separately.

Gopalakrishnan et al. (2001) present a version of the CLSPL with unit produc-

tion costs and item-dependent holding costs, as well as setup costs and times

on a single resource.

Sox and Gao (1999) consider the CLSPL with item-specific setup, holding and

variable production costs, but without setup times. Like Haase (1998) before,

they restrict the number of setup carryovers per item to one in two consecutive

periods. They claim that for a large class of problems the optimal solutions of

the so restricted problem will be close to the optimal solutions of the CLSPL

with any number of consecutive setup carryovers per item.

Sürie and Stadtler (2003) again introduce linked lotsizes into the standard

CLSP. The objective is to minimize setup and holding costs. Production incurs

unit processing and setup times. Setups can be carried over several consecutive

period boundaries. They extend the formulation to the multi-level case by

introducing derived demand into the inventory balance constraint. Several

resources are considered but with a unique assignment of each item to one

specific resource.

Quadt (2004) develops a hierarchical planning approach for the lotsizing and

scheduling problem on flexible flow lines. In this context he first plans the

lotsizes for the bottleneck stage. One of his model formulations is a stan-

Chapter B. Literature Review 19

dard CLSPL with inventory and setup costs, as well as setup times, with the

extensions of backorders (incurring backorder costs) and parallel resources.

Gupta and Magnusson (2005) use a formulation of the CLSPL, which equals

that of Gopalakrishnan et al. (1995), except that they consider sequence-

dependent setup times and costs.

sequence-dependent setup costs (CLSD), which was formulated by Haase and

Goepfert (1996). Due to the sequence-dependency of setups, the sequence of

lots is determined within a period. Thus, we will not present this model here.

As mentioned above, one reason for the extensive work on capacitated lotsizing

problems is their complexity. A mathematical framework to classify compu-

tational problems according to their complexity is provided by the complexity

theory. Two fundamental complexity classes are of interest here, P and N P.

input x in a given domain to an output f (x) in some given range. Then a given

algorithm solves the computational problem by computing f (x). With n = |x|

we denote the length of the somehow encoded input string x. Complexity T (n),

as understood here, then denotes the maximum number of basic arithmetic

operations that is needed to solve any instance of the problem at a given size

n. In other words, we calculate the worst-case-computational effort.2 As it

is difficult to calculate the precise complexity T (n), it is approximated by its

asymptotic order, i. e. T (n) ∈ O(g(n)), if there exist constants c > 0 and a

non-negative integer n0 such that T (n) ≤ c · g(n) for all integers n ≥ n0 . For

example, a maximum computational complexity of 8 · n4 + 27 · n2 is expressed

by O(n4 ).3

For a decision problem there are only two possible outcomes f (x), ‘yes’ or

2 See Neumann and Morlock (1993), page 189.

3 See Brucker (2004), page 37.

20 Multi-Level Capacitated Lotsizing with Setup Carryover

‘no’. In general, ‘yes’ is easier to verify than ‘no’. Class N P, which is the

less restrictive, is defined by the time needed to verify ‘yes’ only. Let y be

a certificate for an input x, which means that f (x) can be determined by

computing y. Then N P denotes the class of decision problems, where for all

‘yes’ inputs x, |y| is bounded by a polynomial in n and there is a polynomial-

time algorithm to verify that y is a valid certificate for x or not.

class P, if it is solvable in polynomial time, i. e. if there exists a polynomial

p such that T (n) ∈ O(p(n)) (or a k such that T (n) ∈ O(nk )) for all inputs x,

where n is the length of the binary encoded input string x.4 This definition

applies to all, ‘yes’ as well as ‘no’, inputs x. As f (x) can be calculated in

polynomial time, it can be used as certificate directly and verified by the

algorithm. Problems belonging to P are called ‘easy’.5

only in exponential time. Furthermore, a decision problem R is called N P-

complete, if every other decision problem Q ∈ N P can be reduced to R in

polynomial time, i. e. there exists a polynomial-time computable function g

that transforms inputs for Q into inputs for R such that x is a ‘yes’ input for

Q, if and only if g(x) is a ‘yes’ input for R.6

CLSP like problems, a decision problem would be whether a feasible solution

exists. Trigeiro et al. (1989) point out, that already for the CLSP this problem

is N P-complete, when positive setup times are considered.7 As every problem

presented in this work can be reduced to the CLSP with setup times by setting

some variables to 0, they are at least as hard to solve and hence also N P-

complete.

5 See Neumann and Morlock (1993), page 191.

6 See Brucker (2004), page 41.

7 See Trigeiro et al. (1989), page 355. They refer to bin packing as a special case of the

CLSP with setup times. For the prove of N P-completeness see Garey and Johnson

(1979).

Chapter B. Literature Review 21

related costs. Hence, they are not decision, but optimization problems. Op-

timization problems are by definition not part of N P. However, an opti-

mization problem is called N P-hard, if the corresponding decision problem

is N P-complete.8 Florian et al. (1980) have proven the single-item CLSP to

be N P-hard. Later, Bitran and Yanasse (1982) showed, that even special

cases solvable in polynomial time become N P-hard with the introduction of

a second item.

such that the solution speed of exact algorithms is enhanced, yet there is

no alternative to heuristic solution approaches, when real-world problems are

considered. Therefore, section B.3 gives an overview on the types of algorithms

that have been presented in the literature.

to tackle CLSP like lotsizing problems. The aim of this review is to give a

structured review on the different types of approaches. A general distinction

can be made between the following categories:

• Lagrangean Heuristics

• Metaheuristics

• Greedy Heuristics

As will be described later, each of these categories can again be divided into

subcategories. Although this review is rather vast, we do not claim to cover

every article published in this area. However, each solution algorithm falls into

8 See Brucker (2004), page 46.

22 Multi-Level Capacitated Lotsizing with Setup Carryover

Manne (1958) SPP DW SL, OV

Dzielinski and Gomory (1965) SPP DW SL

Lasdon and Terjung (1971) SPP CG SL

Bahl (1983) SPP CG SL, ST, OV

Barany et al. (1984) I&L LP B&B, VI SL

Billington et al. (1986) I&L LR B&B ML, ST

Bitran and Matsuo (1986) SPP DW SL, ST

Gelders et al. (1986) I&L LR B&B SL

Eppen and Martin (1987) SR LP RH SL

Cattrysse et al. (1990) SPP CG SL

Chen and Thizy (1990) I&L, SR LR B&B SL

Maes et al. (1991) SPL LP RH ML

Pochet and Wolsey (1991) I&L LP B&B, VI ML, ST

Salomon (1991) SPL LP RH ML

Diaby et al. (1992a) I&L LR B&B SL, ST, OV

Dillenberger et al. (1993) I&L LP F&R SL, PM, BO, SC

Kuik et al. (1993) SPL LP RH ML

Salomon et al. (1993) SPP CG SL, ST

Dillenberger et al. (1994) I&L LP F&R SL, ST, PM, BO, SC

Tempelmeier and Helber (1994) SR LP B&B ML

Clark and Armentano (1995) I&L LP B&C, VI ML, ST

Hindi (1995) SPP CG SL

Harrison and Lewis (1996) I&L CMH ML, ST, BO

Hindi (1996) SPP CG SL

Stadtler (1996) I&L,SPL,SR LP B&B ML, ST, OV

Stadtler (1997) SR LP B&B ML, ST, OV

Hung and Hu (1998) I&L SPB SL, ST, BO

Katok et al. (1998) I&L CMSB ML, ST

Armentano et al. (1999) I&L LP B&B,NFA SL

Belvaux and Wolsey (2000) I&L LP B&C, VI ML, ST, BO

Miller et al. (2000) I&L LP B&C, VI SL, ST

Belvaux and Wolsey (2001) I&L LP B&C, VI ML, ST, BO

Alfieri et al. (2002) SPL,SR LP RH SL

Degraeve and Jans (2003) SPP CG SL, ST

Huisman et al. (2003) SPP CG SL, OV

Mercé and Fontan (2003) I&L LP F&R SL, ST, BO

Stadtler (2003) SPL LP RTW ML, ST, OV

Sürie and Stadtler (2003) SPL LP B&C,C&B, VI SL,ML, ST, SC

Haase (2005) SPP,SPL LP CG,F&R SL

Abbreviations:

FORM Model Formulation (I&L = Inventory&Lotsize, SPL = Simple Plant Location, SPP

= Set Partitioning Problem, SR = Shortest Route)

REL Relaxation (LP = LP Relaxation, LR = Lagrange Relaxation)

SolA Solution Approach (B&B = Branch&Bound, B&C = Branch&Cut, C&B =

Cut&Branch, CG = Column Generation, CMH = Coefficient Modification Heuris-

tic, CMSB = Coefficient Modification Subroutine with Cost Balancing, , DW =

Dantzig-Wolfe Decomposition, F&R = Fix&Relax, NFA = Network Flow Algo-

rithm, RH = Rounding Heuristic, RTW = Rolling Time Window, SPB = Shadow

Price based Heuristic, VI = Valid Inequalities)

MOD Problem solved (ML = Multi-Level, SL = Single-Level, BO = Backorders, OV =

Overtime, PM = Parallel Machines, SC = Setup Carryover, ST = Setup Time)

Chapter B. Literature Review 23

tend to be more flexible than other procedures with respect to model exten-

sions. Furthermore, they usually produce better quality solutions but at the

cost of higher computational complexity.

• Branch&Bound Heuristics

• Fix&Relax Heuristics

• Rounding Heuristics

The former stop with an optimal solution, if an optimal solution exists, regard-

less of the effort in terms of required computational time and memory. The

latter only search parts of the solution space and try to find a good feasible

solution in reasonable time. Introducing a time-limit into an exact method

is an easy way to construct an MP-based heuristic. If the solution obtained

within the time-limit is infeasible, it can serve as a promising lower bound,

from which one can try to construct a feasible solution. Table B.2 gives an

overview on MP-based solution approaches.

The Branch&Bound method (B&B), when left enough time to terminate, be-

longs to the group of exact solution procedures. It enumerates feasible solu-

24 Multi-Level Capacitated Lotsizing with Setup Carryover

tions implicitly. This method consists of two parts, ‘branching’ and ‘bound-

ing’. While during the ‘branching’ part new disjoint subsets of the solution

space are generated, unpromising ones are removed during the ‘bounding’ part.

For Mixed Integer Programs (MIP) with binary variables, as considered here,

branching is based on iteratively fixing the binary variables to 0 and 1. For

bounding a relaxation of the remaining subproblems is solved.

Different possibilities are considered to relax the model formulation. The first

is the LP relaxation of a MIP. Here, the original binary variables of the MIP are

solved with the requirement that all real values of the relaxation are feasible

with respect to their bounds (e. g. 0 and 1 for binary variables). Armentano

et al. (1999) use a B&B procedure to solve the CLSP. Here, the corresponding

LP relaxation is determined by a network flow algorithm. The second relax-

ation method is the Lagrangean relaxation, which is described in detail in the

following section. B&B methods, which are imbedded into a Lagrangean Re-

laxation scheme, are described in Billington et al. (1986), Gelders et al. (1986),

Chen and Thizy (1990) and Diaby et al. (1992a).

As the bounds obtained by relaxing the above mentioned I&L formulation are

quite poor, one major research trend is devoted to tightening the lower bounds.

This in turn makes the B&B method more efficient. The relevant methods are

reformulation of the mathematical model, redefinition of the corresponding

decision variables and the introduction of additional valid inequalities.

B.3.1.1.1 Reformulation

CLSP type are based on the Shortest Route representation (SR) and the Sim-

ple Plant Location respresentation (SPL), respectively. They will be presented

for the MLCLSP in the following. Both are based on the idea, that by adding

information, the quality of the lower bound is increased. In both cases, the

additional information is that production quantities are related to the specific

demand, which they are produced to satisfy.

The SR formulation was introduced by Eppen and Martin (1987) for the

Chapter B. Literature Review 25

SLULSP. Tempelmeier and Helber (1994) extend this formulation with respect

to the capacitated multi-level case. Let

X

Dkt = dkt + akj · Djt ∀ k, t = 1, . . . , T (B.29)

j∈Nk

and

s

X

dkts = Dkτ ∀ k, t = 1, . . . , T, s = t, . . . , T. (B.30)

τ =t

decision variables ∆kts are defined as the fraction of the corresponding cumu-

lated demand dkts , which is satisfied by production in period t. Using these

additional variables and parameters, the MLCLSPSR can be stated as follows:

Model MLCLSPSR

K X

X T

min Z = (sk · γkt + hk · ykt ) (B.31)

k=1 t=1

subject to

T

X

∆k1s = 1 ∀k (B.32)

s=1

t−1

X T

X

∆kp,t−1 = ∆kts ∀k, t = 2, . . . , T (B.33)

p=1 s=t

T

X X T

X

yk,t−1 + dkts · ∆kts = dkt + akj · djts · ∆jts + ykt ∀k, t = 1, . . . , T

s=t j∈Nk s=t

(B.34)

Ã T

!

X X

tbk · dkts · ∆kts + trk · γkt ≤ bm

t ∀m, t = 1, . . . , T (B.35)

k∈Km s=t

T

X

∆kts ≤ γkt ∀k, t = 1, . . . , T, dkts > 0 (B.36)

s=t

26 Multi-Level Capacitated Lotsizing with Setup Carryover

Indices:

s periods, s = 1, . . . , T

Data:

dkts cumulated demand of item k in periods t to s

Variables:

∆kts fraction of the cumulated demand of item k in periods t to s,

which is satisfied by production in period t

The objective function (B.31) minimizes the inventory and setup costs. Con-

straints (B.32) and (B.33) are the flow constraints of the shortest route prob-

lem. While constraints (B.32) ensure that the overall flow is complete, con-

straints (B.33) ensure that at each node input equals output. The inventory

balance equations (B.34), capacity constraints (B.35) and setup enforcement

constraints (B.36) are adapted to the additional variables ∆kts . Constraints

(B.37) to (B.40) give the ranges of the variables.

decreases the number of non-negative coefficients in the constraints’ matrix and

thus the computational effort for solving the corresponding LP relaxation.

In the SPL formulation, the analogy to the facility location problem is that

production periods t are modeled as facilities and demand periods s as cus-

tomers. Only earlier production periods can satisfy demand s ≥ t. Facilities

Chapter B. Literature Review 27

can be used for production, only if they are opened and hence fixed costs

are incurred in the form of setup costs for the respective product. Fulfilling

demand incurs transportation costs, which here equal holding costs.

ever, not cumulated demand, but demand of an individual period s is consid-

ered. The fraction of demand of item k arising in period s that is satisfied

in period t is denoted as δkts . In addition, the net demand Dkt of item k in

period t is defined as the sum of external and derived demands:

X

Dkt = dkt + akj · Djt ∀ k, t = 1, . . . , T (B.41)

j∈Nk

Model MLCLSPSP L

K X

X T

min Z = (sk · γkt + hk · ykt ) (B.42)

k=1 t=1

subject to

t

X

δkst = 1 ∀k, t = 1, . . . , T, Dkt > 0 (B.43)

s=1

T

X X T

X

yk,t−1 + Dks · δkts = dkt + akj · Djs · δjts + ykt ∀k, t = 1, . . . , T

s=t j∈Nk s=t

(B.44)

Ã T

!

X X

tbk · Dks · δkts + trk · γkt ≤ bm

t ∀m, t = 1, . . . , T (B.45)

k∈Km s=t

28 Multi-Level Capacitated Lotsizing with Setup Carryover

Data:

Dkt total (external and derived) demand of item k in period t

Variables:

δkts fraction of demand of item k in period s, which is satisfied by

production in period t

The difference to the SR formulation is that the flow constraints (B.32) and

(B.33) are omitted and instead constraints (B.43) ensure that given demand

is fulfilled without backlogging. The remaining restrictions (B.44) to (B.50)

are merely adapted to the new definition of the production quantity.

The value of the LP relaxation of the SR formulation and of the SPL formu-

lation are identical. The number of decision variables is the same, too, but

not the number of constraints. For the SPL formulation K · T ( 12 (T + 1) − 1)

additional constraints are needed.

Rosling (1986) introduce the SPL formulation for the MLULSP as an extension

to the work of Krarup and Bilde (1977). Furthermore, Maes et al. (1991)

extend the work of Rosling (1986) by introducing capacity constraints.

inequalities. They reduce the size of the solution space by cutting off irrelevant

parts. Three methods can be distinguished. First, if the valid inequalities

Chapter B. Literature Review 29

Second, valid inequalities can be introduced in the course of a B&B algorithm.

This method is called Branch&Cut (B&C). Third, the Cut&Branch procedure

(C&B) incorporates all generated inequalities into the model formulation prior

to starting the B&B algorithm.

Barany et al. (1984) define a set of valid inequalities also known as (l, S)-

inequalities. They include lotsizing variables and inventory variables for the

SLULSP. These additional inequalities describe the convex hull for the single

item uncapacitated lotsizing polytope. They can also be applied to the CLSP.

Furthermore, other valid inequalities are derived by Miller et al. (2000) for the

capacitated problem. Pochet and Wolsey (1991) and Clark and Armentano

(1995) extend the work of Barany et al. (1984) for the multi-level case.

Belvaux and Wolsey (2000) present a general framework for modeling and

solving lotsizing problems, including lotsizing-specific preprocessing, cutting

planes for different aspects of lotsizing problems, general cutting planes, and

a lotsizing-specific primal heuristic. Their work is extended in Belvaux and

Wolsey (2001) to further model extensions, e. g. start-ups, changeovers and

switch-offs.

Sürie and Stadtler (2003) derive valid inequalities for the CLSPL and the ML-

CLSPL. To do so, they introduce extended model formulations by redefining

the setup carryover constraints. In addition, they perform a preprocessing

step, which leads to inequalities that link inventories to setups and single-item

production to capacity. They adopt C&B and B&C with a given time-limit to

find the optimal or the first feasible solution. In order to reduce complexity, a

time-oriented decomposition approach9 is applied.

dividing the problem into several subproblems. Three sets of binary variables

30 Multi-Level Capacitated Lotsizing with Setup Carryover

and the third set is fixed to the values of a previous iteration.

Dillenberger et al. (1993) develop an F&R algorithm for the CLSPL. It essen-

tially consists of a B&B algorithm, where the order of branching is determined

by the sequence of periods. The resulting algorithm is a period-by-period

heuristic with binary setup variables in the first period of the current lotsizing

window, fixed setup variables in all prior periods and relaxed setup variables

in all ongoing periods until the end of the planning horizon. Dillenberger et al.

(1994) again apply the F&R heuristic to an extended model formulation.

lotsizing windows. The binary variables preceding this time window have

been determined in prior iterations and those after the end of the planning

window are neglected. Thus, capacity requirements are only approximated.

This approach is also applied to the CLSPL in Sürie and Stadtler (2003).

Mercé and Fontan (2003) propose a MIP-based heuristic, which relies on the

division of the planning horizon into several sub-horizons and different freezing

methods for past decisions.

In Rounding Heuristics (RH), the LP relaxation of the MIP is solved and the

fractional binary variables are subsequently rounded. In the case of capacitated

lotsizing problems, these solutions are often infeasible. The reason is that

capacity may not be sufficient, if there are too few production periods. Hence,

the fractional binary variables are usually rounded up and rounded down only

within a given threshold.

Using the SPL formulation for the MLCLSP without setup times, Maes et al.

(1991) introduce several RHs. Based on the LP relaxation, the relaxed setup

variable with the highest value is fixed to 1. Subsequently, the reduced LP

relaxation is solved. Furthermore, all relaxed binary variables within a given

range can also be fixed to 1.

Chapter B. Literature Review 31

Eppen and Martin (1987) and Alfieri et al. (2002) present RHs, which solve

the LP relaxation of the SR formulation for the CLSP. All fractional binary

variables are fixed within a given threshold. Then, a limited B&B is started

with the remaining binary variables.

Kuik et al. (1993) present a further RH for the MLCLSP without setup times

and on one resource. Based on the LP relaxation of the SPL formulation,

all binary variables within a given range are fixed either to 0 or to 1. The

remaining binary variables are initially set to 1. A Simulated Annealing or

Tabu Search approach follows, where changes only affect the unfixed binary

variables. A similar approach can be found in Salomon (1991).

These iterative solution methods exploit the fact that omitting the binary

variables leads to a significantly easier problem. Here, the complete setup

pattern is either given or accounted for only implicitly, and the remaining LP

is solved to optimality.

Hung and Hu (1998) present an iterative solution approach for the CLSP, in

which at each iteration the setup pattern is fixed and the remaining LP is

solved. They start with an initial setup pattern assuming a production for ev-

ery item in every period. Hence, all setup variables are fixed to 1. After solving

the resulting linear program, they use the information of the shadow prices

relating to the capacity constraints to identify items and periods, for which it

is beneficial to fix the corresponding setup variables to 0. Subsequently, the

resulting LP is solved.

Heuristic for the MLCLSP. They exploit the fact that in the MLCLSP each

setup variable is linked to a corresponding continuous production variable,

in that both variables must be equal to zero or positive at the same time.

Therefore, the binary variables are omitted and setup times are accounted for

implicitly via modification of the production time coefficients of the related

production variables. This leads to a reduced linear program, as not only the

32 Multi-Level Capacitated Lotsizing with Setup Carryover

binary variables but also the linking constraints to the continuous variables can

be eliminated. In each iteration the respective linear program is solved and the

production time coefficients are modified relating to the capacity consumption

in the previous iteration.

with Cost Balancing, which is an extension to the Coefficient Modification

Heuristic. Here, also the production cost coefficients are modified to account

for setup costs implicitly.

Generation

ferred into a problem of the Set Partitioning type. The objective is to find

a convex combination of single-item schedules, which keeps the capacity con-

straints of the original CLSP and leads to minimal costs. The decision variables

are continuous variables to combine schedules for each item.

To find the optimal solution to the Set Partitioning Problem (SPP), Column

Generation (CG) is applied. CG is an iterative solution procedure, generating

new schedules leading to a better objective function’s value at each iteration.

To do so, the SPP becomes a master problem. The corresponding subproblem

is to find the single-item schedules to feed into the master problem. For the

CLSP, it consists of a set of SLULSPs. The CG procedure starts with a limited

number of schedules for each product. At each iteration, first the SPP is

solved. Then shadow prices related to the capacity and convexity constraints

are updated and used to modify the objective function of the subproblems,

which are subsequently resolved. The resulting schedules are then introduced

into the master problem, only if their objective function’s value is negative

and therefore beneficial. The procedure is repeated until no more schedules

improving the objective function of the SPP can be generated. The approach

11 A solution approach based on CG is not satisfactory for the MLCLSP.

Chapter B. Literature Review 33

stops with a promising lower bound for the CLSP. Afterwards, an additional

solution heuristic has to be applied to generate a feasible solution.

not iterative, but considers only dominant production plans. Manne’s work

is extended by Dzielinski and Gomory (1965), who develop a Column Selec-

tion procedure to handle larger problems. Later, Lasdon and Terjung (1971)

develop a CG approach. Algorithms of this type are also presented by Bahl

(1983), Cattrysse et al. (1990), Salomon et al. (1993), Degraeve and Jans

(2003) and Huisman et al. (2003).

Hindi (1995) presents a heuristic including variable redefinition and CG. They

solve a minimum cost flow problem to compute the upper bound.

Hindi (1996) combines the ideas of LP relaxation, CG, Minimum Cost Network

Flow and Tabu Search into a hybrid algorithm.

Haase (2005) also solves the CLSP via Column Generation. The resulting

lower bound is close to optimality as most of the variables are already integer.

Only a small number of variables are left, which are considered in a following

rolling time window procedure similar to Dillenberger et al. (1993).

grangean Relaxation (LR). This means that the complicating constraints are

relaxed and punished at penalty costs in the objective function. Lagrangean

Heuristics can be divided into

Lagrangean Relaxation or Decomposition, respectively, and given values of the

Lagrangean multipliers. A feasible solution is constructed by some feasibility

34 Multi-Level Capacitated Lotsizing with Setup Carryover

routine and serves as the new upper bound. Finally, the Lagrangean multi-

pliers are updated. The heuristics are based on the convergence of the lower

and the upper bound through the adaptation of the Lagrangean multipliers.

With the right Lagrangean multipliers the solution to the relaxed problem

will be very close to the optimal solution of the original problem. Thus, only

small modifications will have to be made to install a close to optimal feasible

solution.

the case of the CLSP, the only complicating constraints are the capacity con-

straints. When they are relaxed, the remaining problem decomposes into K

problems of the SLULSP type. For the MLCLSP, also the inventory balance

constraints reflect interdependencies because of the derived demand. Relax-

ing both constraints again leads to K problems of the SLULSP type. An

alternative consists of relaxing only the capacity constraints and solving the

remaining MLULSP. For the CLSPL, the capacity constraints and the setup

carryover constraints are relaxed. The setup carryover constraints restrict the

setup carryovers per resource and period to at most one. The remaining prob-

lem is of the SLULSPL type. A number of exact and heuristic algorithms

have been introduced to solve the uncapacitated lotsizing problems efficiently,

e. g. Wagner and Whitin (1958), Evans (1985), Federgruen and Tzur (1991),

Wagelmans et al. (1992), Silver and Meal (1969) and Groff (1979).

Instead of relaxing the complicating constraints, all original constraints are

kept. However, the original problem is decomposed into subproblems via vari-

able duplication. Each subproblem only respects some of the constraints. In

order for the solutions to the subproblems to be a valid solution to the orig-

inal problem, the duplicates must equal the corresponding original variables.

These coupling constraints are finally relaxed.

Various techniques have been developed to generate the upper bound. They

will be described below.

Chapter B. Literature Review 35

stopping criterion is met. The subgradient indicates the direction, in which

the Lagrangean multipliers have to be altered, in order to achieve the greatest

possible improvement of the objective function’s value. Generally, in the case

of optimization problems with multiple constraints, the subgradients can be

based on the violation of the constraints themselves.

Thizy and van Wassenhove (1985) develop a solution algorithm to the CLSP,

which is later extended by Diaby et al. (1992b) to a CLSP with setup times

and overtime. Based on the LR of the capacity constraints the setup decisions

are retained. The basic idea is that once the setup decision has been made, the

resulting problem can be formulated as a TP. A perturbation scheme is devel-

oped based on this formulation. It is applied to schedules, when production

cells do not match the results of the LR.

Trigeiro (1987) solves the CLSP without setup times. Later, Trigeiro et al.

(1989) extend their heuristic to problems with setup times. To generate feasi-

ble solutions, they apply a smoothing procedure. Its objective is to minimize

total opportunity costs divided by the quantity of eliminated overtime. The

smoothing routine consists of up to two backward and two forward passes.

Thus, it shifts lots period-by-period alternately starting from the end of the

planning horizon to earlier periods and from the beginning of the planning

horizon to later periods. Lots are shifted backward either to the next period

or to the next period with a corresponding setup. Either the complete lot or

as much as is needed to eliminate the capacity violation is shifted. For the

forward passes, the target period is always the next later period and the quan-

tity shifted is always the inventory of the respective item. At most one lot per

period may be split. The heuristic moves on to the next period, when all over-

time has been eliminated in the incumbent period. The procedure is guided by

the constraints, the costs and the Lagrangean multipliers. Among the forward

passes, the first considers cumulative overtime only, while the second elimi-

nates overtime in every period. When a feasible solution is found, unnecessary

inventory is eliminated by a simple fix-up, i. e. production postponement.

36 Multi-Level Capacitated Lotsizing with Setup Carryover

Thizy and van Wassenhove (1985) C SLU, DP TP SL

Trigeiro (1987) C SLU, DP FC, B, F SL

Trigeiro et al. (1989) C SLU, DP FC, B, F SL, ST

Campbell and Mabert (1991) C, II CYCSLU, ENUM FC, B SL, CS, II, MINQ

Diaby et al. (1992b) C SLU, DP TP SL, ST, OV

Tempelmeier and Derstroff (1993) C, I SLU, DP FC, B, F ML, ST

Millar and Yang (1994) SET TP, INT, LCINT TP SL, BO

Millar and Yang (1994) DEC TP, SLU, ZAN TP SL, BO

Tempelmeier and Derstroff (1996) C, I SLU, DP FC, B, F ML, ST

Özdamar and Barbarosoglu (1999) PM SLU, LOAD, GR SA ML, ST, PM, OV, BO

Sox and Gao (1999) C, S SLULSPL, DP FC, F, B, SC SL, SC

Özdamar and Barbarosoglu (2000) C, (I) M(S)LULSP, DP SA, LS, GS ML, ST

Moorkanat (2000) C, I SLU, DP F, B, M ML, ST

Chen and Chu (2003) BIN LMLCLSP, LP SET1 ML

Hindi et al. (2003) C SLU , DP FC, B, F SL, ST

Sambasivan and Yahya (2005) C MPSLULSP, B&B F, B, M SL, ST, MultP

Brahimi et al. (2006) C, TW (TW)SLU, DP FC, B, F SL, TW

Abbreviations:

RC Relaxed Constraints (BIN = Binary Setup Variables, C = Capacity Constraints,

DEC = Coupling Constraint between Duplicates and the corresponding original

Variables in Decomposition Approaches, I = Inventory Balance Constraints, II =

Initial Inventory Constraints, PM = Coupling between Production on each Re-

source and total Production for an Item, S = Setup Carryover Constraints, SET

= Coupling Constraint between Production and Setup, TW = Time Window Con-

straints)

SP Subproblem solved in case of Decomposition and Solution Method (CYCSLULSP

= SLULSP with cyclical Schedules, INT = Integer Problem to determine Setup

Scheme, LMLCLSP = Linear MLCLSP, LOAD = Capacitated Loading Problem,

MPSLULSP = Multi-Plant SLULSP, SLU = SLULSP, TP = Transportation Prob-

lem, TWSLU = SLULSP with Time Windows, B&B = Branch&Bound, DP =

Dynamic Programming, ENUM = Enumeration of all possible Schedules, GR =

Greedy, LCINT = Least Cost Search to determine Setups, LP = Linear Program-

ming)

FEAS Type of feasibility Check (B = Backward, BC = Backward Shifting of cumulative

Overtime, F = Forward, FC = Forward Shifting of cumulative Overtime, GS =

Global Search, LS = Local Search, M = Machine-by-Machine, SA = Simulated

Annealing, SC = Feasibility of Setup Carryover Constraints, SET1 = All positive

Setup Variables are set to 1, TP = Fix Setups and solve Transportation Problem)

MOD Problem solved (ML = Multi-Level, SL = Single-Level, BO = Backorders, CS =

Cyclical Schedules, II = Initial Inventory, MINQ = Minimum Production Quan-

tity, MP = Multi-Plant, OV = Overtime, PM = Parallel Machines, SC = Setup

Carryover, SD = Sequence Dependency, ST = Setup Time, TW = Time Windows)

Hindi et al. (2003) solve the CLSP with setup times. Their heuristic consists

of three parts. First, a smoothing heuristic is applied, which is similar to that

of Trigeiro et al. (1989). Second, after each pass, the solution is optimized

further by solving a Capacitated Transshipment Problem by the dual network

complexity method for the given setup schedule. The transshipment model

Chapter B. Literature Review 37

much smaller. The smoothing heuristic is followed by a Variable Neighborhood

Search. The costs of not producing an item in a given period is estimated

as the difference between setup costs and inventory holding costs from the

latest precedent period with a setup for this item. The most beneficial items

are chosen, with their number determined by an adaptive neighborhood size.

Then, the smoothing procedure is reapplied. The move is executed, only if it

leads to a feasible solution, which is better than the incumbent solution. In

this case the number of moves is reset to one.

Tempelmeier and Derstroff (1993) and Tempelmeier and Derstroff (1996) de-

velop a solution procedure to the MLCLSP with setup times and positive lead

times based on the formulation of Billington et al. (1986). To account for the

derived demands of the multi-level structure, a sequence of SLULSPs is solved

starting with the end products and down to the components. Then, they ap-

ply a smoothing procedure similar to Trigeiro et al. (1989) to ensure capacity

feasibility. However, as they solve a multi-level problem, they apply as well

single- as multi-item (subsets of the product structure) shifts.

Moorkanat (2000) solves the MLCLSP with setup times using the single- and

multi-item shifts introduced by Tempelmeier and Derstroff (1996). Addition-

ally to considering every period for a given resource, before moving on to the

next, a second set of heuristics, considering all resources in one period before

moving on is implemented. The major difference to the implementation by

Tempelmeier and Derstroff (1996) is that backlogging is allowed for the end

items. Hence, while the schedule may not be feasible with respect to the in-

ventory balance constraints, it is guaranteed to be feasible with respect to the

capacity constraints.

Campbell and Mabert (1991) develop a LH similar to Trigeiro et al. (1989) for

the CLSP with cyclical schedules. Here, the time between production periods

of an item are constant. To solve the relaxed problems, all combinations of first

production period and cycle length are evaluated and the least cost solution

is chosen.

38 Multi-Level Capacitated Lotsizing with Setup Carryover

Sox and Gao (1999) develop a LR heuristic (which they refer to as Lagrangean

Decomposition), where the capacity and setup carryover constraints are re-

laxed.

and Simulated Annealing to solve a multi-level production system with a serial

product structure for several end items. Production takes place on parallel

machines. By relaxing the coupling constraints between loading quantities

on each machine and the respective total production quantity the problem

decomposes into a set of SLULSPs and a set of capacitated loading problems

for given total production quantities per item and period. Both subproblems

are solved only approximately. From a generated inventory feasible solution,

capacity feasibility is finally achieved through a specialized procedure based

on Simulated Annealing.

Özdamar and Barbarosoglu (2000) apply a similar procedure to solve the stan-

dard MLCLSP with general product structures. They develop two relaxation

schemes. The first is called hierarchical relaxation, as only the capacity con-

straints are relaxed. The second relaxes both the capacity and the inventory

balance constraints. Neither is solved exactly and therefore they are not a

lower bound to the original problem. To generate a feasible and improved

solution to the original problem, they apply three types of procedures. One is

a Simulated Annealing approach. Here, the direct neighborhood is searched,

which is defined by the shifting of a randomly selected partial lot to the next

resource or the next period. The second is based on the same neighborhood

definition, but allows improving moves only. Finally, the third is a global

search procedure to ensure capacity feasibility.

