Analysis and Design of Discrete Part Production Line PDF

ANALYSIS AND DESIGN OF DISCRETE
PART PRODUCTION LINES

Springer Optimization and Its Applications
VOLUME 31
Managing Editor
Panos M. Pardalos (University of Florida)
Editor—Combinatorial Optimization
Ding-Zhu Du (University of Texas at Dallas)
Advisory Board
J. Birge (University of Chicago)
C.A. Floudas (Princeton University)
F. Giannessi (University of Pisa)
H.D. Sherali (Virginia Polytechnic and State University)
T. Terlaky (McMaster University)
Y. Ye (Stanford University)
Aims and Scope

Optimization has been expanding in all directions at an astonishing rate
during the last few decades. New algorithmic and theoretical techniques have
been developed, the diffusion into other disciplines has proceeded at a rapid
pace, and our knowledge of all aspects of the field has grown even more
profound. At the same time, one of the most striking trends in optimization
is the constantly increasing emphasis on the interdisciplinary nature of the
field. Optimization has been a basic tool in all areas of applied mathematics,
engineering, medicine, economics and other sciences.
The Springer Optimization and Its Applications series publishes under-
graduate and graduate textbooks, monographs and state-of-the-art exposi-
tory works that focus on algorithms for solving optimization problems and
also study applications involving such problems. Some of the topics covered
include nonlinear optimization (convex and nonconvex), network flow prob-
lems, stochastic optimization, optimal control, discrete optimization, multi-
objective programming, description of software packages, approximation
techniques and heuristic approaches.
For other titles in this series, go to www.springer.com/series/7393

Analysis and Design of Discrete
Part Production Lines
By
CHRISSOLEON T. PAPADOPOULOS
Aristotle University of Thessaloniki, Greece
MICHAEL E. J. O’KELLY
Waterford Institute of Technology, Ireland
MICHAEL J. VIDALIS
University of the Aegean, Greece
DIOMIDIS SPINELLIS
Athens University of Economics and Business, Greece
123
Chrissoleon T. Papadopoulos Michael E. J. O’Kelly
Aristotle University of Thessaloniki Waterford Institute of Technology
Department of Economics Waterford, Ireland
Thessaloniki, Greece jokelly@eircom.net
hpap@econ.auth.gr
Michael J. Vidalis Diomidis Spinellis

University of the Aegean Athens University of Economics and Business
Department of Business Administration Department of Management Science
Chios, Greece and Technology
mvid@ba.aegean.gr Athens, Greece
dds@aueb.gr
ISSN: 1931-6828
ISBN: 978-0-387-89493-5 e-ISBN: 978-0-387-89494-2
DOI 10.1007/978-0-387-89494-2
Springer Dordrecht Heidelberg London New York
Library of Congress Control Number: 2009926035
Mathematics Subject Classification (2000): 90-02, 90B30, 90B25, 90B50, 91B74, 90C90
c Springer Science+Business Media, LLC 2009

All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY
10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection
with any form of information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are
not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject
to proprietary rights.
Cover illustration: Photo taken by Elias Tyligadas
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

Dedicated to those who are so much more
important to us than mere production lines,
but who could have reasons to inquire
‘why then spend so much time with
production lines?’
Maridora, Katerina, Elisavet — Jane, John,
Eamon, Ita, Paul, David, Jenny — Katerina,
Ioannis, Athanasia, Daphni — Eliza,
Dionysis, Eleana
Preface
Initially, during discussions among the four colleagues about this writing project, we
used “on the optimal design of production lines” as the working title of the book.
However, it must be understood that all models involve assumptions and unless these
assumptions are valid, the results could not be described as optimal. So basically,
what this text is offering is a set of best solutions to the models as described in the
various chapters. The models and the algorithms presented are generally accepted
by internationally respected scholars to give very good solutions following extensive
simulation and comparison with actual systems. We, therefore, see the process of the
optimal design of production lines as a complementary activity between the schol-
ars and the practitioners. The scholars provide models and associated algorithms and
the practitioners, in their turn, ensure the appropriateness of the assumptions of the
models used together with the validity of the data used, and hence, in effect there
is a joint responsibility to achieve the optimal or near optimal design of production
lines. It is our experience that practitioners in industry and consultancy companies
often have considerable difficulties with the academic and research papers which
appear in international journals due to the complexity of the mathematical analy-
sis involved and the lack of readily available efficient algorithms for the solution of
the models presented. The literature consists of a large number of excellent papers
and it is extremely difficult for the practitioner to have an opportunity to examine
the appropriateness of each paper to the design problem on hand. We thought that
this project could assist the practitioner in this regard by providing a set of mod-
els which have been found useful in specific situations. Of course, it is not claimed
that these models cover every conceivable situation, but the authors believe that they
provide an extremely useful starting point for the understanding of production lines.
Furthermore, we thought that it would be very useful to have in one place a collec-
tion of relevant analysis and design material of production lines. For this reason, we
decided to put the algorithms on a web site: http://purl.oclc.org/NET/prodline. Here,
we would like to sincerely acknowledge the generosity of many colleagues across
the world who gave us access to the relevant algorithms. Without such generosity
and cooperation our project would have been a total failure.
vii
viii Preface
Production lines in the context of this work are a subset of general manufacturing
systems. There are various types of manufacturing systems such as job shops, flex-
ible manufacturing systems (FMS), flexible assembly systems (FAS), production or
flow lines and automatic transfer lines. The usual features of a production or flow line
are dedicated work-stations, manual or automatic, usually producing a single product
with a fixed routing and an asynchronous movement of material between the work-
stations and a high mean production rate (throughput). Production lines are complex
systems. Full understanding of such systems requires skilled analysis in order to
facilitate the development of a competent design. Some important design problems
associated with production lines consist of decisions in relation to three main issues,
viz., work-load allocation, buffer allocation and server allocation. The objective of
this book is to provide the reader with a set of models and solutions to these prob-
lems (work-load allocation problem (WAP), buffer allocation problem (BAP) and
server allocation problem (SAP)) which are accepted by experienced researhers and
practitioners to be of value in the design of these systems. To assist in the solution of
these design problems, it is necessary to make use of both evaluative and generative
(optimization) algorithms.
During the course of a project like this, a number of changes of perspective and
vision, as time progresses, are inevitable. Accordingly, we decided to change the
working title of the book to “Discrete Part Production Lines.” It is the authors’ view
that the models presented may be used in either of two modes, viz., analysis and
design. For actual existing lines, the models may be used to predict performance
under existing conditions or if certain changes are made, for example, to the number
of buffer slots before a particular station. If a new design is contemplated, then, of
course, a range of models may be used having in mind the objectives of the design
including cost considerations. We hope that the Analysis and Design Decision Net-
work, given in the book’s web site, will assist the readers in choosing appropriate
models for their investigations. Researchers and practitioners alike have sometimes
questioned the usefulness as well as the benefits derived from very detailed and
somewhat complex analysis of production lines. It is, of course, not always feasi-
ble to adopt in practice what may be the theoretical optimal or near optimal solution
to a design problem in production lines, developed from system modeling. However,
if one knows the optimal or near optimal solution, the theoretical prime cost of adopt-
ing a more ‘practical’ solution would be of interest. Clearly, the software associated
with this text would be of assistance in discussions of these matters.
In Chapter 1, “Manufacturing Systems: Types and Modeling,” an overview of
the evolution and classification of manufacturing systems is given as well as an
introduction to models and modeling.
Chapter 2, “Evaluative Models of Discrete Part Production Lines,” describes four
predictive models of performance evaluation of production or flow lines: the Marko-
vian model, the expansion method, the aggregation method and the decomposition
approach applied both to single-machine station and parallel-machine station pro-
duction lines. A short section on simulation modeling is given at the end of this
chapter.
Preface ix
Chapter 3, “The Design of Production Lines,” introduces the reader to the design
problems of production lines and the concept of improvability.
Chapter 4, “Work-Load and Server Allocation Problems,” describes two separate
problems, viz., the work-load allocation problem and the server allocation problem.
Chapter 5, “The Buffer Allocation Problem,” describes this important problem
within the context of production lines.
Chapter 6, “Double and Triple Optimization,” considers the combinations of the
three pure work-load allocation, server allocation and buffer allocation problems,
taken two at a time or all three together.
Chapter 7, “Cost Considerations,” examines cost considerations in the design
of production lines using profit maximization and cost minimization objective func-
tions.
In Appendix A, a review of some mathematical fundamentals is given, mainly
from linear algebra, probability theory, discrete Markov processes (Markov chains)
and queueing theory.
Appendix B contains details concerning the code available on the book’s web
site. For each algorithm we provide its author, its coder, a short description, the
corresponding output, and key bibliographic references.
Appendix C gives the glossary.
The authors are conscious of the debt of gratitude they owe to a very large num-
ber of researchers and practitioners, much too numerous to list, in the area of the
design of production lines and manufacturing systems in general. We believe that we
must make a special mention of those colleagues who participated either as presen-
ters or attendees at the five Hellenic International Conferences on Analysis, Design
and Optimization of Manufacturing Systems which were held in Greece (four at
the Islands of the Aegean Archipelagos at Samos, Tinos, Tinos and Samos, respec-
tively, and one at Zakynthos Island of the Ionian Sea) and at the 30th Computers &
Industrial Engineering International Conference which was held on Tinos Island and
who assisted us so much in crystallizing our understanding of the research work in
this area. As we are reluctant to list any specific colleagues for special acknowledg-
ment, we give in Appendix D a list of all colleagues who participated as presenters
or attendees at the five Hellenic International Conferences on Analysis, Design and
Optimization of Manufacturing Systems.
Appendix E presents an Arena simulation model of a reliable production line.
In conclusion, the authors hope that the background theory, details of the relevant
algorithms, tabulations of actual computer runs and the provision of the algorithms
at the website associated with this text will together form a reservoir of knowledge to
assist the designers of practical production lines. In particular, the authors hope that
the guides to the use of these algorithms given throughout the text and in Appendix B
will assist the busy designers and practitioners in choosing appropriate computational
tools for their analyses. The individual contributions of the authors are given in the
book, but, of course, the composite contributions of many other researchers which
are included and acknowledged in the text far outweigh what any one of the authors
could hope to contribute.
x Preface
Although Dr. Alexandros Diamantidis’s name appears in both Appendix B and

Appendix D, all the authors wish to make a special acknowledgment of his contribu-
tion to our work particularly in relation to the development of the effective evaluative
decomposition algorithm for solving multi-station multi-server production lines and
for running various problems sets at our request.
Needless to say, as any academic will attest, we are individually very much
in debt to our students who over the years have assisted us in advancing our
understanding of the fascinating subject of production lines.
We wish to acknowledge very sincerely the patience of the publisher with the
delay in producing this text caused inter alia by one of the authors being indisposed
for a relatively long period of time.
Finally, the authors would be very pleased to hear from researchers or practition-
ers who wish to have an algorithm/procedure, developed by them, to be considered
for inclusion at the website. No claim is made, at this point in time, that the algo-
rithms presented can handle all possible realistic design problems for either short or
long production lines and it is in that spirit that the authors invite other researchers
to make available their algorithms so that the issues related to the design of pro-
duction lines are finally closed. Hopefully, in time, a very comprehensive set of
algorithms/procedures for the analysis/design of production lines would become
available for all to use. This could well be the first step to having on a website a
set of algorithms/procedures which have been found to be of value in the design and
analysis of general manufacturing systems.
Thessaloniki, 2009 Chrissoleon T. Papadopoulos

Waterford, 2009 Michael E.J. O’Kelly
Chios, 2009 Michael J. Vidalis
Athens, 2009 Diomidis Spinellis
Contents
1 Manufacturing Systems: Types and Modeling . . . . . . . . . . . . . . . . . . . . . 1

1.1 Manufacturing Systems: Evolution and Classification . . . . . . . . . . . . 1
1.2 Models and Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Classification of Manufacturing Systems . . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Models of Manufacturing Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5 Methods of Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.6 Measures of Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.7 Related Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2 Evaluative Models of Discrete Part Production Lines . . . . . . . . . . . . . . . 25

2.1 Markovian Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1.1 A numerical approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.1.2 The algorithm for the generation of the conservative
matrix A for the reliable exponential production lines with
inter-station buffers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.1.3 A simple non-linear flow model . . . . . . . . . . . . . . . . . . . . . . . . 49
2.2 Decomposition Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.3 The Expansion Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.4 The Aggregation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.5 Modeling of Production Lines with Parallel Reliable Machines
at Each Station . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.5.1 Exact solution to a two-station production line with parallel
machines at each station . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.5.2 Alternative exact Markovian analysis of a two-station line
with parallel machines at each station . . . . . . . . . . . . . . . . . . . 70
2.5.3 Approximate methods for large lines . . . . . . . . . . . . . . . . . . . . 73
2.5.4 Derivation of the decomposition equations . . . . . . . . . . . . . . . 75
2.5.5 The decomposition algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 79
2.5.6 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
2.6 Simulation Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
xi
xii Contents
2.7 General Comment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3 The Design of Production Lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.2 Role of the Design Engineer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.3 Improvability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4 Work-Load and Server Allocation Problems . . . . . . . . . . . . . . . . . . . . . . 113

4.1 The Work-Load Allocation Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.1.1 The bowl phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.1.2 Computational issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
4.2 The Server Allocation Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.3 The Simultaneous Work-Load and Server Allocation: The
L-phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.4.1 Bowl phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.4.2 Reversibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5 The Buffer Allocation Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

5.1 Formulation of the Buffer Allocation Problems . . . . . . . . . . . . . . . . . . 131
5.2 Solution of the Buffer Allocation Problems . . . . . . . . . . . . . . . . . . . . . 132
5.3 Solution Approaches to the BAP in Short Lines . . . . . . . . . . . . . . . . . 134
5.4 Solution Approaches to the BAP in Longer Lines . . . . . . . . . . . . . . . . 145
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6 Double and Triple Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

6.1 Simultaneous Allocation of Work and Buffers, W + B . . . . . . . . . . . . 163
6.2 Simultaneous Allocation of Servers and Buffers, S + B . . . . . . . . . . . 164
6.3 Simultaneous Allocation of Work, Servers and Buffers, W + S + B . 165
6.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
7 Cost Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

7.1 Cost Models: Profit Maximization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.2 Cost Models: Cost Minimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
Contents xiii
A Mathematical Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

A.1 Vectors and Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
A.1.1 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
A.1.2 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
A.2 Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
A.2.1 Bernoulli trials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
A.2.2 Memoryless property of the exponential distribution . . . . . . . 209
A.2.3 Relationship between the exponential distribution and the
Poisson distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
A.2.4 The Coxian distribution with two phases . . . . . . . . . . . . . . . . . 211
A.2.5 Phase-type distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
A.3 Discrete Markov Processes (Markov Chains) . . . . . . . . . . . . . . . . . . . 216
A.4 Data Plotting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
A.5 Well-Known Results of Queueing Theory . . . . . . . . . . . . . . . . . . . . . . 220
A.5.1 M/M/1: First-Come First-Served (FCFS)/∞/∞ queue . . . . . 221
A.5.2 M/M/1: FCFS/N/∞ queue . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
A.5.3 M/M/c: FCFS/∞/∞ queue . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
A.5.4 M/M/c: FCFS/N/∞ queue . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
A.5.5 M/M/c: FCFS/c/∞ queue . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
A.5.6 M/M/∞ queue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
A.5.7 M/M/c: FCFS/K/K–The finite source queue . . . . . . . . . . . . 224
A.5.8 Queueing networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
B Algorithms/Procedures Details and Guide to Use . . . . . . . . . . . . . . . . . . 233

B.1 Markovian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
B.2 Decomposition-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
B.3 Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
B.4 Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
B.5 Decomposition-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
B.6 Two-Level Work-Load Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
B.7 Simulated Annealing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
B.8 Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
B.9 Complete Enumeration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
B.10 Buffer Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
C Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
C.1 General Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
C.2 Production Lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
C.3 Decomposition Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
C.4 Markovian Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
C.5 Expansion Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
C.6 Aggregation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
C.7 Design Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
xiv Contents
C.8 Cost Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244

C.9 Mathematical Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
C.10 Accompanying Algorithms and Procedures . . . . . . . . . . . . . . . . . . . . . 246
D Conference Participants: Presenters and Attendees . . . . . . . . . . . . . . . . 247

D.1 Conference Participants: Presenters . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
D.2 Conference Participants: Attendees . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
E Simulation Model of a Reliable Production Line . . . . . . . . . . . . . . . . . . . 257

E.1 Description of the Production Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
E.2 The Model of the System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
E.2.1 The Arrive module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
E.2.2 The Server modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
E.2.3 The Resource modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
E.2.4 The Depart module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
E.2.5 The Simulate module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
E.2.6 The Statistics module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
E.2.7 The Animate modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277

List of Figures
1.1 Manufacturing transformation process . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 The interrelationship between process choice, plant layout and
technology investment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Process choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Equipment choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Modeling process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6 Synergistic relationship between evaluative and generative models . . 17
1.7 Complexity of the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.8 Flexibility of the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.9 Transparency to the modeler and to the user . . . . . . . . . . . . . . . . . . . . . 19
1.10 Efficiency of model development and evaluation . . . . . . . . . . . . . . . . . 19
1.11 User interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1 A K-work-station production line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.2 The states of si for Pi = 2, Ri = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Structure of A1 , K > 2, B2 = N, B3 , B4 , . . . , BK . . . . . . . . . . . . . . . . . . . 43
2.4 Relationship of sub-matrix E to D∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.5 Structure of A, K > 2, B2 = N, B3 , B4 , . . . , BK . . . . . . . . . . . . . . . . . . . . 45
2.6 Illustration of Rule 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.10 A merge non-linear flow model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.11 A three-station line, L, decomposed into two sub-lines, L1 and L2 . . . 52
2.12 Production line, L, with K = 4 work-stations and 3 intermediate
buffers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.13 Decomposition of the original line, L, into three sub-lines each with
two stations and one buffer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.14 Sub-line Li . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.15 Flow chart for decomposition method . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.16 Expansion of a finite queue M/M/c/K . . . . . . . . . . . . . . . . . . . . . . . . . 61
xv
xvi List of Figures
2.17 A K-work-station production line with Si parallel machines at each

work-station W Si , i = 1, 2, . . . , K . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.18 A two-station, one-buffer production line with parallel machines at
each station . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.19 Algorithm for generation of lower and upper boundary state
transition probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2.20 Algorithm for generation of internal state transition probabilities . . . . 74
2.21 Flow line with K parallel-machine work-stations, K − 1 intermediate
buffers (Line L) and decomposition scheme (Lines L1 , . . . , LK−1 ) . . . . 76
2.22 Decomposition algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.23 A production line with four stations with parallel reliable machines
at each station and three intermediate buffers . . . . . . . . . . . . . . . . . . . . 87
3.1 A typical structure of a complex production line . . . . . . . . . . . . . . . . . . 102
4.1 The work-load allocation over five stations with inter-station buffer
capacities of sizes B2 = B3 = B4 = B5 = 3 slots . . . . . . . . . . . . . . . . . . 116
4.2 Two-level approximation to a bowl phenomenon . . . . . . . . . . . . . . . . . 117
5.1 General process of solution of buffer allocation problems . . . . . . . . . . 133

5.2 Schematic representation of the form of optimal buffer allocation in
terms of E and K of balanced production lines with exponential and
Erlang-2 service times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.3 Throughput as a function of the ordered buffer allocations for K = 5
and N = 5, showing the “self-similarity” phenomenon . . . . . . . . . . . . . 140
5.4 Average WIP, W IP, as a function of the ordered buffer allocations
for K = 5 and N = 5, showing the “self-similarity” phenomenon . . . . 141
5.5 Simulated annealing algorithm for distributing N buffer space,
S servers, and K work-load in a K-station line . . . . . . . . . . . . . . . . . . . . 149
5.6 Performance of simulated annealing S(SA, Deco) compared with
complete S(CE, Deco) and reduced S(RE, Deco) enumerations for 9
stations (left, middle) (Note the log10 scale on the ordinate axis) . . . . 150
complete S(CE, Deco) and reduced S(RE, Deco) enumerations for
15 stations (left, middle) (Note the log10 scale on the ordinate axis) . . 150
5.8 Number of enumerations required for simulated annealing vs. the
number of stations (Note the log10 scale on the ordinate axis) . . . . . . . 151
genetic algorithms S(GA, Deco) for large production lines . . . . . . . . . 152
5.10 Accuracy of simulated annealing S(SA, Deco) compared with
genetic algorithms S(GA, Deco) for large production lines . . . . . . . . . 153
7.1 Production line design: Historical approach . . . . . . . . . . . . . . . . . . . . . . 180

7.2 Production line design: Modern approach . . . . . . . . . . . . . . . . . . . . . . . 181
List of Figures xvii
7.3 Value of F1 as a function of N for a 5-station production line with

R = 50 FU, C = 40 FU, I = 10%, α = 0.5 (r = 5) . . . . . . . . . . . . . . . . 184
R = 50 FU, C = 40 FU, I = 10%, α = 0.5, b = 1000 FU (r = 5) . . . . 185
R = 50 FU, C = 40 FU, I = 10%, α = 0.5, b = 5000 FU (r = 5) . . . . 186
7.6 Value of F3 as a function of N for a 5-station balanced production
line with R − C = 10 FU, I = 10%, α = 0.5, Ch = 2 FU, b = 1000
FU, 2 shifts per day, 5 units per minute maximum mean production
rate of the system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.7 Value of F3 as a function of N for a 5-station balanced production
line with R − C = 4 FU, I = 10%, α = 0.5, Ch = 1 FU, b = 10,000
FU, 1 shift per day, 1 unit per minute maximum mean production
rate of the system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
A.1 Continuous uniform distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

A.2 Five-stage Erlang distribution, Ek=5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
A.3 A two-phase Coxian distribution, C2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
A.4 State space diagram of a three-state machine system . . . . . . . . . . . . . . 216
A.5 A two-station series queueing network with two identical
exponential stations and an intermediate buffer of capacity 1 . . . . . . . 226
E.1 A production line with four stations with parallel machines at each
station and intermediate buffers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
E.2 The Arrive module dialog box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
E.3 The Server module dialog box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
E.4 The Options dialog box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
E.5 The Resource module dialog box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
E.6 The Depart module dialog box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
E.7 The Simulate module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
E.8 The Statistics module dialog box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
E.9 Saving the value of counter No_of_Jobs into file Throughput.DAT . . 265
E.10 The confidence interval (CI = 95%) of throughput . . . . . . . . . . . . . . . . 266
E.11 A snapshot of the evolution of the average level of buffer B3 up to
time equal to 1000 minutes in a production line with 4 stations with
parallel machines at each station and intermediate buffers . . . . . . . . . . 269
List of Tables
1.1 Processes/products/equipment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Specific flexibilities implied by customer interests via enterprise
functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Types of flexibility in manufacturing . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Overview of the evolution of strategic manufacturing systems . . . . . . 11
2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Number of states for P = 1, R = 1 and identical buffer capacities . . . . 32
2.3 Exponential service, repair, and failure, K = 3 . . . . . . . . . . . . . . . . . . . 34
2.4 Erlang service, exponential repair, and failure, K = 4 . . . . . . . . . . . . . . 35
2.5 Erlang service, repair, and exponential failure, K = 4 . . . . . . . . . . . . . . 35
2.6 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.7 States of station i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.8 Number of states of the system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.9 Altered states and their numerical values . . . . . . . . . . . . . . . . . . . . . . . . 41
2.10 Ordering of states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.11 Throughput of a two-work-station system with parallel machines . . . 74
2.12 Comparison of results with Hillier and So (1996) – 5 work-stations . . 80
2.13 Comparison of results with Hillier (1995) – 3, 5 and 7 work-stations . 80
2.14 Sample configurations for long lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2.15 Sample numerical results for long lines . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.16 Configurations for longer lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.17 Numerical results for longer lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.1 The 13 iterations to find the OBA in the production line of example 1 139
5.2 Searching in classes [0, 0], [0, 1], [0, 2] and [0, 3] . . . . . . . . . . . . . . . . . . . 144
5.3 Searching in classes [1, 0], [1, 1] and [1, 2] . . . . . . . . . . . . . . . . . . . . . . . . 145
5.4 Correspondence between annealing in the physical world and
simulated annealing used for production line optimization . . . . . . . . . 148
6.1 Overall plan of experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
xix
xx List of Tables
6.2 Throughput and buffer allocation for 4-, 6- and 8-station lines via CE 168
6.3 Throughput and buffer allocation for 5- and 6-station lines via CE . . . 169
6.4 Throughput and buffer allocation for 5-, 7- and 9-station lines via SA 169
6.5 Throughput and buffer allocation for 5-, 7- and 9-station lines via CE 170
6.6 Throughput and buffer allocation for 10-station lines via SA . . . . . . . 171
6.7 Throughput and buffer allocation for 10-station lines via CE . . . . . . . 171
6.8 Throughput and buffer allocation for 16-station lines via SA . . . . . . . 172
6.9 Throughput and buffer allocation for 5(1)9-, 11(10)61-station lines
via SA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
A.1 Probability mass function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204

A.2 Discrete probability distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
A.3 Continuous probability distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
A.4 Characteristics of single-station queueing systems . . . . . . . . . . . . . . . . 221
A.5 The transition matrix of the queueing network model of example 1 . . 227
E.1 Simulation results: Continuous variables . . . . . . . . . . . . . . . . . . . . . . . . 263

E.2 Simulation results: Discrete variables . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
E.3 Simulation results: Performance measures . . . . . . . . . . . . . . . . . . . . . . . 264
1
Manufacturing Systems: Types and Modeling
Designers in the past were well aware of the need for effective production systems
but were hampered in the development of such systems by a lack of appropriate
manufacturing technology and system design techniques. In the 1950s, the emphasis
of production management was essentially on throughput and standardization. The
economic philosophy was based on the economics of scale with a significant orienta-
tion toward production to stock. Today the situation is very much changed due to the
introduction of new manufacturing technologies and management philosophies. The
focus now is more on the economics of scope, the customization of products and the
preeminence of the market. In the meantime, there has been a considerable advance
in the range of tools available to the designer of production systems.
In this chapter we give a brief overview of the significant technological changes
which have occurred since the 1950s. The importance of information technology in
manufacturing systems and the need for the designer to have performance measures
other than throughput in mind during the design process is treated. A presentation of
some areas in mathematical analysis, which are important for our work, is contained
in Appendix A. In Section 1.1, the evolution and classification of manufacturing sys-
tems is covered. Section 1.2 treats mathematical models and the modeling process.
Section 1.3 attempts a general classification of manufacturing systems with a view
to showing the inherent complexities. Section 1.4 discusses models in the context
of manufacturing systems, whereas Section 1.5 treats methods of analysis of such
models. Finally, Section 1.6 presents measures of performance in manufacturing
systems.
1.1 Manufacturing Systems: Evolution and Classification

Manufacturing is a transformation process as shown in Figure 1.1.
In this model, the inputs (capital, raw materials, energy, educated and trained
personnel, equipment and facilities, tools and software, and customer demand) are
transformed to finished products which are demanded by the market. Inevitably there
is some waste and scrap produced. The management of such a transforming process
C. T. Papadopoulos et al., Analysis and Design of Discrete Part Production Lines, 1

Springer Optimization and Its Applications,
DOI: 10.1007/978-0-387-89494-2_1, © Springer Science+Business Media, LLC 2009
2 1 Manufacturing Systems: Types and Modeling
Capital
Raw materials
Energy
Finished goods meeting
customer needs
Educated & trained
personnel
Manufacturing
system Scrap and waste
Equipment & facilities
Tools and software
Customer demand
Fig. 1.1. Manufacturing transformation process
is quite complex, as it must have regard for the business imperatives as well as the
technical possibilities of the manufacturing process. In the past, the transformation
process tended to take place within the four walls of a factory but this is no longer
the case with the formation of virtual enterprises.
One of the major technological breakthroughs in addition to the abundant supply
of energy which led to modern manufacturing was the development of the concept
of interchangeable production, credited to Eli Whitney in the early 1800s. Up to that
time, craft persons tended to make complete and often individualized products and so
interchangeability of parts was not of major importance. The philosophical concept
of division of labor, developed by Adam Smith, with its associated cost advantages
further added to the development of manufacturing as a form of production. The
interested reader is referred to the very rich literature on operations management
to appreciate the work of such pioneers as Taylor, Gantt, Babbage, and Hawthorn,
among others.
At the strategic level, a company must decide which markets it wishes to compete
in and what will be its competitive advantages in these markets. Clearly, appropri-
ate technology will confer competitive advantage on a company provided the other
essential ingredients for success are also in place. In addition to training and educa-
tion, such additional ingredients include the layout of the plant and the basic choice
of the manufacturing process.
In all types of manufacturing systems there are a number of basic functions or
activities or operations which must be performed during the transformation process.
These activities include:
• Processing operations
• Assembly operations
1.1 Manufacturing Systems: Evolution and Classification 3
• Material handling, transportation and storage

• Product quality assurance, inspection and test
• Process control
It is unfortunate that the word “process” has so many different meanings in the con-
text of manufacturing. Processing operations transform the product from one state
to a more advanced state of completion. Such processing operations (e.g., metal
removal, distillation) may be classified in different ways but these classifications
are unimportant in the context of this work.
Assembly and joining operations (or blending) involve the combination of two
or more separate components. In some systems the operations start at assembly,
because no other processing is involved. Thus, it is possible to describe a particular
manufacturing system as a flexible assembly system (FAS).
Material handling may be manual, semi-automatic or automatic.
Product quality assurance activities are major activities in modern manufactur-
ing. The functions involved may be automated or carried out manually. In some
systems, results from quality assurance are fed back to the production machines.
Process control involves the achievement of performance objectives through the
manipulation of inputs to the process. There exists a significant body of knowledge,
based on statistical methods, to achieve process control.
Speed, reliability, flexibility, cost, rapid product innovation and quality are all
related to process choice. The interrelationship between process choice , plant layout
and investment in technology is clearly shown in Figure 1.2, based on material given
in Brown (1996).
Classically, the layout decision is often described in terms of maximizing the use
of equipment and personnel and is essentially considered to be tactical in nature. For
example, well-known layout techniques exist which minimize the distance traveled
by operatives. However, the layout problem should more properly be conceived as
part of a strategic decision which supports the process choice in serving the chosen
markets.
It is generally agreed that there are five basic process choices as follows:
Investment in
technology
Process Product
technology technology
Process choice/ Volume/

Plant layout Costs
Fig. 1.2. The interrelationship between process choice, plant layout and technology investment
Project
Job
Variety
Batch
Line
Continuous
Volume
Fig. 1.3. Process choice
1. Project: One large and complex unit.

2. Job Shop: Different products in small lot sizes.
3. Batch: Many standard or similar products to customers’ specifications but rela-
tively small volumes.
4. Production/Transfer Line: High-volume repetitive production of discrete units
often associated with a moving assembly line.
5. Continuous Process: Flow-process required by the production technology.
Process types 1 to 4 are often considered to be associated with discrete material flow,
whereas process type 5 is a continuous flow process. The initial choice of process
may be represented in a diagrammatical form as shown in Figure 1.3.
Associated with the choice of process is an appropriate equipment configuration
for specific industries as illustrated in Figure 1.4 (see Phillips, 1997, Figure 4.1).
As may be seen from Figure 1.4, there is a manufacturing spectrum (Phillips,
1997) based on the degree of flexibility (this term will be discussed below) ranging
from high-volume, low-variety production (dedicated and/or automated equipment,
these include continuous flow lines and the well-known production/transfer lines)
to low-volume, high-variety production (standard machinery and equipment, these
include job-shop systemswith process layout and individual project-based systems;
in practice using either stand-alone NC1 or CNC2 machine tools or integrated
machining centers). The difference between production or flow lines and transfer
lines is dependent on the regularity of the movement of material between the sta-
tions. In transfer lines, known also as paced lines, the movement is synchronous,
whereas in production or flow lines, known also as unpaced lines, the movement
is asynchronous. Usually, production lines and transfer lines are one-product lines
with a high output. Continuous flow lines refer to high-volume production systems
where the material process has liquid properties. In a pure job-shop environment, a
1 Numerical control.
2 Computer numerical control.
Degree of Flexibility Required
Job Shop Machinery

& Equipment,
NC Machine Tools
Manufacturing Cell
Group Technology
Variety
Flexible
Manufacturing
Systems
Production or Flow
Lines, Automatic
Transfer Lines
including
Continuous Flow
Lines
Volume
Fig. 1.4. Equipment choice
large variety of products of relatively small volumes are produced. In addition to the
two extremes, there is a need for manufacturing systems with a capability of pro-
ducing mid-range volumes (mid-volume manufacturing) with a significant degree of
flexibility. These manufacturing environments are usually catered for by either flex-
ible manufacturing systems (FMS) or flexible manufacturing cells (FMC) including
a group technology (GT) philosophy of operation.
Essentially, FMS are computer-controlled systems consisting of several stations
each specializing in particular operations with an appropriate transport system for
the movement of the product. The computer system coordinates the activities, and
the essence of FMS is their inherent flexibility. Using an FMS, products may be
produced in a number of variations and in different volumes in different time ranges.
Historically, FMS were developed because of the high cost of production of small
volumes under production line conditions.
As quoted by Schmenner (1990), “In essence group technology (GT) is the con-
version of a job shop layout into a line flow layout. Instead of grouping similar
machines together, group technology may call for grouping dissimilar machines
together into a line flow process all its own. In the new arrangement, a part can
travel from one machine to another without waiting between operations, as would be
customary in the job shop.”
The major benefits of group technology include the rationalization of tooling set-
ups, reduction of set-up times, reduction of throughput times and improvements in
tool design as well as more efficient production planning and scheduling.
Table 1.1. Processes/products/equipment
Process Typical Product Equipment

Project Airplane, space vehicles (one-offs) Standard NC, CNC
Job Instruments, machine tools, Standard NC, CNC
prototypes of future products machining centers by
[Low-volume] manufacturing function
Batch High-end consumer products GT cells,
(e.g. lawn mowers, electric motors, focused mini factories, FMS
furniture, textbooks) [Mid-volume]
Line Telephone screws, light bulbs Automated equipment,
[High-volume] moving assembly lines,
flow lines
Continuous Beer, detergents, chemicals Continuous-flow fully
[Very-high-volume] automated systems
In Table 1.1, a sample list of products is given with the expected associated
process choice.
Flexibility is a term that is widely used in the management and engineering litera-
ture, often without any great degree of precision of language, and is usually assumed
to be a term of approval. In general, whether applied at the overall enterprise or
subenterprise level, the concept implies the ability to cope with change. The impres-
sion is sometimes given that the existence of any degree of flexibility in industrial
organizations, other than zero, is a very modern phenomenon but this is simply not
the case.
In an effort to assist in understanding the need for flexibility in enterprises and
the different dimensions of flexibility, perhaps it is worthwhile to consider the inter-
ests of the customer and the market. An examination of these interests demonstrates
the need for the flexibilities required at the enterprise level and consequently at the
manufacturing systems level.
Table 1.2 lists the customer/market interests or expectations. These lead to the
listed functions of enterprises and these functions imply the five listed flexibilities,
arising, as it were, from a customer focus.
Customers expect a rapid response to their demands. Being late to the market
with a new competitive product is often much more costly to a firm than significant
overruns in either research or development costs. Flexibility in all functional areas of
the firm is the key to “time-based competition,” a competitive advantage term used
to describe efforts to increase innovation, reduce product development time, reduce
delivery time and respond “fully” to the individual needs of customers.
The flexibility of a manufacturing system (automatic or manual) is a function
of the physical system, its associated software, and how it is operated. Table 1.3,
columns 1, 2, and 3 of which are taken with permission from Groover (2001),
defines seven types of flexibility (machine, production, mix, product, routing, vol-
ume, expansion flexibility) exhibited by manufacturing systems and the factors on
Table 1.2. Specific flexibilities implied by customer interests via enterprise functions
Customer/Market Interests Functions of Enterprises Specific Flexibilities Implied
Availability of: Deliver Mix of Quality I Product Mix Flexibility competence to produce
Products/Services including: a wide range or mix of
⭈ Right Products/Services New Products/Services to products
meet customer demand
II Delivery Flexibility capacity to change the
⭈ Right Quantities dynamically by:
time of delivery by
⭈ Right Time rescheduling manu-
facturing priorities or
⭈ Effective Production supply chain
⭈ Right Price
⭈ Effective Delivery III Volume Flexibility ability to change the
⭈ Right Place level of output/time of
⭈ Secure Systems any product produced
IV New Product potential to introduce
Some interests relating Flexibility new products
to environment, ethics ability to meet customer
of business
, etc., are not V Dependability needs in a dependable
considered fashion
Leads to
Implies
7
8
Table 1.3. Types of flexibility in manufacturing
Flexibility Type Definition Depends on Factors Such as Customer Focus Flexi-

bility Type (Table 1.2)
Machine Flexibility Capability to adapt a given machine (work- Setup or changeover time. Ease of machine I, IV
station) in the system to a wide range of produc- reprogramming (ease with which part pro-
tion operations and part styles. The greater the grams can be downloaded to machines).
range of operations and part styles, the greater Tool storage capacity of machines. Skill and
the machine flexibility. versatility of workers in the system.
Production The range of universe of part styles that can be Machine flexibility of individual stations. I, III
Flexibility produced on the system. Range of machine flexibilities of all stations
in the system.
Mix Flexibility Ability to change the product mix while main- Similarity of parts in the mix. Relative work I, II
taining the same total production quantity; that content times of parts produced. Machine
is, producing the same parts only in different flexibility.
proportions.
Product Flexibility Ease with which design changes can be accom- How closely the new part design matches IV
modated. Ease with which new products can be the existing part family. Off-line part pro-
1 Manufacturing Systems: Types and Modeling
introduced. gram preparation. Machine flexibility.

Routing Flexibility Capacity to produce parts through alternative Similarity of parts in the mix. Similar- V
work-station sequences in response to equip- ity of work-stations. Duplication of work-
ment breakdowns, tool failures, and other inter- stations. Cross-training of manual workers.
ruptions at individual stations. Common tooling.
Volume Flexibility Ability to economically produce parts in high Level of manual labor performing produc- II, III
and low quantities of production, given the fixed tion. Amount invested in capital equipment.
investment in the system.
Expansion Flexibility Ease with which the system can be expanded to Expense of adding work-stations. Ease with II, III
increase total production quantities. which layout can be expanded. Type of
part handling system used. Ease with which
properly trained workers can be added.
which they depend. In column 4 we have added the associated customer focus
based flexibilities as defined above. Clearly, engineering and business judgment are
required in accessing the “relative flexibilities” of manufacturing systems.
The development of information technology played a seminal role in the evolu-
tion of manufacturing systems over the past 40 years. Information technology is not
only embedded in the equipment being used and the products being produced but it
also gives the capability to operations managers to focus on the information flows in
the system.
Today, computers are used extensively in design and engineering under the gen-
eral title of computer-aided design (CAD) and computer-aided engineering (CAE).
Similarly, computers are used for production planning and control and tool control
using computer-aided manufacturing systems (CAM). Integrated systems includ-
ing CAD, CAE, CAM, CNC and FMS are referred to as Computer-Integrated
Manufacturing Systems (CIM). The manufacturing systems designer must always
be conscious of the advantages and disadvantages of using humans or machines
for specific operations. It is outside the scope of this work to discuss trends in
human-centered automation.
Of particular importance to the operation of manufacturing systems is the distinc-
tion between push and pull systems of materials management. In the push system,
generally associated with materials requirements planning (MRP), the material is
“pushed” through the manufacturing process by the scheduling system and final
product is often stored until demanded by customers. Pull systems on the other
hand are activated by orders from customers, final and intermediate, and a main
characteristic is reduced work-in-process.
Since the 1990s, there has been a strong interest in accurately assessing the cost
of products, and an accountancy process known as activity-based costing (ABC)
has been developed to accurately determine the manufacturing cost and other costs
associated with a particular product or customer.
In recent times, at least three management philosophies of manufacturing have
been promoted by consultants and academics. These are “lean production,” “agile
manufacturing,” and “intelligent manufacturing.” Womack et al. (1990) described
the characteristics of lean production as integrated production with low invento-
ries using a just in time (JIT) philosophy, and teamworking with a multi-skilled
workforce. In essence the lean manufacturing approach is a combination of JIT and
total quality management (TQM) philosophies. Further information may be found in
Brown (1996).
The term “agile manufacturing” is used to describe a new manufacturing para-
digm to replace existing thinking on mass production. There are four principles of
agile manufacturing (agility), viz., organize to master change, leverage the impact
of people and information, cooperate to enhance competitiveness, and enrich the
customer. Agility may be considered a characteristic of the enterprise rather than
simply of the manufacturing system. The interested reader is referred to Groover
(2001) and Gunneson (1997), among others.
“Intelligent manufacturing” is manufacturing, with the minimum of human
intervention, by equipment in which is embedded the skills and knowledge of
manufacturing experts so that the products produced are indistinguishable from those
produced in conventional manufacturing systems and with similar levels of output
and utilization of raw materials and energy. The skills and knowledge of the manu-
facturing experts (managers, engineers, craft persons and operatives) are embedded
in the system by the use of expert systems, databases and data management systems,
and intelligent machines such as robots with vision and manipulation possibilities.
As it is clear from the above, decisions in relation to process choice, layout and
equipment choice are strategic in nature. These decisions will have a major impact
on the long-term viability of the associated company. Such decisions would normally
be made before the detailed design of a manufacturing system was undertaken. Of
course, any enterprise could make different process choices in relation to different
products within its market range or in relation to the same type of product over the
manufacturing cycle. For example, a particular product may be made using two or
more process choices, i.e., some components of the product may be made on produc-
tion lines, a few components might be produced in a dedicated manufacturing cell,
and the final assembly of the product to customer specification might be performed
under flexible assembly system (FAS) conditions. It is unusual for an enterprise to
use only one specific “pure” type of manufacturing system, e.g., job shop or transfer
line, and a firm may use a mixture of types with the output from one system being
an input to another.
A useful summary of the evolution of strategic manufacturing systems is given
in Table 1.4, which is taken, with permission, from Ostwald and Munoz (1997).
As may be seen from Table 1.4, the driving forces of manufacturing systems have
changed over the years from “cost” in the 1960s to “service and value” in the 2000s.
The associated manufacturing strategies have likewise changed from high-volume,
cost minimization, and product-focused systems to customer-centered global inte-
gration and virtual enterprise systems with a significant concern for the environment
and safety.
Systems used to support the strategies have likewise changed over time from
an emphasis on production and inventory control systems and numerical control
machines in the 1960s to “intelligent” manufacturing systems incorporating flexi-
ble and agile automated systems with emphasis on ergonomics and safety systems in
the 2000s.
The authors believe that the early decades of this century will see the continua-
tion of a very strong customer-driven intelligent manufacturing (CDIM) paradigm,
in which manufacturing will have a strategic focus on catering for the “total” satis-
faction of the customer by delivering enlarged products, i.e., physical products plus
services, perhaps, delivered via a network of virtual enterprises from different sites
on a global basis. Whether the manufacturing community will be satisfied with this
role is an open question. Manufacturing expertise has added significantly to the com-
fort level of human living over the past couple of centuries. It will continue to have
a major role in this regard. However, should its mission be confined to this role or
should it seek a higher perhaps more spiritual role in assisting to ensure the survival
of the human community by realizing the full ambitions of human beings in areas,
among others, such as space travel, health care, infrastructure development, and the
Table 1.4. Overview of the evolution of strategic manufacturing systems
1960s 1970s 1980s 1990s 2000s

Driving Force Cost Market Product Quality Time to Market Service & Value
Manufacturing High Volume Functional Process Control New Product Customer-Centered
Strategies Integration Introduction Mission
Cost Minimization Closed Loop Material Velocity Responsiveness Information
Sharing
Stabilize Automation World Class Manufacturing Global Integration
Manufacturing Metrics
Product Focus Diversification Overhead Cost Reengineering Environmental
Reduction Safety
Rapid prototyping Virtual Enterprise
Systems to Production and Material Manufacturing Computer “Intelligent”
Support the Inventory Control Requirements Resource Integrated Manufacturing
Strategy Systems Planning Planning Manufacturing Systems
Numerical Control Master Production Just in Time Decentralization Flexible and Agile
Scheduling Automated Systems
Computer Statistical Quality Simplification Continuous bench-
Numerical Control Control marking systems
Push Systems Computer-Aided Total Quality Community
Design and Management Involvement
Manufacturing
Simulation Self-directed Continuous
Workforce Infrastructure
Improvement
Pull Systems Activity-Based Paperless
Costing Systems
Ergonomics
Safety Systems
11
reclaiming of the polluted environment? Of course, the manufacturing community

could only take on such a higher mission with general political support and leader-
ship. There may well be a limit to the extent to which any community would allow
its individual members to selfishly consume for their own satisfaction a portion of
the limited manufacturing resources in the context of a customer-driven intelligent
manufacturing paradigm when such resources could with advantage be used else-
where for the benefit of the human community. The resolution of such issues are
well outside the scope of this text.
1.2 Models and Modeling

Models are a means for studying phenomena. A useful model yields information
about the real system it represents at a lower cost and more quickly than if one
undertook experiments on the real system.
Models may be classified in a number of ways. A basic classification is physical
or abstract (mathematical). Physical models may be analogue or iconic. Analogue
models exhibit characteristics in some of their variables which are of interest to the
model builder. This requires that the underlying physical behavior of the real-life
system and of the analogue system are related through a similar set of mathematical
equations. For example, a physical vibrating system may be modeled using an elec-
trical network where current in the electrical network corresponds to motion in the
physical vibrating system.
Iconic models, on the other hand, are physical models where measurements are
made on the physical replicates, often of a reduced scale, of the objects under study.
A good example would be architectural models to assist in understanding space
utilization.
Abstract or mathematical models are sets of equations with mathematical sym-
bols rather than physical devices. Models are in effect a mental image or an intel-
lectual description of a process. This idea was captured by Robert M. Pirsig in his
book (1974) by saying: “An untrained observer will see only physical labour and
often get the idea that physical labour is mainly what the mechanic does. Actually
. . . mechanics don’t like it when you talk to them because they are concentrating on
mental images, hierarchies, and not really looking at you or the physical motorcycle
at all . . . . They are looking at underlying form” (concepts very familiar to Plato).
The initial stage of deriving a model (modeling) is an appropriate simplification
or idealization by extracting from the real-life situation those characteristics, prop-
erties, or features in which we are interested. This stage may be called the problem
formulation stage. Once these significant features have been identified, the next stage
is to assign mathematical terms to them and to formulate relationships (equations)
between these terms. In general, this is not an easy task but it leads to what is nor-
mally called the model. A validation process takes place throughout the problem
formulation stage and after the model itself has been developed. The analyst must
have assurance that the problem has been formulated correctly and that the model
represents reality appropriately. Clearly, as far as the set of mathematical equations
1.2 Models and Modeling 13
Problem
Real Model
Formulation
Physical Situation
Self
Validation Consistency
Fig. 1.5. Modeling process
is concerned, there must be self-consistency. The overall process may be described

as shown in Figure 1.5.
Two quantitative factors, as follows, arise in the development of models:
• Parameter, a characteristic or factor which cannot be changed during the analysis

of a particular specified system (e.g., the number, K, of work-stations required by
process considerations alone in a manufacturing system).
• Variable, a factor which assumes more than one value or state during the period of
interest.
In modeling, the term static implies a relationship that does not change with time,
(e.g., the aggregate production planning problem), whereas dynamic deals with vary-
ing time interactions (e.g., the routing problem in a job shop is a dynamic model).
Models may be linear or non-linear. Linear models have the property of superpo-
sition by which is meant that if an input of X produces an output of Y, an input of
α X produces an output of α Y for any real number α . In non-linear models, such a
relationship would not hold (e.g., the allocation of a certain amount of buffer space
among the work-stations of a production line). Dynamic models may be sub-divided
into stable and unstable models, whereas static models must be stable. A stable
model is one that returns to its original condition after a perturbation, whereas an
unstable model does not return to its original condition.
Other concepts of interest in modeling are steady-state and transient solutions.
Steady-state solutions imply long-term average behavior, whereas transient solutions
are primarily dependent on the initial conditions of the system. These distinctions are
very important in simulation models of manufacturing systems, whereas analytical
models of these systems are generally steady-state.
A final distinction may be made between open and closed models. A closed
model is one that functions without external input and generates the value of vari-
ables through time by the interaction of one variable on another. Open systems are
open to receive inputs from the outside.
Models may be used in two ways, on-line and off-line. In some processes a model
of the process may be used as a decision support system in real time to assist in
control decisions. These on-line models, need to be very reliable but usually are
relatively simple because of the need to assist the decision maker quickly. Off-line
models, which are the only ones considered in this text, do not have such constraints.
1.3 Classification of Manufacturing Systems

Our initial objective was to develop a system of classification/notation of manufac-
turing systems which would be capable of describing comprehensively the essential
elements of manufacturing systems particularly from the point of view of the oper-
ations manager and the designer of such systems. The classification/notation system
would not go into the precise details of each process, i.e., would not go beyond
describing a drilling process as a drilling process.
It should be appreciated that in practice, an enterprise may use a combination
of what in textbooks are described as “pure” types, of manufacturing systems, e.g.,
job shop, transfer line, FMC, FMS, etc., in the production of its products. However,
major elements of the manufacturing systems used by an individual enterprise could
probably be approximately described by some of these “pure” types of manufactur-
ing systems. The detailed classification would be applied in turn to each of these
“pure” types and the overall system would be a combination of the individual classi-
fications, maintaining the individual characteristics of the separate elements making
up the total system.
Manufacturing systems consist of work-stations at which operations take place.
Each work-station may be either automated or manually operated. In some sys-
tems, because of capacity considerations there may be a number of identical or
near-identical work-stations operating in parallel. A basic distinction must be made
between manufacturing systems at which only one product is produced (single-
product systems) and those on which more than one product is produced (multi-
product/mixed-product systems). In mixed-product systems a particular work-station
may carry out a different set of activities depending on the particular product type
(e.g., an FMS work-station).
A significant characteristic of manufacturing systems is the difference between
fixed and variable routing. In fixed routing all piece parts go through the same
route, i.e., are processed by each work-station in turn in the same sequence. In vari-
able routing systems piece parts are processed through a variety of different station
sequences.
Below, a manufacturing systems classification/notation scheme consisting of 12
descriptors, viz., A, B, C, D, E, F, G, H, I, J, K and L, is postulated. This scheme may
be applied only to identifiable sections of an enterprise’s overall manufacturing sys-
tem where the sections in question would approximately correspond to the “pure”
types of manufacturing systems normally described in operations/manufacturing
textbooks. The scheme is essentially numerical in nature using integers. In respect
of some of the descriptors, the information is given in the form of a set of numbers
which may be arranged in a row (row vector) or in an array (matrix). The usual con-
vention about the dimensions and components of row vectors and matrices are used.
A scalar is an ordinary number, in this case, an integer.
A: Number of separate work-stations not including identical or near identical work-
stations in parallel.
B: Number of different products which may be produced by the manufacturing
system.
1.3 Classification of Manufacturing Systems 15
C: A row-vector of dimension 1 × A giving the number of operations that can be

undertaken at each work-station.
D: A set of size A vectors/scalars, one for each work-station, indicating the type
of operations which may be undertaken at each work-station, e.g., 1 = process-
ing operation, 2 = assembly operation, 3 = materials handling, transportation
and storage, 4 = product quality assurance, inspection and test, and 5 = process
control (a finer degree of operation classification is possible). The dimension of
each vector in set D is determined from the corresponding elements of C. The
information contained in descriptors C and D may be given by one descriptor
alone.
E: A row-vector of dimension 1 × A with each element corresponding to each of the
work-stations in the system with components 0 or 1, where 0 indicates manual
and 1 indicates automatic.
F: A row-vector of dimension 1 × A which indicates the total number of identical
or near identical machines or work centers working in parallel associated with
each of the separate work-stations listed in A.
G: A set of row-vectors, one for each product, giving the sequence of work-stations
on the preferred route, according to a numbering scheme for the work-stations in
the manufacturing system. This scheme must be clearly defined.
H: A set of row-vectors each member of which is associated with each element of
vector C which is greater than 1, indicating the micro-route for each product
using the associated work-station. The control systems of FMS and machining
centers could complicate this descriptor.
I: A set of row-vectors, one for each product, giving the expected processing time
spent in each work-station on the preferred route in accordance with the num-
bering scheme of the work-stations as specified in descriptor G. Descriptor I
is obtained using information contained in G and H and the relevant expected
processing times.
J: A set of row-vectors, each of dimension 1 × 2, for each work-station indicating
the number of its incoming and outgoing buffers.
K: A set of row-vectors, one for each work-station, indicating the maximum capac-
ities (sizes) of its associated incoming buffers from other work-station(s) (capac-
ity is based on the mix of products, i.e., a “standard” product).
L: A set of row-vectors for each work-station indicating the capacities (sizes) of its
associated outgoing buffers to other work-station(s) (capacity is based on the mix
of products, i.e., “standard product”).
One needs to be careful when discussing the number of work-stations in a manufac-
turing system from the viewpoint of an individual product.
This classification system, developed above, although perhaps somewhat aca-
demic and which, in some circumstances, could be open to criticism on the basis of
lack of completeness, clearly illustrates the complexity of manufacturing systems in
general and demonstrates the need to confine one’s attention in the context of today’s
computing abilities to well-defined and well-structured manufacturing systems. It is
clear that manufacturing systems are significantly more complex than the queueing
systems found in textbooks often described by Kendall’s well-known notation.
1.4 Models of Manufacturing Systems

Investments in manufacturing are generally considered strategic because of their size
and impact for the enterprise concerned. Once a proposal has been made to invest in
such systems, there is a need to agree on the outline conceptual design of the man-
ufacturing system. Subsequently, the detailed design of the system and modes of
operation of the system must be developed. Finally, appropriate control strategies of
the manufacturing system have to be designed. It is clear therefore that there is a need
for a large number of different types of models to assist the different actors (busi-
ness analysts, design engineers, operations managers, finance managers, marketing
managers and manufacturing personnel) in their decision-making tasks. Among the
reasons why one large overall model of a manufacturing system is not realistic are
• the physical and information flow complexities of such systems as partially illus-
trated in Section 1.3;
• the legitimate different interests of the various actors involved;
• all design is an iterative process with information becoming available as a sequence
of decisions are made. Accordingly, there is, in effect a hierarchy of decisions
which in turn leads to a hierarchy of appropriate models.
Arising from the above considerations, three generic types of models may be
described:
• Planning model
• Design model
• Control/operation model
The planning model is used to test initial assumptions in relation to such issues as
the number and type of work-stations, the type of transportation systems and the
information and control systems to meet the business requirements in relation to
products, finance and return on investments. In planning models the level of detail is
usually low as precise detail of the system has not been developed fully. An economic
justification model falls into this category.
The design model, having as input the information derived from the planning
model, is used to determine such matters as the location and size of inventory buffers,
the number of parallel work-stations in a series-parallel structure, the work-load allo-
cation among the work-stations, the details of the physical transportation system, the
number of pallets and the tool storage capacity. The control strategy of the system
needs to be specified and tested at this stage. The level of detail of these models is
considerably higher than in the planning models.
The control/operation model is normally based on a fully specified physical
manufacturing system. These models are used in the day to day operation of the
manufacturing system to examine such operating issues as the input control (determi-
nation of the sequence and timing of the release of jobs to the system), the scheduling
of work on each work-station and the behavior of the transportation systems between
work-stations.
1.4 Models of Manufacturing Systems 17
Planning models and design models are applicable to a large variety of man-
ufacturing systems, whereas control/operation models are normally developed for
job shops, FMS, flexible manufacturing cells (FMC) and flexible assembly systems
(FAS).
Although the level of detail increases from planning model through design model
to control/operation model, it is unlikely that a model at one level is a simple upgrad-
ing and enhancement by way of greater detail of a model at a previous stage. Models
at the various levels may be used to assist the development and validate the models
at the higher levels of detail.
Modeling practice has given rise in the literature to a description of two types of
models, as follows:
• Evaluative or predictive models in their basic form assume a particular configura-

tion of the manufacturing system under study and performance measure(s) of the
system are determined/evaluated. Such models could be used as planning models
or design models in the context of the classification already given.
• Generative or optimization models have as their basic purpose the determination of
an optimal solution to the system parameters given an overall manufacturing sys-
tem structure and an objective function to be optimized. In particular, two types
of generative models are used in manufacturing systems analysis: (i) models used
to determine specified parameters of the systems to maximize throughput and (ii)
models used to determine specified parameters of the system that minimize a speci-
fied cost objective while achieving a feasible target throughput. Generative models
are particularly useful in determining optimal design specifications.
There is a great potential for a synergistic relationship between these two
types of models as illustrated in Figure 1.6, which is taken, with permission, from
Papadopoulos et al. (1993).
The evaluative model, as noted above, does not necessarily give the user an opti-
mal solution but instead evaluates certain decisions leading to the determination of
performance measures of the system. The generative model, in turn, seeks to max-
imize these performance measures subject to overall constraints that are globally
consistent with the decisions inherent in the evaluative models.
Evaluative Generative
model model
Fig. 1.6. Synergistic relationship between evaluative and generative models

An example may clarify the issues involved. A manufacturing system has been
specified by a certain number of work-stations and a configuration of same with a
given transportation system. In the evaluative model, the capacity of the storage loca-
tions (buffers) associated with each machine is specified as are the product range,
product mix, product routes and production control mechanism. The throughput or
other measures of performance of the system is determined by the evaluative model.
The generative model, in turn, considers a given total amount of storage space allo-
cated to the buffers to be a constraint and finds the maximum possible throughput
with the decision variables being the allocated buffer slots (space) to each buffer
associated with each work-station consistent with the overall amount of storage space
being allocated.
The main thrust and focus of this text is toward design models of production lines,
although a number of the modeling techniques presented and algorithms described
could with some modifications be used as a basis for planning models and to a lesser
extent would be of value in developing operations models of manufacturing systems.
1.5 Methods of Analysis

There are two distinct approaches to the analysis of models of manufacturing sys-
tems, simulation methods and analytical methods. The simulation method involves
the representation of the real manufacturing system in a computer-based model via
the use of an appropriate simulation package such as Arena or eM-plant. Certain
computer packages are more suitable for simulating specific types or parts of man-
ufacturing systems such as FMS or materials handling systems. A major problem
with simulation is the validation of the model, particularly if the manufacturing sys-
tem is not actually built. The solution to evaluative models may be obtained through
simulation.
Analytical methods on the other hand involve formal mathematical solutions
to the problems. Because of the complexity of the mathematical models involved,
two approaches to obtaining a solution are used. An exact, sometimes closed form
solution, may be obtained to a simplified problem or an approximate solution may
be derived often by means of an appropriate and efficient algorithm to the actual
mathematical problem.
It is instructive to consider the diagrams shown in Figures 1.7 to 1.11, devel-
oped from ideas presented in Archetti et al. (1989), which gives a comparison of
simulation and analytical methods of solution of models.
Simulation methods are capable of handling more complex model structures than
are analytical methods, particularly than are those models associated with exact ana-
lytical solutions. In respect of flexibility, it is relatively easy to change the values of
the parameters in analytical models but difficult to change the structure of these mod-
els. As far as simulation models are concerned, parameter values and the structure
may be modified with some degree of difficulty. Transparency, as a characteristic of
solution methods, may be considered from the point of view of the modeler or of the
user. As far as users are concerned there is a low degree of transparency in regard to
1.5 Methods of Analysis 19
High Simulation
Complexity Approximate analytical
Low Exact analytical
Fig. 1.7. Complexity of the model
High Analytical Parameter Values
Flexibility Simulation Parameter Values

Simulation Structure
Low Analytical Structure
Fig. 1.8. Flexibility of the model
High Simulation Modeler

Analytical Modeler
Transparency Simulation Logic Only User
Low Simulation User

Analytical User
Fig. 1.9. Transparency to the modeler and to the user
High Analytical Evaluation
Efficiency Simulation Development

Simulation Evaluation
Low Analytical Development
Fig. 1.10. Efficiency of model development and evaluation

High Simulation, if Visual Display
User Interface
Analytical, if incorporated in DSS
Low
Fig. 1.11. User interface
both solution methods, although the user of simulation models may appreciate the
logic of the model better. The modeler of course in both cases would fully understand
the solution method. Efficiency has two dimensions: (i) in respect of the development
of the model and the solution process and (ii) the efficiency of evaluation. As far as
simulation solutions are concerned, both the evaluation and the development are rel-
atively efficient, whereas in the case of analytical methods the development process
tends to have a low efficiency but the evaluation of the model is often highly efficient.
Finally, in regard to the user interface, the simulation model is highly user friendly
if a visual display unit is part of the simulation solution, and analytical models may
be made more user friendly by incorporation into easy to operate decision support
systems (DSS).
Mathematical techniques found useful in the solution to the models under dis-
cussion include queueing theory (single/multiple station models and queueing net-
work models), Markovian models, graph theory, Petri net models, mathematical
programming models (non-linear programming, stochastic programming, dynamic
programming), search methods, gradient techniques, perturbation methods, simu-
lated annealing methods, Tabu search, various heuristic algorithms, and genetic
programming, among others.
A rather different approach to the development of solutions to problems asso-
ciated with the design and operation of manufacturing systems is the artificial
intelligence (AI) expert systems based techniques. The expert system software is
embedded with knowledge obtained from manufacturing specialists and arranged
according to a rule base. These techniques have been mainly applied to operational
problems particularly where heuristic solutions are the norm. Such methods are
outside the scope of this work.
1.6 Measures of Performance
The typical performance objectives of manufacturing operations from the point of

view of either the operations manager or the customer are generally listed under
the five headings of quality, speed, dependability, flexibility and cost. Arising out
1.6 Measures of Performance 21
of these objectives, the following are some of the (technically based) performance
measures commonly used with respect to manufacturing systems:
• Throughput or mean production rate or mean output rate, X , is the expected

number of parts produced per time unit, given a specified product mix.
• Mean production time, 1/X , is the expected time a “standard” product spends in
the system from the time of its entry to the time of exit. The “standard” prod-
uct concept takes into account the product mix. The mean production time is the
reciprocal of the throughput.
• Mean work-in-process (progress), W IP, is the number of parts present in the
whole system (being processed in the work-stations and awaiting for processing).
For clarification, it should be noted that transportation within the manufacturing
system should be considered a process in this context.
• Utilization, ρi , of a work-station i is the proportion of time that the work-station is
busy (processing parts).
• System or global utilization is the mean of the utilizations of the work-stations
making up the system (including work-stations operating in parallel). For opera-
tional reasons, the utilization of transport systems is often neglected in calculating
global utilization, particularly in FMS.
• Availability, Ai , of a work-station i is the proportion of time the work-station is
capable of processing parts whether required to do so or not. A work-station can
break down and so is not available during the repair period.
• Efficiency, ei , of a work-station i is the ratio of the mean output rate of the
work-station as part of the manufacturing system divided by the mean output rate
achievable without the constraints of being embedded in such a system. When part
of a manufacturing system, a work-station may be blocked from processing parts
or starved because no parts are available for processing thus affecting the arrival
pattern of parts.
• Blocking time proportion for a work-station embedded in a manufacturing system
producing a “standard” mix of products is the proportion of time a particular work-
station, although available for processing, is unable to continue the processing of
parts because its output buffer or the next work-station, in case where there is no
buffer, cannot accept any more product.
• Holding time or completion time of a work-station for a “standard” product is the
product of the “standard” processing time by (1 plus the blocking time proportion).
• Other measures of performance of manufacturing systems include: the mean num-
ber of busy work-stations of any set of work-stations working in parallel, the mean
queue lengths of “standard” products awaiting processing at each work-station,
the mean waiting time for a “standard” product before being processed at a par-
ticular work-station including waiting for transportation between work-stations
and measures of the relative importance of set-up time, particularly in job shop
environments, among others.
1.7 Related Bibliography

There are many excellent textbooks covering the general areas of production and
operations management. Early classics include Wild (1985) and Buffa (1973). These
textbooks tended to emphasize the tactical aspects of operations management and
covered material such as product design, quality, inventory and aggregate produc-
tion planning, line balancing and forecasting, among others. Subsequently, emphasis
was given to quality assurance and materials requirements planning in such texts
as Montgomery (1992), Evans and Lindsay (1996), Orlichy (1975), Waters (1998),
Waters (2006) and Noble (1986). In more recent times, arising from the greater
emphasis on the strategic importance of manufacturing and operations particularly
in MBA executive programs, the texts of Russell and Taylor (2005), Evans (1997),
and Anderson, Sweeney and Williams (1991) were widely read.
An excellent overall text with a strong quantitative orientation is Groover (2001).
Specific texts related to probability and queueing systems include Gross and Harris
(1998), Feller (1961), Feller (1991), and Parzen (1962) as well as classic books in
operations research including Taha (2002), Hillier and Liebermann (2005), Solberg
et al. (1987), and Perros (1994), among others.
There are a large number of texts devoted to the subject of simulation mod-
eling such as Law and Kelton (1999), Pidd (2004) as well as texts devoted to
specific simulation languages such as Arena (Kelton et al., 1998), and eM-Plant
(http://www.plm.automation.siemens.com/en_us/products/
tecnomatix/).
References
1. Archetti, F., Lucertini, M. and Serafini, P. (1989), Operations Research Models in Flexible
Manufacturing Systems, Springer.
2. Anderson, D.R., Sweeney, D.J. and Williams, T.A. (1991), Introduction to Management
Science: Quantitative Approaches to Decision Making, 6th Edition, West Publishing
Company.
3. Brown, S.E. (1996), Strategic Manufacturing for Competitive Advantage, Transforming
Operations from Shop Floor to Strategy, Prentice Hall.
4. Buffa, E.S. (1973), Modern Production Management, John Wiley.
5. Buzacott, J.A. (1989), Flexible models of flexible manufacturing systems, in Operations
Research Models in Flexible Manufacturing Systems, edited by F. Archetti, M. Lucertini
and P. Serafini, Springer.
6. eM-Plant, http://www.ugs.com/products/tecnomatix/plant_design/em_plant.shtml.
7. Evans, J.R. (1997), Production & Operations Management: Quality, Performance and
Value, 5th Edition, West Publishing Company.
8. Evans, J.R. and Lindsay, W.M. (1996), The Management and Control of Quality, 3rd
Edition, West Publishing Company.
9. Feller, W. (1961), An Introduction to Probability Theory and Its Applications, Volume I,
John Wiley.
10. Feller, W. (1991), An Introduction to Probability Theory and Its Applications, Volume II,
6th Edition, John Wiley.
References 23
11. Groover, M.P. (2001), Automation, Production Systems, and Computer Integrated Manu-
facturing, Second Edition, Prentice Hall.
12. Gross, D. and Harris, C.M. (1998), Fundamentals of Queueing Theory, 3rd Edition, John
Wiley.
13. Gunneson, A.O. (1997), Transitioning to Agility; Creating the 21st Century Enterprise,
Addison-Wesley Publications.
14. Hillier, F.S. and Liebermann, G.J. (2005), Introduction to Operations Research,
8th Edition, McGraw-Hill.
15. Kelton, W.D., Sadowski, R.P. and Sadowski, D.A. (1998), Simulation with Arena,
McGraw-Hill.
16. Law, A. and Kelton, W.D. (1999), Simulation Modeling and Analysis, 3rd Edition,
McGraw-Hill.
17. Montgomery, D.C. (1992), Introduction to Statistical Quality Control, Second Edition,
John Wiley.
18. Noble, D.F. (1986), Forces of Production: A Social History of Industrial Automation,
Oxford University Press.
19. Orlichy, J. (1975), Materials Requirements Planning, McGraw-Hill.
20. Ostwald, P.F. and Muñoz, J. (1997), Manufacturing Processes and Systems, Ninth Edition,
Wiley.
21. Papadopoulos, H.T., Heavey, C. and Browne J. (1993), Queueing Theory in Manufactur-
ing Systems Analysis and Design, Chapman & Hall.
22. Parzen, E. (1962), Modern Probability Theory and Its Applications, John Wiley.
23. Perros, H.G. (1994), Queueing Networks with Blocking: Exact and Approximate Solu-
tions, Oxford University Press.
24. Phillips, E.J. (1997), Manufacturing Plant Layout: Fundamentals and Fine Points of
Optimum Facility Design, Society of Manufacturing Engineers (SME).
25. Pidd, M. (2004), Computer Simulation in Management Science, 5th Edition, John Wiley.
26. Pirsig, R.M. (1974), Zen and the Art of Motorcycle Maintenance, William Morrow &
Company and the Bodley Head.
27. Russell, R.S. and Taylor III, B.W. (2005), Operations Management, 4th Edition, Prentice
Hall.
28. Schmenner, R. (1990), Production/Operations Management, Macmillan.
29. Phillips, D.T., Ravindran, A. and Solberg, J.J. (1987), Operations Research, 2nd Edition,
John Wiley.
30. Taha, H.A. (2002), Operations Research: An Introduction, 7th Edition, Prentice Hall.
31. Waters, C.D.J. (1998), A Practical Introduction to Management Science, Addison-Wesley.
32. Waters, C.D.J. (2006), Operations Strategy, Thomson Learning.
33. Wild, R. (1985), Essentials of Production and Operations Management, 2nd Edition,
Holt, Rinehart and Winston.
34. Womack, J., Jones, D. and Roos, D. (1990), The machine That Changed the World, MIT
Press (Rawson Associates, N.Y.).
2
Evaluative Models of Discrete Part Production Lines
The focus here is on discrete part production lines with asynchronous movement
where each part produced is distinct. Production lines processing fluids and other
continuous materials are not considered. From here on, when reference is made to
production lines, discrete part production lines will be understood. In a production or
flow line, all jobs are required to pass through each station in the same sequence once.
These lines are usually associated with scale rather than scope, and a major advantage
of production lines is the associated simple materials handling requirements.
A production line consists of work-stations, materials, human resources, and
inter-work-station storage facilities. Storage facilities have a finite capacity. Ran-
domness is introduced due to random processing times and the random behavior of
work-stations in relation to failure and repair. In terms of classical queueing theory,
production lines would be described as finite buffer tandem queueing systems where
the work-stations are the servers, storage facilities are the buffers or the waiting lines,
and the jobs are the customers.
In Figue 2.1, which depicts a K-work-station production line, W Si , i = 1, 2, . . . , K
represents work-station i and Bi , i = 1, 2, . . . , K denotes the buffer capacity of the
buffer located in front of station W Si . As there are K work-stations, there are K − 1
intermediate buffers. As described in Chapter 1, the goal of evaluative models is to
calculate some performance measures of the system under study. The most usual
performance measure determined is throughput. Each station may consist of a single
perfectly reliable machine or an unreliable machine or a number of identical parallel
reliable or unreliable machines. For notational purposes only, it should be understood
that the word “machines” may cover operators.
In Section 2.1, Markovian analysis of production lines is presented using the
underlying queueing system structure of production lines. It produces an exact anal-
ysis of such lines. In sub-section 2.1.1, a numerical approach is presented for solving
the system of linear equations derived from the Markovian analysis. In sub-section
2.1.2, an algorithm is given for the generation of the conservative matrix A for
the case of an exponential production line with inter-station buffers. In sub-section
2.1.3, a simple merge model of a two-station production line with merge operations
(a non-linear model) is analyzed using exact Markovian methods.

