Force:drawing

Production Planning in Production Networks
Pierluigi Argoneto • Giovanni Perrone

Paolo Renna • Giovanna Lo Nigro
Manfredi Bruccoleri • Sergio Noto La Diega
Production Planning
in Production Networks
Models for Medium and Short-term Planning
123
Pierluigi Argoneto, Dr. Ing. Giovanni Perrone, Prof. Ing.
Dipartimento di Tecnologia Meccanica Dipartimento di Tecnologia Meccanica
Produzione ed Ingegneria Gestionale Produzione ed Ingegneria Gestionale
(DTMPIG) (DTMPIG)
Università degli Studi di Palermo Università degli Studi di Palermo
Viale delle Scienze Viale delle Scienze
90128 Palermo 90128 Palermo
Italy Italy
Paolo Renna, Dr. Ing. Giovanna Lo Nigro, Prof. Ing.
Dipartimento di Ingegneria e Fisica Dipartimento di Tecnologia Meccanica
dell’Ambiente (DIFA) Produzione ed Ingegneria Gestionale
Università degli Studi della Basilicata (DTMPIG)
Macchia Romana Università degli Studi di Palermo
85100 Potenza Viale delle Scienze
Italy 90128 Palermo
Italy
Manfredi Bruccoleri, Dr. Ing. Sergio Noto La Diega, Prof. Ing.
Dipartimento di Tecnologia Meccanica Dipartimento di Tecnologia Meccanica
Produzione ed Ingegneria Gestionale Produzione ed Ingegneria Gestionale
(DTMPIG) (DTMPIG)
Università degli Studi di Palermo Università degli Studi di Palermo
Viale delle Scienze Viale delle Scienze
90128 Palermo 90128 Palermo
Italy Italy
ISBN 978-1-84800-057-5 e-ISBN 978-1-84800-058-2

DOI 10.1007/978-1-84800-058-2
British Library Cataloguing in Publication Data
A catalogue record for this book is available from the British Library
Library of Congress Control Number: 2008921524
© 2008 Springer-Verlag London Limited
Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted
under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or
transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case
of reprographic reproduction in accordance with the terms of licences issued by the Copy-right Licensing
Agency. Enquiries concerning reproduction outside those terms should be sent to the publishers.
The use of registered names, trademarks, etc. in this publication does not imply, even in the absence of a
specific statement, that such names are exempt from the relevant laws and regulations and therefore free for
general use.
The publisher makes no representation, express or implied, with regard to the accuracy of the information
contained in this book and cannot accept any legal responsibility or liability for any errors or omissions that
may be made.
Cover design: eStudio Calamar S.L., Girona, Spain
Printed on acid-free paper
9 8 7 6 5 4 3 2 1
springer.com
Preface
Globalisation is pushing manufacturing companies toward a more distributed

production approach. Indeed, corporate manufacturing firms are spreading their
production all over the world in order to stay close to the customers, while medium
manufacturing firms organise themselves in networks in order to scale their
production to a global level.
This tendency is putting a lot of stress on production planning. Indeed, the more
distributed production facilities are, the more difficult and complex production
planning becomes. Both multi-plant facility and manufacturing networks require to
be coordinated in order to reach effectiveness and efficiency required by the
competitive arena.
Most of the production planning tools, such as Advanced Planning and
Scheduling (APS) tools, are designed to manage a centralised production, i.e. a
production that is accomplished in a single plan. When it comes to managing a
network of plants or a manufacturing network, the application of APS tools
becomes complex, since the complexity of the planning problem scales up and
most of all, because the necessity to recollect data from different sites in a
centralised planner causes several problems of data consistency and updating.
This is the reason why, in managing production networks, decentralised
production planning tools have been recommended by researchers and industrial
managers. Indeed, decentralised production planning has several positive outcomes
when it comes to managing production networks: a) data management is easier and
more trustworthy; b) production planning systems are more robust, scalable,
reliable; c) as a result production planning activities are easier and more reliable.
However, distributed production planning has some drawbacks: a) production
planning outcomes are considered less efficient than centralised production
planning outcomes; b) coordination among the different entities involved in
production planning activities needs to be properly designed; c) properly
commercial tools are not available yet.
vi Preface
This book concerns the above-mentioned issues. It faces the production

planning problem in complex and very structured manufacturing firms such as
those involved in the semiconductor industry. The book presents research work
answering major issues dealing with decentralised production planning; in
particular, for the considered research context it shows:
How to structure and organise decentralised production planning for a
complex multi-national corporate;
How to organise business processes among the decentralised entities
involved in the production planning process;
Which kind of methodological tools can be used to obtain a reliable,
effective and efficient cooperative production plan;
Which kind of technology can be used to develop a distributed
cooperative production planning system;
A benchmark analysis showing how the proposed approach and
methodologies allow obtaining realistic production plans in the considered
research context.
Very briefly (a more detailed outline of this book is given at the end of Chapter 1),
the book is organised as follows: Chapter 1 introduces the research problem and
the research context with reference to the state of the art; Chapters 2 and 3 provide
respectively an overview of Game Theory and Negotiation Theory which are the
methodological tools used to build cooperative production plans in a distributed
environment; Chapter 4 presents an overview of the Agent Theory that is the
technological tool suggested to develop a distributed production planning tool;
Chapter 5 presents our approach for organising and structuring a distributed
production planning system in a complex environment such as the semiconductor
industry. Chapters 6 and 7 present the methodological approach suggested to reach
a cooperative production plan in distributed networks such as described in Chapter
5. Specifically, Chapter 6 presents the methodological approach for planning
production at medium-term level, and Chapter 7 at plant level. Chapter 8 presents
the integration of the methodological approach presented in Chapters 6 and 7 in
order to show how the proposed algorithms integrate with each other in order to
provide a consistent production planning tool; finally, Chapter 9 presents the
conclusions of the research developed in the book.
I wish to thank all the researchers who have been involved in this project.
Special thanks go to Dr. Pierluigi Argoneto who, with his Ph.D. work, has allowed
us to develop a consistent and unitary body of methodological approaches for
planning production in production networks.
Palermo, 25 July 2007
Giovanni Perrone
Contents
1 Introduction and Literature Overview ...........................................................1

1.1 Introduction ................................................................................................1
1.2 Production Planning in High-tech, High-volume Industry.........................2
1.3 Strategic and Tactical Level .......................................................................2
1.4 Operational Models: Optimization and Decision Support..........................5
1.4.1 Mathematical Approach ....................................................................7
1.4.2 Queuing and Stochastic Approaches .................................................8
1.4.3 Heuristics and Simulation-based Approaches ...................................8
1.5 Motivation ..................................................................................................9
1.6 Book Outline ............................................................................................10
1.7 References ................................................................................................10
2 Game Theory: an Overview...........................................................................13

2.1 Introduction ..............................................................................................13
2.2 Game Setup ..............................................................................................14
2.3 Non-cooperative Static Games .................................................................15
2.4 Existence of Equilibrium..........................................................................16
2.5 Multiple Equilibria ...................................................................................17
2.6 Dynamic Games .......................................................................................17
2.7 Simultaneous Moves: Repeated and Stochastic Games ...........................18
2.8 Cooperative Games ..................................................................................18
2.9 N-Person Cooperative Games ..................................................................19
2.10 Characteristic Function and Imputation ...................................................20
2.11 Shapley Value ..........................................................................................21
2.12 The Bargaining Game Model ...................................................................22
2.13 References ................................................................................................23
3 Negotiation: an Overview...............................................................................25
3.1 Introduction ..............................................................................................25
3.2 Negotiation and Rational Self-interested Agents .....................................28
3.3 Negotiation Models ..................................................................................29
viii Contents
3.4 Underlying Principle for Electronic Negotiation......................................30

3.5 Electronic Negotiation Protocols..............................................................31
3.6 Characteristics that Differentiate Negotiations Protocols.........................31
3.7 Modelling Approaches and Solution Concepts ........................................32
3.7.1 Decision Theory ..............................................................................33
3.7.2 Game Theory...................................................................................34
3.7.3 Negotiation Analysis .......................................................................34
3.8 Strategic Negotiation................................................................................35
3.9 Negotiation Strategies ..............................................................................35
3.10 References ................................................................................................36
4 Multiple-agent Systems: an Overview...........................................................41

4.1 Introduction ..............................................................................................41
4.2 Applications .............................................................................................42
4.3 Challenging Issues....................................................................................43
4.4 Individual Agent Reasoning .....................................................................44
4.5 Observable Worlds ...................................................................................45
4.6 Stochastic Transitions and Utilities ..........................................................45
4.7 Distributed Decision Making ...................................................................47
4.8 Recognising and Resolving Conflicts.......................................................48
4.9 Communicating Agents............................................................................48
4.10 References ................................................................................................49
5 Distributed Production Planning in Reconfigurable Production

Networks .........................................................................................................51
5.1 Introduction ..............................................................................................51
5.2 Production Planning in DPS.....................................................................52
5.2.1 Context of the Semiconductor Industry...........................................52
5.2.2 PP in the Considered Industrial Case ..............................................54
5.2.3 IDEF0 Architecture .........................................................................55
5.2.4 Agent Architecture ..........................................................................55
5.3 Top PP Level............................................................................................57
5.4 High PP Level ..........................................................................................57
5.5 Medium PP Level.....................................................................................58
5.6 Low PP Level ...........................................................................................58
5.7 Shop-floor PP Level .................................................................................59
5.8 References ................................................................................................60
6 Distributed Models for Planning Capacity of Reconfigurable Production

Networks at Medium Term............................................................................63
6.1 Introduction ..............................................................................................63
6.2 Initial State ...............................................................................................63
6.3 The Centralised Model .............................................................................64
6.4 The Negotiation Model ............................................................................65
6.5 The Game-theoretical Model....................................................................66
6.5.1 Case 1: Characteristic Function hij > 0.............................................67
6.5.2 Case 2: Characteristic Function hij < 0.............................................67
Contents ix
6.5.3 The Bargaining Solution..................................................................70

6.6 Simulation Case Study .............................................................................71
6.6.1 The Simulation Environment...........................................................71
6.6.2 Simulation Case Study ....................................................................74
6.7 Results ......................................................................................................74
6.7.1 Two-way Analysis of Variance .......................................................75
6.7.2 Design of Experiment (DoE)...........................................................88
6.8 References...................................................................................................95
7 Distributed Models for Plant Capacity Allocation.......................................97

7.1 Introduction ..............................................................................................97
7.2 Initial State ...............................................................................................97
7.3 The Centralised Model .............................................................................98
7.4 The Negotiation Model ............................................................................99
7.4.1 Generative Function ........................................................................99
7.4.2 Reactive Function............................................................................99
7.5 The Game -theoretical Model.................................................................100
7.6 The Simulation .......................................................................................101
7.6.1 The Simulation Environment.........................................................101
7.6.2 The Simulation Case Study ...........................................................104
7.7 Results ....................................................................................................105
7.7.1 Efficiency Performance Analysis: Two-way ANOVA .................105
7.7.2 Efficiency Performance Analysis: DoE.........................................115
7.7.3 Distance Performance Analysis: Two-way ANOVA ....................121
7.7.4 Distance Performance Analysis: DoE ...........................................130
7.7.5 Number of Reconfigurations Performance Analysis: Two-way
ANOVA .................................................................................................136
7.7.6 Number of Reconfigurations Performance Analysis: DoE............147
7.7.7 Absolute Residual Performance Analysis: Two-way ANOVA.....153
7.7.8 Absolute Residual Performance Analysis: DoE ............................164
8 Distributed Production Planning Models: an Integrated Approach........171

8.1 Introduction ............................................................................................171
8.2 The Simulation Case Study ....................................................................172
8.3 Results ....................................................................................................173
8.3.1 Efficiency Performance Analysis: Two-way ANOVA .................173
8.3.2 Efficiency Performance Analysis: DoE.........................................182
8.3.3 Distance Performance Analysis: Two-way ANOVA ....................192
8.3.4 Distance Performance Analysis: DoE ..........................................200
8.3.5 Absolute Residual Performance Analysis: Two-ways ANOVA ...209
8.3.6 Absolute Residual Performance Analysis: DoE ............................214
8.4 Conclusions ............................................................................................223
9 Conclusions....................................................................................................225
9.1 Summary ................................................................................................225
9.2 Major Scientific Contributions of This Book .........................................226
9.3 Directions for Future Work ....................................................................227
x Contents
Appendix A: Simulation Results Related to Chapter 6...................................229
Appendix B: Simulation Input Parameters and Results

Related to Chapter 7 ..........................................................................................233
Appendix C: Simulation Input Parameters and Results

Related to Chapter 8 ..........................................................................................247
Index ....................................................................................................................255
1
Introduction and Literature Overview
1.1 Introduction
In recent years, manufacturing companies have entered a new era in which all
manufacturing enterprises must compete in a global economy. Global competition
increases customers’ purchasing power, which, in turn, drives frequent
introduction of new products and causes large fluctuations in product demand. To
stay competitive, companies must use production systems that not only produce
their goods with high productivity, but also allow for rapid response to market
changes and customers’ needs. A new manufacturing capability that allows for a
quick production launch of new products, with production quantities that might
unexpectedly vary, becomes a necessity. Reconfigurable manufacturing systems
(RMS), offer this capability. If we look at the evolution steps of factories in the last
twenty years, we basically find different factory concepts and production concepts
in each decade depending on different goals and criteria. Stable and well-predicted
markets led to a functional factory in which know-how was concentrated in
departments in order to achieve optimisation with a high flexibility of resources.
Later, the necessity of a more consistent orientation to markets and products led to
segmented factories. These concepts led to an enormous increase of the efficiency
of the total business through independently acting units oriented towards products
and markets. The growing complexity within the value-added chain can no longer
be steered merely by means of hierarchical leadership in competences. In the
future, networks as well as increasing cooperation and decentralisation will be
regarded as a main solution. These concepts are called decentralised production
networks and are regarded as a possible answer to the new challenges of the global
competition in order to achieve shorter response time and stronger orientation
towards customer needs, through a high innovation ability. Prerequisites for the
successful participation in production networks are changeable and reconfigurable
production processes, resources, structures and layouts as well as their logistical
and organisational concepts. This ability is necessary to withstand the permanent
change and the turbulence of the surroundings of production companies and can be
described as changeability.
2 Production Planning in Production Networks
1.2 Production Planning in High-tech, High-volume Industry

In high-tech, high-volume industries such as semiconductors, consumer electronics
and telecommunications, but also in some pharmaceutical industries, the ability of
a firm to manage production capacity is arguably the most critical factor for its
long-term success. Even in a stable economy, the demand for high-tech products is
volatile and difficult to forecast; the rapid rate of technology innovation causes
short product lifecycles, low production yield and, often, long production lead
time, all of which hamper the firm’s ability to respond to market changes.
Uncertain economic times exacerbate these challenges. Whereas in an environment
of sustained demand growth, firms might build inventories or hold excess capacity
to buffer against demand variability, most are reluctant, or unable, to assume such
financial risks in a downward market. Nevertheless, high-tech companies
recognize that in order to sustain their customer base and to seize revenue
opportunities, they must be able to manage successive technological innovations
effectively, e.g., introducing high-margin innovative products at the right moment
while maximising the return on investment for older, more mature products. To do
so, firms must structure capacities in their supply chain so that over time it is
possible to respond to demand surges from new product introduction and market
upturn, and to absorb short-term decline due to technological migration and market
downturn.
The role of capacity management is even more important in industries in which
capital equipment cost is high. For example, in the semiconductor industry,
manufacturers are faced with exorbitant capacity costs, long capacity lead times,
high obsolescence rates and high demand volatility. Moreover, the rapid
technology innovation leads to short product lifecycles and thus to higher
obsolescence rates and increased equipment usage costs. To make the situation
worse, the demand variability during a particular quarter may peak above 80% of
the average sales and the equipment procurement lead times are usually as long as
6–12 months. This means that the demand beyond the capacity lead times is highly
uncertain. The above environment drives semiconductor manufacturers to adopt
exceedingly conservative capacity expansion policies [1]. However, in a fast-
growing global market, the conservative capacity-expansion policy leads to severe
shortfalls in service levels. In this chapter, scientific literature relevant to high-tech
capacity planning and management will be reviewed: we examine the impact of
capacity from strategic as well as tactical and, especially, operational perspectives.
The final goal of the research presented in this book is building up a decision
support model for a specific operational environment, as will be presented in the
following.
1.3 Strategic and Tactical Level

At the strategic level, capacity planning involves not only the firm’s own capacity
investment, but also its supply chain partners’ investments. The capacity
investment of one firm in the supply chain could have enormous impact on the
performance of all upstream and downstream firms; thus, strategic interactions
Introduction 3
between two or more players need to be taken into account. In addition to

monolithic models that employ tools such as expected utility theory and dynamic
programming, the scientific literature increasingly considers settings that model
independent multiple decision makers in the context of supply chain management.
Research in this area utilises game-theoretical models focusing on issues such as
contracting, coordination and risk-sharing mechanisms. At the tactical level,
capacity planning focuses on capacity expansion tactics as related to the
operational aspects of the firm. Demand uncertainty and the short product
lifecycles of high-tech products are two key factors that influence capacity
expansion models. High-tech manufacturers avoid carrying inventory due to high
obsolescence rates. In fact, high-tech products are often treated as perishable
goods. Inventory models developed in this context typically use news vendor
networks settings with single-period and stochastic demand. These models
consider capacity investment by a single or multiple independent decision makers
in a stationary environment; once capacity is built it stays unchanged during the
planning horizon. This literature stream employs aggregate planning for the
acquisition and allocation of resources to satisfy customer demand over a specific
time period. The short product lifecycles in high-tech make such an approach quite
appropriate. In [2], it is observed that firms achieve better financial results by
optimising their capacity and production/inventory decisions simultaneously. In [3]
is proposed a four-way trade-off among capacity, production, subcontracting, and
inventory levels over a finite horizon. In [4], two-stage supply chains have been
considered and the impact of supply/demand uncertainty on capacity and
outsourcing decisions has been analysed. They conclude that greater supply
uncertainty encourages vertical integration. In contrast, outsourcing becomes more
attractive as uncertainty in demand increases. [5] investigates capacity investment
and pricing decisions for a manufacturer and a subcontractor with guaranteed
availability. In [6] the interplay among capacity, inventory and pricing decisions
are examined. The authors study the impact of the timing on these decisions and on
the firm’s profitability; they examine different market settings such as monopoly,
oligopoly and perfect competition. Other news-vendor like game-theoretical
models study sizing [7], timing [8], and allocation [9, 10] of capacity under
competitive settings. Few researchers tackle the capacity expansion problem with
multiple products and multiple resources; however, there is growing interest in
models that consider more general product/resource settings. Using the multi-
dimensional news vendor approach, in [11] a two-product setting in which the firm
has the option to invest in two product-dedicated resources or one flexible resource
that can process both products has been studied. The paper examines the impact of
price, cost, demand uncertainty and demand correlation on the investment
decisions. Later, the research presented in [6] extends these models to news vendor
networks that incorporate multiple products, multiple resources and multiple
storage points. The authors observe that when demand is normally distributed, the
optimal expected investment value is an increasing function of the demand vector
and a decreasing function of any variance term. In [12], news vendor networks
have also been used to investigate investment policies for product-flexible versus
dedicated resources in a two-product setting with correlated demand.
Capacity investment in high-tech industry typically involves substantial cash

exposures, volatile market demand and changing supply (technological)
specifications. In this environment, the return on investment is highly uncertain;
thus, the variability in returns is at least as important as the expected return.
Moreover, most investment expenditures are irreversible and can be considered
sunk costs for the firm once invested. In general, capacity investment implicates
considerable risk for high-tech firms. The traditional approach of maximising net
present value (NPV) when analysing capacity investment is not sufficient to
capture the managerial flexibility that exists throughout the capacity planning
process [13]. A more accurate model should consider not only the purchasing and
installation costs, but also the value of the options related to alternative
investments [14]. Option theory provides a powerful tool to value risky
investments through risk-neutral discounting and to incorporate risk without
explicitly defined utility functions. Since the seminal paper by Black and Scholes
[15], option pricing has become a popular topic in finance. The enormous attention
in the literature on options is summarised in the comprehensive survey of [16].
Although option theory has been primarily studied in finance, the potential benefits
of real options have been recognised by researchers in operations management and
engineering economics in recent years [17]. Real options theory is particularly
relevant to capacity planning because it focuses on the combined importance of
uncertainty and managerial decisions and it offers a dynamic view of a firm’s
investment and operational decisions. In [18] it is asserted that since operational
decisions have the goal of maximising value, the framework of real options can be
used to evaluate decisions under risk. Making the connection between the effect of
capacity on the firm and the pricing of a call option, it is shown that risks can be
incorporated into planning models through capacity adjustments. A model is
proposed that integrates financial risk attitudes into a linear capacity investment
problem. In general, options theory has broad appeal to a variety of application
areas in capacity and production planning. Examples include options to expand,
options to defer production [10], options to abandon a project [19], options to wait
or temporarily shut down production [20], and options to switch [21]. Paper [22]
reports that in recent years, high-tech companies have been making use of real
options to determine their investment and operating strategies, more specifically,
the timing and choice of capacity adjustments. There is a growing espousal of real
options models for high-tech capacity planning in the operations management
literature. A broad description and discussion of real options can be found in [23]
and [24]. Each unit of capacity provides the firm options to produce a certain
quantity of the product throughout its lifecycle; such options are referred to as the
operating options. The investment in capacity is the premium for the option, while
the production cost corresponds to the exercise price. On the other hand, the firm
usually has options to add more capacity, known as growth options. In general,
options are early investments associated with firms’ ability to expand in the future;
the investment may be the acquisition of (or access to) land facility, technology,
know-how or other resources. Following the legacy of option pricing, a majority of
papers in this area employ geometric Brownian motion to model demand changes;
as such, demands in future time periods can be modelled by a lognormal
distribution. As pointed out earlier, high-tech products typically have a short
Introduction 5
lifecycle. Therefore a straightforward adoption of the Black and Scholes formula in

modelling capacity expansion may lead to inaccurate conclusions since the drift
can change direction [25].
In [26], it is demonstrated that simple stochastic processes may not accurately
represent capacity investment in many important manufacturing sectors, such as
the semiconductors because they are characterised by well-defined product
lifecycles with bell-shaped growth patterns. A generalisation of the real options
valuation is proposed that explicitly incorporates stochastic product lifecycle into
the firm’s capacity investment and production planning problem. The product
lifecycle is represented using a regime-switching process in which the planning
horizon starts with a growth regime (increasing demand) and then switches
stochastically to a decay regime (decreasing demand). Both the fixed (irreversible)
capacity and the flexible (reversible) capacity are considered. In the former case,
the capacity must be fixed at the beginning of the project and cannot be adjusted
throughout the lifecycle. The firm chooses the capacity investment that maximises
the project’s NPV, given optimal production policies across periods, the cost of
investment, the demand shift parameter and the discount factor. At the beginning
of each period when it is possible to expand and contract capacity, the firm chooses
the current capacity level based on the trade-off between the cost of adjusting the
capacity and the change in expected future profits. To model the stochastic demand
in the two-regime context, the author employs the Wiener process; however, he
considers different drift values for the demand parameter in the growth and decay
regimes. While the drift is assumed to be positive in the growth regime, negative
drift is used to model the decay regime. In a given period, the probability of
switching from growth to decay is defined by a cumulative normal distribution
function of the time elapsed since the beginning of the project. The author
examines the sensitivity of the project and option values to demand uncertainty and
the project lifecycle through a numerical analysis. He shows that by ignoring the
product lifecycle, traditional approaches may undervalue the contraction option, by
underestimating the probability that the demand may fall at some point in the
future, and may overvalue the expansion option, by implicitly assuming that
demand is expected to grow indefinitely.
1.4 Operational Models: Optimization and Decision Support

We now consider, in a more expansive way, operational models: the literature
focuses on optimisation and decision support models, which can be categorised
according to the level of detail they capture and the length of the planning horizon
they consider. In [27] several fabrication installations (fabs) of different sizes that
can achieve the same throughput levels measured in wafer starts per month have
been considered. They compare multiple fabs of smaller capacity, a single fab of
larger capacity and one or more fabs of expandable capacity, all of which provide
the same total wafer output level. The trade-offs are that smaller and multiple
plants do not achieve economies of scale as good as a single larger plant of the
same total capacity; however, building small plants reduces the risk of
underutilisation. However, long installation times for an individual plant (up to 18
months) may cause lost business opportunities. The decisions are when and how
much capacity to install under demand uncertainty, which is modelled as a
continuous time Brownian motion. The objective is to identify the configuration
that maximises the total expected profits, which is found by applying a stochastic
control approach. This study shows that, in general, it is more costly to install a fab
before demand is realised than to install it late. In that regard, large-capacity
expandable fabs provide a good compromise by delaying the capacity increments
and also achieving economies of scale better than low-capacity fabs.
Paper [28] considers capacity planning across multiple fabs. Microelectronics
technologies are aggregated and distinguished by their capacity consumption rate.
Individual fab capacity is modelled as a single resource whose level is a random
variable and measured in wafer output per period. This approach takes into account
yield variability at an aggregate level. Uncertainty in demand for technologies and
uncertainty in capacity levels are reconciled in the form of discrete scenarios in a
multi-period, multi-stage, stochastic programming model. The objective is to
minimise the expected mismatch between planned and actual capacity allocation as
defined in the scenarios. The authors illustrate various methodologies for preparing
demand scenarios as input to their optimisation model. In [29], based on previous
work, the authors expand their model to include capacity expansion decisions and
aiming at minimising the sum of capacity expansion and expected fab
reconfiguration costs. An efficient solution approach, inspired by the decentralised
decision-making environment is developed. The proposed approach captures the
trade-off between the extent of the information shared between participating
decision makers and the quality of capacity planning decisions. It also quantifies
the value of information sharing.
Another line of research tries to determine whether there is enough tool
capacity to satisfy the anticipated demand through a medium-term planning
horizon and if so, when and in what numbers are additional tools required. These
are very important decisions because tools cost up to several million dollars each
and have long delivery lead times (6 to 12 months).
Further group research in these tactical planning problems comes under three
categories differentiated by the level of detail the models capture and their
respective solution approaches: mathematical programming, queuing networks and
simulation and other heuristics. Mathematical programming approaches distinguish
processes only by production rate which are otherwise treated as generic. Yield and
cycle time variability is modelled as a constant yield factor and as such congestion
effects are not modelled. The planning period is up to two years which reflects the
delivery lead time for newly acquired tools. Some studies also incorporate the
assignment of operations to tool groups that are capable of doing multiple
operations at different rates into their optimisation models. These are also referred
to as routing decisions. The batch-processing nature of semiconductor
manufacturing lends itself well to modelling with queuing networks. This approach
can compute the cycle time and throughput without running computationally
expensive simulation models. Although this approach has been applied extensively
in performance evaluation of manufacturing systems (and semiconductor
manufacturing as well), we restrict our scope to studies that use this approach
within a scheme to make tool capacity planning decisions. In such studies, a
Introduction 7
queuing network model guides the search for a minimum cost tool configuration
that achieves a desired performance level for a wafer fab. One restrictive
assumption of this approach in the literature observed so far is that the
manufacturing flow, the assignment of operations to tools and the routing of
batches is assumed to be predetermined. Detailed simulation models are also
required to test the solutions obtained by optimisation methods or by heuristics.
However, the substantial computational requirements of these models often
prohibit their use for evaluating more than a handful of alternative configurations.
These and some ad hoc approaches for capacity planning at tactical levels (such as
spreadsheet applications) fall into the last category that we consider.
1.4.1 Mathematical Approach
There are a few researchers who study the problem with a deterministic model. In
[30], a linear programming (LP) based decision support system for tactical
planning at an IBM wafer fab which is referred to as the CAPS system has been
developed. The decision problem is to find the most profitable product mix for a
given tool set and an expected demand profile. An important simplifying
assumption is that there is no inventory or back-order between two consecutive
planning periods. In [31], a deterministic integer model that relaxes this
assumption has been developed. The model minimises the total cost of inventory,
tool acquisition and tool operating costs such that the expected demand is satisfied.
The authors develop a Lagrangian relaxation-based heuristic which is empirically
shown to produce solutions with a small duality gap. However, demand volatility
and long manufacturing lead times in the semiconductor industry make it important
to integrate demand uncertainty into any decision model. Hence, most studies use a
stochastic programming approach where the demand distribution is included as a
set of discrete scenarios in the model. Swaminathan [32] studies a tool
procurement problem in a single period under the simplifying assumption that
operation-to-tool assignments are already known (rather than optimised by the
model). Although the underlying model is simplified, the study illustrates the
benefits of incorporating demand scenarios versus optimising with respect to
expected demand only. Later, the same author, in [33], generalises his earlier work
by extending it to a multi-period setting and including operation-tool assignments
as decision variables. Heuristic solution methods are developed and their
performance is investigated. The study reinforces the benefits of the demand
scenario approach and applies the solution methods to an industry data set.
Zhang et al. [34] focus on the large number of demand scenarios that are
required to capture uncertainty accurately, since this aspect is one of the main
reasons for the large size of the resulting model. Their model is developed under
the simplifying assumptions that there is no inventory and there are no backorders
between periods and that the allocation of products to tools follows simple rules
rather than explicitly being optimised by the model. The multi-variate nature of the
demand distribution (for all products combined) complicates the model
considerably.
1.4.2 Queuing and Stochastic Approaches
Chen et al. [35] are among the first to use a queuing network model to study the
performance of a wafer fab as an alternative to a detailed simulation study. The
advantage of a queuing model is that it provides a closed-form solution that relates
control policies and fab configuration to system performance (e.g., manufacturing
cycle time, throughput, work in progress). Such an analytical model makes it easy
to explore the performance consequences of different facility configurations and
operating policies without heavy computational requirements. In [36], the authors
make use of the closed-form solution of a queuing model to develop a set of non-
linear equations that represent the cycle time as a function of the total number of
tools in the system. A non-linear mixed integer optimisation model is developed,
which aims to minimise cycle time with respect to a given demand profile and a
given budget available for tool purchases. The authors develop several heuristic
solution approaches and validate them with a set of industry data.
Hopp et al. [37] combine a queuing network model with simulation. The study
attempts to find a minimum cost capacity configuration while the cycle times are
within acceptable limits and throughput is as desired. The queuing model is used to
guide the search of the optimisation model to achieve this feat Batch processes,
machine changeovers and re-entrant flows are incorporated. In [38], the authors
also use a queuing network model to derive cycle time and production costs as
functions of the number of tools in tool groups and of the throughput for products.
The model is used to determine the number and size of tool groups with the criteria
being the trade-off between manufacturing costs and the product throughput. They
use the model to analyse the impact of fab scale on cycle time and production
costs. They show that the degree of versatility of the tool groups determines the
required fab capacity to achieve a given performance: as the versatility increases
the required fab scale decreases. Agile mini-fabs could therefore be achieved by
installing versatile tools.
1.4.3 Heuristics and Simulation-based Approaches
There are a large number of studies that develop rough-cut heuristic approaches for
finding easy-to-understand practical capacity planning solutions. However,
development complexity and high runtime requirements preclude simulation from
being a mainstream application tool. In [39], the authors apply simulation for
capacity planning in a development wafer fab and look at the work in process
(WIP), cycle time, bottleneck operations and the impact of randomness on the
performance. The model is based on company-wide tool availability, process
requirement and demand data which are used in simple equations to turn first-cut
aggregate results into capacity requirements. In [40], the authors develop a
materials requirement planning model tailored to a twin wafer fab system. The
model keeps track of WIP amounts and current capacity loads and also decides on
wafer releases assuming unlimited capacity. Its main goal is to smooth loading
across fabs and across planning periods. Kotcher and Chance [9] describe an easy-
to-apply methodology for assessing the sensitivity of tool capacity with respect to
product mix. The methodology could be used in connection with a simulation
Introduction 9
model to identify bottleneck tool groups and make tool acquisition decisions. In
[21], the authors evaluate tool and factory floor (shell space) capacity planning and
expansion methodologies that are commonly practiced in industry and highlight
their relative performance under different conditions. They make use of an
optimum-seeking algorithm developed previously by the authors that finds the
optimal tool purchase time and quantities such that the sum of expected lost
demand and tool purchase costs are minimised. The results of this algorithm serves
as a common comparison basis for the heuristic methods that are practised.
1.5 Motivation
As has been shown in the previous sections, different approaches have been
utilised to face the strategic/tactical/operational-level manufacturing decisions of
the high-tech industries. Many of these studies aim to analyse and optimise, where
possible, some performance indicators related to a specific decision level. No
organic study on the capacity planning through all the levels constituting a high-
tech, high-volume company has been detected in the literature. This lack is the
starting point of the research presented in this book: an innovative distributed and
hierarchical approach for planning capacity at different levels of time horizon will
be introduced. A general framework for distributed production planning of
complex and multi-plant manufacturing companies in high-tech, high-volume
industry will be analyzed using different mathematical tools. For each level
bargaining objectives, actors, information and roles will be defined and simulated
using a specific simulation environment based on the multiple-agent-system
(MAS) paradigm. The focus pertains to the study of coordination policies in a
distributed environment. Specifically, the goal is to evaluate their impact in a
company, which we call reconfigurable enterprise (RE), which is a production
network made of different and geographically dispersed plants which can be
reconfigured in order to gather a specific production objective of the enterprise.
The distributed and self-organising nature of REs allows responsiveness with
regard to the unpredictable market changes: quite normally significant variations in
the demand rate arise suddenly and in an unpredictable way. In this situation, the
low degree of flexibility characterising production capacity leads almost inevitably
to some criticalities; however, the RE’effectiveness and efficiency depends on the
extent to which overall performance is obtained; in order to reach this objective
REs require strong coordination tools. As shown in the literature overview of the
previous sections, basically two approaches support coordination in a distributed
organisation: centralised planning tools and decentralised ones. In a centralised
planning perspective decisions are taken by a central planner: the most important
drawback concerns the lack of flexibility of such a solution that leads this approach
to be unusable in the real practice where the scale of the planning problem
increases. The other possible solution is to use decentralised decision-making
approaches. This means that each actor involved in the planning process takes its
own planning decision, while common goals are reached through coordination
mechanisms. This book will introduce a new approach concerning capacity
allocation methodologies, where managerial and organisational tools are proposed
with the aim of supporting human decision making, rather than substituting it. In
particular, the authors combine techniques from cooperative game theory,
negotiations models and multi-agent systems to produce highly reactive,
lightweight, useful tools to be used in a real-time system environment in
distributed enterprises.
1.6 Book Outline

The book is organised as follows:
In Chapter 2 game theory in its most relevant topics for the present research is
introduced. An overview of the scientific literature concerning the applications of
this theory in the capacity planning field is presented.
In Chapter 3 negotiation and its relevant solutions concepts related to the
conflicts among independent agents is discussed. An overview of the scientific
literature concerning the applications of this theory in the capacity planning field is
presented.
In Chapter 4 the multi-agent systems (MAS) approach and its application to the
simulation environment necessary to simulate the coordination approaches
proposed in this is presented.
In Chapter 5 an innovative distributed and hierarchical approach for planning
capacity at different levels of time horizon is introduced. A general framework for
distributed production planning of complex and multi-plant manufacturing
companies in high-tech, high-volume industries is structured. For each level,
bargaining objectives, actors, information and roles are defined.
In Chapter 6 distributed models for capacity planning at medium term are
analysed and simulated. Specifically, the mathematical formulations of the
approaches are shown and then simulated with a specific simulation environment
developed by using the MAS paradigm. A statistical analysis is conducted to
evaluate differences existing between the proposed models and to evaluate the
impact of the input parameters on the estimated performance measures.
In Chapter 7 distributed model for plant allocation is analysed. Also in this
case, the mathematical formulations of the approaches are shown and then
simulated with a specific simulation environment and a statistical analysis is
conducted to evaluate models and input parameters impact on the estimated
performance measures.
In Chapter 8 an innovative integrated approach of the high and medium levels
of capacity planning is simulated. The impact of the input parameters and models
on the estimated performance measures is evaluated using a statistical analysis.
In Chapter 9 conclusions and future work are reported.
1.7 References
[1] Erkoc M, Wu SD (2004) Capacity reservation across multiple buyers. Working paper,
Lehigh University, Bethlehem, PA
Introduction 11
[2] Bradley JR, Arntzen BC (1997) The simultaneous planning of production, capacity
and inventory in seasonal demand environments. Oper Res 47(6):795–806
[3] Atamturk A, Hochbaum DS (2001) Capacity acquisition, subcontracting, and lot
sizing. Manag Sci 47(8):1081–1100
[4] Kouvelis P, Milner JM (2002) Supply chain capacity and outsourcing decisions: the
dynamic interplay of demand and supply uncertainty. IIE Trans 34(8): 717–728
[5] Tan B (2002) Subcontracting with availability guarantees: production control and
capacity decisions,” IIE Trans 36(8):711–724
[6] van Mieghem JA, Rudi N (2002) Newsvendor networks: inventory management and
capacity investment with discretionary activities. M&SOM 4(4):313–335
[7] Bernstein F, DeCroix GA (2004) Decentralized pricing and capacity decisions in a
multitier system with modular assembly. Manag Sci 50(9):1293–1308
[8] Ferguson ME, DeCroix GA, Zipkin PH (2002) When to commit in a multi-echelon
supply chain with partial information updating. Working paper, Georgia Tech,
Atlanta, GA
[9] Kotcher R, Chance F (1999) Capacity planning in the face of product-mix uncertainty.
In: IEEE international symposium on semiconductor manufacturing conference
proceedings, 11–13 October 1999, Santa Clara, CA, pp 73–76
[10] Pindyck RS (1988) Irreversible investment, capacity choice, and the value of the firm.
Am Econ Rev 78(5):969–985
[11] van Mieghem JA (1998) Investment strategies for flexible resources. Manag Sci
44(8):1071–1078
[12] Bish E, Wang Q (2004) Optimal investment strategies for flexible resources,
considering pricing and correlated demands. Oper Res 52(6):954–964
[13] Feinstein SP (2002) A better understanding of why npv undervalues managerial
flexibility. Eng Econ 47(4):418
[14] Dixit A, Pindyck RS (1994) Investment under uncertainty. Princeton University Press,
Princeton, NJ
[15] Black F, Scholes M (1999) The pricing of options and corporate liabilities. J Polit
Econ 81(3):637–654
[16] Broadle M, Detemple JB (2004) Option pricing: valuation models and applications.
Manag Sci 50(9):1145
[17] Miller LT, Park CS (2002) Decision making under uncertainty—real options to the
rescue?. Eng Econ 47(2): 105–150
[18] Birge JR (2000) Option methods for incorporating risk into linear capacity planning
models. Manuf Serv Oper Manag 2(1):19–31
[19] Majd S, Myers S (1990) Abandonment value and project life. Adv Futures Options
Res 4:1–21
[20] McDonald R, Siegel D (1985) Investment and the valuation of firms when there is an
option to shut down. Int Econ Rev 26:331–349
[21] Cakanyildirim M, Roundy RO (2002) Optimal capacity, expansion and contraction
under demand uncertain. Working paper, University of Texas, Dallas, TX
[22] Johnson B, Billington C (2003) Creating and leveraging options in the high
technology supply chain. In: Harrison TP, Lee HL, Neale JJ (eds) The practice of
supply chain management: where theory and application converge. International series
in operations research and management science, Kluwer
[23] Trigeorgis L (1996) Real options: managerial flexibility and strategies in resource
allocation. MIT Press, Cambridge
[24] Amram M, Kulatilaka N (1999) Real options: Managing strategic investment in an
uncertain world. Harvard Business School Press, Boston
[25] Bowman EH, Moskowitz GT (2001) Real options analysis and strategic decision
making. Organ Sci 12(6):772–777
[26] Bollen NP (1999) Real options and product life cycles. Manag Sci 45(5):670
[27] Benavides DL, Duley JR, Johnson BE (1999) As good as it gets: optimal fab design
and deployment. IEEE Trans Semicond Manuf 12(3):281–287
[28] Christie RME, Wu SD (2002) Semiconductor capacity planning: stochastic modelling
and computational studies. IIE Trans 34(2):131–144
[29] Karabuk S, Wu SD (2003) Coordinating strategic capacity planning in the
semiconductor industry. Oper Res 51(6):839–849
[30] Bermon S, Hood SJ (1999) Capacity optimization planning system [caps]. Interfaces
29(5):31–51
[31] Catay B, Erenguc SS, Vakharia AJ (2003) Tool capacity planning in semiconductor
manufacturing. Comput Oper Res 30(9):1349–1366
[32] Swaminathan JM (2000) Tool capacity planning for semiconductor fabrication
facilities under demand uncertainty. Eur J Oper Res 120(3):545–558
[33] Swaminathan JM (2002) Tool procurement planning for wafer fabrication facilities: a
scenario-based approach. IIE Trans 34(2):145–155
[34] Zhang F, Roundy R, Cakanyildirim M, Huh WT (2004) Optimal capacity expansion
for multi-product, multi-machine manufacturing systems with stochastic demand. IIE
Trans 36(1):23–37
[35] Chen H, Harrison JM, Mandelbaum A, Van Ackere A, Vein LM, (1988) Empirical
evaluation of a queuing network model for semiconductor wafer fabrication. Oper Res
36(2):202–215
[36] Bard JF, Srinivasan K, Tirupati D (1999) An optimization approach to capacity
expansion in semiconductor manufacturing facilities. Int J Prod Res 37(15):3359–
3382
[37] Hopp, Wallace J, Spearman ML (1996) Factory Physics, Irwin/McGraw-Hill, Boston
[38] Iwata Y, Taji K, Tamura H (2003) Multi-objective capacity planning for
semiconductor manufacturing. Prod Plan Control 14(3):244–254
[39] Chou W, Everton J (1996) Capacity planning for development wafer fab expansion.
In: Advanced semiconductor manufacturing conference and workshop proceedings,
IEEE/SEMI, Cambridge, MA, pp 17–22
[40] Chen JC, Fan YC, Wang JY, Lin TK, Leea SK, Wu SC, Lan YJ (1999) Capacity
planning for a twin fab. In: IEEE International symposium on semiconductor
manufacturing conference proceedings, 11–13 October 1999, Santa Clara, CA, pp
317–320
2
Game Theory: an Overview
2.1 Introduction
“Game theory is a branch of mathematics that is concerned with the actions of
individuals who are conscious that their actions affect each other”. As such, game
theory (hereafter GT) deals with interactive optimisation problems. While many
economists in the past few centuries have worked on what can be considered
game-theoretical (hereafter G-T) models, John von Neumann and Oskar
Morgenstern are formally credited as the fathers of modern game theory. Their
classic book Theory of Games and Economic Behavior [1] summarises the basic
concepts existing at that time. GT has since enjoyed an explosion of developments,
including the concept of equilibrium [2], games with imperfect information [3],
cooperative games [4, 5], and auctions [6], to name just a few. Citing Shubik [7],
“In the 50s ... game theory was looked upon as a curiosum not to be taken seriously
by any behavioural scientist. By the late 1980s, game theory in the new industrial
organisation has taken over: game theory has proved its success in many
disciplines.”
GT is divided into two branches, called the non-cooperative and cooperative
branches. The two branches of GT differ in how they formalise interdependence
among the players. In the non-cooperative theory, a game is a detailed model of all
the moves available to the players. By contrast, the cooperative theory abstracts
away from this level of detail, and describes only the outcomes that result when the
players come together in different combinations. Though standard, the terms non-
cooperative and cooperative game theory are perhaps unfortunate. They might
suggest that there is no place for cooperation in the former and no place for
conflict, competition etc. in the latter. In fact, neither is the case. One part of the
non-cooperative theory (the theory of repeated games) studies the possibility of
cooperation in ongoing relationships. And the cooperative theory embodies not just
cooperation among players, but also competition in a particularly strong, unfettered
form. The non-cooperative theory might be better termed procedural game theory,
the cooperative theory combinatorial game theory. This would indicate the real
distinction between the two branches of the subject, namely that the first specifies
various actions that are available to the players while the second describes the
outcomes that result when the players come together in different combinations.
The goal of this chapter is to give a brief overview about GT and, specifically,
about G-T concepts and tools. Obviously, due to the need of short explanations, all
proofs will be omitted, and we will only focus on the intuition behind the reported
results.
2.2 Game Setup

To break the ground for next section on non-cooperative games, basic GT notation
will be introduced: the reader can refer to Friedman [8] and Fudenberg and Tirole
[9] if a more deep knowledge is required. A game in the normal form consists of:
players (indexed by i 1, 2,..., n) , a set of strategies (denoted by
xi , i 1, 2,..., n) available to each player and payoffs
( S i ( x1 , x2 ,..., xn ) , i 1, 2,..., n) received by each player. Each strategy is defined
on a set Xi , xi Xi , so we call the Cartesian product X1 u X 2 u ... u Xn the
strategy space (typically the strategy space is R n ). Each player may have a one-
dimensional strategy or a multi-dimensional strategy. However, in simultaneous-
move games each player’s set of feasible strategies is independent from the
strategies chosen by the other players, i.e., the strategy choice of one player does
not limit the feasible strategies of another player. A player’s strategy can be
thought of as the complete instruction for which actions to take in a game. For
example, a player can give his or her strategy to a person that has absolutely no
knowledge of the player’s payoff or preferences and that person should be able to
use the instructions contained in the strategy to choose the actions the player
desires. Because each player’s strategy is a complete guide to the actions that are to
be taken, in the normal form the players choose their strategies simultaneously.
Actions, which are adopted after strategies, are thus chosen and those actions
correspond to the given strategies. The normal form can also be described as a
static game, in contrast to the extensive form which is a dynamic game. If the
strategy has no randomly determined choices, it is called a pure strategy; otherwise
it is called a mixed strategy. There are situations in economics and marketing in
which mixed strategies have been applied: e.g., search models [10] and promotion
models [11]. In a non-cooperative game the players are unable to make binding
commitments regarding which strategy they will choose before they actually
choose their strategies. In a cooperative game players are able to make these
binding commitments. Hence, in a cooperative game players can make side-
payments and form coalitions. The overview here reported starts with non-
cooperative static games.
Game Theory: an Overview 15
2.3 Non-cooperative Static Games