Chen and Chu (2003) develop a solution algorithm to the MLCLSP, in which

they relax the binary constraints for the setup variables. The remaining linear

model is solved approximately by an iterative linear programming algorithm,

respectively obtaining setups for given production quantities and vice versa,

until the same results are obtained in both computations. To update the

Lagrangean multipliers, they use a surrogate subgradient method (SSG). SSG

is a subgradient method, which is adapted to the fact, that the relaxed problem

Chapter B. Literature Review 39

trivially by setting all positive setup variables to 1 and all others to 0. Finally,

the solution is improved on with the help of a local search procedure.

with setup times. This extension is reflected through several resources, item

transfers in the inventory balance constraints and the corresponding cost factor

in the objective function. They also alter the problem to forbid inter-plant

transfers. The remaining problem corresponds to the multi-resource CLSP.

They use a lot shifting-splitting-merging routine to generate a feasible solution.

First, capacity violating lots are shifted within the same period to another

plant, where the corresponding demand arises. If capacity violations persist,

production is then shifted across periods within the same plant. The shifts are

executed forward and backward without incurring capacity violations in the

target periods.

Millar and Yang (1994) develop both a Lagrangean Relaxation and a La-

grangean Decomposition approach to solve the CLSP with backorders. Both

decompose the original problem into two subproblems. In both cases, one is

a TP. In the case of the LR, the other one is an integer problem to determine

setups. It is solved by inspection of the given cost coefficients. In the case of

the LD, the second subproblem is a set of SLULSPs, which is solved by the

algorithm introduced by Zangwill (1966). To compute the upper bound, again

a TP is solved for the setup decisions given by the relaxation.

Brahimi et al. (2006) propose several LR schemes for the CLSP with time

windows. In the standard CLSP, production can take place not later than the

due demand period. Here, production is additionally limited to be not earlier

than a given release period.

Thizy (1991) studies different LD approaches to the CLSP, their values and

complexity.

40 Multi-Level Capacitated Lotsizing with Setup Carryover

of reduced size and then coordinate the individual solutions. Aggregation

approaches reduce problem size by omitting details first and breaking the

solution down later. Decomposition approaches split the original problem into

subproblems and coordinate the schedules later. The following approaches can

be distinguished:

• Time-based decomposition

• Item-based decomposition

• Resource-based decomposition

• Item-based aggregation

• Resource-based aggregation

subproblems. Time-based decomposition is mostly premised on rolling sched-

ules. The idea is an iterative approach, where the planning horizon is split

into shorter (mostly overlapping) time windows. Once a solution is found for

the respective window, the solution is fixed for a number of periods at the

beginning of the window and another problem solved. A closely related set of

iterative algorithms are the F&R algorithms presented above.

approaches.

Newson (1975a) and Newson (1975b) develop a heuristic to the CLSP without

and with overtime, respectively. The idea is to neglect the capacity constraints

first, which leads the problem to decompose into single-item problems of the

SLULSP type. These are solved with the help of the Wagner-Whitin algorithm.

Then, infeasible combinations of production schedules are eliminated from the

set of possible solutions. The reduced set is resolved until a feasible solution

is achieved.

Chapter B. Literature Review 41

Kirca and Kökten (1994) develop an item-by-item heuristic to the CLSP. It-

eratively, one of the yet unscheduled items is selected based on the Least Unit

Cost criterion. Second, production quantities are bounded with the intention

to ensure feasible plans for the remaining items. Third, the modified single-

item CLSP with inventory restrictions is solved for the selected item by an

algorithm based on dynamic programming.

Newson (1975a) DC (K) SLU DP SL, ST

Newson (1975b) DC (K) SLU DP SL, ST, OV

Blackburn and Millen (1984) DC (K) SLU MRP ML

Tempelmeier and Helber (1994) DC (K) SL DS ML

Helber (1994) DC (K) SL DS ML, ST

Kirca and Kökten (1994) DC (K) SL DP SL

Özdamar and Bozyel (2000) AG (K) SI EX SL, ST, OV

Bourjolly et al. (2001) DC (T) SL TS SL, ST, SC

Sambasivan and Schmidt (2002) DC (K) SLU EX SL, ST, MP

Quadt (2004) AG (K, M), DC SLSC EX ML, ST, BO,

(K, T) PM, SC, SD

Quadt and Kuhn (2005) AG (K, M), DC SLSC EX ML, ST, BO,

(K, T) PM, SC, SD

Boctor and Poulin (2005) AG (K) SL GR ML, ST, (OV),

MM

Abbreviations:

AG/DC Aggregation or Decomposition Approach (AG = Aggregation, DC = Decomposi-

tion, K = Itembased, M = Aggregation of parallel Machines, T = Periodbased)

SP Subproblem that is solved in case of Decomposition (SI = Single-Item CLSP, SL =

CLSP, SLSC = CLSPL, SLU = SLULSP)

SOL Solution Method applied to solve the (Sub)problem (DP = Dynamic Programming,

DS = Modified Dixon-Silver Heuristic, EX = Exact Solution, GR = Greedy, MRP

= Capacitated MRP, TS = Tabu Search)

MOD Problem solved (ML = Multi-Level, SL = Single-Level, BO = backorders, MM =

Multi-Machine, MP = Multi-Plant, OV = Overtime, PM = Parallel Machines, SC

= Setup Carryover, SD = Sequence Dependency, ST = Setup Time)

transfers by first solving SLULSPs and then applying a smoothing procedure,

which shifts production quantities across plants and periods to remove capacity

violations. The uncapacitated subproblems are reformulated as shortest-route

problems and then solved exactly.

One decomposition principle for the MLCLSP is based on the idea that (un-

less an item is produced on the same resource as its predecessor) the only link

42 Multi-Level Capacitated Lotsizing with Setup Carryover

lotsizing decisions on one level influence lotsizing on subsequent levels and

hence the arising costs. Thus, one way of reflecting these interdependencies

is via cost modifications. Blackburn and Millen (1984) suggest a set of solu-

tion algorithms to the MLCLSP for assembly processes. First, holding and

setup costs are modified based on the assumption that the order interval of a

component is always an integer multiple of its successor. Second, single-item

problems are solved from the end item and along the product structure. The

applied solution heuristics are capacitated versions of the Wagner-Whitin algo-

rithm, Part-Period Balancing, Silver-Meal, Economic Order Quantity, Periodic

Order Quantity and some modified versions.

CLSPs. They propose four versions, which arise when combining two types

of constructing the CLSPs and two types of cost adjustments. The applied

heuristic is a modified Dixon-Silver heuristic. Helber (1994) extends the algo-

rithm to cover problems with setup times.

planning approach, which is based on the aggregation of the demand of all

items in a period to one type demand. For the resulting single-item CLSP, the

weighted average of process times is used to compute the one aggregate pro-

cess time. Setups are not accounted for explicitly but via a setup allowance

percentage, which reduces capacity. Finally, individual lots are planned re-

specting the aggregate lotsizes by a filling procedure. As a good estimation of

the setup allowance is not trivial, they also propose an iterative approach, in

which it is updated with the help of the respective setup time in the previous

solution.

Boctor and Poulin (2005) solve a serial multi-stage problem with different

resources by making the assumption that lotsizes are the same on each stage.

The remaining CLSP is solved by a Greedy algorithm as described below.

(2001) by applying it in a rolling horizon approach. Based on the idea of

Chapter B. Literature Review 43

dynamic recursion, first the problem for period 1, then 1 and 2, etc. are solved.

The respectively best schedule obtained is used to find the best schedule for

the next subproblem. Setup carryovers are introduced later by linking lots for

an item, if it is produced in two consecutive periods.

Quadt (2004) and Quadt and Kuhn (2005) use a rather sophisticated aggrega-

tion and decomposition approach to solve the lotsizing and scheduling problem

arising at the assembly stage of wafer fabrication. First, they decompose the

multi-level problem by identifying the bottleneck stage. For this stage only,

they solve a CLSPL with backorders and sequence-dependent setups on par-

allel machines by aggregating the parallel machines to one single machine and

products to product families. This so-called ‘heuristic model’ is based on in-

teger variables, which count the number of resources set up for a family. The

assumption is made that no further resource can be set up for a family, which is

partially scheduled on another resource. Setup carryovers can only be installed

on resources that produce only one family. To ensure that a feasible solution

is computed, the heuristic model is again embedded into a period-by-period

heuristic. In each period, the model is solved to optimality for the current and

all ongoing periods using CPLEX. A scheduling subroutine is used to detect

capacity violations in the respectively current period and possible surplus pro-

duction is postponed to the next period. Finally, this procedure may lead to

unsatisfied demand at the end of the planning horizon. For this reason the pro-

cedure ends with a backtracking heuristic, shifting the concerned production

volumes towards the beginning of the planning horizon. When the CLSPL has

been solved, the schedule is rolled out onto the remaining production stages

on the individual machines. Here, the average flow time and the inventory

for intermediate products are minimized. Finally, product families are dis-

aggregated into the individual products. These are subsequently scheduled

within the production slots from the schedule roll-out. Here, the objective is

to minimize setups without exceeding flow time.

44 Multi-Level Capacitated Lotsizing with Setup Carryover

B.3.4 Metaheuristics

Glover (1986) introduces the term ‘metaheuristics’ at first. Till this day, there

exists no agreed definition on the expression ‘metaheuristics’. However, some

characteristics common to all metaheuristic methods can be identified. Meta-

heuristics are non problem-specific strategies, which guide the search process.

They may make use of domain-specific knowledge in the form of heuristics

that are controlled by the upper level strategy. This results in flexibility and

the ability to handle large and complex problems. They are usually non-

deterministic and may incorporate mechanisms to avoid getting trapped in

confined areas of the search space. Furthermore, the search space may also

include infeasible solutions, where the violation of constraints is charged at

penalty costs.

a given initial solution. The two basic principles that largely determine the

behaviour of a metaheuristic are intensification and diversification. The latter

enhances the exploration of the search space while the former allows for the

exploitation of the accumulated search experience.

variables for setups as well as continuous variables for production decisions.

Furthermore, some heuristics are restricted to finding the optimal binary grid.

Here, the respective production plan can be derived via methods of linear pro-

gramming, heuristic approaches or dual reoptimization, as described above.

Alternatively, an indirect encoding can be used, as was done e. g. by Kohlmor-

gen et al. (1999). They use a priority rule for selecting items to build a feasible

production plan.

cal Search procedures to complex learning processes. Today’s more advanced

metaheuristics use search experience to guide the search. The heuristics pre-

sented in the following are:

• Local Search

Chapter B. Literature Review 45

• Simulated Annealing

• Tabu Search

• Genetic Algorithms

Table B.5 gives an overview on papers dealing with metaheuristics for capaci-

tated lotsizing problems.

Salomon et al. (1993) SA, TS ML

Kuik et al. (1993) SA, TS ML

Salomon et al. (1993) SA, TS ML

Haase (1994) LS SL, SC

Helber (1994) SA, TS GA ML, ST

Haase and Kohlmorgen (1995) GA SL

Helber (1995) SA, TS GA ML, ST

Hindi (1996) TS SL

Haase (1998) LS SL, SC

Hung et al. (1999) GA SL, ST, BO

Kohlmorgen et al. (1999) GA SL

Barbarosoglu and Özdamar (2000) SA ML, ST

Hung and Chien (2000) SA, TS GA ML, ST, BO

Özdamar and Barbarosoglu (2000) SA ML, ST

Özdamar and Bozyel (2000) SA ML, ST, OV

Gopalakrishnan et al. (2001) TS SL, SC

Gutierrez et al. (2001) GA ML

Özdamar et al. (2002) SA, TS GA SL, ST, OV

Xie and Dong (2002) GA ML, ST, OV

Chen and Chu (2003) LS ML

Hindi et al. (2003) VNS SL, ST

Hung et al. (2003) TS SL, ST, BO

Berretta and Rodrigues (2004) MA ML, ST

Berretta et al. (2005) SA, TS ML, ST, LT

Pitakaso et al. (2006) ACO ML, ST

Abbreviations:

NHS Neighborhood Search (LS = Local Search, SA = Simulated Annealing, TS = Tabu

Search, VNS = Variable Neighborhood Search)

PB Population based Heuristics (ACO = Ant Colony Optimization, GA = Genetic

Algorithms, MA = Memetic Algorithms)

MOD Problem solved (ML = Multi-Level, SL = Single-Level, BO = Backorders, OV =

Overtime, LT = Lead Times, SC = Setup Carryover, ST = Setup Time)

46 Multi-Level Capacitated Lotsizing with Setup Carryover

is described here as all procedures based on neighborhood search (e. g. Simu-

lated Annealing) follow the principals of LS. LS methods, also called iterative

improvement or hill-climbing methods, are based on the idea that at least a

local optimum can be found by starting from a given solution and iteratively

trying to find a better one in an appropriately defined neighborhood. Here,

the neighborhood is defined as the set of solutions, which can be obtained from

the current by performing simple modifications, called moves. In the field of

lotsizing, possible moves include lotshifting, lotsplitting, as well as elimina-

tion and creation of setup decisions. LS procedures may also vary according

to whether they explore the complete or only part of the neighborhood of a

solution and the stopping criterion applied.

based on a randomized regret measure to solve the CLSPL. Haase (1998)

embeds the same method in a LS procedure, where the regret measure is de-

termined iteratively resulting in the best solution value.

Chen and Chu (2003) combine a LR approach with LS. In every iteration, the

feasible solution constructed during the LR phase is further improved on by

changing the values of two setup variables simultaneously during a following

LS.

As mentioned above, Simulated Annealing (SA), Tabu Search (TS) and Vari-

able Neighborhood Search (VNS) can be seen as intelligent extensions to LS.

They try to escape from local optima in order to find a better solution nearby

the global optimum. However, an appropriate stopping criterion is required,

when the algorithm is designed to overcome local optima.

patrick et al. (1983) and can be seen as a combination of iterative improve-

Chapter B. Literature Review 47

ment and random walk. Neighboring solutions are randomly sampled and

the search is guided by an acceptance probability for worse solutions, which

decreases iteratively.

Kuik et al. (1993) combine their R&H with a SA approach for the MLCLSP

without setup times.

termine a good setup pattern for the MLCLSP.12 Furthermore, he examines

the impact of different initial solutions on the overall solution quality. He

compares a plan generated by the heuristic of Tempelmeier and Helber (1994)

against a plan, where every product is produced in every period.

(1993), Hung and Chien (2000), Barbarosoglu and Özdamar (2000), Özdamar

and Barbarosoglu (2000), Özdamar et al. (2002), Özdamar and Bozyel (2000)

and Berretta et al. (2005).

The distinct characteristic giving the name to TS is that it operates so-called

tabu lists. Tabu lists store information on executed moves to prevent their

reversal. The tabu lists’ length is decisive for the size of the section of the so-

lution space that is searched. More advanced algorithms apply adaptive tabu

lists, which depend on the quality of the visited solutions (improvement or de-

terioration). To prevent the exclusion of beneficial moves, so-called aspiration

criteria can be defined (e. g. if the corresponding solution is the best found so

far), which allow tabu moves to be executed.

of their CG heuristic for the CLSP without setup times.

Hung et al. (2003) solve the CLSP with backlogging with a TS method using

a new candidate list strategy. Neighboring solutions are generated by minor

12 See also Helber (1994).

48 Multi-Level Capacitated Lotsizing with Setup Carryover

and post-optimization informations are used to conduct the next move.

different move types for the sequencing and lotsizing decisions.

TS was also applied by Salomon (1991), Kuik et al. (1993), Helber (1994),

Helber (1995), Hindi (1996), Hung and Chien (2000), Özdamar et al. (2002),

and Berretta et al. (2005).

(1999) and Hansen and Mlaydenović (2001). The process consists of the three

phases. These are shaking, Local Search and move. During the shaking phase,

a neighboring solution is randomly selected from each current solution. Sev-

eral neighborhoods exist and solutions are selected from an increasingly vast

neighborhood, if no improvement is made. Then, the neighboring solution

becomes the starting point of a LS method. The resulting solution is finally

compared to the current. If it is beneficial, the move is executed.

and VNS. Starting with the solution obtained by the LR, a smoothing heuristic

similar to that presented by Trigeiro et al. (1989) is used to eliminate capacity

infeasibility and to improve the incumbent best solution during the VNS.

(EA). They are population-based and incorporate a learning component via

the recombination of solutions. Other methods of this kind are Evolution-

ary Programming (EP) and Evolutionary Strategies (ES). However, they have

rarely been used for lotsizing problems and are thus not of interest here.

GAs are inspired by the principles of natural selection. They evolve a pop-

Chapter B. Literature Review 49

of offsprings through an iterative process until some convergence criteria or

conditions are met. The best chromosome is then decoded. This provides the

corresponding solution. The underlying reproduction process is mainly aimed

at improving the fitness of individuals, i. e. the costs to be minimized while

exploring the solution space. The algorithm applies stochastic operators such

as selection, crossover and mutation on an initially random population in order

to compute a new generation of individuals.

porates SA and GA into an LR scheme.

represented by a string of real values, which control a heuristic method for

generating a production plan.

Further Genetic Algorithms have been proposed for lotsizing by Helber (1994),

Helber (1995), Haase and Kohlmorgen (1995), Hung et al. (1999), Hung and

Chien (2000), Özdamar and Bozyel (2000), Gutierrez et al. (2001), Özdamar

et al. (2002) and Xie and Dong (2002).

Basically, they combine LS with crossover operators. Although population

based, Memetic Algorithms are not restrained by a biological metaphor.

Rather, the aim is to exploit all available knowledge about the studied prob-

lem. Knowledge such as optimal solution properties, heuristics and truncated

exact algorithms is available to all individuals. However, they can rely on

different techniques for its exploitation.

Berretta and Rodrigues (2004) propose a MA for the MLCLSP based on the

work of França et al. (1997).

Ant Colony Optimization techniques (ACO) are inspired by the real world’s be-

havior of ant colonies. During their search for food, ants lay down pheromone

50 Multi-Level Capacitated Lotsizing with Setup Carryover

trails. Following ants usually choose the trail with the highest concentration

of pheromone. The reason is that satisfactory paths possess a higher concen-

tration of pheromone than unsatisfactory ones.

a F&R heuristic for the MLCLSP. During the F&R heuristic only a subset

of the constraints and variables are considered relating to a reduced num-

ber of products and periods. The corresponding MIP is solved to optimality

and the binary variables are fixed. During the ACO part of the heuristic,

pheromone concentrations, which have been built up during previous itera-

tions, are used to single out promising subsets for the next iteration. Subse-

quently, the pheromone information is updated.

Finally, a set of rather intuitive heuristic algorithms are the Greedy Heuristics

(GH). They increase lotsizes successively to achieve cost savings.

select the best candidate. The former ensures feasibility of the overall schedule,

i. e. that all demand is served without backlogging and capacity violation. The

latter serves as a cost criterion to manipulate scheduling decisions.

and look-ahead mechanisms. The former are usually used in backward routines

and push infeasible production to earlier periods. The look-ahead mechanisms,

normally used in forward routines, compute the minimum inventory needed to

avoid future capacity violations a priori and lots are scheduled accordingly.

The priority index is often some criterion taken from uncapacitated dynamic

lotsizing heuristics. These criteria in turn stem from properties of the optimal

production quantity in case of uncapacitated lotsizing with constant demand,

e. g. Silver Meal (SM), Least Unit Cost (LUC), Groff (GR) and Part Period

Balancing (PPB). In this way, the Silver Meal criterion (Silver and Meal (1969)

and Silver and Meal (1973)) is based on the fact that for the respective optimal

Chapter B. Literature Review 51

production quantity the mean costs per time unit are minimal. The Least Unit

Cost Criterion is respectively based on the fact that the mean costs per unit are

minimal. The Groff criterion (Groff (1979)) is based on the fact that marginal

holding costs per time unit equal marginal setup costs per time unit. Finally,

the Part Period Balancing criterion (DeMatteis (1968)) states that additional

holding costs to cover another periods demand equal setup costs.

Eisenhut (1975) C F (FB) PPB SL

van Nunen and Wessels (1978) I (WW) B 1P CPP SL, ST, OV

Lambrecht and Vanderveken (1979) C F FB SM SL

Dixon and Silver (1981) C F LA SM SL

Dogramaci et al. (1981) C F LA SM SL

Dogramaci et al. (1981) I (L4L) G SP LTC SL

Karni and Roll (1982) I (WW) G SP LTC SL

Maes and van Wassenhove (1986c) C F LA SM, LUC, LTC, SL

PPB

Günther (1987) I (L4L) F LA CPP (G) SL

Selen and Heuts (1989) I (L4L) F LA CPP (G) SL

Trigeiro (1989) I (L4L) F, B SP CPP (mod SM) SL, ST,

(OV)

Clark and Armentano (1995) I (WW) F, B SP CPO ML, ST

França et al. (1997) I (WW) F, B SP CPO ML, ST

Özdamar and Barbarosoglu (1999) I (L4L) G MC SL

Gupta and Magnusson (2005) C F FB CD SL, ST, SC

Boctor and Poulin (2005) I (L4L) F FB PPB, MC, AC SL, ST,

(OV), MM

Abbreviations:

TGR Type of Greedy Algorithm (C = Constructive, I = Improving, L4L = Lot-for-Lot,

WW = Wagner-Whitin)

DIR Direction of Greedy Algorithm (B = Backward, F = Forward, G = Considering the

complete Schedule at each Step)

FEAS Type of feasibility Check (1P = Lots are respectively shifted back one Period - Slack

Capacity is shifted forward, FB = Feedback, LA = Look-AheadSP = Smoothing

Procedure)

CC Cost Criterion for Choice of Item (AC = Exact Average Cost Savings, CD = Cu-

mulative Demand, CPO = Cost per Overtime Reduction, CPP = Cost per Setup

and Production Time, G = Groff, LTC = Least Total Cost, LUC = Least Unit

Cost, MC = Marginal Costs, PPB = Part Period Balancing, SM = Silver-Meal)

MOD Problem solved (ML = Multi-Level, SL = Single-Level, MM = Multi-Machine, OV

= Overtime, SC = Setup Carryover, ST = Setup Time, )

itself has terminated. These include enforcing the extreme flow conditions, lot

eliminations and lot merging or interchange.

52 Multi-Level Capacitated Lotsizing with Setup Carryover

A distinction is made between GAs starting from scratch and working period-

by-period and those starting from a given initial solution. Thus, the two types

of GAs are:

• Improvement Heuristics

common-sense or specialized heuristics – generate a solution from scratch by

adding components to an initially empty partial solution, until the solution

is complete. Most constructive GHs work period-by-period either forward or

backward. They are myopic, in that at each step the current lotsize is in-

creased ignoring future costs. The first heuristic of this type was introduced

by Eisenhut (1975). However, the use of a feasibility routine is only suggested,

rather than implemented here. Further heuristics of this type include Lam-

brecht and Vanderveken (1979), Dixon and Silver (1981), Dogramaci et al.

(1981) and Drexl and Haase (1995).

Özdamar and Barbarosoglu (1999) solve the SLULSP type subproblems of the

Lagrangean relaxation algorithm, respecting an upper bound on capacity, with

the help of a greedy heuristic.

Maes and van Wassenhove (1986c) suggest a so-called ABC heuristic. It con-

sists of seventy-two alternative algorithms based on the combination of (a) six

a priori orderings of items, (b) four cost criteria and (c) three search strate-

gies. They suggest applying various combinations and then choosing the best

solution.

Gupta and Magnusson (2005) solve the CLSPL with setup times. First, they

generate a feasible solution to the CLSP without setup times with the help

of a GH. Subsequently, setup times are introduced under the assumption that

a setup carryover is installed for each period and a smoothing procedure is

applied. Finally, a backward oriented improvement procedure is applied.

Chapter B. Literature Review 53

from a usually infeasible inferior starting solution by simple shifting routines.

Heuristics of this type have been presented amongst others by van Nunen and

Wessels (1978), Günther (1987), Selen and Heuts (1989) and Trigeiro (1989).

Selen and Heuts (1989) modify the priority index used by Günther (1987) to

account for setup time savings in case a complete lot is shifted.

Boctor and Poulin (2005) solve a CLSP with multiple capacity restrictions. It

arises, when a serial multi-stage system is aggregated to a single-stage system

by the assumption that lotsizes are the same on each stage. They improve

on a lot-for-lot schedule. First, a feasible schedule is aimed to be achieved in

a forward pass. Through all stages, residual capacity after meeting demand

is calculated. If it is positive, the lot is further increased, if it is negative,

production is shifted into earlier periods. Finally, the solution is improved

improved by by shifting complete lots backward (lot merging).

by which they improve on a lot-for-lot schedule. The four steps consist of:

(the lots leading to the overall greatest cost decrease are chosen) without

incurring overtime,

tions,

while respecting capacity limitations.

Karni and Roll (1982) initially solve the SLULSP optimally for each item.

Then, they apply an algorithm similar to Dogramaci et al. (1981).

Clark and Armentano (1995) solve the MLCLSP. While the starting routine

computes a solution to the MLULSP by solving a series of SLULSPs along

the product structure, the smoothing procedure shifts production backward

54 Multi-Level Capacitated Lotsizing with Setup Carryover

and forward to obtain a feasible schedule. The priority rule is based on ad-

ditional costs (including penalty costs for the overall capacity violations) over

the incurred reduction in capacity violation. Penalty costs for capacity re-

duction are gradually increased at each shifting cycle. França et al. (1997)

extend the algorithm by adding improvement and merging. The smoothing

procedure again shifts production quantities forward and backward, but only

if this decreases costs and without violating the capacity restriction. Merging

shifts lots to periods with an existing setup regardless the capacity violation.

Smoothing, improvement and merging are repeated alternately until a given

number of iterations is reached.

Maes and van Wassenhove (1986a) and Maes and van Wassenhove (1986b)

compare the heuristics of Lambrecht and Vanderveken (1979), Dixon and Silver

(1981) and Dogramaci et al. (1981), when applied on static and rolling horizon

problems. Salomon et al. (1993) suggest a modified GH to solve the remaining

LP problem for a given setup pattern.

Chapter C

Test Instances

well as the parameter combinations and the performance of the overall heuristic

presented in chapter F, we introduce a set of new test instances. Similar to e. g.

Derstroff (1995) and Sürie (2005b) we generated various differing instances as

to diminish the effect of the characteristics of the test set on the results.

times, the relevant distinguishing characteristics are:1

1. the size of the problem (i. e. the number of items, the number of re-

sources, the number of production levels and the length of the planning

horizon),

1 Compare Derstroff (1995), page 90.

55

L. Buschkühl, Multi-Level Capacitated Lotsizing with Setup Carryover,

Edition KWV, https://doi.org/10.1007/978-3-658-24034-9_3

56 Multi-Level Capacitated Lotsizing with Setup Carryover

For our test sample, these characteristics will be described in the remainder

of this chapter.

In the following, the number of production levels will equal the number of

resources. Hence, the problem size is made up of the number of items, the

number of resources and the number of periods. We will consider the six

combinations depicted in table C.1.

1 10 3 4 480

2 10 3 8 480

3 20 6 8 240

4 20 6 16 240

5 40 6 8 240

6 40 6 16 240

assembly and general.2 In serial product structures, every item has at most

one predecessor and at most one successor. In divergent (assembly) product

structures, each item has at most one predecessor (successor), but can have an

unlimited number of successors (predecessors). In general product structures

finally, there is neither a limit on the number of predecessors nor on the number

of successors. Hence, the general product structure is the most complex and

the other three can be seen as simplifications of it.

2 Compare Derstroff (1995), page 21.

Chapter C. Test Instances 57

For our tests, we will consider assembly and general product structures. Fur-

thermore, we choose to set the production coefficient akj to 1 for all items k

and j, where j is a direct successor of k.

As regards the process structures, cyclic and acyclic production processes can

be distinguished. In the latter case, items are produced on a different resource

than their predecessors and their successors. In cyclic process structures, how-

ever, some parents are produced on the same resource as their component. In

the following, we will consider both types of production processes.

The product and process structures are given for 10 items and 3 resources in

figure C.1.

1 2 3 1 A

A A A A

4 5 6 7 2 3 4 B

B B B B B B B

8 9 10 5 6 7 8 9 10 C

C C C C C C C C C

1 2 3 1 A

A A B A

4 5 6 7 2 3 4 B

A B C C C B B

8 9 10 5 6 7 8 9 10 C

B C C C C B B C C

Figure C.1: Product and process structures for 10 items and 3 resources

58 Multi-Level Capacitated Lotsizing with Setup Carryover

The product and process structures for 20 items and 6 resources are depicted

in figure C.2.

1 1 A

A A

2 3 2 3 B

B B B B

4 5 6 7 4 5 6 7 C

C C C C C C C C

8 9 10 8 9 10 D

D D D D D D

11 12 13 14 15 11 12 13 14 15 16 E

E E E E E E E E E E E

16 17 18 19 20 17 18 19 20 F

F F F F F F F F F

1 1 A

A A

2 3 2 3 B

B B B C

4 5 6 7 4 5 6 7 C

C C C C C C D C

8 9 10 8 9 10 D

F C D D D E

11 12 13 14 15 11 12 13 14 15 16 E

F F E E D E E E D F F

16 17 18 19 20 17 18 19 20 F

F F F F E F F E F

Figure C.2: Product and process structures for 20 items and 6 resources

Finally, the process and product structures for 40 items and 6 resources are

given in figure C.3.

Chapter C. Test Instances 59

1 2 3 4 5 6 1 A

A A A A A A A

7 8 9 10 11 12 2 3 4 B

B B B

B B B B B B

13 14 15 16 17 18 19 5 6 7 8 9 C

C C C C C

C C C C C C C

20 21 22 23 24 25 26 10 11 12 13 14 15 16 D

D D D D D D D

D D D D D D D

27 28 29 30 31 32 17 18 19 20 21 22 23 24 25 26 27 28 29 E

E E E E E E E E E E E E E E E E E E E

33 34 35 36 37 38 39 40 30 31 32 33 34 35 36 37 38 39 40 F

F F F F F F F F F F F F F F F F F F F

1 2 3 4 5 6 1 A

A A A A A A A

7 8 9 10 11 12 2 3 4 B

C C B C B B C B B

13 14 15 16 17 18 19 5 6 7 8 9 C

C C C C D C C C C C C C

20 21 22 23 24 25 26 10 11 12 13 14 15 16 D

E D C D D D E D D E D D E D

27 28 29 30 31 32 17 18 19 20 21 22 23 24 25 26 27 28 29 E

F E D D E F E E E E F F E E E F F E E

33 34 35 36 37 38 39 40 30 31 32 33 34 35 36 37 38 39 40 F

F F F E E F F F F F F F F F F F F F F

Figure C.3: Product structures and process structures for 40 items and 6

resources

truncated, such that negative values were set to zero. To generate demand,

we used the ‘Mersenne Twister’ as described in Matsumoto and Nishimura

(1998). The means for the respective end items and the components are given

in table C.2.

0.2, 0.5 and 0.8. They were chosen as to reflect the influence of the volatility

of demand.

60 Multi-Level Capacitated Lotsizing with Setup Carryover

10 (general) 30 (1), 70 (2), 100 (3) 20

10 (assembly) 100 20

20 (general) 100 30

20 (assembly) 80 30

40 (assembly) 90 15

For ease of representation and to ensure feasibility, we set the initial inventory

to 20 for all items. Furthermore, for all predecessor items, preproduction xk0

was computed to match the corresponding demand from all direct successors

in the first period:

X

xk0 = akj · Dj1 (C.1)

j∈Nk

X

Dkt = dkt + akj · Dj,t+1 (C.2)

j∈Nk

10 1 1, 2, 3 4, 5, 6, 7 8, 9, 10

10 2 8, 9, 10 4, 5, 6, 7 1, 2, 3

20 1 1, . . . , 6 7, . . . , 14 15, . . . , 20

20 2 15, . . . , 20 7, . . . , 14 1, . . . , 6

40 1 1, . . . , 13 14, . . . , 27 28, . . . , 40

40 2 28, . . . , 40 14, . . . , 27 1, . . . , 13

Chapter C. Test Instances 61

Setup times are set to 5, 10 and 20 time units in order to reflect the different

impacts of setup time. They are allocated to the different items such that

respectively two different setup time profiles are created for each class. The

individual setup time profiles are shown in table C.3.

When setup and holding costs are considered, the lotsizing problem arises from

the opposing nature of these two cost components. While setup costs decrease

with the lotsize, the holding costs increase and vice versa. Thus, for the nature

of the problem, not the absolute values of the cost components are decisive,

but their ratio. For ease of representation, we set the echelon holding cost

factor ek = 1 for all k. Then, the setup costs result from the target ratio.

We will follow Helber (1994) and Derstroff (1995) in expressing the target cost

ratio via the time between orders (TBO), which results from a given ratio in

the static lotsizing model. With the total demand given in equation (C.2), the

average demand of an item k is computed as:

T

X

Dk = Dkt (C.3)

t=1

s

2 · sk

T BOk = (C.4)

D k · ek

Therefore, for a given TBO profile, the setup costs for item k are calculated

as:

Time between orders are set to 1, 2 and 4. For instances with 10 items (classes

1 and 2) four different TBO profiles are applied. Three profiles consist of the

62 Multi-Level Capacitated Lotsizing with Setup Carryover

same time between orders for all items. The fourth sets the same time between

orders for all items on the same production level. For the instances with 20

and 40 items, we generated two TBO profiles, respectively. The first consists

of a TBO of 2 for all items. The second again consists of different TBOs for

items produced on different levels. The TBO profiles are given in table C.4.

10 1 1, . . . , 10

10 2 1, . . . , 10

10 3 1, . . . , 10

10 (assembly) 4 1 2, . . . , 4 5, . . . , 10

10 (general) 4 1, . . . , 3 4, . . . , 7 8, . . . , 10

20 1 1, . . . , 20

20 2 1, . . . , 3 4, . . . , 10 11, . . . , 20

40 1 1, . . . , 40

40 (assembly) 2 1, . . . , 4 5, . . . , 16 17, . . . , 40

40 (general) 2 1, . . . , 6 7, . . . , 19 20, . . . , 40

utilization profiles. The absolute capacity is hence computed as to achieve the

desired target utilization. We define capacity utilization, Um , of any resource

m as:

T

X X

(trk · γkt + tbk · qkt )

k∈Km t=1

Um = T

(C.6)

X

bm

t

t=1

The problem is, that neither the production quantities qkt nor the setup de-

cisions γkt are known in advance. Therefore, they have to be approximated

based on the exogenous data.