26
B1 WS1 B2 WS2 BK-1 WSK-1 BK WSK
WSi : Work-station i
Bi+1 : Buffer between work-stations WSi

2 Evaluative Models of Discrete Part Production Lines
B1 : Buffer in front of WS1
Fig. 2.1. A K-work-station production line

2.1 Markovian Model 27
However, even for linear production lines with a large number of stations (K > 6)
and reasonable buffer sizes, it is not possible to develop exact numerical results due
to the complexity of the numerical calculations involved. As a result of this restric-
tion, approximate solutions were sought. The decomposition approach is described in
Section 2.2. Essentially, the process involves the decomposition of a large production
line into a number of smaller lines with suitable provision for their inter-connection
so that the behavior of the inter-connected system approximates the behavior of the
original large production line.
Another approximation technique named the expansion method is given in
Section 2.3.
Although the focus up to now has been on prefectly reliable machines at each
station, it should be noted that the Markovian and decomposition methods can
each handle unreliable machines. The aggregation method, which is a different
approximation approach, was specifically developed to analyze transfer lines with
asymptotically reliable stations. The aggregation method is covered in Section 2.4.
Up to this point the models used have been of a serial type, in the reliability
sense, such that if a particular work-station was not operating due to breakdown or
otherwise, the work-stations downstream from that particular work-station would
eventually be starved. Work-stations in parallel were introduced with the result
that the breakdown of a particular work-station would not necessarily lead to the
starvation of stations downstream. The solution of production lines with parallel
machines at each work-station is given in Section 2.5. The exact analysis of a sim-
ple parallel-machine production line consisting of two work-stations is presented in
sub-section 2.5.1. An alternative exact analysis for solving the same two-station pro-
duction line with parallel machines at each station which serves as building block
in decomposing larger lines is given in sub-section 2.5.2. In sub-section 2.5.3 the
approximate solution using decomposition method of large serial production lines
with parallel-machine stations is given.
Simulation is often used when analytical methods prove intractable or are con-
fined to rather simplified assumptions. In the case of production lines, simula-
tion models may be used to assess the results of all approximate models and to
obtain results using distributions for processing, failure, and repair times other than
exponential or phase-type. Such models are explored in Section 2.6.
2.1 Markovian Model
Consider the model as depicted in Figure 2.1. Jobs enter station 1 from buffer B1
of unrestricted capacity according to a Poisson distribution with arrival rate λ . Each
job enters the line at station 1, passes through all stations in order and leaves the K th
station (last) in finished form. All jobs at each station are processed according to a
First-In-First-Out (FIFO) queueing discipline.
28 2 Evaluative Models of Discrete Part Production Lines
The assumptions of the model are summarized below:

• (i) The processing or service times are exponentially or Erlang distributed
random variables with mean rates equal to μi , i = 1, 2, ..., K. In general,
the service rates need not be identical (i.e., μi = μ j for i = j).
• (ii) All buffers between successive stations have finite capacities not necessar-
ily of the same size.
• (iii) Blocking of a station occurs if the downstream buffer is full at the time of
service completion.
• (iv) A station may be assumed to be perfectly reliable or subject to random
failure according to an exponential distribution with mean rates equal to
βi , i = 1, 2, ..., K. In general, the failure rates need not be identical (i.e.,
βi = β j for i = j). However, it is assumed that a failure of a station can
only occur when it is operating, i.e., operational-dependent breakdowns.
• (v) If a station fails, the part which the station was processing remains at the
station, i.e., it is not placed in the preceding buffer.
• (vi) Once a failed station is repaired, it resumes processing at the same phase of
service at which it failed, on the job that was not completed, and as a result
of the memoryless property of the exponential distribution, the remaining
processing time in that phase is exponentially distributed.
• (vii) The repair times are exponentially or Erlang distributed random variables
with mean rates equal to ri , i = 1, 2, ..., K. In general, the repair rates need
not be identical (i.e., ri = r j for i = j).
• (viii) The general rule that deliberate idleness at a station is not allowed applies.
• (ix) A basic assumption is that the first station is never starved and the last
station is never blocked. Although the arrival process is assumed to be
Poisson, it is a necessary assumption of the model that the first station
is never starved. This assumption characterizes the saturated line of the
saturation model. The fact that the last station is never blocked relates to
the storage capacity for final products.
The system under consideration is a two-dimensional stochastic process N(t) =
[N1 (t), N2 (t)]. Both coordinate random variables are integer valued and nonnega-
tive. N1 (t) represents the number of jobs queued up in front of the first station at time
t, and N1 is the expected value of this quantity at equilibrium (the limit of N1 (t), as
t tends to infinity). There is no upper limit for N1 . N2 (t) represents the state of the
sub-network at time t, which consists of stations 1, 2, 3, . . . , K and the intermediate
buffers. In effect, N2 (t) is a vector representing the situation in each station and in
each of the intermediate buffers of the production line at time t. The number of states
in the sub-network equals m, for some finite m. When N1 = 0, the number of states
in the sub-network equals m0 , m0 < m.
The changes in the state of the system are caused by the occurrence of various
events. The occurrence times for all events have negative exponential or Erlang dis-
tributions with strictly positive means. Thus the process is Markovian. Its state-space
is S = {(i, j) : i ≥ 0, 1 ≤ j ≤ m} with the index i specifying the total number of jobs
queued up at the first station. Such customers are called “I-customers.” The index j
determines the state of the sub-network (Unit-II). It is important to note that upon
entering service at station 1, a customer becomes a “II-customer.”
The transition matrix P that describes the model has the following block
tri-diagonal form:

A01 A0

A2 A1 A0

A2 A1 A0
P =
(2.1)
A2 A1 A0
. . .

. . .
and the equilibrium equations π P = 0 can be expressed in matrix-difference form as
πk A0 + πk+1A1 + πk+2 A2 = 0 (2.2)
for k = 0, 1, 2, . . . and
π0 A01 + π1 A2 = 0 (2.3)
for the boundary equations, where,
A0 is an (m × m)-matrix describing the transitions in the sub-network, which
simultaneously produce inputs to the first queue.
A1 is an (m × m)-matrix describing transitions in the sub-network which produce
neither inputs to nor outputs from the first queue assuming that the queue is
not empty.
A2 is an (m × m)-matrix describing transitions in the sub-network which simulta-
neously produce outputs from the first queue, and
A01 is an (m × m)-matrix describing transitions in the sub-network which produce
neither inputs to nor outputs from the first queue, assuming that the queue is
empty.
A Markov chain whose equilibrium equations have the form of equations (2.2) and
(2.3) is known as a Quasi-Birth and Death (QBD) process.
Let P2 be the steady-state probabilities of the sub-network, assuming that the first
queue is never empty. Solving the equations
P2 A = 0
P2 e = 1 (2.4)
where A is the conservative stable matrix given by
A = A0 + A1 + A2 (2.5)
and e is an (m × 1) column-vector, with all elements equal to 1, will give explicit

results for P2 ( j), j = 1, 2, . . . , m.
The equilibrium condition is given by
P2 A2 e > P2 A0 e. (2.6)
Table 2.1. Notation

Symbol Meaning
K Number of stations
Bi Buffer capacity preceding the ith station.
Note: when Bi = B j for all i, then the
buffer capacity is denoted by B
ni Status of buffer i
si Status of station i
Pi Denotes the number of phases of the service
(processing) distribution of the ith station
Ri Denotes the number of phases of the repair
distribution of the ith station
mB,R
K,P Number of states in the sub-network with K
stations, each buffer having the same capacity B,
each service distribution having P phases
and each repair distribution having R phases
2 ,..,BK ,R1 ,R2 ,..,RK
mBK,P 1 ,P2 ,..,PK
Number of states in the sub-network of a K
station system with buffer capacities B2 ,
. . . ,BK . The number of phases of each
station’s service distribution is equal to
P1 , P2 , . . . , PK phases and the number of
phases of each station’s repair distribution
is equal to R1 , R2 , . . . , RK
From this relationship the critical mean input rate (λ ∗ ) to the system can be deter-
mined. In the steady-state, this critical input rate is identical to the maximum
throughput rate of the production line. By calculating the throughput of the sys-
tem as outlined above, we exclude the states of the system where the first station is
empty, i.e., sub-matrix A01 is not included. Therefore, the throughput of the system
is governed by the assumption that the first queue is never empty (saturation model).
The notation used is shown in Table 2.1.
The states of the sub-network are described by the following vector:
(s1 , n2 , s2 , n3 , . . . nK , sK ) (2.7)
where, si can take any value from 0 to (Pi + Ri × Pi ):
si = 0 – station is idle,
si = 1, . . . , Pi – station is in service,
si = (Pi + 1), . . . , (Pi + Pi × Ri ) – station is in repair.
After a station has been repaired, it is assumed that service is resumed at the
phase in which the station was interrupted. Therefore, there is a need to keep a record
of the phase of service in which the station was interrupted. It should be noted that
the need to account for the phase is a modeling requirement and may not have any
Repair
0 1 2 3 4 5 6 7 8
Idle Service Interrupted in Interrupted in

phase 1 of service phase 2 of service
Total number of states = Pi(1 + Ri) = 8
Fig. 2.2. The states of si for Pi = 2, Ri = 3
corresponding physical meaning. This results in using (Pi + Pi × Ri ) states to describe

a station’s repair process. It also necessitates the use of equation (2.8) to transfer
si from a state in service to a state in repair and equation (2.9) to do the reverse.
Beginning repair state = (Pi + ((si − 1) × Ri) + 1) (2.8)

(si − Pi )
Phase to resume service at = . (2.9)
Ri
The states that si can take, for the parameters Pi = 2 and Ri = 3, are illustrated in
Figure 2.2.
ni can take values from 0 to (Bi +1). The values from 0 to Bi denote the number of
items in buffer Bi with Bi also denoting the capacity of buffer Bi . When ni = (Bi + 1),
station (i − 1) is blocked.
The following recursive relationship was obtained (in Heavey, Papadopoulos and
Browne, 1993) to calculate the number of, states of a system with K stations with
parameters Pi , Ri , B j , i = 1, . . . , K, j = 2, . . . , K.
The number of states for a two-station system is first calculated using the
parameters PK−1 , PK , RK−1 , RK , BK .
Two-station system:
Ξ1 = (PK + (PK × RK ) + 1) (2.10)

Ξ2 = (((PK−1 × (BK + 1)) + (PK−1 × RK−1 × (BK + 1)))
+1) (2.11)
Ξ3 = ((BK × (PK−1 + (PK−1 × RK−1))) + 1) (2.12)
Ω2 = ((Ξ1 × Ξ2) − Ξ3). (2.13)
Ω2 will equal the number of states of the system if K = 2. To calculate the number of
states for systems with K > 2, the following recursive scheme is used. Before enter-
ing the loop below, the variable Ω1 is set equal to (PK + PK × RK ) and the variable
Table 2.2. Number of states for P = 1, R = 1 and identical buffer capacities
# of Stations Buffer Size

0 1 2 3 4
2 8 12 16 20 24
3 30 70 126 198 286
4 112 408 992 1,960 3,408
5 418 2,378 7,810 19,402 40,610
6 1,560 13,860 61,488 192,060 483,912
7 5,822 80,782 484,094 1,901,198 5,766,334
8 21,728 470,832 3,811,264 18,819,920 68,712,096
9 81,090 2,744,210 30,006,018 188,119,920 818,778,818
Ω2 takes its value from equation (2.13), with the parameters PK−1 , PK , RK−1 , RK , BK
of the K-station system used in equations (2.10), (2.11) and (2.12).
DO I = (K − 1) to 2, −1
Y1 = ((Ω2 − Ω1 )/(PI × (RI + 1))) + Ω1 (2.14)
Y2 := Using parameters NI , PI , PI−1 ,
RI , RI−1 calculate the number of states
for a two-station system as above, i.e., let
BK = NI , PK−1 = PI−1 , PK = PI ,
RK−1 = RI−1 , RK = RI in equations
(2.10), (2.11), (2.12) above. (2.15)
Y31 = (PI−1 + PI−1 × RI−1 )
Y32 = (PI + PI × RI )
Y33 = (BI × (Y31 × (Y32 − 1)))
Y3 = ((Y31 × Y32 ) + (Y32 − 1) + Y33) × Ω1 (2.16)
Ω1 = Ω2
Ω2 = ((Y1 × Y2 ) − Y3 )
END DO I
2 ,..,BK ,R1 ,R2 ,..,RK

mBK,P 1 ,..,PK
= Ω2 (2.17)
Table 2.2 gives the number of states for K-station systems, with identical buffers
equal to B, P = 1 and R = 1. As one can observe from Table 2.2, the number of states
becomes very large, even for relatively small systems.
2.1.1 A numerical approach
As is well known, there are a number of ways of solving sets of homogeneous lin-
ear equations. To name a few, Gaussian elimination, iterative methods based on
Gauss-Siedel approximation, Jacobian elimination, and matrix recursive methods.

In the solution of such sets of homogeneous linear equations, the analyst is primarily
concerned with efficiency of calculations and rapidity of convergence and estima-
tion of the degree of approximation, if appropriate. Clearly, because of the number
of states, in any realistic model of a production line, there is a need for an efficient
algorithm to determine the steady-state probabilities associated with the states of the
system.
The algorithm outlined below is based on Gaussian elimination with a dynam-
ically adjusted successive over-relaxation factor to achieve rapid convergence. The
essential components of this algorithm are
• Ordering of the states
• Generation of the transition matrix
• Solution of the resulting system of linear equations
This algorithm was coded in C++ by Dr. Cathal Heavey and is based on the work
of Heavey, Papadopoulos and Browne (1993), Papadopoulos, Heavey and O’Kelly
(1989, 1990), Papadopoulos and O’Kelly (1989) and Papadopoulos (1989) and with
appropriate instructions is available at the website associated with this book with
the abbreviated name MARKOV. The user inputs into the algorithm the following
parameters: K the number of stations; B2 , . . . , BK the buffer capacities; P1 , . . . , PK the
number of phases of the service distribution for each station; R1 , . . . , RK , the number
of phases of the repair distribution for each station; μi the mean service rates; ri the
mean repair rates; and βi the mean breakdown rates. Thus, a very general algorithm
has been developed which generates the transition probability matrix, A, and then
solves the set of linear equations via the use of the SOR method and gives as output
the throughput, XK , of a K-station production line with finite intermediate buffers and
with the service and repair times following a phase-type distribution and the times to
failure being exponentially distributed.
The ordering of the states affects the structure of the conservative matrix A. The
objective is to find an ordering of the states such that matrix A will have as simple
a structure as possible from a computational point of view. This will facilitate the
development of a very efficient algorithm for the generation of matrix A. To select
an appropriate ordering of the states, a criterion for the structure of matrix A must
be selected. In the algortithm included at the website associated with this book, the
criterion used was to keep the non-zero elements of the conservative matrix A as close
as possible to the diagonal elements, i.e., a quasi band diagonal matrix. Because of
the increasing number of states, as system complexity increases, it is not possible to
assess how close matrix A is to a strict band diagonal matrix.
A recursive algorithm for generating the conservative matrix A has been devel-
oped based on the generation of a series of sub-matrices (Heavey, Papadopoulos
and Browne, 1993). Specific details of the matrix generation process for the case
of a reliable exponential production line with inter-station buffers (Papadopoulos,
Heavey and O’Kelly, 1989) are given in sub-section 2.1.2.
Table 2.3. Exponential service, repair, and failure, K = 3

B2 = 2, B3 = 4
μ1 = 1.5, μ2 = 2.0, μ3 = 1.9
r1 = 0.1, r2 = 0.02, r3 = 0.15
β1 = 0.02, β2 = 0.01, β3 = 0.09
Analytical Results Simulation Results
95% CI
X3 = 0.7346 0.721 – 0.737 – 0.752
B2 = 5, B3 = 3
μ1 = 2.6, μ2 = 3.0, μ3 = 3.2
r1 = 0.5, r2 = 0.03, r3 = 0.15
β1 = 0.03, β2 = 0.01, β3 = 0.02
95% CI
X3 = 1.2985 1.26 – 1.28 – 1.31
An iterative method was used to solve the system of linear equations. The iter-
ative method used was the Successive Over Relaxation (SOR) method. SOR is
more efficient than the Gauss-Seidel method, but SOR has one main drawback, the
unknown optimal value of the relaxation factor. A process of dynamically adjusting
the relaxation factor has been introduced into the algorithm, which worked well in
practice.
The results of the algorithm have been compared with available analytical results
(systems with a small number of states) and simulation studies on systems with
relatively large number of states and has been found to be satisfactory.
A sample of the throughput rate, XK , from the analytical model, compared with
results from a simulation model are given below. Two arbitrary examples are given
for systems with: (1) Exponential service, repair, and failure (see Table 2.3); (2)
Erlang service, exponential repair, and failure (see Table 2.4); (3) Erlang service,
repair, and exponential failure (see Table 2.5). μi , ri , βi , i = 1, 2, 3 are the mean service
rates, repair rates, and failure rates, respectively.
As can be seen from Tables 2.3, 2.4 and 2.5, the point estimates of the throughput
from the simulation model are very close to the results from the analytical model
and all the analytical results are covered by the 95% confidence intervals (CI). These
results are typical of all the models tested against simulation. Therefore, it can be
safely concluded that the analytical model yields the correct results.
In sub-section 2.1.2, the detailed development of the conservative matrix A for
the reliable exponential case is developed. Details of more general cases (phase-type
distribution of service and repair times) are available in the literature listed at the end
of this chapter.
Table 2.4. Erlang service, exponential repair, and failure, K = 4

P1 = 3, P2 = 2, P3 = 3, P4 = 2
B2 = 3, B3 = 2, B4 = 3
μ1 = 5.0, μ2 = 4.5, μ3 = 5.2, μ4 = 3.7
r1 = 0.05, r2 = 0.03, r3 = 0.07, r4 = 0.1
β1 = 0.02, β2 = 0.001, β3 = 0.003, β4 = 0.05
95% CI
X4 = 2.0025 1.99 – 2.04 – 2.10
P1 = 2, P2 = 3, P3 = 3, P4 = 2
B2 = 3, B3 = 5, B4 = 3
μ1 = 1.5, μ2 = 0.9, μ3 = 0.9, μ4 = 1.5
r1 = 0.1, r2 = 0.03, r3 = 0.2, r4 = 0.3
β1 = 0.02, β2 = 0.01, β3 = 0.09, β4 = 0.2
95% CI
X4 = 0.4591 0.450 – 0.456 – 0.462
Table 2.5. Erlang service, repair, and exponential failure, K = 4
P1 = 3, P2 = 2, P3 = 3, P4 = 2
R1 = 2, R2 = 3, R3 = 4, R4 = 2
B2 = 2, B3 = 1, B4 = 3
μ1 = 2.5, μ2 = 1.9, μ3 = 2.6, μ4 = 3.0
r1 = 0.05, r2 = 0.03, r3 = 0.07, r4 = 0.1
β1 = 0.02, β2 = 0.001, β3 = 0.003, β4 = 0.05
95% CI
X4 = 1.1325 1.10 – 1.12 – 1.14
P1 = 2, P2 = 3, P3 = 3, P4 = 2
R1 = 3, R2 = 2, R3 = 3, R4 = 2
B2 = 1, B3 = 2, B4 = 3
μ1 = 5.0, μ2 = 4.5, μ3 = 5.2, μ4 = 3.7
r1 = 0.1, r2 = 0.03, r3 = 0.2, r4 = 0.3
β1 = 0.02, β2 = 0.01, β3 = 0.09, β4 = 0.2
95% CI
X4 = 1.5958 1.53 – 1.54 – 1.60
Table 2.6. Notation

Symbol Meaning
K Number of stations.
Bi Buffer capacity preceding the ith station. Note: when Bi = B j for all i, then the
buffer capacity is denoted by B.
ni Status of buffer i.
si Status of station i (see Table 2.7).
mBK Number of states in the sub-network of a K station system with identical
buffers, each of capacity B.
mBK2 ,...,BK Number of states in the sub-network of a K station system with non-identical
buffers, with buffer capacities B2 , . . . ,BK .
Table 2.7. States of station i

si Meaning
0 Station is idle.
1 Station is busy.
2 Station is busy and blocking
preceding station.
2.1.2 The algorithm for the generation of the conservative matrix A for the
reliable exponential production lines with inter-station buffers
For purposes of illustration, in this sub-section, the recursive algorithm will be

applied to the case of a reliable exponential production line only (see Papadopoulos,
Heavey and O’Kelly, 1989). Table 2.6 lists the notation used in this sub-section.
The algorithm for generating the conservative matrix A is divided into two parts.
The first part generates sub-matrix Y1∗K=k for the appropriate system. The second
part generates sub-matrix Y2∗K=k from sub-matrix Y1∗K=k , and the non-diagonal ele-
ments P1 μ1 , R1 r1 and β1 . In the first part of the algorithm, the non-zero elements of
Y1∗K=k , the column coefficients, and the number of elements in each row are stored
in separate one dimensional arrays.
The second part of the algorithm is executed during the execution of the solution
procedure. As a consequence, the non-zero elements of sub-matrix Y2∗K=k , and the
non-diagonal elements P1 μ1 , R1 r1 and β1 need not be stored in memory. This can
greatly reduce the amount of memory required to solve a system.
The states of the sub-network are described by the following vector:
(n2 , s2 , n3 , s3 , . . . , nK , sK ) (2.18)
s1 is not included in the state vector because it is always equal to 1, i.e., the first
station is never idle. This does not mean that the first station cannot be blocked by
buffer B2 or work-station W S2 . si can take any of the values listed in Table 2.7, and
ni can take any value from 0 to Bi , as it denotes the number of items in buffer i.
The set of linear equations for the solution of P2 , the marginal p.d.f. for the sub-
network, can be written in the following two ways.
P2 A = 0 (2.19)
AT P2 = 0. (2.20)
In the rest of this sub-section, AT is examined. This is because in order to gener-
ate matrix A efficiently, the relationship between its columns (rows of AT ) needs to
be examined. In order to simplify the notation, A denotes AT in the rest of this
sub-section.
Number of States
A prerequisite to the development of the algorithm is the derivation of an equation

to calculate the number of states in the sub-network. The case where buffers are
identical is investigated first and then the case of buffers being non-identical.
Identical Buffers
For this case, where buffers are of equal capacity, say B = N, the following difference
equation is obtained, in a way analogous to that used for the case where buffers were
not allowed (see Papadopoulos, 1989 and Papadopoulos and O’Kelly, 1989):
mNK+2 − (N + 3) mNK+1 + mNK = 0. (2.21)
Then, its characteristic equation is
x2 − (N + 3)x + 1 = 0,
with two real roots:

(N + 3) + (N + 3)2 − 4 (N + 3) − (N + 3)2 − 4
x1 = , x2 = .
2 2
Therefore the general solution of (2.21) is
mN K K
K = c1 x1 + c2 x2
K K
(N + 3) + (N + 3)2 − 4 (N + 3) − (N + 3)2 − 4
= c1 + c2 .
2 2
The initial conditions: m0 = 0 and m1 = 1 give
c1 + c2 = 0,
and

(N + 3) + (N + 3)2 − 4 (N + 3) − (N + 3)2 − 4
c1 + c2 = 1.
2 2
Hence,
1(N + 3)2 − 4
c1 = −c2 = = ,
(N + 3)2 − 4 (N + 3)2 − 4
and the general solution becomes
⎧ K ⎫
⎨ (N + 3) + (N + 3)2 − 4 K (N + 3) −

(N + 3)2−4 ⎬
mNK = −
⎩ 2 2 ⎭

1
. (2.22)
(N + 3)2 − 4
Equation (2.22) was used to calculate the number of states for the systems in
Table 2.8. It is clear from Table 2.8 that the number of states increases tremendously
with an increase in the size of the buffer and in the number of stations. This places
strict limits on the size of the system for which exact results can be obtained.
Non-identical Buffers
For this case, where buffers are of unequal capacity, say B2 , B3 , . . . , BK , the differ-
ence equation may be shown to be similar to that obtained for Case 1, where buffers
were of equal capacity (equation (2.21)), i.e.,
B ,B ,...,BK+2 B ,B ,...,BK+1 B ,B3 ,...,BK
2 3
mK+2 = (BK+2 + 3) mK+1
2 3
− mK2 . (2.23)
Applying the initial conditions m0 = 0 and m1 = 1 to equation (2.23), for K = 0, 1, . . .,

sequentially,
(1) K = 0:
mB2 2 = (B2 + 3) m1 − m0
= (B2 + 3) (1) − 0
= B2 + 3. (2.24)
(2) K = 1: Combination of equations (2.23) and (2.24) gives

B ,B3
m3 2 = (B3 + 3) mB2 2 − m1
= (B2 + 3) (B3 + 3) − 1. (2.25)
(3) K = 2: Combination of equations (2.23), (2.24) and (2.25) gives
mB4 2 ,B3 ,B4 = (B4 + 3) mB3 2 ,B3 − mB2 2

= (B4 + 3) [(B3 + 3) (B2 + 3) − 1]− (B2 + 3)
= (B2 + 3) [(B3 + 3) (B4 + 3) − 1]− (B4 + 3). (2.26)
Table 2.8. Number of states of the system
# of stations buffer size

0 1 2 3 4 5 6 7 8 9 10
1 1 1 1 1 1 1 1 1 1 1 1
2 3 4 5 6 7 8 9 10 11 12 13
3 8 15 24 35 48 63 80 99 120 143 168
4 21 56 115 204 329 496 711 980 1,309 1,704 2,171
5 55 209 551 1,189 2,255 3,905 6,319 9,701 14,279 20,305 28,055
6 144 780 2,640 6,930 15,456 30,744 56,160 96,030 155,760 241,956 362,544
7 377 2,911 12,649 40,391 105,937 242,047 449,121 950,559
8 987 10,864 60,604 235,416 726,103
9 2,584 40,545 290,376
10 6,765 151,316
11 17,711 564,719
12 46,368
13 121,393
14 317,811
2.1 Markovian Model
39

B ,B3 ,B4 ,B5
m5 2 = (B5 + 3) mB4 2,B3 ,B4 − mB3 2 ,B3
= (B5 + 3) {(B4 + 3) [(B3 + 3) (B2 + 3) − 1]− (B2 + 3)}
− [(B3 + 3) (B2 + 3) − 1]
= [(B2 + 3) (B3 + 3) − 1] [(B4 + 3) (B5 + 3) − 1]
− (B2 + 3) (B5 + 3). (2.27)

B ,...,B6 B ,...,B5
m6 2 = (B6 + 3) m5 2 − mB4 2 ,...,B4
= (B6 + 3) {[(B2 + 3) (B3 + 3) − 1] [(B2 + 3) (B3 + 3) − 1]
− (B2 + 3) (B5 + 3)}
− {(B2 + 3) [(B3 + 3) (B4 + 3) − 1] − (B5 + 3)} ,
and after some algebra,

B ,...,B6
m6 2 = (B4 + 3) [(B2 + 3) (B3 + 3) − 1] [(B5 + 3) (B6 + 3) − 1]
− (B6 + 3) [(B2 + 3) (B3 + 3) − 1]
− (B2 + 3) [(B5 + 3) (B6 + 3) − 1]. (2.28)
The examples illustrated above suggest the following iterative scheme to calcu-
late the number of states of a system with non-identical buffers, i.e., a system with K
stations and buffer capacities B2 , B3 . . . , BK .
Initial Values:
V 1 = 1 = m1
V 2 = 0 = m0
DO J = 2 to K
V = (BJ + 3)V 1 − V 2
V2 = V1
V1 = V
END DO J
mBK2 ,...,BK = V .
The iterative scheme above calculates the number of states of a K station system
B ,B
with buffer capacities B2 , B3 , . . . , BK , by first calculating mB2 2 and then m3 2 3 , i.e., by
using Bi in the following order, i = 2, 3, . . . , K − 1, K. It is interesting to note that the
number of states for a system with non-identical buffers can also be calculated using
B ,B
Bi in the reverse order, i = K, K − 1, . . . , 2, i.e., calculate mB2 K first, then m3 K−1 K and
so on. In the algorithm for the generation of the transition matrix, the latter method
is used.
Ordering of States
Each state is represented by the following vector:
(n2 , s2 , n3 , . . . nK , sK ). (2.29)
Each state is altered by the following rule:

If si equals 2 and i > 2 then
saltered
i−1 = (si−1 − 1).
Then the ‘altered states’ are given a unique numerical value in order to ensure a 1 − 1
correspondence, as follows:
n2 × LE−1 + saltered
2 × LE−2 + · · ·+ nK × LE−(E−1) + saltered
K × LE−E = numerical value
with E equal to the number of elements in the state vector and L given an appropriate
integer value as follows:
L > MAX{B j , 2}, j = 2, . . . , K.
L is the base for the numerical values of the states. The numerical values of the
‘altered states’ are then ordered in increasing value and the states ordered according
to this.
The above procedure will be illustrated with an example. Table 2.9 lists the
states, the ‘altered states’, and the numerical values of the ‘altered states’ for K = 3,
B2 = 0, B3 = 1. E equals 4 and L equals 3.
Only states (0,1,1,2) and (0,2,1,2) were altered (see Table 2.9). Table 2.10 gives
the numerical values of the ‘altered states’ ordered in increasing value and the states
ordered according to this ordering.
The reason for ordering the states is to give matrix A a relatively simple structure
which can be exploited when developing the algorithm to generate matrix A. Matrix
Table 2.9. Altered states and their numerical values

States Altered Numerical
States Value
(0,0,0,0) (0,0,0,0) 0
(0,0,0,1) (0,0,0,1) 1
(0,1,0,0) (0,1,0,0) 9
(0,1,0,1) (0,1,0,1) 10
(0,2,0,0) (0,2,0,0) 18
(0,2,0,1) (0,2,0,1) 19
(0,0,1,1) (0,0,1,1) 4
(0,1,1,1) (0,1,1,1) 13
(0,1,1,2) (0,0,1,2) 5
(0,2,1,1) (0,2,1,1) 22
(0,2,1,2) (0,1,1,2) 14
Table 2.10. Ordering of states
Ordered Numerical
States Value
(0,0,0,0) 0
(0,0,0,1) 1
(0,0,1,1) 4
(0,1,1,2) 5
(0,1,0,0) 9
(0,1,0,1) 10
(0,1,1,1) 13
(0,2,1,2) 14
(0,2,0,0) 18
(0,2,0,1) 19
(0,2,1,1) 22
(2.30) gives matrix A for K = 3, B2 = 0, B3 = 1, with the states ordered according

to Table 2.10. Note that each of the non-diagonal elements, μ3 , μ2 and μ1 , in matrix
(2.30) are always found in the same position relative to the diagonal element, i.e., μ2
is always two columns to the right of the diagonal element.

− μ1 μ3 0 0 0 0 0 0 0 0 0
0 − μ1 − μ3 μ3 μ2
0 0 0 0 0 0 0
0 0 − μ1 − μ3 μ3 0 μ2 0 0 0 0 0
0 − μ1 − μ3 μ2

0 0 0 0 0 0 0 0
μ1 0 0 0 − μ1 − μ2 μ3 0 0 0 0 0
0
A= μ1 0 0 0 − ∑i=1 μi
3
μ3 0 μ2 0 0
0 μ − ∑ 3
μ μ μ
0 1 0 0 0 i=1 i 3 0 2 0
0 0 0 μ1 0 0 0 − μ3 0 0 μ2
0 μ1 0 − μ2 μ3
0 0 0 0 0 0
0 0 0 0 0 μ1 0 0 0 − μ2 − μ3 μ3
0 0 0 0 0 0 μ 0 0 0 −μ − μ

1 2 3
(2.30)
Structure of Matrix A
Matrix A equals the summation of sub-matrices A1 , A2 and A0 . Sub-matrices A0 and

A2 have very simple structures whereas sub-matrix A1 has a relatively complicated
structure. Sub-matrix A1 is examined first and then A0 and A2 .
Description of A1
Matrix A1 for any value K (K > 2) with identical or non-identical buffers was found
to take the form described in Figure 2.3.
C, D, and D∗ are all
B3 ,B4 ,...,BK B3 ,B4 ,...,BK
mK−1 × mK−1
matrices. E and F are

3 ,B4 ,...,BK 4 ,...,BK 3 ,B4 ,...,BK 4 ,...,BK
mBK−1 − mBK−2 × mBK−1 − mBK−2
C F
1 D
F
2 D F
N D F
D* F
Fig. 2.3. Structure of A1 , K > 2, B2 = N, B3 , B4 , . . . , BK
D*
X
F
Fig. 2.4. Relationship of sub-matrix E to D∗
matrices. The number of times sub-matrices D and F appear between sub-matrices

C and D∗ equals B2 = N. The relationships between the sub-matrices are as follows:
1. Sub-matrix C for a K station system with B2 = N, B3 , B4 , . . . , BK is generated from
A1 for K − 1 station system with B3 , B4 , . . . , BK , by: (i) Substituting μi+1 for μi ,
i = K, K − 1, . . . 2 (i.e., backwards) in (A1 )K−1 ; (ii) Subtracting μ1 from the last
4 ,...,BK
mBK−2 diagonal elements of (A1 )K−1 .

B3 ,B4 ,...,BK 4 ,...,BK
2. (i) D is generated from C by subtracting μ2 from the first mK−1 −mBK−2
diagonal elements of C.
(ii) If B2 = 0, then there is no sub-matrix D. Therefore D∗ is generated from C. If
4 ,...,BK
B2 = 0, then μ1 is also added to the last mBK−2 diagonal elements of C.
4 ,...,BK
3. D∗ is generated from D by adding μ1 to the last mBK−2 diagonal elements of D.
If B2 =
0, this relationship does not
hold because there will be no sub-matrix D.
B3 ,B4 ,...,BK B4 ,...,BK B3 ,B4 ,...,BK B4 ,...,BK
4. E is a mK−1 − mK−2 × mK−1 − mK−2 matrix which is gen-
∗
erated from X, a sub-matrix of D (see Figure 2.4), by adding μ1 to all the
diagonal elements of X.

3 ,B4 ,...,BK 4 ,...,BK
5. F is a square diagonal matrix of order mBK−1 − mBK−2 with μ2 in the

4 ,...,BK
diagonal elements. The first sub-matrix F is positioned on the mBK−2 +1

B3 ,B4 ,...,BK
row and the mK−1 + 1 column of matrix A1 . Its position relative to D and
D∗ is the same as its position relative to C.
Once A1 for K = 2, BK is obtained, using the relationships outlined above, A1 for any
value K, B2 = N, B3 , B4 , . . . , BK can be generated. A1 for K = 2 and any value BK is
easy to generate.
Description of A0 and A2

In general A0 is a mBK2 ,B3 ,...,BK × mBK2,B3 ,...,BK matrix with λ in all the diagonal
elements and μK θ in exactly
the same positions as μK θ is in A1 .
B ,B ,...,BK B ,B ,...,BK
In general A2 is a mK2 3 × mK2 3 matrix with μ1 in the Ith column
B B
and the mK−1 + I row with I = 1, 2, . . . , mK − mBK−1 .
Therefore the basic structure of A = A0 + A1 + A2 is given by the structure of
sub-matrix A1 except:
1. A does not contain any λ , i.e., λ in the diagonal elements of A0 cancels − λ in the
diagonal elements of A1 .
2. Instead of μK θ in A1 , there is a μK in A, i.e., (θ + θ ) = 1. This is because μK θ
in A0 is in exactly the same position as μK θ is in A1 .
3. The inclusion of the sub-matrix A2 .
Figure 2.5 gives the structure of sub-matrix A for a K station system (K > 2). Sub-
matrices C, D, D∗ , and E are as described in sub-section 2.1.2 except the changes
outlined above, i.e., their diagonal elements do not contain any λ and μK θ → μK .
Sub-matrices G and H contain the μ1 elements of A2 . G is a square matrix of
3 ,B4 ,...,BK
order mBK−1 with μ1 in the diagonal elements. H is a square matrix of order

B3 ,B4 ,...,BK B4 ,...,BK
mK−1 − mK−2 with μ1 in the diagonal elements.
Algorithm to Generate Matrix A
The following is a description of the algorithm to generate A, which was coded in

C++. The user inputs K the number of stations, B2 , B3 , . . . , BK the buffer capacities,
the mean service rates μ1 , μ2 , . . . μK , and θ the feedback probability.
RULE 1. (i) The first element (row=1,column=1) is equal
to −μ1 and element (row=1,column=2) equals μK .
(ii) This part generates the next (BK + 2) rows.
The next (BK + 2) diagonal elements (i, i) are put equal to
−μ1 − μK and the value μK is placed in element (i, i + 1),
for the (BK + 2) rows, except the last row.
C F
G D F 1
G D F 2
G D F N
G D* F
H E
Fig. 2.5. Structure of A, K > 2, B2 = N, B3 , B4 , . . . , BK
C F
Fig. 2.6. Illustration of Rule 2
(iii) If K=2, then μ1 is added to the last

diagonal element created above. This is matrix A for
K = 2, go to Rule 6.
Rule 1 will create A if K = 2. If K > 2 it will create sub-matrix C for K = 3. Rules

2, 3, 4, and 5, below, are all contained within a loop (see ‘DO T = 3 to K’ below).
‘END DO T ’ denotes the end of the loop. If K = 3, the first iteration of the loop will
create A for K = 3, BK−1 , BK , if not, then sub-matrix C for K = 4, BK−2 , BK−1 , BK is
created and so on until A for K = K is created.
DO T = 3 to K
X = (T − 2)
Y = (X + 1)
W = (K − X )
RULE 2. Place
the top left element of a square
matrix of
B ,...,BK B ,...,BK
order mY W +1 with μW in
− mY W+2
−1

B ,...,BK
its diagonal elements, in the mY W+2 + 1
−1
B ,...,BK
row and the mY W +1 + 1 column of A.
This is sub-matrix F and its position relative to C is
illustrated in Figure 2.6.
C F
D F
1 D F
2 D F
BW D F
D* F
K B ,...,B
RULE 3. C is a square matrix of order mY W+1 . D is
generated
from C by subtracting
μW from the
B ,...,B B ,...,B
first mY W +1 K
− mY W+2
−1
K
diagonal elements
of C. D is positioned as in Figure 2.7. Also, F is
copied onto F (see Figure 2.7). If B = 0 and T = K
B ,...,BK
then μ1 is also added to the last mY W−1+2 diagonal
elements of C, i.e., C will be copied onto D∗ .
Rule 4 is contained within a loop (‘DO Z = 1 to BW ’), and it is executed BW

times. If BW = 0, this rule is not used.
RULE 4. DO Z = 1 to BW
Copy Sub-matrices D and F as described in Figure 2.8.
When T = K and Z = BW , μ1 is added to
B ,...,BK
the last mY W−1+2 diagonal elements of D,
i.e., sub-matrix D is copied onto D∗ .
END DO Z
D*
X
F
RULE 5. The top left of the square sub-matrix

D∗ of dimension
BW +1 ,...,BK BW+2 ,...,BK
mY − mY −1 is copied
on to E, see Figure 2.9 i.e., X is copied onto E.
The position of E is also illustrated in Figure 2.9.
If T = K, μ1 is added to all the diagonal elements of E.
END DO T
Rule 6 below generates the non-zero elements of A2 and is executed after exiting
‘DO T = 3 to K’.

B3 ,...,BK
RULE 6. DO I = 1 to mBK2 ,...,BK − mK−1

3 ,...,BK
Place μ1 in row mBK−1 + I and column I.
END DO I
Application of the Algorithm

Here, the explicit derivation of the conservative matrix A for K = 3, B2 = 1, B3 =
0, with exponentially distributed processing times with mean values μ1i , i = 1, 2, 3,
is developed by applying the algorithm described above.
Rule 1
Applying Rule 1, matrix (2.31) is obtained. Since K > 2, this is matrix C for K = 3.

− λ − μ1 μ3 0

C= 0 − λ − μ1 − μ3 μ3 . (2.31)

0 0 − λ − μ1 − μ3
Rule 1(i) generated the first row of C. Rule 1(ii) generated the next 2 = (B3 + 2)
row(s): note that in the last row generated by Rule 1(ii), μ3 is not placed in the col-
umn next to the diagonal.
Rules 2, 3, 4, 5 are all contained within a loop which is executed (K − 3 + 1)

times. Therefore, for the example illustrated here, only one iteration of the loop is
performed, with T = 3, X = 1,Y = 2 and W = 2.
Rule 2

Rule 2 will generate a square
matrix (sub-matrix F) of order 2 = m02 − m1 =
B ,...,BK B ,...,BK
mY W +1 − mY W+2
−1 with μ2 = μW in the diagonal elements. The top left
B ,...,B
element of F is positioned in the 2nd = (m1 + 1) = (mY W−1+2 K
+ 1) row and the
BW+1 ,...,BK
4 = (m2 + 1) = (mY
th 0 + 1) column of A. The first 3 rows of matrix A are
given in matrix (2.32).

− μ1 μ3 0 0 0 0 0 0 0 0 0

0 − μ1 − μ3 μ3 μ2 0 0 0 0 0 0 0 . (2.32)

0 0 − μ1 − μ3 0 μ2 0 0 0 0 0 0
Rule 3
Rule 3 generates sub-matrix D from sub-matrix C. D is generated from C by subtract-

B ,...,BK B ,...,BK
ing μ2 = μW from the first 2 = (m02 − m1 ) = (mY W +1 − mY W+2
−1 ) diagonal
elements of C. Rule 3 also copies sub-matrix F. The positions of sub-matrices D
and F are illustrated in Figure 2.7. Excluding the non-diagonal μ1 elements that are
generated by Rule 6, matrix (2.33) gives the first 6 rows of A for K = 3, B2 = 1, B3 = 0

− μ1 μ3 0 0 0 0 0 0 0 0 0

0 − μ1 − μ3 μ3 μ2 0 0 0 0 0 0 0

0 0 − μ1 − μ3 0 μ2 0 0 0 0 0 0

0 0 0 − μ1 − μ2 μ3 0 0 0 0 0 0

0 0 0 0 − μ1 − μ2 − μ3 μ3 μ2 0 0 0 0

0 0 0 0 0 − μ1 − μ3 0 μ2 0 0 0
(2.33)
Rule 4
Rule 4 copies sub-matrix D BW = 1 times. Because T = K and Z = BW , μ1 is added

to the last m1 = 1 diagonal elements of D, i.e., sub-matrix D is copied onto D∗ (see
Figure 2.8). Matrix (2.34) gives the first nine rows of A (excluding the non-diagonal
elements generated by Rule 6), which have been generated by rules 1, 2, 3 and 4.

−μ1 μ3 0 0 0 0 0 0 0 0 0

0 −μ1 − μ3 μ μ 0 0 0 0 0 0 0
3 2
0 0 − μ − μ 0 μ 0 0 0 0 0 0
1 3 2
0 0 0 −μ1 − μ2 μ3 0 0 0 0 0 0

0 0 0 0 − ∑ 3 μ μ μ 0 0 0 0
i=1 i 3 2
0 0 0 0 0 −μ1 − μ3 0 μ2 0 0 0

0 0 0 0 0 0 −μ1 − μ2 μ3 0 0 0

0 − ∑3i=1 μi μ3 μ2 0
0 0 0 0 0 0
0 0 0 0 0 0 0 0 −μ3 0 μ2
(2.34)
Rule 5
B ,...,B B ,...,B
E is a square matrix of order 2 = (m02 − m1 ) = (mY W+1 K
− mY W+2
−1
K
) equal to
∗
the top left of sub-matrix D of the said dimension (see Figure 2.9). Since T = K, μ1
is added to all the diagonal elements of E. The position of E is illustrated in Figure
2.9, the top left element of E is positioned on the 10th row and the 10th column of A
(see matrix (2.35)).
Rule 6
The following loop generates the elements of A2 (the non-diagonal μ1 elements of

B2 ,...,BK B3 ,...,BK B3 ,...,BK
A), 8 = (m1,0
3 − m2 ) = (mK
0 − mK−1 ) and 3 = m02 = mK−1 .
DO I = 1 to 8
Place μ1 in row (3 + I) and column I.
END DO I
The required matrix A for K = 3, B2 = 1, B3 = 0 is given in matrix (2.35).

− μ1 μ3 0 0 0 0 0 0 0 0 0
0 − μ1 − μ3 μ3 μ2
0 0 0 0 0 0 0
0 0 − μ1 − μ3 0 μ2 0 0 0 0 0 0
μ1 − μ1 − μ2 μ3

0 0 0 0 0 0 0 0
0 μ1 0 0 − ∑3i=1 μi μ3 μ2 0 0 0 0

A= 0 0 μ1 0 0 − μ1 − μ3 0 μ2 0 0 0
0 μ1 − μ1 − μ2 μ3
0 0 0 0 0 0 0
0 0 0 0 μ1 0 0 − ∑3i=1 μi μ3 μ2 0
0 μ1 − μ3 0 μ2
0 0 0 0 0 0
0 0 0 0 0 0 μ1 0 0 − μ2 μ3
0 0 0 0 0 0 0 μ 0 0 −μ − μ

1 2 3
(2.35)
2.1.3 A simple non-linear flow model
Non-linear flow models have the characteristic that parts may be returned to upstream
stations or skip stations or meet other parts at particular stations for assembly or two
or more parts emerge from a disassembly station. Thus, non-linearity implies some
lack of strict successive continuity of a distinct product going from one station to the
succeeding station in a production line.
In non-linear flow models consideration is given to assembly/disassembly and
merge operations in production lines. These models may also take account of quality
inspection stations and allow for the possibility of rework where a product is returned
to earlier stations. Clearly, the topology of non-linear flow is more complicated than
linear flow models.
Here, consideration is given to the non-linear flow model shown in Figure 2.10.
The merge phenomenon is indicated in Figure 2.10. Two machines upstream
from the buffer perform the same operation and feed the buffer in such a way that
one machine has priority over the other when the buffer is full. The third machine
M1
B(1,2),3 M3
M2
Fig. 2.10. A merge non-linear flow model
removes material from the buffer. The circle indicates a buffer of finite capacity and
the squares indicate the machines. Thus the priority-one buffer is always selected
first unless it is empty, where the priority-two buffer is chosen. The machines may
break down.
Among the major assumptions of the model investigated here, are that the
two upstream machines are never starved, the third machine is never blocked, all
machines have equal and constant processing times, and geometrically distributed
repair times and times to failures and machines can only fail while processing. The
phenomenon of partial and full blocking of the second upstream machine is taken
into account.
In order to analyze the model of Figure 2.10, the following three methodological
steps are required:
(i) Derivation of the transition equations of all states of the system (internal, lower
boundary and upper boundary).
(ii) Development of a recursive algorithm for generating the transition matrix for
any value C of the extended storage level of buffer B(1,2),3 (that is, the capacity
of the original buffer plus 3).
(iii) Numerical computation of the transition probabilities and then of the various
performance measures of the system.
A formula for the number of states, m, for any value C > 4 of the extended storage
level of buffer B(1,2),3 is given by
m = (8 × C) − 4. (2.36)
The expected in-process inventory (average buffer level), W IP, of the system of
Figure 2.10 may be written as follows:
C 1 1 1
W IP = ∑ ∑ ∑ ∑ cp[c, α1 , α2 , α3 ] (2.37)
c=0 α1 =0 α2 =0 α3 =0
where p[c, α1 , α2 , α3 ] denotes the steady-state probability of the system being in state
[c, α1 , α2 , α3 ]. The level of the extended buffer is denoted by c. αi , i = 1, 2, 3, denotes
the status of machine Mi , which may be up (αi = 1) or down (αi = 0).
The blocking probabilities of machines M1 and M2 , denoted by pbl bl
1 , p2 and the
starvation probability of machine M3 , denoted by p3 are st
2.2 Decomposition Approach 51
1 = p[C − 1, 1, 0, 0] + p[C − 1, 1, 0, 1] + p[C, 1, 1, 0] + p[C, 1, 1, 1],

pbl (2.38)
pbl
2 = p[C − 1, 0, 1, 0] + p[C − 1, 0, 1, 1] + p[C − 1, 1, 1, 0]
+ p[C − 1, 1, 1, 1] + p[C, 1, 1, 0] + p[C, 1, 1, 1], (2.39)
pst3 = p[0, 0, 0, 1] + p[0, 0, 1, 1] + p[0, 1, 0, 1] + p[0, 1, 1, 1]. (2.40)
The mean production rates related to each one of the three machines can be read-
ily determined. If βi and ri are the mean rates of failure and repair, respectively,
of machine Mi , then ei = ri /(ri + βi ), i = 1, 2, 3 represents the fraction of time that
machine Mi is operational. Since all processing times are identical and are taken as
the time unit, it is obvious that ei , i = 1, 2, 3, is the isolated mean production rate
of machine Mi , i.e., the mean production rate of machine Mi , if it were working
alone. Since machines Mi , i = 1, 2, 3 are part of the system, blocking and starva-
tion probabilities should be taken into account. Therefore the mean production rates
(throughputs) related to each one of these three machines are
X1 = (1 − pbl
1 )e1 (2.41)
X2 = (1 − pbl
2 )e2 (2.42)
X3 = (1 − pst3 )e3 . (2.43)
In order to determine the throughput of the system shown in Figure 2.10, the
throughput of the third machine should be computed. The throughput, X , of the
system is simply given by X3 .
X = X3 . (2.44)
In Diamantidis, Papadopoulos and Vidalis (2004), a process for the generation
of the transition matrix was developed and an algorithm to evaluate the performance
parameters, including the average buffer level and the throughput of the system, was
presented.
2.2 Decomposition Approach

Queueing networks are a natural way of analyzing production lines. Although there
is a very rich literature in queueing networks, difficulties in analysis arise when finite
buffers are considered because of the associated starving and blocking phenomena.
Classically, queueing networks tended to be investigated through a process of decom-
position into a set of single-server systems. Such an approach is totally valid in the
case of infinite buffers. With finite buffers, exact classical decomposition is inappro-
priate. However, many researchers have developed efficient decomposition methods
for the approximate evaluation of tandem queues which are suitable for the analysis
of production lines. As noted above, numerical techniques based on exact Markovian
analysis are space and computer time consuming for solutions to the exact queueing
problems and are generally only applicable in practice to small production lines.
Considerable effort has been expended on the development of approximate solutions
to large-scale production lines.
M1 B2=0 M2 B3=1 M3 Original line L
u d
M1 B2=0 M1 L1
u d
M2 B3=1 M2 L2
Fig. 2.11. A three-station line, L, decomposed into two sub-lines, L1 and L2
Essentially, the decomposition approach as applied to a K-station line consists of

decomposing the original line into a set of K − 1 sub-lines. Each sub-line normally
consists of two stations and an intermediate buffer which corresponds to a buffer of
the original line. The original method proposed by Gershwin (1987) was initially
used in the analysis of serial production lines. Consider the following example.
Example: Consider a balanced three-station production line where each work-
station consists of only one machine (this is the original line L). All machines are
assumed to be perfectly reliable with exponentially distributed service times and
identical mean service rates, μ1 = μ2 = μ3 = 1. There is no intermediate buffer
between the first two stations, whereas there is an intermediate buffer of size 1
between the second and third station, viz., B2 = 0 and B3 = 1. The original line,
L, is decomposed into two sub-lines, L1 and L2 , as shown in Figure 2.11. XK denotes
the throughput of the K-station line, whereas X1 and X2 denote the throughput of
sub-line L1 and sub-line L2 , respectively.
Following the approach of Gershwin (1994):
The boundary conditions of the decomposition method are
μ1u = μ1 , (2.45)
μK−1
d
= μK . (2.46)
The general steps of the decomposition method are the following:

Step 1: Initialization
μiu = μi , i = 1, 2, . . . , K − 1
μid = μi+1 , i = 1, 2, . . . , K − 1.
Step 2: Iteration
Perform the following steps 2.1 and 2.2 alternately until the termination
condition is satisfied.
Step 2.1: Evaluate quantities
1
μiu = , i = 2, 3, . . . , K − 1
1
Xi−1 + μ1i − μ d1
i−1
Step 2.2: Evaluate quantities

1
μid = , i = K − 2, K − 3, . . ., 1
Xi+1
1
+ μi+1
1
− μ u1
i+1
Step 3: Termination condition

The algorithm is terminated when
|Xi − X1| < ε , i = 2, 3, . . . , K − 1,
where ε is a pre-determined very small positive real number.

Application of the above decomposition algorithm to the example three-station
balanced production line, L:
I NITIALIZATION : μ1u = μ1 , μ2u = μ2 , μ1d = μ2 , μ2d = μ3 .
F IRST I TERATION (I1)
Step 2.1: From the boundary condition: μ1u = μ1 = 1 and
1
μ2u = 1
X1 + μ11 − μ1d
1
μ1d = μ2 1
= 1
X1 + 11 − μ12
1
= = 0.666667,
1
0.666667 + μ11 − 11
where X1 is the throughput of sub-line L1 which is calculated from the Markovian

algorithm with the parameter values μ1u = μ1 = 1, μ1d = μ2 = 1 and B2 = 0 and gives
the value X1 = 0.666667.
Step 2.2: From the boundary condition: μ2d = μ3 = 1 and
1
μ1d = 1
X2 + μ12 − μ1u
2
Step2.1 1
= 1
X2 + 1 − 0.666667
1 1
1
= = 0.826001,
1
0.584573 + μ11 − 0.666667
1

algorithm with the parameter values μ2u = 0.666667, μ2d = μ3 = 1 and B3 = 1 and
gives the value X2 = 0.584573.
T ERMINATION CONDITION : X1 is the throughput of sub-line L1 which is cal-
culated from the Markovian algorithm with the parameter values μ1u = μ1 = 1,
μ1d = 0.826001 and B2 = 0 and gives the values X1 = 0.601320 and X2 = 0.584573
calculated in step 2.2 above. Thus, |X2 − X1 | > ε = 0.0001 and the two-step iteration
continues.
S ECOND I TERATION (I2)
1
μ2u = 1
X1 + μ11 − μ1d
1
Step2.2(I1) 1
= 1
X1 + − 0.826001
1
1
1
1
= = 0.688536.
1
0.601320 + 1 − 0.826001
1
1
μ1d = 1
X2 + μ12 − μ1u
2
Step2.1(I2) 1
= 1
X2 + − 0.688536
1
1
1
1
= = 0.820157,
1
0.598217 + 1 − 0.688536
1

calculated in step 2.2 above. Again |X2 − X1 | > ε = 0.0001 and the two-step iteration
continues.
T HIRD I TERATION (I3)
1
μ2u = 1
X1 + μ11 − μ1d
1
Step2.2(I2) 1
= 1
X1 + 1 − 0.820157
1
1
= = 0.689339.
1
0.598823 + 1 − 0.820157
1

1
μ1d = 1
X2 + μ12 − μ1u
2
Step2.1(I3) 1
= 1
X2 + 11 − 0.689339
1
1
= = 0.819940,
1
0.598707 + 1 − 0.689339
1

calculated in step 2.2 above. Now, |X2 − X1 | = 0.000031 < ε = 0.0001 and the algo-
rithm terminates giving a throughput value, XDECO = 0.5987. This value may be
compared against the throughput value calculated from the Markovian algorithm,
XMARK = 0.613333. The reader may note that if the value of ε was smaller, say
ε = 0.00001, then more iterations would be needed for the algorithm to terminate.
A coded version of the original Gershwin’s decomposition algorithm is not at
the website associated with this book. However, the algorithm developed by Dr.
Diamantidis (Diamantidis, Papadopoulos and Heavey, 2006) for parallel perfectly
reliable machine production lines (given in Section 2.5 and at the website associated
with this text with abbreviated name DECO-2) may be used by setting the number
of parallel machines at each station equal to 1 as an alternative to the Gershwin algo-
rithm for perfectly reliable single-machine stations. The authors have checked that
the equations derived from Diamantidis et al. (2006) work, and setting the number
of servers at each station equal to 1 showed them to be identical to those developed
by Gershwin for the single-machine perfectly reliable production lines.
Dallery and his group undertook considerable work in the area of decomposition
modeling. Available at the website associated with this book with abbreviated name
DECO-1 is a coded version of an algorithm given in Dallery and Frein (1993) for the
analysis of reliable production lines with single machines at each station. To illustrate
the development of Dallery and Frein’s algorithm, consider the four-station produc-
tion line with finite inter-station buffers, B2 , B3 and B4 , with capacities B2 , B3 and
B4 , respectively, shown in Figure 2.12. For simplicity, work-stations are denoted by
Mi , 1 = 1, 2, 3, 4 instead of W Si , i = 1, . . . , 4. Figure 2.13 depicts the decomposition
of this four-station line into three sub-lines, L1 , L2 , and L3 , each consisting of two
stations and one intermediate buffer. All single-machine work-stations are assumed
to be perfectly reliable and the service times at each machine are exponentially
distributed with mean service rates, μi , i = 1, . . . , 4.
The application of queueing network theory to production lines involved the use
of either the open model or the saturated model. In the saturated model, the assump-
tion is that the first station is never starved, whereas in the open model, the first queue
M1 B2 M2 B3 M3 B4 M4
Mi : work-station i, i = 1,2,3,4
Bi : buffer i, i = 2,3,4
Fig. 2.12. Production line, L, with K = 4 work-stations and 3 intermediate buffers
u
B2 d L1
M1 = M1 M1
u L2
M2 B3 M 2d
u B4 d
M3 M 3 = M4 L3
Fig. 2.13. Decomposition of the original line, L, into three sub-lines each with two stations
and one buffer
U B D
Fig. 2.14. Sub-line Li
is finite and the first station may be starved. The model considered here is a saturated
model. Actually, all models examined in this book are saturated models except for
the model solved via the expansion method (Section 2.3).
Some formal results from queueing theory are required in the development of the
decomposition equations and these are given immediately below.
Queueing theory analysis of the sub-lines
Consider the sub-line shown in Figure 2.14.

There are B + 2 states in the system of Figure 2.14, where B is the buffer size
of the intermediate buffer B. Some of the performance parameters of a two-station
system, called decomposition block, are required for the development of the decom-
position equations. The states of the system are characterized by ν = 0, 1, . . . , B + 1,
where ν denotes the number of jobs waiting for service at station 2. ν = B + 1 occurs
when station U is blocked. μU and μD , respectively, denote the mean service rate of
machines U and D. Let p(ν ) be the steady-state probability of the system being in
state ν and let pU (ν ) be the probability that there are ν jobs in the buffer when a job
is completed at station U. Likewise, let pD (ν ) be the probability that on completion
of service at station D, there are ν jobs left in buffer B. The throughput of the system
is denoted by X . Using well-known results, the following relationships apply:
p(ν )
pU (ν ) = , ν = 0, . . . , B (2.47)
1 − p(B + 1)
p(ν + 1)
pD (ν ) = , ν = 0, . . . , B (2.48)
1 − p(0)
X = μU (1 − p(B + 1)) (2.49)
X = μD (1 − p(0)). (2.50)
Two further probabilities are of interest, these being pUbl , the probability that sta-
tion U is blocked, and pstD , the probability that station D is starved. Clearly these are
given by
pUbl = pU (B) (2.51)

pstD = pD (0). (2.52)
Returning to the decomposition process, in Figure 2.13, sub-line Li , i = 1, 2, 3

(= K − 1) approximates the flow of jobs in buffer Bi , i = 2, 3, 4(= K) of the
original line. In sub-lines Li , i = 1, . . . , K − 1, stations Miu , i = 1, . . . , K − 1 and
Mid , i = 2, . . . , K − 1 represent the part of the original line, L, upstream and down-
stream of buffer Bi+1 , i = 1, . . . , K − 1, respectively. The concept is that whereas
the buffer Bi is in all respects identical to the buffer preceding station i in the orig-
inal line L, the two stations Miu and Mid are not identical to stations i − 1 and i,
except that station M1u is identical to station M1 and station M3d is identical to sta-
tion M4 . The characteristics of the remaining stations are so chosen that in effect
they represent the impact of the upstream and downstream parts of the production
line L on the buffer Bi , i = 2, 3, 4. μi , i = 1, . . . , K denotes the mean service rate
of station i, i = 1, . . . , K in the original line, L. Similarly, μiu , i = 1, . . . , K − 1 and
μid , i = 1, . . . , K − 1 denote the mean service rate of stations Miu , i = 1, . . . , K − 1 and
Mid , i = 1, . . . , K − 1, respectively, in the sub-lines, Li , i = 1, . . . , K − 1. All service
times are assumed to be exponentially distributed with the respective mean service
rates given above.
The development of the sets of the decomposition equations starts out by con-
sidering the sub-lines L1 and L3 . This gives the boundary conditions already stated
above that μ1u = μ1 and μ3d = μ4 = μK .
In total, there are in general 2(K − 1) unknowns, and there is a need to obtain
another 2(K − 2) independent equations.
wi = 1/μi , i = 1, . . . , K denotes the mean service time (average work-load) at
station i, i = 1, . . . , K of the original line, L. The mean service time of the downstream
station Mid , i = 1, 2, . . . , K − 2 (K − 2 = 4 − 2 = 2 for the example line of Figure 2.12
and the sub-lines of Figure 2.13), denoted by wdi = 1/μid , is the sum of the service
time at station i in the original line and the possible blocking time of station Mi in the
original line L, which is equivalent to the blocking time of station Mi+1 u in sub-line
Li+1 due to the fact that buffer Bi+2 is full and on the assumption that the station is
perfectly reliable. This gives rise, in general, to the following set of equations for the
reliable exponential production lines:
wdi = wi + pbl
i+1 wi+1 , i = 1, 2, . . . , K − 2,
d
(2.53)
where, pbl u
i+1 denotes the blocking probability of station Mi+1 .
A similar set of equations may be developed for the upstream stations. More
specifically, the mean service time of the upstream station Miu , i = 2, . . . , K − 1
(K − 1 = 4 − 1 = 3 for the example line of Figure 2.12 and the sub-lines of Fig-
ure 2.13), denoted by wui = 1/μiu , is the sum of the service time of station i − 1 in
the original line and the possible starvation time of station i − 1. The latter event in
d in sub-line L
the original line is equivalent to the starvation of station Mi−1 i−1 . This
gives rise, in general, to the following set of equations for the reliable exponential
production lines:
wui = wi−1 + psti−1wui−1 , i = 2, 3, . . . , K − 1, (2.54)
where psti−1 denotes the starvation probability of station Mi−1
d .
The third set of equations is related to the conservation of flow, i.e., the through-
put of all stations in the line is the same and consequently the throughput of the
sub-lines must satisfy the following flow equations:
X1 = X2 = · · · = XK−1 , (2.55)
where Xi denotes the throughput of sub-line Li , i = 1, . . . , K − 1.
As may be noted from the above, there are two sets of K − 2 equations plus
two boundary conditions, so it is not necessary to use all the equations to solve for
the unknowns. This leads to the utilization of the following sub-set of the above
equations:
wui = wi−1 + psti−1 wui−1 , i = 2, 3, . . . , K − 1, (2.56)
wdi = wi + i+1 wi+1 , i = 1, 2, . . . , K − 2,
pbl d
(2.57)
wu1 = w1 , and wdK−1 = wK (2.58)
Dallery and Frein (1993) proved that the above set of equations satisfies the con-
servation of flow criterion. They also proved the existence and uniqueness of the
solution derived from this set of equations and that this symmetrical set of equations
is equivalent to each of the following sets of equations:
wdi = wi + pbl
i+1 wi+1 , i = 1, 2, . . . , K − 2,
d
(2.59)
X1 = X2 = · · · = XK−1 , (2.60)
wu1 = w1 , and wdK−1 = wK . (2.61)
wui = wi−1 + psti−1 wui−1 , i = 2, 3, . . . , K − 1, (2.62)

X1 = X2 = · · · = XK−1 , (2.63)
wu1 = w1 , and wdK−1 = wK . (2.64)
2.3 The Expansion Method 59
Iterative procedures for solving the above three sets of equations have been pro-
posed by Dallery and Frein (1993) and form the basis of the decomposition algorithm
available at the website associated with this book.
The numerical processes involved in the algorithm are relatively straightforward.
The two boundary conditions on the mean service rates of the first and last stations
of the line are set and then the mean service rate of each of the other downstream
stations are set equal to the values of the original line. The starvation and blocking
probabilities are then calculated and values of the upstream and changed values of
the downstream stations mean service rates are developed. This process continues
until satisfactory convergence is achieved. Finally, the throughput of the line may be
determined. The numerical decomposition process is outlined in flow diagram form
in Figure 2.15.
In general, the decomposition method as applied to production lines consists
essentially of three steps as follows:
1. The specification of the sub-lines
2. The determination of a set of equations used to evaluate the unknown parame-
ters of each sub-line in such a way that the flow of material through the sub-lines
resembles the corresponding flow of material in the original line. More specifi-
cally, the following conditions have to be satisfied as explicitly given in Gershwin
(1994):
• The rate of flow into and out of buffer Bi in sub-line Li approximates that of
buffer Bi in the original line L.
• The probability of the buffer of sub-line Li being empty or full is close to
that of Bi in the original line L being empty or full.
• The probability of resumption of flow into and out of the buffer in sub-line
Li in a time interval after a period during which it was interrupted is close to
the probability of the corresponding event in the original line L.
• The average level of material in buffer Bi in sub-line Li approximates the
corresponding material level in buffer Bi in the original line L.
These conditions lead in the case of perfectly reliable production lines to the
conservation of flow equations and in the case of unreliable production lines to the
addition of the resumption of flow equations and interruption of flow equations.
3. The development of an appropriate procedure for solving the set of equations.
2.3 The Expansion Method

The expansion method is an approximation technique developed by Kerbache (1984),
published also in Kerbache and MacGregor Smith (1987) and extended by Jain and
MacGregor Smith (1994). This method is characterized as a combination of repeated
trials and node-by-node decomposition solution procedures. Methodologies for com-
puting performance measures for a finite queuing network use the following two
kinds of blocking:
Initialization Step
Set μ1u = μ1 and μK−1
d
=μK
Set μid = μi+1 for i = 1,2,..,K–2
Set μiu = mi, for i = 2,3,..,K–1
Iteration Step
Step 1: For i = 2,..,K–1