In non-cooperative static games the players choose strategies simultaneously and
are thereafter committed to their chosen strategies. The solution concept for these
games was formally introduced by John Nash [2] although some instances of using
similar concepts date back to a couple of centuries. The concept is best described
through best response functions.
Definition 1. Given the n-player game, player i’s best response (function) to the
strategies x i of the other players is the strategy xi* that maximizes player i’s
payoff S i ( xi , x i ) : x * ( x i ) arg max S i ( xi , x i ) .
i
xi
If S i is quasi-concave in xi the best response is uniquely defined by the first-

order conditions. Clearly, given the decisions of other players, the best response is
the one that the best player i can hope for. Naturally, an outcome in which all
players choose their best responses is a candidate for the non-cooperative solution.
Such an outcome is called a Nash equilibrium (hereafter NE) of the game.
Definition 2. An outcome ( x1* , x2* ,..., xn* ) is a Nash equilibrium of the game if xi* is
a best response to x* i for all i = 1, 2, ..., n.
One way to think about an NE is as a fixed point of the best response mapping
R o R n . Indeed, according to the definition, the NE must satisfy the system of
n
equations wS i wxi 0 , for all i. Recall that a fixed point x of mapping f(x),
R n o R n is any x such that f ( x ) x . Define f i ( x1 , x2 ,..., xn ) wS i wxi xi . By
the definition of a fixed point,
fi ( x1* , x2* ,..., xn* ) wS i ( x1* ,..., xn* ) wxi xi* o wS i ( x1* ,..., xn* ) wxi 0 , all i.
Hence, x * solves the first-order conditions if and only if it is a fixed point of
mapping f(x) defined above. The concept of the NE is intuitively appealing.
Indeed, it is a self-fulfilling prophecy. To explain, suppose a player is able to
guess the strategies of the other players. A guess would be consistent with payoff
maximisation (and therefore reasonable) only if it presumes that strategies are
chosen to maximise every player’s payoff given the chosen strategies. In other
words, with any set of strategies that is not an NE there exists at least one player
that is choosing a non payoff maximizing strategy. Moreover, the NE has a self-
enforcing property: no player wants to unilaterally deviate from it since such
behaviour would lead to lower payoffs. Hence the NE seems to be the necessary
condition for the prediction of any rational behaviour by players.
Although attractive, numerous criticisms of the NE concept exist. Two
particularly vexing problems are the non-existence of equilibrium and the
multiplicity of equilibria. Without the existence of an equilibrium, little can be said
regarding the likely outcome of the game. If there are multiple equilibria, then it is
not clear which one will be the outcome. Indeed, it is possible the outcome is not
even an equilibrium because the players may choose strategies from different
equilibria. In some situations it is possible to rationalise away some equilibria via a
refinement of the NE concept: e.g., trembling hand perfect equilibrium [12],
sequential equilibrium [13] and proper equilibria [14]. In fact, it may even be
possible to use these refinements to the point that only a unique equilibrium
remains.
An interesting feature of the NE concept is that the system optimal solution (a
solution that maximises the total payoff to all players) need not be an NE. In fact,
an NE may not even be on the Pareto frontier: the set of strategies such that each
player can be made better off only if some other player is made worse off. A set of
strategies is Pareto optimal if they are on the Pareto frontier; otherwise a set of
strategies is Pareto inferior. Hence, an NE can be Pareto inferior. The Prisoner’s
Dilemma game is the classic example of this: only one pair of strategies is Pareto
optimal (both “cooperate”), and the unique Nash equilibrium (both “defect”) is
Pareto inferior.
2.4 Existence of Equilibrium

An NE is a solution to a system of n equations (first-order conditions), so an
equilibrium may not exist. Non-existence of an equilibrium is potentially a
conceptual problem since in this case it is not clear what the outcome of the game
will be. However, in many games an NE does exist and there are some reasonably
simple ways to show that at least one NE exists. As already mentioned, an NE is a
fixed point of the best response mapping. Hence fixed-point theorems can be used
to establish the existence of an equilibrium. There are three key fixed point
theorems, named after their creators: Brouwer, Kakutani and Tarski. (see [15] for
details and references.) However, direct application of fixed-point theorems is
somewhat inconvenient and hence generally not done (see [16] for existence proofs
that are based on Brouwer’s fixed-point theorem). Alternative methods, derived
from these fixed-point theorems, have been developed. The simplest (and the most
widely used) technique for demonstrating the existence of an NE is through
verifying concavity of the players’ payoffs, which implies continuous best
response functions.
Theorem 1. [17]. Suppose that for each player the strategy space is compact and
convex and the payoff function is continuous and quasi-concave with respect to
each player’s own strategy. Then there exists at least one pure strategy NE in the
game.
If the game is symmetric (i.e., if the players’ strategies and payoffs are
identical), one would imagine that a symmetric solution should exist. This is
indeed the case, as the next theorem ascertains.
Theorem 2. Suppose that a game is symmetric and for each player the strategy
space is compact and convex and the payoff function is continuous and quasi-
concave with respect to each player’s own strategy. Then there exists at least one
symmetric pure strategy NE in the game.
2.5 Multiple Equilibria

Many games are just not blessed with a unique equilibrium. The next best situation
is to have a few equilibria. (The worst situation is either to have an infinite number
of equilibria or no equilibrium at all.) The obvious problem with multiple
equilibria is that the players may not know which equilibrium will prevail. Hence,
it is entirely possible that a non-equilibrium outcome results because one player
plays one equilibrium strategy while a second player chooses a strategy associated
with another equilibrium. However, if a game is repeated, then it is possible that
the players eventually find themselves in one particular equilibrium. Furthermore,
that equilibrium may not be the most desirable one. If one does not want to
acknowledge the possibility of multiple outcomes due to multiple equilibria, one
could argue that one equilibrium is more reasonable than the others. For example,
there may exist only one symmetric equilibrium and one may be willing to argue
that a symmetric equilibrium is more focal than an asymmetric equilibrium. In
addition, it is generally not too difficult to demonstrate the uniqueness of a
symmetric equilibrium. If the players have one-dimensional strategies, then the
system of n first-order conditions reduces to a single equation and one need only
show that there is a unique solution to that equation to prove the symmetric
equilibrium is unique. If the players have m-dimensional strategies, m > 1, then
finding a symmetric equilibrium reduces to determining whether a system of m
equations has a unique solution (easier than the original system, but still
challenging).
2.6 Dynamic Games

The simplest possible dynamic game was introduced by Stackelberg [18]. In a
Stackelberg duopoly model, player 1 chooses a strategy first (the Stackelberg
leader) and then player 2 observes this decision and makes his own strategy choice
(the Stackelberg follower). To find an equilibrium of a Stackelberg game (often
called the Stackelberg equilibrium) we need to solve a dynamic two-period
problem via backwards induction: first find the solution x2* x1 for the second
player as a response to any decision made by the first player:
* wS 2 ( x2 , x1 )
x ( x1 ) :
2 0.
wx2
Next, find the solution for the first player anticipating the response by the second
player:
*
dS 1 ( x1 , x2* ( x1 )) wS 1 ( x1 , x2* ) wS 1 ( x1 , x2 ) wx 2
0.
dx1 wx1 wx2 wx1
Intuitively, the first player chooses the best possible point on the second player’s
best response function. Clearly, the first player can choose an NE, so the leader is
always at least as well off as he would be in NE. Hence, if a player were allowed to
choose between making moves simultaneously or being a leader in a game with
complete information he would always prefer to be the leader.
2.7 Simultaneous Moves: Repeated and Stochastic Games

A different type of dynamic game arises when both players take actions in multiple
periods. Two major types of this game exist: without and with time dependence. In
the multi-period game without time dependence the exact same game is played
over and over again (hence the term repeated games). The strategy for each player
is now a sequence of actions taken in all periods. Consider one repeated game
version of the competing newsvendor game in which the newsvendor chooses a
stocking quantity at the start of each period, demand is realised and then leftover
inventory is salvaged. In this case, there are no links between successive periods
other than the players’ memory about actions taken in all the previous periods. A
fascinating feature of repeated games is that the set of equilibria is much larger
than the set of equilibria in a static game and may include equilibria that are not
possible in the static game. At first, one may assume that the equilibrium of the
repeated game would be to play the same static NE strategy in each period. This is,
indeed, an equilibrium but only one of many. Since in repeated games the players
are able to condition their behaviour on the observed actions in the previous
periods, they may employ so-called trigger strategies: the player will choose one
strategy until the opponent changes his play, at which point the first player will
change the strategy. This threat of reverting to a different strategy may even induce
players to achieve the best possible outcome (i.e., the centralised solution) which is
called an implicit collusion. Many such threats are, however, non-credible in the
sense that once a part of the game has been played, such a strategy is not an
equilibrium anymore for the reminder of the game. To separate out credible threats
from non-credible, Selten [19] introduced the subgame, a portion of the game (that
is a game in itself) starting from some time period and a related notion of subgame-
perfect equilibrium (this notion also applies in other types of games, not
necessarily repeated), and equilibrium for every possible subgame (see Hall and
Porteus [20] and van Mieghem and Dada [21] for solutions involving subgame-
perfect equilibria in dynamic games).
2.8 Cooperative Games

The idea behind cooperative game theory has been expressed in this way:
“Cooperative theory starts with a formalization of games that abstracts away
altogether from procedures and concentrates, instead, on the possibilities for
agreement. There are several reasons that explain why cooperative games came to
be treated separately. One is that when one does build negotiation and enforcement
procedures explicitly into the model, then the results of a non-cooperative analysis
depend very strongly on the precise form of the procedures, on the order of making
offers and counter-offers and so on. This may be appropriate in voting situations in
which precise rules of parliamentary order prevail, where a good strategist can
indeed carry the day. But problems of negotiation are usually more amorphous; it
is difficult to pin down just what the procedures are. More fundamentally, there is a
feeling that procedures are not really all that relevant; that it is the possibilities for
coalition forming, promising and threatening that are decisive, rather than whose
turn it is to speak. Detail distracts attention from essentials. Some things are seen
better from a distance; the Roman camps around Metzada are indiscernible when
one is in them, but easily visible from the top of the mountain” [22].
The subject of cooperative games first appeared in the seminal work of von
Neumann and Morgenstern [1]. However, for a long time cooperative game theory
did not enjoy as much attention in economics literature as non-cooperative GT.
Cooperative GT involves a major shift in paradigms as compared to non-
cooperative GT: the former focuses on the outcome of the game in terms of the
value created through cooperation of (a subset of) players but does not specify the
actions that each player will take, while the latter is more concerned with the
specific actions of the players. Hence, cooperative GT allows us to model
outcomes of complex business processes that otherwise might be too difficult to
describe (e.g., negotiations) and answers more general questions (e.g., how well is
the firm positioned against competition). In what follows, we will cover
transferable utility cooperative games (including two solution concepts: the core of
the game and the Shapley value).
2.9 N-Person Cooperative Games

Recall that the non-cooperative game consists of a set of players with their
strategies and payoff functions. In contrast, in this case, although players are
autonomous decision makers, they may have an interest in making binding
agreements in order to have a bigger payoff at the end of the game. This agreement
or partnership is the basic ingredient of the mathematical model of a cooperative
game, and it is called a coalition. Mathematically, a coalition is a subset of the set
of players N and we can denote it by S. To form a coalition S, it is required that
agreements take place involving all players in the future coalition S. Whenever all
players approve joining in a new entity called coalition, we can say that the new
coalition is formed. Joining a coalition S also implies that there is no possible
agreement between any member of S and any member not in S (set N\S). In short,
the essential feature of a coalition is its foundational agreement that binds and
reconstitutes the individuals as a coordinated entity. The grand coalition of all n
players will be referred as coalition N (there is a total of 2n 1 possible coalitions);
The empty coalition is a coalition made up of no members (the null set ). A
coalition structure is a means of describing how the players divide themselves into
mutually exclusive coalitions. Any exhaustive partition of the players can be
described by a set S { ^S1 , S2 ,..., Sm ` of the m coalitions that are formed. The set S
is a partition of N that satisfies three conditions:
S j z , j 1,..., m
Si S j , for all i z j , and S j N.

These conditions state that each player belongs to one and only one of the m
non-empty coalitions within the coalition structure, and also specifies that none of
the players in any coalition m is connected to other players not in the coalition;
finally, the mutually exclusive union of all coalitions m forms the grand coalition.
2.10 Characteristic Function and Imputation

von Neumann and Morgenstern [23] introduced the term characteristic function for
the first time. More formally, we can define that:
Definition 3. For each subset S of N, the characteristic function Q of a game gives

the biggest amount Q ( S ) that the members of S can be sure of receiving if they act
together and form a coalition, without any help from other players not in S.
A restriction on this definition is that the value of the game to the empty
coalition is zero, that is, Q ( ). A further requirement that is generally made is
called superadditivity. Superadditivity can be expressed as follows:
Q (S T ) t Q (S ) Q (T ) for all S , T N such that S T .
This means that the total payoff for the grand coalition is collectively rational,
because the total payoff to the players is always as much as what they would get
individually. This suggests the following definition.
Definition 4. A game in characteristic function form consists of a set of players,

together with a function Q defined for all subsets of N, such that
Q ( S T ) t Q ( S ) Q (T ) whenever S and T are disjoint coalitions of players.
Games in which at least one possible coalition can increase the total payoff of
its members are called essential, and those in which there is no coalition that
improves the total payoff are called inessential. Mathematically, an essential game
is one in which at least one of the superadditive inequalities
Q (S T ) t Q (S ) Q (T ) is strict. The specific actions that players have to take to
create this value are not specified: the characteristic function only defines the total
value that can be created by utilising all players’ resources. Hence, players are free
to form any coalitions that are beneficial to them and no player is endowed with
power of any sort. We will further restrict our attention to the transferable utility
games in which the outcome of the game is described by real numbers
S i , i 1,..., N showing how the total created value (or utility or pie)
N
S (N ) ¦ i 1
S i was divided among players. Of course, one could offer a very
simple rule prescribing division of the value; for example, a fixed fraction of the
total pie can be allocated to each player. However, such rules are often too
simplistic to be a good solution concept. A much more frequently used solution
concept of the cooperative game theory is the core of the game. This concept can
be compared to the NE for non-cooperative games:
Definition 5. The utility vector S 1 ,..., S N is in the core (and will be called
imputation) of the cooperative game if it satisfies S ( N ) Q ( N ), group rationality,
and xi t Q (î`), individual rationality.
The core of the game, introduced by Gillies in 1953 [24], can be interpreted
through the added-value principle. Define (N\S) as a set of players excluding those
in coalition S (coalition can include just one player). Then the contribution of a
coalition S can be calculated as Q ( N ) Q ( N \ S ). Clearly, no coalition should be
able to capture more than its contribution to the coalition (otherwise the remaining
N\S players would be better off without the coalition S). Definition 5 clearly
satisfies the added-value principle. Typically, when analysing a game, one has to
calculate an added value from each player: if the value is zero, the player is not in
the core of the game. If the core is non-empty, the added values of all players in the
core comprise the total value that the players create. As is true for NE, the core of
the game may not exist (i.e., it may be empty) and the core is often not unique.
When the core is non-empty, the cooperative demands of every coalition can be
granted, but when the core is empty, at least one coalition will be dissatisfied.
Shubik [7] noted that a game with a non-empty core is sociologically neutral, i.e.
every cooperative demand by every coalition can be granted, and there is no need
to resolve conflicts. On the other hand, in a coreless game, the coalitions are too
strong for any mechanism to satisfy every coalitional demand. However, a core set
with too many elements is not desirable, and it has little predictive power [25].
Imputations in the core, where they exist, have a certain stability, because no
player or subset of players has any incentive to leave the grand coalition. But since
many games have empty core, the core fails to provide a general solution for n-
person games in characteristic form. von Neumann and Morgenstern [24] proposed
a different solution concept more generally applicable than the core. That proposal
is called the von Neumann Morgenstern solution or the stable set. The stable set is
based on the concept of dominance, which is explained as follows. One imputation
is said to dominate another if there is a subset of players who prefer the first to the
second and can enforce it by forming a coalition.
2.11 Shapley Value

The concept of the core, though intuitively appealing, also possesses some
unsatisfying properties. As we mentioned, the core might be empty or quite large
or indeterministic. As it is desirable to have a unique NE in non-cooperative
games, it is desirable to have a solution concept for cooperative games that results
in a unique outcome and hence has a reasonable predictive power. Shapley [26]
offered an axiomatic approach to the solution concept that is based on three rather
intuitive axioms. First, the value of the player should not change due to
permutations of players, i.e., only the role of the player matters and not names or
indices assigned to players. Second, if a player’s added value to the coalition is
zero then this player should not get any profit from the coalition, or in other words
only players generating added value should share the benefits. Finally, the third
axiom requires additivity of payoffs: for any two characteristic functions Q 1 and Q 2
it must be that S (Q 1 Q 2 , N ) S (Q 1 , N ) S (Q 2 , N ).
The surprising result obtained by Shapley is that there is a unique equilibrium
payoff (called the Shapley value) that satisfies all three axioms.
Theorem 3. There is only one payoff function S that satisfies the three axioms.
It is defined by the following expressions for i N and all Q :
|S|!(|N ||S|1)!
S i (Q ) ¦
S N \i |N |!

Q S î` Q S .
The Shapley value assigns to each player his marginal contribution

Q S î` Q S when S is a random coalition of agents preceding i and the
ordering is drawn randomly. To explain further, (see Myerson [14]), suppose
players are picked randomly to enter into a coalition. There are |N|! different
orderings for all players, and for each set S that does not contain player i there are
|S|!(|N ||S|1)! ways to order players so that all of the players in S are picked
ahead of player i. If the orderings are equally likely, there is a probability of
|S|!(|N ||S|1)! |N |! that when player i is picked he will find S players in
the coalition already. The marginal contribution of adding player i to coalition S is
Q S î` Q (S ) . Hence, the Shapley value is nothing more than a marginal
(expected) contribution of adding player i to the coalition. Due to its uniqueness,
the concept of the Shapley value has found numerous applications in economics
and political sciences.
2.12 The Bargaining Game Model

To better understand the negotiational mechanism and theory, which will be shown
more specifically in the next chapter, we here consider the former approach to this
issue showing how to face the problem addressed by the bargaining in cooperative
game theory. In order to do this, consider a group of two or more agents facing
with a set of feasible outcomes, any one of which will be the result if it is accepted
by unanimous agreement of all participants. In the event that no unanimous
agreement is reached, a given disagreement outcome is the result. If the feasible
outcomes are such that each participant can do better than the disagreement
outcome, then there is an incentive to reach an agreement; however, so long as at
least two of the participants differ over which outcome is most preferable, there is
a need for bargaining and negotiation over which outcome should be agreed upon.
Note that since unanimity is required, each participant has the ability to veto any
outcome different from the disagreement outcome. To model this atomic
negotiation process, we use the cooperative bargaining process initiated by Nash
[27]. It is pertinent to mention that experimental bargaining theory indicates
stronger empirical evidence of this bargaining theory than any others. Nash
engaged in an axiomatic derivation of the bargaining solution. The solution refers
to the resulting payoff allocation that each of the participants unanimously agrees
upon. The axiomatic approach requires that the resulting solution should possess a
list of properties. The axioms do not reflect the rationale of the agents or the
process in which an agreement is reached but only attempts to put restrictions on
the resulting solution. Further, the axioms do not influence the properties of the
feasible set. Before listing the axioms, we will now describe the construction of the
feasible set of outcomes. Formally, Nash defined a two-person bargaining problem
(which can be extended easily to more than two players) as consisting of a pair
F , d where F is a closed convex subset of R 2 , and d d1 , d2 is a vector in
R 2 . F is convex, closed, non-empty, and bounded. Here, F, the feasible set,
represents the set of all feasible utility allocations and d represents the
disagreement payoff allocation or the disagreement point. The disagreement point
may capture the utility of the opportunity profit. Nash looked for a bargaining
solution, i.e., an outcome in the feasible set that satisfied a set of axioms. The
axioms ensure that the solution is symmetric (identical players receive identical
utility allocations), feasible (the sum of the allocations does not exceed the total
pie), Pareto optimal (it is impossible for both players to improve their utilities over
the bargaining solutions), the solution be preserved under linear transformations
and be independent of “irrelevant” alternatives. Due to constraints on space, the
reader can refer to Roth [28] for a very good description of the solution approach
and a more detailed explanation of the axioms. The remarkable result due to Nash
is that there is a bargaining solution that satisfies the above axioms and it is unique.
Theorem 4 [27]. There is a unique solution that satisfies all the “axioms”. This
solution, for every two-person bargaining game F , d is obtained by
solving: arg max ( x1 d1 )( x2 d2 ) .
x ( x1 , x2 )F , x t d
The axiomatic approach, though simple, can be used as a building block for
much more complex bargaining problems. Even though the axiomatic approach is
prescriptive, descriptive non-cooperative models of negotiation such as the Nash
demand game [29] and the alternating offer game [30], reach similar conclusions
as Nash bargaining. This somehow justifies the Nash bargaining approach to model
negotiations. In our discussion, we have only provided a description of the
bargaining problem and its solution between two players. However, this result can
easily be generalised to any number of players simultaneously negotiating for
allocations in a feasible set.
2.13 References
[1] von Neumann J, Morgenstern O (1944) Theory of games and economic behaviour.
Princeton University Press
[2] Nash JF (1950) Equilibrium points in n-person games. Proc Nat Acad Sci USA 36,
48–49
[3] Kuhn HW (1953) Extensive games and the problem of information. In Contributions
to the theory of games, vol II, Kuhn HW, Tucker AW, editors. Princeton University
Press. 193–216
[4] Aumann RJ (1959) Acceptable points in general cooperative n-person games. In

Contributions to the theory of games, vol IV, Kuhn HW, Tucker AW, editors.
Princeton University Press. 287–324
[5] Shubik M (1962) Incentives, decentralized control, the assignment of joint costs and
internal pricing. Manag Sci 8:325–343
[6] Vickrey W (1961) Counter speculation, auctions, and competitive sealed tenders. J
Finance, Vol.16, 8–37
[7] Shubik M (2002) Game theory and operations research: some musings 50 years
later. Oper Res 50:192–196
[8] Friedman JW (1986) Game theory with applications to economics. Oxford
University Press
[9] Fudenberg D, Tirole J (1991) Game theory. MIT Press.
[10] Varian H (1980) A model of sales. Am Econ Rev 70:651–659
[11] Lal R (1990) Price promotions: limiting competitive encroachment. Market Sci
9:247–262
[12] Selten R (1975) Reexamination of the perfectness concept for equilibrium points in
extensive games. Int J Game Theory 4:25–55
[13] Kreps D, Wilson R (1982) Sequential equilibria. Econometria 50:863–894
[14] Myerson RB (1997) Game theory. Harvard University Press.
[15] Border KC (1999) Fixed point theorems with applications to economics and game
theory. Cambridge University Press.
[16] Lederer P, Li L (1997) Pricing, production, scheduling, and delivery-time
competition. Oper Res 45:407–420
[17] Debreu D (1952) A social equilibrium existence theorem. Proc Nat Acad Sci USA
38:886–893
[18] Stackelberg H (1934) Markform and gleichgewicht. Vienna: Julius Springer.
[19] Selten R (1965) Spieltheoretische behaundlung eines oligopolmodells mit
nachfragetragheit. Z gesamte staatswiss 12:301–324
[20] Hall J, Porteus E (2000) Customer service competition in capacitated systems.
Manuf Serv Oper Manag 2:144–165
[21] van Mieghem J, Dada M (1999) Price versus production postponement: capacity and
competition. Manag Sci 45:1631–1649
[22] Aumann RJ (1989) Game theory. In Eatwell J, Milgate M, Newman P (eds), The
new Palgrave, New York, Norton, 8–9
[23] von Neumann J, Morgenstern O (1947) Theory of games and economic behaviour.
2nd ed, Princeton, NJ, Princeton University Press
[24] Gillies DB (1953) Some theorems on n-person games. PhD dissertation, Department
of Mathematics, Princeton University, Princeton, NJ
[25] Kahan JP, Rapoport A (1984) Theories of coalition formation, Lawrence Erlbaum,
Hillsdale, NJ
[26] Shapley LL (1953) A value for n-person games. In Kuhn HW, Tucker W (eds),
Contributions to the theory of games II. Ann Math Studies n.28. Princeton NJ,
Princeton University Press
[27] Nash JF (1951) Noncooperative games. Ann Math 54:286–295
[28] Roth A (1979) Axiomatic models in bargaining, Springer-Verlag
[29] Roth A (1995) Handbook of experimental economics, Princeton University Press
[30] Rubinstein A (1982) Perfect equilibrium in a bargaining model. Econometrica
50:97–110
3
Negotiation: an Overview
3.1 Introduction
The variety of involved disciplines and perspectives in the conflict resolution field
has created different terminologies, definitions, notations and formulations about
the concept of negotiation. As a result, interdisciplinary cooperation among the
relevant fields of study would suffer from inconsistencies and contradictions [1].
Yet, negotiation requires an interdisciplinary approach because of their
psychological, social and cultural character; economic, legal and political
considerations involved; quantitative and qualitative aspects; strategic, tactical and
managerial perspectives. Clearly, interdisciplinary approaches provide richer and
more comprehensive models of negotiators and negotiations. A schematic
representation of the different perspectives and influences on negotiation research
is presented in Figure 3.1: law and social sciences are the main contributors to the
prescriptive and descriptive models, heuristics and qualitative studies of
negotiation processes and negotiators’ behaviour [2, 3]; economics and
management science concentrated on the construction of formal models and
procedures of negotiations, rational strategies and the prediction of outcomes [4,
5]. Computer science and information systems contributions include construction
of electronic negotiation tables, decision and negotiation support systems (DSS,
NSS), artificial negotiating software agents (NSA) and software platforms for
bidding and auctioning [6–9]. The four arrows depicted in Figure 3.1 connect areas
of studies with the corresponding common results. The bi-directional arrow
indicates that the negotiation tools, agents and platforms are often based on the
results of the studies in economic and social sciences [10–13], and also that,
increasingly, computational models and systems influence the construction of
negotiation techniques, models and procedures [14, 15]. The latter phenomenon
marks a fairly recent development in negotiation research: increased computing
and networking power and findings from computer science and information
systems feed back into negotiation models and procedures giving them more
flexibility.
Computer Science Economic sciences and

Computational linguistics management
Data mining and KDD Econometrics
Artificial Intelligence Experimental economics
Distributed AI Management science
Autometed negotiation Decision science
Autonomous negotiation agents Bargaining theory
Negotiation expert systems Auction theory
Distributed negotiations Game theory
Negotiation software platforms Negotiation analysis
Negotiation
Negotiation media
procedures and
and systems: tools,
models: strategies,
agents and platforms
tactics and techniques
Information systems Law and social sciences

Decision support Law
Group and negotiation support Psychology
Workflows models Sociology
Electronic commerce Linguistics
Decision support systems Political sciences
Negotiation support systems Mediation and facilitation
Electronic negotiation tables Models of attitude and perceptions
Negotiation support agents Process models
Electronic markets Cultural influences
Electronic auctions Cognitive models
Figure 3.1. Negotiation research areas, their results and reciprocal key influences1
Most traditional negotiations have been conducted face-to-face; others have

been conducted using mail, fax and telephone. Mail-based and email-based
negotiations share many similarities in that they are difficult to manage, are time
consuming, and prone to misunderstanding [16]. Yet, the impact of information
technologies on negotiation is not limited to the use of electronic communication.
Information technology changes the way a negotiation problem can be represented
1
M. Bichler, G. Kersten, S. Strecker: Towards a Structured Design of Electronic
Negotiations. Group Decision and Negotiation, 2003, Vol. 12, No. 4, 311–335.
Negotiation: an Overview 27
and a negotiation process is structured. The use of Internet-based information

systems allows us to implement many more activities undertaken during the
negotiation process, including efficient matching of potential negotiators;
exchange, comparison and categorisation of rich data; and the use of tools for data
collection, problem structuring and analysis, and interpretation of offers. These
new possibilities have led to the birth of formal negotiation procedures and
protocols, which are a prerequisite to extending the use of rich and expressive
information technologies from the messages exchange to the other negotiation
process phases. Initiated by the commercial exploration of the Internet as a global
communication and “negotiation” infrastructure [17], electronic negotiation (e-
negotiation) has started to contribute in manifold ways-from web-based NSS [18],
to on-line auctions [19], to automated agent-based negotiations [20], in both
research studies and business applications [21]. Examples of new negotiation
protocols include auction protocols with combinatorial bids on product bundles
[22–24], automated negotiations among software agents [9, 25, 26] as well as
protocols supporting bi- and multi-lateral negotiations among human negotiators
[27, 28]. The computerisation of negotiation processes increasingly affects the way
organisations (and individuals) interact with each other. E-negotiation promises
higher levels of process efficiency and effectiveness, and most importantly, a
higher quality and faster agreement achievement. The potential economic impact
leads to an increased demand for appropriate electronic negotiations for specific
negotiation situations (e.g. electronic tenders for electronic procurement). Yet, both
the design of suitable electronic negotiation protocols and the implementation of
relevant electronic negotiation media largely lack systematic, traceable and
reproducible approaches and thus they remain more an art than a science. Recent
developments created an opportunity for mutual fertilisation of research studies
and approaches, and for integration of different perspectives on negotiations into
an interdisciplinary research effort to develop an engineering approach to
electronic negotiations, similar to, for example, system or process engineering
which brings together the findings about negotiators and negotiation processes
from the different research areas. The phenomenal growth in computer science and
information systems has led to the design of models and systems that often use
little from the behavioural and economic body of knowledge: people’s biases,
misconceptions and misunderstandings that have been shown to occur very often
are not considered and countered in negotiation design. There is thus a strong need
to consider the many results of economic and social sciences for the design and
development of negotiation media and systems. However, both the new methods
for data processing, representation and presentation, and many pragmatic
approaches to providing user-oriented support that have been proposed by
computer science and information systems should not be discarded by social
scientists. Hence, there should be, as shown in Figure 3.1, a feedback between
results of the two groups of negotiation research areas, that is, between the
negotiation media and systems, and the models and procedures. To achieve this a
multi-disciplinary approach towards the field of electronic negotiation is necessary,
one which is much broader and more encompassing than often implemented.
3.2 Negotiation and Rational Self-interested Agents

One the greatest challenges for computer science is to build computer systems that
can work together. The integration of automated systems has always been a
challenge, but as computers have become more sophisticated, the demands for
coordination and cooperation have become increasingly important and pervasive.
Cooperation and coordination are required for those complex problems that cannot
be solved by a single system in isolation, but require several systems to work
together interactively. Furthermore, there are heterogeneous intelligent systems
that were built in isolation that need to cooperate to achieve a new common goal.
In other situations, several autonomous systems may work in the same
environment, on different goals, and may need to coordinate their activities or to
share resources. They may also benefit from cooperation. Problems of coordination
and cooperation are not unique to automated systems; they exist at multiple levels
of activity in a wide range of populations. There have been several attempts to
define an agent [29–31]. For example Etzioni and Weld [32] require an agent to be
goal-oriented, collaborative, flexible, and capable of making independent decisions
on when to act. In addition, they determined that an agent should be a continuously
running process and be able to engage in complex communication with other
agents, including people. It should automatically customise itself to the preferences
of its user and to changes of the environment. The strategic-negotiation model was
developed to address problems in distributed artificial intelligence (DAI), an area
concerned with how automated agents can be constructed to interact in order to
solve problems effectively. Subrahmanian et al. [30] concentrate on the interaction
of an agent with other agents and the environment. They define a software agent as
a body of software that:
provides one or more useful services that other agents may use under
specified conditions;
includes a description of the service offered by the software, which may
be accessed and understood by agents;
includes the ability to act autonomously without requiring explicit
direction from a human being;
includes the ability to describe succinctly and declaratively how an agent
determines what actions to take even though this description may be kept
hidden from other agents, and
includes the ability to interact with other agents either in a cooperative or
in a adverse manner, as appropriate.
The agent designer faces two critical questions: what is the architecture of each
agent, and how do they interconnect, coordinate their activities, and cooperate?
There are many approaches to the development of a single agent. These approaches
can be divided into three principal categories [31]: deliberative, reactive and hybrid
architectures. A deliberative architecture is one that contains an explicitly
represented, symbolic model of the world, and one in which decisions (e.g., about
what actions to perform) are made via logical reasoning, based on pattern matching
and symbol manipulation. The main criticism of this approach is that the
computational complexity of symbol manipulation is very high, and some key
problems appear to be intractable. A reactive architecture is usually defined as one
that does not include any kind of central symbolic world model and does not use
any complex symbolic reasoning. These types of agents work efficiently when they
are faced with many “routine” activities. Many researchers suggest that neither a
completely deliberate nor a completely reactive approach is suitable for building
agents. They use hybrid systems, which attempt to combine the deliberate an the
reactive approaches.
3.3 Negotiation Models

Negotiation process implementation has been done using DAI application tools
both in distributed problem solving (DPS), where the agents are cooperative, and in
multi agent systems (MAS), where the agents are self-interested. Several studies in
DPS use negotiation for distributed planning and distributed search for possible
solutions for hard problems. For example, Conry et al. [33] suggest multi-stage
negotiation to solve distributed constraint satisfaction problems when no central
planner exists. Moehlman et al. [34] use negotiation as a tool for distributed
planning: each agent has certain important constraints, and it tries to find a feasible
solution using a negotiation process. They applied this approach in the Phoenix
fireman array. Lander and Lesser [35] use a negotiation search, which is a multi-
stage negotiation as a means of cooperation while searching and solving conflicts
among the agents. For the MAS environments, Rosenschein and Zlotkin [9]
identified three distinct domains where negotiation is applicable and found a
different strategy for each domain:
Task-oriented domain: the strategy consists in finding ways in which agents can
negotiate to come to an agreement, and allocating their tasks in a way that is
beneficial to everyone;
State-oriented domain: the strategy consists in finding actions which change the
state of the “world” and serve the agents’ goals; and
Worth-oriented domain: same strategy as State-Oriented Domain above, but, in
this domain, the decision is taken according to the maximum utility the agents gain
from the states.
Sycara [36, 37] presented a model of negotiation that combines case-based
reasoning and optimisation of multi-attribute utilities. In her work agents try to
influence the goals and intentions of their opponents. Kraus and Lehmann [28]
developed an automated diplomacy player that negotiates and plays well in actual
games against human players. Sierra et al. [38] present a model of negotiation for
autonomous agents to reach agreements about the provision of service by one agent
to another. Their model defines a range of strategies and tactics, borrowed from
good behavioural practice in human negotiation, that agents can employ to
generate offers and evaluate proposals. Zeng and Sycara [39] consider negotiation
in a marketing environment with a learning process in which the buyer and the
seller update their beliefs about the opponent’s reservation price using the
Bayesian rule. Sandholm and Lesser [40] discuss issues, such as levels of
commitment, that arise in automated negotiation among self-interested agents
whose rationality is bounded by computational complexity.
3.4 Underlying Principle for Electronic Negotiation

The computerisation of negotiation processes increasingly affects the way
customers, suppliers, and other business partners interact. Traditionally, firms
conducted negotiations with a counterpart in a bilateral manner: face-to-face, or
using media support. Such negotiations are difficult to manage, time consuming,
prone to misunderstanding, and require significant cognitive efforts [16].
Traditional negotiations suffer asymmetry of information about negotiated issues
(e.g. price transparency), acknowledged solution domain limitation (i. e. a shortage
of offers and counter-offers), an ex ante restricted number of potential counterparts
(since, for example, the human capacity for handling multiple telephone calls
simultaneously is limited) and high transaction costs (i.e. implicit transaction costs
such as large bid offer spreads and explicit transaction costs such as labour and
equipment) [41]. Those negotiation processes are rarely efficient and often lead to
inefficient compromises [42]. Meanwhile e-negotiation offers the possibility of
higher levels of process efficiency and effectiveness, including the exchange of
quantitatively and qualitatively improved information during the negotiation
process. Most importantly, e-negotiations promise higher quality and faster
achievement of negotiated agreements. This is because the participants can use
negotiation support tools that are able to support and advise negotiation actors who
can then formulate offers and counter-offers based on a wider knowledge about
negotiation issues. In traditional negotiations, the use of negotiation support tools
is difficult and awkward because all information exchanged face-to-face must be
entered manually during the process. The design of e-negotiation media, support
systems and software agents that matches the diversity of users, and the richness
and complexity of negotiation situations, requires categorisation and structuring of
the latter, and, also, specification of concepts and constructs. This effort led to the
creation of a taxonomy of electronic negotiations comprising types of processes
and terms used to describe different types in detail. Raiffa [43] in his seminal work
discourages “devising a taxonomy of disputes, in which the listing would be
reasonably exhaustive and in which overlaps among categories would be rare. This
was possible, I found, only after developing a host of abstract constructs — and
even then the taxonomy was not very useful”. Noting this caveat, it is obvious that
such efforts need to be made. This is because new information technologies are
increasingly being used to construct media for engagement in social and economic
processes such as negotiation in parallel and independently of the behavioural and
normative models of these processes. Results of social sciences should be taken
into account in the design of these media as well as their implications for the
processes themselves. In addition, a taxonomy allows for the establishment of a
common, unique terminology across disciplines, classification of models and
systems, identification of their possible extensions and for the identification of new
constructs and negotiation protocols.
3.5 Electronic Negotiation Protocols

The implementation of every model in an information system brings forth certain
rules of interaction that those who use this medium must follow. These rules need
to be specified so that agents (human or artificial) know the permissible set of
actions. An e-negotiation protocol is a negotiation protocol in which at least some
activities are supported or performed by information systems and the negotiation
steps are conducted by means of an electronic medium. The e-negotiation protocol
may be complex and with many rules governing the parties as they move through
different stages and phases of the negotiation process. For example, an e-
negotiation may begin with an auction and, after three winning bidders have been
identified, move on to a bilateral bargaining protocol among the three winners.
Typically, designers try to achieve certain goals for the outcome of a negotiation
and for the negotiation process itself, such as Pareto optimality of the result,
maximisation of the bid taker’s revenue/utility, stability, and speed of convergence
[44]. These objectives are achieved through:
specification of the structure of the negotiation problem and process;
specification of rules of feasible activities, and their sequencing and
timing; and
imposition of limitations on the form and content of information
exchange.
Every e-negotiation protocol restricts the negotiators’ freedom in order to meet one
or more of the above objectives. A closed e-negotiation protocol does not allow us
to add new rules throughout the negotiation. A closed negotiation protocol can
cover various negotiation situations but the set of rules is fixed and the rules cannot
be modified. Implementations of traditional auction formats such as the Dutch or
English auction are good examples of a closed e-negotiation protocol. An open e-
negotiation protocol does not contain all rules required for the negotiation; they
may be constructed by the participants or by mechanisms during the negotiation
process. In both cases, this involves learning about the participants, problem and
process; the results of learning are new rules that were not present prior to the e-
negotiation. Complex electronic negotiation protocols often involve a combination
of two or more different classes of negotiation protocols and thus exhibit the
characteristics of multiple negotiation models in either sequential or parallel
execution. For example, in financial markets continuous double auction protocols
have been combined with bilateral chat markets [45] where a trader can select an
offer and engage in a bilateral chat with the respective counterpart.
3.6 Characteristics that Differentiate Negotiations Protocols

Evaluation of the results of negotiations is not easy. Since the agents are self-
interested, when a negotiation is said to be successful we must ask “successful for
whom?” since each agent is concerned only with its own benefits or looses from
the resolution of the negotiation. Nevertheless, there are some parameters that can
be used to evaluate different protocols:
Distribution: the decision-making process should be distributed. There

should be no central unit or agent required to manage the process.
sSmmetry: the coordination mechanism should not treat agents differently
in light of non-relevant attributes. In the considered situations, the agents’
utility functions and their role in the encounter are the relevant attributes.
All other attributes, such as an agent’s name, characteristic, or
manufacturer, are not relevant. That is, symmetry implies that given a
specific situation, the replacement of an agent with another that is
identical with respect to the above attributes will not change the outcome
of the negotiation.
The following parameters can be used to evaluate the results of the negotiation:
Negotiation time: negotiations that end without delay are preferable to
negotiations that are time-consuming. This consideration is based on the
assumption that a delay in reaching an agreement causes an increase in the
cost of communication and computation time spent on the negotiation.
Efficiency: it is preferred that the outcome of the negotiations will be
efficient. It increases the number of agents that will be satisfied by the
negotiation results and the agents’ satisfaction levels from the negotiation
results. Thus it is preferable that the agents reach Pareto optimal
agreements.
Simplicity: negotiation processes that are simple and efficient are
preferable to complex processes. Being a “simple strategy” means that it
is feasible to build it into an automated agent. A “simple strategy” is also
one that an agent will be able to compute in a reasonable amount of time.
Stability: a set of negotiation strategies for a given set of agents is stable
if, given that all the other agents included in the set are following their
strategies, it is beneficial to an agent to follow its strategy too. Negotiation
stable strategies are more useful in multi-agent environments than
unstable ones: if there are stable strategies, we can recommend to all agent
designers to build only relevant strategies into their agents.
3.7 Modelling Approaches and Solution Concepts

Many formal approaches and solution concepts have already been utilised in
negotiation modelling mainly because of the variety of disciplines involved in the
negotiation study. Obviously, further work is needed regarding the comparison and
categorisation of numerous results from behavioural research, which concentrated
on such issues as the relationships between the social and individual context and
the negotiation process and its outcomes, the communication patterns,
interpretations and misinterpretations, and the relationships between individual
characteristics and the process. In the following the most relevant approaches and
solution concepts are briefly reported.
3.7.1 Decision Theory
Negotiations require that the participants make many decisions. The offers,
counter-offers and concessions they make in an effort to search for an agreement
result from the individual decisions they make. The issues of formulating and
solving decision problems, including the specification of feasible alternatives,
formulation of decision criteria and preferences, are the subject of decision theory.
There is a vast literature on decision theory, rationality and decision making. One
of the central tenets of decision theory is rationality and its use in the assessment of
decision alternatives and choice. It is well known that people often do not conform
to the rationality principles even if they are aware of breaking the axioms [46, 47].
Decision theorists assert that even if full rationality is not attainable, instrumental
rationality is, which requires a rigorous approach to gathering and using
information about the problem and about the decisionmaker [48]. This is perhaps
one of the most important contributions of decision theory which also is directly
applicable to negotiations. The three central topics of decision theory are: decision
alternatives, multiple conflicting objectives, and uncertainty of the decision
outcomes. Uncertainty of decision outcomes is an important issue which needs to
be considered in every decision problem, including negotiations. With the
exception of the participants having a different opinion about uncertainty, it is an
issue that is inherent to the problem and not subject to negotiation. The
specification of decision alternatives, either implicitly (i.e., with the use of
constraints) or explicitly, is performed during the structuring of the decision
problem. The development of the problem’s analytic structure involves the
consideration of decision attributes and objectives. The attributes, objectives and
alternatives are key aspects of negotiations; decision theory provides a number of
well-defined approaches for structuring [49] as well as specific techniques,
including decision trees, influence diagrams, and decision tables.
Many decisions are difficult because the decision maker has multiple and
conflicting objectives. Decision theory suggests the of use a preference elicitation
scheme and a function defined on the objectives and preferences [48]. This
function, under rationality conditions, determines the subjective value of a decision
alternative in the case of certainty or the utility, in the case of uncertainty. This
function is a measurement function defined on the set of alternatives and it allows
us to compare them. In particular it allows us to compare offers and counter-offers,
determine differences among them and select concessions that may increase the
utility for the counterpart while limiting the decrease of one’s own utility. The use
of utility functions allows for the achievement of efficient compromises providing
that the participants or a third party knows these functions and the set of feasible
alternatives. Given the set of feasible alternatives, it is the utility function that
assigns a unique value to each alternative, making the choice of the optimal
alternative automatic. Because of the importance of utility, decision analysts
propose that this function be constructed carefully and with the help of experts.
They apply a preference elicitation and utility construction approach so that the
assumptions underlying the multi-attribute utility theory (MAUT) are met. The
strict requirements of MAUT and difficulties in utility construction led to the use
of proxy functions neglecting rationality principles.
3.7.2 Game Theory
Game theory is the most rigorous approach towards conflict resolution and allows
for formal problem analysis and the specification of well-defined solutions. Game-
theoretical bargaining models assume rationality of agents and that each agent’s
choice can be uniquely described by utility function. Linhart [50] summarises
several restrictions regarding game-theoretical models of bargaining. For example,
human agents are often constrained by their bounded rationality, which is one
reason why game-theoretical models of bargaining do not fit in explaining real
human negotiating behaviour. Yet, game-theoretical models of bargaining are seen
as potential candidates for automated negotiations among software agents.
3.7.3 Negotiation Analysis
The limitations of game theory are well known and rooted in its normative
orientation caused mainly by the strict rationality assumptions that are not met in
reality. Therefore, game-theoretica models do not allow for prescriptive advice that
is sought by negotiators and their advisers. From the analyst’s perspective it may
be reasonable to assume that the negotiator whom the analyst advises is rational
however; it is not reasonable to assume the same about the opponent. This
observation, together with a weakened rationality assumption, provided the basis
for negotiation analysis [51]. With the objective of providing advice to one party,
negotiation analysis takes prescriptive/descriptive orientation in that it assumes
rationality of one party but not necessarily of the other [52]. Negotiation analysis
integrates decision analysis and game theory in order to provide formal and
meaningful support. The goal of negotiation analysis is to bridge the gap between
descriptive qualitative models and normative game-theoretical models of
bargaining. It adopts a number of behavioural concepts and incorporates them in
quantitative models. This way significantly extends the expressiveness of the
models and their capability of describing various negotiation situations. It also
allowed analysts and advisors to conduct formal analysis of negotiations and to
provide support. Negotiation analysis tends to adopt the application of game-
theoretical solution concepts or efforts to find unique equilibrium outcomes.
Instead, negotiation analysts generally focus on changes in perceptions in the zone
of possible agreement and the distribution of possible negotiated outcomes,
conditional on various actions [51, 53]. The contributions of negotiation analysis
include: a subjective perspective on the process and outcomes, concentration on the
possible agreements rather than search for one equilibrium point, and acceptance
of goal-seeking rather than game-theoretical rationality. This makes an asymmetric
perspective possible [44, 52]. Other approaches have a symmetrical orientation:
behavioural studies focus on descriptions of the parties and their interactions, game
theory and optimisation assume that the parties are rational hence they have
symmetrically prescriptive orientation. In contrast, negotiation analysis is used to
generate prescriptive advice to the supported party given a descriptive assessment
of the opposing parties. In other words, negotiation analysis reconciled several
important concepts of behavioural research and game theory.
3.8 Strategic Negotiation