For ease of representation, the production time is set to tbk = 1 for all k.

Chapter C. Test Instances 63

As Kimms (1997) points out, it is not a good idea to calculate an overall capac-

ity average. The reason is that neglecting the dynamic nature of demand may

lead to feasibility problems in the early periods. Thus, the minimum average

capacity consumption through production times on resource m in periods t̂ to

t is computed as:4

t

X X

tbk · Dkτ

k∈Km τ =t̂

Dmt̂t = (C.7)

t − t̂ + 1

for m ∈ M do

t=1

while t ≤ Lm do

2 3

X 1 7

bm = 6 (tb · D + tr ) ·

t 6 k kτ k

Um 7

6k∈Km 7

t=t+1

wend

while t ≤ T do

t̂ = t

Dm = 0

Dm = D mt̂t

t=t+1

wend

τ = t̂

while τ ≤ t − 1 do

20 1 3

X 1 7

bm 6

:= 6@Dm + trk A ·

τ

6 k∈Km

Um 7

7

τ =τ +1

wend

wend

end for

Algorithm 1: Generation of the capacity matrix

64 Multi-Level Capacitated Lotsizing with Setup Carryover

During the cumulated lead time Lm = max Lk the minimum average capacity

k∈Km

might not be sufficient as lotsizing is limited by predecessor item availability.

Hence, in these first periods, capacity is computed in each period individually.

Like Derstroff (1995), we consider five capacity profiles, which result from

different combinations of the possible capacity utilizations of 50%, 70% and

90% (see table C.5).

3 1 A, B, C

3 2 A, B, C

3 3 A, B, C

3 4 A B C

3 5 C B A

6 1 A,. . ., F

6 2 A,. . ., F

6 3 A,. . ., F

6 4 A, D B, E C, F

6 5 A, B C, D E, F

To test the performance of the heuristic in the case of parallel machines, the

two profiles of parallel machines given in table C.6 are applied.

] Resources Profile Nm = 1 Nm = 2 Nm = 4

3 1 A, B, C

3 2 A B C

6 1 A,. . ., F

6 2 A, D B, E C, F

The profiles are used on subsets of classes 2, 4 and 6. These includes setup

time profile 1 and TBO profile 2 (4 for class 2).

5 See Kimms (1997), page 84.

Chapter D

Model Formulations

In this chapter we will present model formulations for the MLCLSPL with sin-

gle setup carryover, with consecutive setup carryover and with parallel (iden-

tical) machines.

In section D.1, we will introduce the MLCLSPL with single setup carryover

and give the echelon formulation (MLCLSPLe ). This formulation is used for

the Lagrangean heuristic presented in chapter E.

As stated in section B.3.1.1.1, the standard I&L formulation of the MLCLSP

yields rather poor lower bounds, when a B&B algorithm is applied. The

tests given at the end of this section show that also for the MLCLSPL the

lower bounds of the echelon formulation are rather poor. The reformulations

based on the SR representation and the SPL representation prove to be much

stronger. Thus, we will describe the SR and SPL reformulations,1 which are

used to enhance the performance of the B&B algorithm applied by CPLEX.

In section D.2, the additional restrictions for consecutive setup carryovers are

given. They are the same for the echelon formulation used by the heuristic and

1 Compare section B.3.1.1.1.

65

L. Buschkühl, Multi-Level Capacitated Lotsizing with Setup Carryover,

Edition KWV, https://doi.org/10.1007/978-3-658-24034-9_4

66 Multi-Level Capacitated Lotsizing with Setup Carryover

for the SPL formulation, which is applied by CPLEX.2 Again, we present two

sets of constraints, based on the standard formulation and an extended formu-

lation presented by Sürie and Stadtler (2003) along with their computational

evaluation.

Finally, in section D.3, the constraints for parallel machines are given for the

SPL formulation only. The heuristic applies a ‘heuristic model’, 3 in which

parallel resources are aggregated to one resource type. This will be described

in chapter E. The computational results are given for single and consecutive

setup carryover.

Lotsizing Problem with Linked Lotsizes

the MLCLSPL presented in the following is based on the following assump-

tions:

of each item to a specific resource. All items assigned to the same re-

source are interdependent in that they compete on its scarce capacity.

Furthermore, input-output relationships between some of the items exist

(multi-level production).

a successor item, the corresponding derived demand arises before the

end of a period. To ensure feasibility, hence a constant lead time of one

period is accounted for.

(1,. . . ,T ).

2 In the case of single setup carryovers, the SR and the SPL reformulations yield sim-

ilar results. We chose to implement only the SPL formulation for consecutive setup

carryovers.

3 The approach is similar to the one presented in Quadt (2004).

Chapter D. Model Formulations 67

the total number of items K.

• The objective of the model is to minimize the sum of inventory and setup

costs.

• Setup costs are considered as the out of pocket costs arising from a

changeover. Opportunity costs for lost production time during a setup

will not be regarded, as setup times are directly included in the model.

Inventory costs consist of all quantity-dependent cost factors. Produc-

tion costs are assumed to be constant over time and are thus neglected.

demand are exogenous to the model and assumed to be deterministic.

ages.

time option. Setup and production times are considered. They reduce

capacity in the periods they occur. Other time consuming activities,

such as transportation times are ignored.

setup operation is performed period t. This implies that we leave out

the case, in which a setup state could be preserved over two periods, if

no other item was produced in the middle period.

We assume that, as for the CLSP, external demand arises at the end of a

respective period. Thus, all items produced in a given period t are available

to satisfy the external demand in that period. However, the derived demand

in period t arises from the production of successor items, which starts at some

point before the end of t.

Examples are given in figure D.1. If the predecessor and the successor item are

produced on two different resources in the same period, sequential production

68 Multi-Level Capacitated Lotsizing with Setup Carryover

of the two lots may not be possible, even if the capacity constraints are not

violated (resources A and B). Also, when the setup state is determined at

the beginning and the end of a period, it may be necessary to produce the

component as the final item (resource C) or the successor item as the first

item of a period (resource D). In both cases, secondary demand cannot be

fulfilled in the same period, even if both items are produced on the same

resource.

1

4

6

0 1 t

violation

cessor item in the respectively precedent period. Thus, we will set the lead

time for the derived demand to 1. Therefore, ykt and qkt are available at the

end of period t to serve external demand from the same period and derived

demand from the succeeding period.

The basic model formulation of the MLCLSPL can thus be written as:

Chapter D. Model Formulations 69

Model MLCLSPL

K X

X T

min Z = (sk · γkt + hk · ykt ) (D.1)

k=1 t=1

subject to

X

yk,t−1 + qkt − ykt = dkt + akj · qj,t+1 ∀k, t = 0, . . . , T (D.2)

j∈Nk

X

(tbk · qkt + trk · γkt ) ≤ bm

t ∀m, t = 1, . . . , T (D.3)

k∈Km

X

ωkt ≤ 1 ∀m, t = 1, . . . , T (D.5)

k∈Km

ωk1 = 0 ∀k (D.7)

The model formulation differs from the MLCLSP presented in section B.1.2 in

two ways. First, setup carryovers are introduced in constraints (D.4). There-

fore, a positive production quantity enforces a setup activity, only if the setup

70 Multi-Level Capacitated Lotsizing with Setup Carryover

state was not preserved from the previous period. As in the CLSPL presented

in section B.1.3, at most one setup state can be carried over on each resource

(See constraints (D.5)). However, in this basic model formulation, we allow

for a setup carryover, only if a corresponding setup activity took place in the

previous period. Thus, consecutive setup carryovers are not considered. As

Sox and Gao (1999) point out for the CLSPL, for a large class of problems the

optimal solutions of this restricted problem will be close to the optimal solu-

tions of the MLCLSPL with any number of consecutive setup carryovers per

item. The logic is that in situations, where several items have to be produced

on the same resource in most periods, consecutive setup carryovers cannot be

installed. At the same time, introducing consecutive setup carryovers increases

the according computational effort substantially.

Second, as described above, we consider a constant lead time of one period for

secondary demand. Consequently, production of successor items cannot take

place in the first period of the planning horizon, unless the initial inventories

of their components are sufficiently large. Hence, the inventory balance con-

straints (D.2) are considered from period 0 to link the production quantities of

successor items in period 1 to the inventory of their components in period 0.

Accordingly, initial inventories are defined in period −1 by constraints (D.8).

We will follow Billington et al. (1983), Billington et al. (1986) and Derstroff

(1995) in eliminating the inventory variable ykt from the model. The inventory

of an item k available at the end of period t can be computed as the initial

inventory plus cumulative production less cumulative demand:

t

X t

X X

ykt = (qkτ − dkτ ) − akj · qj,τ +1 + ybk (D.11)

τ =1 j∈Nk τ =0

With the new interpretation of inventory, the sum of inventory holding costs

Chapter D. Model Formulations 71

is calculated as:4

K X

X T K X

X T t

X t

X X

hk · ykt = hk · (qkτ − dkτ ) − akj · qj,τ +1 + ybk

k=1 t=1 k=1 t=1 τ =1 j∈Nk τ =0

(D.12)

K XT

Ã t

!

X X

hk · ybk − dkτ (D.13)

k=1 t=1 τ =1

XK XT t

X t

X X

hk · qkτ − akj · qj,τ +1 (D.14)

k=1 t=1 τ =1 j∈Nk τ =0

(D.13) consists only of exogenous data and is thus the same for all possible

solutions to the model. The value of (D.14) depends on the specific solution.

Thus, we will focus on equation (D.14) as to find a more convenient formula-

tion.

XK X T t

X X X t+1

hk · qkτ − akj · qjτ

k=1 t=1 τ =1 j∈Nk τ =1

K X

X T X

t X K X X

X T

= hk · qkτ − hk · akj · qjτ − (hk · akj · qj,t+1 )

k=1 t=1 τ =1 j∈Nk k=1 j∈Nk t=1

XK XT X t X K X X

X T

hk · qkτ − hk · akj · qjτ − (hk · akj · qjt )

k=1 t=1 τ =1 j∈Nk k=1 j∈Nk t=2

72 Multi-Level Capacitated Lotsizing with Setup Carryover

As all items and all periods are considered, changing from the perspective of

the predecessor to that of the successor results in:5

XK X T X t X K X

X T X

qkτ · hk − hj · ajk − hj · ajk · qkt (D.15)

k=1 t=1 τ =1 j∈Vk k=1 t=2 j∈Vk

Let ek be the marginal holding cost coefficient for item k. Interpreting holding

costs as costs of fixed capital, the difference between full holding costs and

marginal holding costs is that the former account for the complete value of

a product, while the latter only account for its value added. The marginal

holding cost coefficient can thus be derived from the full holding cost coefficient

simply by subtracting the cumulative value of its components:6

X

ek = hk − ajk · hj

j∈Vk

With this notion of the marginal holding cost coefficient equation (D.15) dis-

solves to:

K X

X T X

t K X

X T

ek · qkτ − (hk − ek ) · qkt (D.16)

k=1 t=1 τ =1 k=1 t=2

P

K P

T

The additional component − (hk − ek ) · qkt reduces holding costs. The

k=1 t=2

logic behind this is that the production of a successor item reduces inventory of

its components one period earlier than it adds to its own inventory. Through-

out the lead time, the respective quantity of the components is thus regarded

as work in process inventory and does incur holding costs.7

total (external and secondary) demand, when there is no inventory at the end

5 Compare Derstroff (1995), page 72.

6 Obviously, for all items without predecessors the marginal holding cost coefficient equals

the full holding cost coefficient.

7 Note that the only reason for the earlier production of predecessor items is that derived

demand arises before the end of a period. Remember that this not only effects a reduc-

tion of holding costs in equation (D.16), but also an increase in holding costs via initial

inventory (see equation (D.13)) and earlier production of predecessors.

Chapter D. Model Formulations 73

T

X T

X

qkt = Dkt (D.17)

t=1 t=1

lows:

Dkt =

X t−1

X X

max 0, dkt + akj · Dj,t+1 − max{0, ybk − (dkτ + akj · Dj,τ +1 )}

j∈Nk τ =0 j∈Nk

(D.18)

K X T K

Ã T !

X X X

(hk − ek ) · qkt = (hk − ek ) · Dkt − qk1

k=1 t=2 k=1 t=1

(D.19)

K X

X T K

X

= (hk − ek ) · Dkt − (hk − ek ) · qk1

k=1 t=1 k=1

Again, the first part of equation (D.19) is constant and the second variable.

Adding the former to (D.13) yields:

K XT

Ã t

! K X T

X X X

hk · ybk − dkτ − (hk − ek ) · Dkt

k=1 t=1 τ =1 k=1 t=1

(D.20)

K X

X T K X

X T

=− hk · ((T − t + 1) · dkt − ybk ) − (hk − ek ) · Dkt

k=1 t=1 k=1 t=1

part of equation (D.19) yields

K

Ã T !

X X

ek · (T − t + 1) · qkt + (hk − ek ) · qk1 . (D.21)

k=1 t=1

74 Multi-Level Capacitated Lotsizing with Setup Carryover

Using (D.20) and (D.21) to replace holding costs and equation (D.11) to replace

the inventory variable, we can write the model MLCLSPLe as:

Model MLCLSPLe

K X

X T K

X

min Z = (sk · γkt + ek · (T − t + 1) · qkt )+ (hk −ek )·qk1 −C (D.22)

k=1 t=1 k=1

subject to

t

X t

X X

(qkτ − dkτ ) + ybk ≥ akj · qj,τ +1 ∀k, t = 0, . . . , T (D.23)

τ =1 j∈Nk τ =0

X

(tbk · qkt + trk · γkt ) ≤ bm

t ∀m, t = 1, . . . , T (D.24)

k∈Km

X

ωkt ≤ 1 ∀m, t = 1, . . . , T (D.26)

k∈Km

ωk1 = 0 ∀k (D.28)

K X

X T K X

X T

C= hk · ((T − t + 1) · dkt − ybk ) + (hk − ek ) · Dkt (D.31)

k=1 t=1 k=1 t=1

Chapter D. Model Formulations 75

Data:

ek echelon holding cost coefficient of item k

The objective function (D.22) differs from (D.1) only in the way holding costs

are written. Constraints (D.23) replace the inventory balance constraints

(D.2). The remaining restrictions are the same as for the MLCLSPL.

sented by Stadtler (1997). A SR formulation for the MLCLSP was presented

on page 25. Total demand Dkt is computed as in equation (D.18) on page 73

and cumulated demand dkts is computed as:

s

X

dkts = Dkτ ∀ k, t, s ≥ t

τ =t

Model MLCLSPLSR

K X

T

Ã T

!

X X

min Z = sk · γkt + ek · (T − t + 1) · dkts · ∆kts

k=1 t=1 s=t

K

X T

X

+ (hk − ek ) · dk1s · ∆k1s − C (D.32)

k=1 s=1

subject to

T

X

∆k1s = 1 ∀k (D.33)

s=1

t−1

X T

X

∆kτ,t−1 = ∆kts ∀k, t = 2, . . . , T (D.34)

τ =1 s=t

76 Multi-Level Capacitated Lotsizing with Setup Carryover

T

X T

X X

yk,t−1 + dkts · ∆kts = dkt + akj · dj,t+1,s · ∆j,t+1,s + ykt

s=t j∈Nk s=t+1

∀k, t = 0, . . . , T (D.35)

Ã T

!

X X

tbk · dkts · ∆kts + trk · γkt ≤ bm

t ∀m, t = 1, . . . , T (D.36)

k∈Km s=t

T

X

∆kts ≤ γkt + ωkt ∀k, t = 1, . . . , T (D.37)

s=t

dkts >0

X

ωkt ≤ 1 ∀m, t = 1, . . . , T (D.38)

k∈Km

ωk1 = 0 ∀k (D.40)

To include the effects of lead times, we adapted the objective function of the

MLCLSPLe simply by replacing the production quantities.8 Furthermore, we

8 Note that the effect that holding costs are computed with the help of production quan-

tities rather than inventories remains.

Chapter D. Model Formulations 77

again divide the objective function into a variable and a constant part. The

constant part is given by equation (D.31) on page 74. The flow constraints

(D.33) and (D.34) are the same as for the MLCLSP.

We use a continuous variable ykt to compute the inventory on hand at the end

of each period. Although not needed to compute holding costs, the additional

variables reduce the number of non-zero coefficients in the inventory balance

constraints significantly.9 The capacity and setup enforcement constraints

(D.36) and (D.37) are adapted. ∆kts is a continuous variable within its bounds

(see constraints (D.42)). The remaining constraints of the MLCLSPLe are

unchanged.

sented by Stadtler (2003). Total demand Dkt is again computed as in equation

(D.18) on page 73. Model MLCLSPLSPL

K X

X T T

X

min Z = sk · γkt + ek · (T − t + 1) · D

δkts

k=1 t=1 s=t

Dks >0

K

X T

X

D

+ (hk − ek ) · δk1s −C (D.45)

k=1 s=1

Dks >0

subject to

t

X

D

δkτ t = Dkt ∀k, t = 1, . . . , T (D.46)

τ =1

T

X T

X X

D D

yk,t−1 + δkts = dkt + akj · δj,t+1,s + ykt ∀k, t = 0, . . . , T

s=t j∈Nk s=t+1

Dks >0 Djs >0

(D.47)

9 See Stadtler (1997), page 89.

78 Multi-Level Capacitated Lotsizing with Setup Carryover

X T

X

tbk · D

δkts + trk · γkt m

≤ bt ∀m, t = 1, . . . , T (D.48)

k∈Km s=t

Dks >0

D

δkts ≤ (γkt + ωkt ) · Ω ∀k, t = 1, . . . , T, s = t, . . . , T, Dks > 0 (D.49)

X

ωkt ≤ 1 ∀m, t = 1, . . . , T (D.50)

k∈Km

ωk1 = 0 ∀k (D.52)

T

X

D

δk0s =0 ∀k (D.54)

s=1

Dks >0

D

δkts ≥0 ∀k, t = 1, . . . , T, s = t, . . . , T, Dks > 0 (D.55)

Variables:

D

δkts demand of item k in period s, which is produced in period t

D

(δkts = δkts · Dks )

Chapter D. Model Formulations 79

Again, we adapt the objective function of the MLCLSPLe rather than adopting

the one presented by Stadtler (2003). The constant part is given by equation

(D.31). As in the SPL formulation of the MLCLSP presented on page 27, con-

straints (D.46) and (D.47) ensure that demand is met in each period. However,

again to reduce the number of non-zero elements, we compute the total de-

D

mand in period s, which is produced in period t (δkts ) rather than a fraction

(δkts ).10 The remaining constraints of the MLCLSPLe are adapted only to the

different way of expressing production.

In order to test the performance of our model formulations and also to generate

benchmark results for the heuristic presented in chapter E, we used the stan-

dard solver CPLEX. The 1920 test instances have been presented in chapter

C.

The tests were carried out on a Sun Microsystems compute server of the Sun

Fire 15K type. On this server 72 UltraSPARC-III processors are available.

The processors’ clock speed is 900 Mhz and the total primary storage amounts

to 144 GB. All models were written in ILOG AMPL 10.000 and solved with

the help of ILOG CPLEX 10.000.

We used the CPLEX standard options, eight parallel processors and a time

limit of 3600 seconds.

number of proven optimal solutions found within the time limit, the average

upper bound, the average mipgap and the average computation time.

the time limit is given in table D.1.

80 Multi-Level Capacitated Lotsizing with Setup Carryover

1 MLCLSPLe 461 96.04

MLCLSPLSR 461 96.04

MLCLSPLSPL 450 93.75

2 MLCLSPLe 87 18.13

MLCLSPLSR 354 73.75

MLCLSPLSPL 364 75.83

3 MLCLSPLe 0 0.00

MLCLSPLSR 140 58.33

MLCLSPLSPL 162 67.50

4 MLCLSPLe 0 0.00

MLCLSPLSR 78 32.50

MLCLSPLSPL 78 32.50

5 MLCLSPLe 0 0.00

MLCLSPLSR 46 19.17

MLCLSPLSPL 53 22.08

6 MLCLSPLe 0 0.00

MLCLSPLSR 15 6.25

MLCLSPLSPL 14 5.83

1−6 MLCLSPLe 548 28.54

MLCLSPLSR 1094 56.98

MLCLSPLSPL 1121 58.39

number of optimal solutions

solution for the test instances in classes 3 − 6. On average, the MLCLSPLSPL

and the MLCLSPLSR formulations perform comparably well on all test in-

stances. Overall, the MLCLSPLSPL formulation arrives at 27 optimal solutions

more than the the MLCLSPLSR formulation.

The final feasible solution (U B ∗ ), the best lower bound (LB ∗ ), the absolute

∗ ∗

mipgap (U B ∗ − LB ∗ ), the relative mipgap ( U BU−LB

B∗ · 100) and the average

computation time in CPU seconds11 are given in table D.2.

Remarkably, the average values of the best feasible solutions are quite similar

for all three model formulations. However, as expected, the lower bounds

provided by the MLCLSPLSR and the MLCLSPLSPL formulations are much

tighter than the ones of the standard MLCLSPLe formulation.

11 Note that the time limit of 3600 seconds selected as a CPLEX option is in real time.

The CPU seconds is the cumulated CPU time on all eight processors.

Chapter D. Model Formulations 81

1 MLCLSPLe 2356.7030 2356.6997 0.0034 0.0001 1.6355

MLCLSPLSR 2356.7020 2356.6969 0.0050 0.0001 0.5947

MLCLSPLSPL 2356.7020 2356.6911 0.0109 0.0004 0.3904

2 MLCLSPLe 5019.8844 4848.8784 171.0060 2.0516 6395.9893

MLCLSPLSR 5019.4604 5018.1805 1.2800 0.0121 304.3478

MLCLSPLSPL 5019.4604 5019.3049 0.1555 0.0017 199.4565

3 MLCLSPLe 12478.8188 10797.9802 1680.8387 12.7432 17477.5079

MLCLSPLSR 12470.3297 12469.9901 0.3396 0.0024 28.5733

MLCLSPLSPL 12470.3282 12470.0205 0.3077 0.0022 38.4938

4 MLCLSPLe 27768.5108 19976.8097 7791.7011 26.7639 22456.9721

MLCLSPLSR 27694.9142 27570.7706 124.1435 0.3420 6580.8155

MLCLSPLSPL 27693.2729 27586.2669 107.0060 0.2989 6879.5365

5 MLCLSPLe 14908.3277 10858.4859 4049.8418 25.8289 20151.0491

MLCLSPLSR 14873.6948 14846.9182 26.7766 0.1336 5976.5498

MLCLSPLSPL 14873.7313 14843.7178 30.0135 0.1501 6556.3044

6 MLCLSPLe 33618.4396 19746.1381 13872.3015 38.8995 25823.1900

MLCLSPLSR 33422.4892 33174.2000 248.2891 0.5533 14945.7895

MLCLSPLSPL 33420.1129 33171.1438 248.9691 0.5554 16971.2004

1−6 MLCLSPLe 12940.9090 9473.8212 3467.0877 13.5424 12337.9961

MLCLSPLSR 12901.7191 12851.4542 50.2649 0.1320 3517.7016

MLCLSPLSPL 12901.2213 12852.8926 48.3287 0.1263 3855.6536

mipgap and computation time

Due to the tighter lower bounds, the average relative mipgap of the reformula-

tions is significantly smaller than that of the standard formulation. This leads

both to a greater significance of the results for the quality of any given solution

and also to less required computation time.

formulation requires slightly less computation time, the MLCLSPLSPL formu-

lation leads to marginally better lower and upper bounds. As we will use the

results as benchmark for our heuristic, we decided to use the MLCLSPLSPL

formulation both for the single setup carryover case and as a basis for the

following extensions.

82 Multi-Level Capacitated Lotsizing with Setup Carryover

As pointed out e. g. by Haase (1994) and Sox and Gao (1999), introducing

multi-period setup carryovers (CC) will generally not lead to great improve-

ments of the overall plan. However, this result is highly dependent on the

structure of the specific problem. E. g. if only few items are produced on the

same resource, the option of producing the same item in several consecutive

periods without incurring an additional setup operation is likely to yield a

better solution.

by:

Constraints CC

ωkt + ωk,t+1 ≤ 1 + ϑm

t ∀m ∈ M, k ∈ Km , t = 1, . . . , T − 1 (D.59)

γkt + ϑm

t ≤1 ∀m ∈ M, k ∈ Km , t = 2, . . . , T (D.60)

ϑm

t ≥0 ∀m ∈ M, t = 1, . . . , T (D.61)

A setup carryover can be installed, whether the setup state was evoked

by a setup activity γk,t−1 = 1 or a setup carryover ωk,t−1 = 1 (see con-

straints (D.58)). However, a consecutive carryover can be installed, only if no

changeover takes place in between. This is guaranteed by constraints (D.59)

12 For the other model formulations, the same logic applies and constraints (D.39) and

(D.51) are replaced respectively.

Chapter D. Model Formulations 83

t is a continuous (see constraints (D.61))

dummy variable, which takes the value 1, if a consecutive carryover is in-

stalled, and thus prohibits changeovers. It takes the value 0, if a changeover

is performed and hence prohibits consecutive carryovers.

Stadtler (2003) to strengthen the model formulation in the case of consec-

utive setups. To do so, the resource-specific dummy variable ϑm

t is replaced

by an item-specific variable ϑkt .

Constraints CCext

X

γkt + ωkt + ϑjt ≤ 1 ∀m ∈ M, k ∈ Km , t = 2, . . . , T (D.64)

j∈Km

j6=k

ϑkt ≥ 0 ∀m ∈ M, t = 1, . . . , T (D.65)

ϑk1 = 0 ∀m ∈ M, t = 1, . . . , T (D.66)

Variables:

ϑkt dummy variable to indicate a consecutive carryover for item k

from period t − 1 to t and to t + 1

Constraints (D.62) and (D.63) together replace constraints (D.58). ϑkt can

only take the value 1, if ωkt is 1. Hence, constraints (D.62) impose the same

84 Multi-Level Capacitated Lotsizing with Setup Carryover

carryover at the same time. Constraints (D.64) ensure for each item k, which

is produced on resource m, that a setup activity in t, a setup carryover at the

beginning of t and a consecutive setup carryover for some other item j from

t − 1 to t + 1 respectively exclude one other. As ϑm

t before, ϑkt is defined as a

continuous variable (see constraints (D.65)), but is forced to either 0 or 1 by

the restrictions of the model. Finally, to ensure that a setup must take place

in the first period, if the setup state is carried over to the second, ϑk1 is set to

0 by constraints (D.66).

The computational comparison of the two variants was carried out on the same

test set and with the same computational setup described in section D.1.5 on

page 79.

Again, we compared the number of proven optimal solutions found within the

time limit, the average upper bound, the average mipgap and the average com-

putation time. We also compared the consecutive setup carryover formulations

to the MLCLSPLSPL formulation.13

At least with respect to the overall number of verified optimal solutions found

within the time limit, the MLCLSPLCC outperforms the MLCLSPLSPL . The

MLCLSPLCCEXT performs poorer. However, this result is mainly due to the

fact that the MLCLSPLCC formulation finds the optimal solution for 90.83% of

the test instances in class 2. In class 1, both consecutive carryover formulations

outperform the MLCLSPLSPL . For all other classes, they perform similarly

well and slightly poorer than the single carryover formulation.

13 Remember, that both types of model formulations are defined for the same underlying

problem. Prohibiting consecutive setup carryovers is supposed to reduce computation

time without significant deterioration of the solution quality.

Chapter D. Model Formulations 85

1 MLCLSPLSPL 450 93.75

MLCLSPLCC 458 95.42

MLCLSPLCCEXT 467 97.29

2 MLCLSPLSPL 364 75.83

MLCLSPLCC 436 90.83

MLCLSPLCCEXT 362 75.42

3 MLCLSPLSPL 162 67.50

MLCLSPLCC 139 57.92

MLCLSPLCCEXT 143 59.58

4 MLCLSPLSPL 78 32.50

MLCLSPLCC 65 27.08

MLCLSPLCCEXT 78 32.50

5 MLCLSPLSPL 53 22.08

MLCLSPLCC 47 19.58

MLCLSPLCCEXT 47 19.58

6 MLCLSPLSPL 14 5.83

MLCLSPLCC 18 7.50

MLCLSPLCCEXT 17 7.08

1−6 MLCLSPLSPL 1121 58.39

MLCLSPLCC 1163 60.63

MLCLSPLCCEXT 1114 58.13

number of optimal solutions

The upper bounds, lower bounds, mipgaps and computation times are pre-

sented in table D.4. Notably, the average CPU time required is only

7.4928% (4.9025%) higher for the MLCLSPLCC (MLCLSPLCCEXT ) than for

the MLCLSPLSPL . While the relative mipgaps are respectively 26.6762%

(28.1868%) larger in the consecutive carryover case, both the average up-

per bounds and the average lower bounds are lower (about 3% for the upper

bounds).

all, we will use its results as a benchmark for our heuristic in chapter F.

86 Multi-Level Capacitated Lotsizing with Setup Carryover

1 MLCLSPLSPL 2356.7020 2356.6911 0.0109 0.0004 0.3904

MLCLSPLCC 2250.4603 2250.4515 0.0088 0.0002 0.8933

MLCLSPLCCEXT 2250.4603 2250.4554 0.0049 0.0001 0.4105

2 MLCLSPLSPL 5019.4604 5019.3049 0.1555 0.0017 199.4565

MLCLSPLCC 4528.4330 4527.8398 0.5932 0.0052 163.6972

MLCLSPLCCEXT 4528.4330 4528.0725 0.3604 0.0042 230.1083

3 MLCLSPLSPL 12470.3282 12470.0205 0.3077 0.0022 38.4938

MLCLSPLCC 12097.4853 12097.1265 0.3587 0.0026 57.2062

MLCLSPLCCEXT 12097.4863 12097.1442 0.3421 0.0024 77.3646

4 MLCLSPLSPL 27693.2729 27586.2669 107.0060 0.2989 6879.5365

MLCLSPLCC 26607.6167 26445.3021 162.3146 0.4657 8959.5282

MLCLSPLCCEXT 26605.5829 26452.2365 153.3464 0.4331 7608.3359

5 MLCLSPLSPL 14873.7313 14843.7178 30.0135 0.1501 6556.3044

MLCLSPLCC 14695.6663 14663.0278 32.6384 0.1658 7549.3468

MLCLSPLCCEXT 14695.6544 14652.3530 43.3013 0.2235 6894.5844

6 MLCLSPLSPL 33420.1129 33171.1438 248.9691 0.5554 16971.2004

MLCLSPLCC 33087.4208 32805.5853 281.8355 0.6361 16227.2758

MLCLSPLCCEXT 33087.1675 32808.4020 278.7655 0.6277 17283.2005

1−6 MLCLSPLSPL 12901.2213 12852.8926 48.3287 0.1263 3855.6536

MLCLSPLCC 12505.7470 12445.9530 59.7939 0.1601 4144.6346

MLCLSPLCCEXT 12505.4597 12445.8990 59.5608 0.1619 4044.7787

mipgap and computation time

can be used to produce the same set of items alternately. For each resource m,

a set Pm is defined, which contains the number of parallel resources (at least

1) available.

Dn n n

Hence, now production δkτ t , setup activities γkt and setup carryovers ωkt of

an item k in a period t are all related to one specific parallel resource n.

(D.48) to (D.57) are replaced by constraints (D.67) to (D.77).

Chapter D. Model Formulations 87

Constraints PM

X T

X

tbk · Dn

δkts n

+ trk · γkt m

≤ bt ∀m, n ∈ Pm , t = 1, . . . , T (D.67)

k∈Km s=t

Dks >0

Dn n n

δkts ≤ (γkt +ωkt )·Ω ∀m, k ∈ Km , n ∈ Pm , t = 1, . . . , T, s = t, . . . , T, Dks > 0

(D.68)

X

n

ωkt ≤1 ∀m, n ∈ Pm , t = 1, . . . , T (D.69)

k∈Km

n n

ωkt ≤ γk,t−1 ∀m, k ∈ Km , n ∈ Pm , t = 2, . . . , T (D.70)

n

ωk1 =0 ∀m, k ∈ Km , n ∈ Pm (D.71)

Pm

X

n

γkt = γkt ∀m, k ∈ Km , t = 1, . . . , T (D.72)

n=1

Pm

X

D Dn

δkts = δkts ∀m, k ∈ Km , t = 1, . . . , T, s = t, . . . , T, Dks > 0 (D.73)

n=1

Dn

δkts ≥0 ∀m, k ∈ Km , n ∈ Pm , t = 1, . . . , T, s = t, . . . , T, Dks > 0 (D.74)

D

δkts ≥0 ∀m, k ∈ Km , t = 1, . . . , T, s = t, . . . , T, Dks > 0 (D.75)

n n

γkt , ωkt ∈ {0, 1} ∀m, k ∈ Km , n ∈ Pm , t = 1, . . . , T (D.77)

88 Multi-Level Capacitated Lotsizing with Setup Carryover

Indices:

n parallel machines, n = 1 . . . Pm

Pm number of parallel machines of resource type m

Variables:

n

γkt binary setup variable on machine n

n

ωkt binary setup state variable for item k on machine n at the be-

ginning of period t

Dn

δkts production quantity of item k on machine n in period t to satisfy

demand in period s

more than one machine of each resource type may be used to produce the

same products. Constraints (D.72) and (D.73) cumulate the setup activities

and production quantities respectively over all parallel resources. As a result,

D

production quantities δkts and setups γkt are given per item and period and

the objective function, as well as the inventory balance constraints remain

unchanged. Both variables are continuous (see constraints (D.75) and (D.76)).

The only difference in the remaining constraints is that they are all defined for

each of the parallel resources.

changing the index of ϑm n

t to ϑt and defining each constraint for each indi-

vidual machine n of each resource type m.

Again, the computational study was carried out using the same setup described

in section D.1.5 on page 79. The test set for parallel machines in described in

chapter C.

Chapter D. Model Formulations 89

class model ] feasible sol. % feasible sol. ] optimal sol. % optimal sol.