1.1 Calculate the starvation
st
probability p i−1
1
1.2 Calculate μ i =
u
w iu
Step 2: For i = K–2,…,1

2.1 Calculate the blocking
bl
probability pi +1
1
2.2 Calculate μid = d
wi
Step 3: Go to step 1 until the

changes in μiu ,μdi from one
iteration to the next are
sufficiently small
Exit - Results
Fig. 2.15. Flow chart for decomposition method
Type I: The upstream node i gets blocked if the service on a unit is completed but
it cannot move downstream due to the queue at the downstream node j
being full. This is referred to as blocking after service (BAS) (Onvural and
Perros, 1986, Perros, 1994).
Type II: The upstream node is blocked when the downstream node becomes sat-
urated and service must be suspended on the upstream unit regardless of
whether service is completed or not. This is referred to as blocking before
service (BBS) (Onvural and Perros, 1986, Perros, 1994).
M/M/c/K M/M/c/K
Arrival
1 2 Throughput
Rate
PK´
M/M/c/•
M/M/c/K M/M/c/K
Arrival
1 h 2 Throughput
Rate
PK
1 - PK
Fig. 2.16. Expansion of a finite queue M/M/c/K
The Expansion Method uses Type I blocking, which is prevalent in most production,
manufacturing, transportation and other similar systems.
Consider a single node with finite capacity K (including service). This node
essentially oscillates between two states—the saturated phase and the unsaturated
phase. In the unsaturated phase, node j has at most K − 1 units (in service or in
the queue). On the other hand, when the node is saturated no more units can join
the queue. Refer to Figure 2.16 for a graphical representation of the expansion of a
finite queue M/M/c/K. The reader may note that this model is the only open model
considered in this book. All the other models are saturated models.
The Expansion Method consists of the following three stages:
• Stage I: Network reconfiguration.
• Stage II: Parameter estimation.
• Stage III: Feedback elimination.
The following notation defined by Kerbache and MacGregor Smith (1987) and
Jain and MacGregor Smith (1994) will be used in further discussion regarding this
methodology:
h := The holding node established in the expansion method.
Λ := External Poisson arrival rate to the network.
λ j := Poisson arrival rate to node j.
λ˜j := Effective arrival rate to node j.
μ j := Exponential mean service rate at node j.
μ˜ j := Effective service rate at node j due to blocking.
pK := Blocking probability of finite queue of size K.
pK := Feedback blocking probability in the expansion method.
j
p0 := Unconditional probability that there is no unit in the service channel at node j
(either being served or being held after service).
X := Mean production rate (throughput).
Stage I: Network reconfiguration
Using the concept of two phases at node j, an artificial node h is added for each finite
node in the network to register blocked units. Figure 2.16 shows the additional delay,
caused to units trying to join the queue at node j when it is full, with probability
pK . The units successfully join queue j with a probability (1 − pK ). Introduction of
an artificial node also dictates the addition of new arcs with pK and (1 − pK ) as the
routing probabilities.
The blocked unit proceeds to the finite queue with probability (1 − pK ) once
again after incurring a delay at the artificial node. If the queue is still full, it is re-
routed with probability pK to the artificial node where it incurs another delay. This
process continues until it finds a space in the finite queue. A feedback arc is used
to model the repeated delays. The artificial node is modeled as an M/M/∞ queue.
The infinite number of servers is used simply to serve the blocked unit a delay time
without queuing.
Stage II: Parameter estimation
This stage essentially estimates the parameters pK , pK and μh utilizing known results
for the M/M/c/K model.
• pK : Analytical results from the M/M/c/K model provide the following expression
for pK :
K
1 λ
pK = K−c p0 (2.65)
c c! μ
where for (λ /cμ = 1)
n −1
1c−1
λ (λ /μ )c 1 − (λ /cμ )K−c+1
p0 = ∑ + (2.66)
n=0 n! μ c! (1 − λ /cμ )
and for (λ /cμ = 1),

n −1
1c−1
λ (λ /μ )c
p0 = ∑ + (K − c + 1) . (2.67)
n=0 n! μ c!
• pK : Since there is no closed form solution for this quantity, an approximation
obtained by Labetoulle and Pujolle (1980), using diffusion techniques, is used:
−1
μ j + μh λ [(xK2 − xK1 ) − (x2K−1 − x1K−1 )]
pK = − (2.68)
μh μh [(x2K+1 − x1K+1 ) − (xK2 − xK1 )]
where x1 and x2 are the roots to the polynomial:
λ − (λ + μh + μ j )x + μh x2 = 0 (2.69)
where λ = λ j − λh(1 − pK ) and λ j and λh are the actual arrival rates to the finite and
artificial holding nodes respectively.
In fact, λ j the arrival rate to the finite node is given by:
λ j = λ̃i (1 − pK ) = λ̃i − λh . (2.70)
If an arriving unit is blocked, the queue is full and thus a unit is being serviced, so
the arriving unit to the holding node has to remain in service at the artificial holding
node for the remaining service time interval of the unit in service. The delay distribu-
tion of a blocked unit at the holding node has the same distribution as the remaining
service time of the unit being serviced at the node doing the blocking. Using renewal
theory, one can show that the remaining service time distribution has the following
rate μh :
2μ j
μh = (2.71)
1 + σ 2 μ 2j
where σ 2 is the service time variance given by Kleinrock (1975). Notice that if the
service time distribution at the finite queue doing the blocking is exponential with
rate μ j , then:
μh = μ j
the service time at the artificial node is also exponentially distributed with rate μ j .
Stage III: Feedback elimination
Due to the feedback loop around the holding node, there are strong dependencies in
the arrival processes. Elimination of these dependencies requires reconfiguration of
the holding node which is accomplished by recomputing the service time at the node
and removing the feedback arc. The new service rate is given by:
μh = (1 − pK )μh . (2.72)
The probabilities of being in any of the two phases (saturated or unsaturated) are
pK and (1 − pK ), respectively. The mean service time at a node i, preceding the finite
node is μi−1 if in the unsaturated phase and (μi−1 + μh−1 ) in the saturated phase.
Thus, on average, the mean service time at the node i preceding a finite node, is
given by:
μi˜−1 = μi−1 + pK μh−1 . (2.73)
Similar equations can be established with respect to each of the finite nodes.
Ultimately, a set of simultaneous non-linear equations in variables pK , pK , μh−1
along with auxiliary variables such as μ j and λ̃i is developed. Solvingthese equations
simultaneously, all the performance measures of the network can be computed:
λ = λ j − λh (1 − pK ) (2.74)
λ j = λ̃i (1 − pK ) (2.75)
λ j = λ̃i − λh (2.76)
−1
μ j + μh λ [(xK2 − xK1 ) − (x2K−1 − x1K−1 )]
pK = − (2.77)
μh μh [(x2K+1 − x1K+1 ) − (xK2 − xK1 )]
z = (λ + 2 μh)2 − 4λ μh (2.78)
1
[(λ + 2 μh) − z ]
2
x1 = (2.79)
2 μh
1
[(λ + 2 μh) + z 2 ]
x2 = (2.80)
2 μh
K
1 λ
pK = K−c p0 . (2.81)
c c! μ
Equations (2.74) to (2.77) are related to the arrivals and feedback in the holding
node. Equations (2.78) to (2.80) are used for solving equation (2.77) with z used
as a dummy parameter for simplicity of the solution. Finally, equation (2.81) gives
the approximation to the blocking probability derived from the exact model for the
M/M/c/K queue. Hence, essentially there are five equations to solve, viz. (2.74)
to (2.77) and (2.81).
To recapitulate, first the network is expanded with an artificial holding node; this
stage is followed by the approximation of the routing probabilities, due to blocking,
and the service delay in the holding node; and, finally, the feedback arc at the holding
node is eliminated. Once these three stages are completed, an expanded network has
been developed which can then be used to compute the performance measures for the
original network. As a decomposition technique, this approach allows the successive
addition of a holding node for every finite node, estimation of the parameters and
subsequent elimination of the holding node.
The expansion algorithm is available on the website associated with this text
with the abbreviated name EXPAN. Not many practitioners are aware of the expan-
sion method and there is little guidance in the published literature as to the accuracy
achieved using the method in the analysis of realistic systems of interest to the
designers of production lines. However, it must be recognized that in a historical
context, the expansion method was used as a first serious attempt to computationally
solve systems with parallel machines at each station.
2.4 The Aggregation Method

Lim, Meerkov and Top (1990) published an approximation approach used in the anal-
ysis of transfer lines which has come to be known as the aggregation method. This
very powerful method begins by combining the first two machines of the transfer
2.4 The Aggregation Method 65
line into a new combined machine. This aggregated combined machine is then com-
bined with the third machine and this forward aggregation process is continued until
the last machine is reached. A backward aggregation process is then applied. The
algorithm which is available at the website associated with this book stops when the
results of both aggregations (forward and backward) coincide.
Assumptions of the model
A serial transfer line is considered consisting of K machines and K − 1 intermediate

buffers. Machines are assumed to have identical cycle time and the time axis is slot-
ted with the slot/period duration equal to the cycle time. It is assumed that the first
machine is never starved and the last machine is never blocked. It is further assumed
that a certain machine i, i = 1, . . . , K produces a part during any time slot/period with
probability qi and fails to do so with probability 1 − qi , provided that machine i is
neither blocked nor starved. Mathematically, qi is defined as follows:
q i = 1 − ε Λi ,
where 0 < ε << 1, which characterizes the asymptotically reliable line, and Λi ,
i = 1, . . . , K is independent of ε . The Λi ’s were defined by Lim, Meerkov and Top
(1990) as the loss parameters. Lim et al. (1990) also defined the following function:
1−α
Q(α , ν ) = , α ∈ R + , ν ∈ [1, ∞). (2.82)
1 − αν
The two-machine, one-buffer transfer line
A two-machine, one-buffer transfer line in steady state is equivalent to a single

aggregated machine characterized by

Λ2
qaggregation = 1 − Λ2 + Λ1 Q ,ν ε (2.83)
Λ1
Thus, the loss parameter of the equivalent aggregated machine is

Λ2
Λaggregation = Λ2 + Λ1 Q ,ν , (2.84)
Λ1
where Λ1 and Λ2 are, respectively, the loss parameters of the first and second
machine. The mean production rate, X2 , of the two-machine, one-buffer system is
given by

Λ2
X2 = 1 − Λ2 + Λ1 Q ,ν ε + O(ε 2 )
Λ1
(2.85)

Λ1
= 1 − Λ 1 + Λ2 Q ,ν ε + O(ε 2 ).
Λ2
It is obvious that
Λ1
Λaggregation = Λ1 + Λ2 Q ,ν . (2.86)
Λ2
Equations (2.84) and (2.86) show that

Λ1 Λ2
Λaggregation = Λ1 + Λ2 Q ,ν = Λ2 + Λ1 Q ,ν .
Λ2 Λ1
Longer lines
The above process can be generalized to the case of homogeneous asymptotically

reliable serial transfer lines consisting of K machines and K − 1 intermediate buffers.
The first two machines are combined into an aggregated machine with the loss
parameter, Λ2f , defined by (2.84), i.e.,

f Λ2
Λ 2 = Λ2 + Λ 1 Q , ν1 .
Λ1
Superscript ‘f’ indicates that during the aggregation, one moves forward (from the
first to the last machine). The aggregated machine, characterized by Λ2f , is now com-
bined with the third machine, defined by the loss parameter Λ3 . The new aggregated
machine is characterized by the loss parameter:

Λ3
f f
Λ 3 = Λ3 + Λ 2 Q , ν2 .
Λ2f
At the ith step of this multi-stage aggregation process, one may obtain:

Λi
f f
Λi = Λi + Λi−1 Q f
, νi−1 (2.87)
Λi−1
and at the final step:

f f ΛK
ΛK = ΛK + ΛK−1 Q f
, νK−1 .
ΛK−1
The estimate of the mean production rate (throughput) obtained as a result of this
aggregation is:

ΛK
f f
XK = 1 − ΛK + ΛK−1 Q f
, νK−1 ε. (2.88)
ΛK−1
f
Because there is no proof that XK is close to the real throughput of the production line
with K machines in series and K − 1 intermediate buffers, another set of equations
should be supplemented, but this time directed backwards instead of forwards. This
2.5 Modeling of Production Lines with Parallel Reliable Machines at Each Station 67
scheme is called backward aggregation and aggregates the line moving from the last
machine to the first machine. The respective loss paramaters are:
f
ΛK−1
ΛK−1 = ΛK−1 + ΛK Q
b
, νK−1
ΛK
f
Λ
ΛbK−2 = ΛK−2 + ΛbK−1 Q K−2
, νK−2
ΛbK−1
f
Λj
Λ j = Λ j + Λ j+1 Q
b b
,νj
Λbj+1

Λ1
Λb1 = Λ1 + Λb2 Q , ν1 .
Λb2
By repeating the process and constructing a new forward aggregation based on the
backward aggregation, the following iterative algorithm is obtained:

f f Λbi (s)
Λi (s + 1) = Λi + Λi−1 (s + 1) Q f
, νi−1 , i = 2, . . . , K
Λi−1 (s + 1)
s = 0, 1, . . . , Λbi (0) = Λi , Λ1f (s) = Λ1 , ΛbK (s) = ΛK , ∀s. (2.89)
f
Λ j (s + 1)
Λbj (s + 1) = Λ j + Λbj+1(s + 1) Q , ν j , j = 1, . . . , K − 1.
Λbj+1 (s + 1)
Procedure (2.89) generates the following two sequences of throughput estimates:
XKf (s) = 1 − ΛKf (s) ε

(2.90)
XKb (s) = 1 − Λb1(s) ε .
The properties of these sequences are described in Lim, Meerkov and Top (1990).
The aggregation algorithm is available at the website associated with this text with
abbreviated name AGGRE.
2.5 Modeling of Production Lines with Parallel Reliable

Machines at Each Station
The throughput of production lines may be increased by adding extra machines at
stations. It should be understood that all machines at the stations are used provided
there is work available. Here, attention is confined to lines with reliable machines
and with exponential processing times. A K-work-station line with Si , i = 1, 2, . . . , K
parallel machines at work-station i, denoted by W Si , and with intermediate buffers
B j , j = 2, . . . , K of capacities B j is depicted in Figure 2.17.
68
Work-station WS1 Work-station WS2 Work-station WSK
Machine 1 Machine 1 Machine 1
Machine 2 Machine 2 Machine 2

B2 B3 BK
Machine S1 Machine S2 Machine SK

Fig. 2.17. A K-work-station production line with Si parallel machines at each work-station W Si , i = 1, 2, . . . , K
In sub-section 2.5.1, the exact solution to a two-station line with multiple mac-
hines at each station is presented. It might be noted that it is possible to develop the
conservative matrix A of these systems with a view of developing exact numerical
solutions along the lines already explained in Section 2.1. Interested readers might
refer to Vidalis and Papadopoulos (2001). However, computational complexities con-
siderably reduce the value of developing the conservative matrix A of such systems.
In sub-section 2.5.2, an alternative exact solution to the the two-station line with par-
allel machines at each station is presented as given in Diamantidis, Papadopoulos
and Heavey (2006). This solution is used as a building block for a decomposition
analysis of larger production lines with parallel machines at each station. Details of
the latter analysis are given in sub-section 2.5.3.
2.5.1 Exact solution to a two-station production line with parallel machines

at each station
Consider a system consisting of two stations with S1 and S2 parallel machines at

station 1 and station 2, respectively. It is assumed that the first station is always
busy, i.e., it is saturated and the intermediate buffer is of capacity B2 which includes
the number of parallel machines at station 2, i.e., B2 ≥ S2 . The processing times
at each station are exponentially distributed with mean rates μi , i = 1, 2. Buzacott
and Shanthikumar (1993) (pages 205–206) and Perros (1994) (pages 64–65), among
others, considered this problem. By forming the Markovian chain of this system, the
random variable of interest is N(t), the number of jobs which have been processed
by the first station at time t and have not finished their processing at station 2 (at
time t). N(t),t ≥ 0 is a birth-death process with state space s = {0, 1, . . . , S2 , S2 +
1, . . . , B2 , B2 + 1, . . . , B2 + S1 }. The birth rate is μ1 (min S1 , B2 + S1 − ν ), whereas the
death rate is μ2 (min S2 , ν ), where, ν = 0, 1, . . . , S2 , S2 + 1, . . . , B2 , B2 + 1, . . . , B2 + S1 .
Let p(ν ) be the probability that there are ν jobs in the second station including the
jobs in the first station that are blocked. The steady-state (flow balance) equations
associated with p(ν ) are (see Perros, 1994, pp. 64–65):
S1 μ1 p(0) = μ2 p(1)
(S1 μ1 + ν μ2 )p(ν ) = (ν + 1)μ2 p(ν + 1) + S1 μ1 p(ν − 1),
for ν = 1, 2, . . . , S2 − 1,
(S1 μ1 + S2 μ2 )p(ν ) = S2 μ2 p(ν + 1) + S1 μ1 p(ν − 1),
for ν = S2 , . . . , B2 , (2.91)
[(S1 + B2 − ν )(μ1 + S2 μ2 )] p(ν ) = S2 μ2 p(ν + 1)
+[S1 + B2 − (ν − 1)]μ1 p(ν − 1),
for ν = B2 + 1, . . ., B2 + S1 − 1,
S2 μ2 p(B2 + S1 ) = μ1 p(B2 + S1 − 1).
In steady state, the throughput of this system may be shown to be:

S2 −1 B2 +S1
X= ∑ ν μ2 p(ν ) + S2 μ2 ∑ p(ν ) (2.92)
ν =0 ν =S2
B2 B2 +S1
X= ∑ S1 μ1 p(ν ) + ∑ (B2 + S1 − ν )μ1 p(ν ) (2.93)
ν =0 ν =B2 +1
where p(ν ), ν = 0, 1, . . . , S2 , S2 + 1, . . . , B2 , B2 + 1, . . . , B2 + S1 are obtained from

equations (2.91), above, by iteration:
⎧ Sν ν
⎪ μ1
⎪ 1
⎪ ν ! μ2 p(0), ν = 0, . . . , S2
⎪
⎨ Sν ν
μ1
1
p(ν ) = S2 !Sν −S2 μ2 p(0), ν = S2 + 1, . . ., B2 (2.94)
⎪
⎪ 2

⎪
⎪
B2
μ1
ν
⎩ S1 S1 !
p(0), ν = B2 + 1, . . ., B2 + S1
ν −S2
S2 !S2 (B2 +S1 −ν )! μ2
where probability p(0) is obtained from the normalizing condition that the sum of
all the steady-state probabilities is equal to 1.
As part of the development of a decomposition method (sub-section 2.5.3), Dia-
mantidis, Papadopoulos and Heavey (2006) also solved the above problem exactly
and the algorithm formulated by them is given below in sub-section 2.5.2.
This algorithm is available at the website associated with this book as special
case of the 1184 2 (with the abbreviated name DECO-2) for K = 2 and in this case it
gives the exact solution.
2.5.2 Alternative exact Markovian analysis of a two-station line with parallel

machines at each station
The motivation for the development of this solution to the two-station multiple server
line was to have available a building block for use in a decomposition approach to
the solution of larger lines.
m1 m2
. .
. .
m1 m2
. B .
. .
. .
. .
m1 m2
S1 servers S2 servers
M1 M2
Fig. 2.18. A two-station, one-buffer production line with parallel machines at each station
Consider a system with two work-stations, which, for simplicity, are denoted
by M1 and M2 (instead of W S1 and W S2 , respectively) as shown in Figure 2.18
consisting of S1 and S2 parallel machines, respectively. This system is used as the
decomposition block in the decomposition approach given in sub-section 2.5.2. It is
assumed that an inexhaustive supply of workpieces is available upstream of work-
station 1, and an unlimited storage area is present downstream of work-station 2, viz.,
work-station 1 is never starved and work-station 2 is never blocked. Work-station
i, i = 1, 2 consists of Si reliable and identical machines, arranged in parallel and S1
need not equal S2 . Each parallel server has an exponentially distributed service time
with mean 1/ μi . The size or capacity of the intermediate buffer is denoted by B.
The total storage capacity of the system is the physical storage of buffer B as well
as the service positions at both work-stations 1 and 2. Therefore, the total storage
capacity of the system, C, is C = S1 + S2 + B. Thus work-station 1 can be either
partially or fully blocked. More specifically, if the current inventory of parts of the
system (including those on the machines) equals S2 + B + 1, then only one machine
at work-station 1 is blocked and the remaining S1 − 1 machines are not blocked.
In this case, work-station 1 is partially blocked. If the storage of the system equals
S1 + S2 + B, then all S1 machines of the first work-station are blocked and, therefore,
this work-station is fully blocked.
Because of the exponentially distributed service times, during the time interval
[t,t + dt] it is assumed only one event can occur at each work-station. Thus during
the time interval [t,t + dt] only one machine among the S1 machines of work-station
1 can produce a part, or only one machine among the S2 machines of work-station 2
can remove a part from buffer B. The total number of units in the system varies from
0 to S1 + S2 + B. It is straightforward that the total number of states is S1 + S2 + B + 1.
Let y = (c) denote the state of the system, where c = 0, . . . ,C.
To solve this two-station system using exact Markovian analysis, the transition
matrix must be derived. The following sub-section gives the transition equations.
Then in sub-section 2.5.2 an algorithm for generating the transition matrix for any
value C is presented.
Transition equations
The system states can be divided with respect to the storage level, c, into three sets:
(i) lower boundary states; (ii) internal states; (iii) upper boundary states. It is further
assumed that S1 ≥ 1, S2 ≥ 1 and B ≥ 0.
Lower boundary state equation
The transition equation for state y with c = 0 (referred to as lower boundary state)
has the following structure:
p0 = (1 − S1 μ1 )p0 + μ2 p1 . (2.95)
Internal state equations
The transition equations for states y = (c) with 0 < c < C can be sub-classified as
follows:
Case 1: If c > 0 and c < S2 , then:
pc = S1 μ1 pc−1 + (1 − S1 μ1 − cμ2 )pc + (c + 1)μ2 pc+1 . (2.96)
Case 2: If c > S2 − 1 and c < S2 + B + 1, then:
pc = S1 μ1 pc−1 + (1 − S1μ1 − S2 μ2 )pc + S2 μ2 pc+1 . (2.97)
Case 3: If c > S2 + B and c < C, then:
pc = S2 μ2 pc+1 + (C + 1 − c)μ1 pc−1

+ (1 − (C − c)μ1 − S2 μ2 )pc . (2.98)
Upper boundary state equation
The state with storage level c = C is called an upper boundary state. It holds that:
pC = μ1 pC−1 + (1 − S2 μ2 )pC . (2.99)
The algorithm for generating the transition matrix
Based on the above classification of the steady-state equations, an algorithm has

been developed to generate the transition matrix of the two-station system (called
decomposition block in the context of the decomposition approach, given below, in
sub-section 2.5.2). Let Pi j , i, j = 0, . . . ,C be the element that is located in the ith
row and jth column of the transition matrix P. The algorithm generates the transition
probabilities in three stages: (i) transition probabilities of the lower boundary states
(see Figure 2.19); (ii) transition probabilities of the internal states (Figure 2.20);
(iii) transition probability of the upper boundary state (see Figure 2.19).
The transition matrix for the decomposition block can be generated using the
algorithms presented in Figures 2.19 and 2.20. The Gaussian elimination method
implemented in C++ is used to solve for the steady-state probabilities. The mean
production rate (throughput) (X ) of the decomposition block is calculated by using
either of the following two formulas:
c=C c=S2 −1
X = S 2 μ2 ∑ p c − μ2 ∑ (S2 − c)pc (2.100)
c=0 c=0
or
c=C c=C
X = S 1 μ1 ∑ p c − μ1 ∑ (c − S2 − B)pc . (2.101)
c=0 c=S2 +B+1
{Lower boundary states}
P0,0 = 1 − S1 μ1
P0,1 = S1 μ1
for c = 2 to C do
P0,c = 0.0
end for
{Upper boundary states}
PC,C−1 = S2 μ2
PC,C = 1 − S2 μ2
for c = 0 to C − 2 do
PC,c = 0.0
end for
Fig. 2.19. Algorithm for generation of lower and upper boundary state transition probabilities
The expected in-process inventory (average storage level), W IP, of the system can
be calculated as follows (the reader is referred to Gershwin, 1994 and Helber, 1999)
c=C
W IP = ∑ c pc . (2.102)
c=0
The method for solving the decomposition block was validated using simula-
tion. Sample results are given Table 2.11. For comparison purposes, a simulation
model was developed in Arena V3.0 and the simulation results were found to be
close enough to those obtained from the analytical model. Ninety-five percent confi-
dence intervals were computed for any value B. The length of the simulation time is
identical for all cases and equals 1100 time units.
For the experiments presented in Table 2.11, the processing times at both work-
stations are assumed to be exponentially distributed with mean service rates, μ1 =
μ2 = 1. In Table 2.11, the first column gives the number of parallel machines at the
first work-station (S1 ), the number of parallel machines at work-station 2 (S2 ) and
the buffer size, B. All these three values are represented by a vector (S1 , S2 , B). The
second column gives the throughput obtained by the numerical solution of the exact
analytical algorithm proposed by Diamantidis, Papadopoulos and Heavey (2006),
described above, while Xalgorithm and the third column gives the estimated 95%
confidence intervals for the simulated mean production rates.
2.5.3 Approximate methods for large lines
Using the solution of the two-station line as a building block, the decomposition
approach was applied by Diamantidis, Papadopoulos and Heavey (2006) to solve
large-scale production lines consisting of K parallel-machine work-stations as those
shown in Figure 2.21. Each work-station i, denoted for simplicity by Mi in the rest of
for i = 1 to C − 1 do
for j = 0 to j = C do
if i > j and i − j = 1 and i < S2 then
Pi, j = iμ2
end if
if i > j and i − j = 1 and i ≥ S2 then
Pi, j = S2 μ2
end if
if i = j and j < S2 and i < S2 + B + 1 then
Pi, j = 1 − S1 μ1 − j μ2
end if
if i = j and j ≥ S2 and i < S2 + B + 1 then
Pi, j = 1 − S1 μ1 − S2 μ2
end if
if i = j and j ≥ S2 and i ≥ S2 + B + 1 then
K = C−i
Pi, j = 1 − K μ1 − S2 μ2
end if
if j > i and j − i = 1 and i < S2 + B + 1 then
Pi, j = S1 μ1
end if
if j > i and j − i = 1 and i ≥ S2 + B + 1 then
m = C−i
Pi, j = mμ1
end if
if i > j and i − j > 1 then
Pi, j = 0.0
end if
if j > i and j − i > 1 then
Pi, j = 0.0
end if
end for
end for
Fig. 2.20. Algorithm for generation of internal state transition probabilities
Table 2.11. Throughput of a two-work-station system with parallel machines
(S1 , S2 , B) Xalgorithm 95% CI for Simulated Throughput

(4,4,3) 3.52593 (3.48, 3.57)
(5,5,5) 4.54631 (4.47, 4.60)
(10,15,7) 9.99439 (9.80, 10.15)
(15,20,15) 14.99740 (14.75, 15.14)
(10,10,10) 9.45416 (9.25, 9.56)
this sub-section, consists of multiple identical reliable parallel machines with service
rates μi , i = 1, . . . , K and intermediate buffers Bi , i = 2, . . . , K. The number of parallel
machines at station i is Si , i = 1, . . . , K, with each Si an integer. Service times are
exponentially distributed with mean 1/ μi . It is also assumed that when any one of
the Si parallel machines at work-station Mi completes a part, that part is placed in the
buffer Bi+1 downstream of the work-station immediately, provided the buffer is not
full.
Markovian analysis of flow lines with moderate to large sized K is compu-
tationally expensive or impossible due to the enormous resulting state space (see
Vidalis and Papadopoulos, 2001). Approximate methods are required to solve large
systems. The work reported here uses the decomposition algorithm developed by
Diamantidis, Papadopoulos and Heavey (2006). This decomposition method actu-
ally extends the work by Gershwin (1987) to solve flow lines with parallel servers at
each work-station.
The solution approach for solving large lines with parallel machines at each
work-station is as follows:
Following the derivation of the transition equations of the two-station system
using exact Markovian analysis, an algorithm for generating the transition matrix for
any two-station parallel system is developed. Thereafter, decomposition equations
are derived using the well-known two-step methodology of obtaining the conserva-
tion flow equations and the flow rate idle time equations. Finally, a decomposition
algorithm as outlined in Figure 2.22 was developed.
In the sequence, first, the decomposition equations are derived and then the
decomposition algorithm is presented.
2.5.4 Derivation of the decomposition equations
In general, the decomposition method makes use of the four sets of equations (see
Gershwin, 1994 where the decomposition method is described in great detail): (i) the
conversation of flow equations; (ii) the flow rate idle equations; (iii) the resumption
of flow equations; (iv) the interruption of flow equations. As the system analyzed
here is reliable, only the first two sets of equations are used.
Conservation of flow equations
Let XiL be the mean production rate of the two-work-station, one-buffer sub-line
Li and Xiu (Xid ) be the mean production rate of the virtual upstream (downstream)
pseudo work-station Miu (Mid ), i = 1, . . ., K − 1. The mean production rate of each
work-station Mi in the original line is denoted by Xi . The conservation of flow
equations states that the production rates of all the two-work-station, one-buffer
sub-systems Li are the same.
Because the flow is conserved, it holds:
XiL = Xiu = Xid = Xi , i = 1, . . ., K − 1. (2.103)
The flow rate idle time equations
Each virtual upstream pseudo work-station Miu of sub-line Li , i = 1, . . ., K − 1, con-

sists of Si parallel machines, while each virtual downstream pseudo work-station Mid
of line Li consists of Si+1 parallel machines. The service times of the Si parallel
76
S1 servers S2 servers S3 servers SK-1 servers SK servers
M1 B2 M2 B3 M3 …. MK–1 BK MK Line L
u d
M1 B2 M1 Line L1
M u2 B3 M d2 Line L2
.
.
.
SK–1 servers SK servers

u d
M K=1 BK M K=1 Line LK–1
Fig. 2.21. Flow line with K parallel-machine work-stations, K − 1 intermediate buffers (Line L) and decomposition scheme (Lines L1 , . . . , LK−1 )
{Step 1: Initialization}
for i = 1 to K − 1 do
μiu = μi
μid = μi+1
ε = small positive number for terminating condition
end for
{Step 2: Calculate μiu and μ dj }
Calculate μiu using equation (2.116)
Evaluate the two-work-station, one buffer sub-line Li−1 , using the most recent values of
μi−1
u and μ d in the algorithm presented in sub-section 2.5.1.
i−1
end for
j = K −i
Calculate μ dj using equation (2.117)
Evaluate the two-work-station, one buffer sub-line Li+1 , using the most recent values of
μi+1
u and μ d in the algorithm presented in sub-section 2.5.1.
i+1
end for
{Step 3: Terminating Conditions}

if XiL − X1L < ε , i = 2, . . . , K − 1 then
GOTO Step 4
else
GOTO Step 2
end if
{Step 4: Output Results}
X = XiL , i = 1, . . . , K − 1
Fig. 2.22. Decomposition algorithm

1
machines of pseudo work-station Miu are exponentially distributed with mean μiu ,
while the service times of the Si+1 parallel machines of pseudo work-station Mid are
also exponentially distributed with mean μ1d , i = 1, . . ., K −1. Also, define pic to be the
i
steady-state probability of state y = (c) for sub-line Li , where c = 0, . . . ,Ci . Defining
Ci = Si + Si+1 + Bi and taking into account equations (2.100) and (2.101), the mean
production rate of work-stations Mi−1d and M u is given by the following formulae:
i
c=Ci−1 c=Si −1
d
Xi−1 = Si μi−1
d
∑ c − μi−1
pi−1 d
∑ (Si − c)pi−1
c , i = 2, . . ., K (2.104)
c=0 c=0
c=Ci c=Ci
Xiu = Si μiu ∑ pic − μiu ∑ (c − Si+1 − Bi )pic , i = 1, . . ., K − 1. (2.105)
c=0 c=Si+1 +Bi +1
Rewriting equations (2.104) and (2.105):

c=Ci−1 c=Si −1
d
Xi−1 = μi−1
d
Si ∑ c −
pi−1 ∑ (Si − c)pi−1
c (2.106)
c=0 c=0
and

c=Ci c=Ci
Xiu = μiu Si ∑ pic − ∑ (c − Si+1 − Bi )pic . (2.107)
c=0 c=Si+1 +Bi +1
It is straightforward to show that:

c=Ci c=Ci−1
∑ pic = ∑ c = 1.
pi−1 (2.108)
c=0 c=0
The blocking probability pbl

i of each virtual two-work-station, one-buffer sub-line
Li is given by (derivation is omitted and the reader is addressed to Diamantidis,
Papadopoulos and Heavey, 2006):
c=Ci
i =
pbl ∑ (c − Si+1 − Bi )pic , i = 1, . . ., K − 1 (2.109)
c=Si+1 +Bi +1
whereas the starvation probability psti−1 of each virtual two-work-station, one-buffer

sub-line Li−1 is given by (derivation is omitted and the reader is addressed to
Diamantidis, Papadopoulos and Heavey, 2006):
c=Si −1
psti−1 = ∑ (Si − c)pi−1
c , i = 2, . . ., K. (2.110)
c=0
Substituting equations (2.108) and (2.109) into equation (2.107), the mean produc-
tion rate of the upstream work-station Miu is:
Xiu = μiu (Si − pbl

i ). (2.111)
Similarly, substituting equations (2.108) and (2.110) into equation (2.106), the mean
production rate of the downstream work-station Mi−1d is:
d
Xi−1 = μi−1
d
(Si − psti−1 ). (2.112)
The mean production rate of work-station i, Xi , of the original production line L is

given by:
Xi = μi (Si − psti−1 − pbl
i ). (2.113)
Calculating probabilities pbl s
i and pi−1 from equations (2.111) and (2.112), respec-
tively, one obtains:
Xiu
i = Si − u
pbl (2.114)
μi
and
d
Xi−1
psti−1 = Si − . (2.115)
μi−1
d
Substituting equations (2.114) and (2.115) into equation (2.113) and taking into
account conservation of flow in equation (2.103), the following two equations for
calculating μiu and μid can be derived:
1
μiu = Si
, i = 2, ..., K − 1 (2.116)
1
μi + Xi−1 − μ d1
i−1
1
μid = Si+1
, i = K − 2, ..., 1. (2.117)
1
μi+1 + Xi+1 − μ u1
i+1
Finally, because the virtual work-station M1u corresponds to the input work-
station M1 and the virtual work-station MK−1d corresponds to the output machine
MK of the original line L, the following boundary conditions are used:
μ1u = μ1 and μK−1

d
= μK . (2.118)
2.5.5 The decomposition algorithm
Using the above derived equations, a decomposition algorithm shown in Figure 2.22
was developed. The ε value used in all the numerical examples given here was
0.00001.
2.5.6 Numerical results
In order to evaluate the performance and the accuracy of the proposed decomposi-
tion algorithm, several numerical experiments have been conducted by Diamantidis,
Papadopoulos and Heavey (2006) for various configurations of production lines with
parallel machines at each work-station. Here, a few representative sample numeri-
cal results are given. First, results for short lines of up to 7 stations are presented
and compared to published results. Then, to illustrate the efficiency of the solution
method, sample results for long production lines are presented.
Comparison with published exact results—Short lines
In Diamantidis, Papadopoulos and Heavey (2006), results for short lines with up to
7 work-stations were compared against those reported in Hillier and So (1989, 1995,
1996). Hillier and So applied exact Markovian analysis to calculate the throughput
of small production lines with up to 7 work-stations in series. Here, in Table 2.12
and Table 2.13 sample results are given for lines with 5 stations, unbalanced lines
(processing rates of machines at different stations are not the same but the process-
ing rates of machines at any station are the same) and 3, 5 and 7 stations, balanced
lines (all machines in all stations have the same processing rate) with different
Table 2.12. Comparison of results with Hillier and So (1996) – 5 work-stations
s μ XDECO XHS96 % Error Time

(1,1,1,1,4) (2.0876, 2.7624, 3.0120, 3.4013, 0.2831) 1.0192 1.0240 0.469 0.01
(4,1,1,1,1) (0.2831, 3.4013, 3.0120, 2.7624, 2.0876) 1.0192 1.0210 0.176 0.01
(1,1,1,4,1) (2.0920, 2.8248, 3.2786, 0.2894, 2.4509) 0.9965 1.0120 1.532 0.01
(1,1,4,1,1) (2.1739, 3.1347, 0.2905, 3.1347, 2,1739) 0.9913 1.0100 1.851 0.01
(2,1,1,1,3) (0.6877, 2.9585, 2.9673, 3.1250, 0.3920) 0.9890 0.9790 1.021 0.01
(1,2,1,1,3) (2.0533, 0.7710, 3.0581, 3.1250, 0.3892) 0.9745 0.9730 0.154 0.01
(1,1,2,3,1) (1.9569, 2.6881, 0.7987, 3.2467, 0.3910) 0.9667 0.9690 0.237 0.01
(1,2,1,3,1) (2.0408, 0.7686, 3.2154, 0.4103, 2.1691) 0.9514 0.9610 0.999 0.01
(1,1,1,1,6) (3.0211, 4.0000, 4.4444, 5.1020, 0.2501) 1.4149 1.4130 0.134 0.01
(6,1,1,1,1) (0.2501, 5.1020, 4.4444, 4.0000, 3.0211) 1.4149 1.4090 0.419 0.01
Table 2.13. Comparison of results with Hillier (1995) – 3, 5 and 7 work-stations
K s XDECO XHS95 %Error Time

3 (15,16,16) 13.5500 13.5400 0.074 0.05
3 (1,2,2) 0.8846 0.8873 0.304 0.01
3 (30,32,30) 27.7800 27.6800 0.361 0.27
5 (1,1,2,1,1) 0.5541 0.5638 1.720 0.02
5 (1,2,1,2,1) 0.6677 0.6637 0.603 0.02
5 (1,2,2,2,2) 0.8716 0.8752 0.411 0.02
7 (1,1,2,1,2,1,1,) 0.5575 0.5613 0.677 0.05
7 (1,2,1,2,1,2,1) 0.6334 0.6320 0.222 0.05
servers allocation per station and zero buffer levels for all the intermediate buffers,
respectively.
In all tables, columns labeled by vectors s = (S1 , . . ., SK ), and μ = (μ1 , . . ., μK )
denote, respectively, the server allocation and the mean service rate allocation at
the respective work-stations of a production line with N work-stations. The column
labeled XDECO gives the estimated mean production rate using the proposed decom-
position algorithm by Diamantidis, Papadopoulos and Heavey (2006), while the
column labeled XHSXX (XX = 95, 96) is the published results given in Hillier and So
(1995, 1996). The percentage error between the results obtained from decomposition
and those reported in Hillier (1995, 1996) is computed using the following formula:
|XDECO − XHSXX |
% Error = × 100%, (2.119)
XHSXX
where XX denotes results reported in year 19XX.
Table 2.12 presents numerical results obtained for a production line consisting
of 5 work-stations with all the buffer capacities equal to zero. Column 4 presents
the mean production rate reported in Hillier (1996), (XHS96 ). Column 5 (% Error)
gives the percentage error between the results obtained from decomposition and
those reported in Hillier (1996). Column 6 gives the time (in seconds) taken by
decomposition to estimate the throughput. The decomposition algorithm was run

on a Pentium III at 450MHz with 256MB RAM.
Table 2.13 presents numerical results for production lines where the number of
work-stations K are 3, 5 and 7. It is also assumed that the processing rates of all
machines at each work-station are equal to 1 and that the buffer level for all the
intermediate buffers is 0. The first column (K) gives the number of work-stations,
whereas the fourth column gives the estimated throughput reported in Hillier (1995),
(XHS95 ). The fifth and sixth columns present the percentage error and the required
time by decomposition to obtain the results, respectively.
Comparison with simulation results—Long lines
The decomposition algorithm proposed in Diamantidis, Papadopoulos and Heavey

(2006) has also been used to estimate the throughput of large balanced and unbal-
anced production lines (up to 1000 and even more work-stations). This algorithm was
developed and coded by Dr. Alexandros Diamantidis in C++. To our knowledge,
there is no exact analytical method that can estimate the throughput of such large
lines. Diamantidis, Papadopoulos and Heavey (2006) applied their algorithm and
solved a large system with 1000 work-stations, each with 3 servers and 999 interme-
diate buffers each of buffer size equal to 10 buffer slots in approximately 50 minutes
on a Pentium IV computer. For comparison purposes, Diamantidis, Papadopoulos
and Heavey (2006) developed a simulation model in eM-Plant
(http://www.ugs.com/products/tecnomatix/plant_design/em_
plant.shtml) in order to compare the results obtained from the decomposition
algorithm for large production lines. Recall that this decomposition algorithm for
K = 2 stations specifies the two-station line, with parallel machines at each work-
station and an intermediate buffer,which is solved exactly and the throughput of the
system is calculated.
Si , and μi , i = 1, . . ., K, denote the number of parallel machines at work-station i
and the service rate of each one of the Si parallel machines at work-station i, respec-
tively. Bi , i = 2, . . ., K, denote the storage capacity of the buffer located in front of
work-station i, i = 2, . . . , K.
Table 2.14 presents configurations for 12 sample production lines varying from
10 to 120 stations in steps of 10. The results given for each example in Table 2.15
are: (i) throughput of the system calculated using simulation (XSIM ); (ii) throughput
of the system calculated using the decomposition algorithm (XDECO ); (iii) average
inventory for the system calculated using simulation (c̄SIM ); (iv) average inventory
for the system calculated using the decomposition algorithm (c̄DECO ). Table 2.15 also
presents 95% confidence intervals calculated for XSIM and c̄SIM . The % Error for X
was calculated using equation (2.119) with the % Error for c̄ calculated similarly.
The computing time, in seconds, to execute the decomposition algorithm is given
in the two columns labeled “X-Time” and “c̄-Time,” respectively. These columns
give the time to calculate the throughput and the average inventory, respectively. The
computing time, in seconds, to execute the simulation experiments is given in the last
column. The simulation model was run for 20,000 units/customers before statistics
Table 2.14. Sample configurations for long lines
Production Line 1 Production Line 2 Production Line 3

K = 10 K = 20 K = 30,
Si = 2, i = 1, . . ., 4 Si = 3, i = 1, . . ., 7 Si = 3, i = 1, . . ., 10
S5 = S6 = 3 S j = 4, j = 8, . . ., 15 S j = 4, j = 11, . . ., 20
Si = 2, i = 7, . . ., 10 Sk = 3, k = 16, . . ., 20 Sk = 3, k = 21, . . ., 30
Bi = 3, i = 1, . . ., 9 Bi = 4, i = 1, . . ., 19, Bi = 4, i = 1, . . ., 29 μi =
μi = 1, i = 1, . . ., 10 μi = 1, i = 1, . . ., 20 1, i = 1, . . ., 30
K = 40 K = 50 K = 60
Si = 2, i = 1, . . ., 15 Si = 2, i = 1, . . ., 10 Si = 4, i = 1, . . ., 20
S j = 3, j = 16, . . ., 25 S j = 1, j = 11, . . ., 40 S j = 5, j = 21, . . ., 40
Sk = 2, k = 26, . . ., 40 Sk = 2, k = 41, . . ., 50 Sk = 4, k = 41, . . ., 60
Bi = 3, i = 1, . . ., 49 Bi = 2, i = 1, . . ., 49 Bi = 4, i = 1, . . ., 59
μi = 1, i = 1, . . ., 50 μi = 1, i = 1, . . ., 50 μi = 1, i = 1, . . ., 60
K = 70 K = 80 K = 90
Si = 3, i = 1, . . ., 25 Si = 2, i = 1, . . ., 30 Si = 2, i = 1, . . ., 30
S j = 2, j = 26, . . ., 45 S j = 3, j = 31, . . ., 50 S j = 3, j = 31, . . ., 60
Sk = 3, k = 46, . . ., 70 Sk = 2, k = 51, . . ., 80 Sk = 2, k = 61, . . ., 90
Bi = 2, i = 1, . . ., 69 Bi = 2, i = 1, . . ., 79 Bi = 2, i = 1, . . ., 89
μi = 1, i = 1, . . ., 70 μi = 1, i = 1, . . ., 80 μi = 1, i = 1, . . ., 90
K = 100 K = 110 K = 120
Si = 2, i = 1, . . ., 40 Si = 5, i = 1, . . ., 40 Si = 3, i = 1, . . ., 50
Sk = 3, k = 41, . . ., 60 S j = 6, j = 41, . . ., 70 S j = 4, i = 51, . . ., 70
Sm = 2, m = 61, ..., 100 Sk = 5, i = 71, . . ., 110 Sk = 3, i = 71, . . ., 120
Bi = 2, i = 1, . . ., 99 Bi = 4, i = 1, . . ., 110 Bi = 3, i = 1, . . ., 120
μi = 1, i = 1, . . .100 μi = 1, i = 1, . . ., 110 μi = 1, i = 1, . . ., 120
were collected. The batch means method was used to collect 30 independent samples
within a single run. A batch size of 5000 units/customers was used.
From examination of Table 2.15 it can be observed that the maximum error for
the throughput is 1.72%. The accuracy of the decomposition algorithm for average
inventory is not as good with a maximum error of 16.27% observable in Table 2.15.
The convergence of the algorthm was found, for the majority of cases, to be very fast.
However, the convergence speed can vary considerably and is system dependent, as
can be observed in Table 2.15 where Line # 9 and 10 took approximately 50% of the
time it took to obtain results using simulation. For all the results of the decomposition
algorithm, a Pentium III at 450MHz with 256MB RAM was used. The simulation
experiments were carried out on a Pentium IV at 2992MHz with 1000MB of RAM.
Table 2.15. Sample numerical results for long lines
Line # Throughput Average Inventory Decomposition Time Simulation

XSIM XDECO % Error 95% CI c̄SIM c̄DECO % Error 95% CI X-Time c̄-Time Time
1 1.5242 1.5423 1.1739 1.5198−1.5285 32.95679 33.4985 1.6171 30.7799−35.1336 0.09 0.39 87.46
2 2.3884 2.3808 0.3183 2.3805−2.3961 87.02922 77.5526 12.2196 83.4917−90.5666 0.61 0.66 136.83
3 2.3315 2.3533 0.9269 2.3402−2.3506 143.4842 139.6802 2.7234 138.9421−148.0263 0.22 0.22 183.85
4 1.4387 1.4447 0.4184 1.4353−1.4419 135.6619 130.001 4.3545 131.2453−140.0785 54.65 56.35 252.26
5 0.6092 0.6036 0.9335 0.6081−0.6102 101.4165 116.9982 13.3179 97.5978−105.2352 0.98 0.99 303.57
6 3.1274 3.1662 1.2250 3.1204−3.1343 347.8995 321.6019 8.1771 340.8268−354.9722 1.43 1.26 366.95
7 1.3440 1.3412 0.2071 1.3407−1.3472 214.3427 255.9992 16.2721 208.7912−219.8942 13.46 1.27 437.96
8 1.3311 1.3389 0.5814 1.3283−1.3339 239.463 215.59 11.0733 233.5951−245.3308 42.24 62.47 477.1
9 1.3319 1.339 0.5311 1.329 −1.3347 262.4053 230.581 13.8018 256.2628−268.5477 244.69 249.64 562.71
10 1.3241 1.3366 0.9315 1.3212−1.327 279.8781 274.1059 2.1058 273.5344−286.2217 260.68 266.6 616.71
11 3.9448 4.0066 1.5418 3.9357−3.9538 768.0851 701.1589 9.5451 757.5761−778.5941 2.64 1.27 704.81
12 2.1926 2.2309 1.7175 2.1883−2.1968 529.9708 503.4035 5.2775 521.2414−538.7001 2.97 3.02 752.33
2.5 Modeling of Production Lines with Parallel Reliable Machines at Each Station
83
Table 2.16. Configurations for longer lines

K = 200 K = 300 K = 400
Si = 3, i = 1, . . ., 50 Si = 3, i = 1, . . ., 50 Si = 2, i = 1, . . ., 100
Si = 2, i = 51, . . ., 100 Si = 1, i = 51, . . ., 100 Si = 3, i = 101, . . ., 200
Si = 4, i = 101, . . ., 150 Si = 4, i = 101, . . ., 150 Si = 4, i = 201, . . ., 300
Si = 1, i = 151, . . ., 200 Si = 2, i = 151, . . ., 200 Si = 1, i = 301, . . ., 400
Bi = 2, i = 1, . . ., 199 Si = 3, i = 201, . . ., 300 Bi = 2, i = 1, . . ., 399
μi = 1, i = 1, . . ., 200 Bi = 3, i = 1, . . ., 299 μi = 1, i = 1, . . ., 400
μi = 1, i = 1, . . ., 300
K = 500 K = 600 K = 700
Si = 3, i = 1, . . ., 150 Si = 2, i = 1, . . ., 200 Si = 4, i = 1, . . ., 250
Si = 4, i = 151, . . ., 300 Si = 4, i = 201, . . ., 400 Si = 2, i = 251, . . ., 450
Si = 2, i = 301, . . ., 500 Si = 3, i = 401, . . ., 600 Si = 4, i = 451, . . ., 700
Bi = 3, i = 1, . . ., 499 Bi = 2, i = 1, . . ., 599 Bi = 3, i = 1, . . ., 699
μi = 1, i = 1, . . ., 500 μi = 1, i = 1, . . ., 600 μi = 1, i = 1, . . ., 700
K = 800 K = 900 K = 1000
Si = 4, i = 1, . . ., 200 Si = 4, i = 1, . . ., 400 Si = 4, i = 1, . . ., 400
Si = 3, i = 201, . . ., 400 Si = 3, i = 401, . . ., 500 Si = 3, i = 401, . . ., 600
Si = 2, i = 401, . . ., 600, Si = 4, i = 501, . . ., 900 Si = 4, i = 601, . . ., 1000
Si = 3, i = 601, . . ., 800 Bi = 3, i = 1, . . ., 899 Bi = 3, i = 1, . . ., 999
Bi = 3, i = 1, . . ., 799 μi = 1, i = 1, . . ., 900 μi = 1, i = 1, . . ., 1000
μi = 1, i = 1, . . ., 800
In Table 2.17 throughput results are given for the configurations shown in Table
2.16 which include even longer production lines (with K = 200(100)1000 work-
stations). As it can be seen from Table 2.17, the maximum error was found to be
2.5%. The parameters s, number of parallel stations, n, the buffer sizes, and μ , the
mean service times, vary arbitrarily, so as to illustrate the versatility of the algorithm.
Run times, in seconds, for the decomposition algorithm and the simulation model
are given in the last two columns. From Table 2.17 it can be noted that XDECO
lies outside the 95% CI for Line # 16–21. In general it was found that XDECO was
outside the 95% CI for configurations with % Error greater than 1.00.
The numerical results presented in Table 2.15 and in Table 2.17 indicate that
the decomposition algorithm is very accurate. The average percentage error of the
throughput obtained from the proposed decomposition algorithm and simulation for
lines with up to 100 stations is less than 1%, whereas the results presented for lines
with up to 1000 stations indicate that the percentage error is less than 2.5%. The
convergence of the algorithm is very fast and reliable. Diamantidis, Papadopoulos
and Heavey (2006) claim that they have not found a case in which the algorithm
does not converge.
2.6 Simulation Modeling 85
Table 2.17. Numerical results for longer lines
Line # XSIM XDECO % Error 95% CI Decomposition Simulation

Time Time
13 0.6047 0.6014 0.5444 0.6036–0.6058 7.58 1340.72
14 0.6620 0.6678 0.8688 0.6610–0.6631 15.05 1478.35
15 0.6001 0.6004 0.0546 0.5994–0.6008 131.77 3104.08
16 1.3984 1.4287 2.1651 1.3963–1.4012 369.16 3673.55
17 1.3089 1.3335 1.8763 1.3073–1.3105 347.08 3188.14
18 1.3988 1.4287 2.1411 1.3956–1.4012 361.09 4618.73
19 1.3982 1.4287 2.1790 1.3960–1.4003 571.55 5626.19
20 2.1802 2.2292 2.2470 2.1766–2.1838 239.97 5887.03
21 2.1759 2.2287 2.4276 2.1722–2.1796 804.60 7162.45
The above algorithm is available at the website associated with this book, with
abreviated name DECO-2. As noted above if K, the number of stations is equal
to 2, the exact numerical solution to the two-station production line with identical
perfectly reliable parallel machines at each station may be obtained. In addition, if
the number of machines at each station is set equal to 1, the authors have shown that
the results obtained for large production lines with single machine stations, expo-
nential service times, perfectly reliable machines and intermediate buffers of finite
capacity replicate the decomposition equations originally given by Gershwin (1994).
2.6 Simulation Modeling
Simulation of production lines is a powerful tool in obtaining the performance mea-

sures where analytical methods are either difficult or impossible to use. In the past,
simulation was often considered to be an expensive and time consuming approach to
the solution of system type problems. However, with the increase of computer power
and the availability of special-purpose simulation languages, such constraints are less
severe. Usually, in simulation studies of production lines what is technically involved
is Monte Carlo simulation because of the inherent stochastic variability of these sys-
tems. The combination of simulation studies with analytical studies is probably the
way of the future in the design of production lines.
Ideally, the production line analyst and designer requires a discrete event simu-
lation package with Monte Carlo simulation capabilities, graphical and other output
reporting facilities together with a relatively easy method for the statistical assess-
ment of results. In simulation modeling, the modeler must specify very carefully how
the production line is meant to operate and the various disciplines and rules which
are involved. The basis of discrete event simulation is that the system state at any
time t is stored and that the state only changes when a particular event occurs. The
specification of the state of the system depends on the detail required by the modeler
in respect to the performance characteristics of the system. In all simulation stud-
ies there is a need to consider the time horizon of the process under investigation.
Short term system performance analysis requires that data be taken from the simula-
tion during a short time horizon. On the other hand, steady-state simulation models
are appropriate for the analysis of systems which in theory could run indefinitely.
Usually, in production line modeling the modeler is interested in steady-state behav-
ior of the system by which time the precise initial state of the system has little impact.
It is normal in these cases to have a “warm-up” period before recording data for the
calculation of the steady-state behavior. Graphical output of the performance param-
eters of the system can be extremely useful in determining when the warm-up period
is ended. In some simulation packages it is possible to specify in advance the time
to allow the system to settle down. The appropriateness of this time may in fact be
checked from the results of the simulation. Finally, it is usual to place confidence
limits on the values of the permormance parameters of the system, based on certain
assumptions, and such limits may usually be incorporated into modern simulation
models.
As an illustration of the power of simulation, Arena, a simulation software pack-
age available to the authors, has been used to model a system of the type depicted in
Figure 2.23, with the following characteristics:
• The line consists of K = 4 work-stations with identical parallel machines at each
work-station. The number of machines at station i, i = 1, . . . , 4 are 3, 2, 2, 3,
respectively. Service or processing times are exponentially distributed and the
mean service rates of the identical machines at each work-station are μi = 1,
i = 1, . . . , 4. Thus, the probability that a service is completed on a machine at station
i in a time interval Δt is μi Δt.
• All machines are assumed to be perfectly reliable.
• The interstation buffers Bi , i = 2, 3, 4 have capacities of 4, 2, 4, respectively.
• All products produced conform to specifications.
• Transfer times between stations and buffers and between buffers and stations are
considered negligible.
• Any particular machine may be blocked after service due to the finite capacities of
the buffers excluding the last set of parallel machines.
• Arrangements are made to ensure that the first station is always busy, i.e., never
starved or it is saturated; any machine at any other station may be starved.
• No machine of any of the group of machines at any particular station is given prior-
ity in relation to being unblocked when unblocking occurs. Likewise in relation to
the resumption of production at a work-station following the removal of starvation.
The following performance measures of the line were determined:

• Throughput (jobs exiting from the production line per unit time).
• Utilization of each work-station (the limit of the time average of the number of
busy machines over time divided by the total number of machines in the station).
• Average buffer level for each intermediate buffer.
• Average work-in-process, W IP, excluding the buffer before the first station.
• Average job waiting time at each of the intermediate buffers.
m1 m4
m2 m3
B2 B3 B4
m1 m4
m2 m3
m1 m4
S1 = 3 servers S2 = 2 servers S3 = 2 servers S4 = 3 servers
Fig. 2.23. A production line with four stations with parallel reliable machines at each station and three intermediate buffers
2.6 Simulation Modeling
87
To ensure saturation of the first station of the line, the capacity of the queue in front
of the first station was set at twenty (20) units and batches arrive at sizes of five (5),
and the arrival rate into the system, λ , was taken to be greater than 3 μ1 . Finally, a
warm-up period of 100 minutes was specified to ensure that steady-state conditions
were obtained. (More details of this Arena simulation model and numerical results
are given in Appendix E.)
It may be of interest to discuss briefly possible extensions of the relatively simple
simulation model developed above. Clearly, the number of stations and the number
of parallel machines at each station could be modified. Buffer capacities could be
changed, and the restriction of identical machines at each work-station could be
removed which would impact on the scheduling rules. Unreliable machines could
be incorporated. Processing times, repair times, times to failures and transport times
from stations to buffers and from buffers to stations may be modeled using the Monte
Carlo simulation. There is no need to confine the distributions to be exponential as
Erlang, phase-type, normal, uniform or deterministic distributions may be modeled
using Monte Carlo simulation. The production of defective items at any station could
be incorporated into the model via a feedback mechanism for rework if necessary to
earlier stations.
2.7 General Comment

The reader will note that the Markovian method gives an exact evaluation of the state
probabilities from which an exact evaluation of the throughput may be obtained,
whereas the other three methods (aggregation, decomposition and expansion) give
approximate results for the throughput. Clearly, the limitation in using Marko-
vian methods is the smaller number of states which may be handled effectively in
contrast to the approximate methods at the expense of accuracy. The existence of
parallel-machine stations introduced complications on two fronts, viz., significantly
increasing the number of states and the difficulties of such elements being incor-
porated in the early classical decomposition methods. The expansion method gave
early hope that some of the problems with the modeling of parallel machines would
be reduced, but issues relating to accuracy of results still remain. It appears that the
results reported in Section 2.5 overcome a number of the problems associated with
earlier attempts to use decomposition methods in systems with parallel structures. In
principle, it is possible to derive the states for Markovian type of solution for such
systems, but unfortunately this would be computationally very inefficient for any
realistic production lines with parallel-machine stations. All methods presented are,
of course, in principle applicable to unreliable as well as prefectly reliable machines.