The strategic-negotiation model is based on Rubinstein’s model of alternating
offers [54]. In the strategic model there are N agents, Agents ^ A1 ,..., AN ` . The
agents need to reach an agreement on a given issue. It is assumed that the agents
can take actions in the negotiation only at certain times in the set T ^0,1, 2...` that
are determined in advance and are known to the agents. In each period t T of the
negotiation, if the negotiation has not terminated earlier, the eligible agent makes
an offer suggesting a possible agreement (with respect to the specific negotiation
issue), and each of the other agents may either accept the offer (choose Yes), reject
it (choose No), or opt out of the negotiation (choose Opt). If an offer is accepted by
all the agents (i.e., all of them choose Yes), then the negotiation ends, and this offer
is implemented. If at least one of the agents opts out of the negotiation, then the
negotiation ends and a conflictual outcome results. If no agent has chosen “Opt”,
but at least one of the agents has rejected the offer, the negotiation proceeds to
period t + 1, and the next eligible agent makes a counter-offer, the other agents
respond, and so on. We assume that an agent responding to an offer is not informed
of the other responses during the current negotiation period. We call this protocol a
simultaneous response protocol. In the strategic-negotiation model there are no
rules which bind the agents to any specific strategy. We do not make any
assumptions about the offers the agents make during the negotiation. In particular,
the agents are not bound to any previous offers that have been made. After an offer
is rejected, an agent can choose whether to make the same offer again, or to
propose a new offer. The protocol only provides a framework for the negotiation
process and specifies the termination condition, but there is no limit on the number
of periods. A fair and reasonable method for deciding on the order in which agents
will make offers is to arrange them randomly in a specific order before the
negotiation begins. The agents’ time preferences and the preferences between
agreements and opting out are the driving force of the model. They will influence
the outcome of the negotiation. In particular, agents will not reach an agreement
which is not at least as good as opting out for all of them. Otherwise, the agent who
prefers opting out over the agreement, will opt out.
3.9 Negotiation Strategies

In the literature special cases of negotiation strategies are discussed [55, 56]. One
important family of strategies consists of functions which are only time-dependent,
i.e. strategy changes as the time goes on. Other families of strategies which could
be employed by the agents are resource-dependent and behaviour-dependent
strategies. In a resource-dependent strategy the scarcer the resource the more
urgent the need for an agreement to be reached. Taking a behaviour-dependent
strategy an agent has to analyze the behaviour of other agents involved in the
negotiation and generate his counter-offers according to the dynamics of how the
values are changed in the offers he has received. In general, a strategy of an agent
can be very complex, switching over time between time-dependent, resource-

dependent, behaviour-dependent, and other strategies. From a strategic point of
view, an agent’s negotiation strategy specifies for the agent what to do next, for
each sequence of offers s0 ,..., st . In other words, for the agent whose turn it is to
make an offer, it specifies which offer to make next. That is, it indicates to the
agent which offer to make at t 1 , if in periods 0 until t the offers s0 ,..., st had been
made and were rejected by at least one of the agents, but none of them has opted
out. Similarly, in time periods when it is the agent’s turn to respond to an offer, the
strategy specifies whether to accept the offer, reject it or opt out of the negotiation.
A strategy profile is a collection of strategies, one for each agent [57]. The main
question is how a rational agent will choose its negotiation strategy. A useful
notion is the Nash equilibrium [58, 59]. This means that if all the agents use the
strategies specified for them in the strategy profile of the Nash equilibrium, then no
agent has a motivation to deviate and use another strategy. However, the use of the
Nash equilibrium in a model of alternating offers leads to absurd Nash equilibria
[60]: an agent may use a threat that would not be carried out if the agent were put
in the position to do so, since the threat fulfilment would give to the agent lower
payoff than it would get by not undertaking the threatened action. This is because
Nash equilibrium strategies may be in equilibrium only in the beginning of the
negotiation, but may be unstable in intermediate stages. The concept of subgame
perfect equilibrium [57], which is a stronger concept, is defined as a strategy
profile if the strategy profile induced in every subgame is a Nash equilibrium of
that subgame. This means that at any step of the negotiation process, no matter
what the history is, no agent has motivation to deviate and use any strategy other
than that defined in the strategy profile. In situations of incomplete information
there is no proper subgame. The sequential equilibrium [61], which takes the
beliefs of the agents into consideration, can be used in the incomplete information
situations.
3.10 References
[1] Gulliver PH (1979) Disputes and negotiations: a cross-cultural perspective. Orlando,
FL, Academic
[2] Bell DE, Raiffa H (1991) Decision making: descriptive, normative, and prescriptive
interactions. Cambridge, Cambridge University
[3] Pruitt DG (1981) Negotiation behaviour. New York, Academic
[4] Nash JF (1950) The bargaining problem. Econometrica 18: 155–162
[5] Roth AE (1979) Axiomatic models of bargaining. Berlin, Springer-Verlag
[6] Holsapple CW, Lai H (1998) A formal basis for negotiation support system research.
Group Decis Negot 7(3): 199–202
[7] Kersten GE (1997) Support for group decisions and negotiations: an overview. In
multicriteria analysis, Climaco J (ed), Springer Verlag, pp 332–346
[8] Maes P, Guttman R (1999) Agents that buy and sell: transforming commerce as we
know it. Commun ACM 42(3): 81
[9] Rosenschein JS, Zlotkin G (1994) Rules of encounter: designing conventions for
automated negotiation among computers. Cambridge, MA, MIT Press
[10] Bui T, Yen J (2001) A multi-attribute negotiation support system with market
signaling for electronic markets. Group Decis Negot 10(6): 515–537
[11] Hamalainen RP (1995) Special issue on dynamic game modeling in bargaining and
environmental negotiations. Group Decis Negot 4(1)
[12] Raiffa H (1996) Lectures on negotiation analysis. Cambridge, MA, PON
[13] Teich JE, Wallenius H. (1994) Advances in negotiation science. Trans Oper Res 6:
55–94
[14] Kumar M, Feldman S (1998) Internet auctions. Proc 3rd USENIX Workshop on
Electronic Commerce. Boston, MA: 49–60
[15] Pennock DM, Horvitz E (2000). Collaborative filtering by personality diagnosis: a
hybrid memory and model-based approach. Proc 16th conference on uncertainty in
artificial intelligence, San Francisco, USA
[16] Thompson L (2001) The mind and heart of the negotiator. Upper Saddle River, NJ,
Prentice Hall
[17] Raisch W (2000) The emarketplace: strategies for success in B2B ecommerce,
McGraw-Hill
[18] Shim J Hsiao N (1999) A literature review on web-based negotiation support system.
Documentation for web-based negotiation training
[19] Vakali A, Angelis L (2001) Internet based auctions: a survey on models and
applications. SIGe-com, Exch Newsl ACM Spec Interest Group E-Commerce 2(2): 6–
15
[20] Jennings NR, Faratin P (2001). Automated negotiations: prospects, methods and
challenges. Group Decis Negot 10(2): 199–215
[21] Edwards J (2001) Working the wiggle room. Line 56(April): 50–55
[22] Parkes D (1999) iBundle: an efficient ascending price bundle auction. Proc 1st ACM
conference on electronic commerce (EC–99): 148–157
[23] Rassenti S, Smith VL (1982) A combinatorial auction mechanism for airport time slot
allocations. Bell J Econ 13: 402–417
[24] Rothkopf MH, Pekec A (1998). Computationally manageable combinatorial auctions.
Manag Sci 44(8): 1131–1147
[25] Guttman R (1998) Agent-mediated integrative negotiation for retail electronic
commerce. Proc workshop on agent-mediated electronic trading (AMET'98)
[26] Wurman P, Wellman M (1998). The michigan internet auctionbot: a configurable
auction server for human and software agents. Proc 2nd int conf on autonomous agents
(Agents–98)
[27] Stroebel M (2002). A design and implementation framework for multi-attribute
negotiation: intermediation in electronic markets. University of St. Gallen,
Switzerland.
[28] Kraus S, Lehmann D (1995) Designing and building a negotiating automated agent.
Comput Intell 11(1):132–171
[29] Teich JE, Wallenius H (2001) Designing electronic auctions: an internet-based hybrid
procedure combining aspects of negotiations and auctions. Electron Commerce Res
1(1): 301–314
[30] Subrahmanian VS, Bonatti P Dix J, Eiter T, Kraus S, Ozean F, Ross R (2000).
Heterogenous agent systems: theory and implementation. MIT Press, Cambridge,
Massachusetts
[31] Wooldrige MJ, Jennings NR (1995). Intelligent agents. Springer-Verlag, Berlin
[32] Etzioni O, Weld DS (1995) Intelligent agents on the internet: fact, fiction and forecast.
IEEEE Expert 10(4):44–49
[33] Conry SE, Kuwabara K, Lesser VR, Meyer RA (1991) Multistage negotiation for
distributed satisfaction. IEEE Trans Syst Man Cybern, Special issue on distributed
artificial intelligence, 21(6):1462–1477
[34] Moehlman T, Lesser V, Buteau B (1992) Decentralized negotiation: An approach to

the distributed planning problem. Group Decision and Negotiation, 2:161–191
[35] Lander SE, Lesser VR (1992). Customizing distributed search among agents with
heterogeneous knowledge. Proc. 1st int conf information knowledge management
Baltimore pp 335–344
[36] Sycara KP (1987) Resolving adversarial conflicts: an approach to integrating case-
based and analytic methods. PhD thesis, School of Information and Computer Science,
Georgia Institute of Technology
[37] Sycara KP (1990). Persuasive argumentation in negotiation. Theory Decis, 28:203–
242
[38] Sierra C, Faratin P, Jennings N (1997) A service-oriented negotiation model between
autonomous agents. Proc. 8th european workshop on modeling autonomous agents in
a multi-agent world (MAAMAW–97), Ronneby, Sweden pp 17–35
[39] Zeng D., Sycara K (1998) Bayesian learning in negotiation. Int J Hum Comput Studi,
48:125–141
[40] Sandholm TW, Lesser VR (1995) Issues in automated negotiation and electronic
commerce: extending the contract net framework. 1st int conf multiagent systems
(ICMAS–95), San Fransisco pp 328–335
[41] Weinhardt C, Gomber P (1999) Agent-mediated off-exchange trading. Proc 32nd
Hawaii conference on system sciences
[42] Kersten G, Mallory G (1999) Rational inefficient compromises in negotiations.
Ottawa, Canada, Interneg, Carleton University
[43] Raiffa H (1982) The Art and Science of Negotiation. Cambridge, MA, Harvard
University Press
[44] Jackson MO (2000) Mechanism theory. Pasadena, CA, Humanities and social
sciences, California institute of technology
[45] Budimir M, Holtmann C (2001) The design of innovative securities markets: the case
of asymmetric information. In e-Finance: innovative problemlösungen für
Informationssysteme in der Finanzwirtschaft. Buhl HU, Kreyer N, Steck W (eds)
Berlin, Heidelberg, New York, Springer: 175–196
[46] Kahneman D, Slovic P (1982) Judgement under uncertainty: heuristics and biases.
Cambridge, MA:, Cambridge Univ Press
[47] Nozick R (1993) The nature of rationality. Princeton, NJ, Princeton University Press.
[48] French S (1998) Decision theory: an introduction to the mathematics of rationality.
New York, Ellis Hoerwood
[49] Keeney R (1992) Value-focused thinking: a path to creative decision making.
Cambridige, Harvard University Press
[50] Linhart PB, Radner R (1992). Bargaining with incomplete information. San Diego,
CA, Academic Press
[51] Sebenius JK (1992) Negotiation analysis: a characterization and review. Manag Sci
38(1): 18–38
[52] Young HP (1991) Negotiation analysis. Ann Arbor, The University of Michigan Press
[53] Clyman DR (1995) Measures of joint performance in dyadic mixed-motive
negotiations. Organ Behav Hum Decis Proces 64(1): 38–48
[54] Rubinstein A (1982) Perfect equilibrium in a bargaining model. Econometrica,
50(1):97–109
[55] Jennings NR, Parsone S, Sierra C, Faratin P (2000) Automated negotiation. Proc 5th
int conf on the practical application of intelligent agents and multi-agent technology
(PAAM–2000), pp 23–30
[56] Teuteberg F, Kurbel K (2002) Anticipating agents negotiation strategies in an e-
marketplace using belief models. Proc 5th int conf business information systems BIS
2002, Poznan, Poland
[57] Osborne MJ, Rubinstein A (1994) A course in game theory. MIT Press, Cambridge,
Massachusetts
[58] Nash JF (1953) Two-person cooperative games. Econometrica, 21:128–140
[59] Luce RD, Raiffa H (1957) Games and decisions. Wiley, New York
[60] Tirole J (1988) The theory of industrial organization. MIT Press, Cambridge, MA
[61] Kreps D, Wilson R (1982) Sequential equilibria. Econometrica, 50:863–894
This page intentionally blank
4
Multiple-agent Systems: an Overview
4.1 Introduction
Most researchers in artificial intelligence (AI) to date have dealt with developing
theories, techniques, and systems to study and understand the behaviour and
reasoning properties of a single cognitive entity. AI has matured, and it endeavours
to face more complex, realistic, and large-scale problems. Such problems are
beyond the capabilities of an individual agent. The capacity of an intelligent agent
is limited by its knowledge, its computing resources, and its perspective. This
bounded rationality is one of the underlying reasons for creating problem-solving
organisations. The most powerful tools for handling complexity are modularity and
abstraction. Multi-agent systems (MASs) offer modularity. If a problem domain is
particularly complex, large, or unpredictable, then the only way it can reasonably
be addressed is to develop a number of functionally specific and (nearly) modular
components (agents) that are specialised at solving a particular problem aspect.
This decomposition allows each agent to use the most appropriate paradigm for
solving its particular problem. When interdependent problems arise, the agents in
the system must coordinate with each other one to ensure that interdependencies
are properly managed. Research in MASs is concerned with the study, behaviour,
and construction of a collection of possibly pre-existing autonomous agents that
interact with each other and their environments. Study of such systems goes
beyond the study of individual intelligence to consider, in addition, problem
solving that has social components. A MAS can be defined as a loosely coupled
network of problem solvers that interact to solve problems that are beyond the
individual capabilities or knowledge of each problem solver [1]. These problem
solvers, often called agents, are autonomous and can be heterogeneous or
homogenous. The characteristics of MASs are that:
each agent has incomplete information or capabilities for solving the
problem and, thus, has a limited viewpoint;
there is no system global control;
data are decentralised; and
computation is asynchronous.
The motivations for the increasing interest in MAS research include the ability
of MASs to do the following.
First, MASs are able to solve problems that are too large for a centralized
systems because of resource limitations or the sheer risk of having one centralized
system that could be a performance bottleneck or could fail at critical times.
Second, MASs can allow interconnection and interoperation of multiple
existing legacy systems. To keep pace with changing business needs, legacy
systems must periodically be updated. Completely rewriting such software tends to
be prohibitively expensive and is often simply impossible. Therefore, in the short
to medium term, the only way that such legacy systems can remain useful is to
incorporate them into a wider cooperating agent community in which they can be
exploited by other pieces of software.
Third, MASs provide solutions to problems that can naturally be regarded as a
society of autonomous interacting components agents. For example, in meeting
scheduling, a scheduling agent that manages the calendar of its user can be
regarded as autonomous and interacting with other similar agents that manage
calendars of different users [2, 3]. Such agents also can be customised to reflect the
preferences and constraints of their users. Other examples include air-traffic
control [4] and multi-agent bargaining for buying and selling goods on the internet.
Fourth, MASs could provide solutions that efficiently use information sources
that are spatially distributed. Examples of such domains include sensor networks
[5], seismic monitoring [6], and information gathering from the Internet [7].
Fifth, MASs could provide solutions in situations where expertise is distributed.
Examples of such problems include concurrent engineering [8], health care, and
manufacturing.
Sixth, MASs could enhance performance along the dimensions of
computational efficiency because concurrency of computation is exploited (as long
as communication is kept minimal, for example, by transmitting high-level
information and results rather than low-level data); reliability, that is, graceful
recovery of component failures, because agents with redundant capabilities or
appropriate interagent coordination are found dynamically (for example, taking up
responsibilities of agents that fail); extensibility because the number and the
capabilities of agents working on a problem can be altered; robustness, the
system’s ability to tolerate uncertainty, because suitable information is exchanged
among agents; maintainability because a system composed of multiple components
agents is easier to maintain because of its modularity; responsiveness because
modularity can handle anomalies locally, not propagate them to the whole system;
flexibility because agents with different abilities can adaptively organise to solve
the current problem; and reuse because functionally specific agents can be reused
in different agent teams to solve different problems.
4.2 Applications
MASs are now a research reality and are rapidly having a critical presence in many
human computer environments. A very challenging application domain for MASs
technology is the Internet. Today the Internet has developed into a highly
Multiple-agent Systems: an Overview 43
distributed open system where heterogeneous software agents come and go, there
are no well-established protocols or languages on the agent level (higher than
TCP/IP), and the structure of the network itself keeps on changing. In such an
environment, MAS technology can be used to develop agents that act on behalf of
a user and are able to negotiate with other agents in order to achieve their goals.
Auctions on the Internet and electronic commerce are such examples [9, 10]. One
can also think of applications where agents can be used for distributed data mining
and information retrieval. MASs can also be used for traffic control where agents
(software or robotic) are located in different locations, receive sensor data that are
geographically distributed, and must coordinate their actions in order to ensure
global system optimality [11]. Other applications are in social sciences where MAS
technology can be used for simulating interactivity and other social phenomena
[12], in robotics where a frequently encountered problem is how a group of robots
can localise themselves within their environment [13], and in virtual reality and
computer games where the challenge is to build agents that exhibit intelligent
behaviour [14].
4.3 Challenging Issues

The transition from single-agent systems to MASs offers many potential
advantages but also raises challenging issues. Some of these are:
how to decompose a problem, allocate subtasks to agents, and synthesise
partial results;
how to handle the distributed perceptual information. How to enable
agents to maintain consistent shared models of the world;
how to implement decentralised control and build efficient coordination
mechanisms among agents;
how to design efficient multi-agent planning and learning algorithms;
how to represent knowledge. How to enable agents to reason about the
actions, plans and knowledge of other agents;
how to enable agents to communicate. What communication languages
and protocols to use. What, when, and with whom should an agent
communicate;
how to enable agents to negotiate and resolve conflicts;
how to enable agents to form organisational structures like teams or
coalitions;
how to assign roles to agents;
how to ensure coherent and stable system behaviour.
Clearly the above problems are interdependent and their solutions may affect each
other. For example, a distributed planning algorithm may require a particular
coordination mechanism, learning can be guided by the organisational structure of
the agents, and so on. In the following sections we will try to provide answers to
some of the above questions.
4.4 Individual Agent Reasoning

Sophisticated individual agent reasoning can increase MAS coherence because
each individual agent can reason about non-local effects of local actions, form
expectations of the behaviour of others, or explain and possibly repair conflicts and
harmful interactions. Numerous works in AI research try to formalise a logical
axiomatisation for rational agents. This axiomatisation is accomplished by
formalising a model for agent behaviour in terms of beliefs, desires, goals, and so
on. An agent will be considered rational if it always selects an action that
optimises an appropriate performance measure, given what the agent knows so far.
The performance measure is typically defined by the user (the designer of the
agent) and reflects what the user expects from the agent in a specific task. The
problem of optimal decision making of an agent was first studied in optimal
control [15]. For the purpose of this discussion, it will be considered a discrete set
of time steps t 1,..., n , in each of which the agent must choose an action from a
finite set of actions A that it has available. Intuitively, in order to act rationally, an
agent should take both the past and the future into account when choosing an
action. The past refers to what the agent has perceived and what actions it has
taken until time t, and the future refers to what the agent expects to perceive and do
after time t. If we denote by oW the perception of an agent at time W , then the above
implies that in order for an agent to optimally choose an action at time t, it must in
general use its complete history of perceptions oW and actions aW for W d t. The
function
S ( o1 , a1 , o2 , a2 ,..., ot ) at
that maps the complete history of perception action pairs up to time t to an optimal
action at is called the policy of the agent. As long as we can find a function S that
implements the above mapping, the part of optimal decision making that refers to
the past is solved. However, defining and implementing such a function is
problematic; the complete history can consist of a very large (even infinite) number
of perception action pairs, which can vary from one task to another. Merely storing
all perceptions would require very large memory, aside from the computational
complexity for actually computing S . This fact calls for simpler policies. One
possibility is for the agent to ignore all its percept history except for the last
perception ot . In this case its policy takes the form
S ( ot ) at
which is a mapping from the current perception of the agent to an action. An agent
that simply maps its current perception ot to a new action at , thus effectively
ignoring the past, is called a reflex agent, and its policy is called reactive or
memoryless.
4.5 Observable Worlds

From the discussion above it is clear that the terms agent and environment are
coupled, so that one cannot be defined without the other. In fact, the distinction
between an agent and its environment is not always clear, and it is sometimes
difficult to draw a line between these two [16]. To simplify things it will be
assumed hereafter the existence of a world in which one or more agents are
embedded, and in which they perceive, think, and act. The collective information
that is contained in the world at any time step t, and that is relevant for the task at
hand, will be called a state of the world and denoted by st . The set of all states of
the world will be denoted by S. Depending on the nature of problem, a world can
be either discrete or continuous. A discrete world can be characterised by a finite
number of states. Examples are the possible board configurations in a chess game.
On the other hand, a continuous world can have infinitely many states. A
fundamental property that characterises a world from the point of view of an agent
is related to the perception of the agent. We will say that the world is (fully)
observable to an agent if the current perception ot of the agent completely reveals
the current state of the world, that is, st ot . On the other hand, in a partially
observable world the current perception ot of the agent provides only partial
information about the current world state in the form of a conditional probability
distribution P(st |ot ) over states. This means that the current perception ot does
not fully reveal the true world state, but to each state st the agent assigns
probability P(st |ot ) that st is the true state (with 0 d P(st |ot ) d 1 and
¦ st S
P(st |ot ) 1). Here we treat st as a random variable that can take all
possible values in S. Partial observability can be attributed, for instance, to an
inherent property of the environment referred to as perceptual aliasing: different
states may produce identical perceptions to the agent at different time steps. In
other words, two states may look the same to an agent, although the states are
different from each other. Partial observability is much harder to handle than full
observability, and algorithms for optimal sequential decision making in a partially
observable world can easily become intractable [17].
4.6 Stochastic Transitions and Utilities
As mentioned above, at each time step t the agent chooses an action at from a
finite set of actions A. When the agent takes an action, the world changes as a
result of this action. A transition model (sometimes also called world model)
specifies how the world changes when an action is executed. If the current world
state is st and the agent takes action at , we can distinguish the following two
cases:
in a deterministic world, the transition model maps a state action pair

( st ; at ) to a single new state st 1 . In chess for example, every move
changes the configuration on the board in a deterministic manner;
in a stochastic world, the transition model maps a state action pair ( st ; at )
to a probability distribution P(st 1 |st , at ) over states. As in the partial
observability case above, st 1 is a random variable that can take all
possible values in S, each with corresponding probability P(st 1 |st , at ) .
Most real-world applications involve stochastic transition models, for
example, robot motion is inaccurate because of wheel slip and other
effects.
In classical AI, a goal for a particular task is a desired state of the world.
Accordingly, planning is defined as a search through the state space for an optimal
path to the goal. When the world is deterministic, planning comes down to a graph
search problem for which a variety of methods exist, see for example [17]. In a
stochastic world, however, planning can not be done by simple graph search
because transitions between states are non-deterministic. The agent must now take
the uncertainty of the transitions into account when planning. To see how this can
be realised, note that in a deterministic world an agent prefers by default a goal
state to a non-goal state. More generally, an agent may hold preferences between
any world states. A way to formalise the notion of state preferences is by assigning
to each state s a real number U(s) that is called the utility of state s for that
particular agent. Formally, for two states s and s’ U(s) > U(s’) holds if and only if
the agent prefers state s to state s’, and U(s) = U(s’) if and only if the agent is
indifferent between s and s’. Intuitively, the utility of a state expresses the
desirability of that state for the particular agent; the larger the utility of the state,
the better the state is for that agent. Equipped with utilities, the question now is
how an agent can efficiently use them for its decision making. Let us assume that
the world is stochastic with transition model P(st 1 |st , at ), ,and is currently in state
st , while an agent is pondering how to choose its action at . Let U(s) be the utility
of state s for the particular agent (we assume there is only one agent in the world).
Utility-based decision making is based on the premise that the optimal action a* of t
the agent should maximise expected utility, that is,

a * arg max ¦ P(st 1 |st , at )
t
at A
st 1
where it is summed over all possible states st 1 S the world may transition to,
given that the current state is st and the agent takes action at . In words, to see
how good an action is, the agent has to multiply the utility of each possible
resulting state with the probability of actually reaching this state, and sum up the
resulting terms. Then the agent must choose the action a * that gives the highest
t
sum. If each world state has a utility value, then the agent can do the above
calculations and compute an optimal action for each possible state. This provides
the agent with a policy that maps states to actions in an optimal sense. In particular,
given a set of optimal (i.e., highest attainable) utilities U*(s) in a given task, the
greedy policy
S * (s ) arg max ¦ P(s '|s , a)U * (s ')
a
s'
is an optimal policy for the agent.
4.7 Distributed Decision Making

As previously discussed, a distinguishing feature of a multi-agent system is the fact
that the decision making of the agents can be distributed. This means that there is
no central controlling agent that decides what each agent must do at each time step,
but each agent is to a certain extent responsible for its own decisions. The main
advantages of such a decentralised approach over a centralised one are efficiency,
due to the asynchronous computation, and robustness, in the sense that the
functionality of the whole system does not rely on a single agent. In order for the
agents to be able to take their actions in a distributed fashion, appropriate
coordination mechanisms must be additionally developed. Coordination can be
regarded as the process by which the individual decisions of the agents result in
good joint decisions for the group. A typical situation where coordination is needed
is among cooperative agents that form a team, and through this team they make
joint plans and pursue common goals. In this case, improving coherence by
planning their actions ensures that the agents do not obstruct each other when
taking actions, and moreover that these actions serve the common goal. Planning
for a single agent is a process of constructing a sequence of actions considering
only goals, capabilities, and environmental constraints. However, planning in a
MAS environment also considers the constraints that the other agents’ activities
place on an agent’s choice of actions, the constraints that an agent’s commitments
to others place on its own choice of actions, and the unpredictable evolution of the
world caused by other unmodeled agents. Most early work in distributed artificial
intelligence (DAI) has dealt with groups of agents pursuing common goals (for
example, [18-21]). Agent interactions are guided by cooperation strategies meant
to improve their collective performance. Most work on multi-agent planning
assumes an individual sophisticated agent architecture that enables them to do
rather complex reasoning. Early work on distributed planning took the approach of
complete planning before action. To produce a coherent plan, the agents must be
able to recognise subgoal interactions and avoid them or resolve them. Another
direction of research in cooperative multi-agent planning has been focused on
modelling teamwork explicitly. Explicit modelling of teamwork is particularly
helpful in dynamic environments where team members might fail or be presented
with new opportunities. In such situations, it is necessary that teams monitor their
performance and reorganise based on the situation. In many practical applications,
however, we have to deal with self-interested agents, for example, agents that act
on behalf of some owner that wants to maximise his or her own profit. A typical
example is a software agent that participates in an electronic auction. Developing
an algorithm or protocol for such a system is a much more challenging task than in
the cooperative case. First, an agent has to be motivated to participate in the
protocol, which is not a priori the case. Second, we have to take into account the
fact that an agent may try to manipulate the protocol for its own interest, leading to
suboptimal results. The latter includes the possibility that the agent may lie, if
needed. The development of protocols that are stable (non-manipulable) and
individually rational for the agents is the subject of mechanism design or
implementation theory.
4.8 Recognising and Resolving Conflicts

Because MASs lack global viewpoints, global knowledge, and global control, there
is the potential for disparities and inconsistencies in agents’ goals, plans,
knowledge, beliefs, and results. To achieve coherent problem solving, these
disparities must be recognised and resolved. Disparities can be resolved by making
an agent omniscient so it can see the states of all agents and determine where the
disparities lie and how to resolve them. This approach is limiting because it makes
this agent a bottleneck and a single point of failure. To detect and correct
disparities and conflicts using only local perspective is difficult. To facilitate
detection and resolution of conflicts, agents can rely on models of the world and
other agents. Disparity resolution can be influenced by the organisational structure
of the agent society and an agent’s role within it, the kinds of models an agent has,
and the agent’s reasoning algorithms. The main approach for resolving disparities
in a MAS is negotiation. Negotiation is seen as a method for coordination and
conflict resolution (for example, resolving goal disparities in planning, resolving
constraints in resource allocation, resolving task inconsistencies in determining
organisational structure). Negotiation has also been used as a metaphor for
communication of plan changes, task allocation, or centralised resolution of
constraint violations. The main characteristic of negotiation that is necessary for
developing applications in the real world is the presence of some sort of conflict
that must be resolved in a decentralised manner by self-interested agents under
conditions of bounded rationality and incomplete information. Furthermore, the
agents communicate and iteratively exchange proposals and counterproposals.
4.9 Communicating Agents

Multi-agent interaction is often associated with some form of communication that
involves several levels of abstraction. On the lower, network level, one would like
to make sure that the messages that are communicated among the agents arrive
safely and promptly at their destination. For that, several well-studied formalisms
and protocols exist in the distributed systems literature [22]. On an intermediate,
language level, one would like to have a basic set of language primitives and a
standardised format for exchanging these primitives, so that agents that speak the
same language can easily understand each other. Finally, on a high, application
level, one would like to effectively use communication for solving standard multi-
agent problems, like coordination or negotiation. A formal way to describe
communication is by treating each communication primitive as an action that

updates the knowledge of an agent about aspects of the state like those described
above. The communication primitives that are exchanged among agents are
typically referred to as communicative acts or speech acts. Some of the most
common types of communicative acts are the following:
informing about some aspects of the current state,
querying aspects of the state that are hidden,
committing to a particular action,
directing an agent to do an action.
Each communicative act can affect the knowledge of the agents in a different way.
For example, an informing act can reduce the uncertainty of an agent about the
current state by eliminating candidate states from his information set. A
committing communicative act, on the other hand, can be used for informing about
the chosen course of action of an agent. Similar interpretations can be given to the
other communicative acts. For instance, a prohibiting act can have the same effect
as a role, by enforcing the deactivation of some actions of an agent in a particular
situation. Several agent communication languages have been proposed in the agent
community, aiming at standardizing the multi-agent communication process [23].
The two most notable ones are Knowledge Query and Manipulation Language
(KQML) and Foundation for Intelligent Physical Agents (FIPA), each using a
slightly different syntax and set of communicative acts. Unfortunately, many
dialects of these languages have already appeared, and the languages seem not to
conform with other standards (for example, in the Internet). As is often the case,
we might see in the future new language standards emerging directly from
applications.
4.10 References
[1] Durfee EH, Lesser V (1989) Negotiating task decomposition and allocation using
partial global planning. In Gasser L, Huhns M (eds) Distributed artificial intelligence,
Vol 2, ,. San Francisco, CA, Morgan Kaufmann pp 229–244
[2] Dent L, Boticario J, McDermott J, Mitchell T, Zabowski D (1992) A personal learning
apprentice. Proc 10th national conference on artificial intelligence. Menlo Park, CA,
American Association for Artificial Intelligence, pp 96–103
[3] Garrido L, Sycara K (1996) Multiagent meeting scheduling: preliminary experimental
results. Proc 2nd int conf on multiagent systems, Menlo Park, CA, American
Association for Artificial Intelligence, pp 95–102.
[4] Kinny D, Ljungberg M, Rao A, Sonenberg E, Tidhard G, Werner E (1992) Planned
team activity. In Castelfranchi C, Werner E (eds) Artificial social systems, New York,
Springer-Verlag
[5] Corkill DD, Lesser VR (1983) The use of metalevel control for coordination in a
distributed problem solving network. Proc 8th int joint conf on artificial intelligence
(IJCAI–83), Menlo Park, CA, pp 767–770
[6] Mason C, Johnson R (1989) DATMS: A framework for distributed assumption-based
reasoning. In Huhns M, Gasser L (eds) Distributed artificial intelligence, Vol 2,. San
Francisco, CA, Morgan Kaufmann, pp 293–318
[7] Sycara K, Decker K, Pannu A, Williamson M, Zeng D (1996) Distributed intelligent
agents. IEEE Expert 11(6): 36–46
[8] Lewis CM, Sycara K (1993) Reaching informed agreement in multispecialist

cooperation. Group Decis Negot 2(3): 279–300
[9] Sandholm T (1999) Distributed rational decision making. In Weiss G (ed) Multiagent
systems: a modern introduction to distributed artificial intelligence, MIT Press pp
201–258
[10] Noriega P, Sierra C (1999) Agent mediated electronic commerce. Lecture notes in
Artificial Intelligence 1571, Springer
[11] Lesser VR, Erman LD (1980) Distributed interpretation: a model and experiment.
IEEE Trans Comput 29(12):1144–1163
[12] Gilbert N, Doran J (1994) Simulating societies: the computer simulation of social
phenomena. UCL Press, London
[13] Roumeliotis SI, Bekey GA (2002) Distributed multi-robot localization. IEEE Trans
Rob Autom 18(5):781–795
[14] Terzopoulos D (1999) Artificial life for computer graphics. Commun ACM, 42(8):
32–42
[15] Bellman R (1961) Adaptive control processes: a guided tour. Princeton University
Press.
[16] Sutton RS, Barto AG (1998) Reinforcement learning: an introduction. MIT Press,
Cambridge, MA
[17] Russell SJ, Norvig P (2003) Artificial intelligence: a modern approach. Prentice Hall,
2nd edition
[18] Conry SE, Meyer RA, Lesser VR (1988) Multistage negotiation in distributed
planning. In Bond AH, Gasser L, (eds) Readings in distributed artificial intelligence.
San Francisco, CA, pp 367–384
[19] Durfee EH (1988) Coordination of distributed problem solvers. Boston, Kluwer
Academic
[20] Lesser VR (1991) A retrospective view of FA/C distributed problem solving. IEEE
Trans Syst Man Cybern 21(6): 1347–1363
[21] Lesser VR, Durfee EH, Corkill D (1989) Trends in cooperative distributed problem
solving. IEEE Trans Knowl Data Eng 1(1): 63–83
[22] Tanenbaum AS, van Steen M (2001) Distributed systems: principles and paradigms.
Prentice Hall
[23] Labrou Y, Finin T, Peng Y (1999). Agent communication languages: the current
landscape. IEEE Intell Syst 14(2):45–52
5
Distributed Production Planning in Reconfigurable

Production Networks
5.1 Introduction
Mass customisation consequences, such as shorter product life cycles and low-cost
variety, have brought critical pressures to improve production efficiency,
responsiveness to market changes, and substantial cost reduction. Scientific papers,
as well as government and industry expectations, seem to acknowledge that the
challenge keyword is “reconfiguration” [1]. Specifically, two major industrial
responses to mass customisation can be acknowledged: reconfigurable
manufacturing and reconfigurable enterprise (RE). From a manufacturing
perspective, the most agreed response to mass customisation is connected to the
concepts of modularity and reconfigurability of production systems. Thanks to
their modularity reconfigurable manufacturing systems (RMS) allow us to achieve
low customisation cost [2]. On the other hand, from an organizational point of
view, reconfigurable production networks or enterprises are nowadays considered
the industrial response to the great challenges conveyed by this new era
characterised by the global market and the impressive advances of information and
communication technology [3]. Market globalisation, indeed, has offered to
companies the possibility to split geographically their production capacity;
business opportunities lead companies to work together in temporary
organisations; in the same firm, business units behave as autonomous profit centres
and compete with each other for the production capacity allocation. In other words,
REs represent production networks made of different and geographically dispersed
plants which can be reconfigured in order to gather a specific production process or
product family. However, the RE members need to be properly coordinated to
achieve reduction in lead times and costs, alignment of interdependent decision-
making processes, and improvement in the overall performance of each member,
as well as of the RE. In this context, operations management and coordination of
RE and RMS are challenging issues involving distributed problem-solving tasks.
Specifically, in production planning (PP) and control activities, the concern with
internal production planning is replaced by the complexity of the external supply
chain. Indeed, as soon as a manufacturing unit tries to achieve coordination with its
partners, it quickly faces difficulties associated with different operational
conventions, locally specific constraints, software legacy and properties,
conflicting objectives and misaligned incentives. This task becomes even more
difficult when each RE member or plant consists of reconfigurable production
systems; indeed, the reconfiguration capability makes planning and scheduling
processes even more complex. The distributed production system (DPS), which has
been considered here, is linked to the semiconductor component industry; indeed,
usually, companies belonging to such an iIndustry sector are often organised into
reconfigurable manufacturing units geographically dispersed. In particular,
companies have a divisional organisation and the top management is called a
corporate; each division is called a group and is in charge of a product family
commercialisation. Each group can fulfil the collected orders thanks to the
reconfigurable characteristics of the manufacturing plants. Reconfiguration skills
offer clear advantages to the enterprise while increasing the complexity of
production planning activities. Multi-agent systems (MAS) techniques have been
largely used for their suitability in modelling complex systems involving multiple
autonomous agents with internal knowledge and reasoning engines which
communicate and interact with each other by exchanging messages according to
specific protocols. This technology is strictly required for approaching this kind of
complex problem, as will be shown in the following sections.
5.2 Production Planning in DPS

As a DPS aims at achieving goals deriving from localisation, the production
management policies cannot adopt a classical hierarchical approach. System
complexity and time-based competition push towards sharing the decision making
process. Production planning and control usually adopt different tools depending
on the considered time horizon maintaining, anyway, a centralised vision. This
chapter presents a decision support system for production planning activities
structured in different steps (time horizons) according to a classical viewpoint,
providing for each step the active involvement of the interested actors. This last
requirement promotes decision autonomy: autonomy becomes a key issue in
achieving the dynamic dimension required in the actual market scenario. Also, as
stated above, the decision-making process adopted in production planning
activities should always guarantee a global satisfaction level by means of
coordination and this is accomplished by negotiation and game theory as explained
in the following sections.
5.2.1 Context of the Semiconductor Industry
Production capacity is the most significant portion of capital investment in

semiconductor wafer manufacturing. Effective utilisation and expansion of
production capacity have significant cost implications, and arguably drive the
profitability of the operation. Capacity management in the industry typically entails
long-term strategic planning and short-term operational planning organised in a
Distributed Production Planning in Reconfigurable Production Networks 53
hierarchical manner. Strategic planning decisions include how much of the

aggregate microelectronics technology is produced in what facilities, and which
capacity element to expand within what timeframe so as to meet the projected
demands. Operational planning determines capacity adjustment or reconfigurations
when more accurate demand and capacity information becomes available.
Operational planning is frequent and dynamic so as to accommodate weekly
production wafer “starts” to be released to manufacturing. One important
characteristic of semiconductor capacity planning is that both product demands and
manufacturing capacity are sources of uncertainty. As is the case in most hi-tech
industries, the semiconductor market has a demand structure that is intrinsically
volatile. A microelectronic chip that faces high demands today, may be quickly
outdated in a few months with the introduction of a next-generation chip requiring
an enhanced manufacturing process. New manufacturing processes create high
variability in the yields, and consequently uncertainty on the manufacturing
throughput, which in turn lead to uncertainty in capacity estimation. Since the
production volumes are typically high (for the interests of achieving economies of
scale), extreme outcomes on demand and capacity realisations can lead to very
undesirable business consequences. Therefore, while capacity configuration and
allocation are important decisions for any manufacturing company, a few factors
make this problem especially crucial to the semiconductor industry.
First is the high cost and long lead time for equipment procurement and clean-
room construction. The semiconductor wafer fabrication process requires state-of-
the-art manufacturing equipment, costing many millions and which must be
ordered up to twelve months in advance. Because of the long lead time involved,
capacity expansion decisions must be made far in advance. A wrong decision,
either over- or under- estimation, could have major impacts on profitability. On the
other hand, if the capacity for a certain technology is not expanded promptly to
meet market demands, a significant loss of market share may result.
A second factor that exacerbates the impact of capacity planning in the
semiconductor industry is the rapid advancement of fabrication technologies and
the pace of transition from old technologies to new. Semiconductor technologies
can be defined in several ways, one of which is the space between features on a
semiconductor die, known in the industry as line width. The most expensive and
crucial pieces of equipment in the wafer fabrication line are used in the
photolithography process, where the chip features are defined on a silicon wafer.
With each advancement in photolithography technology, new and more expensive
equipment must be purchased so that features with smaller line widths can be
produced. Although the equipment is more expensive, it allows the manufacturer to
either make smaller chips or fit more features on the same-size chip, effectively
reducing manufacturing costs for a given chip functionality. Another factor in the
advancement of semiconductor manufacturing technology is the size of the wafers.
Equipment manufacturers are continuously trying to increase the wafer size, which
increases the number of chips to be made at once and produces higher yields,
which in turn reduces the unit manufacturing cost. As semiconductor technologies
advance, the company must be prepared to switch manufacturing capability to the
newer technologies. These transitions take time and they must be anticipated
correctly. A premature transition will lead to costly underutilisation, or forcing
manufacturing to use newer, more expensive equipment to manufacture older

technologies that do not generate expected revenue. An overdue transition leads to
missed market opportunities, which also lead to lower ROI for the capital
investment.
A third factor unique to the semiconductor industry is that manufacturing
capacity often suffers high variability. The aggregate notion of manufacturing
capacity used during strategic planning is in reality an approximation at best.
Given a particular capacity configuration for each clean room, the manufacturing
manager still has much flexibility in how that capacity is utilised, and his/her
decision will determine what the “effective capacity” ultimately is. For instance,
newer equipment can typically be used to manufacture older technologies, albeit at
a lower cost efficiency. Further, the “effective capacity” to manufacture the same
technology is different in each location, depending upon the technology mixture
(capacity configuration), the wafer size made in that facility, skill level of the
labour, and myriad other factors. This is further complicated by contractual
requirements with the customer, which may dictate that certain products must be
made in certain locations. Finally, if there is not enough capacity in the company-
owned facilities, outsourced foundry capacity can be bought for some technologies,
at a higher cost. If wafers are made at a foundry, there may be contractual
requirements that specify a minimum purchase. All of these factors come into play
to make manufacturing capacity a significant source of uncertainty at the point of
strategic planning.
5.2.2 PP in the Considered Industrial Case
Semiconductor manufacturing operations consist of two main stages: the “front-

end” operation of wafer fabrication, and the “back-end” operation of assembly and
testing. The front-end operation is typically the bottleneck as the process involves a
6–12-week manufacturing lead time, while the back-end requires 2–4 days.
Moreover, the wafer fabrication facilities are extremely capital intensive while
requiring significant lead-time to build. Demand in high-tech industry is known to
be volatile and particularly sensitive to economic cycles. Managing wafer
fabrication capacity is one of the most crucial activities for semiconductor firms. In
the considered case, every year, the corporate level assigns to the company groups
(responsible for product families such as electronic memories) a certain level of the
total production capacity (called capacity ownership) based on long-term demand
forecasting and product strategic positioning. Every three months the groups, after
having collected backlog and forecast orders coming from the regional divisions,
according to the ownership they hold and to the demand they have to supply, make
their capacity allocation plan. If the group capacity ownership is not enough to
supply the demand orders, then the group can negotiate a portion of capacity with
the other groups whose assigned ownership exceeds their actual demand [4]. In
practice, such negotiation and consequent possible exchange of capacity turn out
into a re-assignment of some production plants to a different group. Plants assigned
at the beginning of the year to the production of components belonging to a
specific product family, could be reconfigured throughout the year for producing
different types of components, i.e. different product families. Also, within the
annual quarter in which the assignment of plants to a group remains fixed, orders
of products belonging to a product family (group) must be allocated to the different
plants temporary assigned to the group [5]. Plants represent reconfigurable
production systems able to be reconfigured in the short period (within the three
months) in order to manufacture different types of product of the same part family.
Such a brief description demonstrates how the PP process in a distributed
organisation can become complex, multi-period, multi-decision and multi-issue
when a somewhat reconfigurable capability is considered.
Figure 5.1 reports an IDEF0 view of the PP process in reconfigurable
enterprises made up of DPS.
5.2.3 IDEF0 Architecture
Each PP activity reported in Figure 5.1 is related to a different time horizon and
concerns different PP levels.
The top PP level (activity A1) starting from information on groups’ one-year
demand forecast and priorities for each group deriving from strategy
considerations, assigns the global ownership to groups. The output of activity A1
represents an input for A2 together with groups’ priorities and demand forecast for
the considered time horizon of three months. Reducing the forecast horizon, future
estimates are more reliable and at the high PP level a tuning in ownership
allocation is allowed. The medium PP level focuses on the same time horizon of
the previous level but it aims at assigning each plant to a group basing on the
output of activity A2. Low level and shop-floor level both concern real-time
planning issues: the first one allocates, for each group, orders to one of the plants
assigned to that group, while the second one is responsible, for each plant, for the
allocation of jobs to resources.
5.2.4 Agent Architecture
Actors involved in the distributed production planning (DPP) are: corporate,

groups, plants, orders, jobs, resources. They have characteristics proper of agents:
goal-oriented, collaborative, flexible and capable of making independent decisions
on when to act [6]. Actually, involved agents pursue their own goals, interact with
other agents to bargain a common action plan, get different roles along the process
and know when and how to act in accordance with the concerted protocol. In the
following sections we will characterise the collaboration/cooperation among those
actors by identifying for each level of PP:
the objective of the cooperation, that is, the goal to be reached;
the bargain issue;
the actors involved;
the roles actors play in the bargaining.
56
Global Ownership
Groups
1 year Ownership
demand
TOP PP Level
forecast
1
Groups Tools to allocate
priorities ownership
HIGH PP Level Re-modulated
ownership Parts
production caacity
2
Tools to re-
3-months Groups modulate Plant/Group
demand forecasts ownership allocation matrix
MEDIUM PP Reseource technology characteristics
Level
3
Production Planning in Production Networks
Order/Plant
LOW PP Level allocation
Parts
characteristics 4
Tools to allocate Job/Resources

orders to plants allocation
SHOP FLOOR
PP Level
Group Orders
5
Tools to allocate
Resources to Jobs
NODE: A0 TITLE: Production Planning activities NO.:
Figure 5.1. A layered architecture for the PP problem