2/1 MLCLSPLPM 60 100.00 27 45.00

MLCLSPLPMCC 60 100.00 20 33.33

2/2 MLCLSPLPM 60 100.00 15 25.00

MLCLSPLPMCC 60 100.00 17 28.33

4/1 MLCLSPLPM 60 100.00 3 5.00

MLCLSPLPMCC 59 98.33 1 1.67

4/2 MLCLSPLPM 60 100.00 2 3.33

MLCLSPLPMCC 57 95.00 0 0.00

6/1 MLCLSPLPM 60 100.00 0 0.00

MLCLSPLPMCC 55 91.67 0 0.00

6/2 MLCLSPLPM 60 100.00 0 0.00

MLCLSPLPMCC 18 30.00 0 0.00

all MLCLSPLPM 360 100.00 47 13.06

MLCLSPLPMCC 309 85.83 38 10.56

number of optimal solutions

Because of the increased complexity of the models with parallel machines, only

few optimal solutions are found for instances outside class 2.14 Furthermore,

for consecutive setup carryovers, the time limit is repeatedly met with no

feasible solution.

2/1 MLCLSPLPM 4173.8950 4146.0824 27.8126 0.4426 4200.9245

MLCLSPLPMCC 3452.9543 3376.3566 76.5977 1.4823 7073.6980

2/2 MLCLSPLPM 4078.3925 3974.8070 103.5855 2.1383 8879.9980

MLCLSPLPMCC 2619.5473 2425.3135 194.2338 5.4332 13738.5947

4/1 MLCLSPLPM 31477.3183 30437.6641 1039.6542 2.8738 22544.7356

MLCLSPLPMCC 29075.5847 26268.5335 2807.0513 9.1934 25098.4803

4/2 MLCLSPLPM 30860.4767 28442.9821 2417.4946 6.8579 19499.2483

MLCLSPLPMCC 21062.8318 15616.3403 5446.4914 23.0250 24530.5735

6/1 MLCLSPLPM 38601.8583 37125.1021 1476.7563 3.4456 27184.9962

MLCLSPLPMCC 38067.6582 35036.8236 3030.8346 7.2807 22248.6933

6/2 MLCLSPLPM 35548.8033 32890.7854 2658.0180 6.3960 21202.7968

MLCLSPLPMCC 42570.0278 27798.5472 14771.4806 25.1894 17086.1506

all MLCLSPLPM 24123.4574 22836.2372 1287.2202 3.6924 17252.1166

MLCLSPLPMCC 22808.1007 18420.3191 4387.7815 11.9340 18296.0317

mipgap and computation time

15 They are based on the feasible solutions only.

Chapter E

Lagrangean Relaxation

The basic idea is to solve a complex optimization problem with the help of

more easily tractable subproblems. To do so, complicating constraints of the

original (primal) problem are relaxed and accounted for through a penalty cost

term - the Lagrangean multipliers - in the objective function.

are to reflect the scarcity of the respectively correspondent constraint in the

optimal solution of the original problem. The purpose is to adequately ma-

nipulate the solutions to the subproblems. As the scarcity of the constraints

and hence the optimal values of the Lagrangean multipliers are unknown, they

are adapted in an iterative procedure. At each iteration, the subproblems are

solved with the current values of the Lagrangean multipliers. Then, the latter

are updated according to the magnitude of violation of the corresponding con-

straints in the current solution to the original problem. The dual problem of

Lagrangean relaxation is to maximize the lower bound through a good choice

91

L. Buschkühl, Multi-Level Capacitated Lotsizing with Setup Carryover,

Edition KWV, https://doi.org/10.1007/978-3-658-24034-9_5

92 Multi-Level Capacitated Lotsizing with Setup Carryover

The solution to the relaxed problem is at least as good as the optimal solution

to the original problem as the solution space of the latter is part of the solution

space of the former. Still, even the optimal Lagrangean multipliers cannot

guarantee the feasibility of the solution to the original problem, which is why

the solution of the subproblems is a lower bound to the optimal solution of the

original problem only.2 With increasing quality of the Lagrangean multipliers,

however, the solution is likely to be close to the optimal solution, which is why

it is taken as the basis for constructing a good feasible solution, i. e. the upper

bound.

are3

that the relaxation should make the problem easier, but not too easy.4

In the case of the MLCLSPL and its variants, the constraints to be re-

laxed are the capacity constraints, reflecting the items’ competition on one

resource’s scarce capacity, the inventory balance constraints, reflecting de-

rived demand, and the setup carryover constraints, limiting the number of

setup carryovers to one per resource and period. The resulting subproblem

is the dynamic single-item single-level uncapacitated lotsizing problem with

setup carryover (SLULSPL). Refraining from relaxing the inventory balance

1 See Beasley (1995), page 249.

2 For a deduction of the characterstics of the lower bound and the existing duality gap

for capacitated dynamic lotsizing problems see Derstroff (1995), pages 67–71.

3 Compare Fisher (1981), page 4.

4 See Fisher (1985), page 11.

Chapter E. Solution Procedure based on Lagrangean Relaxation 93

with setup carryover (MLULSPL). However, even in the case without setup

carryover, this problem can still not be solved efficiently.5 Solving the sub-

problem heuristically would lead to lower bounds, which may exceed the

optimal function value. Not only may the optimum not be found, but it is

also impossible to deduct any information on the quality of the solution. 6

level uncapacitated lotsizing problem without setup carryover, based on

the Wagner-Whitin dynamic recursion from Wagner and Whitin (1958),

e. g. by Federgruen and Tzur (1991) and Evans (1985). These have been

adapted to solve the SLULSPL in polynomial time.7

These are Column Generation, Multiplier Adjustment and Subgradient Op-

timization. Subgradient Optimization has proven to be the dominant of the

three methods as far as computational complexity is concerned.8 It is also

the only method applied in the literature presented in chapter B.3.2.

ing ones are based on finite loading procedures, which extend the work of

Trigeiro et al. (1989).

In all iterations but the first, the lower bound is computed. This will be de-

scribed in chapter E.2. Subsequently, the Lagrangean multipliers are updated

according to the magnitude of violation of the corresponding constraints as

described in chapter E.3. Finally, a feasible solution is constructed in each

5 See Derstroff (1995), page 65 and Tempelmeier (2006), pages 233–234.

6 See Özdamar and Barbarosoglu (2000), page 322.

7 See Sox and Gao (1999) and Briskorn (2006).

8 For more details, see Derstroff (1995), pages 79–81.

9 See chapter B.3.2.

94 Multi-Level Capacitated Lotsizing with Setup Carryover

iteration, if possible. To do so, first the SLULSPLs are resolved from the end

products along the bill of materials and down to the components. From the

production quantities of the successors, demand of the predecessors is derived

and hence the inventory balance constraints are enforced. Subsequently, all but

the most beneficial setup carryover for each resource and period are eliminated

from the production plan. Hence, the setup carryover constraints are enforced.

Finally, a finite loading heuristic is applied to enforce the capacity constraints,

while both the inventory balance and the setup carryover constraints remain

in place. If a new feasible schedule is found, its objective function’s value is

compared to the incumbent upper bound and the latter may be updated. This

procedure is described in chapter E.4.

l = 0; U Bmin = +∞; LBmax = −∞; initialize all Lagrangean multipliers at 0

repeat

if l > 0 then

compute the current lower bound LBl

if LBl > LBmax then

LBmax = LBl

end if

update all Lagrangean multipliers

end if

resolve the SLULSPLs along the bill of materials to enforce the inventory balance

constraints

enforce the setup carryover constraints

enforce the capacity constraints

compute the new upper bound U Bl

if U Bl < U Bmin then

U Bmin = U Bl

end if

l=l+1

until (l = lmax or all Lagrangean multipliers have converged)

with

l current iteration

lmax maximum number of iteration

LBl current lower bound

LBmax best lower bound

U Bl current upper bound

U Bmin best upper bound

Chapter E. Solution Procedure based on Lagrangean Relaxation 95

In the case of the MLCLSP already, but especially in the case of the ML-

CLSPL, the duality gap is significant. Thus, there is no benefit in defining a

stopping criterion based on the convergence of the lower and the upper bound.

Therefore, the procedure terminates, when the Lagrangean multipliers have

converged or after a given number of iterations. The maximum number of

iterations is determined by the parameter lmax .10

The adaptation of the Lagrangean heuristic for parallel machines will finally

be presented in section E.5.

Mind, that in this chapter parameters will be introduced, for which the best

values have been computed only in the computational study. In these cases,

no values are given here. Instead, all tested and the best values for each of

these parameters will be given in chapter F.

MLCLSPLe . Then, we will introduce the resulting subproblems along with

the solution procedures applied. Finally, we will deduct the computation of

the lower bound. In chapter E.2.2, we will then describe the adjustments

required to compute the lower bound to the MLCLSPLCC .

In the case of the MLCLSPLe , relaxing the inventory balance constraints, the

capacity constraints and the setup carryover constraints will lead to model

LR-MLCLSPL.

10 In this chapter, we will introduce the parameters of the heuristic. The computational

tests to analyze their impact on the solution quality will be presented in chapter F. The

values tested for lmax will be presented in section F.3.

96 Multi-Level Capacitated Lotsizing with Setup Carryover

Model LR-MLCLSPL

K

Ã T !

X X

min Z = (sk · γkt + ek · (T − t + 1) · qkt ) + (hk − ek ) · qk1

k=1 t=1

K X

X T K X

X T

− hk · ((T − t + 1) · dkt − ybk ) − (hk − ek ) · Dkt

k=1 t=1 k=1 t=1

K X

X T t

X t

X X

+ ukt · (dkτ − qkτ ) + akj · qj,τ +1 − ybk (E.1)

k=1 t=1 τ =1 j∈Nk τ =0

M X

T

Ã !

X X

+ vtm · (tbk · qkt + trk · γkt ) − bm

t

m=1 t=1 k∈Km

M X

T

Ã !

X X

+ wtm · ωkt − 1

m=1 t=1 k∈Km

subject to

t

X

(qkτ − Dkτ ) ≥ 0 ∀k, t = 1, . . . , T (E.2)

τ =1

dk0 = 0 ∀k (E.6)

v0m = 0 ∀m (E.7)

Chapter E. Solution Procedure based on Lagrangean Relaxation 97

Variables:

ukt Lagrangean multiplier for the inventory balance constraint

vtm Lagrangean multiplier for the capacity constraint

wtm Lagrangean multiplier for the constraint of total number of setup

carryovers

Equation (E.1) is the objective function of the relaxation. The first two lines

are the objective function of the original problem (D.22). The remaining three

lines represent the costs for violating the relaxed constraints - first for the

inventory balance constraints (D.23), then for the capacity constraints (D.24)

and finally for the setup carryover constraints (D.26).

ensure that for each item k the cumulative production until the current period

t is not less than the corresponding cumulative total demand. Note that as

in the objective function total demand is calculated according to equation

(D.18) on page 73. Hence, initial inventory is accounted for. It is important

to ensure that at least overall the entire demand is produced. Without this

constraint nothing would be produced, as every production is costly. However,

the derived demand here is computed. It is based on the total demands of

the successor items and not their production quantities. Thus, total demand

is given exogenously. This implies that these inventory balance constraints

reflect no interdependency between the production decisions of any items.

the SLULSPL type. Again, to speed up computation, we separate the constant

parts from the objective function (E.1). To see the relevant parts more easily,

we thus rearrange (E.1) to:

98 Multi-Level Capacitated Lotsizing with Setup Carryover

M

X T ·

X X µX X

T T

X ¶ ¸

ek · (T − t + 1) + ajk · ujτ − ukτ + vtm · tbk · qkt

m=1 k∈Km t=1 j∈Vk τ =t−1 τ =t

τ >0

M

X T µ

X X ¶ XM X

+ (sk + vtm · trk ) · γkt + wtm · ωkt + (hk − ek ) · qk1 − c

m=1 k∈Km t=1 m=1 k∈Km

(E.10)

K X

X T

c= (hk − ukt ) · ((T − t + 1) · dkt − ybk )

k=1 t=1

K X

X T M X

X T

+ (hk − ek ) · Dkt + (vtm · bm m

t + wt ) (E.11)

k=1 t=1 m=1 t=1

XK XT t

X X X t

ukt · qkτ − akj · qj,τ +1

k=1 t=1 τ =1 j∈Nk τ =0

K X

X T T

X X T

X

= qkt ·

ukτ − ajk · ujτ

(E.12)

k=1 t=1 τ =t j∈Vk τ =t−1

τ >0

the beginning of each iteration to compute the lower bound. As will be seen,

at each iteration the values of the Lagrangean multipliers ukt , vtm and wtm are

calculated prior to the solution of the SLULSPLs and are thus given. Also,

the full holding cost coefficient hk , external demand dkt , initial inventory yk0

and capacity bm

t are exogenous to the model. Therefore, using (E.10), the

objective function of the single-item problems simplifies to:

Chapter E. Solution Procedure based on Lagrangean Relaxation 99

Model SLULSPL

T

X

min Zk = (ηkt · γkt + θkt · qkt + wtm · ωkt ) m = 1, . . . , M (E.13)

k∈Km

t=1

subject to

t

X

(qkτ − Dkτ ) ≥ 0 t = 1, . . . , T (E.14)

τ =1

µXX

T T

X ¶

θk1 = ek · T + ajk · ujτ − ukτ + v1m · tbk + hk − ek (E.22)

j∈Vk τ =1 τ =1

µX X

T T

X ¶

θkt = ek · (T − t + 1) + ajk · ujτ − ukτ + vtm · tbk

j∈Vk τ =t−1 τ =t

t = 2, . . . , T (E.23)

100 Multi-Level Capacitated Lotsizing with Setup Carryover

Variables:

Vk set of direct predecessors of item k

eases solution significantly, was discovered by Wagner and Whitin (1958).

They found that in the optimal solution to such a problem a setup occurs

only in those periods, which no inventory is brought into. All lotsizes are ei-

ther zero or the sum of demand from the respective period to the last period

before the next lot of the same item. Finally, they find that, if it is optimal

for a lot in t to cover demand in period t + n, then it is also optimal for it to

cover all demands in τ = t, . . . , t + n − 1.

The advantage of these findings is that the T period problem can be divided

into smaller subproblems. Wagner and Whitin (1958) exploit this advantage

and develop a solution algorithm based on dynamic programming. The under-

lying assumption is that once a solution to a t period problem (t = 1, . . . , T ) is

found this solution remains optimal, even if later periods are considered. Thus,

the solution for the t + 1 problem can be solved independently of all previous

problems 1, . . . , t. This implies that the exact optimal solution to the T period

problem can be computed by solving one subproblem after the other. Other

efficient implementations have been presented by Evans (1985), Wagelmans

et al. (1992) and Federgruen and Tzur (1991). We have chosen to adapt the

algorithms presented by Wagner and Whitin (1958) and by Federgruen and

Tzur (1991) to the SLULSPL.

The difference between the standard SLULSP and the SLULSPL, as stated

also in Sox and Gao (1999), lies in the presence of setup carryovers. Thus, the

decision for the t + 1 problem is independent only of the problems for periods

1, . . . , t − 1. The following cases have to be distinguished here:12

τ < t and there is no setup carryover installed from τ − 1 to τ .

Chapter E. Solution Procedure based on Lagrangean Relaxation 101

in the same period, and there is no setup carryover from period t − 1.

τ < t and there is a setup carryover installed from period τ − 1 to period

τ.

schedule the demand in period t in the same period without additional

changeover.

With fkt as the optimal value for the t period problem the dynamic program-

ming recursion to solve the SLULSPL is:

½ Xt

fkt = min min {ηkτ + θkτ · Dks + fk,τ −1 },

τ ≤t∧ωkτ =0

s=τ

t

X ¾

min {ηk,τ −1 + θk,τ −1 · Dk,τ −1 + θkτ · Dks + wτm + fk,τ −2 } (E.24)

τ ≤t∧ωkτ =1

s=τ

Note that cases 1. and 2. are covered by the first line of the recursion and cases

3. and 4. by the second line of it.

With the solution of the single-item problems, the value of the lower bound is

then given as:

M

X X

LB = Zk − const (E.25)

m=1 k∈Km

The outline of the procedure to compute the lower bound is given in algorithm

3.

setup carryovers is discarded. The only purpose of the procedure is to deduct

the lower bound. The production plan is then generated from scratch during

the computation of the upper bound.

lead to the same optimal solution, the only difference lies in the computation

102 Multi-Level Capacitated Lotsizing with Setup Carryover

that of the Federgruen-Tzur algorithm is O(T log T ). However, the constant

increase in computational effort is higher due to more complex calculations in

the Federgruen-Tzur algorithm. Therefore, it is beneficial only for instances

with a greater number of periods.13

LBl = 0

for k ∈ K do

compute Dkt , ηkt , θkt and wtm ∀t

solve SLULSPLk

LBl = LBl + Zk

end for

LBl = LBl − const

SLULSPLs are altered only in that equation (E.16) is replaced by equation

(E.26):

multiplier as in the single setup carryover case. Further costs for whether or

not another item is produced in between a consecutive carryover cannot be

accounted for here. Thus, we will omit restriction (D.59) for the relaxation

without adding a corresponding cost term to the objective function. By simply

replacing constraint (E.16), we guarantee to compute a lower bound.

13 During initial tests, we found that it is beneficial to apply the Federgruen-Tzur algorithm

for test instances with more than 20 periods.

Chapter E. Solution Procedure based on Lagrangean Relaxation 103

½ ½ t

X ¾

fkt = min min ηkτ + θkτ · Dks + fk,τ −1 ,

τ ≤t∧ωkτ =0

s=τ

½ τ2

X

min ηk,τ1 −1 + wτm + fk,τ1 −2

τ1 ≤τ2 ≤t

τ2 τ =τ1

P

∧ ωkτ =τ2 −τ1 +1

τ =τ1

½ τX

3 −1 t

X ¾¾¾

+ min θkτ · Dkτ + θkτ3 · Dks (E.27)

τ1 ≤τ3 ≤τ2

τ =τ1 −1 s=τ3

The first line of recursion (E.27) equals the first line of recursion (E.24). The

remaining two lines now account for the fact that cases 3. and 4. of the single

setup carryover case have to be adapted, in that more than one consecutive

setup carryover can take place.14 If production costs were constant for each

period, it would be optimal to produce lot for lot, as long as the setup state

is preserved. However, as the modified production costs θkτ are time variant,

depending on the respective Lagrangean multiplier for each period, any combi-

nation of lots may be optimal. For complexity reasons, we reduce the number

of possible lots to two. The first lot is scheduled in period τ1 . The second lot

can be scheduled in any period τ3 (τ1 ≤ τ3 ≤ τ2 ).

The resulting lower bound, may not be a true lower bound. However, in the

case of multi-item production with regular demand of all items, single-item

production over several consecutive periods is rather unlikely. Furthermore,

as mentioned above, there will be a duality gap for most instances, as the

Wagner-Whitin property does not apply in the capacitated case. Hence, the

lower bound provided by the recursion above is very unlikely to exceed the

objective function’s value of the optimal solution.

For the consecutive setup carryover case, only one solution procedure based

on the Wagner-Whitin algorithm is applied.

14 Note that recursions (E.24) and (E.27) are the same, when τ1 = τ2 .

104 Multi-Level Capacitated Lotsizing with Setup Carryover

As pointed out above, several methods are applicable to update the La-

grangean multipliers. These are Column Generation, Multiplier Adjustment

and Subgradient Optimization. Column Generation leads to better results, but

at the cost of a higher computational effort.15 Multiplier Adjustment meth-

ods are developed for a specific application and exploit the special structure

of the problem. Therefore, they may perform better than the general Subgra-

dient Optimization,16 but this is not guaranteed.17 Also, as often only one

multiplier is adjusted at an iteration,18 it is likely to encounter an increased

computational effort. To our knowledge, this has not been tried in the field of

capacitated dynamic lotsizing.19

tipliers. A gradient is a first-order differential operator that maps functions

to vector fields. It is a generalization of the ordinary derivative, and as such

conveys information about the rate of change of a function relative to small

variations in the independent variables. Hence, the gradient also indicates the

direction, in which the decision variables (here the Lagrangean multipliers)

have to be altered, in order to achieve the greatest possible improvement of

the objective function’s value. However, as first-order differential, the gradient

method is applicable only for continuously differentiable functions. The ob-

jective functions of relaxed dynamic capacitated lotsizing problems, however,

are not continuously differentiable due to the binary variables. For this type

of problem, Subgradient Optimization is an applicable solution method. A

subgradient is also defined at points, where the objective function is not dif-

ferentiable. Here, the gradient is calculated as a combination of the derivatives

16 Compare Fisher (1985), page 15.

17 See Guignard and Rosenwein (1989), page 198.

18 See Reith-Ahlemeier (2002), page 121.

19 Still, we chose to implement a Multiplier Adjustment method for the Lagrangean multi-

plier of the setup carryover constraint. Computational tests showed, that in our case it

performed inferior to Subgradient Optimization both in computation time and solution

quality. Hence, it will not be considered further here.

Chapter E. Solution Procedure based on Lagrangean Relaxation 105

starts from an initial solution. Here, we initialize all Lagrangean multipliers

at zero (u0kt = 0, vtm0 = 0 and wtm0 = 0).21

Then, the Lagrangean multipliers for the respectively next iteration l are up-

l

dated with the help of the corresponding subgradient (ζkt , ξtml and ςtml ) and

the corresponding stepsize (λlu , λlv and λlw ) as follows:

½ ¾

ulkt = max 0, ul−1

kt + λ l

u · ζ l

kt k = 1, . . . , K; t = 1, . . . , T (E.28)

½ ¾

m,l−1

vtml = max 0, vt l ml

+ λv · ξ t m = 1, . . . , M ; t = 1, . . . , T (E.29)

½ ¾

wtml = max 0, wtm,l−1 + λlw · ςtml m = 1, . . . , M ; t = 1, . . . , T (E.30)

Variables:

l

ζkt subgradient of the inventory balance constraint for item k in

period t at iteration l

ξtml subgradient of the capacity constraint for resource m in period

t at iteration l

ςtml subgradient of the setup carryover constraint for resource m in

period t at iteration l

l

λu stepsize for the update of the Lagrangean multiplier of the in-

ventory balance constraint for item k in period t at iteration l

λlv stepsize for the update of the Lagrangean multiplier of the ca-

pacity constraint for resource m in period t at iteration l

λlw stepsize for the update of the Lagrangean multiplier of the setup

carryover constraint for resource m in period t at iteration l

21 We also tested the initialization of the multiplier of the setup carryover constraints at

the minimum setup costs on the resource considered plus the small number ² (wtm0 =

min {sk } + ²). However, the results were clearly inferior.

k∈Km

106 Multi-Level Capacitated Lotsizing with Setup Carryover

subgradients can be based on the violation or slack of the constraints them-

selves.22 The subgradients of the three constraints are thus computed as fol-

lows:

t

X t

X X

¡ ¢

l

ζkt = αl · l

dkτ − qkτ + l

akj · qj,τ bk + (1 − αl ) · ζkt

+1 − y

l−1

τ =1 j∈Nk τ =0

k = 1, . . . , K; t = 1, . . . , T ; l = 1, . . . , lmax (E.31)

Ã !

X

ξtml l

=α · (tbk · l

qkt + trk · l

γkt ) − bm

t + (1 − αl ) · ξtm,l−1

k∈Km

m = 1, . . . , M ; t = 1, . . . , T ; l = 1, . . . , lmax (E.32)

Ã !

X

ςtml = αl · l

ωkt −1 + (1 − αl ) · ςtm,l−1

k∈Km

m = 1, . . . , M ; t = 1, . . . , T ; l = 1, . . . , lmax (E.33)

Variables:

αl smoothing coefficient to stabilize convergence of the Lagrangean

multipliers

When αl is set to one, the Lagrangean multipliers are updated only according

to the magnitude of violation of the corresponding constraint in the directly

precedent iteration. However, to stabilize convergence, Crowder (1976) pro-

poses to apply exponential smoothing,23 where the smoothing parameter αl is

computed as:

αl−1

αl = l = 1, . . . , lmax (E.34)

αl−1 + αred

22 See Fisher (1985), page 13.

23 See also Baker and Sheasby (1999).

Chapter E. Solution Procedure based on Lagrangean Relaxation 107

with the help of parameter αred .24

The subgradients will take negative values, and thus reduce the Lagrangean

multipliers, when the corresponding constraint is kept with additional slack.

However, the Lagrangean multipliers can only take non-negative values. If the

update would lead to a negative Lagrangean multiplier, the latter is forced to

25

zero.

While the subgradients point the direction of the update, convergence is depen-

dent on the choice of stepsize.26 The reason is that changing the Lagrangean

multipliers in the direction of the subgradient may lead to deterioration of the

result. Convergence can be achieved, only if the steps are sufficiently small. 27

We will follow Derstroff (1995) and Moorkanat (2000) in that we will apply

the same stepsize λlu = λlv for ulkt and vtml and we will follow Sox and Gao

(1999) in applying a different stepsize λlw for wtml .

Thus, the stepsizes to update the multipliers of the inventory balance and the

capacity constraints are calculated as:28

Z ∗ − LB l

λlu = λlv = δ l · T ³ ´2 T ³ ml ´2

(E.35)

P

K P l P

M P

ζ kt + ξt

k=1 t=1 m=1 t=1

The stepsize for the update of the Lagrangean multipliers of the setup carryover

constraints is computed as:

Z ∗ − LB l

λlw = δ l · T ¡

(E.36)

P

M P ¢2

ς ml

t

m=1 t=1

24 The tested values for α0 and αred will be given in section F.1.1.1.

25 To facilitate computation in cases, where constraints cannot possibly be kept, we decided

to restrict the maximum value of the Lagrangean multipliers to 10, 000.

26 Compare Bazaraa and Sherali (1981).

27 See Derstroff (1995), pages 84 – 85.

28 We tried the iterative procedure suggested by Poljak (1969) to calculate the sequence

of stepsizes. However, we found that in the case of the MLCLSPL and its variants, the

gaps between the upper and the lower bounds (Z ∗ − LB l ) are rather great. Hence, we

adapted the equation by increasing the denominator.

108 Multi-Level Capacitated Lotsizing with Setup Carryover

The factor δ l cannot be smaller than zero nor greater than two. It is initialized

at δ 0 and updated after lδ non-improving iterations according to:29

grangean multipliers of the previous iterations. Z ∗ is an estimate of the optimal

objective function’s value. As Z ∗ is unknown, it is approximated with the help

of the known values of the lower and upper bounds.

Several procedures to update the stepsize have been proposed in the literature.

We implemented three types of procedures, which follow the same principles,

but differ in the values of the parameters. The applied procedure is selected

by the parameter MODUS, which can take the values 10, 11 and 12.

applied.30 In this case, the factor δ l is initialized at two and halved every n

iterations. n in turn is initialized at the maximum number of iterations lmax

and also halved along with δ l , until n = 1.

the current upper bound U B l−1 .

Especially in the first iterations of the procedure, when the Lagrangean mul-

tipliers are far from their optimal values, it may not be possible, to construct

a feasible solution. Thus, as long as no upper bound has been found, it is esti-

mated as the best known lower bound max{LB ∗ , LB l } and a markup, which

is in turn determined by the parameter markupU B .

LB λ to be the current lower bound or the best known lower bound.

parameter δ 0 . It is multiplied with the parameter δ red after lδ non-improving

29 The tested values for δ l and δ 0 will be given in section F.1.1.2 along with all other

parameters relevant for the updating of the stepsizes.

30 See Derstroff (1995), page 86.

31 Compare Derstroff (1995), page 87.

Chapter E. Solution Procedure based on Lagrangean Relaxation 109

iterations. Thus, if δ 0 = 2 and δ red = 0.5, the only difference between MODUS

10 and MODUS 12 is that the number of non-improving iterations is given by

the parameter lδ and is not adapted in the course of the procedure.

The difference between MODUS 10 and MODUS 11 lies in the estimate of the

optimal solution Z ∗ . For MODUS 11, the optimal solution is estimated as

the minimum of the current upper bound U B l−1 and the current lower bound

LB l−1 with an added markup. The factor for the markup is in turn selected

by the parameter markupLB .

obviously an upper bound to its optimal solution.32 As described above, even

with the optimal Lagrangean multipliers, the solution to the relaxed prob-

lem LR-MLCLSPL may not be a feasible solution to the original problem.

Therefore, a feasible solution is generated from scratch using the information

provided by the relaxation, i. e. the Lagrangean multipliers.

steps. First, a production plan without backorders is generated as described

in section E.4.1. From the resulting production plan a feasible solution to

the MLULSPL is deducted by enforcing compliance with the setup carryover

constraints (see section E.4.2). Finally, a feasible solution to the MLCLSPL

is constructed by installing the capacity constraints through a finite loading

procedure (see section E.4.3).

straints can be secured quite trivially. We reinterpret total demand Dkt as:

110 Multi-Level Capacitated Lotsizing with Setup Carryover

Dkt =

X t−1

X X

max 0, dkt + akj · qj,t+1 − max{0, ybk − (dkτ + akj · qj,τ +1 )}

j∈Nk τ =0 j∈Nk

For later periods, solving the corresponding K problems of the SLULSPL type

from the end products along the bill of materials down to the component parts

automatically leads to a solution without backlogging for the multi-level prob-

lem. In the first period, however, the production quantity of a successor item

is restricted by the initial inventory of its predecessors. In general product

structures, the successor items compete on the available predecessor items.

Hence, to guarantee feasibility, lot-for-lot production, i. e. qkt = Dkt , is in-

stalled during the cumulated lead time. The cumulated lead time of an item k

with low-level code i is computed as Lk = I − i, presuming that the low-level

code of end items is 0 and I is the highest low-level code. The procedure is

outlined in algorithm 4.

for u ∈ U do

for k ∈ Ku do

for t ∈ T do

compute total demand Dkt

end for

for t ∈ Lk do

qkt = Dkt

end for

solve SLULSPLk

end for

end for

with

u level of bill of materials structure

U highest level of bill of materials structure

Ku set of products belonging to low-level code u

only if the solution to the MLULSPL is new.33

33 We tested the impact on the MLCLSPL with single setup carryover on the test set

described in chapter C. On average, the results remained the same. Solution speed

Chapter E. Solution Procedure based on Lagrangean Relaxation 111

resource has not been coordinated. Therefore, from the solution without back-

orders a feasible solution to the MLULSPL and the MLULSPLCC respectively

has to be generated by imposing the setup carryover constraints. Lotsizes are

not changed in the process. Additional setups may lead to capacity violations,

but this is repaired later.

P

i. e. ωkt < 1.

k∈Km

P

i. e. ωkt = 1.

k∈Km

P

i. e. ωkt > 1.

k∈Km

Generally, for each of the three cases stated above, the following steps have to

be applied:34

both in period t − 1 and in period t and a setup is performed in both

periods, i.e. ωk,t−1 = 0, the second setup can be replaced by a setup

carryover. This saves both setup time and costs.

Let Kγ = {k : k ∈ Km ∧ γk,t−1 + γkt = 2} denote the set of items, for

which it is possible to install a setup carryover to period t. If at least

one such item exists, i. e. Kγ 6= ∅, let k ∗ = argmaxk∈Kγ sk be the one

with the highest setup costs. Then, a setup carryover is installed for this

item, i. e. ωk∗ t = 1 and γk∗ t = 0.

increased by 7.51%.

34 See also Sox and Gao (1999), page 180.

112 Multi-Level Capacitated Lotsizing with Setup Carryover

for the same item have been prevented in the recursion. Hence, no further

adaptation is needed.

3. More than one setup carryover has been installed on m in t. In this case

all but one setup carryovers must be removed.

With the notion of Kω = {k : k ∈ Km ∧ ωkt = 1} as the set of items,

for which a setup carryover is installed in period t, find again the item

k ∗ = argmaxk∈Kω sk with the highest setup costs. For this item the

setup carryover is kept, while for all other items the setup carryover is

removed, i. e. ωkt = 0 and γkt = 1, ∀k ∈ Kω \ k ∗ .

This implies that, in cases 2 and 3, the setup carryover installed may not be

the overall best possible one. The reason is that only items are considered, for

which a setup carryover is already scheduled. If another item k 0 incurs higher

setup costs than the ones with a setup carryover, i. e. sk0 > sk∗ , additional

costs could be saved. However, the resulting schedule would differ more from

the one computed for the relaxed problem.

In the case of single setup carryovers, compliance with constraints (D.26) needs

to be achieved. The outline of the procedure is given in algorithm 5.

s

K

P m

On each resource m, the procedure runs from period 1 to period T . ωit > 1

i=1

s

K

P m

corresponds to case 3. ωit = 0 corresponds to case 1.35

i=1

s

As the set Km is sorted in decreasing order of the items’ modified setup costs,

the first setup carryover, which can be chosen, is the best. If one setup car-

P

k−1

ryover is installed, ωit becomes one. Thus, further setup carryovers are

i=1

prohibited.

Chapter E. Solution Procedure based on Lagrangean Relaxation 113

In each period, information on the number of setups and on the setup carry-

overs into the period is stored and preserved for the next period.

for m ∈ M do

for t ∈ T do

s

Km

P

if ωit > 1 then

i=1

s

for k ∈ Km do

if qkt > 0 then

P

k−1

if ωit = 1 then

i=1

ωkt = 0

γkt = 1

end if

end if

end for

s

Km

P

else if ωit = 0 then

i=1

s

for k ∈ Km do

if qkt > 0 then

P

k−1

if ωit = 0 and γk,t−1 = 1 then

i=1

ωkt = 1

γkt = 0

end if

end if

end for

end if

end for

end for

with

s

Km set of the products produced on resource m in decreasing order of their

modified setup costs

to (D.60) needs to be achieved. The outline of the procedure is given in

algorithm 6.

114 Multi-Level Capacitated Lotsizing with Setup Carryover

Here, the solution is generated in a single pass. To do so, only the production

quantities are taken over from the solutions to the SLULSPLk .