Researchers and practitioners alike should appreciate that there is a very rich litera-
ture applicable to the general area of the performance evaluation of production lines.
Here an attempt is made to highlight the main strands of thought in this area. It should
2.8 Related Bibliography 89
be noted that work intitiated for application areas quite diverse from manufacturing
has been found to be fruitful when applied to the analysis of production lines. Cases
in point are analyses originally oriented toward computer performance modeling and
communication networks have subsequently given insights into problems germane
to the analysis of production lines. Basically, the mathematical underlined theory of
production line analysis is queueing theory, in particular, queueing networks with
blocking. An exceptional reference in this area is the book by Perros (1994). How-
ever, care must be taken to ensure the validity of the model of the production system
in that for example blocking in a communication system tends to occur before the
service starts, whereas in a production line blocking occurs after the service has been
completed. What is offered below is a classification by the authors of what they
believe to be significant and somewhat distinct areas of research of value in the anal-
ysis of the performance of production lines. As far as Markovian analysis approaches
are concerned, some five areas of work have been identified and are described below.
It should be noted that there is nothing unique about this categorization and indeed
some authorities might well question the relative influence accorded to the work of
particular researchers.
The exact solution of small production lines was initiated by Hunt (1956) fol-
lowed by Buzacott (1972), Gershwin and Berman (1981) and Gershwin and Schick
(1983), among others. Solutions were obtained for two/three stations with limited
inter-station buffers, and methods of solution used included matrix recursive and
matrix geometric methods applied to the underlying Markov chains. Initially, expo-
nentially distributed processing times were only considered, but the work of Buzacott
and Kostelski (1987) extended the distribution of processing times to phase-type
distributions.
Altiok (1997) in a seminal work summarized and developed the earlier work by
Altiok and Ranjan (1989), Buzacott and Kostelski (1987) and Perros (1994), among
others, and brought phase-type modeling to its present position. Exact analysis of
small-scale production lines with any type of processing and repair time distributions
may be undertaken. Arising out of Altiok’s work it is possible to perform approx-
imate analysis of larger systems with any general distributions of processing and
repair times by the approximation of these distributions by phase-type distributions
by matching their first two or three moments.
When faced with the analysis of relatively large production lines, there is a
need for efficient computational procedures due essentially to the large number
of associated states of the underlying Markov chain of such systems. Hillier and
Boling (1967) developed a numerical approach for solving reliable exponential and
Erlang production lines. Papadopoulos and O’Kelly (1989), Papadopoulos, Heavey
and O’Kelly (1989, 1990) and Heavey, Papadopoulos and Browne (1993) further
developed this work by producing efficient numerical algorithms for generating the
transition matrices for reliable and unreliable production lines with exponential and
Erlang processing and repair time distributions and efficient solution methods. Fur-
ther extensions in this area are included in the book of Altiok (1997), as noted above,
by using the mixed generalized exponential distributions (phase-type distributions).
The algorithm included at the website associated with this book with abbreviated
name MARKOV for the generation of the transition matrix and the solution of the
associated steady-state Markov equations is based on the work of Papadopoulos,
O’Kelly and Heavey.
In contrast to continuous parameter discrete state Markov process analysis of
production lines, Muth (1984) introduced the concept of the holding time model
where the focus is on the three possible states of each station, viz., the station is
idle, busy or blocked. Alkaff and Muth (1987) extended Muth’s model to solve
K-station production lines with an arbitrary number of stations. A major advantage of
the holding time model is that the number of separate non-linear equations that have
to be solved is significantly reduced in comparison to the Markovian situation. The
price paid for this reduction is the need to solve non-linear equations that are of the
form of a fixed point problem. Holding time models can accommodate Erlang and
phase-type distributions more readily than can Markovian methods again because
of the reduction in the number of states. However, the holding time model cannot
accommodate intermediate buffers of non-zero capacity.
It is of great assistance to designers to have simple closed form formulae to
determine the throughput of production lines. A number of such formulae have
been developed based on insights from general queueing theory, considerations and
sometimes curve-fitting. Hunt (1956) was an early developer of such a closed form
expression. Makino (1964), Muth (1984) and Muth (1987) offered formulas for the
exponential and two-phase Erlang and distribution-free cases with no intermedi-
ate buffers between successive stations. Freeman (1968) and Anderson and Moodie
(1969) obtained empirical formulas for utilization of the production line, based on
regression analysis of various sets of simulation data. Knott (1970) offered a formula
based on theoretical and intuitive reasoning. Blumenfeld (1990) extended Muth’s
formula for throughput of a production line with variable processing times and
buffers of finite capacities. Haydon (1973) dealt with approximations in his Ph.D.
dissertation and he provided approximate throughput formulae that perform quite
satisfactorily. Papadopoulos (1996) using Muth’s holding time model developed an
analytical formula for the throughput of a K-station production line with no interme-
diate buffers and exponential processing times which may be different at the various
stations of the line. A particular simpler formula was developed for the balanced line.
The limitations of seeking exact solutions to production line problems are related
to problems arising from the number of states of such systems and the difficulties
associated with a numerical approach. Thus, there has been considerable interest in
developing approximate methods of analysis. Most approximate methods are based
on decomposition and an esssential element of this approach is that the sub-lines
used have exact solutions. Decomposition methods are approximations as the sub-
lines used are simpler than the original line and the equations used to develop the
parameters may also be approximated to facilitate the numerical analysis. Earlier
work on decomposition methods include Zimmern (1956), Sevast’yanov (1962) and
Hillier and Boling (1967). Queueing networks with blocking were decomposed by
Caseau and Pujolle (1979), Takahashi et al. (1980), and Boxma and Konheim (1981).
Altiok and Perros and their teams have made significant contributions by working on
decomposition to solve large systems with exponential and phase-type distributed
processing times. This work is reported in papers including Altiok (1982), Perros
and Altiok (1986), Jun and Perros (1987), Brandwajn and Jow (1988), Altiok (1989),
Altiok and Ranjan (1989) and Gun and Makowski (1989). Excellent expositions of
this work are given in the book of Altiok (1997). Gershwin (1987) in a well-known
article offered an efficient decomposition method for the approximate evaluation of
tandem queues with finite intermediate buffers and blocking. Dallery, David and Xie
(1988) improved the convergence of Gershwin’s algorithm. An excellent review of
flow line models is given in Dallery and Gershwin (1992), and the decomposition
approaches are treated in detail in the book of Gershwin (1994). Decomposition
models of various types of manufacturing systems are also included in the seminal
work of Buzacott and Shanthikumar (1993). Dallery and Frein (1993) classified the
various decomposition methods for solving production lines into one of three classes
according to the sets of decomposition equations used by the various authors.
Many papers concerned with the analysis of production lines have reported
results on simulation studies. It is virtually impossible to give an adequate review
of such papers from the perspective of the use of simulation method in the deter-
mination of production line performance. Nevertheless, there are a number of books
and research papers which are certainly worth further detailed study and investigation
by analysts specifically interested in simulation. These include the books by Altiok
and Melamed (2001), Kouikoglou and Phillis (2001), Law and Kelton (2000), Guide
to Arena Standard Edition by Systems Modeling Corporation (1999), Banks et al.
(1999), Kelton et al. (1998), Benson (1996), Khoshnevis (1994), Papadopoulos et al.
(1993), Brateley et al. (1987), Pritsker (1986) and Fishman (1973), among others.
Decomposition techniques have also been applied not only to manufacturing sys-
tems but also to computer systems (see Perros, 1994 and many references therein),
among others, and to more general manufacturing systems. For example,
Tempelmeier and Burger (2001a) examined non-homogeneous asynchronous flow
production systems and presented an analytical approximation for the performance
of such systems. They assumed generally distributed stochastic processing times
as well as breakdowns and imperfect production. The proposed approximation was
based on the decomposition of an K-station-line into (K − 1) two-station-lines that
were analyzed using a GI/G/1/Nmax queueing model. They also presented numeri-
cal comparisons with exact and simulation results which indicated that the procedure
provides accurate results. In Kuhn (2003) an analytical approach was given for
performance evaluation of an automated flow line system which considers the depen-
dency between the production and the repair system. The proposed model and
solution approach may be used in the initial design phase as well as during a redesign
process in order to evaluate various configurations of the production and repair
systems.
Tolio and Matta (1998) presented an elegant decomposition approach for the
performance evaluation of automated flow lines with multiple failure modes. The
decomposition block that was used in their analysis was solved exactly by a method
that is independent of the buffer size. An extension of the decomposition approach
for the performance evaluation of a flow line with linear flow of material and two part
types was presented by Nemec (1999). A different efficient decomposition analysis
for serial flow lines with two part types, deterministic identical processing times
and multiple failure modes was proposed by Colledani, Matta and Tolio (2003).
Flow lines with single machine work-stations and non-linear flow of material are
examined in Helber (1999), where a detailed analysis of flow lines with split and
merge operations is presented, Gershwin et al. (2001), Helber and Mehrtens (2003),
Tan (2001) and Helber and Jusic (2004).
Tolio, Matta and Gershwin (2002) presented an analytical method for the per-
formance evaluation of production lines with two unreliable machines and one
intermediate buffer of finite capacity. Each machine can fail in more than one way.
Levantesi, Matta and Tolio (1999a, b) developed an efficient decomposition
method for the performance evaluation of production lines with exponential pro-
cessing times, multiple failure modes and finite buffer capacities. The different types
of failures are distributed according to different exponential distributions as are the
times to repair.
Levantesi, Matta and Tolio (2003) provided an approximate analytical method
for the performance evaluation of asynchronous production lines with deterministic
processing times, multiple failure modes and finite buffer capacity. In their analysis,
the authors approximated the discrete flow of parts by a continuous flow of material.
Literature is relatively scarce on the analysis of flow lines with multiple identi-
cal parallel-machine work-stations. Friedman (1965) presented a reduction method
that reduces a queueing system with parallel-machine work-stations to corresponding
problems for a system of fewer stages. It was also assumed that for any sequence of
customer arrival times, the time spent in the system was independent of the order of
stages. Forestier (1980) examined automated flow lines where each station consists
of two parallel machines. Dubois and Forestier (1982) considered similar systems
using Markovian analysis. Iyama and Ito (1987) considered a flow line where some
work-stations have different numbers of parallel machines and unequal service rates.
They presented the effects of server allocation on the maximum average production
rate by using a Markovian model.
The exact solution of the two-station production line with the first station satu-
rated is based on queueing theory and a good exposition of this analysis may be found
in the book by Perros (1994), in the book by Buzacott and Shanthikumar (1993) and
in the book by Neuts (1981), among others. Details of the generation of the asso-
ciated conservative matrix, A, and a method for the calculation of the throughput
of such systems based on the elements of A are given in the paper by Vidalis and
Papadopoulos (2001). In addition, the recursive relationship for the number of states
of a general production line with K ≥ 2 parallel stations is derived in this paper.
With respect to the approximate solutions of larger systems there are a few research
studies of interest. These include the book by Buzacott and Shanthikumar (1993),
where an iterative procedure is applied to calculate the throughput of the long line
using the solution to the two-station line described above. In the paper by Jain and
MacGregor Smith (1994), the expansion method was used to approximate the per-
formance measures of each parallel station of the production line. In this paper, apart
from the series system, merge and splitting topologies were also analyzed.
Another paper of major interest is that by Patchong and Willaeys (2001), where
each set of parallel machines is replaced by an equivalent single machine at each
station of the production line. Then, existing methods may be used to derive the
performance measures of the original system. A similar approximation method was
applied by Jeong and Kim (1999) for performance analysis of assembly/disassembly
systems with parallel machine stations. Earlier, Caseau and Pujolle (1979) derived
the throughput of some specialized telecommunication models using repeated trials
methods.
In van Dijk and van der Wal (1989) computationally attractive lower and upper
bounds for finite multi-server exponential tandem queues were presented. A proof of
the bounds and related monotonicity results were also presented, which were based
on Markov reward theory. Gosavi and MacGregor Smith (1995) developed computa-
tionally efficient bounds and approximations for the performance measures of series
parallel queueing networks. They approximated analytically the throughput of a sys-
tem with two tandem exponential queues and extended their analysis to elementary
merge and split queuing networks.
Ancelin and Semery (1987) described a method that replaces each parallel-
machine work-station by an equivalent single machine work-station. The processing
rate of the equivalent work-station equals the sum of the processing rates of all
parallel machines in the work-station. The failure rate and repair rate of the equiv-
alent work-station are given by a formula which incorporates the failure and repair
parameters of the parallel machines in the work-station.
Burman (1995) applied a similar method that replaces each parallel server
work-station by a single equivalent work-station for the case of continuous flow of
material. The author assumed that the equivalent work-station has a maximum pro-
cessing rate which equals the sum of the processing rates of the parallel machines.
The failure and repair parameters of the equivalent work-station are calculated by
using the assumption that all parallel machines at a specific work-station operate
independently.
Cheah and MacGregor Smith (1994) showed how a M/G/C/C state dependent
queuing model is embodied into the modeling of large-scale facilities where the
blocking phenomenon can be or cannot be controlled. They also presented an approx-
imation technique based on the expansion method to incorporate the M/G/C/C
queuing models into series, merge and splitting topologies of production lines. Jain
and MacGregor Smith (1994) presented an analytical technique to calculate sys-
tem performance measures of M/M/C/K queuing networks. They analyzed series,
merge and splitting topologies and in addition they explored the optimal order of the
M/M/C/K servers in such systems.
In Diamantidis, Papadopoulos and Heavey (2006), a flow line with parallel
machines at each work-station is analyzed via the decomposition method which was
presented in Section 2.5.2. The proposed approach differs from those of Ancelin and
Semery (1987), Burman (1995), Jeong and Kim (1999) and Patchong and Willaeys
(2001), in that each parallel-machine work-station is not replaced by an equivalent
work-station. That is, the decomposition approach is applied directly to each one of
the parallel machines for each work-station without using replacement techniques.
It is expected that this direct approach will provide more accurate results than do the
replacement techniques.
Regarding the non-linear flow lines, the material in the text is based on the paper
by Diamantidis, Papadopoulos and Vidalis (2004). An excellent exposition of this
area is given in the book by Helber (1999) in which various non-linear flow models
are analyzed. Models using continuous variables are given by Tan (2001) and by
Helber and Mehrtens (2003), in which times to failure and repair are exponentially
distributed. Other relevant papers include Gershwin (1991), Jeong and Kim (1998),
Yu and Bricker (1993), Ammar and Gershwin (1989), Dallery, Liu and Towsley
(1994), Di Mascolo et al. (1991), Frein et al. (1996), Helber (1998), among others.
References
1. Alkaff, A. and Muth, E.J. (1987), The throughput rate of multistation production lines
with stochastic servers, Probability in the Engineering and Informational Sciences,
Vol. 1, pp. 309–326.
2. Altiok, T. (1982), Approximate analysis of exponential tandem queues with blocking,
European Journal of Operational Research, Vol. 11, pp. 390–398.
3. Altiok, T. (1989), Approximate analysis of queues in series with phase-type service
times and blocking, Operations Research, Vol. 37, pp. 601–610.
4. Altiok, T. (1997), Performance Analysis of Manufacturing Systems, Springer.
5. Altiok, T. and Ranjan, R. (1989), Analysis of production lines with general service
times and finite buffers: A two-node decomposition approach, Engineering Costs and
Production Economics, Vol. 17, pp. 155–165.
6. Altiok, T. and Melamed, B. (2001), Simulation Modeling and Analysis with Arena,
Cyber Research, Inc. and Enterprise Technology Solutions, Inc.
7. Ammar, M. and Gershwin, S.B. (1989), Equivalence relations in queueing models of
fork/join queueing networks with blocking, Performance Evaluation, Vol. 10, pp. 233–
245.
8. Ancelin, B. and Semery, A. (1987), Calcul de la productivite d’une ligne integree de
fabrication: CALIF, une methode analytique industrielle, RAIRO, APII, Vol. 21, No. 3,
pp. 209–238.
9. Anderson, D.R. and Moodie, C.L. (1969), Optimal buffer storage capacity in production
line systems, International Journal of Production Research, Vol. 7, pp. 233–240.
10. Banks, J., Carson, J.S., and Nelson, B.L. (1999), Discrete-Event System Simulation,
2nd Edition, Prentice Hall.
11. Benson, D. (1996), Simulation Modeling and Optimization using PROMODEL,
PROMODEL Corporation, Orem, Utah.
12. Blumenfeld, D.E. (1990), A simple formula for estimating throughput of serial pro-
duction lines with variable processing times and limited buffer capacity, International
Journal of Production Research, Vol. 28, No. 6, pp. 1163–1182.
13. Boxma, O.J. and Konheim, A.G. (1981), Approximate analysis of exponential queueing
systems with blocking, Acta Informatica, Vol. 15, pp. 19–66.
14. Brandwajn, A. and Jow, Y.L. (1988), An approximation method for tandem queues with
blocking, Operations Research, Vol. 36, pp. 73–83.
15. Brateley, P., Fox, B.L., and Schrage, L.E. (1987), A Guide to Simulation, Springer-
Verlag.
References 95
16. Burman, M.H. (1995), New Results in Flow Line Analysis, Ph.D. Thesis, Massachusetts
Institute of Technology (M.I.T.).
17. Buzacott, J.A. (1972), The effect of station breakdowns and random processing times
on the capacity of flow lines with in-process storage, AIIE Transactions, Vol. 4, No. 4,
pp. 308–313.
18. Buzacott, J.A. and Kostelski, D. (1987), Matrix-geometric and recursive algorithm
solution of a two-stage unreliable flow line, IIE Transactions, Vol. 19, pp. 429–438.
19. Buzacott, J.A. and Shanthikumar, J.G. (1993), Stochastic Models of Manufacturing
Systems, Prentice Hall.
20. Caseau, P. and Pujolle, G. (1979), Throughput capacity of a sequence of transfer lines
with blocking due to finite waiting room, IEEE Transactions Software Eng., SE-5,
pp. 631–642.
21. Cheah, J.Y. and MacGregor Smith, J. (1994), Generalized M/G/C/C state dependent
queueing models and pedestrial traffic flows, Queueing Systems, Vol. 15, pp. 365–386.
22. Colledani, M., Matta, A., and Tolio, T. (2003), Performance evaluation of production
lines with finite buffer capacity producing two different products, In Proceedings of
the Fourth Aegean International Conference on Analysis of Manufacturing Systems, pp.
231–240, Samos Island, Greece.
23. Dallery, Y., David, R., and Xie, X.L. (1988), An efficient algorithm for analysis of
transfer lines with unreliable machines and finite buffers, IIE Transactions, Vol. 20,
pp. 280–283.
24. Dallery, Y. and Frein, Y. (1993), On decomposition methods for tandem queueing
networks with blocking, Operations Research, Vol. 41, No. 2, pp. 386–399.
25. Dallery, Y. and Gershwin, S.B. (1992), Manufacturing flow line systems: A review of
models and analytical results, Queueing Systems Theory and Applications, Vol. 12,
pp. 3–94.
26. Dallery, Y., Liu, Z., and Towsley, D. (1994), Equivalence, reversibility, symmetry
and concavity properties in fork-join queuing networks with blocking, Journal of the
Association for Computing Machinery, Vol. 41, No. 5, pp. 903–942.
27. Diamantidis, A.C., Papadopoulos, C.T., and Heavey, C. (2006), Approximate analysis
of serial flow lines with multiple parallel-machine stations, IIE Transactions, Vol. 39,
No. 4, pp. 361–375.
28. Diamantidis, A.C., Papadopoulos, C.T., and Vidalis, M.I. (2004), Exact analysis of
a discrete material three station one buffer merge system with unreliable machines,
International Journal of Production Research, Vol. 42, No. 4, pp. 651–675.
29. Di Mascolo, M., David, R., and Dallery, Y. (1991), Modeling and analysis of assembly
systems with unreliable machines and finite buffers, IIE Transactions, Vol. 23, No. 4,
pp. 315–330.
30. Dubois, D. and Forestier, J.P. (1982), Productivité et en-cours moyens d’un ensemble de
deux machines separées par une zone de stockage, RAIRO Automatique, Vol. 16, No. 2,
pp. 105–132.
31. Fishman, G.S. (1973), Concepts and Methods in Discrete Event Simulation, Wiley.
32. Forestier, J.P. (1980), Modelisation stochastique et comportement asymptotique d’un
systéme automatise de production, RAIRO Automatique, Vol. 14, No. 2, pp. 127–143.
33. Freeman, D.R. (1968), A general line balancing model, Proceedings of the 19th Annual
Conference, AIIE, Tampa, Florida, Norcross, Georgia: American Institute of Industrial
Engineers, pp. 230–235.
34. Frein, Y., Commault, C., and Dallery, Y. (1996), Modeling and analysis of closed-loop
production lines with unreliable machines and finite buffers, IIE Transactions, Vol. 28,
pp. 545–554.
35. Friedman, H.D. (1965), Reduction methods for tandem queueing systems, Operations
Research, Vol. 13, pp. 121–131.
36. Gershwin, S.B. (1987), An efficient decomposition method for the approximate eval-
uation of tandem queues with finite storage space and blocking, Operations Research,
Vol. 35, pp. 291–305.
37. Gershwin, S.B. (1991), Assembly/disassembly systems: An efficient decomposition
algorithm for tree structured networks, IIE Transactions, Vol. 23, No. 4, pp. 302–314.
38. Gershwin, S.B. (1994), Manufacturing Systems Engineering, Prentice Hall.
39. Gershwin, S.B. and Berman, O. (1981), Analysis of transfer lines consisting of two
unreliable machines with random processing times and finite storage buffers, AIIE
Transactions, Vol. 13, No. 1, pp. 2–11.
40. Gershwin, S.B., Maggio, N., Matta, A., Tolio, T., and Werner, L. (2001), Analysis of loop
networks by decomposition, In Proceedings of the Third Aegean International Confer-
ence on Analysis and Modeling of Manufacturing Systems, pp. 239–248, Tinos Island,
Greece.
41. Gershwin, S.B. and Schick, I.C. (1983), Modeling and analysis of three-stage transfer
lines with unreliable machines and finite buffers, Operations Research, Vol. 31, No. 2,
pp. 354–380.
42. Gosavi, H.D. and MacGregor Smith, J. (1995), Asymptotic bounds of throughput in
series-parallel queueing networks, Computers & Operations Research, Vol. 22, No. 10,
pp. 1057–1073.
43. Gun, L. and Makowski, A. (1989), An approximation method for general tandem queue-
ing systems subject to blocking, Proc. 1st Int’l Workshop on Queueing Networks with
Blocking, Raleigh, NC.
44. Haydon, B.J. (1973), The behaviour of systems of finite queues, Ph.D. Thesis, The
University of New South Wales, Kensington, New South Wales, Australia.
45. Heavey, C., Papadopoulos, H.T., and Browne, J. (1993), The throughput rate of multi-
station unreliable production lines, European Journal of Operational Research, Vol. 68,
pp. 69–89.
46. Helber, S. (1998), Decomposition of unreliable assembly/disassembly networks with
limited buffer capacity and random processing times, European Journal of Operational
Research, Vol. 109, No. 1, pp. 24–42.
47. Helber, S. (1999), Performance Analysis of Flow Lines with Non-Linear Flow of
Material, In Volume 473 of Lecture Notes in Economics and Mathematical Systems,
Springer-Verlag.
48. Helber, S. and Jusic, H. (2004), A new decomposition approach for non-cyclic continu-
ous material flow lines with a merging flow of material, Annals of Operations Research,
Vol. 124, No. 1–4, pp. 117–139.
49. Helber, S. and Mehrtens, N. (2003), Exact analysis of a continuous material merge
system with limited buffer capacity and three stations, In S.B. Gershwin, Y. Dallery,
C.T. Papadopoulos, and J. MacGregor Smith, Editors, Analysis and Modeling of
Manufacturing Systems, pp. 85–121, Kluwer Academic Publishers.
50. Hillier, F.S. and Boling, R.W. (1967), Finite queues in series with exponential or Erlang
service times – A numerical approach, Operations Research, Vol. 15, pp. 286–303.
51. Hillier, F.S. and So, K.C. (1989), The assignment of extra servers to stations in tandem
queueing systems with small or no buffer, Performance Evaluation, Vol. 10, pp. 219–
231.
52. Hillier, F.S. and So, K.C. (1995), On the optimal design of tandem queueing systems
with finite buffers, Queueing Systems, Vol. 21, pp. 245–266.
References 97
53. Hillier, F.S. and So, K.C. (1996), On the simultaneous optimization of server and
work allocations in production line systems with variable processing times, Operations
54. Hunt, G.C. (1956), Sequential arrays of waiting lines, Operations Research, Vol. 4,
pp. 674–683.
55. Iyama, T. and Ito, S. (1987), The maximum production rate for an unbalanced multi-
server flow line system with finite buffer storage, International Journal of Production
56. Jain, S. and Smith, J.M. (1994), Open finite queueing networks with M/M/C/K parallel
servers, Computers & Operations Research, Vol. 21, No. 3, pp. 297–317.
57. Jeong, K.-C. and Kim, Y.-D. (1998), Performance analysis of assembly/disassembly sys-
tems with unreliable machines and random processing times, IIE Transactions, Vol. 30,
pp. 41–53.
58. Jeong, K.-C. and Kim, Y.-D. (1999), An approximation method for performance anal-
ysis of assembly/disassembly systems with parallel-machine stations, IIE Transactions,
Vol. 31, pp. 391–394.
59. Jun, K. and Perros, H.G. (1987), An approximate analysis of open tandem queueing
networks with blocking and general service times, Computer Science, NCSU, Raleigh,
NC.
60. Kelton, W.D., Sadowski, R.P., and Sadowski, D.A. (1998), Simulation with Arena,
McGraw-Hill.
61. Kerbache, L. (1984), Analysis of open finite queueing networks, Ph.D. thesis, Depart-
ment of Industrial Engineering and Operations Research, University of Massachusetts,
Amherst, MA.
62. Kerbache, L. and MacGregor Smith, J. (1987), The generalized expansion method for
open finite queueing networks, European Journal of Operational Research, Vol. 32,
pp. 448–461.
63. Khoshnevis, B. (1994), Discrete Systems Simulation, McGraw-Hill.
64. Kleinrock, L. (1975), Queueing Systems, Vol. I, Wiley.
65. Knott, A.D. (1970), The inefficiency of a series of work-stations – a simple formula,
International Journal of Production Research, Vol. 8, pp. 109–119.
66. Kouikoglou, V.S. and Phillis, Y.A. (2001), Hybrid Simulation Models of Production
Networks, Kluwer Academic Publishers.
67. Kuhn, H. (2003), Analysis of automated flow line systems with repair crew interference,
In S.B. Gershwin, Y. Dallery, C.T. Papadopoulos, and J. MacGregor Smith, Editors,
Analysis and Modeling of Manufacturing Systems, pp. 155–179, Kluwer Academic
Publishers.
68. Labetoulle, J. and Pujolle, G. (1980), Isolation method in a network of queues, IEEE
Transact. Software Eng., Vol. SE-6, No. 4.
69. Law, A.M. and Kelton, W.D. (2000), Simulation Modeling and Analysis, McGraw-Hill.
70. Levantesi, R., Matta, A., and Tolio, T. (1999a), Exponential two machine lines with
multiple failure modes and finite buffer capacity, In Proceedings of the Second
Aegean International Conference on Analysis and Modeling of Manufacturing Systems,
pp. 103–112, Tinos Island, Greece.
71. Levantesi, R., Matta, A., and Tolio, T. (1999b), A decomposition method for the per-
formance evaluation of production lines with random processing times, multiple failure
modes and finite buffer capacity, In Proceedings of the Second Aegean International
Conference on Analysis and Modeling of Manufacturing Systems, pp. 113–122, Tinos
Island, Greece.
72. Levantesi, R., Matta, A., and Tolio, T. (2003), Performance evaluation of continuous
production lines with machines having different processing times and multiple failure
modes, Performance Evaluation, Vol. 51, pp. 247–268.
73. Lim, J.-T., Meerkov, S.M., and Top, F. (1990), Homogeneous, asymptotically reliable
serial production lines: Theory and a case study, IEEE Transactions on Automatic
Control, Vol. 35, No. 5, pp. 524–534.
74. Magazine, M.J. and Stecke, K.E. (1996), Throughput for production lines with serial
work-stations and parallel service facilities, Performance Evaluation, Vol. 25, No. 3,
pp. 211–232.
75. Makino, T. (1964), On the mean passage time concerning some queueing problems of
the tandem type, J. Opns Res. Soc. Japan, Vol. 7, pp. 17–47.
76. Muth, E.J. (1984), Stochastic processes and their network representations associated
with a production line queueing model, European Journal of Operational Research, Vol.
15, pp. 63–83.
77. Muth, E.J. (1987), An update on analytical models of serial transfer lines, Research
Report, No. 87-15, Gainesville, FL: Department of Industrial and Systems Engineering,
University of Florida.
78. Nemec, J.E. (1999), Diffusion and Decomposition Approximations of Stochastic Models
of Multiclass Processing Networks, Ph.D. Thesis, Massachusetts Institute of Technology
(M.I.T.).
79. Neuts, M.F. (1981), Matrix-Geometric Solutions in Stochastic Models – An Algorithmic
Approach, The Johns Hopkins University Press.
80. Onvural, R.O. and Perros, H.G. (1986), On equivalencies of blocking mechanisms in
queueing networks with blocking, Operations Research Letters, Vol. 5, No. 6, pp. 293–
298.
81. Papadopoulos, H.T. (1989), Mathematical modelling of reliable production lines, Ph.D.
Thesis, University College Galway, Ireland.
82. Papadopoulos, H.T. (1996), An analytic formula for the mean throughput of K-station
production lines with no intermediate buffers, European Journal of Operational
Research, Vol. 91, pp. 481–494.
83. Papadopoulos, H.T., Heavey, C., and Browne, J. (1993), Queueing Theory in Manufac-
turing Systems Analysis and Design, Chapman & Hall.
84. Papadopoulos, H.T., Heavey, C., and O’Kelly, M.E.J. (1989) Throughput rate of
multistation reliable production lines with inter-station buffers (I) Exponential Case,
Computers in Industry, Vol. 13, No. 3, pp. 229–244.
85. Papadopoulos, H.T., Heavey, C., and O’Kelly, M.E.J. (1990), Throughput rate of multi-
station reliable production lines with inter station buffers: II Erlang case, Computers in
Industry, Vol. 13, No. 4, pp. 317–335.
86. Papadopoulos, H.T. and O’Kelly, M.E.J. (1989) A recursive algorithm for generating
the transition matrices of multistation series production lines, Computers in Industry,
Vol. 12, pp. 227–240.
87. Patchong, A. and Willaeys, D. (2001), Modeling and analysis of an unreliable flow line
composed of parallel-machine stages, IIE Transactions, Vol. 33, pp. 559–568.
88. Perros, H. (1994), Queueing Networks with Blocking - Exact and Approximate Solutions,
Oxford University Press.
89. Perros, H.G. and Altiok, T. (1986), Approximate analysis of open networks of queues
with blocking, tandem configurations, IEEE Trans. Soft. Eng., SE-12, pp. 450–461.
90. Pritsker, A.A.B. (1986), Introduction to Simulation and Slam II, 3rd Edition, Halsted.
References 99
91. Sevast’yanov, B.A. (1962) Influence of stage bin capacity on the average standstill time
of a production line, Theory of Probability and its Applications, Vol. 7, No. 4, pp. 429–
438.
92. Systems Modeling Corporation (1999), Guide to Arena Standard Edition, Sewickley.
93. Takahashi, Y., Miyahara, H., and Hasegawa, T. (1980), An approximation method for
open restricted queueing network, Operations Research, Vol. 28, pp. 594–602.
94. Tan, B. (2001), A three-station merge system with unreliable stations and a shared buffer,
Mathematical and Computer Modeling, Vol. 33, pp. 1011–1026.
95. Tempelmeier, H. and Burger, M. (2001), Performance evaluation of unbalanced flow
lines with general distributed processing times, failures and imperfect production, IIE
Transactions, Vol. 33, No. 4, pp. 293–302.
96. Tolio, T. and Matta, A. (1998), A method for performance evaluation of automated flow
lines, Annals of the CIRP, Vol. 47, No. 1, pp. 373–376.
97. Tolio, T., Matta, A., and Gershwin, S.B. (2002), Analysis of two-machine lines with
multiple failure modes, IIE Transactions, Vol. 34, pp. 51–62.
98. Van Dijk, N. and van der Wal, J. (1989), Simple bounds and monotonicity results
for finite multi-server exponential tandem queues, Queueing Systems Theory and
Applications, Vol. 4, pp. 1–16.
99. Vidalis, M.I. and Papadopoulos, H.T. (2001), A recursive algorithm for generating
the transition matrices of multistation multiserver exponential queueing networks,
Computers & Operations Research, Vol. 28, No. 9, pp. 853–883.
100. Yu, K.-Y.C. and Bricker, D.L. (1993), Analysis of a Markov chain model of a multistage
manufacturing system with inspection, rejection, and rework, IIE Transactions, Vol. 25,
No. 1, pp. 109–112.
101. Zimmern, B. (1956), Etudes de la propagation des arrets aleatoires dans les chaines de
production, Review Statistical Applications, Vol. 4, pp. 85–104.
3
The Design of Production Lines
3.1 Introduction
This chapter is essentially a prelude to the rest of the text and its objective is to assist
the reader to understand the main initial design problems that arise with produc-
tion lines. It is important for the reader to clarify the context of any design problem
related to any production line, e.g., is it a green fields situation, a modification of an
existing production line to enhance performance or the adaptation of an existing line
to produce products not produced already?
Once the strategic decision to use a production line to manufacture the products
has been made, the design of the line must be undertaken. To remind the reader of the
complexities involved, in Figure 3.1, an example of a relatively complex production
line, adapted from Li (2003), is shown.
In Figure 3.1, the rectangles represent machines and the circles represent buffers.
Although it is traditional in analysis to indicate machines by rectangles, it must be
remembered that associated with many such machines are human operators and that
human operators may in fact form a separate work-station without any machines.
Indeed, it is these human operators that add variability to the production line in that
many machine processes are essentially deterministic in practice. Here, it is assumed
that the ergonomic design of the systems is undertaken by relevant specialists while
the physical requirements of the system are being fully specified by others. All these
specialists are of course in a position to contribute to an understanding of the vari-
ability involved in production lines on an ongoing basis during the design process of
the production line.
Quality is a major performance characteristic of modern manufacturing and in
particular there are inspection and test stations embedded in production lines. The
precise arrangement for handling rework of defective material is generally dependent
on the materials handling arrangements. Sometimes feedback is possible resulting in
the reuse of the inspection and testing facilities whereas in other situations rework
and further inspection are effectively performed off the main line. Either of these
cases may be handled in most models.
102
Line 12
Line 2
Line 9
M11 B M1M 1
Ma1 Ba 2 Ma A Ba A+1 12
Bf1 Mf1 B BfF MfF BfF+1

Assembly f2
M1 B2 Mα Bα+1 Bjr Mjr Bjr+1 Mp Bp+1 Bq Mq Bf Mf Br Mr Bj f Mj f BK MK
Line 7 Line 8
Line 1 Line 3 Line 4 Line 5 Line 6
Br 1
3 The Design of Production Lines
Mk1 Bk2 MkMk Line 10 Mr 1

BrR+1
Line 11 Brs
Feedback
Scrap
MrR Mrs+1 Brs+1 Mrs
Main Line: M1,..,Mp,Mq,..,MK, B1,..,Bp,Bq,..,BK

Feeder Line: Ma1,..,MaA, Ba2,..,BaA+1
Feed-Forward Line: Mf1,..,MfF, Bf1,..,BfF+1
Rework Loop: Mr1,..,MrR, Br1,..,BrR+1
Parallel Lines: Mij,i=1,..,k, j=1,..,Mi, Bij, i=1,..,k, j=2,..,Mi
Merge machines: Ma (assembly merge), Mjr (rework merge), Mjf (feed-forward merge)
Split machines: Mr (rework split), Mf (feed-forward split), Mrs (scrap)
Split/merge buffers: Bp (parallel split), Bq (parallel merge)
Fig. 3.1. A typical structure of a complex production line

3.1 Introduction 103
As shown in Figure 3.1, there is one main production line (Line 1, Line 3, Line
4, Line 12, Line 5, Line 6, Line 7 and Line 8), with Line 12 being a parallel-machine
line consisting of κ sub-lines, a feeder line (Line 2), a feed-forward line (Line 9)
as well as a rework line/loop (Line 10 and Line 11). Machine Ma is an assembly
merge machine, machine M jr is a rework merge machine and machine M j f is a feed-
forward merge machine, while machine Mr is a rework split machine, machine M f
is a feed-forward split machine and machine Mrs is a scrap split machine. There are
split/merge buffers associated with the parallel line (Line 12).
By design is meant the specification of some of the parameters (structure of the
production system) to achieve a specific objective. The approach is quite different to
the use of methods to evaluate the performance of a specified system which has been
already discussed in Chapter 2.
In this chapter, it is assumed that the production processes at each machine are
specified. To arrive at this situation may have involved considerable engineering
work. In addition, the sequencing of the machines/layout of the production line has
been determined. For the purposes of this chapter, the details of the transportation
system between the machine stations are assumed to be given and the information
and control systems are not of specific interest. Essentially, what is being said is
that a flow diagram of type Figure 3.1 has been developed in outline form where
the production rate of each individual machine, the details of the buffer sizes and
the number of parallel machines have yet to be determined. Further details of the
considerations involved may be found in Buzacott and Shanthikumar (1993), Altiok
(1997) and Groover (2001), among others.
In general, there are three methods of increasing the throughput of an individual
work-station: (a) increasing the production rate of an individual machine, (b) using
machines in parallel, or (c) a combination of both. These involve technological and
managerial choices. The design of production lines as understood here is confined to
the following issues:
1. Work-load at each station: There are well-known design guidelines, discussed
below in Chapter 4, which result in increased throughput of the line (units pro-
duced per unit time over the entire line). The application of these guidelines will
specify the mean production rates of each of the work-stations. These design
problems are referred to as work-load allocation problems, WAP. In such prob-
lems it is normal to assume fixed specified buffer sizes and single-machine
work-stations.
Readers will be aware that research results of interest to manufacturing systems
designers may arise in work not specifically oriented towards manufacturing sys-
tems. This is particularly true in relation to the work allocation problem where a
series of papers have developed quite strong results mainly using mathematical
analysis. Interested readers are referred to the papers listed in Chapter 4.
2. Determination of the number of machines at each work-station: The use of
parallel systems will affect the throughput of the line. The associated design prob-
lem is referred to as the server allocation problem, SAP, and is treated also in
104 3 The Design of Production Lines
Chapter 4. Normally, in such design problems it is assumed that there are fixed
station specific buffer sizes between the parallel machine stations.
3. Specification of the sizes of the buffers: It is more usual to have machine or station
specific buffers but occasionally common buffers for more than one machine or
station are sometimes used. Such designs are referred to as the buffer allocation
problem, BAP, which is the subject of Chapter 5.
The design problem from the point of view of the systems engineer is as follows
Given:
• Fixed number of work-stations (K). This number is determined by tech-
nological, precedence and economic considerations. Servers at these K
work-stations may consist of machines only, of human operatives only or
of a feasible and necessary combination of these two types of resources.
• Number of servers S(S ≥ K).
• Total work-load of the line, normalized to K (time units).
• Total number of buffer spaces (N).
The design problem in general is to do the following meet a specified objective,
usally expressed in throughput, work-in-process or cost terms:
(i) Allocate the number of servers S over the given K stations; clearly there
must be at least one server at each station;
(ii) Allocate the normalized work-load to each of the given fixed K stations;
(iii) Allocate the total number of buffer spaces N over the K − 1 buffer stor-
age areas. Usually, the buffer in front of the first station is assumed to be
of adequate size (theoretically infinite) to accommodate the flow of work
and these buffer spaces are not included in the N buffer spaces which are
considered as a parameter of the design problem. Likewise, the storage
spaces after the last (Kth) station are excluded from consideration leaving
just K − 1 storage areas among the K stations.
Needless to say, it is possible to consider the design problem of maximizing the
throughput of production lines in which none of the following are specified a priori:
the production rate at each station, the inter-station buffer sizes and the number of
parallel servers at each station. This leads to a very general design problem with con-
siderable computational complexities. In practice however, it is more usual, initially,
to consider simpler design problems with two of the three decisions listed above
already made, and these simpler design problems may be considered to be “pure”
allocation problems.
It should be noted that usually the word ‘allocation’ has a very specific meaning.
In the pure work-load allocation problem, the objective is to allocate a total capacity
of K time units over K work-stations so as to maximize throughput given the machine
specific buffers in the system. In the pure buffer allocation problem, the objective is
to maximize throughput by allocating an overall buffer space of size N among the
K − 1 buffer locations, where each station has a fixed production rate. Finally, in
the pure server allocation problem the total number of servers in the system is fixed
3.1 Introduction 105
and the objective is to maximize throughput of the system by allocating an integer

number of servers to each station given fixed station specific buffers.
The words ‘work-stations’ and ‘machines’ are used interchangeably in produc-
tion line design problems. However, it should be noted that here ‘machines’ is
a generic term which includes the following meanings: physical machines alone,
operators alone or a combination of these two resources or more generally, servers.
Usually, designers are concerned with maximizing throughput. There are a few
other possible objective functions which may be of interest. These include the
minimization of average work-in-process, W IP, having in mind current operations
philosophies of lean production. In such models, a threshold throughput, X0 , must be
achieved and W IP is minimized in the context of this achievement while satisfying
other constraints in relation to buffer allocation, server allocation and work-load allo-
cation. Finally, a more specific cost/financial objective function may be developed to
include machining cost and buffer space and inventory holding costs.
The two performance measures mentioned above, viz., throughput, X , and aver-
age work-in-process, W IP, may be characterized as efficiency and effectiveness
performance measures, respectively. Increasing the throughput of the line is normally
associated with increasing average WIP and vice versa. Usually, other measures of
performance such as mean flow or production time, utilization of individual stations,
often a favorite of earlier generations of production engineers and managers, etc.,
may be easily obtained from the computer results.
In production lines, machines may be considered to be reliable or unreliable.
Unreliable machines have an associated reliability or survival curve from which the
mean time to failure (MTTF) may be determined. Failed machines may be repaired
in accordance with a repair time distribution from which the mean time to repair
(MTTR) may be determined.
The processing time at a machine may be assumed to be deterministic or stochas-
tic. If stochastic, the mean service time and its coefficient of variation may be
determined from the associated processing time distribution. Often the exponential
distribution is used, resulting in a coefficient of variation of 1. In practice, it has been
observed that the coefficient of variation is less than 1 and thus a strict exponential
distribution of processing times is inappropriate. However, phase type distributions
(e.g., Coxian distribution with two phases) can be used to accommodate situations
where the coefficient of variation of service time is less than 1 while retaining the
analytical benefits of the exponential distribution.
The reader might note that the justification of any particular design of a man-
ufacturing system raises complex issues, particularly in the case of systems which
have some inherent flexibility. In the past, finished designs tended to be costed and
evaluated on the basis of either meeting or not meeting a specified interest (hur-
dle) rate in a discounted cash flow analysis (dcf). Many criticisms have been leveled
at this approach (see for example Noble and Tanchoco, 1993). As research in this
area has progressed, the methodology for the concurrent evaluation of design perfor-
mance and economic evaluation has been developed. In production line design, the
full realization of this approach to design evaluation can only be achieved through
the holistic integration of the work of the detailed engineering designers specifying
the outline of the initial system and the work of the system specialists involved in
performance analysis. Chapter 7 is concerned with the costing of various designs.
3.2 Role of the Design Engineer

The design of a production line essentially involves the allocation of the follow-
ing resources after decisions about the number of separate work-stations and the
sequence of such work-stations have been made:
• Servers (operators and/or machines).
• Buffer slots/space between stations.
• Work-load at each station (expected service time of individual parts at each
station).
Clearly, decisions in relation to resources have cost implications which must be taken
into account in addition to the performance measures of the production line. From
a purely production engineering and operations management point of view, there is
considerable attraction to the concept of a symmetrically balanced production line
which would be characterized by the following features:
• Servers with identical mean service rates.
• Same number of servers at each work-station.
• Identical inter-station buffer capacities.
• Identical expected service time of each part at each station (balanced work-load).
The reader might note that in the technical literature, generally, balanced production
lines merely implies that mean service rates (number of servers at each station by
identical individual mean service rate) are equal over all the stations. In fact, there
are a number of algorithms and heuristics in existence which assume that a balanced
work-load leads to the maximum throughput of a production line. The symmetri-
cal balanced line described above has considerable intuitive appeal but very simple
examples will show that if there is variability in service, a balanced line will not lead
to the maximum throughput. For example, consider a balanced production line with
K = 4 stations with exponentially distributed service/processing times at each sta-
tion with equal mean service rates: μi = 1, for i = 1, . . . , 4, and buffer sizes: B j = 2,
for j = 2, 3, 4. The throughput of such a line is 0.7007 compared with the higher
throughput 0.7051 obtained by an unbalanced line with μ1 = 1/1.069, μ2 = 1/0.931,
μ3 = 1/0.931 and μ4 = 1/1.069 (the summation of 1/μi ’s is equal to K = 4) and
buffer sizes again all equal to 2, i.e., the percentage increase in throughput is 0.63%.
Here, consideration is given to the performance measure throughput only. The engi-
neering economist/operations manager should ask the question at what cost was this
increase in throughput achieved? For instance: What about comparison between the
mean work-in-process of the two systems? What was the cost of utilizing the servers
with the specified service rates in each case? Are such servers in fact available? What
utilization of each server was achieved? A conclusion from this simple case is that
improvement in engineering/operations performance measures alone may not assist
3.3 Improvability 107
in answering the real question which is how to achieve the minimum cost of produc-
tion using a production line system of work organization. It is possible to debate that
because many analytical studies have shown that the optimal throughput is not ‘sig-
nificantly’ different from the throughput of a ‘balanced’ production line, it may not
be worthwhile pursuing the optimal solution. However, it must be noted that gen-
erally production lines are developed for high volume and relatively long life and
so a small improvement in throughput may have a significant economic advantage.
Another criticism of analytical models is that they fail to capture the complexities
of real-life systems. This is true. For example, variability in service times is often
described by a phase-type distribution such as the exponential, Erlang or Coxian,
whereas in practice the distribution of service times may follow a very different dis-
tribution. The same criticism can be made of simulation studies. In passing, it may
be noted that in practice the coefficient of variation of service times has been found
to be of the order 0.2 to 0.4. However, if an analytical representation of a proposed
production line indicates that the performance of the system would improve if some
imbalance was introduced, then the designer would be well advised to take this into
account. It is unrealistic to expect at this stage of the development of analytical mod-
eling, including simulation, that the designer can produce designs of production lines
which are incapable of being improved. The viewpoint of this book is to give every
possible assistance to the designer to investigate different design configurations and
to arrive at a design that is feasible, economic and has an acceptable performance.
After the implementation of the design, further improvement is generally possible by
way of special studies, simulations and analytical work following actual experience
in operation.
3.3 Improvability
A different design problem arises when modifications to an existing system are con-
templated. A production line may for example not be achieving desired production
levels due to a deterioration in service levels or to a changed product or product mix.
Clearly, in such cases a total re-design and physical re-construction of the production
line may not be justified. Using evaluative models it may be possible to determine the
throughput with the parameters derived from measurements on the existing system
and hopefully confirming its current performance. Such models might point to the
existence of a bottleneck station through for instance the starving of a downstream
station and/or the blocking of an upstream station and so the design effort could be
concentrated on alleviating the bottleneck station. In other cases it may be possible to
design for optimal throughput and to determine how far the existing system is from
optimal in terms of such measures as work-load allocation, machine specific buffers
and number of parallel machines at each work-station. Clearly, in all such cases there
would be a concern to achieve maximum impact on the performance measure desired
at minimum cost in re-designing the current system.
It should be noted that in practice, the concept of buffers has several meanings.
For example, a buffer between two single-machine stations might be considered to
be in series or in shunt (parallel). The discipline for the series buffer would normally
be First-In, First-Out (FIFO), whereas the discipline for the shunt buffer would be
Last-In, First-Out (LIFO). Where a station with parallel machines is concerned, the
buffer discipline can be quite complex, in that an idle machine may not be in a posi-
tion to service a waiting unit due to the materials handling protocol. So, the usual
assumption of queueing theory that an idle server would immediately serve a wait-
ing job may be violated. Clearly, care should be taken by the analyst to ensure that
the buffer protocols used in any modeling work are in accordance with the actual
situation.
The reader might note that the design problems specified above are not the same
as those problems faced by operations management in their quest for continuous
improvement (KAIZAN). When issues of improving the performance of an exist-
ing system arise, the work of Meerkov and his colleagues is particularly relevant.
Meerkov defines a production system to be improvable if the limited resources
involved in its operation can be redistributed so that a performance measure is
improved. It must be understood that in practice there may be constraints on the
redistribution process and improvability as such may not reach the optimality achiev-
able in a mathematical sense. Performance measures involved here could relate to
throughput, work-in-process (WIP), workforce (WF) allocation and due-time per-
formance. Details of improvement strategies may be found in Jacobs and Meerkov
(1995a). The role of bottlenecks in production systems is well known and in the
paper cited above, Jacobs and Meerkov, gave a very precise definition of a bottleneck
machine or buffer, as follows.
Let PI(μ1 , . . . , μK , N1 , . . . , NK ) be the performance index of interest, e.g., the
throughput, the due-time performance, the workforce allocation, product quality, and
so forth.
A production system is called improvable with respect to WIP if there exists a
sequence N1∗ , . . . , NK∗ such that ∑Ki=1 Ni∗ = N and
PI(μ1 , . . . , μK , N1∗ , . . . , NK∗ ) > PI(μ1 , . . . , μK , N1 , . . . , NK ),
where, ∑Ki=1 Ni = N.
A production system is called improvable with respect to workforce (WF) if there
exists a sequence μ1∗ , . . . , μK∗ such that ∏Ki=1 μi∗ = μ ∗ and
PI(μ1∗ , . . . , μK∗ , N1 , . . . , NK ) > PI(μ1 , . . . , μK , N1 , . . . , NK ),
where, ∏Ki=1 μi = μ ∗ .
The reader will note that the second equation above is in product form and is
a bound on the workforce (WF). The assignment of the workforce defines the pro-
duction rate (machine operators) and the average up-time (repair personnel) of each
machine. The available workforce can be assigned to the work-stations in accordance
with the constraint given by the second equation. This constraint may be referred
to as the machine efficiency constraint and changes in the allocation of resources
within the production line are required to maintain this overall constraint. In con-
trast, the design problem in the earlier paragraphs of this chapter was formulated
3.3 Improvability 109
using the work-load allocation, where the usual summation constraint was used,
i.e., ∑Ki=1 wi = 1.
A production system is called improvable with respect to WIP and WF simulta-
neously if there exist sequences N1∗ , . . . , NK∗ and μ1∗ , . . . , μK∗ such that ∑Ki=1 Ni∗ = N,
∏Ki=1 μi∗ = μ ∗ and
PI(μ1∗ , . . . , μK∗ , N1∗ , . . . , NK∗ ) > PI(μ1 , . . . , μK , N1 , . . . , NK ),
where, ∑Ki=1 Ni = N.
Machine i is the bottleneck machine if
∂ PI(μ1 , . . . , μK , N1 , . . . , NK ) ∂ PI(μ1 , . . . , μK , N1 , . . . , NK )
> , ∀ j = i.
∂ μi ∂ μj
Buffer i is the bottleneck buffer if

PI(μ1 , . . . , μK , N1 , . . . , Ni + 1, . . . , NK ) > PI(μ1 , . . . , μK , N1 , . . . , N j + 1, . . . , NK ), ∀ j = i.
Other definitions of bottlenecks exist, for example, the bottleneck machine is

the machine with the lowest isolated production rate. A buffer is considered to be
a bottleneck buffer if the expected size of the work-in-process (WIP) at that buffer
is larger than the expected size of the work-in-process, W IP, at the other buffers
on the assumption that all buffer capacities are equal. More detailed examination of
issues related to bottlenecks are covered in Goldratt and Cox (1986) in the context
of the theory of constraints. As is clear from the precise definition given by Jacobs
and Meerkov (1995), the bottleneck machine is not necessarily the machine with the
lowest isolated mean production rate nor is the bottleneck buffer that buffer with
the smallest capacity. The reader is referred to an interesting example given in the
work by Jacobs and Meerkov (1995a). Further extensions in the general area on the
topic of improvability are contained in the following papers: Jacobs and Meerkov
(1995b) Kuo, Lim and Meerkov (1996), Chiang, Kuo and Meerkov (1998), Chiang,
Kuo and Meerkov (2000), Li and Meerkov (2000), Li and Meerkov (2001), Chiang,
Kuo, Lim and Meerkov (2000a) and Chiang, Kuo, Lim and Meerkov (2000b). Other
references of interest include Enginarlar, Li and Meerkov (2003a) and Enginarlar,
Li and Meerkov (2003b). Collectively, these papers contain a rich source of infor-
mation to determine bottleneck stations and buffers and insightful design guidelines
which would enhance the performance of existing systems and could also be used
to check the appropriateness of systems design using the methods that are more
germaine to the main stream of the methods proposed in this text.
In Chapters 4, 5, and 6, design problems of particular importance to production
line designers are presented. The objectives of these chapters are to assist designers in
the solution of practical problems using software available at the website associated
with this text. Where possible, design guide rules are given with respect to specific
situations. It must be understood that these guidelines were developed by researchers
following, in most cases, extensive experimentation over a wide range of parameters.
However, although useful, these guidelines must be treated with respect, particularly
if applied to situations not covered by the original experimentation. In this regard, the
reader is advised to consult the original papers which are usually given in the rele-
vant bibliography. Finally, the authors would urge the designer to carry out, using the
software provided, a series of experiments, if at all possible, over the range of param-
eters of interest, so that the appropriateness of the set of the design guidelines may
be tested. It should also be remembered that it is important to develop some expe-
rience of the relative accuracy of some of the algorithms being used by researchers
generally in this area and that perhaps it is true to say that algorithms developed
more recently tend to be more accurate and more efficient. Nevertheless, it is vital
to be fully familiar with the assumptions of any particular model being used because
although most models will give a result, the really important issue is how realistic is
the result obtained when applied to the problem in hand.
References
1. Altiok, T. (1997), Performance Analysis of Manufacturing Systems, Springer-Verlag.
3. Chiang, S.-Y., Kuo, C.-T., Lim, J.-T., and Meerkov, S.M. (2000a), Improvability of assem-
bly systems I: Problem formulation and performance evaluation, Mathematical Problems
in Engineering, Vol. 6, pp. 321–357.
4. Chiang, S.-Y., Kuo, C.-T., Lim, J.-T., and Meerkov, S.M. (2000b), Improvability of
assembly systems II: Improvability indicators and case study, Mathematical Problems
in Engineering, Vol. 6, pp. 359–393.
5. Chiang, S.-Y., Kuo, C.-T., and Meerkov, S.M. (1998), Bottlenecks in Markovian produc-
tion lines: A systems approach, IEEE Transactions on Robotics and Automation, Vol. 14,
No. 2, pp. 352–359.
6. Chiang, S.-Y., Kuo, C.-T., and Meerkov, S.M. (2000), DT-Bottlenecks in serial production
lines: Theory and application, IEEE Transactions on Robotics and Automation, Vol. 16,
No. 5, pp. 567–580.
7. Enginarlar, E., Li, J., and Meerkov, S.M. (2003a), How lean can lean be? The Univer-
sity of Michigan, Systems Science and Engineering Division, Department of Electrical
Engineering and Computer Science, Report No. CGR-03-10, September 2003.
8. Enginarlar, E., Li, J., and Meerkov, S.M. (2003b), Lean buffering in serial production lines
with non-exponential machines, The University of Michigan, Systems Science and Engi-
neering Division, Department of Electrical Engineering and Computer Science, Control
Group Report No. CGR-03-13, November 2003.
9. Goldratt, E., and Cox, J. (1986), The Goal, North Rivers Press.
10. Groover, M.P. (2001), Automation, Production Systems, and Computer Integrated Manu-
facturing, Second Edition, Prentice Hall.
11. Jacobs, D., and Meerkov, S.M. (1995a), Mathematical theory of improvability for
production systems, Mathematical Problems in Engineering, Vol. 1, pp. 95–137.
12. Jacobs, D. and Meerkov, S.M. (1995b), A system-theoretic property of serial produc-
tion lines: improvability, International Journal of Systems Science, Vol. 26, No. 4,
pp. 755–785.
13. Kuo, C.-T., Lim, J.-T., and Meerkov, S.M. (1996), Bottlenecks in serial production lines:
A system-theoretic approach, Mathematical Problems in Engineering, Vol. 2, pp. 233–
276.
References 111
14. Li, J. (2003), Modeling and analysis of complex production systems, Published in the
Proceedings of the Fourth Aegean International Conference on “Analysis of Manufactur-
ing Systems,” pp. 203–212, Samos Island, Greece, July 1–4.
15. Li, J. and Meerkov, S.M. (2000), Bottlenecks with respect to due-time performance in
pull serial production lines, Mathematical Problems in Engineering, Vol. 5, pp. 479–498.
16. Li, J. and Meerkov, S.M. (2001), Customer demand satisfaction in production systems:
A due-time performance approach, IEEE Transactions on Robotics and Automation,
Vol. 17, No. 4, pp. 472–482.
17. Noble, J.S. and Tanchoco, J.M.A. (1993), Design justification of manufacturing
systems – A review, The International Journal of Flexible Manufacturing Systems, Vol. 5,
pp. 5–25.
4
Work-Load and Server Allocation Problems
In this chapter, two separate design problems are considered, viz., the work-load
allocation problem and the server allocation problem in production lines. In a broad
sense both design problems are related to the allocation of work from the point of
view of the operators. Section 4.1 of the chapter describes what is classically known
as the work-load allocation problem, i.e., the allocation of work to each station of
the line so that all the required work is undertaken having in mind any precedence
requirements. A well-known empirically observed phenomenon, namely the bowl
phenomenon, is described. Some computational issues are then discussed. In Sec-
tion 4.2, the server allocation problem is described. In Section 4.3, the simultaneous
optimization of the work allocation and server allocation problems is considered.
Associated with this double optimal problem is the so-called L-phenomenon.
4.1 The Work-Load Allocation Problem

The work-load allocation problem in production lines is analogous to the assembly
line balancing problem in that one is assigning to each work-station a certain amount
of the work, in terms of time, which has to be done. The assembly line balancing
problem is concerned with how much work should be done at each station given the
precedence requirements. In the work-load allocation problem for production lines,
the overall constraint is that the sum of the expected service times is a fixed constant
and the work-load allocation problem essentially is to allocate this total time among
the stations so as to optimize a given objective function, usually, throughput, XK , or
average work-in-process, W IP.
In mathematical terms, the work-load allocation problem, WAP, may be stated as
follows:
max X (w),
subject to:
K
∑ wi = W, (4.1)
i=1
114 4 Work-Load and Server Allocation Problems
where w = (w1 , w2 , . . . , wK ) denotes the vector of wi ’s, i = 1, 2, . . . , K. wi > 0,

i = 1, 2, . . . , K, in turn, denotes the mean service time of the i, i = 1, . . . , K stations
and W is a fixed constant, which may be normalized and set equal to K, the num-
ber of work-stations of the line. This normalized work-load is to be divided among
the K stations. X(w) denotes the throughput of the production line as a function
of the mean service times. The line is configured with buffers of finite capacities
between any two successive stations. Throughput is also a function of other parame-
ters such as the buffer sizes, the number of servers at each work-station as well as the
moments of higher than the first order of the service time distributions at each station.
The latter set of three parameters are not decision variables in this section and are
assumed to be fixed as specified in the production line layout. Also, in place of max-
imizing throughput, other performance measures may be used such as minimizing
average WIP or minimizing the average flow (production) time. It is quite practical
to minimize average WIP given that a specified level of throughput is achieved. The
essence of the work-load allocation problem is the assignment of service rates to the
machine stations to meet the objectives required. In practice, this would involve the
appropriate use of operators and associated machines to achieve the desired service
rates.
The distribution of service times used in unpaced (asynchronous) production
lines is a matter of practical importance. Many models have used the exponential
distribution for which the associated coefficient of variation is one. Other studies
have used the normal (generally truncated) distribution with a range of coefficients
of variation. There is a tendency to use the same service time distribution at each
of the stations but maybe with different parameters. Generally speaking, the per-
ceived view is that a more appropriate practical service time distribution is skewed
to the right (see Dudley, 1968, Knott and Sury, 1987, Murrell, 1962, Buzacott and
Shanthikumar, 1993, among others). Also, it may be noted that there seems to be
a consensus among relevant experts, using experimental data, that the coefficient of
variation of practical service time distributions should be in the range 0.2 to 0.5 (see
Slack, 1982, Knott and Sury, 1987, Pike and Martin, 1994, among others). Reliance
is also placed on Muth’s observation (Muth 1977) that for practical purposes the
throughput of a production line is a function of only the first two moments of the
underlying distributions of service time. In sophisticated models it is possible to
arrange specified values for the mean service time and the coefficient of variation at
each station.
As far back as 1977, Hillier and Boling (1977) initiated a theoretical investigation
into the existence of the bowl phenomenon for production lines with service times
distributed according to exponential and phase-type distributions. They postulated
three conjectures based on three properties, namely, the reversibility property, the
symmetricity property and the monotonicity property, which if they could be shown
to hold for a particular system, a bowl phenomenon would apply which would lead
to an optimal throughput solution. As stated, these conjectures would be a suffi-
cient condition for the existence of an optimal solution having the symmetrical bowl
phenomenon.
4.1 The Work-Load Allocation Problem 115
A production line with intermediate buffers of equal capacities is said to be

reversible if the throughput of the line remains the same if the order of work assigned
to the stations is reversed, i.e.,
X (w1 , w2 , . . . , wK ) = X (wK , wK−1 , . . . , w1 ).
A work-load allocation is said to be symmetrical if w j = wK+1− j for j = 1, . . . , K.

The monotonicity property implies that the work-load w j satisfies the following
condition: w j > w j+1 for j ≤ [(K − 1)/2] and w j < w j+1 for j > [K/2] + 1, where
[x] is the largest integer not exceeding x, x ∈ R.
It may be noted that the work-load allocation problem could be considered in
the context of a fixed configuration of server allocation (but not specifying work
undertaken by any server) and buffer space allocation which classically has resulted
in the discovery of the bowl phenomenon. Alternatively, the work-load allocation
problem could be considered in the context of a fixed buffer space configuration
but with the server configuration not explicitly given. This latter simultaneous work-
load and server allocations has resulted in the discovery of the L-phenomenon. In
a broader sense, the work-load allocation problem is concerned not only with the
allocation of work to each station but also with the allocation of capacity to each
station through the allocation of the servers when S > K.
4.1.1 The bowl phenomenon
One of the most interesting aspects of the design of production lines is the role of the
bowl phenomenon. Originally observed in experimental work by Hillier and Boling
(1966), most researchers would now agree that optimizing throughput in a production
line requires that the work-load be not uniformly allocated (equally balanced) among
the stations of the line. A typical illustration of the bowl phenomenon is shown in
Figure 4.1. This particular bowl arose in maximizing the throughput in a five-station
production line with equal inter-station buffers each of size 3 slots.
It may be noted that because of the size of the system, an exact solution is pos-
sible. As may be seen from Figure 4.1, the curve connecting the service times is
concave and is representative of the cross section of a bowl from which the title
comes. To date no satisfactory theoretical proof of the required existence of the bowl
phenomenon has been presented. Readers should be aware that whereas Figure 4.1
gives a relatively dramatic illustration of the bowl phenomenon, in many cases in
practice the deviation from a uniform balance may be small or non-existent. As
the reader will realize, the concept of reducing the work-load on a station in effect
implies a smaller mean time to service and requires a more powerful station and so
the station in effect could be described in “colloquial terms” as being a quicker or a
more powerful station where the best operators may be assigned. Despite the absence
of a rigorous mathematical proof of the existence under given conditions (buffer
size allocation, service time distributions, number of stations and number of parallel
servers at each station) of the bowl phenomenon, most experimental work confirms
its existence. Also, some theoretical work in serial production lines involving the
1.080
1.040
1.000
0.960
0.920
0.880
w1 w2 w3 w4 w5
Fig. 4.1. The work-load allocation over five stations with inter-station buffer capacities of sizes
B2 = B3 = B4 = B5 = 3 slots
appropriate ordering of machines of different service rates capabilities to achieve

optimal performance does not lead to the conclusion that it does not exist even though
this work is concerned with a different formulation of the problem as the service
rates are specified and the order is unimportant from the point of view of produc-
tion on the line. Relevant references include Tembe and Wolff (1974), Whitt (1985),
Greenberg and Wolff (1988), Huang and Weiss (1990), Suresh and Whitt (1990),
Shanthikumar, Yamazaki and Sakasegawa (1991), Ding and Greenberg (1991), Liao
and Rosenshine (1992), Yamazaki, Sakasegawa and Shanthikumar (1992), Cheng
and Zhu (1993) and Wan and Wolff (1993), among others.
The classical concept of the bowl phenomenon is that the work-load should be
allocated among the work-stations according to a strictly concave function. How-
ever, in the literature, approximations to the bowl phenomenon have been made
using piecewise linear approximations. Two such approximations may be noted:
one-level and two- or more-level allocations (the reader is referred to Buzacott and
Shanthikumar, 1993, formulas (5.94) and (5.95) for equal buffer sizes, on page 202).
The one-level allocation corresponds to the perfectly balanced line, whereas the two-
level allocation consists of stations with two different levels of mean service times:
the outer (first and last) stations equally having a higher mean service time and the
intermediate stations each having equally lower mean service times (see Figure 4.2
for a two-level approximation to a bowl phenomenon).
However, it is important for practitioners to note that in many studies where the
bowl phenomenon was found to exist, the coefficient of variation of the service times
and the impact of different levels of buffers between the stations were not explicitly
considered. The controversy about the practical value of the bowl phenomenon to
production line designers and production managers still exists. There are at least
three aspects to this controversy: (i) Does the bowl phenomenon exist? On balance,
there are sufficient carefully executed studies to support the contention that it does
exist and is important particularly in situations where there are limited buffer sizes
1.080
1.068 1.068
1.040
1.000
0.957 0.950 0.957
0.960
0.920
0.880
w1 w2 w3 w4 w5
Fig. 4.2. Two-level approximation to a bowl phenomenon
and the coefficients of variation of service times are relatively close to one another.
(ii) Is the bowl phenomenon a mirage in systems with a large number of states where
exact analysis is not possible? Basically, this question cannot be really answered
because generally the throughput in such systems is approximated by an algorithm
and it is impossible to know the limits of the approximation to the throughput. (iii) Is
it of value in practice? Some analysts might be of the view that 1% or 2% change in
optimal throughput of a production system is of little relevance particularly having in
mind how difficult it would be in practice to ensure that the service rates designated
by the bowl phenomenon would be achieved. However, it should be remembered that
production lines are designed for high volume and even a small change in throughput
may well be worthwhile in commercial terms. In any case, it is probably as difficult
practically to uniformly balance a line as to configure the line in accordance with
the bowl phenomenon. So, why attempt a second best solution? The real issue is, of
course, whether those mathematical models in which a bowl phenomenon is shown to
be associated with maximum throughput are accurate models of the realities of actual
production lines. This brings into question the validity of specifying processing times
in stochastic terms generally using phase-type distributions. It might be the case that
in production lines where there is a significant human operator involvement, as well
as machine involvement, the bowl phenomenon is more relevant. The authors would
encourage readers to make up their own minds about this controversy.
4.1.2 Computational issues
Here, the objective is to acquaint the reader with a few numerical techniques which
have been found useful in relation to production line analysis regarding the work-
load allocation problem. It is by no means a comprehensive survey of numerical
analysis approaches.
Because of the general belief in the existence of the bowl phenomenon, there
is often a need to obtain the maximum throughput through a process of numerical
iteration. The usual efficient search procedures such as gradient search procedures
are used. Given below is an algorithm for the steepest ascent method of parallel
tangents (PARTAN method) which has been found useful. The approach is described
in Buehler, Shah and Kempthorne (1964).
The steepest ascent method of parallel tangents (PARTAN method) for solving
the work-load allocation problem
Step 1
Develop an appropriate initial feasible work-load allocation, (w01 , . . . , w0K ) (uni-
form if there is no other information available). Then determine the throughput
X 0 (w1 , . . . , wK ) for this work-load after calculating the steady-state probabilities.
Step 2
Choose a small quantity h (based on experience) and determine the partial derivatives
of X 0 (w01 , . . . , w0K ), numerically, as follows:
∂ X 0 X(w01 , . . . , w0i + h, . . ., w0K ) − X 0(w01 , . . . , w0K )
= , i = 1, . . . , K.
∂ w0i h
It should be noted that throughput, X , must be evaluated K times at this step (from
the associated probabilities) and the value X 0 is also used.
Step 3
Evaluate X 1 (w11 , . . . , w1K ), where,

∂X0
w1i = w0i + , i = 1, . . . , K,
∂ w0i
for values of such that the work-load constraint: ∑Ki=1 w1i = K is satisfied.
Step 4
Repeat Step 3 to obtain in a similar fashion X 2 , which is the optimal value of X along
the line of the steepest ascent from Step 3.
Step 5
Knowing the starting point at Step 2 and the optimal point reached at Step 4, proceed
along the line joining these two points again in steps until an optimal value of X is
obtained. To clarify, if the initial point is given as (w01 , . . . , w0K ) and the point reached
in Step 4 is (w21 , . . . , w2K ), evaluate X at the following point
(w21 + (w21 − w01 ), . . . , w2K + (w2K − w0K ))
for values of > 0 until an optimal for X 3 is reached.
Step 6
Return to Step 2 and continue on to Steps 3, 4 and 5 until a satisfactory convergence

to an optimal X is achieved.
General remark
Any special features of the network, e.g., symmetry, may be explored to reduce
computational effort.
Another numerical approach which is useful for solving the work-load allocation
problem is described in Baruh and Altiok (1991). The authors used the first- and
second-order numerical perturbations to determine the optimal work-load allocation
in production lines.
Next, an approximate method, the two-level work-load allocation algorithm, pro-
posed by Buzacott and Shanthikumar (1993), for obtaining near optimal throughputs
and work-load allocations in production lines with single-machine work-stations
is given. The algorithm is available at the website associated with this book with
abbreviated name TLWLA. This is a stand-alone optimization algorithm.
If the inter-station buffers, B2 , . . . , BK , have the same size B (B2 = B3 = · · · =
BK = B), where (K − 1)B = N, the total number of buffer slots in the system, and
the total work-load has been normalized to equal the total number of stations, K,
Buzacott and Shanthikumar’s near optimal approximations are as follows:
K(B + 1) + 2
X ∗ (B, K) = . (4.2)
K(B + 3)
The associated work-load allocation is a two-level bowl approximation, given by:
K(B + 2)
w∗1 = w∗K = (4.3)
K(B + 1) + 2
K(B + 1)
w∗i = , i = 2, . . . , K − 1. (4.4)
K(B + 1) + 2
If the inter-station buffers, B j , j = 2, 3, . . . , K, are unequal with respective capac-

ities B j , j = 2, 3, . . . , K, and again the total work-load has been normalized to K,
Buzacott and Shanthikumar’s near optimal approximations are:
K
2 1
X ∗ (B2 , . . . , BK , K) = 1 −
K ∑ Bi + 3 . (4.5)
i=2
The associated work-load allocation is a multi-level and not necessarily symmet-

rical bowl approximation, given by:
K αi
w∗i = , i = 1, . . . , K, (4.6)
∑Ki=1 αi
where
B2 + 2
α1 = , (4.7)
B2 + 3
1 1
αj = 1 − − , j = 2, . . . , K − 1, (4.8)
B j + 3 B j+1 + 3
BK + 2
αK = . (4.9)
BK + 3
The following summary may be of value to the reader who wishes to use the
software available at the website associated with this book in solving work-load
Work-Load Allocation Problem (WAP)
1. TLWLA
• Exponential service time distributions.
• Any distribution of buffers.
The TLWLA procedure will give an approximate work-load allocation and
approximate optimal throughput of the production line.
2. MARKOV and SA/GA
• For short reliable or unreliable production lines with Erlang-k (k ≥ 1) service
and repair times and exponential times to failure.
3. DECO-1 and simulated annealing/genetic algorithms SA/GA
• For large reliable exponential production lines with single machine stations.
• Finite intermediate buffers.
4. DECO-2 and simulated annealing/genetic algorithms SA/GA1
• For large reliable exponential production lines with multiple parallel identi-
cal machine stations.
• Finite intermediate buffers.
4.2 The Server Allocation Problem
One of the allocation issues for the systems designer is to allocate the number of
servers, S (when S > K), over the given fixed K stations. Conceptually what is
involved is the assignment of service capacity to each of the K stations to meet the
objectives of maximizing throughput or minimizing average WIP.
If s = (S1 , S2 , . . . , SK ) denotes the vector of servers allocated to the i stations,
i = 1, 2, . . . , K, in mathematical terms the server allocation problem, SAP, is as
follows:
max X (s)
1 Details of simulated annealing (SA) and genetic algorithms (GA) as optimization proce-
dures are given in Chapter 5, Section 5.4.
4.2 The Server Allocation Problem 121
subject to:
K
∑ Si = S
i=1
for fixed allocation of work to each station and fixed buffer alocation.
Interesting papers in this area include Hillier and So (1989), Futamura (2000),
Hillier and So (1995) and Magazine and Stecke (1996), with the latter two papers
dealing with various combinations of the work-load, server and buffer allocations.
Hillier and So (1989) considered production lines with exponential, Erlang with
two phases of service and Coxian with two phases of service at each station, no
intermediate buffers or with just one buffer slot among the stations and equal work-
load allocation over all the stations. Define n = [S/K] as the greatest integer ≤S/K
and E = S − nK, where S > K is the total number of servers available for allocation
over the stations. The main results of this study may be represented as rules for
maximizing the throughput and are as follows:
• Rule 1: If S/K is an integer, allocate the servers uniformly among the K stations.
• If S/K is not an integer, initially allocate n = [S/K] to each of the K stations mak-
ing a total initial allocation of Kn servers, and the balance of the servers, E, are
allocated according to the following rules:
– Rule 2: If E = 1 and K is odd, then allocate the extra server to the center
station.
– Rule 3: If E = 1 and K is even, then allocate the extra server to one of the
two central stations. If a lower W IP is of interest, choose the left central
station, i.e., the station nearest the beginning of the line.
– Rule 4: If E = K − 1 > 1, then allocate an extra server to each of the stations
except station 1, but for extremely large n in which case more than one server
may be assigned to a single station.
– Rule 5: If E = K − 2 > 1, then allocate an extra server to every station except
the first and last stations.
– Rule 6: If 1 < E < K − 2, then allocate the extra servers “almost uniformly”
over the interior stations.
Although the work-load allocation is uniform, these design rules generally sup-
port the concept of allocating extra servers to the interior stations and therefore in
accordance with the idea of an inverse bowl of service capacity.
Futamura (2000) considered the case where the coefficient of variation of the
servers was not identical. The general rule of thumb is to allocate more servers to the
stations with a higher coefficient of variation although, as the author indicated, more
research is needed to derive precise design rules.
It is possible to solve any server allocation problem using the algorithms available
at the website associated with this book for the following type of serial production
line:
• Parallel exponential reliable machines at each station,
• Number of stations: No practical limit (over 1000 stations),
• Number of buffer slots: 5000.
Specifically, one uses the evaluative decomposition algorithm for solving serial
production lines with multiple parallel-machine stations in conjunction with an opti-
mization algorithm such as simulated annealing and genetic algorithm to find an
optimal or near optimal solution to the server allocation problem.
The following detailed summary may be of value to the reader who wishes to
use the software available at the website associated with this book in solving server
1. Server Allocation Problem (SAP)
• Apply the rules given above for the allocation of servers and use DECO-2
to determine the throughput with the specified allocated work-load.
• Exponential service time distributions and reliable machines.
2. DECO-2 and simulated annealing/genetic algorithms SA/GA∗
• For large reliable exponential production lines with multiple parallel iden-
tical machine stations and finite intermediate buffers with the specified
work-load allocation.
∗ Details of simulated annealing (SA) and genetic algorithms (GA) as optimization
procedures are given in Chapter 5, Section 5.4.
4.3 The Simultaneous Work-Load and Server Allocation:

The L-phenomenon
Up to now, design of production lines was considered in the context of one dimen-
sion only, i.e., work-load allocation or server allocation. Now consideration is given
to the simultaneous allocation of work-load and servers. Investigations on the simul-
taneous allocations of both work-load and servers have led to the discovery by Hillier
and So (1995, 1996) of the so called L-phenomenon, when the objective is to maxi-
mize throughput. The design rule here is to allocate just one server to each one of the
stations and all the remaining servers (S − K + 1) to one of the two end stations of the
line. If one considers the smallest WIP, then extra servers must be allocated to the
first station, although this will slightly reduce the throughput from the optimal value
which occurs when the extra servers are allocated to the last station. This allocation
resembles the shape of the capital letter L and thus the name of the phenomenon.
The associated work-load allocation at optimal illustrates the well-known bowl phe-
nomenon but of course it is not symmetrical with a significantly higher work-load
allocation to the station to which the extra servers have been allocated. Hillier and
So (1996) also derived the marginal contributions in terms of throughput of addi-
tional servers and showed that in fact it increases with the addition of extra servers.
This was designated by Hillier and So as “the multiple-server phenomenon” (type 1).
Hillier and So (1995) also observed that a “second multiple-server phenomenon” is
in operation in the absence of the L-phenomenon when the number of servers is an
integer multiple of the number of stations, i.e., S = sK, where s denotes the number
of servers allocated per station. In that case, as s is increased, the marginal increase
4.3 The Simultaneous Work-Load and Server Allocation: The L-phenomenon 123
in throughput using the corresponding optimal work allocation increases monotoni-

cally in s. It should be noted that Hillier and So used a model consisting of K stations
and that the service times are independent and identically distributed (i.i.d.) random
variables following an exponential distribution with fixed means that are normalized
in the analysis.
The primary concept behind the bowl phenomenon is that the interior stations
are given preferential treatment, i.e., they are given less work to do in an expected
sense. So, the situation is that if you have stations with equal service rate capacities,
the designer should allocate work so that the processing time of a part in the interior
stations is lower than the processing time of the part at the outer stations. Now, how-
ever, when work-load and servers are allocated simultaneously to achieve optimal
throughput, it is the end stations that are given preferential treatment with respect
to server allocation, although a non-symmetrical bowl phenomenon of work-load
allocation still exists, at the optimal throughput. More precisely, the corresponding
optimal work-load allocation assigns by far the largest amount of work per server to
the station with the largest number of servers and the work allocations per server (the
work assigned to the station) are monotonically decreasing from the station with the
largest number of servers (either the first or the last station) to the next to the end
station. The end station (either the first or the last station, to which only one server
is assigned) has an increased work-load per server allocation over that allocated to
the next to end station. These results show that unbalancing the servers and work-
loads can provide substantial improvements in throughput of unpaced production
lines with service rate variability over the balanced as possible allocations which, of
course, would be optimal in paced lines or deterministic lines or lines with variable
service rates with infinite inter-station buffer capacities.
The L-phenomenon which, as noted above, occurs under the simultaneous alloca-
tion of servers and work-load may be contrasted with the results obtained when only
server allocation is being considered given a fixed work-load allocation. It may be
pointed out that in our cases no evidence that the bowl phenomenon does not exist is
forthcoming, although it might have been conjectured that even in the simultaneous
allocation of servers and work-load, the allocation of the servers would give prefer-
ential treatment to the inner stations of the line contributing in a simplistic manner
to the bowl phenomenon. Clearly, much fundamental research effort is required for
a full understanding of the bowl phenomenon and its ramifications.
Hillier and So (1995) gave a heuristic allocation scheme for the case where there
are upper and lower bounds on the number of servers to be allocated at each station.
The steps of this scheme are as follows:
Step 1: Allocate the minimum number of servers required at each station.
Step 2: Allocate as many as possible extra servers at the last station.
Step 3: If there are extra servers left, allocate as many as possible to the first station.
Step 4: If there are still servers remaining, allocate as many as possible to the next
to the last station, then as many as possible to the second station, etc.
Step 5: The procedure is concluded when there are no remaining servers.
Hillier and So also indicated that in their experience, if the number of stations is itself
a decision variable, throughput would be maximized if K = [S/Smax], where K is the
optimal number of stations, [x] indicates the maximum integer less than or equal to
x, S is the number of available servers and Smax is the upper bound on the servers that
can be assigned to any station. From Hiller and So’s empirical studies, the optimal
server allocation to achieve maximum throughput would be in accordance with the
scheme of the L-phenomenon given above.