5.3 Top PP Level

At this level, the corporate behaves as a seller while groups act as buyers. In fact,
the corporate behaves like the owner that splits ownership to groups, based on the
priorities assessed at the strategic level and communicated to all the groups. Each
group aims at maximising the negotiated ownership and in order to limit this
approach and to achieve company goals, the corporate concedes ownership
obliging groups to assure to the company a certain profit level according to their
priorities. The conditions of the contracts signed in this phase would be used as
control for groups’ performance at the end of the year. These considerations
represent guidelines to keep in mind to set, in related future research, strategies and
tactics. Protocol variables to be set (the others are common for all the levels)
concern the communication channel (cc): there exist a cc for offer and counter-
offer and an informational cc to broadcast groups’ priorities. Summarising, at this
level the objective is the ownership assignment; this goal is obtained through a
bargaining process among the corporate and the groups, which are of course the
actors involved in the bargaining process acting respectively as seller and buyers.
Such a situation is depicted in Table 5.1.
Table 5.1. DPP variables for top PP level
Objective Bargaining issue Actors Roles

Ownership Ownership and Buyer (Groups)
Corporate & groups
assignment requested credits Seller (Corporate)
5.4 High PP Level

Once ownership is assigned and the time horizon reduced, it might happen that
some groups have been assigned an excess of capacity when compared with the
actual demand they actually need to face, while other groups might be, for opposite
reasons, in a capacity defecient position. At this level then, groups can assume an
opportunistic behaviour changing over ownership in the current quarter to contract
options for receiving it back in the future. To do that, it is necessary to introduce a
lateral payment using an expedient: credits. Credits correspond to a virtual
production capacity and are equally distributed to the groups at the beginning of
each year; they are used to buy capacity, then groups with a great number of credits
have a great contractual power to obtain capacity. At this level, based on the
quarter forecast and on the ownership assigned at the previous level, groups can
take on a buyer or a seller behaviour. If the workload related to the forecasted
demand is higher than the ownership, the group wants to buy production capacity;
in the opposite condition it is interested in selling the extra capacity and receiving
credits. Sellers and buyers adopt a time-dependent tactic and use different
generative functions for the order and counter-order formulation. Summarising, at
this level the objective is to balance capacity among groups; this goal is obtained
through a bargaining process among the groups, which acting as sellers and
buyers, depending on the available capacity, negotiate by exchanging capacity
ownership and credits. DPP characterisation at this level is reported in Table 5.2.
Table 5.2. DPP variables for high PP level

Balance capacity Capacity ownership Buyer (Groups)
Groups
among groups and credits Seller (Groups)
5.5 Medium PP Level

Once groups have balanced capacity with their workload, they need to acquire
actual production capacity by buying it from plants. This level has the same time
horizon as the previous one, but involves plants as active participants which have
to be assigned to groups, based on the ownership obtained at the end of the HIGH
PP level. Assuming that each plant is a cost and profit centre, it is interested to be
assigned to the most promising group in terms of future profit; the plant, to avoid
imperfect commitment, asks a price to assure its availability. Moreover it can
happen that group catches plants at different rounds offering an increasing price,
but giving it a lower priority in term of guaranteed workload. So plants assess a
risk attitude, which is therefore used in their reactive function formulation.
Summarising, at this level the objective is to acquire actual production capacity
through plant assignment to groups; this goal is obtained through a bargaining
process among the groups, which act as buyers, and the plants, which act as sellers
of capacity. They bargain by exchanging plant production capacity and price the
buyers pay to gain it. DPP characterisation at this level is reported in Table 5.3.
Table 5.3. DPP variables for medium PP level

Plant production
Buyer (Groups)
Plant assignment capacity Groups and plants
Seller (Plants)
and price
5.6 Low PP Level

This level considers a real time horizon; each group collects orders from its
divisions, and each order is assigned to one of the plants gained at the previous
level. Therefore, at this level the problem is to assign each order to one of the
plants of the group in order to satisfy orders’ and groups’ objectives. The
assignment will obviously consider, as in the previous level, logistic (the distance
between the plant and the final customer), economic (the demand elasticity of each
product) and technology issues (plant/product efficiency matrix and
reconfiguration costs) but, at this level, time is most of all a scarce resource.
Moreover, each plant does not know the price agreed with other plants, nor their
priorities; therefore, two situations can occur: orders or plants can represent the
scarce resource. It can happen that workload is greater than available capacity and
vice versa; however, this information is not known because it results as a private
one. Because a third part (a mediator) is not present, it can be argued that plants
behave as sellers and order as buyers because of the promises deriving from
previous level (price and priority). Summarising, at this level the objective is to
allocate actual orders to plants assigned to groups (order assignment); this goal is
obtained through a bargaining process among plants, which act as capacity sellers,
and the orders, which act as buyers. They bargain by exchanging production
capacity and price. DPP characterisation at this level is reported in Table 5.4.
Table 5.4. DPP variables for low PP level
Assign orders to Plant capacity/order Buyer (Orders)

Orders and plants
plant assignment and price Seller (Plants)
5.7 Shop-floor PP Level

This last level concerns resources assignment within each plant. Indeed, once each
order has been assigned to a plant it is decomposed into jobs characterised by
volumes and due-date. Therefore this is a common scheduling problem, which is
faced in a distributed fashion. Indeed, scheduling is obtained through a bargaining
process among the plant resources (sellers), which sell their production availability
to jobs (buyers) obtaining a price in exchange. At the shop-floor level the global
perspective is the plant perspective and the local interests arise from the opposite
goal pursued by jobs and resources. Table 5.5 summarises the DPP at this level.
Table 5.5. DPP variables for shop-floor level
Objective Bargaining issues Actors Roles

Scheduling
Job/resources Buyer (jobs)
problem: jobs to
assignment and Job and resources
resources Seller (resources)
price
Assignment
Table 5.6 summarises the proposed hierarchical distributed production planning

approach. As the reader can see it associates the planning horizon, the planning
objective, the makers involved and information/issues considered in the bargaining

mechanism.
Table 5.6. Production planning levels
Planning Planning Planning Information/issues in the

Makers
level horizon objective bargaining mechanism
Corporate
Assign capacity Long-term demand
agent and
Top 1 year ownership to forecasting and product
groups’
groups strategic positioning
agents
Unbalancing level among
Remodulate
Groups’ medium-term forecasting
High 3 months ownership
agents plus backorder and
among groups
assigned ownership
Plants’ skills in producing a
Groups’
Assign plants to given part family,
Medium 3 months agents and
group geographical position of
plants’ agent
the plants
Order agents Plant’s skills in processing
Real- Allocate orders
Low and plants’ the order and geographical
time to plant
agent position of the plants
Job agents Resources’ skills in
Allocate jobs to
Shop- Real- and processing the job,
production
floor time resources’ reconfiguration time
resources
agent required to process the job
The specific focus of this book concerns the design of a negotiational and game
theoretical mechanism for solving bargaining at the high and medium-level
production planning. This results will be benchmarked with a centralised model
with full knowledge. Low and shop-floor levels will be neglected because several
decentralised approaches have been proposed in the literature which can find
application in our proposal.
5.8 References
[1] US National Research Council (1998) Visionary manufacturing challenges for 2020.
National Academy, Washington, DC: 13–36
[2] Koren Y, Heisel U, Jovane F, Moriwaki T, Pritschow G, Ulsoy G, Van Brussel H
(1999) Reconfigurable manufacturing systems. Ann CIRP, 48/2: 527–541
[3] Wiendahl HP, Lutz S (2002) Production networks. Ann CIRP 51/2: 573–587
[4] Argoneto P, Bruccoleri M, Lo Nigro G, Perrone G, Noto La Diega S, Renna P,
Sudhoff W (2006) High level planning of reconfigurable enterprises: a game theoretic
approach. Ann CIRP Vol. 55/1
[5] Argoneto P, Perrone G, Renna P, 2006. Medium level planning of reconfigurable

enterprises: a game theoretical approach. MITIP, 11–12 September, Budapest,
Hungary
[6] Etzioni O, Weld D (1995) Intelligent agents on the internet: fact, fiction, and forecast.
IEEE Expert, August 1995, 44–49
6
Distributed Models for Planning Capacity of

Reconfigurable Production Networks at Medium Term
6.1 Introduction
Production planning at medium term consists in what we have referred to as high-
level production planning in the previous chapter. As mentioned, at this level the
planning objective is to balance capacity among groups; this goal is obtained
through a bargaining process among groups, which, by acting as sellers and buyers
and accordin to the available capacity, negotiate for exchanging capacity
ownership and credits. In the following sections a set of distributed production
planning models for such a purpose will be discussed.
6.2 Initial State

Let us assume that the total planning horizon T is divided into N sub-periods t (t =
1,2,…,N); at the beginning of each sub-period, each group receives a new
workload from the respective divisions and a group priority level from the
corporate level; on the other hand, it has been assumed that the group capacity
ownership, assigned to the top level phase, remains the same along the whole
planning horizon T. Each group (g = 1,…,G) receives, from its divisions, data
about the workload WL it needs to sustain and, from the corporate agent, data
regarding the priority level, P , that the specific group has been assigned to
according to the customer importance. Let us define with Pg [0, 1] the priority
index that the corporate agent has assigned to the g-th group. Groups compute their
respective available capacity C g according to the ownership they have got.
Specifically:
C g (Og Ctot )/100 (6.1)
where Og is the ownership of the group g and Ctot is the total capacity the
company holds. At this point each group computes an index S g measuring the
difference between workloads and capacity:
Sg WLg C g (6.2)
Subsequently groups are identified and classified as overloaded
OG ^1,..., i ,..., N` , i.e. groups with Sg ! 0, or underloaded
UG ^1,..., j ,..., M` , with Sg 0. Afterwards, OG and UG groups compute
respectively the required capacity RC g and the offered capacity OC g :
RC g WLg C g (6.3)
OC g C g WLg (6.4)
6.3 The Centralised Model

First of all, it is very important to distinguish between the centrality of the problem
and the centrality of the solution. In the following centralised model, the
assumption is that all the relevant planning data can be centrally located and can be
analysed by a centralised algorithm. If the approach is distributed, like in the
following sections, the statement is that it is physically impossible (or unwanted) to
locate all the actual data at one physical host (e.g. logistics planning). Actually, a
given problem can be centralised only to some extent (e.g. only meta-information
such as from where to where is the semi-product transported is centrally available
while the exact location of the truck is unknown). Obviously, centralised problems
are more likely to be solved by centralised algorithms, while distributed problems
would require more negotiation-based approaches. In this model, after having
classified groups in OG and UG, the corporate-centralised decision maker’s goal is
to minimise the capacity backlog by minimising the OG groups requested
capacities. That is achieved through the following steps:
1. a matrix L(OG u UG ), whose elements cij RC i OC j are the
differences between the requested capacity RC i by the i-th group and

residual capacity OC j which is offered by the j-th group is computed:
§ ... ... ... ·
¨ ¸
L ¨ ... cij ... ¸ ; (6.5)
¨ ... ... ... ¸
© ¹
2. in order to achieve the best capacity allocation solution, the following
MILP model is solved:
§ N M ·
min ¨ ¦¦ c ij xij ¸ , (6.6)
ï 1 j 1 ¸
© ¹
Distributed Models for Planning Capacity at Medium Term 65
subject to:
N M
¦x
i 1
ij 1; ¦x
j 1
ij 1, if N = M (6.7)
N M
¦x
i 1
ij 1; ¦x
j 1
ij d 1, if N < M (6.8)
N M
¦ xij d 1;
i 1
¦x
j 1
ij 1, if N >M (6.9)
where xij is a binary variable such that

1, if group j capacity is assigned to i
xij® (6.10)
¯0, otherwise
Once each group i OG is associated to a group j UG , i.e. xij 1, their requested
and offered capacities are respectively updated as
RC i ' RC i min( RC i , OC j ) (6.11)
OC j ' OC j min( RCi , OC j ) (6.12)
6.4 The Negotiation Model

The negotiational mechanism has the same goal as the centralised one: to minimise
the unallocated capacity. The method, this time, is a decentralised one and is
referred to as a one-to-many negotiation approach. During this process, negotiators
are allowed to communicate their respective desires and they will compromise to
reach mutually beneficial agreements. This strategic negotiation is a process that
will include several iterations of offers and counter-offers. In particular, the
algorithm works as follows:
1. for each step s the group i * OG , which is in charge of starting the
negotiation, is the one with the maximum value of workload excess, i.e
for s = 1:
RC s* / i *
i
max RC s ;
i
^ i
` (6.13)
2. each group j UG computes the capacity to be offered and its own credits
threshold according to
M rj min( RC r* , OC r ), i j
(6.14)
§ r 1·
t rj M rj ¨ 1 PL j ¸, (6.15)
© r ¹
where: PL j C j WL j /C j ;
3. the i*-th group formulates its bid to gain the requested capacity;
specifically, it offers a certain number of credits to each
j UG , according to
OC ir* j ^ `
min CC ir* , M rj (1 r PLi * ) , (6.16)
r
where CC are the i*-th group current credits. Indeed, each group
i*
i OG is requested to provide a certain number of credits to j in order to

bargain some amount of capacity;
4. the negotiation continues until r rmax , unless group i* and one group j
belonging to UG come to an agreement; the agreement condition assures a
maximum allocation capacity from i*-th’s point of view, i.e.:
^
max OC ir* j t t rj ,
j
` (6.17)
5. if at r rmax no agreement is reached, a sub-optimal accord is sought:
group i and j come together if OC ir t t rj for a generic j. If there exist
several j groups able to satisfy the previous condition, the one minimising
the residual i’s requested capacity is selected;
6. if no agreement is reached, accordingly step s is increased and a new
i * OG , corresponding to the group having the maximum value of
workload excess for the new step s, is selected, according to Condition
6.18, and the negotiation starts again from point 2;
i * /max ^RCi ( RCmax
1 s1
,..., RCmax )` (6.18)
i
If, also in this case, no agreement is reached, the process begins again
from point 6 until the end of i OG ;
7. the residual capacity and credits are uploaded for each group and the
mechanism starts from point 6.
6.5 The Game-theoretical Model

In this case, the problem can be formalised like an assignment game. The outline is
a quadruple AG r (OG r , UG r , T r , O r ) , where:
OG r ^1,..., i ,..., N ` r
is a vector whose element i represents an
overloaded group at round r;
UG r ^
1,..., j ,..., M r `
is a vector whose element j represents an
underloaded group at round r;
T r is a vector whose element, t j r , represents the credits (threshold of)
that the j-th group, belonging to UG r , assigns to its own residual
capacity. This value is calculated through Equation 6.15;
O r is a matrix whose generic element OC ij r is the value the
group i OG r assigns to the capacity offered by each player j UGr
using Equation 6.16.
A transfer utility game [1] is associated to the quadruple AG r ; in that, the whole
set of players are given by * r OG r UG r and the characteristic function Ȟ is
defined, given that groups i and j make a coalition, as follows:
°OCij r t j r if OC ij r t j r ! 0
v (i , j ) hij r ® (6.19)
°̄0 if OC ij r t j r d 0
The set of hij r defines the following assignment problem:
§ N M ·
max ¨ ¦¦ hij r zij ¸ (6.20)
ï 1 j 1 ¸
© ¹
where,
zij ^
1, if i and j makes a coalition
0, otherwise , (6.21)
subject to, i OG r , j UG r , the following constraints:

N M
¦z
i 1
ij 1; ¦z
j 1
ij 1 , if N = M (6.22)
N M
¦z
i 1
ij 1; ¦z
j 1
ij d 1 , if N < M (6.23)
N M
¦
i 1
zij d 1; ¦z
j 1
ij 1 , if N >M (6.24)
After having created the assignment among groups, it is necessary to consider two
different situations depending on the value of the characteristic function.
6.5.1 Case 1: Characteristic Function hij > 0
In this case, the coalitional players are required to split the game surplus. In order
to do this, the optimising charge belonging to the core game [2] is computed.
According to Owen’s theorem [3], a core allocation, in a linear assignment game,
can be obtained starting from an optimal solution of the dual programming model:
§ N M ·
min ¨ ¦ yi ¦ y j ¸ , (6.25)
ï 1 ¸
© j 1 ¹
subject to:
yi y j t hij r . (6.26)
Once the solution is calculated, the credits and the capacities are uploaded for all
the coalitional players.
6.5.2 Case 2: Characteristic Function hij < 0
Here, in comparison to that explained above, different considerations have to be

made. This is the situation in which the group i offers a non-sufficient (scarce)
number of credits to the group j. Obviously, there is no surplus to be shared among
players but, on the contrary, it is necessary, if possible, to reach an agreement, to

share the penalty given by the spread existing by the offered and requested credits:
OCij r t j r . This is the typical situation that can be formalised like a bargaining
situation between two players. The adopted solution concept is referred to as the
Nash formulation [4], which satisfies a set of axioms (invariance to equivalent
utility representations, symmetry, independence of irrelevant alternatives and
Pareto efficiency) for a “fair” bargain. In order to reach this goal, it is necessary to
build up two different kinds of utility functions for the involved players.
6.5.2.1 Utility Function for the OG Group

From the perspective of this group, two different parameters have to be taken into
consideration: one taking into account their unallocated capacity (whit respect to
the one needed), and the other one considering the numbers of credits they are
willing to pay in order to obtain it.
The first parameter is indicated by:
ci callir 1
[i , (6.27)
ci
callir 1 being the capacity already allocated to the group until the round
r 1.
The higher is [ i , the more the group is willing to increase the numbers of
credits to pay to obtain the capacity it needs. Having defined with ' the
percentage of the difference OC ij r t j r the group i-th has still to fill up to
reach the agreement, it is obvious that the more ' increases, the more its
own associated utility decreases with a slope depending on [ i ;
The second parameter to be considered is:
Cri
Ki , (6.28)
Crtot
Cri being the residual credits owned by the i-th group at step r and Crtot
the credits it owned at the beginning of the game. Also in this case,
increasing ' means decreasing utility, with slope depending on Ki .
Consequently, the utility function of the groups belonging to the OG set will be
a parametric function that will be modelled by using a Bézier’s curve. In particular,
the Bezier curve control points (Pi), those constituting the characteristic polygon,
are: P1 { 0 , 1 , P2 { P3 { P4 { Xi ,Xi , P5 { 1 , 0 , , where:
Xi Ki [ i . (6.29)
This choice allows us to obtain a utility function equal to the minimum value,
0, when the penalty to pay is ' 100% , and equal to 1 when ' 0% Alsothe
choice to have three overlapping points is useful to obtain a curve that best
approximates the characteristic polygon. Therefore, the utility function for the OG
group will be written as:
5
UiOG B u ¦Pb
t 0
i t ,5 u , 0 d u d 1, (6.30)
where the polynomials

§ 5· 5 t
bt ,5 u ¨ ¸ ut 1 u ,t 0 ,..., 5 , (6.31)
t
© ¹
are known as Bernstein basis polynomials of degree t.
U iOG
1
Qi
0,8
0,6
0,4
0,2
'
0
0 0,2 0,4 0,6 0,8 1 1,2
Figure 6.1. Utility function of the groups belonging to the OG set
6.5.2.2 Utility Function of the Groups Belonging to the UG Set

From the perspective of the groups belonging to the UG set, the reasoning to build
up their utility function is rather more intuitive. Given the above assumptions, they
are required to assign their own capacity M rj to the OG groups. Therefore, they
just wish to gain more credits as possible. This means that the more the offered

payment OC ij r
is close to the requested one t j r , the more the utility value
tends, with linear approximation, to the maximum value. Using the same notation
as above, we simply have:
U jUG ' . (6.32)
U jUG
0
0 ' 1
Figure 6.2. Utility function of the groups belonging to the UG set
6.5.3 The Bargaining Solution
Figure 6.3 shows the bargaining set. Graphically it is the area under the curve and
is easy to see that, as requested by the Nash solution, it is convex and bounded.
1 U iOG
Bargaining set
0
0
U jUG 1
Figure 6.3. The bargaining set

The Nash bargaining solution is given by the following equation:

NBS max U
¬
max ª U iOG U iOG U jUG U jUG º ,
¼ (6.33)
U iOG , U jUG being the fall-back outcome, i.e. the outcome that can result if the
bargaining breaks down. In this specific case, it is easy to demonstrate that
U iOG U jUG 0 . The result will be the additional credits (OC ij r t j r ) ' the
group i has to pay to the group j and, complementarily, the number of credits the j-
th group has to give up, (OC ij r t j r ) (1 ' ), in order to reach the agreement.
Finally, the actual exchanged credits will be OC ij r ' :
OC ij r ' OC ij r (1 ' ) (t rj ' ), (6.34)
If the j-th group do not have a sufficient amount of credits to allow the exchange,
the bargaining falls and the i-th group does not assign the requested capacity.
Finally, the residual capacity and credits are uploaded for each group involved in
the game and the algorithm starts again (r = r + 1), formulating a new assignment
game, until one of the following two conditions (no more workload excess or no
more extra capacity) comes true:
N M
r r
¦ RC
i 1
i
0; ¦ OC
j 1
j
0; (6.35)
6.6 Simulation Case Study

A discrete event simulator, based on a multi-agent system (MAS) approach, has
been developed and used for testing the proposed models. Optimisation has been
performed through the LINGO® package conveniently linked to the simulation
environment. The simulation environment has been entirely developed using
JAVA. The choice of using JAVA has led to obtaining high flexibility in the
development of the simulator, especially thanks to the advantages related to the
object-oriented philosophy. To reduce development time and cost related to the
simulator development, the open source Java Development Kit (JDK) has been
utilised. The adopted formalism is related to all possible interactions existing
between the considered agents. The simulation carries on considering an events
timetable of the agents’ actions: this schedule is utilised to define the sequence of
the events. In the following, UML class and sequence diagrams will be utilised to
describe the simulation environment.
6.6.1 The Simulation Environment
For all different approaches above explained, the simulation environment consists
of the following objects, as also depicted in the class diagram of Figure 6.4:
the Schedule: this object is in charge of managing the sequence of events
during the simulation time;
the Model: this object handles the signals related to the orders, coordinates
the activities among groups and selects the strategies;
the Group: this object possesses all groups’ information and functions
necessary to the implementation;
the Statistical Analysis: this last object is dedicated to the analysis of the
groups’ performance.
The sequence diagram shown in Figure 6.5 refers to a generic order o, o =
1,..,O:
Figure 6.4. Class diagram of the high production planning level

Figure 6.5. Sequence diagram, high production planning level
The Schedule object activates all other objects as explained below.

o-th order: order number o is communicated to Model object;
Model object updates information related to o-th order (workload, ownership,
priority);
Initialize o-th order: information related to o-th order is communicated to
groups;
Groups calculate the capacity they need or their capacity excess: this evaluation
leads the groups to choose the OG or the UG set;
Select set (OG or UG): the selected set, for each group, is communicated
to the Model;
Depending on the selected set, Model needs information by the groups (e.g. offered
or required capacity, utility values);
Values request: needed values are requested to groups;
Groups calculate the requested values;
Values communication: requested values are communicated by Group to
Model;
Model applies one of the three different approaches, depending on the considered
model: centralised, negotiational, game theoretical;
Performances communication: groups’ performances are communicated
to Statistical Analysis object;
Assignment communication: new capacity adjustment among groups is
communicated to Group object;
Statistical Analysis object uploads performances and carries on with the simulation
until the defined confidence interval is reached;
Statistical performances: performances are communicated to Schedule
object;
Schedule verifies the performances’ statistical significance;
Significance communication: Schedule communicates to Statistical
Analysis object the result of the previous check;
Final report is written.
6.6.2 Simulation Case Study
The test environment consists of 27 different combinations of input parameters:

workloads (WL), priorities (P) and ownerships C indexes. In particular, the o-th
order is given by a vector (WL,P,C): each index is considered randomly drawn
from three normal distributions N(µ;ı) whose parameters are reported in Table 6.1
as Small (S), Bounded (B) or Wide (W) with relation to the standard deviations
size:
Table 6.1. Distributions of parameters
Small (S) Bounded (B) Wide (W)

Workload (WL) N(17; 0.85) N(17; 5.1) N(17; 10.2)
Ownership (C) N(17; 3.4) N(17; 6.8) N(17; 10.2)
Priority (P) N(0.5; 0.025) N(0.5; 0.3) N(0.5; 0.5)
Simulation experiments have been conducted with an infinite number of credits

and, for each replication, the average of unallocated capacity (Avg UC) with a
confidence interval equal to 95% has been evaluated.
6.7 Results
The results of 27 different combinations of input parameters for the centralised
(C.A.), negotiational (N.A.) and game-theoretical (G.A.) approaches are reported in
Appendix A. They are represented in tables indicating: the average value of the
unallocated capacity (Avg), the (half-width) confidence interval (Half), the number
of replications needed by the simulator to calculate it (Rep), the extreme values
(Min, Max) of the estimated parameter and the variance (Var). In order to analyse
the results, two different techniques have been used: the two-way Analysis of
Variance (ANOVA) and the Design of Experiment (DoE). The first one leads us to
understand the dependence of the results from the variance of input parameters,
from the three different proposed models and, if there is, from their interaction.
Instead, DoE will lead us to comprehend the direction of this dependence.
6.7.1 Two-way Analysis of Variance
The two-way Analysis of Variance is an extension of the one-way ANOVA: here,

there are two independent variables. The classical assumptions are the following:
the population from which the data are drawn must be normally or
approximately normally distributed;
data must be independent;
the variances of the populations must be equal;
the groups must have the same sample size.
In the analysed case, the independent variables are: the different models, C.A., N.A.
and G.A., and the different input parameters, WL, P and C. The obtained results,
rearranged in the following tables, can be considered like a set of data drawn from
three normal distributions with means equal to the three different grand means
(depending on the utilised model) and the same (unknown) variance. This last
assumption assures us that all the above constraints are satisfied. Each factor (i.e.
each independent variable) has three different levels. Respectively, the factor
“model” has levels corresponding to C.A., N.A., G.A.; the factor “input parameters”
has levels equal to W, B, S. The goal is to understand in which way the input
parameters affect the estimated performance. In order to do this, it is necessary to
consider only one parameter at time. The first one will be the workload.
6.7.1.1 Workload
Table 6.2 reports the result of the unallocated capacity (given in Appendix A)
rearranged to highlight the workload parameter. Indeed, in the rows of Table 6.2
we have the different levels of workload for each set of model; in the columns, on
the other hand, we have the levels for combinations of the other two input
parameters, P and C. Such combinations have been obtained as indicated in Table
6.3.
Table 6.4 shows all the mean values for each model and for each level of WL.
Data from this table suggest that the two decentralised approaches perform slightly
worse than the centralised approach, there being a reduction in performance of
about a 20% on average. This is quite a good performance considering the
distributed nature of the two approaches. Also no great difference is evident
between the two distributed approaches, namely the negotiation and the game
theoretical one.
Table 6.2. Replications values for WL
1 2 3 4 5 6 7 8 9
W 51.04 39.16 34.10 51.54 34.31 34.65 49.98 38.84 34.48
C.A. B 41.88 27.78 21.69 42.72 28.20 20.66 41.86 26.32 21.48
S 36.43 18.49 8.74 35.88 18.67 8.45 35.22 18.59 8.76
W 67.73 43.64 35.61 64.86 35.10 34.89 67.27 43.12 36.15
G.A. B 59.13 31.56 21.57 58.88 31.65 21.73 57.86 30.51 21.74
S 52.23 20.90 8.53 55.79 21.77 8.35 53.60 20.53 9.05
W 65.83 43.88 36.76 66.14 43.32 34.80 67.61 44.58 36.66
N.A. B 58.57 31.37 22.00 57.72 30.19 21.93 57.64 30.40 21.06
S 51.28 20.53 8.56 53.39 21.07 8.60 53.23 20.61 8.69
Table 6.3. Combination of P and C

1 2 3 4 5 6 7 8 9
Priority (P) variance W W W B B B S S S
Ownership (C) variance W B S W B S W B S
Table 6.4. Mean values for WL

W B S Mean
C.A. 40.90 30.29 21.03 30.74
G.A. 47.60 37.18 27.86 37.55
N.A. 48.84 36.76 27.33 37.65
Mean 45.78 34.74 25.41 35.31
Let us investigate deeply the obtained results. In particular, there are three sets of
hypotheses to consider; indeed, the null and alternative hypotheses for each set are
the following:
1. for the population means of the first factor (the model)

° H 0 : PC PN PG
®H : ; (6.36)
°̄ 1 not all Pi^C , N ,G` are equal
2. for the population means of the second factor (the workload):

° H 0 : P W P B PS
®H : ; (6.37)
°̄ 1 not all P j^W , B ,S` are equal
3. considering the interdependencies:

H0 : there are no interactions
® . (6.38)
¯ H1 : there are interactions
Before explaining the components of the two-way ANOVA table, it is necessary to
the notation generalise a little bit. Let us assume that:
the main effect A has a levels (and a 1 degrees of freedom),
the main effect B has b levels (and b 1 degrees of freedom),
c is the sample size of each treatment, and
N = abc is the total sample size,
then we have:
Table 6.5. Two-way ANOVA table

Sum square Degree of Mean square
Source of Variation F-test
(SS) freedom (df) (MS)
Main Effect A SSA A1 s 2A SSA /( a 1) s 2A / sr2
Main Effect B SSB b1 sB2 SSB /(b 1) sB2 / sr2
Interaction effect SSi (a 1)(b 1) si2 SSi /( a 1)(b 1) si2 / sr2
Within SSr ab(c 1) sr2 SSr / ab(c 1) –
Total SST abc 1 –
The main effect involves the independent variables one at a time. The
interaction is ignored for this part. Just the rows or just the columns are
used, not a mixture. This is the part which is similar to the one-way
ANOVA.
The interaction effect is the effect that one factor has on the other factor.
The degrees of freedom here are the product of the two degrees of
freedom for each factor.
The within variation is the sum of squares within each treatment group.
The total number of groups’ treatment is given by the product of the
number of the levels multiplied for each factor. The within variance is the
within variation divided by its degrees of freedom. The within group is
also called the error.
At the last column, there is an F-test for each of the hypotheses: it is the
mean square for each main effect and the interaction effect divided by the
within variance. The numerator degrees of freedom come from each
effect, and the denominator degrees of freedom are the degrees of
freedom for the within variance in each case.
In the specific case we have:
Table 6.6. Two-way ANOVA table for WL
SS df MS F-test F-crit D 5%
First factor 846.45 2 423.23 1.87 F2 , 72 3,13
Second factor 5616.96 2 2808.48 12.43* F2 , 72 3,13
Interaction effect 10.90 4 2.72 0.01 F4 , 72 2 , 51
Within 16272.65 72 226.01 – –

Total 22746.96 80 –
From the above results, we can see that only the main effect of WL is significant,
but the interaction between the input factors is not. That is, only the second null
hypothesis can be rejected: the workload means are not all equal. Often, the best
way of interpreting and understanding an interaction is by a graph. A two-factor
ANOVA with no significant interaction can be represented by two approximately
parallel lines, whereas a significant interaction results in a graph with non-parallel
lines. Because two lines will rarely be exactly parallel, the significance test on the
interaction is also a test of whether the two lines diverge significantly from being
parallel. If there is no interaction between the considered factors, each mean
reported in Table 6.7 is influenced only by the two factors in an additive way. For
example, if we consider A (factor in the column) and B (factor in the row) without
any interaction, we should obtain:
X ai b j X ai X b j X (6.39)
i.e.: if no interaction exists between A and B, each single considered value of the
table of means results from the sum of the means in the row and in the column
minus the grand mean. In other words, the values shown in Table 6.4 are the
observed values, while the estimated values calculated using Equation 6.40 are
reported in Table 6.7.
Table 6.7. Estimated values for WL

W B S Mean
C.A. 41.21 30.17 20.83 30.74
G.A. 48.02 36.98 27.64 37.55
N.A. 48.12 37.08 27.74 37.65
Mean 45.78 34.74 25.41 35.31
Therefore, it is possible to plot these values and verify that the obtained lines are
all parallel in both cases; this will confirm that there is no interaction between the
factors.
50
45
40
Centralised
35 GameTheory
Negotiation
30
25
20
0 W
1 B
2 S
3 4
Figure 6.6. Observed values
50
45
40
Centralised
35 Game Theory
Negotiation
30
25
20
0 1
W 2B 3S 4
Figure 6.7. Estimated values
The ANOVA test has been used to find out if there is a significant difference
between the considered groups of means. However, the ANOVA analysis simply
indicates if there is a difference between two or more groups of means, but it does
not tell us what mean is significantly different from the others. To understand it, in
the following, the Tukey test has been used. This is a test properly designed to
perform a pairwise comparison of the means to see where the significant difference
is; according to it, it is sufficient to simply calculate the T-critical value and the
difference between all possible pairs of means. Each difference is then compared to
the T-critical value: if this difference is larger than the Tukey value, the
comparison is significant (and it will be indicated by “*”), otherwise it is not.
sr2
T critical QD , p , ab( c 1) , (6.40)
k
being:
D is the selected probability (5%);
p is the simultaneously number of compared means;
ab(c 1) is the df of the within variance;
sr2 is the within variance;
Q is the studentised range statistic;
k is the number of the data necessary to calculate the means we are going
to compare.
Referring to the analysed case, we have:
Table 6.8. T-values for WL
Factor Q-value T-value
226.66
Row factor Q0.05 ,3,72 3.39 T 3.39 9.80
27
226.66
Column factor Q0.05 ,3,72 3.39 T 3.39 9.80
27
226.66
Interaction Q0.05 ,9 ,72 4.53 T 4.53 22.70
9
Using data reported in Table 6.4, we obtain:
Table 6.9. Differences of means for row factors
Row factor C.A. G.A. N.A.

G.A. 6.81 0 0
N.A. 6.91 0.10 0
Table 6.10. Differences of means for column factors
Column factor W B S
B 11.04* 0 0
S 20.37* 9.34 0
It is easy to see that, also with this test, the only influence on means we observe is
due to the column factor, i.e. to the main effect of WL, like already observed by
refusing the only null hypothesis 6.38.
Table 6.11 is referred to the groups’ interaction analysis:
Table 6.11. Groups’ means differences for column factors
W-C.A. B-C.A. S-C.A. W-G.A. B-G.A. S-G.A. W-N.A. B-N.A.

B-C.A. 10.61 – – – – – – –
S-C.A. 19.87 9.26 – – – – – –
W-G.A. 6.70 17.31 26.57* – – – – –
B-G.A. 3.72 6.89 16.15 10.42 – – – –
S-G.A. 13.04 2.43 6.83 19.73 9.32 – – –
W-N.A. 7.94 18.55 27.82* 1.25 11.66 20.98 – –
B-N.A. 4.14 6.47 15.74 10.83 0.42 8.90 12.08 –
S-N.A. 13.57 2.96 6.30 20.27 9.85 0.53 21.52 9.44
Finally, given these results with a confidence interval equal to 5%, we can draw the
following conclusions:
there is no statistical difference between the performance estimated using
the centralised model and the two decentralised ones;
the variability of the input data (WL) significantly influences the response
of the systems;
the input data variability and the different models does not interact
significantly.
This considerations lead us to a very important result: given a fixed variability of
input data (WL) there is, statistically speaking, no difference between the three
proposed models. This is a really good result considering the distributed nature of
the negotiational and game-theoretical approaches.
6.7.1.2 Priority
Similarly to the workload analysis, we start filling in Table 6.12. In the rows we
have the levels of both considered factors (priority and type of model) and, in the
columns, the values of nine different replications obtained combining the other two
input parameters (WL, C) in a combinatorial way, as explained in Table 6.13.
Table 6.12. Replication values for P
1 2 3 4 5 6 7 8 9
W 51.04 39.16 34.10 41.88 27.78 21.69 36.43 18.49 8.74
C.A. B 51.54 34.31 34.65 42.72 28.20 20.66 35.88 18.67 8.45
S 49.98 38.84 34.48 41.86 26.32 21.48 35.22 18.59 8.76
W 67.73 43.64 35.61 59.13 31.56 21.57 52.23 20.90 8.53
G.A. B 64.86 35.10 34.89 58.88 31.65 21.73 55.79 21.77 8.35
S 67.27 43.12 36.15 57.86 30.51 21.74 53.60 20.53 9.05
W 65.83 43.88 36.76 58.57 31.37 22.00 51.28 20.53 8.56
N.A. B 66.14 43.32 34.80 57.72 30.19 21.93 53.39 21.07 8.60
S 67.61 44.58 36.66 57.64 30.40 21.06 53.23 20.61 8.69
Table 6.13. Combination of WL and C
1 2 3 4 5 6 7 8 9
Workload (WL) variance W W W B B B S S S
Ownership (C) variance W B S W B S W B S
The next step is again to refer to a table with all the mean values for each model
and for each level of P. Again, observing Table 6.12 and 6.14, it is possible to
conclude that the two distributed approaches perform slightly worse than the
centralised one by about 20%. Also in this case the performances of the two
decentralised approaches are very close each other.
Table 6.14. Mean values for P
W B S Mean
C.A. 31.03 30.57 30.62 30.74
G.A. 37.88 37.00 37.76 37.55
N.A. 37.64 37.46 37.83 37.65
Mean 35.52 35.01 35.40 35.31
Also in this case, there are three sets of hypotheses to consider, as shown above for
WL. The null and alternative hypothesis for each sets are given, also in this case, by
Equations 6.37 –6.39. Using the same notation asTable 6.5, we have:
Table 6.15. Two-way ANOVA table for P
SS df MS F-Test F-crit D 5%
First factor 846.45 2 423.23 1.39 F2.72 3.13
Second factor 3.81 2 1.91 0.006 F2.72 3.13
Interaction effect 2.03 4 0.51 0.002 F4.72 2.51
Within 21894.67 72 304.09 – –

Total 22746.96 80 –
From the above results, we can see that no factor is significant, including the
interaction between them. Therefore, in this case, it is pointless to use a Tukey test,
because it is impossible to reject all null hypotheses reported in Equations 6.37–
6.39. To further support this consideration, the observed and the estimated values2
are plotted in Figures 6.8 and 6.9.
Table 6.16. Estimated values for P
W B S Mean
C.A. 30.95 30.44 30.83 30.74
G.A. 37.75 37.25 37.64 37.55
N.A. 37.85 37.35 37.74 37.65
Mean 35.52 35.01 35.40 35.31
It is possible verify that the obtained lines are all parallel in both cases: this
confirms that there is no interaction between factors. These conclusions lead us to
understand, considering a confidence interval equal to 5%, that:
there is no statistical difference between the performance estimated using
the centralised model and the two decentralised ones;
the variability of the input data (P) does not significantly influence the
response of the systems;
the input data variability and the different used models does not interact
significantly.
These considerations lead us to another very important result: regardless the
variability of input data (P) there is, statistically speaking, difference between the
three proposed models.
2 This values are calculated using (6.40).

39,00
38,00
37,00
36,00 Centralised
35,00 Game theory
34,00 Negotiation
33,00
32,00
31,00
30,00
W B S
39
38
37
36 Centralised
35 Game theory
34 Negotiation
33
32
31
30
W B S
6.7.1.3 Ownership
The last parameter to investigate is the ownership. Table 6.17 reports the data as
usual. In the rows we have the levels of both factors and, in the columns, the values
of nine different replications obtained combining the other two input parameters
(WL, P) in a combinatorial way, like explained in Table 6.18.
Table 6.17. Replication values for C
1 2 3 4 5 6 7 8 9
W 51.04 51.54 49.98 41.88 42.72 41.86 36.43 35.88 35.22
C.A. B 39.16 34.31 38.84 27.78 28.20 26.32 18.49 18.67 18.59
S 34.10 34.65 34.48 21.69 20.66 21.48 8.74 8.45 8.76
W 67.73 64.86 67.27 59.13 58.88 57.86 52.23 55.79 53.60
G.A. B 43.64 35.10 43.12 31.56 31.65 30.51 20.90 21.77 20.53
S 35.61 34.89 36.15 21.57 21.73 21.74 8.53 8.35 9.05
W 65.83 66.14 67.61 58.57 57.72 57.64 51.28 53.39 53.23
N.A. B 43.88 43.32 44.58 31.37 30.19 30.40 20.53 21.07 20.61
S 36.76 34.80 36.66 22.00 21.93 21.06 8.56 8.60 8.69
Table 6.18. Combination of WL and P
1 2 3 4 5 6 7 8 9
Workload (WL) variance W W W B B B S S S
Priority (P) variance W B S W B S W B S
Table 6.19 shows the means for the data reported in Table 6.17. Also in this case
decentralised approaches perform slightly worse than the centralised and no great
difference can be measured between the two approaches.
Table 6.19. Mean values for C
Ownership W B S Mean
C.A. 42.95 27.82 21.45 30.74
G.A. 59.70 30.97 21.96 37.55
N.A. 59.05 31.77 22.12 37.65
Mean 53.90 30.19 21.84 35.31
Again, the null hypothesis sets to consider is given by Equations 6.37–6.39.