On each resource m, the best possible setup carryover is installed. For all

other items with a positive production quantity, a setup activity is scheduled

in t. Thus, here the overall best setup carryover is selected. The request that

consecutive setup carryovers can only be installed, if no other item is produced

in between, cannot be considered in the solutions to the SLULSPLk . Hence,

we chose to reconsider the complete setup pattern here.

carryovers are permitted, if no other item is produced in period t − 1.

for m ∈ M do

for t ∈ T do

s

for k ∈ Km do

if qkt > 0 then

P

k−1 P

if ωit = 0 and (γk,t−1 = 1 or (ωk,t−1 = 1 and γi,t−1 = 0)) then

i=1 i∈Km

ωkt = 1

else

γkt = 1

end if

end if

end for

end for

end for

with

s

Km set of the products produced on resource m in decreasing order of their

modified setup costs

the capacity constraints. To do so, we will use a finite loading heuristic.

Violations of the capacity constraints are interpreted as prohibited overtime,

which is charged by the Lagrangean multipliers. Then, in several alternating

Chapter E. Solution Procedure based on Lagrangean Relaxation 115

forward and backward passes, production quantities are shifted across periods.

While the primary objective is to find a feasible production plan, the solution

quality with respect to the costs incurred may also be improved on. A feasible

production plan is found, when all overtime has been eliminated. Through-

out the procedure, the inventory balance constraints and the setup carryover

constraints are respected.

This type of procedure was first introduced by Trigeiro (1987) and Trigeiro

et al. (1989) for the CLSP without and with setup times, respectively. It

has later been adapted and extended by several authors also to the MLCLSP

and the CLSPL.36 Work of this type includes Campbell and Mabert (1991),

Tempelmeier and Derstroff (1996), Sox and Gao (1999), Moorkanat (2000),

Hindi et al. (2003).

overall overtime for all resources and all periods SO is computed as:

T

X X

SO = max{Ntm − bm

t , 0} (E.38)

m∈M t=1

P

Ntm is the total capacity requirement (tbk · qkt + trk · γkt ) on resource

k∈Km

m in period t. If SO takes a positive value, the finite loading procedure is

started, applying four different shifts. Production quantities are shifted on

the current resource m from the current period t, until either all overtime has

been eliminated (SO = 0) or no further production is shiftable.

First, a forward pass is started. In each period, starting in period 1 and moving

forward, first single-item forward shifts are executed. If overtime prevails in

the respective period and no further production can be shifted independently

of other items, multi-item forward shifts are applied next.

116 Multi-Level Capacitated Lotsizing with Setup Carryover

ff

P P

T P

t

SO = max{Ntm − bm m

t , 0}; COt = max (Nτm − bm

τ ), 0

m∈M t=1 τ =1

while m ∈ M O do

t=1

while t ≤ T do

if l = 1 and COtm > 0 then

single-item forward shifts

if COtm > 0 then

multi-item forward shifts

end if

else if l > 1 and Ntm > bm

t then

single-item forward shifts

if Ntm > bm

t then

multi-item forward shifts

end if

end if

wend

wend

if SO > 0 then

O

while m ∈ M do

t=T

while t > 0 do

if l > 1 and Ntm > bm

t then

single-item backward shifts

if Ntm > bm

t then

multi-item backward shifts

end if

end if

wend

wend

end if

wend

with

COtm cumulated excess capacity requirement on resource m until period t

l current iteration

lmax maximum number of iteration

MO set of ordered resources

O

M resources in reverse order

Ntm capacity requirement (setup and production times) on resource m in period t

SO overall overtime

Chapter E. Solution Procedure based on Lagrangean Relaxation 117

During the first forward pass only cumulated overtime COtm is considered in

order to leave more flexibility to the subsequent backward pass. The cumulated

overtime is computed as:

( t )

X

m m m

COt = max (Nτ − bτ ), 0 (E.39)

τ =1

When the end of the planning horizon is reached and SO is still positive, a

backward pass is undertaken. The same principle applies, but the orders of

resources and periods are reversed and production is shifted backward. Again,

single-item shifts are applied first in each period and multi-item shifts are

thereafter applied, if necessary.

The procedure is terminated, when either all overtime has been eliminated or

the maximum number of iterations lmax has been reached.37

The outline in algorithm 7 shows one possible ordering of resources and peri-

ods.38 The order, in which resources and periods with overtime are considered,

influences the solution. The reason is that the multi-level structure may lead

to interdependencies between the shifts executed on a given resource in a given

period and the shiftable quantities in others.

same period before proceeding to the next period or in all periods on the same

resource and then on the next resource. If all periods are considered for the

same resource, parameter SEQ determines, whether both the forward and the

backward pass are carried out together, or else first all forward passes and

then all backward passes are carried out.

38 The tested values for the parameters for the ordering of shifts will be given in section

F.1.3.

118 Multi-Level Capacitated Lotsizing with Setup Carryover

Regardless the sequence of resources and periods selected, the order in which

resources are considered is determined a priori. It is computed individually for

forward and backward passes in ascending or descending order of the selected

criterion by the parameters FOR and BACK. The criteria are based either on

the resources’ capacity requirements (KAP) or on the number of successors

and predecessors of the corresponding items (M).

The logic behind the first criterion is that resources with high capacity re-

quirements are dealt with first. This approach was applied also by Moorkanat

(2000). Possible criteria for sorting are:

The second criterion was proposed by Derstroff (1995) and is based on the

idea that the number of multi-item shifts required to find a feasible solution

can be affected by the order, in which resources are considered. If, during

forward (backward) passes, resources are considered first, on which primarily

items with many successors (predecessors) are produced, this will increase the

number of multi-item shifts due to the derived demands.

To sort resources, two criteria can serve to compute the priority value for the

ordering. The first is the difference of all successor relations and all predecessor

relations of all items assigned to the resource. These relations are computed

as the number of indirect and direct successors and predecessors produced on

a different resource than the one considered.

corresponding items. Here, the maximum low-level code of all items produced

on the resource is used.

E.4.3.3 Shifts

Chapter E. Solution Procedure based on Lagrangean Relaxation 119

period

cessors to a future period

period

of its predecessors to an earlier period

The general structure is the same for all types of shifts. Each shift is started

on a specific resource m in a specific period t. The shifts are executed until

either all overtime has been eliminated on m in t or no further shift can be

executed. As described above, during the first iteration of forward shifts only

cumulated overtime is remedied. In all other iterations and during backward

passes the absolute capacity violation of period t is considered.

Apart from the multi-item backward shifts, which consider only the directly

precedent period, the target period is found by evaluating the possible shifts

for each item. The difference for single- and multi-item shifts is that multi-

item shifts have to consider the same changes to production, inventory and

setups for all items shifted. This will be explained below.

At each step, all items are evaluated to find the best feasible shift. If such

a shift exists, it is executed. Thereafter, capacity consumption is updated

and the procedure is repeated, if necessary. If no favorable shift exists, the

procedure is terminated.

Feasible shifts are always determined by the request that no shortages and no

violations of the setup restrictions are incurred. Additionally, the aim of the

procedure is to avoid overtime. However, preventing the build-up of overtime

in the target period may limit the solution space of the procedure too much.

This is especially the case, when capacity utilization is high.

120 Multi-Level Capacitated Lotsizing with Setup Carryover

Thus, the heuristic is divided into two parts by the parameter lO .39 Additional

overtime is allowed in the target period at the beginning of the procedure. In

iterations later than lO , however, shifts are allowed, only if they do not incur

additional overtime in the target period. This has different implications for

forward and for backward passes, as will be explained below.

To find the best possible shift, four computations have to be made, which are

interdependent:

In the following, we will first present the changes to the setup pattern incurred

with a shift. In the case of the MLCLSP, production always incurs a setup in

the same period. Hence, the setup state is changed in the original period, only

if the complete lot is shifted and hence the setup activity becomes obsolete.

In the target period, the setup state is changed, only if there has been no

production of the item in that period and hence an additional setup activity is

undertaken. However, as setup carryovers are considered here, the changes to

the setup pattern are generally more complicated. Therefore, we will describe

the changes to the setup pattern in more detail. The way the setup pattern is

affected is the same for all types of shifts.

The shiftable quantities ∆qk are generally restricted by the capacity and the

inventory balance constraints. Their computation is different for each type of

shift and will thus be described in detail for each shift.

39 The tested values for lO and all other parameters for the shifts (apart from the compu-

tation of the priority values) will be given in section F.1.5.

Chapter E. Solution Procedure based on Lagrangean Relaxation 121

Period

are independent of the actual quantity of the shift. Hence, the changes to

the setup pattern and their impact on the capacity consumption in the target

period can be determined a priori. In the single setup carryover case, the

following four cases have to be distinguished:

1. ωkτ = 1

2. γkτ = 1

The first two cases are trivial. The production can be added without changes

to the setup pattern.

be preserved from the precedent period.

In the fourth case, two options exist. The first option is to install a setup in τ

(γkτ = 1). The second option is to preserve the setup state from the previous

period, i. e. ωkτ = 1. From γk,τ −1 = 1 follows that ωk,τ −1 = 0 and from

γkτ = 0 that ωkτ +1 = 0. Hence, a setup carryover can be introduced in τ ,

unless there is a setup carryover for another item.

The decision is based on the costs incurred by either option. If no setup carry-

over is installed in period τ ,40 the setup will incur costs, which the carryover

will not. Hence, a setup carryover is installed. If, however, a setup carryover

had been installed for a different item k 0 , the setup costs of the two items are

compared. For the item incurring greater setup costs, the setup carryover is

installed. For the other item, a setup activity is undertaken also in period τ .

The selection follows the same principle presented in section E.4.2.

40 This is the case, only if no other item is scheduled in τ and in τ − 1.

122 Multi-Level Capacitated Lotsizing with Setup Carryover

carryover for another item k 0 from period τ − 1 to period τ + 1, when no other

item is scheduled in period τ . Then, the setup carryover for item k 0 into period

τ is replaced by a setup activity, i. e. γk0 τ = 1 and ωk0 τ = 0.

carryover can be installed, if no other item is produced in τ − 1. Otherwise,

again a setup activity is introduced in τ .

Period

When production is shifted from a period, the changes to the setup pattern

depend primarily on the shifted quantity. If less than the complete lot is

shifted, the setup pattern remains unchanged. Hence, the changes to the

setup pattern are determined after the computation of the shiftable quantity.

For single setup carryovers, the following three cases are distinguished:

3. ωkt = 1

Case 2 implies that the setup activity cannot be canceled. However, it can be

shifted to the next period t + 1, if there is another item k 0 , for which a setup

activity is undertaken both in period t and in period t + 1. Then, the setup

activity for this item can be replaced by a setup carryover in period t + 1. If

Chapter E. Solution Procedure based on Lagrangean Relaxation 123

several items exist with setups in both periods, the one with the greatest setup

costs is chosen. The selection with respect to the cost criterion again follows

the same principles presented in section E.4.2. Still, when the setup for item

k is pushed forward, compliance with the capacity constraints also in period

t + 1 must be secured. If the setup time of item k is greater than that of item

k 0 , the difference trk − trk0 must be less than the incumbent spare capacity in

t + 1. Otherwise, the setup pattern is left unchanged.

Again, a setup carryover is installed for a different item k 0 , if possible. If

several items have a setup scheduled in periods t − 1 and t, the one with the

highest setup costs is chosen along the same lines explained above.

For consecutive setup carryovers again case three has to be further distin-

guished into

Case 3.a again corresponds to case 3 for single setup carryovers. In case 3.b,

several setup carryovers may be superfluous, if the setup is preserved from

some period τ 1 < t to some other period τ 2 > t without production in periods

τ 1 to τ 2 − 1. Then, if shifting the setup activity from period τ 1 to period τ 2

is possible, later shifts can install setup carryovers for other items.

is invoked, if the capacity (cumulated or absolute) is exceeded for the given

resource in the given period. As long as overtime persists, the best possible

shift is computed.

considered. To find the best possible for an item k from the original period

124 Multi-Level Capacitated Lotsizing with Setup Carryover

max

t to a given target period τ , first the maximum shiftable quantity ∆qktτ is

computed.

while (l > 1 and Ntm > bm m

t ) or (l = 1 and COt > 0) do

P R∗ = −∞

for k ∈ Km do

τ =t

while τ ≤ T do

max

compute ∆qktτ

max

if ∆qktτ = 0 then

next k

end if

compute ZR, ∆qktτ , QR, P R

if P R > P R∗ then

P R∗ = P R, ∆qktτ

∗

= ∆qktτ , τ ∗ = τ , QR∗ = QR, ZR∗ = ZR

end if

τ =τ +1

wend

end for

if P R∗ > −∞ then

execute shift (take over t, τ ∗ , ZR∗ , ∆qktτ

∗

, QR∗ )

else

break

end if

wend

with

∆qktτ shiftable quantity for item k from period t to period τ

∗

∆qktτ shiftable quantity for item k from period t to period τ , which belongs to the

shift with the incumbent highest priority value

max

∆qktτ maximum shiftable quantity for item k from period t

PR priority value

P R∗ incumbent highest priority value

QR changes to setup pattern in t

QR∗ changes to setup pattern in t, which belong to the shift with the incumbent

highest priority value

τ target period of shift

t original period of shift

ZR changes to setup pattern in τ

ZR∗ changes to setup pattern in τ , which belong to the shift with the incumbent

highest priority value

Obviously, not more than the lotsize can be shifted. Furthermore, as produc-

Chapter E. Solution Procedure based on Lagrangean Relaxation 125

period. As no shortages are to be incurred, the maximum shiftable quantity is

thus further limited by the minimum inventory on hand in the periods between

t and τ .41

½ ¾

max

∆qktτ = min qkt , min yks (E.40)

t≤s≤τ

period between t and τ , the next item is considered.

If the maximum shiftable quantity is equal to the lotsize qkt , it is taken as the

shiftable quantity. Following Derstroff (1995), this measure aims at eliminating

setups.42 If the maximum shiftable quantity is positive but less than the

lotsize, it is further limited by the capacity requirements. Thus, first the

changes to the setup pattern in the target period are computed as described

in section E.4.3.3.1. Thereafter, the remaining available capacity for shifting

production into the target period can be calculated.

From iteration lO on, the shiftable quantity of item k is limited by the request

that no additional overtime may be incurred in τ . In iterations prior to lO ,

additional overtime may be incurred in τ as long as further overtime eliminated

in t. Thus, the shiftable quantity for item k is calculated as:

n n m oo

min ∆q max

, max

Cτ

, OV m

l < lO

ktτ tbk t

∆qktτ = n o (E.41)

m

min ∆q max , Cτ l ≥ lO

ktτ tbk

Cτm is the free capacity in period τ after the changes to the setup pattern

and OVtm is the overtime in period t. This implies that more production

than is needed to eliminate overtime in the original period t may be shifted.

Hereby, inventory holding costs are reduced and more flexibility is left to the

subsequent backward pass.

42 See Derstroff (1995), page 114.

126 Multi-Level Capacitated Lotsizing with Setup Carryover

section E.4.3.3.2 and the priority value of the shift is calculated as will be

described in section E.4.3.3.7.

If any feasible shift is found, the one with the highest priority value is executed

and the procedure is repeated, if necessary. If no such shift is found, the

procedure is terminated.

If the single-item forward shifts fail to eliminate all overtime from a given pe-

riod t, multi-item forward shifts are applied. The outline is given in algorithm

9.

others by the request that shortages may not be incurred. Therefore, it may be

extended, by simultaneously shifting production quantities of successor items

and hence the corresponding derived demands.

putational effort, we will follow Derstroff (1995) in considering serial subsets

ST R of the product structure. To do so, all immediate and mediate succes-

sors i ∈ NkC are considered and all alternative subsets ST Rki of the product

structure linking k to i are evaluated.43

Applied to the complete serial subset ST Rki of the products structure, the

multi-item forward shift follows the same principle as the single-item forward

shift. Again, all future target periods are evaluated. First, the maximum

shiftable quantity is computed for the final item of the serial subset i according

to equation (E.41). The cumulated lead time is accounted for on all stages

of the serial subset. Hence, the production quantities of successor items are

shifted from the corresponding periods tu = t + uk − ui0 to τu = τ + uk − ui0 .

43 For an illustration of the serial subsets see Derstroff (1995), page 117.

Chapter E. Solution Procedure based on Lagrangean Relaxation 127

t ) or (l = 1 and COt > 0) do

P R∗ = −∞

for k ∈ Km do

τ =t

while τ ≤ T do

for i ∈ NkC do

for ST R ∈ ST Rki do

max

compute ∆qit u τu

max

if ∆qit u τu

= 0 then

next ST R

end if

for i0 ∈ ST R do

compute ZR, ∆qi0 tu τu , QR, P RF OR

P R = P R + P RF OR

end for

if P R > P R∗ then

P R∗ = P R, ∆qk

∗

= ∆qk , τ ∗ = τ , QR∗ = QR, ZR∗ = ZR

end if

end for

end for

τ =τ +1

wend

end for

if P R∗ > −∞ then

execute shift (take over t, τ ∗ , ZR∗ , ∆qk

∗

, QR∗ )

else

break

end if

wend

with

max

∆qktτ maximum shiftable quantity for item k from period t

∗

∆qk shiftable quantities, which belong to the multi-item shift with the incumbent

highest priority value

NkC set of all direct and indirect successors of k

PR priority value

P R∗ incumbent highest priority value

P RF OR priority value a single item adds to the priority value of the multi-item shift

QR changes to setup pattern in t

QR∗ changes to setup pattern in t, which belong to the shift with the incumbent

highest priority value

ST Rki complete serial subset of successor items of k, starting with k and ending

with i

∆qktτ shiftable quantity for item k from period t to period τ

τ target period of shift

t original period of shift

ZR changes to setup pattern in τ

ZR∗ changes to setup pattern in τ , which belong to the shift with the incumbent

highest priority value

128 Multi-Level Capacitated Lotsizing with Setup Carryover

the shiftable quantity of the final item i is extrapolated along the serial subset

down to item k.

shiftable quantities.44 If only the postponement of derived demand is used to

shift production of the predecessors of i (VARFOR=1), the maximum shiftable

quantity of an item i0 is computed as:

n o

∆qimax

0t τ

u u

= min qi0 tu , qnmax

i 0 t u τ u

· a i 0n 0

i

(E.42)

FOR=0). The maximum shiftable quantity is then computed as:

½ ¾

∆qimax

0t τ

u u

= min qi 0t ,

u min y ks + q max

ni0 tu τu · a i 0n 0

i

(E.43)

tu ≤s≤τu

When all maximum shiftable quantities have been computed, the actually

shiftable quantities are adapted to capacity requirements similar to the single-

item shifts. Again, the procedure starts with the final item i.

lot could be shifted, but careful shifts are applied, the shiftable quantity is

restricted by the free capacity of the respective resource Cτmu after the changes

to the setup pattern have been installed.

If the complete lot cannot be shifted, the shiftable quantity is further limited

by the quantity necessary to eliminate overtime on the resource of item k in

period t. Aki0 represents the number of items of product k needed to produce

one item of product i0 . Again, in iterations later than lO , the shiftable quantity

is further restricted by the free capacity of the respective resource in period

τu after the changes to the setup pattern have been installed.

Chapter E. Solution Procedure based on Lagrangean Relaxation 129

q i0 t u l < lO ∧ qimax

0t τ = q i0 t u

u u

n o

min q ,

Cτ u

l ≥ lO ∧ qimax = q i0 t u

0

i tu tb 0 0t τ

i u u

ff

max OVt

min ∆qktτ , tb

m

− min yks

k

min ∆qimax 0t τ ,

t≤s≤τ

u u Aki0

∆qi0 tu τu =

l < lO ∧ qimax < q i0 t u

0t τ

u u

ff

min ∆qktτ

m

m

max OVt

, tb − min yks

min ∆qimax

0 t τ , min

k t≤s≤τ C

, tbτu0

u u

Aki0 i

l ≥ lO ∧ qimax < q i0 t u

0t τ

u u

(E.44)

For each item, the maximum shiftable quantity is updated with the informa-

tion on the shiftable quantity of its successor. For the first item k, the same

computations are applied with the new inventory as in single item shifts.

When all shiftable quantities have been determined, the priority value is cal-

culated. The best feasible shift is executed, if such a shift exists.

The outline of single-item backward shifts is given in algorithm 10. The struc-

ture is the same as for single-item forward shifts with the exception that the

periods before t are considered here.45

When shifting production backward, the inventory balance constraints for the

respective item hold in any case. However, if the item is a successor of other

items, shifting production will change their derived demand. Thus, to ensure

that the inventory balance constraints remain kept for all items, the maximum

45 See algorithm 8.

130 Multi-Level Capacitated Lotsizing with Setup Carryover

yi,s−1

max

∆qktτ = min qkt , min (E.45)

i∈Vi ∧ aik

t≤s≤τ

max

If ∆qktτ = 0, the next item is considered. Otherwise, the changes to the setup

pattern in the target period and the actual shiftable quantities are computed

reflecting the capacity requirements.

qkt l < lO ∧ ∆qktτ

max

= qkt ∧ P R ≥ 0

n o

OVtm

min qkt , tbk

l < lO ∧ ∆qktτ

max

= qkt ∧ P R < 0

∆qktτ =

0 l < lO ∧ ∆qktτ

max

< qkt ∧ l < lLS ∧ (LS = 0 ∨ P R < 0)

n o

min ∆q max , Cτ

m

max

< qkt ∧ l ≥ lLS ) ∨ l ≥ lO

(E.46)

Before iteration lO , the complete lot is shifted, if this is possible and the

priority value is non-negative. If the complete lot could be shifted, but the

priority value is negative, only the quantity corresponding to the overtime on

m in t is shifted.47

Partial lots may be shifted, only if this is not prohibited. The logic is that shifts

eliminating setups are preferred in the beginning of the procedure. Thus, in the

first lLS − 1 iterations, partial lots are either shifted, only if the corresponding

priority value is positive (LS=1), or not at all (LS=0).

max

minimum of free capacity in τ and the maximum shiftable quantity ∆qktτ is

shifted.

46 Compare Derstroff (1995), page 126. Note, that here the lead time is one period. Hence,

derived demand arises one period earlier than the production of the successor.

47 Note, that shifting production backward incurs additional holding costs rather than

saving them. Thus, it may not be beneficial to shift more production than necessary.

Chapter E. Solution Procedure based on Lagrangean Relaxation 131

t do

P R∗ = −∞

for k ∈ Km do

τ =t

while τ ≥ 1 do

max

compute ∆qktτ

max

if ∆qktτ = 0 then

next k

end if

compute ZR, ∆qktτ , QR, P R

if P R > P R∗ then

P R∗ = P R, ∆qktτ

∗

= ∆qktτ , τ ∗ = τ , QR∗ = QR, ZR∗ = ZR

end if

τ =τ −1

wend

end for

if P R∗ > −∞ then

execute shift (take over t, τ ∗ , ZR∗ , ∆qktτ

∗

, QR∗ )

else

break

end if

wend

with

∆qktτ shiftable quantity for item k from period t to period τ

∗

∆qktτ shiftable quantity for item k from period t to period τ , which belongs to the

shift with the incumbent highest priority value

max

∆qktτ maximum shiftable quantity for item k from period t

PR priority value

P R∗ incumbent highest priority value

QR changes to setup pattern in t

QR∗ changes to setup pattern in t, which belong to the shift with the incumbent

highest priority value

τ target period of shift

t original period of shift

ZR changes to setup pattern in τ

ZR∗ changes to setup pattern in τ , which belong to the shift with the incumbent

highest priority value

Finally, adaptations are made to reflect the changes in the cumulated overtime

incurred with the shift according to parameter SAV.48 If costs are weighted,

(SAV=1 and SAV=3) positive priority values are divided by the additional

48 See Derstroff (1995), page 126.

132 Multi-Level Capacitated Lotsizing with Setup Carryover

cumulated overtime in the target period, negative ones are multiplied likewise.

If additional cumulated overtime is incurred, the shift with the least additional

overtime is selected, if SAV=2 or SAV=3.49

If a feasible shift exists, the best one is executed and the procedure is repeated.

If the procedure is terminated before all overtime has been eliminated, multi-

item backward shifts are applied next.

to the other three types, here, shifts are restricted to the directly precedent pe-

riod. Similar to multi-item forward shifts, multi-item backward shifts extend

the shiftable quantity of an item by simultaneously shifting its predecessors.

However, instead of considering serial subsets of the product structure, all im-

mediate and mediate predecessor items i ∈ VkC are shifted, if this is necessary

to avoid shortages.

First, the shiftable quantity of the considered item k is computed as the min-

imum of the lotsize of k and the quantity, which would eliminate overtime on

m in t.

½ ¾

OVtm

∆qkt,t−1 = min qkt , (E.47)

tbk

Then, the shift is extrapolated down the low-level codes. Note, that the subset

contains all predecessors of k, so that each item may have several predeces-

sors. However, it has an assembly structure, so that each item has only one

successor.50 For each item, the shifted derived demand, which cannot be met

by the inventory on hand, has to be shifted:51

49 The same modifications are applied to multi-item backward shifts. The reason that it is

not used in forward shifts is that shifting production forward cannot lead to additional

cumulated overtime.

50 See Derstroff (1995), page 130.

51 Note here, that this quantity cannot exceed the items lotsize. The part of derived

demand not satisfied by inventory on hand had to be met by production in tu .

Chapter E. Solution Procedure based on Lagrangean Relaxation 133

t do

P R∗ = −∞

for k ∈ Km do

if qkt > 0 and Vk 6= ∅ then

compute ∆qkt,t−1

for u > uk do

for i ∈ Ku do

if i ∈ VkC then

compute ∆qitu ,tu −1

if tu = 1 and ∆qitu ,tu −1 > y

bi then

next k

end if

end if

end for

end for

for u ∈ U do

for i ∈ Ku do

τ =1

while τ ≤ t do

if ∆qiτ,τ −1 > 0 then

compute ZR, QR, P R

end if

wend

end for

end for

if P R > P R∗ then

P R∗ = P R, ∆qkt,t−1

∗

= ∆qkt,t−1 , QR∗ = QR, ZR∗ = ZR

end if

end if

end for

if P R∗ > −∞ then

execute shift (take over t, ZR∗ , ∆qk

∗

, QR∗ )

else

break

end if

wend

with

Ku set of products belonging to low-level code u

P RF OR priority value a single item adds to the priority value of the multi-item shift

ST Rki complete subset of successor items of k, ending with i

u level of bill of materials structure

uk level of bill of materials structure of item k

U highest level of bill of materials structure

VkC set of all direct and indirect predecessors of k

134 Multi-Level Capacitated Lotsizing with Setup Carryover

shifted exceeds the corresponding initial inventory, the shift is infeasible.

Hence, the next item k is considered.

changes to the setup patterns in the respective original and target periods,

as well as the priority value, are calculated. The shifts are not restricted by

the capacity. Additional cumulated overtime is again considered depending on

the parameter SAV.

executed in iterations prior to lLS .

The priority values of the shifts are computed as the incremental costs associ-

ated with the step. The relevant costs generally consist of setup and holding

costs. This means that the priority value is computed as the costs saved

through the elimination of setups and the reduction of inventory less the costs

incurred by additional setup activities and held inventory. For multi-item

shifts, the priority value is computed as the sum of the incremental costs over

all items shifted.

Depending on the parameter PRC,52 either the original cost factors or the

modified cost factors, including the Lagrangean multipliers, are applied. The

advantage of utilizing modified cost factors is that the schedule may be closer

to the optimal solution to the relaxed problem.

52 The

P tested values for

P the parameters for the priority values will be given in section F.1.6.

53 ∆γiτ and ∆γit represent the changes of the setup pattern in the target

i∈Km i∈Km

period and the original period, respectively. See sections E.4.3.3.1 and E.4.3.3.2.

Chapter E. Solution Procedure based on Lagrangean Relaxation 135

P

∆qktτ · (θkt − θkτ ) − (∆γit · ηit + ∆γiτ · ηiτ ) PRC = 0

i∈Km

P

∆qktτ · ((τ − t) · ek + (hk − ek )) − (∆γit − ∆γiτ ) · si

i∈Km

P Rfor =

PRC = 1 ∧ t = 1

P

∆qktτ · (τ − t) · ek − (∆γit − ∆γiτ ) · si PRC = 1 ∧ t > 1

i∈Km

(E.49)

If the modified cost factors are applied, during backward shifts the produc-

tion and setup time eliminated from the original period is computed. If it

exceeds overtime in that period, i. e. additional free capacity is generated in

the original period, the excess is punished at the Lagrangean multiplier of the

corresponding capacity constraint. For backward shifts, the priority value is

thus computed as:

P

∆qktτ · (θkt − θkτ ) − (∆γit · ηit + ∆γiτ · ηiτ )

i∈Km

− max {0, (∆qktτ · tbk − ∆γit · trk − OVtm ) · vtm } PRC = 0

P

P Rback = ∆qktτ · ((t − τ ) · ek − (hk − ek )) − (∆γit − ∆γiτ ) · si

i∈Km

PRC = 1 ∧ τ = 1

P

∆qktτ · (t − τ ) · ek − (∆γit − ∆γiτ ) · si PRC = 1 ∧ τ > 1

i∈Km

(E.50)

Also, the parameter PRW determines, whether the priority values are weighted

to reflect the total overtime eliminated in the original period. Although the

aim of the heuristic is to find the best possible schedule, the primary aim of

the finite loading heuristic is to find a feasible one. If PRW is set to one, the

priority value is weighted with the minimum of overtime in the original period

t and the shifted production time:54

54 See Derstroff (1995), page 135.

136 Multi-Level Capacitated Lotsizing with Setup Carryover

PR

min{∆qktτ ·tbk ,OVtm } PR ≥ 0

PR = (E.51)

P R · min {∆q · tb , OV m }

ktτ k t PR < 0

To ensure that the best possible schedule is found in case the relaxed schedule

is close to feasibility already, only cost-improving shifts are allowed in the first

iterations of the heuristic. However, if such shifts do not lead to a feasible

solution until iteration lP R , the overall least cost feasible shift is selected.

Following Derstroff (1995), we apply tabu lists to prevent the finite loading

heuristic from cycling and thus to stabilize convergence.55

From iteration lTABU of the finite loading heuristic on, a tabu list is kept

for each resource. At each iteration, executed shifts are stored in the tabu

list.56 Each new shift is entered at the beginning of the list and the others

pushed forward. The shifts remain tabu active, i. e. stored in the tabu list,

until either their position exceeds the length of the tabu list nTABU or the

maximum duration of the tabu status tTABU is exceeded.57

While a shift is tabu active, its reversal, i. e. shifting the same quantity ∆qktτ

for item k back from period τ to period t, is prohibited unless the aspiration

criterion is met. The aspiration criterion is applied in iterations later than lASP

and enables the execution of tabu active moves with the probability ASP.

E.4.4 Postoptimization

improve on a given feasible schedule. It is applied, only if the incumbent

55 See Derstroff (1995), pages 137-140. For Tabu Search see B.3.4.3.

56 The information stored comprises the product k, the target period τ , the original period

t and the shifted quantity ∆qktτ .

57 The tested values for the parameters for tabu lists will be given in section F.1.7.

Chapter E. Solution Procedure based on Lagrangean Relaxation 137

schedule is the new best schedule. The postoptimization consists of the three

steps

The first two steps have been presented by Derstroff (1995).58 They are alter-

nated, until no further improvement can be achieved. Then, the third step is

applied.

to shift production quantities to future periods in order to save holding

costs. Equation E.52 expresses a modification of the Wagner-Whitin prop-

erty presented on page 100 for capacitated lotsizing problems. It basically

states that in a given period t either all available capacity is consumed

¡ P ¢

( bm

t − i∈Km (trk · γkt + tbk · qkt ) = 0) or the Wagner-Whitin property for

uncapacitated lotsizing problems holds (qkt · yk,t−1 = 0).

Ã !

X

qkt · yk,t−1 · bm

t − (tri · γit + tbi · qit ) =0 ∀m, k ∈ Km , t = 1 . . . T

i∈Km

(E.52)

available capacity on the corresponding resource m, it can be used to avoid the

holding costs. Hence, for all resources m and periods t the shiftable quantities

of all items k ∈ Km are computed as:

( Ã ! )

X 1

m

min yk,t−1 , bt − (tri · γit + tbi · qit ) · (E.53)

tbk

i∈Km

58 See Derstroff (1995), page 141. See also Dixon and Silver (1981), page 26.

138 Multi-Level Capacitated Lotsizing with Setup Carryover

If the shiftable quantity is positive for at least one item, it is shifted from the

next earlier production periods into period t. If several possible shifts exist,

the respectively best shift is executed until E.52 holds.

avoided by shifting complete lots to the next earlier period, in which the item

is already produced, if possible. In other words, qkt is shifted to the next

period τ < t, in which equation E.54 holds.

Ã !

X

m

qkt · tbk < bτ − (tri · γiτ + tbi · qiτ ) · (γkτ + ωkτ ) (E.54)

i∈Km

Again, if several shifts are possible from period t, the respectively best are

selected.

improve the setup pattern. As explained in sections E.4.3.3.1 and E.4.3.3.2,

the best setup carryover decision is reflected for each shift of the finite loading

procedure.59 However, this is done sequentially, first for the target period and

then for the original period. In some cases, e. g. when production is shifted to

the directly precedent period, the optimal setup decisions in both periods may

be interrelated. Considering these interdependencies in the evaluation of each

possible shift would lead to an unreasonable increase in computational effort.

Thus, for the given plan, the setup carryover decisions are reconsidered. The

production quantities are not changed, nor are the capacity constraints vio-

lated. In each period, the best possible setup carryover, respecting the setup

and the carryover decisions in the adjacent periods, as well as the capacity

constraint of the current period, is selected.

three parts, the computation of the lower bound, the update of the Lagrangean

59 This is also true for steps 1 and 2 of the postoptimization procedure.

Chapter E. Solution Procedure based on Lagrangean Relaxation 139

Instead of relaxing the MLCLSPLP M the same way as we did the MLCLSPLe

and the MLCLSPLCC , we decided to apply the procedure presented above on

aggregated resource types. The rationale behind this is that the MLCSLPL

is decomposed into uncapacitated single-item problems for the computation

of the lower bounds. As capacity is unlimited, only the best resource of each

type would be used in each period.

This in turn implies that applying an individual Lagrangean multiplier for each

resource of the same type would not lead to additional benefits. Instead, it

might even reduce convergence, as respectively only one resource is considered

in the solution of each iteration.