4.4.1 Bowl phenomenon
Hillier and Boling (1966) first observed and conjectured the existence of the bowl
phenomenon by examining two-, three- and four-station lines with exponentially
distributed processing times.
Rao (1975a) analyzed a two-station production line and showed that for moderate
coefficients of variation of the processing times, the mean service rates for Erlang and
normal density functions of the service times differ only marginally. Rao showed that
throughput improves by allocating a slightly higher work-load to the less variable
station.
Rao (1975b) concluded that at high values of coefficient of variation, the type of
service time distribution has a considerable effect on the efficiency of a two-station
series system when the variability of service times at the stations is not the same.
Rao (1976) analyzed a three-station production line and showed that the variabil-
ity imbalance plays a decisive role and outweighs the bowl phenomenon.
De La Wyche and Wild (1977) investigated via simulation the imbalance in ser-
vice time variability, the imbalance in buffer storage and the interaction of service
time and buffer imbalance.
Hillier and Boling (1977) considered short lines with Erlang service times and
proposed three conjectures implying the bowl phenomenon.
Magazine and Silver (1978) studied the effect on throughput from different
choices of design parameters. They proposed some heuristics to find approximate
values for the optimal work-load allocation and for the throughput.
El-Rayah (1979b) examined the effect of inequality of interstage buffer capacities
and operation time variability on the throughput of production lines.
Hillier and Boling (1979) studied the change of the optimal allocation of work
between stations with respect to (i) the number of work-stations in the line, (ii) the
limit on the amount of work-in-process (WIP) and (iii) the variance of station service
or processing times. They re-confirmed the existence of the bowl phenomenon.
El-Rayah (1979a) conducted computer simulations and confirmed the bowl
phenomenon too.
Mishra et al. (1985) showed that in systems consisting of a hyperexponential
station, the guideline of allocating more work-load to the stations with less variability
is violated due to the fact that hyperexponential is a composite distribution.
Lau and Martin (1986) developed a decision support system for the design of
production lines, incorporating a bowl phenomenon.
Muth and Alkaff (1987) presented a method for analyzing distribution-free three-
station production lines and offered a bibliography on the work-load allocation and
the bowl phenomenon.
Thompson and Burford (1988) showed that the bowl phenomenon is associated
with an imbalance in absolute variability and that the bowl effect vanishes in cases
where a minimal level of in-process buffer stock is provided.
So (1989) conducted simulation experiments in production lines with normally
distributed processing times and showed that throughput can be improved by appro-
priately unbalancing work allocations.
Pike and Martin (1994) studied the bowl phenomenon in production lines under
realistic operating conditions and found out that bowl-shaped configurations perform
better than perfectly balanced lines for systems of at least 30 stations in length and
with inter-station buffer capacities of up to one unit. They also showed that the opti-
mal two-level allocation of mean service times performs no worse than the optimal
multi-level allocation. In addition, they discovered that the amount of imbalance in
a line can generally be double the imbalance in an optimal bowl and still perform
at least as well as the balanced line. Finally, the authors showed that optimal bowl
configurations are not particularly sensitive to coefficient of variation or distribution
shape within a realistic range.
Lau (1994) simulated production lines with different buffer sizes and differ-
ent combinations of station service times’ means and variances and concluded that
throughput is maximized when the means and variances are both balanced. Lau also
found out that variance imbalance has a very small effect on throughput, among other
findings.
Hillier and So (1995) considered combinations of the three design problems
in production lines: the work-load allocation, the buffer allocation and the server
Spinellis, Papadopoulos and MacGregor Smith (2000) also examined combi-
nations of the above three design problems using a robust generalized queueing
network algorithm as an evaluative procedure and simulated annealing for optimizing
production line configurations.
Shanthikumar and Yao (1988) dealt with the server allocation problem in mul-
tiple center manufacturing systems. They formulated the problem as a nonlinear
integer program of allocating servers in a closed queueing network to maximize the
throughput of the system.
Dallery and Stecke (1990) addressed the problem of the optimal allocation of
servers and work-loads in closed queueing networks. They used decomposition to
obtain results for the subnetworks in isolation and then to solve the optimal configu-
ration problem. The authors also recommended applications of their results to design
and planning of flexible manufacturing systems.
4.4.2 Reversibility
Dattatreya (1978) Defined C- and D-reversibility, as follows. A tandem queueing

system is said to be C-reversible if the original system has the same throughput as its
reversed system. A blocking system, on the other hand, is defined as D-reversible if
the distributions of times of the departure epochs from both systems are all identical.
Makino (1964) proved C-reversibility for simple systems.
Yamazaki and Sakasegawa (1975) proved D-reversibility, whereas Dattatreya
(1978) and Muth (1979) proved independently the reversibility property.
Yamazaki, Kawashima and Sakasegawa (1985) proved C-reversibility in two-
station blocking systems with parallel machines at each station and stochastic service
times. The same authors proved that this property cannot be extended to larger similar
systems (with parallel machines at each station, stochastic service times and finite
intermediate buffers).
Melamed (1986) provided some results on the reversibility and duality of some
tandem blocking queueing systems.
Dallery, Liu and Towsley (1991) considered reversibility in fork/join queueing
networks with blocking after servcie (manufacturing blocking).
References
2. Baruh, H. and Altiok, T. (1991), Analytical perturbations in Markov chains, European
Journal of Operational Research, Vol. 51, pp. 210–222.
3. Buehler, R.J., Shah, B.V. and Kempthorne, O. (1964), Methods of parallel tangents,
Chemical Engineering Progress Symp. Series, Vol. 60, No. 50, pp. 1–7.
5. Cheng, D.W. and Zhu, Y. (1993), Optimal order of servers in a tandem queue with general
blocking, Queueing Systems, Vol. 14, pp. 427–437.
6. Dallery, Y., Liu, Z. and Towsley, D. (1991), Reversibility in fork/join queueing networks
with blocking after service, Laboratoire MASI, Universite P. et M. Curie, Paris, France.
7. Dallery, Y. and Stecke, K.E. (1990), On the optimal allocation of servers and work-loads
in closed queueing networks, Operations Research, Vol. 38, No. 4, pp. 694–703.
8. Dattatreya, E.S. (1978), Tandem Queueing Systems with Blocking, Ph.D. Dissertation,
Department of Industrial Engineering and Operations Research, University of California,
Berkeley.
9. De La Wyche, P. and Wild, R. (1977), The design of imbalanced series queue flow lines,
Operational Research Quarterly, Vol. 28, No. 3, ii, pp. 695–702.
10. Ding, J. and Greenberg, B. (1991), Bowl shapes are better with buffers–sometimes,
Probability in the Engineering and Informational Sciences, Vol. 5, pp. 159–169.
11. Dudley, N.A. (1968), Work Measurement, Some Research Studies, Macmillan.
12. El-Rayah, T.E. (1979a), The efficiency of balanced and unbalanced production lines,
References 127
13. El-Rayah, T.E. (1979b), The effect of inequality of interstage buffer capacities and oper-
ation time variability on the efficiency of production line systems, International Journal
of Production Research, Vol. 17, No. 1, pp. 77–89.
14. Futamura, K. (2000), The multiple server effect: Optimal allocation of servers to sta-
tions with different service-time distributions in tandem queueing networks, Annals of
Operations Research, Vol. 93, pp. 71–90.
15. Greenberg, B. and Wolff, R.W. (1988), Optimal order of servers for tandem queues in
light traffic, Management Science, Vol. 34, No. 4, pp. 500–508.
16. Helber, S. (1999), Performance Analysis of Flow Lines with Non-Linear Flow of Material,
Springer-Verlag.
17. Hillier, F.S. and Boling, R.W. (1966), The effect of some design factors on the efficiency
of production lines with variable operation times, The Journal of Industrial Engineering,
Vol. 17, pp. 651–658.
18. Hillier, F.S. and Boling, R.W. (1977), Toward characterizing the optimal allocation of
work in production line systems with variable operation times, Advances in Operations
Research (Marc Roubens, Ed.), North-Holland, Amsterdam, pp. 649–658.
19. Hillier, F.S. and Boling, R.W. (1979), On the optimal allocation of work in symmetrically
unbalanced production line systems with variable operation times, Management Science,
Vol. 25, No. 8, pp. 721–728.
20. Hillier, F.S. and So, K.C. (1989), The assignment of extra servers to stations in tan-
dem queueing systems with small or no buffers, Performance Evaluation, Vol. 10,
pp. 219–231.
21. Hillier, F.S. and So, K.C. (1995), On the optimal design of tandem queueing systems with
finite buffers, Queueing Systems, Vol. 21, pp. 245–266.
23. Huang, C.C. and Weiss, G. (1990), On the optimal order of M machines in tandem,
Operations Research Letters, Vol. 9, pp. 299–303.
24. Iyama, T. and Ito, S. (1987), The maximum production rate for an unbalanced multi-server
flow line system with finite buffer storage, International Journal of Production Research,
Vol. 25, No. 8, pp. 1157–1170.
25. Knott, K. and Sury, R.J. (1987), A study of work-time distributions in unpaced tasks, IIE
Transactions, Vol. 19, pp. 50–55.
26. Lau, H.-S. (1994), Allocating work to a two-stage production system with interstage
buffer, International Journal of Production Economics, Vol. 36, pp. 281–289.
27. Lau, H.-S. and Martin, G.E. (1986), A decision support system for the design of
unpaced production lines, International Journal of Production Research, Vol. 24, No. 3,
pp. 599–610.
28. Liao, C.-J. and Rosenshine, M. (1992), A heuristic for determining the optimal order in a
tandem queue, Computers and Operations Research, Vol. 19, No. 2, pp. 133–137.
29. Magazine, M.J. and Silver, G.L. (1978), Heuristics for determining output and work allo-
cations in serial flow lines, International Journal of Production Research, Vol. 16, No. 3,
pp. 169–181.
30. Magazine, M.J. and Stecke, K.E. (1996), Throughput for production lines with serial work
stations and parallel service facilities, Performance Evaluation, Vol. 25, pp. 211–232.
31. Makino, T. (1964), On the mean passage time concerning some queueing problems of the
tandem type, Journal of the Operations Research Society of Japan, Vol. 7, pp. 17–47.
32. Melamed, B. (1986), A note on the reversibility and duality of some tandem blocking
queueing systems, Management Science, Vol. 32, No. 12, pp. 1648–1650.
33. Mishra, A., Acharya, D., Rao, N.P., and Sastry, G.P. (1985), Composite stage effects in
unbalancing of series production systems, International Journal of Production Research,
Vol. 23, No. 1, pp. 1–20.
34. Murrell, K.F.H. (1962), Operator variability and its industrial consequence, International
Journal of Production Research, Vol. 1, pp. 39–55.
35. Muth, E.J. (1977), Numerical methods applicable to a production line with stochastic
servers, TIMS Studies in the Management Sciences, 7:143–159.
36. Muth, E.J. (1979), The reversibility property of production lines, Management Science,
Vol. 25, No. 2, pp. 152–158.
37. Muth, E.J. and Alkaff, A. (1987), The bowl phenomenon revisited, International Journal
38. Noble, J.S. and Tanchoco, J.M.A. (1993), Design justification of manufacturing systems –
A review, The International Journal of Flexible Manufacturing Systems, Vol. 5, pp. 5–25.
39. Papadopoulos, H.T., Heavey, C., and Browne, J. (1993), Queueing Theory in Manufac-
turing Systems Analysis and Design, Chapman & Hall.
40. Pike, R. and Martin, G.E. (1994), The bowl phenomenon in unpaced lines, International
Journal of Production Research, Vol. 32, No. 3, pp. 483–499.
41. Rao, N.P. (1975a), Two-stage production systems with intermediate storage, AIIE Trans-
actions, Vol. 7, No. 4, pp. 414–421.
42. Rao, N.P. (1975b), On the mean production rate of a two-stage production system of the
tandem type, International Journal of Production Research, Vol. 13, No. 2, pp. 207–217.
43. Rao, N.P. (1976), A generalization of the ‘bowl phenomenon’ in series production
systems, International Journal of Production Research, Vol. 14, No. 4, pp. 437–443.
44. Shanthikumar, J.G., Yamazaki, G., and Sakasegawa, H. (1991), Characterization of opti-
mal order of servers in a tandem queue with blocking, Operations Research Letters,
Vol. 10, pp. 17–22.
45. Shanthikumar, J.G. and Yao, D.D. (1988), On server allocation in multiple center
manufacturing systems, Operations Research, Vol. 36, No. 2, pp. 333–342.
46. Slack, N. (1982), Work time distributions in production system modelling, Research
paper, Oxford Centre for Management Studies.
47. Smunt, T.L. and Perkins, W.C. (1990), The efficiency of unbalancing production lines: an
alternative interpretation, International Journal of Production Research, Vol. 28, No. 6,
pp. 1219–1220.
48. So, K.C. (1989), On the efficiency of unbalancing production lines, International Journal
49. So, K.C. (1990), On the efficiency of unbalancing production lines: a reply to an alter-
native interpretation, International Journal of Production Research, Vol. 28, No. 6,
pp. 1221–1222.
50. Spinellis, D., Papadopoulos, C., and MacGregor Smith, J. (2000), Large production line
optimization using simulated annealing, International Journal of Production Research,
Vol. 38, No. 3, pp. 509–541.
51. Suresh, S. and Whitt, W. (1990), Arranging queues in series: A simulation experiment,
Management Science, Vol. 36, No. 9, pp. 1080–1091.
52. Tembe, S.V. and Wolff, R.W. (1974), The optimal order of service in tandem queues,
53. Thompson, W.W. and Burford, R.L. (1988), Some observations on the bowl phenomenon,
54. Yamazaki, G., Kawashima, T., and Sakasegawa, H. (1985), Reversibility of tandem
blocking systems, Management Science, Vol. 31, No. 1, pp. 78–83.
References 129
55. Yamazaki, G. and Sakasegawa, H. (1975), Properties of duality in tandem queueing

systems, Annals of the Institute of Statistical Mathematics, Vol. 27, pp. 201–212.
56. Yamazaki, G., Sakasegawa, H., and Shanthikumar, J.G. (1992), On optimal arrangement
of stations in a tandem queueing system with blocking, Management Science, Vol. 38,
No. 1, pp. 137–153.
57. Wan, Y.-W. and Wolff, R.W. (1993), Bounds for different arrangents of tandem queues
with nonoverlapping service times, Management Science, Vol. 39, pp. 1173–1178.
58. Whitt, W. (1985), The best order for queues in series, Management Science, Vol. 31,
No. 4, pp. 475–487.
5
The Buffer Allocation Problem
The buffer allocation problem, BAP, is concerned with the allocation of a certain
fixed number of buffer slots, N, among the K − 1 intermediate buffer locations of a
production line in order to meet some specified objective. The number of stations of
the line is fixed at K, the number of servers assigned to each station is fixed and the
work allocation w = (w1 , w2 , . . . , wK ) is also fixed.
The buffer allocation problem is of particular interest to operations management
in that in many practical production line situations, the allocation of buffer space
may be the primary flexibility available to the organization. Clearly, buffer space is
an expensive resource and so, ideally models involving cost considerations are very
desirable. Of course, there are also plant layout issues involved.
At least three buffer allocation problems have been identified in the literature and
these are described in Section 5.1. Solutions of the buffer allocation problems are
discussed in Section 5.2. Special solution approaches to buffer allocation problems
in short lines are the subject of Section 5.3, whereas solution approaches to buffer
allocation problems in longer lines are treated in Section 5.4.
5.1 Formulation of the Buffer Allocation Problems

The formulation of the buffer allocation problems depends on the objective function
chosen. These objective functions may be concerned with maximizing throughput,
minimizing average work-in-process, or minimizing the total number of buffer slots,
subject in each case to appropriate constraints. In detail:
Problem BAP-A:
max X (n) = max X (N2 , . . . , NK )
subject to
K
∑ Nj = N
j=2
Nj ≥ 0
132 5 The Buffer Allocation Problem
where n = (N2 , N3 , . . . , NK ) denotes the vector of the buffer sizes, N j , j = 2, . . . , K,

which are integer numbers and X (n) = X (N2 , . . . , NK ) is the throughput of the
K-station production line as a function of the buffers’ sizes vector.
Problem BAP-B:
min L(n) = min L(N2 , . . . , NK )
subject to
X (n) = X (N2 , . . . , NK ) ≥ X0
K
∑ Nj ≤ N
j=2
Nj ≥ 0
where L(n) = L(N2 , . . . , NK ) denotes the average WIP, W IP, as a function of N j ,

j = 2, . . . , K, which are integer numbers and X0 is a specified throughput level.
Problem BAP-C:
K
min N = ∑ Nj
j=2
subject to
X (n) = X (N2 , . . . , NK ) ≥ X0
Nj ≥ 0
where N j , j = 2, . . . , K are integer numbers and X0 is a specified throughput level.

As a consequence of the proof by Meester and Shanthikumar (1990) of the con-
cavity of the throughput of tandem queueing systems with finite buffer storage space,
it is clear that problem BAP-A is an increasing function of the total buffer space N.
Hence the results obtained for problem BAP-A can be used to solve problem BAP-C.
Thus, the above three problems really are reduced to two problems. Generally speak-
ing, because of the discrete nature of the buffer allocation and the unavailability of
expressions in closed form, numerical approaches to the solution of the problems are
inevitable even in situations with relatively small number of states. Such approaches
are discussed below.
5.2 Solution of the Buffer Allocation Problems

A useful division of buffer allocation problems is based on the concepts of short and
longer lines. Although precise definitions are impossible, in our experience, short
lines might be designated as production lines with up to six stations with a maxi-
mum of up to 20 buffer slots in total. Lines with specifications outside these ranges
may be classified as longer lines. Either of these lines consisting of single-machine
stations may be balanced or unbalanced. By balanced is meant a line with equal
mean service or processing times at each of the K stations. Unbalanced lines may
5.2 Solution of the Buffer Allocation Problems 133
Evaluative Optimization
Methods methods
Fig. 5.1. General process of solution of buffer allocation problems
be classified as (i) μ -unbalanced, where the allocation of work is unbalanced across

the stations, (ii) c.v.-unbalanced, where the c.v. (coefficient of variation) of the ser-
vice times at the stations are not identical and (iii) fully unbalanced lines which are
both μ -unbalanced and c.v.-unbalanced. A further characteristic of production lines
is reliable and unreliable. Our definition of reliable lines is that each station of the
line cannot fail, i.e., if free, the single server at the station is available to immediately
serve a waiting part and all servers are perfectly reliable.
The method of solution of the buffer allocation problem follows the process as
indicated in Figure 5.1. This solution process, not unique to the buffer allocation
problem, basically consists of a loop process which leads to an optimal solution after
a finite number of iterations. The analyst initiates the process by specifying an initial
configuration of the system assuming values for the decision variables, in this case,
the buffer allocation. Clearly, the experience of the analyst is of value at this stage.
The evaluative method determines the value of the performance objective for the
system as specified. The optimization or generative method (search algorithm) takes
over and presents to the evaluative method a sequence of candidate systems with
new values for the decision variables. The evaluative method calculates for each sys-
tem presented the value of the performance measure. The effectiveness of the overall
process depends on the efficiency with which the generative method generates suit-
able candidate systems for evaluation as well as on the effictiveness of the evaluative
method itself. Clearly, a very efficient internal accounting process is required.
Evaluative methods, which predict the performance measures of the system, are
based on aggregation approaches, decomposition methods, expansion or any approx-
imate methods, Markovian exact models and simulation. Optimization methods, on
the other hand, which lead to the optimal values of the decision variables and work on
results of the evaluative methods, are very varied as will be described further below
in Sections 5.3 and 5.4. A clear distinction should be made between an exact solution
to the specified problem (however idealized initially) and an approximate solution to
the same problem. For example, a Markovian model gives an exact solution to series
production lines, whereas a decomposition approach to the same system attempts
to develop a solution that is very close to the exact solution. However, decompo-
sition methods because of their computational efficiency may be the only practical
approach for the solution of systems with a large number of states. Again, as far as
optimization is concerned, it is clear that enumeration, if possible, will give an exact
optimal solution. Enumeration is generally only possible for very small systems and
other approaches to optimization are required for large systems. Such other numer-
ical approaches may have some difficulty in actually reaching the precise optimal
solution and so, the analyst should exercise caution in ascertaining that the optimal
solution has been achieved.
In the following Sections 5.3 and 5.4 an attempt is made to assist the reader in
understanding the methods used for the solution of the buffer allocation problems in
the specified production lines.
5.3 Solution Approaches to the BAP in Short Lines

The methods used for the solution of the buffer allocation problems in long lines are
also applicable to short lines. Here, however, the objective is to indicate to the reader
those methods which are practically applicable only to short lines. Buffer alloca-
tion problems in short production lines whether balanced or unbalanced, reliable or
unreliable are generally solved using Markovian and expansion evaluative methods
(see Sections 2.1 and 2.3 in Chapter 2) with complete enumeration as the optimizing
method. The latter algorithm is available at the website associated with this text with
abbreviated name CE.
An example of the results obtained are given in Papadopoulos and Vidalis (1998).
Their model consists of single-machine K stations which are perfectly reliable with
processing times following an exponential or Erlang-k distribution with k phases of
service (k = 2, 3, 4), with K = 13 for the case of the exponential distribution and
K = 6 for the case of the Erlang-2 distribution. The objective function was to maxi-
mize the throughput of the line, by allocating a given total number, N, of buffer slots
among the K − 1 intermediate buffers of the line (problem BAP-A). Using essentially
an optimization technique based on the Hooke-Jeeves algorithm, the authors derived
results on the form of the optimal buffer allocation for the exponential and Erlang-k
service times. A few sample results for production lines with up to K = 11 stations
and exponential and Erlang-2 service times are shown in Figure 5.2.
From an analysis of their results the authors derived two basic design rules, viz.,
• Rule 1: This is a confirmation and an extension of a rule earlier stated by Conway
et al. (1998) to the effect that “For the optimal buffer allocation of N buffer slots
among the K − 1 buffers of a K-station line, first allocate equally these N slots
to all the K − 1 buffers, i.e., allocate n = [N/(K − 1)] slots to each of the K − 1
buffers, where [x] is the maximum integer ≤ x, and the remaining E buffer slots so
that E + 1 sub-lines with equal “distance” to be produced.”
Explanation of Rule 1: As there are E privileged buffers denoted by B1 , . . . , BE
(which receive one extra buffer slot above the uniform allocation), the original line
may be divided into E + 1 buffer sub-lines as follows:
B1 → B1 ; B1 → B2 ; · · · ; BE → BK .
The measure “distance,” D(Bi , B j), is defined as the number of buffers in between
buffer Bi and buffer B j , not including buffer Bi and buffer B j .
K N=0mod(K-1) N=1mod(K-1) N=2mod(K-1) N=3mod(K-1) N=4mod(K-1) N=5mod(K-1) N=6mod(K-1) N=7mod(K-1) N=8mod(K-1)
3 (q,q) (r,q)
4 (q,q,q) (q,r,q) (q,r,r)
5 (q,q,q,q) (q,q,r,q) (q,r,r,q) (q,r,r,r)
6 (q,q,q,q,q) (q,q,r,q,q) (q,r,q,r,q) (q,r,r,r,q) (q,r,r,r,r)
7 (q,q,q,q,q,q) (q,q,q,r,q,q) (q,q,r,r,q,q) (q,r,q,r,r,q) (q,r,r,r,r,q) (q,r,r,r,r,r)
8 (q,q,q,q,q,q,q) (q,q,q,r,q,q,q) (q,q,r,q,r,q,q) (q,r,q,r,q,r,q) (q,r,r,q,r,r,q) (q,r,r,r,r,r,q) (q,r,r,r,r,r,r)
9 (q,q,q,q,q,q,q,q) (q,q,q,q,r,q,q,q) (q,q,r,q,q,r,q,q) (q,q,r,q,r,r,q,q) (q,r,q,r,r,q,r,q) (q,r,r,q,r,r,r,q) (q,r,r,r,r,r,r,q) (q,r,r,r,r,r,r,r)
10 (q,q,q,q,q,q,q,q,q) (q,q,q,q,r,q,q,q,q) (q,q,q,r,q,r,q,q,q) (q,q,r,q,r,q,r,q,q) (q,q,r,r,q,r,r,q,q) (q,r,q,r,r,r,q,r,q) (q,r,r,r,q,r,r,r,q) (q,r,r,r,r,r,r,r,q) (q,r,r,r,r,r,r,r,r)
11 (q,q,q,q,q,q,q,q,q,q) (q,q,q,q,q,r,q,q,q,q) (q,q,q,r,q,q,r,q,q,q) (q,q,q,r,q,r,q,r,q,q) (q,q,r,q,r,r,q,r,q,q) (q,q,r,r,q,r,r,q,r,q) (q,r,q,r,r,r,r,q,r,q) (q,r,r,r,q,r,r,r,r,q) (q,r,r,r,r,r,r,r,r,q)
q : denotes the uniform allocation of the n = [N/K_1] slots whereas r denotes one extra buffer slot, i.e., r = q + 1
Given the integers N, i and j, the expression N = i (mod j) means that N has as remainder the integer i when divided by the integer j
Fig. 5.2. Schematic representation of the form of optimal buffer allocation in terms of E and K of balanced production lines with exponential and
Erlang-2 service times
5.3 Solution Approaches to the BAP in Short Lines
135
• Rule 2: Each buffer that is allocated an extra slot must be closer to a central buffer
that has been also allocated extra slot(s) rather than to a buffer lying toward the
end stations which has also been alloted an extra slot.
However, Hillier and So (1995) noted that in BAP-A the optimal buffer allocation
may begin to deviate from the uniform as possible allocation when the number of
buffer slots available increases.
An allocation routine based on Rule 1 and Rule 2 (which have been obtained
empirically) was developed and may be used to obtain the optimal buffer configura-
tion.
With respect to μ -balanced unreliable production lines, Papadopoulos and Vidalis
(1999) considered the buffer allocation problem, BAP-A, and in particular the effects
of the distribution of service times, the availability (assumed identical) of the m ≤ K
unreliable stations and of the repair rates on the throughput and the optimal buffer
allocation. The assumptions of the model include single-machine stations, exponen-
tial or Erlang-k service times at each station and times to failure and repair times are
all exponentially distributed with different mean rates. Complete enumeration was
the search procedure used initially, but as the experimentation continued an efficient
reduction search procedure was developed.
With the usual definition of the availability, Ai , of unreliable stations
ri
Ai = ,
ri + βi
where 1/ri is the mean time to repair station i and 1/βi is the mean time to failure of
station i, some of the conclusions may be given as follows:
1. As far as the optimal buffer allocation (OBA) is concerned, there are three sep-
arate cases. For small values of the availability of the m unreliable stations
(m ≤ K):
(i) when m < K and even, the OBA resembles the shape of a bowl;
(ii) when m < K and odd, the OBA resembles the shape of a non-symmetric
bowl and
(iii) when m = K, the OBA resembles the shape of an “inverted bowl.” This
observation is in contrast with the well-known result about the uniformity
of the optimal buffer allocation in a balanced line.
In all three cases, as the availability (assumed identical) of the unreliable stations
tends to unity, all the buffers are allocated evenly the buffer slots, at the optimal
situation.
2. As the number of service phases increase (from exponential to Erlang-k (k > 1)
distribution) then
(i) the coefficient of variation (c.v.) of the effective service time decreases and
this results in an increase in the throughput of the line;
(ii) it becomes more difficult to justify economically the provision of extra
buffer spaces, i.e., the marginal increase in throughput per buffer slot is
decreasing;
5.3 Solution Approaches to the BAP in Short Lines 137
(iii) the shape of the OBA as given in conclusions 1((i), (ii), (iii)) above become
more pronounced and
(iv) there is a linear relationship between the value (assumed identical) of the
c.v. of the service time distribution and the number of buffer slots required
to achieve a given throughput.
In a further paper, Papadopoulos and Vidalis (2001a) have considered the buffer
allocation in short unreliable and unbalanced production lines with K ≤ 6 stations.
Times to failure are assumed to be exponential, whereas service and repair times are
assumed to follow any Erlang-k distribution, with k ≥ 1. Single-machine stations are
also assumed. An algorithm was developed to solve the buffer allocation problem in
this type of production line. The algorithm is available at the website associated with
this text with abbreviated name BA . This algorithm is a stand-alone optimization
algorithm which cooperates with the MARKOV evaluative algorithm and consists of
the following steps:
1. Preparation for a ‘good’ initial buffer allocation
Order the stations M1 , . . . , MK of the production line from the slower to the faster
based on the value of the mean effective service rate, ei = μi Ai = μi ri +riβi , where
μi is the mean service rate of station i, ri is the mean repair rate of station i and βi
is the mean failure rate of station i. Let this arrangement be: M1 , . . . , MK .
2. Determination of a ‘good’ initial buffer allocation
Apply the following linear buffer allocation scheme (LBAS):
(i) The buffer that is located toward the center of the actual line and next to
station M j is assigned a weight of 2(K + 1 − j);
(ii) The buffer that is located toward the end of the line and next to station M j is
assigned a weight of 2(K + 1 − j) − 1;
(iii) The central buffer is assigned a weight of K (if K is odd) and when there are
two central buffers (if K is even) these are equally weighted K/2.
3. Search for the OBA using a sectioning method Starting with the ‘good’ initial
buffer vector determined in Step 2, above, use the sectioning (segmentation)
routine to find the optimal or near optimal buffer vector. More specifically this
search loop operates by increasing or decreasing by one unit each of the K − 1
intial buffer decision variables and evaluating the throughput for the correspond-
ing buffer allocation. The buffer allocation that gives the maximum throughput
during this process is the initial buffer allocation for the next cycle of searches. A
usual stopping criterion is adopted.
Example 1
A numerical example is given below to show the application of the above algorithm,
taken from Papadopoulos and Vidalis (2001a).
Consider a four-station unreliable production line with the service and repair
times following the two-stage Erlang distribution. The various parameters of this
system have the following values: mean service rates: μ1 = 3.7, μ2 = 1.5, μ3 = 1.1,
μ4 = 3; mean repair rates: r1 = 0.17, r2 = 0.37, r3 = 0.78, r4 = 0.5; mean failure

rates: β1 = 0.07, β2 = 0.11, β3 = 0.49, β4 = 0.19. Find the optimal buffer alloca-
tion of N = 9 total buffer slots among the three intermediate buffer locations of the
production line which maximizes its throughput.
Step 1: Ordering of stations
By application of the relevant formulae, ei = μi Ai , where Ai , ei denote, respectively,

the availability and the mean effective service rate (efficiency) of the unreliable sta-
tion i, i = 1, . . . , 4, one finds: e1 = 3.0683, e2 = 1.3059, e3 = 0.8371, e4 = 2.521.
Therefore, the new ordering of the stations from the bottleneck station to the faster
station is:
M1 = M3 , M2 = M2 , M3 = M4 , M4 = M1 .
Step 2: Application of the LBAS
(i & ii) Give preferential treatment to the buffers that are close to the bottleneck
stations. Buffer B3 is assigned 2(4 + 1 − 1) = 8 points, whereas buffer B4 is
assigned 2(4 + 1 − 1) − 1 = 7 points. Buffer B3 is assigned 2(4 + 1 − 2) = 6
points, whereas buffer B2 is assigned 2(4 + 1 − 2) − 1 = 5 points. Buffer B4
is assigned 2(4 + 1 − 3) = 4 points and buffer B2 is assigned 2(4 + 1 − 4) = 2
points.
(iii) Give preferential treatment to the central buffer(s). There is only one central
buffer, the B3 , which is assigned K = 4 points.
Adding all these points one may see that buffer B2 is assigned 7 points (19.44%),
buffer B3 18 points (50%) and buffer B4 11 points (30.56%). These percentages split
the total number of buffer slots, N = 9 as follows:
. . .
B2 = 1.75 = 2, B3 = 4.5 = 5, B4 = 2.75 = 3.
However, this allocation gives a total ∑i=2 K=4

Bi = 10 > 9 = N. To overcome this, we
subtract the one extra buffer slot from the buffer with the least priority, which in this
case is buffer B2 . Thus, the initial buffer allocation vector is the (1, 5, 3).
Step 3: Search for the optimal buffer allocation, OBA, via the application of the
sectioning approach
This search process consists of 13 iterations, which are given in Table 5.1, to find
the OBA which is the (0, 6, 3) allocation. By complete enumeration, 55 iterations are
needed to find the OBA.
Comments on Step 3: The initial buffer allocation is the (1, 5, 3). The first cycle
of iterations consists of three two-step searches (two searches per component of the
buffer vector), i.e., #1 and #2 for buffer B2 , #3 and #4 for buffer B3 and #5 and #6
for buffer B4 (see Table 5.1). While keeping the buffer combination corresponding
to highest throughput, in the first step, the value of the buffer component is increased
by 1, whereas in the second step, this value is decreased by 1. The buffer allocation
Table 5.1. The 13 iterations to find the OBA in the production line of example 1
Iteration Buffer allocation Throughput

# B2 B3 B4 XK=4
0 1 5 3 0.6453
1 2 4 3 0.6408
2 0 6 3 0.6471
3 0 7 2 0.6470
4 0 5 4 0.6431
5 0 5 4 0.6431
6 0 7 2 0.6470
7 0 6 3 0.6471
8 1 5 3 0.6453
9 0 7 2 0.6470
10 0 5 4 0.6431
11 0 5 4 0.6431
12 0 7 2 0.6470
13 0 6 3 0.6471
that gives the maximum throughput, (0, 6, 3), is kept and forms the initial buffer allo-
cation for the second cycle of iterations. This cycle consists of only five searches
(#8 – #12) as the first element of the initial buffer allocation is zero and it can only
take the value 1. The resulting allocations do not give higher throughput than allo-
cation (0, 6, 3) and thus the iteration procedure terminates at this point. There is no
need to go further as the initial buffer allocation for the third cycle of searches is
identical to that of the second cycle and therefore would lead to the same result.
This algorithm was found to give the exact optimal allocation in over 97% of the
373 experiments undertaken.
With respect to the problem of allocating buffer space with the objective of the
minimization of the average work-in-process, WIP, subject to a minimum required
throughput (problem BAP-B, stated above), Papadopoulos and Vidalis (2001b)
extended the work by So (1997) and showed:
• the choice of minimum throughput level has a critical impact on the minimum WIP
achievable;
• A “self-similarity” phenomenon prevails in the case of balanced lines which sig-
nificantly reduces the search space required to determine the buffer allocation
associated with minimum WIP.
Figure 5.3 and Figure 5.4 give the throughput and the WIP, respectively, as a function
of the ordered buffer allocations for a five-station production line with N = 5 buffer
slots to be allocated among the 4 intermediate buffers, showing the “self-similarity”
phenomenon. The details of the sequence of the ordered buffer allocations are given
in Papadopoulos and Vidalis (2001b).
0,64
140
0,62
0,6
0,58
5 The Buffer Allocation Problem
X5(N = 5)
Throughput
0,56
0,54
0,52
0,5
)
)
)
)
)
)
)
)
3)
3)
4)
1)
2)
0)
1)
0) 3) 1) 2) 0) 2) 0) 1) 0)
-1
-0
0) 0)
-3
-2
-5
-2
-0
-0
-
-1
-1
1-
0-
0-
0-
2-
2-
0- 1- 2- 0- 2- 1- 0- 2- 1- 0- 0-
-1
-0
-0
-4
-2
-3
-2
-1
-0
-4
-3
2-
0-
1-
2
4-
2
1-
0-
5- 0- 0- 1- 1- 2- 0- 0- 1- 0- 1-
-3
-
-
-1
-3
-2
-1
-0
-
-
-0
-
-1
-0
- - - - - - - - - - -
1-
0-
(1
(0
(1
(2
(0
(0
(1
(1
(0
(
(1
(0
(0
(0
(0
(0
(0
(1 (2 (2 (2 (2 (3 (3 (3 (4 (4
Fig. 5.3. Throughput as a function of the ordered buffer allocations for K = 5 and N = 5, showing the “self-similarity” phenomenon
WIP Average
2
3
4
5
6
7
8
9
(0-0-0-5)
(0-0-2-3)
(0-0-4-1)
(0-1-0-4)
(0-1-2-2)
(0-1-4-0)
(0-2-1-2)
(0-2-3-0)
(0-3-1-1)
(0-4-0-1)
(0-5-0-0)
(1-0-1-3)
(1-0-3-1)
(1-1-0-3)
(1-1-2-1)
(1-2-0-2)
(1-2-2-0)
(1-3-1-0)
(2-0-0-3)
(2-0-2-1)
(2-1-0-2)
(2-1-2-0)
(2-2-1-0)
(3-0-0-2)
(3-0-2-0)
(3-1-1-0)
(4-0-0-1)
(4-1-0-0)
Fig. 5.4. Average WIP, W IP, as a function of the ordered buffer allocations for K = 5 and N = 5, showing the “self-similarity” phenomenon
141 5.3 Solution Approaches to the BAP in Short Lines
The authors found that in many cases the optimal buffer allocation (OBA) has
a monotonic increasing characteristic of the buffer allocations, i.e., BK ≥ BK−1 ≥
BK−2 ≥ · · · and made observations on how to allocate the buffer slots when this
property does not hold. An algorithm was developed to reduce the search space.
To understand the details of this algorithm, some definitions developed by the
authors are required. For a production line with K stations and N buffer slots that
are to be allocated among the K − 1 intermediate buffers, let B denote the set of all
possible buffer allocations.
B = {B1 , B2 , . . . , BL },
where

N +K−2 (N + 1)(N + 2) · · ·(N + K − 2)
L= =
K−2 (K − 2)!
The Bi ’s are vectors with K − 1 elements which are nonnegative integer numbers of
the form
Bi = {Bi2 , Bi3 , . . . , BiK },
where Bi j , 2 ≤ j ≤ K expresses the capacity of the jth buffer.
Set B is split into N + 1 equivalence buffer classes which are characterized as
classes of first generation, the following: [0], [1], [2], . . . , [N]. A class of first genera-
tion, say the [I], 0 ≤ I ≤ N, consists of all the allocations Bi with Bi2 = I. All classes
of first generation I, 0 ≤ I ≤ N, are divided into N + 1 − I classes which are defined
as classes of second generation. In turn, each class of second generation, e.g., class
[I, J], 0 ≤ I ≤ N, 0 ≤ J ≤ N + 1 − I, consists of all the buffer allocations with the
first two elements equal to I and J, respectively, and is divided into N + 1 − (I + J)
classes of third generation and so forth. Each element of the (K − 2) generation class
specifies the contents of the last buffer.
Definition of subsequent buffer classes: Let [ζ1 , ζ2 , . . . , ζκ ] and [η1 , η2 , . . . , ηκ ]
be two buffer classes of the same generation κ , 1 ≤ κ ≤ K − 3. We say that
[η1 , η2 , . . . , ηκ ] is subsequent to [ζ1 , ζ2 , . . . , ζκ ] if ηi = ζi for i = 1, 2, . . . , κ − 1 and
ηκ = ζκ + 1.
The input data to the algorithm consists of:
• K, the number of stations of the line,
• μi ’s, i = 1, . . . , K, the mean service rates,
• N, the number of total buffer slots to be allocated among the K − 1 buffers of the
line and
• X0 , the minimum throughput level that has to be achieved.
The algorithm finds the integer values i2 , i3 , . . ., iK−2 which correspond to the
classes of first, second, . . . , (K − 3)rd generation up to which the average throughput
increases and at which the throughput attains its maximum value. The steps of the
algorithm are as follows:
Step 1: (Initialization phase)

Step 1.1: Put B2 = B3 = . . . = BK−2 = 0 and search for the maximum value
i = iK−1 that buffer BK−1 can take such that for any j = 0, 1, . . . , N:
XK (0, 0, . . . , i, N − i) ≥ XK (0, 0, . . . , j, N − j).
i.e., throughput is maximized.

Step 1.2: Put B2 = B3 = . . . = BK−3 = 0 and search for the maximum
value i = iK−2 that buffer BK−2 can take such that for any j =
0, 1, . . . , iK−1 :
max XK (0, 0, . . . , i) ≥ max XK (0, 0, . . . , j).
i.e., throughput is maximized.

Step 1.3: Find the upper values of the remaining buffers, i.e., the values
iK−3 , iK−4 , . . . , i3 , i2 ([x] denotes the maximum integer less than or
equal to x):

N
B2 = 1, . . . , i2 = ,
K −1
B2 = 0, . . . , i3 (= i4 − 1),
.. ..
.=.
BK−4 = 0, . . . , iK−4 (= iK−3 − 1),
BK−3 = 0, . . . , iK−3 (= iK−2 − 1),
K−1
BK = N − ∑ Bj
j=2
Step 2: (Search phase) The algorithm searches for the optimal buffer allocation
which minimizes the average WIP, W IP, and gives a throughput that is
greater than or equal to the specified level, X0 in the reduced search space
given by the values of iK−1 , . . . , i2 .
Example 2
A numerical example is given below to show the application of the above algorithm,
taken from Papadopoulos and Vidalis (2001b).
Consider a five-station reliable balanced production line with the service times
following the exponential distribution. Find the optimal buffer allocation of N = 5
total buffer slots among the four intermediate buffer locations of the production line
which minimizes the average WIP, W IP, and gives a throughput that exceeds a given
level X0 = 0.5961.
Step 1.1: The upper value of buffer BK−1 = B4 , iK−1 = i4 = 4, as the buffer
allocation (0, 0, 4, 1) gives the maximum throughput, 0.5597, after 6 searches (of the
respective number) of buffer allocations of class [0, 0] (see Table 5.2).
Table 5.2. Searching in classes [0, 0], [0, 1], [0, 2] and [0, 3]
Iteration Equivalence Buffer Throughput Average WIP

# buffer class allocation XK=5 W IP
1 (0, 0, 0, 5) 0.5146
2 (0, 0, 1, 4) 0.5441
3 [0, 0] (0, 0, 2, 3) 0.5550
4 (0, 0, 3, 2) 0.5590
5 (0, 0, 4, 1) 0.5597
6 (0, 0, 5, 0) 0.5557
7 (0, 1, 0, 4) 0.5580
8 (0, 1, 1, 3) 0.5872
9 [0, 1] (0, 1, 2, 2) 0.5974 4.1518
10 (0, 1, 3, 1) 0.5990 4.3964
11 (0, 1, 4, 0) 0.5887
12 (0, 2, 0, 3) 0.5800
13 (0, 2, 1, 2) 0.6061 4.5340
14 [0, 2] (0, 2, 2, 1) 0.6114 4.7960
15 (0, 2, 3, 0) 0.5982 5.5007
16 (0, 3, 0, 2) 0.5895
17 [0, 3] (0, 3, 1, 1) 0.6096 5.1276
18 (0, 3, 2, 0) 0.5982 5.7633
Step 1.2: The upper value of buffer BK−2 = B3 , iK−2 = i3 = 2, as class [0, 2] gives
the maximum throughput, 0.6114, after 12 searches: 5 in class [0, 1], 4 in class [0, 2]
and 3 in class [0, 3] (see Table 5.2).
Step 1.3: The upper value of buffer BK−3 = B2 , iK−3 = i2 = 1 = [5/(5 − 1)] ([x]
is the largest integer less than or equal x). This is the end of the initialization phase.
Step 2: The values of the buffers are determined:
4
B2 = 1, B3 = 0, 1, 2, B4 = 0, 1, 2, 3, 4, B5 = N − ∑ Bi .
i=2
Comments
Since class [0] was already checked in Steps 1.1 and 1.2, B2 takes only the value 1.
The number of iterations (searches) in Step 2 are the following 12: 5 in class [1, 0], 4
in class [1, 1] and 3 in class [1, 2] (see Table 5.3).
Again none of all these iterations gives W IP less than 4.1518, the minimum
value found so far. Thus, the buffer allocation that minimizes the W IP is (0, 1, 2, 2)
and the corresponding minimum W IP is 4.1518 for the selected throughput level
X0 = X0,2 = 0.5961 that has to be exceeded. The total numer of searches from all
steps of the algorithm is 30 as compared with the 56 allocations from enumeration.
This means that the algorithm leads to a 46% reduction in the number of searches to
find the optimal buffer allocation.
From experimental work, the percentage reduction of the search space was found
to be generally over 50%.
5.4 Solution Approaches to the BAP in Longer Lines 145
Table 5.3. Searching in classes [1, 0], [1, 1] and [1, 2]
Iteration Equivalence Buffer Throughput Average WIP

# buffer class allocation XK=5 W IP
19 (1, 0, 0, 4) 0.5438
20 (1, 0, 1, 3) 0.5801
21 [1, 0] (1, 0, 2, 2) 0.5935
22 (1, 0, 3, 1) 0.5963 4.8100
23 (1, 0, 4, 0) 0.5860
24 (1, 1, 0, 3) 0.5857
25 (1, 1, 1, 2) 0.6202 5.1638
26 [1, 1] (1, 1, 2, 1) 0.6275 5.4941
27 (1, 1, 3, 0) 0.6096 6.1794
28 (1, 2, 0, 2) 0.6012 5.5889
29 [1, 2] (1, 2, 1, 1) 0.6275 5.8978
30 (1, 2, 2, 0) 0.6114 6.5231
5.4 Solution Approaches to the BAP in Longer Lines

It should be noted that optimization procedures used in long lines (with K > 6
stations in series) may also be used for shorter lines. Such procedures were not con-
sidered in Section 5.3 because, with respect to shorter lines, exact results are available
by way of enumeration. The optimization techniques used for longer lines in effect
are approximate techniques. Limitations of the evaluative technique used are cru-
cial in determining which technique to use in short or longer lines. Basically, as far
as short lines are concerned, Markov-based evaluative techniques are adequate and
accurate, whereas in longer lines approaches based on Markovian analysis are not
applicable because of the required computer storage issues. Decomposition methods
extend greatly the range of cases which may be evaluated but at the cost of some
reduction in accuracy. However, this reduction in accuracy may not be significant
in practical situations. To achieve a solution to the system in an efficient manner,
it is necessary for both the evaluative technique and the optimization technique to
be efficient (see Figure 5.1 which shows the feedback loop between the two tech-
niques used in determining the solution to the system). This leads to the exclusion of
simulation as a possible optimization technique in large systems. The decomposition
approach (described in Chapter 2) is the main evaluative technique used in the anal-
ysis of large lines. Associated with this evaluative approach, there are a number of
different choice possibilities with respect to the optimization techniques. The latter
include:
• Gradient methods
• Dynamic programming
• Simulated annealing
• Genetic algorithms and
• Tabu search
Gradient methods
The well-known numerical analysis approach, gradient method (see Ho et al., 1979
and Gershwin and Schor, 2000), may be adopted to solve problem BAP-A, and the
steps of an appropriate algorithm are given below:
Step 1: Specify an admissible set of initial buffer allocations and use the evaluative
technique to determine the throughput XK (B2 , B3 , . . . , BK ) of the line.
Step 2: Calculate the gradient g = (g2 , g3 , . . . , gK ) given by
X (B2 , . . . , Bi + δ Bi , . . . , BK ) − X (B2, . . . , Bi , . . . , BK )
gi =
δ Bi
where δ Bi is a step size of integer value.
Step 3: Obtain the search direction v by projecting the gradient g = (g2 , g3 , . . . , gK )
on to the hyperplane such that:
K
∑ δ Bi = 0
i=2
giving vi = gi − ḡ, where,
1 K
ḡ = ∑ gi .
K − 1 i=2
Step 4: Find Λ such that X (B + Λ v) is maximized.

Step 5: Define B = B + Λ v.
Step 6: If B is close to B, Stop (B is the optimal buffer allocation), otherwise, let
B = B and go to Step 2 and continue.
A difficulty of the above procedure arises if Bi = 0 and vi < 0 for some i. In that case,
the new direction is calculated by deleting the ith component of g and setting vi = 0.
The other components of v are determined as before and the process is continued
until a feasible v is determined or all components of g are deleted.
In Gershwin and Schor (2000), an algorithm for the solution of the BAP-B prob-
lem, designated primal by the authors, using the solution of the BAP-A problem is
given. Gershwin and Schor noted that the maximum throughput as a function of the
total buffer space, N, for a three-station two-buffer system is monotonically increas-
ing in N and may be approximated by two linear functions with different slopes.
This observation is crucial to the development of an algorithm for the solution of the
BAP-B problem, the steps of which are given below.
Step 1: Let N 0 = (0, 0, . . . , 0). Calculate Xmax (0, 0, . . . , 0) using an appropriate eval-
uative method (Markovian for short lines and decomposition for longer
lines).
Step 2: Specify a new estimate of the total buffer slots to be allocated, N 1 and solve
problem BAP-A (designated as dual problem by Gershwin and Schor) using
the gradient algorithm given above. Set j = 2.
Step 3: Calculate N j from

N j = aN j−1 + bN j−2,
where a and b are determined from the assumed linear approximation
between maximum throughput and total buffer size. Note that because of
the two linear approximations, the slope of the line will be different from
one range of total buffer size to another.
Step 4: Use the gradient algorithm to determine Xmax (N j ).
Step 5: If Xmax is sufficiently close to the desired throughput level, stop, otherwise
return to Step 3.
The solution specifies N, the total number of buffer slots, and the distribution of
N among the intermediate K − 1 buffers of the production line.
Dynamic programming
In general, dynamic programming, DP, is a multi-stage decision process where the

objective is to allocate a limited resource sequentially over the stages so as to opti-
mize the objective function based on Bellman’s principle of optimality in that, at any
stage in the analysis, the optimal solution involves being on an optimal path from
that stage onwards. In general terms, with respect to problem BAP-B the stations
are considered to be the stages and the total number of buffer slots to be allocated
are considered to be the states. An appropriate recursive relationship must be devel-
oped having in mind the overall objective function of minimizing the total number
of buffer slots to be allocated among the intermediate buffers of the line subject to a
minimum throughput. The structure of the BAP-B problem may be utilized to effect
computational efficiencies with a dynamic programming approach.
Simulated annealing
Simulated annealing, SA, is an adaptation of the simulation of physical thermo-

dynamic annealing principles to combinatorial optimization problems. It follows
a logical improvement paradigm having regard to the exponential complexity of
the solution space. The algorithm is based on randomization and as a result there
is no certainty in relation to achieving the precise optimal solution. Spinellis and
Papadopoulos (2000a) used a simulated annealing approach for the solution of the
buffer allocation problem BAP-A in reliable (large) production lines. The authors
assumed exponential processing times at each station with mean service rates μi ,
i = 1, . . . , K. The exact numerical algorithm of Heavey, Papadopoulos and Browne
(1993) and the decomposition algorithm A3 of Dallery and Frein (1993) were used
in conjunction with the simulated annealing algorithm developed by the authors.
In the improvement process of the simulated annealing approach, the probabilistic
“uphill” energy movement avoids the entrapment of the solution in a local mini-
mum. The authors gave the correspondence between annealing in the physical world
and simulated annealing used in the optimal buffer allocation as shown in Table 5.4.
The associated simulated annealing algorithm is given in Figure 5.5.
Table 5.4. Correspondence between annealing in the physical world and simulated annealing
used for production line optimization
Physical World Production Line Optimization

Atom placement Line configuration
Random atom movements Buffer space, server, work-load movement
Energy, E Throughput, X
Energy differential, Δ E Configuration throughput differential, Δ X
Energy state probability distribution Changes according to the Metropolis criterion,
exp( −Δ
T ) > rand(0 . . . 1), implementing the
E
Boltzmann probability distribution

Temperature Variable for establishing configuration acceptance
termination
The SA procedure was run by Spinellis and Papadopoulos (2000a) with the
following characteristics based on the number of stations K:
Maximum trials at given temperature: 100 × K
Maximum successes at given temperature: 10 × K
Initial temperature: 0.5
Cooling schedule: Exponential: TJ+1 = 0.9TJ
Initial line configuration: Equal division of buffers and servers among stations with
any remaining resources placed on the station(s) in the middle
Reported time: Elapsed wall clock time in seconds.
As the reader will note, the algorithm given in Figure 5.5 is a simulated annealing
algorithm for not only distributing N total buffer slots but also simultaneously S(S ≥
K) servers and work-load normalized to K.
The algorithm is available at the website associated with this text with abbre-
viated name SA and may be accessed by selecting the corresponding genera-
tive/optimization algorithm.
The authors demonstrated the accuracy of the simulated annealing approach com-
pared to the complete enumeration for shorter lines with up to 9 stations, and in
the case of longer lines (15 stations), a comparison was made between the solutions
obtained using reduced enumeration and decomposition and simulated annealing and
decomposition. For comparing the results, the authors used the formalism S(G, E)
to describe a closed loop system using the generative method G and the evaluative
method E. A major contribution of this work was that solutions (buffer allocations)
were achieved using a simulated annealing for very large lines (with up to 400 sta-
tions). Clearly, only decomposition may be used as evaluative model in these cases.
With respect to these large lines it should be noted that the time requirements of the
simulated annealing method becomes competitive with other methods as the num-
ber of stations and the available buffer size increases as indicated in Figure 5.6 and
Figure 5.7. Figure 5.8 shows that the number of enumerations required for simulated
annealing solutions increases linearly with the number of stations.
1. [Set initial line configuration.] Set Bi ← N/(K − 1),1 set BK/2 ← BK/2 + N − ∑K i=2 N/(K − 1), set si ← S/K, set sK/2 ← sK/2 + S −
∑K i=1 S/K, set w i ← 1/K.
2. [Set initial temperature T0 .] Set T ← 0.5.
3. [Initialize step and success count.] Set I ← 0, set U ← 0.
4. [Create new line with a random redistribution of buffer space, servers, or work-load.]
(i) [Create a copy of the configuration vectors.] Set n ← n, set s ← s, set w ← w.
(ii) [Determine which vector to modify.] Set ← rand[0 . . . 2].
(iii) if = 0 [Create new line with a random redistribution of buffer space.] Move rn space from a source buffer Brs to a destination buffer Brd : set
rs ← rand[2 . . . K), set rd ← rand[2 . . . K), set rn ← rand[1 . . . Brs + 1), set Brs ← Brs − rn , set Brd ← Brd + rn .
(iv) if = 1 [Create new line with a random redistribution of server allocation.] Move rn servers from source station srs to a destination station
srd : set rs ← rand[1 . . . K), set rd ← rand[1 . . . K), set rn ← rand[1 . . . srs + 1), set srs ← srs − rn , set srd ← srd + rn .
(v) if = 2 [Create new line with a random redistribution of work-load.] Move rn work-load from source station srs to a destination station srd :
set rs ← rand[1 . . . K), set rd ← rand[1 . . . K), set rn ← rand(0 . . . wrs ], set wrs ← wrs − rn , set wrd ← wrd + rn .
5. [Calculate energy differential.] Set ΔE ← X(n, s, w − X(n , s , w ).
6. [Decide acceptance of new configuration.] Accept all new configurations that are more efficient and, following the Boltzmann probability
distribution, some that are less efficient: if ΔE < 0 or exp( −ΔE
T ) > rand(0 . . . 1), set n ← n , set s ← s , set w ← w , set U ← U + 1.
7. [Repeat for current temperature.] Set I ← I + 1. If I < maximum number of steps, go to step 4.
8. [Lower the annealing temperature.] Set T ← cT (0 < c < 1).
9. [Check if progress has been made.] If U > 0, go to step 3; otherwise the algorithm terminates.
1 We employ the notation for open intervals (a . . . b), closed intervals [a . . . b], and the floor function x to create sets of random numbers in a given
range. For this purpose we use a function rand() generating random real numbers in the range specified by an interval. Specifically, rand(a . . . b)
will produce a random integer x : a < x ≤ b, and rand[a . . . b] will produce a random integer x : a ≤ x ≤ b. The half closed intervals, [a . . . b) and
(a . . . b], work in a similar way.
5.4 Solution Approaches to the BAP in Longer Lines
Fig. 5.5. Simulated annealing algorithm for distributing N buffer space, S servers, and K work-load in a K-station line
149
9 stations; 1–20 buffers

1e+006
Evaluated configurations
100000
10000
1000
100 S(CE, Deco-1)

S(RE, Deco-1)
10 S(SA, Deco-1)
1
0 5 10 15 20
Buffer space
Fig. 5.6. Performance of simulated annealing S(SA, Deco) compared with complete S(CE,
Deco) and reduced S(RE, Deco) enumerations for 9 stations (left, middle) (Note the log10
scale on the ordinate axis)
15 stations; 1–30 buffers

1e+011
1e+010
1e+009
1e+008
1e+007
1e+006
100000
10000
1000 S(CE, Deco-1)
S(RE, Deco-1)
100
S(SA, Deco-1)
10
0 5 10 15 20 25 30
Buffer space
Fig. 5.7. Performance of simulated annealing S(SA, Deco) compared with complete S(CE,
Deco) and reduced S(RE, Deco) enumerations for 15 stations (left, middle) (Note the log10
scale on the ordinate axis)
Genetic algorithms
Another optimization method for solving the buffer allocation problem is based on
genetic algorithms. These are global stochastic optimization techniques that avoid a
number of the shortcomings exhibited by local search techniques on difficult search
spaces and are based on the mechanics of natural selection and genetics. Spinellis
and Papadopoulos (2000b) used this approach for reliable lines with exponentially
SA for 20–400 stations; #buffers = 3 * stations

1.4e+006
1.2e+006
1e+006
800000
600000
400000
200000
0
0 50 100 150 200 250 300 350 400
Station number
Fig. 5.8. Number of enumerations required for simulated annealing vs. the number of stations
(Note the log10 scale on the ordinate axis)
distributed service times with mean service rates μi , i = 1, . . . , K for the solution of
the BAP-A problem.
Genetic algorithms rely on modeling the buffer allocation problem as a popula-
tion of organisms. Each organism represents a possible valid solution to the problem.
Organisms are composed of alleles representing parts of a given solution. In each
iteration which corresponds to a generation, a new population is created by retaining
all solutions and generating new solutions from the previous population using three
basic genetic operators, viz., reproduction operator, crossover operator and mutation
operator.
The genetic algorithm for solving the BAP-A problem is available at the website
associated with this text with abbreviated name GA. This algorithm may be described
in the following steps:
1. [Initialize a population of size M.] Set P0...M,0...N ← rand[0 . . . K − 1).
2. [Evaluate population members creating throughput vector X.] For i ← 0 . . . M: set
Xi ← XK (Pi ).
3. [Create roulette selection probability vector R.] Set Ri ← ∑ij=0 (X j / ∑M k=0 Xk ).
4. [Create new population using crossovers from the previous population.] For

i ← 0 . . . M: if rand[0 . . . 1) < crossover rate, set c ← rand[0 . . . M), set Pi,0...c ←

PRr ,0...c , set Pi,c+1...N ← PRr ,c+1...N ; otherwise set Pi ← PRr by selecting each r
using the roulette selection probability vector so that Rr ≤ rand[0 . . . 1) < Rr+1 .
5. [Introduce mutations.] For i ← 0 . . . M: for j ← 0 . . . N: if rand[0 . . . 1)
< mutation rate, set Pi, j ← rand[0 . . . K − 1).
6. [Keep fittest organism for elitist selection strategy.] Select k so that Xk ≥ X0...M ,
set Prand[0...M) ← Pk .
7. [Make new population the current population.] Set P ← P .
8. [Loop based on the population’s variance.] If ∑Pi=0 |Xk − Xi | > maximum variance
go to step 2; otherwise the algorithm terminates with the optimal line setup in Pk .
The implementation of genetic algorithms may be tuned using different parameters.

Spinellis and Papadopoulos used the parameters that Grefenstette (1986) derived:
• a populations size of 50,
• a crossover rate of 0.6,
• a mutation rate of 0.0001,
• a generation gap of 1 (the entire population is replaced during each generation),
• no scaling window, and
• an elitist selection strategy (the organism with the best performance survives intact
into the next generation).
The crossover points, the mutation rates and the selection of organisms are produced
using the subtractive method algorithm described by Knuth (1981). The decompo-
sition method was used as the evaluative method for calculating throughput. From
this work it appears that simulated annealing approach is slower than the genetic
algorithm in terms of computer time, but as the simulated annealing results in a
higher throughput using the same evaluative method it could be argued that simu-
lated annealing is the more accurate method. Details are shown in Figures 5.9 and
5.10, respectively.
Tabu search algorithm
According to Glover and Laguna (1998): “The word tabu or taboo comes from
Tongan, a language of Polynesia, where it was used by the aborigines of Tonga
island to indicate things that cannot be touched because they are sacred. According to
SA, GA for 20–400 stations; #buffers = 3 * stations

1.4e+006
1.2e+006 S(SA, Deco-1)

S(GA, Deco-1)
1e+006
800000
600000
400000
200000
0
0 50 100 150 200 250 300 350 400
Station number
Fig. 5.9. Performance of simulated annealing S(SA, Deco) compared with genetic algorithms
S(GA, Deco) for large production lines
Throughput: 20–400 stations; #buffers = 3 * stations

0.69
S(SA, Deco-1)
0.685 S(GA, Deco-1)
Line throughput
0.68
0.675
0.67
0.665
0.66
0 50 100 150 200 250 300 350 400
Number of stations
Fig. 5.10. Accuracy of simulated annealing S(SA, Deco) compared with genetic algorithms
S(GA, Deco) for large production lines
Webster’s dictionary, the word now means a prohibition imposed by social custom
as a protective measure or of something banned as constituting a risk. These current
more pragmatic senses of the word accord well with the theme of tabu search. The
risk to be avoided in this case is that of following a counter-productive course, includ-
ing one which may lead to entrapment without hope of escape. On the other hand, as
in a broader social context where protective prohibitions are capable of being super-
seded when the occasion demands, the tabus of tabu search are to be overruled when
evidence of a preferred alternative becomes compelling.”
Tabu search was first introduced by Glover (1986) and a few of its basic ideas
were also given by Hansen (1986).
Let x be a certain initial solution from a set of solutions, Λ, and G(x) be a
neighborhood of this solution. The solution is feasible if it satisfies a certain set
of constraints. Let also x∗ denote the best solution found so far, i be an iteration
counter and Λ∗ be a subset of solutions in the neighborhood G(x). f (x) denotes a
function which is sought to be optimized, e.g., to be minimized. Tabu search is an
iterative method more sophisticated than the ordinary descent method in at least two
dimensions, as follows:
(i) It makes systematic use of memory in order to avoid re-visiting the same
solutions considered previously.
(ii) It uses an elaborate exploration process in order to escape from a local minimum
by using intensification and diversification. Intensification’s role is to ensure that
the next solution in the search process is close enough to the current solution
when both of them have common features. This can be achieved by adding an
extra term in the objective function and penalizing solutions which are far from
the current solution. On the other hand, the role of diversification is exactly the
opposite one, viz., to guarantee that the next solution is far from the current
one when it is discovered that this solution does not have the desired features.
Mathematically, diversification is carried out by inserting another term in the
objective function penalizing solutions that are close to the current solution. By
performing intensification and diversification, the initial objective function, f , is
modified to the following modified function, fm , which may be written as:
fm (x) = f (x) + intensification + diversification.
When a solution visited at a stage is not acceptable, then it has to be removed from
the neighborhood G(x) as it is considered a tabu solution which should be avoided
in the future iterations of the search process. That way a tabu list, T , = 1, . . . ,t, is
created which is a collection of tabu conditions, t (x, h) which are some constituents
t (x, h) that are given a tabu status to indicate that these constituents are currently
not allowed to be involved in the next solution. t are functions of solution x or
h, where h is a move applied to the solution x to direct it to a new solution, say,

y = x h (symbol denotes the application of move h to x to obtain y). This
move h is said to be a tabu move if all conditions are satisfied. (Usually, reversible
moves are used. A move h is called a reversible move when there exists a move

h−1 such that: (x h) h−1 = x.) However, sometimes, tabu move h may appear
attractive because it gives a solution better than the best solution found so far in the
search process. In that case, one would like to accept h in spite of its status. This may
be done if it has an aspiration level, α (x, h), which is better than A (x, h), a given
threshold value. If at least one of the following conditions:
α (x, h) ∈ A (x, h), = 1, . . . , α ,
is satisfied by the tabu move h applied to solution x, then move h will be accepted.
Tabu search uses a tabu list with variable size to prevent cycling of the solutions.
Following the lines of Hertz, Taillard and de Werra (1995), the steps of the Tabu
search, TS, are summarized below:
Step 1: Choose an initial solution x in Λ. Set x∗ = x and i = 0.
Step 2: Set i = i + 1 and generate a subset Λ∗ of solution in G(x, i) such that either
one of the tabu conditions t (x, h) ∈ T is violated ( = 1, . . . ,t) or at least
one of the aspiration conditions α (x, h) ∈ A (x, h), = 1, . . . , α , holds.

Step 3: Choose a best y = x h in Λ∗ with respect to f (x) or to the modified
function fm (x) and set x = y.
Step 4: If f (x) < f (x∗ ) then set x = x∗ .
Step 5: Update tabu and aspiration conditions.
Step 6: If a stopping condition is met then stop. Else go to Step 2.
Many applications of tabu search are given in the book by Glover, Taillard, Laguna
and de Werra (1992).
Summary
The following summary may be of value to the reader who wishes to use the soft-
ware available at the website associated with this book in solving buffer allocation
problems.
Buffer Allocation Problem (BAP)
1. BA
• For short unreliable production lines with Erlang-k (k ≥ 1) service and repair
times and exponential times to failure and intermediate buffers.
2. DECO-1 and SA/GA
• For large reliable exponential production lines with single machine stations
and finite intermediate buffers.
3. DECO-2 and SA/GA
• For large reliable exponential production lines with multiple parallel identi-
cal machine stations and finite intermediate buffers.

Conway et al. (1988), considering problem BAP-A and based on simulation, conjec-
tured, among other results, that the loss of throughput due to interference between
stations in production lines occurs in the first few stations and that the variability of
the service times has a major impact on this phenomenon and there is a decreasing
marginal advantage with placing buffers between work-stations. They demonstrated
the existence of the “inverted bowl” phenomenon and suggested that buffers in
unbalanced lines should be placed toward the bottleneck station.
Powell (1994) dealt with problem BAP-A for three-station production lines with
the service time probability distributions differing in either or both mean and vari-
ance. He developed rules of thumb for the optimal buffer allocation and showed
that a balanced or nearly balanced allocation is optimal for many highly unbalanced
lines. The author also demonstrated that an imbalance in means has a greater impact
on buffer allocation than does an imbalance in variances. He used simulation as an
evaluative technique.
A knowledge-based approach to problem BAP-A was given in Vouros and
Papadopoulos (1998).
Ho et al. (1979) dealt with a gradient method for solving the BAP-A problem.
Chow (1987) proposed a dynamic programming algorithm to solve the buffer
allocation problem BAP-A.
Jensen et al. (1991) also dealt with the buffer optimization problem in serial and
diverging-branch (non-linear) configurations of production systems. They aplied a
classical dynamic programming algorithm for solving the problem by taking into
account production system costs.
Yamashita and Altiok (1998) solved the BAP-B problem by applying a dynamic
programming algorithm associated with Altiok’s (1989) decomposition method for
analyzing the production line.
Kubat and Sumita (1985) and Jafari and Shanthikumar (1989 ) also used dynamic
programming approaches for solving the BAP in automatic transfer lines.
Papadopoulos and Karagiannis (2001) and Spinellis and Papadopoulos (2000b)

developed a genetic algorithm for solving the buffer allocation problem in unreli-
able and reliable production lines, respectively, with exponentially distributed service
completion times.
Colledani, Matta, Grasso and Tolio (2005) developed a new analytical method
for the buffer space allocation problem in production lines.
Levantesi, Matta and Tolio (2001) presented a new search algorithm which in
conjunction with a decomposition method for the performance evaluation of the
production lines solved the buffer allocation problem very fast.
Singh and MacGregor Smith (1997) dealt with the buffer allocation problem in
production lines with multiple parallel machines at each work-station.
Additional works on tabu search are given in Glover (1989), de Werra and Hertz
(1989),Glover (1990). Theoretical aspects of tabu search are presented in Faigle and
Kern (1992), Glover (1992) and Fox (1993).
References
1. Altiok, T. (1989), Approximate analysis of queues in series with phase-type service times
and blocking, Operations Research, Vol. 37, pp. 601–610.
2. Altiok, T. and Stidham, S. Jr. (1983), The allocation of interstage buffer capacities in
production lines, IIE Transactions, Vol. 15, No. 4, pp. 292–299.
3. Anderson, D.R. and Moodie, C.L. (1989), Optimal buffer storage capacity in production
line systems, International Journal of Production Research, Vol. 7, No. 3, pp. 233–240.
4. Bulgak, A.A., Diwan, P.D., and Inozu, B. (1995), Buffer size optimization in asyn-
chronous systems using genetic algorithms, Computers and Industrial Engineering, Vol.
28, No. 2, pp. 309–322.
5. Cheng, D.W. (1994), On the design of a tandem queue with blocking: Modeling analysis,
and gradient estimation, Naval Research Logistics, Vol. 41, pp. 759–770.
6. Chow, W.-M. (1987), Buffer capacity analysis for sequential production lines with
variable process times, International Journal of Production Research, Vol. 25, No. 8,
pp. 1183–1196.
7. Colledani, Matta, A., Grasso, M., and Tolio, T. (2005), A new analytical method for buffer
space allocation in production lines, CIRP–Journal of Manufacturing Systems, Vol. 34,
No. 4.
8. Conway, R., Maxwell, W., McClain, J.O., and Thomas, L.J. (1988), The role of work-
in-process inventory in serial production lines, Operations Research, Vol. 36, No. 2,
pp. 229–241.
9. Dallery, Yves and Frein, Yannick (1993), On decomposition methods for tandem
queueing networks with blocking, Operations Research, Vol. 41, No. 2, pp. 386–399.
10. de Werra, D. and Hertz, A. (1989), Tabu Search techniques: A tutorial and an application
to neural networks, OR Spektrum, pp. 131–141.
11. Enginarlar, E., Li, J., Meerkov, S., and Zhang, Q. (2002), Buffer capacity for accommo-
dating machine downtime in serial production lines, International Journal of Production
12. Erel, E. (1993), Effect of discrete batch WIP transfer on the efficiency of production lines,
References 157
13. Faigle, U. and Kern, W. (1992), Some convergence results for probabilistic Tabu search,
ORSA Journal on Computing, Vol. 4, pp. 32–37.
14. Fox, B.L. (1993), Integrating and accelerating tabu search, simulated annealing and
genetic algorithms, Annals of Operations Research, Vol. 41, pp. 46–67.
15. Glover, F. (1986), Future paths for integer programming and links to artificial intelligence,
Computers and Operations Research, Vol. 13, pp. 533–549.
16. Glover, F. (1989), Tabu search, Part I, ORSA Journal on Computing, Vol. 1, pp. 190–206.
17. Glover, F. (1990), Tabu search, Part II, ORSA Journal on Computing, Vol. 2, pp. 4–32.
18. Glover, F. (1992), Private communication.
19. Glover, F. and Laguna, M. (1998), Tabu search, Kluwer Academic Publishers.
20. Glover, F., Taillard, E., Laguna, M., and de Werra, D. (1993), Tabu search, Annals of
21. Grefenstette, J.J. (1986), Optimization of control parameters for genetic algorithms, IEEE
Transactions on Systems, Man, and Cybernetics, Vol. 16, No. 1, pp. 122–128.
22. Gershwin, S.B. and Schor, J.E. (2000), Efficient algorithms for buffer space allocation,
Annals of Operations Research, Vol. 93, pp. 117–144.
23. Grefenstette, J. (1986), Optimization of control parameters for genetic algorithms, IEEE
Transactions on Systems, Man and Cybernetics, Vol. 16, Issue 1, pp. 122–128.
24. Hansen, P. (1986), The steepest ascent mildest descent heuristic for combinatorial
programming, Presented at the Congress on Numerical Methods in Combinatorial
Optimization, Capri, Italy.
25. Harris, J.H. and Powell, S.G. (1999), An algorithm for optimal buffer placement in
reliable serial lines, IIE Transactions, Vol. 31, pp. 287–302.
26. Hatcher, J.M. (1969), The effect of internal storage on the production rate of a series of
stages having exponential service times, AIIE Transactions, Vol. 1, No. 2, pp. 150–156.
No. 1, pp. 69–89.
28. Hertz, A., Taillard, E., and de Werra, D. (1995), A tutorial on Tabu search, In Proc. of
Giornate di Lavoro AIRO ’95, (Enterprise Systems: Management of Technological and
Organizational Changes), pp. 13–24.
29. Hillier, F.S. and So, K.C. (1991a), The effect of machine breakdowns and interstage stor-
age on the performance of production line systems, International Journal of Production
30. Hillier, F.S. and So, K.C. (1991b), The effect of the coefficient of variation of operation
times on the allocation of storage space in production line systems, IIE Transactions,
Vol. 23, No. 2, pp. 198–206.
32. Hillier, F.S., So, K.C., and Boling, R.W. (1993), Notes: Towards characterizing the opti-
mal allocation of storage space in production line systems with variable processing times,
Management Science, Vol. 39, No. 1, pp. 126–133.
33. Ho, Y.C., Eyler, M.A., and Chien, T.T. (1979), A gradient technique for general buffer
storage design in a production line, International Journal of Production Research,
Vol. 17, No. 6, pp. 557–580.
34. Jafari, M.A. and Shanthikumar, J.G. (1989), Determination of optimal Buffer storage
capacities and optimal allocation in multistage automatic transfer lines, IIE Transactions,
Vol. 21, No. 2, pp. 130–135.
35. Jensen, P.A., Pakath, R., and Wilson, J.R. (1991), Optimal buffer inventories for multi-
stage production systems with failures, European Journal of Operational Research, Vol.
51, pp. 313–326.
36. Knuth, D.E. (1981), The Art of Computer Programming, Vol. 2/Seminumerical Algo-
rithms, Second Edition, pp. 171–173, Addison-Wesley.
37. Koulamas, C.P. (1989), Optimal buffer space in two-stage machining systems with
Markovian or non-Markovian tool life processes, International Journal of Production
38. Kraemer, S.A. and Love, R.F. (1970), A model for optimizing the buffer inventory storage
size in a sequential production system, AIIE Transactions, Vol. 2, No. 1, pp. 64–69.
39. Kubat, P. and Sumita, U. (1985), Buffers and backup machines in automatic transfer lines,
International Journal of Production Research, Vol. 23, pp. 1259–1270.
40. Levantesi, R., Matta, A., and Tolio, T. (2001), A new algorithm for buffer alloca-
tion in production lines, Proceedings of the Third Aegean International Conference on
Design and Analysis of Manufacturing Systems, May 19–22, 2001, Tinos Island, Greece,
pp. 279–288.
41. Liu, C. and Tu, F.-S. (1994), Buffer allocation via genetic algorithm, Proceedings of
the 33rd Conference on Decision and Control, Lake Buena Vista, FL, December, 1994,
pp. 609–610.
42. Masso, J. and Smith, M.L. (1974), Interstage storages for three stage lines subjet to
stochastic failures, AIIE Transactions, Vol. 6, No. 4, pp. 354–358.
43. Martin, G.E. (1994), Optimal design of production lines, International Journal of
Production Research, Vol. 32, No. 5, pp. 989–1000.
44. Meester, L.E. and Shanthikumar, J.G. (1990), Concavity of the throughput of tandem
queueing systems with finite buffer storage space, Advances in Applied Probability,
Vol. 22, pp. 764–767.
45. Okamura, K. and Yamashita, H. (1977), Analysis of the effect of buffer storage capacity
in transfer line systems, AIIE Transactions, Vol. 9, No. 2, pp. 127–135.
46. Papadopoulos, H.T. and Karagiannis, T.I. (2001), A genetic algorithm approach for
the buffer allocation problem in unreliable production lines, International Journal of
Operations and Quantitative Management, Vol. 7, No. 1, pp. 23–35.
47. Papadopoulos, H.T. and Vidalis, M.I. (1998), Optimal buffer storage allocation in bal-
anced reliable production lines, International Transactions in Operational Research,
Vol. 5, No. 4, pp. 325–339.
48. Papadopoulos, H.T. and Vidalis, M.I. (1999), Optimal buffer allocation in short μ -
balanced unreliable production lines, Computers & Industrial Engineering, Vol. 37,
pp. 691–710.
49. Papadopoulos, H.T. and Vidalis, M.I. (2001a), A heuristic algorithm for the buffer allo-
cation in unreliable unbalanced production lines, Computers & Industrial Engineering,
Vol. 41, pp. 261–277.
50. Papadopoulos, H.T. and Vidalis, M.I. (2001b), Minimizing WIP inventory in reliable
production lines, International Journal of Production Economics, Vol. 70, pp. 185–197.
51. Park, T. (1993), A two-phase heuristic algorithm for determining buffer sizes of produc-
tion lines, International Journal of Production Research, Vol. 31, No. 3, pp. 613–631.
52. Powell, S.G. (1994), Buffer allocation in unbalanced three-station serial lines, Interna-
tional Journal of Production Research, Vol. 32, No. 9, pp. 2201–2217.
53. Powell, S.G. and Pyke, D.F. (1994), Optimal allocation of buffers in serial produc-
tion lines with a single bottleneck, The Amos Tuck School of Business Administration,
Dartmouth College, Working Paper No. 301.
References 159
54. Sevast’yanov B.A. (1962), Influence of storage bin capacity on the average standstill time
of a production line, Theory of Probability and its Applications, Vol. 7, pp. 429–438.
55. Singh, A. and Smith, MacGregor, J. (1997), Buffer allocation for an integer nonlin-
ear network design problem, Computers and Operations Research, Vol. 24, No. 5,
pp. 453–472.
56. Smith, MacGregor, J., and Chikhale, N. (1995), Buffer allocation for a class of nonlinear
stochastic knapsack problems, Annals of Operations Research, Vol. 58, pp. 323–360.
57. Smith, MacGregor, J., and Daskalaki, S. (1988), Buffer space allocation in automated
assembly lines, Operations Research, Vol. 36, No. 2, pp. 343–358.
58. So, K.C. (1990), The impact of buffering strategies on the perfomance of production line
systems, International Journal of Production Research, Vol. 28, No. 2, pp. 2293–2307.
59. So, K.C. (1997), Optimal buffer allocation strategy for minimizing work-in-process
inventory in unpaced production lines, IIE Transactions, Vol. 29, No. 1, pp. 81–88.
60. Spinellis, D.D. and Papadopoulos, C.T. (2000a), A simulated annealing approach for
buffer allocation in reliable production lines, Annals of Operations Research, Vol. 93,
pp. 373–384.
61. Spinellis, D.D. and Papadopoulos, C.T. (2000b), Stochastic algorithms for buffer allo-
cation in reliable production lines, Mathematical Problems in Engineering, Vol. 5,
pp. 441–458.
62. Vouros, G.A. and Papadopoulos, H.T. (1998), Buffer allocation in unreliable produc-
tion lines using a knowledge based system, Computers & Operations Research, Vol. 25,
No. 12, pp. 1055–1067.
63. Yamashita, H. and Altiok, T. (1998), Buffer capacity allocation for a desired throughput
in production lines, IIE Transactions, Vol. 30, pp. 883–891.
6
Double and Triple Optimization
There are three pure allocation problems, viz., the work-load allocation problem,
the server allocation problem and the buffer allocation problem, all concerned
with maximizing throughput. Mathematically, these problems may be described as
follows:
The work-load allocation problem, WAP:
max X (w) = max X (w1 , w2 , . . . , wK )
subject to:
K
∑ wi = 1, for wi > 0
i=1
for normalized total work-load equal to unity and fixed allocation of servers and fixed
buffer allocation.
The server allocation problem, SAP:
max X(s) = max X (S1 , S2 , . . . , SK ))

subject to:
K
∑ Si = S, for Si ≥ 1 and integer
i=1
for fixed allocation of work to each station and fixed buffer allocation.
The buffer allocation problem, BAP:
max X (n) = X (N2 , . . . , NK )
subject to:
K
∑ Ni = N, for Ni ≥ 0 and integer
i=2
for fixed allocation of work to each station and fixed allocation of servers.
162 6 Double and Triple Optimization
As indicated above, there are three single-variable decision problems. Combining

these problems into two-variable problems leads to the following three problems
which may be mathematically described as follows:
The combined work-load allocation and server allocation problems, W + S:
max X (w, s)
subject to:
K
∑ wi = 1, for wi > 0 and normalized work-load
i=1
and
K
i=1
and for fixed buffer allocation.