Referring to Table 6.5, we have:
Table 6.20. Two-way ANOVA table for C
First factor 846.45 2 423.23 4.99* F2.72 3.13
Second factor 14937.46 2 7468.73 88.04* F2.72 3.13
Interaction effect 854.84 4 213.71 2.52* F4.72 2.51
Within 6108.21 72 84.84 – –

Total 22746.96 80 –
From the above results, we can see that all three null hypotheses can be rejected.
This means that not only the main effects of considered factors are significant, but
also their interaction. In this case, if we consider the observed values plot, we will
not obtain parallel lines, but divergent ones. Obviously, this consideration is not
true for the estimated values.
Table 6.21 Estimated values for C
Ownership W B S Mean
C.A. 49.33 25.62 17.27 30.74
G.A. 56.14 32.42 24.08 37.55
N.A. 56.24 32.52 24.18 37.65
Mean 53.90 30.19 21.84 35.31
70
60
50
Centralised
40 Game theory
Negotiation
30
20
10
0
W B S

60
50
40 Centralised
Game theory
30 Negotiation
20
10
0
W B S
To better understand where this significant difference is located, we will use the
Tukey test. Referring to Formula 6.40, we have:
Table 6.22. T-values for C
84.84
Row factor Q0.05 ,3,72 3.39 T 3.39 6.01
27
84.84
Column factor Q0.05 ,3,72 3.39 T 3.39 6.01
27
84.84
Interaction Q0.05 ,9 ,72 4.53 T 4.53 13.91
9

G.A. 6,81* 0 0
N.A. 6,91* 0,10* 0
Column factor W B S
B 23.71* 0 0
S 32.06* 8.35* 0
It is easy to see that, also using this test, the influence on means we observe is due
by either the column and row factor, i.e. the main effect of O and the main effect of
the models, like already view refusing all hypotheses 6.37-6.39. Table 6.25 refers
to the interaction:
Table 6.25. Groups means’ differences for column factors

B-C.A. 15.13 * – – – – – – –
S-C.A. 21.50* 6.37 – – – – – –
W-G.A. 16.75* 31.89* 38.26* – – – – –
B-G.A. 11.98 3.16 9.53 28.73* – – – –
S-G.A. 20.99* 5.86 0.51 37.75* 9.02 – – –
W-N.A. 16.09* 31.23* 37.60* 0.66 28.07* 37.09* - -
B-N.A. 11.18 3.95 10.33 27.93* 0.80 9.81 27.27* -
S-N.A. 20.83* 5.70 0.67 37.59* 8.86 0.16 36.93* 9.65
At the end, considering a confidence interval equal to 5%, it is possible to conclude

that:
there is statistical difference between the performance estimated using the
centralised model and the two decentralised ones and also between them;
the variability of the input data (C) significantly influences the response of
the systems;
the input data variability and the different models used interact
significantly.
These results lead us to consider that, in this case, it makes sense to select the most
favourable combination of input parameters to obtain the minimisation of the UC.
In particular, with a small variability of the ownership, the centralised model
returns the minimum value of unallocated capacity.
6.7.2 Design of Experiment (DoE)
In this section, DoE is used to investigate in which way the input variables
influence the response, i.e. the unallocated capacity. In particular, a full factorial
experiment has been conducted: responses have been measure3 at all combinations
of the experimental factor levels. Factorial design allows us to understand the
simultaneous effects that several factors may have on the results. MinitabTM has
been used to analyse the obtained values and to build the following graphs.
3Each response is the estimated average of the unallocated capacity as explained in the
Section 6.6. Relative tables are reported in Appendix A.
6.7.2.1 Centralised Model

In the centralised model, our benchmark, a full factorial analysis has been
conducted to investigate the influence of the input parameters. In Figures 6.12 and
6.13 it is possible to observe the main effects and the interaction between WL, P
and C. Priority main effect is not significant: the output parameter is scarce
conditioned by it. Also, its interaction with others (WL, C) is irrelevant. Different
consideration can be made for workload and ownership. Both their main effect and
their interaction strongly influence the unallocated capacity (UC). In particular, if
we observe the output parameter minimum values, reported in the Table 6.26, we
can see that reached results are indifferent to P but require WL and C at their low
level:
Table 6.26. UC minimum values
Centralised WL P C UC
S W S 8.744
S B S 8.451
S S S 8.762
WL P
40
35
Mean of Unallocated Capacity
30
25
20
W B S W B S
C
40
35
30
25
20
W B S
Figure 6.12. Main effects plot

W B S W B S
50 WL
W
B
30 S
WL
10
50 P
W
B
30 S
P
10
Figure 6.13. Interaction plot
Finally, Figure 6.14 shows the surface response of the models. The first pair of
plots shows how when the workload increases in variance, passing from S to W, the
performance deteriorates since the unallocated capacity grows. This trend is
confirmed observing the second pair of plots: in this case the two factors act
jointly.
In the last situation only the Ownership, when its variance increases, leads to
bad performance; Priority does not have any influence on the surface response.
6.7.2.2 Negotiation
Adopting the same methodology, it is possible to analyse the negotiational model
performance as summarized in Figures 6.15 and 6.16.
S
C W
Figure 6.14. Surface plot for UC (C.A.)
Priority, also in this case, plays a very marginal role. Its main effect and its
interaction is unimportant; therefore, the considerations made in the previous case
are still valid. In the following table the minimum values of UC are indicated and
the surfaces plots are reported in Figure 6.17. Also in this case, as the uncertainty
of the workload and the ownership increases, the negotiation model performance
deteriorates. It is to be noticed that performance deteriorates in a worse way than in
the centralised model. Therefore, the negotiation model is more sensitive to the
uncertainty.
WL P
60
50
40
30
20
W B S W B S
C
60
50
40
30
20
W B S
W B S W B S
WL
W
50
B
S
WL
25
0
P
W
50
B
S
P
25

Negotiation WL P C UC
S W S 8.557
S B S 8.603
S S S 8.688
S
C W
Figure 6.17. Surface plot for UC (N.A.)
6.7.2.3 Game Theory

In this last case, we obtain the effects and interactions reported in the following
figures and tables. Also in this case the game-theoretical model is very sensitive to
variance of input data.
WL P
60
50
40
30
20
W B S W B S
C
60
50
40
30
20
W B S
W B S W B S
WL
W
50
B
S
WL
25
0
P
W
50
B
S
P
25
Game theory WL P C UC
S W S 8.528
S B S 8.353
S S S 9.052
S
C W
Figure 6.20. Surface plot for UC (G.A.)
6.8 References
[1] Fernandez FR, Hinojosa MA, Puerto J (2004) Set-valued TU-games. Eur J Oper Res
159 181–195
[2] Shapley LS, Shubik M (1972) The assignment game: the core. Int J Game Theory
1:111–130
[3] Owen G (1975) On the core of linear production games. Math Progr 9 358–37
[4] Nash J (1950) The bargaining problem. Econometrica, 18(2):155–162
7
Distributed Models for Plant Capacity Allocation
7.1 Introduction
Plant capacity allocation consists in what we have referred as medium-level
production planning in the general overview discussed in Chapter 5. As mentioned,
at this level the objective is to acquire actual production capacity through plant
assignment to groups; this goal is obtained through a bargaining process among the
groups, which act as buyers, and the plants, which act as a sellers of capacity. They
bargain by exchanging plant production capacity and price, which the buyers pay
to gain capacity.
7.2 Initial State

Let us suppose that the company produces g different (g = 1,…,G) product
families, each associated to a group. Such products are produced in m plants (m =
1,…,M) each geographically dispersed and reconfigurable in order to produce all
of the different product families. The reconfiguration is allowed once every three
months and, at that time, the m plants are assigned to the g groups for
accomplishing the annual quarter production of the associated product family. The
group–plant association is addressed, as in the previous chapter, using a centralised
model and two decentralised ones: a negotiational and a game-theoretical
approach. The group–plant association can be interpreted as an internal
transaction. The group will be requested to pay a certain number of fictitious
credits to the plant in order to receive the needed capacity. The number of credits,
pgm Vgm , is the product of the negotiated transaction price, pgm, that the g-th group
pays to the plant m at the end of the quarter, times the volume of product family g
that will be produced by the plants m, Vgm. Indeed, credits might be thought of as a
virtual production capacity and they are equally distributed to the groups at the
beginning of each year.
98 Production Planning in Production Network
7.3 The Centralised Model

In the following centralised model, we will keep the assumption that all the
relevant planning data can be located centrally and can be analysed by a centralised
algorithm. Given the annual quarter capacity ownership Cg , assigned to each
group as an outcome of the high planning level, the agents of groups (hereafter
GAs) look for an agreement with the agents of plants (hereafter PAs) in order to
gain a sufficient number of plants for supplying the required C g . In particular, we
consider that the goal is to minimise the incurring costs of groups using the
following objective function:
§ ·
¦¦ u
ª¬ xgm Qm ( Dgm V gm ) ( Rc wgm ) º¼ ¦ Wg ¨¨ Rg ¦ xgm Qm ¸¸ (7.1)
g m g © m ¹
where:
Qm is the production capacity of the m-th plant;
u
Dgm is a matrix whose elements are the unit transport cost needed to
sustain products transfer from the m-th plant to the g-th regional division;
V gm is a measure of the efficiency of plant m in producing products for
group g; it represents the plant m variable costs for producing this specific
product;
Rc is the reconfiguration cost of plant m when it is allocated to a different
group with respect to that of the previous quarter;
Wg is the unit cost the g-th group needs to sustain when the acquired
capacity is different with respect to the one it needs. It can be interpreted
like a production failure cost when the obtained capacity is less than the
required one and like a unutilised capacity cost on the opposite situation;
R is the capacity requested by group g.
g
The constraints to take into account are:

x gm ^ 1, if plant m is assigned to g
0, otherwise
(7.2)
G
¦x
g 1
gm 1 (7.3)
wgm t x gm k gm (7.4)
where:
k gm ^ 1, if in the antecedent annual quarter m was allocated to g
0, otherwise
(7.5)
Constraint 7.2 defines the decision variables; Constraint 7.3 expresses that each
plant can be assigned only to one group, while Constraints 7.4 and 7.5 are binary
variables that lead us to take into account the reconfiguration costs.
Distributed Models for Plant Capacity Allocation 99
7.4 The Negotiation Model

The negotiational mechanism has the same goal as the centralised one: minimise
the incurring costs of groups. The negotiation is bilateral and simultaneous: all of
the GAs, every three months, concurrently submit their own offer to the PAs.
During the negotiation, the GAs make offers, based on their generative functions,
to the PAs and these, based on their reactive functions, evaluate such proposals and
decide whether they want to sign the contract with a given group or to ask for a
new offer. There is a fixed number of rounds, rmax, for the negotiation being
concluded. If at the end of rmax some of the plants are not allocated to any group,
they are automatically allocated to the groups that in the last round proposed the
best offer. A given plant can be assigned only to one group. The negotiation issue
is the price p gm that the g-th group offers to the m-th plants at each round r. The
GA’ generative functions for the price offer computation and the PA’ reactive
functions for the price offer evaluation are described in what follows.
7.4.1 Generative Function
The generative function allows us to compute the price as defined below:

1 § r 1 Rg callgr 1 Dgmax Dgm
u
·
pr p min ( p gmax p r 1 ) ¨ max ¸, (7.6)
gm g
3 gm¨ rmax 1 Rg Dg Dg ¸¹
min
©
where:
r is the negotiation round;
pgmax is the fixed maximum price (reservation price) that each group can
offer;
p min is the minimum offered price. It is a constant guaranteeing fair
g
(based on common value) offer by the groups to the plants;

Rg , Du have been already defined;
gm
r 1
call g is the capacity already allocated by the g-th group until round
r 1;
Dgmax ^ `
u
max Dgm
m
and Dgmin u
^ `
min Dgm
m
.
As can be observed in Equation 7.6, the price the group g offers to the plant m
increases according to the round of the negotiation (time-dependent tactic), while it
decreases according to the capacity already allocated by the group itself (resource-
dependent tactic).
7.4.2 Reactive Function
The reactive function is represented as:

Wrgm
pm* U mr V gm Rc, (7.7)
where:
r
W gm is the reactive function threshold. The plant m accepts the offer of
r r
group g if and only if pgm t W gm ;
pm* is a constant defined by plant m;
Rc is the reconfiguration cost, already defined;
V gm is the plant m variable costs, already defined;
the last term U mr can be expressed as:
ª r º
U mr U m «1 Imr » , where (7.8)
¬ rmax ¼
Gr r
Igm
Imr ¦G
g 1
r
, and (7.9)
r r 1
r
p gm pgm
Igm r 1 (7.10)
pgm
.
It represents the risk attitude of plant m at round r; it increases according to the
round itself (time-dependent tactic) and decreases with the average gradient
Imr (imitative tactic) between the prices that the groups offer at rounds r and r 1.
G r is the number of groups that are still negotiating at round r;
U m is a constant depending on the plant and it characterizes the decisional
parameter related to its own risk attitude.
The risk is related to a first in, first out (FIFO) discipline in allocating orders to the
plants; that is, throughout the annual quarter the groups allocate the actual
production orders first to those plants which have signed earliest the contract with
the group itself. Thus, plants that are allocated last have a higher risk of remaining
underloaded. Equation 7.7 shows that the utility threshold of the plant agents
increases according with their risk attitude (agent state), their efficiency in
producing product g, and the reconfiguration cost.
7.5 The Game-theoretical Model

The game-theoretical model is quite similar to the one already discussed for
production planning at the high level, which has been presented in Chapter 6. Here
we will put into evidence the main differences with that discussed before. This
time, the outline of the game is a quadruple AG r (G r , P r , T r , O r ) , where:
Gr ^1,..., i ,..., G` is a vector whose element g represents a group at
round r;
P r ^1,..., m ,..., M` is a vector whose element m represents a plant at
round r;
T r is a vector whose elements W gm

r
, represent the credits requested by the
plant m to sell its own capacity to the g-th group. This value is calculated
by the reactive function 7.7;
Or is a matrix whose generic element prgm is the value the group g assigns
to the capacity offered by each player m Pr using the generative
function of Equation 7.6.
A transfer utility game can be associated to the quadruple AG r ; here the whole
players are given by G r P r and the characteristic function Ȟ is defined, given that
group g and plant m make a coalition, as in Equation 7.11:
° p gm r W gmr if p gm r W gmr ! 0.
r
v ( g , m) h gm ® (7.11)
°̄0 if p gm r W gm r d 0.
The consequent assignment problem is solved using, also this time, the core game
solution, and the related Owen’s theorem, if hgm > 0, or the Nash bargaining
approach otherwise. Obviously, the utility functions which has been used is
reported in Equation 7.12 for the GAs and Equation 7.13 for the PAs:
5
Ug ¦P b
t 0
t t ,5 u , 0 d u d 1, (7.12)
Um ' (7.13)
Ug being the Bernstein basis polynomials of degree t with Pt points and ' a
parameter that has been already defined in Equation 6.33.
7.6 The Simulation

As in the high-level production planning case of Chapter 6, a discrete event
simulator, based on a multi-agent system (MAS), has been developed using JAVA.
Optimisation has been performed through the LINGO® package conveniently
linked to the simulation environment. The simulation carries on considering an
events timetable of the agents’ actions: this schedule is utilised to define the
sequence of the events. In the following, class and sequence diagrams will be
utilised to show the simulation environment.
7.6.1 The Simulation Environment
The simulation environment considers, for all different approaches explained

above, the following objects:
Schedule: this object is in charge of managing the sequence of events
during the simulation time;
Model: this object handles the signals related to the orders, coordinates the
activities among groups and plants and selects the strategies;
Group: this object represents the GAs set; it has all the information,
algorithms and functions needed for their implementation;
Plant: this object represents the PAs set; it has all the information,
algorithms and functions needed for their implementation;
Statistical Analysis: this last object is dedicated to the analysis of groups’
and plants’ performance analysis.
The class diagram of Figure 7.1 shows the most relevant methods utilised in the
Java code to obtain the described simulation environment.
Figure 7.1. Class diagram of the medium production planning level
On the other hand, Figure 7.2 shows the sequence diagram; it is referred to a
generic order o, o = 1,..,O:
Figure 7.2. Sequence diagram of the medium production planning level
Schedule object activates all other objects shown here.

o-th order: order number o is communicated to the Model object;
Model object updates information related to o-th order;
Initialsze o-th order: information related to o-th order is communicated to
groups;
Groups calculate the capacity they need;
Needed capacity: groups’ requested capacity is communicated to the
Model;
Requested capacity: requested capacity is communicated to the plants;
Plants calculate their utilities values related to the offered credits;
Utilities values: utilities values are communicated by Plant to Model;
Model applies one of the three different approaches, depending on the considered
model: centralized, negotiational, game-theoretical;
Group/Plant Assignments: Group/Plant assignment is communicated to
Group and Plants objects;
Performance communication: groups’ and plants’ performance measures
are communicated to Statistical Analysis object;
Statistical Analysis object uploads performance measures and carries on with the
simulation until the defined confidence interval is reached;
Statistical performances: performance measures are communicated to
Schedule object;
Schedule verifies the performance measures statistical significance;

Significance communication: Schedule communicates to Statistical
Analysis object the result of the previous check;
Final report is written.
7.6.2 The Simulation Case Study
The test environment consists of 27 different combination of input parameters:

capacity ownership of groups (C), risk attitude ( R { U m ) and capacities (Q) of
plants. In particular, the o-th order is given by a vector (C,R,Q): each index is
considered randomly drawn from three normal distributions N(µ;ı) whose
parameters are reported in Table 7.1 as, Small (S), Bounded (B), Wide (W) with
relation to the standard deviations size.
Table 7.1. Distributions of input parameters
Small (S) Bounded (B) Wide (W)

Ownership (C) N(25;5) N(25;10) N(25;15)
Risk attitude (R) N(0.5;0.025) N(0.5;0.3) N(0.5;0.5)
Capacity (Q) N(10;0.5) N(10;3) N(10;8)
An infinite number of credits have been supposed for each simulation experiment;
for each replication the following performance variables have been computed with
a confidence interval equal to 95%:
G
V ¦¦ Q
g 1 mg
m V gm , (7.14)
G
u
G ¦¦ Q
g 1 mg
m Dgm , (7.15)
G M
- ¦¦ w
g 1m 1
gm , (7.16)
G § ·
Z ¦W
g 1
¨ Rg ¦ Qm ¸ .
g
¨ mg
¸
(7.17)
© ¹
Equation 7.14 expresses the efficiency of the models in assigning plants to groups.
The lower its value, the higher the ability of the models to properly assign plants to
the respective group. On the other hand, Equation 7.15 takes into account the
transportation costs related to the distances existing between plants and regional
divisions, while Equation 7.16 provides a measure of the mean number of
reconfigurations obtained during the simulation. The last performance, i.e. the one
reported in Equation 7.17, is the cost of the difference between the requested and
the actual allocated capacity.
7.7 Results
The simulation parameters as well as the results of the 27 different combinations of
input parameters for the centralised (C.A.), negotiational (N.A.) and game-
theoretical (G.A.) approaches are reported in Appendix B. Also in this case, a
proper analysis of the results has been carried out through the two-way Analysis of
Variance (ANOVA) and the Design of Experiment (DoE) for each estimated
performance.
The obtained results, rearranged in the following tables, can be considered like
a set of data drawn from three normal distributions with mean equal to the three
different grand mean (depending on the utilized model) and the same variance.
Each factor (i.e. independent variable) has three different levels; respectively, the
factor model has three levels corresponding to C.A., N.A., G.A. and factor input
parameter has also three levels corresponding to the variance wideness of the input
data, i.e. wide (W), bounded (B) and small (S) variance. The goal is to understand
in which way the input parameters affect the estimated performance measures.
7.7.1 Efficiency Performance Analysis: Two-way ANOVA
7.7.1.1 Ownership
Table 7.2 reports in the rows the levels of the input factors and in the columns the
values of the nine different combinations of the other two input parameters (R, Q)
as explained in Table 7.3.
Table 7.2. Replication values of V
1 2 3 4 5 6 7 8 9
W 14.45 14.43 12.18 14.36 14.44 12.16 14.43 14.50 14.47
C.A. B 14.67 14.95 12.07 14.75 15.05 12.03 14.77 14.95 12.06
S 15.16 15.59 12.68 15.07 15.66 12.92 15.12 15.10 12.65
W 17.34 17.56 17.72 17.35 17.58 17.67 17.34 17.59 17.31
G.A B 17.64 17.85 18.04 17.59 17.96 17.89 17.48 17.96 17.99
S 17.94 18.77 18.77 17.90 18.65 19.16 17.91 17.92 18.94
W 15.60 15.74 15.95 15.64 15.70 15.71 15.68 15.60 16.50
N.A B 15.90 16.02 15.86 15.54 15.62 15.62 16.93 17.32 17.09
S 15.91 15.92 15.91 15.56 15.62 15.26 17.17 17.03 17.50
Table 7.3. Combination of R and Q
1 2 3 4 5 6 7 8 9
Risk attitude (R) variance W W W B B B S S S
Production capacity (Q) variance W B S W B S W B S
Table 7.4 reports all the mean values for each model and for each level of the
ownership variance. This results will be deeply investigated in what follows.
Table 7.4. Average values of V

W B S Mean
C.A. 13.94 13.92 14.44 14.10
G.A. 17.49 17.82 18.44 17.92
N.A. 15.79 16.21 16.21 16.07
Mean 15.74 15.98 16.36 16.03
Specifically, there are three sets of hypotheses to consider. The null and alternative
hypotheses for each set are the following:
1. for the population means of the first factor (the model):

H0 : PC PN PG
°
®H : not all P are equal
; (7.18)
°̄ 1 i^C ,N ,G`
2. for the population means of the second factor (the variance):

H0 : PW P B PS
°
®H : not all P are equal
; (7.19)
°̄ 1 j^ W ,B,S`
3. and, finally, for the interdependencies:

°H 0 : there is no interaction
® ; (7.20)
°̄H1 : there is interaction
Table 7.5 reports the results of the ANOVA test for V.
Table 7.5. Two-way ANOVA table for V
First factor 197.021 2 98.510 143.081* F2.72 3.13
Second factor 5.302 2 2.651 3.850* F2.72 3.13
Interaction Effect 1.458 4 0.364 0.529 F4.72 2.51
Within 49.572 72 0.688 – –
Total 253.352 80 –
From the analysis of Table 7.4, that only the main effects are significant, while the
interaction between them is not. That is, only the first two null hypotheses 7.18 and
7.19 can be rejected. Figures 7.3 and 7.44 report respectively the observed and
estimated values, whose numerical values are reported in Table 7.6, from which it
is possible to observe the absence of interaction.
Table 7.6. Estimated values for V

W B S Mean
C.A.
13.81 14.05 14.43 14.10
G.A.
17.63 17.87 18.25 17.92
N.A.
15.78 16.03 16.40 16.07
Mean
15.74 15.98 16.36 16.03
19,00
18,00
17,00
Centralised
16,00
Game Theory
15,00
Negotiation
14,00
13,00
12,00
0 W
1 B
2 S3 4
19,00
18,00
17,00
Centralised
16,00 Game Theory
Negotiation
15,00
14,00
13,00
0 W
1 B2 3S 4
As also done in Chapter 6, Table 7.7 shows the T-critical test values.
Table 7.7. T-values

Row factor Q0.05 ,3,72 3.39 T 0.541
Column factor Q0.05 ,3,72 3.39 T 0.541
Interaction Q0.05 ,9 ,72 4.53 T 1.253

G.A. 3.819* 0 0
N.A. 1.972* 1.848* 0

Column factor W B S
B 0.244 0 0
S 0.622* 0.377 0
Last Table is referred to the group’s interaction analysis.
Table 7.10. Groups means differences for column factors
W–C.A. B–C.A. S–C.A. W–G.A. B–G.A. S–G.A. W–N.A. B–N.A.

B–C.A. 0.014 – – – – – – –
S–C.A. 0.504 0.518 – – – – – –
W–G.A. 3.559* 3.573* 3.056* – – – – –
B–G.A. 3.886* 3.900* 3.382* 0.327 – – – –
S–G.A. 4.503* 4.517* 4.000* 0.944 0.617 – – –
W–N.A. 1.855* 1.869* 1.352* 1.704* 2.031* 2.648* – –
B–N.A. 2.276* 2.290* 1.773* 1.283* 1.609* 2.227* 0.421 –
S–N.A. 2.273* 2.288* 1.770* 1.286* 1.612* 2.229* 0.418 0.003
By comparing data from Table 7.7 and Tables 7.8, 7.9 and 7.10 it appears quite
evident that, given a confidence interval equal to 5%,
there is statistical difference between the value of efficiency estimated
using the centralised model and the two decentralised ones and between them
also;
the system passing from W to S;
the input data variability and the different models used do not interact
significantly.
This considerations lead us to conclude that the centralised model really
behaves better with respect to the decentralised ones in the optimisation of the
efficiency. At the same time, a high variability of C seems to have a positive
influence on the estimated performance.
7.7.1.2 Risk Attitude

Similarly to the ownership analysis, Table 7.11 reports the values of the
performance V depending on the two factors levels on the rows, i.e. model
typology and risk attitude variance wideness, and on the combinations of the other
two input parameters (C,Q) in the columns. The association of the combination of
C and Q to the number reported in the columns of Table 7.11 is reported in Table
7.12.
Table 7.11. Replication values for V
1 2 3 4 5 6 7 8 9
W 14.45 14.43 12.18 14.67 14.95 12.07 15.16 15.59 12.68
C.A. B 14.36 14.44 12.16 14.75 15.05 12.03 15.07 15.66 12.92
S 14.43 14.50 14.47 14.77 14.95 12.06 15.12 15.10 12.65
W 17.34 17.56 17.72 17.64 17.85 18.04 17.94 18.77 18.77
G.A. B 17.35 17.58 17.67 17.59 17.96 17.89 17.90 18.65 19.16
S 17.34 17.59 17.31 17.48 17.96 17.99 17.91 17.92 18.94
W 15.60 15.74 15.95 15.64 15.70 15.71 15.68 15.60 16.50
N.A. B 15.90 16.02 15.86 15.54 15.62 15.62 16.93 17.32 17.09
S 15.68 15.60 16.50 16.93 17.32 17.09 17.17 17.03 17.50
Table 7.12. Combination of C and Q
1 2 3 4 5 6 7 8 9
Ownership (C) variance W W W B B B S S S
Production capacity (Q) variance W B S W B S W B S
Table 7.13 reports the mean values for the performance V. As the reader can notice
from Table 7.13 also in this case efficiency seems to be smaller with the
centralised models. This conclusion will be deeply investigated by performing a
two-way ANOVA.

W B S Mean
C.A.
14.02 14.05 14.23 14.10
G.A.
17.96 17.97 17.83 17.92
N.A.
15.79 16.21 16.76 16.25
Mean
15.92 16.08 16.27 16.09
Specifically, there are three sets of hypothesis to consider. The null and alternative
hypothesis for each sets are given, also in this case, by 7.18–7.20. Table 7.14
reports the results of the ANOVA analysis in this case.
Table 7.14. Two-way ANOVA table for R
SS Df MS F-Test F-crit D 5%
First factor 198.034 2 99.017 133.140* F2.72 3.13
Second factor 1.646 2 0.823 1.107 F2.72 3.13
Within 53.547 72 0.744 – –
Total 256.166 80 –
From the above results it is quite evident that only one factor is significant for the
estimated parameter. In order to strongly support this conclusion, also in this case,
the Tukey test has been carried out. The T-values of the test are reported in Table
7.15, while Table 7.16 reports the differences of the means. The Tukey test
confirms the result of the two-way ANOVA. Finally, Table 7.17 reports the
interaction analysis.
Table 7.15. T-values
Row factor Q0.05 ,3,72 3.39 T 0.563
Column factor Q0.05 ,3,72 3.39 T 0.563
Interaction Q0.05 ,9 ,72 4.53 T 1.302
Table 7.16. Differences of the means for row factors

G.A. 3.820* 0 0
N.A. 2.155* 1.665* 0
The last table refers to the group’s interaction analysis.
Table 7.17. Group’mean differences for column factors
W–C.A. B–C.A. S–C.A. W–G.A. B–G.A. S–G.A. W–N.A. B–N.A.
B–C.A. 0.029 – – – – – – –
S–C.A. 0.208 0.179 – – – – – –
W–G.A. 3.936* 3.907* 3.729* – – – – –
B–G.A. 3.953* 3.924* 3.745* 0.017 – – – –
S–G.A. 3.806* 3.777* 3.598* 0.130 0.147 – – –
W–N.A. 1.771* 1.742* 1.563* 2.165* 2.182* 2.035* – –
B–N.A. 2.192* 2.163* 1.984* 1.744* 1.761* 1.614* 0.421 –
S–N.A. 2.739* 2.710* 2.531* 1.197 1.214 1.067 0.968 0.547
The obtained results, with a confidence interval equal to 5%, lead us to the
there is a statistical difference between the performance estimated using
the centralised model and the two decentralised ones and between them also;
the variability of the input data (R) does not significantly influence the
response of the system;
the input data variability and the different models used do not interact
significantly.
These considerations lead us to conclude that, also in this case, high variability of
input data and the centralised model minimize the efficiency.
7.7.1.3 Plants Capacity

The last parameter to be investigated is the plant capacity Q. The performance
measure V values are reported in Table 7.18, while average values are reported in
Table 7.20, which is the basis of the two-way ANOVA.
As the reader can notice, also for the this input variable, the centralised
approach seems to perform better that the decentralised ones. The two null
hypotheses are 7.18 and 7.20. ANOVA analysis results are reported in Table 7.21.
The next step is to calculate a table with all the mean values for each model and
for each level of Q.
Again, the null hypothesis set is given by 7.18–7.20.
In this case, from Table 7.21, we can see that all three null hypotheses can be
rejected. This means that not only the main effects of considered factor are
significant, but also their interaction. In this case, if we consider the observed
values plot, we will not obtain parallel lines, but divergent ones. Obviously, this
consideration is not true for the estimated values reported in Table 7.22. This
graphical analysis is reported in Figures 7.5 and 7.6.
1 2 3 4 5 6 7 8 9
W 14.45 14.36 14.43 14.67 14.75 14.77 15.16 15.07 15.12
C.A. B 14.43 14.44 14.50 14.95 15.05 14.95 15.59 15.66 15.10
S 12.18 12.16 14.47 12.07 12.03 12.06 12.68 12.92 12.65
W 17.34 17.35 17.34 17.64 17.59 17.48 17.94 17.90 17.91
G.A. B 17.56 17.58 17.59 17.85 17.96 17.96 18.77 18.65 17.92
S 17.72 17.67 17.31 18.04 17.89 17.99 18.77 19.16 18.94
W 15.60 15.64 15.68 15.90 15.54 16.93 15.91 15.56 17.17
N.A. B 15.74 15.70 15.60 16.02 15.62 17.32 15.92 15.62 17.03
S 15.95 15.71 16.50 15.86 15.62 17.09 15.91 15.26 17.50
Table 7.19. Combination of C and R
1 2 3 4 5 6 7 8 9
Ownership (C) variance W W W B B B S S S
Risk attitude (R) variance W B S W B S W B S

W B S Mean
C.A.
14.75 14.96 12.58 14,10
G.A.
17.61 17.98 18.17 17,92
N.A.
15.99 16.06 16.16 16,07
Mean
16,12 16,33 15,63 16,03
Table 7.21. Two-way ANOVA table for Q
SS df MS F-Test F-crit D 5%
First factor
197.021 2 98.510 302.927* F2.72 3.13
Second factor
6.961 2 3.480 10.703* F2.72 3.13
Interaction effect
25.956 4 6.489 19.954* F4.72 2.51
23.414 72 0.325 - -
Within
253.352 80 -
Total
Table 7.22. Estimated values of V

W B S Mean
C.A.
14.19 14.40 13.70 14.10
G.A.
18.01 18.22 17.52 17.92
N.A.
16.16 16.38 15.68 16.07
Mean
16.12 16.33 15.63 16.03
Also in this case, the Tukey test, whose T-values are reported in Table 7.23, is
carried out to better understand where this significance is located. Mean
differences are reported in Tables 7.24 and 7.25 and the test brings us to conclude
that the influence on means is due to either the column or the row factor; this
allows us to reject the main effects null hypothesis. On the other hand, Table 7.26
reports the interaction data and it allows us to reject also the interaction null
hypotheses.
19
18
17
Centralised
16 Game theory
Negotiation
15
14
13
12
W B S
19
18
17
Centralised
16 Game theory
Negotiation
15
14
13
12
W B S

Table 7.23. T-values for Q

Row factor Q0 , 05 ,3,72 3, 39 T 0 , 372
Column factor Q0 , 05 ,3,72 3, 39 T 0 , 372
Interaction Q0 , 05 ,9 ,72 4, 53 T 0 , 861

G.A. 3,820* 0 0
N.A. 1,972* 1,848* 0

Column factor W B S
B 0,216 0 0
S 0,485* 0,701* 0
It is easy to see that, also using this test, the influence on means we observe is due
to either the column or the row factor, as already viewed refusing all hypotheses
7.18–7.20. The last table refers to the interaction.
Table 7.26. Group’ mean differences for column factors
W – C.A. B –C.A. S –C.A. W –G.A. B –G.A. S –G.A. W –N.A. B –N.A.
B –C.A. 0.209 – – – – – – –
S –C.A. 2.175* 2.384* – – – – – –
W –G.A. 2.857* 2.648* 5.032* – – – – –
B –G.A. 3.225* 3.016* 5.400* 0.368 – – – –
S –G.A. 3.411* 3.202* 5.586* 0.554 0.186 – – –
W –N.A. 1.238* 1.029* 3.413* 1.619* 1.987* 2.173* – –
B –N.A. 1.309* 1.100* 3.483* 1.548* 1.916* 2.102* 0.071 –
S –N.A. 1.403* 1.194* 3.578* 1.454* 1.822* 2.008* 0.165 0.094
With a confidence interval equal to 5%. it has been highlighted that:

there is a statistical difference between the efficiency estimated using the
centralised model and the two decentralised ones and it is relevant also
incomparing N.A. and G.A.;
the variability of the input data (Q) significantly influences the response
of the systems, specifically when the variance diminishes passing from W
to S and from B to S;
the input data variability and the different models interact significantly
combining their effects on the estimated performance.
7.7.2 Efficiency Performance Analysis: DoE
In the following analysis, DoE is used to investigate in which way the input
variables influence the efficiency V .

In the centralised model, our benchmark, a full factorial analysis has been
conducted in order to investigate the influence of the input parameters. Figure 7.7
reports the main effect’ impact on the efficiency, Figure 7.8 the interaction effect.
C R
11,5
11,0
10,5
Mean of Distances
10,0
9,5
W B S W B S
Q
11,5
11,0
10,5
10,0
9,5
W B S
Figure 7.7. Main effects plot (C.A.)

W B S W B S
C
15,0
W
B
C 13,5
S
12,0
R 15,0
W
B
R 13,5
S
12,0
Figure 7.8. Interaction plot (C.A.)
As the reader can notice, efficiency, V, rapidly decreases when the production
capacity of plants (Q) has a small variance. The other two parameters, and their
interaction, have less influence. These results are reported in the surface response
curve of Figure 7.9.
Figure 7.9. Surface plot for efficiency (C.A.)
7.7.2.2 Negotiation Model

Adopting the same approach, it is possible to analyse the negotiational model. The
main effects are reported in Figure 7.10, and the interaction effect in Figure 7.11.
As the reader can notice, a wide variance of the three input variable can cause a
significant reduction of the efficiency. On the other hand, if variance wideness
decreases the performance approach rises. Figure 7.12 shows the surface curves in
this case.
C R
16,8
16,5
16,2
15,9
Mean Efficiency
15,6
W B S W B S
Q
16,8
16,5
16,2
15,9
15,6
W B S
Figure 7.10. Main effects plot (N.A.)
W B S W B S
C
W
16,8
B
S
C 16,2
15,6
R
W 16,8
B
R
S
16,2
15,6
Figure 7.11. Interaction plot (N.A.)

S
R W
Figure 7.12. Surface plot for efficiency (N.A.)
7.7.2.3 Game Theory

In this last case we obtain the effects shown in Figure 7.13, and interactions in
Figure 7.14. Groups’ ownership and plants’ production capacity play, within this
model, a very significant role: by diminishing the variance of this data input, the
estimated parameters increase in a quasi linear way. Also their interaction leads to
a very strong influence on the output. Finally, the surface plots are reported in
Figure 7.15.
C R
18,50
18,25
18,00
17,75
Mean of Efficiency
17,50
W B S W B S
Q
18,50
18,25
18,00
17,75
17,50
W B S
Figure 7.13. Main effects plot (G.A.)
W B S W B S
C
W 18,6
B
C
S 18,0
17,4
R
18,6
W
B
R 18,0
S
17,4
Figure 7.14. Interaction plot (G.A.)

B S
R W
Figure 7.15. Surface plot for efficiency (G.A.)
7.7.3 Distance Performance Analysis: Two-way ANOVA
7.7.3.1 Ownership
Table 7.27 reports, in the rows, the levels of the input factors while in the columns
the values of the nine different combinations of the other two input parameters (R,
Q) as explained in Table 7.3.
Table 7.28 reports all the mean values for each model and for each level of the
ownership variance. This results will be deeply investigated in what follows.
The null and alternative hypotheses have been already defined in 7.18–7.20.
The ANOVA table we obtain is reported in Table 7.29.
Table 7.27. Replication values of G
1 2 3 4 5 6 7 8 9
W 11.09 11.13 9.48 11.02 11.09 9.47 11.04 11.14 11.08
C.A. B 11.30 11.50 9.40 11.32 11.61 9.35 11.26 11.54 9.30
S 11.67 11.99 9.59 11.58 12.07 9.89 11.61 11.62 9.64
W 13.23 13.36 13.47 13.24 13.52 13.61 13.23 13.40 13.26
G.A. B 13.52 13.58 13.59 13.41 13.65 13.70 13.33 13.62 13.81
S 13.50 13.57 13.67 13.47 13.60 13.69 13.74 13.48 13.60
W 12.15 12.22 12.30 12.19 12.25 12.18 12.19 12.07 13.08
N.A. B 12.37 12.48 12.47 12.18 12.24 12.35 13.39 13.59 13.54
S 12.46 12.63 12.60 12.29 12.38 12.24 13.57 13.54 13.80
Table 7.28. Average values of G

W B S Mean
C.A.
10.73 10.73 11.07 10.84
G.A.
13.37 13.58 13.59 13.51
N.A.
12.29 12.74 12.83 12.62
Mean
12,13 12,35 12,50 12,33
Table 7.29. Two-way ANOVA table for C
First factor 99.713 2 49.856 126.228* F2.72 3.13
Second factor 1.873 2 0.936 2.371 F2.72 3.13
Within 28.438 72 0.395 – –
Total 130.640 80 –
From the analysis of Table 7.28, only the main effect related to the first factor is
significant: the first null hypothesis 7.18 can be rejected. Considering the values
reported in Table 7.28 we can graphically observe the absence of interaction (lines
are approximately parallel) in the Figure 7.16.
14,00
13,00
Centralised
12,00 Game Theory
Negotiation
11,00
10,00
0 1
W 2
B 3S 4
Referring to the analysed case we can calculate the T-critical values and the
following table.
Table 7.30a. T-values

Row factor Q0.05 ,3,72 3.39 T 0.410
Interaction Q0.05 ,9 ,72 4.53 T 0.949
Table 7.30b. Differences of means for row factors

G.A. 2.69* 0 0
N.A. 1.777* 0.892* 0
Table 7.31 refers the group’s interaction analysis.

Finally, given these results with a confidence interval equal to 5%, we can draw
the following conclusions:
there is a statistical difference between the performance estimated using the
centralised model and the two decentralised ones and between them also;
the variability of the input data (C) does significantly influence the
the input data variability and the different models do not interact
significantly.
Statistically speaking, there are differences between the three proposed models.
Ample discrepancy is present considering the distance existing between the
benchmark (C.A.) and the two decentralised approaches, but no great difference
can be measured among G.A. and N.A.
Table 7.31. Group’ means differences for column factors
B -C.A. 0.004 - - - - - - -
S -C.A. 0.346 0.342 - - - - - -
W -G.A. 2.641* 2.637* 2.295* - - - - -
B -G.A. 2.851* 2.847* 2.505* 0.210 - - - -
S -G.A. 2.866* 2.862* 2.519* 0.225 0.014 - - -
W -N.A. 1.566* 1.562* 1.220* 1.075* 1.285* 1.300* - -
B -N.A. 2.010* 2.006* 1.664* 0.631 0.841 0.856 0.444 -
S -N.A. 2.106* 2.102* 1.760* 0.535 0.745 0.759 0.540 0.096

Similarly to the ownership analysis, Table 7.32 reports the values of the
performance G depending on the two factor levels in the rows, i.e. model typology
and risk attitude variance wideness, and on the combinations of the other two input
parameters ( C , Q ) in the columns. The association of the combination of C and Q
to the number reported in the columns of Table 7.32 is reported in Table 7.12.
Table 7.32. Replication values for G
1 2 3 4 5 6 7 8 9
W 11.09 11.13 9.48 11.30 11.50 9.40 11.67 11.99 9.59
C.A. B 11.02 11.09 9.47 11.32 11.61 9.35 11.58 12.07 9.89
S 11.04 11.14 11.08 11.26 11.54 9.30 11.61 11.62 9.64
W 13.23 13.36 13.47 13.52 13.58 13.59 13.50 13.57 13.67
G.A. B 13.24 13.52 13.61 13.41 13.65 13.70 13.47 13.60 13.69
S 13.23 13.40 13.26 13.33 13.62 13.81 13.74 13.48 13.60
W 12.15 12.22 12.30 12.37 12.48 12.47 12.46 12.63 12.60
N.A. B 12.19 12.25 12.18 12.18 12.24 12.35 12.29 12.38 12.24
S 12.19 12.07 13.08 13.39 13.59 13.54 13.57 13.54 13.80
Table 7.33 reports the mean values for the performanceG. As the reader can notice
from Table 7.33 also in this case distance seems to be smaller with the centralised
models. This conclusion will be deeply investigated by performing a two-way
ANOVA.

W B S Mean
C.A.
10.79 10.82 10.92 10.84
G.A.
13.50 13.54 13.50 13.51
N.A.
12.41 12.26 13.20 12.62
Mean
12.23 12.21 12.54 12.33
The hypothesis set is the same as given by Equations 7.18 –7.20.
First factor 99.713 2 49.856 136.771* F2.72 3.13
Second factor 1.800 2 0.900 2.469 F2.72 3.13
Within 26.246 72 0.365 – –
Total 130.640 80 –
From the above results it is quite evident that only one factor is significant for the
estimated parameter. In order to strongly support this conclusion, also in this case,
the Tukey test has been carried out. The T-values of the test are reported in Table
7.35, while Table 7.35 reports the differences of the means. The Tukey test
confirms the result of the two-way ANOVA. Finally, Table 7.37 reports the
interaction analysis.

Row factor Q0.05,3,72 3.39 T 0.394
Interaction Q0.05,9 ,72 4.53 T 0.912

G.A. 2.669* 0 0
N.A. 1.777* 0.892* 0
The last table refers to the groups’ interaction analysis.

Table 7.37. Group’ means differences for column factors

B –C.A. 0.031 – – – – – – –
S –C.A. 0.123* 0.092* – – – – – –
W –G.A. 2.707* 2.676* 2.584* – – – – –
B –G.A. 2.752* 2.721* 2.628* 0.044 – – – –
S –G.A. 2.703* 2.672* 2.580* 0.004 0.049 – – –
W –N.A. 1.618* 1.586* 1.494* 1.090* 1.134* 1.086* – –
B –N.A. 1.463* 1.432* 1.340* 1.244* 1.288* 1.240* 0.154 –
S –N.A. 2.405* 2.374* 2.282* 0.302 0.346 0.298 0.788 0.942
By plotting data shown in Table 7.33 we obtain:
14,00
13,00
Centralised
12,00 Game Theory
Negotiation
11,00
10,00
0 1
W B2 S3 4
It is easy to see that there is some interaction between factors but, with the utilised
confidence interval, it cannot be considered relevant for our analysis.
Consequently, with a confidence interval equal to 5%, we can draw the same
conclusions as reported in the previous section.
7.7.3.3 Plant Capacity

The last parameter to be investigated is the plant capacity Q. As usual the
performance measure G values are reported in Table 7.38, while average values
are reported in Table 7.39, which is the basis of the two-way ANOVA. Also for
this input variable, the centralised approach seems to perform better than the
decentralised ones. The two null hypotheses are 7.18 and 7.20. ANOVA analysis
results are reported in Table 7.40.
1 2 3 4 5 6 7 8 9
W 11.09 11.02 11.04 11.30 9.40 9.35 11.67 9.59 9.89
C.A. B 11.13 11.02 11.04 11.50 11.32 11.26 11.99 11.58 11.61
S 9.48 9.47 11.14 9.40 9.35 11.54 9.59 9.89 11.62
W 13.23 13.24 13.23 13.52 13.41 13.33 13.50 13.47 13.74
G.A. B 13.36 13.52 13.40 13.58 13.65 13.62 13.57 13.60 13.48
S 13.47 13.61 13.26 13.59 13.70 13.81 13.67 13.69 13.60
W 12.15 12.19 12.19 12.37 12.18 13.39 12.46 12.29 13.57
N.A. B 12.22 12.25 12.07 12.48 12.24 13.59 12.63 12.38 13.54
S 12.30 12.18 13.08 12.47 12.35 13.54 12.60 12.24 13.80
The next step is to create a table with all the mean values.
W B S Mean
C.A.
10.48 11.38 10.16 10.68
G.A.
13.41 13.53 13.60 13.51
N.A.
12.53 12.60 12.73 12.62
Mean
12.14 12.50 12.16 12.27
In this case, from Table 7.40, we can see that all three null hypotheses can be
rejected. This means that not only the main effects of the considered factor are
significant, but also their interaction. In this case, if we consider the observed
values plot, we will not obtain parallel lines, but divergent ones. Obviously, this
consideration is not true for the estimated values reported in Table 7.41. This
graphical analysis is reported in Figures 7.18 and 7.19.
Table 7.40. Two-way ANOVA table for Q
First factor
113.571 2 56.786 174.283* F2.72 3.13
Second factor
2.237 2 1.118 3.432* F2.72 3.13
Interaction Effect
5.306 4 1.326 4.071* F4.72 2.51
23.459 72 0.326 – –
Within
144.573 80 –
Total
Table 7.41. Estimated values of G

W B S Mean
C.A.
10.55 10.91 10.57 10.68
G.A.
13.38 13.75 13.41 13.51
N.A.
12.49 12.86 12.52 12.62
Mean
12.14 12.50 12.16 12.27
carried out to better understand where this significance is located. Mean
differences are reported in Tables 7.43 and 7.44 and the reader can notice how the
test brings us to conclude that the influence on means is due to either the column or
the row factor; this allows us to reject the main effects null hypothesis. On the
other hand, Table 7.45 reports the interaction data and it allows us to reject also the
interaction null hypotheses.
Table 7.42. T-values for Q

Row factor Q0.05 ,3,72 3.39 T 0.372
Column factor Q0.05 ,3,72 3.39 T 0.372
Interaction Q0.05 ,9 ,72 4.53 T 0.861
14,00
13,00
Centralised
12,00 Game Theory
Negotiation
11,00
10,00
0 W
1 2B 3S 4
14,00
13,00
Centralised
12,00 Game Theory
Negotiation
11,00
10,00
0 W
1 2B S
3 4

G.A. 2.836* 0 0
N.A. 1.944* 0.892* 0

Column factor W B S
B 0.374* 0 0
S 0.024 0.340 0
Table 7.45 refers to the interaction. In this case, the conclusions we can draw are:
the centralised model and the two decentralised ones. There is a difference
also considering the G.A. and the N.A.;
of the systems, especially passing from W to B;
the input data variability and the different models interact significantly
and, therefore, influence the value of the estimated performance.
Specifically, the centralised approach with a small variance of input data leads
us to obtain the minimum value of the performance. The negotiational approach
works better with respect to the game-theoretical one: the variability of Q does not
significantly influence the results.
Table 7.45. Group’ mean differences for column factors
B-C.A. 0.90* – – – – – – –
S-C.A. 0.32 1.22* – – – – – –
W-G.A. 2.93* 2.02* 3.24* – – – – –
B-G.A. 3.05* 2.15* 3.37* 0.12 – – – –
S-G.A. 3.12* 2.22* 3.44* 0.19 0.07 – – –
W-N.A. 2.05* 1.15* 2.37* 0.87* 0.99* 1.07* – –
B-N.A. 2.12* 1.22* 2.44* 0.81 0.93* 1.00* 0.07 –
S-N.A. 2.25* 1.35* 2.57* 0.68 0.80 0.87* 0.19 0.13
7.7.4 Distance Performance Analysis: DoE
variables influence the distance.