For the construction of the upper bound, each individual resource is relevant as

a feasible solution to the MLCLSPLP M has to be constructed. We tried apply-

ing a modified feasibility routine, which assigned production to the individual

resources in each iteration. However, we found that the different computation

of the upper bound destabilized convergence of the Lagrangean multipliers and

thus lead to poor results. The schedules incurred increased holding costs as the

effect of inferior Lagrangean multipliers could not be offset by the feasibility

routine.

which applies the complete procedure presented above to an aggregate model,

the MLCLSPLP M AGG , in each iteration. After the stopping criterion of the

Lagrangean heuristic is met, the lots are distributed among the individual

resources of each type. Subsequently, a modified feasibility routine is used to

eliminate the capacity violations that stem from the underestimation of the

number of setups.

sources is adapted and the Lagrangean heuristic is restarted. If a feasible

schedule is generated, postoptimization is applied.

140 Multi-Level Capacitated Lotsizing with Setup Carryover

12.

l=0

repeat

solve the aggregate model using the Lagrangean heuristic

assign production to the parallel machines

enforce the capacity constraints

if solution is feasible then

postoptimization

break

else

adapt available capacities

end if

l=l+1

PM

until (l = lmax )

with

l current iteration

PM

lmax maximum number of iterations

In section E.5.1, we will present the adaptations made to apply the Lagrangean

heuristic as presented above. In section E.5.2, we will then present the routine

to compute a feasible schedule to the MLCLSPLP M . Finally, the adaptation

of the postoptimization procedure presented in section E.4.4 will be described

in section E.5.3.

the CLSPL with backorders and parallel machines, in which he aggregates

the capacities and production quantities across the machines belonging to the

same type of resource. We will apply the same principle here.

in a given period will be distributed among as few resources as possible. The

reason is that each additional machine used for the same item will require a

Chapter E. Solution Procedure based on Lagrangean Relaxation 141

approximates the number of setups required by the production on parallel

machines by assuming that at most one resource will be partially loaded by

this item. The number of resources required is computed according to equation

(E.55):

» ¼

qkt

Nkt = m (E.55)

bt

Consider a small example with three items 1, 2 and 3, which can be produced

on resources A and B. As figure E.1 shows, although the time required by each

of the items is less than a full resource capacity, at least one of the items has

to be split across the two resources.

1 2 t

Underestimating the number of setups will lead to at least equally good sched-

ules as the original problem during the computation of the lower bound.60

However, during the computation of the upper bound, the reserved setup time

and remaining available capacity may not suffice to construct a feasible so-

lution to the MLCLSPLPM . Thus, at the beginning of each iteration, the

60 To ensure that the solution is a lower bound to the MLCLSPLPM , we will not share the

assumption that resources can only carry over their setup state, if no other item has

been produced meanwhile.

142 Multi-Level Capacitated Lotsizing with Setup Carryover

t is updated. bt is initialized at the

unmodified available capacity bm

t . If the available capacity of any resource

n ∈ Pm is violated in any period t at the end of the overall procedure, it is

reduced by the overtime SOtn . The available capacity in iteration l is thus

computed as:

X

ebml = ebm,l−1 − SOtn,l−1 (E.56)

t t

n∈Pm

( )

X

SOtn,l−1

= max 0, n n em,l−1

(trk · γkt + tbk · qkt ) − bt (E.57)

k∈Km

To ensure that a feasible solution exists for the problem solved by the La-

grangean heuristic, the cumulative capacity remains unchanged. To do so,

P

n∈Pm SOt

n,l−1

is added in the next earlier period, in which ebm,l−1

t was not

completely consumed. If no such period exists, it is added in the first period.

Finally, the MLCLSPLP M AGG is not able to reflect that in the case of parallel

machines, the setup state can be carried over for one item on each machine in

each period. Hence, in the solution of the MLCLSPLP M , it will be beneficial

for more than one item to be produced in two adjacent periods on the same

resource type. Consequently, to manipulate the solution and hence increase

the overall solution quality, we chose to modify the setup costs sek . More

specifically, sek is computed by multiplying the original cost factors sk with the

parameter SCMOD:

K

Ã T !

X X

min Z = (e

sk · γkt + ek · (T − t + 1) · qkt ) + (hk − ek ) · qk1 − C

k=1 t=1

Chapter E. Solution Procedure based on Lagrangean Relaxation 143

(E.59)

subject to

t

X t

X X

(qkτ − dkτ ) + ybk ≥ akj · qj,τ +1 ∀k, t = 0, . . . , T (E.60)

τ =1 j∈Nk τ =0

X

(tbk · qkt + trk · γkt ) ≤ Nm · ebm

t ∀m, t = 1, . . . , T (E.61)

k∈Km

t

qkt ≤ γkt · + ωkt · t ∀m, k ∈ Km , t = 1, . . . , T (E.62)

tbk tbk

X

ωkt ≤ Nm ∀m, t = 1, . . . , T (E.63)

k∈Km

tbk · qk,t−1

ωkt ≤ + γk,t−1 ∀m, k ∈ Km , t = 2, . . . , T (E.65)

ebm

t−1

ωk1 = 0 ∀k (E.66)

Data:

sek modified setup cost coefficient of item k

ebm modified available capacity of resource type m in period t

t

The variables γkt and ωkt are now integers, which count the number of re-

sources with a setup operation or carryover for the item k in period t respec-

tively. The objective function (E.59) and the inventory balance constraints

144 Multi-Level Capacitated Lotsizing with Setup Carryover

(E.60) equal those of the MLCLSPLe .62 The capacity constraints (E.61) take

into account that not one but Nm resources of type m are available.

Equations (E.62) ensure that, under the assumption that items are not pro-

duced partially on more than one resource, items can only be produced, when

the resources are setup for the item.

straints (E.63)). And setups can be carried over, only if there has been a setup

in the previous period or if the setup state has been carried over already to the

prior period (constraints (E.64)). In the latter case, setups can be carried over

consecutively, only if production exploits the complete capacity of a resource

(constraints (E.65)).63 Finally, the setup state can only be carried over to

the first period, if there has been preproduction for the item in period zero

(constraints (E.66)).

Then, as γkt and ωkt are integer, they automatically become binary. Thus,

the single-item problems and the dynamic recursions remain the same as in-

troduced before. The fact that the production of an item may now require

several parallel machines with corresponding setup states will be accounted

for during the computation of the upper bound.

The lower bound is again computed according to equation (E.25),64 but the

constant part is now given as:

K X

X T

const = (hk − ukt ) · ((T − t + 1) · dkt − ybk )

k=1 t=1

K X

X T M X

X T

+ (hk − ek ) · Dkt + (vtm · ebm m

t + w t ) · Nm . (E.68)

k=1 t=1 m=1 t=1

62 See page 74. Thus, the constant part C is computed according to equation (D.31) on

page 74.

63 In the case of single setup carryovers, ωk,t−1 is excluded from constraints (E.64) and

constraints (E.65) are neglected.

64 See page 101.

Chapter E. Solution Procedure based on Lagrangean Relaxation 145

Machines

to the MLCLSPLP M 65 is constructed in two steps. First, the production

quantities are assigned to the individual resources. This will be described in

section E.5.2.1. If the solution is not feasible, a feasibility routine is applied

next to enforce the capacity constraints. This will be described in section

E.5.2.2.

capacities are adapted and the Lagrangean Heuristic is restarted.

Resources

to assign the production quantities computed by the Lagrangean heuristic to

the individual resources. Here, each resource type m is considered separately.

There is no interdependency between different resource types, as each produc-

tion quantity qkt computed for an item k ∈ Km in period t is assigned to the

resources n ∈ Pm in that period. As the inventory balance constraints are kept

in the solution of the aggregate model, they remain kept during the assignment

procedure. The goal of the procedure is to assign the production quantities

such that as few setup activities as possible are incurred. The outline of the

procedure is given in algorithm 13.

t = T to period t = 1. The reason is that, in the presence of setup carryovers,

setup activities can be shifted into the previous period, if the capacity limit

is met. The backward orientation permits installing setup carryovers while

considering only the production in the current period t and not the adjacent

periods t − 1 and t + 1.

65 The individual resources n, the original setup cost factors sk and available capacities

bm

t are considered.

146 Multi-Level Capacitated Lotsizing with Setup Carryover

for m ∈ M do

t=T

while t > 0 do

qbkt = qkt , ∀k ∈ Km ; b

bn m

t = bt , ∀n ∈ Pm

for n ∈ Pm do

for k ∈ Km do

if γkn > 0 then

n

compute qkt , qbkt and b

bn

t

end if

wend

end for

P

while qbkt > 0 do

k∈Km

max {bqkt ·tbk +trk }

k∈Km

compute i, qbit = tbi

n0

compute n, b bm

t = max {bb t }

n0 ∈Pm

m b

if qbit · tbi > bt and bt = bm

n

t then

compute qkt n

, qbkt and bbn

t

else if b bn

t ≥ (b

qit · tbi + tri ) then

for n0 ∈ Pm do

0

if b

bn

t ≥ (b

qit · tbi + tri ) then

add n0 to N 0

end if

end for

∗ 0

compute n∗, b

bn

t = min b

bn

t

n0 ∈N 0

n∗

compute qkt , qbkt and b

bn∗

t

else

n

compute qkt , qbkt and b

bn

t

end if

wend

for n ∈ Pm do

install setup carryovers

end for

t=t−1

wend

end for

with

qbkt quantity to be assigned on parallel machines for item k in period t

b

bn available capacity of an individual resource n in the case of parallel

t

machines in period t

Chapter E. Solution Procedure based on Lagrangean Relaxation 147

lot from the Lagrangean heuristic, qkt . The available capacity of each resource

n, bbn , is initialized at the capacity of the resource type, bm .

t t

First, existing setup activities are exploited. On a given resource n at most one

item k may have been scheduled with a setup carryover in t + 1 and the setup

activity pushed back into period t.66 If this is the case, the item is scheduled

on n also in t.

n

qkt is determined as

n bm

t − trk

qkt = min{b

qkt , }. (E.69)

tbk

n bm

t

qkt = min{b

qkt , }. (E.70)

tbk

Subsequently, qbkt and bbnt are reduced according to equations (E.71) and (E.72).

n

qbkt = qbkt − qkt (E.71)

t t kt kt

When all resources have been considered, the items are ordered according

to the remaining capacity requirement, qbkt · tbk + trk . The item i with the

maximum capacity requirement is considered. Resources are considered in

descending order of their available capacities bbn . Here, three cases are distin-

t

guished:

type, bm

t , and for at least one resource n the remaining available capacity

66 As the setup decisions are reconsidered here, the corresponding setup times are consid-

ered only after the scheduling decision. See constraint (E.72).

148 Multi-Level Capacitated Lotsizing with Setup Carryover

corresponds to bm

t .

67

Then, the complete capacity of n is consumed by

i. The production quantity is computed as:

n 1

qit = bm

t · (E.73)

tbi

the quantity for i is updated according to equation (E.71).

available capacity of at least one resource n may suffice to produce the

complete remaining quantity of k in t:

t

tified. Within this subset, resource n∗ with the least free capacity is

computed:

t 0 0 t

n ∈N

∗

with the greatest remaining capacity:

t 0 t

n ∈Pm

Again, qbkt and bbnt are updated according to equations (E.71) and (E.72).

In this case, overtime is incurred on n.

P

When all production quantities have been assigned, i. e. qbkt = 0, the

k∈Km

corresponding setup decisions are considered.

67 This implies that neither a setup nor production have been scheduled on n in t.

68 This implies that the capacity requirements for all other items are also less than a full

period capacity.

Chapter E. Solution Procedure based on Lagrangean Relaxation 149

On the resources, which are fully loaded by only one item, a setup carryover is

installed for k in t. On the remaining resources a setup carryover is installed,

n

if possible. This is the case, if an item k is scheduled on n in t (qkt > 0) and it

will be scheduled in t − 1 (qk,t−1 > 0). If several items fulfill this requirement,

the one with the highest setup costs sk is chosen.

tion is applied next. Otherwise, a feasibility routine is applied.

The resulting schedule is feasible with respect to the inventory balance con-

straints and the setup decisions. However, the capacity constraints of indi-

vidual resources may be violated, i. e. bbnt < 0. Then, the feasibility rou-

tine described in this section is applied. It consists of three steps. The first

step splits lots on resources of the same type in the same period (see section

E.5.2.2.1). The second step shifts setup activities to adjacent periods (see sec-

tion E.5.2.2.2). The third step shifts production to other periods (see section

E.5.2.2.3). The outline is given in algorithm 14.

least one machine with spare capacity exists, i. e. max {bbn } ≥ 0.

0

t

n0 ∈Pm

To shift production within m and t, all resources n, for which the capacity

constraint is violated, are considered in descending order of the the magnitude

of violation −bbn . All resources l with spare capacity are sorted in descending

t

order of their spare capacity bblt .

150 Multi-Level Capacitated Lotsizing with Setup Carryover

for m ∈ M do

t=1

while t ≤ T do

0

while b

bn min {b

bn bl bn0

t = t } < 0 and bt = max {bt } > 0 do

n0 ∈Pm n0 ∈Pm

shift production from n to l

update bbn and b

t bl t

wend

t=t+1

wend

for n ∈ Pm do

t=1

while t ≤ T do

if b

bn < 0 then

t

shift setup activity to t + 1

update b bn and b

t bn t+1

end if

if b

bn

t < 0 then

shift setup activity to t − 1

update b bn and b

t bn t−1

end if

t=t+1

wend

end for

t=1

while t ≤ T do

0

if min {b

bn

t } < 0 then

n0 ∈Pm

single-item shifts forward

single-item shifts backward

multi-item shifts backward

wend

end if

t=t+1

wend

end for

For each resource n, the item k 0 produced on n in t is selected, which incurs the

shortest setup time. The shiftable quantity is then computed as the minimum

of the lot for item k 0 , the overtime incurred on n and the spare capacity left

Chapter E. Solution Procedure based on Lagrangean Relaxation 151

on l:69

( )

−bbnt bblt − trk0

∆qknl0 = min qkn0 t , , (E.77)

tbk0 tbk0

sidered. If a feasible schedule is generated, postoptimization is applied next.

Otherwise, setup activities are shifted to adjacent periods.

In the presence of setup times and setup carryovers, capacity violations may

be reduced by shifting a setup activity from a period t with overtime to the

previous period t − 1 or the next period t + 1.

The setup activity for item k can be shifted from period t to the next period

t + 1, if it is not produced in t. This is the case, if it is only installed to

preserve the setup state for production in the next period. Thus, all items are

n n

considered, for which γkt = 1 and qkt = 0. Furthermore, the capacity left in

the next period must suffice to cover the setup time, i. e. bbnt+1 ≥ trk . If both

n n n

requirements are met for an item k, γkt and ωk,t+1 are set to zero and γk,t+1

is set to one. The available capacities of periods t and t + 1 are subsequently

updated and the procedure repeated, if necessary.

If the capacity violation persists, shifting setup activities into the previous

period is tried next. This is possible, only if no setup carryover is installed

in period t. All items k are considered, for which a setup is undertaken in t

n

and hence γkt = 1. Additionally, in the case of single setup carryovers there

n

may be no setup carryover to period t + 1, i. e. ωk,t+1 = 0. If consecutive

n

setup carryovers are allowed and ωk,t+1 is one, no other item than k may be

produced on n in t.

For the relevant items, the setup can be shifted into period t − 1, if the item

is already produced in t − 1. Else, the spare capacity has to suffice to under-

69 Note, that items have not been partially loaded on more than one machine before.

Hence, l has to be set up for k0 in any case.

152 Multi-Level Capacitated Lotsizing with Setup Carryover

distinguished:

n

trk ωk,t−1 =0

bbn ≥ (E.78)

t−1

tr − tr ∗ n

k k ωk,t−1 =1

If the setup state of item k has not been carried over to period t − 1, only

the setup time of k is considered. Otherwise, the setup carryover to t − 1 is

deleted. It becomes superfluous, when a setup activity is undertaken in t − 1.

Then, a different setup carryover may be installed for the most beneficial item

k ∗ , if possible. Then, the corresponding setup time is saved.

If a setup carryover could be installed for several items, the one with the

highest setup costs is chosen. The setup variables are adapted accordingly

and the available capacities of periods t and t − 1 are updated. Again, the

procedure is repeated if necessary.

timization is applied next. Otherwise, a modified finite loading heuristic is

applied, which is based on the procedure presented in section E.4.3.

The modified finite loading heuristic consists of one forward pass for each re-

source type m. In each period with capacity violations, three types of shifts

are considered, single-item forward, single item backward and multi-item back-

ward.

µ ¶

If the capacity of resource type m is violated in t max bn0

{bt } > 0 , all pos-

0

n ∈Pm

max

sible single-item shifts are evaluated. The shiftable quantity ∆qktτ from the

current period t to period τ ,70 is computed in equation (E.79). It is restricted

by the complete production quantity for k in t, the minimum inventory ∆y min

between t and τ and the free capacity in period τ . Again, changes to the setup

70 For backward shifts, all period τ < t are considered, for forward shifts all periods τ < t.

Chapter E. Solution Procedure based on Lagrangean Relaxation 153

½ m

¾

max min Cτ

∆qktτ = min qkt , ∆y , (E.79)

tbk

the shift. If production is shifted forward, the inventory balance constraints

of k have to be considered. If production is shifted backward, the inventory

balance constraints of predecessors of k have to be considered.71 Thus, the

minimum inventory is given as:

min yks

t≤s≤τ τ >t

∆y min = y (E.80)

min i,s−1 τ <t

i∈Vi ∧ aik

t≤s≤τ

Only shifts that eliminate all overtime in t are considered. Hence, the shiftable

quantity is computed as

bn P bn

−b max

tbkt

− bt ≥ ∆qktτ · tbk

n∈Pm

∆qktτ = P bn (E.81)

max

0 − bt < ∆qktτ · tbk

n∈Pm

uled in order to construct a feasible solution. This in turn may lead to ad-

ditional setups. Thus, the changes to the setup pattern are reflected in the

P bn

computation of − bt .

n∈Pm

If several shifts are possible, the one with the highest priority value is se-

lected.72 If no single-item shift can be executed, multi-item backward shifts

are applied next.

For multi-item backward shifts, the minimum inventory ∆y min is not consid-

max

ered for the computation of the shiftable quantity ∆qktτ in equation (E.79).

72 See section E.4.3.3.7.

154 Multi-Level Capacitated Lotsizing with Setup Carryover

Instead, when ∆qktτ has been computed, the minimum production quantity

of the predecessors is shifted backward to reinforce the inventory balance con-

straints.73

Otherwise, capacities are adapted and the Lagrangean heuristic is restarted.

E.5.3 Postoptimization

The same three postoptimization steps described in section E.4.4 are applied

here. Additionally, lots are shifted on the parallel resources within a period to

save additional setup costs. The four postoptimization steps undertaken here

are thus

The difference to the case without parallel machines is that for the first two

steps, not only the individual resource n is considered. Rather, all resources,

on which a given item k is produced in the relevant periods, are considered

simultaneously.

When steps 1 and 2 have been completed, lots are rescheduled within the same

period, if setups can be eliminated. This is the case, if an item is produced on

more than one resource in the same period and the capacity of one resource

suffices to produce the lot scheduled on a different resource.

section E.4.4.

Chapter F

Computational Study

Our computational tests were carried out on the instances presented in chapter

C. To evaluate the performance of our heuristic, we used the benchmark results

described in section D.1.5.

lower bound computed by CPLEX. Thus, the results are given as:

UBheuristic − LBCPLEX

gapCPLEX = · 100 (F.1)

LBCPLEX

The tests of the heuristic were carried out on an Intel Core 2 Duo processor

with 2.66 GHz and 2 GB RAM.

First, we carried out extensive tests on the MLCLSPL to find the best param-

eter configuration for the given test instances. As will be described in sections

F.1 and F.2, the performance of the Lagrangean heuristic is highly dependent

on the chosen parameters.

test set, which we used to evaluate the performance of our heuristic for the

155

L. Buschkühl, Multi-Level Capacitated Lotsizing with Setup Carryover,

Edition KWV, https://doi.org/10.1007/978-3-658-24034-9_6

156 Multi-Level Capacitated Lotsizing with Setup Carryover

MLCLSPL with single setup carryover, with consecutive setup carryover and

with parallel machines. This will be described in section F.3.

Various Parameter Combinations

First, we tested different parameter combinations for the MLCLSPL with sin-

gle setup carryover.

initial tests. The mean result1 over all instances for the standard configuration

is:

gapCPLEX

standard = 9.4659 (F.2)

For the standard configuration, four instances of class six could not be solved

and are thus neglected in the evaluation of the solution quality.

For the tests presented in the remainder of section F.1, the parameter com-

binations were tested with all other parameters set to the standard values

to isolate the effect of the individual parameters analyzed. However, the pa-

rameters presented in the same section are interdependent, but were tested

sequentially. Then, the relevant parameters were set to the best values. 2

The individual test scenarios and the results will be explained in the following.

Each section gives the standard values and a table with the other tested values

for the respective parameters. In the table with the tested values, the standard

combination, if contained in the test set, is marked bold.

Finally, a table with the results for each tested scenario is given. The results

are computed as the overall mean deviation of the results3 of the respective

parameter combination from that of the standard configuration:

standard (F.3)

1 The result is computed according to equation (F.1).

2 This is respectively indicated in the text.

3 Again, they are computed according to equation (F.1).

Chapter F. Computational Study 157

We chose to exclude all instances from the analysis of a given parameter, if they

could not be solved by at least one scenario. If not indicated otherwise, the

four instances, which the standard configuration could not solve, are excluded.

In each of the tables containing the results, the respectively best scenario is

marked bold.

Subgradients

E.34 on page 106, α is initialized at a given value α0 and updated with the

help of αred . In the standard configuration, α0 is set to 1 and αred to 0.25.

Furthermore, we tested the combinations given in table F.1.4

scenario 1.1 1.2 1.3 1.4 1.5 1.6 2.1 2.2 2.3 2.4 2.5 2.6

α0 0.25 0.25 0.25 0.25 0.25 0.25 0.5 0.5 0.5 0.5 0.5 0.5

αred 0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25

scenario 3.1 3.2 3.3 3.4 3.5 3.6 4.1 4.2 4.3 4.4 4.5 4.6

α0 0.75 0.75 0.75 0.75 0.75 0.75 1 1 1 1 1 1

αred 0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25

Table F.1: Tested values for the smoothing parameters of the subgradients

Table F.2 shows that it is beneficial to choose a high initial value for α. How-

ever, the results are not systematically ameliorating with a higher initial value

α0 . The effect of the reduction rate αred depends on the initial value. It is not

systematic either.

The overall best results were generated with α0 = 1 and αred = 0.15. Thus, we

used these values for the remaining tests of the Lagrangean multiplier update.

158 Multi-Level Capacitated Lotsizing with Setup Carryover

∆gapCPLEX 0.4161 0.4613 0.2896 0.2321 0.2752 0.3354

∆gapCPLEX 0.4736 0.3428 0.3229 0.3946 0.1152 0.2317

∆gapCPLEX 0.2506 0.2193 0.0961 0.1526 0.0816 0.0853

∆gapCPLEX 0.1647 0.2096 0.2318 -0.0088 0.103 0

Multipliers

For the update of the Lagrangean multipliers5 , three procedures have been

implemented. They are described in table F.3.

after lδ non-improving iterations δ l = δ l · δ red

estimate of the optimal solution Z ∗ = U B l−1

11 initialization of the factor δ l at δ 0

after lδ non-improving iterations δ l = δ l · δ red

estimate of the optimal solution

Z ∗ = min{U B l−1 , LB l · (1 + markupLB )}

12 initialization of the factor δ l at 2

after n non-improving iterations δ l = δ l · 12 and n = n

2

n is initialized at lmax

estimate of the optimal solution Z ∗ = U B l−1

Table F.3: Tested Combinations for the Parameters for the Lagrangean

multipliers

will first present our computations for the parameters δ 0 , δ red and lδ based on

5 See equations (E.35) and (E.36) on page 107.

Chapter F. Computational Study 159

MODUS 10. Then, different values for markupLB to compute the estimate of

the optimal solution for MODUS 11 will be given. Here, the dominant values

for δ 0 , δ red and lδ computed for MODUS 10 are applied. Finally, the three

procedures are compared along with different estimates for the upper bound

U B l−1 before the first feasible solution is found and different lower bounds

LB l (best or current).

First, we tested the different parameter combinations to update the factor δ for

MODUS 10.6 Their values in the standard configuration are δ 0 = 2, δ red = 0.5

and lδ = 4. Furthermore, we tested the combinations given in table F.4.7

δ0 1 1 1 1 1 1 1 1

δ red 0.5 0.5 0.5 0.5 0.75 0.75 0.75 0.75

lδ 2 4 8 10 2 4 8 10

δ0 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5

δ red 0.5 0.5 0.5 0.5 0.75 0.75 0.75 0.75

lδ 2 4 8 10 2 4 8 10

δ0 2 2 2 2 2 2 2 2

δ red 0.5 0.5 0.5 0.5 0.75 0.75 0.75 0.75

lδ 2 4 8 10 2 4 8 10

The results are given in table F.5. They ameliorate systematically with a

higher number of non-improving iterations lδ until the next adaptation. Also,

the results become systematically better with a higher initial value and a

higher rate of reduction δ red . Consequentially, the best results were achieved

with δ 0 = 2, δ red = 0.75 and lδ = 10. They were used for the remaining

computations in this subsection.

6 See equation (E.37) on page 108.

7 Remember that δ can take values between 0 and 2 only.

160 Multi-Level Capacitated Lotsizing with Setup Carryover

∆gapCPLEX 0.9885 0.6302 0.4045 0.2902 0.7777 0.4823 0.2864 0.2165

∆gapCPLEX 0.7777 0.4344 0.0159 -0.0284 0.5501 0.1607 -0.1202 -0.2538

∆gapCPLEX 0.4512 -0.0088 -0.2644 -0.4084 0.1388 -0.2563 -0.5459 -0.7266

Solution

For MODUS 11, the optimal solution is estimated by the lower bound and a

markup.8 We tested the values for markupLB as depicted in table F.6.9

scenario 1 2 3 4

markupLB 0.2 0.4 0.6 0.8

Table F.6: Tested values for the estimate of the upper bound

Table F.7 shows that the results ameliorate with a higher markup in the given

interval. The best results were generated with a markup of 0.8.

scenario 1 2 3 4

∆gapCPLEX 0.664 0.1955 0.0084 -0.2143

9 We kept δ 0 = 2, δ red = 0.75 and lδ = 10 from the computations for MODUS 10.

Chapter F. Computational Study 161

With the best configurations for MODUS 10 and MODUS 11, we tested dif-

ferent choices of the lower bound LB l10 and markupU B on the upper bound.

Before the first feasible solution is found, the upper bound is estimated as

U B = max{LB ∗ , LB l } · markupU B .

The lower bound LB l is either the current lower bound (LB λ = current) or the

maximum lower bound (LB λ = maximum) found so far. The tested parameter

combinations are given in table F.8.

MODUS 10 10 10 10

markupU B 0 0 2 2

LB λ current maximum current maximum

MODUS 11 11 11 11

markupU B 0 0 2 2

LB λ current maximum current maximum

MODUS 12 12 12 12

markupU B 0 0 2 2

LB λ current maximum current maximum

Table F.8: Tested values for the stepsize update: upper and lower bounds

The results are given in table F.9. The markup on the upper bound has no

effect on the solution quality.11

The choice of the current lower bound as opposed maximum lower bound leads

to better results for all procedures applied.

11 We also tested markups of 4, 6 and 8 with the same result.

162 Multi-Level Capacitated Lotsizing with Setup Carryover

∆gapCPLEX -0.7266 -0.6239 -0.7266 -0.6239

∆gapCPLEX -0.2143 -0.3025 -0.2143 -0.3025

∆gapCPLEX -0.9418 -0.8571 -0.9418 -0.8571

Loading Heuristic

The maximum number of iterations lmax is the stopping criterion of the finite

loading heuristic.12 In the standard configuration it is set to 50. We tested

various values between 25 and 150. For our test sample, the overall solution

quality was not affected by the choice of lmax within this interval. As the

stopping criterion is relevant, only if no feasible solution has been found earlier

and due to the solution speed of the heuristic, we chose to keep lmax = 50.

Resources can be sorted independently for forward passes (FOR) and for back-

ward passes (BACK). They are respectively considered in order of capacity uti-

lization (KAP), according to the items they produce (M) or in reverse order

of M (M). When the resources are ordered with respect to the corresponding

items, the order is determined either by the number of predecessor and succes-

sor relationships of the corresponding items with items on different machines

(0) or by the low-level codes (1). The capacity criterion is the relative capacity

utilization (0), the Lagrangean multipliers of the capacity constraints (1) or

the absolute overtime (2). Resources are respectively considered in descending

order of the criterion.13

12 See algorithm 7 on page 116.

13 See section E.4.3.2.

Chapter F. Computational Study 163

FOR KAP KAP KAP

BACK KAP KAP KAP

KAP 0 1 2

FOR KAP KAP KAP KAP KAP KAP

BACK M M M M M M

M 0 0 0 1 1 1

KAP 0 1 2 0 1 2

FOR KAP KAP KAP KAP KAP KAP

BACK M M M M M M

M 0 0 0 1 1 1

KAP 0 1 2 0 1 2

FOR M M M M M M

BACK KAP KAP KAP KAP KAP KAP

M 0 0 0 1 1 1

KAP 0 1 2 0 1 2

FOR M M

BACK M M

M 0 1

FOR M M

BACK M M

M 0 1

FOR M M M M M M

BACK KAP KAP KAP KAP KAP KAP

M 0 0 0 1 1 1

KAP 0 1 2 0 1 2

FOR M M

BACK M M

M 0 1

FOR M M

BACK M M

M 0 1

164 Multi-Level Capacitated Lotsizing with Setup Carryover

their relative capacity utilization both during forward and during backward

passes. Furthermore, we tested the combinations given in table F.10.

The results are given in table F.11. It shows that, when the criterion for

the forward passes is set to the product structure, changing from M to M

affects the solution. All other parameters have no influence. When resources

are ordered according to their capacity requirement during the forward pass,

sorting them according to the product structure during the backward pass is

beneficial. Changing the order of consideration for the backward pass14 does

not influence the solution quality. Changing the capacity criterion affects the

solution.

∆gapCPLEX 0 0.0407 0.0069

∆gapCPLEX -0.0006 0.0403 0.0064 -0.0006 0.0403 0.0064

∆gapCPLEX -0.0006 0.0403 0.0064 -0.0006 0.0403 0.0064

∆gapCPLEX 0.0474 0.0474 0.0474 0.0474 0.0474 0.0474

∆gapCPLEX 0.0474 0.0474

∆gapCPLEX 0.0474 0.0474

∆gapCPLEX 0.0054 0.0054 0.0054 0.0054 0.0054 0.0054

∆gapCPLEX 0.0054 0.0054

∆gapCPLEX 0.0054 0.0054

Chapter F. Computational Study 165

Overall, the best results were achieved, when during forward passes resources

are considered in order of the relative capacity utilization, while during back-

ward passes they are considered in some way according to the product struc-

ture.

Periods

In the standard configuration, all resources are considered for each period,

then the next period. We also tried the combinations given in table F.12.

scenario MP/SEQ

1 all resources are considered for each period, then the next period

2 all periods are considered forward and backward for each resource,

forward and backward passes are executed separately

3 all periods are considered forward for each resource, then backward,

forward and backward passes are executed together

Table F.12: Tested values for the order of consideration of resources and

periods

The results are given in table F.13. Six instances could not be solved by

at least one parameter combination and have been excluded. Overall, the

standard configuration proofed to be most beneficial.

scenario 1 2 3

∆gapCPLEX 0 1.8983 2.2667

Table F.13: Results for the order of consideration of resources and periods

166 Multi-Level Capacitated Lotsizing with Setup Carryover

During backward shifts, shifting deteriorating (P R < 0) or all (all) partial lots

is prohibited until iteration lLS .15

lLS 10 10 10 10 10 10 10 10

LS all all all all PR < 0 PR < 0 PR < 0 PR < 0

SAV 0 1 2 3 0 1 2 3

lLS 12 12 12 12 12 12 12 12

LS all all all all PR < 0 PR < 0 PR < 0 PR < 0

SAV 0 1 2 3 0 1 2 3

lLS 14 14 14 14 14 14 14 14

LS all all all all PR < 0 PR < 0 PR < 0 PR < 0

SAV 0 1 2 3 0 1 2 3

lLS 16 16 16 16 16 16 16 16

LS all all all all PR < 0 PR < 0 PR < 0 PR < 0

SAV 0 1 2 3 0 1 2 3

lLS 18 18 18 18 18 18 18 18

LS all all all all PR < 0 PR < 0 PR < 0 PR < 0

SAV 0 1 2 3 0 1 2 3

lLS 20 20 20 20 20 20 20 20

LS all all all all PR < 0 PR < 0 PR < 0 PR < 0

SAV 0 1 2 3 0 1 2 3

reflected in the choice of the best shift or not. In the first case, additional

cumulated overtime is not reflected at all (0). In the second case, the priority

values are weighted with additional cumulated overtime (1). In the third case,

15 See section E.4.3.3.5.

Chapter F. Computational Study 167

overtime is executed (2). Finally, in the fourth case, the priority values are

weighted and the best shift with the least additional overtime is selected (3).

In the standard configuration, partial shifts are prohibited, only if their priority

value is negative. This is done until iteration lLS = 4. Additional cumulated

overtime is not reflected. We also tested the combinations given in table F.14.

The results are given in tables F.15. No combination lead to better results

than the standard configuration.