The reader may note that this problem has already been discussed in Chapter 4.
The combined work-load allocation and buffer allocation) problems, W + B:
max X (w, n)
subject to:
K
i=1
and
K
i=2
and for fixed server allocation.

The combined server allocation and buffer allocation problems, S + B:
max X (s, n)
subject to:
K
i=1
and
K
i=2
and for fixed work-load allocation.

If all three decision variables are considered together, the following combined
problem, W + S + B, may be described mathematically as:
max X (w, s, n)
6.1 Simultaneous Allocation of Work and Buffers, W + B 163
subject to:
K
i=1
and
K
i=1
and
K
∑ Ni = N, for Ni ≥ 0 and integer.
i=2
In practice, it is likely that the design of production line systems would involve
decisions on at least two if not on all three of the decision variables, i.e., work-load,
servers and buffers.
A further consideration in the above problems would be the service time distri-
bution, given the number of servers. Usually, the service distribution variability is
captured through the concepts of the mean of the service time distribution and the
coefficient of variation of the distribution itself with the usual assumption of phase-
type distributions. Even given the restriction of phase-type distribution (which is used
for ease of computation), it must be realized that in a set of K work-stations there
could be a difference in the coefficient of variation (cv) of the work-stations.
In Section 6.1, the W + B problem is treated and some design guide rules are
given. Likewise, in Section 6.2 the S + B problem is discussed and some design
guidelines are presented. Finally, in Section 6.3 the W + S + B problem is discussed
and a range of results are presented.
6.1 Simultaneous Allocation of Work and Buffers, W + B
In the W + B problem the number of servers at each station is fixed.

Buzacott and Shanthikumar (1993) showed analytically that if the buffer slots
are an integer multiple of the K − 1 intermediate buffers, a balanced buffer allocation
is associated with maximum throughput. Hillier and So (1995) obtained empirical
results for the optimal buffer and work allocations for maximum throughput for
K = 3, 4 and 5 stations and N = 0, 1, . . . , 8 total buffer slots. The reader will note that
in this paper, Hillier and So considered saturated lines, i.e., the first station is never
starved and results are only given for lines with perfectly reliable machines. Buza-
cott and Shanthikumar’s results were confirmed and generally a bowl phenomenon
of work allocation was associated with maximum throughput although in two cases
a type of double bowl phenomenon of work allocation was indicated (in the cases of
K = 5, N = 2 and K = 5, N = 6). Symmetrical bowl allocations of work-load are asso-
ciated with strictly uniform allocation of buffers. With respect to buffer allocation the
optimal pattern is as follows:
Step 1: Allocate the same maximum number of buffer slots to each buffer of the
line.
Step 2: Allocate the extra buffer slots over the above uniform buffer allocation to the
interior buffers, particularly to the central buffers rather than the end buffers.
Hillier and So (1995) results for the W + B problem do not extend to large buffer
slots and the published results assumed that there is only one server at each station
although the authors stated that their conclusions apply for multi-server work-
stations. Table 7 of Hillier and So (1995) gives a timely warning to researchers and
designers alike by showing the relatively small increase in throughput achieved by
unbalancing the lines and using the optimal allocation of buffers slots as derived
from the W + B optimization.
Over the range of experiments undertaken by Hillier and So (1995) it may be
concluded that the work allocation and the corresponding buffer allocation solutions
for the W + B optimization problem produce the following general rules:
• Larger work-loads are associated with stations with larger buffer capacities.
• Higher work-loads are assigned to stations at both ends of the line.
It is clear from this work that as uniform as possible buffer allocation will gener-
ally lead to better throughput independently of the actual work-load allocation in
production lines, where only the server allocation is faced.
Suggested solution procedures for the W + B problem using the algorithms
available at the website associated with this text will be discussed at the end of
Section 6.3.
6.2 Simultaneous Allocation of Servers and Buffers, S + B

In the S + B problem the work-load allocation is given.
Hillier and So (1995) studied the S + B problem under uniform work-load allo-
cation for K = 3, 4 and 5 stations and total number of servers, S up to twice the
number of stations and buffer spaces, N = 1, . . . , 4. Their results indicate that the
optimal server allocation is given by the as uniform as possible allocation and the
extra servers assigned to the interior stations with the two end stations having the
lowest priority in receiving the extra servers with the last station receiving the extra
server before the first station.
More specifically, in the cases published by Hillier and So (1995) the following
pattern of the optimal server allocation has been observed:
1. If S = aK, a > 1 and integer, a uniform server allocation is optimal.
2. If S = aK − 1, a > 1 and integer, a server allocation in which the first station
has a − 1 servers and all the others have a servers is optimal and as would be
expected the mirror image of this server allocation where the deficit in the number
of servers is at the last station results in a near optimal solution.
3. If S = aK − 2, a > 1 and integer, the optimal server allocation is to allocate a − 1
servers to both end stations and a servers to the other K − 2 stations.
6.3 Simultaneous Allocation of Work, Servers and Buffers, W + S + B 165
Beyond the above three rather definite patterns it is difficult to develop a server allo-
cation rule from the Hillier and So’s published results for other levels of total number
of servers.
It appears from Hillier and So’s results for the S + B problem that the optimal
buffer allocation does not follow any particular pattern. It is well known that an extra
server at a station in effect provides an extra buffer slot to that station. The general
guideline therefore is to provide extra buffer slots to stations which have received
fewer servers as a result of the optimal allocation of servers as outlined above. For
example, in the cases where S = 2K − 2 and K = 3, 4 and 5, the buffer slots are
assigned to the two end stations which are deficient by one server in comparison to all
the other stations using the optimal server allocation procedure. It would be fair to say
that arising from Hillier and So’s results, a designer would be well advised to allocate
the servers first according to the heuristic indicated above and then to allocate the
buffer slots to near stations that have received fewer servers. Clearly, this guideline
must be understood in the context of the relatively small number of stations, servers
and buffer slots involved in Hillier and So’s pioneering studies (1995).
Suggested solution procedures for the S + B problem using the algorithms
available at the website associated with this text will be discussed at the end of
Section 6.3.
6.3 Simultaneous Allocation of Work, Servers and Buffers,

W+S+B
In this problem the number of stations, K, the total number of servers, S, and the
total number of buffer slots, N, are given. Because of the computational complexity
involved, Hillier and So (1995) studied systems for K = 3, 4, 5 stations, S = 4, . . . , 8
servers and total buffer slots, N up to 4 maximum.
The most persistent result is the existence of the L-phenomenon for the server
allocation, i.e., to allocate all extra servers to the end stations. For the buffer alloca-
tion the tendency is to allocate the buffer slots as uniformly as possible with the extra
buffers being assigned toward the last station of the line. Thus, it appears that results
from the W + S problem and the W + B problem somewhat dominate the results from
the S + B problem. The optimal work-load allocation in all cases follows a bowl pat-
tern with a significantly increased work-load assigned to the last station. Indeed, the
optimal work-load per server has a much higher value for the last station than do any
of the other stations and the first station has the next highest work-load per server. In
the cases studied by Hillier and So (1995), the work-load allocation to the stations
other than the two end stations tended to be almost uniform resulting in what might
be described as a three-level bowl of work-load allocation.
Hillier and So (1995) proposed the following heuristic allocation procedure:
Step 1: Determine the server allocation by following the L-phenomenon procedure
already discussed in Chapter 4, taking into account any adjustments required
for upper and lower bounds on the number of servers per station.
Step 2: Determine the buffer allocation as follows:

Step 2.1: If the server allocation is balanced, select as uniform as possible
buffer allocation with buffers closest to the center having a higher
priority to the buffers away from the center, in accordance with
the results obtained from the W + S problem. Any tie should be
resolved in favor of buffer slots toward the end of the line.
Step 2.2: If the server allocation is unbalanced, select as uniform as possi-
ble buffer allocation with extra slots assigned to buffers that are
closest to stations with the most servers.
Step 3: Determine the work-load allocation once the server and buffer allocations
are assigned.
In Spinellis, Papadopoulos and MacGregor Smith (2000), simulated annealing was
used in conjunction with the expansion algorithm to obtain results for both short and
long lines with perfectly reliable stations consisting of parallel machines at each sta-
tion. It should be noted that the application here is with respect to quasi-saturated
lines in contrast to that treated by Hillier and So (1995) and described above where
saturated lines were considered. In quasi-saturated lines the number of buffer slots
in front of the first station in the Spinellis et al. (2000) application may not have
ensured that the first station is always occupied which was the case in the Hillier
and So (1995) application. The analytical approach also differs in that Spinellis et al.
(2000) used the novel search method of simulated annealing and the approximate
expansion algorithm for both short and long lines with single-machine and multiple-
machine stations, whereas Hillier and So (1995) used complete enumeration and
an exact Markovian method for single-machine and multiple-machine station short
lines. Additionally, in the Spinellis et al. (2000) model, and within the context of the
expansion method, the contents of the buffer in front of the first station are taken into
consideration when allocating the total number of buffer slots, which is not the case
in the model of Hillier and So (1995). Despite these differences, it is illuminating to
compare the results of Spinellis et al. (2000) with those of Hillier and So (1995) in
relation to short lines. Of course, as the topologies of the systems are slightly differ-
ent one could expect different results. However, the main point must surely be the
fact that Hillier and So (1995) used exact methods (Markovian and complete enu-
meration), whereas Spinellis et al. (2000) used the approximate expansion algorithm
and simulated annealing.
Comparing the results obtained for short lines in both papers, the following
observations may be made:
1. The work-load allocation in Spinellis et al. (2000) does not follow the bowl phe-
nomenon as in Hillier and So (1995), but in general continues to diminish toward
the end of the line.
2. The buffer allocation in Spinellis et al. (2000) does not follow the inverse bowl
phenomenon as in Hillier and So (1995), but increases monotonically across the
line toward the end stations. As the number of total buffer slots increases in
Spinellis et al. (2000), a certain fixed number of slots are allocated in front of
the first station and the remaining slots are allocated almost uniformly among
the interstation buffers with some preference given to the downstream buffer
locations.
3. The server allocation in Spinellis et al. (2000) for a small number of servers
follows a pattern similar to the one presented in Hillier and So (1995). How-
ever, as the number of servers increases in Spinellis et al. (2000), servers tend to
accumulate toward the beginning of the line.
4. The server and work-load allocation in Spinellis et al. (2000) do not exhibit the
L-phenomenon shown in Hillier and So (1995).
5. The buffer and work-load allocation results are different in the two papers. As
far as the buffer allocation is concerned, in Spinellis et al. (2000) buffers tend
to accumulate toward the end of the line, whereas in Hillier and So (1995), the
buffer allocations are more uniform. The allocation of work, however, does not
exhibit the symmetrical bowl phenomenon presented in Hillier and So (1995) and
follows the usual descending rate across the line.
6. The buffer and server allocation results of Spinellis et al. (2000) are roughly sim-
ilar to those presented in Hillier and So (1995) in both the server and the buffer
allocation vectors. In both models servers are rather uniformly allocated, but in
Spinellis et al. (2000), servers tend to accumulate toward the beginning and mid-
dle of the line, whereas in Hillier and So (1995) they are allocated toward the
middle and end of the line. In both models, buffers tend to accumulate toward the
line ends.
7. Finally, the buffer, server, and work-load allocation roughly follows the shape
of work-load allocation presented in Hillier and So (1995), but the allocation of
buffers and servers is quite dissimilar.
In the Spinellis et al. (2000) paper, the expansion method was always used in devel-
oping the results presented. Recently, the authors decided to investigate saturated
lines using the expansion method (EXPA), the decomposition method (DECO) and
the Markovian method (MARK) as evaluative tools. The expansion method and
the Markovian method can accommodate parallel-machine stations, whereas the
decomposition method is usually used in single-machine station lines. The following
experiments, given in Table 6.1, were run using the above three evaluative methods
in conjunction with complete enumeration (CE) and simulated annealing (SA) as the
search methods.
Below, in Tables 6.2–6.3, are the numerical results of these experiments in all of
which the buffer in front of the first station was assigned 5 slots, i.e., B1 = 5. For long
lines it is unlikely that B1 = 5 will lead to saturated lines. The Markovian model used
by the authors is based on the algorithm given in Heavey, Papadopoulos and Browne
(1993), the decomposition algorithm used is based on one of the algorithms given in
Dallery and Frein (1993) and the expansion algorithm is based on the method given
in Kerbache and MacGregor Smith (1987) and Jain and MacGregor Smith (1994).
All of these methods are described in Chapter 2.
Table 6.1. Overall plan of experiments
Evaluative Search No. of Stations Total # of Buffer Slots

Method Method K N
MARK CE 4,6,8 1–26
DECO CE 4,6,8 1–26
EXPA CE 4,6,8 1–26
DECO SA 5,7,9 9–30
EXPA SA 5,7,9 9–30
DECO CE 5,7,9 9–30
EXPA CE 5,7,9 9–30
DECO SA 10 11–21
EXPA SA 10 11–21
DECO CE 10 11–18
EXPA CE 10 11–18
DECO SA 16 16–45
EXPA SA 16 16–45
DECO SA 5–9, 11(10)61 8–16, 20–120
EXPA SA 5–9, 11(10)61 8–16, 20–120
DECO CE 5–6 8–10
EXPA CE 5–6 8–10
Table 6.2. Throughput and buffer allocation for 4-, 6- and 8-station lines via CE
(Evaluative-Search) No. of Stations No. of Buffer Slots Throughput Buffer Allocation

(Method-Method) K N XK (B1 , B2 , · · · , BK )
(MARK-CE) 8 1 0.464806 (5,0,0,0,1,0,0,0)
(MARK-CE) 8 2 0.485547 (5,0,0,1,0,1,0,0)
(MARK-CE) 8 3 0.503378 (5,0,1,0,1,0,1,0)
(MARK-CE) 8 4 0.521864 (5,0,1,1,0,1,1,0)
(MARK-CE) 8 5 0.542124 (5,0,1,1,1,1,1,0)
(MARK-CE) 8 6 0.554479 (5,1,1,1,1,1,1,0)
(MARK-CE) 8 7 0.572014 (5,1,1,1,1,1,1,1)
(MARK-CE) 6 6 0.60923 (5,1,1,2,1,1)
(MARK-CE) 6 7 0.624668 (5,1,2,1,2,1)
(MARK-CE) 4 21 0.844477 (5,7,7,7)
(MARK-CE) 4 24 0.858043 (5,8,8,8)
(MARK-CE) 4 25 0.862382 (5,8,9,8)
(MARK-CE) 4 26 0.866002 (5,8,10,8)
(DECO-CE) 8 1 0.428197 (5,0,0,0,0,0,1,0)
(DECO-CE) 8 2 0.45005 (5,0,0,0,1,0,1,0)
(DECO-CE) 8 3 0.472079 (5,0,0,1,0,1,1,0)
(DECO-CE) 8 4 0.493967 (5,0,1,0,1,1,1,0)
(DECO-CE) 8 5 0.515725 (5,0,1,1,1,1,1,0)
(DECO-CE) 8 6 0.535165 (5,0,1,1,1,1,1,1)
(DECO-CE) 8 7 0.550428 (5,1,1,1,1,1,1,1)
(continued)
Table 6.2. (Continued)

(DECO-CE) 6 6 0.587922 (5,1,1,2,1,1)
(DECO-CE) 6 7 0.606644 (5,1,1,2,2,1)
(DECO-CE) 4 21 0.842941 (5,7,8,6)
(DECO-CE) 4 24 0.857248 (5,8,9,7)
(DECO-CE) 4 25 0.861613 (5,8,9,8)
(DECO-CE) 4 26 0.865664 (5,8,10,8)
(EXPA-CE) 8 1 0.204673 (5,0,0,0,0,0,0,1)
(EXPA-CE) 8 2 0.224199 (5,0,0,0,0,0,1,1)
(EXPA-CE) 8 3 0.247024 (5,0,0,0,0,1,1,1)
(EXPA-CE) 8 4 0.273771 (5,0,0,0,1,1,1,1)
(EXPA-CE) 8 5 0.306172 (5,0,1,0,1,1,1,1)
(EXPA-CE) 8 6 0.343134 (5,0,1,1,1,1,1,1)
(EXPA-CE) 8 7 0.379798 (5,1,1,1,1,1,1,1)
(EXPA-CE) 6 6 0.456404 (5,1,1,1,1,2)
(EXPA-CE) 6 7 0.480015 (5,1,1,1,2,2)
(EXPA-CE) 4 21 0.830446 (5,7,7,7)
(EXPA-CE) 4 24 0.846169 (5,8,8,8)
(EXPA-CE) 4 25 0.850921 (5,8,8,9)
(EXPA-CE) 4 26 0.855233 (5,8,9,9)
Table 6.3. Throughput and buffer allocation for 5- and 6-station lines via CE

(DECO-CE) 5 8 0.665763 (5,2,2,2,2)
(DECO-CE) 6 10 0.653177 (5,2,2,2,2,2)
(EXPA-CE) 5 8 0.590449 (5,2,2,2,2)
(EXPA-CE) 6 10 0.555666 (5,2,2,2,2,2)
Table 6.4. Throughput and buffer allocation for 5-, 7- and 9-station lines via SA

(DECO-SA) 9 9 0.556706 (5,1,1,1,1,1,2,1,1)
(DECO-SA) 9 10 0.56886 (5,1,1,1,2,1,2,1,1)
(DECO-SA) 9 11 0.581225 (5,1,1,2,1,2,1,2,1)
(DECO-SA) 9 12 0.593139 (5,1,1,2,1,2,2,2,1)
(DECO-SA) 9 13 0.604882 (5,1,1,2,2,2,2,2,1)
(DECO-SA) 9 14 0.617087 (5,1,2,2,2,2,2,2,1)
(DECO-SA) 9 15 0.626453 (5,1,2,2,2,2,2,2,2)
(DECO-SA) 7 12 0.644706 (5,2,2,2,2,2,2)
(DECO-SA) 7 13 0.655936 (5,2,2,2,3,2,2)
(continued)

(DECO-SA) 5 25 0.819322 (5,6,7,7,5)
(DECO-SA) 5 28 0.833229 (5,6,8,8,6)
(DECO-SA) 5 29 0.836937 (5,7,8,8,6)
(DECO-SA) 5 30 0.840874 (5,7,8,8,7)
(EXPA-SA) 9 9 0.372453 (5,1,1,1,1,1,1,1,2)
(EXPA-SA) 9 10 0.386788 (5,1,1,1,1,1,1,2,2)
(EXPA-SA) 9 11 0.402043 (5,1,1,1,1,1,2,2,2)
(EXPA-SA) 9 12 0.418135 (5,1,1,1,1,2,2,2,2)
(EXPA-SA) 9 13 0.434875 (5,1,1,1,2,2,2,2,2)
(EXPA-SA) 9 14 0.451915 (5,1,1,2,2,2,2,2,2)
(EXPA-SA) 9 15 0.468776 (5,1,2,2,2,2,2,2,2)
(EXPA-SA) 7 12 0.527447 (5,2,2,2,2,2,2)
(EXPA-SA) 7 13 0.541095 (5,2,2,2,2,2,3)
(EXPA-SA) 5 25 0.789839 (5,6,6,6,7)
(EXPA-SA) 5 28 0.806145 (5,6,7,7,8)
(EXPA-SA) 5 29 0.811007 (5,7,7,7,8)
(EXPA-SA) 5 30 0.815795 (5,7,7,8,8)
Table 6.5. Throughput and buffer allocation for 5-, 7- and 9-station lines via CE

(DECO-CE) 9 9 0.556706 (5,1,1,1,1,1,2,1,1)
(DECO-CE) 9 10 0.56886 (5,1,1,1,2,1,2,1,1)
(DECO-CE) 9 11 0.581225 (5,1,1,2,1,2,1,2,1)
(DECO-CE) 9 12 0.593139 (5,1,1,2,1,2,2,2,1)
(DECO-CE) 9 13 0.604882 (5,1,1,2,2,2,2,2,1)
(DECO-CE) 9 14 0.617087 (5,1,2,2,2,2,2,2,1)
(DECO-CE) 9 15 0.626453 (5,1,2,2,2,2,2,2,2)
(DECO-CE) 7 12 0.644706 (5,2,2,2,2,2,2)
(DECO-CE) 7 13 0.655936 (5,2,2,2,3,2,2)
(DECO-CE) 5 25 0.819322 (5,6,7,7,5)
(DECO-CE) 5 28 0.833229 (5,6,8,8,6)
(DECO-CE) 5 29 0.836937 (5,7,8,8,6)
(DECO-CE) 5 30 0.840874 (5,7,8,8,7)
(EXPA-CE) 9 9 0.372453 (5,1,1,1,1,1,1,1,2)
(EXPA-CE) 9 10 0.386788 (5,1,1,1,1,1,1,2,2)
(EXPA-CE) 9 11 0.402043 (5,1,1,1,1,1,2,2,2)
(EXPA-CE) 9 12 0.418135 (5,1,1,1,1,2,2,2,2)
(EXPA-CE) 9 13 0.434875 (5,1,1,1,2,2,2,2,2)
(EXPA-CE) 9 14 0.451915 (5,1,1,2,2,2,2,2,2)
(EXPA-CE) 9 15 0.468776 (5,1,2,2,2,2,2,2,2)
(continued)

(EXPA-CE) 7 12 0.527447 (5,2,2,2,2,2,2)
(EXPA-CE) 7 13 0.541095 (5,2,2,2,2,2,3)
(EXPA-CE) 5 25 0.789839 (5,6,6,6,7)
(EXPA-CE) 5 28 0.806145 (5,6,7,7,8)
(EXPA-CE) 5 29 0.811007 (5,7,7,7,8)
(EXPA-CE) 5 30 0.815795 (5,7,7,8,8)
Table 6.6. Throughput and buffer allocation for 10-station lines via SA

(DECO-SA) 10 11 0.561615 (5,1,1,1,1,1,2,1,2,1)
(DECO-SA) 10 12 0.572518 (5,1,1,2,1,2,1,2,1,1)
(DECO-SA) 10 13 0.583461 (5,1,1,2,1,2,1,2,2,1)
(DECO-SA) 10 14 0.593987 (5,1,1,2,1,2,2,2,2,1)
(DECO-SA) 10 15 0.604453 (5,1,1,2,2,2,2,2,2,1)
(DECO-SA) 10 16 0.615375 (5,1,2,2,2,2,2,2,2,1)
(DECO-SA) 10 17 0.623801 (5,1,2,2,2,2,2,2,2,2)
(DECO-SA) 10 18 0.630954 (5,1,2,2,2,2,2,3,2,2)
(DECO-SA) 10 19 0.637935 (5,1,2,2,2,3,2,3,2,2)
(DECO-SA) 10 20 0.645319 (5,2,2,2,3,2,2,3,2,2)
(DECO-SA) 10 21 0.65282 (5,2,2,3,2,3,2,3,2,2)
(EXPA-SA) 10 11 0.365613 (5,1,1,1,1,1,1,1,2,2)
(EXPA-SA) 10 12 0.378946 (5,1,1,1,1,1,1,2,2,2)
(EXPA-SA) 10 13 0.393029 (5,1,1,1,1,1,2,2,2,2)
(EXPA-SA) 10 14 0.407759 (5,1,1,1,1,2,2,2,2,2)
(EXPA-SA) 10 15 0.422938 (5,1,1,1,2,2,2,2,2,2)
(EXPA-SA) 10 16 0.438231 (5,1,1,2,2,2,2,2,2,2)
(EXPA-SA) 10 17 0.453198 (5,1,2,2,2,2,2,2,2,2)
(EXPA-SA) 10 18 0.4664 (5,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 10 19 0.475801 (5,2,2,2,2,2,2,2,2,3)
(EXPA-SA) 10 20 0.485532 (5,2,2,2,2,2,2,2,3,3)
(EXPA-SA) 10 21 0.495542 (5,2,2,2,2,2,2,3,3,3)
Table 6.7. Throughput and buffer allocation for 10-station lines via CE

(DECO-CE) 10 11 0.561615 (5,1,1,1,1,1,2,1,2,1)
(DECO-CE) 10 12 0.572518 (5,1,1,2,1,2,1,2,1,1)
(DECO-CE) 10 13 0.583461 (5,1,1,2,1,2,1,2,2,1)
(DECO-CE) 10 14 0.593987 (5,1,1,2,1,2,2,2,2,1)
(continued)

(DECO-CE) 10 15 0.604453 (5,1,1,2,2,2,2,2,2,1)
(DECO-CE) 10 16 0.615375 (5,1,2,2,2,2,2,2,2,1)
(DECO-CE) 10 17 0.623801 (5,1,2,2,2,2,2,2,2,2)
(DECO-CE) 10 18 0.630954 (5,1,2,2,2,2,2,3,2,2)
(EXPA-CE) 10 11 0.365613 (5,1,1,1,1,1,1,1,2,2)
(EXPA-CE) 10 12 0.378946 (5,1,1,1,1,1,1,2,2,2)
(EXPA-CE) 10 13 0.393029 (5,1,1,1,1,1,2,2,2,2)
(EXPA-CE) 10 14 0.407759 (5,1,1,1,1,2,2,2,2,2)
(EXPA-CE) 10 15 0.422938 (5,1,1,1,2,2,2,2,2,2)
(EXPA-CE) 10 16 0.438231 (5,1,1,2,2,2,2,2,2,2)
(EXPA-CE) 10 17 0.453198 (5,1,2,2,2,2,2,2,2,2)
(EXPA-CE) 10 18 0.4664 (5,2,2,2,2,2,2,2,2,2)
Table 6.8. Throughput and buffer allocation for 16-station lines via SA
(Evaluative-Search) No. of No. of Buffer Throughput Buffer Allocation

(Method-Method) Stations Slots XK (B1 , B2 , · · · , BK )
K N
(DECO-SA) 16 16 0.531584 (5,1,1,1,1,1,1,1,1,1,1,1,1,1,2,1)
(DECO-SA) 16 17 0.538177 (5,1,1,1,1,1,1,1,1,1,1,1,2,1,2,1)
(DECO-SA) 16 18 0.544701 (5,1,1,1,1,1,1,1,1,1,1,2,1,2,2,1)
(DECO-SA) 16 19 0.551151 (5,1,1,1,1,1,1,1,1,1,2,1,2,2,2,1)
(DECO-SA) 16 20 0.557554 (5,1,1,1,1,1,1,1,1,2,1,2,2,2,2,1)
(DECO-SA) 16 21 0.564139 (5,1,1,2,1,2,1,1,2,1,2,1,2,1,2,1)
(DECO-SA) 16 22 0.570869 (5,1,1,2,1,2,1,2,1,2,1,2,1,2,2,1)
(DECO-SA) 16 23 0.577393 (5,1,1,2,1,2,1,2,1,2,1,2,2,2,2,1)
(DECO-SA) 16 24 0.583771 (5,1,1,2,1,2,1,2,1,2,2,2,2,2,2,1)
(DECO-SA) 16 25 0.590093 (5,1,1,2,1,2,1,2,2,2,2,2,2,2,2,1)
(DECO-SA) 16 26 0.596421 (5,1,1,2,1,2,2,2,2,2,2,2,2,2,2,1)
(DECO-SA) 16 27 0.602862 (5,1,2,1,2,2,2,2,2,2,2,2,2,2,2,1)
(DECO-SA) 16 28 0.609608 (5,1,2,2,2,2,2,2,2,2,2,2,2,2,2,1)
(DECO-SA) 16 29 0.614872 (5,1,2,2,2,2,2,2,2,2,2,2,2,2,2,2)
(DECO-SA) 16 30 0.619337 (5,1,2,2,2,2,2,2,2,2,2,2,2,3,2,2)
(DECO-SA) 16 31 0.623678 (5,1,2,2,2,2,2,2,2,2,2,3,2,3,2,2)
(DECO-SA) 16 32 0.627994 (5,1,2,2,2,2,2,2,2,2,2,3,2,3,3,2)
(DECO-SA) 16 33 0.632321 (5,1,2,2,2,2,2,2,2,2,3,2,3,3,3,2)
(DECO-SA) 16 34 0.636609 (5,1,2,2,2,2,2,2,2,3,2,3,3,3,3,2)
(DECO-SA) 16 35 0.640958 (5,2,2,3,2,3,2,2,3,2,2,3,2,3,2,2)
(DECO-SA) 16 36 0.645451 (5,2,2,3,2,3,2,3,2,3,2,3,2,3,2,2)
(DECO-SA) 16 37 0.649859 (5,2,2,3,2,3,2,3,2,3,2,3,2,3,3,2)
(DECO-SA) 16 38 0.654271 (5,2,2,3,2,3,2,3,2,3,2,3,3,3,3,2)
(continued)
(Evaluative-Search) No. of No. of Buffer Throughput Buffer Allocation

(Method-Method) Stations Slots XK (B1 , B2 , · · · , BK )
K N
(DECO-SA) 16 39 0.658581 (5,2,2,3,2,3,2,3,2,3,3,3,3,3,3,2)
(DECO-SA) 16 40 0.662864 (5,2,2,3,2,3,3,2,3,3,3,3,3,3,3,2)
(DECO-SA) 16 41 0.6672 (5,2,2,3,3,2,3,3,3,3,3,3,3,3,3,2)
(DECO-SA) 16 42 0.671569 (5,2,2,3,3,3,3,3,3,3,3,3,3,3,3,2)
(DECO-SA) 16 43 0.676039 (5,2,3,3,3,3,3,3,3,3,3,3,3,3,3,2)
(DECO-SA) 16 44 0.679259 (5,2,3,3,3,3,3,3,3,3,3,3,3,3,3,3)
(DECO-SA) 16 45 0.682455 (5,2,3,3,3,3,3,3,3,3,3,3,3,4,3,3)
(EXPA-SA) 16 16 0.277368 (5,1,1,1,1,1,1,1,1,1,1,1,1,1,1,2)
(EXPA-SA) 16 17 0.284163 (5,1,1,1,1,1,1,1,1,1,1,1,1,1,2,2)
(EXPA-SA) 16 18 0.291307 (5,1,1,1,1,1,1,1,1,1,1,1,1,2,2,2)
(EXPA-SA) 16 19 0.298814 (5,1,1,1,1,1,1,1,1,1,1,1,2,2,2,2)
(EXPA-SA) 16 20 0.30669 (5,1,1,1,1,1,1,1,1,1,1,2,2,2,2,2)
(EXPA-SA) 16 21 0.314933 (5,1,1,1,1,1,1,1,1,1,2,2,2,2,2,2)
(EXPA-SA) 16 22 0.323534 (5,1,1,1,1,1,1,1,1,2,2,2,2,2,2,2)
(EXPA-SA) 16 23 0.332466 (5,1,1,1,1,1,1,1,2,2,2,2,2,2,2,2)
(EXPA-SA) 16 24 0.341681 (5,1,1,1,1,1,1,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 16 25 0.351103 (5,1,1,1,1,1,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 16 26 0.360616 (5,1,1,1,1,2,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 16 27 0.370053 (5,1,1,1,2,2,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 16 28 0.379188 (5,1,1,2,2,2,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 16 29 0.38777 (5,1,2,2,2,2,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 16 30 0.39505 (5,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 16 31 0.400635 (5,2,2,2,2,2,2,2,2,2,2,2,2,2,2,3)
(EXPA-SA) 16 32 0.406398 (5,2,2,2,2,2,2,2,2,2,2,2,2,2,3,3)
(EXPA-SA) 16 33 0.412335 (5,2,2,2,2,2,2,2,2,2,2,2,2,3,3,3)
(EXPA-SA) 16 34 0.41844 (5,2,2,2,2,2,2,2,2,2,2,2,3,3,3,3)
(EXPA-SA) 16 35 0.424701 (5,2,2,2,2,2,2,2,3,2,2,3,3,3,3,3)
(EXPA-SA) 16 36 0.4311 (5,2,2,2,2,2,2,2,2,2,3,3,3,3,3,3)
(EXPA-SA) 16 37 0.437612 (5,2,2,2,2,2,2,2,2,3,3,3,3,3,3,3)
(EXPA-SA) 16 38 0.444201 (5,2,2,2,2,2,2,2,3,3,3,3,3,3,3,3)
(EXPA-SA) 16 39 0.450818 (5,2,2,2,2,2,2,3,3,3,3,3,3,3,3,3)
(EXPA-SA) 16 40 0.457398 (5,2,2,2,2,2,3,3,3,3,3,3,3,3,3,3)
(EXPA-SA) 16 41 0.463855 (5,2,2,2,2,3,3,3,3,3,3,3,3,3,3,3)
(EXPA-SA) 16 42 0.470077 (5,2,2,2,3,3,3,3,3,3,3,3,3,3,3,3)
(EXPA-SA) 16 43 0.475921 (5,2,2,3,3,3,3,3,3,3,3,3,3,3,3,3)
(EXPA-SA) 16 44 0.481208 (5,2,3,3,3,3,3,3,3,3,3,3,3,3,3,3)
(EXPA-SA) 16 45 0.485701 (5,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3)
As may be seen from these tables, the results obtained by Hillier and So (1995)
are confirmed in all cases considered, using the Markovian model of Heavey et al.
(1993) and complete enumeration. This result confirms that the Markovian model
used by the authors gives the same results as the Markovian model used by Hillier
and So (1995). As may be noted from the tables, the same lines were analyzed
174
Table 6.9. Throughput and buffer allocation for 5(1)9-, 11(10)61-station lines via SA
(Evaluative- No. of No. of Buffer Buffer
Search) Stations Slots Throughput Allocation
(DECO-SA) 5 8 0.665763 (5,2,2,2,2)

(DECO-SA) 6 10 0.653177 (5,2,2,2,2,2)
(DECO-SA) 7 12 0.644706 (5,2,2,2,2,2,2)
(DECO-SA) 8 14 0.63876 (5,1,2,2,2,3,2,2)
(DECO-SA) 9 16 0.634412 (5,1,2,2,2,2,3,2,2)
(DECO-SA) 11 20 0.628133 (5,1,2,2,2,2,2,2,3,2,2)
(DECO-SA) 21 40 0.614733 (5,1,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,3,2,2)
6 Double and Triple Optimization
(DECO-SA) 31 60 0.60998 (5,1,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,3,2,2)

(DECO-SA) 41 80 0.607546 (5,1,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,3,2,2)
(DECO-SA) 51 100 0.606066 (5,1,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2, 2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,3,2,2)
(DECO-SA) 61 120 0.605072 (5,1,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2, 2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,3,2,2)
(EXPA-SA) 5 8 0.590449 (5,2,2,2,2)
(EXPA-SA) 6 10 0.555666 (5,2,2,2,2,2)
(EXPA-SA) 7 12 0.527447 (5,2,2,2,2,2,2)
(EXPA-SA) 8 14 0.503894 (5,2,2,2,2,2,2,2)
(EXPA-SA) 9 16 0.483812 (5,2,2,2,2,2,2,2,2)
(EXPA-SA) 11 20 0.451099 (5,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 21 40 0.358377 (5,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 31 60 0.311407 (5,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 41 80 0.28145 (5,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 51 100 0.260082 (5,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2, 2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2)
(EXPA-SA) 61 120 0.243786 (5,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2, 2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2)
using the decomposition method and simulated annealing and separately complete
enumeration. These lines were also analyzed using the expansion method and again
with simulated annealing and separately complete enumeration. With respect to each
line the results were independent of the search method used but were dependent on
the evaluative method. The decomposition method in all cases gave a higher through-
put than that obtained using the expansion method and in some cases significantly
higher when the number of buffer slots were very small. However, both approx-
imate evaluative methods gave lower throughput rates than that calculated on an
exact basis using the Markovian method and either complete enumeration or simu-
lated annealing in the cases presented. With respect to the buffer allocation it would
appear from the above results that neither approximate evaluative method achieves
the optimal buffer allocation associated with the optimal throughput. The buffer allo-
cation obtained using the decomposition method is closer to the optimal. There is a
tendency for the expansion method to allocate the buffer slots toward the end buffers
of the line.
6.4 Concluding Remarks

Taking all the results of this chapter into account, it would appear that simulated
annealing is potentially a powerful optimizing technique and, as would be expected,
can lead to the optimal solution. Such a technique is of course required in situations
such as long lines where complete enumeration is infeasible. Despite the attractive-
ness of the expansion method in handling parallel-machine station lines, it must be
recognized that its use results in a maximum throughput less than what the real
line is capable of achieving with the given resources. Decomposition methods indi-
cate higher throughput levels than those indicated by the expansion method but are
still less than the exact optimal throughput determined through Markov methods
where comparisons are computationally possible. As the line becomes larger and
the number of buffer slots to be allocated increases, the throughput indicated using
the expansion method tends to the exact result. A major advantage of the expansion
method over the Markovian method when treating lines with parallel-machine sta-
tions is the faster computation time involved. Given this situation it would appear that
the appropriate advice to give to practitioners at this time would be to use simulated
annealing and decomposition methods for long lines with single-machine stations
and to use simulated annealing or complete enumeration and Markov methods for
short lines. If lines with parallel-machine stations are involved, consideration should
be given to the use of decomposition methods particularly if the Markovian methods
become computationally slow.

Lau and Martin (1986) developed a decision support system for the design of pro-
duction lines and considered the work-load and buffer allocation problems (WAP
and BAP) using simulation.
Hillier and So (1989) considered the pure server allocation problem in short
production lines with small or no intermediate buffers and with processing times
following the exponential, Erlang and Coxian-2 distributions.
Hillier and So (1996) treated the W + S problem in production lines with variable
processing times.
Magazine and Stecke (1996) dealt with the work-load, buffer and server alloca-
tion problems in production lines with K = 2 and 3 stations with parallel servers at
each station. They have listed several conjectures.
Futamura (2000) considered the optimal allocation of servers to stations with
different service time distributions in tandem queueing networks. The author showed
that the coefficient of variation (cv) of the service time distribution converges to unity
as the number of servers increases independently of the cv of the individual servers.
He examined the server-cv interaction.
Tempelmeier (2003) examined the W + B optimization problem by considering it
as a work-load optimization problem that includes the pure buffer allocation problem
as a sub-problem. The treatment of the service facility is somewhat unusual in that
the number of servers is not specified. His objective function is weighted with three
factors to minimize the total number of buffer slots, to maximize the mean processing
times at each station having in mind upper and lower bounds in the work-load on the
stations and to equalize the mean processing times between stations so as to have
as balanced a line as possible. A range for the buffer sizes at each buffer is given
and a minimum throughput is specified. The author solved first the buffer allocation
problem using an algorithm given by Gershwin and Schor (2000) and subsequently
he used a greedy heuristic to solve the work-load allocation problem.
References
3. Diamantidis, A.C., Papadopoulos, C.T., and Heavey, C. (2006), Approximate analysis of
serial flow lines with multiple parallel-machine stations, IIE Transactions, Vol. 39, issue
4, pp. 361–375.
4. Futamura, K. (2000), The multiple server effect: Optimal allocation of servers to sta-
tions with different service-time distributions in tandem queueing networks, Annals of
5. Gershwin, S.B. and Schor, J.E. (2000), Effcient algorithms for buffer space allocation,
Annals of Operations Research, Vol. 93, 1/4, pp. 117–144.
pp. 69–89.
7. Hillier, F.S. and So, K.C. (1989), The assignment of extra servers to stations in tandem
queueing systems with small or no buffers, Performance Evaluation, Vol. 10, pp. 219–
231.
References 177
pp. 448–461.
12. Lau, H.-S. and Martin, G.E. (1986), A decision support system for the design of unpaced
production lines, International Journal of Production Research, Vol. 24, No. 3, pp. 599–
610.
13. Magazine, M.J. and Stecke, K.E. (1996), Throughput for production lines with serial work
stations and parallel service facilities, Performance Evaluation, Vol. 25, pp. 211–232.
14. Papadopoulos, C.T. and Karagiannis, T.I. (2001), A genetic algorithm approach for
pp. 373–384.
16. Spinellis, D.D. and Papadopoulos, C.T. (2000b), Stochastic algorithms for buffer alloca-
tion in reliable production lines, Mathematical Problems in Engineering, Vol. 5, issue 6,
pp. 441–458.
Vol. 38, No. 3, pp. 509–541.
18. Tempelmeier, H. (2003), Simultaneous buffer and work-load optimization for asyn-
chronous flow production systems, in Proceedings of the Fourth Aegean International
Conference on the Analysis of Manufacturing Systems, July, 1–4, 2003, Samos Island,
Greece, pp. 31–39.
7
Cost Considerations
A major consideration in the management of production is to understand the cost

impact of various designs. Much of the work relating to overall production man-
agement including the design of facilities would seem to indicate that the decision
processes are serial rather than concurrent or iterative feedback. The classical idea
would appear to be that the engineering designers decide on the layout of the line
with primary interest in the engineering performance measures of the stations and
subsequently this design is costed and justified on some concept closely related to
discounted cash flow. There are a number of papers discussing the inadequacy of
the approach just outlined, particularly in relation to systems with inherent flexibility
and the justification of which may be more strategic rather than tactical.
An appropriate philosophy for world-class companies is for the company to
re-invent itself from time to time. This is in contrast to the view that a company
can retain its competitive advantage by simply being effective in existing markets.
Although it is important to retain or improve one’s position in existing markets,
it is often in the development of new markets, new customers and new products
or services that the health of the company is ensured. Re-invention therefore is
a questioning philosophy of how the company can do better today and what it
should be doing tomorrow. Clearly, it is a mixture of continuous improvement in all
aspects of its activities and new product development. Thus, manufacturing strategy
should support marketing strategy and give competitive advantage to the organiza-
tion. In the light of this strategic trust, investments in production facilities should
be carefully assessed from the point of view, of quality, flexibility, time-to-market,
dependability, market positioning as well as cost. It is of course difficult to put
all these diverse tangible and intangible benefits into one overall metric with the
view to choosing particular marketing and manufacturing strategies including pro-
duction systems designs. From the decision theory point of view what is involved
is a multiple-criteria decision problem. However, the authors are unaware of any
published work which applies techniques of multiple-criteria decision making such
as ranking/scaling methods, analytical hierarchical processes to the selection, at a
strategic level of a particular production system. There is an obvious need to recon-
cile the point of view of the economists, who would be concerned with opportunity
180 7 Cost Considerations
costs, and the approach of accountants who are generally more interested in using
actual historical or projected tangible costs, with the imperatives of manufacturing
engineering and the operating philosophies and desires of production management,
perhaps through a process of brainstorming and Delphi methods. Arising out of the
strategic analysis, outline and broadly based decisions on the required manufactur-
ing facilities and capabilities of the company would be developed. For example,
decisions might be made to invest in automation to develop flexible manufactur-
ing systems (FMS) or to use production lines. At a level below the strategic level,
tactical level decisions must be undertaken to specify, for example, location, size
and general layout of manufacturing facility, production machine processes, effi-
ciency/effectiveness, required throughput, capital cost targets, quality targets, degree
of flexibility envisaged and operating cost targets for each of the manufacturing
systems specified in outline form at the strategic level. Finally, it is at the detailed
design level that the production line, if one is required, is completely specified with
regard to the number of stations, the equipment at each station, the number of oper-
atives at each station, the inter-station buffer capacities and the work-load allocation
to achieve the required throughput. Issues related to maintenance and reliability of
equipment need to be considered. The overall objective is to meet the target cost per
unit produced. The focus of this book is at this detailed design level.
In general, there are two approaches to the incorporation of cost parameters into
the decision process at the detailed design stage of any manufacturing system. One
approach is to do all the engineering work essentially without specific reference to
costs relying on the experience of the engineers to develop an economical design and
subsequently to submit this design for cost evaluation which is generally undertaken
by accountants perhaps with some assistance from the engineers. The other approach
is to involve implicit decisions in relation to costs concurrently at all stages of the
engineering design. To some extent, the different approaches are adopted because of
the organization structure of the company, engineers doing the engineering, accoun-
tants doing the costing. Clearly, the concurrent evaluation of the cost implication of
engineering decisions is the preferable approach. Figures 7.1 and 7.2 illustrate the
difference between these two approaches.
Known Engineering Performance Determine

Specification
Constraints Design Evaluation Cost
Management/Tactical
Decisions
Final
Decision
Fig. 7.1. Production line design: Historical approach

7 Cost Considerations 181
Concurrent Engineering, Design, Costing

Specification Constraints
and Performance Evaluation
Management/Tactical
Decisions
Final
Decision
Fig. 7.2. Production line design: Modern approach
It is important to note that following operational experience of a particular imple-

mented design, some re-adjustment and improvements are inevitable which in some
cases may even require re-design of sections of the production line in certain circum-
stances. It is imperative that designers listen carefully to the views of the shop floor
personnel and conduct appropriate industrial experiments if there is evidence that
some re-engineering is required. The philosophy of continuous improvement should
apply to the design process as to all other activities of the organization.
Basically, there are two approaches in developing cost objective functions. The
first approach is concerned with the maximization of profit arising from production,
while the second approach is concerned with the minimization of costs incurred in
production. Because in the analysis one is dealing with stochastic processes, it is the
expected values of the random variables of the costs and profits that are optimized
based on the expected value of the throughput. More advanced analysis, particularly
in relation to risk, would involve the use of measures of variance and higher moments
of these random variables. Which particular objective function to use in a given situ-
ation is essentially a matter of managerial judgment. One of the advantages of using
cost minimization is that there is no need to consider revenue issues or equivalently
the price at which the product will be sold. Such minimum cost objective functions
are generally used in association with a constraint on the target level of production.
In other words, the production line is required to produce a target level of production
at minimum cost and so, normally there will be no incentive to go beyond the target
production level. The other objective function tends to look upon the production line
more as a generator of profit and the production level is an outcome rather than a
constraint on the system. Of course, even in the latter case, a minimum production
rate might be required in practice, but the optimal profit level may be achieved at a
higher production level. Some of the financial figures in cost objective functions are
known and deterministic. Others may have to be assessed using estimating proce-
dures such as regression and various types of curve fitting. Other financial figures in
these objective functions are essentially stochastic in nature and it is usual to treat
them in terms of their estimated values. Of course, if the financial figures are truly
uncertain, then an analytical treatment becomes virtually impossible. The financial
parameters placed in cost objective functions are those associated more with man-
agement accountancy rather than cost accountancy. This distinction is made here to
emphasize that allocated cost data, e.g., depreciation charges or distribution of over-
head costs, needs to be subject to careful review before being included as genuine
costs in these models.
Below, in Section 7.1, a number of appropriate cost models for profit maximiza-
tion with increasing degrees of sophistication are presented. Likewise, in Section
7.2, cost minimization models are presented. In all of these models, the assumption
is made that a decision has already been taken in regard to the specific type of pro-
duction system which will be used. Thus, the models are, in effect, tactical level
models.
7.1 Cost Models: Profit Maximization

Usually, in cost objective functions there is a need to consider the time value of cash
flows. The usual approach is through a discounted cash flow in which cash flow
in the future is discounted downwards to the present day. Although the concept of
this cash flow is well understood and is analytically tractable, nevertheless a seri-
ous technical/managerial issue arises in regard to what interest rate (I%, sometimes
called the hurdle rate) to assume in the discounted cash flow calculations. The nature
of the different elements in the objective function may be such that some parts of
these elements are integer valued only, whereas others may be continuous. This gives
rise to the usual computational difficulties associated with such objective functions.
Because of the different considerations mentioned in this section, there are a number
of different possible formulations of production line design problems involving cost
considerations. A number of profit objective functions will be considered below in
increasing order of sophistication.
A relatively simple profit objective function would be as follows:
max F1 = (R − C)XK∗ − ChW IP, (7.1)
where
R is the deterministic selling price of a unit of the product.
C is the direct cost associated with each unit produced. This direct cost should
include the operative wages cost per unit produced, the cost of material used per unit
and the recurrent machine costs per unit produced. In some situations, the latter costs
may not be included.
XK is the normalized throughput of the specified production line consisting of K
work-stations and K − 1 intermediate buffers, whereas XK∗ is the normalized through-
put multiplied by the maximum physical throughput of the system, i.e., the XK∗ is the
actual physical output per unit time.
W IP is the average work-in-process (WIP) summed over all the K − 1 buffers.
7.1 Cost Models: Profit Maximization 183
Ch is the inventory holding cost, reminiscent of the same factor, which appears
in inventory models, and is a measure of the cost of holding an item in inventory for
the same time period as is in the throughput.
Clearly, this rather simple objective function has a number of associated difficult
measurement/estimation problems. As indicated above, C, the direct cost, may not
be very inclusive because some of the direct costs associated with the production
of one unit of product are fuzzy. For example, it may be quite difficult to estimate
the per unit product cost of machine repair. Needless to remark, there is no specific
assignment of overhead costs and F1 could not be considered to be a normal profit
function as understood by accountants. Sometimes, the term “contribution to profit
and overhead” is used to indicate the term (R − C) above. Moreover, it is necessary
to specify the system to which objective function (7.1) is to be applied. What is fixed
and what may be considered decision variables?
To illustrate the use of cost objective functions in general, the following exper-
iment was undertaken. Perfectly reliable production lines with K = 3, 4, 5 and 6
stations were considered with the same average processing time at each station (bal-
anced lines). The number of total buffer slots to be allocated among the K − 1 buffers
varied from 1 up to 65, depending on the size of the production line.
Consider the output of a production line with K stations in which the finished
product is sold at a value of 50 financial units (FU) and costs 40 FU to produce. The
contribution margin per unit produced is therefore 10 FU. The assessment of Ch , the
average holding cost per unit held in work-in-process (WIP), may be approximated
as follows:
Ch = α CI,
where 0 < α < 1, C is the cost of production, defined above, and I is the relevant
or assigned interest rate per annum. The function of α is to take into account the
fact that the value of an item in WIP increases as the item progresses down the
line. In the example here, if α is assumed to be 0.5 and I = 10% per annum, then
Ch = 0.5C(0.1) = 0.05C. In general, the value of α would be estimated based on
material, labor and machine content of the item in inventory.
In this case, define the ratio, r, of the marginal contribution divided by the average
unit per annum holding cost as
f1 R −C 50 − 40
r= = = = 5.
f2 Ch (0.05)(40)
To continue the illustration, assume the isolated throughput of all the stations is
1 unit per minute (balanced line) and that the facility operates 2 shifts of 8 hours per
day for 250 working days per annum. The total maximum output is therefore
60 ∗ 2 ∗ 8 ∗ 250 = 240,000 units/annum.
The objective function now becomes
max F1 = f1 XK∗ − f2W IP = f1 (240,000)XK − f2W IP.

F1
1200000
1100000
1000000
900000
800000 F1
700000
600000
500000
400000
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58
Ntotal
Fig. 7.3. Value of F1 as a function of N for a 5-station production line with R = 50 FU, C = 40
FU, I = 10%, α = 0.5 (r = 5)
This is equivalent to maximizing
(240,000)rXK − W IP = 1,200,000XK − W IP.
This latter equation indicates that in any conceivable practical situation (where
W IP < 1,200,000XK ), an objective function of type (7.1), given above, leads to the
same optimal operating strategy as the more simple objective function of maximizing
throughput for the given number of buffer slots, N. Of course, equation (7.1) assumes
that R and C are constants independent of the value of XK∗ and so, for example, there
is no overtime premium and all product produced is sold. In Figure 7.3, the value of
the objective function for each N for these parameter values is given.
A modification of the profit objective function given in (7.1) would be
max F2 = max F1 − bN = (R − C)XK∗ − ChW IP − bN (7.2)
where bN represents the cost on an annual basis of the buffer space used, i.e., each
buffer slot costs b financial units (FU) per annum, where physical output, XK∗ , and
inventory costs, Ch in FU, are on a per annum basis. The term bN gives the plan-
ner some scope for financial justification of a proposed design of a production line
in which the number of stations is fixed, the cost of the proposed machines at the
stations is given but the decision in relation to the total number of buffer slots, N, is
still open. In a practical situation, where the effect of the W IP term is small and the
buffer allocation associated with maximum throughput for any N is known, it is only
necessary to determine the lowest value of N at which the marginal contribution to
the first term of F2 is lower than the cost of providing an extra buffer slot. A slightly
F2
1100000
1000000
900000
800000
F2
700000
600000
500000
400000
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59
Ntotal
FU, I = 10%, α = 0.5, b = 1000 FU (r = 5)
more appropriate objective function might replace the bN term with ∑Ki=2 bi Ni where
bi is the cost per annum of providing a buffer slot of type i for each of the Ni slots,
i = 2, 3, . . . , K.
It is clear from expression (7.2) that in a practical system, a situation will arise
where the marginal advantage of increasing the throughput by the optimal allocation
of an additional buffer slot will result in a reduction in the objective function, F2 . This
is illustrated in Figures 7.4 and 7.5. Figure 7.4 refers to the following system with
parameters: K = 5 stations, R = 50 FU, C = 40 FU, I = 10%, α = 0.5, b = 1000 FU,
whereas, Figure 7.5 refers to a system with parameters as follows: K = 5 stations,
R = 50 FU, C = 40 FU, I = 10%, α = 0.5, b = 5000 FU.
As indicated in Figure 7.4, F2 increases monotonically as N increases rather sim-
ilarly as F1 , shown in Figure 7.3. On the other hand, Figure 7.5 indicates that an
optimal value of F2 is achieved and that increasing the total number of slots, N,
above a certain level leads to a reduction in the value of the objective function F2
(from the value in this case of N = 27 onwards).
Extending the profit objective function further, it is desirable to bring into con-
sideration discounted cash flows or the time value of money. If one considers a
production line to be a generator of cash flows, there are three types of cash flows
involved:
• Initial investment in production line facility, i.e., machines, stations, buffer slots,
all at time t = 0. These flows are considered to be negative.
• Cash flows during the useful life of the production line. These flows would nor-
mally be considered to occur with regular frequency and consist of such flows
F2
900000
MAX
850000
800000
750000 F2
700000
650000
600000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35
Ntotal
FU, I = 10%, α = 0.5, b = 5000 FU (r = 5)
as revenue from sales, wages paid, energy used, materials purchased and used,
maintainance/repair costs, etc. In a net sense, these flows would be expected to be
positive.
• End of life flows, such as salvage value of machines and buffers, human resources
consequences and final disposal of remaining work-in-process (WIP). Normally,
these flows could be positive or negative.
In discounted cash flow analysis, the three different types of flows listed above must
be treated separately. The basic concept is to develop a present value of all the flows
at a chosen time t, usually, t = 0. If t = 0, the initial investment in the production
line facility does not have to be discounted but it is considered negative and is here
denoted by Ω1 . On the other hand, the other two cash flows have to be discounted.
Taking the end of life flows first and assume that they occur at time t = T , each of
these flows must be discounted using the following Present Worth Factor, P.W.F.∗ ,
of a cash flow received at time T from now (t = 0):
1
P.W.F.∗ =
(1 + I)T
where I is the interest rate per unit time (usually, per annum).
The above P.W.F.∗ is difficult to use in analysis and as a result the concept of
continuous discounting is introduced. The idea is basically as follows.
Assume an element of time Δ. The appropriate interest rate during this period
of time, Δ, would be (Δ)I and hence the continuous Present Worth Factor would be
approximated by
1
P.W.F.∗ = T .
(1 + (Δ)I) Δ
As there are in effect T /Δ periods of discounting of duration Δ, as Δ tends to 0,
(continuous discounting)
P.W.F.∗ → e−IT .
Each end of life flow is discounted by using a factor e−IT . The notation Ω2 is
used for the Present Worth Value, P.W.V., of the end of life cash flows, i.e.,
Ω2 = P.W.V. = (P.W.F.∗)(Net end of life cash flows) = e−IT S
where S is the net end of life cash flows.

On the other hand, each cash flow during the useful life of the production line is
discounted using the following formulation:
T
Ω3 = f (t)e−It dt
0
where f (t)dt is the cash flow between t and t + dt and T is the life of the production
line and I is the interest rate. If f (t), 0 ≤ t ≤ T , the cash flow during the operating
life of the production line is equal to f , a constant,
f
Ω3 = [1 − e−IT ].
I
The reader might note that it may be necessary to modify the functional form
of Ω3 in cases where there are discrete rather than continuous cash flows during the
life of the equipment. The modification would entail the addition of terms such as
∗
F · e−It , where a discrete cash flow of net value F occurred at time t ∗ .
A further development of the profit objective functions would be as follows:
max F3 = P.W.V. = −Ω1 + Ω2 + Ω3 (7.3)
where the different cash flows are now discounted. This model can be adopted to
incorporate all reasonable cost objective functions.
As an illustration, it may be instructive to modify the above numerical examples
as follows:
Scenario 1: Each of the machines costs 200,000 FU including tooling and has a
salvage value of 20,000 FU at the end of the fifth year of its operation. Each buffer
slot costs 1000 FU including capitalized rent and has a salvage value after five years
of 200 FU. R − C = 10 FU is revenue per unit less materials and energy costs and
Ch = 2 FU, I = 10%, α = 0.5, and 80,000 FU is a per station annual cost based on
a working year of 4000 hours and 20 FU per hour to cover fixed wages and fixed
routine repair costs per station. These 4000 hours represent 250 working days of two
eight-hours shifts per day. The mean production rate of the last station is 5 units per
minute and therefore, 1,200,000 is the maximum mean production rate of the system
per annum working on the basis of two eight-hour shifts per day and 250 working
days per annum.
Scenario 2: Each of the machines costs 100,000 FU and has a salvage value of
20,000 FU at the end of the fifth year of its operation. Each buffer slot costs 10,000
FU and has a salvage value after five years of 2000 FU. R − C = 4 FU, Ch = 1
FU, I = 10%, α = 0.5, and 40,000 FU is the per station annual cost based on a
working year of 2000 hours and 20 FU per hour to cover fixed wages and fixed
routine repair costs per station. These 2000 hours represent 250 working days of one
eight-hour shift per day. The mean production rate of the last station is 1 unit per
minute and therefore, 120,000 is the maximum mean production rate of the system
per annum working on the basis of one eight-hour shift per day and 250 working
days per annum.
For scenario 1, the following may be determined, for a five-station production
line (K = 5):
Ω1 = K(200,000) + 1000N + Ch (W IP)

= 5(200,000) + 1000N + 2(W IP)
= 1,000,000 + 1000N + 2(W IP).
In the above expression, the term 2(W IP) arises from valuing the beginning inven-
tory at the same level as it is valued during the life of the production line and at the
end of the life of the production line. This assumption could of course be modified,

Ω2 = K(20,000) + 200N + Ch (W IP) (e−5I )

= 5(20,000) + 200N + 2(W IP) (e−5I )

= 100,000 + 200N + 2W IP (e−0.5 )

= 100,000 + 200N + 2W IP (0.6065)

1 − e−5I
Ω3 = {(R − C)1,200,000XK − 80,000K}
I

1−e −5(0.1)
= [12,000,000XK − 400,000]
0.1

= [120,000,000XK − 4,000,000] 1 − e−0.5
= [120,000,000XK − 4,000,000](0.3935)
= (47.22XK − 1.574)106
leading to the following objective function:
max F3 = −Ω1 + Ω2 + Ω3
= −2,513,350 − 878.7N − 0.787WIP + (47.22)(106)XK .
For scenario 2, F3 may be determined in a similar fashion leading to F3 in this

case being as follows:
max F3 = −Ω1 + Ω2 + Ω3
= −1,226,350 − 8787N − 0.3935WIP + 1,888,800XK
Figure 7.6 indicates the value of the objective function, F3 , for scenario 1, above.
As may be seen, F3 is a monotonically increasing function of N. Figure 7.7 shows
the value of the objective function, F3 , for scenario 2, which attains its optimal value
at N = 26.
It is interesting to note the form of F1 , F2 and F3 as developed for the above
examples. In general, for this type of analysis the form of these objective func-
tions is:
max F = α XK − β W IP − γ N
45000000
40000000
35000000
F3
F3
30000000
25000000
20000000
1 6 11 16 21 26 31 36 41 46 51 56 61
Ntotal
Fig. 7.6. Value of F3 as a function of N for a 5-station balanced production line with R−C = 10
FU, I = 10%, α = 0.5, Ch = 2 FU, b = 1000 FU, 2 shifts per day, 5 units per minute maximum
mean production rate of the system
150000
100000
50000
0
_ 50000 1 6 11 16 21 26 31 36 41 46 51 56
F3
_100000 F3
_150000
_200000
_250000
_300000
Ntotal
Fig. 7.7. Value of F3 as a function of N for a 5-station balanced production line with R −C = 4
FU, I = 10%, α = 0.5, Ch = 1 FU, b = 10,000 FU, 1 shift per day, 1 unit per minute maximum
mean production rate of the system
where α , β , γ > 0. In many practical situations, α is large while the β and γ are much
less than the α . Clearly, W IP ≤ N. So, for these lines, it is reasonable to say that the
objective function is a function of XK and N. If there is no cost associated with the
provision of buffer space (γ = 0, the case of F1 above), clearly, there is really no need
to use a cost-based objective function as a simpler objective function of maximizing
throughput will give the same result, always assuming that the contribution of β W IP
does not lead to a restriction on N. It would appear that the only case of significant
practical interest is when there is a limit on the size of N from a cost aspect, in other
words, buffer sizes beyond this N in effect reduce the value of the cost objective
function.
The reader may have some observations about the values of the parameters used
in some of the examples presented above, but the intention is to illustrate the differ-
ent cases which may arise rather than any particular practical example. The reader
will note that an underlying assumption in the above examples is that the lines are
balanced.
7.2 Cost Models: Cost Minimization

Up to now, the maximization of profit was the objective function. Attention is now
turned to cost minimization objective functions where the objective function consists
of a sum of costs (perhaps including present worth value terms) which is to be mini-
mized and the set of constraints includes a production throughput target which must
be at least achieved.
Consider a balanced production line with K stations. A total buffer capacity of N
slots are available for allocation. The production line has to meet a target production
level of at least X0 at overall minimum cost.
The problem formulation is as follows:

K
min G1 = min ∑ biNi + ChW IP + γ (XK − X0) (7.4)
i=2
s.t.
XK ≥ X0
Ni,min ≤ Ni ≤ Ni,max
∑Ki=2 Ni ≤ N
Each Ni , i = 2, . . . , K is an integer,
where
bi , i = 2, . . . , K, may be a net present value type cost coefficient associated with
each buffer slot.
Ch could also incorporate a net present value coefficient because of the cost of
holding average work-in-process, W IP, that is incurred throughout the life of the
project.
γ is a linear penalty cost associated with going above the required production
target, X0 , and it could include considerations of present value. γ would normally
7.2 Cost Models: Cost Minimization 191
be very large in comparison to Ch , in practical situations as a deviation of the mean

production rate from the required target level, X0 , is heavily penalized.
In this formulation, the concept of a constraint on the size of the buffer at each
location i is incorporated, where Ni,min is the lower bound and Ni,max the upper bound,
respectively, of Ni , i = 2, . . . , K. The usual assumption of integer Ni is made.
It should be observed that the objective function assumes that the N available
buffer slots are not necessarilly all to be utilized, whereas it may be possible to
achieve the desired production level at a lower overall cost while not utilizing all
available buffer slots. The implication is, of course, that available buffer slots not
used are not part of the objective function.
The computational procedure is as follows (assuming for simplicity a balanced
line with K stations):
1. Select ∑Ki=2 Ni = N ∗ ≤ N.
2. With this N ∗ determine the value of G1 for different buffer allocations which meet
the throughput requirement that XK ≥ X0 .
3. Determine the minimum G1 from those obtained in step 2, above. It should be
noted that in step 2 the objective function, in fact, reduces to minimization of
ChW IP + γ (XK − X0 ) if bi = b for all i = 2, . . . , K.
4. Search the next value of N ∗ and determine if an improvement can be made in
the value of G1 and continue until ∑Ki=2 Ni = N ∗ ≤ N is the minimum N ∗ which
results in the minimum of the objective function G1 while meeting the throughput
constraint.
Should there be a tie, i.e., when the minimum cost is achieved for two different
values of N ∗ , break ties arbitrarily.
A deviation on the above balanced production line would be to assume a non-

balanced production line, where wi , the work-load at station i, is specified at each i,
i = 1, . . . , K, and is outside the control of the system designer. In this latter case, sim-
ilar observations would lead to the same conclusion, viz., the cost objective function
could be reduced to one of minimizing W IP.
Below, two different examples are considered.
Example 1: K = 5 stations, bi = b = 1000 FU, i = 2, 3, 4, 5. Ch = 2 FU, X0 = 0.80,
γ = 180,000 FU. The value of γ was determined on the basis of a one shift system of
8 hours per day, 250 working days per annum and a maximum throughput of the last
station of 30 units per hour. A nominal value of 3 FU for each unit of excess product
produced was assumed. This nominal value would include such costs as material and
labor costs, scrap values and disposal costs. A maximum of N = 32 buffer slots are
available for allocation among the the four (4) intermediate buffers of the system.
The problem may now be formulated as follows:

5
min G1 = min ∑ 1,000Ni + 2W IP + 180,000(XK − 0.80)
i=2
s.t.
XK ≥ 0.80
2 ≤ Ni ≤ 10, i = 2, 3, 4, 5
∑Ki=2 Ni ≤ 32
Each Ni , i = 2, . . . , 5 is an integer.
Numerical investigation of the above problem using the objective function, G1 , as
specified above, leads to the following results: N2 = 10, N3 = 3, N4 = 7 and N5 = 10;
XK = 0.806982; W IP = 17.588877. As expected, the BAP-B problem gives identical
results, but it has a simpler objective function, i.e., minW IP.
Example 2: K = 5-stations balanced line,
b2 = 10,000, b3 = 1,000, b4 = 1,500, b5 = 1,400 FU.

Ch = 2FU, X0 = 0.80, γ = 180,000 FU,
as in example 1. Again, a maximum of N = 32 buffer slots are available for allocation

among the four (4) intermediate buffers of the system.
The problem may now be formulated as follows:
min G1 = min {(10,000N2 + 1,000N3 + 1,500N4 + 1,400N5)

+ 2W IP + 180,000(XK − 0.80)
s.t.
XK ≥ 0.80
2 ≤ Ni ≤ 10, i = 2, 3, 4, 5
∑Ki=2 Ni ≤ 32
Each Ni , i = 2, . . . , K is an integer.
Numerical investigation of the above problem using the objective function, G1 , as
specified above, leads to the following results: N2 = 3, N3 = 10, N4 = 7 and N5 = 10;
XK = 0.814949; W IP = 13.914076.
If the design involves the allocation of work to each station, wi , i = 1, . . . , K,
and the buffer allocation has already been decided, viz., Ni , i = 2, . . . , K, are given, a
minimization cost objective function and associated constraints may be formulated
as follows:

K
min G2 = min ChW IP + ∑ [εi |wi − wi | + ηi (wi − 1)δ (wi − 1)] (7.5)
i=1
s.t.
XK ≥ X0
∑Ki=1 wi = K,
wi,min ≤ wi ≤ wi,max ,
Ni , i = 2, . . . , K are a feasible set of specified integer values which will allow for
the achievement of the required throughput, X0 . wi,min and wi,max are, respectively,
the lower and upper bounds of the wi ’s, i = 1, . . . , K, acceptable to the respective
stations.
7.2 Cost Models: Cost Minimization 193
A set of feasible Ni ’s may be obtained from the solution of the BAP-C

problem, i.e.,
min ∑Ki=2 Ni
s.t.
XK ≥ X0
∑Ki=2 Ni ≤ N,
Ni,min ≤ Ni ≤ Ni,max , i = 2, . . . , K
Each Ni , i = 2, . . . , K is an integer.
Clearly, other feasible sets of Ni ’s could be specified knowing the results of this
sub-problem.
δ (wi − 1) is the Kronecker delta function defined as follows:

1, if a > 0
δ (a) =
0, if a ≤ 0.
The term in the objective function involving the coefficient εi is a present worth
value of costs associated with operating machines away from their natural design
speeds, wi , i = 1, . . . , K. The concept is that machines have a natural design speed
and that running a machine above or below this design speed incurs a linear penalty.
Although the format of the penalty in the above objective function involves an abso-
lute magnitute expression, it is possible for analytical reasons to use a term such as
(wi − wi )2 in place of |wi − wi |, particularly in the case of small deviations. If desired,
the factor εi can be adjusted by simple numerical means to make the two expressions
equivalent at a particular deviation level. The term involving the coefficient ηi cap-
tures the present worth value of the extra wages which in some circumstances may
be paid to the operators arising out of the unequal work-load at the stations. The
payment would be only applicable to those operators working in stations where the
work-load was greater than the average normalized value of 1.
As may be seen from the above, it becomes increasingly more difficult to be
precise about what costs to include in the objective function and the problem of the
assessment of these costs is by no means trivial. Consequently, a different approach
may be adopted, whereby, in conjunction with management, the designer specifies
what is essentially a wish list of objectives and places weights on these objectives.
For example, an objective function could be formed using the following:
• Minimize the total number of buffers.
• Maximize the service rates.
• Equalize the service times.
• Minimize the average work-in-process, W IP.
• Minimize the deviation of the throughput from the target throughput.
The objective function in this case could be:

K K K
min G3 = min f1 ∑ Ni + f2 ∑ (wi,max − wi ) + f3 ∑ (wi − 1)2
i=2 i=1 i=1

+ f4W IP + f5 (XK − X0 ) (7.6)
s.t.
XK ≥ X0 ,
∑Ki=1 wi = K,
where
Ni , i = 2, . . . , K are integer and wi,max is the maximum wi acceptable to station i,
i = 1, . . . , K.
The optimization problem could be formulated using the usual constraints. The
values of weights fi , i = 1, . . . , 5, would be determined in consultation between the
designers and the managers of the system with some regard given to the magnitude of
the different costs involved using perhaps brainstorming or a Delphi approach. AHP,
other paired comparison methods and other multicriterion decision methodologies
would also have a place in this approach. It might be also noted that this approach
could be formulated via a goal programming format.
Generally speaking, it must be remembered that the coefficient of the average
WIP term would tend to be small relative to the coefficients of other terms in any
of the above objective functions. This should not lead to the conclusion that mini-
mizing W IP is never a sensible objective function in the design of production lines
as has already been shown in the examples presented above. In all cases where cost
considerations are involved, it would be very desirable for the designer to test the
sensitivity of the design to changes in the costs figures used in the objective function.
References
2. Altiok, T. and Stidham, S. Jr. (1983), The allocation of interstage buffer capacities in
production lines, AIIE Transactions, Vol. 15, No. 4, pp. 292–299.
3. Beged Dov, A.G., Carmichael, C.D., Ferguson, S.T., Mitchel, I.E., and Strube, W.H.
(1968), Production programming by revenue curve analysis, Proceedings of the Second
Winter Simulation Conference on Applications of Simulations, pp. 53–57.
4. Blank, L. and Carrasco, H. (1985), The economics of new technology: System design
and development methodology, Annual International Industrial Engineering Conference
Proceedings, pp. 161–168.
5. Boer, C.R. and Metzler, V. (1986), Economic evaluation of advanced manufacturing by
means of simulation, Material Flow, Vol. 3, pp. 215–224.
6. Boothroyd, G. (1982), Economics of Assembly Systems, Journal of Manufacturing
Systems, Vol. 1, No. 1, pp. 111–126.
7. Canada, J.R. and Sullivan, W.G. (1990), Persistent pifalls and applicable approaches
for justification of advanced manufacturing systems, Engineering Costs and Production
Economics, Vol. 18, pp. 247–253.
8. Carrasco, H. and Blank, L. (1987), Prototype development of an investment tracking sys-
tem for life cycle costing, World Productivity Forum & 1987 International Industrial
Engineering Conference Proceedings, pp. 211–216.
9. Christy, P.D. and Kleindorfer, B.G. (1990), Simultaneous cost and production analysis of
manufacturing systems, Proceedings of the Winter Simulation Conference, pp. 582–589.
10. Gustavson, R.E. (1983), Choosing manufacturing systems based on unit cost, Proceedings
of the 13th International Symposium on Industrial Robots and Robots, Vol. 1, pp. 4-85–
4-104.
References 195
11. Haider, S.W. and Blank, L.T. (1983), A role for computer simulation in the economic
analysis of manufacturing systems, Proceedings of the Winter Simulation Conference,
pp. 199–206.
12. Hedge, G.G., Ramamurthy, K., Tadikamalla, P.R., and Kekre, S. (1988), Capacity choice,
work-in-process inventory and throughput: A simulation study, Proceedings of the Winter
Simulation Conference, pp. 662–666.
13. Helber, S. (1999), Performance Analysis of Flow Lines with Non-Linear Flow of Material,
Springer-Verlag.
14. Hutchinson, G.K. and Holland, J.R. (1982), The economic value of flexible automation,
Journal of Manufacturing Systems, Vol. 1, No. 2, pp. 215–227.
15. Jensen, P.A., Pakath, R., and Wilson, J.R. (1988), Optimal buffer inventories for mul-
tistage production systems with failures, SMM 88-1, School of Industrial Engineering,
Purdue University.
16. Jeong, K.C. and Kim, Y.-D. (2000), Heuristics for selecting machines and determining
buffer capacities in assembly systems, Computers & Industrial Engineering, Vol. 38,
pp. 341–360.
17. Karmarkar, U. (1987), Manufacturing configuration, capacity and mix decisions consid-
ering operational costs, Journal of Manufacturing Systems, Vol. 6, Part 4, pp. 315–324.
18. Meredith, J.R. and Suresh, N.C. (1986), Justification techniques for advanced manu-
facturing technologies, International Journal of Production Research, Vol. 24, No. 5,
pp. 1043–1057.
19. Moerman, P.A. (1988), Economic evaluation of investments in new production technolo-
gies, Engineering Costs and Production Economics, Vol. 13, pp. 241–262.
20. Monga, A. and Zuo, M.J. (2001), Optimal design of series-parallel systems considering
maintenance and salvage value, Computers & Industrial Engineering, Vol. 40, No. 4,
pp. 323–337.
21. Noble, J.S. and Tanchoco, J.M.A. (1993), Design justification of manufacturing systems –
A review, The International Journal of Flexible Manufacturing Systems, Vol. 5, pp. 5–25.
22. Smith, J.M. and Chikhale, N. (1994), Buffer allocation for a class of non-linear stochas-
tic knapsack problems, Department of Industrial Engineering and Operations Research,
University of Massachusetts, Amherst.
23. Smith, J.M. and Daskalaki, S. (1988), Buffer space allocation in automated assembly
lines, Operations Research, Vol. 36, pp. 343–358.
24. Son, Y.K. (1991), A cost estimation model for advanced manufacturing systems, Interna-
tional Journal of Production Research, Vol. 29, No. 3, pp. 441–452.
Vol. 38, No. 3, pp. 509–541.
26. Tempelmeier, H. (2003), Simultaneous buffer and work-load optimization for asyn-
chronous flow production systems, In Proceedings of the Fourth Aegean International
Conference on the Analysis of Manufacturing Systems, July, 1–4, 2003, Samos Island,
Greece, pp. 31–39.
27. Yeralan, S., Dieck, A.J., and Darwin, R.F. (1986), Economically optimum maintenance,
repair and buffering operations in manufacturing systems, The Engineering Economist,
Vol. 31, No. 4, pp. 279–292.
A
Mathematical Fundamentals
In this appendix, the objective is to give an outline review of those areas of mathemat-
ics including probability and statistics which are essential background for a complete
understanding of the material covered in the main text. It is assumed that many read-
ers would have no need for this appendix. Of necessity, the treatment is rather brief
and has an engineering rather than a pure mathematical orientation and with some
emphasis on numerical examples.
A.1 Vectors and Matrices

A.1.1 Vectors
The definition of a vector is an ordered set of numbers which in applications to

manufacturing systems are generally real numbers, e.g., vector v = (v1 , v2 , v3 , v4 ) is
a vector with four components or elements and each component is a member of a set
of real numbers. The ordering relates to the position of the components in the array.
A numerical example would be v = (4, 7, 8, 5) where the third component of the
[1 × 4] vector is 8. The vector φ = (0, 0, . . . , 0) is known as the null vector. Specific
vector operations are shown numerically below.
Given α = 3 and v = (v1 , v2 , v3 , v4 ) = (3, 5, 8, 2), then α v = (9, 15, 24, 6) is a
[1 × 4] vector. The transpose of a [1 × n] row vector, v, is a [n × 1] column vector,
denoted by vT , as illustrated below.
⎡ ⎤
2
⎢1⎥
v = (2, 1, 4, 7), v = ⎢
T
⎣4⎦
⎥
7
If
c = (4, 3, 2, 1)
198 A Mathematical Fundamentals
and
d = (5, 6, 9, 12)
both [1 × 4] row vectors, then the sum of the vectors is
c + d = (9, 9, 11, 13)
and the subtraction of the vectors is
c − d = (−1, −3, −7, −11).
A.1.2 Matrices
A matrix R of dimension (m × n) is a rectangular array of real numbers arranged in

m rows and n columns as follows:
⎡ ⎤
r11 r12 · · · r1n
⎢ r21 r22 · · · r2n ⎥
⎢ ⎥
R=⎢ . . . . ⎥
⎣ . . . . ⎦
. . . .
rm1 rm2 · · · rmn
a (m × n) matrix.
If R and S are (3 × 3) matrices as given below:
⎡ ⎤
123
R = ⎣4 5 6⎦
789
⎡ ⎤
111
S = ⎣3 4 5⎦
131
then the matrix R + S is the following (3 × 3) matrix
⎡ ⎤
2 3 4
R + S = ⎣ 7 9 11 ⎦
8 11 10
and the matrix R − S is the following (3 × 3) matrix

⎡ ⎤
012
R−S = ⎣1 1 1⎦.
658
If α = 2, then the matrix T = α R is given by

⎡ ⎤
2 4 6
T = α R = ⎣ 8 10 12 ⎦
14 16 18
A.1 Vectors and Matrices 199
The transpose AT of matrix A of dimension (m × n) is a matrix of dimension (n × m)

derived as follows:
⎡ ⎤
a11 a12 · · · a1n
⎢ a21 a22 · · · a2n ⎥
⎢ ⎥
A=⎢ . .. .. ⎥
⎣ .. . ··· . ⎦
am1 am2 · · · amn
⎡ ⎤
a11 a21 · · · am1
⎢ a12 a22 · · · am2 ⎥
⎢ ⎥
AT = ⎢ . .. .. ⎥ .
⎣ .. . ··· . ⎦
a1n a2n · · · amn
For example, if A is the following (4 × 2) matrix:
⎡⎤
2 4
⎢7 9 ⎥
A=⎢
⎣ 8 10 ⎦
⎥
11 −5
AT is the (2 × 4) matrix:

2 7 8 11
AT = .
4 9 10 −5
The product CD of matrices C and D is defined only if the number of columns of C

is equal to the number of rows of D. If C has dimension (m × r) and D has dimension
(r × n), then CD has dimesnion (m × n) with the (i, j)th element of CD given by
n
∑ Cik Dk j .
κ =1
Conceptually, the elements of CD are the inner or dot products of the appropriate
row and column of C and D, respectively.
Given the (2 × 3) matrix, C:

246
C=
712
and the (3 × 4) matrix D:
⎡ ⎤
9214
D = ⎣1 2 2 1⎦
8236
CD is the following (2 × 4) matrix:

70 24 28 48
CD = .
80 20 15 41
Note that DC given C and D above has no meaning.

Consider the first row of C as a row vector E and the second column of D as a
column vector F where
E = (2, 4, 6)
⎡ ⎤
2
F = ⎣ 2 ⎦,
2
then the inner or dot product of the two vectors E · F = 24 which is the (1,2) element
of CD.
Matrix operations have the following general properties:
(C + D) + B = C + (D + B) = C + D + B
C + B = B +C
(C + D)B = CB + DB
(CD)B = C(DB) = CDB
(B + C)T = BT + CT
(AB)T = BT AT .
If CT = C, then C is said to be a symmetric matrix.

The identity matrix, I, is an (n × n) matrix as follows:
⎡ ⎤
1 0 0 0 ··· 0
⎢0 1 0 0 ··· 0 ⎥
⎢ ⎥
⎢0 0 1 0 ··· 0 ⎥
⎢ ⎥
I = ⎢0 0 0 1 ··· 0 ⎥.
⎢ ⎥
⎢ .. .. .. .. .⎥
⎣ . . . . · · · .. ⎦
0 0 0 0 ··· 1
The elements of I are either 1 or 0 and the elements with value 1 are placed on what
is termed the principal diagonal of the matrix.
A.1 Vectors and Matrices 201
Given a square matrix C of dimension (n × n) and another square matrix D of the

same dimension, then, if
CD = I
D is said to be the inverse of C and C is said to be the inverse of D.
If C and D are as follows:

15 1 2 −5
C= ; D=−
42 18 −4 1
CD = I; DC = I
and so D is the inverse of C and C is the inverse of D.
There are a number of efficient computer packages available to determine the
inverse of matrices. The inverse of a matrix may be determined by way of elementary
transformations/operations as indicated below.
To determine the inverse of C:

15
C=
42
1 0 1 5 (i)
0 1 4 2 (ii)
1 −2.5 −9 0 (iii) = (i) − 2.5(ii)
− 19 2.5
9 1 0 (iv) = (iii)/(−9) .
− 49 10
9 4 0 (v) = 4(iv)
4
9 − 1
9 0 2 (vi) = (ii) − (v)
2
9 − 1
18 0 1 (vii) = (vi)/2
D from (iv) and (vii) is

− 19 2.5 1 2 −5
D= 9 =−
9 − 18
2 1
18 −4 1
which is the inverse of C, as shown above.

If a square matrix C has an inverse, the inverse may be shown to be unique.
A square matrix which does not have an inverse is said to be singular. For example,
the following (2 × 2) matrix C:
12
C=
24
does not have an inverse. To show that this is the case by contradiction, assume that
D is the inverse:
ab
D=
cd
As CD = I then
a + 2c =1 (i)
b + 2d =0 (ii)
.
2a + 4c =0 (iii)
2a + 4c =1 (iv)
(i) and (iii) are incompatible as are (ii) and (iv) and so C does not have an inverse.
Determining the inverse of matrices of high order using classical methods is
tedious; one such method involves a well-known mathematical process of obtaining
the determinant of the square matrix and then deriving the determinants of the adjoint
matrices. Interested readers are referred to standard textbooks on linear algebra.
A simple example of the process involved is given below:

15
C=
42
The determinant of C, denoted by ||C||, is given by
||C|| = 1 × 2 − 5 × 4 = −18.
Determinants of the adjoint matrices are obtained by deleting appropriate rows and
columns of the parent matrix:
||C11 || = 2; ||C12 || = 4; ||C21 || = 5; ||C22 || = 1

1 (−1)2 2 (−1)3 5 1 2 −5
D = C−1 = − = −
18 (−1)3 4 (−1)4 1 18 −4 1
as above.
In the analysis of queues it is sometimes useful to use the notation eAt , where,
1 22 1 33
eAt = I + At + A t + A t + ···
2! 3!
It may be easily shown that
d At
{e } = AeAt or [eAt ]A
dt
where A is a square matrix.
A.2 Probability 203
Readers are referred to standard texts on linear algebra to determine eAt .

The following particular procedure may be used to determine eAt .
Given

12
A= ,
03
let [F(t)] = [e[A]t ] and F(0) = I. Knowing that dtd e[A]t = [A]e[A]t and taking the Laplace
transforms of both sides
& − I] = [A][F(s)]
[sF(s) &
& =I
[sI − A]F(s)
&
[F(s)] = [sI − A]−1
&
where [F(s)] is the Laplace transform of [F(t)].
&
[eAt ] = L −1 [F(s)] = L −1 [sI − A]−1
1 2

s−1 (s−1)(s−3)
= L −1
1
0 s−3
t 3t
e e − et
= .
0 e3t
Note that
2 1 1
= −
(s − 1)(s − 3) s − 3 s − 1
and

1
L −1 = eα t .
s−α
A.2 Probability
In probability theory, an experiment is a well-defined process, the observed outcome
of which is not known in advance. The set of all possible outcomes is the sample
space. Depending on the experiment, the sample space can consist of a countable
number of outcomes, in which case, the sample space is described as discrete. If the
sample space does not consist of a countable number of outcomes, it is described as
continuous. A random variable is a function that assigns a value to every element of
the sample space. Thus random variables are generally either discrete or continuous,
but may be hybrid. Associated with each discrete value of a random variable is a
probability mass function and associated with each continuous value of a random
variable is a probability density function . An example of a discrete random variable
would be the outcome of an experiment involving the determination of the sum of
two throws of two fair dice. The sample space, here, would be the discrete integer
values from 2 to 12 and the associated probability mass function is given below,
where T1 is the random variable indicating the outcome of the first throw and T2 is
the random variable indicating the outcome of the second throw and X is the random
variable which is the sum of T1 and T2 .
The mean value or the expected value of any random variable is a measure of its
central tendency and in the case of a discrete distribution is defined as follows
n
EX = ∑ xi pi ,
i=1
where xi is a value of the discrete random variable and pi is the associated probability
mass function (probability of occurence).
In the above example
11
EX = ∑ xi pi = 7.
i=1
There are 11 different values of the random variable X = T1 + T2 and the respective
probabilities pi , i = 2, . . . , 11 are given in Table A.1.
The probability mass function of X = T1 + T2 is symmetric (see Table A.1). The
reader should note that whereas the probability mass functions of T1 and T2 are both
uniform (probabilities of all possible events in each case being 1/6 as the die is fair
and the throws are independent), the probability mass function of the discrete random
variable X = T1 + T2 has a triangular form. The process of deriving the distribution
of a random variable which is the sum of other random variables is known as convo-
lution. The example given above, in a very basic way, illustrates the fundamentals of
the Central Limit Theorem in that a rudimentary form of a normal distribution has
evolved from the convolution of two uniform probability mass functions.
The variance of a discrete random variable, X , is defined as
n
Var X = σX2 = E(X − EX)2 = ∑ (xi − EX)2 pi
i=1
Table A.1. Probability mass function
X = T1 + T2 Probability mass function

2 1/36
3 2/36
4 3/36
5 4/36
6 5/36
7 6/36
8 5/36
9 4/36
10 3/36
11 2/36
12 1/36
A.2 Probability 205
where σX is the standard deviation of the random variable X . For the above example,
the variance of X = T1 + T2 is
11 11
Var X = σX2 = ∑ (xi − EX)2 pi = ∑ x2i pi − (EX)2 = EX 2 − (EX)2 = 35/6 = 5.833.
i=1 i=1
The reader will note that with respect to discrete random variables
n
∑ pi = 1
i=1
and the corresponding results for the continuous random variables is

∞
f (x)dx = 1
−∞
where x is a value of X , a continuous random variable with associated probability

density function f (x). The reader will note that the range of x does not necessarily
extend from −∞ to +∞. For example, if X is a random variable indicating service
time, the values of X are strictly greater than or equal to zero. The variance is a
measure of the spread of a random variable.
A measure of significance in production line analysis is the coefficient of varia-
tion (c.v.) which gives the dimenionless value of the standard deviation of a random
variable in terms of the value of its mean and is defined as follows:
'
standard deviation σX2
c.v.(X) = = .
mean EX
The c.v. of X = T1 + T2 is 0.345.
To illustrate the use of probability density functions in association with contin-
uous random variables, assume that the service time of an operation is uniformly
distributed between a minimum of 1 minute and a maximum of 3 minutes as
illustrated in Figure A.1.
Probability
Density
Function
1/2
0 1 2 3
Time (minutes)
Fig. A.1. Continuous uniform distribution

Table A.2. Discrete probability distributions
Random variable Probability mass Mean Variance Range of Parameter

name function variable values
Bernoulli px (1 − p)1−x p p(1 − p) x = 0, 1 0< p<1
Binomial n! px (1 − p)n−x np np(1 − p) x = 0, 1, . . . , n 0< p<1

x!(n − x)!
n = 1, 2, . . .
e−λ λx!
x
Poisson λ λ x = 0, 1, 2, . . . λ >0
1 1− p
Geometric p(1 − p)x−1 p x = 1, 2, . . . 0< p<1
p2
The probability density function (pdf), f (t), is defined as follows

1
, if 1 ≤ t ≤ 3;
f (t) = 2
0. otherwise.
Clearly, as expected
3
f (t)dt = 1.
1
The expected value of this random variable, T , is given by
3 3
1
ET = t f (t)dt = tdt = 2
1 2 1
and the variance of T is

3 3
1 1
Var T = (t − 2)2 f (t)dt = (t − 2)2dt = .
1 2 1 3
The c.v.(T) is given by '
1
3
c.v.(T ) = = 0.289
2
Next, a number of important probability distributions which are applicable to
the analysis of production lines are considered. Table A.2 tabulates some of the dis-
crete probability distributions and Table A.3 gives some of the continuous probability
distributions.
A.2.1 Bernoulli trials
If a discrete random variable Y may only assume two values 0 or 1 with probabilities
p and q = 1 − p, respectively, such a random variable is known as a Bernoulli random
variable and each sample of the experiments is known as a Bernoulli trial.
A.2 Probability 207
Table A.3. Continuous probability distributions

Random Probability density Mean Variance Range of Parameter
variable name function variable values

1 (x − μ )2
Normal 1/2 exp − μ σ2 −∞ < x < +∞ −∞ < μ < +∞
σ (2π ) 2σ 2
σ >0
Exponential λ e−λ x 1 1 0<x<∞ λ >0

λ λ2
Erlang λ κ xκ −1 e−λ x κ κ 0<x<∞ λ >0

(κ − 1)! λ λ2
integer κ > 0
Gamma λ n xn−1 e−λ x n n 0<x<∞ λ,n > 0

Γ(n) λ λ2
1 α +b (b − α )2
Uniform b−α 2 12 α <x<b −∞ < α , b < +∞
Clearly, EY = p and VarY = p − p2 = p(1 − p) = pq and c.v.2 (Y ) = q/p.

If there are n Bernoulli trials and X is the discrete random variable with values
x = 0, 1, . . . , n, then the probability mass function of X which is the sum of the values
obtained over the n Bernoulli trials is given by:

n
p(x) = px (1 − p)n−x , x = 0, 1, . . . , n
x
where
n n!
= .
x x!(n − x)!
This distribution is known as the Binomial distribution.
q
EX = np; VarX = npq; c.v.2 (X ) = .
np
Distributions (probability mass functions) of random variables related to
the binomial distribution are the geometric distribution and the negative binomial
distribution.
The geometric distribution is the distribution of the discrete random variable Y
which is the number of the Bernoulli trials which occur up to and including the trial
at which the random variable equals 1 for the first time. The mean and the variance of
the geometric distribution are given in Table A.2 with associated squared coefficient
of variation c.v.2 (Y ) = 1 − p = q.
The binomial and geometric distributions are two-parameter distributions in that
two parameters n and p are required to specify completely the distributions. A ran-
dom variable R follows the negative binomial distribution, if R is the number of the
Bernoulli trials which occur until the rth random variable equals 1 (r ≥ 1). The reader
should note that the geometric distribution is a special case of the negative binomial
distribution when r = 1. The negative binomial distribution is a three-parameter dis-
tribution. Details of the functional form, mean and variance are given in standard
probability textbooks. For completeness, the reader might note that the hypergeo-
metric distribution, a discrete distribution, often associated with sampling of lots of
finite size, is a three-parameter distribution.
The Poisson distribution which may be derived analytically, under given hypothe-
ses, has the following probability mass function:
λx
p(x) = e−λ , x = 0, 1, 2, . . .
x!
where X is a discrete random variable with values x = 0, 1, 2, . . . It is easy to show that
EX = λ , VarX = λ and c.v.(X ) = 1. The Poisson distribution is often used to specify
the distribution of arrivals in queueing systems. The expected number of arrivals λ is
generally known as the arrival rate per unit time. In queueing theory if λ is the mean
arrival rate, the following Poisson distribution as a function of time is generally used:
(λ t)x
p(x,t) = e−λ t , x = 0, 1, 2, . . .
x!
where p(x,t) is the probability of x arrivals in time t. In the derivation of the Poisson
distribution, an assumption is made that the number of arrivals in time element δ t is
λ δ t, λ a constant.
Perhaps the most important continuous probability density function (p.d.f.) is the
normal distribution. This well-known distribution is used in a variety of applica-
tions in all disciplines. It is a two-parameter distribution with the mean and variance
appearing explicitly in its functional form. The distribution has a number of very sig-
nificant properties including, for example, that the sum (convolution) of two or more
normally distributed random variables is itself normally distributed. Under certain
very broad conditions, the sum of a large number of independent and arbitrarily dis-
tributed random variables may be shown to be approximately normally distributed
(generalized central limit theorem). Further details about the normal distribution are
available in most standard probability textbooks. An associated distribution of the
normal distribution is the so-called lognormal distribution which is the distribution
of a random variable whose natural logarithm follows a normal distribution. In con-
trast to the normal distribution, the lognormal distribution has applications where the
product of nonnegative random variables is involved.
The exponential distribution of a continuous random variable T has the following
functional form:
f (t) = λ e−λ t , t ≥ 0.
A.2 Probability 209
The mean and variance of this distribution are, respectively:
∞
1
ET = t f (t)dt =
0 λ
2
2 1 1
Var T = ET − (ET ) = 2 −
2 2
= 2
λ λ λ
and the coefficient of variation is
c.v.(T ) = 1.
There is an interesting relationship between the discrete Poisson distribution and

the continuous exponential distribution as shown below.
Given a Poisson process
(λ t)n
p(n,t) = e−λ t , n = 0, 1, 2, . . . , ∞, t ≥ 0
n!
let Ω be a random variable of the time between the events, then f (ω )Δω , the prob-
ability of no event up to time ω and one event between ω and ω + Δω , is given
by

(λ ω )0 −λ Δω (λ Δω )1
f (ω )Δω = e−λ ω e
0! 1!
= λ Δω e−λ ω to order Δω .
Therefore, f (ω ) = λ e−λ ω , ω ≥ 0, is the pdf of Ω. Thus, in a Poisson process the

time between events follows an exponential distrribution with the same parameter λ .
The converse is also true. To summarize, if arrivals into a process occur according
to a Poisson distribution with arrival rate λ (mean value λ ), the inter-arrival time is
distributed according to an exponential distribution with mean value 1/λ .
A.2.2 Memoryless property of the exponential distribution
Random variables distributed according to an exponential distribution exhibit an

interesting property known as the “memoryless” property.
Given T , a continuous random variable distributed according to an exponential
distribution, i.e., with a pdf
f (t) = λ e−λ t , t ≥ 0
the cumulative distribution function (CDF), F(t) of T is defined by
F(t) = Prob(T ≤ t) = 1 − e−λ t , t ≥ 0.

Assume T > s, a given value, determine the pdf of T given T ≥ s.

Let A be the event T ≥ s and B be the event T being between t1 and t1 + Δt1 ;
t1 > s. Then from a well-known result from the axioms of probability theory:
P(B ∩ A)
Prob(B|A) =
P(A)
where ∩ is the intersection operator and since event B ∩ A = event B,
f (t1 ) Δt1
Prob(B|A) = = λ e−λ (t1 −s) Δt1 , t1 ≥ s.
e− λ s
In other words, if T is known to be greater than or equal to s, the pdf of T |T ≥ s
is an exponential distribution, starting out at t = s, with parameter λ . Thus, if the
time between successive occurrences of an event is exponentially distributed, then
the time to the occurrence of the next event does not depend on how long ago the
previous event occured. The time to the next occurrence of an event at any point in
time under these circumstances is distributed according to an exponential distribution
starting out at this point in time.
A.2.3 Relationship between the exponential distribution and the Poisson

distribution
The reader may note the important relationship between the exponential distribution
and the Poisson distribution: When the arrivals to a service system in a time interval
(0,t] follow the Poisson distribution with mean arrival rate λ units per unit time then
the inter-arrival times between any two successive arrivals follow the exponential
distribution with the same parameter, i.e., with mean inter-arrival time 1/λ .
A convolution of κ independent exponentially distributed random variables is
distributed according to an Erlang distribution with functional form as given in Table
A.3. The squared coefficient of variation of an Erlang distribution with κ exponential
phases is given by
κ
1
c.v.2 (X ) = κλ 2 =
2
(λ ) κ
so the c.v. of an Erlang distribution with κ > 1 is always less than 1.
A generalization of the Erlang distribution is the so-called Gamma distribution
with its functional form given in Table A.3, where Γ(n) is the gamma function
given by ∞
Γ(n) = xn−1 e−x dx
0
where in general n is a positive real number. When n is integer, Γ(n) = (n − 1)! and
in this case the gamma distribution reduces to the Erlang distribution. In general,
it may be shown that Γ(n) = (n − 1)Γ(n − 1) and Γ(1) = 1. Tables of the Gamma
function are available over the interval (0, 1).
The c.v.2 of the gamma distribution is 1/n and as n > 0, the c.v. of a gamma
distribution may be any value greater than 0.
A.2 Probability 211
The reader will note that the Gamma and Erlang distributions like the normal
and lognormal distributions are two-parameter distributions whereas the exponential
distribution is a one-parameter distribution.
A.2.4 The Coxian distribution with two phases
The random variable Ω is said to be distributed according to the Coxian distribution

with two phases, denoted by C2 , when it is equal to X with probability d1 and equal
to X +Y with probability d2 , where X and Y are both random variables exponentially
distributed with parameters μ1 and μ2 , respectively, i.e.,

X, with probability d1 = 1 − d
Ω=
X + Y, with probability d2 = d.
where d1 + d2 = 1, with the d being the branching probability. Without loss of

generality, it may be assumed that μ1 > μ2 .
The probability density function of C2 is obtained by considering the two ways
in which the random variable Ω will be in the interval t to t + Δt:
μ1 μ2
fΩ (t) = d1 μ1 e−μ1t + d2 (e−μ1t − e−μ2t ), t ≥ 0
μ2 − μ1
as the probability density function of the convolution of exponentially distributed
random variables X and Y with parameters μ1 and μ2 , respectively, is given by
μ1 μ2 −μ1t
fX+Y (t) = e − e−μ2t , t ≥ 0.
μ2 − μ1
Another more general form of this probability density function is
fΩ (t) = β1 μ1 e−μ1t + β2 μ2 e−μ2t , t ≥ 0
where the coefficients β1 and β2 are given by

d μ2 d μ1
β1 = 1 − d − = 1− ,
μ1 − μ2 μ1 − μ2
d μ1
β2 = = 1 − β1 .
μ1 − μ2
The cumulative distribution function of C2 is
d2
FΩ (t) = 1 − d1e−μ1t − (μ2 e−μ1t − μ1 e−μ2t ), t ≥ 0
μ2 − μ1
d 2 μ1
= 1 − e−μ1t − (e−μ1t − e−μ2t ), t ≥ 0
μ2 − μ1
and in a more general form, this may be written as follows
FΩ (t) = 1 − β1e−μ1t − β2 e−μ2t , t ≥ 0.

The variance of this distribution is

1 d2 d1 d2
Var [Ω] = + 2+ 2 .
μ1 μ2
2 μ2
If d1 = 1 or (d2 = 1 and μ1 = μ2 ), the exponential or Erlang distribution with two

phases of service, denoted by E2 , are obtained, respectively.
The first three moments, ψ1 , ψ2 , ψ3 , of the Coxian distribution with two phases
of service, C2 , are given by (see Papadopoulos et al., 1993):
1 d2
ψ1 = E[Ω] = + (A.1)
μ1 μ2
2 2d2(μ1 + μ2 )
ψ2 = 2 + (A.2)
μ1 μ1 μ22
6 6d2(μ12 + μ1 μ2 + μ22 )
ψ3 = + . (A.3)
μ13 μ12 μ23
The squared coefficient of variation is
μ22 + d2(1 + d1)μ12

c.v.2 (Ω) = .
(μ2 + d2 μ1 )2
The C2 distribution may be used to approximate any general distribution by

matching their first three moments. For a general distribution with its first three
momentsknown, ψ1 , ψ2 and ψ3 , the Coxian-2 parameters may be derived from
equations A.1, A.2 and A.3. If μ1 + μ2 = ν and μ1 μ2 = ξ , then:
6 ψ1 ψ2 − 2 ψ3
ν= (A.4)
3ψ22 − 2ψ1ψ3
12ψ12 − 6ψ2
ξ = (A.5)
3ψ22 − 2ψ1ψ3
1( ) 1( )
μ1 = ν + ν 2 − 4 ξ , μ2 = ν − ν 2 − 4ξ (A.6)
2 2
−1
d2 = μ1 μ2 (ψ1 μ1 − 1), μ1 > 0. (A.7)
For the Coxian-2 parameters to have real values, the coefficient of variation of
the general distribution must be greater than or equal to 1 and also the condi-
tion ψ3 /ψ13 > 1.5(c.v.2 + 1)2 must be satisfied. Although the C2 distribution has
three independent parameters, it is ideal for a two-moment approximation of general
distributions with a squared coefficient of variation greater than or equal to 0.5.
When 0.5 ≤ c.v.2 ≤ 1, the parameters of the C2 distribution may be shown to be
1 2
μ1 = , μ2 = , d2 = 2(1 − c.v.2). (A.8)
ψ1 c.v.2 ψ1
A.2 Probability 213
µ1 µ2 µ3 µ4 µ5
µi = µ, i = 1,2,...,5
Fig. A.2. Five-stage Erlang distribution, Ek=5
A.2.5 Phase-type distributions
The virtues of having the Markovian property in queueing problems is well known.
In many cases it may be reasonable to assume Poisson arrivals but service time dis-
tributions may not be exponential. To overcome this latter difficulty, there is a rich
literature related to the so-called phase method. The Erlang distribution discussed
above may be considered as a convolution of k identical exponential distributions
and may be depicted as in Figure A.2.
The individual stages themselves may have no physical interpretation and the
concept is that the customer leaves the system having passed through all k stages.
Considerable development of the phase method approach is possible, for example,
instead of having identical exponential distributions of time spent in each stage
it would be possible to have a convolution of non-identical distributions in series
(sometimes called the generalized Erlang distribution). A further simple extension
would be to have exponential distributions in parallel with different means and ser-
vice time of a particular customer chosen at random. Such an arrangement would
lead to a hyper-exponential distribution with k parallel channels. Early work with
respect to the phase method stressed the advantage of being able to model distri-
butions with squared coefficients of variation other than 1 (which is the pure single
exponential case). For example, the coefficient of variation of Erlang is between
0 and 1((0,1]) and for the hyper-exponential distribution the squared coefficient of
variation is greater than 1 provided that all the μi ’s are not equal. However, mod-
eling distributions just to match the squared coefficient of variation is a relatively
weak match as it is possible to have many different shapes of distributions with the
same squared coefficient of variation. This observation has led to further develop-
ments of the phase method where, for example, one could consider (i) a convolution
of identical or non-identical Erlang distributions in series format or (ii) a parallel
series arrangement of identical or non-identical exponential distributions. A partic-
ular phase type distribution is the Coxian distribution with two phases of service,
already described above, and it may be of interest to the reader to consider its deriva-
tion via the use of transition probability matrices in contrast to the direct derivation
used or at least implied above.
Consider the following system as depicted in Figure A.3:
A unit enters the system at the first station. After completion of service at this
station, the unit may either exit the system with probability 1 − d2 or enter the
second station with probability d2 , 0 ≤ d2 ≤ 1. Both stations have exponentially
distributed service times with parameters μ1 , μ2 , respectively. Arrivals to the system
are assumed to follow a Poisson distribution with mean arrival rate λ .
d2
µ1 µ2
Exit
1 – d2
There are 3 states as follows:
State 1: being served in station 1

State 2: being served in station 2
State 3: Exit from the system (absorbing state)
Fig. A.3. A two-phase Coxian distribution, C2
The transition probability matrix, T (Δt), from states at time t to states at time
t + Δt (to the order Δt) is given below:
1 2 3
⎛ ⎞
1 1 − μ1Δt d2 μ1 Δt (1 − d2)μ1 Δt
2⎝ 0 1 − μ2Δt μ2 Δt ⎠.
3 0 0 1
Note that state 3 is an absorbing state.

Let fΩ (t)dt = P[t ≤ t ≤ t + dt] and t is the time to exit from the system. If Π(t) is
the probability vector giving the probabilities of being in each of the non-absorbing
states at time t (states 1 and 2, above), then it holds, in general:
d
Π(t) = Π(t)[Θ ]
dt
where

− μ1 d 2 μ1
Θ=
0 − μ2
derived from T (Δt), above. T (Δt) is a 3 × 3 matrix but here only a 2 × 2 matrix
is needed. [Θ ] is obtained from T (Δt) by deleting the third row and third column
(corresponding to the absorbing state 3), subtracting 1 from the diagonal elements
and deleting the Δt factors from all terms.
Taking Laplace transforms:
sΠ(s) − Π(0) = Π(s)[Θ ]
where Π0 is a row vector giving the probabilities of being in each non-absorbing

state (1 or 2, above) at time t = 0. Assume Π(0) = (1, 0), therefore
Π(s)[sI − [Θ ]] = Π(0)
Π(s) = Π(0)[sI − [Θ ]]−1

A.2 Probability 215
Now fΩ (t)dt = Π(t){A}dt, where {A} is a column vector giving the probability of
being absorbed from each nonabsorbing state and
.
(1 − d2)μ1
{A} =
μ2
obtained from T (Δt), above.
F(s) = Π(s){A}
= Π(0)[sI − [Θ ]]−1 {A}
where F(s) is the Laplace transform of f (t). Then,
f (t) = Π(0)[eΘ t ]{A}, t ≥ 0.
Let eΘ t = [G(t)], then

dG(t)
= [G(t)][Θ ].
dt
Taking Laplace transforms
s[G(s)] − I = [G(s)][Θ ]
[G(s)] = [sI − Θ ]−1

d2 μ1

1
s+ μ1 (s+ μ1 )(s+ μ2 )
= 1 .
0 s+ μ2
By inversion

e−μ1t − (μd2−μμ1 ) e−μ1t + (μd2−μμ1 ) e−μ2t
[G(t)] = 1 2 1 2 = e[Θ ]t .
0 e−μ2t
Note that
d 2 μ1 d 2 μ1 d 2 μ1
=− + .
(s + μ1 )(s + μ2 ) (s + μ1 )(μ1 − μ2 ) (s + μ2 )(μ1 − μ2 )
Then, given Π(0) = (1, 0),

.
[Θ ]t (1 − d2)μ1
f (t) = {10}[e ]
μ2
which on substitution reduces to

d 2 μ1 d 2 μ1
f (t) = 1 − μ1 e−μ1t + μ2 e−μ2t , t ≥ 0,
μ1 − μ2 μ1 − μ2
which may easily be converted to the expression for fΩ (t), given above.
A.3 Discrete Markov Processes (Markov Chains)

Usually, production machines may be described as residing in a discrete and iden-
tifiable state, e.g., operating satisfactorily (up state) or in repair (down state). The
analysis of the behavior of systems of such machines involves the determination
of the state of the system (each machine) at selected times (technically known as
stages). These stages or times may be represented by either discrete or continuous
variables. If discrete, the actual time between each stage may be regular or irregular.
If the states and stages of the system are both discrete and the future states of the
system are characterized by a lack of memory, i.e., the state of the system at stage
t + 1 depends only on the state of the system at stage t and not on the history of
the system up to stage t, the underlying process is described as a discrete Markov
process or equivalently as a Markov chain.
Although, in general, it is satisfactory to describe the states of machines as dis-
crete and distinct, it may be desirable to consider the time variable as either discrete
or continuous. There is a rich literature on discrete state Markov processes with dis-
crete or continuous time stages. Here, by way of an example a simple discrete state,
discrete stage Markov process (Markov chain) is considered:
A machine as shown in Figure A.4 may be in one of three states, viz., operating
satisfactorily, state 1; operating in a derated condition, state 2; and broken down,
state 3. Assume the stages of the system are discrete and regular.
0.8
0.4
0.1
0.1
0.1
2 3
0.4
0.5 0.6
Fig. A.4. State space diagram of a three-state machine system

A.3 Discrete Markov Processes (Markov Chains) 217
The system may be described by the following transition probability matrix T :
1 2 3
⎛ ⎞
1 0.8 0.1 0.1
T = 2 ⎝ 0.4 0.5 0.1 ⎠ = [pi j ].
3 0 0.4 0.6
Note that the sum of the elements in each row of T add to 1 and all elements are
probability masses, where 0 ≤ pi j ≤ 1, i, j = 1, 2, 3.
If the system is in state 2 at stage t, for example, it will be in state 1 at stage t + 1
with probability 0.4, it will remain in state 2 with probability 0.5, while it will move
to state 3 with probability 0.1.
An interesting question is to track the behavior of the machine system over time.
For example, if the system starts out in state 2, where will it be after three stages
(periods)? Answering such questions involves matrix multiplication as follows:
Initial state of system:
{P} = {P10 , P20 , P30 }
as a probability vector. After one stage, the state of the system is:
{P1 } = {P0 } T.
After n stages, the state of the system is
{Pn } = {P0 } [T ]n .
So, for example, the state of the system which if it starts out in state 2 initially will
be as follows after three stages:
{P0 } = {0, 1, 0}
⎡ ⎤
0.612 0.213 0.175
[T ]3 = ⎣ 0.548 0.277 0.175 ⎦ .
0.304 0.396 0.300
The state of the system at stage 3 will be:
{P3 } = {P0 } [T ]3 = {0.548, 0.277, 0.175}.
Note {P3 } is a probability vector.

Some transition probability matrices have the following property:
lim {Pn } = {P} exists,

n→∞
such that
{P} = {P} [T ],
where {P} is termed the steady state probability vector.
Continuing the example:
{P} = {P1, P2 , P3 }
and {P} = {P} [T ] leads to the following three simultaneous, but not independent,
linear equations:
P1 = 0.8P1 + 0.4P2
P2 = 0.1P1 + 0.5P2 + 0.4P3
P3 = 0.1P1 + 0.1P2 + 0.6P3.
To obtain the values of P1 , P2 and P3 it is necessary to include the normalizing

condition:
P1 + P2 + P3 = 1.
The solution to the four equations, above, involving P1 , P2 and P3 is
.
8 4 3
{P} = {P1 , P2 , P3 } = , , .
15 15 15
An interpretation of this result is that in steady state or in the long term, in every 15
stages the machine is operating satisfactorily in 8 of these stages, is operating in the
derated state in 4 of these stages, while it is broken down in 3 of these stages.
In Markov chains, a state is said to be an absorbing state if it has the property that
once the system enters the absorbing state it remains in that state for all subsequent
stages of the system. The concept of the absorbing state is very useful in determining
the mean number of periods a system remains outside a particular state.
To illustrate the use of the concept of absorbing states, consider the above exam-
ple machine system with a view to determining the mean (expected) number of stages
to the first breakdown of the system (state 3) starting out in any of the other two states
(state 1 or state 2, now called “transient” states).
Define [T ∗ ], obtained from [T ] as follows:
⎡ ⎤
0.8 0.1 0.1
T ∗ = ⎣ 0.4 0.5 0.1 ⎦ .
0 0 1
Note that in T ∗ , state 3 is an absorbing state and that T ∗ may be partitioned as

follows:
∗ AB
T =
0 I
where A, B, 0 and I have dimensions 2 × 2, 2 × 1, 1 × 2 and 1 × 1, respectively.
The matrix 0 has all elements equal to 0 while I is an identity matrix. This general
structure of T ∗ will hold even if there are more than one absorbing states and the
dimension of T ∗ is greater than 3 × 3.
A.4 Data Plotting 219
In the following analysis a matrix F, the fundamental matrix, plays a central role:
F = [I − A]−1.
A as noted above represents the transient states prior to being absorbed.
F = I + A + A2 + · · ·
gives the expected number of stages the system stays in each transient (non-
absorbing) state before the system reaches the absorbing state(s). This may be shown
from a consideration of the structure of [T ∗ ], [T ∗ ]2 , [T ∗ ]3 , . . . , [T ∗ ]n as follows:

AB
[T ∗ ] =
0 I
2
[A] AB + B
[T ∗ ]2 =
0 I
3 2
[A] A B + AB + B
[T ∗ ]3 =
0 I
..
. n n−1
∗ n [A] [A + · · · + A + I] B
[T ] = .
0 I
In the above example

8.33 1.67
F = [I − A]−1 = .
6.67 3.33
So, starting out in state 1, the system would on average spend 8.33 stages in state
1 and 1.67 stages in state 2 or a total of 10.0 stages before failure for the first time.
Thus, the mean time to failure (complete breakdown) of the system starting out in
state 1 is 10 stages.
The above result could be obtained in a more efficient manner as follows:
Let EPi , i = 1, 2 be the expected number of stages before reaching state 3 having
set out in state i, i = 1, 2, respectively,
EP1 = 1 + 0.8EP1 + 0.1EP2

EP2 = 1 + 0.4EP1 + 0.5EP2
from which EP1 = 10; EP2 = 10.
A.4 Data Plotting

In practice, it may be necessary to assess if a particular data set comes from a selected
underlying distribution. For example, the inter-arrival times of a process have been
obtained from an actual observation of the process. The issue is how one would test
if it was reasonable to assume that the underlying distribution would be, for example,
exponential. Basically, there are two different approaches. One is to plot the data on
appropriate probability paper and assess usually by eye if the data is close enough to
be on a straight line. Probability paper exists for a number of distributions including
normal, exponential and gamma. An issue arises in plotting experimental data as
to how to transfer the data on to the plotting coordinates. There are a number of
different procedures in existence and the reader is referred to Shapiro (1980) and
Montgomery (1996), among others. One of the benefits of using probability plots is
that the parameters of the distribution in question may be read off the plot or derived
by simple calculations. The other approach is to use statistical methods such as the
chi-squared or the Kolmogorov/Smirnov goodness of fit tests. In goodness of fit tests
the null hypothesis is that the candidate distribution is correct. Hence, there is a
strong bias in favor of whatever distribution is chosen. For this reason the modeler
testing physical data should have in mind appropriate distributions to test. In some
work, it might be more appropriate to use the data and to estimate moments of the
underlying distribution and from these moments to develop appropriate parameters
for selected distributions as discussed earlier in relation to the Coxian distribution.
A.5 Well-Known Results of Queueing Theory
Here important results for single-station queueing systems are given.

Kendall’s notation for single station queues, A/B/C : D/E/F, where A is a
descriptor of the statistics of the arrival into the single-station system, B is a descrip-
tor of the service time distribution of each of the servers in the system, C refers to the
number of servers, D is a descriptor of the queueing discipline, i.e., how the arriving
units are called into the service, E is a specification of the overall size of the system
which is a limit to the number of those waiting for service plus those being served
and F is a descriptor of the population from which the arriving units come, speci-
fies fully any single-station queueing system. In queueing theory the terms units and
customers are used interchangeably and there is a basic assumption that customers
are not allowed to wait if there is a free server available. The queueing discipline
affects the waiting time of classes of customers arriving into the system. There are
two types of solutions to queueing systems, one called time-dependent and the other
steady-state. The steady-state solution is a limit as time goes to infinity of the time-
dependent solution, which of course is the full solution. Many of the analytical results
of queueing theory are derived on the assumption that the arrival process is a Poisson
process and that the distributions of service times of servers are independent identi-
cally distributed (i.i.d.) exponential distribution. Because of the relationship between
the Poisson and exponential distributions and the well-known Markov process the-
ory, it is normal to let A = M and B = M if the arrival process is Poisson and the
service time distribution is exponential.
Table A.4 illustrates the types of characteristics of a single-station queueing
system that have been developed analytically.
A.5 Well-Known Results of Queueing Theory 221
Table A.4. Characteristics of single-station queueing systems
Description Notation
Probability that there are n units in the system at time t Pn (t)
Probability that there are n units in the system in steady state Pn
The expected number of units in the system in steady state LS
The expected number of units waiting for service in steady state Lq
The expected time spent in the system WS
The expected time spent waiting for service Wq
Probability density function (pdf) of the total time spent in the system under a fWS (t)
given queueing discipline
pdf of the time spent waiting for service under a given queueing discipline fWq (t)
A.5.1 M/M/1: First-Come First-Served (FCFS)/∞/∞ queue
(
Pn (t) = e−(λ +μ )t ρ (n−i)/2In−i (2 λ μ t) + ρ (n−i−1)/2In+i+1 (2 λ μ t)

∞
+ (1 − ρ )ρ n
∑ ρ I j (2 λ μ t) , (n ≥ 1),
− j/2
j=n+i+2
where ρ = λ /μ , λ is the mean arrival rate and μ is the mean service rate, and Ik (x)
is the modified Bessel function of first kind and kth order.
For ρ < 1:
Pn = (1 − ρ )ρ n
ρ
LS =
1−ρ
ρ2
Lq =
1−ρ
ρ /μ
Wq =
1−ρ
1/ μ
WS =
1−ρ
fWS (t) = μ (1 − ρ )e−μ (1−ρ )t , t ≥ 0
fWq (t) = (1 − ρ )δ (t) + λ (1 − ρ )e−μ (1−ρ )t , t ≥ 0
where δ (t) is the unit impulse function occurring at time t = 0 called the Dirac delta
function.
A.5.2 M/M/1: FCFS/N/∞ queue

N+1 ,
1
ρ =1 λ
Pn = 1−ρ , 0 ≤ n ≤ N, ρ = ,
1−ρ N+1
ρ , ρ = 1
n
μ

ρ 1 − (N + 1)ρ N + N ρ N+1
LS = , ρ = 1,
(1 − ρ N+1)(1 − ρ )
N
LS = , ρ = 1,
2
ρ (1 − ρ N )
Lq = LS − , ρ = 1,
1 − ρ N+1
N N N−1
Lq = LS − = , ρ = 1,
N+1 2 N+1
1
WS = LS ,
λ (1 − PN )
1
Wq = Lq .
λ (1 − PN )
A.5.3 M/M/c: FCFS/∞/∞ queue
⎧
⎨ (cρ )n P0 , n ≤ c
n!
Pn =
⎩ cc ρ n P , n ≥ c,
c! 0
where ρ = λ /cμ < 1 is the utilization factor and

−1
c−1
(ρ c) j (ρ c)c 1
P0 = ∑ j! + c! 1−ρ
.
j=0
(ρ c)c ρ
Lq = P0 ,
c! (1 − ρ )2
LS = Lq + ρ c,
.
(ρ c)c
fWq (t) = 1 − δ (t)+
c!(1 − ρ )
(cμ − λ )(ρ c)c
+ P0 e−(cμ −λ )t , t ≥ 0,
c!(1 − ρ )
(ρ c)c
Wq = P0 ,
c!cμ (1 − ρ )2
1 −μt
fWS (t) = μ e (λ − cμ + μ A)
λ − (c − 1)μ
)
− (1 − A)(λ − cμ )μ e−(cμ −λ )t , t ≥ 0,
(ρ c)c
A = 1− P0 ,
c!(1 − ρ )
1
WS = Wq + .
μ
A.5.4 M/M/c: FCFS/N/∞ queue

⎧
⎨ (cρ ) P0 , 1 ≤ n ≤ c − 1
n
n!
Pn =
⎩ cc (ρ )n P , c ≤ n ≤ N,
c! 0
where ρ = λ /cμ < 1 and

−1
c−1
(ρ c) j cc N
P0 = ∑ j! + c! ∑ρj .
j=0 j=c
(ρ c)c ρ
P 1 − ρ N−c+1 − (1 − ρ )(N − c + 1) ρ N−c ,
(1−ρ )2 0
ρ = 1;
Lq = c!
(N−c)(N−c+1)
cc
c! P0 2 , ρ = 1,
LS = Lq + ρ c (1 − pN ),
ρ (1 − PN ) = utilization factor,
λe = 1 − PN (the mean effective arrival rate)
1
Wq = Lq ,
λ (1 − PN )
1
WS = LS .
λ (1 − PN )
A.5.5 M/M/c: FCFS/c/∞ queue
(ρ c)n
Pn = n!
, ρ = λ /cμ , n = 0, 1, . . . , c,
c (ρ c)i
∑i=0 i!
(ρ c)c
Pc = c!
, n = c,
c (ρ c)i
∑i=0 i!
which is the well-known Erlang’s loss formula and corresponds to the probability of
a full system in the steady state.
A.5.6 M/M/∞ queue
This is the self-service queueing system where the number of servers is equal to the
number of units in the system.
1 n −ρ λ
Pn = ρ e , ρ=
n! μ
LS = ρ
Lq = 0
Wq = 0
1
WS = .
μ
The formulae above hold for the M/G/∞ queue, where G is a general service time
distribution too, except of course the last one, which applies to the exponential
service time distribution.
A.5.7 M/M/c: FCFS/K/K–The finite source queue
In this model, the population from which the arrivals come is finite, say of size K.
Classically known as the machine interference problem, it is concerned with the
modeling of the repair of a bank of machines of size K with c ≤ K repair persons
⎧
⎪
⎪ K λ n
⎨ n ( μ ) P0 ,
⎪ 0≤n<c
Pn =
⎪
⎪ K n!cc λ n
⎪
⎩
n c! ( cμ ) P0 , c ≤ n ≤ K,
where,
c−1 ν K ν −1
K λ K ν !cc λ
P0 = ∑ ν μ
+∑
ν c! cμ
ν =0 ν =c
and

K K!
= . (A.9)
ν ν !(K − ν )!
Using the following notation:

• LS equals the average number of machines ‘down’ in the system,
• Lq equals the average number of machines ‘down’ that are waiting in the queue to
be repaired,
• Ld,r equals the average number of machines ‘down’ that are under repair, and
• Lo equals the average number of machines that are operating,
then these measures, all relating to steady-state conditions, are given by the following
expressions:

μ c−1
LS = K − c − ∑ (c − ν ) Pν
λ ν =0

μ c−1
Lq = K − (1 + ) c − ∑ (c − ν ) Pν
λ ν =0
c−1
Ld,r = LS − Lq = c − ∑ (c − ν ) Pν
ν =0

μ c−1
Lo = K − LS = c − ∑ (c − ν ) Pν
λ ν =0
λ (K − LS ) λ Lo
= = utilization factor.
cμ cμ
The total service capacity of the system is cμ and the mean effective arrival rate is
λ Lo = λ (K − LS ). Little’s formula may be used to determine the average time spent
in the system, WS , and the average time spent waiting for repair, Wq .
Little’s formulae, LS = λeWS and Lq = λeWq , where λe is the mean effective
arrival rate into the system, which differs from λ if there are constraints preventing
an arrival unit entering the system, e.g., if there is a maximum size N of the system.
This formula may be shown to apply to most queueing systems that are organized so
that a server is never idle if there is a unit waiting for service. Generally speaking,
Little’s formulae may be used in the analysis of manufacturing systems.
Analytical results are also available for queues where the arrivals and/or service
times do not follow a Poisson distribution and an exponential distribution, respec-
tively. The reader is referred to the specialists’ textbooks on queueing theory for
further information.
A.5.8 Queueing networks
A queueing network is a network of service stations each of which has at least one
server and with storage capacity of finite size greater than or equal to zero between
the service stations. This storage capacity is generally referred to as inter-station
buffer capacity and its function is to allow queues to form before the associated
service stations. Units in general may enter the queueing system at any particular
station, if necessary wait for service, leave that station after service and go through
the network along a route which may not be the same for other arriving units. It is
possible for units to return to stations at which they were previously served, to leave
the network at some point or to remain in the network indefinitely. Fundamental work
on queueing networks was undertaken by Erlang and Jackson.
In queueing networks with K stations there are two fundamental concepts,
namely, the probability of arriving from outside the network to station i, i = 1,
2, . . . , K, and the probability that a unit which has completed its service at station
i will go immediately to station j. The former probability, in the earlier work was
assumed to follow a Poisson process with mean rate λi,e , i = 1, 2, . . . , K, and the lat-
ter, the routing probability, is given by qi, j , i, j = 0, . . . , K, where K is the total number
of stations in the network and i or j = 0 indicates arrival from outside or departure
from the system, respectively. Open networks are those networks which have contact
with the outside, i.e., λi,e = 0 for all i. Closed networks, on the other hand, have no
contact with the outside and so λi,e = 0 for all i and qi,0 = 0 for all i, i.e., no units
are allowed to enter from outside the system or to leave the system. A cyclical queue
is a closed queueing network in which the units follow a path from station i through
the network and back to station i and they repeat this route ad infinitum. Series or
tandem queues are open networks with arrivals only at the first station, processing of
all items in a specific order and exit from the system from the last station with some
probability to return to the first station via a process called feedback. Once at the first
station, the unit goes through all stations in turn again.
Jackson devised results for open queueing networks where there is in effect no
constraint in relation to the flow of units arising out of a shortage of inter-station
buffer space. A sufficient condition for such a netwok would be that there is an
infinite buffer capacity between stations. For such open Jackson networks, Jackson
derived well-known product-form solutions for the joint probabilities. In relation
to closed Jackson networks, Gordon and Newell derived a formula for the joint
probabilities and Buzen provided an efficient way of determining the normalization
coefficient.
Example 1
A two-station series model with capacity of the intermediate buffer equal to 1.

Consider the system depicted in Figure A.5. Items arrive at the first station
according to a Poisson distribution with mean arrival rate equal to λ units per unit
time. Items get service first at the first station and then move on to the second station
via the intermediate buffer. The general rule is that an item is served if a server is free
to give service. It should be noted that station 2 is never blocked and station 1 can
be blocked when an item has finished its service at the first station, the intermediate
Entry M1 B2 = 1 M2 Exit
µ1 = µ2 = µ is the mean service rate at each of the stations
Fig. A.5. A two-station series queueing network with two identical exponential stations and
an intermediate buffer of capacity 1
Table A.5. The transition matrix of the queueing network model of example 1
(0, 0, 0) (0, 0, 1) (0, 1, 1) (b, 1, 1) (1, 0, 0) (1, 0, 1) (1, 1, 1)
(0, 0, 0) 1 − λ Δt 0 0 0 λ Δt 0 0
(0, 0, 1) μ Δt 1 − μ Δt − λ Δt 0 0 0 λ Δt 0
(0, 1, 1) 0 μ Δt 1 − μ Δt − λ Δt 0 0 0 λ Δt
(b, 1, 1) 0 0 μ Δt 1 − μ Δt 0 0 0
(1, 0, 0) 0 μ Δt 0 0 1 − μ Δt 0 0
(1, 0, 1) 0 0 μ Δt 0 μ Δt 1 − 2μ Δt 0
(1, 1, 1) 0 0 0 μ Δt 0 μ Δt 1 − 2μ Δt
buffer is full and the second station is busy. Both stations are assumed to be perfectly
reliable and service times at the stations are identical and exponentially distributed,
i.e., the average service rates are the same, viz., μ1 = μ2 = μ . There is no waiting
space in front of the first station and so, during the period the first station is blocked,
all incoming items are lost to the system.
The possible states of this system are labeled (a, B, c), where a represents the
state of the first station, B represents the state of the buffer and c represents the state
of the second station. a may have three values: 0, 1 and b, where 0 indicates that the
station is free, 1 indicates that the station is busy and b indicates that the station is
blocked. Likewise, c can take two values: 0 and 1, whereas B may take two values,
indicating the number of units in the buffer, i.e., 0 or 1. There are 7 feasible states
as shown in the transition matrix given in Table A.5. The transition probabilities
are obtained noting that a service will be completed in time interval (t,t + Δt) with
probability μ Δt and an arrival will enter the system if station 1 is not blocked in time
(t,t + Δt) with probability λ Δt. Using the usual assumptions of arrival and service
completion, the transition matrix given in Table A.5 from states at time t to states at
time t + Δt (to the order Δt) may be determined.
The steady-state probabilities of the 7 states of the system may be derived from
the solution of a system of 7 linear simultaneous equations, the following:
P000 (t + Δt) = P000 (t)[1 − λ Δt] + P001(t)μ Δt

P001 (t + Δt) = P001 (t)[1 − (μ + λ )Δt] + P011(t)μ Δt + P100(t)μ Δt
P011 (t + Δt) = P011 (t)[1 − (μ + λ )Δt] + Pb11(t)μ Δt + P101(t)μ Δt
Pb11 (t + Δt) = Pb11 (t)[1 − μ Δt] + P111(t)μ Δt
P100 (t + Δt) = P000 (t)λ Δt + P100(t)[1 − μ Δt] + P101(t)μ Δt
P101 (t + Δt) = P001 (t)λ Δt + P101(t)[1 − 2 μ Δt] + P111(t)μ Δt
P111 (t + Δt) = P011 (t)λ Δt + P111(t)[1 − 2 μ Δt].
These equations lead to the following set of steady-state equations:
λ P000 = μ P001
(λ + μ )P001 = μ P011 + μ P100
(λ + μ )P011 = μ Pb11 + μ P101
μ Pb11 = μ P111
μ P100 = λ P000 + μ P101

2μ P101 = λ P001 + μ P111
2μ P111 = λ P011.
By using the boundary equation that the sum of the probabilities of all 7 states
equals one, viz.,
∑ PaBc = 1
∀a,B,c
in conjunction with any six out of the above 7 equations, one may derive:
P000 = (4α 3 + α 2 )A
P001 = (4α 2 + α )A
P011 = (2α )A
P100 = (4α 2 + 3α + 1)A
Pb11 = A
P101 = (1 + 2α )A
P111 = A
∑ PaBc = [4 + 8α + 9α 2 + 4α 3 ]A = 1
∀a,B,c
where
A = [4 + 8α + 9α 2 + 4α 3 ]−1
and
μ
α=
.
λ
The throughput of the system, XK , may be calculated as follows:
XK = μ P[the last station is busy]

= μ [P001 + P011 + Pb11 + P101 + P111]
= μ [(4α 2 + α )A + (2α )A + A + (1 + 2α )A + A]
= μ A [4α 2 + 5α + 3].
The average queue length of the system, LS , may be derived as follows:
LS = 0P000 + 1(P001 + P100) + 2(P011 + P101) + 3(Pb11 + P111)

= (4α 2 + α )A + (4α 2 + 3α + 1)A + 2(2α )A
+2(1 + 2α )A + 3A + 3A
= A[8α 2 + 12α + 9].
The mean effective arrival rate to the system, λe , is given by:
λe = λ P[the first station is idle]

= λ [P000 + P001 + P011]
= λ [(4α 3 + α 2 )A + (4α 2 + α )A + (2α )A]
= λ A[4α 3 + 5α 2 + 3α ]
= μ A[4α 2 + 5α + 3]
and the average waiting time in the system, WS , may be obtained from Little’s
formula:
LS
WS =
λe
A[8α 2 + 12α + 9]
=
μ A[4α 2 + 5α + 3]
8α 2 + 12α + 9
= .
μ [4α 2 + 5α + 3]
It may be also noticed that the average waiting time in the system consists of the
following times:
1 1 2
WS = + Wb + WB + = + Wb + WB
μ μ μ
where Wb and WB are the average blocking time and the average waiting time at the
intermediate buffer, respectively.
The average waiting time in the queue, Wq , consists of two elements, viz., the
waiting time while an item is blocked at the first station, Wb , and the waiting time at
the intermediate buffer, WB . Applying Little’s formula, one may obtain:
Lq 1(P011 + P111) + 2Pb11
Wq = Wb + WB = = .
λe λe
Example 2: A production machine subject to failure and repair
Consider a machine with a mean service (production) rate of μ units per unit time
which is subject to failure and repair at rates of β and r per unit time, respectively.
The failure and repair mechanism is assumed to follow a Markov process. There
are two states in the system, viz., 0 when the machine is operating and 1 when the
machine is down.
Let Pi (t) equal the probability of the system being in state i at time t, i = 0, 1.
Then, the equations of state of the system may easily be derived using the Markov
properties and are as follows:
dP0 (t)
+ β P0(t) = rP1 (t)
dt
dP1 (t)
+ rP1 (t) = β P0 (t).
dt
Using the normalizing condition P0 (t) + P1 (t) = 1, with the initial condition
P0 (0) = 1, leads to
r β −(β +r)t
P0 (t) = + e
β +r β +r
β
P1 (t) = [1 − e−(β +r)t ]
β +r
and the steady-state solutions are as follows:
r
P0 =
β +r
β
P1 = .
β +r
It may be noted that these steady-state results could be obtained by considering that
the machine operates for a mean time of 1/β and is down for a mean time of 1/r
during repair.
The steady-state availability, A, of the system is given by
1
MT T F β r
A= = =
MT T F + MT T R 1
β + 1
r
β +r
where MT T F is the mean time to failure and MT T R is the mean time to repair.
So, as μ is the mean service rate of the machine, the average steady-state output,
X , would be
MT T F
X =μA=μ .
MT T F + MT T R
References
2. Barnett, S. (1994), Matrices: Methods and Applications, Oxford Applied Mathematics
and Computing Science Series.
3. Bellman, R. (1960), Matrix Analysis, McGraw-Hill.
4. Billinton, R. and Allan, R.N. (1983), Reliability Evaluation of Engineering Systems:
Concepts and Techniques, Longman Scientific & Technical.
5. Brogan, W.L. (1974), Modern Control Theory, Quantum Publishers.
6. Cramer, H. (1946), Mathematical Methods of Statistics, Princeton University Press.
7. Feller, W. (1957), An Introduction to Probability Theory and Its Applications, Vol. I, 2nd
Edition, John Wiley & Sons.
8. Feller, W. (1966), An Introduction to Probability Theory and Its Applications, Vol. II, 2nd
Edition, John Wiley & Sons.
9. Gross, D. and Harris, C.M. (1985), Fundamentals of Queueing Theory, Second Edition,
John Wiley & Sons.
10. Hahn, G.J. and Shapiro, S.S. (1967), Statistical Models in Engineering, John Wiley &
Sons.
References 231
11. Horn, R.A. and Johnson, C.J. (1991), Topics in Matrix Analysis, Cambridge University
Press.
12. Kemeny, J.G. and Snell, J.L. (1960), Finite Markov Chains, Van Nostrand.
13. Kleinrock, L. (1975), Queueing Systems, Vol. I: Theory, John Wiley & Sons.
14. Meyer, P.L. (1965), An Introduction to Probability Theory and Its Applications, Addison-
Wesley.
15. Montgomery, C.D. (1996), Introduction to Statistical Quality Control, John Wiley &
Sons.
16. Mood, A.M. and Graybill, F. (1963), Introduction to the Theory of Statistics, McGraw-
Hill.
17. Neuts, M.F. (1981), Matrix-Geometric Solutions in Stochastic Models – An Algorithmic
Approach, John Hopkins University Press.
18. Neville, A.M. and Kennedy, J.B. (1964), Basic Statistical Methods for Engineers and
Scientists, International Textbook Co.
19. Papadopoulos, H.T., Heavey, C. and Browne, J. (1993), Queueing Theory in Manufactur-
ing Systems Analysis and Design, Chapman & Hall.
20. Parzen, E. (1960), Modern Probability Theory and Its Applications, John Wiley & Sons.
21. Ross, S.M. (1985), Introduction to Probability Models, John Wiley & Sons.
22. Shapiro, S.S. (1980), How to Test Normality and Other Distributional Assumptions,
Vol. 3, The ASQC Basic References in Quality Control: Statistical Techniques, ASQC,
Milwaukee, WI.
B
Algorithms/Procedures Details and Guide to Use
The following material appears at the website associated with this book.
The details of the algorithms available at the website associated with this book
are given below:
1. Abbreviation is the name by which the algorithm/procedure is named at the
web-site associated with this book
2. Author name
3. Coder name
4. The type of algorithm or procedure used, e.g., evaluative/predictive or genera-
tive/optimization
5. Description of system to which the algorithm/procedure may be applied including
size restrictions, if any
6. Output of the algorithm/procedure
7. Reference
Only one algorithm is capable of handling open loop (unsaturated) serial production
lines (EXPAN).
234 B Algorithms/Procedures Details and Guide to Use
B.1 Markovian
Abbreviation: MARKOV
Author: Cathal Heavey, University of Limerick, Ireland
Coder: Cathal Heavey
Algorithm: Evaluative/Predictive
Description: Given a detailed specification of a reliable or unreliable produc-
tion line with single machines at each station with service and
repair times distributed according to an Erlang-k (k ≥ 1) distribu-
tion and the times to failure following an exponential distribution.
Intermediate buffers of finite capacity are allowed between any
two successive stations of the saturated line. With current com-
puter capabilities the algorithm is able to handle systems with up
to 300,000 states/equations in reasonable time.
Output: Exact throughput of the specified production line
Reference: Heavey, Papadopoulos and Browne (1993)
B.2 Decomposition-1
Abbreviation: DECO-1
Author: Yves Dallery (Ecole Centrale Paris) and Yannick Frein (Institut
Polytechnique de Grenoble, France)
Coder: Michael Vidalis (University of the Aegean, Greece)
Description: The algorithm is capable of handling any size of serial single
machine station reliable saturated production lines with exponen-
tial service times and intermediate buffers of finite capacity using
the decomposition approach.
Output: Throughput of the specified production line
Reference: Dallery and Frein (1993), among other papers
B.3 Expansion
Abbreviation: EXPAN
Author: Laoucine Kerbache and James MacGregor Smith
Coder: Suchant Jain and James MacGregor Smith
Description: The algorithm is capable of handling unsaturated reliable serial
production lines with parallel machines at each station with finite
intermediate buffers using a decomposition methodology.
Output: Throughput of the specified production line
Reference: Kerbache and MacGregor Smith (1987) and Jain and MacGregor
Smith (1994)
B.5 Decomposition-2 235
B.4 Aggregation
Abbreviation: AGGRE
Author: Jonh-Tae Lim, Semyon Meerkov and Ferudun Top
Coder: Jonh-Tae Lim, Semyon Meerkov and Ferudun Top
Description: The algorithm is capable of handling asymptotically reliable satu-
rated transfer lines (with the machines having identical cycle times)
of any size using the aggregation approach and involving forward
and backward loops to obtain convergence.
Output: Throughput of the specified transfer line
Reference: Jonh-Tae Lim, Semyon Meerkov and Ferudun Top (1990)
B.5 Decomposition-2
Abbreviation: DECO-2
Author: Alexandros Diamantidis (Aristotle University of Thessaloniki,
Greece)
Coder: Alexandros Diamantidis
Description: The algorithm is capable of handling saturated long lines (with
over 1000 stations in series) with exponential service times, paral-
lel identical machines at each station and finite intermediate buffers
using a decomposition methodology.
Output: Throughput of the specified production line. Note 1: For the num-
ber of stations, K = 2, the algorithm gives the exact equations
and numerical results of the two-station production line with par-
allel machines at each station. Note 2: For the number of parallel
machines at each station, si = 1, i = 1, 2, . . . , K, the algorithm
gives the same equations and numerical results as those originally
developed by Gershwin (1987, 1994).
Reference: Diamantidis, Papadopoulos and Heavey (2007)
B.6 Two-Level Work-Load Allocation

Abbreviation: TLWLA
Author: John Buzacott and George J. Shanthikumar
Coder: Michael Vidalis and Alexandros Diamantidis
Algorithm: Stand-alone Optimization
Description: It is a self-contained algorithm which develops an approximate two-
level work-load allocation for saturated production lines with single
machine reliable stations and specified identical or non-identical
buffer sizes.
Output: Throughput and two-level work-load approximation of the specified
production line
Reference: Buzacott and Shanthikumar (1993)
B.7 Simulated Annealing

Abbreviation: SA
Author: Diomidis Spinellis (Athens University of Economics and Busi-
ness) and Chrissoleon Papadopoulos (Aristotle University of
Thessaloniki, Greece)
Coder: Diomidis Spinellis
Algorithm: Generative/Optimization
Description: It is an optimizing search algorithm based on the methodology of
simulated annealing which communicates with appropriate evalua-
tive/predictive algorithm(s) to solve large production lines.
Output: Work-load-, Buffer-, and Server-allocations, in single or double or
triple combinations
Reference: Spinellis and Papadopoulos (2000a)
B.8 Genetic Algorithm

Abbreviation: GA
Author: Diomidis Spinellis and Chrissoleon Papadopoulos
Coder: Fanis Karagiannis and Diomidis Spinellis
Description: It is an optimizing search algorithm based on the methodology
of genetic programming which communicates with appropriate
evaluative/predictive algorithm(s) to solve large production lines.
Output: Work-load-, Buffer-, and Server-allocations, in single or double or
triple combinations
Reference: Papadopoulos and Karagiannis (2001) and Spinellis and
Papadopoulos (2000b)
References 237
B.9 Complete Enumeration

Abbreviation: CE
Author: Michael Vidalis and Chrissoleon Papadopoulos
Coder: Michael Vidalis and Diomidis Spinellis
Description: It is an optimizing search algorithm based on enumeration which
communicates with appropriate evaluative/predictive algorithm(s)
to solve only small production lines with constraints with respect to
total number of buffer slots and total number of servers.
Output: Buffer- and Server-allocations, in single or double combinations
Reference: enumeration, CE—
B.10 Buffer Allocation

Abbreviation: BA
Author: Chrissoleon Papadopoulos and Michael Vidalis
Coder: Michael Vidalis and Diomidis Spinellis
Algorithm: Stand-alone optimization
Description: It is a self-contained algorithm which initially specifies a near opti-
mal buffer allocation and being directly connected to the Markovian
algorithm develops via the Hooke and Jeeves search mechanism
the optimal buffer allocation and the associated optimal throughput.
It solves small reliable or unreliable production lines.
Output: Buffer allocation and throughput of the specified production line
Reference: Papadopoulos and Vidalis (2001a)
The authors would be very pleased to hear from researchers or practitioners who
wish to have an algorithm/procedure developed by them to be included at the web-
site. Hopefully in time a very comprehensive set of algorithms/procedures for the
analysis/design of serial production lines would become available for all to use. This
could well be the first step to having at the website a set of algorithms/procedures
which have been found to be of value in design and analysis of general manufacturing
systems.
References
3. Diamantidis, A.C., Papadopoulos, C.T., and Heavey, C. (2007), Approximate analysis
of serial flow lines with multiple parallel-machine stations, IIE Transactions, Vol. 39,
issue 4, pp. 361–375.
4. Gershwin, S.B. (1987), An efficient decomposition method for the approximate evalu-
ation of tandem queues with finite storage space and blocking, Operations Research,
Vol. 35, pp. 291–305.
5. Gershwin, S.B. (1994), Manufacturing Systems Engineering, Prentice Hall.
pp. 69–89.
pp. 448–461.
9. Lim, J.-T., Meerkov, S.M., and Top, F. (1990), Homogeneous, asymptotically reliable
serial production lines: Theory and a case study, IEEE Transactions on Automatic Control,
Vol. 35, No. 5, pp. 524–534.
10. Papadopoulos, C.T. and Karagiannis, T.I. (2001), A genetic algorithm approach for
11. Papadopoulos, H.T. and Vidalis, M.I. (2001a), A heuristic algorithm for the buffer allo-
cation in unreliable unbalanced production lines, Computers & Industrial Engineering,
Vol. 41, pp. 261–277.
pp. 373–384.
13. Spinellis, D.D. and Papadopoulos, C.T. (2000b), Stochastic algorithms for buffer
allocation in reliable production lines, Mathematical Problems in Engineering,
Vol. 5, pp. 441–458.
C
Glossary
C.1 General Acronyms
Symbol Meaning
ABC ABC analysis in inventory/stock control
ABC Actvity-based costing
Arena A simulation package
CDIM Customer-driven intelligent manufacturing
DSS Decision support system
eM-plant A simulation package
PL Production line
CI Confidence interval
FIFO First-In, First-Out
FCFS First-Come, First-Served
LIFO Last-In, First-Out
FMS Flexible manufacturing system
FMC Flexible manufacturing cell
FAS Flexible assembly system
GT Group technology
CAD Computer-aided design
CAM Computer-aided manufacturing
CAE Computer-aided engineering
CNC Computer numerically controlled
NC Numerically controlled
CIM Computer-integrated manufacturing
JIT Just-In-Time
TQM Total quality management
WIP Work-In-Process or Work-In-Progress
W IP Average WIP
WF Workforce
(continued)
240 C Glossary
General Acronyms — (Continued)

Symbol Meaning
MRP Materials requirements planning
BAS Blocking after service
BBS Blocking before service
WAP Work-load allocation problem
BAP Buffer allocation problem
SAP Server allocation problem
W+S Simultaneous work-load and server allocation
W+B Simultaneous work-load and buffer allocation
S+B Simultaneous server and buffer allocation
W+S+B Simultaneous work-load and server and buffer allocation
PARTAN The steepest ascent method of parallel tangents
SOR Successive over relaxation factor or method
DP Dynamic programming
SA Simulated annealing
GA Genetic algorithms
TS Tabu search algorithm
w.r.t. With respect to
r.v. Random variable
c.v.(X) Coefficient of variation of the r.v. X
C.2 Production Lines
Symbol Meaning
K Number of stations in a production line
Bi Buffer i, i = 2, 3, . . . , K placed before station i in a K-station
production line
Bi Capacity of buffer Bi , i = 2, 3, . . . , K
N Total number of buffer slots to be allocated among the
K − 1 intermediate buffers, Bi , i = 2, 3, . . . , K of a K-station
production line
Ni Number of buffer slots allocated to buffer Bi , i = 2, 3, . . . , K
(0 ≤ Ni ≤ Bi ) in the buffer allocation problem
W Si or Mi Work-station i in a K-station line. This may be single- or
multi-machine work-station
μi Mean service or processing rate of station i
wi = 1/μi Mean service or processing time (work-load) of station i
βi Mean failure rate of station i
1/βi = MT T F mean time to failure (MTTF) of station i
ri Mean repair rate of station i
1/ri = MT T R Mean repair time or mean time to repair (MTTR) of station i
(continued)
C.3 Decomposition Approach 241
Production Lines — (Continued)

Symbol Meaning
Ai Availability of station i
XK Throughput or mean production (output) rate of a K-station
production line
1/XK Mean production time of a K-station production line
Xi Throughput or mean production (output) rate of the ith station
of a production line
ρi Utilization of the ith work-station of a production line
ei Efficiency or mean effective service rate of the ith station of
a production line (equal to μi Ai )
C.3 Decomposition Approach
Symbol Meaning
L Original production line that is decomposed in the decomposition approach
Li Sub-line i, i = 1, 2, . . . , K − 1, in the decomposition approach
Miu The part of the original line, L, upstream buffer Bi+1 . It is a pseudo work-
station for i = 2, . . . , K − 1. For i = 1 it holds: M1u = M1
d
Mi−1 The part of the original line, L, downstream buffer Bi . It is a pseudo work-
station for i = 2, . . . , K − 1. Special case: It holds: MKd = MK
wi The mean service time (work-load) of station i, i = 1, 2, . . . , K, in the
original line, L (wi = 1/μi )
wdi The sum of the mean service time and the possible blocking time at station
i in the original line, L (wdi = 1/μid )
wui−1 The sum of the mean service time at station i − 1 and the possible starvation
time of station i − 1, i = 2, . . . , K in the original line, L (wui−1 = 1/μi−1
u )
μi The mean service rate of station i, i = 1, 2, . . . , K in the original line, L

μiu The mean service (processing) rate of the upstream station of buffer
Bi+1 , i = 1, . . . , K − 1 in the decomposition approach
μid The mean service (processing) rate of the downstream station of buffer
Bi , i = 2, . . . , K in the decomposition approach
pbl
i The blocking probability of a station i
pst
i The starvation probability of a station i
pbl
i The blocking probability of sub-line Li , i = 1, . . . , K − 1
pst
i−1 The starvation probability of sub-line Li−1 , i = 2, . . . , K
XDECO Throughput or mean production (output) rate of a K-station production line
obtained from application of the decomposition method
XSIM Simulated throughput or mean production (output) rate of a K-station
production line obtained from application of a simulation package
242 C Glossary
C.4 Markovian Model
Symbol Meaning
λ Mean arrival rate
N(t) = A two-dimensional stochastic process in the context of a
[N1 (t), N2 (t)] queueing network (q.n.)
N1 (t) The number of jobs queued up in front of the first station
of the q.n.
N2 (t) The state of the sub-network of the q.n. at time t
QBD Quasi birth and death process
e A (m × 1) column-vector with all elements equal to 1
P Steady-state probability
A = A0 + A1 + A2 The conservative matrix in the stochastic process model
ni Status of buffer i in the Markovian model
si Status of station i in the Markovian model
mBK Number of states in the sub-network of a K-station line
with identical buffers, each of capacity B
mBK2 ,...,BK Number of states in the sub-network of a K-station
line with non-identical buffers, with buffer capacities
B2 , . . . ,BK
Pi The number of phases of the service time distribution of
the ith station in the Markovian model
Ri The number of phases of the repair time distribution of
the ith station in the Markovian model
mB,R
K,P The number of states in the sub-network with K-stations,
each buffer having the same capacity B, each service
time distribution having P phases and each repair time
distribution having R phases in the Markovian model
2 ,..,BK ,R1 ,R2 ,..,RK
mBK,P 1 ,P2 ,..,PK
The number of states in the sub-network of a K-station
system with buffer capacities B2 , . . . ,BK . The number
of phases of each station’s service time distribution is
equal to P1 , P2 , . . . , PK phases and the number of phases
of each station’s repair time distribution is equal to
R1 , R2 , . . . , RK in the Markovian model
C.5 Expansion Method
Symbol Meaning
h The holding node established in the expansion method
Λ External Poisson arrival rate to the network
λj Poisson arrival rate to node j
λ˜j Effective arrival rate to node j
(continued)
C.7 Design Problems 243
Expansion Method — (Continued)

Symbol Meaning
μj Exponential mean service rate at node j
μ˜ j Effective service rate at node j due to blocking
pK Blocking probability of finite queue of size K
pK Feedback blocking probability in the expansion method
j
p0 Unconditional probability that there is no unit in the service
channel at node j (either being served or being held after service)
X Throughput (mean production rate)
C.6 Aggregation Method
Symbol Meaning
Λi The loss parameter of the ith machine
qi = 1 − ε Λi , ε << 1
The probability machine i produces a part during a time slot/period
f
Λi The loss parameter of the ith machine in the forward aggregation
Λbi The loss parameter of the ith machine in the backward aggregation
f
XK The throughput of the K-machine line in the forward aggregation
XKb The throughput of the K-machine line in the backward aggregation
C.7 Design Problems
Symbol Meaning
PI(μ1 , . . . , μK , N1 , . . . , NK ) Performance index
MARKO Markovian algorithm
DECO Decomposition algorithm
EXPAN Expansion algorithm
SA Simulated annealing algorithm
GA Genetic algorithm
CE Complete enumeration
RE Reduced enumeration
OBA Optimal buffer allocation
OSA Optimal server allocation
LBAS Linear buffer allocation scheme
w := (w1 , w2 , . . . , wK ) = The mean service times
(work-load) vector in WAP
(continued)
244 C Glossary
Design Problems — (Continued)

Symbol Meaning
s := (S1 , S2 , . . . , SK ) = The servers vector in SAP
n := (N2 , N3 , . . . , NK ) = The buffers vector in BAP
g := (g2 , g3 , . . . , gK ) = The gradient vector in the gradient
method
x The floor function, denoting the largest integer less than
or equal to x
(a . . . b) An open interval containing all values from a to b
excluding the two endpoints a and b
[a . . . b] A closed interval containing all values from a to b
including the two endpoints a and b
[a . . . b) A half closed interval containing all values from a to b
including a but excluding b
(a . . . b] A half closed interval containing all values from a to b
excluding a but including b
[I] Buffer classes of first generation (I = 0, . . . , N). It con-
sists of all buffer allocations with the first element of the
buffer vector equal to I
[I, J] Buffer classes of second generation, (I = 0, . . . , N, J =
0, . . . , N + 1 − I), and so on. It consists of all buffer allo-
cations with the first two elements of the buffer vector
equal to I and J, resp.
C.8 Cost Considerations
Symbol Meaning
AHP Analytical hierarchical processes
R Selling price of a unit of the product
C Product unit cost
Ch Inventory unit holding cost
I Interest annual rate
FU Financial units
bi A net present value coefficient associated
with each buffer slot
P.W.F.∗ Present worth factor
P.W.V. Present worth value
δ (a) The Kronecker
delta function defined by:
1, if a > 0
δ (a) =
0, if a ≤ 0.
Fi , i = 1, 2, 3 Profit maximization objective functions
G j , j = 1, 2, 3 Cost minimization objective functions
C.9 Mathematical Fundamentals 245
C.9 Mathematical Fundamentals
Symbol Meaning
I The identity matrix of dimension n × n
AT The transpose matrix of dimension n × m of matrix A of
dimension m × n
&
[F(s)] The Laplace transform of the function [F(t)]
EX The mean value or expected value of the r.v. X
Var X The variance of the r.v. X
Ek Erlang distribution with k phases
C2 Coxian distribution with two phases
ψi , i = 1, 2, 3 The first three moments of a probability distribution
T The transition probability matrix in Markov chains
F = [I − A]−1 The fundamental matrix in Markov chains
QN Queueing network
q.s. Queueing system
A/B/c : D/E/F Kendall’s notation of a queueing system (q.s.), where:
A A descriptor of the statistics of the arrival into the q.s.
If A = M The arrival process is Poisson
If B = M The service time is an exponential distribution
If A = B = D Both the inter-arrival and the service time distributions
are deterministic
If A = GI General independent arrival process
If B = G General service time distribution
B A descriptor of the service time distribution of each of
the servers of the q.s.
c The number of servers of the q.s.
D A descriptor of the queueing discipline
E The overall size of the q.s.
F A descriptor of the population from which the arriving
units to the q.s. come
i.i.d. Independent identically distributed
p.d.f. Probability density function
C.D.F. Cumulative distribution function
Pn (t) Probability that there are n units in the q.s. at time t
Pn Probability that there are n units in the q.s. in steady state
LS The expected number of units in the q.s. in steady state
Lq The expected number of units waiting for service in
steady state
WS The expected time spent in the q.s.
Wq The expected time spent waiting for service
fWS (t) p.d.f. of the total time spent in the q.s. under a given
queueing discipline
(continued)
246 C Glossary
Mathematical Fundamentals — (Continued)

Symbol Meaning
fWq (t) p.d.f. of the time spent waiting for service under a given
queueing discipline
λe Mean effective arrival rate
Ik (x) The modified Bessel function of first kind and kth order
λ Mean arrival rate
μ Mean service rate
δ (t) The Dirac delta function in the q.s. M/M/1 : FCFS/∞/∞
Pc The Erlang’s loss formula in the q.s. M/M/c : FCFS/c/∞
LS := λe WS , Little’s formula
Lq := λe Wq , Little’s formula
C.10 Accompanying Algorithms and Procedures
Symbol Meaning
MARKOV Markovian algorithm
DECO-1 Decomposition algorithm for solving reliable exponential
single-machine station production lines
EXPAN Expansion algorithm
AGGRE Aggregation algorithm
DECO-2 Decomposition algorithm for solving reliable exponential
parallel-machine station production lines
TLWLA Two-level work-load allocation algorithm
SA Simulated annealing algorithm
GA Genetic algorithm
CE Complete enumeration
BA Buffer allocation procedure in unreliable production lines
D
Conference Participants: Presenters and Attendees
These are the lists of participants:Presenters and attendees at the five Hellenic
(Aegean and Ionian) International Conferences on Analyis, Design and Optimiza-
tion of Manufacturing Systems. The conferences were held in Greece in the islands
of Samos, Tinos, Tinos, Samos, and Zakynthos in 1997, 1999, 2001, 2003, and 2005,
respectively.
D.1 Conference Participants: Presenters
L /I Participant Name Affiliation
1 A LLON G. Technion, Haifa, Israel

2 A LTIOK Tayfur Rutgers University, USA
3 A SKIN Ronald University of Arizona, USA
4 Avsar Zeynep Muge Middle East Technical University, Turkey
5 A XSATER Sven Lund University, Sweden
6 BATZIAS F.A. University of Piraeus, Greece
7 B ENJAAFAR Saif University of Minnesota, USA
8 B ILALIS Nikos Tecnhical University of Crete, Greece
9 B OHORIS Giorgos University of Piraeus, Greece
10 B ROUNS Gido Eindhoven University of Technology,
The Netherlands
11 B UITENHEK R. Technical University of Twente, Holland
12 B URMAN Mitchell Analytics, Inc., USA
13 B UZACOTT John York University, Toronto, Canada
14 C ARAMANIS Michael Boston University, USA
15 C OLLEDANI Marcello Politecnico di Milano, Italy
(continued)
248 D Conference Participants: Presenters and Attendees
Conference Participants: Presenters — Continued
16 DALLERY Yves Ecole Centrale, Paris and Groupe HEC, France

17 DASKALAKI Sophia University of Patras, Greece
18 D IAMANTIDIS Alexandros
Aristotle University of Thessaloniki, Greece
19 Di M ASCOLO Maria Laboratoire d’ Automatique de Grenoble,
France
20 D OGRU Mustafa Kemal Eindhoven Technical University, The Nether-
lands
21 D OUNIAS Georgios University of the Aegean, Greece
22 D UENYAS Izak University of Michigan
23 E MIRIS Dimitrios University of Piraeus, Greece
24 FADILOGLU Murat Bilkent University
25 F URMANS Kai IFL - University of Karlsruhe, Germany
26 G EORGIADIS Patroklos Aristotle University of Thessaloniki, Greece
27 G ERAGHTY John Dublin City University, Ireland
28 G ERSHWIN Stan MIT , USA
29 G OH Thiam Hock Jason National University of Singapore
30 G RUBBSTROM Robert Department of Production Economics,
Linkoping Institute of Technology, Sweden
31 G UPTA Omprakash Nirma Institute of Management, India
32 G URKAN Gul Tilburg University, Holland
33 H AGHNEVIS Moeed University of Tehran, Iran
34 H EAVEY Cathal University of Limerick, Ireland
35 H ELBER Stephan University of Hannover, Germany
36 H ONGLER Max Oliver EPFL , Switzerland
37 H U Jian-Quiang Boston University, USA
38 I OANNIDIS Stratos University of the Aegean, Greece
39 I OANNOU Georgios Athens University of Economics and Busi-
ness, Greece
40 I SHIKURA Hiroki Osaka Gakuin University, Japan
41 JAFARI Mohsen Rutgers University, USA
42 K ARAESMEN Fikri Koc University, Turkey
43 K ERBACHE Laoucine HEC School of Management, France
44 K HODADADEGAN Yasaman University of Tehran, Iran
45 K NIKER T. Analytics Inc., USA
46 KOCK Ad Eindhoven University of Technology,
The Netherlands
47 KOUIKOGLOU Vassilis Technical University of Crete, Greece
48 KOUKOUMIALOS S. University of Thessaly, Grrece
(continued)
D.1 Conference Participants: Presenters 249
49 K RISHNAMURTHY Ananth Rensselaer Polytechnic Institute, Troy, New

York, USA
50 K UHN Heinrich Catholic University of Eichstaett, Germany
51 K YRIAKIDIS Nondas University of the Aegean, Greece
52 L AIOS Labros University of Piraeus, Greece
53 L EVANTESI Raniero Politecnico di Milano, Italy
54 L I Jingshan
55 L IBEROPOULOS Giorgos University of Thessaly, Greece
56 L OU Sheldon CSU San Marcos
57 L OUKIS E. Univeristy of the Aegean, Greece
58 M AC G REGOR S MITH James University of Massachusetts, Amherst, USA
59 M AKRIS Spilios Telcordia Technologies, USA
60 M AKROKANIS Georgios Neoset, Greece
61 M ALHAME Roland Ecole Polytechnique de Montreal and GERAD,
Canada
62 M ATTA Andrea Politecnico di Milano, Italy
63 M EERKOV Semyon University of Mishigan, USA
64 M ITTMAN Stefanos Intracom, Greece
65 M OCANU Stephane Laboratoire d’Automatique de Grenoble,
France
66 M ODESTOU Thomas BSHG , Greece
67 M OUSTAKIS Vassilis Tecnhical University of Crete, Greece
68 N ENES George Aristotle University of Thessaloniki, Greece
69 N ONAKA Youichi Hitachi, Japan
70 O’K ELLY M.E.J. National University of Ireland, Galway,
Ireland
71 PANAGIOTOU N. NTUA , Greece
72 PAPACHRISTOS Sotiris University of Ioanina, Greece
73 PAPADOPOULOS Aristotle University of Thessaloniki, Thessa-
Chrissolleon T. loniki, Greece
74 PAPPIS Kostas University of Piraeus, Greece
75 PARDALOS Panos University of Florida, USA
76 P ERROS Harry North Carolina State University, USA
77 P ETERING Matthew National University of Singapore, Singapore
78 P OLLOCK Stephen University of Michigan, USA
79 P RASTACOS Gregory Athens University of Economics and Busi-
ness, Greece
80 R AI Sudhendu Xerox Corporation, USA
81 R AVIV Tal Technion, Haifa, Israel
(continued)
82 S AWIK Tadeusz AGH University of Science and Technology,

Poland
83 S BITI Naval University Mohammed V-Agdal, Maroc
84 S ELCUK Baris Eindhoven University of Technology, The
Netherlands
85 S HANTHIKUMAR, J. George University of California, Berkeley, USA
86 S KINTZI Georgia Athens University of Economics and Busi-
ness, Greece
87 S OFIANOPOULOU Stella University of Piraeus, Greece
88 S OLOMOS Lefteris Delta Dairy S.A., Greece
89 S PINELLIS Diomidis Athens University of Economics and Busi-
ness, Greece
90 TAGARAS Georgios Aristotle University of Thessaloniki, Greece
91 TAN Baris Koc University, Turkey
92 TAN Tarkan Eindhoven University of Technology,
The Netherlands
93 TATSIOPOULOS Ilias NTUA , Greece
94 T EMPELMEIER Horst University of Cologne, Germany
95 T OLIO Tullio Politecnico di Milano, Italy
96 T RIANTAPHYLLOU Louisiana State University, USA
Evangelos
97 T SAKONAS Athanassios University of the Aegean, Greece
98 VAN DER WAL Jan Eindhoven University of Technology,
The Netherlands
99 VAN H OUTUM Gert-Jan Eindhoven University of Technology,
The Netherlands
100 VAN N YEN Pieter Eindhoven University of Technology,
The Netherlands
101 VAN O OIJEN Henny Eindhoven University of Technology,
The Netherlands
102 VAN V UUREN Marcel Eindhoven University of Technology,
The Netherlands
103 VAN W OENSEL Tom Eindhoven University of Technology,
The Netherlands
104 VANDAELE Nico University of Antwerp, Belgium
105 V IDALIS Michalis University of the Aegean, Greece
106 V LACHOS Dimitrios Aristotle University of Thessaloniki,
Greece
107 VOUROS George University of the Aegean, Greece
(continued)
D.2 Conference Participants: Attendees 251
108 W INANDS Erik Eindhoven University of Technology,

The Netherlands
109 X ANTHAS Alekos Tecnhical University of Crete, Greece
110 YAO David Columbia University, USA
111 Y ERALAN Sencer University of Florida, Gainesville, USA
112 Z AZANIS Michael Athens University of Economics and Busi-
ness, Greece
113 Z IJM W.H.M. University of Twente, The Netherlands
114 Z IKOPOULOS Christos Aristotle University of Thessaloniki, Greece
115 Z OGRAFOS Kostas Athens University of Economics, Greece
D.2 Conference Participants: Attendees
1 A DAMIS Ioannis I NTRACOM, Greece

2 A LAYALIS Thanos D ELTA, Greece
3 A NDROUTSOPOULOS K. Athens University of Economics and
Business, Greece
4 A PERGI A. University of Piraeus, Greece
5 A RAVOSSITAS Giorgos Univeristy of the Aegean, Greece
6 BALASKAS P. University of Piraeus, Greece
7 B EKOS N. University of Piraeus, Greece
8 B ELESIOTOU F. University of Piraeus, Greece
9 B LESSIOS N. University of Piraeus, Greece
10 C ARAMANI A. University of the Aegean, Greece
11 C HALIKAS N. University of Piraeus, Greece
12 C HRISTOFIDES Tony University College, Galway, Ireland
13 C HRISTOPOULOS K. University of Piraeus, Greece
14 DARZENTAS Jenny University of the Aegean, Greece
15 DARZENTAS John University of the Aegean, Greece
16 DASKALOPOULOS I NTRACOM, Greece
Konstantinos
17 D ESSOUKY Mohamed University of Southern California, USA
18 D IMOPOULOS Ilias Business Consultant, Theesaloniki, Greece
19 D IMOPOULOU Helen GSRT , Greece
20 E MERY Helen Rigel Corporation, USA
(continued)
Conference Participants: Attendees — Continued
21 F ILIPPIDIS D. Kraft Foods Hellas, Greece

22 F LESSAS George University of the Aegean, Greece
23 F OLVEN Gary Kluwer Academic Publishers, USA
24 F ONIADAKI A. University of Piraeus, Greece
25 F RANTZESKAKIS Kyriakos ELVAL , Greece
26 G ALANI V. BSHG , Greece
27 G ALIATSATOS Christos Coca Cola HBC, Greece
28 G AMBROGIANNI C. Univeristy of Piraeus, Greece
29 G ANAS G. University of Ioannina, Greece
30 G AONKAR R. National University of Singapore, Singapore
31 G EORGIADIS Efstathios BSHG , Greece
32 G ERARDOU A. BSHG , Greece
33 G ERSHWIN A LLOU Francis USA
34 G ERSHWIN J. USA
35 G IASTA E. University of Piraeus, Greece
36 G IONI, S. University of Piraeus, Greece
37 G LAROS A. BSHG , Greece
38 G LYKAS Michalis University of the Aegean, Greece
39 G OTSIAS A. University of the Aegean, Greece
40 H ADJISAVVAS Nikolaos University of the Aegean, Greece
41 I LIADIS A. Athens 2004 Olympic Games, Greece
42 K AKOS A. University of Piraeus, Greece
43 K AKOUDAKIS J. University of Piraeus, Greece
44 K ALLIANOS Theodoros European Commission
45 K APSIS George Metaxa S.A.
46 K ARAGIANNIS F. University of the Aegean, Greece
47 K ARAKAPILIDIS N. University of Patras, Greece
48 K ARAKOSTAS Christos Eurometal, Greece
49 K ARAKOSTAS Georgios University of the Aegean, Greece
50 K ARAPATAKIS S. Greece
51 K ARAYANNIS George Perefetti Van Melle, Greece
52 K ATRAMADOS Y. University of Piraeus, Greece
53 K ATRAVAS D. BSHG , Greece
54 K ATSIKAS Socratis University of the Aegean , Greece
55 K AYAVA Irene Univeristy of the Aegean, Greece
56 K EHAGIAS J. University of the Aegean, Greece
57 K LOUTSINIOTI A. University of Piraeus, Greece
58 KONTORIS P. University of Piraeus, Greece
59 KOUSSENIDIS Dimitrios Aristotle University of Thessaloniki, Greece
(continued)
60 KOUTSOPODIOTIS T. University of Piraeus, Greece

61 L ABIRIS K. University of Piraeus, Greece
62 L APSATIS George BSHG , Greece
63 L EKKAS Themis University of the Aegean, Greece
64 L ETSIOS Y. University of Piraeus, Greece
65 L IBEROPOULOS K. University of the Aegean, Greece
66 L ITINAS Nikos University of the Aegean, Greece
67 L ORETZATOS J. University of Piraeus, Greece
68 L OUKAKIS E. Aristotle University of Thessaloniki, Greece
69 M AGINA Chryssanthi Univeristy of the Aegean, Greece
70 M ALKOTZOGLOU A. University of Piraeus, Greece
71 M ALTSINIOTIS K. I NTRACOM
72 M ANIADAKIS E. Hellenic Scientific, Greece
73 M ANOLAKOS T. BSP ABE , Greece
74 M ARKOULAKI Effie University of Piraeus, Greece
75 M AVRAKAKIS Michalis I NTRACOM, Greece
76 M ERETAKI Sofia B RAVO S.A., Greece
77 M ERIKAS A. University of the Aegean, Greece
78 M ETAXAS N. Philkeram - Johnson S.A., Greece
79 M IDDLE John Loughborough University, United Kingdom
80 M OSCHOURIS Socratis University of Piraeus, Greece
81 N EGAKIS C. Aristotle University of Thessaloniki, Greece
82 N IKOLAOU N. University of Piraeus, Greece
83 PANOUSSOPOULOU P. Univeristy of the Aegean, Greece
84 PANTAZIS P. BSHG , Greece
85 PAPADAKIS E. University of Piraeus, Greece
86 PAPADOPOULOS V. University of Piraeus, Greece
87 PAPADOPOULOU N. University of Macedonia, Greece
88 PAPOUTSI C. University of the Aegean, Greece
89 PARIKAKIS George Univeristy of the Aegean, Greece
90 PASCHALIDIS Apostolos Ministry of Education, Greece
91 PATSILINAKOS T. University of Piraeus, Greece
92 PAVLIS N. University of Piraeus, Greece
93 PAVLOU Eleni Univeristy of the Aegean, Greece
94 P IRGIANAKI E. University of the Aegean, Greece
95 P YROMALLIS Spiros Ministry of Education, Greece
96 R AFTOPOULOS John Athenian Brewery, Greece
97 R EKLITIS E. University of the Aegean, Greece
98 R EVELAKIS M. Coca Cola HBC, Greece
(continued)
99 ROGDAKI E. University of the Aegean, Greece

100 S AINI A. University of Piraeus, Greece
101 S ANTRI M. Neoset, Greece
102 S APIDIS N. University of the Aegean, Greece
103 S ARANTELIS A. Athens 2004 Olympic Games, Greece
104 S ARDIS Frantzeskos Univeristy of the Aegean, Greece
105 S IMITZIS Theodoros Unilever Hellas, Greece
106 S KENTOS Nikos Univeristy of the Aegean, Greece
107 S OCRATOUS P. University of Piraeus, Greece
108 S OTIROPOULOU I. BSHG , Greece
109 S PYROY Thomas University of the Aegean, Greece
110 S TAMATIS Dimitris Univeristy of the Aegean, Greece
111 S TASSIS I. University of Piraeus, Greece
112 TARTA S. University of Piraeus, Greece
113 TATARAKIS Antonios Oracle Hellas A.E., Greece
114 T ETOROU P. University of Piraeus, Greece
115 T HEODORAS D. University of Piraeus, Greece
116 T HEODOSIOU K. University of the Aegean, Greece
117 T HILAKERATHNE C. University of the Aegean, Greece
118 T OPOUZAS A. Neoset, Greece
119 T RIANTAPHYLLOY ELVAL , Greece
Georgios
120 T RIANTZI Z. University of Piraeus, Greece
121 T RIMERITI K. University of the Aegean, Greece
122 T SABAS G. University of Piraeus, Greece
123 T SAFARAS P. BSHG , Greece
124 T SAGARIS Christos University of the Aegean, Greece
125 T SAVDARIDOU Z. University of the Aegean, Greece
126 T SEKREKOS E. Ikonomikos Tachidromos Journal,
Greece
127 T SIOTRAS Georgios University of Macedonia, Thessaloniki,
Greece
128 T SOULFAS G. University of Piraeus, Greece
129 T ZARATZOUNI N. University of Piraeus, Greece
130 VALASSIADIS T. Philkeram - Johnson S.A., Greece
131 VOUTSINAS T. University of Piraeus, Greece
132 V RYSSAGOTIS Vassilis Univeristy of the Aegean, Greece
133 X IDIAS D. University of Piraeus, Greece
134 Z ACHAROPOULOU C. Aristotle University of Thessaloniki, Greece
(continued)
135 Z ORZOS Iakovos Univeristy of the Aegean, Greece

136 Z OTOS E. University of Piraeus, Greece
E
Simulation Model of a Reliable Production Line
E.1 Description of the Production Line

Consider a production line consisting of four stations S1 , S2 , S3 , S4 and three inter-
mediate buffers B2 , B3 , B4 shown in Figure E.1. At each station there are identical
machines: 3 machines at station 1, 2 machines at stations 2 and 3, and 3 machines at
station 4. The service times at each station are assumed to be exponential distributed
with service rates μi , i = 1, 2, 3, 4 (may be identical or different). Between the sta-
tions there are three buffers to reduce the starvation and blocking phenomena. The
capacities of the buffers are 4, 2, and 4 slots in buffers Bi+1 , i = 1, 2, 3, respectively.
Due to the finiteness of the buffers, blocking may occur. On completion of the ser-
vice at a machine at station i, the job tries to enter the next buffer Bi+1 or station i + 1.
If the buffer Bi+1 is full, the job is forced to stay at the machine in station i until a
space becomes available at buffer Bi+1 and then that machine can initiate the service
on the next job, when available. The system (PL) is saturated, i.e., in front of the first
station there is adequate raw material so that the first station is never starved. Also
the last (fourth) station is never blocked.
E.2 The Model of the System

The model of the system has been constructed using Arena 3.0 simulation software.
As may be seen from Figure E.1, it contains a collection of modeling constructs
(called modules). Particularly there are one Arrive module, four Server modules,
three Resource modules, one Simulate module, one Statistics module, one Depart
module and three Animate modules.
Each module represents a part of the original system (PL) and gives information
about the evolution of the system.
E.2.1 The Arrive module

The Arrive module represents the generation of the entities (jobs) that enter
the system. On doubleclicking on the Arrive module the following screen shown in
258 E Simulation Model of a Reliable Production Line
Fig. E.1. A production line with four stations with parallel machines at each station and
intermediate buffers
Figure E.2 will appear. The entities (jobs) are generated in batches of 15 jobs every
1 time unit (minute) and enter the queue of station 1. So a large number of jobs are
waiting for service in front of station 1 (saturation of station 1).
E.2.2 The Server modules
The Server modules represent the stations of the original system. On doubleclicking
on a Server module, e.g., station 1, the screen shown in Figure E.3 will appear.
In the dialog box are the name of the station, the capacity of the station (i.e.,
the number of identical machines, 3 in this case), information about the distribution
of service time (the exponential distribution with mean service time equal to 1 time
unit (minute)), and information about the flow of the jobs (connection to the next
station). There is also information about blocking. Clicking on the options button of
the server 2 (station 2) dialog, the next screen appears as shown in Figure E.4
E.2 The Model of the System 259
Fig. E.2. The Arrive module dialog box
Fig. E.3. The Server module dialog box

Fig. E.4. The Options dialog box
Here there is information order about the job process before the service of a job
at station 2. As one can see at the left bottom corner there is an order to release one
unit of buffer B2 (in front of server 2) as soon as the job seizes a machine at station
S2 (a random selection between the available machines). After completion of service
of a job at station 2 there is a request to seize a slot unit at the next buffer (i.e.,
B3 ) before freeing the machine at station S2 . If B3 is full, the jobs remain at station
2 (blocking after service) until a space becomes free at B3 . Then a machine in S2
becomes free – a random selection between the blocked machines.
E.2.3 The Resource modules
The Resource modules represent the buffers at the original system. On doubleclick-
ing on a Resource module, the screen shown in Figure E.5 will appear. In the dialog
box are the name of the buffer and the capacity of the buffer, i.e., the number of
buffer slots.
E.2.4 The Depart module
The Depart module represents jobs leaving the system. On doubleclicking on the
Depart module, the screen shown in Figure E.6 will appear.
In the count area of the dialog box, the individual counter button is selected to
obtain the total number of jobs that have passed through this module, with counter
name: No_of_Jobs. This is what creates the number above the icon for the Depart
module (initially at 0) which will clock up as jobs pass through this module.
E.2.5 The Simulate module
The Simulate module does not represent any part of the original system but gives
information about the length of simulation time run, the number of replications, the
Fig. E.5. The Resource module dialog box
length of warm-up period, etc. In this case, as shown in Figure E.7, the model is to be
run for 10 replications each run having a length of 50,000 time units (minutes) and a
warm-up period of length 100 minutes. The warm-up period for each run is needed
to ensure that the queue in front of station 1 is always full.
Running the model for one replication only by clicking the run button in the run
toolbar, and clicking Yes on the message to see the results we obtain the numeri-
cal summary results listed in Tables E.1 and E.2. The performance measures of the
system are listed in Table E.3.
E.2.6 The Statistics module
The Statistics module defines additional statistics to be collected as well as specify-

ing which data are to be saved to files. To create a confidence interval (CI = 95%)
for the main performance measure (Throughput), we need to run the model for more
than one replication (specifically for 10 replications) and to use the Statistics module
to save statistical data. Doubleclicking on the Statistics module, the screen shown in
Figure E.8 appears.
On clicking in the Statistics module the Edit button in the counter area, the screen
shown in Figure E.9 appears.
Fig. E.6. The Depart module dialog box
The result from this option is that the value of the counter No_of_Jobs for each
replication is saved at Throughput.DAT. This information is used by the Output
Analyzer of Arena to create the confidence interval as shown in Figure E.10.
Statistical analysis
Statistical analysis may be used to estimate some characteristic of a large population,

too large to enumerate completely. Consider, for example, the throughput of a pro-
duction line during a particular time period. Under the usual assumptions in regard
to the input to the line and the service time of the servers, the throughput is a ran-
dom variable with a particular but often unknown distribution. A run of a simulation
model may be considered to be an experiment involving the taking of a random sam-
ple of some variable of interest, in this case the throughput. If the experiment is run
a number of times n with throughput values equal to X1 , X2 , . . . , Xn , then
n
1
X=
n ∑ Xi , (E.1)
i=1
Fig. E.7. The Simulate module
Table E.1. Simulation results: Continuous variables

Identifier Average Half Width Minimum Maximum Observations
Station_1_Queue Time 11.555 0.06817 4.7310 22.969 82959
is the sample mean

n
1
s2 =
n−1 ∑ (Xi − X)2 , (E.2)
i=1
is the sample variance, where the set X1 , X2 , . . . , Xn is defined as the sample. It is well
known that
E(X) = μ , (E.3)
σ2
Var(X) = , (E.4)
n
Table E.2. Simulation results: Discrete variables

Identifier Average Half Width Minimum Maximum Final Value
# in Station_1_Queue 19.171 0.00685 13.000 20.000 20.000
# in Station_2_Queue 3.7697 0.01331 0.000 4.000 4.000
# in Station_3_Queue 1.0105 0.02009 0.000 2.000 2.000
# in Station_4_Queue 0.2714 0.01599 0.000 4.000 0.000
Buffer_2 Busy 3.7697 0.01331 0.0000 4.000 4.000
Buffer_3 Busy 1.0105 0.02009 0.0000 2.000 2.000
Buffer_4 Busy 0.27137 0.01599 0.0000 4.000 0.000
Station_1 Busy 3.0000 0.00000 2.0000 3.000 3.000
Station_2 Busy 1.9955 9.6872E-04 0.0000 2.000 2.000
Station_3 Busy 1.6698 0.01114 0.0000 2.000 2.000
Station_4 Busy 1.6588 0.01656 0.0000 3.000 1.000
Station_1 Available 3.00 N/A 3.0000 3.000 3.000
Buffer_4 Available 4.00 N/A 4.0000 4.000 4.000
Table E.3. Simulation results: Performance measures

Measure Value
Throughput 82956 jobs/50000 minutes = 1.65912 jobs/min
System efficiency 1.65912/3 = 55.30%
Utilization of S1 3.00/3 machines = 100%
Utilization of S2 1.9955/2 machines = 99.775%
Average level of B2 3.7697 jobs
Waiting time at B2 2.2720 minutes
where E and Var are the expected value and variance operators, respectively, and μ
and σ 2 are the mean and variance of the underlying distribution of X , the throughput.
√ of X is normal, X ∼ N(μ , σ ), then the distribution of X
If the underlying distribution
is normal, X ∼ N(μ , σ / n).
E(s2 ) = σ 2 (E.5)
Fig. E.8. The Statistics module dialog box
Fig. E.9. Saving the value of counter No_of_Jobs into file Throughput.DAT
where σ 2 is the variance of the underlying distribution. Should the underlying distri-
bution of X be normal, X ∼ N(μ , σ ), then the statistic (n − 1)s2 /σ 2 has a chi-squared
distribution with (n − 1) degrees of freedom, denoted by χn−1 2
.
Fig. E.10. The confidence interval (CI = 95%) of throughput
In simulation experiments, it is usually necessary to estimate the parameters of

the underlying distributions such as the mean and variance by using the output rate
from the experiments. One approach is to use a point estimate to give a single numer-
ical value for the parameter of interest. A statistic acting as a point estimate is said
to be unbiased if the expected value of the point estimate is equal to the value of the
parameter of interest. Clearly from equations (E.1), (E.2), (E.3) and (E.4) above, X
and s2 are unbiased estimates of the mean and variance, respectively, of the under-
lying distribution of X, the throughput. The variance of an estimator is an important
characteristic in that, for example, if two unbiased estimators were available, the
estimator with the lower variance is preferable in that the point estimate using that
estimator is more likely to be closer to the parameter being estimated. The lower-
variance estimator is said to be more ‘efficient.’ A further and related property is the
concept of the ‘consistency’ of an estimator. Basically, the idea is that if an estimator
is consistent, as the number in the sample, n, increases the estimator improves in
some sense, for example, the variance becomes smaller.
In simulation experiments, good point estimates of parameters are generally
available. However, some quantitative information about the variance of the esti-
mator used is required if the analyst is to have confidence in the estimate. A more
formal approach is to develop a ‘confidence interval’ for the parameter of inter-
est. The objective of determining a confidence interval is to form an interval with
specified end points that will contain the parameter of interest with a pre-specified
probability level. Clearly, the parameter of the underlying distribution has a fixed
value, i.e., it is not itself a random variable, but the estimate used is a random vari-
able. To illustrate the development of a confidence interval, consider the following
simple example.
Assume X ∼ N(μ , σ ) with σ known and μ fixed but unknown, given X = √ ∑ni=1
where Xi , i = 1, 2, . . . , n is a set of sample values of X . Now X ∼ N(μ , σ / n).
Therefore,

X −μ
Prob −zα /2 ≤ √ ≤ zα /2 = (1 − α ),
σ/ n
where zα /2 is obtained from the standard normal tables, Z ∼ N(0, 1) and

Prob −zα /2 ≤ Z ≤ zα /2 = (1 − α ).
From

X −μ
Prob −zα /2 ≤ √ ≤ zα /2 = (1 − α ),
σ/ n
by algebraic manipulation:

σ σ
Prob X − √ z ≤μ ≤X+ √ z = (1 − α ).
n α /2 n α /2
Thus the interval X − √σn zα /2 to X + √σn zα /2 is said to be the (1 − α ) confidence

interval for μ .
There were two constraints in the development of the above confidence interval
for μ , the mean of the underlying distribution, viz., X was distributed according to
a normal distribution and σ 2 , the variance of X was known. Generally speaking, if
the sample size is large, say n ≥ 30, the central limit theorem may be used to assume
that X follows a normal distribution.
As a working rule, if the number of sample, n, is 30 or more and even if the
underlying distribution of X is unknown and σ , the variance of the underlying distri-
bution is unknown, the following is a (1 − α ) confidence interval for the mean, μ , of
the underlying distribution:
σ 1 n
X ± zα /2 √ ;
n
σ2 =
n−1 ∑ (Xi − X)2
i=1
where zα /2 is obtained from N(0, 1) tables as indicated above. However, if n, the

sample size is less than 30 and the underlying distribution of X can be assumed to
be normal but the value of σ , the variance of the underlying distribution is unknown,
the following (1 − α ) confidence interval for μ obtains:
σ 1 n
X ± tn−1,α /2 √ ;
n
σ2 =
n−1 ∑ (Xi − X)2
i=1
where tn−1,α /2 is the value from the t distribution providing an area of α /2 in the
upper tail of the t distribution, with n − 1 degrees of freedom.
It should be noted that the term ‘Half Width’ in the Arena tables refers to half
the confidence interval, e.g., tn−1,α /2 in the third case discussed above. In reference
to the simulation experiments discussed above, Figure E.10 gives the confidence
interval (CI = 95%; α = 0.05) of the mean value of the throughput.
The average of 10 observations of the throughput is given in Figure E.10 as
8.32e + 004 or 83200 units, s, the sample standard deviation being 310. As n, the
sample size is less than 30 and σ 2 , the variance of the throughput is unknown, the
following 95% confidence interval for the mean of the throughput applies, provided
it may be assumed that the underlying distribution of the throughput is normal:
s
X ± t9,0.025 √
n
t9,0.025 = 2.26 from t-tables.
Thus the confidence interval for the mean value of the throughput is

310
83200 ± 2.26 √ = 83200 ± 222.
10
Note that Figure E.10 gives a ‘Half Width’ value of 222. The information on
‘Half Width’ in the Arena dialogs may be used to give what might be described
as ‘reasonable’ upper and lower bounds for the throughput. It must be noted that
some round-off error is present in the Arena calculations. For example, the number
of observations in the queue time tableau (Table E.1) of the first replication of the
above model is given to five significant figures, whereas in the confidence interval
dialog (Figure E.10) the throughput is given to three significant figures only. Using
the data given in Figure E.10, the following range of values of throughput per minute
may be obtained:
Upper and lower limits (95% confidence interval):
Upper Limit: = 83400/50000 = 1.668 jobs/minute.
Lower Limit: = 82900/50000 = 1.658 jobs/minute.
Mean Value: = 83200/50000 = 1.664 jobs/minute.
It might be noted that because of the round-off error procedures in Arena the
mean value of the distribution in this case is not exactly equal to the average of the
upper limit and lower limit values.
E.2.7 The Animate modules
The Animate modules give temporary information (dynamic plots during the sim-
ulation running) of selected parts of the system. Here the Animate module gives
information about the average level of buffer B3 . In Figure E.11, a snapshot of the
average level of buffer B3 up to a specific time (time = 1000 minutes) is shown. One
can see from this picture that the average level of buffer B3 is 1 job.
References 269
Fig. E.11. A snapshot of the evolution of the average level of buffer B3 up to time equal to
1000 minutes in a production line with 4 stations with parallel machines at each station and
intermediate buffers
References
1. Kelton, W.D., Sadowski, R.P., and Sadowski, D.A. (1998), Simulation with Arena,
McGraw-Hill.
2. Anderson, D.R., Sweeney, D.J., and Williams, T.A. (1997), Statistics for Business and
Economics, Fourth Edition, West Publishing Company.
Subject Index
absorbing state, 218 gradient method, 146

abstract or mathematical model, 12 BAP-B
accompanying algorithms, 246 formulation, 132
activity based costing (ABC), 9 heuristic algorithm, 146
aggregation BAP-C
algorithm, 67 formulation, 132
backward, 67 design rules, 134
forward, 66 formulation, 132
loss parameters, 65 heuristic algorithm
aggregation algorithm (AGGRE), 67, 235, max XK , 137
246 minW IP, 143
aggregation method, 243 software
agile manufacturing, 9 BA, 137, 155, 237, 246
algorithms and procedures guide, 233–237 DECO-1 and SA/GA, 155
analogue physical model, 12 DECO-2 and SA/GA, 155
analytical hierarchical process (AHP), 179 MARKOV, 137
artificial intelligence expert systems, 20 solution, 132
asymptotically reliable line, 65 long lines, 145
availability, 136 short lines, 134
buffers, 107
blocking after service (BAS), 60

blocking before service (BBS), 60 closed model, 13
bottleneck, 109 coefficient of variation, 105
bottleneck buffer, 109 complete enumeration (CE), 134, 136, 175,
bottleneck machine, 108, 109 237, 246
bottleneck station, 107 complex production line, 102
bowl phenomenon, 115, 118 computational issues, 117
related bibliography, 124–126 computer aided engineering (CAE), 9
brainstorming method, 180 computer aided manufacturing (CAM), 9
buffer allocation problem (BAP), 131–156, computer integrated manufacturing (CIM), 9
161, 175 computer-aided design (CAD), 9
BAP-A, 132, 134, 136, 147, 151, 155 conference participants, 247–255
formulation, 131 attendees, 251
271
272 Subject Index
presenters, 247 improvability, 107–109

continuous improvement, 108 role of the design engineer, 106–107
continuous probability distributions, 207 server allocation problem (SAP), 103
Coxian with two phases, C2 , 105, 176, work-load allocation problem (WAP),
211–214 103
Erlang, Eκ , 134–144, 176, 207, 210 diffusion technique, 62
exponential, 105, 136, 207, 208 discounted cash flow, 179, 182, 185
memoryless property, 209–210 during the useful life, 185
gamma, 207, 210 end of life, 186
lognormal, 208 initial investment, 185
normal, 207, 208, 267 discounted cash flow analysis, 105, 186
phase-type, 213–215 discrete Markov processes (Markov chains),
Student’s t, 268 216–219
uniform, 205, 207 discrete probability distributions, 206
control model, 16 Bernoulli, 206
cost considerations, 179–194, 244 binomial, 206, 207
cost minimization objective function geometric, 206, 207
G1 , 190 Poisson, 206, 208
solution algorithm, 191 double and triple optimization, 161–176
G2 , 192 dynamic model, 13
G3 , 193
cost models, 182–194
critical mean input rate, 30 economic evaluation, 105
customer driven intelligent manufacturing equilibrium condition, 29
(CDIM), 10 equivalence buffer classes, 142
customer/market interests, 6 Erlang’s loss formula, 223
evaluative methods, 133
aggregation method, 27, 64–67, 133, 235
data plotting, 219–220 decomposition approach, 27, 51–59, 133,
decision support systems (DSS), 20 145, 175, 234, 235
decomposition expansion method, 27, 59–64, 133, 134,
equations, 56, 57 234
parallel-machine station line Markovian analysis, 27–51, 69, 133, 134,
decomposition algorithm, 79 145, 234
decomposition block, 70 simulation, 85–88, 133
decomposition equations, 75 evaluative models, 17, 25–94
single-machine station line evolution of strategic manufacturing
decomposition algorithm, 59 systems, 10
decomposition block, 56 expansion algorithm, EXPAN, 64, 233, 234,
decomposition equations, 57 246
decomposition algorithm 1, DECO-1, 120, expansion flexibility, 8
234, 246 expansion method, 242
decomposition algorithm 2, DECO-2, 70, exponential and Poisson distribution
85, 120, 122, 235, 246 relationship, 210–211
decomposition approach, 241
Delphi method, 180, 194
design problems, 101, 243–244 failure
allocation, 104 exponential, 33, 34
buffer allocation problem (BAP), 104 first generation buffer class, 142
Subject Index 273
first-in-first-out, FIFO, 27 Kronecker delta function, 193

flexibility, 6
flexibility types, 8
fundamental matrix, 219 L-phenomenon, 115, 122, 123
lean production, 9
linear model, 13
Gauss-Seidel method, 32–34
Gaussian elimination, 32, 33, 72
general acronyms, 239–240 machine efficiency constraint, 108
generation of the conservative matrix, 36–49 machine flexibility, 8
generative methods, 133, 148 machines, 105
dynamic programming, DP, 145, 147, manufacturing strategy, 179
155 manufacturing systems
genetic algorithms, GA, 120, 122, 145, classification, 14–15
150, 151, 236, 246 evolution, 11
gradient method, 145, 146 methods of analysis, 18–20
Hooke-Jeeves algorithm, 134 analytical, 18
simulated annealing, SA, 120, 122, 145, simulation, 18, 20
147, 148, 175, 236, 246 models, 12–18
tabu search algorithms, TS, 145, design, 16
152–154, 156 planning, 16
generative model, 17 operations, 2
glossary, 239 process choices, 3
accompanying algorithms, 246 types, 1–12
aggregation method, 243 continuous flow lines, 4
cost considerations, 244 flexible manufacturing cells (FMC), 5
decomposition approach, 241 flexible manufacturing systems
design problems, 243–244 (FMS), 5, 180
expansion method, 242–243 group technology (GT), 5
general acronyms, 239–240 job-shop systems, 4
Markovian model, 242 and modeling, 5
mathematical fundamentals, 245–246 production or flow or unpaced lines, 4
production lines, 240–241 transfer or paced lines, 4
marketing strategy, 179
Markovian algorithm (MARKOV), 33, 234,
I-customers, 28 246
iconic physical model, 12 Markovian analysis, 25, 51, 120
II-customers, 29 conservative matrix, 25
improvability, 108 Markovian model, 242
improvable production system, 108–109 materials management, 9
with respect to WF, 108 pull systems, 9
with respect to WIP, 108 push systems, 9
with respect to WIP and WF, 109 materials requirements planning (MRP), 9
information technology, 9 mathematical fundamentals, 245
intelligent manufacturing, 9, 10 matrices, 198–203
investment in technology, 3 inverse, 201
product, 199
singular, 201
Jacobian elimination, 33 transpose, 199
274 Subject Index
matrix recursive methods, 33 parameter model, 13

maufacturing systems PARTAN method, 118
evolution and classification, 1–12 performance index, 108
maximization of profit, 181 physical model, 12
mean effective service rate, 137 plant layout, 3
mean service time, 105 predictive model, 17
mean time to failure (MTTF), 105 present worth factor, P.W.F.∗ , 186
mean time to repair (MTTR), 105 present worth value, P.W.V., 187
measures of performance, 20–21 probability, 203–215
availability, 21 density function, 203, 205
blocking time proportion, 21 mass function, 203
efficiency, 21 process, 3
holding or completion time, 21 process choice, 3
mean number of busy work-stations, 21 product flexibility, 8
mean production time, 21, 105 production
mean queue lengths, 21 high-volume, 4
mean waiting time, 21 low-variety, 4
mean WIP, 105 production flexibility, 8
mean work-in-progress, 21 production lines, 240–241
set up time, 21 production or flow line, 25, 26
system or global utilization, 21 balanced, 79, 132
throughput, 21, 105 balanced unreliable, 136, 137
utilization, 21, 105 parallel machines, 67–85, 103
merge line, 49 reliable exponential, 58
minimization of costs, 181 unbalanced, 79, 132
mix flexibility, 8 production rate, 103
modeling, 12 profit maximization objective function
F, 189
problem formulation, 12
F1 , 182
self-consistency, 13
F2 , 184
steady-state, 13
F3 , 187
transient, 13
validation, 12
modeling process, 13 quality, 101
models, 12–18 quasi-birth and death (QBD) process, 29
monotonicity property, 114, 115 queueing networks, 51
movement closed, 226
asynchronous, 4 cyclical, 226
open, 226
series or tandem queues, 226
non-linear model, 13, 49 queues, 220
M/M/1: FCFS/N/∞, 222
M/M/1: FCFS/∞/∞, 221
off-line model, 13 M/M/∞, 224
on-line model, 13 M/M/c: FCFS/c/∞, 223
open model, 13, 55 M/M/c: FCFS/K/K, 224
operation model, 16 M/M/c: FCFS/N/∞, 223
optimal buffer allocation (OBA), 136–144 M/M/c: FCFS/∞/∞, 222
optimization model, 17 Kendall’s notation, 220
Subject Index 275
Little’s formulae, 225 static model, 13

queueing networks, 225 stochastic process, 28
single-station systems, 220 Poisson, 209
strategic level decisions, 180
subsequent buffer classes, 142
random variable successive over relaxation factor, 33
coefficient of variation, 205 successive over relaxation method, 34
continuous, 205 survival curve, 105
discrete, 204 symmetricity property, 114
mean or expected value, 181, 204
variance, 181, 204
reliability, 105 tactical level decisions, 180
repair, 28 the multiple-server phenomenon
Erlang, 28 type 1, 122
exponential, 28, 34 type 2, 122
reversibility property, 114, 126 theory review
related bibliography, 125 data plotting, 219
routing flexibility, 8 discrete Markov processes (Markov
chains), 216
matrices, 198
saturated line, 28 inverse, 201
saturated model, 55 product, 199
second generation buffer class, 142 singular, 201
self-similarity phenomenon, 139–141 transpose, 199
serial type, 27 probability, 203
server allocation problem (SAP), 103, 120, density function, 203, 205
122, 161 mass function, 203
design rules, 121 queues, 220
heuristic algorithm, 123 M/M/1: FCFS/N/∞, 222
server and buffer allocation problem, S + B, M/M/1: FCFS/∞/∞, 221
162–165 M/M/∞, 224
heuristic algorithm, 165 M/M/c: FCFS/c/∞, 223
service, 28 M/M/c: FCFS/K/K, 224
Coxian, 107 M/M/c: FCFS/N/∞, 223
Erlang, 28, 34, 107 M/M/c: FCFS/∞/∞, 222
exponential, 28, 34, 107, 147 Kendall’s notation, 220
simulation Little’s formulae, 225
Arena, 18, 73, 86, 88 queueing networks, 225
eM-plant, 18, 81 single-station systems, 220
example, 81 vectors, 197
simulation model of a reliable production time based competition, 6
line, 257–269 transition probability matrix, 217
confidence interval, 267, 268 two-level work-load allocation algorithm
half width, 268 (TLWLA), 119, 120, 236
estimator consistency, 266
sample mean, 263 unstable model, 13
sample variance, 263
unbiased estimates, 266
stable model, 13 variable model, 13
276 Subject Index
vectors, 197 work-load and server allocation problem,

volume flexibility, 8 W + S, 162, 165
work-load and server allocation, W + S
L-phenomenon, 122
work-load allocation, 118 work-load and server and buffer allocation
work-load allocation problem (WAP), problem, W + S + B, 162, 163, 165
113–115, 161, 176 heuristic algorithm, 165
work-load and buffer allocation problem, work-stations, 105
W + B, 162–164, 175 work-stations in parallel, 27
Author Index
Adam, S., 2 Chiang, S. Y., 109

Alkaff, A., 90, 125 Chow, W. M., 155
Altiok, T., 89–91, 103, 119, 155 Colledani, M., 92, 156, 248
Ammar, M., 94 Conway, R., 134, 155
Ancelin, B., 93 Cox, J., 109
Anderson, D. R., 22, 90
Archetti, F. L., 18
Dallery, Y., 55, 58, 59, 91, 94, 125, 126, 147,
167
Babbage, Ch., 2 Dattatreya, E. S., 126
Banks, J., 91 David, R., 91
Baruh, H., 119 De La Wyche, P., 124
Benson, D., 91 de Werra, D., 154, 156
Berman, O., 89 Diamantidis, A. C., 51, 55, 69, 73, 75,
Blumenfeld, D. E., 90 78–81, 84, 93, 94
Boling, R. W., 89, 90, 114, 115, 124 Di Mascolo, M., 94
Boxma, O. J., 90 Ding, J., 116
Brandwajn, A., 91 Dubois, D., 92
Brateley, P., 91 Dudley, N. A., 114
Bricker, D. L., 94
Browne, J., 31, 33, 89, 147, 167
Brown, S. E., 3, 9 El-Rayah, T. E., 124
Buehler, R. J., 118 Enginarlar, E., 109
Buffa, E. S., 22 Evans, J. R., 22
Burford, R. L., 125
Burger, M., 91
Burman, M. H., 93 Faigle, U., 156
Buzacott, J. A., 69, 89, 91, 92, 103, 114, Feller, W., 22
116, 119, 163 Fishman, G. S., 91
Forestier, J. P., 92
Fox, B. L., 156
Caseau, P., 90, 93 Freeman, D. R., 90
Cheah, J. Y., 93 Frein, Y., 55, 58, 59, 91, 94, 147, 167
Cheng, D. W., 116 Friedman, H. D., 92
277
278 Author Index
Futamura, K., 121, 176 Kern, W., 156

Khoshnevis, B., 91
Kim, Y. D., 93, 94
Gant, H. L., 2 Knott, A. D., 90
Gershwin, S. B., 52, 55, 59, 73, 75, 85, 89, Knott, K., 114
91, 92, 94, 146, 176 Knuth, D. E., 152
Glover, F., 152–154, 156 Konheim, A. G., 90
Goldrat, E., 109 Kostelski, D., 89
Grasso, M., 156 Kouikoglou, V. S., 91
Greenberg, B., 116 Kubat, P., 155
Grefenstette, J., 152 Kuhn, H., 91
Groover, M. P., 6, 9, 22, 103 Kuo, C. T., 109
Gross, D., 22
Gun, L., 91
Gunneson, A. O., 9 Labetoulle, J., 62
Laguna, M., 152, 154
Lau, H. S., 125, 175
Hansen, P., 153 Law, A. M., 22, 91
Harris, C. M., 22 Levantesi, R., 92, 156
Hawthorn, 2 Liao, C. J., 116
Heavey, C., 31, 33, 36, 55, 69, 70, 73, 75, Liebermann, G. J., 22
78–81, 84, 89, 90, 93, 147, 167, Li, J., 101, 109
173 Lim, J. T., 64, 65, 67, 109
Helber, S., 73, 92, 94 Lindsay, W. M., 22
Hertz, A., 154, 156 Liu, Z., 94, 126
Hillier, F. S., 22, 79–81, 89, 90, 114, 115,
121–125, 136, 163–167, 173, 176 Magazine, M. J., 121, 124, 176
Ho, Y. C., 146, 155 Makino, T., 90, 126
Huang, C. C., 116 Makowksi, A., 91
Hunt, G. C., 89, 90 Martin, G. E., 114, 125, 175
Matta, A., 91, 92, 156
Meerkov, S. M., 64, 65, 67, 108, 109
Ito, S., 92
Meester, L. E., 132
Iyama, T., 92
Mehrtens, N., 92, 94
Melamed, B., 91, 126
Jacobs, D., 108, 109 Mishra, A., 124
Jafari, M. A., 155 Montgomery, D. C., 22, 220
Jain, S., 59, 61, 92, 93, 167 Moodie, C. L., 90
Jensen, P. A., 155 Munoz, J., 10
Jeong, K. C., 93, 94 Murrel, K. F. H., 114
Jow, Y. L., 91 Muth, E. J., 90, 114, 125, 126
Jusic, H., 92
Nemec, J. E., 91
Neuts, M. F., 92
Karagiannis, T. I., 156
Noble, D. F., 22
Kawashima, T., 126
Noble, J. S., 105
Kelton, W. D., 22, 91
Kempthorne, O., 118
Kerbache, L., 59, 61, 167 O’Kelly, M. E. J., 33, 36, 37, 89, 90
Author Index 279
Onvural, R. O., 60 Stecke, K. E., 121, 125, 176

Orlichy, J., 22 Sumita, U., 155
Ostwald, Ph. F., 10 Suresh, S., 116
Sury, R. J., 114
Sweeny, D. J., 22
Papadopoulos, H. T., 17, 31, 33, 36, 37,
51, 56, 69, 70, 73, 75, 78–81, 84,
89–94, 125, 134, 136, 139, 143, Taha, H. A., 22
147, 148, 150, 152, 155, 156, 166, Taillard, E., 154
167, 212 Takahashi, Y., 90
Parzen, E., 22 Tan, B., 92, 94
Patchong, A., 93 Tanchoco, J. M. A., 105
Perros, H. G., 22, 60, 69, 89–92 Taylor, F. W., 2, 22
Phillips, E. J., 4 Tembe, S. V., 116
Phillis, Y. A., 91 Tempelmeier, H., 91, 176
Pidd, M., 22 Thompson, W. W., 125
Pike, R., 114, 125 Tolio, T., 91, 92, 156
Plato, 12 Top, F., 64, 65, 67
Powel, S. G., 155 Towsley, D., 94, 126
Pritsker, A. A. B., 91
Pujolle, G., 62, 90, 93
Vidalis, M. I., 51, 69, 75, 92, 94, 134, 136,
137, 139, 143
Ranjan, R., 89, 91
Vouros, G. A., 155
Rao, N. P., 124
Robert, P. M., 12
Rosenshine, M., 116 Wan, Y. W., 116
Russell, R. S., 22 Waters, C. D. J., 22
Weiss, G., 116
Sakasegawa, H., 116, 126 Whitney, E., 2
Schick, I. C., 89 Whitt, W., 116
Schmenner, R., 5 Wild, R., 22, 124
Schor, J. E., 146, 176 Willaeys, D., 93
Semery, A., 93 Williams, T. A., 22
Sevast’yanov, B. A., 90 Wolff, R. W., 116
Shah, B. V., 118 Womack, J. J., 9
Shanthikumar, J. G., 69, 91, 92, 103, 114,
116, 119, 125, 132, 155, 163
Xie, X. L., 91
Silver, G. L., 124
Singh, A., 156
Slack, N., 114 Yamashita, H., 155
Smith, M. J., 59, 61, 92, 93, 125, 156, 166, Yamazaki, G., 116, 126
167 Yao, D. D., 125
So, K. C., 79, 80, 121–125, 136, 139, Yu, K. Y. C., 94
163–167, 173
Solberg, J. J., 22
Spinellis, D., 125, 147, 148, 150, 152, 156, Zhu, Y., 116
166, 167 Zimmern, B., 90

Analysis and Design of Discrete Part Production Line PDF

Uploaded by

Analysis and Design of Discrete Part Production Line PDF

Uploaded by

ANALYSIS AND DESIGN OF DISCRETE

PART PRODUCTION LINES

Aims and Scope

For other titles in this series, go to www.springer.com/series/7393

Michael J. Vidalis Diomidis Spinellis

c Springer Science+Business Media, LLC 2009

Cover illustration: Photo taken by Elias Tyligadas

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Although Dr. Alexandros Diamantidis’s name appears in both Appendix B and

Thessaloniki, 2009 Chrissoleon T. Papadopoulos

1 Manufacturing Systems: Types and Modeling . . . . . . . . . . . . . . . . . . . . . 1

2 Evaluative Models of Discrete Part Production Lines . . . . . . . . . . . . . . . 25

2.7 General Comment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

3 The Design of Production Lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

4 Work-Load and Server Allocation Problems . . . . . . . . . . . . . . . . . . . . . . 113

5 The Buffer Allocation Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

6 Double and Triple Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

7 Cost Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

A Mathematical Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

B Algorithms/Procedures Details and Guide to Use . . . . . . . . . . . . . . . . . . 233

C.8 Cost Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244

D Conference Participants: Presenters and Attendees . . . . . . . . . . . . . . . . 247

E Simulation Model of a Reliable Production Line . . . . . . . . . . . . . . . . . . . 257

Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277

1.1 Manufacturing transformation process . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.1 A K-work-station production line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.17 A K-work-station production line with Si parallel machines at each

3.1 A typical structure of a complex production line . . . . . . . . . . . . . . . . . . 102

5.1 General process of solution of buffer allocation problems . . . . . . . . . . 133

7.1 Production line design: Historical approach . . . . . . . . . . . . . . . . . . . . . . 180

7.3 Value of F1 as a function of N for a 5-station production line with

A.1 Continuous uniform distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

6.1 Overall plan of experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

A.1 Probability mass function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204

E.1 Simulation results: Continuous variables . . . . . . . . . . . . . . . . . . . . . . . . 263

1.1 Manufacturing Systems: Evolution and Classification

C. T. Papadopoulos et al., Analysis and Design of Discrete Part Production Lines, 1

Tools and software

Fig. 1.1. Manufacturing transformation process

• Material handling, transportation and storage

Process choice/ Volume/

Fig. 1.3. Process choice

1. Project: One large and complex unit.

Degree of Flexibility Required

Job Shop Machinery

Fig. 1.4. Equipment choice

Table 1.1. Processes/products/equipment

Process Typical Product Equipment

Flexibility Type Definition Depends on Factors Such as Customer Focus Flexi-

introduced. gram preparation. Machine flexibility.

1960s 1970s 1980s 1990s 2000s

reclaiming of the polluted environment? Of course, the manufacturing community

1.2 Models and Modeling

Fig. 1.5. Modeling process

is concerned, there must be self-consistency. The overall process may be described

• Parameter, a characteristic or factor which cannot be changed during the analysis

1.3 Classification of Manufacturing Systems

C: A row-vector of dimension 1 × A giving the number of operations that can be

1.4 Models of Manufacturing Systems

• Evaluative or predictive models in their basic form assume a particular configura-

Fig. 1.6. Synergistic relationship between evaluative and generative models

1.5 Methods of Analysis

Complexity Approximate analytical

Low Exact analytical

Fig. 1.7. Complexity of the model

High Analytical Parameter Values

Flexibility Simulation Parameter Values

Low Analytical Structure

Fig. 1.8. Flexibility of the model

High Simulation Modeler

Transparency Simulation Logic Only User

Low Simulation User

Fig. 1.9. Transparency to the modeler and to the user

High Analytical Evaluation