Figure 7.20 reports the main effects impact on the efficiency, and Figure 7.21 the
interaction effect.
C R
11,5
11,0
10,5
10,0
Mean of Distances
9,5
W B S W B S
Q
11,5
11,0
10,5
10,0
9,5
W B S

W B S W B S
12
C
W
B 11
S
10
12
R
W
B 11
R S
10
As the reader can notice, distance rapidly diminishes when the production capacity
of plants (Q) have a small variance. The other two parameters, and their
interaction, have less influence on the estimated performance. In Figure 7.22 the
surfaces response of the model are depicted.

Adopting the same approach, it is possible to analyse the negotiational model. The
main effects are reported in Figure 7.23, interaction effects in Figure 7.24. A wide
variance of the three input variables can cause a significant reduction of the
distance performance. Figure 7.25 shows the surface curves in this case.
S
B S
R W
Figure 7.22. Surface plot for distance (C.A.)

C R
13,2
12,9
12,6
Mean of Distances
12,3
W B S W B S
Q
13,2
12,9
12,6
12,3
W B S
W B S W B S
C 13,5
W
B
13,0
S
C
12,5
R 13,5
W
B 13,0
R S
12,5

B S
R W
7.7.4.3 Game Theory

In this last case we obtain the following effects, Figure 7.26, and interactions,
Figure 7.27. Groups’ ownership and plants’ production capacity play, within this
model, a very significant role: by diminishing the variance of this data input, the
estimated parameters increase. Also their interaction leads to a very strong
influence on the output. Finally, the surface plots are reported in Figure 7.28.
C R
13,60
13,55
13,50
13,45
Mean of Distances
13,40
W B S W B S
Q
13,60
13,55
13,50
13,45
13,40
W B S
W B S W B S
C
W 13,6
B
S
C
13,4
13,2
R
W 13,6
B
R S
13,4
13,2

S
R W
Figure 7.28. Surface plot for distance (G.A.)
7.7.5 Number of Reconfigurations Performance Analysis: Two-way ANOVA
7.7.5.1 Ownership
Table 7.46 reports, in the rows, the levels of the input factors and in the columns
the values of the nine different combinations of the other two input parameters
(R,Q) as explained in Table 7.3.
Next step is to calculate a table with all the mean values for each model and for
each level of C : in this way it will be easier to formulate and verify the
hypotheses 7.18 –7.20 related to the two-way ANOVA.
From the analysis of Table 7.48 results that not only the main effects are
significant, but the interaction between them also. That is, all the three null
hypothesis 7.18 –7.20 can be rejected. Figures 7.29 and 7.30 report respectively the
observed and estimated values, whose numerical values are reported in Table 7.49,
from which it is possible to observe the absence of interaction.
The estimated values are reported in Table 7.49.
Table 7.46. Replication values of -
1 2 3 4 5 6 7 8 9
W 8.63 8.83 6.68 8.58 8.77 6.76 8.61 8.74 8.51
C.A. B 8.44 8.57 5.70 8.48 8.62 5.88 8.51 8.65 5.67
S 8.36 8.72 4.53 8.48 8.69 4.54 8.35 8.37 4.59
W 7.31 7.30 7.06 7.30 7.17 7.05 7.31 7.11 7.35
G.A. B 6.09 5.47 5.20 6.02 5.51 5.41 6.08 5.39 5.21
S 4.37 3.03 2.44 4.37 3.01 2.72 4.38 4.41 2.49
W 8.39 8.52 8.62 7.29 7.56 7.37 7.51 7.29 6.04
N.A. B 7.22 7.35 7.16 6.00 5.97 6.09 4.60 4.35 4.16
S 6.25 6.12 6.18 4.83 4.56 4.71 3.19 3.15 2.30
Table 7.47. Mean values of -

W B S Mean
C.A.
8.23 7.61 7.18 7.68
G.A.
7.22 5.60 3.47 5.43
N.A.
7.62 5.88 4.59 6.03
Mean
7.69 6.36 5.08 6.38
Table 7.48. Two-way ANOVA table for -
First factor 73.123 2 36.561 27.633* F2.72 3.13
Second factor 92.136 2 46.068 34.817* F2.72 3.13
Within 95.265 72 1.323 - -
Total 278.809 80 -
Table 7.49. Estimated values for -

W B S Mean
C.A.
8.99 7.66 6.38 7.68
G.A.
6.74 5.41 4.13 5.43
N.A.
7.34 6.01 4.73 6.03
Mean
7.69 6.36 5.08 6.38
Referring to the analysed case we can calculate the T-critical values reported in
Table 7.50.

Row factor Q0.05 ,3,72 3.39 T 0.750
Column factor Q0.05 ,3,72 3.39 T 0.750
Interaction Q0.05 ,9 ,72 4.53 T 1.374
Using data reported in Table 7.47, we obtain what is reported in Table 7.51 and
7.52.

G.A. 2.247* 0 0
N.A. 1.648* 0.599 0
7
Centralised
6 Game Theory
Negotiation
5
3
0 W
1 2B 3S 4
10
7 Centralised
Game Theory
6
Negotiation
5
3
0 1
W 2
B 3S 4

Column factor W B S
B 1.329* 0 0
S 2.612* 1.284* 0
Last table is referred to the groups’ interaction analysis.
Table 7.53. Groups’ means differences for column factors

B-C.A. 0.621 - - - - - - -
S-C.A. 1.053 0.432 - - - - - -
W-C.A. 1.016 0.395 0.038 - - - - -
B-G.A. 2.637* 2.016* 1.583 1.621 - - - -
S-G.A. 4.764* 4.143* 3.710* 3.748* 2.127* - - -
W-N.A. 0.613 0.008 0.441 0.403 2.024* 4.151* - -
B-N.A. 2.357* 1.736 1.304 1.342 0.279 2.407* 1.744* -
S-N.A. 3.648* 3.027* 2.595* 2.633* 1.011 1.116 3.035* 1.291
Given these results with a confidence interval equal to 5%, we can draw the
C.A. and the two decentralized ones but not between G.A. and N.A.;
the system;
the input data variability and the different models used interact
significantly in the performance evaluation.
These considerations lead us to a very important result: there is, statistically
speaking, no difference between the two proposed decentralised models. The better
one proves to be the G.A., and in a particular case, i.e. for small variability of input
data, this approach leads us to obtain the minimum value of the number of
reconfigurations. This is a really good result considering the distributed nature of
this approach: it can be utilised to reduce to the minimum value the global number
of reconfigurations.

Table 7.54 reports the values of the performance depending on the two factor
levels in the rows, i.e. model typology and risk attitude variance wideness, and on
the combinations of the other two input parameters (C,Q) on the columns. The
association of the combination of C and Q to the number reported in the columns
of Table 7.54 is reported in Table 7.12.
Table 7.55 reports the mean values for the performance - . As the reader can
notice, in this case, - seems to be higher with the centralised model. This
conclusion will be deeply investigated by performing a two-way ANOVA.
Table 7.54. Replication values for -
1 2 3 4 5 6 7 8 9
W 8.63 8.83 6.68 8.44 8.57 5.70 8.36 8.72 4.53
C.A. B 8.58 8.77 6.76 8.48 8.62 5.88 8.48 8.69 4.54
S 8.61 8.74 8.51 8.51 8.65 5.67 8.35 8.37 4.59
W 7.31 7.30 7.06 6.09 5.47 5.20 4.37 3.03 2.44
G.A. B 7.30 7.17 7.05 6.02 5.51 5.41 4.37 3.01 2.72
S 7.31 7.11 7.35 6.08 5.39 5.21 4.38 4.41 2.49
W 8.39 8.52 8.62 7.22 7.35 7.16 6.25 6.12 6.18
N.A. B 7.29 7.56 7.37 6.00 5.97 6.09 4.83 4.56 4.71
S 7.51 7.29 6.04 4.60 4.35 4.16 3.19 3.15 2.30
Table 7.55. Average values of -

W B S Mean
C.A.
7.61 7.65 7.78 7.68
G.A.
5.36 5.40 5.53 5.43
N.A.
7.31 6.04 4.73 6.03
Mean
6.76 6.36 6.01 6.38
The null and alternative hypotheses to consider for each set are given, also in this
case, by 7.18 –7.20.
First factor 73.123 2 36.561 15.009* F2.72 3.13
Second factor 7.626 2 3.813 1.565 F2.72 3.13
Within 175.391 72 2.436 – –
Total 278.809 80 –
From the above results it is evident that only one factor is significant for the
estimated parameter, but also the interaction plays a very important role. In order
to strongly support this conclusion, also in this case, the Tukey test has been
carried out. The T-values of the test are reported in Table 7.58, while Table 7.59
reports the differences of the means. The Tukey test confirms the result of the two-
way ANOVA. Finally, Table 7.60 reports the interaction analysis.

W B S Mean
C.A.
8.06 7.66 7.31 7.68
G.A.
5.81 5.41 5.06 5.43
N.A.
6.41 6.01 5.66 6.03
Mean
6.76 6.36 6.01 6.38
Figure 7.32, whose numerical values are reported in Table 7.57, shows the absence
of interaction given by the plotted parallel lines.

Row factor Q0.05 ,3,72 3.39 T 1.018
Column factor Q0.05 ,3,72 3.39 T 1.018
Interaction Q0.05 ,9 ,72 4.53 T 2.357
Using data reported in Table 7.55, Table 7.59 can be created.


G.A. 2.247* 0 0
N.A. 1.648* 0.599 0
Centralised
6 Game Theory
Negotiation
4
0 W
1 B
2 3S 4
7 Centralised
Game Theory
6 Negotiation
4
0 W
1 2B 3S 4
Table 7.60 refers to the groups’ interaction analysis.

Table 7.60. Groups’ mean differences for column factors

B-C.A. 0.038 - - - - - - -
S-C.A. 0.169 0.130 - - - - - -
W-G.A. 2.243 2.281 2.412* - - - - -
B-G.A. 2.210 2.248 2.379* 0.033 - - - -
S-G.A. 2.082 2.120 2.250 0.161 0.128 - - -
W-N.A. 0.293 0.332 0.462 1.949 1.916 1.788 - -
B-N.A. 1.567 1.605 1.735 0.676 0.643 0.515 1.273 -
S-N.A. 2.876* 2.915* 3.045* 0.634 0.666 0.795 2.583* 1.310
the centralised model and the two decentralised ones but not between G.A.
and N.A.;
the variability of the input data (R) does not significantly influence the
the input data variability and the different models interact significantly.
These considerations lead us to the same conclusions as shown in the previous
section: G.A. is the best approach in optimising the global number of
reconfiguration.

The last parameter to be investigated is the plant capacity Q . The performance
measure - values are reported in Table 7.61, while average values are reported in
Table 7.61, which is the basis of the two-way ANOVA. Also for this input
variable, the centralised approach seems to perform worst that the decentralised
ones. However, also in this case the two null hypotheses are 7.18–7.20. ANOVA
analysis results are reported in Table 7.63.
Table 7.61. Replication values of -

1 2 3 4 5 6 7 8 9
W 8.63 8.58 8.61 8.44 8.48 8.51 8.36 8.48 8.35
C.A. B 8.83 8.77 8.74 8.57 8.62 8.65 8.72 8.69 8.37
S 6.68 6.76 8.51 5.70 5.88 5.67 4.53 4.54 4.59
W 7.31 7.30 7.31 6.09 6.02 6.08 4.37 4.37 4.38
G.A. B 7.30 7.17 7.11 5.47 5.51 5.39 3.03 3.01 4.41
S 7.06 7.05 7.35 5.20 5.41 5.21 2.44 2.72 2.49
W 8.39 7.29 7.51 7.22 6.00 4.60 6.25 4.83 3.19
N.A. B 8.52 7.56 7.29 7.35 5.97 4.35 6.12 4.56 3.15
S 8.62 7.37 6.04 7.16 6.09 4.16 6.18 4.71 2.30
The next step is to calculate a table with all the mean values for each model and for
each level of Q .
Table 7.62. Average values of -

W B S Mean
C.A.
8.49 8.66 5.87 7.68
G.A.
5.92 5.38 4.99 5.43
N.A.
6.14 6.10 5.85 6.03
Mean
6.85 6.71 5.57 6.38
Again, the nulls and alternative hypotheses for each set are given by 7.18 –7.20.
From the results reported in figure 7.63, we can see that only the main effects
are significant, but the interaction between them is not (even if its value is very
close to the one of the threshold. This nearness has its effect on the graphs of mean
values). That is, only the first two null hypotheses 7.18 and 7.19 are rejected. The
estimated values are reported in Table 7.64, while the graphical representations of
Tables 7.62 and 7.643 are reported, respectively, in Figures 7.33 and 7.34.
Table 7.63. Two-way ANOVA table for -
First factor 73.123 2 36.561 16.731* F2.72 3.13
Second factor 26.631 2 13.316 6.093* F2.72 3.13
Within 157.340 72 2.185 – –
Total 278.809 80 –

W B S Mean
C.A.
8.15 8.01 6.87 7.68
G.A.
5.90 5.76 4.62 5.43
N.A.
6.50 6.36 5.22 6.03
Mean
6.85 6.71 5.57 6.38
carried out to better understand where this significance is located.

Row factor Q0.05 ,3,72 3.39 T 0.964
Column factor Q0.05 ,3,72 3.39 T 0.964
Interaction Q0.05 ,9 ,72 4.53 T 2.232
Mean differences are reported in Tables 7.66 and 7.67 and the reader can notice
how the test allows us to conclude that the influence on means is due to either the
column or the row factors; this brings us to reject main effects null hypotheses.

G.A. 2.247* 0 0
N.A. 1.648* 0.599 0

Column factor W B S
B 0.138 0 0
S 1.279* 1.141* 0
7 Centralised
Game Theory
6 Negotiation
4
0 1
W 2
B 3S 4
7 Centralised
Game Theory
6 Negotiation
4
0 W
1 B
2 S
3 4

Table 7.68 refers to group’s interaction analysis.
Table 7.68. Groups’ mean’ differences for column factors

B-C.A. 0.168 - - - - - - -
S-C.A. 2.621* 2.789* - - - - - -
W-G.A. 2.579* 2.747* 0.042 - - - - -
B-G.A. 3.115* 3.283* 0.494 0.536 - - - -
S-G.A. 3.502* 3.670* 0.881 0.923 0.387 - - -
W-N.A. 2.352* 2.520* 0.269 0.227 0.763 1.150 - -
B-N.A. 2.398* 2.566* 0.223 0.180 0.716 1.103 0.046 -
S-N.A. 2.647* 2.815* 0.026 0.068 0.468 0.855 0.295 0.248
Also in this last case, the conclusions we can highlight are the same as the previous
two sections. The only difference consists in the influence that the variability of
input data plays in the estimation of the performance. Passing from W to S and
from B to S there is a relevant reduction of the number of reconfigurations.
7.7.6 Number of Reconfigurations Performance Analysis: DoE
variables influence the number of reconfigurations.

In the centralised model, regarding the main effects on the estimated performance,
we have the main effect impact reported in Figure 7.35, while in Figure 7.36 the
interaction effects are shown.
C R
9
8
Mean of Reconfigurations
W B S W B S
Q
9
W B S
W B S W B S
C
8
W
B
C
S 6
4
R 8
W
R
B
6
S
The number of reconfigurations rapidly diminishes when the production capacity

of plants (Q) and the ownership of groups have a small variance. The risk attitude
of plants influences less the estimated parameter. Given this result, it is interesting
to show some graphs representing the surface response of the model reported in
Figure 7.37.
B B S
R W
Figure 7.37. Surface plot for reconfiguration (C.A.)

Adopting the same approach, it is possible to analyse the negotiational model.
Main effects are reported in Figure 7.38, and interaction effects in Figure 7.39. A
wide variance of the C and R input variables can cause a significant increase in the
performance. Figure 7.40 shows the surface curves in this case.
C R
8
6
Mean of Reconfiguration
W B S W B S
Q
8
W B S
W B S W B S
C 8
W
B 6
C S
4
R 8
W
B 6
R
S
4

S
B S
R W
Figure 7.40. Surface plot for reconfiguration (N.A.)
7.7.6.3 Game Theory

In this last case we obtain the effects shown in Figure 7.41 and interactions shown
in Figure 7.42. Groups’ ownership and plants’ production capacity play, within this
model, a very significant role: by increasing the variance of this data input, the
estimated parameters increase in a quasi linear way. Also their interaction leads to
a very strong influence on the output. Finally, the surface plots are reported in
Figure 7.43.
C R
7
5
Mean of Reconfiguration
3
W B S W B S
Q
7
3
W B S
W B S W B S
C
W 6
B
S
C
4
2
R
W 6
B
R S
4

B S
R W
Figure 7.43. Surface plot for reconfiguration (G.A.)
7.7.7 Absolute Residual Performance Analysis: Two-way ANOVA
7.7.7.1 Ownership
To start with the analysis, Table 7.69 has been taken into consideration. In the rows
we have the levels of both considered factors and, in the columns, the values of the
nine different replications obtained combining the other two input parameters
(R,Q) in a combinatorial way.
The next step is to calculate a table with all the mean values for each model and
for each level of C : in this way it will be easier to formulate and verify the
hypothesis related to the two-way ANOVA.
Table 7.69. Replication values of Z
1 2 3 4 5 6 7 8 9
W 52.84 46.06 54.88 53.34 45.95 56.81 53.36 46.86 54.36
C.A. B 43.52 33.20 44.26 42.49 32.36 43.91 43.39 32.65 44.52
S 31.64 19.68 35.95 31.06 19.80 36.80 32.12 32.09 35.35
W 86.18 93.29 86.18 116.24 92.72 86.32 114.49 93.14 115.08
G.A. B 97.42 71.64 64.13 96.17 71.42 64.39 96.28 71.70 65.64
S 86.26 57.34 51.57 84.64 57.83 52.35 84.89 85.35 51.77
W 99.45 78.21 72.40 99.62 79.56 73.29 100.04 79.07 100.97
N.A. B 90.71 66.76 60.80 90.96 67.99 60.64 91.44 69.22 62.96
S 80.35 54.68 48.36 80.85 54.54 48.01 80.98 79.68 50.35
Table 7.70. Average values of Z

W B S Mean
C.A. 51,61 40,03 30,50 40,71
G.A. 98,18 77,64 68,00 81,28
N.A. 86,96 73,50 64,20 74,89
Mean 78,92 63,72 54,23 65,62
The centralised model seems to perform better in comparison to the two

decentralised ones. The nulls and alternative hypotheses are given by 7.18–7.20.
The ANOVA table we obtain is reported in Table 7.71.
Table 7.71. Two-way ANOVA table for Z
First factor 25685.835 2 12842.917 86.727* F2.72 3.13
Second factor 8370.568 2 4185.284 28.263* F2.72 3.13
Within 10662.069 72 148.084 – –
Total 44992.935 80 –
From the results reported in Table 7.71, we can see that only both main effects are
significant, their interaction is not. That is, only first two null hypothesis 7.18 and
7.19 are rejected. The estimated values are reported in the following Table 7.72.
Table 7.72. Estimated values for Z
W B S Mean
C.A.
54.00 38.81 29.32 40.71
G.A.
94.57 79.38 69.88 81.28
N.A.
88.18 72.99 63.49 74.89
Mean
78.92 63.72 54.23 65.62
By plotting these values, see Figures 7.44 and 7.45, it is easy to observe the
absence of interaction as highlighted above.
Referring to the analyzed case we can calculate the T-critical values reported in
Table 7.73.
Table 7.73. T-Values

Row factor Q0.05 ,3,72 3.39 T 7.939
Column factor Q0.05 ,3,72 3.39 T 7.939
Interaction Q0.05 ,9 , 72 4.53 T 18.375
105
95
85
75 Centralised
65 Game Theory
55 Negotiation
45
35
25
0 W
1 2B 3S 4
105
95
85
75 Centralised
65 Game Theory
55 Negotiation
45
35
25
0 W
1 2B 3S 4

G.A. 40.56* 0 0
N.A. 34.17* 6.39 0

Column factor W B S
B 15.19* 0 0
S 24.68* 9.49* 0
Table 7.76 reports group’s interaction analysis.
following conclusions.
the centralised model and the two decentralised ones. Between the last
two, the existing difference is not relevant;
the variability of the input data (C) significantly influence the response of
the systems;
the input data variability and the different used models do not interact
significantly.
The centralised model proves to be the best one in the optimization of this
performance. N.A. acts, statistically speaking, in the same way as G.A.. For all
three approaches, the reduction of variability of C leads to the best values of

absolute residual.
B-C.A. 11 .574 – – – – – – –
S-C.A. 21.106* 9.532 – – – – – –
W-G.A. 46.577* 58.151* 67.683* – – – – –
B-G.A. 26.037* 37.611* 47.144* 20.540* – – – –
S-G.A. 16.395 27.969* 37.501* 30.182* 9.643 – – –
W-N.A. 35.352* 46.926* 56.458* 11.225 9.314 18.957* – –
B-N.A. 21.892* 33.466* 42.998* 24.685* 4.145 5.497 13.460 –
S-N.A. 12.594 24.168* 33.700* 33.983* 13.444 3.801 22.758* 9.298

We start filling in Table 7.77. The values on the columns are the nine different
replications obtained combining the other two input parameters (C, Q) in a
combinatorial way.
Table 7.77. Replication values for Z
1 2 3 4 5 6 7 8 9
W 52.84 46.06 54.88 43.52 33.20 44.26 31.64 19.68 35.95
C.A. B 53.34 45.95 56.81 42.49 32.36 43.91 31.06 19.80 36.80
S 53.36 46.86 54.36 43.39 32.65 44.52 32.12 32.09 35.35
W 86.18 93.29 86.18 97.42 71.64 64.13 86.26 57.34 51.57
G.A. B 116.24 92.72 86.32 96.17 71.42 64.39 84.64 57.83 52.35
S 114.49 93.14 115.08 96.28 71.70 65.64 84.89 85.35 51.77
W 99.45 78.21 72.40 90.71 66.76 60.80 80.35 54.68 48.36
N.A. B 99.62 79.56 73.29 90.96 67.99 60.64 80.85 54.54 48.01
S 100.04 79.07 100.97 91.44 69.22 62.96 80.98 79.68 50.35
The next step is to create Table 7.78 with all the mean values for each model and
for each level of R : in this way it will be easier to formulate and verify the
hypothesis related to the two-way ANOVA whose results are reported in Table
7.79.

W B S Mean
C.A.
40.23 40.28 41.63 40.71
G.A.
77.11 80.23 86.48 81.28
N.A.
72.41 72.83 79.41 74.89
Mean
63.25 64.45 69.18 65.62
There are three sets of hypothesis to consider, like above showed for Ownership.
The null and alternative hypothesis for each set are given, also in this case, by
Equations 7.18–7.20.
First factor 25685.803 2 12842.902 49.692* F2.72 3.13
Second factor 530.221 2 265.116 1.027 F2.72 3.13
Within 18608,295 72 258,448 – –
Total 44992,934 80 –
From the above results, we can see that only the first factor is significant with its
main effect. Given this result we can plot the (means of) observed values and their
estimated ones (see Figures 7.46 and 7.47).
W B S Mean
C.A. 38.34 39.54 44.26 40.71
G.A. 78.90 80.10 84.83 81.28
N.A. 72.51 73.71 78.44 74.89
Mean 63.25 64.45 69.18 65.62
90
80
70
Centralised
60 Game Theory
Negotiaton
50
40
30
0 W
1 B
2 S3 4
90
80
70
Centralised
60 Game Theory
Negotiation
50
40
30
0 W
1 2B 3S 4
Referring to the analysed case we can calculate the T-critical values reported in
Table 7.81.

Row factor Q0.05 ,3.72 3.39 T 10.488
Interaction Q0.05 ,9.72 4.53 T 24.275
Using data reported in Table 7.78, we obtain data of Table 7.88.


G.A. 40.56* 0 0
N.A. 34.17* 6.39 0
Table 7.83 reports groups’ interaction analysis.

B-C.A. 0.055 - - - - - - -
S-C.A. 1.408 1.353 - - - - - -
W-G.A. 36.887* 36.833* 35.479* - - - - -
B-G.A. 40.006* 39.951* 38.598* 3.119 - - - -
S-G.A. 46.259* 46.204* 44.851* 9.372 6.253 - - -
W-N.A. 32.188* 32.134* 30.781* 4.699 7.818 14.070 - -
B-N.A. 32.605* 32.551* 31.198* 4.282 7.401 13.653 0.417 -
S-N.A. 39.187* 39.132* 37.779* 2.300 0.819 7.072 6.999 6.581
Given these results with a confidence interval equal to 5%, we can conclude that
the only relevant difference is between the performance estimated using the
centralised model and the two decentralised ones. G.A. and N.A. are statistically
equivalent.

In the rows of Table 7.84 we have the levels of both considered factors and, in the
columns, the values of the nine different replications obtained combining the other
two input parameters (C , R ) in a combinatorial way.
The next step is to create Table 7.85 with all the mean values for each model
and for each level of Q : in this way it will be easier to formulate and verify the
hypotheses related to the two-way ANOVA whose results are reported in Table
7.86.
1 2 3 4 5 6 7 8 9
W 52.84 53.34 53.36 43.52 42.49 43.39 31.64 31.06 32.12
C.A. B 46.06 45.95 46.86 33.20 32.36 32.65 19.68 19.80 32.09
S 54.88 56.81 54.36 44.26 43.91 44.52 35.95 36.80 35.35
W 86.18 116.24 114.49 97.42 96.17 96.28 86.26 84.64 84.89
G.A. B 93.29 92.72 93.14 71.64 71.42 71.70 57.34 57.83 85.35
S 86.18 86.32 115.08 64.13 64.39 65.64 51.57 52.35 51.77
W 99.45 99.62 100.04 90.71 90.96 91.44 80.35 80.85 80.98
N.A. B 78.21 79.56 79.07 66.76 67.99 69.22 54.68 54.54 79.68
S 72.40 73.29 100.97 60.80 60.64 62.96 48.36 48.01 50.35

W B S Mean
C.A.
42.64 34.29 45.20 40.71
G.A.
95.84 77.16 70.83 81.28
N.A.
90.49 69.97 64.20 74.89
Mean
76.32 60.47 60.08 65.62
There are three sets of hypotheses to consider. The null and alternative hypotheses
for each set are given by Equations 7.18–7.20. The ANOVA table we obtain is:
First factor 25685.83 2 12842.92 75.55* F2.72 3.13
Second factor 4638.21 2 2319.11 13.64* F2.72 3.13
Within 12239.67 72 170.00 - -
Total 44992.93 80 -
From the above results, we can see that interactions between factors and their main
effects are relevant. That means that all three null hypothesis 7.18–7.20 are
rejected. In Table 7.87 the estimated values are reported.

W B S Mean
C.A.
51.41 35.56 35.16 40.71
G.A.
91.97 76.13 75.73 81.28
N.A.
85.58 69.73 69.34 74.89
Mean
76.32 60.47 60.08 65.62
Referring to the analysed case, the Tukey test results are reported in Table 7.87.

Row factor Q0.05 ,3.72 3.39 T 7.939
Column factor Q0.05 ,3.72 3.39 T 7.939
Interaction Q0.05 ,9.72 4.53 T 18.375
Using data reported in Table 7.85, Table 7.89 and 7.90 can be built.

G.A. 40.56* 0 0
N.A. 34.17* 6.39 0
100
90
80
70 Centralised
Game Theory
60 Negotiation
50
40
30
0 W
1 B
2 3S 4

100
90
80
70 Centralised
Game Theory
60 Negotiation
50
40
30
0 W
1 2B 3S 4

Column factor W B S
B
15.85* 0 0
S 16.25* 0.40 0
Table 7.91 reports groups’ interaction analysis.

B-C.A. 8.345 - - - - - - -
S-C.A. 2.564 10.910 - - - - - -
W-G.A. 53.202* 61.547* 50.638* - - - - -
B-G.A. 34.520* 42.865* 31.956* 18.682 - - - -
S-G.A. 28.186* 36.532* 25.622* 25.016* 6.333 - - -
W-N.A. 47.850* 56.195* 45.286* 5.352 13.330 19.663 - -
B-N.A. 27.328* 35.673* 24.764* 25.874* 7.192 0.858 20.522* -
S-N.A. 21.559* 29.904* 18.995 31.643* 12.961 6.628 26.291* 5.769
the only relevant difference is between the performance estimated using
the centralised model and the two decentralised ones. G.A. and N.A. are
statistically equivalent;
of the systems, especially passing from W to B and from W to S;
the input data variability and the different used models interact
significantly in the estimation of the performance.
7.7.8 Absolute Residual Performance Analysis: DoE
In this analysis, DoE is used to investigate in which way the input variables
influence the absolute residual.

In the centralised model, regarding the main effects on the estimated performance,
simulation results are reported in Figures 7.50 and 7.51.
C R
50
45
40
Mean of Absolute Residual
35
30
W B S W B S
Q
50
45
40
35
30
W B S

W B S W B S
C
50
W
B
C 35
S
20
R 50
W
R
B
35
S
20
The absolute residual rapidly diminishes when the capacity ownership of groups
has a small variance. The dependence by the risk attitude of plants is bounded
while, regarding the capacity of plants, the minimum value of the estimated
parameter is reached when the variance is at level B. This trend is also observable
in the interaction plot. The surface responses of the model are reported in Figure
7.52.
S
B S
R W
Figure 7.52. Surface plot for absolute residual (C.A.)

Adopting the same methodology, it is possible to analyse the negotiational model
(Figures 7.53 and 7.54).
C R
90
80
70
W B S W B S
Q
90
80
70
W B S
W B S W B S
100
C
W
B
S 75
C
50
100
R
W
B
S 75
R
50
Wide variance of ownership and capacity of plants leads us to obtain the higher
values of the absolute residual. The diminishing wideness of this data input means
that in an approximately linear way, the value of the estimated performance

diminishes. Surface plots are reported in Figure 7.55.
S
R W
Figure 7.55. Surface plot for absolute residual (N.A.)

7.7.8.3 Game Theory

In this last case we obtain the effects and interactions reported in Figures 7.56 and
7.57.
C R
100
90
80
70
W B S W B S
Q
100
90
80
70
W B S
W B S W B S
C 100
W
B
C 75
S
50
R 100
W
R
B
75
S
50

Wide variance of ownership and capacity of plants leads us to obtain higher values
of the absolute residual. Diminishing wideness of this data input means that, in an
approximately linear way, the value of the estimated performance diminishes.
Surface plots are reported in Figure 7.58.
S
R W
Figure 7.58. Surface plot for absolute aesidual (G.A.)

8
Distributed Production Planning Models:

an Integrated Approach
8.1 Introduction
In Chapter 5 a hierarchical model for planning production through a distributed
approach in a complex reconfigurable production network has been presented. That
approach is articulated in five hierarchical steps of planning, each one referring to a
specific time horizon and a specific planning objective.
It is also quite evident how each lower level receives input from the upper level
and therefore how a functional integration is needed between each couple of
models. The aim of this chapter is to show how this integration is actually possible
between, at least, two levels of the hierarchical structure presented in Chapter 5.
This will be obtained by carrying out an integrated simulation of two models.
Therefore, the final aim of this chapter is to show how by integrating distributed
production planning approaches is possible to obtain a complete tool for planning
production at different levels of a corporation. The IDEF0 diagram of Figure 8.1
shows the functional integration of the two production planning models at high and
medium levels. As the reader can notice the interdependence between the two
models is represented by the “balanced ownership”, that is, the output of activity
A1. Indeed, as the reader will remember, the aim of A1 is to balance the
ownership, initially located at the Top level, to the groups. As described in Chapter
6 this activity is obtained by bargaining ownership between overloaded and
underloaded groups. Once the balanced ownership is defined, this becomes the
input of the planning at medium level, whose objective is to allocate plants to
groups as described in Chapter 7.
Global Ownership
Re-modulated
Groups Ownership
Ownership
3 Months groups High PP Level
demand forecast
A1
Tools to re-modulate
ownership
Medium PP Level
Plants/Groups
Plants A2 allocation matrix
characteristics
Tools to allocate plants

to groups
Figure 8.1. The integration of the high PP and medium PP levels
As already mentioned, both the high level and medium level focus on the same
time horizon. In this chapter a simulation for the two distributed models, the game-
theoretical and the negotiational one, and for the benchmark, the centralised one, is
carried out. The simulation has been carried out through the agent-based simulation
environment described in the previous chapters. The case study consists of six
groups and 16 plants whose characteristics have been reported in Appendix C.
8.2 The Simulation Case Study

The test environment consists of 16 different combination of four input parameters,
i.e.: groups’ workloads ( WLg ) and ownerships ( C g ), plants’ production capacities
( Qm ) and the risk attitude ( Rm ). Specifically, each input parameter is considered
randomly drawn from a normal distribution N(µ;ı) whose parameters are reported
in Table 7.1. Again, like in the other chapters, small, S, and wide, W, refer to the
standard deviation size of the normal distribution.
Table 8.1. Distributions of input parameters
Small (S) Wide (W)
Workloads WL g N(27; 5.4) N(27; 21.6 )
Ownership C g N(27;1.35) N(27; 16.2)
Capacity Qm N(10; 0.5) N(10; 6)
Risk attitude Rm N(0.5; 0.025) N(0.5; 0.5 )

Distributed Production Planning Models: an Integrated Approach 173
Simulation experiments have been conducted with an infinite number of credits.

The aim of each replication is to evaluate some performance measures such as: the
efficiency (V), the distance (G) and the absolute residual ( Z ), respectively defined
in the previous chapter in Section 7.6.2. Each performance measure has been
evaluated with a confidence interval equal to 90%.
8.3 Results
Appendix C reports the fixed coefficients used for the simulation and the results
grouped for the 16 different combinations of the input parameters. To better
analyse this results, also in this case, the two-way Analysis of Variance (ANOVA)
and the Design of Experiment (DoE) have been used.
8.3.1 Efficiency Performance Analysis: Two-way ANOVA
Two factors have been used, i.e. the “kind of model”, whose levels are C.A.
(centralised model), N.A. (negotiation model), G.A. (game-theoretical model) and
the “input parameter” whose levels refer the wideness of the variance range, i.e.
wide (W) and small (S). The “input parameters” are those reported in Table 8.1.
8.3.1.1 Workload WLg

Table 8.2 reports the value of the parameter efficiency, V, obtained when the levels
of the input parameters varies, in the rows, and when the other input parameters
( Cg , Qm , Rm ) are combined as in the columns. The reader should refer to Table
8.3 for the association of the combinations to the numbers in the rows of Table 8.2.
1 2 3 4 5 6 7 8
W 7.928 8.297 7.903 8.332 7.980 8.534 7.454 7.577
C.A.
S 7.984 8.369 8.026 8.470 7.901 8.546 7.474 7.317
W 8.880 8.940 8.974 8.943 9.434 9.418 9.200 9.200
G.A.
S 8.717 8.865 8.967 8.859 9.428 9.298 9.200 9.200
W 7.956 8.460 7.928 8.296 7.928 8.437 7.476 7.546
N.A.
S 7.998 8.248 7.869 8.427 7.950 8.536 7.430 7.318
Table 8.3. Combination of C g , Qm , Rm
1 2 3 4 5 6 7 8
Cg W W W W S S S S
Qm W W S S W W S S
Rm W S W S W S W S
Table 8.4 reports the mean values of the numbers reported in Table 8.2, while the
last column and row of Table 8.2 report the average of the means.
W S Mean
C.A. 8.00 8.01 8.01
G.A. 9.12 9.07 9.10
N.A. 8.00 7.97 7.99
Mean 8.38 8.35 8.36
Table 8.4 shows that efficiency in the negotiation model is quite similar to the
centralised one, while the game-theoretical one seems to perform worse than the
other two models. This conclusion will be deeply investigated with the two-way
ANOVA. Specifically, there are three sets of hypotheses to be considered. As in
the previous chapters, we consider the null H 0 and the alternative hypothesis
H1 for each set:
for the population means of the factor “kind of model”
° H 0 : PC PN PG
®H : ; (8.1)
°̄ 1 not all Pi^C , N ,G` are equal
for the population means of the factor “workload”:
° H 0 : P W P B PS
®H : ; (8.2)
°̄ 1 not all P j^W , B ,S` are equal
for the interdependencies:
H0 : there is no interaction
® ; (8.3)
¯ H1 : there is interaction
Table 8.5 reports the ANOVA results. Only the main effect relative to the first
factor is significant: therefore, only the first null hypothesis 8.1 is rejected. This
actually means that the model influences the performance parameter efficiency,
while the workload does not. This also means that the difference between the
performance of the C.A. and N.A. models and the G.A. one is quite significant, i.e.
the game theoretical model performs worse than the other two, while the
negotiation model performs similarly to the centralised one. Also in this case the T-
test has been carried out; Table 8.6 reports the T-values, and Table 8.7 the
differences.
First factor 12.873 2 6.437 49.352* F2.42 3.252
Second factor 0.008 1 0.008 0.062 F1.42 4.102
Within 5.478 42 0.130 – –

Total 18.368 47 –
Table 8.6. T-values
Row factor Q0.05 ,3.42 3.44 T 0.254
Column factor Q0.05 , 2.42 2.86 T 0.258
Interaction Q0.05 , 6.42 4.23 T 0.540

G.A. 1.089* 0 0
N.A. 0.018 1.107* 0
W-C.A. S-C.A. W-G.A. S-G.A. W-N.A.

S–C.A. 0.010 – – – –
W–G.A. 1.123* 1.113* – – –
S–G.A. 1.066* 1.056* 0.057 – –
W–N.A. 0.003 0.008 1.120* 1.064* –
S–N.A. 0.029 0.039 1.152* 1.095* 0.031
From the analysis of the T-test it is possible to conclude that the estimated
differences among the models are relevant only between C.A. and G.A. and
between G.A. and N.A., but not among the centralised approach and the
negotiational one. Finally, Table 8.8 reports the interaction analysis. It confirms
that the models, combined with the WL input parameter, influence less the
estimated performance.
8.3.1.2 Capacity Ownership Cg

Similarly to the workload analysis, Table 8.9 reports the efficiency performance, V,
results obtained when the input parameter levels vary as shown in the table rows
and when the other parameters ( WLg , Qm , R ) vary their combination according to
m
Table 8.3 where WLg takes the place of Cg. Table 8.10 reports the mean values of
the numbers in Table 8.9.
1 2 3 4 5 6 7 8
W 7.928 8.297 7.903 8.332 7.984 8.369 8.026 8.470
C.A.
S 7.980 8.534 7.454 7.577 7.901 8.546 7.474 7.317
W 8.880 8.940 8.974 8.943 8.717 8.865 8.967 8.859
G.A.
S 9.434 9.418 9.200 9.200 9.428 9.298 9.200 9.200
W 7.956 8.460 7.928 8.296 7.998 8.248 7.869 8.427
N.A.
S 7.928 8.437 7.476 7.546 7.950 8.536 7.430 7.318
W S Mean
C.A. 8.16 7.85 8.01
G.A. 8.89 9.30 9.10
N.A. 8.15 7.83 7.99
Mean 8.40 8.32 8.36
As the reader can notice also in this case the centralised and the negotiational
model perform similarly and better than the game-theoretical one. This conclusion
will be deeply investigated with the ANOVA, whose results are reported in Table
8.11. Also in this case the first null hypothesis can be rejected, leading to
considerations very similar to those of the previous section, i.e. the model
typologies influence the efficiency. In this case, however, also the third null
hypothesis is rejected meaning that some interactions between the two factors are
present. Such interactions can be evaluated observing data from Table 8.12 and
Figures 8.2 and 8.3.
First factor 12.873 2 6.437 67.017* F2.42 3.252
Second factor 0.071 1 0.071 0.744 F1.42 4.102
Within 4.034 42 0.096 – –

Total 18.368 47 –
Table 8.12. Estimated values for V
W S Mean
C.A. 8.04 7.97 8.01
G.A. 9.13 9.06 9.10
N.A. 8.03 7.95 7.99
Mean 8.40 8.32 8.36
10,00
9,50
9,00
Centralised
8,50 Game Theory
Negotiation
8,00
7,50
7,00
W
1 2S
10,00
9,50
9,00
Centralised
8,50 Game Theory
Negotiation
8,00
7,50
7,00
W
1 2S
Also in this case a T-test analysis has been carried out and results have been
reported in Tables 8.13–8.14. The T-test analysis confirms the ANOVA results.
Finally, Table 8.15 reports the groups’ interaction analysis whose results support
the previous conclusions.
Row factor Q0.05 ,3.42 3.44 T 0.218
Column factor Q0.05 , 2.42 2.86 T 0.222
Interaction Q0.05 , 6.42 4.23 T 0.463

G.A. 1.089* 0 0
N.A. 0.018 1.108* 0
W–C.A. S–C.A. W–G.A. S–G.A. W–N.A.