∆gapCPLEX 0.0576 0.0692 0.0694 0.0696 0.0577 0.0693 0.0694 0.0696

∆gapCPLEX 0.0569 0.0621 0.0621 0.0624 0.0569 0.0621 0.0621 0.0624

∆gapCPLEX 0.0569 0.0611 0.0612 0.0615 0.0569 0.0611 0.0612 0.0615

∆gapCPLEX 0.0608 0.0652 0.0653 0.0656 0.0608 0.0652 0.0653 0.0656

∆gapCPLEX 0.0537 0.0609 0.0609 0.0612 0.0537 0.0609 0.0609 0.0612

∆gapCPLEX 0.0604 0.0661 0.0662 0.0665 0.0604 0.0661 0.0662 0.076

For multi-item forward shifts (VARFOR) either only the shiftable quantity

of the final item is extrapolated (min) or inventory is used to shift further

production (max).16 Multi-item backward shifts (VARBACK) are either ne-

glected in iterations prior to lLS (min) or the quantity causing overtime is

shifted (max).17

17 See section E.4.3.3.6.

168 Multi-Level Capacitated Lotsizing with Setup Carryover

In the standard configuration, only the shiftable quantity of the final item

is extrapolated. During backward shifts, multi-item shifts are neglected in

iterations prior to lLS . Furthermore, we tested the combinations given in table

F.16.

VARFOR min min max max

VARBACK min max min max

The results are given in table F.17. Again, the standard configuration proofed

most beneficial. Changing the criterion for forward passes leads to a rather in-

significant deterioration of the overall solution quality. Changing the criterion

for backward passes has no significant effects on the solution quality.

∆gapCPLEX 0 1.2096 0.014 1.2632

As described in section E.4.3.3, the heuristic is divided into two parts by the

parameter lO . At the beginning of the procedure, additional overtime may be

incurred in the target period. By contrast, in iterations later than lO , shifts

are only allowed, if they do not incur additional overtime in the target period.

In the standard configuration, lO is set to 10. Furthermore, we tested the

values given in table F.18.

Chapter F. Computational Study 169

scenario 1 2 3 4 5 6

lO 10 12 14 16 18 20

Table F.18: Tested values for the first iteration of careful shifts

Five instances have been excluded from the evaluation. Table F.19 shows that

increasing lO has a deteriorating effect on the solution quality. However, the

effect is not significant.

scenario 1 2 3 4 5 6

∆gapCPLEX 0 0.0013 0.0017 0.0025 0.03 0.04

Values

The parameters for the priority values consist of the cost factors applied

(PRC), the weight (PRW) and the iteration, until which only improving shifts

are allowed (lP R ). Either the modified cost factors of the SLULSPLs (mod-

ified) or the original cost factors (original) are applied. They are either not

weighted (no) or weighted with overtime (yes).18

In the standard configuration, the modified cost factors of the SLULSPLs are

applied. The priority values are not weighted with overtime. Furthermore,

lP R is set to 0, which means that deteriorating shifts are always allowed.

The results are given in table F.21. The difference in solution quality is signif-

icant. The greatest impact stems from prohibiting deteriorating shifts during

the first iterations. Using the original cost factors leads to better results than

170 Multi-Level Capacitated Lotsizing with Setup Carryover

PRW no no no no no no

lP R 0 2 4 6 8 10

PRW yes yes yes yes yes yes

lP R 0 2 4 6 8 10

PRW no no no no no no

lP R 0 2 4 6 8 10

PRW yes yes yes yes yes yes

lP R 0 2 4 6 8 10

Table F.20: Tested values for the computation of the priority values

using modified once. The impact of weighting the priority value with overtime

is less strong.19

∆gapCPLEX 0 -1.7016 -2.2524 -2.2384 -2.2278 -2.1887

∆gapCPLEX 0.1191 -1.5095 -2.1664 -2.1725 -2.159 -2.173

∆gapCPLEX 0.0033 -1.7258 -2.2718 -2.257 -2.2481 -2.2113

∆gapCPLEX 0.0832 -1.4983 -2.177 -2.1837 -2.1857 -2.1751

19 Due to the importance of the choice of priority values, we redid the test in combination

with the maximum number of iterations of the overall procedure and the optimal values

for the remaining parameters. This will be described in section F.2.

Chapter F. Computational Study 171

Overall, the best results are achieved when the original cost factors are used,

the priority values are not weighted with overtime and deteriorating shifts are

prohibited until period 4.

In the standard configuration, tabu lists are not applied. We tested different

combinations for the first iteration the tabu lists are applied lTABU , the max-

imum duration of a tabu entry tTABU and the length of the tabu lists nTABU

as depicted in table F.22.

TABU

l 8 8 8 8 8 8 8 8

tTABU 3 3 3 3 6 6 6 6

nTABU 5 10 15 20 5 10 15 20

lTABU 16 16 16 16 16 16 16 16

tTABU 3 3 3 3 6 6 6 6

nTABU 5 10 15 20 5 10 15 20

TABU

l 24 24 24 24 24 24 24 24

tTABU 3 3 3 3 6 6 6 6

nTABU 5 10 15 20 5 10 15 20

The results are given in table F.23. All instances could be solved by at least

one parameter combination, but six test instances could not be solved by each

one. Thus, they were neglected in the evaluation.

172 Multi-Level Capacitated Lotsizing with Setup Carryover

∆gapCPLEX 0.0416 0.0539 0.0473 0.0503 0.0416 0.0523 0.0474 0.0475

∆gapCPLEX 0 0 0 0 0 0 0 0

∆gapCPLEX 0 0 0.0006 0.0001 -0.0001 0.0007 0.0004 0.0008

The best results were achieved with lTABU = 24, tTABU = 6 and nTABU = 5.

However, the benefit is marginal. Also, results were most stable with the first

iteration of the application of tabu lists set to 16. Thus, we will keep the result

of a different test set, in which the best results were found for lTABU = 16,

tTABU = 3 and nTABU = 5.

Finally, we tested the aspiration criterion. The combinations of the first it-

eration lASP , in which the aspiration criterion is applied, and the probability

ASP of accepting a tabu move are given in table F.24.

lASP 24 24 24 24 26 26 26 26

ASP 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8

lASP 28 28 28 28 30 30 30 30

ASP 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8

In our test set, the application of an aspiration criterion did not change the

results. Thus, we will again keep the results of prior testing, in which the best

combination was lASP = 21 and ASP = 0.6.

Chapter F. Computational Study 173

Further Tests for Selected Parameters

Parameters

From the tests described in section F.1, we deducted the dominant parameter

configurations for the different aspects of the heuristic. They will be given in

section F.2.1.

With all parameters set to their dominant values, we conducted further tests

for the total number of iterations and the computation of the priority values.

The test of the latter was included due to the significance of the results. The

tests are described in section F.2.2.

Finally, we tested the setup cost reduction factor for parallel machines. This

will be described in section F.2.3.

αred 0.15 factor for the update of the smoothing parameter αl

MODUS 12 initialization of the factor δ l at 2

after n non-improving iterations δ l = δ l · 12 and n = n

2

n is initialized at lmax

estimate of the optimal solution Z ∗ = U B @−1

δ0 - not relevant for MODUS 12

δ red - not relevant for MODUS 12

lδ - not relevant for MODUS 12

markupLB - not relevant for MODUS 12

markupU B 0 markup to estimate the upper bound before the first feasible solution

U B = max{LB ∗ , LB l } · markupU B

LB λ 0 use the current lower bound to the stepsize λ

Table F.25: Dominant configuration for the parameters for the Lagrangean

multipliers

174 Multi-Level Capacitated Lotsizing with Setup Carryover

In table F.25, the dominant parameter configuration for the Lagrangean mul-

tipliers are given.

The stopping criterion for the finite loading heuristic will be kept from the

standard configuration, as testing lead to no differing results.20

Table F.26: Dominant Configuration for the Parameter for the termination

of the finite loading heuristic

MP 0 all resources are considered for one period, then the next period

SEQ 0 forward and backward passes are executed separately

FOR 0 forward shifts: order of resources determined by KAP

BACK 1 backward shifts: order of resources determined by M

KAP 0 sorting in descending order of relative capacity utilization

M 1 sorting in descending order of the low-level codes of the corresponding items

Table F.27: Dominant configuration for the parameters for the ordering of

shifts

VARFOR 0 multi-item forward shifts: only the shiftable quantity of the final item is

extrapolated

lLS 4 backward shifts: iteration until which shifts

of partial lots according to LS are prohibited

LS 1 backward shifts: partial shifts with negative priority values are prohibited

SAV 0 backward shifts: no adaptation to additional cumulated overtime

VARBACK 0 multi-item backward shifts: are neglected in iterations prior to lLS

Table F.28: Dominant configuration for the parameters for the shifts

Chapter F. Computational Study 175

The dominant values for the ordering of resources and periods are given in

table F.27.

Table F.28 gives the dominant values for the parameters to determine the

shiftable quantities for the different types of shifts.

Finally, table F.29 contains the values for the parameters of the tabu lists.

They are not deducted from the test set as the results were not significant.

Instead, we selected the values obtained in prior tests.

nTABU 5 length of tabu list

tTABU 3 maximum duration a tabu entry is kept

lASP 21 first iteration the aspiration criterion is applied

ASP 0.6 probability of accepting a tabu shift when the aspiration criterion is applied

Table F.29: Dominant configuration for the parameters for the application of

tabu lists

The maximum number of iterations lmax is the stopping criterion of the La-

grangean heuristic. In the standard configuration, it is set to 70. As the choice

of the computation of the priority values had such a high impact on the solu-

tion quality,21 we decided to test the stopping criterion of the overall heuristic

together with the computation of the priority values. From the tests described

above we will keep the iteration lP R , until which only improving shifts are

allowed, at 4. The further parameter combinations are given in table F.30.

The results are given in table F.31. For all types of computation for the priority

value, the quality of the solution systematically increases with the maximum

number of iterations. Furthermore, weighting the priority values with overtime

is beneficial regardless the number of iterations.

176 Multi-Level Capacitated Lotsizing with Setup Carryover

PRC modified modified modified modified modified modified

PRW no no no no no no

lmax 70 80 90 100 110 120

PRC modified modified modified modified modified modified

PRW yes yes yes yes yes yes

lmax 70 80 90 100 110 120

PRC original original original original original original

PRW no no no no no no

lmax 70 80 90 100 110 120

PRC original original original original original original

PRW yes yes yes yes yes yes

lmax 70 80 90 100 110 120

Table F.30: Tested values for the computation of the priority values

∆gapCPLEX -2.8294 -2.9021 -2.9795 -3.027 -3.1468 -3.1887

∆gapCPLEX -2.9208 -2.9974 -3.0714 -3.1244 -3.2395 -3.2946

∆gapCPLEX -2.8831 -2.9576 -3.0458 -3.0843 -3.2226 -3.2714

∆gapCPLEX -2.965 -3.0562 -3.1187 -3.1737 -3.27 -3.3352

Table F.31: Results for the maximum number of per TBO profile

From these further tests, we deducted the dominant parameters for the com-

putation of the priority values as shown in table F.32.

Chapter F. Computational Study 177

PRC 1 use modified cost factors of the SLULSPLs for priority values

PRW 1 priority values are not weighted with overtime

lP R 4 iteration until which only improving shifts are allowed

Table F.32: Dominant configuration for the parameters for the computation

of priority values

The test sample for parallel machines is described in section C.7. For the

subsets of classes 2, 4 and 6 with setup time profile 1 and TBO profile 2 (4 for

class 2), we introduced to profiles of parallel machines. We will refer to them

as PM1 and PM2.22

There are two parameters for parallel machines. The first is the number of

P M 23

iterations lmax . Remember, that for parallel machines the next iteration

PM

is started, only if no feasible solution is found. Here, we set the lmax to 20.

Some further testing showed that instances, which could be solved after 20

PM

iterations, could not be solved with a higher lmax either.

The second parameter is SCMOD. It is used to modify the setup costs for the

aggregate model, MLCLSPLP M AGG .24

We conducted two tests for SCMOD for the MLCLSPLP M with single setup

carryovers, as well as for the MLCLSPLP M CC with consecutive setup carry-

overs. The reason is that in the consecutive setup carryover case, it may be

beneficial to produce an item in even more periods, as more setup carryovers

can be installed for the same item.

First, we tested the values given in table F.33 for the MLCLSPLP M with single

setup carryovers.

23 See algorithm 12 on page 140.

24 See equation E.58 on page 142.

178 Multi-Level Capacitated Lotsizing with Setup Carryover

scenario 1 2 3

SCMOD 0.25 0.5 0.75

Table F.33: Tested values for the setup cost modification - test 1

The results are given in table F.34 for the two profiles PM1 and PM2 of parallel

machines as described in sectionC.7. For each profile, three instances could not

be solved by at least one scenario and have been excluded. For ∆gapCPLEX

PM ,

the best scenario (SCMOD = 0.5) of each profile was selected as reference

value.

PM1 PM2

scenario 1 2 3 1 2 3

∆gapCPLEX

PM 1.7374 0 0.3693 0.5846 0 1.6162

Table F.34: Results for the setup cost modification for the MLCLSPLP M -

test 1

As the best results for both profiles were generated with SCMOD set to 0.5,

we further conducted the tests given in table F.35.

scenario 1 2 3 4 5 6 7 8 9

SCMOD 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7

Table F.35: Tested values for the setup cost modification for the

MLCLSPLP M - test 2

The results are contained in tables F.36. Again, for each profile three instances

could not be solved by at least one scenario and have been excluded.

Chapter F. Computational Study 179

PM1 PM2

scenario 1 2 3 1 2 3

∆gapCPLEX

PM 1.2609 1.1042 0.4135 0.9135 0.3571 0.2563

scenario 4 5 6 4 5 6

∆gapCPLEX

PM 0.2019 0 0.1723 0.0425 0 0.3317

scenario 7 8 9 7 8 9

∆gapCPLEX

PM 0.1978 0.0968 0.2388 0.7451 1.2106 1.307

Table F.36: Results for the setup cost modification for the MLCLSPLP M -

test 2

Again, SCMOD = 0.5 lead to the best results for both profiles.

Also for the MLCLSPLP M CC with consecutive setup carryovers, we tested the

values given in table F.33 first.

The results are given in table F.37 for the two profiles PM1 and PM2. For

each profile, two instances could not be solved by at least one scenario and

have been excluded. Again, the best scenario was selected as reference for

∆gapCPLEX

PM . For PM2, the results did not differ. Thus, SCMOD = 0.25 was

selected for both profiles.

PM1 PM2

scenario 1 2 3 1 2 3

∆gapCPLEX

PM 0 0.0096 0.0096 0 0 0

Table F.37: Results for the setup cost modification for the MLCLSPLP M CC

- test 1

180 Multi-Level Capacitated Lotsizing with Setup Carryover

scenario 1 2 3 4 5 6 7 8

SCMOD 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Table F.38: Tested values for the setup cost modification for the

MLCLSPLP M CC - test 2

The results are contained in tables F.39. For each profile two instances could

not be solved by at least one scenario and have been excluded.

PM1 PM2

scenario 1 2 3 1 2 3

∆gapCPLEX

PM 0 0.0169 0.0169 0.0117 0.0117 0.0115

scenario 4 5 6 4 5 6

∆gapCPLEX

PM 0.0169 0.0169 0.0169 0.0115 0.0115 0 0.0117

scenario 7 8 7 8

∆gapCPLEX

PM 0.0169 0.0169 0.0115 0.0115

Table F.39: Results for the setup cost modification for the MLCLSPLP M CC

- test 2

The best results for the MLCLSPLP M CC were computed with SCMOD = 0.05

for PM1 and SCMOD = 0.25 for PM2.

the Heuristic

With the best parameter combinations found during our tests, we finally tested

the performance of the overall heuristic for the different model formulation.

iterations, or earlier, if the subgradients become sufficiently small. In section

Chapter F. Computational Study 181

F.2, we tested values up to 250 for lmax . As table F.31 shows, the quality of

the solution increases with the number of iterations in the given interval.

40000

35000

30000

25000

20000

15000

10000

5000

0

0 120 240 360 480 600 720 840 960 1080 1200

Figure F.1: Development of the upper and the lower bound for a given test

instances

Figure F.1 gives the development of the best upper bound and the best lower

bound for one test instance of the MLCLSPLe . The horizontal line represents

the optimal solution obtained with CPLEX. Although this holds true for some

rather than all test instances, F.1 shows that significant improvements both

of the upper and of the lower bound are found after more than 120 iterations.

Thus, we decided to test different values for the stopping criterion lmax of the

overall procedure.

tested the values given in table F.40.

182 Multi-Level Capacitated Lotsizing with Setup Carryover

scenario 1 2 3 4 5

lmax 120 240 360 480 1200

Table F.40: Tested values for the maximum number of iterations lmax for the

MLCLSPL without parallel machines

For the MLCLSPLP M with parallel machines, we tested the values given in

table F.41. We did not conduct further with higher values for lmax , because

PM

the Lagrangean heuristic may be executed up to lmax = 20 times.

scenario 1 2

lmax 70 120

Table F.41: Tested values for the maximum number of iterations lmax for the

MLCLSPL with parallel machines

four criteria:

2. the average deviation of the best upper bound found by the heuristic

from the lower bound of CPLEX as computed by equation F.1,

3. the average deviation of the upper bound and the lower bound given by

the heuristic and

The average deviation of the upper bound and the lower bound given by the

heuristic is computed as:

UBheuristic − LBheuristic

gapheuristic = · 100 (F.4)

LBheuristic

Chapter F. Computational Study 183

our analysis into the following four categories:

The first type of table gives the results per class. This demonstrated the

solution quality for different problem sizes. The second type of table gives

the results per setup time profile and capacity profile. We chose to consider

the two together as the setup time profile influences the capacity utilization

of a given resource. The third type of table gives the results per TBO profile

and thus per setup and holding cost ratio. TBO profiles 3 and 4 are only

relevant for classes 1 and 2. Finally, the fourth type of table gives the results

per product and process structure and per demand profile.

In section F.3.1, the results for the MLCLSPLe will be given. Then, sec-

tion F.3.2 gives the results for the MLCLSPLCC . Finally, the results for the

MLCLSPLP M will be presented in section F.3.3.

Carryover

scenario 1 2 3 4 5

] feasible solutions 1917 1916 1917 1916 1917

% feasible solutions 99.8438 99.7917 99.8438 99.7917 99.8438

25 See chapter C.

184 Multi-Level Capacitated Lotsizing with Setup Carryover

table F.42.

Remember, that MODUS is set to 12.26 Thus, lmax affects the update of the

Lagrangean multipliers. Therefore, a higher lmax may lead to inferior results.

Tables F.43 to F.47 contain the solution quality and computation time for the

1916 test instances, which could be solved by all scenarios.

scenario 1 2 3 4 5

gapCPLEX 6.1307 5.7222 5.5225 5.2739 4.8884

gapheuristic 14.2426 12.7222 12.0175 11.2989 9.8812

CPU sec. 0.147 0.2766 0.4075 0.5408 1.3459

Table F.43 shows that the deviations of the optimal solutions provided by

the heuristic from the best lower bounds computed by CPLEX ameliorate

significantly with the maximum number of iterations. The best results were

computed for lmax = 1200. Here, the average deviation from the lower bound

provided by CPLEX is 4.8884% of the latter.

Also, the average deviation of the results from the lower bound provided by

the heuristic become better with higher maximum numbers of iterations. For

lmax = 1200, the average deviation from the lower bound provided by the

heuristic is 9.8812% of the latter.

Not surprisingly, the average computation time in CPU seconds increases with

higher values for lmax . Still, even for lmax = 1200 only 1.3459 CPU seconds

on average are required.

Information on the different classes of our test sample is given in table C.1

on page 56. Table F.44 shows that for the MLCLSPLe the solution quality

deteriorates, when the number of periods increases relative to the number of

items. It is not surprising that the solution quality for the small problems of

Chapter F. Computational Study 185

classes 1 and 2 is rather good. Notably, however, the heuristic performs better

for classes 5 and 6 than for classes 3 and 4, although the former comprise 40

items while the latter only comprise 20 items.

scenario 1 2 3 4 5

CPLEX

class 1 gap 1.1708 1.0750 1.0696 1.0696 1.0639

gapheuristic 3.9063 3.3263 3.0652 2.9297 2.6091

CPU sec. 0.0065 0.0095 0.0141 0.0186 0.0437

gapheuristic 10.1007 8.2464 7.3086 6.6624 5.1032

CPU sec. 0.0216 0.0407 0.0577 0.0775 0.1966

gapheuristic 14.1548 12.9454 12.839 12.4932 11.9357

CPU sec. 0.0901 0.1721 0.2606 0.3500 0.8920

gapheuristic 34.9442 33.0065 32.0457 30.4812 25.8122

CPU sec. 0.4004 0.7518 1.1159 1.4830 3.6735

gapheuristic 13.6532 12.2558 11.8458 10.5095 9.9711

CPU sec. 0.1337 0.2563 0.3822 0.5230 1.4116

gapheuristic 23.3260 20.5551 18.7740 17.8321 16.0079

CPU sec. 0.5020 0.9431 1.3737 1.7994 4.3596

The five capacity profiles are given in table C.5 on page 64. The two setup

time profile are given in table C.3 on page 60. Table F.45 shows that the

results are better for instances with high capacity utilization profiles. The

best results were computed for capacity profiles 4 and 5, where the capacity

utilization of the resources differed. The influence of the setup time profiles is

neither important nor systematic.

186 Multi-Level Capacitated Lotsizing with Setup Carryover

scenario 1 2 3 4 5

gapheuristic 15.4882 14.5081 14.1308 13.0331 12.5519

CPU sec. 0.0442 0.0842 0.1186 0.1679 0.4469

gapheuristic 15.2185 14.1051 13.7395 12.7780 12.3180

CPU sec. 0.0481 0.0822 0.1165 0.1608 0.4256

gapheuristic 17.1471 14.7468 13.9589 13.2659 11.6662

CPU sec. 0.0792 0.1460 0.2162 0.2936 0.7257

gapheuristic 14.515 13.1829 12.6413 11.5614 9.8267

CPU sec. 0.0854 0.1432 0.2157 0.2831 0.7184

gapheuristic 19.2396 17.2527 16.0245 15.2228 12.2287

CPU sec. 0.2739 0.5213 0.7700 1.0320 2.5186

gapheuristic 14.2277 12.6792 11.8474 11.1289 9.3954

CPU sec. 0.2797 0.5287 0.7791 1.0332 2.5578

gapheuristic 11.2879 9.9278 9.3034 8.9400 7.8309

CPU sec. 0.1722 0.3394 0.5001 0.6340 1.5608

gapheuristic 10.4824 9.1559 8.4378 8.1308 7.0452

CPU sec. 0.1805 0.3349 0.4966 0.6511 1.5867

gapheuristic 14.4725 12.5972 11.7374 11.0093 9.1471

CPU sec. 0.1506 0.2903 0.4224 0.5710 1.4366

gapheuristic 10.4511 9.1608 8.4369 8.0007 6.8503

CPU sec. 0.1594 0.3008 0.4472 0.5919 1.5063

Table F.45: Results for the MLCLSPLe per capacity and setup time profile

The setup and holding cost ratios are depicted in table C.4 on page 62. TBO

profiles 3 and 4 are considered for classes 1 and 2 only. Table F.46 shows that

the heuristic performs significantly better, when the setup and holding cost

ratio differs for various items than when they are equal for all items.

Chapter F. Computational Study 187

scenario 1 2 3 4 5

gapheuristic 25.8248 23.2169 21.9903 20.4969 17.8689

CPU sec. 0.2098 0.3990 0.5882 0.7765 1.9041

gapheuristic 8.1525 7.4839 7.1642 6.9080 6.1954

CPU sec. 0.1723 0.3214 0.4750 0.6343 1.6112

gapheuristic 7.5803 5.6617 4.8494 4.4878 3.6221

CPU sec. 0.0169 0.0286 0.0406 0.0530 0.1251

gapheuristic 4.4742 4.0571 3.8694 3.7289 3.2702

CPU sec. 0.0139 0.0243 0.0317 0.0441 0.1029

Table F.46: Results for the MLCLSPLe per setup and holding cost ratio

The demand profiles are described in section C.3. Table F.47 shows that the

results are inferior for general product structures than for assembly ones. Also,

the best results are generated for demand profile 1, where the coefficient of

variation is 0.2. In general, the results for cyclic process structures are slightly

inferior to those for acyclic process structures.

188 Multi-Level Capacitated Lotsizing with Setup Carryover

scenario 1 2 3 4 5

gapheuristic 3.6692 3.177 2.9907 2.9337 2.634

CPU sec. 0.0906 0.1764 0.2446 0.3117 0.7460

gapheuristic 10.2661 8.5992 8.0255 7.0456 6.3397

CPU sec. 0.1356 0.2559 0.3731 0.5003 1.3096

gapheuristic 12.8835 10.9682 9.9278 9.0908 7.0694

CPU sec. 0.1535 0.2954 0.4159 0.5211 1.2567

gapheuristic 3.3339 3.0294 2.9041 2.8646 2.5467

CPU sec. 0.0777 0.1378 0.2069 0.2772 0.7199

gapheuristic 10.1388 8.8790 8.3689 7.2578 6.387

CPU sec. 0.1244 0.2347 0.3509 0.4732 1.1905

gapheuristic 13.8567 11.9717 10.7567 9.8005 7.4059

CPU sec. 0.1295 0.2319 0.3389 0.4519 1.1270

gapheuristic 11.9798 10.7959 10.5647 10.2283 9.5651

CPU sec. 0.1351 0.2523 0.3616 0.4819 1.1981

gapheuristic 20.2535 18.3686 17.3339 16.9108 15.0234

CPU sec. 0.2878 0.5505 0.8155 1.0724 2.6814

gapheuristic 23.7647 21.5896 19.8398 18.5890 17.2503

CPU sec. 0.2190 0.4258 0.6421 0.8611 2.1797

gapheuristic 12.7047 11.4197 11.1577 10.3987 8.0189

CPU sec. 0.0991 0.1727 0.2578 0.3506 0.8141

gapheuristic 19.8962 18.3908 17.4317 16.9602 15.8866

CPU sec. 0.1428 0.2614 0.3912 0.5238 1.3266

gapheuristic 28.4607 25.7509 25.1703 23.7537 20.6741

CPU sec. 0.1707 0.3268 0.4951 0.6703 1.6151

Table F.47: Results for the MLCLSPLe per product and process structure

and demand profile

Chapter F. Computational Study 189

Setup Carryover

contained in table F.48.

scenario 1 2 3 4 5

] feasible solutions 1912 1913 1912 1913 1914

% feasible solutions 99.5833 99.6354 99.5833 99.6354 99.6875

Tables F.49 to F.53 give the solution quality and computation time for the

1912 test instances, which could be solved by all scenarios.

scenario 1 2 3 4 5

gapCPLEX 2.3402 2.2594 2.2221 2.2081 2.1459

gapheuristic 13.0406 11.949 11.4025 11.0337 10.0337

CPU sec. 1.8696 3.2652 4.5844 5.8660 13.0627

Table F.49 shows that the deviations of the optimal solutions provided by the

heuristic from the best lower bounds computed by CPLEX and by the heuristic

ameliorate with the maximum number of iterations also for the MLCLSPLCC .

However, the effect is less important than for the MLCLSPLe .

For the best results, the average deviation from the lower bound provided

by CPLEX is 2.1459% of the latter. This implies that the results for the

MLCLSPLCC are better than the ones for the MLCLSPLe in two ways. First,

better solutions are incurred due to the mere fact that consecutive setup car-

ryovers are allowed.27 Second, the solution provided by the heuristic is closer

27 See table D.4 on page 86.

190 Multi-Level Capacitated Lotsizing with Setup Carryover

to the exact optimal solution for the MLCLSPLCC than for the MLCLSPLe .

The average deviation from the lower bound provided by the heuristic is

10.0337% of the latter for lmax = 1200. Hence, the lower bounds pro-

vided by the heuristic are slightly inferior for the MLCLSPLCC than for the

MLCLSPLe .

Again, the average computation time in CPU seconds increases with higher

values for lmax . For lmax = 1200, 13.0627 CPU seconds on average are re-

quired.

scenario 1 2 3 4 5

CPLEX

class 1 gap 1.7782 1.7359 1.712 1.7167 1.6856

gapheuristic 9.5023 8.6721 8.2626 8.0292 7.4599

CPU sec. 0.0158 0.0247 0.0373 0.0430 0.1175

gapheuristic 17.0670 15.3376 14.3841 13.7955 12.1950

CPU sec. 0.0836 0.1493 0.2060 0.2710 0.6259

gapheuristic 10.8220 10.4416 10.1899 10.0001 9.4371

CPU sec. 0.7126 1.2750 1.7838 2.2912 5.0047

gapheuristic 19.8613 18.1381 17.4702 16.7840 14.9775

CPU sec. 3.3004 6.0584 8.7577 11.4322 27.0004

gapheuristic 9.5353 9.0711 8.8212 8.6825 8.3328

CPU sec. 1.8156 3.0277 4.0967 5.0827 10.2357

gapheuristic 10.9727 9.9226 9.4471 9.1542 8.2076

CPU sec. 9.1226 15.7478 22.0211 28.0957 62.1165

Table F.50 shows that also for the MLCLSPLCC the solution quality deterio-

rates, when the number of periods increases relative to the number of items.

Again, the heuristic performs better for classes 5 and 6 than for classes 3 and

4. Overall, the solution quality for the different classes deviates less than it is

Chapter F. Computational Study 191

scenario 1 2 3 4 5

gapheuristic 8.5926 8.2737 8.0378 7.9091 7.6330

CPU sec. 0.4697 0.6155 0.7604 0.9130 1.8037

gapheuristic 8.5179 8.1441 7.9265 7.8047 7.5535

CPU sec. 0.4811 0.6106 0.7646 0.9146 1.8123

gapheuristic 13.2869 12.4488 11.951 11.6340 10.8696

CPU sec. 0.8298 1.2481 1.6342 2.0151 4.0987

gapheuristic 10.3759 9.7108 9.3984 9.1557 8.6111

CPU sec. 0.8861 1.3050 1.6793 2.0725 3.9273

gapheuristic 21.5707 19.1089 18.2113 17.2177 14.8787

CPU sec. 3.9557 7.3944 10.7684 14.1459 33.9950

gapheuristic 15.7813 14.2323 13.3844 12.9775 11.6028

CPU sec. 3.5965 6.6378 9.4844 12.3308 29.0833

gapheuristic 12.1237 11.0385 10.4039 10.1559 9.2797

CPU sec. 2.2498 3.9618 5.5613 7.0979 15.5643

gapheuristic 11.3558 10.2420 9.7110 9.4181 8.7505

CPU sec. 2.3169 4.0524 5.6002 7.0708 15.3910

gapheuristic 16.6424 15.1030 14.2674 13.6219 11.5979

CPU sec. 1.9300 3.3395 4.6984 5.8898 12.0262

gapheuristic 12.2911 11.2999 10.838 10.5370 9.6264

CPU sec. 2.0064 3.5374 4.9678 6.3070 13.1407

Table F.51: Results for the MLCLSPLCC per capacity and setup time profile

Table F.51 shows that here the results are better for instances with low capacity

utilization profiles. The best results were computed for capacity profile 1,

192 Multi-Level Capacitated Lotsizing with Setup Carryover

where the capacity utilization of the resources is 50%. Again, the influence of

the setup time profiles is neither important nor systematic.

scenario 1 2 3 4 5

gapheuristic 15.0342 13.7959 13.1931 12.7142 11.3991

CPU sec. 2.3867 4.2714 6.0143 7.7021 16.9994

gapheuristic 11.2699 10.2264 9.6743 9.3424 8.4580

CPU sec. 2.5670 4.3797 6.1336 7.8390 17.5955

gapheuristic 16.1041 14.8262 14.2685 13.9231 13.1380

CPU sec. 0.0715 0.1276 0.1758 0.2365 0.5480

gapheuristic 9.2961 8.6861 8.3357 8.1620 7.5441

CPU sec. 0.0453 0.0764 0.1056 0.1320 0.3124

Table F.52: Results for the MLCLSPLCC per setup and holding cost ratio

Table F.52 shows for TBO profiles 3 and 4 that the heuristic performs signif-

icantly better, when the setup and holding cost ratio differs for various items

than when they are equal for all items. For profiles 1 and 2 the results for the

MLCLSPLCC differ from those for the MLCLSPLe . However, as the results

here are better in general this effect may be due to TBO profile 1 meaning

TBO = 1 and TBO profile 2 meaning TBO = 2 for classes 1 and 2.

Table F.53 shows that the results are again inferior for general product struc-

tures than for assembly ones. Also, the best results are generated for demand

profile 1. Overall, the heuristic performs comparably well for cyclic and for

acyclic process structures.

Chapter F. Computational Study 193

scenario 1 2 3 4 5

gapheuristic 3.2018 3.1330 3.0849 3.0457 2.9795

CPU sec. 1.2506 1.9096 2.5110 3.0534 5.9262

gapheuristic 8.3028 7.6797 7.3294 7.1307 6.6188

CPU sec. 1.5199 2.5910 3.6449 4.6738 10.4449

gapheuristic 13.2781 11.8247 10.8301 10.1443 8.3244

CPU sec. 1.8996 3.2526 4.5487 5.7826 13.2365

gapheuristic 3.5900 3.4739 3.4388 3.4266 3.3701

CPU sec. 1.0894 1.6005 2.0546 2.4707 4.7584

gapheuristic 9.1652 8.6475 8.3160 8.0913 7.2777

CPU sec. 1.5333 2.4874 3.4166 4.3482 9.8267

gapheuristic 14.5789 13.0075 12.0920 11.5932 9.5557

CPU sec. 1.5294 2.4729 3.3857 4.2930 9.8351

gapheuristic 7.3489 6.6996 6.3460 6.1436 5.8204

CPU sec. 2.0499 3.5181 4.7506 5.8999 11.3684

gapheuristic 18.0765 15.6358 14.5870 13.9870 12.5401

CPU sec. 4.2395 7.9700 11.4875 14.9294 33.6112

gapheuristic 18.6133 16.7020 15.8064 15.1950 13.6440

CPU sec. 3.1205 6.0588 8.9712 11.9154 29.2312

gapheuristic 15.3115 14.5770 14.2795 14.0954 13.6537

CPU sec. 1.0568 1.7295 2.3481 2.8998 5.6132

gapheuristic 22.1288 20.8624 20.4171 20.0212 18.9981

CPU sec. 1.7564 3.0470 4.2039 5.2815 11.0205

gapheuristic 23.2781 21.5015 20.6419 19.8491 17.8949

CPU sec. 1.3715 2.5166 3.6544 4.803 11.8362

Table F.53: Results for the MLCLSPLCC per product and process structure

and demand profile

194 Multi-Level Capacitated Lotsizing with Setup Carryover

The results for the MLCLSPLP M with single setup carryover will be presented

in section F.3.3.1. Thereafter, the results for the MLCLSPLP M CC with con-

secutive setup carryover will be presented in section F.3.3.2.

in table F.54.