S-C.A. 0.316 – – – –
W-G.A. 0.729* 1.045* – – –
S-G.A. 1.134* 1.449* 0.404 – –
W-N.A. 0.016 0.300 0.746* 1.150* –
S-N.A. 0.336 0.020 1.065* 1.470* 0.320
8.3.1.3 Plant Production Capacity (Qm)

In this section the influence of the input factor Qm is investigated. Of course, in this
case the other input combinations reported in Table 8.16 refer to the input
parameters WLg , C g , Rm ; combinations are obtained as indicated in Table 8.3.
1 2 3 4 5 6 7 8
W 7.928 8.297 7.980 8.534 7.984 8.369 7.901 8.546
C.A.
S 7.903 8.332 7.454 7.577 8.026 8.470 7.474 7.317
W 8.880 8.940 9.434 9.418 8.717 8.865 9.428 9.298
G.A.
S 8.974 8.943 9.200 9.200 8.967 8.859 9.200 9.200
W 7.956 8.460 7.928 8.437 7.998 8.248 7.950 8.536
N.A.
S 7.928 8.296 7.476 7.546 7.869 8.427 7.430 7.318
Also in this case the two-way ANOVA is carried out, where the null hypotheses
are the same of the previous section. Data for the ANOVA analysis are reported in
Tables 8.17–8.18. The two first null hypotheses are rejected; this means that both
the input factors, model typology and Qm, impact the efficiency means. Again it is
confirmed that the negotiation and centralised approaches perform quite similarly
and somewhat better than the game-theoretical.
W S Mean
C.A. 8.19 7.82 8.01
G.A. 9.12 9.07 9.10
N.A. 8.19 7.79 7.99
Mean 8.50 8.22 8.36
First factor 12.873 2 6.437 63.204* F2.42 3.252
Second factor 0.920 1 0.920 9.029* F1.42 4.102
Within 4.277 42 0.102 – –

Total 18.368 47 –
Row factor Q0.05 ,3.42 3.44 T 0.224
Column factor Q0.05 , 2.42 2.86 T 0.228
Interaction Q0.05 , 6.42 4.23 T 0.477
Column factor W S
S 0.277* 0

G.A. 1.089* 0 0
N.A. 0.018 1.108* 0

S–C.A. 0.373 – – – –
W–G.A. 0.930* 1.303* – – –
S–G.A. 0.876* 1.249* 0.055 – –
W–N.A. 0.003 0.370 0.933* 0.879* –
S–N.A. 0.406 0.033 1.336* 1.282* 0.403
T-test analysis, whose results are reported in Tables 8.19–8.21, confirms these
conclusions. Finally, interactions analysis reported in Table 8.22 confirms that
there is not any significant interaction between the two factors.
8.3.1.4 Risk Attitude Rm

Finally, the influence of the Rm input parameter has been analysed. Of course, in
this case, the other input parameters, whose combinations are reported in the
columns of Table 8.23, are WLg , Cg , Qm . Also in this case a two-way ANOVA
analysis has been carried out whose data and results are reported in Tables 8.23–
8.25. In this case the first and third null hypotheses can be rejected. This means
again that model typologies influence efficiency; in particular C.A. and N.A.
models perform quite similarly and better than the G.A. one. Furthermore, an
interaction between the two input factors is present. T-test analysis (see Tables
8.26–8.27) and interaction analysis, Table 8.28, confirm this conclusion.
Referring to efficiency it is possible to conclude that model typology impacts
efficiency performance in the integrated uses of high and medium levels models.
The negotiation distributed approach (N.A.) performs quite similarly to the
centralised one (C.A.), while the game-theoretical shows performances a bit worse
than the other two approaches.
1 2 3 4 5 6 7 8
W 7.928 7.903 7.980 7.454 7.984 8.026 7.901 7.474
C.A.
S 8.297 8.332 8.534 7.577 8.369 8.470 8.546 7.317
W 8.880 8.974 9.434 9.200 8.717 8.967 9.428 9.200
G.A.
S 8.940 8.943 9.418 9.200 8.865 8.859 9.298 9.200
W 7.956 7.928 7.928 7.476 7.998 7.869 7.950 7.430
N.A.
S 8.460 8.296 8.437 7.546 8.248 8.427 8.536 7.318
W S Mean
C.A. 7.83 8.18 8.01
G.A. 9.10 9.09 9.10
N.A. 7.82 8.16 7.99
Mean 8.25 8.48 8.36
First factor 12.873 2 6.437 59.540* F2.42 3.252
Second factor 0.619 1 0.619 5.721* F1.42 4.102
Within 4.541 42 0.108 – –

Total 18.368 47 –
Row factor Q0.05 ,3.42 3.44 T 0.235
Column factor Q0.05 , 2.42 2.86 T 0.235
Interaction Q0.05 , 6.42 4.23 T 0.492

G.A. 1.089* 0 0
N.A. 0.018 1.108* 0

S–C.A. 0.349 – – – –
W–G.A. 1.269* 0.920* – – –
S–G.A. 1.259* 0.910* 0.010 – –
W–N.A. 0.014 0.364 1.283* 1.274* –
S–N.A. 0.327 0.022 0.942* 0.932* 0.342
8.3.2 Efficiency Performance Analysis: DoE
DoE is used to investigate in which way the input variables influence the estimated
efficiency.

Figure 8.4 reports the main effects, i.e. how input parameter levels influence the
efficiency performance, V. As the reader can notice, the workload variance has no
influence on the performance, while capacity ownership and plant production
capacity variances negatively influence the efficiency; on the other hand, risk
attitude variance positively influences the efficiency. Figure 8.5 shows the input
parameter interactions. Only C g vs. Q shows a discordant interaction.
WL Cg
8,2
8,1
8,0
Mean of Efficiency
7,9
7,8
S W S W
Q R
8,2
8,1
8,0
7,9
7,8
S W S W
S W S W S W
8,5
WL
S
8,0
WL W
7,5
8,5
Cg
S
8,0
Cg W
7,5
8,5
Q
S
8,0
Q W
7,5

To better understand which factor is actually relevant, a normal probability plot

and a Pareto chart have been considered. The first one, see Figure 8.6, is a
graphical method, based on a subjective visual examination of the data, for
acknowledging whether estimated data match the hypothesised (normal)
distribution. In the normal probability plot of the effects, points not falling near the
line usually signal important effects. To understand not only the importance but
also the magnitude of such effects, a Pareto analysis has also been considered
(Figure 8.7). The related chart displays the absolute value of the effects by setting a
reference line: any effect extending its influence beyond this reference line is
potentially important. Both the analyses confirm that significant and important
effects are Q, Cg, R and the joint effect of Cg and Q.
99
Effect Type
Not Significant
95 C Significant
90 B Factor Name
80 A WL
B Cg
70
C Q
Percent
60
50 D R
40
30
20
10 D
5 BC
1
-0,5 -0,4 -0,3 -0,2 -0,1 0,0 0,1 0,2 0,3 0,4
Effect
Figure 8.6. Normal probability plot of the effects (C.A.)

0,1785
BC Factor Name
C A WL
D B Cg
C Q
B
D R
BCD
CD
AB
Term
BD
ACD
ABCD
ABC
ABD
AD
A
AC
0,0 0,1 0,2 0,3 0,4

Effect
Figure 8.7. Pareto chart of the effects (C.A.)
The above results are also confirmed by the surface curves of Figure 8.8.
Concluding, for the centralised model, the efficiency, V, is negatively influenced
by the capacity ownership, Cg, and by the plant capacity, Q, positively by the risk
attitude, R, and by the interaction effects of the Cg and Q variances.

By adopting the same methodology, the impact of the input factor variance on the
efficiency performance has been evaluated within the negotiation model.
By observing Figure 8.9, which reports the main effects, Figure 8.10, which
reports the interactions, the normal probability plot (Figure 8.11), the Pareto chart
(Figure 8.12) and the interaction response curves of Figure 8.13 it is possible to
conclude that the efficiency performance in the negotiation model is negatively
influenced by the capacity ownership, Cg, and by the plant capacity variances (the
most important parameter according to Figure 8.12), Q, while it is positively
influenced by the risk attitude variance, R. Also, significant and important
interactions are those among Q, Cg and R and between Q and Cg.
S
R W
Figure 8.8. Surface plot for efficiency (C.A.)

WL Cg
8,2
8,1
8,0
7,9
Mean of Efficiency
7,8
L H L H
Q R
8,2
8,1
8,0
7,9
7,8
L H L H
L H L H L H
8,5
WL
L
8,0
WL H
7,5
8,5
Cg
L
8,0
Cg H
7,5
8,5
Q
L
8,0
Q H
7,5

99
Effect Type
Not Significant
95 C Significant
90 B Factor Name
80 BCD A WL
B Cg
70
C Q
Percent
60
50 D R
40
30
20
10 D
5 BC
1
-0,4 -0,3 -0,2 -0,1 0,0 0,1 0,2 0,3 0,4 0,5
Effect
Figure 8.11. Normal probability plot of the effects (N.A.)
0,1203
C Factor Name
BC A WL
D B Cg
B C Q
BCD D R
CD
ABCD
Term
ABC
BD
A
ACD
AD
AC
AB
ABD
0,0 0,1 0,2 0,3 0,4

Effect
Figure 8.12. Pareto chart of the effects (N.A.)

S
R W
8.3.2.3 Game Theory

As the reader can notice by observing Figures 8.14–8.18 a different situation
emerges for the game-theoretical model; indeed, only the capacity ownership
variance has a significant and also an important (according to the Pareto chart)
effect on efficiency performance, V. Furthermore, also the interaction effect of Cg
and Q has some significant and important effect on the efficiency.
WL Cg
9,3
9,2
9,1
9,0
Mean of Efficiency
8,9
S W S W
Q R
9,3
9,2
9,1
9,0
8,9
S W S W
S W S W S W
9,4
WL
9,2 S
WL W
9,0
9,4
Cg
9,2 S
Cg W
9,0
9,4
Q
9,2 S
Q W
9,0

99
Effect Type
Not Significant
95 Significant
90 Factor Name
80 A WL
B Cg
70
C Q
Percent
60
50 D R
40
30
20
10 BC
5 B
1
-0,4 -0,3 -0,2 -0,1 0,0 0,1
Effect
Figure 8.16. Normal probability plot of the effects (G.A.)
0,0980
B Factor Name
BC A WL
BCD B Cg
A
C Q
D R
C
ABCD
AC
Term
BD
CD
AB
ABD
AD
D
ACD
ABC
0,0 0,1 0,2 0,3 0,4

Effect
Figure 8.17. Pareto chart of the effects (G.A.)
This allows us to draw an interesting conclusion about the use of the proposed
models in an integrated way at high and medium level planning. Although the
game-theoretical approach seems to perform a bit worse of the centralised and the
negotiation one, it is the less sensible to the input factors uncertainty.
Therefore, when the input factors are known in a quite certain way, the
negotiation approach should be used to plan in a distributed way production at high
and medium level; on the other hand, when input factors uncertainty is high, the
game-theoretical approach should be considered.
S
R W
Figure 8.18. Surface plot for efficiency (G.A.)
8.3.3 Distance Performance Analysis: Two-way ANOVA
8.3.3.1 Workload ( WLg )

Table 8.29 reports the values of the distance performance when input factors “kind
of model” and “workload variance” vary according to their levels (table rows) and
when the other input parameters, namely Cg , Qm , Rm , are combined according to
Table 8.3.
1 2 3 4 5 6 7 8
W 5.957 6.294 5.937 6.328 6.150 6.540 5.855 6.282
C.A.
S 6.004 6.387 6.075 6.482 6.107 6.498 5.790 5.936
W 6.707 6.677 6.792 6.729 6.897 6.690 6.600 6.600
G.A.
S 6.522 6.711 6.791 6.747 6.818 6.590 6.600 6.600
W 6.033 6.357 5.984 6.353 6.090 6.482 5.790 6.257
N.A.
S 5.957 6.242 5.954 6.492 6.121 6.632 5.831 5.936
Table 8.30 reports the average value of numbers in Table 8.29, while Table
8.31 reports the results of the ANOVA. Tables 8.32 the T-test and Table 8.33 the
interaction analysis. From the analysis of the results it is possible to conclude that
only the null hypothesis 1 can be rejected: i.e. the factor “kind of model” has a
significant impact on the distance performance. Specifically. again the model C.A.
and N.A. seem to perform quite similarly. while the model G.A. performs a bit
worse than the others.
W S Mean
C.A. 6.17 6.16 6.16
G.A. 6.71 6.67 6.69
N.A. 6.17 6.15 6.16
Mean 6.35 6.33 6.34
Table 8.31. Two-way ANOVA table for G
First factor 3.013 2 1.507 31.818* F2.42 3.252
Second factor 0.006 1 0.006 0.137 F1.42 4.102
Within 1.989 42 0.047 – –

Total 5.010 47 –
8.3.3.2 Capacity Ownership Cg

Tables 8.34–8.41 show the usual data for the capacity ownership factor. Again
only the null hypothesis 8.1 can be rejected, that is, the factor “kind of model”
influences the distance performance. Also in this case models C.A. and N.A.
perform quite similarly, while the model G.A. performs a bit worse than the others.
Row factor Q0.05 ,3.42 3.44 T 0.153
Column factor Q0.05 , 2.42 2.86 T 0.1556
Interaction Q0.05 , 6.42 4.23 T 0.325

S-C.A. 0.008 - - - -
W-G.A. 0.544* 0.552* - - -
S-G.A. 0.504* 0.512* 0.039 - -
W-N.A. 0.000 0.008 0.543* 0.504* -
S-N.A. 0.022 0.,014 0.566* 0.527* 0.023
Table 8.34. Replications values of G
1 2 3 4 5 6 7 8
W 5.957 6.294 5.937 6.328 6.004 6.387 6.075 6.482
C.A.
S 6.150 6.540 5.855 6.282 6.107 6.498 5.790 5.936
W 6.707 6.677 6.792 6.729 6.522 6.711 6.791 6.747
G.A.
S 6.897 6.690 6.600 6.600 6.590 6.600 6.600 6.577
W 6.033 6.357 5.984 6.353 5.957 6.242 5.954 6.492
N.A.
S 6.090 6.482 5.790 6.257 6.121 6.632 5.831 5.936
W S Mean
C.A. 6.18 6.14 6.16
G.A. 6.71 6.64 6.68
N.A. 6.17 6.14 6.16
Mean 6.35 6.31 6.33
First factor 2.845 2 1.423 30.408* F2.42 3.252
Second factor 0.023 1 0.023 0.498 F1.42 4.102
Within 1.965 42 0.047 – –

Total 4.836 47 –
Table 8.37. Estimated values for G
W S Mean
C.A. 6.18 6.14 6.16
G.A. 6.71 6.64 6.68
N.A. 6.17 6.14 6.16
Mean 6.35 6.31 6.33
Row factor Q0.05 ,3.42 3.44 T 0.152
Column factor Q0.05 , 2.42 2.86 T 0.155
Interaction Q0.05 , 6.42 4.23 T 0.323


G.A. 0.521* 0 0
N.A. 0.007 0.522* 0
Column factor W S
S 0.004 0

S–C.A. 0.038 – – – –
W–G.A. 0.527* 0.565* – – –
S–G.A. 5.462* 5.499* 4.935* – –
W–N.A. 0.011 0.027 0.538* 5.473* –
S-N.A. 0.040 0.002 0.567* 5.502* 0.029
8.3.3.3 Plant Production Capacity ( Qm )

Tables 8.42–8.47 show the usual data for the plant production capacity factor.
Again only the null hypothesis 8.1 can be rejected, that is, the factor “kind of
model” influences the distance performance. Also in this case models C.A. and
N.A. perform quite similarly, while the model G.A. performs a bit worse than the
others.
1 2 3 4 5 6 7 8
W 5,957 6,294 6,150 6,540 6,004 6,387 6,107 6,498
C.A.
S 5,937 6,328 5,855 6,282 6,075 6,482 5,790 5,936
W 6,707 6,677 6,897 6,690 6,522 6,711 6,818 6,590
G.A.
S 6,792 6,729 6,600 6,600 6,791 6,747 6,600 6,600
W 6,033 6,357 6,090 6,482 5,957 6,242 6,121 6,632
N.A.
S 5,984 6,353 5,790 6,257 5,954 6,492 5,831 5,936
W S Mean
C.A. 6,24 6,09 6,16
G.A. 6,70 6,68 6,69
N.A. 6,24 6,07 6,16
Mean 6,39 6,28 6,34
First factor 3.013 2 1.507 35.359* F2.42 3.252
Second factor 0.154 1 0.154 3.619 F1.42 4.102
Within 1.789 42 0.043 – –

Total 5.010 47 –
Row factor Q0.05 ,3.42 3.44 T 0.224
Column factor Q0.05 , 2.42 2.86 T 0.228
Interaction Q0.05 , 6.42 4.23 T 0.477

G.A. 0.528* 0 0
N.A. 0.007 0.535* 0
8.3.3.4 Risk Attitude ( Rm )

Finally, tables 8.48–8.55 show data for the risk attitude factor. This time all the
three null hypotheses 8.1 can be rejected. This means: a) the factor “kind of model”
influences the distance performance; b) the factor “risk attitude variance”
influences the distance performance; c) there is an interaction effect of the input
factors on performance values.

S-C.A. 0.157 – – – –
W-G.A. 0.459* 0.616* – – –
S-G.A. 0.440* 0.597* 0.019 – –
W-N.A. 0.003 0.154 0.462* 0.443* –
S-N.A. 0.167 0.011 0.627* 0.608* 0.164
1 2 3 4 5 6 7 8
W 7.928 7.903 7.980 7.454 7.984 8.026 7.901 7.474
C.A.
S 8.297 8.332 8.534 7.577 8.369 8.470 8.546 7.317
W 8.880 8.974 9.434 9.200 8.717 8.967 9.428 9.200
G.A.
S 8.940 8.943 9.418 9.200 8.865 8.859 9.298 9.200
W 7.956 7.928 7.928 7.476 7.998 7.869 7.950 7.430
N.A.
S 8.460 8.296 8.437 7.546 8.248 8.427 8.536 7.318
W S Mean
C.A. 5.98 6.34 6.16
G.A. 6.72 6.67 6.69
N.A. 5.97 6.34 6.16
Mean 6.22 6.45 6.34
First factor 3.013 2 1.507 69.240* F2.42 3.252
Second factor 0.625 1 0.625 28.732* F1.42 4.102
Within 0.914 42 0.022 – –

Total 5.010 47 –
Table 8.51. Estimated values for G
W S Mean
C.A. 6.05 6.28 6.16
G.A. 6.58 6.81 6.69
N.A. 6.04 6.27 6.16
Mean 6.22 6.45 6.34
Row factor Q0.05 ,3.42 3.44 T 0.104
Column factor Q0.05 , 2.42 2.86 T 0.105
Interaction Q0.05 , 6.42 4.23 T 0.221

G.A. 0.528* 0 0
N.A. 0.007 0.535* 0
Column factor W S
S 0.228249* 0
6,80
6,70
6,60
6,50
6,40 Centralised
6,30 Game Theory
6,20 Negotiation
6,10
6,00
5,90
5,80
W
1 2S

6,80
6,70
6,60
6,50
6,40 Centralised
6,30 Game Theory
6,20 Negotiation
6,10
6,00
5,90
5,80
1W S
2

S-C.A. 0.359* – – – –
W-G.A. 0.731* 0.373* – – –
S-G.A. 0.683* 0.325* 0.048 – –
W-N.A. 0.014 0.373* 0.746* 0.698* –
S-N.A. 0.359* 0.001 0.372* 0.324* 0.374*
8.3.4 Distance Performance Analysis: DoE
Again, DoE is used to investigate in which way the input variables influence the
estimated efficiency.

distance performance. The workload and capacity ownership variances have no
influence on the performance; on the other hand, plant production capacity
variance negatively influences the efficiency and the risk attitude variance
positively influences the distance parameter. From the interaction side, Figure 8.22,
it seems that the interaction of Q and Cg might have influence on the distance
performance. This is confirmed by the normal probability plot, Figure 8.23, and by
the Pareto chart, Figure 8.24, which shows how the risk attitude variance is the
most important factor in this case. Finally, Figure 8.25 reports the response curves
analysis.
WL Cg
6,3
6,2
6,1
Mean of Distance
6,0
S W S W
Q R
6,3
6,2
6,1
6,0
S W S W
S W S W S W
6,4 WL
S
6,2
WL W
6,0
6,4 Cg
S
6,2
Cg W
6,0
6,4 Q
S
6,2
Q W
6,0

99
Effect Type
Not Significant
95 C Significant
90 Factor Name
80 A WL
B Cg
70
C Q
Percent
60
D R
50
40
30
20
10 BC
5 D
1
-0,4 -0,3 -0,2 -0,1 0,0 0,1 0,2
Effect
0,1298
D Factor Name
BC A WL
C B Cg
C Q
AB
D R
ABC
ABD
ACD
Term
B
BCD
ABCD
AD
AC
BD
CD
A
0,0 0,1 0,2 0,3 0,4

Effect

S
R W
Figure 8.25. Surface plot for distance (C.A.)
8.3.4.2 Negotiation
distance performance when using the N.A. model. Again, the workload and
capacity ownership variances have no influence on the performance; on the other
hand, plant production capacity variance negatively influences the efficiency and
the risk attitude variance positively influences the distance parameter. From the
interaction side, Figure 8.27, it seems that the interaction of Q and Cg might have
influence on the distance performance. This is confirmed by the normal probability
plot, Figure 8.28, and by the Pareto chart, Figure 8.29, that puts on evidence how
the risk attitude variance is the most important factor in this case. Finally, Figure
8.30 reports the response curves analysis.
WL Cg
6,4
6,3
6,2
6,1
Mean of Distance
6,0
S W S W
Q R
6,4
6,3
6,2
6,1
6,0
S W S W
S W S W S W
6,4 WL
6,2
S
WL
W
6,0
6,4 Cg
6,2
S
Cg
W
6,0
6,4
Q
6,2
S
Q
W
6,0

99
Effect Type
Not Significant
95 C Significant
90 Factor Name
80 A WL
B Cg
70
C Q
Percent
60
50 D R
40
30
20
10 BC
5 D
1
-0,4 -0,3 -0,2 -0,1 0,0 0,1 0,2
Effect
0,0996
D Factor Name
BC A WL
C B Cg
C Q
ABC
D R
ABCD
BCD
ABD
Term
ACD
B
A
AC
AD
BD
CD
AB
0,0 0,1 0,2 0,3 0,4

Effect

S
R W
Figure 8.30. Surface plot for distance (N.A.)
8.3.4.3 Game Theory

Figure 8.31 reports the main effects when using the G.A. model. This time, all the
input parameter variances have a significant influence on the performance. Also
from the interaction side, Figure 8.32, it seems that the interaction of WL and Q, Q
and Cg, and of Cg and R might have influence on the distance performance.
However, the normal probability plot, Figure 8.33, and the Pareto chart show us
that only the interaction among Q and Cg is significant and important.
WL Cg
6,71
6,70
6,69
6,68
Mean of Distance
6,67
S W S W
Q R
6,71
6,70
6,69
6,68
6,67
S W S W
S W S W S W
WL
6,72
S
WL W
6,66
6,60
Cg
6,72
S
Cg W
6,66
6,60
Q
6,72
S
Q W
6,66
6,60

99
Effect Type
Not Significant
95 Significant
90 Factor Name
80 A WL
B Cg
70
C Q
Percent
60
50 D R
40
30
20
10
5 BC
1
-0,15 -0,10 -0,05 0,00 0,05 0,10
Effect
0,1157
BC Factor Name
BCD A WL
BD B Cg
D
C Q
D R
AC
A
B
Term
ABD
ABCD
AD
ACD
CD
C
AB
ABC
0,00 0,02 0,04 0,06 0,08 0,10 0,12 0,14

Effect

S
S S
R W R W
Figure 8.35. Surface plot for distance (G.A.)
The analysis allows us to draw a conclusion that is quite similar to that drawn for
the efficiency performance. Indeed, also for the distance performance although the
game-theoretical approach, G.A., seems to perform a bit worse of the centralised,
C.A., and the negotiation, N.A., one, it is the last sensitive to the input factor
uncertainty. Therefore, when the input factors are known in a quite certain way, the
negotiation approach should be used to plan in a distributed way production at high
and medium level; on the other hand, when input factors uncertainty is high, the
game-theoretical approach should be used.
8.3.5 Absolute Residual Performance Analysis: Two-ways ANOVA
8.3.5.1 Workload WLg

Table 8.57 reports the values of the absolute residual performance when input
factors “kind of model” and “workload variance” vary according to their levels
(Table rows) and when the other input parameters, namely Cg , Qm , Rm , are
combined according to Table 8.3. Table 8.57 reports the average value of numbers
in Table 8.56, while Table 8.58 reports the results of the ANOVA. As the reader
can notice, none of the null hypotheses can be rejected, therefore the T-test analysis
is pointless. None of the input factors, nor their interaction influences the
performance values.
1 2 3 4 5 6 7 8
W 51.761 57.385 42.226 43.216 36.641 36.289 22.450 23.298
C.A.
S 53.674 52.864 46.744 49.030 36.917 37.535 22.347 23.184
W 58.791 59.694 53.668 47.153 38.927 40.474 24.000 25.500
G.A.
S 58.485 58.392 49.213 50.741 40.141 39.182 23.667 24.750
W 53.296 55.043 46.694 47.156 36.947 37.937 22.336 24.037
N.A.
S 50.415 54.864 44.223 45.007 36.096 39.126 23.353 23.168
W S Mean
C.A. 39.16 40.29 39.72
G.A. 43.53 43.07 43.30
N.A. 40.43 39.53 39.98
Mean 41.04 40.96 41.00
Table 8.58. Two–way ANOVA table for Z
SS df MS F–test F–crit D 5%
First factor 127.251 2 63.626 0.391 F2.42 3.252
Second factor 0.068 1 0.068 0.000 F1.42 4.102
Within 6827.921 42 162.570 – –

Total 6964.330 47 –
8.3.5.2 Capacity Ownership ( Cg )

Tables 8.59–8.64 show the usual data for the capacity ownership factor. This time,
only the null hypothesis 8.2 can be rejected, that is, the factor “capacity ownership
variance” influences the absolute residual performance. Also in this case, however,
it is possible to conclude that the factor “kind of model” does not influence the
performance under investigation. Of course, T-test analysis refers in this case to the
second input factor.
1 2 3 4 5 6 7 8
W 51.761 57.385 42.226 43.216 53.674 52.864 46.744 49.030
C.A.
S 36.641 36.289 22.450 23.298 36.917 37.535 22.347 23.184
W 58.791 59.694 53.668 47.153 58.485 58.392 49.213 50.741
G.A.
S 38.927 40.474 24.000 25.500 40.141 39.182 23.667 24.750
W 53.296 55.043 46.694 47.156 50.415 54.864 44.223 45.007
N.A.
S 36.947 37.937 22.336 24.037 36.096 39.126 23.353 23.168
W S Mean
C.A. 49.61 29.83 39.72
G.A. 54.52 32.08 43.30
N.A. 49.59 30.38 39.98
Mean 51.24 30.76 41.00
First factor 127.251 2 63.626 1.500 F2.42 3.252
Second factor 5031.374 1 5031.374 118.585* F1.42 4.102
Within 1781.995 42 42.428 – –

Total 6964.330 47 –
Row factor Q0.05 ,3.42 3.44 T 4.657
Column factor Q0.05 , 2.42 2.86 T 4.574
Interaction Q0.05 , 6.42 4.23 T 9.741

Column factor W S
S 20.476* 0

S-C.A. 19.780* – – – –
W-G.A. 4.904 24.684* – – –
S-G.A. 17.532* 2.247 22.437* – –
W-N.A. 0.025 19.755* 4.930 17.507* –
S-N.A. 19.237* 0.542 24.142* 1.705 19.212*
8.3.5.3 Plants Production Capacity ( Qm )

Tables 8.65–8.69 show the usual data for the plants production capacity factor.
Also in this case, only the null hypothesis 8.2 can be rejected. However, it is
possible to conclude that the factor “kind of model” does not influence the
performance under investigation. Of course, T-test analysis, also in this case, refers
to the second input factor.
Table 8.65. Replications values of Z
1 2 3 4 5 6 7 8
W 51.761 57.385 36.641 36.289 53.674 52.864 36.917 37.535
C.A.
S 42.226 43.216 22.450 23.298 46.744 49.030 22.347 23.184
W 58.791 59.694 38.927 40.474 58.485 58.392 40.141 39.182
G.A.
S 53.668 47.153 24.000 25.500 49.213 50.741 23.667 24.750
W 53.296 55.043 36.947 37.937 50.415 54.864 36.096 39.126
N.A.
S 46.694 47.156 22.336 24.037 44.223 45.007 23.353 23.168
W S Mean
C.A. 45.38 34.06 39.72
G.A. 49.26 37.34 43.30
N.A. 45.47 34.50 39.98
Mean 46.70 35.30 41.00
First factor 127.251 2 63.626 0.507 F2.42 3.252
Second factor 1560.857 1 1560.857 12.429* F1.42 4.102
Within 5274.354 42 125.580 – –

Total 6964.330 47 –
Row factor Q0.05 ,3.42 3.44 T 7.868
Column factor Q0.05 , 2.42 2.86 T 8.012
Interaction Q0.05 , 6.42 4.23 T 16.759
Column factor W S
S 11.405* 0
8.3.5.4 Risk Attitude ( Rm )

Finally, tables 8.70–8.72 show the usual data for the risk attitude factor. Also in
this case, like for the workload, none of the null hypotheses can be rejected. This
means, in particular that the performance Z is not influenced by the typology of
the model used. Of course, in this case T-test analysis is pointless.
1 2 3 4 5 6 7 8
W 51.761 42.226 36.641 22.450 53.674 46.744 36.917 22.347
C.A.
S 57.385 43.216 36.289 23.298 52.864 49.030 37.535 23.184
W 58.791 53.668 38.927 24.000 58.485 49.213 40.141 23.667
G.A.
S 59.694 47.153 40.474 25.500 58.392 50.741 39.182 24.750
W 53.296 46.694 36.947 22.336 50.415 44.223 36.096 23.353
N.A.
S 55.043 47.156 37.937 24.037 54.864 45.007 39.126 23.168
W S Mean
C.A. 39.10 40.35 39.72
G.A. 43.36 43.24 43.30
N.A. 39.17 40.79 39.98
Mean 40.54 41.46 41.00
First factor 127.251 2 63.626 0.392 F2.42 3.252
Second factor 10.094 1 10.094 0.062 F1.42 4.102
Within 6820.189 42 162.385 – –

Total 6964.330 47 –
8.3.6 Absolute Residual Performance Analysis: DoE

Z performance. As the reader can notice, the workload and the risk attitude
variances have no influence on the performance; on the other hand, capacity
ownership and plant production capacity variances positively influence the
performance. From the interaction side, Figure 8.37, it seems that no interaction
has influence on the performance. This is confirmed by the normal probability plot,
Figure 8.38, and by the Pareto chart, Figure 8.39. Finally, Figure 8.40 reports the
response curves analysis.
WL Cg
50
45
40
35
30
S W S W
Q R
50
45
40
35
30
S W S W
S W S W S W
50 WL
S
40
WL W
30
50 Cg
S
40
Cg W
30
50 Q
S
40
Q W
30

99
Effect Type
Not Significant
95 B Significant
90 C Factor Name
80 A WL
B Cg
70
C Q
Percent
60
50 D R
40
30
20
10
5
1
0 5 10 15 20
Effect
3,25
B Factor Name
C A WL
BC B Cg
ABC C Q
AC
D
A
Term
ABCD
ACD
AB
BD
ABD
AD
BCD
CD
0 5 10 15 20
Effect

S
R W
Figure 8.40. Surface plot for absolute residual (C.A.)

Z performance. Also in this case, the workload and the risk attitude variances have
no influence on the performance; on the other hand, capacity ownership and plant
production capacity variances negatively influence the performance. From the
interaction side, Figure 8.42, it seems that none interaction has influence on the
performance. This is confirmed by the normal probability plot, Figure 8.43, and by
the Pareto chart, Figure 8.44. Finally, Figure 8.45 reports the response curves
analysis.
WL Cg
50
45
40
35
30
S W S W
Q R
50
45
40
35
30
S W S W
L H L H L H
50 WL
L
40
WL H
30
50 Cg
L
40
Cg H
30
50 Q
L
40
Q H
30

99
Effect Type
Not Significant
95 B Significant
90 C Factor Name
80 A WL
B Cg
70
C Q
60
Percent
50
D R
40
30
20
10 D
5 BC
1
0 5 10 15 20
Effect
1,46
B Factor Name
C A WL
BC B Cg
C Q
D
D R
AB
CD
A
Term
ACD
AD
ABD
BCD
BD
AC
ABCD
ABC
0 5 10 15 20
Effect

S
R W
Figure 8.45. Surface plot for absolute residual (N.A.)
8.3.6.3 Game Theory

Finally, Figure 8.46 reports the main effects influence on the Z performance. Also
in this case, the workload and the risk attitude variances have no influence on the
performance; on the other hand, capacity ownership and plant production capacity
variances positively influence the performance. From the interaction side, Figure
8.47, it seems that no interaction has influence on the performance. This is not
confirmed by the normal probability plot, Figure 8.48, which indicates that there is
a significant interaction between Q and Cg, even if its importance is just beyond the
threshold, which as reported in the Pareto chart, Figure 8.49. Finally, Figure 8.50
reports the response curves analysis.
WL Cg
55
50
45
40
35
S W S W
Q R
55
50
45
40
35
S W S W
S W S W S W
WL
50
S
WL W
35
20
Cg
50
S
Cg W
35
20
Q
50
S
Q W
35
20

99
Effect Type
Not Significant
95 B Significant
90 C Factor Name
80 A WL
B Cg
70
C Q
60
Percent
50
D R
40
30
20
10
5 BC
1
0 5 10 15 20 25
Effect
1,91
B Factor Name
C A WL
BC B Cg
C Q
ACD
D R
ABD
BCD
BD
Term
ABCD
AD
CD
A
ABC
AB
D
AC
0 5 10 15 20 25
Effect

S
R W
Figure 8.50. Surface plot for absolute residual (G.A.)
As far as absolute residual performance, Z, is concerned we need to stress that

there is no statistical evidence that the model typology influences the performance.
The influence of the input parameter variance on the models is, on the other hand,
quite similar among the three models.
8.4 Conclusions
This chapter shows the integrated use of the planning models we have discussed in
the previous chapters for planning production at high and medium levels. The aim
of the chapter was to understand what kind of indications could be drawn from an
integrated use of the models.
The main conclusion is that the negotiation model, N.A., performs quite
similarly to the centralised one, C.A., as far as efficiency and distance are
concerned, while the game-theoretical model, G.A., performs a bit worse than the
others.
However, the G.A. model is less sensitive to input parameters uncertainty.

Therefore, when the uncertainty is low, the N.A. model should be used; conversaly,
if the uncertainty is high the G.A. model should be used. The absolute residual
performance measure is not influenced by the model typology.
9
Conclusions
9.1 Summary
This book presents a study that has been conducted about the opportunity to utilise
a set of decentralised models to face the difficulty emerging in the capacity
allocation problem which concerns the semiconductor industry or, more generally,
high-tech, high-volume Industries with a relevant numbers of groups (product
families) and plants geographically dispersed. We modelled the decision problem
using the viewpoints of the managers of the organisation. We first showed that
decentralisation requires the additional restriction of maintaining private
information, which creates unavoidable degradation of overall performance. Under
the private information assumption, we showed that coordination is achievable
between groups’ and plants’ agents using an information exchange scheme. We
proposed two coordination mechanisms using this framework and proved that these
approaches converge to the optimal global solution defined by the centralised
(benchmark) model. From a mathematical standpoint, we modelled the information
exchange as an auction between the competing decision-makers. This approach
reflects the actual opposite agents’ local utility functions and it tries to regulate the
local decisions to match the other side’s needs as closely as possible. The
negotiational and game-theoretical theories provided a flexible analytical
framework to develop this coordination device. Finally, we tested the proposed
mechanisms in an agent-based ad hoc developed discrete-event simulation
environment by using generated numerical data. As far as we know, this research is
one of the very few studies that utilises the application of the game-theory and
negotiational framework to study this kind of issue. It appears that future research
will needed to apply this approach to both optimisation and coordination problems
that can be solved by mathematical-simulative composition.
In what follows, we summarise the main scientific contributions of this book
and some directions for future work.
9.2 Major Scientific Contributions of This Book

The main contributions of this book are:
In Chapter 5 an innovative distributed and hierarchical approach for
planning capacity at different levels of time horizon has been introduced.
For each level bargaining objective, actors, information and roles have
been located. In synthesis the model presented in Chapter 5 represents a
general framework for distributed production planning of complex and
multi-plant manufacturing companies in high-tech, high-volume
industries;
A structured model of agents with bounded rationality that interact to
solve, in a competitive way, a capacity allocation problem. We presented
models framework for the high-level production planning (Chapter 6) and
for the medium-level production planning (Chapter 7). We defined the
utility functions introducing performance analysis and concepts that are
feasible to be adopted in practice and that can lead to better drive the
agents’ coordination and the control decisions compared to other
approaches from the scientific literature.
A formal game-theoretical and negotiational model for this goal has been
developed. We proposed incorporating the deliberation actions of agents
in a mathematical setting. In the first part of Chapters 6 and 7 we
presented definitions of strategies for agents with resources limitations,
and introduced the concept of equilibrium using a fictitious credits.
We also discussed new strategic behaviours which may arise among
agents. In particular, we introduced the Bargaining solution in an
allocation cooperative game that drastically increases the performance of
this approach.
Analysis of negotiation mechanisms has also been conducted. In Chapters
6 and 7 we also analysed negotiation protocols, using the Rubinstein
equilibrium as the solution concept.
A simulation environment has been utilised to simulate the behaviour of
the defined agents and to estimate the established performance measures.
In particular, the last part of Chapter 6 showed that, regarding the
unallocated capacity of the high level, no relevant differences can be
highlighted among the three different proposed approaches. Considering a
confidence interval equal to 5% the existing discrepancies are not
significantly relevant: this is a really good result considering the
decentralised nature of the negotiational and game-theoretical proposed
framework.
In Chapter 7 we highlighted the importance of each model with respect to
different performance indicators. In this case the centralised one proves to
be the better one in optimising the utility function and no difference,
statistically speaking, can be considered between the two decentralised
approaches. Instead, by considering one performance measure at time, we
noticed that game theory models approaches better the global plants’
number of reconfigurations.
Conclusions 227
Chapter 8 showed how the system acts when the proposed models work in
an integrated way. This simulation leads us to highlight a statistically
relevant difference between the two decentralised models and between the
centralised and the game-theoretical one, but not between the centralised
and negotiational one. This result leads us to a very important conclusion:
in optimising the capacity allocation and the allocation to the groups of
the production plants the negotiation model acts very well, quite reaching
the same performance obtained in the centralised case. The noticed
differences are only due to the randomness of data input. However, when
the data uncertainty is high, the game-theoretical approach is more robust
in its performance than the negotiational one.
9.3 Directions for Future Work

This book opens up new interesting directions for future research in this area.
In Chapters 6 and 7 we proposed a set of properties which we believed
that the designed mechanisms should exhibit. In particular, we believed
that mechanisms should be provided with an infinite number of credits,
that agents should act in a cooperative way and that agents should not
have incentives to deceive others about their computed results.
Unfortunately, it is realistic to consider that these properties are, together,
not attainable. Therefore, at least one property must be relaxed.
Regarding the number of credits, it might be possible to consider some
incentives schemes that can be introduced in such a model. For example,
appropriate incentives will be needed to get the agent to solve the
problems, and to share the information with all agents to which it pertains.
Naturally, some problems that need to be addressed include:
o What sort and how much information should the agents reveal
through the mechanism?
o Is there a minimal amount of information that needs to be
revealed to the mechanism so that strategic-advantage
disappears?
o What sort of incentives are required before agents will truthfully
reveal this information?
o What type of feedback should the mechanism provide to the
agents?
Regarding the integrated model it might be interesting to evaluate the
influence of the mathematical models on the results. This can be obtained
by modifying the proposed mechanism considering not only the
input/output relationship between high and medium levels working
sequentially, but the mathematical integration too.
A
Simulation Results Related to Chapter 6
Table A.1. Centralised model results
Input
Avg Half Min Max Rep Var Check
parameter levels
WL P C 51.040 2.552 0 343 3541 58.973 5.00%
W W W 39.160 1.958 0 227 2369 37.009 5.00%
W W B 34.098 1.704 0 136 1785 27.957 5.00%
W W S 51.539 2.577 0 354 3645 60.417 5.00%
W B W 34.315 1.715 0 143 1670 27.218 5.00%
W B B 34.649 1.732 0 149 1662 27.421 5.00%
W B S 49.983 2.498 0 340 3725 59.218 5.00%
W S W 38.840 1.942 0 228 2389 36.857 5.00%
W S B 34.475 1.723 0 175 1751 28.006 5.00%
W S S 41.878 2.094 0 409 4621 55.275 5.00%
B W W 27.777 1.389 0 177 3100 30.026 5.00%
B W B 21.694 1.085 0 96 1791 17.826 5.00%
B W S 42.722 2.136 0 344 4529 55.816 5.00%
B B W 28.200 1.410 0 169 3165 30.796 5.00%
B B B 20.659 1.033 0 118 1984 17.863 5.00%
B B S 41.863 2.093 0 355 4703 55.744 5.00%
B S W 26.324 1.316 0 183 3359 29.624 5.00%
B S B 21.485 1.074 0 111 1875 18.062 5.00%
B S S 36.434 1.822 0 361 6000 54.796 5.00%
230 Appendix A
Table A.1 (continued)
S W W 18.486 0.924 0 184 5363 26.280 5.00%

S W B 8.744 0.437 0 52 2616 8.683 5.00%
S W S 35.884 1.794 0 388 5962 53.790 5.00%
S B W 18.671 0.933 0 214 5418 26.683 5.00%
S B B 8.451 0.422 0 50 2475 8.163 5.00%
S B S 35.216 1.761 0 339 6094 53.379 5.00%
S S W 18.595 0.930 0 173 5269 26.207 5.00%
S S B 8.762 0.438 0 45 2396 8.327 5.00%
S S S 51.040 2.552 0 343 3541 58.973 5.00%
Table A.2. Negotiational model results
Input parameter
levels
WL P C 65.826 3.291 0 368 1937 56.246 5.00%
W W W 43.880 2.193 0 216 1826 36.400 5.00%
W W B 36.765 1.838 0 137 1413 26.826 5.00%
W W S 66.142 3.306 0 357 1875 55.596 5.00%
W B W 43.321 2.165 0 268 1817 35.823 5.00%
W B B 34.803 1.739 0 187 1551 26.604 5.00%
W B S 67.615 3.380 0 352 1913 57.413 5.00%
W S W 44.582 2.228 0 211 1628 34.912 5.00%
W S B 36.657 1.832 0 150 1520 27.732 5.00%
W S S 58.570 2.928 0 340 2330 54.883 5.00%
B W W 31.366 1.568 0 185 2554 30.778 5.00%
B W B 21.999 1.099 0 97 1642 17.301 5.00%
B W S 57.717 2.886 0 412 2361 54.453 5.00%
B B W 30.194 1.509 0 167 2487 29.229 5.00%
B B B 21.926 1.096 0 149 1779 17.952 5.00%
B B S 57.637 2.881 0 357 2382 54.603 5.00%
B S W 30.400 1.519 0 208 2582 29.983 5.00%
B S B 21.064 1.053 0 99 1812 17.403 5.00%
Appendix A 231
Table A.2. (continued)
B S S 51.281 2.563 0 322 2619 50.947 5.00%

S W W 20.532 1.026 0 146 3941 25.021 5.00%
S W B 8.557 0.428 0 51 2549 8.388 5.00%
S W S 53.387 2.668 0 360 2497 51.784 5.00%
S B W 21.071 1.054 0 181 3940 25.681 5.00%
S B B 8.603 0.430 0 51 2573 8.472 5.00%
S B S 53.231 2.660 0 354 2600 52.679 5.00%
S S W 20.605 1.030 0 204 4022 25.365 5.00%
S S B 8.688 0.434 0 67 2632 8.651 5.00%
S S S 65.826 3.291 0 368 1937 56.246 5.00%
Table A.3. Game-theoretical model results
Input parameter levels Avg Half Min Max Rep Var Check
WL P C 67.729 3.386 0 383 1838 56.378 5.00%
W W W 43.636 2.181 0 261 1840 36.328 5.00%
W W B 35.612 1.780 0 173 1530 27.038 5.00%
W W S 64.861 3.241 0 338 1951 55.600 5.00%
W B W 35.100 1.754 0 149 1535 26.692 5.00%
W B B 34.894 1.744 0 146 1452 25.811 5.00%
W B S 67.269 3.362 0 387 1898 56.885 5.00%
W S W 43.118 2.155 0 195 1757 35.081 5.00%
W S B 36.148 1.806 0 150 1486 27.035 5.00%
W S S 59.126 2.956 0 433 2125 52.916 5.00%
B W W 31.559 1.578 0 191 2436 30.463 5.00%
B W B 21.575 1.078 0 91 1771 17.621 5.00%
B W S 58.884 2.944 0 357 2346 55.138 5.00%
B B W 31.647 1.582 0 185 2483 30.616 5.00%
B B B 21.730 1.086 0 98 1840 18.092 5.00%
B B S 57.858 2.892 0 367 2251 53.292 5.00%
B S W 30.507 1.525 0 210 2497 29.600 5.00%
B S B 21.739 1.087 0 90 1760 17.708 5.00%
232 Appendix A
Table A.3. (continued)
B S S 52.230 2.611 0 300 2486 50.555 5.00%

S W W 20.899 1.045 0 155 3866 25.223 5.00%
S W B 8.528 0.426 0 59 2671 8.556 5.00%
S W S 55.789 2.789 0 416 2561 54.811 5.00%
S B W 21.772 1.089 0 184 3904 26.413 5.00%
S B B 8.353 0.418 0 49 2523 8.145 5.00%
S B S 53.596 2.679 0 427 2552 52.552 5.00%
S S W 20.534 1.027 0 190 4091 25.500 5.00%
S S B 9.052 0.452 0 57 2367 8.549 5.00%
S S S 67.729 3.386 0 383 1838 56.378 5.00%
B
Simulation Input Parameters and Results Related

to Chapter 7
Table B.1. Values of V gm parameter
Group 1 Group 2 Group 3

Plant 1 0.9 0.2 0.5
Plant 2 0.8 0.2 0.9
Plant 3 0.4 0.5 0.7
Plant 4 0.6 0.7 0.9
Plant 5 1.0 0.2 0.8
Plant 6 0.8 0.5 0.6
Plant 7 0.3 0.4 0.6
Plant 8 0.8 0.9 0.3
Table B.2. Values of Dgm parameter
Group 1 Group 2 Group 3