PM1 PM2

scenario 1 2 1 2

] feasible solutions 178 178 178 177

% feasible solutions 98.8889 98.8889 98.8889 98.3333

Tables F.55 to F.58 contain the solution quality and computation time for the

178 test instances, which could be solved by all scenarios for PM1 and the 177

test instances, which could be solved by all scenarios for PM2.

PM1 PM2

scenario 1 2 1 2

gapCPLEX 12.2873 12.5241 19.6529 19.6938

gapheuristic 155.8932 156.3991 170.7263 170.9209

CPU sec. 0.4118 0.666 0.3045 0.4724

Table F.55 shows that the deviations of the optimal solutions provided by the

heuristic from the best lower bounds computed by CPLEX even deteriorate

Chapter F. Computational Study 195

from lmax = 70 to lmax = 120. For lmax = 70, the average deviation from

the lower bound provided by CPLEX is 12.2873% for PM1 and 19.6529% for

PM2. Again, the basis is the lower bound of CPLEX.

PM1 PM2

scenario 1 2 1 2

gapheuristic 137.1999 138.5020 151.5586 151.5294

CPU sec. 0.0295 0.0442 0.0245 0.0353

gapheuristic 195.6645 196.2222 204.1324 204.7429

CPU sec. 0.5364 0.9193 0.3833 0.5893

gapheuristic 134.0883 133.7170 155.7386 155.7309

CPU sec. 0.6783 1.0474 0.5164 0.8094

Also, the average deviation of the results from the lower bounds provided by

the heuristic deteriorates with higher maximum numbers of iterations. Due to

the heuristic model applied here, the lower bounds provided by the heuristic

are very poor. The gapheuristic is 155.8932 for PM1 and 170.7263 for PM2,

when lmax is set to 70.

PM1 and 0.3045 for PM2.

As table F.56 shows, for the MLCLSPLP M the solution quality is best for

class 2 and worst for class 4. This corresponds to the findings for the problems

without parallel machines.

196 Multi-Level Capacitated Lotsizing with Setup Carryover

PM1 PM2

scenario 1 2 1 2

gapheuristic 128.9103 128.9001 125.2153 124.8244

CPU sec. 0.0564 0.0733 0.0539 0.0698

gapheuristic 135.6981 136.0142 146.4494 146.6906

CPU sec. 0.1285 0.1991 0.1014 0.1697

gapheuristic 204.8347 206.3645 228.5725 229.4054

CPU sec. 1.4071 2.2758 0.7161 0.9103

gapheuristic 156.581 156.3743 183.3609 182.859

CPU sec. 0.3108 0.552 0.2928 0.4744

gapheuristic 156.1608 157.1183 173.598 174.4058

CPU sec. 0.2114 0.3194 0.381 0.7621

Table F.57 shows that the best results were achieved for a capacity utilization

of 70%.

As table F.58 shows, the results are generally better for cyclic than for acyclic

process structures. The best results were computed for demand profile 1, i. e.

a coefficient of variation of 0.2. Overall, the heuristic performs similarly well

for assembly and general product structures.

Chapter F. Computational Study 197

PM1 PM2

scenario 1 2 1 2

gapheuristic 125.3703 125.4144 141.2038 140.2517

CPU sec. 0.2407 0.274 0.1408 0.1708

gapheuristic 142.8128 142.6344 153.7664 155.3708

CPU sec. 0.5136 0.9 0.5562 0.8803

gapheuristic 170.313 170.995 179.0855 178.8954

CPU sec. 0.6239 0.9322 0.251 0.4195

gapheuristic 123.6882 122.6257 123.757 124.1578

CPU sec. 0.1407 0.2124 0.0738 0.1197

gapheuristic 131.6212 132.743 136.3008 135.7699

CPU sec. 0.5596 0.8634 0.3458 0.5262

gapheuristic 153.267 153.1216 163.3441 161.5849

CPU sec. 0.4311 0.7 0.203 0.2938

gapheuristic 134.376 137.0806 151.0279 150.9412

CPU sec. 0.0842 0.1344 0.1354 0.1749

gapheuristic 204.507 204.7252 228.1956 225.6701

CPU sec. 0.7634 1.1938 0.8769 0.8947

gapheuristic 208.9916 209.4996 225.3671 230.8918

CPU sec. 0.5859 1.1306 0.489 1.1951

gapheuristic 123.8548 125.4706 154.0371 154.2448

CPU sec. 0.0697 0.0989 0.052 0.0728

gapheuristic 169.0137 169.3972 192.7239 193.8236

CPU sec. 0.3823 0.6188 0.2417 0.3302

gapheuristic 188.6245 188.7806 208.5986 208.4507

CPU sec. 0.5682 0.9865 0.3005 0.6646

Table F.58: Results for the MLCLSPLP M per product and process structure

and demand profile

198 Multi-Level Capacitated Lotsizing with Setup Carryover

given in table F.59.

PM1 PM2

scenario 1 2 1 2

] feasible solutions 178 179 176 177

% feasible solutions 98.8889 99.4444 97.7778 98.3333

As described in table D.5 on page 89, not all instances for the MLCLSPLP M CC

could be solved by CPLEX within the given time limit. Hence, they were

excluded from the evaluation of the solution quality and the computation time

presented in tables F.60 to F.63. Overall, 169 test instances for PM1 and 134

test instances for PM2 are considered.

PM1 PM2

scenario 1 2 1 2

gapCPLEX 20.2278 20.2417 51.8062 51.8499

gapheuristic 3835.5764 3871.4291 1189.2057 1190.3529

CPU sec. 1.0281 1.8910 0.7206 1.2373

Again, the best results were generated for lmax = 70 (see table F.60). Here,

the average deviation from the lower bound provided by CPLEX is 20.2278%

for PM1 and 51.8062% for PM2. When evaluating the performance of the

heuristic, remember that the lower bounds computed by CPLEX for the

MLCLSPLP M CC are poorer than for the other problems. As given in ta-

Chapter F. Computational Study 199

ble D.5 on page 89, the relative mipgap28 for all instances is 11.9340. Still, in

comparison to the other model formulations the results are poorer due to the

complexity of the model. This is especially true for PM2.

The average gapheuristic is 3835.5764 for PM1 and 1189.2057 for PM2, when

lmax is set to 70. Thus, this measure is not adapt for the evaluation of the

performance of the heuristic.

The average computation time in CPU seconds is 1.0281 for PM1 and 0.7206

for PM2.

PM1 PM2

scenario 1 2 1 2

CPLEX

class 2 gap 25.1934 25.3150 45.3636 46.1045

gapheuristic 2290.5867 2307.8341 800.0819 804.9221

CPU sec. 0.0543 0.0747 0.0551 0.1050

gapheuristic 7261.4046 7350.9037 1859.0146 1857.6327

CPU sec. 1.8611 3.4824 1.5018 2.5837

gapheuristic 1809.1971 1807.1087 316.7537 313.3473

CPU sec. 1.2000 2.1703 0.4503 0.7197

Surprisingly, for the MLCLSPLP M CC the solution quality is best for class 6

(see table F.61). The reason could be that we generated the test sample by

adapting the resource capacities such that the available capacity of the resource

type remained roughly the same.29 Hence, the resource capacities are smaller

relative to the setup times, which have not been changed.

29 See section C.7.

200 Multi-Level Capacitated Lotsizing with Setup Carryover

PM1 PM2

scenario 1 2 1 2

gapheuristic 1782.6936 1780.5312 400.6897 397.923

CPU sec. 0.2187 0.3143 0.205 0.3139

gapheuristic 2907.9077 3039.8772 729.4522 732.8126

CPU sec. 0.2921 0.4762 0.2494 0.4059

gapheuristic 6934.6087 6973.1903 2119.3302 2118.3252

CPU sec. 3.1232 5.8048 2.0954 2.8806

gapheuristic 4140.7352 4125.2299 1651.8378 1654.9885

CPU sec. 1.2725 2.1073 0.9107 1.7247

gapheuristic 3931.0955 3957.6725 1637.3254 1641.294

CPU sec. 0.5706 1.3907 0.6686 1.6981

Table F.62 shows that the best results were achieved for a capacity utilization

of 50%.

As depicted in table F.63, here the results are generally better for acyclic

than for cyclic process structures. The best results were again generated for

demand profile 1. Finally, the heuristic performed better for general product

structures.

Chapter F. Computational Study 201

PM1 PM2

scenario 1 2 1 2

gapheuristic 1888.94 1888.5977 521.1553 523.7871

CPU sec. 0.3875 0.5824 0.3125 0.5132

gapheuristic 2068.2216 2078.8995 964.6491 967.883

CPU sec. 0.5891 0.9889 0.5042 1.0187

gapheuristic 3445.2126 3499.6488 1385.1478 1397.1971

CPU sec. 1.0691 1.797 0.7005 0.8829

gapheuristic 2052.6906 2058.9487 737.0596 737.0359

CPU sec. 0.2615 0.3718 0.1847 0.2785

gapheuristic 2116.0329 2118.787 868.6673 869.8006

CPU sec. 0.9365 1.5523 0.2699 0.3949

gapheuristic 1939.2012 1946.913 790.0722 803.6714

CPU sec. 0.5444 0.8091 0.2343 0.3704

gapheuristic 2421.7686 2403.9384 1008.8838 1007.973

CPU sec. 0.2084 0.3376 0.1648 0.273

gapheuristic 8144.1136 8158.3717 1576.6568 1594.8862

CPU sec. 2.4556 4.7391 1.8281 4.4579

gapheuristic 7513.7897 7518.3956 2383.2161 2391.9095

CPU sec. 2.0904 4.2905 2.2033 3.8782

gapheuristic 2556.9017 2566.8836 680.7309 660.2237

CPU sec. 0.2761 0.376 0.1692 0.2409

gapheuristic 6212.6266 6209.7121 1489.1525 1494.9448

CPU sec. 1.5324 2.4522 0.4687 0.7671

gapheuristic 5774.4507 6154.7353 1882.2037 1856.8552

CPU sec. 2.1107 4.7217 1.5535 1.7893

Table F.63: Results for the MLCLSPLP M CC per product and process

structure and demand profile

Chapter G

In this work, we presented new model formulations for the well-known MLC-

SLP with setup time and positive lead time, which also account for setup

carryovers and parallel machines. We presented different reformulations as

well as a heuristic solution procedure.

solved by CPLEX. As in the case of the standard MLCLSP the SR and the

SPL formulations proved to yield much tighter bounds than the standard I&L

formulations. Despite the solution quality achieved during the tests, exact

solution algorithms are yet unlikely to succeed in solving real life problems.

This is indicated also by the fact that in the case of consecutive setup carryovers

and parallel machines several instances could not be solved within the given

time limit. Having said that, the superiority of the reformulations introduced

here shows that heuristics based on exact solution procedures may be a field

of future work.

ation. An extensive computational study for different parameter combinations

presented in chapter F shows how sensitive the quality of the solutions is to

203

L. Buschkühl, Multi-Level Capacitated Lotsizing with Setup Carryover,

Edition KWV, https://doi.org/10.1007/978-3-658-24034-9_7

204 Multi-Level Capacitated Lotsizing with Setup Carryover

The heuristic was tested on a new test set of 1920 instances for the problems

without parallel machines and 360 instances for the problems with parallel ma-

chines. The instances were generated by systematically varying different prob-

lem aspects.1 Overall, the heuristic performed well with respect to feasibility,

solution quality and solution speed for all model formulations introduced.

heuristic improvement procedure. Still, hybrid heuristics of this kind may well

be worth future work. One further aspect is that metaheuristic algorithms are

more flexible than Lagrangean heuristics. Therefore, they can be used to in-

clude other practically relevant problem aspects. These include rolling horizon

settings and interactive production planning.

1 See chapter C.

Bibliography

the capacitated lot-sizing problem: the interaction of model formulation and

solution algorithm. International Journal of Production Research 40 (2),

441–458.

flow model for the capacitated lot-sizing problem. Omega 27 (2), 275–284.

item scheduling. IIE Transactions 15 (2), 136–141.

and resource requirements: A review. Operations Research 35 (3), 329–345.

gradient optimisation. European Journal of Operational Research 117 (1),

136–144.

Barany, I., T. J. van Roy, and L. A. Wolsey (1984). Strong formulations for

multi-item capacitated lot sizing. Management Science 30 (10), 1255–1261.

performance of simulated annealing: the case of the multi-level capacitated

lot sizing problem. Computers & and Operations Research 27, 895–903.

L. Buschkühl, Multi-Level Capacitated Lotsizing with Setup Carryover,

Edition KWV, https://doi.org/10.1007/978-3-658-24034-9

206 Multi-Level Capacitated Lotsizing with Setup Carryover

dient optimization. European Journal of Operational Research 7, 380–388.

ern Heuristic Techniques for Combinatorial Problems, pp. 243–303. London:

McGraw-Hill.

system for lot-sizing problems. Management Science 46 (5), 724–738.

as mixed-integer programs. Management Science 47 (7), 993–1007.

approaches for the multilevel resource-constrained lot-sizing problem with

setup and lead times. Asia-Pacific Journal of Operational Research 22 (2),

261–286.

stage capacitated lot-sizing problem. International Journal of Production

Economics 87 (1), 67–81.

gramming approaches to capacity-constrained MRP systems: review, for-

mulation and problem reduction. Management Science 29 (10), 1126–1141.

multilevel lot-sizing with a bottleneck. Management Science 32 (8), 989–

1006.

Bitran, G. R. and H. Matsuo (1986). The multi-item capacitated lot size prob-

lem: error bounds of Manne’s formulations. Management Science 32 (3),

350–359.

capacitated lot size problem. Management Science 28 (10), 1174–1186.

Bibliography 207

planning in multi-stage assembly processes. European Journal of Operational

Research 16, 84–93.

nomic lot sizing and scheduling problem with dynamic demand. Interna-

tional Journal of Production Research 43 (13), 2809–2828.

han (2001). On a tactical/operational production planning problem in sup-

ply chain management: Balancing inventory and setup costs. In MIC’2001

- 4th Metaheuristics International Conference, pp. 569–571.

item lot-sizing problems with time windows. Operations Research 54 (5),

951–967.

item lot sizing problems. European Journal of Operational Research 168 (1),

1–16.

Briskorn, D. (2006). A note on capacitated lot sizing with setup carry over.

IIE Transactions 38 (11), 1045–1047.

lot sizing with dynamic demands. Management Science 37 (4), 409–427.

Cattrysse, D., J. Maes, and L. N. van Wassenhove (1990). Set partitioning and

column generation heuristics for capacitated dynamic lotsizing. European

Journal of Operational Research 46 (1), 38–47.

chain planning with order/setup costs and capacity constraints. Journal of

Systems Science and Systems Engineering 12 (1), 98–110.

208 Multi-Level Capacitated Lotsizing with Setup Carryover

Chen, W.-H. and J.-M. Thizy (1990). Analysis of relaxations for the multi-item

capacitated lot-sizing problem. Annals of Operations Research 26, 29–72.

capacitated multi-stage lot-sizing problem with lead times. Journal of the

Operational Research Society 46, 1208–1222.

tion. Symposia Mathematica 19, 357–372.

grams. Operations Research 8 (1), 101–111.

Branch-and-Price algorithm for the Capacitated Lot Sizing Problem with set

up times. Erasmus Research Institute of Management (ERIM).

algorithm. IBM Systems Journal 7 (1), 30–38.

beschränkungen. Produktion und Logistik. Heidelberg: Physica-Verlag.

lot-sizing and scheduling by Lagrangean relaxation. European Journal of

Operational Research 59, 444–458.

relaxation approach for very-large-scale capacitated lot-sizing. Management

Science 38 (9), 1329–1340.

solving a large-scale resource allocation problem in production planning.

In G. Fandel, T. Gulledge, and A. Jones (Eds.), Operations Research in

Production Planning and Control, Berlin, pp. 105–119. Springer.

practical resource allocation for production planning and scheduling with

Bibliography 209

275–286.

multi-item, single-level, limited capacity, lot-sizing problem. Journal of Op-

erations Management 2 (1), 23–39.

lot-sizing problem for multiple items under limited capacity. AIIE Transac-

tions 13 (4), 294–303.

national Journal of Production Economics 40 (1), 73–87.

Drexl, A. and A. Kimms (1997). Lot sizing and scheduling - survey and

extensions. European Journal of Operational Research 99 (2), 221–235.

inventory and labor allocations. Management Science 11 (9), 874–890.

straints. AIIE Transactions 7 (2), 170–176.

sizing problems using variable redefinition. Operations Research 35 (6), 832–

848.

gorithm for dynamic lot-sizing. Journal of Operations Management 5 (2),

229–235.

general dynamic lot sizing models with n periods in o(n log n) or o(n) time.

Management Science 37 (8), 909–925.

programming problems. Management Science 27 (1), 1–18.

210 Multi-Level Capacitated Lotsizing with Setup Carryover

Interfaces 15 (2), 10–21.

ropean Journal of Operational Research 44 (3), 337–348.

planning: Algorithms and complexity. Management Science 26 (7), 669–

679.

heuristic method for lot-sizing in multi-stage systems. Computers & Oper-

ations Research 24 (9), 861–874.

the Theory of NP-Completeness. San Francisco: Freeman and Co.

bound algorithm for the multi item single level capacitated dynamic lotsizing

problem, pp. 92–108. Berlin.

Glover, F. (1986). Future paths for integer programming and links to artificial

intelligence. Computers & Operations Research 13 (5), 533–549.

ryover in the capacitated lotsizing problem. International Journal of Pro-

duction Research 38 (14), 3421–3424.

search heuristic for the capacitated lot-sizing problem with set-up carryover.

Management Science 47 (6), 851–863.

modelling setup carryover in the capacitated lot sizing problem. Interna-

tional Journal of Production Research 33 (7), 1973–1988.

Groff, G. (1979). A lot sizing rule for time-phased component demand. Pro-

duction and Inventory Management 20 (1), 47–53.

Bibliography 211

designing Lagrangean dual ascent algorithms. European Journal of Opera-

tional Research 43 (2), 197–205.

Günther, H.-O. (1987). Planning lot sizes and capacity requirements in a single

stage production sytem. European Journal of Operational Research 31 (2),

223–231.

problem with sequence-dependent setup costs and setup times. Computers

& Operations Research 32 (4), 727–747.

International Journal of Operations & Production Management 10 (9), 57–

73.

capacitated lot sizing decisions. In Computer Research Conference.

Notes in Economics and Mathematical Systems. Berlin: Springer.

adjacent periods, pp. 127–146. Berlin.

to Optimality. University Dresden. Working Paper.

reihenfolgeabhängigen rüstvorgängen in einem unternehmen der satz- und

drucktechnik. Zeitschrift für Betriebswirtschaft 66 (12), 1511–1526.

Haase, K. and U. Kohlmorgen (1995). Parallel genetic algorithm for the ca-

pacitated lot-sizing problem. In P. e. a. Kleinschmidt (Ed.), Operations

Research Proceedings 1995, pp. 370–375. Springer.

borhood search. In S. Voß, S. Martello, I. Osman, and C. Roucairol (Eds.),

212 Multi-Level Capacitated Lotsizing with Setup Carryover

tion, Dordrecht, pp. 433–458. Kluwer Academic Publishers.

ciples and applications. European Journal of Operational Research 130,

449–467.

with multiple constrained resources. Management Science 42 (1), 19–36.

Stuttgart: M und P.

systems. OR Spektrum 17, 5–18.

optimization. Mathematical Programming 6, 62–88.

pacitated lot-sizing problem. Computers and Industrial Engineering 28 (4),

709–719.

the Operational Research Society 47, 151–161.

the CLSP with set-up times. Journal of the Operational Research Society 54,

490–498.

Column Generation and Lagrangean Relaxation. Erasmus Research Institute

of Management (ERIM).

Hung, Y.-F., C.-P. Chen, C.-C. Shih, and M.-H. Hung (2003). Using tabu

search with ranking candidate list to solve production planning problems

with setups. Computers & Industrial Engineering 45 (4), 615–634.

Bibliography 213

sizing model. Journal of the Operational Research Society 51, 1309–1318.

Hung, Y.-F. and Y.-C. Hu (1998). Solving mixed integer programming produc-

tion planning problems with setups by shadow price information. Computers

& Operations Research 25 (12), 1027–1042.

Hung, Y.-F., C.-C. Shih, and C.-P. Chen (1999). Evolutionary algorithms for

production planning problems with setup decisions. Journal of the Opera-

tional Research Society 50, 857–866.

review and comparison of solution approaches. European Journal of Opera-

tional Research 177, 1855–1875.

sizing problem: a review of models and algorithms. Omega 31, 365–378.

cycling problem. Operations Research 33 (2), 326–345.

Karni, R. and Y. Roll (1982). A heuristic algorithm for the multi-item lot

sizing problem with capacity constraints. IIE Transactions 14 (4), 249–256.

assembly systems with setup costs, setup times, and multiple constrained

resources. Management Science 44 (6), 859–877.

Logistics. Heidelberg: Physica-Verlag.

Kirca, Ö. and M. Kökten (1994). A new heuristic approach for the multi-item

dynamic lot sizing problem. European Journal of Operational Research 75,

332–341.

simulated annealing. Science 220 (4598), 671–680.

214 Multi-Level Capacitated Lotsizing with Setup Carryover

grained parallel genetic algorithms. Annals of Operations Research 90, 203–

219.

Krarup, J. and O. Bilde (1977). Plant Location, Set Covering and Economic

Lot Size: An O(mn)-Algorithm for structured Problems, pp. 155–180. Basel,

Stuttgart: L. Collatz and G. Meinardus and W. Wetterling.

structure and models. European Journal of Operational Research 75, 243–

263.

programming, simulated annealing and tabu search heuristics for lotsizing

in bottleneck assembly systems. IIE Transactions 25 (1), 62–72.

single operation, multi-item loading problem. AIIE Transactions 11 (4),

319–326.

scheduling. Operations Research 19 (4), 946–969.

itated lotsizing complexity and LP-based heuristics. European Journal of

Operational Research 53, 131–148.

dynamic lot-sizing heuristics: A general review. Journal of the Operational

Research Society 39 (11), 991–1004.

Maes, J. and L. N. van Wassenhove (1986a). Multi item single level capacitated

dynamic lotsizing heuristics: A computational comparison (part i: Static

case). IIE Transactions 18, 114–123.

Maes, J. and L. N. van Wassenhove (1986b). Multi item single level capacitated

dynamic lotsizing heuristics: A computational comparison (part ii: Rolling

horizon). IIE Transactions 18, 124–129.

Bibliography 215

Maes, J. and L. N. van Wassenhove (1986c). A simple heuristic for the multi

item single level capacitated lotsizing problem. Operations Research Let-

ters 4 (6), 265–273.

ence 4 (2), 115–135.

dimensionally equidistributed uniform pseudorandom number generator.

ACM Transactions on Modeling and Computer Simulation 8 (1), 3–30.

problems. International Journal of Production Economics 85, 97–111.

tinuierliche Produktionslnien. Gabler Edition Wissenschaft: Produktion und

Logistik. Wiesbaden: Gabler.

multi-item lot-sizing problem with backordering. International Journal of

Production Economics 34, 1–15.

Multi-Item Capacitated Lot-Sizing Problems with Setup Times by Branch-

and-Cut. Université Catholique de Louvain, Belgium. Core Discussion Paper

00/39.

production control problems in multi-stage production control problems in

multi-stage production inventory systems. Ph. D. thesis, Indian Institute of

Technology, Bombay.

with fixed resources. Management Science 21 (10), 1186–1193.

216 Multi-Level Capacitated Lotsizing with Setup Carryover

with variable resources. Management Science 21 (10), 1194–1203.

stage capacitated lot sizing and loading problem. Journal of the Operational

Research Society 50, 810–825.

relaxation-simulated annealing approach to the multi-level multi-item ca-

pacitated lot sizing problem. International Journal of Production Eco-

nomics 68 (3), 319–331.

with overtime decisions and setup times. IIE Transactions 32, 1043–1057.

Özdamar, L., Şevket Ilker Birbil, and M.-C. Portmann (2002). Technical note:

New results for the capacitated lot sizing problem with overtime decisions

and setup times. Production Planning & Control 13 (1), 2–10.

exact and population-based methods for the constrained multilevel lot sizing

problem. International Journal of Production Research 44 (22), 4755–4771.

using strong cutting planes. Management Science 37 (1), 53–67.

Mathematics and Mathematical Physics 9, 14–29.

Quadt, D. (2004). Lot-Sizing and Scheduling for Flexible Flow Lines. Lecture

Notes in Economics and Mathematical Systems. Springer.

scheduling of flexible flow lines. International Journal of Production Re-

search 43 (11), 2291–2308.

Lieferantenauswahl. Norderstedt: Books on Demand.

Bibliography 217

119–131. Berlin.

ning, Volume 355 of Lecture Notes in Economics and Mathematical Systems.

Berlin, Heidelberg, New York, Bunkyo-ku: Springer.

extensions of the discrete lotsizing and scheduling problem. Management

Science 37 (7), 801–812.

methods for lotsizing problems. Annals of Operations Research 41, 453–468.

multi-plant, multi-item, multi-period capacitated lot-sizing problems. Asia-

Pacific Journal of Operational Research 19, 87–105.

plant, multi-item, multi-period capacitated lot-sizing problems with inter-

plant transfers. Computers & Operations Research 32 (3), 537–555.

lot-sizing heuristic under capacitated single stage production. European

Journal of Operational Research 41, 181–185.

Silver, E. A. and H. C. Meal (1969). A simple modification of the EOQ for the

case of a varying demand rate. Production and Inventory Management 10,

51–55.

Silver, E. A. and H. C. Meal (1973). A heuristic for selecting lot size quantities

for the case of a deterministic time-varying demand rate and discrete oppor-

tunities for replenishment. Production and Inventory Management 14 (2),

64–74.

Sox, C. R. and Y. Gao (1999). The capacitated lot sizing problem with setup

carry-over. IIE Transactions 31, 173–181.

218 Multi-Level Capacitated Lotsizing with Setup Carryover

namic multi-item multi-level capacitated lotsizing. European Journal of Op-

erational Research 94, 561–581.

multi-item multi-level capacitated lotsizing. OR Spektrum 19 (2), 87–96.

Stadtler, H. (2003). Multilevel lot sizing with setup times and multiple con-

strained resources: Internally rolling schedules with lot-sizing windows. Op-

erations Research 51 (3), 487–502.

ing models. In 23rd Annual Symposium of the Brazilian Operational Re-

search Society (SOBRAPO), Campos do Jordao SP, Brazil, pp. 938–947.

International Journal of Production Research 43 (1), 49–66.

Lecture Notes in Economics and Mathematical Systems. Berlin: Springer.

linked lot sizes. Management Science 49 (8), 1039–1054.

Springer.

Losgrößenplanung bei beschränkten Ressourcen und genereller Erzeugnis-

struktur. OR Spektrum 15, 63–73.

dynamic multilevel multiitem constrained lotsizing with setup times. Man-

agement Science 42 (5), 738–757.

multi-level capacitated lotsizing for general product structures. European

Journal of Operational Research 75 (2), 296–311.

Bibliography 219

Thizy, J.-M. (1991). Analysis of the Lagrangean decomposition for the multi-

item capacitated lot-sizing problem. INFOR 29 (4), 271–283.

Thizy, J.-M. and L. N. van Wassenhove (1985). Lagrangean relaxation for the

multi-item capacitated lot-sizing problem: A heuristic implementation. IIE

Transactions 17, 308–313.

problem. IIE Transactions 19 (3), 67–72.

Trigeiro, W. W. (1989). A simple heuristic for lot sizing with setup times.

Decision Sciences 20 (2), 294–303.

sizing with setup times. Management Science 35, 353–366.

and scheduling under capacity constraints. European Journal of Operational

Research 2 (1), 36–41.

Wagelmans, A., S. van Hoesel, and A. Kolen (1992). Economic lot sizing:

An o(n log n) algorithm that runs in linear time in the wagner-whitin case.

Operations Research 40 (1), 145–156.

lot size model. Management Science 5 (1), 89–96.

of Operational Research 86, 395–401.

Xie, J. and J. Dong (2002). Heuristic genetic algorithms for general capacitated

lot-sizing problems. Computers & Mathematics with Applications 44 (1-2),

263–276.

model with backlogging. Management Science 13 (1), 105–119.

- Set TheoryEnviado porKamal Lohia
- Operation ResearchEnviado porsurajrohera
- An Introduction to Optimization Chong Solution Manual PDFEnviado porBhanu Prakash
- Reference Material I MEE437 or T Module 2 WS2017 18 Tutorial ProblemEnviado porMustafa Jawad wala
- MATH-162Enviado porSurafel L. Tilahun
- A Survey of Modern Data Classification TechniquesEnviado porInternational Journal for Scientific Research and Development - IJSRD
- Model Predictive ControlEnviado porzohaib319
- Centrifugal PumpEnviado porMuhamed Badawi
- Vector Expansion TechniqueEnviado porGuilherme Gomes
- Undergraduate 14[1]Enviado porThakur Aryan Singh
- DSP-sharif universityEnviado porsahar
- Imperialist Competitive AlgorithmEnviado porMochamad Avid Fassamsi
- Robust design optimisation .pdfEnviado porbiswajit roy
- Literature Survey for Face Detection under Illumination VariationEnviado porijsret
- Integrated Framework of Optimization Technique and Information Theory Measures for Modeling Neutrosophic VariablesEnviado porMia Amalia
- KOMAT---ARRAY & LINSPACE.docxEnviado porhikmah1998
- iberamia02pp69.pdfEnviado porPrashanth Vaidya
- 1.pdfEnviado porDilruk Gallage
- threedetailedexamplesofsystemsofequationsEnviado porapi-227793309
- notes le3Enviado porapi-373634605
- Stress MethodEnviado porGautham Shankar
- Heuristic Optimization Applied on Dynamic Matrix Control and Cascade PI Control Tunning for Speed Control of DC MotorEnviado porHanzo Hasashi
- Lamp IranEnviado por'ahmad' Faiz 'Lewer'
- Backtracking ProgramsEnviado porSudhakar Singh
- TQC TUGASEnviado porArga Saputra
- Ch12 The Art of Modelling with SpreadsheetEnviado porrajkzo
- Tomic_2015.pdfEnviado porSeda Toy
- Exam April 2014Enviado porzaur
- Jia 2015Enviado porVitthal Patnecha
- Jouve_Lecture1_2.pdfEnviado porNirmalya Mishra

- Enabling-the-next-production-revolution-the-future-of-manufacturing-and-services-interim-report.pdfEnviado porGerson Schaffer
- lab-glpk-enEnviado porGerson Schaffer
- Or Executive GuideEnviado porGerson Schaffer
- Lectura01Enviado porqwert132i
- Industrie4.0 Smart Manufacturing for the Future EnEnviado porgorfnesh
- TaxaSelic_Diaria_20040101_20071231Enviado porGerson Schaffer
- Manual Ade4Enviado porGerson Schaffer
- Cashflow CaseEnviado porGerson Schaffer
- OreBlend.xlsEnviado porGerson Schaffer
- ExamplesEnviado porGerson Schaffer
- Chap3DataEnviado porGerson Schaffer
- Trans ExEnviado porGerson Schaffer
- RubiaEnviado porCLEM_REIS
- Patent Mining - A SurveyEnviado porGerson Schaffer
- Book Sustainabily Updated 3a Final06072012Enviado porGerson Schaffer
- AlgDesign.pdfEnviado porGerson Schaffer
- Lingo 14Enviado porGerson Schaffer

- Petronasproceduresandguidelinesforfinance 150630052159 Lva1 App6891Enviado porslamet4riadi
- 3533616 Financial MathematicsEnviado porSAMRAT PAUL
- Jack Ma ProfileEnviado porHeena Dua
- Sepp-AKEnviado pornieschopwit
- Comparitive Study of Reliance Mutual Funds and Uti Mutual FundsEnviado porRichard Gaurab
- SAPM PPTEnviado porVijay Deshpande
- (2014 digital library._ Supply and operations management collection) Olson, David Louis-Supply chain risk management _ tools for analysis-Business Expert Press (2014).pdfEnviado porUsama Waly
- Introduction to Business Processes, BPM, and BPM SystemsEnviado porMiguel Angel
- project report on Mutual fundEnviado porRaghavendra yadav KM
- Business Law 4Enviado porJavid
- resume-ez 1Enviado porapi-314057353
- parle gEnviado porAbhilash Nox
- UT Dallas Syllabus for entp6378.501.11f taught by Joseph Picken (jcp016300)Enviado porUT Dallas Provost's Technology Group
- 005 Municipality of Paoay vs Manaois.docxEnviado porcharmssatell
- MBA Course @ AMU (Full) All SemestersEnviado porSyed Asad
- How Social Enables Brand Resilience and Why It Matters - Complete SetEnviado poriGo2 Group
- Operation strategy KPSEnviado porKARTIKEYA MISRA
- ProntoEnviado porJustin Carder
- Cosumer Behavior Salt BangladeshEnviado porMahmudul Hasan Rabby
- Comprehensive Charterers Liability CoverEnviado porMurat Yilmaz
- Brand Preference MobileEnviado porSandeep Gupta
- Sage Timberline PDFsEnviado porRoxx Cerise
- Pacific Timber vs. CAEnviado porAiza Giltendez
- Follow Up Letter Debt CollectionEnviado porkokoripollo
- Perspectec Survey Uncovers Netflix Subscriber Churn and Penetration in Germany and JapanEnviado porPR.com
- Head+-+Power%2C+Facility+Planning.docxEnviado porJohn Cabrera
- El Cluster Biotecnológico de Munich kEnviado portitov pensez
- BOM Vs BOOEnviado pormanoj2288890
- ER420-130328-5003-CV-ULT - Method Statement of Box Culvert ConstructionEnviado porWr Ar
- Lecture 8 Performance MeasuresEnviado porZaeed Huq