Plant 1 0.6 0.3 0.8
Plant 2 0.8 1.1 0.3
Plant 3 0.9 1.2 0.7
Plant 4 1.1 0.4 0.9
Plant 5 0.5 0.7 1.2
Plant 6 0.6 0.5 0.7
Plant 7 0.6 1.2 1.0
Plant 8 1.1 0.9 0.9
234 Appendix B
Table B.3. Utility functions data
p min
g
p max
g
pm* Rc
0.5 2 0.5 0.6
Table B.4. Centralised model results for efficiency
Input parameter
levels
C R Q 14.447 0.127 5.7 23.6 3068 2.725 < 5%
W W W 14.435 0.143 6.4 22.4 1698 2.283 < 5%
W W B 12.180 0.238 6.2 18.6 565 2.194 < 5%
W W S 14.363 0.125 4.1 23.0 3008 2.665 < 5%
W B W 14.435 0.147 7.0 21.8 1552 2.242 < 5%
W B B 12.156 0.224 6.3 19.0 571 2.076 < 5%
W B S 14.430 0.129 5.8 23.1 2903 2.706 < 5%
W S W 14.503 0.144 7.0 21.8 1588 2.235 < 5%
W S B 14.473 0.130 4.7 22.8 2938 2.745 < 5%
W S S 14.670 0.112 4.6 23.9 3908 2.729 < 5%
B W W 14.945 0.128 8.4 22.8 1813 2.121 < 5%
B W B 12.070 0.265 7.3 17.8 370 1.978 < 5%
B W S 14.754 0.111 6.0 24.0 3964 2.713 < 5%
B B W 15.051 0.128 8.6 22.5 1918 2.171 < 5%
B B B 12.030 0.270 7.0 18.1 342 1.937 < 5%
B B S 14.771 0.112 4.1 23.7 3996 2.745 < 5%
B S W 14.946 0.132 6.9 21.9 1670 2.102 < 5%
B S B 12.056 0.289 7.3 19.1 350 2.103 < 5%
B S S 15.161 0.106 4.8 24.0 4743 2.841 < 5%
S W W 15.592 0.121 8.8 22.8 1830 2.015 < 5%
S W B 12.679 0.432 7.8 18.2 149 2.049 < 5%
S W S 15.070 0.102 6.2 24.0 4828 2.760 < 5%
S B W 15.664 0.124 8.9 22.4 1826 2.051 < 5%
Appendix B 235
Table B.4. (continued)
S B B 12.917 0.386 7.6 18.0 172 1.968 < 5%

S B S 15.121 0.109 5.5 24.4 4642 2.879 < 5%
S S W 15.096 0.104 6.0 24.4 4783 2.806 < 5%
S S B 12.653 0.402 8.8 17.3 147 1.891 < 5%
S S S 14.447 0.127 5.7 23.6 3068 2.725 < 5%
Table B.5. Centralised model results for distance
Input parameter
levels
C R Q 11.091 0.101 4.1 18.4 3068 2.167 < 5%
W W W 11.130 0.110 5 18.4 1698 1.764 < 5%
W W B 9.475 0.184 4.4 15.5 565 1.696 < 5%
W W S 11.016 0.100 2.9 18 3008 2.129 < 5%
W B W 11.093 0.111 5 18.2 1552 1.697 < 5%
W B B 9.473 0.178 4.3 15.6 571 1.653 < 5%
W B S 11.043 0.102 3.3 17.7 2903 2.135 < 5%
W S W 11.140 0.113 5.3 17.1 1588 1.745 < 5%
W S B 11.078 0.103 2.4 18.1 2938 2.173 < 5%
W S S 11.296 0.088 2.8 18.6 3908 2.148 < 5%
B W W 11.496 0.100 5.3 17.6 1813 1.654 < 5%
B W B 9.397 0.213 4.9 14.5 370 1.592 < 5%
B W S 11.320 0.089 4.6 18.9 3964 2.180 < 5%
B B W 11.611 0.101 5.9 17.5 1918 1.718 < 5%
B B B 9.354 0.222 4.7 14.4 342 1.591 < 5%
B B S 11.264 0.089 2.3 18.8 3996 2.190 < 5%
B S W 11.545 0.105 5.6 17.8 1670 1.659 < 5%
B S B 9.295 0.226 4.9 13.9 350 1.641 < 5%
B S S 11.669 0.086 3.4 19.2 4743 2.289 < 5%
S W W 11.988 0.096 4.9 18.4 1830 1.603 < 5%
S W B 9.586 0.346 5.4 14.4 149 1.642 < 5%
S W S 11.584 0.081 2.4 18.8 4828 2.192 < 5%
236 Appendix B
S B W 12.070 0.099 6.1 17.7 1826 1.637 < 5%

S B B 9.887 0.305 5.7 14.3 172 1.552 < 5%
S B S 11.607 0.086 2.8 19.6 4642 2.273 < 5%
S S W 11.625 0.084 3.5 19.1 4783 2.254 < 5%
S S B 9.641 0.301 6.1 12.7 147 1.417 < 5%
S S S 11.091 0.101 4.1 18.4 3068 2.167 < 5%
Table B.6. Centralised model results for reconfiguration
C R Q 8.632 0.148 0 22 3068 3.180 < 5%
W W W 8.830 0.189 1 19 1698 3.019 < 5%
W W B 6.681 0.290 1 19 565 2.681 < 5%
W W S 8.575 0.148 0 19 3008 3.150 < 5%
W B W 8.771 0.194 1 19 1552 2.970 < 5%
W B B 6.764 0.289 1 17 571 2.680 < 5%
W B S 8.607 0.149 0 25 2903 3.125 < 5%
W S W 8.738 0.183 1 19 1588 2.837 < 5%
W S B 8.511 0.148 0 19 2938 3.114 < 5%
W S S 8.444 0.139 0 20 3908 3.380 < 5%
B W W 8.574 0.183 1 20 1813 3.023 < 5%
B W B 5.697 0.269 0 16 370 2.757 < 5%
B W S 8.483 0.137 0 22 3964 3.357 < 5%
B B W 8.619 0.178 0 20 1918 3.030 < 5%
B B B 5.883 0.259 0 18 342 2.576 < 5%
B B S 8.511 0.139 0 20 3996 3.401 < 5%
B S W 8.646 0.188 0 20 1670 2.985 < 5%
B S B 5.666 0.175 0 15 350 2.724 < 5%
B S S 8.357 0.151 0 23 4743 4.029 < 5%
S W W 8.721 0.213 0 20 1830 3.531 < 5%
S W B 4.530 0.052 0 15 149 2.481 < 5%
Appendix B 237
S W S 8.484 0.150 0 23 4828 4.040 < 5%

S B W 8.692 0.205 0 20 1826 3.404 < 5%
S B B 4.541 0.182 0 14 172 2.962 < 5%
S B S 8.354 0.154 0 25 4642 4.086 < 5%
S S W 8.367 0.152 0 23 4783 4.073 < 5%
S S B 4.585 0.189 0 13 147 2.774 < 5%
S S S 8.632 0.148 0 22 3068 3.180 < 5%
Table B.7. Centralised model results for absolute residual
C R Q 52.838 1.788 0 291 3068 38.459 < 5%
W W W 46.064 1.967 0 215 1698 31.479 < 5%
W W B 54.876 2.437 8 170 565 22.500 < 5%
W W S 53.339 1.810 1 276 3008 38.541 < 5%
W B W 45.952 1.962 0 173 1552 30.020 < 5%
W B B 56.807 2.581 12 185 571 23.949 < 5%
W B S 53.358 1.802 0 279 2903 37.706 < 5%
W S W 46.858 2.000 0 216 1588 30.946 < 5%
W S B 54.362 1.869 1 301 2938 39.337 < 5%
W S S 43.519 1.372 1 332 3908 33.309 < 5%
B W W 33.201 1.420 0 161 1813 23.475 < 5%
B W B 44.257 2.015 13 110 370 15.054 < 5%
B W S 42.485 1.319 1 270 3964 32.239 < 5%
B B W 32.359 1.374 0 255 1918 23.362 < 5%
B B B 43.912 1.995 15 89 342 14.326 < 5%
B B S 43.392 1.370 0 249 3996 33.642 < 5%
B S W 32.646 1.393 1 134 1670 22.109 < 5%
B S B 44.517 2.011 16 94 350 14.613 < 5%
B S S 31.641 1.180 0 250 4743 31.562 < 5%
S W W 19.685 0.961 0 108 1830 15.963 < 5%
238 Appendix B
S W B 35.946 1.792 15 62 149 8.495 < 5%

S W S 31.063 1.159 0 241 4828 31.275 < 5%
S B W 19.797 0.970 0 129 1826 16.090 < 5%
S B B 36.802 1.831 13 59 172 9.326 < 5%
S B S 32.122 1.201 0 213 4642 31.768 < 5%
S S W 32.088 1.200 0 207 4783 32.230 < 5%
S S B 35.354 1.758 14 68 147 8.276 < 5%
S S S 52.838 1.788 0 291 3068 38.459 < 5%
Table B.8. Negotiation model results for efficiency
C R Q 15.600 0.124 9.1 21.5 2001 2.149 < 5%
W W W 15.738 0.145 7.9 21.1 1330 2.054 < 5%
W W B 15.954 0.153 9.8 21.5 1102 1.971 < 5%
W W S 15.638 0.130 7.6 22.3 2001 2.253 < 5%
W B W 15.697 0.136 8.9 21.1 1358 1.942 < 5%
W B B 15.710 0.151 8.7 21.1 1107 1.954 < 5%
W B S 15.677 0.131 6.7 22.1 2001 2.283 < 5%
W S W 15.599 0.147 8.6 22.0 1315 2.073 < 5%
W S B 16.504 0.136 9.1 23.4 2001 2.367 < 5%
W S S 15.901 0.120 8.1 22.1 2001 2.087 < 5%
B W W 16.020 0.121 10.9 21.6 1456 1.788 < 5%
B W B 15.864 0.140 9.9 20.9 1048 1.757 < 5%
B W S 15.542 0.115 8.5 21.4 2001 1.996 < 5%
B B W 15.621 0.121 10.0 21.5 1368 1.744 < 5%
B B B 15.617 0.128 10.2 20.9 1147 1.684 < 5%
B B S 16.931 0.125 7.8 22.6 2001 2.179 < 5%
B S W 17.319 0.135 9.9 22.8 1335 1.910 < 5%
B S B 17.093 0.150 11.4 22.7 1000 1.848 < 5%
B S S 15.910 0.113 8.8 22.3 2001 1.960 < 5%
Appendix B 239
S W W 15.919 0.099 10.8 21.0 1714 1.586 < 5%

S W B 15.913 0.137 11.6 20.6 800 1.506 < 5%
S W S 15.556 0.113 7.2 21.5 2001 1.962 < 5%
S B W 15.618 0.099 11.4 21.2 1660 1.562 < 5%
S B B 15.260 0.132 12.3 21.1 743 1.394 < 5%
S B S 17.173 0.124 9.1 22.6 2001 2.162 < 5%
S S W 17.035 0.120 10.4 23.2 2001 2.088 < 5%
S S B 17.499 0.174 12.8 22.6 728 1.823 < 5%
S S S 15.600 0.124 9.1 21.5 2001 2.149 < 5%
Table B.9. Negotiation model results for distance
C R Q 12.150 0.095 5.8 17.1 2001 1.652 < 5%
W W W 12.222 0.106 6.2 16.8 1330 1.507 < 5%
W W B 12.304 0.117 7.1 17 1102 1.505 < 5%
W W S 12.187 0.096 5.8 17.1 2001 1.667 < 5%
W B W 12.251 0.104 6.9 16.5 1358 1.483 < 5%
W B B 12.182 0.114 7.6 17.3 1107 1.476 < 5%
W B S 12.191 0.097 5.2 17.5 2001 1.681 < 5%
W S W 12.072 0.110 5.6 17.4 1315 1.553 < 5%
W S B 13.077 0.101 6.9 18.7 2001 1.746 < 5%
W S S 12.374 0.090 5.4 17.7 2001 1.568 < 5%
B W W 12.482 0.088 7.8 16.7 1456 1.303 < 5%
B W B 12.472 0.099 7.5 16.1 1048 1.240 < 5%
B W S 12.183 0.088 6.4 16.8 2001 1.524 < 5%
B B W 12.240 0.088 6.9 16.6 1368 1.262 < 5%
B B B 12.349 0.099 7.4 16.5 1147 1.300 < 5%
B B S 13.393 0.091 7 18.7 2001 1.589 < 5%
B S W 13.594 0.098 8.1 18 1335 1.391 < 5%
B S B 13.544 0.112 8.3 18.3 1000 1.372 < 5%
240 Appendix B
B S S 12.460 0.082 5.9 17.3 2001 1.426 < 5%

S W W 12.627 0.071 8.4 16.6 1714 1.136 < 5%
S W B 12.597 0.091 9.3 15.8 800 0.997 < 5%
S W S 12.293 0.083 5.9 17 2001 1.438 < 5%
S B W 12.375 0.070 8.8 16.2 1660 1.111 < 5%
S B B 12.239 0.094 9.8 16.1 743 0.996 < 5%
S B S 13.570 0.084 6.7 18.2 2001 1.467 < 5%
S S W 13.536 0.086 8.1 19.5 2001 1.501 < 5%
S S B 13.799 0.112 10.2 17.1 728 1.170 < 5%
S S S 12.150 0.095 5.8 17.1 2001 1.652 < 5%
Table B.10. Negotiation model results for reconfiguration
C R Q 8.392 0.166 0 22 2001 2.889 < 5%
W W W 8.521 0.195 2 19 1330 2.755 < 5%
W W B 8.624 0.218 1 19 1102 2.812 < 5%
W W S 7.290 0.162 0 23 2001 2.813 < 5%
W B W 7.557 0.191 1 19 1358 2.740 < 5%
W B B 7.371 0.204 0 16 1107 2.640 < 5%
W B S 7.508 0.163 0 19 2001 2.825 < 5%
W S W 7.290 0.187 0 20 1315 2.628 < 5%
W S B 6.041 0.155 0 17 2001 2.689 < 5%
W S S 7.224 0.149 1 16 2001 2.584 < 5%
B W W 7.350 0.171 0 18 1456 2.537 < 5%
B W B 7.163 0.190 0 14 1048 2.391 < 5%
B W S 6.000 0.139 0 16 2001 2.414 < 5%
B B W 5.966 0.154 1 16 1368 2.210 < 5%
B B B 6.085 0.168 0 14 1147 2.214 < 5%
B B S 4.600 0.125 0 12 2001 2.163 < 5%
B S W 4.352 0.142 0 13 1335 2.016 < 5%
Appendix B 241
B S B 4.155 0.160 0 12 1000 1.960 < 5%

B S S 6.246 0.144 0 15 2001 2.499 < 5%
S W W 6.122 0.148 0 15 1714 2.375 < 5%
S W B 6.183 0.223 0 15 800 2.451 < 5%
S W S 4.831 0.127 0 14 2001 2.206 < 5%
S B W 4.556 0.128 0 13 1660 2.026 < 5%
S B B 4.708 0.188 0 13 743 1.990 < 5%
S B S 3.187 0.103 0 11 2001 1.786 < 5%
S S W 3.148 0.103 0 14 2001 1.783 < 5%
S S B 2.297 0.149 0 10 728 1.560 < 5%
S S S 8.392 0.166 0 22 2001 2.889 < 5%
Table B.11. Negotiation model results for absolute residual
C R Q 99.455 1.930 17 285 2001 33.534 < 5%
W W W 78.208 1.915 18 224 1330 27.127 < 5%
W W B 72.400 1.876 23 154 1102 24.189 < 5%
W W S 99.618 1.911 16 266 2001 33.194 < 5%
W B W 79.564 1.960 15 214 1358 28.053 < 5%
W B B 73.294 1.940 20 170 1107 25.064 < 5%
W B S 100.044 1.970 25 282 2001 34.220 < 5%
W S W 79.071 1.917 27 200 1315 27.003 < 5%
W S B 100.966 1.946 28 301 2001 33.806 < 5%
W S S 90.706 1.662 31 262 2001 28.873 < 5%
B W W 66.760 1.299 25 171 1456 19.246 < 5%
B W B 60.804 1.328 21 130 1048 16.700 < 5%
B W S 90.959 1.642 27 224 2001 28.520 < 5%
B B W 67.991 1.366 17 160 1368 19.617 < 5%
B B B 60.643 1.260 23 121 1147 16.569 < 5%
B B S 91.443 1.630 27 222 2001 28.320 < 5%
B S W 69.216 1.384 24 177 1335 19.645 < 5%
242 Appendix B
B S B 62.961 1.405 23 132 1000 17.253 < 5%

B S S 80.350 1.562 19 197 2001 27.142 < 5%
S W W 54.680 0.901 18 120 1714 14.481 < 5%
S W B 48.356 1.032 17 93 800 11.341 < 5%
S W S 80.850 1.603 23 205 2001 27.839 < 5%
S B W 54.541 0.908 20 135 1660 14.364 < 5%
S B B 48.015 0.998 13 83 743 10.563 < 5%
S B S 80.982 1.629 22 229 2001 28.296 < 5%
S S W 79.678 1.619 17 213 2001 28.130 < 5%
S S B 50.346 1.105 18 85 728 11.579 < 5%
S S S 99.455 1.930 17 285 2001 33.534 < 5%
Table B.12. Game-theoretical model results for efficiency
C R Q 17.336 0.135 8.6 23.9 1973 2.321 < 5%
W W W 17.556 0.148 10.4 23.7 1307 2.079 < 5%
W W B 17.721 0.166 11.4 23.2 1018 2.061 < 5%
W W S 17.351 0.137 9.4 23.5 1933 2.341 < 5%
W B W 17.578 0.149 11.8 23.8 1242 2.037 < 5%
W B B 17.675 0.157 11.1 23.5 1071 1.997 < 5%
W B S 17.342 0.129 8 24.1 2104 2.305 < 5%
W S W 17.585 0.150 11.3 23.7 1284 2.081 < 5%
W S B 17.310 0.142 7.3 24.4 1921 2.413 < 5%
W S S 17.639 0.121 7.9 24.2 2341 2.277 < 5%
B W W 17.846 0.134 10.7 23 1334 1.907 < 5%
B W B 18.038 0.160 11.3 23.6 962 1.925 < 5%
B W S 17.594 0.122 8.4 23.6 2426 2.326 < 5%
B B W 17.955 0.132 10.6 24 1405 1.915 < 5%
B B B 17.891 0.153 12.1 23.6 948 1.834 < 5%
B B S 17.484 0.122 7.2 23.6 2409 2.331 < 5%
B S W 17.959 0.138 11.8 23.7 1294 1.931 < 5%
Appendix B 243
B S B 17.986 0.148 11.5 24.1 974 1.790 < 5%

B S S 17.936 0.106 7.6 24.1 2889 2.213 < 5%
S W W 18.767 0.116 13.7 24.1 1649 1.826 < 5%
S W B 18.766 0.145 14.6 23.3 692 1.486 < 5%
S W S 17.905 0.109 7.6 23.8 3053 2.333 < 5%
S B W 18.649 0.118 12.5 23.5 1583 1.826 < 5%
S B B 19.160 0.173 14.5 23.3 632 1.688 < 5%
S B S 17.912 0.098 8.9 24.1 3060 2.115 < 5%
S S W 17.915 0.101 8.4 23.8 3285 2.250 < 5%
S S B 18.940 0.182 14.2 23.2 645 1.794 < 5%
S S S 17.336 0.135 8.6 23.9 1973 2.321 < 5%
Table B.13. Game-theoretical model results for distance
C R Q 13.229 0.108 6.7 19 1973 1.862 < 5%
W W W 13.357 0.112 8.3 18.9 1307 1.571 < 5%
W W B 13.471 0.124 8.1 18.6 1018 1.537 < 5%
W W S 13.237 0.104 5.8 19 1933 1.783 < 5%
W B W 13.517 0.116 8.3 18.6 1242 1.588 < 5%
W B B 13.613 0.120 8.6 19 1071 1.525 < 5%
W B S 13.229 0.104 5.3 18.8 2104 1.848 < 5%
W S W 13.401 0.116 6.4 17.9 1284 1.614 < 5%
W S B 13.257 0.111 5.1 19 1921 1.891 < 5%
W S S 13.518 0.093 5.3 18.7 2341 1.745 < 5%
B W W 13.578 0.098 8 17.8 1334 1.386 < 5%
B W B 13.592 0.111 10 17.6 962 1.341 < 5%
B W S 13.412 0.094 6.2 18.8 2426 1.795 < 5%
B B W 13.654 0.101 8.9 17.8 1405 1.472 < 5%
B B B 13.695 0.111 9.2 17.9 948 1.323 < 5%
B B S 13.327 0.092 5.3 18.3 2409 1.761 < 5%
B S W 13.620 0.099 9.4 17.4 1294 1.384 < 5%
244 Appendix B
B S B 13.807 0.111 10 17.9 974 1.350 < 5%

B S S 13.505 0.077 6.1 18.8 2889 1.612 < 5%
S W W 13.571 0.072 9.9 16.6 1649 1.129 < 5%
S W B 13.674 0.099 10.1 16.6 692 1.013 < 5%
S W S 13.473 0.079 4.9 17.9 3053 1.701 < 5%
S B W 13.597 0.073 8.9 16.9 1583 1.126 < 5%
S B B 13.695 0.099 9.9 16.3 632 0.966 < 5%
S B S 13.735 0.079 7.1 18.9 3060 1.689 < 5%
S S W 13.485 0.073 6.3 18.5 3285 1.631 < 5%
S S B 13.598 0.094 10.5 16.4 645 0.929 < 5%
S S S 13.229 0.108 6.7 19 1973 1.862 < 5%
Table B.14. Game-theoretical model results for reconfiguration
C R Q 7.313 0.170 0 18 1973 2.933 < 5%
W W W 7.301 0.213 0 17 1307 2.997 < 5%
W W B 7.059 0.225 0 16 1018 2.792 < 5%
W W S 7.305 0.172 0 16 1933 2.935 < 5%
W B W 7.173 0.206 0 19 1242 2.814 < 5%
W B B 7.049 0.229 0 16 1071 2.905 < 5%
W B S 7.306 0.165 0 17 2104 2.937 < 5%
W S W 7.114 0.206 0 21 1284 2.862 < 5%
W S B 7.349 0.172 0 18 1921 2.928 < 5%
W S S 6.090 0.154 0 19 2341 2.897 < 5%
B W W 5.473 0.187 0 15 1334 2.648 < 5%
B W B 5.201 0.216 0 14 962 2.604 < 5%
B W S 6.019 0.152 0 18 2426 2.907 < 5%
B B W 5.506 0.185 0 16 1405 2.692 < 5%
B B B 5.412 0.225 0 16 948 2.695 < 5%
B B S 6.078 0.155 0 18 2409 2.947 < 5%
B S W 5.389 0.190 0 15 1294 2.653 < 5%
Appendix B 245
B S B 5.213 0.217 0 15 974 2.634 < 5%

B S S 4.372 0.137 0 19 2889 2.868 < 5%
S W W 3.035 0.131 0 13 1649 2.070 < 5%
S W B 2.436 0.116 0 10 692 1.701 < 5%
S W S 4.373 0.131 0 17 3053 2.817 < 5%
S B W 3.015 0.137 0 13 1583 2.120 < 5%
S B B 2.725 0.136 0 11 632 1.811 < 5%
S B S 4.384 0.130 0 17 3060 2.802 < 5%
S S W 4.409 0.127 0 16 3285 2.821 < 5%
S S B 2.488 0.106 0 10 645 1.733 < 5%
S S S 7.313 0.170 0 18 1973 2.933 < 5%
Table B.15. Game-theoretical model results for absolute residual
C R Q 86.182 1.974 29 185 1018 33.084 < 5%
W W W 93.291 2.012 26 219 1307 28.247 < 5%
W W B 86.182 1.974 29 185 1018 24.458 < 5%
W W S 116.245 2.039 42 302 1933 34.808 < 5%
W B W 92.717 2.009 26 203 1242 27.502 < 5%
W B B 86.317 2.001 31 204 1071 25.425 < 5%
W B S 114.494 1.937 38 260 2104 34.506 < 5%
W S W 93.139 2.024 30 211 1284 28.170 < 5%
W S B 115.081 2.005 30 295 1921 34.134 < 5%
W S S 97.415 1.633 28 252 2341 30.692 < 5%
B W W 71.642 1.434 22 158 1334 20.343 < 5%
B W B 64.127 1.397 25 140 962 16.822 < 5%
B W S 96.167 1.538 30 251 2426 29.415 < 5%
B B W 71.421 1.341 30 179 1405 19.514 < 5%
B B B 64.395 1.396 20 123 948 16.690 < 5%
B B S 96.281 1.552 28 240 2409 29.581 < 5%
B S W 71.702 1.382 22 156 1294 19.307 < 5%
246 Appendix B
B S B 65.642 1.439 27 146 974 17.443 < 5%

B S S 86.261 1.424 22 207 2889 29.718 < 5%
S W W 57.344 1.030 18 158 1649 16.243 < 5%
S W B 51.568 1.062 21 91 692 10.851 < 5%
S W S 84.639 1.342 20 217 3053 28.795 < 5%
S B W 57.830 0.980 22 116 1583 15.141 < 5%
S B B 52.350 1.157 19 89 632 11.295 < 5%
S B S 84.892 1.343 24 221 3060 28.850 < 5%
S S W 85.350 1.309 19 222 3285 29.131 < 5%
S S B 51.774 1.122 23 85 645 11.069 < 5%
S S S 86.182 1.974 29 185 1018 33.084 < 5%
C
Simulation Input Parameters and Results Related to

Chapter 8
Table C.1. Values of V gm parameter
Group 1 Group 2 Group 3 Group 4 Group 5 Group 6

Plant 1 0.9 0.2 0.5 – – –
Plant 2 0.8 0.2 0.9 – – –
Plant 3 0.4 0.5 0.7 – – –
Plant 4 0.6 0.7 0.9 – – –
Plant 5 1.0 0.2 0.8 – – –
Plant 6 0.8 0.5 0.6 – – –
Plant 7 0.3 0.4 0.6 – – –
Plant 8 0.8 0.9 0.3 – – –
Plant 9 – – – 0.9 0.2 0.5
Plant 10 – – – 0.8 0.2 0.9
Plant 11 – – – 0.4 0.5 0.7
Plant 12 – – – 0.6 0.7 0.9
Plant 13 – – – 1.0 0.2 0.8
Plant 14 – – – 0.8 0.5 0.6
Plant 15 – – – 0.3 0.4 0.6
Plant 16 – – – 0.8 0.9 0.3
248 Appendix C
Table C.2. Values of V gm parameter
Group 1 Group 2 Group 3 Group 4 Group 5 Group 6

Plant 1 0.6 0.3 0.8 – – –
Plant 2 0.8 1.1 0.3 – – –
Plant 3 0.9 1.2 0.7 – – –
Plant 4 1.1 0.4 0.9 – – –
Plant 5 0.5 0.7 1.2 – – –
Plant 6 0.6 0.5 0.7 – – –
Plant 7 0.6 1.2 1.0 – – –
Plant 8 1.1 0.9 0.9 – – –
Plant 9 – – – 0.6 0.3 0.8
Plant 10 – – – 0.8 1.1 0.3
Plant 11 – – – 0.9 1.2 0.7
Plant 12 – – – 1.1 0.4 0.9
Plant 13 – – – 0.5 0.7 1.2
Plant 14 – – – 0.6 0.5 0.7
Plant 15 – – – 0.6 1.2 1.0
Plant 16 – – – 1.1 0.9 0.9
Table C.3. Utility functions data
p min
g
p maxg
pm* Rc
0.5 2 0.5 0.6
Table C.4. Centralised model results for efficiency
Input parameters’levels Avg Half Min Max Rep Var Check
WLg Cg Qm Rm
W W W W 7.928 0.205 2.2 11.7 301 1.38 < 10%

W W W S 8.297 0.236 0.6 11.3 301 1.59 < 10%
W W S W 7.903 0.223 0.8 11 301 1.51 < 10%
W W S S 8.332 0.248 1.4 11.3 301 1.67 < 10%
W S W W 7.980 0.125 5.7 10.6 301 0.84 < 10%
Appendix C 249
Table C.4. (continued)
W S W S 8.534 0.112 6.7 10.4 301 0.76 < 10%

W S S W 7.454 0.182 6.2 9.6 120 0.77 < 10%
W S S S 7.577 0.045 6.2 7.6 84 0.16 < 10%
S W W W 7.984 0.231 1.2 11.2 301 1.56 < 10%
S W W S 8.369 0.203 2 11.3 301 1.37 < 10%
S W S W 8.026 0.209 2.8 10.9 301 1.41 < 10%
S W S S 8.470 0.231 0 11.6 301 1.55 < 10%
S S W W 7.901 0.128 5.5 10.3 301 0.86 < 10%
S S W S 8.546 0.129 6.1 10.7 301 0.87 < 10%
S S S W 7.474 0.175 5.9 9.4 121 0.75 < 10%
S S S S 7.317 0.053 6.7 7.4 87 0.19 < 10%
Table C.5. Centralised model results for distance
WLg Cg Qm Rm
W W W W 5.957 0.164 1.3 8.1 301 1.10 < 10%

W W W S 6.294 0.175 0.5 8.2 301 1.18 < 10%
W W S W 5.937 0.173 0.7 9 301 1.17 < 10%
W W S S 6.328 0.185 1.1 8.6 301 1.25 < 10%
W S W W 6.150 0.081 4.7 8 301 0.55 < 10%
W S W S 6.540 0.076 4.7 8 301 0.51 < 10%
W S S W 5.855 0.105 5.2 7.3 120 0.45 < 10%
W S S S 6.282 0.036 5.2 6.3 84 0.13 < 10%
S W W W 6.004 0.172 0.8 9.2 301 1.16 < 10%
S W W S 6.387 0.157 1.9 8.1 301 1.06 < 10%
S W S W 6.075 0.162 1.3 8.4 301 1.09 < 10%
S W S S 6.482 0.169 0 8.2 301 1.14 < 10%
S S W W 6.107 0.077 4.2 7.8 301 0.52 < 10%
S S W S 6.498 0.077 4.4 7.9 301 0.52 < 10%
S S S W 5.790 0.101 5.2 7 121 0.43 < 10%
S S S S 5.936 0.041 5.6 6 87 0.15 < 10%
250 Appendix C
Table C.6. Centralised model results for absolute residual
WLg Cg Qm Rm
W W W W 51.761 3.812 10 172 301 25.68 < 10%

W W W S 57.385 4.936 1 188 301 33.26 < 10%
W W S W 42.226 3.988 4 158 301 26.87 < 10%
W W S S 43.216 4.199 1 158 301 28.29 < 10%
W S W W 36.641 1.748 12 73 301 11.78 < 10%
W S W S 36.289 1.687 11 66 301 11.37 < 10%
W S S W 22.450 0.914 12 39 120 3.89 < 10%
W S S S 23.298 0.844 17 32 84 3.01 < 10%
S W W W 53.674 4.659 8 165 301 31.39 < 10%
S W W S 52.864 4.257 6 165 301 28.69 < 10%
S W S W 46.744 4.550 8 161 301 30.66 < 10%
S W S S 49.030 4.293 7 202 301 28.92 < 10%
S S W W 36.917 1.849 8 86 301 12.46 < 10%
S S W S 37.535 1.898 12 83 301 12.79 < 10%
S S S W 22.347 1.005 14 42 121 4.29 < 10%
S S S S 23.184 1.116 16 34 87 4.04 < 10%
Table C.7. Game-theoretical model results for efficiency
WLg Cg Qm Rm
W W W W 8.880 0.153 3.3 11.6 301 1.35 < 10%

W W W S 8.940 0.152 3.9 11.1 301 1.35 < 10%
W W S W 8.974 0.148 3.8 12 301 1.31 < 10%
W W S S 8.943 0.165 2 11.8 301 1.46 < 10%
W S W W 9.434 0.110 6.7 11.8 179 0.75 < 10%
W S W S 9.418 0.132 7.4 10.9 152 0.83 < 10%
W S S W 9.200 0.000 9.2 9.2 22 – < 10%
W S S S 9.200 0.000 9.2 9.2 14 – < 10%
Appendix C 251
S W W W 8.717 0.170 2.2 11.8 301 1.51 < 10%

S W W S 8.865 0.134 4 11.5 301 1.18 < 10%
S W S W 8.967 0.147 3.5 11.3 301 1.31 < 10%
S W S S 8.859 0.145 4.1 11.3 301 1.28 < 10%
S S W W 9.428 0.124 7.5 10.8 142 0.75 < 10%
S S W S 9.298 0.141 6.7 11.3 121 0.79 < 10%
S S S W 9.200 0.000 9.2 9.2 12 – < 10%
S S S S 9.200 0.000 9.2 9.2 16 – < 10%
Table C.8. Game-theoretical model results for distance
WLg Cg Qm Rm
W W W W 6.707 0.126 2.1 8.8 301 1.12 < 10%

W W W S 6.677 0.143 1.9 8.9 301 1.27 < 10%
W W S W 6.792 0.130 1.8 8.6 301 1.15 < 10%
W W S S 6.729 0.132 0.8 9.2 301 1.16 < 10%
W S W W 6.897 0.076 5.3 8.2 179 0.52 < 10%
W S W S 6.690 0.092 4.8 8 152 0.58 < 10%
W S S W 6.600 0.000 6.6 6.6 22 – < 10%
W S S S 6.600 0.000 6.6 6.6 14 – < 10%
S W W W 6.522 0.147 1.4 8.7 301 1.30 < 10%
S W W S 6.711 0.117 2.2 9.4 301 1.03 < 10%
S W S W 6.791 0.126 2.4 9.2 301 1.12 < 10%
S W S S 6.747 0.137 2.4 9 301 1.21 < 10%
S S W W 6.818 0.097 4.6 8.3 142 0.59 < 10%
S S W S 6.590 0.079 5.2 7.7 121 0.44 < 10%
S S S W 6.600 0.000 6.6 6.6 12 – < 10%
S S S S 6.600 0.000 6.6 6.6 16 – < 10%
252 Appendix C
Table C.9. Game-theoretical model results for absolute residual
WLg Cg Qm Rm
W W W W 58.791 3.304 16 214 301 29.24 < 10%

W W W S 59.694 3.349 11 182 301 29.65 < 10%
W W S W 53.668 3.563 7 165 301 31.54 < 10%
W W S S 47.153 3.160 7 156 301 27.97 < 10%
W S W W 38.927 1.922 11 73 179 13.12 < 10%
W S W S 40.474 1.991 10 75 152 12.52 < 10%
W S S W 24.000 2.131 10 35 22 – < 10%
W S S S 25.500 2.472 19 36 14 – < 10%
S W W W 58.485 3.267 12 184 301 28.92 < 10%
S W W S 58.392 3.478 11 253 301 30.79 < 10%
S W S W 49.213 2.964 7 143 301 26.24 < 10%
S W S S 50.741 3.238 7 150 301 28.66 < 10%
S S W W 40.141 2.066 13 74 142 12.56 < 10%
S S W S 39.182 1.988 13 64 121 11.16 < 10%
S S S W 23.667 2.267 19 34 12 – < 10%
S S S S 24.750 2.329 15 33 16 – < 10%
Table C.10. Negotiation model results for efficiency
WLg Cg Qm Rm
W W W W 7.956 0.220 0.6 11.2 301 1.48 < 10%

W W W S 8.460 0.223 1.6 12.2 301 1.51 < 10%
W W S W 7.928 0.215 1.9 11.2 301 1.45 < 10%
W W S S 8.296 0.225 0.8 12 301 1.52 < 10%
W S W W 7.928 0.124 5.5 10.3 301 0.84 < 10%
W S W S 8.437 0.126 6.2 10.6 301 0.85 < 10%
W S S W 7.476 0.186 5.9 9.5 113 0.77 < 10%
W S S S 7.546 0.077 6.2 7.6 54 0.22 < 10%
Appendix C 253
S W W W 7.998 0.222 1.8 11.2 301 1.49 < 10%

S W W S 8.248 0.233 0 12 301 1.57 < 10%
S W S W 7.869 0.207 2.5 10.9 301 1.39 < 10%
S W S S 8.427 0.222 3.1 11.9 301 1.49 < 10%
S S W W 7.950 0.123 5.9 10.2 301 0.83 < 10%
S S W S 8.536 0.124 5.8 10.7 301 0.84 < 10%
S S S W 7.430 0.145 5.9 9.4 153 0.70 < 10%
S S S S 7.318 0.053 6.2 7.4 125 0.23 < 10%
Table C.11. Negotiation model results for distance
WLg Cg Qm Rm
W W W W 6.033 0.167 0.6 8 301 1.13 < 10%

W W W S 6.357 0.173 1.1 8.5 301 1.17 < 10%
W W S W 5.984 0.173 1.7 8.6 301 1.16 < 10%
W W S S 6.353 0.164 0.6 8 301 1.10 < 10%
W S W W 6.090 0.085 4.2 7.7 301 0.57 < 10%
W S W S 6.482 0.079 5.1 7.9 301 0.53 < 10%
W S S W 5.790 0.099 5.2 6.9 113 0.41 < 10%
W S S S 6.257 0.061 5.2 6.3 54 0.17 < 10%
S W W W 5.957 0.174 0.8 8.6 301 1.17 < 10%
S W W S 6.242 0.181 0 9 301 1.22 < 10%
S W S W 5.954 0.159 1.4 7.9 301 1.07 < 10%
S W S S 6.492 0.173 1.9 9.1 301 1.17 < 10%
S S W W 6.121 0.078 4 7.5 301 0.53 < 10%
S S W S 6.632 0.075 4.8 8.1 301 0.51 < 10%
S S S W 5.831 0.096 5.2 7.6 153 0.46 < 10%
S S S S 5.936 0.041 5.2 6 125 0.18 < 10%
254 Appendix C
Table C.12. Game-theoretical model results for absolute residual
WL g Cg Qm Rm
W W W W 53.296 4.546 3 204 301 30.63 < 10%

W W W S 55.043 5.093 6 276 301 34.31 < 10%
W W S W 46.694 5.109 5 210 301 34.42 < 10%
W W S S 47.156 4.654 1 155 301 31.35 < 10%
W S W W 36.947 1.824 11 82 301 12.29 < 10%
W S W S 37.937 1.752 12 75 301 11.81 < 10%
W S S W 22.336 1.079 12 38 113 4.45 < 10%
W S S S 24.037 1.034 17 32 54 2.95 < 10%
S W W W 50.415 4.249 11 217 301 28.63 < 10%
S W W S 54.864 4.590 11 197 301 30.92 < 10%
S W S W 44.223 4.266 1 150 301 28.74 < 10%
S W S S 45.007 4.335 7 166 301 29.21 < 10%
S S W W 36.096 1.861 7 77 301 12.54 < 10%
S S W S 39.126 1.708 8 76 301 11.51 < 10%
S S S W 23.353 0.941 14 42 153 4.52 < 10%
S S S S 23.168 0.986 13 33 125 4.28 < 10%
Index
added-value principle, 21 decision support system, 5

advanced planning and scheduling decentralised decision-making, 6
(APS), v demand uncertainty, 3
agent mechanism design, 49 design of experiment (DoE), 77, 90, 107
ANOVA, 77, 107 distributed artificial intelligence, 29, 48
artificial intelligence, 42 distributed decision making, 48
axiomatic approach, 21, 23 distributed production planning, v, 9, 52,
56, 173
behaviour-dependent strategy, 37 distributed production system, 53
Bézier curve, 70
brownian motion model, 5 electronic negotiation, 31
electronic procurement, 28
capacity allocation, 99
capacity balancing, 60 fall-back outcome, 73
capacity configuration, 55 Foundation for Intelligent Physical
capacity expansion, 2, 54 Agents (FIPA), 50
capacity management, 2
capacity ownership, 55, 60 game theory, 13, 68, 102
capacity planning, 54 cooperative game, 18
centralised planning, 66, 100 core game, 69
characteristic function, 20, 69, 103 dynamic game, 17
class diagram, 74, 104 game payoff, 14, 15
coalition, 19 non-cooperative games, 15
empty coalition, 19 transfer utility game, 68, 103
grand coalition, 19 trigger strategy, 18
coalition structure, 19 generative function, 60, 101
communication channel (cc), 59 graph search problem, 47
cooperation, 29
coordination, v, 3, 9, 29, 48, 52, 227 heuristic models, 8
high-tech industry, 2, 55, 227
decentralised production network, 1 high-volume industry, 2, 9, 227
decision making, 34
corporate level, 53 IDEF0, viii, 56, 173
group level, 53 imitative tactic, 102
implicit collusion, 18
256 Index
Knowledge Query and Manipulation queuing network model, 8

Language (KQML), 50
reactive function, 60, 101
linear programming, 7 real options theory, 4
LINGO® package, 73, 103 reconfigurable enterprise, 9, 52
reconfigurable manufacturing systems,
mathematical programming, 6 1, 52
multi-agent systems, 30, 42, 103 reconfigurable production network, 52,
agent architecture, 56 173
agent policy, 45 reconfiguration cost, 100
agent reasoning, 45 reservation price, 101
multi-attribute utility theory (MAUT), 34 resource assignment, 62
resource-dependent strategy, 36
Nash equilibrium, 15, 37 response function, 15
negotiation, 26, 67, 101 risk attitude, 4, 60, 102, 106
credits, 59, 99 risk sharing, 3
negotiation protocol, 32 Rubinstein model, 25, 36, 39, 40
round, 101
news vendor model, 3 semiconductor industry, 2, 54
sequence diagram, 75, 105
on-line auctions, 28 Shapley value, 19, 21
operational planning, 54 simulation, 73, 103
optimal control, 45 simultaneous response protocol, 36
order assignement, 61 software agent, 29
ownership assigment, 59 stable set, 21
stochastic programming, 6
Pareto chart, 186, 187 stochastic transition model, 47
Pareto efficiency, 70 superadditivity, 20
Pareto frontier, 16
Pareto optimal, 23, 32
perceptual aliasing, 46 time-dependent tactic, 102
plant assigment, 60 Tukey test, 82
production capacity, 54 utility function, 47, 70, 71
production network, v
production planning, 53 wafer fabrication, 12, 54, 55

Force:drawing

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Force:drawing

Hochgeladen von

Copyright:

Verfügbare Formate

Production Planning in Production Networks

Pierluigi Argoneto • Giovanni Perrone

ISBN 978-1-84800-057-5 e-ISBN 978-1-84800-058-2

Globalisation is pushing manufacturing companies toward a more distributed

This book concerns the above-mentioned issues. It faces the production

Palermo, 25 July 2007

1 Introduction and Literature Overview ...........................................................1

2 Game Theory: an Overview...........................................................................13

3.4 Underlying Principle for Electronic Negotiation......................................30

4 Multiple-agent Systems: an Overview...........................................................41

5 Distributed Production Planning in Reconfigurable Production

6 Distributed Models for Planning Capacity of Reconfigurable Production

6.5.3 The Bargaining Solution..................................................................70

7 Distributed Models for Plant Capacity Allocation.......................................97

8 Distributed Production Planning Models: an Integrated Approach........171

Appendix A: Simulation Results Related to Chapter 6...................................229

Appendix B: Simulation Input Parameters and Results

Appendix C: Simulation Input Parameters and Results

Introduction and Literature Overview

1.2 Production Planning in High-tech, High-volume Industry

1.3 Strategic and Tactical Level

between two or more players need to be taken into account. In addition to

Capacity investment in high-tech industry typically involves substantial cash

lifecycle. Therefore a straightforward adoption of the Black and Scholes formula in

1.4 Operational Models: Optimization and Decision Support

1.4.1 Mathematical Approach

1.4.2 Queuing and Stochastic Approaches

1.4.3 Heuristics and Simulation-based Approaches

1.6 Book Outline

Game Theory: an Overview

2.2 Game Setup

2.3 Non-cooperative Static Games

If S i is quasi-concave in xi the best response is uniquely defined by the first-

2.4 Existence of Equilibrium

2.5 Multiple Equilibria

2.6 Dynamic Games

2.7 Simultaneous Moves: Repeated and Stochastic Games

2.8 Cooperative Games

2.9 N-Person Cooperative Games

Si  S j , for all i z j , and S j N.

2.10 Characteristic Function and Imputation

Definition 3. For each subset S of N, the characteristic function Q of a game gives

Definition 4. A game in characteristic function form consists of a set of players,

2.11 Shapley Value

The Shapley value assigns to each player his marginal contribution

2.12 The Bargaining Game Model

[4] Aumann RJ (1959) Acceptable points in general cooperative n-person games. In

Computer Science Economic sciences and

Information systems Law and social sciences

Most traditional negotiations have been conducted face-to-face; others have

and a negotiation process is structured. The use of Internet-based information

3.2 Negotiation and Rational Self-interested Agents

3.3 Negotiation Models

3.4 Underlying Principle for Electronic Negotiation

3.5 Electronic Negotiation Protocols

3.6 Characteristics that Differentiate Negotiations Protocols

 Distribution: the decision-making process should be distributed. There

3.7 Modelling Approaches and Solution Concepts

3.7.1 Decision Theory

3.7.2 Game Theory

3.7.3 Negotiation Analysis

3.8 Strategic Negotiation

3.9 Negotiation Strategies

Si S j , for all i z j , and S j N.

Distribution: the decision-making process should be distributed. There

in a deterministic world, the transition model maps a state action pair

i OG is requested to provide a certain number of credits to j in order to

subject to, i OG r , j UG r , the following constraints: