Simulating Social Complexity - A Handbook-Springer-Verlag Berlin Heidelberg (2013)

Understanding Complex Systems
Bruce Edmonds
Ruth Meyer Editors
Simulating
Social Complexity
A Handbook
Springer Complexity
Springer Complexity is an interdisciplinary program publishing the best research and
academic-level teaching on both fundamental and applied aspects of complex systems
— cutting across all traditional disciplines of the natural and life sciences, engineering,
economics, medicine, neuroscience, social and computer science.
Complex Systems are systems that comprise many interacting parts with the ability to
generate a new quality of macroscopic collective behavior the manifestations of which are
the spontaneous formation of distinctive temporal, spatial or functional structures. Models
of such systems can be successfully mapped onto quite diverse “real-life” situations like the
climate, the coherent emission of light from lasers, chemical reaction-diffusion systems,
biological cellular networks, the dynamics of stock markets and of the internet, earthquake
statistics and prediction, freeway traffic, the human brain, or the formation of opinions in
social systems, to name just some of the popular applications.
Although their scope and methodologies overlap somewhat, one can distinguish the
following main concepts and tools: self-organization, nonlinear dynamics, synergetics,
turbulence, dynamical systems, catastrophes, instabilities, stochastic processes, chaos,
graphs and networks, cellular automata, adaptive systems, genetic algorithms and compu-
tational intelligence.
The three major book publication platforms of the Springer Complexity program are the
monograph series “Understanding Complex Systems” focusing on the various applications
of complexity, the “Springer Series in Synergetics”, which is devoted to the quantitative
theoretical and methodological foundations, and the “SpringerBriefs in Complexity”
which are concise and topical working reports, case-studies, surveys, essays and lecture
notes of relevance to the field. In addition to the books in these two core series, the program
also incorporates individual titles ranging from textbooks to major reference works.
Editorial and Programme Advisory Board

Henry Abarbanel, Institute for Nonlinear Science, University of California, San Diego, USA
Dan Braha, New England Complex Systems Institute and University of Massachusetts Dartmouth, USA
Péter Érdi, Center for Complex Systems Studies, Kalamazoo College, USA and Hungarian Academy of
Sciences, Budapest, Hungary
Karl Friston, Institute of Cognitive Neuroscience, University College London, London, UK
Hermann Haken, Center of Synergetics, University of Stuttgart, Stuttgart, Germany
Viktor Jirsa, Centre National de la Recherche Scientifique (CNRS), Université de la Méditerranée,
Marseille, France
Janusz Kacprzyk, System Research, Polish Academy of Sciences,Warsaw, Poland
Kunihiko Kaneko, Research Center for Complex Systems Biology, The University of Tokyo, Tokyo,
Japan
Scott Kelso, Center for Complex Systems and Brain Sciences, Florida Atlantic University, Boca Raton,
USA
Markus Kirkilionis, Mathematics Institute and Centre for Complex Systems, University of Warwick,
Coventry, UK
Jürgen Kurths, Nonlinear Dynamics Group, University of Potsdam, Potsdam, Germany
Andrzej Nowak, Department of Psychology, Warsaw University, Poland
Linda Reichl, Center for Complex Quantum Systems, University of Texas, Austin, USA
Peter Schuster, Theoretical Chemistry and Structural Biology, University of Vienna, Vienna, Austria
Frank Schweitzer, System Design, ETH Zurich, Zurich, Switzerland
Didier Sornette, Entrepreneurial Risk, ETH Zurich, Zurich, Switzerland
Stefan Thurner, Section for Science of Complex Systems, Medical University of Vienna, Vienna, Austria
Understanding Complex Systems
Founding Editor: S. Kelso
Future scientific and technological developments in many fields will necessarily

depend upon coming to grips with complex systems. Such systems are complex in
both their composition - typically many different kinds of components interacting
simultaneously and nonlinearly with each other and their environments on multiple
levels – and in the rich diversity of behavior of which they are capable.
The Springer Series in Understanding Complex Systems series (UCS) promotes
new strategies and paradigms for understanding and realizing applications of
complex systems research in a wide variety of fields and endeavors. UCS is
explicitly transdisciplinary. It has three main goals:
First, to elaborate the concepts, methods and tools of complex systems at all
levels of description and in all scientific fields, especially newly emerging areas
within the life, social, behavioral, economic, neuro- and cognitive sciences (and
derivatives thereof); second, to encourage novel applications of these ideas in
various fields of engineering and computation such as robotics, nano-technology
and informatics; third, to provide a single forum within which commonalities and
differences in the workings of complex systems may be discerned, hence leading to
deeper insight and understanding.
UCS will publish monographs, lecture notes and selected edited contributions
aimed at communicating new findings to a large multidisciplinary audience.
For further volumes:

http://www.springer.com/series/5394
Bruce Edmonds • Ruth Meyer
Editors
Simulating
Social Complexity
A Handbook
Editors
Bruce Edmonds
Ruth Meyer
CPM - Centre for Policy Modelling
Manchester Metropolitan University
Manchester
United Kingdom
ISSN 1860-0832 ISSN 1860-0840 (electronic)

ISBN 978-3-540-93812-5 ISBN 978-3-540-93813-2 (eBook)
DOI 10.1007/978-3-540-93813-2
Springer Heidelberg New York Dordrecht London
Library of Congress Control Number: 2013934710
# Springer-Verlag Berlin Heidelberg 2013

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief excerpts
in connection with reviews or scholarly analysis or material supplied specifically for the purpose of being
entered and executed on a computer system, for exclusive use by the purchaser of the work. Duplication
of this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from
Springer. Permissions for use may be obtained through RightsLink at the Copyright Clearance Center.
Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

Contents
Part I Introductory Material
1 Introduction to the Handbook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Bruce Edmonds and Ruth Meyer
2 Historical Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Klaus G. Troitzsch
3 Types of Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Paul Davidsson and Harko Verhagen
Part II Methodology
4 Informal Approaches to Developing Simulation Models . . . . . . . . . 39

Emma Norling, Bruce Edmonds, and Ruth Meyer
5 A Formal Approach to Building Compositional Agent-Based
Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Catholijn M. Jonker and Jan Treur
6 Checking Simulations: Detecting and Avoiding Errors
and Artefacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
José M. Galán, Luis R. Izquierdo, Segismundo S. Izquierdo, José I. Santos,
Ricardo del Olmo, and Adolfo López-Paredes
7 Documenting Social Simulation Models: The ODD Protocol
as a Standard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Volker Grimm, Gary Polhill, and Julia Touza
8 Validating Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Nuno David
9 Understanding Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . 173
Andrew Evans, Alison Heppenstall, and Mark Birkin
v
vi Contents
10 Participatory Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

Olivier Barreteau, Pieter Bots, Katherine Daniell, Michel Etienne,
Pascal Perez, Cécile Barnaud, Didier Bazile, Nicolas Becu,
Jean-Christophe Castella, William’s Daré, and Guy Trebuil
11 Combining Mathematical and Simulation Approaches to
Understand the Dynamics of Computer Models . . . . . . . . . . . . . . . 235
Luis R. Izquierdo, Segismundo S. Izquierdo, José M. Galán,
and José I. Santos
12 Interpreting and Understanding Simulations: The Philosophy
of Social Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
R. Keith Sawyer
Part III Mechanisms
13 Utility, Games, and Narratives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293

Guido Fioretti
14 Social Constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
Martin Neumann
15 Reputation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Francesca Giardini, Rosaria Conte, and Mario Paolucci
16 Social Networks and Spatial Distribution . . . . . . . . . . . . . . . . . . . . 401
Frédéric Amblard and Walter Quattrociocchi
17 Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
Michael W. Macy, Stephen Benard, and Andreas Flache
18 Evolutionary Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
Edmund Chattoe-Brown and Bruce Edmonds
Part IV Applications
19 Agent-Based Modelling and Simulation Applied to

Environmental Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
Christophe Le Page, Didier Bazile, Nicolas Becu, Pierre Bommel,
François Bousquet, Michel Etienne, Raphael Mathevet,
Véronique Souchère, Guy Trébuil, and Jacques Weber
20 Assessing Organisational Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
Virginia Dignum
21 Distributed Computer Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563
David Hales
Contents vii
22 Simulating Complexity of Animal Social Behaviour . . . . . . . . . . . . 581

Charlotte Hemelrijk
23 Agent-Based Simulation as a Useful Tool for the Study
of Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 617
Juliette Rouchier
24 Movement of People and Goods . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
Linda Ramstedt, Johanna Törnquist Krasemann, and Paul Davidsson
25 Modeling Power and Authority: An Emergentist View from
Afghanistan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 667
Armando Geller and Scott Moss
26 Human Societies: Understanding Observed Social Phenomena . . . . 709
Bruce Edmonds, Pablo Lucas, Juliette Rouchier, and Richard Taylor
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 749
Part I
Introductory Material
Chapter 1
Introduction to the Handbook
Bruce Edmonds and Ruth Meyer
Why Read This Chapter? To understand some of the background and motivation
for the handbook and how it is structured.
1.1 Simulating Social Complexity
As the title indicates, this book is about Simulating Social Complexity. Each of
these three words is important.
Simulating – the focus here is on individual- or agent-based computational simulation
rather than analytic or natural language approaches (although these can be involved).
In other words, this book deals with computer simulations where the individual
elements of the social system are represented as separate elements of the simulation
model. It does not cover models where the whole population of interacting
individuals is collapsed into a single set of variables. Also, it does not deal with
purely qualitative approaches of discussing and understanding social phenomena, but
just those that try to increase their understanding via the construction and testing of
simulation models.
Social – the elements under study have to be usefully interpretable as interacting
elements of a society. The focus will be on human society, but can be extended
to include social animals or artificial agents where such work enhances our
understanding of human society. Thus this book does not deal with models of
single individuals or where the target system is dealt with as if it were a single
entity. Rather it is the differing states of the individuals and their interactions that
are the focus here.
B. Edmonds (*) • R. Meyer

Manchester Metropolitan University, Centre for Policy Modelling, Manchester Metropolitan
University Business School, All Saints Campus, Oxford Road, M15 6BH Manchester, UK
e-mail: [email protected]; [email protected]
B. Edmonds and R. Meyer (eds.), Simulating Social Complexity, 3

Understanding Complex Systems, DOI 10.1007/978-3-540-93813-2_1,
4 B. Edmonds and R. Meyer
Complexity – the phenomena of interest result from the interaction of social actors in
an essential way and are not reducible to considering single actors or a representative
actor and a representative environment. It is this complexity that (typically) makes
analytic approaches infeasible and natural language approaches inadequate for
relating intricate cause and effect. This complexity is expressed in many different
ways, for example: as a macro/micro link; as the social embedding of actors within
their society; as emergence. It is with these kinds of complexity that a simulation
model (of the kind we are focussing on) helps, since the web of interactions is too
intricate and tedious to be reliably followed by the human mind. The simulation
allows emergence to be captured in a formal model and experimented upon.
Since this area is relatively new it involves researchers from a wide variety of
backgrounds, including: computer scientists, sociologists, anthropologists, geographers,
engineers, physicists, philosophers, biologists and even economists. It is only in the past
few years that the elements and shape of the field are starting to be discernable. It is for
this reason that a handbook is particularly timely. We hope that it will help to introduce
and guide newcomers into the field so as to involve more minds and effort in this
endeavour, as well as inform those who enter it from one perspective to learn about other
sides and techniques.
1.2 The Context, Going Back to Herbert Simon
This handbook is in memory of Herbert Simon, since he initiated several key

strands that can be found in the work described here.
He observed how people behave in a social system instead of following some
existing framework of assumptions as to how they behave (Simon 1947). That is, he
tried to change the emphasis of study from a normative to a descriptive approach –
from how academics think people should be behaving to how people are observed to
behave. Famously he criticised “arm chair” theorising, the attempt to make theories
about social phenomena without confronting the theory with observation. There is
still a lot of “arm chair” theorising in the field of simulating social complexity, with a
“Cambrian Explosion” of simulation models, which are relatively unconstrained by
evidence from social systems. If the development of this work is seen as a sort of
evolutionary process then the forces of variation are there in abundance but the
forces of selection are weak or non-existent (Edmonds 2010).
Importantly for the simulation of complex social systems, Simon observed that
people act with a procedural rather than substantive rationality – they have a proce-
dure in the form of a sequence of actions that they tend to use to deal with tasks and
choices rather than try to find the best or ideal sequence of actions (Simon 1947,
1976). With the advent of computational simulation it is now fairly common to
represent the cognition of agents in a model with a series of rules or procedures.
This is partly because implementing substantive rationality is often infeasible due to
the computational expense of doing do, but more importantly it seems to produce
results with a greater “surface validity” (i.e. it looks right). It turns out that adding
1 Introduction to the Handbook 5
some adaptive or learning ability to individuals and allowing the individuals to

interact can often lead to effective “solutions” for collective problems (e.g. the entities
in Chap. 21). It is not necessary to postulate complex problem solving and planning
by individuals for this to occur.
Herbert Simon observed further that people tend to change their procedure only if it
becomes unsatisfactory; they have some criteria of sufficient satisfaction for judging a
procedure and if the results meet this they do not usually change what they do. Later
Simon (1956) and others (e.g. Sargent 1993) focused on the contrast between
optimisers and satisficers, since the prevailing idea of decision making was that
many possible actions are considered and compared (using the expected utility of
the respective outcomes) and the optimal action was the one that was chosen.
Unfortunately it is this later distinction that many remember from Simon, and not
the more important distinction between procedural and substantive rationality.
Simon’s point was that he observed that people use a procedural approach to tasks;
the introduction of satisficing was merely a way of modelling this (Sent 1997).
However, the idea of thresholds, that people only respond to a stimulus when it
become sufficiently intense, is often credible and is seen in many simulations (for
some examples of this, see Chaps. 22 and 25).
Along with Alan Newell, Simon made a contribution of a different kind to the
modelling of humans. He produced a computational model of problem solving in the
form of a computer program, which would take complex goals and split them into
sub-goals until the sub-goals were achievable (Newell and Simon 1972). The
importance of this, from the point of view of this book, is that it was a computational
model of an aspect of cognition, rather than one expressed in numerical and analytic
form. Not being restricted to models that can be expressed in tractable analytic forms
allows a much greater range of possibilities for the representation of human individ-
ual and social behaviour. Computational models of aspects of cognition are now
often introduced to capture behaviours that are difficult to represent in more tradi-
tional analytic models. Computational power is now sufficiently available to enable
each represented individual to effectively have its own computational process,
allowing a model to be distributed in a similar way to that of the social systems we
observe. Thus the move to a distributed and computational approach to modelling
social phenomena can be seen as part of a move away from abstract models divorced
from what they model towards a more descriptive type of representation.
This shift towards a more straightforward (even ‘natural’) approach to modelling
also allows for more evidence to be applied. In the past anecdotal evidence, in the
form of narrative accounts by those being modelled was deemed as “unscientific”.
One of the reasons that such evidence was rejected is that it could not be used to help
specify or evaluate formal models; such narrative evidence could only be used within
the sphere of rich human understanding and not at the level of a precise model.
Computational simulation allows some aspects of individual’s narratives to be used to
specify or check the behaviour of agents in a model, as well as the results being more
readily interpretable by non-experts. This has let such computational simulations to be
used in conjunction with stakeholders in a far more direct way than was previously
possible. Chapter 10 looks at this approach.
Herbert Simon did not himself firmly connect the two broad strands of his work:
the observation of people’s procedures in their social context and their algorithmic
modelling in computer models. This is not very surprising as the computational
power to run distributed AI models (which are essentially what agent-based
simulations are) was not available to him. Indeed these two strands of his work are
somewhat in opposition to each other, the one attempting to construct a general
model of an aspect of cognition (e.g. problem solving) and the other identifying quite
specific and limited cognitive procedures. I think it is fair to say that whereas Simon
did reject the general economic model of rationality, he did not lose hope of a general
model of cognitive processes, which he hoped would be achieved starting from good
observation of people. There are still many in the social simulation community who
hope for (or assume) the existence of an “off-the-shelf” model of the individuals’
cognition which could be plugged into a wider simulation model and get reasonable
results. Against any evidence, it is often simply hoped that the details of the
individuals’ cognitive model will not matter once embedded within a network of
interaction. This is an understandable hope, since having to deal with both individual
cognitive complexity and social complexity makes the job of modelling social
complexity much harder – it is far easier to assume that one or the other does not
matter much. Granovetter (1985) addressed precisely this question arguing against
both the under-socialised model of behaviour (that it is the individual cognition that
matters and the social effects can be ignored) and the over-socialised model (that it is
the society that determines behaviour regardless of the individual cognition).
Herbert Simon did not have at his disposal the techniques of individual- and agent-
based simulation discussed in this handbook. These allow the formal modelling of
socially complex phenomena without requiring the strong assumptions necessary to
make an equation-based approach (which is the alternative formal technique) analyti-
cally tractable. Without such simulation techniques modellers are faced with a
dilemma: either to “shoe-horn” their model into an analytically tractable form,
which usually requires them to make some drastic simplifications of what they are
representing, or to abandon any direct formal modelling of what they observe. In the
later case, without agent-based techniques, they then would have two further choices:
to simply not do any formal modelling at all remaining in the world of natural
language, or to ignore evidence of the phenomena and instead model their idea
concerning the phenomena. In other words, to produce an abstract but strictly analog-
ical model – a way of thinking about the phenomena expressed as a simulation. This
latter kind of simulation does not directly relate to any data derived from observation
but to an idea, which, in turn, relates to what is observed in a rich, informal manner.
Of course there is nothing wrong with analogical thinking, it is a powerful source of
ideas, but such a model is not amenable to scientific testing.
The introduction of accessible agent-based modelling opens up the world of social
complexity to formal representation in a more natural and direct manner. Each entity in
the target system can be represented by a separate entity (agent or object) in the model,
each interaction between entities as a set of messages between the corresponding
entities in the model. Each entity in the model can be different, with different
behaviours and attributes. The behaviour of the modelled entities can be realised in
terms of readily comprehensible rules rather than equations, rules that can be directly
compared to accounts and evidence of the observed entities’ behaviour. Thus the
mapping between the target system and model is simpler and more obvious than when
all the interactions and behaviour is “packaged up” into an analytic or statistical model.
Formal modelling is freed from its analytical straight jacket, so that the most appropri-
ate model can be formulated and explored. It is no longer necessary to distort a model
with the introduction of overly strong assumptions simply in order to obtain analytic
tractability. Also, agent-based modelling does not require high levels of mathematical
skill and thus is more accessible to social scientists. The outcomes of such models can
be displayed and animated in ways that make them more interpretable by experts and
stakeholders (for good and ill).
It is interesting to speculate what Herbert Simon would have done if agent-based
modelling was available to him. It is certainly the case that it brings together two of the
research strands he played a large part in initiating: algorithmic models of aspects of
cognition; and complex models that are able to take into account more of the available
evidence. We must assume that he would have recognised and felt at home with such
kinds of model. It is possible that he would not have narrowed his conception of
substantive rationality to that of satisficing if he had other productive ways of formally
representing the processes he observed in the way he observed them occurring.
It is certainly true that the battle he fought against “armchair theorising” (working
from a neat set of assumptions that are independent of evidence) is still raging. Even in
this volume you will find proponents (let’s call them the optimists) that still hope that
they can find some short-cut that will allow them to usefully capture social complexity
within abstract and simple models (theory-like models), and those (the pessimists) that
think our models will have to be complex, messy and specific (descriptive models) if
they are going to usefully represent anything we observe in the social world. However,
there is now the possibility of debate, since we can compare the results and success of
the optimistic and pessimistic approaches and indeed they can learn from each other.
It seems that research into social complexity has reached a cusp, between the
“revolutionary” and “normal” phases described by Kuhn (1962). A period of explor-
atory growth, opposed to previous orthodoxies, has occurred over the last 15–20 years,
where it was sufficient to demonstrate a new kind of model, where opening up new
avenues was more important than establishing or testing ideas about observed systems.
Now attention is increasingly turning to the questions such as: how to productively and
usefully simulate social complexity; how to do it with the greatest possible rigour; how
to ensure the strongest possible relation to the evidence; how to compare different
simulations; how to check them for unintentional errors; how to use simulation
techniques in conjunction with others (analytic, narrative, statistical, discourse analy-
sis, stakeholder engagement, data collection etc.). The field – if it is that – is maturing.
This handbook is intended to help in this process of maturation. It brings together
summaries of the best thinking and practice in this area, from many of the top
researchers. In this way it aims to help those entering into the field so that they do
not have to reinvent the wheel. It will help those already in the field by providing
accessible summaries of current thought. It aims to be a reference point for best current
practice and a standard against which future methodological advances are judged.
1.3 The Structure of the Handbook
The material in this book is divided into four sections: Introductory, Methodology,
Mechanisms and Applications. We have tried to ensure that each chapter within these
sections covers a clearly delineated set of issues. To aid the reader each chapter starts
with a very brief section called “Why read this chapter?” that sums up the reasons
you would read it in a couple of sentences. This is followed by an abstract, which
summarises the content of the chapter. Each chapter also ends with a section of
“Further Reading” briefly describing things that a newcomer might read next if they
are interested. This is separate from the list of references, which contains all the
references mentioned in the chapter.
1.3.1 Introductory Section
The introductory section includes three chapters: this chapter, a historical introduc-
tion (Chap. 2) that reviews the development of social simulation providing some
context for the rest of the book, and an overview of the different kinds of simulation
(Chap. 3).
1.3.2 Methodology Section
The next section on methodology consists of nine chapters that aim to guide the
reader through the process of simulating complex social phenomena. It starts with
two approaches to designing and building simulation models: formal (Chap. 5) and
informal (Chap. 4). The former being more appropriate where the goals and
specification of the proposed simulation are known and fixed, the latter more
appropriate in the case where possible models are being explored, in other words
when the simulation model one wants cannot be specified in advance.
However carefully a modeller designs and constructs such simulations they are
complex entities, which are difficult to understand completely. The next chapter (6)
guides the reader through the ways in which a simulation model can be checked to
ensure that it conforms to the programmer’s intentions for it. All of the approaches
described in these three chapters are aided by good, clear documentation. Chapter 7
describes a way of structuring and performing such documentation that helps to
ensure that all necessary information is included without being an overly heavy
burden.
Three chapters in this section are concerned with the results of simulations.
Chapter 8 concentrates on the validation of simulation models: the many ways in
which a model and the possible outputs from simulation runs can be related to data as
a check that it is correct for its purpose. Chapter 9 explores ways of analysing and
visualising simulation results, which is vital if the programmer or a wider audience is

to understand what is happening within complex simulations. Chapter 12 looks at the
wider question of the meaning and import of simulations, in other words the
philosophy of social simulation including what sort of theorising they imply.
Two other chapters consider separate aspects, but ones that will grow in importance
over time. Chapter 10 looks at participatory approaches to simulation, that is ways of
involving stakeholders more directly in the model specification and/or development
process. This is very different to an approach where the simulation model is built by
expert researchers who judge success by the correspondence with data sets, and can
almost become an intervention within a social process rather than a representation of it.
Chapter 11 investigates how analytic approaches can be combined with simulation
approaches, both using analytics to approximate and understand a simulation model as
well as using simulation to test the assumptions within an analytic model.
1.3.3 Mechanisms Section
The third section considers types of social mechanisms that have been used and
explored within simulations. It does not attempt to cover all such approaches, but
concentrates upon those with a richer history of use, where knowing about what has
been done might be important and possibly useful.
Chapter 13 takes a critical look at mechanisms that may be associated with
economics. Although this handbook is not about economic simulation1 mechanisms
from economics are often used within simulations with a broader intent. Unfortunately,
this is often done without thinking so that, for example, an agent might be programmed
using a version of economic rationality (i.e. considering options for actions and rating
them as to their predicted utility) just because that is what the modellers know or
assume. However, since economic phenomena are a subset of social phenomena this
chapter does cover these.
Chapter 14 surveys a very different set of mechanisms, those of laws, conventions
and norms. This is where behaviour is constrained from outside the individual in some
way (although due to some decision to accept the constraint from the inside to differing
degrees). Chapter 15 focuses on trust and reputation mechanisms; how people might
come to judge that a particular person is someone they want to deal with.
Chapter 16 looks at a broad class of structures within simulations, those that
represent physical space or distribution in some way. This is not a cognitive or social
mechanism in the same sense of the other chapters in this section, but has implications
for the kinds of interactions that can occur, and indeed facilitates some kinds of
interaction due to partial isolation of local groups.
The last two chapters in this section examine ways in which groups and individuals
might adapt. Learning and evolution are concepts that are not cleanly separable;
evolution is a kind of learning by the collection of entities that are evolving and has
1
There is an extensive handbook on this (Tesfatsion and Judd 2006).
been used to implement learning within an individual (e.g. regarding the set of
competing strategies an individual has) as well as within a society. However,
Chap. 17 investigates these concepts primarily from the point of view of algorithms
for an individual to learn, while Chap. 18 looks at approaches that explicitly take a
population and apply some selective pressures upon it, along with adding some sources
of variation.
1.3.4 Applications Section
The last section looks at eight areas where the techniques that have been described
are being applied. We chose areas where there has been some history of application
and hence some experience. Areas of application that are only just emerging are not
covered here.
Chapter 19 reviews applications to ecological management. This is one of the
oldest and most productive areas where simulation approaches have been applied.
Since it is inevitable that the interaction of society and the environment is complex,
analytic approaches are usually too simplistic and approaches that are better suited
are needed.
Chapter 20 discusses how simulation approaches have begun to inform the
design of organisations or its processes. Chapter 21 explores how a simulation-
based understanding of ICT systems can enable new kinds of distributed systems to
be designed and managed, while Chap. 22 looks at how simulation can help us
understand animal interaction. Chapter 23 describes agent-based simulations as a
useful tool to come to a complex understanding of how markets actually work (in
contrast to their economic idealisations). Chapter 24 considers systems where
people and/or goods are being moved within space or networks including logistics
and supply chains.
The last two chapters look at understanding human societies. Chapter 25 focuses
on a descriptive modelling approach to structures of power and authority, with
particular reference to Afghanistan, whereas Chap. 26 reviews the different ways in
which simulations have been used to understand human societies, briefly describing
examples of each.
References
Edmonds B (2010) Bootstrapping knowledge about social phenomena using simulation models.
J Artif Soc Soc Simulat 13(1). http://jasss.soc.surrey.ac.uk/13/1/8.html
Granovetter M (1985) Economic action and social structure: the problem of embeddedness. Am J
Sociol 91(3):481–510
Kuhn TS (1962) The structure of scientific revolutions. University of Chicago Press, Chicago
Newell A, Simon HA (1972) Human problem solving. Prentice-Hall, Englewood Cliffs
Sargent TJ (1993) Bounded rationality in macroeconomics: the Arne Ryde memorial lectures.
Clarendon, Oxford
Sent E-M (1997) Sargent versus Simon: bounded rationality unbound. Camb J Econ 21:323–338
Simon HA (1947) Administrative behavior: a study of decision-making processes in administrative

organizations. The Free Press, New York
Simon HA (1956) Rational choice and the structure of the environment. Psychol Rev 63(2):129–138
Simon HA (1976) Administrative behavior, 3rd edn. The Free Press, New York
Tesfatsion L, Judd KL (eds) (2006) Handbook of computational economics, vol 2: Agent-based
computational economics, vol 13: Handbooks in economics. North Holland, Amsterdam
Chapter 2
Historical Introduction
Klaus G. Troitzsch
Why Read This Chapter? To understand the historical context of simulation in

the social sciences and thus to better comprehend the developments and
achievements of the field.
Abstract This chapter gives an overview of early attempts at modelling social
processes in computer simulations. It discusses the early attempts, their successes
and shortcomings and tries to identify some of them as forerunners of modern
simulation approaches.
2.1 Overview
The chapter is organised as follows: the next section will discuss the early attempts
at simulating social processes, mostly aiming at prediction and numerical simula-
tion of mathematical models of social processes. Section 2.3 will then be devoted to
the non-numerical and early agent-based approaches, while Sect. 2.4 will give a
short conclusion, followed by some hints at further reading.
2.2 The First Two Decades
Simulation in the social sciences is nearly as old as computer simulation at large. This
is partly due to the fact that some of the pioneers of computer science – such as John
von Neumann, one of the founders of game theory – were at the same time pioneers in
the formalisation of social science. In addition, Herbert A. Simon, one of the pioneers
K.G. Troitzsch (*)

Institut für Wirtschafts- und Verwaltungsinformatik, Universität Koblenz-Landau,
Koblenz, Germany
e-mail: [email protected]

14 K.G. Troitzsch
in formalising social science, was an early adopter of computer-assisted methods of

building social theories. Thus the first two decades of computational social science
saw mathematical models and their inelegant solutions, microsimulation and even the
first agent-based models before the name of this approach was coined.
Among the first problems tackled with the help of computer simulation were
predictions of the future of companies (“industrial dynamics”, Forrester 1961),
cities (“urban dynamics”, Forrester 1969) and the world as a whole (“world
dynamics”, Forrester 1971) in the early 1960s and 1970s by Jay W. Forrester as
well as predictions of the consequences of tax and transfer laws for both the
individual household and the national economy in microanalytical simulation, an
attempt that started as early as 1956 (Orcutt 1957). Other early attempts at the
prediction of elections and referendum campaigns also became known in the 1960s,
such as Abelson and Bernstein’s simulation analysis of a fluoridation referendum
campaign within the Simulmatics project directed by de Sola Pool (Abelson and
Bernstein 1963). What all these early simulations have in common is that they were
aimed at predicting social and economic processes in a quantitative manner, and
that computer simulation was seen as a “substitute for mathematical derivations”
(Coleman 1964, p. 528). Despite Simon and others having already taught computers
to deal with non-numerical problems as early as 1955 (“Logic Theorist, the first
computer program that solved non-numerical problems by selective search”,
Simon 1996, pp. 189–190), 10 years later Coleman still believed that “the computer
cannot solve problems in algebra; it can only carry out computations when actual
numbers are fed into it (Coleman 1964, p. 529).
The remainder of this section will give a short overview of system dynamics and
microanalytic simulation – simulation approaches that continue to be promoted by
learned societies such as the System Dynamics Society and the International
Microsimulation Association, each celebrating their 50th anniversary with interna-
tional conferences held in Boston in July 2007 and Vienna in August 2007,
respectively –, before going into the details of some other early models that
remained more or less isolated and are now all but forgotten.
System dynamics was developed by Jay W. Forrester in the mid-1950s as a tool
to describe systems which could be modelled with large sets of difference and
differential equations containing functions whose mathematical treatment would
have been difficult or impossible. The general idea behind system dynamics was,
and is, that a system, without considering its components individually, could be
described in terms of its aggregate variables and their changes over time. The best
known examples of system dynamics models are Forrester’s (1971) and Meadows
et al. (1974) world models which were inspired by the Club of Rome and won
public attention in the 1970s when they tried to forecast the world population, the
natural resources, the industrial and agricultural capital and the pollution until the
end of the twenty-first century by describing the annual change of these aggregate
variables as functions of their current states and numerous parameters which had
some empirical background.
Microsimulation was first described in papers by Orcutt (1957) who designed a
simulation starting with a (sample of a) given population and simulating the individ-
ual fate of all the members of this population (sample) with the help of transition
2 Historical Introduction 15
probabilities empirically estimated from official statistics. Transitions represent

changes in the circumstances of an individual, e.g. switching to a different job,
achieving a higher educational level, marriage, birth of a child or death. These models
have mainly been used for predicting demographic changes and the effects of tax and
transfer rules. Usually they do not take into account that the overall changes of the
aggregated variables of the population (or the sample) may affect individual
behaviour. Thus in the sense of Coleman (1990, p. 10) these models neglect the
“downward causation” (i.e. the influence of the aggregate on the individual) and focus
only on the “upward causation”, namely the changes on the macro level, which are the
result of the (stochastically simulated) behaviour of the individuals.
The fluoridation referendum campaign model already mentioned above was one
of the first models that can be classified as an early predecessor of today’s agent-
based models. It consisted of a large number of representatives of people living in a
community faced with the option of compulsory fluoridation of drinking water – an
issue often discussed in the 1960s – which they would have to vote upon at the end
of a longish campaign, in which the media and local politicians were publishing
arguments in favour of or against this issue. In this model, 500 individuals are
exposed to information spread by several communication channels or sources and
additionally, they also exchange information among themselves. It depends on their
simulated communication habits to which extent they actually receive this infor-
mation and, moreover, to which extent this leads to changes in their attitudes
towards the referendum issue. Abelson and Bernstein defined 51 rules of behaviour,
22 of which are concerned with the processing of information spread over the
communication channels, whereas 27 rules are related to the information exchange
among individuals. Another two rules determine the final voting behaviour at the
end of the referendum campaign. The rules for processing the information from
public channels and those for processing the information exchanged among indi-
vidual citizens are quite similar; rule A3 and rule B2 read, for instance, “Receptivity
to [source] s is an inverse function of the extremity of [individual] i’s attitude
position.”
While this early model did not endow the model individuals with an appropriate
repertoire of behaviours, it nevertheless displays a relatively broad range of com-
munication possibilities among the model individuals – something that was neither
aimed at in the classical microanalytical simulation approach, nor in the cellular
automata approach adopted in the early 1970s in Thomas Schelling’s seminal paper
on segregation. One of the shortcomings of Abelson and Bernstein’s model in the
eyes of its critics was the fact that it “has never been fully tested empirically” (Alker
1974, p. 146). They also contested the adequacy of its “static representations of
citizen belief systems defined primarily in terms of assertions held, assertions
acceptance predispositions, with associated, more general, conflict levels” (Alker
1974, p. 146). Moreover, the assertions were modelled numerically (not a problem
with the proponents of a mathematical sociology who would even have used a large
system of differential equations to model the citizens’ attitude changes) where
obviously real citizens’ attitudes were never mapped on to the set of integer or
real numbers. Nowak et al. (1990, p. 371) give further reasons for the fact that this
approach was dropped for decades, “the ad hoc quality of many of the assumptions
16 K.G. Troitzsch
of the models, perhaps because of dissatisfaction with the plausibility of their

outcomes despite their dependence on extensive parameter estimation, or perhaps
because they were introduced at a time when computers were still cumbersome and
slow and programming time-consuming and expensive.”
Simulmatics suffered basically the same fate as Abelson and Bernstein’s model:
Simulmatics was set up “for the Democratic Party during the 1960 campaign. . . .
The immediate goal of the project was to estimate rapidly, during the campaign, the
probable impact upon the public, and upon small strategically important groups
within the public, of different issues which might arise or which might be used by
the candidates” (de Sola Pool and Abelson 1961, p. 167). The basic components of
this simulation model were voter types, 480 of them, not individual voters, with
their attitudes towards a total of 48 so-called “issue clusters”, i.e. “political
characteristics on which the voter type would have a distribution”. Voter types
were mainly defined by region, agglomeration structure, income, race, religion,
gender and party affiliation. From different opinion polls and for different points of
time these voter types were attributed four numbers per “issue cluster”: the number
of voters in this type and “the percentages pro, anti and undecided or confused on
the issue” (168). For each voter type empirical findings about cross-pressure
(e.g. anti-Catholic voters who had voted for the Democratic Party in the 1958
congressional elections and were likely to stay at home instead of voting for the
Catholic candidate of the Democrats) were used during a simulation run to re-adjust
the preferences of the voters, type by type. It is debatable whether this would
classify as a simulation in current social simulation communities, but since this
approach at least in some way resembles the classical static microsimulation, where
researchers are interested in the immediate consequences of new tax or transfer
laws with no immediate feedback, one could argue that Simulmatics was a simula-
tion project – though with as little sophistication as static microsimulation.
Thus the first two decades of computer simulation in the social sciences were
mainly characterised by two beliefs: that computer simulations were nothing but the
numerical solution of more adequate mathematical models, and that they were most
useful for predicting the outcome of social processes whose first few phases had
already been observed. This was also the core of the discussion opened in 1968 by
Hayward Alker who analysed, among others, the Abelson-Bernstein community
referendum model and came to the conclusion that this “simulation cannot be
‘solved’: one must project what will be in the media, what elites will be doing,
and know what publics already believe before even contingent predictions are made
about community decisions. In that sense an open simulation is bad mathematics
even if it is a good social system representation.” (Alker 1974, p. 153)
2.3 Computer Simulation in Its Own Right
The Simulmatics Corporation mentioned in the previous subsection did not only
work in the context of election campaigning, but later on also as a consulting
agency in other political fields. Their Crisiscom model is another example of an
early forerunner of current simulation models of negotiation and decision making

processes. At the same time it is an early example of a simulation not aimed at
prediction but at “our understanding of the process of deterrence by exploring how
far the behaviour of political decision makers in crisis can be explained by psycho-
logical mechanisms.” (de Sola Pool and Kessler 1965, p. 31) Crisiscom dealt with
messages of the type “actor one is related to actor two”, where the set of relations
was restricted to just two relations: affect and salience. In some way, Crisiscom
could also be used as part of a gaming simulation in which one or more of the actors
were represented by human players, whereas the others were represented by the
computer program – thus it can also be classified as a predecessor of participatory
simulation (see Chap. 10 in this volume).
The 1970s and 1980s saw a number of new approaches to simulate abstract
social processes, and most of them now were actual computer simulations, as – in
terms of Thomas Ostrom – they used the “third symbol system” (Ostrom 1988,
p. 384) directly by translating their ideas from the first symbol system, natural
language, into higher level programming languages instead of using it as a machine
to manipulate symbols of the second symbol system, mathematics. Although this
was already true for Herbert Simon’s Logic Theorist, the General Problem Solver
and other early artificial intelligence programs, the direct use of the “third symbol
system” in social science proper was not introduced before the first multilevel
models and cellular automata that integrated at least primitive agents in the sense
of software modules with some autonomy.
Cellular automata (Farmer et al. 1984; Ilachinski 2001) are a composition of
finite automata which all follow the same rule, are ordered in a (mostly) two-
dimensional grid and interact with (receive input from) their neighbours. The
behaviour of the individual cells is usually quite simple: they only have a small
number of states among which they switch according to relatively simple transition
rules. Prime example is the famous game of life (Gardener 1970), where the cells
are either alive or dead and change state according to two simple rules: (a) a cell
stays alive if it has exactly two or three live neighbouring cells, otherwise it dies;
(b) a dead cell bursts into life if there are exactly three live cells among its eight
neighbours. The great variety of outcomes on the level of the cellular automaton as
a whole enthused researchers in complexity science and laid the headstone for
innumerable cellular automata in one or two dimensions.
One of the first applications of cellular automata to problems of social science is
Thomas Schelling’s (1971) segregation model, demo versions of which are nowa-
days part of any distribution of simulation tools used for programming cellular
automata and agent-based models. This model shows impressively that segregation
and the formation of ghettos is inevitable even if individuals tolerate a majority of
neighbours different from themselves.
Another example is Bibb Latané’s Dynamic Social Impact theory with the
implementation of the SITSIM model (Nowak and Latané 1994). This model,
similar to Schelling’s, also ends up in clustering processes and in the emergence
of local structures in an initially randomly distributed population, but unlike
Schelling’s segregation model (where agents move around the grid of a cellular
18 K.G. Troitzsch
automaton until they find themselves in an agreeable neighbourhood) the clustering

in SITSIM comes from the fact that immobile agents adapt their attitudes to the
attitudes they find in their neighbourhood according to the persuasive strength of
their neighbours.
Other cellular automata models dealt with n-person cooperation games and
integrated game theory into complex models of interaction between agents and
their neighbourhoods. These models, too, usually end up in emergent local structures
(Hegselmann 1996).
Another game-theory-related computer simulation, run by Axelrod (1984),
showed the Tit-For-Tat strategy in the iterated prisoner’s dilemma as superior to
all other strategies represented in a computer tournament. The prisoner’s dilemma
had served game theorists, economists and social scientists as a prominent model of
decision processes under restricted knowledge. The idea stems from the early
1950s, first written down by Albert Tucker, and is about “two men, charged with
a joint violation of law, [who] are held separately by the police. Each is told that
(1) if one confesses and the other does not, the former will be given a reward . . . and
the latter will be fined . . . (2) if both confess, each will be fined . . . At the same time,
each has good reason to believe that (3) if neither confesses, both will go clear.”
(Poundstone 1992, pp. 117–118) In the non-iterated version the rational solution is
that both confess – but if they believe they can trust each other, they can both win,
as both will go clear if neither confesses. Axelrod’s question was under which
conditions a prisoner in this dilemma would “cooperate” (with his accomplice, not
with the police) and under which condition they would “defect” (i.e. confess, get a
reward and let the accomplice alone in prison). Strategies in this tournament had to
define which choice – cooperate or defect – each player would make, given the
history of choices of both players, but not knowing the current decision of the
partner. Then every strategy played the iterated game against every other strategy,
with identical payoff matrices – and the Tit-For-Tat strategy proved to be superior
to 13 other strategies proposed by economists, game theorists, sociologists,
psychologists and mathematicians (and it was the strategy that had the shortest
description in terms of lines-of-code). Although later on several characteristics of a
number of the strategies proposed could be analysed mathematically, the tourna-
ment had at least the advantage of easy understandability of the outcomes – which,
by the way, is another advantage of the “third symbol system” over the symbol
system of mathematics.
Cellular automata later on became the environment of even more complex
models of abstract social processes. They serve as a landscape where moving,
autonomous, pro-active, goal-directed software agents harvest food and trade
with each other. Sugarscape is such a landscape functioning as a laboratory for a
“generative social science” (Epstein and Axtell 1996, p. 19) in which the researcher
“grows” the emergent phenomena typical for real-world societies in a way that
includes the explanation of these phenomena. In this artificial world, software
agents find several types of food which they need for their metabolism, but in
different proportions, which gives them an incentive to barter one kind of food,
of which they have plenty, for another kind of food, which they urgently need.
Table 2.1 Overview of important approaches to computational social science

Approach Used since Characteristics
System Mid-1950s Only one object (the system) with a large number of
dynamics attributes
Microsimulation Mid-1950s A large number of objects representing individuals
that do not interact, neither with each other nor
with their aggregate, with a small number of
attributes each, plus one aggregating object
Cellular Mid-1960s Large number of objects representing individuals
automata that interact with their neighbours, with a very
restricted behaviour rule, no aggregating object,
thus emergent phenomena have to be visualised
Agent-based Early 1990s, with some Any number of objects (“agents”) representing
models forerunners in the individuals and other entities (groups, different
1960s kinds of individuals in different roles) that
interact heavily with each other, with an
increasingly rich repertoire of changeable
behaviour rules (including the ability to learn
from experience and/or others, to change their
behavioural rules and to react differently to
identical stimuli when the situation in which
they are received is different)
This kind of laboratory gives an insight under which conditions skewed wealth
distributions might occur or be avoided; with some extensions (König et al. 2002)
agents can even form teams led by agents who are responsible for spreading the
information gained by their followers among their group.
2.4 Conclusion
This short guided tour through early simulation models tried to show the optimism
of the early adopters of this method: “If it is possible to reproduce, through
computer simulation, much of the complexity of a whole society going through
processes of change, and to do so rapidly, then the opportunities to put social
science to work are vastly increased.” (de Sola Pool and Abelson 1961, p. 183) 35
years later, Epstein and Axtell formulate nearly the same optimism when they list a
number of problems that social sciences have to face – suppressing real-world
agents’ heterogeneity, neglecting non-equilibrium dynamics and being preoccupied
with static equilibria – and claim that “the methodology developed [in Sugarscape]
can help to overcome these problems” (Epstein and Axtell 1996, p. 2).
To complete this overview, Table 2.1 lists the approaches touched in this
introductory chapter with their main features.
As one can easily see from this table, only the agent-based approach is able to
“cover all the world” (Brassel et al. 1997), as only this approach can (a) include the
features of all the other approaches, and (b) meet the needs of social science for
20 K.G. Troitzsch
models of individuals which are able to exchange symbolic messages that have to be
interpreted by the recipients before they can take effect. When investigating large-
scale social phenomena involving large numbers of individuals in more or less similar
situations, then microsimulation, cellular automata, including sociophysics models
(Chakrabarti et al. 2006; Ball 2005), or even system dynamics may provide a good
(enough) approximation of what happens in human societies. But if we deal with
small communities – including the local communities Abelson and Bernstein
analysed –, then the process of persuasion, which needs at least one persuasive person
and one or more persuadable persons, has to be taken into account, and this calls for
agents of a richer structure than the early approaches could provide.
Further Reading
Most of the literature suggested for further reading has already been mentioned.
Epstein and Axtell’s (1996) work on generating societies gives a broad overview of
early applications of agent-based modelling. Epstein (2006) goes even further as he
defines this approach as the oncoming paradigm in social science. For the state of
the art of agent-based modelling in the social sciences at the onset of this approach,
the proceedings of early workshops and conferences on computational social
science are still worth reading (Gilbert and Doran 1994; Gilbert and Conte 1995;
Conte et al. 1997; Troitzsch et al. 1996).
And many early papers on computational social science were recently
republished (Gilbert 2010).
References
Abelson RP, Bernstein A (1963) A computer simulation of community referendum controversies.

Public Opin Q 27:93–122
Alker HR Jr (1974) Computer simulations: inelegant mathematics and worse social science. Int J
Math Educ Sci Technol 5:139–155
Axelrod R (1984) The evolution of cooperation. Basic Books, New York
Ball P (2005) Critical mass – how one thing leads to another. Arrow Books, London
Brassel KH, Möhring M, Schumacher E, Troitzsch KG (1997) Agents cover all the world? In:
Conte R, Hegselmann R, Terna P (eds) Simulating social phenomena (Lecture notes in
economics and mathematical systems), vol 456. Springer, Berlin, pp 55–72
Chakrabarti BK, Chakraborti A, Chatterjee A (eds) (2006) Econophysics and sociophysics: trends
and perspectives. Wiley, Weinheim
Coleman JS (1964) Introduction to Mathematical Sociology. The Free Press of Glencoe,
New York
Coleman JS (1990) The foundations of social theory. Harvard University Press, Boston
Conte R, Hegselmann R, Terna P (eds) (1997) Simulating social phenomena (Lecture notes in
economics and mathematical systems), vol 456. Springer, Berlin
de Sola Pool I, Abelson RP (1961) The simulmatics project. Public Opin Q 25:167–183
de Sola Pool I, Kessler A (1965) The Kaiser, the Czar, and the computer: information processing in
a crisis. Am Behav Sci 8:32–38
Epstein JM (2006) Generative social science: studies in agent-based computational modeling.
Princeton University Press, Princeton/Oxford
Epstein JM, Axtell R (1996) Growing artificial societies: social science from the bottom up.
Brookings/MIT Press, Washington/Cambridge, MA
Farmer D, Toffoli T, Wolfram S (eds) (1984) Cellular automata: proceedings of an interdisciplin-
ary workshop, Los Alamos, 7–11 Mar 1983. North-Holland, Amsterdam
Forrester JW (1961) Industrial dynamics. MIT/Wright Allen, Cambridge, MA
Forrester JW (1969) Urban dynamics. MIT/Wright Allen, Cambridge, MA
Forrester JW (1971) World dynamics. MIT/Wright Allen, Cambridge, MA
Gardener M (1970) The game of life. Sci Am 223(4):120–123
Gilbert N (ed) (2010) Computational Social Science Four-Volume Set. Sage, Los Angeles
Gilbert N, Conte R (eds) (1995) Artificial societies: the computer simulation of social life.
UCL Press, London
Gilbert N, Doran JE (eds) (1994) Simulating societies: the computer simulation of social phenom-
ena. UCL Press, London
Hegselmann R (1996) Cellular automata in the social sciences: perspectives, restrictions, and
artefacts. In: Hegselmann R, Mueller U, Troitzsch KG (eds) Modelling and simulation in the
social sciences from the philosophy of science point of view. Kluwer, Dordrecht, pp 209–234
Ilachinski A (2001) Cellular automata: a discrete universe. World Scientific, Singapore/River
Edge/London/Hong Kong
König A, Möhring M, Troitzsch KG (2002) Agents, hierarchies and sustainability. In: Billari F,
Prskawetz-Fürnkranz A (eds) Agent based computational demography. Physica, Berlin/
Heidelberg, pp 197–210
Meadows DL et al (1974) Dynamics of growth in a finite world. Wright-Allen, Cambridge, MA
Nowak A, Latané B (1994) Simulating the emergence of social order from individual behaviour.
In: Gilbert N, Doran JE (eds) Simulating societies: the computer simulation of social processes.
University College of London Press, London, pp 63–84
Nowak A, Szamrej J, Latané B (1990) From private attitude to public opinion: a dynamic theory of
social impact. Psychol Rev 97:362–376
Orcutt G (1957) A new type of socio-economic system. Rev Econ Stat 58:773–797
Ostrom TM (1988) Computer simulation: the third symbol system. J Exp Soc Psychol 24:381–392
Poundstone W (1992) Prisoner’s dilemma: John von Neumann, game theory, and the puzzle of the
bomb. Oxford University Press, Oxford
Schelling TC (1971) Dynamic models of segregation. J Math Sociol 1:143–186
Simon HA (1996) Models of my life. MIT Press, Cambridge, MA
Troitzsch KG, Mueller U, Gilbert N, Doran JE (eds) (1996) Social science microsimulation.
Springer, Berlin
Chapter 3
Types of Simulation
Paul Davidsson and Harko Verhagen
Why Read This Chapter? To understand the different ways that computer simu-
lation can differ in terms of (a) purpose, (b) targets for simulation, (c) what is
represented, and (d) its implementation; and subsequently, to be more aware of the
choices to be made when simulating social complexity.
Abstract This chapter describes the main purposes of computer simulation and
gives an overview of the main issues that should be regarded when developing
computer simulations. While there are two basic ways of representing a system in a
simulation model – the equation-based or macroscopic approach and the individual-
based or microscopic approach – this chapter (as the rest of the handbook) focuses
on the latter. It discusses the various options a modeller faces when choosing how
to represent individuals, their interactions and their environment in a simulation
model.
3.1 Introduction
Simulation concerns the imitation of some aspects of the reality (past, present, or
future) for some purpose. We should contrast computer simulation to physical
simulation in which physical objects are substituted for the real thing. These
physical objects are often chosen because they are smaller or cheaper than the
actual object or system. When (some of) the objects in a physical simulation are
humans, we may refer to this as human simulation. However, the focus of this book
is on computer simulation, and in particular, computer simulation of social
P. Davidsson (*)
Department of Computer Science, Malmö University, Malmö, Sweden
H. Verhagen
Stockholm University, Stockholm, Sweden

24 P. Davidsson and H. Verhagen
complexity, which concerns the imitation of the behaviour of one or more groups of
social entities and their interaction.
Computer simulation, as any other computer program, can be seen as a tool,
which could be used professionally, or used in the user’s spare time, e.g., when
playing computer games. It is possible to distinguish between different types of
professional users, e.g. scientists who use simulation in the research process to gain
new knowledge, policy makers who use it for making strategic decisions, managers
(of a system) who use it to make operational decisions, and engineers who use it
when developing systems. We can also differentiate two user situations, namely the
user as participant in the simulation and the user as observer of the simulation.
Computer games and training settings are examples of the former, where the user is
immerged in the simulation. In the case of using simulation as a tool for, say,
scientific research or decision support, the user is an outside observer of the
simulation. (In other words, we may characterize this difference as that between
interactive simulations and batch simulations.)
The main task of computer simulation is the creation and execution of a formal
model of the behaviour and interaction (of the entities) of the system being
simulated. In scientific research, computer simulation is a research methodology
that can be contrasted to empirically driven research.1 As such, simulation belongs
to the same family of research as analytical models. One way of formally modelling
a system is to use a mathematical model, and then attempt to find analytical
solutions enabling the prediction of the system’s behaviour from a set of parameters
and initial conditions. Computer simulation, on the other hand, is often used when
simple closed form analytic solutions are not possible. Although there are many
different types of computer simulation, they typically attempt to generate a sample
of representative scenarios for a model in which a complete enumeration of all
possible states would be prohibitive or impossible.
It is possible to make a general distinction between two ways of modelling the
system to be simulated. One is to use mathematical models and is referred to as
equation-based (or system dynamics or macro-level) simulation. In such models the
set of individuals (the population of the system) is viewed as a structure that can be
characterized by a number of variables. In the other way of modelling, which is
referred to as individual-based (or agent-based or micro-level) simulation, the
specific behaviours of specific individuals are explicitly modelled. In contrast to
equation-based simulation, the structure is viewed as emergent from the
interactions between the individuals and thus exploring the standpoint that complex
effects need not have complex causes. We will here, as well as in the remainder of
this book, focus on individual-based simulation.
In this chapter we will describe the main purposes of computer simulation and
also give an overview of the main issues that should be regarded when developing
computer simulations.
1
This distinction is of course not set in stone. For an example of an evidence-driven approach to
computer simulation see Chap. 25 in this volume (Geller and Moss 2013).
3 Types of Simulation 25
3.2 Purposes of Simulation
We can identify a number of distinct purposes of simulation. In general terms,

simulation is almost always used for analyzing (some aspects of) a system, typically
by predicting future states. More specifically, we may say that in the case when the
user is observing the simulation, the purpose is often one of the following:
– Management of a system, where simulation of (parts of) this system is used to
support operational decisions, i.e. which action to take, or strategic decisions, i.e.
which policy to use. The chapters on application areas in this book provide some
examples of this purpose; e.g., Chap. 19 addresses environmental management
(Le Page et al. 2013).
– Design or engineering of a system, where simulation is used as a tool to support
design decisions when developing a system. Chapter 21 illustrates how simula-
tion can help in the design of distributed computer systems (Hales 2013). In
fact, many new technical systems are distributed and involve complex interac-
tion between humans and machines, which make individual-based simulation a
suitable approach. The idea is to model the behaviour of the human users which
is useful in situations where it is too expensive, difficult, inconvenient, tiresome,
or even impossible for real human users to test out a new technical system. An
example of this is the simulation of “intelligent buildings” where software
agents model the behaviour of the people in the building (Davidsson 2000).
– Evaluation and verification, where simulation is used to evaluate a particular
theory, model, hypothesis, or system, or compare two or more of these. More-
over, simulation can be used to verify whether a theory, model, hypothesis,
system, or software is correct. An example of this purpose is found in Chap. 20
of this book on the assessment of (changes in) organizational design (Dignum
2013). More generally, in the context of social theory building, simulations can
be seen as an experimental method or as theories in themselves (Sawyer 2003).
In the former case, simulations are run e.g. to test the predictions of theories,
whereas in the latter case the simulations themselves are formal models of
theories. Formalizing the ambiguous, natural language-based theories of the
social sciences helps to find inconsistencies and other problems, and thus
contributes to theory building.
– Understanding, where simulation is used to gain deeper knowledge of a certain
domain. In such explorative studies, there is no specific theory, model, etc. to
be verified, but we want to study different phenomena (which may however
lead to theory refinement). Chapter 22 in this volume provides a number of
examples how simulation has helped in understanding animal social behaviour
(Hemelrijk 2012).
The focus of this book is on the user as an observer, the role of the user as
participant is just touched upon in Chap. 10 on participatory approaches (Barreteau
et al. 2013). However, to give a more complete picture, we have identified the
following purposes in the case when the user is participating in the simulation:
– Education, where simulation is used to explain or illustrate a phenomenon and

deepen the user’s theoretical knowledge. An example of this is the recently
developed SimPort,2 a multiplayer serious game where the players have to
construct a port area in the vicinity of Rotterdam. One aim of this simulation-
based tool is to give its users better insight into any unforeseen, undesirable and
unintentional effects of one or more development strategies and design variations
in the medium term (10–30 years) as a result of exogenous uncertainties (eco-
nomic, market, technological) and due to strategic behaviour of the parties
involved. Another example of individual-based simulation for educational pur-
pose is the PSI agent (Künzel and Hämmer 2006) that supports acquiring
theoretical insights in the realm of psychological theory. It enables students to
explore psychological processes without ethical problems.
– Training, where simulation is used to improve a person’s practical skills in a
certain domain. The main advantage of using simulation for training purposes is
to be part of a real-world-like situation without real-world consequences. An
early work in this area was a tool to help train police officers to manage large
public gatherings, such as crowds and protest marches (Williams 1993). Another
example of agent-based simulation for training purposes is Steve, an agent
integrated with voice synthesis software and virtual reality software providing
a very realistic training environment. For instance, it has been applied to
maintenance tasks in nuclear power plants (Méndez et al. 2003).
– Entertainment, where simulation is used just to please the user. There are a large
number of popular simulation games available. These belong to genres like
construction and management simulations, where players experience managing
a government, a sports team, a business, or a city, life simulations, where players
manage a life-form or ecosystem, such as the well-known “Sims” and its sequels,
vehicle simulations, where players experience driving a vehicle, such as an
airplane or a racing car, and of course different types of war games.
3.3 Types of Systems Simulated
It is possible to categorize the systems being simulated:

1. Human-centered systems, such as:
– Human societies, consisting of a set of persons with individual goals. That is,
the goal of different individuals may be conflicting. In Chap. 26 of this book,
more information on the simulation of human societies is given (Edmonds
et al. 2013).
2
http://www.simport.eu/
– Organizations, which we here define as structures of persons related to each

other in order to purposefully accomplishing work or some other kind of
activity. That is, the persons of an organization share some of their goals.
Further details on the modelling and simulation of organizations are provided
in Chap. 20 (Dignum 2013).
– Economic systems, which are organized structures in which actors
(individuals, groups, or enterprises) are trading goods or services on a market.
Chapter 23 (Rouchier 2013) takes a closer look at markets.
2. Natural systems, such as:
– Animal societies, which consist of a number of interacting animals, such as an
ant colony or a colony of birds. Chapter 22 (Hemelrijk 2013) is devoted to
simulation of animal societies.
– Ecological systems, in which animals and/or plants are living and evolving in
a relationship to each other and in dependence of the environment (even if
humans also are part of the ecological system, they are often not part of these
simulation models). In Chap. 19 (Le Page et al. 2013) more details on the
simulation of ecological systems are discussed.
3. Socio-technical systems, which are hybrid systems consisting of both living
entities (in most cases humans) and technical artefacts interacting with each
other. Examples of this type of system are transportation and traffic systems
concerning the movement of people, or goods in a transportation infrastructure
such as a road network. Chapter 24 (Ramstedt et al. 2013) provides a review of
simulation studies in these areas.
4. Artificial societies, which consist of a set of software and/or hardware entities,
i.e. computer programs and/or robots, with individual goals. One type of artifi-
cial societies, namely distributed computer systems, is treated in Chap. 21
(Hales 2013).
In addition, there are systems that are interesting to simulate using a micro-level
approach, but that we do not regard as social systems and are therefore not treated in
this book. One class of such systems are physiological systems, which consist of
functional organs integrated and co-operating in a living organism, e.g. subsystems
of the human body. Physical systems, which are collections of passive entities
following only physical laws, constitute another type of non-social systems.
3.4 Modelling
Let us now focus on how to model the system to be simulated. This depends on the
type of system and the purpose of the simulation study. An individual- or agent-
based model of a system consists of a set of entities and an environment in which
the entities are situated. The entities are either individuals (agents) that have some
decision-making capabilities, or objects (resources) that have no agency and are
purely physical. There are a number of characteristics that can be used to differen-
tiate between different types of models. We will first look at how individuals are
being modelled, then on the interaction between the individuals, and finally how the
environment is being modelled.
3.4.1 Individuals
A model of an individual can range from being very simple, such a one binary
variable (e.g. alive or dead) that is changed using only a single rule, to being very
complex. The complexity of the model for a given simulation should be determined
by the complexity of the individuals being simulated. Note, however, that very
complex collective behaviour could be achieved from very simple individual
models, if the number is sufficiently large.
We can distinguish between modelling the state of an individual and the
behaviour of the individual, i.e. the decisions and actions it takes. The state of an
individual, in turn, can be divided into the physical and the mental state. The
description of the physical state may include the position of the individual, and
features such as age, sex, and health status. The physical state is typically modelled
as a feature vector, i.e. a list of attribute/value pairs. However, this is not always the
case as in some domain the physical state of individual is not modelled at all. An
example is the PSI agent mentioned earlier that was used to give students theoreti-
cal insights in the area of psychological theory.
Whereas the physical state is often simple to model, representing the mental
state is typically much more complex, especially if the individuals modelled are
human beings. A common approach is to model the beliefs, desires, and intentions
of the individual, for instance by using the BDI model (Bratman 1987; Georgeff
et al. 1998). Such a model may include the social state of the individual, i.e. which
norms it adheres to, which coalitions it belongs to, etc. Although the BDI model is
not based on any experimental evidence of human cognition it has proven to be
quite useful in many applications. There has also been some work on incorporating
emotions in models of the mental state of individuals (cf. Bazzan and Bordini 2001)
as well as obligations, like the BOID model (Broersen et al. 2001), which extends
the BDI with obligations.
Modelling the behaviours (and decisions) of the individuals can be done in a
variety of ways, from simple probabilities to sophisticated reasoning and planning
mechanisms. As an example of the former we should mention dynamic micro-
simulation (Gilbert and Troitzsch 2005), which was one of the first ways of
performing individual-based simulation and is still frequently used. The purpose
is to simulate the effect the passing of time has on individuals. Data (feature
vectors) from a random sample from the population is used to initially characterize
the simulated individuals. A set of transition probabilities is then used to describe
how these features will change over a time period, e.g. there is a probability that an
employed person becomes unemployed during a year. The transition probabilities
are applied to the population for each individual in turn, and then repeatedly
re-applied for a number of simulated time periods. In traditional micro-simulation,
the behaviour of each individual is regarded as a “black box”. The behaviour is
modelled in terms of probabilities and no attempt is made to justify these in terms of
individual preferences, decisions, plans, etc. Thus, better results may be gained if
also the cognitive processes of the individuals were simulated.
Opening the black box of individual decision-making can be done in several
ways. A basic and common approach is to use decision rules, for instance, in the
form of a set of situation-action rules: If an individual and/or the environment is in
state X then the individual will perform action Y. By combining decision rules and
the BDI model quite sophisticated behaviour can be modelled. Other models of
individual cognition used in agent-based social simulation include the use of Soar,
a computer implementation of Allen Newell’s unified theory of cognition (Newell
1994), which was used in Steve (discussed above). Another unified theory of
individual cognition, for which a computer implementation exists, is ACT-R
(Anderson et al. 2004), which is realized as a production system. A less general
example is the Consumat model (Janssen and Jager 1999), a meta-model combining
several psychological theories on decision making in a consumer situation. In
addition, non-symbolic approaches such as neural networks have been used to
model the agents’ decision making (Massaguer et al. 2006).
As we have seen, the behaviour of individuals could be either deterministic or
stochastic. Also, the basis for the behaviour of the individuals may vary. We can
identify the following categories:
– The state of the individual itself: In most social simulation models the physical
and/or mental state of an individual plays an important role in determining its
behaviour.
– The state of the environment: The state of the environment surrounding the
individual often influences the behaviour of an individual. Thus, an individual
may act differently in different contexts although its physical and mental state is
the same.
– The state of other individuals: One popular type of simulation model, where the
behaviour of individuals is (solely) based on the state of other individuals, is
those using cellular automata (Schiff 2008). Such a simulation model consists of
a grid of cells representing individuals, each in one of a finite number of states.
Time is discrete and the state of a cell at time t is a function of the states of a
finite number of cells (called its neighbourhood) at time t 1. These neighbours
are a fixed selection of cells relative to the specified cell. Every cell has the same
rule for updating, based on the values in its neighbourhood. Each time the rules
are applied to the whole grid a new generation is created. In this case, informa-
tion about the state of other individuals can be seen as gained through
observations. Another possibility to gain this information is through communi-
cation, and in this case the individuals do not have to be limited to the
neighbours.
– Social states (norms etc.) as viewed by the agent: For simulation of social
behaviour the agents need to be equipped with mechanisms for reasoning at
the social level (unless the social level is regarded as emergent from individual
behaviour and decision making). Several models have been based on theories
from economy, social psychology, sociology, etc. Guye-Vuillème (2004)
provides an example of this with his agent-based model for simulating human
interaction in a virtual reality environment. The model is based on sociological
concepts such as roles, values, and norms and motivational theories from social
psychology to simulate persons with social identities and relationships.
In most simulation studies, the behaviour of the individuals is static in the sense
that decision rules or reasoning mechanisms do not change during the simulation.
However, human beings and most animals do have an ability to adapt and learn. To
model dynamic behaviour of individuals through learning/adaptation can be done in
many ways. For instance, both ACT-R and Soar have learning built in. Other types
of learning include the internal modelling of individuals (or the environment) where
the models are updated more or less continuously.
Finally, there are some more general aspects to consider when modelling
individuals. One such aspect is whether all agents share the same behaviour or
whether they behave differently, in other words, representation of behaviour is either
individual or uniform. Another general aspect is the number of individuals modelled,
i.e. the size of the model, which may vary from a few individuals to billions of
individuals. Moreover, the population of individuals could be either static or
dynamic. In dynamic populations, changes in the population are modelled, typically
births and deaths.
3.4.2 Interaction Between Individuals
In dynamic micro-simulation simulated individuals are considered in isolation

without regard to their interaction with others. However, in many situations the
interaction between individuals is crucial for the behaviour at system level. In such
cases better results will be achieved if the interaction between individuals was
included in the model. Two important aspects of interaction are (a) who is
interacting with whom, i.e. the interaction topology, and (b) the form of this
interaction.
A basic form of interaction is physical interaction or interaction based on spatial
proximity. As we have seen, this is used in simulations based on cellular automata,
e.g. in the well-known Game of Life (Gardner 1970). The state of an individual is
determined by how many of its neighbours are alive. Inspired by this work
researchers developed more refined models, often modelling the social behaviour
of groups of animals or artificial creatures. One example is the Boid model by
Reynolds (1987), which simulates coordinated animal motion such as bird flocks
and fish schools in order to study emergent phenomena. In these examples,
the interaction topology is limited to the individuals immediately surrounding an

individual. In other cases, as we will see below, the interaction topology is defined
more generally in terms of a (social) network. Such a network can be either static,
i.e. the topology does not change during a simulation, or dynamic. In these networks
interaction is typically language-based. An example is the work by Verhagen
(2001), where agents that are part of a group use direct communication between
the group members to form shared group preferences regarding the decisions they
make. Communication is steered by the structure of the social network regardless of
the physical location of the agents within the simulated world. For a more detailed
discussion of the different options to model interaction topologies see Chap. 16 in
this volume (Amblard and Quattrochiocchi 2013).
3.4.3 The Environment
The state of the environment is usually represented by a set of (global) parameters,

e.g. temperature. In addition, there are a number of important aspects of the
environment model, such as:
– Spatial explicitness: In some models, there is actually no notion of physical
space at all. An example of a scenario where location is of less importance are
“innovation networks” (Gilbert et al. 2001). Individual agents are high-tech
firms that each have a knowledge base used to develop artefacts to launch on a
simulated market. The firms are able to improve their products through research
or by exchanging knowledge with other firms. However, in many scenarios
location is very important, thus each individual (and sometimes objects) is
assigned a specific location at each time step of the simulation. In this case, the
individuals may be either static (the entity does not change location during the
simulation) or mobile. The location could either be specified as an absolute
position in the environment, or in terms of relative positions between entities.
In some areas the simulation software is integrated with a Geographical Infor-
mation System (GIS) in order to achieve closer match to reality (cf. Schüle
et al. 2004).
– Time: There are in principle two ways to address time, and one is to ignore it. In
static simulation time is not explicitly modelled; there is only a “before” and an
“after” state. However, most simulations are dynamic, where time is modelled as
a sequence of time steps. Typically, each individual may change state between
each time step.
– Exogenous events: This is the case when the state of the environment, e.g. the
temperature, changes without any influence/action from the individuals. Exoge-
nous events, if they are modelled, may also change the state of entities, e.g.
decay of resources, or cause new entities to appear. This is a way to make the
environment stochastic rather than deterministic.
3.4.4 Factors to Consider When Choosing a Model
In contrast to some of the more traditional approaches, such as system dynamics,

individual-based modelling does not yet have any standard procedures that can
support the model development. (Although some attempts in this direction have
been made, e.g. by Grimm et al. (2006) in the area of ecological systems.) In
addition, it is often the case that the only formal description of the model is the
actual program code. However, it may be useful to use the Unified Modelling
Language (UML) to specify the model.
Some of the modelling decisions are determined by the features of the system to
be simulated, in particular those regarding the interaction model and the environ-
ment model. The hardest design decision is often how the mental state and the
behaviour of individuals should be modelled, in particular when representing
human beings. For simpler animals or machines, a feature vector combined with
a set of transitions rules is often sufficient. Depending on the phenomena being
studied, this may also be adequate when modelling human beings. Gilbert (2006)
provides some guidelines whether a more sophisticated cognitive model is neces-
sary or not. He states that the most common reason for ignoring other levels is that
the properties of these other levels can be assumed constant, and exemplifies this by
studies of markets in equilibrium where the preferences of individual actors are
assumed to remain constant. (Note, however, that this may not always be true).
Another reason for ignoring other levels, according to Gilbert, is when there are
many alternative processes at the lower level, which could give rise to the same
phenomenon at the macro level. He illustrates this with the famous study by
Schelling (1971) regarding residential segregation. Although Schelling used a
very crude model of the mental state and behaviour of the individuals, i.e. ignoring
the underlying motivations for household migration, the simulation results were
valid (as the underlying motivations were not relevant for the purpose of
Schelling’s study).
On the other hand, there are many situations where a more sophisticated
cognitive model is useful, in particular when the mental state or behaviour of the
individual constraints or in other ways influences the behaviour at the system level.
However, as Gilbert concludes, the current research is not sufficiently mature in
order to give advice on which cognitive model to use (BDI, Soar, ACT-R, or other).
Rather, he suggests that more pragmatic considerations should guide the selection.
The model of the environment is mostly dictated by the system to be simulated,
with the modeller having to decide on the granularity of the values the environ-
mental attributes can take. The interaction model is often chosen based on the
theory or practical situation that lies at the heart of the simulation, but sometimes
the limitations of the formal framework used restrict the possibilities. Here the
modeller also has to decide upon the granularity of attribute values.
3.5 Implementation
We will now discuss some issues regarding the implementation (programming and
running) of a simulator.
A simulator can be time-driven, where the simulated time is advanced in
constant time steps, or event-driven, where the time is advanced based on the
next event. In an event-driven simulation, a simulation engine drives the simulation
by continuously taking the first event out of a time-ordered event list, and then
simulating the effects on the system state caused by this event. Since time segments
where no event takes place are not regarded, event-driven simulation is often more
efficient than time-driven simulation. On the other hand, since time is incremented
at a constant pace during a simulation in time-driven mode, this is typically a better
option if the simulation involves user participation.
There are a number of platforms or toolkits for agent-based simulation available,
such as Swarm, NetLogo and RePast (see Railsback et al. (2006) for a critical
review of these and some other platforms). These are freely available, simplify the
programming and can be of great help, in particular for modellers that are not
skilled programmers. However, they all impose some limitations on what can be
modelled, which may or may not be crucial for the application at hand. An approach
without such limitation is of course to program the simulator from scratch using
ordinary programming languages like Java or C, which is more difficult and time
consuming. In some cases, e.g. if you want to distribute the simulation on a number
of computers, it may be appropriate to use an agent platform, such as JADE. In this
case, the individuals may be implemented as actual software agents. In particular,
when the number of individuals simulated is large and/or the models of individuals
are complex, it may be too time consuming to run the simulation on a single
computer. Instead, one may distribute the computational load on several computers
in order to get reasonable running times. It should be mentioned that there are some
efforts on making agent-based simulation platforms run on large-scale computer
networks such as Grids, see e.g. the work by Chen et al. (2008).
It is worth noting that the resulting software is an approximation of a simulation
model, which in turn is an approximation of the actual system. Thus, there are
several steps of verification and validation that need to be addressed in the devel-
opment of a simulation model, as discussed in Chap. 8 (David 2013).
3.6 Conclusion
As we have seen, there are many different types of individual-based social simula-
tion. In the table below, we provide a summary.
Focus Aspect Options

Usage Users Scientists
Policy makers
Managers
Non-professionals
Purposes Management of a system
Design or engineering of a system
Evaluation and verification
Understanding
Education
Training
Entertainment
System simulated Human-centered systems Human societies
Organizations
Economic systems
Natural systems Animal societies
Ecological systems
Socio-technical systems
Artificial systems
Individual model Individual physical state Feature vector
Individual mental state Feature vector
BDI
Individual behaviour Transition probabilities
Decision rules
Cognitive model (Soar, ACT-R, etc.)
Basis of behaviour Own state
State of the environment
State of other individuals
Social states
Uniformity Uniform/non-uniform
Population Static/dynamic
Interaction model Form of interaction No interaction
Physical
Language-based
Interaction topology Static/dynamic
Neighbourhood/network
Environment model Spatial explicitness None
Relative positions
Absolute positions
Time Static/dynamic
Exogenous events Yes/no
Implementation Simulation engine Time-driven/event-driven
Programming MABS platform (NetLogo, Repast, etc.)
MAS platform (JADE, etc.)
From scratch (C, Java, etc.)
Distributedness Single computer/distributed
Further Reading
Gilbert and Troitzsch (2005) also have sections that describe the different kinds of
simulation available. Railsback and Grimm (2011) present a complementary anal-
ysis, coming from ecological modelling. The introductory chapters in (Gilbert and
Doran 1994) and (Conte and Gilbert 1995) map out many of the key issues and
aspects in which social simulation has developed.
References
Amblard F, Quattrociocchi W (2013) Social networks and spatial distribution. Chapter 16 in this
volume
Anderson JR et al (2004) An integrated theory of the mind. Psychol Rev 111(4):1036–1060
Barreteau O et al (2013) Participatory approaches. Chapter 10 in this volume
Bazzan ALC, Bordini RH (2001) A framework for the simulation of agents with emotions: report
on experiments with the iterated prisoners dilemma. In fifth international conference on
autonomous agents, Montreal. ACM Press, New York, pp 292–299
Bratman ME (1987) Intentions, plans, and practical reason. Harvard University Press, Cambridge,
MA
Broersen J, Dastani M, Huang Z, Hulstijn J, Van der Torre L (2001) The BOID architecture:
conflicts between beliefs, obligations, intentions and desires. In fifth international conference
on autonomous agents, Montreal. ACM Press, New York, pp 9–16
Chen D et al (2008) Large-scale agent-based simulation on the grid. Future Gen Comput Syst 24
(7):658–671
Conte R, Gilbert N (eds) (1995) Artificial societies: the computer simulation of social life. UCL
Press, London
David N (2013) Validating simulations. Chapter 8 in this volume
Davidsson P (2000) Multi agent based simulation: beyond social simulation. In: Moss S,
Davidsson P (eds) Multi agent based simulation (Lecture notes in computer science, 1979).
Springer, Berlin, pp 98–107
Dignum V (2013) Organisational design. Chapter 20 in this volume
Edmonds B, Lucas P, Rouchier J, Taylor R (2013) Human societies: understanding observed social
phenomena. Chapter 26 in this volume
Gardner M (1970) Mathematical games: the fantastic combinations of John Conway’s new
solitaire game “Life”. Sci Am 223(4):120–124
Geller A, Moss S (2013) Modeling power and authority: an emergentist view from Afghanistan.
Chapter 25 in this volume
Georgeff M, Pell B, Pollack M, Tambe M, Wooldridge M (1998) The belief-desire-intention
model of agency. In: Muller J, Singh M, Rao A (eds) Intelligent agents V (Lecture notes in
artificial intelligence, 1555). Springer, Berlin, pp 1–10
Gilbert N (2006) When does social simulation need cognitive models? In: Sun R (ed) Cognition
and multi-agent interaction: from cognitive modelling to social simulation. Cambridge
University Press, Cambridge, pp 428–432
Gilbert N, Doran J (eds) (1994) Simulating societies. UCL Press, London
Gilbert N, Troitzsch KG (2005) Simulation for the social scientist, 2nd edn. Open University
Press/McGraw Hill Education, Maidenhead
Gilbert N, Pyka A, Ahrweiler P (2001) Innovation networks: a simulation approach. J Artif Soc
Soc Simulat 4(3). http://jasss.soc.surrey.ac.uk/4/3/8.html
Grimm V et al (2006) A standard protocol for describing individual-based and agent-based

models. Ecol Model 198:115–126
Guye-Vuillème A (2004) Simulation of nonverbal social interaction and small groups dynamics in
virtual environments (Ph.D. thesis). Ècole Polytechnique Fédérale de Lausanne, No 2933
Hales D (2013) Distributed computer systems. Chapter 21 in this volume
Hemelrijk C (2013) Animal social behaviour. Chapter 22 in this volume
Janssen MA, Jager W (1999) An integrated approach to simulating behavioural processes: a case
study of the lock-in of consumption patterns. J Artif Soc Soc Simulat 2(2). http://jasss.soc.
surrey.ac.uk/2/2/2.html
Künzel J, Hämmer V (2006) Simulation in university education: the artificial agent PSI as a
teaching tool. Simulation 82(11):761–768
Le Page C et al (2013) Environmental management. Chapter 19 in this volume
Massaguer D, Balasubramanian V, Mehrotra S, Venkatasubramanian N (2006) Multi-agent
simulation of disaster response. In: Jennings NR, Tambe M, Ishida T, Ramchurn SD (eds)
First international workshop on agent technology for disaster management. Hakodate, 8 May
2006, pp 124–130. http://users.ecs.soton.ac.uk/sdr/atdm/ws34atdm.pdf
Méndez G, Rickel J, de Antonio A (2003) Steve meets jack: the integration of an intelligent tutor
and a virtual environment with planning capabilities. In: intelligent virtual agents (Lecture
notes on artificial intelligence, 2792). Springer, Berlin, pp 325–332
Newell A (1994) Unified theories of cognition. Harvard University Press, Cambridge, MA
Railsback SF, Grimm V (2011) Agent-based and individual-based modeling: a practical introduction.
Princeton University Press, Princeton
Railsback SF, Lytinen SL, Jackson SK (2006) Agent-based simulation platforms: review and
development recommendations. Simulation 82(9):609–623
Ramstedt L, Törnquist Krasemann J, Davidsson P (2013) Movement of people and goods.
Reynolds CW (1987) Flocks, herds, and schools: a distributed behavioural model. Comput Graph
21(4):25–34
Rouchier J (2013) Markets. Chapter 23 in this volume
Sawyer RK (2003) Artificial societies: multi-agent systems and the micro–macro link in sociolog-
ical theory. Sociol Methods Res 31(3):325–363
Schiff JL (2008) Cellular automata: a discrete view of the world. Wiley, Oxford
Schüle M, Herrler R, Klügl F (2004) Coupling GIS and multi-agent simulation: towards infra-
structure for realistic simulation. In: Lindemann G, Denzinger J, Timm IJ, Unland R (eds)
Multiagent system technologies, second German conference, MATES 2004 (LNCS, 3187).
Verhagen H (2001) Simulation of the learning of norms. Soc Sci Comput Rev 19(3):296–306
Williams R (1993) An agent based simulation environment for public order management training.
In: Western simulation multiconference, object-oriented simulation conference, Hyatt Regency,
La Jolla, 17–20 Jan 1993, pp 151–156
Part II
Methodology
Chapter 4
Informal Approaches to Developing
Simulation Models
Emma Norling, Bruce Edmonds, and Ruth Meyer
Why Read This Chapter? To get to know some of the issues, techniques and tools
involved in building simulation models in the manner that probably most people in the
field do this. That is, not using the “proper” computer science techniques of specifica-
tion and design, but rather using a combination of exploration, checking and
consolidation.
Abstract This chapter describes the approach probably taken by most people in the
social sciences when developing simulation models. Instead of following a formal
approach of specification, design and implementation, what often seems to happen in
practice is that modellers start off in a phase of exploratory modelling, where they
don’t have a precise conception of the model they want but a series of ideas and/or
evidence they want to capture. They then may develop the model in different
directions, backtracking and changing their ideas as they go. This phase continues
until they think they may have a model or results that are worth telling others about.
This then is (or at least should be) followed by a consolidation phase where the model
is more rigorously tested and checked so that reliable and clear results can be
reported. In a sense what happens in this later phase is that the model is made so
that it is as if a more formal and planned approach had been taken.
There is a danger of this approach: that the modeller will be tempted by apparently
significant results to rush to publication before sufficient consolidation has occurred.
There may be times when the exploratory phase may result in useful and influential
personal knowledge but such knowledge is not reliable enough to be up to the more
exacting standards expected of publicly presented results. Thus it is only with careful
consolidation of models that this informal approach to building simulations should
be undertaken.
E. Norling (*) • B. Edmonds (*) • R. Meyer

Manchester Metropolitan University, Centre for Policy Modelling, Business School, Oxford Road,
Manchester M15 6BH, United Kingdom
e-mail: [email protected]; [email protected]; [email protected]

40 E. Norling et al.
4.1 Introduction: Exploration and Consolidation

Modelling Phases
Formal approaches to the development of computer programs have emerged through the
collective experience of computer scientists (and other programmers) over the past half-
century. The experience has shown that complex computer programs are very difficult
to understand: once past a certain point, unless they are very careful, programmers lose
control over the programs they build. Beyond a certain stage of development, although
we may understand each part, each micro-step, completely we can lose our understand-
ing of the program as a whole: the effects of the interactions between the parts of a
program are unpredictable; they are emergent. Thus computer science puts a big
emphasis on techniques that aim to ensure that the program does what it is intended to
do as far as possible. However, even with the most careful methodology it is recognised
that a large chunk of time will have to be spent debugging the program – we all know that
a program cannot be relied on until it has been tested and fixed repeatedly.
However, it is fair to say that most computational modellers do not follow such
procedures and methodologies all the time (although since people don’t readily admit
to how messy their implementation process actually is, we cannot know this, just as
one does not know how messy people’s homes are when there are no visitors). There
are many reasons for this. Obviously those who are not computer scientists may simply
not know these techniques. Then there are a large number of modellers who know of
these techniques to some degree but judge that they are not necessary or not worth the
effort. Such a judgement may or may not be correct. Certainly it is the case that people
have a tendency to underestimate the complexity of programming and so think they
can get away with not bothering with a more careful specification and analysis stage.
There may also be times when there are good reasons not to follow such techniques.
A specification and design approach is simply not possible if you don’t have a very
clear idea of your goal. When modelling some target phenomena, one simply does not
know beforehand which parts of the system will turn out to be important to the
outcomes or even possible to computationally model. One of the big benefits of
modelling phenomena computationally is that one learns a lot about what is crucial
and possible in the process of building a simulation model. This is very unlike the case
where one has a functional goal or specification for a program that can be analysed into
sub-goals and processes etc. In (social) simulation, the degree to which formal
approaches are useful depends somewhat on the goal of modelling. If the goal is
very specific, for example understanding the effect of the recovery rate on the change
in the number of infections in an epidemic, and the basic model structure is known then
what is left is largely an engineering challenge. However, if the goal is general
understanding of a particular process then there is no possible way of systematically
determining what the model should be. Here the modelling is essentially a creative
process, and the development of the model proceeds in parallel with the development
of the understanding of the process; the model is itself a theory under development.
Thus what often seems to happen in practice is that modellers start off in a phase
of exploratory modelling, where they don’t have a precise conception of the model
4 Informal Approaches to Developing Simulation Models 41
they want but a series of ideas and/or evidence they want to capture. They then may
develop the model in different directions, backtracking and changing their ideas as
they go. This phase continues until they think they may have a model or results that
are worth telling others about. This then is (or at least should be) followed by a
consolidation phase where the model is more rigorously tested and checked so that
reliable and clear results can be reported. In a sense what happens in this later phase
is that the model is made so that it is as if a more formal and planned approach had
been taken.
There is nothing wrong with using an exploratory approach to model development.
Unfortunately, it is common to see models and results that are publicly presented
without a significant consolidation phase being undertaken. It is very understandable
why a researcher might want to skip the consolidation phase: they may have discov-
ered a result or effect that they find exciting and not wish to go through the relatively
mundane process of checking their model and results. They may feel that they have
discovered something that is of more general importance – however this personal
knowledge, that may well inform their understanding, is not yet of a standard that
makes it worthwhile for their peers to spend time understanding, until it has been more
rigorously checked.
One of the problems with the activity of modelling is that it does influence how the
modeller thinks. Paradoxically, this can also be one of the advantages of this approach.
After developing and playing with a model over a period of time it is common to “see”
the world (or at least the phenomena of study) in terms of the constructs and processes
of that model. This is a strong version of Kuhn’s “theoretical spectacles” (Kuhn 1969).
Thus it is common for modellers to be convinced that they have found a real effect or
principle during the exploration of a model, despite not having subjected their own
model and conception to sufficient checking and testing – what can be called modelling
spectacles. Building a model in a computer is almost always in parallel with the
development of one’s ideas about the subject being modelled. This is why it is almost
inevitable that we think about the subject in terms of our models – this is at once a
model’s huge advantage but also disadvantage. As long as one is willing to be aware of
the modelling spectacles and be critical of them, or try many different sets of modelling
spectacles, the disadvantage can be minimised.
Quite apart from anything, presenting papers with no substantial consolidation is
unwise. Such papers are usually painfully obvious when presented at workshops and
easily criticised by referees and other researchers if submitted to a journal. It is socially
acceptable that a workshop paper will not have as much consolidation as might be
required of a journal article, since the criticism and evaluation of ideas and models at a
workshop is part of its purpose, but presenting a model with an inadequate level of
consolidation just wastes the other participants’ time.
What steps then should modellers who follow such an informal approach take to
ensure that their model is sufficiently consolidated to present to a wider audience? The
first step is to ensure that they have a clear purpose for their model, as described below.
Secondly, the modeller must be careful to identify the assumptions that are made
during the construction of the model. Thirdly, the modeller must maintain control of
the model while exploring different possibilities. And fourthly – and this is perhaps the
Fig. 4.1 The exploration and

consolidation approach to Purpose Assumptions
model development
Exploration
Model
Consolidation
Simulation
Documentation
results
most difficult – the modeller must maintain an understanding of the model. The
following sections of this chapter discuss these points in more detail. Then there is
the all-important consolidation phase (which may proceed in parallel with the former
steps, rather than strictly sequentially), during which the modeller formalises the
model in order to ensure that the results are sound and meaningful. Figure 4.1
illustrates this approach to model building.
4.2 Knowing the Purpose of the Model
There are many possible purposes for constructing a model. Although some models
might be adapted for different purposes without too much difficulty, at any one time a
model will benefit from having a clear purpose. One of the most common criticisms
of modelling papers (after a lack of significant consolidation) is that the author has
made a model but is unclear as to its purpose. This can be for several reasons, such as:
• The author may have simply modelled without thinking about why. (For exam-
ple having vague ideas about a phenomenon, the modeller decides to construct a
model without thinking about the questions one might want to answer about that
phenomenon.)
• The model might have been developed for one purpose but is being presented as
if it had another purpose.
• The model may not achieve any particular purpose and so the author might be
forced into claiming a number of different purposes to justify the model.
The purpose of a model will affect how it is judged and hence should influence
how it is developed and checked.
The classic reason for modelling is to predict some unknown aspect of observed
phenomena – usually a future aspect. If you can make a model that does this for
unknown data (data not known to the modellers before they published the model) then
there can be no argument that such a model is (potentially) useful. Due to the fact that
predictive success is a very strong test of a model for which the purpose is prediction,
this frees one from an obligation as to the content or structure of the model.1
In particular the assumptions in the model can be quite severe – the model can be
extremely abstract as long as it actually does predict.
However, predictive power will not always be a measure of a model’s success.
There are many other purposes for modelling other than prediction. Epstein (2008)
lists 16 other purposes for building a model, e.g. explanation, training of practitioners
or education of the general public, and it is important to note that the measure of
success will vary depending on the purpose.
With an explanatory model, if one has demonstrated that a certain set of
assumptions can result in a set of outcomes (for example by exhibiting an acceptable
fit to some outcome data) this shows that the modelled process is a possible explanation
for those outcomes. Thus the model generates an explanation, but only in terms of the
assumptions in the setup of the simulation. If these assumptions are severe ones, i.e. the
model is very far from the target phenomena, the explanation it suggests in terms of
the modelled process will not correspond to a real explanation in terms of observed
processes. The chosen assumptions in an explanatory model are crucial to its purpose
in contrast to the case of a predictive model – this is an example of how the purpose of a
model might greatly influence its construction.
It does sometimes occur that a model made for one purpose can be adapted for
another, but the results are often not of the best quality and it almost always takes
more effort than one expects. In particular, using someone else’s model is usually
not very easy, especially if you are not able to ask the original programmer
questions about it and/or the code is not very well documented.
4.3 Modelling Assumptions
Whilst the available evidence will directly inform some parts of a model design, other
parts will not be so informed. In fact it is common that a large part of a simulation
model is not supported by evidence. The second source for design decisions is the
conceptions of the modeller, which may have come from ideas or structures that are
available in the literature. However, this is still not sufficient to get a model working.
In order to get a simulation model to run and produce results it will be necessary to add
in all sorts of other details: these might include adding in a random process to “stand in”
1
Of course a successfully predictive model raises the further question of why it is successful,
which may motivate the development of further explanatory models, since a complete scientific
understanding requires both prediction and explanation, but not necessarily from the same models
(Cartwright 1983).
for an unknown decision or environmental factor, or even be a straight “kludge”

because you don’t know how else to program something. Even when evidence supports
a part of the design there will necessarily be some interpretation of this evidence. Thus
any model is dependent upon a whole raft of assumptions of different kinds.
If a simulation depends on many assumptions that are not relatable to the object
or process it models, it is unlikely to be useful. However, just because a model has
some assumptions in it this does not mean it should be disregarded. Any modelling
is necessarily a simplification of reality, done within some context or other. Hence
there will be the assumption that the details left out are not crucial to the aspect of
the results deemed important, as well as those assumptions that are inherent in the
specification of the context. This is true for any kind of modelling, not just social
simulation. It is not sufficient to complain that a model has assumptions or does
simplify, since modelling without this is impossible; one has to argue or show why
the assumptions included will distort the results. (Equally, the author of a model
should be able to justify the assumptions that have been made.)
What one can do is to try to make the assumptions as transparent, as clear and as
explicit as possible. Thus future researchers will be better able to judge what the
model depends upon and adapt the model if any of the assumptions turns out to be
considered bad. The most obvious technique is to try to document and display the
assumptions. This not only helps to defend the model against criticism but also
helps one to think more clearly about the model.
Particularly in the early stages of constructing a model, it is common to make a
number of “assumptions” about various processes that are involved. In a sense these
are not strictly assumptions – they are just necessary simplifications made in order
to get something running – but nevertheless are included here. The model builder
might for example include a random term to substitute for an unknown process, or a
particular value might be chosen for a constant without knowing if it is a suitable
value. The modeller must carefully document such decisions and be prepared to
revisit them and adjust them as necessary.
The next type of assumption to consider is that which is “forced” by the constraints of
the programming system. This might be the simplification of a process due to computa-
tional power limitations, restrictions forced upon the modeller due to the data structures
and/or algorithms available, or the desire to reuse another (sub-)model. Again, such
decisions must be documented. While the modeller may feel that these decisions have
been forced, their documentation can serve two purposes. Firstly, other modellers may
have insights into the same programming system that will allow them to suggest
alternate approaches. Secondly, modellers who wish to replicate the model using an
alternate system may be able to better demonstrate the impact of these assumptions.
The third type of assumption to consider is the choice of relevant objects and
processes. As mentioned previously, any modelling exercise is necessarily an
abstraction, and one must leave out much of the detail of the real world. Of course
it is impractical to document every detail that has been omitted, but the modeller
should consider carefully which objects and processes may be relevant to the
model, and document those that have been included and those that have been
omitted. This documentation will then prove invaluable in the consolidation phase
(see Sect. 4.6), when the modeller should explicitly test these assumptions.
The most difficult type of assumption to track and document is that which derives
from the modeller’s own personal biases. For example, the modeller may have an
innate “understanding” of some social process that is used in the model without
question. The modeller may also have been trained within a particular school that
embraces a traditional set of assumptions. Such traditional assumptions may be so
deeply ingrained that they are treated as fact rather than assumption, making them
difficult to identify from within.
This final class of assumption may be difficult for the modeller to identify and
document, but all others should be carefully documented. The documentation can then
be used in the exploration and consolidation phases (see below), when the modeller checks
these assumptions as much as possible, refining the model as necessary. The assumptions
should also be clearly stated and justified when reporting the model and results.
4.4 Maintaining Control of the Model While Exploring
The second biggest problem in following the exploration and consolidation approach
to model building (after that of giving in to the temptation to promote your results
without consolidation) is that one loses control of the model while exploring, resulting
in a tangle of bugs. Exploration is an essential step, testing the impact of the
assumptions that have been made, but if not carefully managed can result in code
that achieves nothing at all. Bugs can creep in, to an extent that fixing one merely
reveals another, or the model can become so brittle that any further modifications are
impossible to integrate, or the model becomes so flaky that it breaks in a totally
unexpected manner. Although interactions between processes might be interesting
and the point of exploration, too much unknown interaction can just make the model
unusable. Thus it is important to keep as many aspects as possible under control as you
explore, so you are left with something that you can consolidate!
The main technique for maintaining control of a model is doing some planning
ahead and consolidation as you explore. This is a very natural way for a modeller to
work – mixing stages of exploration and consolidation as they feel necessary and as
matches their ambitions for the model. Each programmer will have a different balance
between these styles of work. Some will consolidate immediately after each bit of
development or exploration; some will do a lot of exploration, pushing the model to its
limits and then reconstruct a large part of the model in a careful and planned way. Some
will completely separate the two stages, doing some exploration, then completely
rebuild their ideas in a formal planned way but now having a better idea (if they are
correct) of: what they are aiming to achieve, what needs to go into the model (and what
not), what is happening in the model as it runs, and which results they need to collect
from it.
There is no absolute rule for how careful and planned one should be in developing a
model, but roughly the more complex and ambitious it is the more careful one should be.
Whilst a “quick and dirty” implementation may be sufficient for a simple model, for
others it is unlikely to get the desired results: it is too easy to lose understanding and
control of the interactions between the various parts, and also the model loses the
flexibility to be adapted as needed later on. At the other end of the spectrum, one can
spend ages planning and checking a model, building in the maximum flexibility and
modularity, only to find that the model does not give any useful results. This might be a
valuable experience for the programmer, but does not produce interesting knowledge
about the target phenomenon. This is the fundamental reason why exploration is so
important: because one does not know which model to build before it is tried. This is
particularly so for models that have emergent effects (like most of the ones discussed in
this volume), and also for those where there is no benchmark (either formal or observed)
against which to check them.
One important thing about the activity of modelling is that one has to be willing to
throw a lot of model versions away. Exploratory modelling is an inherently difficult
activity; most of the models built will either be the wrong structure or just not helpful
with regard to the phenomena we seek to understand. Further, the modelling is
constrained in many ways: in the time available for developing and checking them,
in the amount of computational resources they require, in the evidence available to
validate the model, in the necessary compromises that must be made when making a
model, and in the need to understand (at least to some extent) the models we make.
Thus the mark of a good modeller is that he or she throws away a lot of models and only
publishes the results of a small fraction of those he or she builds. There is a temptation
to laziness, to trying to ‘fix’ a model that is basically not right and thus save a lot of
time, but in reality this often only wastes time. This relates to the modelling spectacles
mentioned above: one becomes wedded to the structure one is developing and it takes a
mental effort to start afresh. However, if it is to be effective, a corollary of an
exploratory approach is being highly selective about what one accepts – junking a
lot of models is an inevitable consequence of this.
Whatever balance you choose between exploration and consolidation, it is probably
useful to always pause before implementing any key algorithm or structure in your
model, thinking a little ahead to what might be the best way. This is an ingrained habit
for experienced programmers but may take more effort for the beginner. The beginner
may not know of different ways of approaching a particular bit of programming and so
may need to do some research. This is why developing some knowledge of common
algorithms and data structures is a good idea. There is a growing body of work on
documenting programming ‘patterns’ – which seek to describe programming solutions
at a slightly general level – which can be helpful, although none of these pattern
catalogues have yet been written specifically with models of social complexity in mind
(but see Grimm et al. 2005 for examples from ecology). Increasingly too researchers
within this field are making their code, or at least descriptions of the algorithms used,
available to wider audiences.
There are dangers of using someone else’s code or algorithm though. There is the
danger of assuming that one understands an algorithm, relying on someone else’s
description of it.2 It is almost inconceivable that there will not be some unforeseen results
of applying even a well-known algorithm in some contexts. When it comes to reusing
code, the risk is even higher. Just as there are assumptions and simplifications in one’s
own code, so there will be in the code of others, and it is important to understand their
implications. Parameters may need adjustment, or algorithms tweaking, in order for the
code to work in a different context. Thus one needs to thoroughly understand at the very
least the interface to the code, and perhaps also its details. In some cases the cost of doing
this may well outweigh the benefits of reuse.
It is important to note that even though the approach presented here deviates from
formal approaches to software development, this does not mean one should ignore the
standard ‘good practices’ of computer programming. Indeed, due to the complexity of
even the simplest models in this field, it is advisable to do some planning and design
before coding. In particular, the following principles should be applied:
• Conceptualisation: any model will benefit greatly from developing a clear under-
standing of the model structure and processes before starting to program. This is often
called a conceptual model and usually involves some diagramming technique. While
computer scientists will tend to use UML for this purpose, any graphical notation that
you are familiar with will do to sketch the main entities and their relationships on
paper, such as mind maps or flow diagrams. Apart from helping a modeller to better
understand what the model is about this will form a valuable part of the overall model
documentation. See (Alam et al. 2010; appendix) for an example of using UML class
and activity diagrams.
• Modularity: it is not always possible to cleanly separate different functions or
parts of a model but where it is possible it is hugely advantageous to separate
these into different modules, classes or functions. In this way the interactions
with the other parts of your model are limited to what is necessary. It makes it
much easier to test the module in isolation, facilitates diagnostics, and can make
the code much simpler and easier to read.
• Clear structures/analogies: it is very difficult to understand what code does, and to
keep in mind all the relevant details. A clear idea or analogy for each part of the
simulation can help you keep track of the details as well as being a guide to
programming decisions. Such analogies may already be suggested by the
conceptions that the programmer (or others) have of the phenomena under study,
but it is equally important not to assume that these are always right, even if this was
your intention in programming the model.
• Clear benchmarks: if there is a set of reference data, evidence, theory or other model
to which the simulation is supposed to adhere this can help in the development of a
model, by allowing one to know when the programming has gone astray or is not
working as intended. The clearest benchmark is a set of observed social phenomena,
since each set of observations provides a new set of data for benchmarking.
2
Of course this danger is also there for one’s own programming: it is more likely, but far from
certain, that you understand some code you have implemented or played with.
Similarly if a part of the model is supposed to extend another model then restricting
the new model should produce the same outcomes as the original.3
• Self-documentation: if one is continuously programming a simulation that is not
very complex then one might be able to recall what each chunk of code does.
However when developing this type of simulation it is common to spend large
chunks of time focusing on one area of a model before returning to another. After
such a lapse, one will not necessarily remember the details of the revisited code,
but making the code clear and self-documenting will facilitate it. This sort of
documentation does not necessarily have to be full documentation, but could
include: using sensible long variable and module names, adding small comments
for particularly tricky parts of the code, keeping each module, class, function or
method fairly simple with an obvious purpose, and having some system for
structuring your code.
• Building in error checking: errors are inevitable in computer code. Even the best
programmer can inadvertently introduce errors to his or her code. Some of these
will be obvious but some might be subtle; difficult to isolate and time-consuming
to eliminate. Detecting such errors as early as possible is thus very helpful and
can save a lot of time. Including safeguards within your code that automatically
detect as many of these errors as possible might seem an unnecessary overhead,
but in the long run can be a huge benefit. Thus you might add extra code to check
that all objects that should exist at a certain time do in fact exist, or that a message
from one object to another is not empty, or that a variable that should only take
values within a certain range does stay within this range. This is especially
important in an exploratory development, where one might develop a section
of code for a particular purpose, which then comes to be used for another
purpose. In other words the computational context of a method or module has
altered.
There are also many techniques that computer scientists may exhort you to use
that are not necessarily useful, that may be more applicable to the development of
software with more clearly defined goals. Thus do evaluate any such suggested
techniques critically and with a large dose of common sense.
4.5 Understanding the Model
Understanding a model is so intertwined with controlling a model that it is difficult to

cleanly separate the two. You cannot really control a complex model if you do not at
least partially understand it. Conversely, you cannot deeply understand a model until
you have experimented with it, which necessitates being able to control it to a
considerable extent. However, since modelling complex social phenomena requires
3
What the “same outcomes” here means depends on how close one can expect the restricted new
model to adhere to the original, for example it might be the same but with different pseudo-random
number generators.
(at least usually and probably always) complex models, complete understanding and/
or control is often unrealistic. Nevertheless, understanding your model as much as is
practical is key to useful modelling. This is particularly true for exploratory modelling
because it is the feedback between trying model variations and building an under-
standing of what these variations entail that makes this approach useful.
Understanding one’s model is a struggle. The temptation is to try shallow
approaches by only doing some quick graphs of a few global measures of output,
hoping that this is sufficient to give a good picture of what is happening in a complex
social simulation. Although information about the relationship of the set-up of a
simulation and its global outcomes can be useful, this falls short of a full scientific
understanding, which must explain how these are connected. If you have an idea of
what the important features of your simulation model are, you might be able to design a
measure that might be suitable for illustrating the nature of the processes in your
model. However, a single number is a very thin indication of what is happening – this is
OK if you know the measure is a good reflection of what is crucial in the simulation, but
can tend to obscure the complexity if you are trying to understand what is happening.
To gain a deeper understanding, one has to look at the details of the interactions
between the parts of the simulation as well as the broader picture. There are two main
ways of doing this: case studies using detailed traces/records and complex visualisations.
A case study involves choosing a particular aspect of the simulation, say a
particular individual, object or interaction; then following and understanding it,
step-by-step, using a detailed trace of all the relevant events. Many programming
environments provide tracing tools as an inbuilt feature, but not all social simulation
toolkits have such a feature. In this latter case, the modeller needs to embed the
tracing into the model, with statements that will log the relevant data to a file for later
analysis. This “zooming in” into the detail is often very helpful in developing a good
understanding of what is happening and is well worth while, even if you don’t think
you have any bugs in your code. However, in practice many people seek to avoid this
mundane and slightly time-consuming task.
The second way to gain an understanding is to program a good dynamic visualisation
of what is happening in the model. What exactly is “good” in this context depends
heavily on the nature of the model: it should provide a meaningful view of the key
aspects of the model as the simulation progresses. Many social simulation toolkits
provide a range of visualisation tools to assist this programming, but the key is
identifying the relevant processes and choosing the appropriate visualisation for
them – a task that is not amenable to generic approaches. Thus you could have a 2D
network display where each node is an individual, where the size, shape, colour, and
direction of each node all indicate different aspects of its state, with connections drawn
between nodes to indicate interactions, and so on. A good visualisation can take a while
to design and program but it can crucially enhance the understanding of your
simulation and in most cases is usable even when you change the simulation set-
up. The chapter by Evans et al. (2013) in this volume discusses a range of
visualisation techniques aimed at aiding the understanding of a simulation model.
4.6 The Consolidation Phase
The consolidation phase should occur after one has got a clear idea about what simula-
tion one wants to run, a good idea of what one wants to show with it, and a hypothesis
about what is happening. It is in this stage that one stops exploring and puts the model
design and results on a more reliable footing. It is likely that even if one has followed a
careful and formal approach to model building some consolidation will still be needed,
but it is particularly crucial if one has developed the simulation model using an informal,
exploratory approach. The consolidation phase includes processes of: simplification,
checking, output collection and documentation. Although the consolidation phase has
been isolated here, it is not unusual to include some of these processes in earlier stages of
development, intermingling exploration and consolidation. In such circumstances, it is
essential that a final consolidation pass is undertaken, to ensure that the model is truly
robust.
Simplification is where one decides which features/aspects of the model you need
for the particular paper/demonstration you have in mind. In the most basic case, this
may just be a decision as to which features to ignore and keep fixed as the other features
are varied. However this is not very helpful to others because (a) it makes the code and
simulation results harder to understand (the essence of the demonstration is cluttered
with excess detail) and (b) it means your model is more vulnerable to being shown to be
brittle (there may be a hidden reliance on some of the settings for the key results).
A better approach is to actually remove the features that have been explored but turned
out to be unimportant, so that only what is important and necessary is left. This not only
results in a simpler model for presentation, but is also a stronger test of whether or not
the removed features were irrelevant.
The checking stage is where one ensures that the code does in fact correspond to the
original intention when programming it and that it contains no hidden bug or artefact.
This involves checking that the model produces “reasonable” outputs for both “standard”
inputs and “extreme” inputs (and of course identifying what “standard” and “extreme”
inputs and “reasonable” outputs are). Commonly this involves a series of parameter
sweeps, stepping the value of each parameter in turn to cover as wide a combination as
possible (limited usually by resources). When possible, the outputs of these sweeps
should be compared against a standard, whether that is real world data on the target
phenomenon, or data from a comparable (well-validated) model.
The output collection stage is where data from the various runs is collected and
summarised in such a way that (a) the desired results are highlighted and (b) sufficient
“raw” data is still available to understand how these results have been achieved. It
would be impractical to record the details of every variable for every run of the
simulation, but presenting results in summary form alone may hide essential details.
At the very least, it is essential to record the initial parameter settings (including
random seeds, if random numbers are used) so that the summary results may be
regenerated. It may also be informative to record at least a small number of detailed
traces that are illustrative of the simulation process (once one has determined which
parameter configurations produce “interesting” results).
Documentation is the last stage to be mentioned here, but is something that should
develop throughout the exploration and consolidation of a model. Firstly, as mentioned
above, the code should be reasonably self-documenting (through sensible naming and
clear formatting) to facilitate the modeller’s own understanding. Secondly, the
consolidated model should be more formally documented. This should include any
assumptions (with brief justifications), descriptions of the main data structures and
algorithms, and if third-party algorithms or code have been used, a note to their source.
This may seem like unnecessary effort, particularly if the modeller has no intention of
publicly releasing the code, but if questions arise some months or years down the track,
such documentation can be invaluable, even for the original author’s understanding.
Finally, the modeller must present the model and its results to a wider audience.
This is essential to the process of producing a model, since one can only have some
confidence that it has been implemented correctly when it has been replicated,
examined and/or compared to other simulations by the community of modellers.
The distribution of the model should include a description of the model with sufficient
detail that a reader could re-implement it if desired. It should present the typical
dynamics of the system, with example output and summaries of detailed output. The
relevant parameters should be highlighted, contrasting those deemed essential to the
results with those with little or no impact. The benchmark measurements should be
summarised and presented. To maximise a simulation’s use in the community the
simulation should be appropriately licensed to allow others to analyse, replicate and
experiment with it (Polhill and Edmonds 2007).
4.7 Tools to Aid Model Development
As indicated previously, there is now a variety of systems for aiding the development
of complex simulations. These range from programming-language-based tracing and
debugging tools through to frameworks designed explicitly for social simulation,
which include libraries of widely used patterns. Learning to use a particular system or
framework is a substantial investment and because of this, most people do not swap
from system to system readily once they have mastered one (even when an alternate
system may provide a far more elegant solution to a problem). Ideally, a modeller
would evaluate a range of systems when embarking on a new project, and decide
upon the most appropriate one for that project. In practice, most modellers simply
continue to use the same system as they have used on previous projects, without
considering alternatives. There is no simple answer as to which system is “best”. The
available options are constantly changing as new systems are developed and old ones
stop being supported. The type of modelling problem will influence the decision.
And indeed it is partly a personal decision, depending on the modeller’s own
personal style and preferences. However given that such an investment is involved
in learning a new system, it is a good idea to make this investment in one that will
have large payoffs; that will be useful for developing a wide range of models.
Systems for developing and testing simulations range from the very specific to
those that claim to be fairly generally applicable. At the specific end there are
simulators that are designed with a restricted target in mind – such as a grid-based
simulation of land use change (e.g. FEARLUS,4 Polhill et al. 2001, or SLUDGE,
Parker and Meretzky 2004) – where most of the structures, algorithms and outputs are
already built in. The user has some latitude to adapt the simulation for their own
modelling ends, but the ease with which one can make small changes and quickly get
some results may be at the cost of being stuck with inbuilt modelling assumptions,
which may not be appropriate for the task at hand. The specificity of the model means
that it is not easy to adapt the system beyond a certain point; it is not a universal system,
capable in principle, of being adapted to any modelling goal. Thus such a specific
modelling framework allows ease of use at the cost of a lack of flexibility.
At the other end of the spectrum are systems that aim to be general systems to
support simulation work; that can, at least in principle, allow you to build any
simulation that can be conceived. Such systems will usually be close to a computer
programming language, and usually include a host of libraries and facilities for the
modeller to use. The difficulty with this type of system is that it can take considerable
effort to learn to use it. The range of features, tools and libraries that they provide take
time to learn and understand, as does learning the best ways to combine these features.
Furthermore, even if a system in principle makes it possible to implement a modelling
goal, different systems have different strengths and weaknesses, making any particular
system better for some types of models, and less good for others. Thus modellers will
sometimes “fight the system,” implementing workarounds so that their model can be
implemented within the system in which they have invested so much time, when in fact
the model could more efficiently be implemented in an alternative system.
Between these two extremes lie a host of intermediate systems. Because they are
often open source, and indeed more specific modelling frameworks are commonly
built within one of these generic systems, it is usually possible (given enough time
and skill) to ‘dig down’ to the underlying system and change most aspects of these
systems. However the fundamental trade-offs remain – the more of a simulation
that is ‘given,’ the more difficult it will be to adapt and the more likely it is that
assumptions that are not fully understood will affect results.
Thus it is impossible to simply dictate which the best system is to use for
developing simulation models of social complexity; indeed there is no single system
that is best under all circumstances. However the sorts of questions one should
consider are clearer. They include:
• Clear structure: Is the way the system is structured clear and consistent? Are
there clear analogies that help ‘navigate’ your way through the various choices
you need to make? Is it clear how its structures can be combined to achieve more
complex goals?
• Documentation: Is there a good description of the system? Is there a tutorial to
lead you through learning its features? Are there good reference documents
4
http://www.macaulay.ac.uk/fearlus/.
where you can look up individual features? Are there lots of well-documented
examples you can learn from?
• Adaptability: Can the system be adapted to your needs without undue difficulty?
Is the way it is structured helpful to what you want to do? Are the structures
easily adaptable once implemented in your model? Does the system facilitate the
modularisation of your model so that you can change one aspect without having
to change it all?
• Speed: How long does it take to run a model? Speed of execution is particularly
important when a variety of scenarios or parameters need to be explored, or
when several runs are necessary per parameter configuration due to random
processes in the model.
• User community: Do many people in your field use the system? Are there active
mailing lists or discussion boards where you can ask for help? If you publish a
model in that system is it likely that it will be accessible to others?
• Debugging facilities: Does the system provide inbuilt facilities for debugging
and tracing your simulation? If not, are there perhaps generic tools that could be
used for the purpose? Or would you have to debug/trace your model by manually
inserting statements into your code?
• Visualisation facilities: Does the system provide tools and libraries to visualise
and organise your results? Are there dynamic visualisation tools (allowing one to
view the dynamics of the system as it evolves)? How quickly can you develop a
module to visualise the key outputs of a simulation?
• Batch processing facilities: Is there a means of running the model a number of
times, collecting and perhaps collating the results? Is it possible to automatically
explore a range of parameters whilst doing this?
• Data collection facilities: Are the results collected and stored systematically so
that previous runs can easily be retrieved? Is it possible to store them in formats
suitable for input into other packages (for example for statistical analysis, or
network analysis)?
• Portability: Is the system restricted to a particular platform or does it require
special software to run? Even if all your development will be done on one
particular machine, in the interests of reusability it is desirable to use a system
that will run on multiple platforms, and that is not dependent on specialised
commercial software.
• Programming paradigm: Different programming paradigms are more appropriate
to different types of modelling problems. If for example you think of things in
terms of “if-then” statements, a rule-based system might be the most appropriate
for your modelling. If instead you visualise things as series of (perhaps branching)
steps, a procedural one might be more appropriate. In practice most systems these
days are not purely one paradigm or another, but they still have leanings one way
or another, and this will influence the way you think about your modelling.
• Timing: How will time be handled in the simulation? Will it be continuous or
stepped, or perhaps event-driven? Will all agents act “at once” (in practice,
unless each agent is run on a separate processor they will be executed in some
sense sequentially, even if conceptually within the model they are concurrent),
or do they strictly take turns? Will it be necessary to run the simulation in real
time, or (many times) faster than real time?
Once one has considered these questions, and decided on the answers for the
particular model in mind, the list of potential systems will be considerably shortened,
and one should then be able to make an informed choice over the available options.
The temptation, particularly when one is beginning to write models, is to go for the
option that will produce the quickest results, but it is important to remember that
sometimes a small initial investment can yield long-term benefits.
4.8 Conclusion
It is easy to try and rationalise bad practice. Thus it is tempting to try and prove that
some of the more formal techniques of computer science are not applicable to
building social simulations just because one cannot be bothered to learn and master
them. It is true however that not all the techniques suggested by computer scientists
are useful in an exploratory context, where one does not know in advance precisely
what one wants a simulation to do. In these circumstances one has to take a looser
and less reliable approach, and follow it with consolidation once one has a more
precise idea of what one wants of the simulation. The basic technique is to mix bits
of a more careful approach in with the experimentation in order to keep sufficient
control. This has to be weighed against the time that this may take given one does
not know which final direction the simulation will take. There is a danger of this
approach: that the modeller will be tempted by apparently significant results to rush
to publication before sufficient consolidation has occurred. There may be times
when the exploratory phase may result in useful and influential personal knowledge
but such knowledge is not reliable enough to be up to the more exacting standards
expected of publicly presented results. Thus it is only with careful consolidation of
models that this informal approach to building simulations should be undertaken.
Further Reading
Outside the social sciences, simulation has been an established methodology for
decades. Thus there is a host of literature about model building in general. The biggest
simulation conference, the annual “Winter Simulation Conference”, always includes
introductory tutorials, some of which may be of interest to social scientists. Good
examples are (Law 2008) and (Shannon 1998).
For a comprehensive review of the currently existing general agent-based simula-
tion toolkits see (Nikolai and Madey 2009); other reviews focus on a smaller selection
of toolkits (e.g. Railsback et al. 2006; Tobias and Hofmann 2004; Gilbert and Bankes
2002).
The chapters on checking your simulation model (Galán et al. 2013), documenting
your model (Grimm et al. 2013) and model validation (David 2013) in this volume
should be of particular interest for anyone intending to follow the exploration and
consolidation approach to model development. However, if you would rather attempt a

more formal approach to building an agent-based simulation model, the subsequent
chapter (Jonker and Treur 2013) discusses one such approach in detail. You could also
consult textbooks on methodologies for the design of multi-agent systems, such as
(Luck et al. 2004), (Bergenti et al. 2004) or (Henderson-Sellers and Giorgini 2005).
After all, any agent-based simulation model can be seen as a special version of a multi-
agent system.
References
Alam SJ, Geller A, Meyer R, Werth B (2010) Modelling contextualized reasoning in complex societies
with “endorsements”. J Artif Soc Soc Simul 13(4). http://jasss.soc.surrey.ac.uk/13/4/6.html
Bergenti F, Gleizes M-P, Zambonelli F (eds) (2004) Methodologies and software engineering for
agent systems: the agent-oriented software engineering handbook. Kluwer, Boston
Cartwright N (1983) How the laws of physics lie. Clarendon, Oxford
Epstein J (2008) Why model? J Artif Soc Soc Simul 11(4). http://jasss.soc.surrey.ac.uk/11/4/12.html
Evans A, Heppenstall A, Birkin M (2013) Understanding simulation results. Chapter 9 in this volume
Galán J et al (2013) Detecting and avoiding errors and artefacts. Chapter 6 in this volume
Gilbert N, Bankes S (2002) Platforms and methods for agent-based modelling. Proc Natl Acad Sci
USA 99(3):7197–7198
Grimm V et al (2005) Pattern-oriented modeling of agent-based complex systems: lessons from
ecology. Science 310:987–991
Grimm V, Polhill G, Touza J (2013) Documenting simulations using ODD. Chapter 7 in this volume
Henderson-Sellers B, Giorgini P (eds) (2005) Agent-oriented methodologies. Idea Group, Hershey
Jonker C, Treur J (2013) A formal approach to building compositional agent-based simulations.
Kuhn T (1969) The structure of scientific revolutions. University of Chicago Press, Chicago
Law AM (2008) How to build valid and credible simulation models. In: Mason SJ et al. (eds)
Proceedings of the 2008 winter simulation conference, Miami, FL. http://www.informs-sim.org/
wsc08papers/007.pdf
Luck M, Ashri R, d’Inverno M (2004) Agent-based software development. Artech House, London
Nikolai C, Madey G (2009) Tools of the trade: a survey of various agent based modelling
platforms. J Artif Soc Soc Simul 12(2). http://jasss.soc.surrey.ac.uk/12/2/2.html
Parker DC, Meretsky V (2004) Measuring pattern outcomes in an agent-based model of edge-
effect externalities using spatial metrics. Agric Ecosyst Environ 101(2–3):233–250
Polhill JG, Edmonds B (2007) Open access for social simulation. J Artif Soc Soc Simul 10(3).
http://jasss.soc.surrey.ac.uk/10/3/10.html
Polhill G, Gotts N, Law ANR (2001) Imitative versus non-imitative strategies in a land-use
simulation. Cybern Syst 32(1–2):285–307
development recommendations. Simulation 82:609–623
Shannon RE (1998) Introduction to the art and science of simulation. In: Medeiros DJ, Watson EF,
Carson JS, Manivannan MS (eds) Proceedings of the 1998 winter simulation conference,
Washington, D.C. http://www.informs-sim.org/wsc98papers/001.PDF
Tobias R, Hofmann C (2004) Evaluation of free Java-libraries for social-scientific agent based
simulation. J Artif Soc Soc Simul 7(1). http://jasss.soc.surrey.ac.uk/7/1/6.html
Chapter 5
A Formal Approach to Building Compositional
Agent-Based Simulations
Catholijn M. Jonker and Jan Treur
Why Read This Chapter? To be introduced to a more formal “computer-science”

style of simulation design, especially suited to simulations of multi-level systems
(e.g. firms, departments, and people).
Abstract This chapter is an introduction to a more formal approach to designing
agent-based simulations of organisations (in the widest sense). The basic method is
the iterative refinement of structure, process and knowledge, decomposing each
abstraction into near-decomposable components that can be (for the most part) then
considered separately. Within this over all framework there are two complementary
approaches: designing the organisation first, and designing the individual agents first.
5.1 Introduction
This chapter outlines a more formal approach to designing an agent system, in this
case an agent-based simulation. Its approach comes from computer science, and
shows how one can develop a design for a simulation model in a staged and
cautious manner. This is particularly appropriate when the simulation is complex,
requiring the interaction of complex and intelligent entities, and the intended design
of the model is essentially known or accessible. Thus it contrasts with and
complements the previous chapter describing more informal and exploratory
approaches to developing simulations (Norling et al. 2013).
C.M. Jonker (*)

Man-Machine Interaction, Delft University of Technology, Mekelweg 4, Delft 2628 CD,
The Netherlands
J. Treur
Department of Artificial Intelligence, Vrije Universiteit Amsterdam, De Boelelaan 1081a,
Amsterdam 1081 HV, The Netherlands

58 C.M. Jonker and J. Treur
This chapter draws on approaches which were designed for engineering agent-
based software systems (e.g. air traffic control) and applies them to the engineering
of agent-based simulations since a simulation model is an example of a complex
piece of software. Such a cautious approach will result in the development of the
model taking more time and effort but also should result in a simulation model that
is easier to maintain and adapt and create fewer bugs, or simulations artefacts
(as described in Chap. 6, Galán et al. 2013).
The chapter is structured into three main sections, covering: compositional
design, organisations and agents. The section on compositional design is the most
general, and “works” for both the design of organisations and agents. However,
there is no need to reinvent the wheel – the sections on designing organisations and
agents raise more specific issues and aspects that experience has shown to be
helpful when designing simulations of these entities. Although they do not recapit-
ulate all in the section on compositional design they are essentially examples of that
general approach, though they each only concentrate on part of the overall process.
The differences are, roughly, that the organisational view starts with the
organisation, then the roles within that organisation, working “downwards”. It
highlights the constraints from above that organisations place on their members,
dealing with the nature of the agents after. The agent viewpoint starts with the
agents and their properties first and then moves on to how they interact (including
maybe as an organisation). Social phenomena often do involve both the “down-
ward” actions of constraint and “immergence” as well as the upwards actions of
emergence and collective outcomes. Thus both organisational and agent views are
often necessary to be explicitly considered in simulations.
5.2 Principles of Compositional Design of Multi-agent Systems
Although any principled design method can be used for the development of
simulations, we are concentrating in this chapter on those social systems which
are compositional in nature, since compositional actors are common in the social
world and this case also covers the principles in the simpler, non-compositional,
case. The principles described are quite general in nature, but will be illustrated
with respect to a particular compositional multi-agent development method, the
Design and Specification of Interacting Reasoning Components (DESIRE) (Brazier
et al. 1998).
The approach described here considers the design of autonomous interactive
agents and explicitly models both the intra-agent functionality and the inter-agent
functionality. Intra-agent functionality concerns the expertise required to perform
the tasks for which an agent is responsible in terms of the knowledge, and reasoning
and acting capabilities. The inter-agent functionality concerns the expertise
required to perform and guide co-ordination, co-operation and other forms of social
interaction in terms of knowledge, reasoning and acting capabilities.
5 A Formal Approach to Building Compositional Agent-Based Simulations 59
In the approach here described, both the individual agents and the overall system
are considered as compositional structures – hence all functionality is designed in
terms of interacting, compositionally structured components. Complex distributed
processes are the result of tasks performed by agents in interaction with their
environment.
The process starts at the most general, aggregate and abstract level, works in ever
increasing detail down through the specification of individual agents and finally the
code that determines their behaviour. At each stage there are a number of decisions
to be made, which correspond to specifying the features and properties that are
described below. The decisions at the higher levels help frame those made for the
next level, etc. until everything has been described. This should leave a documented
“trace” of the decisions that are made during the simulation design, that will help
with the description of the simulation and the explicit logging of assumptions made
during the design process.
Once this has been done and a roughly working simulation obtained, the verifi-
cation and validation procedure happens in the other direction, starting at testing
and validating the smallest processes and components and building upwards to
higher groups and units until the simulation as a whole has been checked.
5.2.1 The Design Process
The design of a multi-agent simulation is an iterative process, which aims at the

identification of the parties involved (i.e., human agents, system agents, external
worlds), and the processes, in addition to the types of knowledge needed. Initially
broad conceptual descriptions of specific processes and knowledge are attained.
Further explication of these conceptual design descriptions results in more detailed
design descriptions, most often in parallel with the development of the conceptual
design. During the design of these models, prototype implementations or parts or
sections of the overall model may be used to analyse or verify the resulting behaviour.
On the basis of examination of these partial prototypes, new designs and prototypes
are generated and examined, and so on and so forth. This evolutionary development
of systems is characteristic to the whole approach. Thus, as with any idealised design
methodology, in practice there is a lot of iterating back and forth between stages and
levels until a satisfactory design is obtained. The concepts presented in this chapter
are to help focus what might otherwise be an unstructured process.
We distinguish the following kinds of descriptions within the development
process:
• Problem description – Sect. 5.2.3
• Conceptual design – Sect. 5.2.4
• Detailed design – Sect. 5.2.4
• Design rationale – Sect. 5.2.5
• Operational design
The problem description is a description of the target system to be simulated and

the main issues and questions to be investigated, this usually includes the
requirements imposed on the design – what the simulation must do to be useful
in this regard. Starting from the problem description, the design rationale specifies
the choices made during each of the levels of the design process, the reasons for
those choices, and the assumptions behind those choices that will impinge on its
use. In other words, the design rationale is the strategy for “solving” the problem
description using the design along with the reasons and justification for that
strategy.
The actual design process roughly proceeds from conceptual and abstract, down
to the more concrete until one has almost written the simulation code itself. The
conceptual design includes conceptual models (the main design ideas and
structures) for each entity in the model: the organisation, its roles and groups, the
individual agents, the external world, the interaction between agents, and the
interaction between agents and the external world. In a sense a conceptual design
could apply as much to a human who will have to fulfil a role as a programmed
agent, it does not concern itself with exactly how these aspects are to be achieved,
but more about how it relates to other key structures. The detailed design of a
system, based on the conceptual design, specifies all aspects of a system’s knowl-
edge and behaviour. It describes how the agent’s processes will achieve its role.
This can be thought of as the step where one is thinking about how a computational
agent might achieve what is described in the conceptual design, but it is probably
independent of which computer language, or system the agent is destined to be
implemented in. A detailed design is an adequate basis for the operational design,
which deals with some of the nitty-gritty of a specific implementation in a particular
system. It stops short of actual programming code, but would be enough for a
programmer to implement the system. This final stage will not be discussed in this
chapter since it will be different for each programming language or system that is
used to implement the final simulation.
The sequence tends to progress roughly from the conceptual towards the con-
crete, however this is only a general rule; there is no immutable sequence of design:
depending on the specific situation, different types of knowledge are available at
different points during system design which means that stages need to be iterated or
even that, at times, lower level necessities might drive the higher levels of design.
5.2.2 Compositionality of Processes and Knowledge
Compositionality is a general principle that refers to the use of components to

structure a design. This process of composition can be extended downwards as far
as it is useful, for example, components can themselves be compositional structures
in which a number of other, more specific components are grouped. During the
design all the components at the different levels are identified. Processes at each of
these levels (except the lowest level) are modelled as (process) components
composed of entities at the level one lower to the process. Clearly this approach
depends upon the possibility of decomposing what is being modelled into separate
entities that are somewhat independent, so that the interaction of these components
can be considered in turn. In other words it is necessary that what is being modelled
is a near-decomposable system (Simon 1962). Although this is not always the case,
there are many social actors that, prima facie, are themselves composed of other
actors, for example political parties or firms. Thus this is often a reasonable
approach to take in the field of social simulation. Even when it is not clear that
what is being modelled does divide so neatly, then this method provides a system-
atic approach to attempting to identify and analyse those parts that are amenable to
design so that when they are all put together the desired behaviour will emerge from
their interaction during the simulation. Such emergence is an indication of non-
decomposability and can never be guaranteed – one has to find out by running the
simulation model, i.e. performing simulation experiments. If the desired behaviour
does not emerge then one has to try and work out the possible reasons for the failure
and go back to earlier stages of the design and rethink the approach.
In this compositional approach, each process within a multi-agent system may
be viewed as the result of interaction between more specific processes. A complete
multi-agent system may, for example, be seen to be one single component respon-
sible for the performance of the overall process. Within this one single component a
number of agent components within a common internal environment may be
distinguished, each responsible for a more specific process. Each agent component
may, in turn, have a number of internal components responsible for more specific
parts of this process. These components may themselves be composed, again
entailing interaction between other more specific processes.
The knowledge and information that is stored, produced, and communicated is
as important as the processes. The set of all terms, labels, types, structures etc. that
is used to encode and process the knowledge needed in a specific domain or for the
purposes of a particular agent may also be seen as a component, a knowledge
structure. This knowledge structure can be composed of a number of more specific
knowledge structures which, in turn, may again be composed of other even more
specific knowledge structures.
Compositionality of processes and compositionality of knowledge are two inde-
pendent dimensions of design. Thus a set of processes might be summarised as a
single process when appropriate or, in the other direction, broken down into a
system of sub-processes. The knowledge structures at one level might be adequate
for the purposes of the compositional processes or, maybe, a finer system of
description might be needed to be utilised by a finer grain of process representation.
For example, some simulations might represent the spread of beliefs through a
group as a simple contagion process, with simple entities representing the beliefs
and who has them. A finer grained model might include more of a cognitive
representation of the knowledge structures involved and how others are persuaded
to accept these as the result of a dialogue process.
Compositionality is a means to achieve information and process hiding within a
model: by defining processes and knowledge at different levels of abstraction,
unnecessary detail can be hidden at those stages, allowing the broader

considerations to be considered separately from the component details. Clearly in
the realm of social simulation being able to satisfactorily express social processes
without always going down to the details is necessary if the resulting model is to be
feasible – clearly we cannot simulate social actors, going all the way down to the
atoms they are made of. Compositionality also makes it possible to integrate
different types of components in one agent, providing the structure and means by
which they work together.
5.2.3 Problem Description
There are many ways to write a problem description. Techniques vary in their
applicability, depending on, for example, the situation, the task, or the type of
knowledge on which the system developer wishes to focus. Therefore, no particular
method will be described here. However, whichever way the problem description is
developed it is crucial to capture the key requirements to be imposed on the system –
that is, what one wants to gain from building and using the simulation. These
requirements are part of the initial problem definition, but may also evolve during
the development of a system. Different simulations of the same phenomena might
well be appropriate because each might have different requirements. For example, a
simulation model to predict where queues will form on a certain stretch of motor-
way will probably be different from one to predict whether different proportions of
lorries might affect the throughput of traffic, even if both simulations are of traffic
on the same stretch of road at the same times.
5.2.4 Conceptual and Detailed Design
A conceptual and detailed design consists of specifications of the following three

types:
• Process composition;
• Knowledge composition;
• The relation between process composition and knowledge composition.
These are discussed in turn in more detail below.
5.2.4.1 Process Composition
Process composition identifies the relevant processes at different levels of (process)

abstraction, and describes how a process can be defined in terms of lower level
processes. Depending on the context in which a system is to be designed two
different approaches can be taken: a task perspective, or a multi-agent perspective.
The task perspective refers to the approach, in which the processes needed to
perform an overall task are distinguished first, which are then delegated to appro-
priate agents and the external world. In this approach the agents and the external
world are designed later. The multi-agent perspective refers to the approach in
which agents and an external world are distinguished first and afterwards the
processes within each agent and within the external world.
Identification of Processes at Different Levels of Abstraction
Processes can be described at different levels of abstraction; for example, the

processes for the multi-agent system as a whole, processes within individual agents
and the external world, processes within task-related components of individual
agents. Thus in a traffic simulation system processes might include the introduction
and removal of vehicles, the collection of statistics and the visualisations of the
system state; individual agents representing vehicles might have processes for
monitoring their speed and for deciding when to change lane; within these agents
might be a reactive component that monitors and adjusts the speed reacting when
other traffic gets too close, a learning component that remembers which lanes were
faster in the past, and a reasoning component that decides when to change lanes.
Relevant Aspects of a Process
The processes identified are modelled as components. For each process the types of
information used by it for input and resulting as output are identified and modelled
as input and output interfaces of the component (an interface is a protocol for the
information and maybe some specification of a process to deal with it, translating or
storing it). So in a traffic simulation the process in an agent may need the distance
and the relative speed of any object in its path to be passed to it as an input and the
reactions (accelerating, braking) may be passed as its outputs. Clearly in a simple
simulation these interfaces will be trivial, but in more complex simulations or
where a degree of modularity is required some effort in designing these interfaces
might well pay off.
Modelling Process Abstraction Levels
Each level of process is either an abstraction of lower levels of component and/or a

specialisation of the levels above. These layers of abstraction only go down so far
since processes are either composed of other components or they may be primitive.
Primitive components may be either reasoning components (for example based on a
knowledge base), or, alternatively, components capable of performing tasks such as
calculation, information retrieval, optimisation, etc.
The identification of processes at different levels of abstraction results in the

specification of components that can be used as building blocks, and which
components are sub-components of which other component. The distinction of
different levels of process abstraction results in hiding detail from the processes
at the higher levels. Thus a process to decide whether to change lane in the traffic
example might be composed of a process to access the memory of how fast each
lane was in the past, an estimate of the average speed of the current lane, and how
fast the traffic ahead is moving.
Composition
The way in which processes at one level of abstraction in a system are composed of
processes at the adjacent lower abstraction level in the same system is called
composition. This composition of processes is described not only by the
component/sub-component relations, but in addition by the (possibilities for) infor-
mation exchange between processes (the static aspects), and task control knowledge
used to control processes and information exchange (the dynamic of the
composition).
Information Exchange
A specification of information exchange defines which types of information can be

transferred between components and the ways by which this can be achieved, called
information links. Within each of the components private information links are
defined to transfer information from one component to another. In addition,
mediating links are defined to transfer information from the input interfaces of
encompassing components to the input interfaces of the internal components, and to
transfer information from the output interfaces of the internal components to the
output interface of the encompassing components. That is the mediating links are
those which pass information “up” and “down” the structure of the agent: to their
components or up to the entity that they are a component of. Thus in the traffic
example there might well be mediating information links from each vehicle up to
the simulation to pass information about its current speed and position.
5.2.4.2 Knowledge Composition
Knowledge composition identifies the knowledge structures at different levels of

abstraction, and describes how a knowledge structure can be defined in terms of
lower level knowledge structures. The levels of knowledge abstraction may corre-
spond to the levels of process abstraction, but this is not necessarily the case.
Identification of Knowledge Structures at Different Abstraction Levels
The two main structures used as building blocks to model knowledge are: informa-
tion types and knowledge bases. These knowledge structures can be identified and
described at different levels of abstraction. At the higher levels the details can be
hidden. The resulting levels of knowledge abstraction can be distinguished for both
information types and knowledge bases.
Information Types
An information type defines the sorts of terms that will be used describe objects or
other terms, their kinds, and the relations or functions that can be defined on these
objects.1 Information types can be specified in graphical form, or in formal textual
form. Thus the speed of objects is a type of knowledge in the traffic example,
relatable to other speeds in terms of relative speed.
Knowledge Bases
Knowledge bases are structured collections of information held by agents. The
specification of the knowledge bases use the information types just described. To
specify a knowledge base one needs to say which information types are used in as
well as the relationships between the concepts specified in the information types.
Thus in a (somewhat complex) driver memory there might be three kinds of
information: days of the week, times of the day, lane label and categories of
speed. Each lane might relate to a set of past occasions composed of day of the
week, time of day and speed category.
Composition of Knowledge Structures
Information types can be composed of more specific information types, following

the principle of compositionality discussed above. Similarly, knowledge bases can
be composed of more specific knowledge bases. Thus in the example of memory
about past lane speeds the sets of past occasions might be a list of limited size
ordered first by level of annoyance and secondly by recency.
5.2.4.3 Relation Between Process Composition and Knowledge Composition
Each process in a process composition uses knowledge structures. These will be

involved in the building and maintenance of the knowledge structures at their level,
but could involve knowledge structures from higher or, occasionally, lower levels.
1
Such sets of agreed terms are often called an “ontology” in computer science.
So the processes that comprise the cognitive processes of a traffic agent might well
be involved in maintaining the memory of past lane speeds but might also relate to
the positional clues that are associated with the highest levels of the simulation.
5.2.5 Design Rationale
The design rationale is important because it makes explicit the reasoning “glue”
that underpins some of the other parts. Essentially it answers the question “given the
modelling target and goals why have the design decisions been made?” Thus it
describes the relevant properties of the design in relation to the requirements
identified in the problem description. It also documents the verification of the
design – that is how one will check that the implemented system does in fact
meet its specification, including the assumptions under which the desired properties
will hold. All the important design decisions are made explicit, together with some
of the alternative choices that could have been made, and the arguments in favour of
and against the different options. At the operational level the design rationale
includes decisions based on operational considerations, such as the choice to
implement an agent’s cognitive process in a particular way in order to make the
simulation run at a reasonable speed.
5.2.6 Multi-agent Systems in the Simulation of Social Phenomena
The method described above deals with the design process in terms of components
and the interactions between those components. In this light, multi-agent systems
are not considered specifically. However, in the context of simulating social
phenomena, it comes out naturally that in many instances the appropriate
“components” are the correlates of observed social actors. In other words it is
almost always overwhelmingly sensible to model the target system as a multi-agent
system, where agents in the model are representations of the actors (people, firms,
etc.) that are known to exist. In a sense, in the social sphere almost everything is an
agent, or conversely, agents are nothing special. It is simply that a component that is
naturally thought of as having elements of cognition (learning, reasoning etc.) is an
agent and will be endowed, as part of the simulation process, with many of the
attributes that agents are expected to have (and are discussed later in this chapter).
Representations of humans in a simulation will not include all aspects of cognition
but, dependent on the modelling goals, might well be much simpler. On the other
hand some non-human components, such as a firm, might be represented as an
agent, being able to learn, react, and reason in an agent-like way.
Simulations are growing in complexity, not in the least because agents are asked
to fulfil different roles over time, and to change their behaviour according to both
their own internal learning mechanisms and changing role descriptions. Within the
described approach it has become good practice to first design the organisation, and
then the agents and their interaction in such a way that the agents realize the
organisation. The next section explicitly considers organisations. The chapter on
“Assessing Organisational Design” (Dignum 2013) follows the application of the
ideas that are described here.
5.3 Organisations
The organisational approach to simulation design takes the observed and inferred
organisational structures as the starting point and considers individual action and
agency at a later stage. This is particularly suitable for situations that seem to be
structured in this way, that is to say the roles and the requirements significantly
characterise and constrain individual action. Clearly in many observed cases there
is a complex mix of organisational constraint and emergence from individual action
so the decision to adopt a primarily organisational approach is a pragmatic one. In
many cases a mixture of organisation-based and agent-based approaches will be
necessary.
Societies are characterised by complex dynamics involving interaction between
many actors and groups of actors. If such complex dynamics take place in an
completely unstructured, incoherent manner, then the actors involved will probably
not be able to predict much, and not able to use and exploit any knowledge that they
have in a useful way. However in many social situations this is not the case, social
phenomena are full of structure, and even in initially unstructured situations social
actors will often quickly develop norms, rules, habits etc. – effectively creating
structure. Some sociologists (e.g. Luhman) have suggested that the purpose of
human institutional structure is to manage the complexity, in other words to
simplify social action and make planning possible. Organisational structure
provides co-ordination of the processes in such a manner that the agents involved
can function in a more adequate manner. The dynamics in many organisational
structures are much more dependable and understood than in apparently entirely
unstructured situations.
One key assumption of the organisational approach to simulation design is that
the organisational structure itself is relatively stable, i.e., the structure may change,
but the frequency and scale of change are assumed low compared to the more
standard dynamics through the structure. Within the field of Organisation Theory
such organisational structures regulating societal dynamics are studied (see e.g.
Kreitner et al. 2001; Mintzberg 1979). In summary, organisational structure is used
to help specify the dynamics (or organisational behaviour) of a desired type. A
crucial issue for further analysis is how exactly structure is able to affect dynamics.
A number of organisation modelling approaches have been developed to simu-
late and analyse dynamics within organisations in society (e.g. Ferber and
Gutknecht 1998; Hannoun et al. 1998, 2000; Hübner et al. 2002a b; Lomi and
Larsen 2001; Moss et al. 1998; Prietula et al. 1997). Some of these approaches
explicitly focus on modelling organisational structure, abstracting from the detailed

dynamics. Other approaches put less emphasis on organisational structure but focus
on the dynamics in the sense of implementing and experimenting with simulation
models. Often these simulation models are based on some implementation environ-
ment and not specified in an implementation-independent manner using a formally
defined conceptual language. However, there are some exceptions to this where the
specification approach is supported by an implementation framework.2 The Agent/
Group/Role (AGR) approach (previously called Aalaadin) introduced in (Ferber
and Gutknecht 1998) is a good example of the organisational approach. It focusses
on organisational structure, abstracting from the details of the dynamics. It helps
define a formal relation between the dynamic properties and the organisational
structure (Ferber et al. 1999, 2000). The relevance for this chapter is that it shows
how dynamics of the organisational structure itself can be modelled: agents can
dynamically create, join, or quit groups. This is particularly relevant for simulating
situations where the organisational structure is somewhat fluid.
In this section the “dynamics specification” approach exemplified by AGR is
presented.3 The organisational structure is discussed and its parts defined. Then the
dynamics of the organisation is discussed in terms of dynamic properties that can be
associated to each element of the organisational structure. These dynamic
properties can help the simulation and analysis of empirical or simulated traces.
The various compositional levels within an organisation are related to the
organisational dynamics via a series of relationships. Finally, as a prerequisite to
realising an organisation the requirements of the agents are specified from their
roles within the organisation model.
5.3.1 Specification of Organisation Structure
In this approach, an organisation is viewed as a framework for activity and

interaction through the definition of groups, roles and their relationships. By
avoiding an agent-oriented viewpoint, an organisation is regarded as a structural
relationship between agents. In this way the organisation is described solely on the
basis of its structure, i.e. by the way groups and roles are arranged to form a whole,
without being concerned with the way agents actually behave. That is the systems
will be analysed from the outside, as a set of interaction modes. The specific
architecture of the agents is purposely not addressed in the organisational model.
2
The Strictly Declarative Modelling Language SDML (Moss et al. 1998) and the use of the agent-
oriented modelling approach DESIRE in social simulation as presented in (Brazier et al. 2001) are
two examples.
3
For more information on the use of AGR, see (Jonker and Treur 2003).
The three primitive definitions are:

• The agents. The model places no constraints on the internal architecture of
agents. An agent is only specified as an active communicating entity which
plays roles within groups. This agent definition is intentionally general to allow
agent designers to adopt the most accurate definition of agent-hood relative to
their application. In other words, the specification of the agent is left as flexible
as possible, given the organisational constraints upon its roles.
• Groups are sets of agents. Each agent is part of one or more groups. In its most
basic form, the group is only a way to tag a set of agents. An agent can be a
member of several groups at the same time. A major point of these groups is that
they can freely overlap.
• A role is an abstract representation of an agent function, service or identification
within a group. Each agent can have multiple roles and each role handled by an
agent is local to a group. Roles could be assigned beliefs; that is, they could reason
about whether they should have a particular belief given a certain role. These
beliefs can be seen as an additional requirement on the agents playing that role.
Organisation structure is often show as a diagram (for example, as kind of
labelled graph; see Fig. 5.3 in Sect. 5.3.5) consisting of roles, groups, and
interactions, and of relationships between these elements.
Within AGR an organisation structure consists of a set of groups, the roles in
each group and the agents fulfilling those roles. To complete the picture
relationships between roles can be specified.
5.3.2 Organisation Structure
An AGR specification of an organisation structure is defined by the following:

groups, roles, (intergroup) interactions, transfers (intra-group interactions), which
roles are in which groups, the roles that are the source of interactions, the roles that
are the destination of interactions, the roles that are the source of transfers, and the
roles that are the destination of transfers. Transfers, under this scheme, are within a
group as opposed to interactions which may be between groups. Thus it is necessary
that the source and destination of all transfers belong to the same group. Although
intergroup interactions are defined above as between two roles, this can easily be
generalised to intergroup interactions involving more than two roles.
5.3.3 Dynamic Properties of an Organisation
After the foundation of an organisation structure has been defined, the foundations
for specification of dynamic properties in an organisation are addressed. The aim is
not only to cover simple types of dynamics, such as simple reactive behaviour, but
also more complex dynamics, necessary for the simulation of realistic

organisations. The challenge here is to incorporate somehow the organisational
structure within the formal description of the organisation’s internal dynamics. To
this aim, the following approach is introduced:
For each element within the organisational structure characterise its dynamics by a
specific set of dynamic properties.
This is based on the structural relations between elements in an organisational
structure. Then:
Identify relationships between the sets of dynamic properties corresponding with
these elements;
In general, the dynamics of an element within an organisation structure can be
characterised by describing how the states of the elements change over time. For a
role the ‘state’ needs to include descriptions of for both the input and the output of
the role. Transfers and intergroup interactions are assumed to operate only on input
and output states of roles. These roles do not have their own internal state, so no
further state is needed to be described for such transfers and intergroup interactions.
An organisational structure defines relations between different elements in an
organisation. The dynamics of these different elements are characterised by their
dynamic properties. An organisational structure has the aim of keeping the overall
dynamics of the organisation manageable. For this reason the structural relations
between the different elements within the organisational structure have to somehow
impose constraints on or dependencies between their dynamics. Logical relations
defined between sets of dynamic properties allow the use of logical methods to
analyse, verify and validate organisation dynamics in relation to organisation
structure. Within AGR organisation models three aggregation levels are involved:
• The organisation as a whole (the highest aggregation level)
• The level of a group
• The level of a role within a group
A general pattern for the dynamics in the organisation as a whole in relation to
the dynamics in groups is as follows:
Dynamic properties for the groups AND dynamic properties for intergroup role
interaction
⇨ Dynamic properties for the organisation
Moreover, dynamic properties of groups can be related to dynamic properties of
roles as follows:
Dynamic properties for roles AND dynamic properties for transfer between roles
⇨ Dynamic properties for a group
The idea is that these are properties dynamically relating a number of roles
within one group.
Fig. 5.1 Overview of inter-

level relations between
dynamic properties within an
AGR organisation model
Table 5.1 Types of dynamic Property type Relating

properties for an AGR
organisation model Role r Role r input Role r output
Transfer from r1 to r2 Role r1 output Role r2 input
Group G Input or output of roles in G
Intragroup interaction Role r1 output Role r2 output
Intergroup interaction Role r1 input Role r2 output
Organisation Input or output of roles in O
An overview of the logical relationships between dynamic properties at different

aggregation levels is depicted as an AND-tree in Fig. 5.1.4
To define states the notion of state property is useful. The notion of trace as a
sequence of states over a time frame is used to formalise the dynamics. To formally
specify dynamic properties that are essential in an organisation, an expressive
language is needed. One can do this using a formal language,5 however in this
chapter this will not be used to retain its accessibility.
We distinguish five kinds of dynamic properties that might be described during a
specification (or at least thought about). Not all of these are always necessary. A
summary of them is displayed in Table 5.1.
5.3.3.1 Role Dynamic Properties
The role dynamic properties relate input to output of that role. This is a subset of the
dynamic properties of that role; it is a concern of that role only. For example, the
gossip role behaviour: ‘whenever somebody tells you something, you will tell it to
everybody else’ is expressed in terms of input of the role leading to output of the
role in a reactive manner.
4
For formalisation details of the logical relationships put forward above, see (Jonker and Treur
2003).
5
E.g. the Temporal Trace Language (TTL), which defines the dynamics in terms of a “leads to”
relation (Jonker et al. 2001). A specification of dynamic properties in leads to format has as
advantages that it is executable and that it can often easily be depicted graphically.
5.3.3.2 Transfer Dynamic Properties
Transfer properties relate output of the source roles to input of the destination roles.
That is they represent the dynamic properties of transfers from one role to another.
Typically, these sets contain properties such as: information is indeed transferred
from source to destination, transfer is brought about within x time, arrival comes
later than departure, and information departs before other information also arrives
before that other information.
5.3.3.3 Group Dynamic Properties
Group dynamic properties relate input and/or output of roles within a group, it
relates the roles within the group. An example of a group property is: “if the
manager asks anyone within the group to provide the secretary with information,
then the secretary will receive this information”.
A special case of a group property is an intragroup interaction relating the
outputs of two roles within a group. A typical (informal) example of such an
intragroup interaction property is: “if the manager says ‘good afternoon’, then the
secretary will reply with ‘good afternoon’ as well”. Other examples may involve
statistical information, such as “3 out of the 4 employees within the organisation
never miss a committed deadline”.
5.3.3.4 Intergroup Interaction Dynamic Properties
Intergroup interaction properties relate the input of the source role in one group to
the output of the destination role in another group. Note that intergroup interaction
is specified by the interaction of roles within the group, and not the groups
themselves. Sometimes there are specialist roles for such intergroup interaction.
For example, a project leader is asked by one of the project team members (input of
role ‘project leader’ within the project group) to put forward a proposal in the
meeting of project leaders (output of role ‘member’ within the project leaders
group).
5.3.3.5 Organisation Dynamic Properties
Organisation dynamic properties relate to input and/or output of roles within the
organisation. A typical (informal) example of such a property is: “if within the
organisation, role A promises to deliver a product, then role B will deliver this
product”.
The different types of dynamic properties all relate to different combinations of
input and output. Table 5.1 provides an overview of these combinations. Note that
with respect to simulation, the above dynamics definition can contain elements that
are redundant: a smaller subset of dynamical properties could form an executable

specification of the dynamics of an AGR type organisation – not all of the above are
always needed.
For example, an organisation could be simulated on the basis of the role dynamic
properties, the transfer dynamic properties and the intergroup interactions. The
group dynamic properties, including the intragroup role interaction properties, and
the organisation properties should emerge in the execution. However specifying
them in advance can be used to check what is expected and help verify the
simulation.
In order to make an executable organisation model the dynamical properties
need to be chosen from those properties that can be executed.
5.3.4 Organisation Realisation
In this section criteria are discussed when allocation of a set of agents to roles is
appropriate to realize the organisation dynamics, illustrated for the AGR approach.
One of the advantages of an organisation model is that it abstracts from the specific
agents fulfilling the roles. This means that all dynamic properties of the
organisation remain the same, independent of the particular allocated agents.
However, the behaviours of these agents have to fulfil the dynamic properties of
the roles and their interactions that have been already specified. The organisation
model can be (re)used for any allocation of agents to roles for which:
• For each role, the allocated agent’s behaviour satisfies the dynamic role
properties,
• For each intergroup role interaction, one agent is allocated to both roles and its
behaviour satisfies the intergroup role interaction properties, and
• The communication between agents satisfies the respective transfer properties.
To satisfy the relationships specified above there needs to be a relevant overlap
between the agent’s ontologies and the role ontologies,6 i.e. there must be some
common referents so that the interactions of the agents are well defined with respect
to their roles. Moreover, note that if one agent performs two roles in the group then
dynamic properties of communication from itself to itself are required, i.e. that it
will receive (at its input state) what it communicates (at its output state): ‘it hears
itself talking’. The logical relationships can be depicted as in the extension of
Fig. 5.1 shown as Fig. 5.2.
Alternatively, if the roles in an intergroup interaction would not be fulfilled by
one agent, but by several, this would create a mystery, since input to one agent
creates output for another agent, even though the agents are not connected by any
6
For a more detailed discussion on this issue, see (Sichman and Conte 1998).
Fig. 5.2 Inter-level relations

between dynamic properties
for a realised organisation
model
transfer since the roles they fulfil are from separate groups. This would suggest that
the organisation structure is not complete. The whole idea of specifying the
organisational approach through roles is that all communication and interaction is
somehow made explicit – in an AGR organisation model it is assumed that the roles
in an intergroup interaction are fulfilled by one agent.
5.3.5 Organisational Example
Here the organisational approach to simulation specification is illustrated. This

shows how an organisation that is the target of simulation can be analysed into
groups, roles and processes. This analysis can then be the basis for the design of the
simulation implementation and finally its code. As described this is essentially a
top-down analytic approach (in contrast to the more bottom-up agent approach
described at the end of this chapter).
In this example, a factory and some of its components are considered. This
factory is organised at the highest level according to two divisions: the division that
produces certain components (division A) and the division that assembles these
components into products (division B). At the lower level, division A is organised
in two departments: the work planning department for division A (dep. A1) and the
component production department (dep. A2). Similarly, division B is organised in
two department roles: one for assembly work planning (dep. B1) and one for
product production (dep. B2). This example is illustrated in Fig. 5.3.
Here the two divisions are modelled as groups (depicted by the larger ovals),
with the departments as their roles (depicted by smaller ovals within larger ones). A
third group, the Connection Group C, models the communication between the two
divisions. This group consists of the two roles ‘division A representative’ and
‘division B representative’. Intergroup role interactions (depicted by pairs of
dashed lines) are modelled between the role ‘department A1’ in the division A
Connection Group C
divA rep divB rep
dep A2 dep B2
dep A1 dep B1
Group divA Group divB
Fig. 5.3 Organisational example. The smaller ovals indicate roles, the bigger ovals groups.
Connections are indicated by the two types of lines (dashed indicates an intergroup interaction,
solid arrow indicates a transfer). Membership of a role to a group is indicated by drawing the
smaller role oval within the bigger group oval
group and the role ‘division A representative’ within the connection group, and
between the role ‘department B1’ in the division B group and the role ‘division B
representative’ within the connection group. Intragroup role transfers model com-
munication between the two roles within each of the groups (depicted by the
arrows).
Connections have destination roles (indicated by the arrow points) and source
roles (where the arrow originates). Based on the semantic structures of many-sorted
predicate logic a more precise formal definition is the following.
5.3.5.1 Groups and Roles in Organisational Example
The example has the following groups, roles, and relationships between them:
• Groups ¼ {divA, divB, C},
• Roles ¼ {depA1, depA2, depB1, depB2, divArep, divBrep},
• Intergroup_interactions ¼ {iAC, iCA, iBC, iCB}
• Transfers ¼ {tA12, tA21, tB12, tB21},
• Some of the relationships are:
Within divA Organisation level

Role_in(depA1, divA) Source_of_interaction(divA, iAC)
Role_in(depA2, divA) Destination_of_interaction(C, iAC)
Source_of_transfer(depA1, tA12) Source_of_interaction(C, iCA)
Destination_of_transfer(depA2, tA12) Destination_of_interaction(divA, iCA)
Source_of_transfer(depA2, tA21)
Destination_of_transfer(depA1, tA21)
5.3.5.2 Dynamic Properties in Organisational Example
To get the idea, consider the special case of an intragroup role interaction from role
r1 to role r2, characterised by dynamic properties that relate output of one role r1 to
output of another role r2. Assuming that transfer from output of r1 to input of r2 is
adequate and simply copies the information, this property mainly depends on the
dynamics of the role r2. Therefore in this case the relationship has the form:
Dynamic properties for role r2 AND
Dynamic properties for transfer from role r1 to role r2
⇨ Dynamic properties of intragroup interaction from r1 to r2
Role Dynamic Properties
DP(depA1) Progress Information Generates Planning in depA1

If within division A department A1 receives progress information on component
production,then an updated planning will be generated by department A1 taking
this most recent information into account.
Group Dynamic Properties
DP(A) A Progress Information Generation

This property is the conjunction of the following two properties.
DP1(A) Initial A Progress Information Generation
Department A1 receives initial progress information on component production
processes, involving already available components.
DP2(A) Subsequent A Progress Information Generation
Within the division A group, for any component production planning gene-
rated by department A1, incorporating a specific required set of components,
progress information on the production of these components will be received by
department A1.
Intergroup Interaction Dynamic Properties
Intergroup Role Interaction between A and C: IrRI(A, C)

For the connectivity between the groups A and C, the following intergroup role
interaction properties are considered, one from A to C, and one from C to A.
IrRI(depA1, divArep) Progress Information Provision A to B
If within division A progress information on component production is received

by department A1, then within the connection group this will be communicated by
the division A representative to the division B representative.
IrRI(divArep, depA1) B Progress Information Incorporation by A
If within the connection group the division A representative receives informa-
tion from the division B representative on which components are needed, then
within division A a component production planning will be generated by depart-
ment A1 taking these into account.
Organisational Dynamic Properties
DP(F) Overall Progress Notification

If a request for a product is made (by a client), then progress information will be
provided (for the client).
Realisation
The following allocation of agents agentA1, agentA2, agentB1, agentB2 to roles is

possible:
agentA1 – depA1 agentB1 – depB1 agentA1 – divArep

agentA2 – depA2 agentB2 – depB2 agentB1 – divBrep
To realise the organisation model, for example agentA1 has to satisfy the
following dynamic properties:
DP(agentA1)
If agent A1 receives progress information on component production,
Then an updated planning will be generated by agent A1 taking this most recent
information into account.
IrRI(agentA1)
If progress information on component production is received by agent A1,
Then this will be communicated by agent A1 to agent B1
If agent A1 receives information on which components are needed,
Then a component production planning will be generated by agent A1 taking
these components into account
5.3.5.3 Conclusion of Organisational Example
One can see how the above analysis is getting us closer to the implementation of a
simulation of this organisation. Given details of an organisation this could continue
down the organisational structure. Clearly this kind of analysis is more appropriate
when the structure of the organisation is known, and much less appropriate when
the structure is only partially known or, indeed, emergent. However, even in those
cases it could guide the documentation capturing how and which aspects of an
organisation’s official structure was translated into a simulation.
5.4 Organisation Design by Requirements Refinement
The previous sections address the question of how the structure and the behaviour
of a given organisation can be modelled. This section takes the design perspective.
This perspective does not assume a given organisation, but aims at creating a new
organisation in silica. Whilst on the whole in simulation we are aiming to capture
aspects of existing organisations one might want to design one in a number of
circumstances.
For example, one may not know the internal structure of an organisation which is
part of the system one is trying to model but only how it communicates or relates to
the actors around it. In this case to get the simulation to run one would have to
invent the organisation. Obviously the danger in this case is that the organisational
structure that is chosen might subtly affect how it interacts with other agents and
thus have an impact upon the simulation outcomes. However in some cases the
internal workings of an organisation are effectively insulated from how it interacts
with the outside world by regulation and self-interest – in these cases one might
well have no choice but to invent its workings on the basis of its external constraints
and common knowledge of how such things are arranged.
Another case is where one is not attempting to represent anything that is
observed but rather exploring the space of possible organisations. For example
one might wish to know which of several possible organisational structures might
be best according to some specified criteria. Such “artificial societies” or “thought
experiments” are reasonable common, however their relevance is questionable. If a
small change in the environment or other aspect (e.g. reliability of internal commu-
nication) means that what was a good organisational structure now fails, then the
results of such artificial studies are difficult to apply to other cases. In other words
the general applicability of such studies is hard to establish. On the other hand if one
has a good knowledge of the environment and characteristics where the results of
the simulation experiments are to be applied and one does extensive ‘what if’
experiments testing the robustness of the designs to such small changes then this
can be a helpful way forward.
Such an design process starts by specifying requirements for the overall
organisation behaviour. The requirements express the dynamic properties that
should ‘emerge’ if appropriate organisational building blocks, such as roles, groups,
transfers, group interactions, role behaviours, and so on, are glued together in an
appropriate manner in an organisation model. In addition, other requirements on
behavioural and structural aspects of the organisation to be created may be
imposed. Given these requirements on overall organisation behaviour (and, per-

haps, some additional requirements), organisational structure and organisational
behaviour are designed in such a manner that the requirements are fulfilled. The
approach described in this section is the method of requirements refinement, which
is illustrated for an example.
5.4.1 Designing by Requirements Refinement
In Sect. 5.3.3 a scheme for specifying the dynamic properties and relationships at
different levels of aggregation was described; overall organisational behaviour can
be related to dynamic group properties and group interaction properties via the
following pattern:
Dynamic properties for the groups AND dynamic properties for group interaction
⇨ Dynamic properties for the organisation
This scheme is also useful for the design perspective. Consider a design problem
for which the requirements of the overall behaviour are given in the form of
dynamic properties. This scheme says that to fulfil these overall dynamic
properties, dynamic properties of certain groups and group interactions together
imply the organisation behaviour requirements. This process is called requirements
refinement in that the requirements for the whole organisation are reduced to that of
its constituent groups and the interactions between these groups. It thus provides a
new, refined set of requirements in terms of the behaviour of groups and group
interaction.
Clearly if one has a multi-level organisation with component sub-organisations
as well as groups one has a choice as to how best to fill in the detail of one’s design.
One can decide to reduce it first to the behaviour of its constituent groups but it is
also possible to first refine requirements for the behaviour of the organisation as a
whole to the requirements on the behaviour of parts of the organisation, before
further refinement is made to refinements for groups. In each case this is a
pragmatic decision and will depend on the organisation being designed.
Subsequently, the required dynamic properties of groups can be refined to
dynamic properties of certain roles and transfers, making use of:
Dynamic properties for roles AND dynamic properties for transfer between roles
) Dynamic properties for a group
This provides requirements on role behaviour and transfer that together imply
the requirements on the behaviour of the group. Again it is possible to first refine
requirements on the behaviour of a group to requirements of the behaviour of parts
of the group, before further refinement to role behaviour requirements is made,
depending on what is best in each case.
An overview of the inter-level relationships between these dynamic properties at
different aggregation levels is depicted in Fig. 5.1, repeated for your convenience
Fig. 5.4 Overview of organisation properties

interlevel relationships
between dynamic properties
within an organisation model
group interaction
group properties
properties
transfer properties role properties
here as Fig. 5.4. In summary, from the design perspective, a top-down refinement
approach can be followed. That is, the requirements on overall organisational
behaviour can be first refined to requirements on behaviour of groups and group
interaction, and then the requirements on behaviour of groups can be refined to
requirements on roles and transfers. Notice that as part of this refinement process
the organisational structure (e.g., the groups and roles) is defined.
A design problem statement consists of:
• A set of requirements (in the form of dynamic properties) that the overall
organisational behaviour has to fulfil
• A partial description of (prescribed) organisational structure that has to be
incorporated
• A partial description of (prescribed) dynamic properties of parts of the
organisation that have to be incorporated; e.g., for roles, for transfers, for groups,
for group interactions.
A solution specification for a design problem is a specification of an organisation
model (both structure and behaviour) that fulfils the imposed requirements on
overall organisation behaviour, and includes the given (prescribed) descriptions
of organisation structure and behaviour. Here ‘fulfilling’ the organisation behaviour
requirements means that the dynamic properties for roles, transfers, and group
interactions within the organisation model imply the behaviour requirements.
In specific circumstances, part of the organisational structure and/or behaviour
may already be prescribed by requirements. For example, the organisational struc-
ture may already be prescribed; in such a case only the organisation dynamics is
designed, for the given organisational structure. Other, more specific cases are, for
example, role behaviour design and interaction protocol design.
5.4.1.1 Role Behaviour Design
For role behaviour design the organisational structure and the transfers and
interactions are completely prescribed. However, appropriate dynamic properties
for the different roles have yet to be found, to satisfy the requirements for the
Fig. 5.5 Role behaviour organisation properties

design
group interaction
group properties
properties
Fig. 5.6 Organisation design organisation properties
group interaction
group properties
properties
organisational behaviour that are imposed; see Fig. 5.5. Here (and in Fig. 5.6) the
grey rectangles indicate what is already given as prescribed and the transparent
rectangle what has to be designed.
5.4.1.2 Interaction Protocol Design
For interaction protocol design the organisational structure and role dynamics are
completely prescribed, but appropriate transfer and interaction dynamics have to be
found to satisfy given requirements for the organisational behaviour that are
imposed; see Fig. 5.6.
5.5 The Agent Approach
The agent approach is contrary to the organisational approach. It starts with the
agents and its properties and attempts to work upwards towards the whole system.
This is more useful in situations where the “top down” social constraints upon
action are weak or non-existent, and it is the “upwards” emergence of outcome from
the many micro-level interactions that is more important. Clearly, if one was in a
situation where top down social and/or organisational constraints were severe then
one would have no guarantee that working bottom-up in this manner one would be
able to meet those constraints at the higher levels of the structure. It would be like
trying to organise the production of a kind of car by the random meeting of people
with particular parts and skills without any planning. However, especially in the
development of new social structure such “bottom-up” processes can be crucial, so
that the agent approach can be appropriate for investigating such issues. Often, for
social phenomena some mix of both approaches is necessary, first a bit of one, then
a bit of the other etc.
5.5.1 Some Agent Notions
The term agent has been used for a wide variety of applications, including: simple
batch jobs, simple email filters, mobile applications, intelligent assistants, and
large, open, complex, mission critical systems (such as systems for air traffic
control).7 Some of the key concepts concerning agents lack universally accepted
definitions. In particular, there is only partial agreement on what an agent is. For
example, simple batch jobs are termed agent because they can be scheduled in
advance to perform tasks on a remote machine, mobile applications are termed
agent because they can move themselves from computer to computer, and intelli-
gent assistants are termed agents because they present themselves to human users as
believable characters that manifest intentionality and other aspects of a mental state
normally attributed only to humans. Besides this variety in different appearances of
agents, the only precise description of the agents involved is their implementation
code. As a result, existing agent architectures are only comparable in an informal
manner – just because something is called an agent-architecture or an agent does
not mean that it is suitable for simulating a human or social actor. Especially if the
goal of the agent-based system is a complex simulation, a principled, design-
oriented description of the organisation, and of the agents in it at a conceptual
and logical level is of the essence, since the control, testing, verification and
validation of such complex simulations is an issue. Spending time on a more formal
and staged approach to simulation design can make life a lot easier later. Due to the
organisational nature, and the complexity of intelligent agents and their interaction,
a more formal compositional design method for agents is necessary.
As agents show a variety of appearances, perform a multitude of tasks, and their
abilities vary significantly, attempts have been made to define what they have in
common. The weak notion of agent is seen as a reference. The weak notion of agent
is a notion that requires the behaviour of agents to exhibit at least the following four
types of behaviour:
7
Many of the notions discussed in this and the following section are adopted from (Wooldridge
and Jennings 1995), (Nwana 1996), (Nwana and Ndumu 1998) and (Jennings and Wooldridge
1998).
• Autonomous behaviour
• Responsive behaviour (also called reactive behaviour)
• Pro-active behaviour
• Social behaviour
Autonomy relates to control: although an agent may interact with its environ-
ment, the processes performed by an agent are in full control of the agent itself.
Autonomous behaviour is defined as:
. . . where the system is able to act without the direct intervention of humans (or other
agents) and should have control over its own actions and internal state.
This means that an agent can only be requested to perform some action, and:
The decision about whether to act upon the request lies with the recipient.
Examples of autonomous computer processes are: process control systems (e.g.,

thermostats, missile guiding systems, and nuclear reactor control systems), soft-
ware daemons (e.g., one that monitors a user’s incoming email and obtains their
attention by displaying an icon when new, incoming email is detected), and
operating systems.
Many processes that exhibit autonomous behaviour are called agents. However,
if such agents do not exhibit flexible behaviour, they are not, in general, considered
to be intelligent agents. An intelligent agent is a computer system that is capable of
flexible autonomous actions in order to meet its design objectives – indeed Randall
Beer (1990) defined intelligence as “the ability to display adaptive behaviour”.
Intelligence requires flexibility with respect to autonomous actions, meaning that
intelligent agents also need to exhibit responsive, social, and pro-active behaviour.
An agent exhibits responsive (or reactive) behaviour if it reacts or responds to
new information from its environment. Responsive behaviour is where:
Agents perceive their environment (which may be the physical world, a user, a collection of
agents, the Internet, etc.) and respond in a timely fashion to changes that occur in it.
A barometer is a simple example of a system that exhibits responsive behaviour:

It continually receives new information about the current air pressure and responds
to this new information by adjusting its dial.
Pro-active behaviour is where:
Agents do not simply act in response to their environment, but are be able to exhibit
opportunistic, goal-directed behaviour and take the initiative where appropriate.
Pro-active behaviour in some sense is the most difficult of the required types of
behaviour for an agent defined according to the weak agent notion. For example,
pro-active behaviour can occur simultaneously with responsive behaviour. It is
possible to respond to incoming new information in an opportunistic manner
according to some goals. Also initiatives can be taken in response to incoming
new information from the environment, and thus this behaviour resembles respon-
sive behaviour. However, it is also possible to behave pro-actively when no new
information is received from the environment. This last behaviour can by no means
be called responsive behaviour.
An agent exhibits social behaviour if it communicates and co-operates with
other agents. Jennings and Wooldridge define social behaviour as when:
Agents are able to interact, when they deem appropriate, with other artificial agents and
humans in order to complete their own problem solving and to help others with their
activities.
An example of an agent that exhibits social behaviour is a car: it communicates

with its human user by way of its dials (outgoing communication dials: speed,
amount of fuel, temperature) and its control mechanisms (incoming communication
control mechanisms: pedals, the steering wheel, and the gears). It co-operates with
its human user, e.g., by going in the direction indicated by the user, with the speed
set by that user.
Agents can also be required to have additional characteristics. Here three of
these characteristics are discussed: adaptivity, pro-creativity, and intentionality.
Adaptivity is a characteristic that is vital in some systems. An adaptive agent
learns and improves with experience. This behaviour is vital in environments that
change over time in ways that would make a non-adaptive agent obsolete or give it
no chance of survival. This characteristic is modelled in simulations of societies of
small agents, but also, for example, in adaptive user interface agents.
Pro-creativity is of similar importance to find agents that satisfy certain
conditions. The chance of survival is often measured in terms of a fitness function.
This characteristic is found in various simulations of societies of small agents (see
the literature in the area of Artificial Life). A computer virus is an infamous form of
a pro-creative agent.
An intentional system is defined by Dennett to be an entity
. . . whose behaviour can be predicted by the method of attributing beliefs, designs and
rational acumen.
Mentalistic and intentional notions such as beliefs, desires, intentions,

commitments, goals, plans, preference, choice, awareness, may be assigned to
agents. The stronger notion of agenthood, in which agents are described in terms
of this type of notions, provides additional metaphorical support for the design of
agents.
5.5.2 Representative Agents
Of course a software agent need not have all the characteristics of something it
represents. Thus, depending on the model purpose, it is quite possible to represent
an intelligent actor by a relatively simply object, that might not even be meaning-
fully called an agent. For example, if simulating the movement of a crowd or traffic,
it might not be necessary to include much, if any, of the features discussed above.
So sometimes something is called an agent because it represents an agent, but this

usage conflates what is being modelled and the model, which may cause confusion.
How and when it is necessary to include the various features that are known about
the actor being modelled in the model is a crucial modelling question. However it is
not a question to which there is a general answer since this goes to the heart of the
difficulty of making sense of the complexity that we observe.
The power of the agent idea and approach to programming, clearly comes from
the apparent efficacy of observed social actors that seem to be able to organise
themselves in useful and adaptive ways, that they have the characteristics listed
above. This differs qualitatively from more traditional computer science
approaches to programming. It also comes from the power of the analogy with
humans to guide the direction of programming – we have a deep mundane knowl-
edge of how humans work (c.f. Dennett 1996) and this can help in design decisions,
for example whether a certain feature or ability is necessary to model a particular
social system. In this sense the idea of an agent can be thought of as a stance, in
comparison to the “intentional” stance mentioned above – it may be useful to think
of a computational object as an agent, having some of the sort of properties we
know real human actors have. However there are clearly more and less useful
applications of this stance: I may think of a light switch as an agent, but it has
limited usefulness in terms of understanding or modelling it. In the contrary
direction on can think of a fully autonomous and intelligent agent such as a
human as a merely a physical particle for some circumstances, however this is
open to question, depending upon the purpose and target of the exercise. Clearly, in
many circumstances and for many purposes, treating complex social actors as if
they were simple (for example acted upon in the manner of a simple linear influence
plus some randomness) is insufficient since individual complexity can impinge
upon the social complexity that results.
5.5.3 Agent Properties
The notions of agency discussed above are highly abstract notions. In order to
design agents, it is necessary to be familiar with a number of primitive agent
concepts.8 These primitive concepts serve as an ontology or vocabulary used to
express analyses and designs of applications of agents and multi-agent systems.
Two classes of primitive notions are distinguished: those used to describe the
behaviour of agents in terms of their external (or public) states and interactions
(Sect. 5.5.3.1), and those used to describe the behaviour of agents in terms of their
internal (or private) states, and processes (Sect. 5.5.3.2). To illustrate these
concepts, some example agents are discussed in Sect. 5.5.4.
8
The material in this section is based on (Brazier et al. 2000).
5.5.3.1 External Primitive Concepts
Two types of interaction of an agent with its environment are distinguished,

depending on whether the interaction takes place with an agent or with something
else (called an external world), for example a database, or the material world. For
each of these two types of interaction specific terminology is used.
Interaction with the External World
Two primitive types of interaction with the external world are distinguished. The
first type of interaction, observation, changes the information the agent has about
the world, but does not change the world state itself, whereas the second type,
performing an action, does change the world state, but does not change the
information the agent has about the world. Combinations of these primitive types
of interaction are possible; for example, performing an action, and observing its
results.
Observation
In which ways is the agent capable of observing or sensing its environment? Two
types of observation can be distinguished: the agent passively receives the results of
observations without taking any initiative or control to observe (passive observa-
tion), or the agent actively initiates and controls which observations it wants to
perform; this enables the agent to focus its observations and limit the amount of
information acquired (active observation).
Execution of Actions in the External World
An agent may be capable of making changes to the state of its environment by

initiating and executing specific types of actions.
Communication with Other Agents
Two directions of communication are distinguished, which can occur together:

outgoing communication (is the agent capable of communicating to another
agent; to which ones?), and incoming communication (is the agent capable of
receiving communication from another agent; from which ones?).
5.5.3.2 Internal Primitive Concepts
A description in terms of the external primitive concepts abstracts from what is

inside the agent. In addition to descriptions of agents in terms of the external
concepts, descriptions in terms of internal concepts are useful. The following
internal primitive agent concepts are distinguished.
World and Agent Models
An agent may create and maintain information on (a model of) external world based
on its observations of that world, on information about that world communicated by
other agents, and its own knowledge about the world. The agent may also create and
maintain information on (models of) other agents in its environment based on its
observations of these agents as they behave in the external world, on information
about these agents communicated by other agents, and knowledge about the world.
Self Model and History
Some agents create and maintain information on (a model of) their own
characteristics, internal state, and behaviour. Or the agent creates and maintains a
history of the world model, or agent models, or self model, or own and group
processes.
Goals and Plans
To obtain pro-active, goal-directed behaviour, an agent represents, generates, and

uses explicit goals and its own plans of action in its processing.
Group Concepts
Besides individual concepts, agents can use group concepts that allow it to
co-operate with other agents. For example, joint goals: is the agent capable of
formulating or accepting and using goals for a group of agents, i.e., goals that can
only be achieved by working together? Or joint plans: is the agent capable of
representing, generating, and using plans of action for joint goals, i.e., involving
which actions are to be performed by which agents in order to achieve a certain joint
goal? Also commitments to joint goals and plan, negotiation protocols and
strategies can be useful group concepts for agents, depending on their role and
function.
Table 5.2 External primitive concepts for an elevator

External primitive concepts Elevator
Interaction with the world
Passive observations Presence of objects between doors (optically)
Total weight
Its position
Active observations Presence of objects between the doors
(mechanically)
Performing actions Moving
Opening and closing doors
Communication with other agents
Incoming communication From users in the elevator
Where they want to go (pushing button in elevator)
From users outside
Where they want to be picked up (pushing button
outside elevator)
Outgoing communication To users in the elevator
Where we are (display)
There is overweight (beep)
To users outside
Where is the elevator (display)
In which direction it moves (display)
5.5.4 Example of the Agent Approach: An Elevator
Let us illustrate the agent concepts introduced above by an example: an elevator is

analysed from the agent perspective using these basic concepts (Table 5.2). This
might be an element in the simulation of how people move around a building. The
advantage of using an elevator as an example is that it does interact with users as an
agent, but it is well known and simple enough to make a clear example of
specifying an agent using the agent-oriented approach.
5.5.4.1 External Primitive Concepts (Table 5.2)
Observation
Observations are performed continually. However, it receives passive observation

results on the presence of objects between the doors (an optical sensor), the total
weight of its contents, and its position in the building (at which floor). Besides it is
able to perform active observation: the presence of objects between the doors
(a mechanical sensor which is moved in the door opening just ahead of the doors
themselves).
Table 5.3 Internal primitive concepts for an elevator

Internal primitive concepts Elevator
World model The current floor, max load, current load
Agent models A user wants to be picked up from floor X
A user wants to go to floor Y
Self model When maintenance is next due
History When maintenance was last performed
Goals To go to floor X to pick up somebody
To go to floor Y to deliver somebody
Plans The order in which the required floors are visited
Sometimes: the speed that is taken
Group concepts
Joint goals With other elevators to transport people and
goods as efficiently as possible
Joint plans Some elevators are capable of distributing the work
Commitments The elevators then commit to their part of the work
Negotiation protocol To reach a good distribution, they may have to negotiate
Negotiation strategies To reach a good distribution, they may have to negotiate
Performing Actions
Its actions are moving itself (and people) vertically from one position to another and
opening and closing doors.
Incoming Communication
The elevator receives communication from users by buttons that have been pressed
outside (to call the lift and indicate the direction they wish to go in) and inside the
lifts (providing information about the floor to which they wish to be transported).
Outgoing Communication
The elevator communicates to a user by indicating which floor the lift is on (both
inside and outside the lifts) and sounding beeps (information about overload)
(Table 5.2).
5.5.4.2 Internal Primitive Concepts (Table 5.3)
World and Agent Models
Elevators know where they are, to do this they keep track of which floor they are on
based on their actions (going two floors up, going one floor down) they perform.
Table 5.4 Types of behaviour for an elevator

Types of behaviour Elevator
Autonomy Yes
Responsiveness In reaction to user requests
In immediate reaction to observed objects between the doors
Pro-activeness Taking the initiative to go to a normally busy floor, if empty
and not being called by a user
Social behaviour Co-operation with users, and, sometimes, with other elevators
Own adaptation and learning Often not possible
Furthermore, the elevator knows if the weight in the lift is over its maximum limit.
The agent information of the user goals (where they want to go) may be maintained
as well.
Self Model and History
The agent does not know what actions it previously performed to perform its
current task. It might have an explicit representation of when it has last received
maintenance.
Goals and Plans
Modern elevators make use of the explicit goals (adopted from the goals
communicated by the users). The goals are used to determine which actions to
perform. They may even make plans for reaching these goals: determine the order
of actions, for example when one of the users has the goal to be at a higher floor and
another on a lower floor.
Group Concepts
The elevator co-operates with its users. The elevator might also be designed to co-
operate with other elevators so that they could strategically distribute themselves
over the floors. The goals adopted from the goals communicated by the users are
joint goals (joint with the users), and sometimes even joint with the other elevators.
Modern elevators are capable of distributing the work load, and thus of making joint
plans. To achieve the joint goals an elevator must commit to its part of the work as
specified in the joint plans. To make a joint plan, the elevators might negotiate using
a particular strategy as to which elevator goes where. Negotiation is only possible if
a negotiation protocol is followed.
5.5.4.3 Types of Behaviour (Table 5.4)
Autonomy
As soon as it is activated, no system or human is controlling its machinery, and

(normally) it is not switched off and on by the user. The elevator has full control of
its motor, doors, and lights.
Pro-activeness
The simplest elevators stay where they are (some take the initiative to close their
doors) when no longer in use, but more intelligent elevators go to a strategic floor
(e.g., the ground floor).
Reactiveness
The elevator reacts to the immediate stimuli of buttons pressed, therefore, it is

shows reactive behaviour. People often have to wait for the elevator as the elevator
picks up people on other floors, however, the elevator does not forget a signal and
will, eventually, come to the requested floor.
Social Behaviour
The elevator co-operates with users and, sometimes, with other elevators.
Own Adaptation and Learning
Simple elevators are not capable of adjusting their own behaviour to new situations,
nor are they capable of learning. However, it is possible to conceive of more
intelligent elevators that can learn the rush hours for the different floors.
5.5.4.4 Conclusion of Elevator Example
One can see that the above analysis has clarified what is needed to implement a
model of this element within a simulation. The objects, properties and processes of
the simulation code is an easy step from here. One also sees that some of the
complexities of the agent have been considered before the implementation starts, in
this way cleaner and more maintainable code might be produced, fewer mistakes
made in the implementation and some of the complexities in terms of user-lift

interaction considered and anticipated. Of course in the case of simulating a human
agent there are likely to be many unknowns in terms of their goals, group concepts
etc. – unless the simulators simply add in their informed guesses they will have a
considerable job finding evidence to guide them in the answers to fill in within such
an analysis. Thus this kind of analysis is not a total solution when simulating
complex social actors whose attributes and internal states may be largely unknown.
5.6 Conclusion
More formal approaches to simulation design can help make complex

implementations manageable and can probably save one time in the longer-run. It
also makes the simulation easier to check, validate, re-implement and further
develop. These approaches do this in a three principled ways. Firstly, by encourag-
ing the more complete documentation of the intentions and decisions of a designer/
implementer. One can see a lot of this chapter as a check-list of all the aspects one
might think about and record. Secondly, it helps encourage doing this in an explicit
and exact manner. We have not displayed some of the formal notation that can be
used in this chapter, since we did not want to overwhelm the reader, however for
those who wish to utilise this style of design to the fullest will naturally find them
themselves adopting a variety of formal languages and diagrams in pursuit of
precision. Thirdly, it suggests a method by which a complex specification can be
iteratively refined from abstract and high-level entities towards a detailed imple-
mentation. In this way the design decisions do not all have to be made simulta-
neously but can be made in stages.
Further Reading
For readers interested in software engineering approaches Bergenti et al. (2004)

give a thorough introduction to and overview of current methodologies. Gilbert and
Terno (2000) offer suggestions on techniques for building and structuring agent-
based simulation models, particularly geared towards use in the social sciences.
In addition to methodologies, a lot of work has been done in the development of
programming languages and platforms to support the implementation of multi-
agent systems and models. Bordini et al. (2010) focus on a comprehensive presen-
tation of MAS programming, including four approaches that are based on formal
methods, whereas Railsback et al. (2006) provide a review of platforms for agent-
based simulations.
References
Beer RD (1990) Intelligence as adaptive behavior: an experiment in computational neuroethology.

Academic, Boston
Bergenti F, Gleizes M-P, Zambonelli F (eds) (2004) Methodologies and software engineering for
agent systems: the agent-oriented software engineering handbook. Kluwer, Boston
Bordini RH, Dastani M, Seghrouchni AEF (eds) (2010) Multi-agent programming: languages,
platforms and applications. Springer, Berlin
Brazier FMT, Jonker CM, Treur J (1998) Principles of compositional multi-agent system devel-
opment. In: Cuena J (ed) IT & KNOWS – information technology and knowledge systems,
IFIP world computer congress 1998 (Schriftenreihe der Österreichischen Computer Gesell-
schaft, 122). OCG, Wien, pp 347–360
Brazier FMT, Jonker CM, Treur J (2000) Compositional design and reuse of a generic agent
model. Appl Artif Intell J 14:491–538
Brazier FMT, van Eck PAT, Treur J (2001) Modelling a society of simple agents: from conceptual
specification to experimentation. J Appl Intell 14:161–178
Daniel C. Dennett (1996), The Intentional Stance (6th printing), Cambridge, Massachusetts: The
MIT Press, ISBN 0-262-54053-3 (First published 1987)
Dignum V (2013) Assessing organisational design. Chapter 20 in this volume
Ferber J, Gutknecht O (1998) A meta-model for the analysis and design of organisations in multi-
agent systems. In: Demazeau Y (ed) Proceedings of the third international conference on multi-
agent systems (ICMAS’98), Paris, 3–7 July 1998. IEEE Computer Society Press, pp 128–135
Ferber J, Gutknecht O (1999) Operational semantics of a role-based agent architecture. In:
Jennings NR, Lespérance Y (eds) Intelligent agents VI, proceedings of the 6th international
workshop on agent theories, architectures and languages, ATAL’99 (Lecture notes in
Computer Science), vol 1757. Springer, Berlin, pp 205–217
Ferber J, Gutknecht O, Jonker CM, Müller JP, Treur J (2000) Organization models and
behavioural requirements specification for multi-agent systems. In: Proceedings of the fourth
international conference on MultiAgent systems (ICMAS), Boston, 10–12 July 2000. IEEE
Computer Society, pp 387–388
Galán J et al (2013) Checking simulations: detecting and avoiding errors and artefacts. Chapter 6
in this volume
Gilbert N, Terno P (2000) How to build and use agent-based models in social science. Mind Soc
1(1):57–72
Hannoun M, Sichman JS, Boissier O, Sayettat C (1998) Dependence relations between roles in a
multi-agent system: towards the detection of inconsistencies in organization. In: Sichman JS,
Conte R, Gilbert N (eds) Multi-agent systems and agent-based simulation, proceedings of the
first international workshop on multi-agent based simulation, MABS’98. (Lecture notes in
artificial intelligence), vol 1534. Springer, Berlin, pp 169–182
Hannoun M, Boissier O, Sichman JS, Sayettat C (2000) MOISE: an organizational model for
multi-agent systems. In: Monard MC, Sichman JS (eds) Advances in artificial intelligence,
international joint conference 7th Ibero-American conference on AI, 15th. Brazilian sympo-
sium on AI (IBERAMIA-SBIA 2000), Atibaia, 19–22 Nov 2000. Proceedings (Lecture notes in
computer science), vol 1952. Springer, Berlin, pp 156–165
Hübner JF, Sichman JS, Boissier O (2002a) A model for the structural, functional and deontic
specification of organizations in multiagent systems. In: Bittencourt G, Ramalho GL (eds)
Advances in artificial intelligence. Proceedings of 16th Brazilian symposium on artificial
intelligence (SBIA’02), Porto de Galinhas/Recife, Brazil, 11–14 Nov 2002. (Lecture notes in
Hübner JF, Sichman JS, Boissier O (2002b) MOISE+: towards a structural, functional and deontic
model for MAS organizations. In: Castelfranchi C, Johnson WL (eds) Proceedings of the first
international joint conference on autonomous agents and multi-agent systems, AAMAS 2002,
Bologna, 15–19 July 2002. ACM Press, pp 501–502
Jennings NR, Wooldridge M (eds) (1998) Agent technology: foundations, applications, and
markets. Springer, Berlin
Jonker CM, Treur J (2003) Relating structure and dynamics in organisation models. In: Sichman JS,
Bousquet F, Davidsson P (eds) Multi-agent-based simulation II, third international workshop,
MABS 2002, Bologna, July 2002, revised papers (Lecture notes in AI), vol 2581. Springer,
Berlin, pp 50–69
Jonker CM, Treur J, Wijngaards WCA (2001) Temporal languages for simulation and analysis of
the dynamics within an organisation. In: Dunin-Keplicz B, Nawarecki E (eds) From theory to
practice in multi-agent systems, proceedings of the second international workshop of central
and eastern Europe on multi-agent systems, CEEMAS’01 (Lecture notes in computer science),
vol 2296. Springer, Berlin, pp 151–160
Kreitner R, Kunicki A (2001) Organisational behavior. McGraw-Hill, New York
Lomi A, Larsen ER (2001) Dynamics of organizations: computational modeling and organization
theories. AAAI Press, Menlo Park
Mintzberg H (1979) The structuring of organisations. Prentice Hall, Englewood Cliffs
Moss S, Gaylard H, Wallis S, Edmonds B (1998) SDML: a multi-agent language for organiza-
tional modelling. Comp Math Org Theory 4(1):43–70
Norling E, Edmonds B, Meyer R (2013) Informal approaches to developing simulation models.
Nwana HS (1996) Software agents: an overview. Knowl Eng Rev 11(3):205–244
Nwana HS, Ndumu DT (1998) A brief introduction to software agent technology. In: Jennings M,
Wooldridge NR (eds) Agent Technology: Foundations, Applications and Markets. Springer
Verlag, Berlin, pp 29–47
Prietula M, Gasser L, Carley K (1997) Simulating organizations. MIT Press, Cambridge, MA
Sichman JS, Conte R (1998) On personal and role mental attitudes: a preliminary dependence-
based analysis. In: Oliveira F (ed) Advances in AI. Proceedings of the 14th Brazilian sympo-
sium on artificial intelligence, SBIA’98 (Lecture notes in artificial intelligence), vol 1515.
Simon HA (1962) The architecture of complexity. Proc Am Philos Soc 106(6):467–482
Wooldridge M, Jennings NR (1995) Agent theories, architectures, and languages: a survey. In:
Wooldridge M, Jennings NR (eds) Intelligent agents, proceedings of the first international
workshop on agent theories, architectures and languages, ATAL’94 (Lecture notes in AI),
vol 890. Springer, Berlin, pp 1–39
Chapter 6
Checking Simulations: Detecting and Avoiding
Errors and Artefacts
José M. Galán, Luis R. Izquierdo, Segismundo S. Izquierdo, José I. Santos,

Ricardo del Olmo, and Adolfo López-Paredes
Why Read This Chapter? Given the complex and exploratory nature of many
agent-based models, checking that the model performs in the manner intended by
its designers is a very challenging task. This chapter helps the reader to identify
some of the possible types of error and artefact that may appear in the different
stages of the modelling process. It will also suggest some activities that can be
conducted to detect, and hence avoid, each type.
Abstract The aim of this chapter is to provide the reader with a set of concepts and
a range of suggested activities that will enhance his or her ability to understand
agent-based simulations. To do this in a structured way, we review the main
concepts of the methodology (e.g. we provide precise definitions for the terms
‘error’ and ‘artefact’) and establish a general framework that summarises the
process of designing, implementing, and using agent-based models. Within this
framework we identify the various stages where different types of assumptions are
usually made and, consequently, where different types of errors and artefacts may
appear. We then propose several activities that can be conducted to detect each type
of error and artefact.
J.M. Galán (*) • L.R. Izquierdo • J.I. Santos • R. del Olmo

Departamento de Ingenierı́a Civil, Universidad de Burgos, Burgos E-09001, Spain
e-mail: [email protected]; [email protected]; [email protected]; [email protected]
S.S. Izquierdo • A. López-Paredes
Departamento de Ingenierı́a Civil, Universidad de Burgos, Valladolid E-47011, Spain

96 J.M. Galán et al.
6.1 Introduction
Agent-based modelling is one of multiple techniques that can be used to conceptu-

alise social systems. What distinguishes this methodology from others is the use of
a more direct correspondence between the entities in the system to be modelled and
the agents that represent such entities in the model (Edmonds 2001). This approach
offers the potential to enhance the transparency, soundness, descriptive accuracy,
and rigour of the modelling process, but it can also create difficulties: agent-based
models are generally complex and mathematically intractable, so their exploration
and analysis often require computer simulation.
The problem with computer simulations is that understanding them in reason-
able detail is not as straightforward an exercise as one could think (this also applies
to one’s own simulations). A computer simulation can be seen as the process of
applying a certain function to a set of inputs to obtain some results. This function is
usually so complicated and cumbersome that the computer code itself is often not
that far from being one of the best descriptions of the function that can be provided.
Following this view, understanding a simulation would basically consist in
identifying the parts of the mentioned function that are responsible for generating
particular (sub)sets of results.
Thus, it becomes apparent that a prerequisite to understand a simulation is to
make sure that there is no significant disparity between what we think the computer
code is doing and what is actually doing. One could be tempted to think that, given
that the code has been programmed by someone, surely there is always at least one
person – the programmer – who knows precisely what the code does. Unfortu-
nately, the truth tends to be quite different, as the leading figures in the field report:
You should assume that, no matter how carefully you have designed and built your
simulation, it will contain bugs (code that does something different to what you wanted
and expected). (Gilbert 2007)
An unreplicated simulation is an untrustworthy simulation – do not rely on their results,
they are almost certainly wrong. (‘Wrong’ in the sense that, at least in some detail or other,
the implementation differs from what was intended or assumed by the modeller). (Edmonds
and Hales 2003)
Achieving internal validity is harder than it might seem. The problem is knowing
whether an unexpected result is a reflection of a mistake in the programming, or a surprising
consequence of the model itself. [. . .] As is often the case, confirming that the model was
correctly programmed was substantially more work than programming the model in the first
place. (Axelrod 1997a)
In the particular context of agent-based simulation, the problem tends to be

exacerbated. The complex and exploratory nature of most agent-based models
implies that, before running a model, there is almost always some uncertainty
about what the model will produce. Not knowing a priori what to expect makes it
difficult to discern whether an unexpected outcome has been generated as a legiti-
mate result of the assumptions embedded in the model or, on the contrary, it is due
to an error or an artefact created in its design, its implementation, or in the running
process (Axtell and Epstein 1994: 31; Gilbert and Terna 2000).
6 Checking Simulations: Detecting and Avoiding Errors and Artefacts 97
Moreover, the challenge of understanding a computer simulation does not end

when one is confident that the code is free from errors; the complex issue of
identifying what parts of the code are generating a particular set of outputs remains
to be solved. Stated differently, this is the challenge of discovering what
assumptions in the model are causing the results we consider significant. Thus, a
substantial part of this non-trivial task consists in detecting and avoiding artefacts:
significant phenomena caused by accessory assumptions in the model that are
(mistakenly) deemed irrelevant. We explain this in detail in subsequent sections.
The aim of this chapter is to provide the reader with a set of concepts and a range
of suggested activities that will enhance his ability to understand simulations. As
mentioned before, simulation models can be seen as functions operating on their
inputs to produce the outputs. These functions are created by putting together a
range of different assumptions of very diverse nature. Some assumptions are made
because they are considered to be an essential feature of the system to be modelled;
others are included in a somewhat arbitrary fashion to achieve completeness – i.e. to
make the computer model run –, and they may not have a clear referent in the target
system. There are also assumptions – e.g. the selection of the compiler and the
particular pseudo-random number generator to be employed – that are often made,
consciously or not, without fully understanding in detail how they work, but
trusting that they operate in the way we think they do. Finally, there may also be
some assumptions in a computer model that not even its own developer is aware of,
e.g. the use of floating-point arithmetic, rather than real arithmetic.
Thus, in broad terms, understanding simulations requires identifying what
assumptions are being made, and assessing their impact on the results. To achieve
this, we believe that it is useful to characterise the process by which assumptions
accumulate to end up forming a complete model. We do this in a structured way by
presenting a general framework that summarises the process of creating and using
agent-based models through various stages; then, within this framework, we char-
acterise the different types of assumptions that are made in each of the stages of the
modelling process, and we identify the sort of errors and artefacts that may occur;
we also propose activities that can be conducted to avoid each type of error or
artefact.
The chapter is structured as follows: the following section is devoted to
explaining what we understand by modelling, and to argue that computer simula-
tion is a useful tool to explore formal models, rather than a distinctively new
symbolic system or a uniquely different reasoning process, as it has been suggested
in the literature. In Sect. 6.3 we explain what the essence of agent-based modelling
is in our view, and we present the general framework that summarises the process of
designing, implementing, and using agent-based models. In Sect. 6.4 we define the
concepts of error and artefact, and we discuss their relevance for validation and
verification. The framework presented in Sect. 6.3 is then used to identify the
various stages of the modelling process where different types of assumptions are
made and, consequently, where different types of errors and artefacts may appear.
We then propose various activities aimed at avoiding the types of errors and
artefacts previously described, and we conclude with a brief summary of the
chapter.
6.2 Three Symbolic Systems Used to Model Social Processes
Modelling is the art of building models. In broad terms, a model can be defined as
an abstraction of an observed system that enables us to establish some kind of
inference process about how the system works, or about how certain aspects of the
system operate.
Modelling is an activity inherent to every human being: people constantly
develop mental models, more or less explicit, about various aspects of their daily
life. Within science in particular, models are ubiquitous. Many models in the “hard”
sciences are formulated using mathematics (e.g. differential equation models and
statistical regressions), and they are therefore formal, but it is also perfectly
feasible – and acceptable – to build non-formal models within academia; this is
often the case in disciplines like history or sociology – consider e.g. a model written
in natural language that tries to explain the expansion of the Spanish Empire in the
sixteenth century, or the formation of urban “tribes” in large cities.
We value a model to the extent that it is useful – i.e. in our opinion, what makes a
model good is its fitness for purpose. Thus, the assessment of any model can only be
conducted relative to a predefined purpose. Having said that, there is a basic set of
general features that are widely accepted to be desirable in any model, e.g.
accuracy, precision, generality, and simplicity (see Fig. 6.1). Frequently some of
these features are inversely related; in such cases the modeller is bound to compro-
mise to find a suitable trade-off, considering the perceived relative importance of
each of these desirable features for the purpose of the model (Edmonds 2005).
Some authors (Gilbert 1999; Holland and Miller 1991; Ostrom 1988) classify the
range of available techniques for modelling phenomena in which the social dimen-
sion is influential according to three symbolic systems.
One possible way of representing and studying social phenomena is through
verbal argumentation in natural language. This is the symbolic system traditionally
used in historical analyses, which, after a process of abstraction and simplification,
describe past events emphasising certain facts, processes, and relations at the
expense of others. The main problem with this type of representation is its intrinsic
lack of precision (due to the ambiguity of natural language) and the associated
difficulty of uncovering the exact implications of the ideas put forward in this way.
In particular, using this symbolic system it is often very difficult to determine the
whole range of inferences that can be obtained from the assumptions embedded in
the model in reasonable detail; therefore it is often impossible to assess its logical
consistency, its scope, and its potential for generalisation in a formal way.
A second symbolic system that is sometimes used in the Social Sciences,
particularly in Economics, is the set of formal languages (e.g. leading to models
expressed as mathematical equations). The main advantage of this symbolic system
derives from the possibility of using formal deductive reasoning to infer new facts
from a set of clearly specified assumptions; formal deductive reasoning guarantees
that the obtained inferences follow from the axioms with logical consistency.
Formal languages also facilitate the process of assessing the generality of a
Simplicity
Generality
Specificity (makes
precise predictions)
Lack of error
(accuracy of results)
Fig. 6.1 The trade-off between various desirable features depends on the specific case and model.
There are not general rules that relate, not even in a qualitative fashion, all these features. The
figure shows a particular example from Edmonds (2005) that represents the possible equilibrium
relationships between some features in a particular model
model and its sensitivity to assumptions that are allowed to change within the
boundaries of the model (i.e. parameter values and non-structural assumptions).
However, the process of reducing social reality to formal models is not exempt
from disadvantages. Social systems can be tremendously complex, so if such
systems are to be abstracted using a formal language (e.g. mathematical equations),
we run the risk of losing too much in descriptiveness. To make things worse, in
those cases where it appears possible to produce a satisfactory formal model of the
social system under investigation, the resulting equations may be so complex that
the formal model becomes mathematically intractable, thus failing to provide most
of the benefits that motivated the process of formalisation in the first place. This is
particularly relevant in the domain of the Social Sciences, where the systems under
investigation often include non-linear relations (Axtell 2000). The usual approach
then is to keep on adding simplifying hypotheses to the model – thus making it
increasingly restrictive and unrealistic – until we obtain a tractable model that can
be formally analysed with the available tools. We can find many examples of such
assumptions in Economics: instrumental rationality, perfect information, represen-
tative agents, etc. Most often these concepts are not included because economists
think that the real world works in this way, but to make the models tractable (see for
instance Conlisk (1996), Axelrod (1997a), Hernández (2004), Moss (2001, 2002)).
It seems that, in many cases, the use of formal symbolic systems tends to increase
the danger of letting the pursuit for tractability be the driver of the modelling
process.
But then, knowing that many of the hypotheses that researchers are obliged to
assume may not hold in the real world, and could therefore lead to deceptive
conclusions and theories, does this type of modelling representation preserve its
advantages? Quoting G.F. Shove, it could be the case that sometimes “it is better to
be vaguely right than precisely wrong”.
The third symbolic system, computer modelling, opens up the possibility of
building models that somewhat lie in between the descriptive richness of natural
language and the analytical power of traditional formal approaches. This third type
of representation is characterized by representing a model as a computer program
(Gilbert and Troitzsch 1999). Using computer simulation we have the potential to
build and study models that to some extent combine the intuitive appeal of verbal
theories with the rigour of analytically tractable formal modelling.
In Axelrod’s (1997a) opinion, computational simulation is the third way of doing
science, which complements induction –the search for patterns in data– and deduc-
tion –the proof of theorems from a set of fixed axioms. In his opinion, simulation,
like deduction, starts from an explicit set of hypotheses but, rather than generating
theorems, it generates data that can be inductively analysed.
While the division of modelling techniques presented above seems to be
reasonably well accepted in the social simulation community –and we certainly
find it useful–, we do not fully endorse it. In our view, computer simulation does not
constitute a distinctively new symbolic system or a uniquely different reasoning
process by itself, but rather a (very useful) tool for exploring and analysing formal
systems. We see computers as inference engines that are able to conduct algorith-
mic processes at a speed that the human brain cannot achieve. The inference
derived from running a computer model is constructed by example and, in the
general case, reads: the results obtained from running the computer simulation
follow (with logical consistency) from applying the algorithmic rules that define the
model on the input parameters1 used.
In this way, simulations allow us to explore the properties of certain formal
models that are intractable using traditional formal analyses (e.g. mathematical
analyses), and they can also provide fundamentally new insights even when
such analyses are possible. Like Gotts et al. (2003), we also believe that mathemat-
ical analysis and simulation studies should not be regarded as alternative and even
opposed approaches to the formal study of social systems, but as complementary.
They are both extremely useful tools to analyse formal models, and they are
complementary in the sense that they can provide fundamentally different insights
on one same model.
To summarise, a computer program is a formal model (which can therefore be
expressed in mathematical language, e.g. as a set of stochastic or deterministic
equations), and computer simulation is a tool that enables us to study it in ways that
go beyond mathematical tractability. Thus, the final result is a potentially more
realistic – and still formal – study of a social system.
1
By input parameters in this statement we mean “everything that may affect the output of the
model”, e.g. the random seed, the pseudo-random number generator employed and, potentially,
information about the microprocessor and operating system on which the simulation was run, if
these could make a difference.
6.3 Agent Based Modelling
6.3.1 Concept
As stated before, modelling is the process of building an abstraction of a system for

a specific purpose (see Edmonds (2001) and Epstein (2008) for a list of potential
applications). Thus, in essence, what distinguishes one modelling paradigm from
another is precisely the way we construct that abstraction from the observed system.
In our view, agent-based modelling is a modelling paradigm with the defining
characteristic that entities within the target system to be modelled – and the
interactions between them – are explicitly and individually represented in the
model (see Fig. 6.2). This is in contrast to other models where some entities are
represented via average properties or via single representative agents. In many other
models, entities are not represented at all, and it is only processes that are studied
(e.g. a model of temperature variation as a function of pressure), and it is worth
noting that such processes may well be already abstractions of the system.2 The
specific process of abstraction employed to build one particular model does not
necessarily make it better or worse, only more or less useful for one purpose or
another.
The specific way in which the process of abstraction is conducted in agent-based
modelling is attractive for various reasons: it leads to (potentially) formal yet more
natural and transparent descriptions of the target system, provides the possibility to
model heterogeneity almost by definition, facilitates an explicit representation of
the environment and the way other entities interact with it, allows for the study of
the bidirectional relations between individuals and groups, and it can also capture
emergent behaviour (see Epstein 1999; Axtell 2000; Bonabeau 2002). Unfortu-
nately, as one would expect, all these benefits often come at a price: most of the
models built in this way are mathematically intractable. A common approach to
study the behaviour of mathematically intractable formal models is to use computer
simulation. It is for this reason that we often find the terms “agent-based modelling”
and “agent-based simulation” used as synonyms in the scientific literature (Hare
and Deadman 2004).
Thus, to summarise our thoughts in the context of the classification of modelling
approaches in the social sciences, we understand that the essence of agent-based
modelling is the individual and explicit representation of the entities and their
interactions in the model, whereas computer simulation is a useful tool for studying
the implications of formal models. This tool happens to be particularly well suited
to explore and analyse agent-based models for the reasons explained above. Run-
ning an agent-based model in a computer provides a formal proof that a particular
micro-specification is sufficient to generate the global behaviour that is observed
2
The reader can see an interesting comparative analysis between agent-based and equation-based
modelling in (Parunak et al. 1998).
Fig. 6.2 In agent based modelling the entities of the system are represented explicitly and
individually in the model. The limits of the entities in the target system correspond to the limits
of the agents in the model, and the interactions between entities correspond to the interactions of
the agents in the model (Edmonds 2001)
during the simulation. If a model can be run in a computer, then it is in principle

possible to express it in many different formalisms, e.g. as a set of mathematical
equations. Such equations may be very complex, difficult to interpret and impossi-
ble to solve, thus making the whole exercise of changing formalism frequently
pointless, but what we find indeed useful is the thought that such an exercise could
be undertaken, i.e. an agent-based model that can be run in a computer is not that
different from the typical mathematical model. As a matter of fact, it is not difficult
to formally characterise most agent-based models in a general way (Leombruni and
Richiardi 2005).
6.3.2 Design, Implementation, and Use of an Agent-Based Model
Drogoul et al. (2003) identify three different roles in the design, implementation,
and use of a typical agent-based model: the thematician (domain expert), the
modeller, and the computer scientist. It is not unusual in the field to observe that
Fig. 6.3 Different stages in

the process of designing,
implementing and using an
agent-based model
one single person undertakes several or even all of these roles. We find that these
three roles fit particularly well into the framework put forward by Edmonds (2001)
to describe the process of modelling with an intermediate abstraction. Here we
marry Drogoul et al.’s and Edmonds’ views on modelling by dissecting one of
Drogoul et al.’s roles and slightly expanding Edmonds’ framework (Fig. 6.3). We
then use our extended framework to identify the different types of assumptions that
are made in each of the stages of the modelling process, the errors and artefacts that
may occur in each of them, and the activities that can be conducted to avoid such
errors and artefacts. We start by explaining the three different roles proposed by
Drogoul et al. (2003).
The role of the thematician is undertaken by experts in the target domain. They
are the ones that better understand the target system, and therefore the ones who
carry out the abstraction process that is meant to produce the first conceptualisation
of the target system. Their job involves defining the objectives and the purpose of
the modelling exercise, identifying the critical components of the system and the
linkages between them, and also describing the most prominent causal relations.
The output of this first stage of the process is most often a non-formal model
expressed in natural language, and it may also include simple conceptual diagrams,
e.g. block diagrams. The non-formal model produced may describe the system
using potentially ambiguous terms (such as e.g. learning or imitation, without fully
specifying how these processes actually take place).
The next stage in the modelling process is carried out by the role of the modeller.
The modeller’s task is to transform the non-formal model that the thematician aims
to explore into the (formal) requirement specifications that the computer scientist –
the third role – needs to formulate the (formal) executable model. This job involves
(at least) three major challenges. The first one consists in acting as a mediator
between two domains that are very frequently fundamentally different (e.g. sociol-
ogy and computer science). The second challenge derives from the fact that in most
cases the thematician’s model is not fully specified, i.e. there are many formal
models that would conform to it.3 In other words, the formal model created by
the modeller is most often just one of many possible particularisations of the
thematician’s (more general) model. Lastly, the third challenge appears when the
thematician’s model is not consistent, which may perfectly be the case since his
model is often formulated using natural language. Discovering inconsistencies in
natural language models is in general a non-trivial task. Several authors (e.g.
Christley et al. (2004), Pignotti et al. (2005), and Polhill and Gotts (2006)) have
identified ontologies to be particularly useful for this purpose, especially in the
domain of agent-based social simulation. Polhill and Gotts (2006) write:
An ontology is defined by Gruber (1993) as “a formal, explicit specification of a shared
conceptualisation”. Fensel (2001) elaborates: ontologies are formal in that they are machine
readable; explicit in that all required concepts are described; shared in that they represent an
agreement among some community that the definitions contained within the ontology
match their own understanding; and conceptualisations in that an ontology is an abstraction
of reality. (Polhill and Gotts 2006, p. 51)
Thus, the modeller has the difficult –potentially unfeasible– task of finding a set
of (formal and consistent) requirement specifications4 where each individual
requirement specification of that set is a legitimate particular case of the
thematician’s model, and the set as a whole is representative of the thematician’s
specifications (i.e. the set is sufficient to fully characterise the thematician’s model
to a satisfactory extent).
Drogoul et al.’s third role is the computer scientist. Here we distinguish between
computer scientist and programmer. It is often the case that the modeller comes up
with a formal model that cannot be implemented in a computer. This could be, for
example, because the model uses certain concepts that cannot be operated by
present-day computers (e.g. real numbers, as opposed to floating-point numbers),
or because running the model would demand computational requirements that are
not yet available (e.g. in terms of memory and processing capacity). The job of the
computer scientist consists in finding a suitable (formal) approximation to the
modeller’s formal model that can be executed in a computer (or in several
computers) given the available technology. To achieve this, the computer scientist
may have to approximate and simplify certain aspects of the modeller’s formal
model, and it is his job to make sure that these simplifications are not affecting the
results significantly. As an example, Cioffi-Revilla (2002) warns about the poten-
tially significant effects of altering system size in agent-based simulations.
The Navier–Stokes equations of fluid dynamics are a paradigmatic case in point.
They are a set of non-linear differential equations that describe the motion of a fluid.
Although these equations are considered a very good (formal and fully specified)
3
Note that the thematician faces a similar problem when building his non-formal model. There are
potentially an infinite number of models for one single target system.
4
Each individual member of this set can be understood as a different model or, alternatively, as a
different parameterisation of one single –more general– model that would itself define the whole
set.
model, their complexity is such that analytical closed-form solutions are available
only for the simplest cases. For more complex situations, solutions of the
Navier–Stokes equations must be estimated using approximations and numerical
computation (Heywood et al. 1990; Salvi 2002). Deriving such approximations
would be the task of the computer scientist’s role, as defined here.
One of the main motivations to distinguish between the modeller’s role and the
computer scientist’s role is that, in the domain of agent-based social simulation, it is
the description of the modeller’s formal model what is usually found in academic
papers, even though the computer scientist’s model was used by the authors to
produce the results in the paper. Most often the modeller’s model (i.e. the one
described in the paper) simply cannot be run in a computer; it is the (potentially
faulty) implementation of the computer scientist’s approximation to such a model
what is really run by the computer. As an example, note that computer models
described in scientific papers are most often expressed using equations in real
arithmetic, whereas the models that actually run in computers almost invariably
use floating-point arithmetic.
Finally, the role of the programmer is to implement the computer scientist’s
executable model. In our framework, by definition of the role computer scientist,
the model he produces must be executable and fully specified, i.e. it must include all
the necessary information so given a certain input the model always produces the
same output. Thus, the executable model will have to specify in its definition
everything that could make a difference, e.g. the operating system and the specific
pseudo-random number generator to be used. This is a subtle but important point,
since it implies that the programmer’s job does not involve any process of abstrac-
tion or simplification; i.e. the executable model and the programmer’s
specifications are by definition the same (see Fig. 6.3). (We consider two models
to be the same if and only if they produce the same outputs when given the same
inputs.) The programmer’s job consists “only” in writing the executable model in a
programming language.5 If the programmer does not make any mistakes, then the
implemented model (e.g. the code) and the executable model will be the same.
Any mismatch between someone’s specifications and the actual model he passes
to the next stage is considered here an error (see Fig. 6.3). As an example, if the
code implemented by the programmer is not the same model as his specifications,
then there has been an implementation error. Similarly, if the computer scientist’s
specifications are not complete (i.e. they do not define a unique model that produces
a precise set of outputs for each given set of inputs) we say that he has made an error
since the model he is producing is necessarily fully specified (by definition of the
role). This opens up the question of how the executable model is defined: the
executable model is the same model as the code if the programmer does not
5
There are some interesting attempts with INGENIAS (Pavón and Gómez-Sanz 2003) to use
modelling and visual languages as programming languages rather than merely as design languages
(Sansores and Pavón 2005; Sansores et al. 2006. These efforts are aimed at automatically
generating several implementations of one single executable model (in various different simula-
tion platforms).
make any mistakes. So, to be clear, the distinction between the role of computer
scientist and programmer is made here to distinguish (a) errors in the implementa-
tion of a fully specified model (which are made by the programmer) from (b) errors
derived from an incomplete understanding of how a computer program works
(which are made by the computer scientist). An example of the latter would be
one where the computer scientist’s specifications stipulate the use of real arithmetic,
but the executable model uses floating-point arithmetic.
It is worth noting that in an ideal world the specifications created by each role
would be written down. Unfortunately the world is far from ideal, and it is often the
case that the mentioned specifications stay in the realm of mental models, and never
reach materialisation.
The reason for which the last two roles in the process are called ‘the computer
scientist’ and the ‘programmer’ is because, as mentioned before, most agent-based
models are implemented as computer programs, and then explored through simula-
tion (for tractability reasons). However, one could also think of e.g. a mathemati-
cian conducting these two roles, especially if the formal model provided by the
modeller can be solved analytically. For the sake of clarity, and without great loss of
generality, we assume here that the model is implemented as a computer program
and its behaviour is explored through computer simulation.
Once the computer model is implemented, it is run, and the generated results are
analysed. The analysis of the results of the computer model leads to conclusions on
the behaviour of the computer scientist’s model and, to the extent that the computer
scientist’s model is a valid approximation of the modeller’s formal model, these
conclusions also apply to the modeller’s formal model. Again, to the extent that the
formal model is a legitimate particularisation of the non-formal model created by
the thematician, the conclusions obtained for the modeller’s formal model can be
interpreted in the terms used by the non-formal model. Furthermore, if the
modeller’s formal model is representative of the thematician’s model, then there
is scope for making general statements on the behaviour of the thematician’s
model. Finally, if the thematician’s model is satisfactorily capturing social reality,
then the knowledge inferred in the whole process can be meaningfully applied to
the target system.
In the following section we use our extended framework to identify the different
errors and artefacts that may occur in each of the stages of the modelling process
and the activities that can be conducted to avoid such errors and artefacts.
6.4 Errors and Artefacts
6.4.1 Definition of Error and Artefact and Their Relevance for

Validation and Verification
Since the meanings of the terms validation, verification, error, and artefact are not
uncontested in the literature, we start by stating the meaning that we attribute to
each of them. For us, validation is the process of assessing how useful a model is for
a certain purpose. A model is valid to the extent that it provides a satisfactory range
of accuracy consistent with the intended application of the model (Kleijnen 1995;
Sargent 2003).6 Thus, if the objective is to accurately represent social reality, then
validation is about assessing how well the model is capturing the essence of its
empirical referent. This could be measured in terms of goodness of fit to the
characteristics of the model’s referent Moss et al. (1997).
Verification – sometimes called “internal validation”, e.g. by Taylor (1983),
Drogoul et al. (2003), Sansores and Pavón (2005), or “internal validity”, e.g. by
Axelrod (1997a) – is the process of ensuring that the model performs in the manner
intended by its designers and implementers (Moss et al. 1997). Let us say that a
model is correct if and only if it would pass a verification exercise. Using our
previous terminology, an expression of a model in a language is correct if and only
if it is the same model as the developer’s specifications. Thus, it could well be the
case that a correct model is not valid (for a certain purpose). Conversely, it is also
possible that a model that is not correct is actually valid for some purposes. Having
said that, one would think that the chances of a model being valid are higher if it
performs in the manner intended by its designer. To be sure, according to our
definition of validation, what we want is a valid model, and we are interested in its
correctness only to the extent that correctness contributes to make the model valid.
We also distinguish between errors and artefacts (Galán et al. 2009). Errors
appear when a model does not comply with the requirement specifications self-
imposed by its own developer. In simple words, an error is a mismatch between
what the developer thinks the model is, and what it actually is. It is then clear that
there is an error in the model if and only if the model is not correct. Thus,
verification is the process of looking for errors. An example of an implementation
error would be the situation where the programmer intends to loop through the
whole list of agents in the program, but he mistakenly writes the code so it only runs
through a subset of them. A less trivial example of an error would be the situation
where it is believed that a program is running according to the rules of real
arithmetic, while the program is actually using floating-point arithmetic (Izquierdo
and Polhill 2006; Polhill and Izquierdo 2005; Polhill et al. 2005, 2006).
In contrast to errors, artefacts relate to situations where there is no mismatch
between what the developer thinks a model is and what it actually is. Here the
mismatch is between the set of assumptions in the model that the developer thinks
are producing a certain phenomenon, and the assumptions that are the actual cause
of such phenomenon. We explain this in detail. We distinguish between core and
accessory assumptions in a model. Core assumptions are those whose presence is
believed to be important for the purpose of the model. Ideally these would be the
only assumptions present in the model. However, when producing a formal model it
is often the case that the developer is bound to include some additional assumptions
for the only purpose of making the model complete. We call these accessory
6
See a complete epistemic review of the validation problem in Kleindorfer et al. (1998).
assumptions. Accessory assumptions are not considered a crucial part of the model;
they are included to make the model work. We also distinguish between significant
and non-significant assumptions. A significant assumption is an assumption that is
the cause of some significant result obtained when running the model. Using this
terminology, we define artefacts as significant phenomena caused by accessory
assumptions in the model that are (mistakenly) deemed non-significant. In other
words, an artefact appears when an accessory assumption that is considered non-
significant by the developer is actually significant. An example of an artefact would
be the situation where the topology of the grid in a model is accessory, it is believed
that some significant result obtained when running the model is independent of the
particular topology used (say, e.g. a grid of square cells), but it turns out that if an
alternative topology is chosen (say, e.g. hexagonal cells) then the significant result
is not observed.
The relation between artefacts and validation is not as straight-forward as that
between errors and verification. For a start, artefacts are relevant for validation only
to the extent that identifying and understanding causal links in the model’s referent
is part of the purpose of the modelling exercise. We assume that this is the case, as
indeed it usually is in the field of agent-based social simulation. A clear example is
the Schelling-Sakoda model of segregation, which was designed to investigate the
causal link between individual preferences and global patterns of segregation
(Sakoda 1971; Schelling 1971, 1978). The presence of artefacts in a model implies
that the model is not representative of its referent, since one can change some
accessory assumption (thus creating an alternative model which still includes all the
core assumptions) and obtain significantly different results. When this occurs, we
run the risk of interpreting the results obtained with the (non-representative) model
beyond its scope (Edmonds and Hales 2005). Thus, to the extent that identifying
causal links in the model’s referent is part of the purpose of the modelling exercise,
the presence of artefacts decreases the validity of the model. In any case, the
presence of artefacts denotes a misunderstanding of what assumptions are
generating what results.
6.4.2 Appearance of Errors and Artefacts
The dynamics of agent-based models are generally sufficiently complex that model
developers themselves do not understand in exhaustive detail how the obtained
results have been produced. As a matter of fact, in most cases if the exact results and
the processes that generated them were known and fully understood in advance,
there would not be much point in running the model in the first place. Not knowing
exactly what to expect makes it impossible to tell whether any unanticipated results
derive exclusively from what the researcher believes are the core assumptions in the
model, or whether they are due to errors or artefacts. The question is of crucial
importance since, unfortunately, the truth is that there are many things that can go
wrong in modelling.
Errors and artefacts may appear at various stages of the modelling process
(Galán and Izquierdo 2005). In this section we use the extended framework
explained in the previous section to identify the critical stages of the modelling
process where errors and artefacts are most likely to occur.
According to our definition of artefact – i.e. significant phenomena caused by
accessory assumptions that are not considered relevant –, artefacts cannot appear in
the process of abstraction conducted by the thematician, since this stage consists
precisely in distilling the core features of the target system. Thus, there should not
be accessory assumptions in the thematician’s model. Nevertheless, there could still
be issues with validation if, for instance, the thematician’s model is not capturing
social reality to a satisfactory extent. Errors could appear in this stage because the
thematician’s specifications are usually expressed in natural language, and rather
than being written down, they are often transmitted orally to the modeller. Thus, an
error (i.e. a mismatch between the thematician’s specifications and the non-formal
model received by the modeller) could appear here if the modeller misunderstands
some of the concepts put forward by the thematician.
The modeller is the role that may introduce the first artefacts in the modelling
process. When formalising the thematician’s model, the modeller will often have to
make a number of additional assumptions so the produced formal model is fully
specified. By our definition of the two roles, these additional assumptions are not
crucial features of the target system. If such accessory assumptions have a signifi-
cant impact on the behaviour of the model and the modeller is not aware of it, then
an artefact has been created. This would occur if, for instance, (a) the thematician
did not specify any particular neighbourhood function, (b) different neighbourhood
functions lead to different results, and (c) the modeller is using only one of them and
believes that they all produce essentially the same results.
Errors could also appear at this stage, although it is not very likely. This is so
because the specifications that the modeller produces must be formal, and they are
therefore most often written down in a formal language. When this is the case, there
is little room for misunderstanding between the modeller and the computer scien-
tist, i.e. the modeller’s specifications and the formal model received by the com-
puter scientist would be the same, and thus there would be no error at this stage.
The role of the computer scientist could introduce artefacts in the process. This
would be the case if, for instance, his specifications require the use of a particular
pseudo-random number generator, he believes that this choice will not have any
influence in the results obtained, but it turns out that it does. Similar examples could
involve the arbitrary selection of an operating system or a specific floating-point
arithmetic that had a significant effect on the output of the model.
Errors can quite easily appear in between the role of the computer scientist and
the role of the programmer. Note that in our framework any mismatch between
the computer scientist’s specifications and the executable model received by the
programmer is considered an error. In particular, if the computer scientist’s
specifications are not executable, then there is an error. This could be, for instance,
because the computer scientist’s specifications stipulate requirements that cannot
be executed with present-day computers (e.g. real arithmetic), or because it does not
specify all the necessary information to be run in a computer in an unequivocal way

(e.g. it does not specify a particular pseudo-random number generator). The error
then may affect the validity of the model significantly, or may not.
Note from the previous examples that if the computer scientist does not provide a
fully executable set of requirement specifications, then he is introducing an error,
since in that case the computer program (which is executable) would be necessarily
different from his specifications. On the other hand, if he does provide an execut-
able model but in doing so he makes an arbitrary accessory assumption that turns
out to be significant, then he is introducing an artefact.
Finally, the programmer cannot introduce artefacts because his specifications
are the same as the executable model by definition of the role (i.e. the programmer
does not have to make any accessory assumptions). However, he may make
mistakes when creating the computer program from the executable model.
6.4.3 Activities Aimed at Detecting Errors and Artefacts
In this section we identify various activities that the different roles defined in the
previous sections can undertake to detect errors and artefacts. We consider the use
of these techniques as a very recommendable and eventually easy to apply practice.
In spite of this, we should warn that, very often, these activities may require a
considerable human and computational effort.
6.4.3.1 Modeller’s Activities
• Develop and analyse new formal models by implementing alternative accessory

assumptions while keeping the core assumptions identified by the thematician.
This exercise will help to detect artefacts. Only those conclusions which are not
falsified by any of these models will be valid for the thematician’s model. As an
example, see Galán and Izquierdo (2005), who studied different instantiations of
one single conceptual model by implementing different evolutionary selection
mechanisms. Takadama et al. (2003) conducted a very similar exercise
implementing three different learning algorithms for their agents. In a collection
of papers, Klemm et al. (2003a, b, c, 2005) investigate the impact of various
accessory assumptions in Axelrod’s model for the dissemination of culture
(Axelrod 1997b). Another example of studying different formal models that
address one single problem is provided by Kluver and Stoica (2003).
• Conduct a more exhaustive exploration of the parameter space within the
boundaries of the thematician’s specifications. If we obtain essentially the
same results using the wider parameter range, then we will have broadened
the scope of the model, thus making it more representative of the thematician’s
model. If, on the other hand, results change significantly, then we will have
identified artefacts. This type of exercise has been conducted by e.g. Castellano
et al. (2000) and Galán and Izquierdo (2005).
• Create abstractions of the formal model which are mathematically tractable. An
example of one possible abstraction would be to study the expected motion of a
dynamic system (see the studies conducted by Galán and Izquierdo (2005),
Edwards et al. (2003), and Castellano et al. (2000) for illustrations of mean-
field approximations). Since these mathematical abstractions do not correspond
in a one-to-one way with the specifications of the formal model, any results
obtained with them will not be conclusive, but they may suggest parts of the
model where there may be errors or artefacts.
• Apply the simulation model to relatively well understood and predictable
situations to check that the obtained results are in agreement with the expected
behaviour (Gilbert et al. 2000).
6.4.3.2 Computer Scientist’s Activities
• Develop mathematically tractable models of certain aspects, or particular cases,

of the modeller’s formal model. The analytical results derived with these models
should match those obtained by simulation; a disparity would be an indication of
the presence of errors.
• Develop new executable models from the modeller’s formal model using alter-
native modelling paradigms (e.g. procedural vs. declarative). This activity will
help to identify artefacts. As an example, see Edmonds and Hales’ (2003)
reimplementation of Riolo et al. (2001) model of cooperation among agents
using tags. Edmonds reimplemented the model using SDML (declarative),
whereas Hales reprogrammed the model in Java (procedural).
• Rerun the same code in different computers, using different operating systems,
with different pseudo-random number generators. These are most often accessory
assumptions of the executable model that are considered non-significant, so any
detected difference will be a sign of an artefact. If no significant differences are
detected, then we can be confident that the code comprises all the assumptions
that could significantly influence the results. This is a valuable finding that can be
exploited by the programmer (see next activity). As an example, Polhill et al.
(2005) explain that using different compilers can result in the application of
different floating-point arithmetic systems to the simulation run.
6.4.3.3 Programmer’s Activities
• Re-implement the code in different programming languages. Assuming that the

code contains all the assumptions that can influence the results significantly, this
activity is equivalent to creating alternative representations of the same execut-
able model. Thus, it can help to detect errors in the implementation. There are
several examples of this type of activity in the literature. Bigbee et al. (2007)
reimplemented Sugarscape (Epstein and Axtell 1996) using MASON. Xu et al.
(2003) implemented one single model in Swarm and Repast. The reim-
plementation exercise conducted by Edmonds and Hales (2003) applies here too.
• Analyse particular cases of the executable model that are mathematically tracta-
ble. Any disparity will be an indication of the presence of errors.
• Apply the simulation model to extreme cases that are perfectly understood
(Gilbert et al. 2000). Examples of this type of activity would be to run
simulations without agents or with very few agents, explore the behaviour of
the model using extreme parameter values, or model very simple environments.
This activity is common practice in the field.
6.5 Summary
The dynamics of agent-based models are usually so complex that their own
developers do not fully understand how they are generated. This makes it difficult,
if not impossible, to discern whether observed significant results are legitimate
logical implications of the assumptions that the model developer is interested in or
whether they are due to errors or artefacts in the design or implementation of the
model.
Errors are mismatches between what the developer believes a model is and what
the model actually is. Artefacts are significant phenomena caused by accessory
assumptions in the model that are (mistakenly) considered non-significant. Errors
and artefacts prevent developers from correctly understanding their simulations.
Furthermore, both errors and artefacts can significantly decrease the validity of a
model, so they are best avoided.
In this chapter we have outlined a general framework that summarises the
process of designing, implementing, and using agent-based models. Using this
framework we have identified the different type of errors and artefacts that may
occur in each of the stages of the modelling process. Finally, we have proposed
several activities that can be conducted to avoid each type of error or artefact. Some
of these activities include repetition of experiments in different platforms,
reimplementation of the code in different programming languages, reformulation
of the conceptual model using different modelling paradigms, and mathematical
analyses of simplified versions or particular cases of the model. Conducting these
activities will surely increase our understanding of a particular simulation model.
Acknowledgements The authors have benefited from the financial support of the Spanish
Ministry of Education and Science (projects DPI2004–06590, DPI2005–05676 and
TIN2008–06464–C03–02), the Spanish Ministry for Science and Innovation (CSD2010–00034)
within the framework of CONSOLIDER-INGENIO 2010 and of the JCyL (projects VA006B09,
BU034A08 and GREX251–2009). We are also very grateful to Nick Gotts, Gary Polhill, Bruce
Edmonds and Cesáreo Hernández for many discussions on the philosophy of modelling.
Further Reading
Gilbert (2007) provides an excellent basic introduction to agent based modelling.

Chapter 3 summarizes the different stages involved in an agent-based modelling
project, including verification and validation. The paper entitled “Some myths and
common errors in simulation experiments” (Schmeiser 2001) discusses briefly
some of the most common errors found in simulation from a probabilistic and
statistical perspective. The approach is not focused specifically on agent based
modelling but on simulation in general. Yilmaz (2006) presents an analysis of the
life cycle of a simulation study and proposes a process-centric perspective for the
validation and verification of agent-based computational organization models.
Finally, Chap. 8 in this volume (David 2013) discusses validation in detail.
References
Axelrod RM (1997a) Advancing the art of simulation in the social sciences. In: Conte R,
Hegselmann R, Terna P (eds) Simulating social phenomena, vol 456, Lecture notes in
economics and mathematical systems. Springer, Berlin, pp 21–40
Axelrod RM (1997b) The dissemination of culture: a model with local convergence and global
polarization. J Confl Resolut 41(2):203–226
Axtell RL (2000) Why agents? On the varied motivations for agent computing in the social
sciences. In: Macal CM, Sallach D (eds) Proceedings of the workshop on agent simulation:
applications, models, and tools. Argonne National Laboratory, Argonne, pp 3–24
Axtell RL, Epstein JM (1994) Agent based modeling: understanding our creations. The Bulletin of
the Santa Fe Institute, Winter, pp 28–32
Bigbee T, Cioffi-Revilla C, Luke S (2007) Replication of sugarscape using MASON. In: Terano T,
Kita H, Deguchi H, Kijima K (eds) Agent-based approaches in economic and social complex
systems IV: post-proceedings of the AESCS international workshop 2005. Springer, Tokyo, pp
183–190
Bonabeau E (2002) Agent-based modeling: methods and techniques for simulating human
systems. Proc Natl Acad Sci U S A 99(2):7280–7287
Castellano C, Marsili M, Vespignani A (2000) Nonequilibrium phase transition in a model for
social influence. Phys Rev Lett 85(16):3536–3539
Christley S, Xiang X, Madey G (2004) Ontology for agent-based modeling and simulation. In:
Macal CM, Sallach D, North MJ (eds) Proceedings of the agent 2004 conference on social
dynamics: interaction, reflexivity and emergence. Argonne National laboratory/The University
of Chicago, Chicago. http://www.agent2005.anl.gov/Agent2004.pdf
Cioffi-Revilla C (2002) Invariance and universality in social agent-based simulations. Proc Natl
Acad Sci U S A 99(3):7314–7316
Conlisk J (1996) Why bounded rationality? J Econ Lit 34(2):669–700
Drogoul A, Vanbergue D, Meurisse T (2003) Multi-agent based simulation: where are the agents?
In: Sichman JS, Bousquet F, Davidsson P (eds) Proceedings of MABS 2002 multi-agent-based
simulation, vol 2581, Lecture notes in computer science. Springer, Bologna, pp 1–15
Edmonds B (2001) The use of models: making MABS actually work. In: Moss S, Davidsson P
(eds) Multi-agent-based simulation, vol 1979, Lecture notes in artificial intelligence. Springer,
Berlin, pp 15–32
Edmonds B (2005) Simulation and complexity: how they can relate. In: Feldmann V, Mühlfeld K
(eds) Virtual worlds of precision: computer-based simulations in the sciences and social
sciences. Lit-Verlag, Münster, pp 5–32
Edmonds B, Hales D (2003) Replication, replication and replication: some hard lessons from
model alignment. J Artif Soc Soc Simulat 6(4). http://jasss.soc.surrey.ac.uk/6/4/11.html
Edmonds B, Hales D (2005) Computational simulation as theoretical experiment. J Math Sociol
29:1–24
Edwards M, Huet S, Goreaud F, Deffuant G (2003) Comparing an individual-based model of
behaviour diffusion with its mean field aggregate approximation. J Artif Soc Soc Simulat 6(4).
Epstein JM (1999) Agent-based computational models and generative social science. Complexity
4(5):41–60
Epstein JM (2008) Why model? J Artif Soc Soc Simul 11(4). http://jasss.soc.surrey.ac.uk/11/4/12.
html
Epstein JM, Axtell RL (1996) Growing artificial societies: social science from the bottom up.
Brookings Institution Press/MIT Press, Cambridge, MA
Fensel D (2001) Ontologies: a silver bullet for knowledge management and electronic commerce.
Springer, Berlin
Galán JM, Izquierdo LR (2005) Appearances can be deceiving: lessons learned re-implementing
Axelrod’s ‘Evolutionary approach to norms’. J Artif Soc Soc Simulat 8(3). http://jasss.soc.
Galán JM et al (2009) Errors and artefacts in agent-based modelling. J Artif Soc Soc Simulat 12(1).
Gilbert N (1999) Simulation: a new way of doing social science. Am Behav Sci 42(10):1485–1487
Gilbert N (2007) Agent-based models. Sage, London
Gilbert N, Terna P (2000) How to build and use agent-based models in social science. Mind Soc 1
(1):57–72
Gilbert N, Troitzsch KG (1999) Simulation for the social scientist. Open University Press,
Buckingham
Gotts NM, Polhill JG, Adam WJ (2003) Simulation and analysis in agent-based modelling of land
use change. In: Online proceedings of the first conference of the European Social Simulation
Association, Groningen, 18–21 Sept 2003. http://www.uni-koblenz.de/~essa/ESSA2003/
gotts_polhill_adam-rev.pdf
Gruber TR (1993) A translation approach to portable ontology specifications. Knowl Acquis 5
(2):199–220
Hare M, Deadman P (2004) Further towards a taxonomy of agent-based simulation models in
environmental management. Math Comput Simulat 64(1):25–40
Hernández C (2004) Herbert A Simon, 1916–2001, y el Futuro de la Ciencia Económica. Revista
Europea De Dirección y Economı́a De La Empresa 13(2):7–23
Heywood JG, Masuda K, Rautmann R, Solonnikov VA (eds) (1990) The Navier–Stokes equations:
theory and numerical methods. In: Proceedings of a conference held at Oberwolfach, FRG,
18–24, Sept 1988 (Lecture notes in mathematics), vol 1431. Springer, Berlin
Holland JH, Miller JH (1991) Artificial adaptive agents in economic theory. Am Econ Rev 81
(2):365–370
Izquierdo LR, Polhill JG (2006) Is your model susceptible to floating point errors? J Artif Soc Soc
Simulat 9(4). http://jasss.soc.surrey.ac.uk/9/4/4.html
Kleijnen JPC (1995) Verification and validation of simulation models. Eur J Oper Res 82
(1):145–162
Kleindorfer GB, O’Neill L, Ganeshan R (1998) Validation in simulation: various positions in the
philosophy of science. Manage Sci 44(8):1087–1099
Klemm K, Eguı́luz V, Toral R, San Miguel M (2003a) Role of dimensionality in Axelrod’s model
for the dissemination of culture. Phys A 327:1–5
Klemm K, Eguı́luz V, Toral R, San Miguel M (2003b) Global culture: a noise-induced transition in
finite systems. Phys Rev E 67(4):045101
Klemm K, Eguı́luz V, Toral R, San Miguel M (2003c) Nonequilibrium transitions in complex
networks: a model of social interaction. Phys Rev E 67(2):026120
Klemm K, Eguı́luz V, Toral R, San Miguel M (2005) Globalization, polarization and cultural drift.
J Econ Dyn Control 29(1–2):321–334
Kluver J, Stoica C (2003) Simulations of group dynamics with different models. J Artif Soc Soc
Leombruni R, Richiardi M (2005) Why are economists sceptical about agent-based simulations?
Phys A 355:103–109
Moss S (2001) Game theory: limitations and an alternative. J Artif Soc Soc Simulat 4(2). http://
jasss.soc.surrey.ac.uk/4/2/2.html
Moss S (2002) Agent based modelling for integrated assessment. Integr Assess 3(1):63–77
Moss S, Edmonds B, Wallis S (1997) Validation and verification of computational models with
multiple cognitive agents (Report no. 97–25). Centre for Policy Modelling, Manchester. http://
cfpm.org/cpmrep25.html
Ostrom T (1988) Computer simulation: the third symbol system. J Exp Soc Psychol 24(5):381–392
Parunak HVD, Savit R, Riolo RL (1998) Agent-based modeling vs. equation-based modeling: a
case study and users’ guide. In: Sichman JS, Conte R, Gilbert N (eds) Multi-agent systems
and agent-based simulation, vol 1534, Lecture notes in artificial intelligence. Springer, Berlin,
pp 10–25
Pavón J, Gómez-Sanz J (2003) Agent oriented software engineering with INGENIAS. In: Marik
V, Müller J, Pechoucek M (eds) Multi-agent systems and applications III, 3rd international
central and eastern European conference on multi-agent systems, CEEMAS 2003 (Lecture
notes in artificial intelligence), vol 2691. Springer, Berlin, pp 394–403
Pignotti E, Edwards P, Preece A, Polhill JG, Gotts NM (2005) Semantic support for computational
land-use modelling. In: Proceedings of the 5th international symposium on cluster computing
and the grid (CCGRID 2005). IEEE Press, Piscataway, pp 840–847
Polhill JG, Gotts NM (2006) A new approach to modelling frameworks. In: Proceedings of the first
world congress on social simulation, vol 1. Kyoto, 21–25 Aug 2006, pp 215–222
Polhill JG, Izquierdo LR (2005) Lessons learned from converting the artificial stock market to
interval arithmetic. J Artif Soc Soc Simulat 8(2). http://jasss.soc.surrey.ac.uk/8/2/2.html
Polhill JG, Izquierdo LR, Gotts NM (2005) The ghost in the model (and other effects of floating
point arithmetic). J Artif Soc Soc Simulat 8(1). http://jasss.soc.surrey.ac.uk/8/1/5.html
Polhill JG, Izquierdo LR, Gotts NM (2006) What every agent based modeller should know about
floating point arithmetic. Environ Model Software 21(3):283–309
Riolo RL, Cohen MD, Axelrod RM (2001) Evolution of cooperation without reciprocity. Nature
411:441–443
Sakoda JM (1971) The checkerboard model of social interaction. J Math Sociol 1(1):119–132
Salvi R (2002) The Navier–Stokes equation: theory and numerical methods, Lecture notes in pure
and applied mathematics. Marcel Dekker, New York
Sansores C, Pavón J (2005) Agent-based simulation replication: a model driven architecture
approach. In: Gelbukh AF, de Albornoz A, Terashima-Marı́n H (eds) Proceedings of MICAI
2005: advances in artificial intelligence, 4th Mexican international conference on artificial
intelligence, Monterrey, 14–18 Nov 2005. Lecture notes in computer science, vol 3789.
Sansores C, Pavón J, Gómez-Sanz J (2006) Visual modeling for complex agent-based simulation
systems. In: Sichman JS, Antunes L (eds) Multi-agent-based simulation VI, international
workshop, MABS 2005, Utrecht, 25 July 2005, Revised and invited papers. Lecture notes in
computer science, vol 3891. Springer, Berlin, pp 174–189
Sargent RG (2003) Verification and validation of simulation models. In: Chick S, Sánchez PJ,
Ferrin D, Morrice DJ (eds) Proceedings of the 2003 winter simulation conference. IEEE,
Piscataway, pp 37–48
Schelling TC (1971) Dynamic models of segregation. J Math Sociol 1(2):47–186

Schelling TC (1978) Micromotives and macrobehavior. Norton, New York
Schmeiser BW (2001) Some myths and common errors in simulation experiments. In: Peters BA,
Smith JS, Medeiros DJ, Rohrer MW (eds) Proceedings of the winter simulation conference, vol
1. Arlington, 9–12 Dec 2001, pp 39–46
Takadama K, Suematsu YL, Sugimoto N, Nawa NE, Shimohara K (2003) Cross-element valida-
tion in multiagent-based simulation: switching learning mechanisms in agents. J Artif Soc Soc
Taylor AJ (1983) The verification of dynamic simulation models. J Oper Res Soc 34(3):233–242
Xu J, Gao Y, Madey G (2003) A docking experiment: swarm and repast for social network
modeling. In: seventh annual swarm researchers conference (SwarmFest 2003), 13–15 Apr
2003, Notre Dame
Yilmaz L (2006) Validation and verification of social processes within agent-based computational
organization models. Comput Math Organ Theory 12(4):283–312
Chapter 7
Documenting Social Simulation Models:
The ODD Protocol as a Standard
Volker Grimm, Gary Polhill, and Julia Touza
Why Read This Chapter? To learn about the importance of documenting your
simulation model and discover a lightweight and appropriate framework to guide you
in doing this.
Abstract The clear documentation of simulations is important for their communica-
tion, replication, and comprehension. It is thus helpful for such documentation to
follow minimum standards. The “Overview, Design concepts and Details” document
protocol (ODD) is specifically designed to guide the description of individual- and
agent-based simulation models (ABMs) in journal articles. Popular among ecologists,
it is also increasingly used in the social simulation community. Here, we describe the
protocol and give an annotated example of its use, with a view to facilitating its wider
adoption and encouraging higher standards in simulation description.
7.1 Introduction and History
A description protocol is a framework for guiding the description of something, in

this case a social simulation model. It can be thought of as a check-list of things that
need to be covered and rules that should be followed when specifying the details of
a simulation (in a scholarly communication). Following such a protocol means that
readers can become familiar with its form and that key elements are less likely to be
V. Grimm (*)
UFZ, Helmholtz Centre of Environmental Research – UFZ, Department of Ecological
Modelling, Permoserstr. 15, 04318 Leipzig, Germany
G. Polhill
The James Hutton Institute, Craigiebuckler, Aberdeen AB15 8QH, United Kingdom
J. Touza
Applied Economics Department, University of Vigo, Vigo C.P. 36.310, Spain

118 V. Grimm et al.
forgotten. This chapter describes a particular documentation protocol, the ODD

(pronounced: “odd”, or “oh dee dee”) protocol.
The ODD protocol (Grimm et al. 2006, 2010; Polhill et al. 2008; Polhill 2010) is a
standard layout for describing individual- and agent-based simulation models
(ABMs), especially for journal articles, conference papers, and other academic
literature. It consists of seven elements which can be grouped into three blocks:
Overview, Design concepts, Details (hence, “ODD”; see Table 7.1). The purpose of
ODD is to facilitate writing and reading of model descriptions, to better enable
replication of model-based research, and to establish a set of design concepts that
should be taken into account while developing an ABM. It does this in a relatively
lightweight way, avoiding over-formal approaches whilst ensuring that the essentials
of a simulation are explicitly described in a flexible yet appropriate manner.
Originally, ODD was formulated by ecologists, where the proportion of ABMs
described using ODD is increasing fast and might cross the 50 % margin in the near
future. In social simulation, the acceptance of ODD has been slower. A first test, in
which three existing descriptions of land-use models were re-formulated according
to ODD, demonstrated the benefits of using ODD but also revealed that some
refinements were needed to make it more suitable for social simulation (Polhill
et al. 2008). In 2010, an update of ODD was released (Grimm et al. 2010), which is
based on users’ feedback and a review of more than 50 ODD-based model
descriptions in the literature. In this update, ODD itself was only slightly modified
but the explanation of its elements completely rewritten, with the specific intention
of making it more suitable for social simulation.
Currently in social simulation, interest in ODD is also increasing (Polhill 2010).
An indicator for this is the inclusion of ODD chapters in recent reference books
(this volume; Heppenstall et al. 2012). Moreover, a recent textbook of agent-based
modelling uses ODD consistently (Railsback and Grimm 2012), so that the next
generation of agent-based modellers is more likely to be familiar with ODD, and
hence to use it themselves.
7.2 The Purpose of ODD
Why is ODD (or a protocol very much like it) needed? There are a number of
endeavours in agent-based social simulation that are facilitated through having a
common approach to describing the models that is aimed at being readable and
complete1:
1
Many of these endeavours have been covered in submissions to the “model-to-model” series of
workshops, organised by members of the social simulation community (Hales et al. 2003;
Rouchier et al. 2008. The second workshop was held as a parallel session of the ESSA 2004
conference: see http://www.insisoc.org/ESSA04/M2M2.htm).
Table 7.1 The seven elements of the ODD protocol. Descriptions of ABMs are compiled by answering the questions linked to each element
Overview 1. Purpose What is the purpose of the model?
2. Entities, state variables, scales What kind of entities are in the model? Do they represent managers, voters, landowners, firms or
something else? By what state variables, or attributes, are these entities characterized? What
are the temporal and spatial resolutions and extents of the model?
3. Process overview, scheduling What entity does what, in what order? Is the order imposed or dynamic? When are state variables
updated? How is time modelled: as discrete steps or as a continuum over which both
continuous processes and discrete events can occur?
Design concepts 4. Design Basic Which general concepts, theories or hypotheses are included in the model’s design? How were
concepts principles they taken into account? Are they used at the level of submodels
or at the system level?
Emergence What key results are emerging from the adaptive traits, or behaviours of individuals? What
results vary in complex/unpredictable ways when particular characteristics change? Are there
other results that are more tightly imposed by model rules and hence less dependent on what
individuals do?
Adaptation What adaptive traits do the individuals have? What rules do they have for making decisions or
changing behaviour in response to changes in themselves or their environment? Do agents
seek to increase some measure of success or do they reproduce observed behaviours that they
perceive as successful?
Objectives If agents (or groups) are explicitly programmed to meet some objective, what exactly is that and
how is it measured? When individuals make decisions by ranking alternatives, what criteria
do they use? Note that the objective of such agents as group members may not refer to
themselves but the group
Learning May individuals change their adaptive traits over time as a consequence of their experience? If
so, how?
Prediction Prediction can be part of decision-making; if an agent’s learning procedures are based on
7 Documenting Social Simulation Models: The ODD Protocol as a Standard
estimating future consequences of decisions, how they do this? What internal models do
agents use to estimate future conditions or consequences? What ‘tacit’ predictions are
implied in these internal model’s assumptions?
Sensing What aspects are individuals assumed to sense and consider? What aspects of which other
entities can an individual perceive (e.g. displayed ‘signals’)? Is sensing local, through
networks or global? Is the structure of networks imposed or emergent? Are the mechanisms
by which agents obtain information modelled explicitly in a process or is it simply ‘known’?
119
(continued)
Table 7.1 (continued)
120
Interaction What kinds of interactions among agents are assumed? Are there direct interactions where
individuals encounter and affect others, or are interactions indirect, e.g. via competition for a
mediating resource? If the interactions involve communication, how are such
communications represented?
Stochasticity What processes are modelled by assuming they are random or partly random? Is stochasticity
used, for example, to reproduce variability in processes for which it is unimportant to model
the actual causes of the variability, or to cause model events or behaviours to occur with a
specified frequency?
Collectives Do the individuals form or belong to aggregations that affect, and are affected by, the
individuals? Such collectives can be an important intermediate level of organization. How are
collectives represented – as emergent properties of the individuals or as a separate kind of
entity with its own state variables and traits?
Observation What data are collected from the ABM for testing, understanding, and analyzing it, and how are
they collected? Are all output data freely used, or are only certain data sampled and used, to
imitate what can be observed in an empirical study?
Details 5. Initialization What is the initial state of the model world, i.e., at time t ¼ 0? How many entities of what type
are there initially, and what are the values of their state variables (or how were they set)? Is
initialization always the same, or is it varied? Are the initial values chosen arbitrarily or based
on available data?
6. Input data Does the model use input from external sources such as data files or other models to represent
processes that change over time?
7. Submodels What are the submodels that represent the processes listed in ‘Process overview and scheduling’?
What are the model parameters, their dimensions, and reference values? How were
submodels designed or chosen, tested, and parameterized?
V. Grimm et al.
7 Documenting Social Simulation Models: The ODD Protocol as a Standard 121
• Communication is the most basic aim of anyone trying to publish their results.
For agent-based modellers, this can pose a particular challenge, as our models
can be complicated, with many components and submodels. As a critical mass of
papers using ODD develops, so readers of agent-based modelling papers will
find themselves sufficiently more familiar with papers structured using ODD
than those using an arbitrary layout devised by the authors that they will find the
former easier to read and understand than the latter.
• Replication, as we discuss later in this chapter, is a pillar of the scientific
endeavour. If our model descriptions are inadequate, our results are not repeat-
able, and the scientific value of our work commensurately reduced. ODD helps
to encourage the adequacy of descriptions by saving authors having to ‘reinvent
the wheel’ each time they describe a model, by providing a standard layout
designed to ensure that all aspects of a model needed to replicate it are included
in the account.
• Comparing models is likely to become increasingly important as work in agent-
based modelling continues. If two or more research teams produce similar
models with different outcomes, comparing the models will be essential to
identifying the cause of the variance in behaviour. Such comparisons will be
much easier if all teams have used the same protocol to describe the models. At a
conceptual level, the design concepts also enable comparison of models with
greater differences and application domains.
• Dialogue among disciplines can be encouraged through a standard that is used by
both the ecological and social simulation communities. This is especially useful
for those developing coupled socio-ecosystem models (Polhill et al. 2008),
which is a rapidly growing area of research (Polhill et al. 2011).
In the following, we briefly describe the rationale of ODD and how it is used,
provide an example model description, and finally discuss benefits of ODD, current
challenges, and its potential future development.
7.3 The ODD Protocol
A core principle of ODD is that first an ‘Overview’ of a model’s purpose, structure

and processes should be provided, before ‘Details’ are presented. This allows
readers to quickly get a comprehensive overview of what the model is, what it
does, and for what purpose it was developed. This follows the journalistic ‘inverted
pyramid’ style of writing, where a summary is provided in the first one or two
paragraphs, and progressively further detail is added on the story the further on you
read (see, e.g. Wheeler 2005). It allows the reader to easily access the information
they are interested in at the level of detail they need. For experienced modellers, this
overview part is sufficient to understand what the model is for, to relate it to other
models in the field, and to assess the overall design and complexity.
122 V. Grimm et al.
Before presenting the ‘Details’, ODD requires a discussion of whether, and how,
ten design concepts were taken into account while designing the model. This
‘Design concepts’ part of ODD does not describe the model itself but the principles
and rationale underlying its design. ‘Design concepts’ is thus not needed for model
replication but for making sure that important design decisions were made con-
sciously and that readers are fully aware of these decisions. For example, it is
important to be clear about what model output is designed to emerge from the
behaviour the model’s entities and their interactions, and what, in contrast, is
imposed by fixed rules and parameters. Ideally, key behaviours in a model emerge,
whereas other elements might be imposed. If modellers are not fully aware of this
difference, which is surprisingly often the case, they might impose too much so that
model output is more or less hard-wired into its design, or they might get lost in a too
complex model because too much emergence makes it hard to understand anything.
Likewise, the design concept ‘stochasticity’ requires that modellers explicitly say
what model processes include a stochastic component, why stochasticity was used,
and how it was implemented. Note that, in contrast to the seven elements of ODD,
the sequence in which design concepts are described can be changed, if needed, and
design concepts that are not relevant for the model can be omitted.
The ‘Details’ part of ODD includes all details that are needed to re-implement
the model. This includes information about the values of all model entities’ state
variables and attributes at the begin of a simulation (‘Initialisation’), the external
models or data files that are possibly used as ‘Input data’ describing the dynamics of
one or more driving contextual or environmental variables (e.g., rainfall, market
price, disturbance events), and ‘Details’ where the submodels representing the
processes listed in ‘Process overview and scheduling’ are presented. Here, it is
recommended for every submodel to start with the factual description of what the
submodel is and then explain its rationale.
Model parameters should be presented in a table, referred to in the ‘Submodels’
section of ODD, including parameter name, symbol, reference value, and – if the
model refers to real systems – unit, range, and references or sources for choosing
parameter values. Note that the simulation experiments that were carried out to
analyse the model, characterized by parameter settings, number of repeated runs,
the set of observation variables used, and the statistical analyses of model output, is
not part of ODD but ideally should be presented in a section ‘Simulation
experiments’ directly following the ODD-based model description.
7.4 How to Use ODD
To describe an ABM using ODD, the questions listed in Table 7.1 have to be
answered. The identifiers of the three blocks of ODD elements – Overview, Design
concepts, Details – are not used themselves in ODD descriptions (except for
‘Design concepts’, which is the only element of the corresponding block). Rather,
the seven elements are used as headlines in ODD-based model descriptions. For
experienced ODD users, the questions in Table 7.1 are sufficient. For beginners,
however, it is recommended to read the more detailed description of ODD in
Grimm et al. (2010) and to use the template, which provides additional questions
and examples, and which is available via download.2
7.5 An Example
In the supplementary material of Grimm et al. (2010), publications are listed which
use ODD in a clear, comprehensive, and recommendable way. Many further
examples are provided in the textbook by Railsback and Grimm (2012). In Grimm
and Railsback (2012), Schelling’s segregation model, as implemented in the model
library of the software platform NetLogo (Wilensky 1999), is used as an example.
Here, we demonstrate the process of model documentation using ODD by describing
a model developed by Deffuant et al. (2002), which explores the emergence of
extreme opinions in a population. We choose this model because it is simple but
interesting and opinion dynamics models are quite well-known in the social simula-
tion community. It is also one of the introductory examples in Gilbert (2007). The
ODD for the Deffuant et al. model is interspersed with comments on the information
included, with a view to providing some guidelines for those applying ODD to their
own model. Clearly this is a very simple example and many models would require
more extensive description. The parts of ODD are set in italics and indented to
distinguish them from comments. Normally the ODD description would simply
form part of the text in the main body of a paper or in an appendix.3
7.5.1 Purpose
The model’s purpose is to study the evolution of the distribution of opinions in a population
of interacting individuals, which is under the influence of extremists’ views. Specifically, it
aims to answer how marginal extreme opinions can manage to become the norm in large
parts of a population. The central idea of the model is that people who have more extreme
opinions are more confident than people with moderate views. More confident people are,
however, assumed to more easily affect the opinion of others, who are less confident.
Comments: The purpose section is deliberately brief. Even for more sophisticated
models than this, we would not expect to see much more text here. This would
2
E.g. http://www.ufz.de/index.php?de¼10466.
3
It is often the case that a substantial description needs to be included in the main text so readers
can get an idea of what is being discussed, but maybe a more complete description might be added
in an appendix.
124 V. Grimm et al.
otherwise repeat information in the rest of the paper. However, since the ODD, to
some extent, needs to stand alone and be comprehensive, the summary of the
purpose is included as here.
7.5.2 Entities, State Variables, and Scales
The model includes only one type of entity: individuals. They are characterised by two continu-
ous state variables, opinion x and uncertainty u. Opinions range from 1 to 1. Individuals with
an opinion very close to x ¼ 1 or +1 are referred to as “extremists”, all other individuals are
“moderates”. Uncertainty u defines an interval around an individuals’ opinion and determines
whether two individuals interact and, if they do, on the relative agreement of those two
individuals which then determines how much opinion and uncertainty change in the interaction.
One time step of the model represents the time in which all individuals have randomly chosen
another individual and possibly interacted with it. Simulations run until the distribution of
opinions becomes stationary.
Comments: For larger models, this section has the potential to get quite long if
written in the same style as this example, which has only one type of entity, with
two state variables. Other articles have taken the approach of using tables to express
this information; one table per entity, with one row per state variable associated
with that entity (see, e.g. Polhill et al. 2008). Other articles have used UML class
diagrams (e.g., Bithel et al. 2009), as suggested in the original ODD article (Grimm
et al. 2006); however, these do not provide a means for giving any description,
however brief, of each state variable. Simply listing the entities and the data types
of the state variables does not provide all the information that this element of ODD
should provide. This, together with the fact that UML is focused on Object-
Oriented Design (which is used to implement the majority of ABMs, but by no
means all: NetLogo, for example, is not an object-oriented language, and many,
particularly in agent-based social simulation, use declarative programming
languages), meant that the recommendation to use UML was retracted in the recent
ODD update (Grimm et al. 2010).
In declarative programming languages, the entities and their state variables may
not be so explicitly represented in the program code as they are in object-oriented
languages. For example, this information may be implicit in the arguments to rules.
However, many declarative programs have a database of knowledge that the rules
operate on. This database could be used to suggest entities and state variables. For
example, a Prolog program might have a database containing the assertions person
(volker) and nationality(volker, german). This suggests that ‘person’ is an entity,
and ‘nationality’ a state variable. (It might be reasonable to suggest in general that
assertions with one argument suggest entities, and those with two, state variables.)
7.5.3 Process Overview and Scheduling
In each time step each individual chooses randomly one other individual to interact with,
then the relative agreement between these two agents is evaluated, and the focal
individual’s opinion and uncertainty are immediately updated as a result of this opinion
interaction. Updating of state variables is thus asynchronous. After all individuals have
interacted, a convergence index is calculated which captures the level of convergence in the
opinions of the population; additionally, and output is updated (e.g.: draw histogram of the
population’s opinions; write each individual’s opinion to a file.)
Comments: This section briefly outlines the processes (or submodels) that the
model runs through in every time step (ignoring initialisation), and in what order.
Notice how each process is given an emphasized label, which corresponds to
subsection headings in the Submodels section. Whilst the ODD protocol does not
make such precise stipulations as to formatting, there should be a clear one-to-one
correspondence between the brief outlines of processes here, and the details
provided on each in the Submodels section.
In describing larger models than Deffuant et al.’s, it may be appropriate to simply
present the process overview as a list. Many models have a simple schedule structure
consisting of a repeated sequence of actions; such a list would clearly show this
schedule. However, others use more complicated scheduling arrangements (e.g.
dynamic scheduling). In such cases, the rules determining when new events are
added to the schedule would need to be described, as well as an (unordered) list of
event types, each corresponding to a subsection of ‘Submodels’.
The ‘schedule’ in a declarative model may be even less clear, as it will depend on
how the inference engine decides which rules to fire. However, declarative programs
are at least asked a query to start the model, and this section would be an appropriate
place to mention that. Some declarative programs also have an implied ordering to
rule firing. For example, in Prolog, the rule a :- x, y, z. will, in the event that the
inference engine tries to prove a, try to prove x, then y, then z. Suppose the model is
started with the query ?- a. In describing the model here, it might suffice simply to
summarise how x, y and z change the state of the model. Any subrules called by the
inference engine trying to prove these could be given attention in the Details section.
The declarative programmer may also use language elements (such as cuts in
Prolog) to manage the order of execution. In deciding which rules to describe here,
a declarative modeller might focus on those changing the value of a state variable
over time. The key point is that the program will do something to change the values
of state variables over time in the course of its execution. Insofar as that can be
described in a brief overview, it belongs here.
7.5.4 Design Concepts
Basic principles. – This model extends earlier stylised models on opinion dynamics, which
either used only binary opinions instead of a continuous range of opinions, or where
126 V. Grimm et al.
interactions only depended on whether opinion segments overlapped, but not on relative
agreement (for references, see Deffuant et al. 2002).
Emergence. – The distribution of opinions in the population emerges from interactions
among the individuals.
Sensing. – Individuals have complete information of their interaction partner’s opinion
and uncertainty.
Interaction. – Pairs of individuals interact if their opinion segments, [x u, x + u],
overlap.
Stochasticity. – The interaction between individuals is a stochastic process because
interaction partners are chosen randomly.
Observation. – Two plots are used for observation: the histogram of opinions, and the
trajectories of each individual’s opinion. Additionally, a convergence index is calculated.
Comments: Note that the design concepts are only briefly addressed. This would be
expected in larger models too. Note also that several design concepts have been
omitted because they are not appropriate to the model. Specifically, adaptation,
objectives, learning, prediction, and collectives have been left out here: individuals
change their opinion after interaction, but this change is not adaptive since it is not
linked to any objective; there also no collectives since all individuals act on their
own. Nevertheless, most models should be able to relate to some basic principles,
emergence, interactions, and observation, and most often also stochasticity. Small
models might use the option of concatenating the design concepts into a single
paragraph to save space.
7.5.5 Initialization
Simulations are run with 1,000 individuals, of which a specified initial proportion, pe, are
extremists; p+ denotes the proportion of ‘positive’ extremists, and p are the proportion of
‘negative’ extremists. Each moderate individual’s initial opinion is drawn from a random
uniform distribution between 1 and +1 (not inclusive). Extremists have on opinion of
either 1 or +1. Initially, individuals have a uniform uncertainty, which is larger for
moderates than for extremists.
Comments: This explains how the simulation is set up before the main schedule
starts. In other models, this might include empirical data of various kinds from, for
example, surveys. The key question to ask here, particularly given the potential for
confusion with the next section (‘input data’), is whether the data are used only to
provide a value for a state variable before the schedule runs.
7.5.6 Input Data
The model does not include any input of external data.
Comments: These are time-series data used to ‘drive’ the model. Some of these
data may specify values for variables at time 0 (i.e. during initialisation); however,
if a data series specifies values for any time step other than during initialisation,
then it is input data rather than initialisation. It is also important not to confuse
‘Input data’ with parameter values.
7.5.7 Submodels
All model parameters are listed in the following table.
Parameter Description
N Number of individuals in population
U Initial uncertainty of moderate individuals
μ Speed of opinion dynamics
pe Initial proportion of extremists
p+ Initial proportion of positive extremists
p Initial proportion of negative extremists
ue Initial uncertainty of extremists
Opinion interaction. – This is run for an agent j, whose ‘opinion segment’ sj is defined in
terms of its opinion xj and uncertainty uj as:

s j ¼ x j uj ; x j þ uj
The length of the opinion segment is 2uj and characterizes an individual’s overall
uncertainty.
In opinion interaction, agent j (the influenced, focal, or ‘calling’ individual) is paired with a
randomly chosen agent, i, the influencing individual. The ‘overlap’ of their opinion segments,
hij, is then computed as:

hij ¼ min xi þ ui ; xj þ uj max xi ui ; xj uj
This overlap determines whether an opinion interaction will take place or not: Agent j will
change its opinion if hij > ui, which means that overlap of opinions is higher than the
uncertainty of the influencing agent (see Fig. 7.1).
For opinion interactions, the relative agreement of the two agents’ opinions, RA, is
calculated by dividing the overlap of their opinion segments (hij) minus the length of the
non-overlapping part of influencing individual’s opinion segment, (2ui – hij), and this
difference divided by agent i’s opinion segment length, 2ui (Fig.7.1 depicts these terms
graphically):

RA ¼ hij 2ui hij =2ui ¼ 2 hij ui =2ui ¼ hij =ui 1
The opinion and uncertainty of agent j are then updated as follows:

xj ¼ xj þ μRA xi þ xj

uj ¼ uj þ μRA ui þ uj
128 V. Grimm et al.
Fig. 7.1 Visualisation of the

individual’s opinions,
uncertainties, and overlap in
opinions in the model of
Deffuant et al. (2002)
Thus, the new values are determined by the old values and the sum of the old values of both
interacting individuals multiplied by the relative agreement, RA, and by parameter μ, which
determines how fast opinions change.
The main features of this interaction model are, according to Deffuant et al. (2002):
• Individuals not only influence each other’s opinions but also each other’s uncertainties.
• Confident agents, who have low uncertainty, are more influential. This reflects the
common observation that confident people more easily convince more uncertain people
than the other way round – under the conditions that their opinions are not too different
at the beginning.
Calculate convergence index. – This index, y, is used as a summary model output for
sensitivity analysis and an exploration of the model’s parameter space. It is defined as:
y ¼ qþ þ q
where q+ and q are the proportions of initially moderate agents which become extremists in
the positive extreme or negative extreme, respectively. If after reaching the steady state none
of the initially moderate agents became extremist the index would take a value of zero. If half
of them become positive extremists and the other half becomes negative extremists, the index
would be 0.5. Finally, if all the initially moderate agents converge to only one extreme, the
index would be one. Note that for calculating y, “positive” or “negative” extreme has to be
defined via an interval close the extreme, with a width of, for example, 0.15.
Comments: Here, details on the two processes described in Sect. 7.3 are provided,
in sufficient depth to enable replication, i.e. opinion interaction and calculate
convergence index. Note how these names match with those used in the process
overview in Sect. 7.3.
Authors describing larger models may find journal editors protesting at the length
of the ODD if all submodels are described in the detail required. There are various
ways such constraints can be handled. One is to include the submodels in an appendix
or supplementary material to the paper. Another is to provide them as a technical
report accessible separately (e.g. on a website), and referred to in the text. If space is
not too limited, a summary of each submodel could be provided in the main text,
longer than the brief description in the process overview, but shorter than the full
detail; the latter being provided separately. For very large models, or where space is
highly constrained, there may be little room for much more than the three Overview
sections in the journal article; again, making the full ODD available separately is a
possible solution. Nevertheless, excluding the ‘Submodels’ element entirely from the
main text should be avoided because this would mean to ask readers to accept, in the
main text of the article, the model as a black box. Description of the most important
processes should therefore be included also in the main text.
7.6 Discussion
Since the example model by Deffuant et al. (2002) is very simple, using ODD here
comes with the cost of making the model description longer than the original one,
through requiring the ODD labels. The original model is actually relatively clear
and easy to replicate (which might partly explain this model’s success). However,
easy replication is much more the exception than the rule (Hales et al. 2003;
Rouchier et al. 2008), and the more complex an ABM, the higher the risk that not
all information is provided for unambiguous replication.
ODD facilitates writing comprehensive and clear documentations of ABMs.
This does not only facilitate replication, it also makes writing and reading model
documentations easier. Modellers no longer have to come up with their own format
for describing their model, and readers know, once they are familiar with the
structure of ODD, exactly where to look for what kind of information.
Whether or not to use ODD as a standard format for model descriptions might
look like a rather technical question, but it has fundamental consequences, which go
far beyond the issue of replication. Once ODD is used as a standard, it will be
become much easier to compare different models addressing similar questions.
Even now, ODD can be used to review models in a certain field, by rewriting
existing model descriptions according to ODD (Grimm et al. 2010). Building
blocks of existing models, in particular specific submodels, which seem to be useful
in general, will be much easier to identify and re-use in new models.
Most importantly, however, using ODD affects the way we design and formulate
ABMs in the first place. After having used ODD for documenting two or three
models, you start formulating ABMs by answering the ODD questions: What
‘things’, or entities, do I need to represent in my model? What state variables and
behavioural attributes do I need to characterize these entities? What processes do I
want to represent explicitly, and how should they be scheduled? What are the
spatial and temporal extent and resolution of my model, and why? What do I
want to impose, and what to let emerge? What kind of interactions does the
model include? For what purposes should I include stochasticity? How should the
model world be initialized, what kinds of input data do I need, and how should I, in
detail, formulate my submodels?
130 V. Grimm et al.
These questions do not impose any specific structure on simulation models, but
they provide a clear checklist for both model developers and users. This helps
avoiding “ad hoc-ery” in model design (Heine et al. 2005). Modellers can also more
easily adopt designs of existing models and don’t have to start from scratch all the
time, as in most current social simulation models.
Criticisms of ODD include Amouroux et al. (2010), who, acknowledging its
merits, find the protocol ambiguous and insufficiently specified to enable replica-
tion. This article pertained to the Grimm et al. (2006) first description of ODD. The
update in Grimm et al. (2010) endeavoured to address issues such as these.
However, the success of the latter article in so doing, and indeed any future
revisions of ODD, can only be measured by comparing replication efforts based
on ODD descriptions with those not conforming to any protocol – the norm prior to
2006 when ODD was first published. As suggested above, the record for articles not
using ODD has not been particularly good: Rouchier et al. (2008) observe in their
editorial to a special section of JASSS on the third Model-2-Model workshop that
several researchers attempting replications have to approach the authors of the
original articles to disambiguate model specifications. If the models were ade-
quately described in the original articles, this should not be necessary.
Polhill et al. (2008) also observed that those used to object-oriented designs for
modelling will find the separation of what will for them effectively amount to
instance variables and methods (state variables and processes respectively) counter-
intuitive, if indeed not utterly opposed to encapsulation: one of the key principles of
object orientation. For ODD, however, it is the reader who is important rather than
programming principles intended to facilitate modularity and code reuse. It is also
important that, as a documentation protocol, ODD does not tie itself to any
particular ABM implementation environment. From the perspective of the human
reader, it is illogical (to us at least) to discuss processes before being informed what
it is the processes are operating on. Encapsulation is about hiding information;
ODD has quite the opposite intention.
The main issue with ODD in social simulation circles as opposed to ecology, from
which it originally grew, pertains to its use with declarative modelling environments.
This matter has been raised in Polhill et al. (2008), and acknowledged in Grimm et al.
(2010). Here we have tried to go further towards illustrating how a declarative
modeller might prepare a description of their model that conforms to ODD. However,
until researchers using declarative environments attempt to use ODD when writing an
article, and feedback on their findings, this matter cannot be properly addressed.
Certainly, ODD is not the silver bullet regarding standards for documenting
ABMs. Nevertheless, even at the current stage its benefits by far outweigh its
limitations, and using it more widely is an important condition for further
developments. Still, since ODD is a verbal format, not all ambiguities can
be prevented. Whilst a more formal approach using, for example XML or UML
(e.g. Triebig and Klügl 2010, and for ABMs of land use/cover change, the
MRPOTATOHEAD framework – Livermore 2010; Parker et al. 2008) might address
such ambiguities, we consider it important that written, natural language formulations
of ABMs exist (Grimm and Railsback 2005). This is the only way to make modelling,
as a scientific activity, independent of technical aspects of mark-up or programming
languages and operating systems. Further, verbal descriptions force us to think about a
model, to try to understand what it is, what it does, and why it was designed in that
way and not another (J. Everaars, pers. comm.). We doubt that a ‘technical’ standard
for documenting ABMs – one that can be read by compilers or interpreters, would
ever initiate and require this critical thinking about a model.
Nevertheless, it is already straightforward to translate ODD model description to
NetLogo programs because much of the way models are written in NetLogo
corresponds to the structure of ODD: the declaration of ‘Entities, state variables,
and scales’ is done via NetLogo’s globals, turtles-own, and patches-own primitives,
‘Initialization’ is done via the setup procedure, ‘Process overview and scheduling’
corresponds to the go procedure, ‘Details’ are implemented as NetLogo procedures,
and ‘Design concepts’ can be included, (as indeed can the entire ODD model
description), on the ‘Information’ tab of NetLogo’s user interface.
7.7 Conclusion
Clearly describing simulations well, so that other researchers can understand a

simulation is important for the scientific development and use of complex
simulations. It can help in: the assessment and comprehension of simulation results
by readers; replicating simulations for checking and analysis by other researchers;
transferring knowledge embedded within simulations from one domain to another;
and allowing simulations to be better compared. It is thus an important factor for
making the use of simulations more rigorous and useful. A protocol such as ODD is
useful in standardising such descriptions and encouraging minimum standards. As the
field of social simulation matures it is highly likely that the use of a protocol such as
ODD will become standard practice.
The investment in learning and using ODD is minimal but the benefits, both for
its user and the scientific community, can be huge. We therefore recommend
learning and testing ODD by re-writing the model description of an existing,
moderately complex ABM, and, in particular, using ODD to formulate and docu-
ment the next ABM you are going to develop.
Acknowledgements We are grateful to Bruce Edmonds for inviting us to contribute this chapter,
and for his helpful comments and suggested amendments to earlier drafts. Gary Polhill’s contri-
bution was funded by the Scottish Government.
Further Reading
Railsback and Grimm (2012) is a textbook which introduces agent-based modelling

with examples described using ODD. The OpenABM website (http://openabm.org) is
a portal specifically designed to facilitate the dissemination of simulation code and
132 V. Grimm et al.
descriptions of these using the ODD protocol. The original reference document for
ODD is (Grimm et al. 2006) with the most recent update being (Grimm et al. 2010).
Polhill (2010) is an overview of the 2010 update of ODD written specifically with the
social simulation community in mind.
References
Amouroux E, Gaudou B, Desvaux S, Drogoul A (2010) O.D.D.: a promising but incomplete

formalism for individual-based model specification. In: Ho TB, Zuckerman DN, Kuonen P,
Demaille A, Kutsche R-D (eds) 2010 IEEE-RIVF international conference on computing and
communication technologies: research, innovation and vision for the future, Vietnam National
University, Hanoi, 1–4 Nov 2010
Bithell M, Brasington J (2009) Coupling agent-based models of subsistence farming with
individual-based forest models and dynamic models of water distribution. Environ Model
Software 24:173–190
Deffuant G, Amblard F, Weisbuch G, Faure T (2002) How can extremism prevail? A study based
on the relative agreement interaction model. J Artif Soc Soc Simulat 5(4):1. http://jasss.soc.
Gilbert N (2007) Agent-based models. Sage, London
Grimm V, Railsback SF (2005) Individual-based modeling and ecology. Princeton University
Press, Princeton
Grimm V, Railsback SF (2012) Designing, formulating, and communicating agent-based models.
In: Heppenstall A, Crooks A, See LM, Batty M (eds) Agent-based models of geographical
systems. Springer, Berlin, pp 361–377
Grimm V et al (2010) The ODD protocol: a review and first update. Ecol Model 221:2760–2768
Hales D, Rouchier J, Edmonds B (2003) Model-to-model analysis. J Artif Soc Soc Simulat 6(4):5.
Heine B-O, Meyer M, Strangfeld O (2005) Stylised facts and the contribution of simulation to the
economic analysis of budgeting. J Artif Soc Soc Simulat 8(4):4. http://jasss.soc.surrey.ac.uk/8/
4/4.html
Heppenstall A, Crooks A, See LM, Batty M (eds) (2012) Agent-based models of geographical
systems. Springer, Berlin
Livermore M (2010) MR POTATOHEAD framework: a software tool for collaborative land-use
change modeling. In: Swayne DA, Yang W, Voinov AA, Rizzoli A, Filatova T (eds) Interna-
tional Environmental Modelling And Software Society (iEMSs) 2010 international congress
on environmental modelling and software: modelling for environment’s Sake, Fifth Biennial
Meeting, Ottawa. http://www.iemss.org/iemss2010/index.php?n¼Main,Proceedings
Parker DC et al (2008) Case studies, cross-site comparisons, and the challenge of generalization:
comparing agent-based models of land-use change in frontier regions. J Land Use Sci
3(1):41–72
Polhill JG (2010) ODD updated. J Artif Soc Soc Simulat 13(4):9. http://jasss.soc.surrey.ac.uk/13/
4/9.html
Polhill JG, Parker D, Brown D, Grimm V (2008) Using the ODD protocol for describing three
agent-based social simulation models of land use change. J Artif Soc Soc Simulat 11(2):3.
Polhill JG, Gimona A, Aspinall RJ (2011) Agent-based modelling of land use effects on ecosystem
processes and services. J Land Use Sci 6(2–3):75–81
Railsback SF, Grimm V (2012) Agent-based and individual-based modeling: a practical introduc-
tion. Princeton University Press, Princeton
Rouchier J, Cioffi-Revilla C, Polhill JG, Takadama K (2008) Progress in model-to-model analysis.
J Artif Soc Soc Simulat 11(2):8. http://jasss.soc.surrey.ac.uk/11/2/8.html
Triebig C, Klügl F (2010) Elements of a documentation framework for agent-based simulation.
Cybern Syst 40(5):441–474
Wheeler S (2005) Beyond the inverted pyramid: developing news-writing skills. In: Keeble R (ed)
Print journalism: a critical introduction. Routledge, Abingdon, pp 84–93
Wilensky U (1999) NetLogo. http://ccl.northwestern.edu/netlogo
Chapter 8
Validating Simulations
Nuno David
Why Read This Chapter? To help you decide how to check your simulation –
both against its antecedent conceptual models (verification) and external standards
such as data (validation) – and in this way help you to establish the credibility of
your simulation. In order to do this the chapter will point out the nature of these
processes, including the variety of ways in which people seek to achieve them.
Abstract Verification and validation are two important aspects of model building.
Verification and validation compare models with observations and descriptions of
the problem modelled, which may include other models that have been verified and
validated to some level. However, the use of simulation for modelling social
complexity is very diverse. Often, verification and validation do not refer to an
explicit stage in the simulation development process, but to the modelling process
itself, according to good practices and in a way that grants credibility to using the
simulation for a specific purpose. One cannot consider verification and validation
without considering the purpose of the simulation. This chapter deals with a com-
prehensive outline of methodological perspectives and practical uses of verification
and validation. The problem of verification and validation is tackled in three main
topics: (1) the meaning of the terms verification and validation in the context of
simulating social complexity; (2) methods and techniques related to verification,
including static and dynamic methods, good programming practices, defensive
programming, and replication for model alignment; and (3) types and techniques
of validation as well as their relationship to different modelling strategies.
N. David (*)
ISCTE – Lisbon University Institute, Av. das Forças Armadas, Lisbon 1649-026, Portugal

136 N. David
8.1 Introduction
The terms verification and validation (V&V) are commonly used in science but
their meaning is often controversial, both in the natural and the social sciences.
The purpose of this chapter is not to describe any general theory of model V&V.
A general theory of that kind does not exist.
Besides the epistemological underpinnings of the terms, their use in simulation has
a pragmatic nature. In disciplines that make use of computerised models, the role of
V&V is related to the need of evaluating models along the simulation development
process. Basically, the very idea of V&V is comparing models with observations and
descriptions of the problem modelled, and this may include other models that have
been verified and validated to some level. This chapter describes methodological
perspectives and practical uses of the terms as well as different strategies and
techniques to verify and validate models of social complexity, mostly in social
simulation.
The use of simulation for modelling social complexity is very diverse. Often,
V&V do not refer to an explicit stage in the simulation development process, but to
the modelling process itself according to good practices and in a way that grants
credibility to using the simulation for a specific purpose. Normally, the purpose is
dependent on different strategies or dimensions, along which simulations can be
characterized, related to different kinds of claims intended by the modeller, such as
theoretical claims, empirical claims or subjunctive theoretical claims. The term
subjunctive is used when simulations are used for talking about scenarios in possible
worlds, such as describing ‘what would happen if something were the case’. There
cannot be V&V without considering the purpose of the simulation.
In the next section of the chapter, I will deal with the meaning of the terms V&V
in the context of the simulation development process. While the terms are often
used with the same meanings, or even interchangeably, practical reasons exist for
distinguishing between them. Whereas verification concerns the evaluation of the
implementation of the model in terms of the researchers’ intentions, validation
refers to the evaluation of the credibility of the model as a representation of the
subject modelled. In Sect. 8.3, methods and techniques related to verification are
described. Subsequently, in Sect. 8.4, validation types and techniques are described,
as well as their relationship to different modelling strategies.
8.2 The Simulation Development Process
Several chains of intermediate models are developed before obtaining a satisfactory

verified and validated model. What does it mean to verify and validate a model in
social simulation? Is there a fundamental difference between verifying and
validating models? The purpose of this section is to define the role of V&V within
the scope of the simulation development process.
8 Validating Simulations 137
Target
theory or
phenomenon
Validation Validation
Conceptualization and
Publication / Theory
model construction
dissemination / Application
Post-computerized Pre-computerized
models models
Implementation
Conceptualization and (physical construction of
model construction computerised models)
Executing
Verification Verification
computerized
model
Fig. 8.1 Verification and validation related to the model development process (David 2009)
The most common definitions of V&V are imported from computer science, as
well as from technical and numerical simulation,1 having intended distinct – although
often overlapping – meanings. The reason for distinguishing the terms is the need to
determine the suitability of certain models for representing two distinct subjects of
inquiry. This is represented in Fig. 8.1, in which V&V are related to a simplified
model development process. Two conceptual models mediate between two subjects
of inquiry. The latter are (1) the target theory or phenomenon and (2) the executing
computerised model. The conceptual model on the right, which is designated here as
the pre-computerised model, is basically a representation in the minds and writing of
the researchers, which presumably represents the target. This model must be
implemented as a computerised executable model, by going through a number of
intermediate models such as formal specification or textual programs written in high-
level programming languages.
The analysis of the executing model gives rise to one or more conceptual models
on the left, which are designated here as the post-computerised models. They are
constructed based on the output of the computerised model, often with the aid of
statistical packages, graphing and visualisation. The whole construction process
results in categories of description that may not have been used for describing the
pre-computerised model. This is the so-called idea of emergence, when interactions
among objects specified through pre-computerised models at some level of descrip-
tion give rise to different categories of objects at macro levels of description observed
in the executing model, which are later specified through post-computerised models.
1
Numerical simulation refers to simulation for finding solutions to mathematical models, normally
for cases in which mathematics does not provide analytical solutions. Technical simulation stands
for simulation with numerical models in computational sciences and engineering.
138 N. David
As an example consider the culture dissemination model of Axelrod (1997a)

which has a goal of analysing the phenomena of social influence. At a micro-level of
description, a pre-computerised model defines the concept of actors distributed on a
grid, the concept of culture of each actor as a set of five features and the interaction
mechanisms specified with a bit-flipping schema, in which the probability of inter-
action between two actors is set proportionately to measure the similarity between
two cultures.
The executing model is then explored and other categories of objects resulting
from the interaction of individual cultures may be defined, associated with macro
properties of interest and conditions in which they form, such as the concepts of
regions and zones on the grid. A great deal of the simulation proposed by Axelrod
concerns investigating properties of regions and zones in the context of the concep-
tual, post-computerised, model proposed, such as the relation between the size of a
region formed and the number of features per individual culture. These concepts
are interpreted in relation to the target social phenomena of social influence.
I will now situate the role of V&V in the modelling process of social simulation.
8.2.1 What Does It Mean to Verify a Computerised Model?
Computerised model verification is defined as checking the adequacy among con-

ceptual models and computerised models (see also Chap. 6 in this volume, Galán
et al. 2013). Consider the lower quadrants of Fig. 8.1. They are concerned with
ensuring that the pre-computerised model has been implemented adequately as an
executable computerised model, according to the researchers’ intentions in the
parameter range considered, and also that the post-computerised model adequately
represents the executing model in the parameter range considered.2
At this point you might question the meaning of adequately. There is no formal
definition of adequately. Actually, it is a rather difficult epistemological problem,
considered in the series of EPOS meetings about epistemological perspectives on
simulation (see Frank and Troitzsch 2005; Squazzoni 2009; David et al. 2010).
A practical and minimal definition could be the following: adequateness means that
the relationship between inputs and outputs of the computerised model is consistent
with the semantics of both the pre- and post-computerised models, in accordance with
the researcher’s intentions. However, the outcomes of computer programs in social
simulation are often unintended or not known a priori and thus the verification process
requires more than checking that the executing model does what it was planned to do.
The goal of the whole exercise is to assess logical links within, as well as between, the
pre- and the post-computerised models. This requires assessing whether the post-
computerised model – while expressing concepts that the pre-computerised model
does not express – is consistent with the latter. From a methodological point of view
2
Verification in the left quadrant of Fig. 8.1 is sometimes known as “internal validation”.
this is a complicated question, but from a practical perspective one might operationally
define the verification problem with the following procedures:
(a) For some pre-computerised model definable as a set of input/output pairs in a
specified parameter range, the corresponding executing model is verified for
the range considered if the corresponding post-computerised model expresses
the same set of inputs/outputs for the range considered.
(b) For some pre-computerised model defined according to the researcher and/or
stakeholders’ intentions in a specified parameter range, the corresponding
executing model is verified for the range considered if the corresponding
post-computerised model complies with the researchers and/or stakeholders’
intentions for the range considered.
Note that both procedures limit the verification problem to a clearly defined
parameter range. The first option is possible when extensive quantitative data is
available from the target with which to test the executing model. This is normally
not the case and the verification problem often amounts to the last option. This is
possible since the aim of verification is to assess the appropriateness of the logical
links that may be established between micro-levels of description specified in the
pre-computerised model and macro-levels of description specified through post-
computerised models, which should be amenable to evaluation by researchers and
stakeholders. Nevertheless, as we will discuss further, a computerised model should
only be qualified as verified with reasonable confidence if it has been successfully
subjected to a procedure known in software engineering as N-version program-
ming. A synonym commonly used in social simulation consists of replicating
implementations for model alignment. This is described in Sect. 8.2.4.
It is important to observe that, unlike many areas of computer science and formal
logics, the verification process should not be understood as a mere process of
evaluating the correctness of formal inference steps. The coherence among concep-
tual and computerised models, in the light of the researchers’ and stakeholders’
intentions, is, to a large extent, narratively evaluated. Were we to envisage computers
and programming languages as mere formal machines then computers and program-
ming languages would have limited expressiveness for representing inferences upon
semantically rich descriptions of social processes. After defining the meaning of
validation in the following section, this will be a point where the meanings of
verifying and validating a simulation model will overlap, and it may not be trivial
to distinguish between the two.3
3
From a technical point of view, in classical computer theory, verification amounts to ascertaining
the validity of certain output as a function of given input, regardless of any interpretation given in
terms of any theory or any phenomenon not strictly computational – it is pure inference in a closed
world. But this would require us to assume that social processes are computational in a Church-
Turing sense, which seems difficult to conceive. For an elaboration on this see (David et al. 2007).
140 N. David
8.2.2 What Does It Mean to Validate a Model?
Model validation is defined as ensuring that both conceptual and computerised

models are adequate representations of the target. The term “adequate” in this sense
may stand for a number of epistemological perspectives. From a practical point of
view we could assess whether the outputs of the simulation are close enough to
empirical data.
We could also assess various aspects of the simulation, such as if the mechanisms
specified in the simulation are well accepted by stakeholders involved in a partici-
pative-based approach. In Sect. 8.4 we will describe the general idea of validation as
the process that assesses whether the pre-computerised models – put forward as
models of social complexity– can be demonstrated to represent aspects of social
behaviour and interaction able to give rise to post-computerised models that are, at
some given level, consistent with the subjacent theories or similar to real data.
Given the model development process described, is there any practical differ-
ence between verifying and validating simulations? Rather than being a sharp
difference in kind it is a distinction that results from the computational method.
Whereas verification concerns the assessment of the logical inferences that are
established between micro and macro concepts with close reference to the
computerised model, validation concerns the evaluation of such inferences and
concepts with a closer reference to the target.
In paraphrasing Axelrod (1997b), at first sight, we could say that the problem is
whether an unexpected result is a reflection of the computerised model, due to a
mistake in the implementation of the pre-computerised model, or is a surprising
consequence of the pre-computerised model itself. Unfortunately, the problem is
more complicated than that. In many cases mistakes in the code may not be
qualified simply as mistakes, but only as one interpretation among many other
possible interpretations for implementing a conceptual model. Nevertheless, from a
practical viewpoint there are still good reasons to make the distinction between
V&V. A number of established practices exist for the corresponding quadrants of
Fig. 8.1. We will address these in the following sections.
8.3 Verification Methods and Techniques
The need to transform conceptual models into computerised models and back to
conceptual models is one of the substratums of simulation. This process is illustrated
in Fig. 8.2. Several conceptual models are constructed and specified in order to obtain
a proper implementation as a computerised model. The intended computerised model
is specified as a computer program in a readable form with a high-level language. The
compilation of the program into a low-level program, followed by its execution with
a set of inputs, results in outputs described with data and visualization models. Given
the use of pseudo-random generators, the simulation will be run many times with the
Pre-Computerized models Computerized models Post-Computerized models

Inputs:
- parameters execution Conceptual
Conceptual Conceptual
Conceptual … - random seeds Outputs …
models models models
models compilation
High level interpretation
programs
VERIFICATION
Fig. 8.2 Implementation of pre-computerised models and construction of post-computerised

models
same inputs. At any rate, the complexity of social theories and phenomena implies
that the simulation must be run many times in order to observe its behaviour in a
variety of conditions. A post-computerised model is built, i.e. a conceptual model
describing the simulation behavior, possibly incorporating new concepts and knowl-
edge not captured with the pre-computerised model.
8.3.1 Static and Dynamic Methods
In simulation, verification techniques involve showing that the process of transla-

tion among conceptual and computerised models is carried out according to the
researchers’ intentions. Computer science, mostly in the field of software engineer-
ing, has developed numerous approaches and techniques for writing programs and
verifying their execution in computers. Most of these are not particular to simula-
tion and not all are well suited to social simulation. The techniques and terminology
that we describe in this section are based on Sommerville (1995) and Sargent
(1999), and are adapted to social simulation whenever appropriate. There are two
basic approaches for verification: static and dynamic methods.
Dynamic methods involve exercising the computerised model with sets of inputs
and the results obtained are used to determine whether the computerised model has
been programmed appropriately. A possible simulation failure occurs when the
simulation is executing and does not behave as expected. Whether the unexpected
output is indeed a failure or a legitimate consequence of the pre-computerised
model concerns the use of static methods. Static methods involve showing that
the computerised model appropriately implements the pre-computerised models
without software faults, i.e. without programming errors whereby the computerised
model does not conform to the researchers’ intentions.
Software faults are static and their existence may be inferred from simulation
failures and further inspections to the high-level program code. A software fault
causes a simulation failure when the faulty code is executed with a set of inputs that
expose the simulation fault (cf. Sommerville 1995).
We can never conclusively demonstrate that a simulation is fault free, but we can
increase our confidence in a program within the range of the parameter space tested
by adopting good programming and testing practices in order to minimize faults.
142 N. David
Good practices common to any kind of programming include commenting your

code and testing the program with parameter values whose outputs are known or
with values that are at the extremes of the legal values.
Nevertheless, not all programming and testing techniques in computer science
can be applied to social simulation. For the most part, simulation is used to explore
concepts that are not anticipated during the specification of pre-computerised
models. Virtually, the software notion of “functional requirement” does not exist
in social simulation. If we are dealing with complex social models, it is virtually
impossible to enumerate a priori an exhaustive list of requirements that a program
should satisfy (David et al. 2003). To the extent that the researcher does not know
what to expect, program testing is not enough to show that the program properly
reflects the intended model.
In any case, the verifiability of a simulation is influenced by the process used to
develop that simulation. The adoption of good programming practices for designing
and implementing the model is fundamental. Defensive programming methodologies,
like contract-based programming with use of assertions, are well suited for the
explorative nature of simulation. Defensive programming consists of including
extra, redundant, code to test for errors and to aid in general debugging. In addition,
the need to produce several implementations of the same conceptual model is
increasingly playing a role. These techniques are described below along four major
topics: good programming practices, defensive programming, replication, and a very
brief reference to participative-based methods.
8.3.2 Good Programming Practices
Good programming in social simulation includes several approaches, techniques or

mechanisms to improve verifiability in general and, specifically, to decrease the
number of expected simulation faults, to ease debugging and to improve code
readability as well as faster and more flexible development. Available techniques
and mechanisms include modularity, encapsulation, high-level memory manage-
ment and, more generally, software reuse. Most object-oriented programming
languages include built-in mechanisms that simplify their use. The following
deserve particular attention.
8.3.2.1 Modularity and Encapsulation
Programs should be written and verified in modules or subprograms.4 In the words

of Kleijen (1995), this is a kind of “divide and conquer” process, where one verifies
the whole simulation module by module. Modular programming simplifies the
4
We consider “modules” and “sub-programs” as synonymous.
location of faults in the program. A module is any component that provides one or
more services to other modules and which separates its public interface from its
implementation details. The public interface of a module concerns the specification
for how other modules may request services from it. The implementation concerns
particular design decisions hidden within the module and not accessible to the other
modules, which are more liable to faults or design changes along the simulation
development process. If the implementation of a module is changed this should not
affect how other modules request services from it. A typical kind of module is a
class in object-oriented programming.
A class is a model for object instantiation. Objects are independent, loosely coupled
entities in which the particular implementation of the object state (information
representation) and of the services provided by the object (information processing)
should not affect the public interface through which other objects request services
from it. An object comprises a set of private attributes – which define its state – and a
set of public operations that check or act on those attributes – which define its
interface. Insofar as the public interface remains the same along the development
process, the internal data representation of the class may be changed along the way
without affecting the way other classes request services from it. The grouping of
related classes within a single file is another kind of module, usually called a physical
module. Yet another kind is the concept of package, defined as a group of related files,
such as in the Java programming language. The public interface of a package is the set
of all public services available in all files. The package design details most likely to
change are hidden behind the interface, possibly implemented with the aid of other
private classes in the package.
It is essentially up to the programmer to decide which data structures, classes and
files should be grouped together. According to Sommerville (1995, pp. 218–219),
the cohesion of a module is a measure of the closeness of the relationship between
its components. Conversely, the coupling among modules measures the strength of
interconnections among them. As a rule, modules are tightly coupled if they make
use of shared variables or if they interchange many types of control information.
A good design calls for high cohesion within a module and low coupling among
modules. For instance, an agent implemented as a group of classes should have
its public interface clearly defined and be as independent as possible from the group
of classes that implement the interaction environment of the agents, such as a
bi-dimensional torus grid. Whether the grid is implemented as a torus or as a
network should not imply future changes to the implementation of the agent
architecture, which in any case would request the same services from the grid,
such as moving to the left or right. Modularity encourages the production of
readable and testable simulations. Moreover, insofar as agent-based modelling
can be easily mapped to the programming of interacting modules, the adoption of
bottom-up design and testing approaches is a natural way to proceed. Small
modules at the lower levels in the hierarchy are tested first and the other modules
are worked up the hierarchy until the whole simulation is tested.
144 N. David
8.3.2.2 High-Level Memory Management
The use of pointers and pointer arithmetic in a program are low-level constructs
available in some high-level languages that manipulate variables referring directly to
the machine memory. They are inappropriate in social simulation programs because
they allow memory leaks and foment aliasing, which makes programs prone to bugs.
A memory leak occurs when a program fails to release memory no longer needed,
which may have the effect of consuming more memory than necessary and decreasing
efficiency. We say there is aliasing when an object is changed through a reference
variable with unwanted indirect effects on another variable. Both leaks and aliasing
make programs harder to understand and faults harder to find. Whenever possible,
high-level languages with pointer arithmetic should be avoided. In programming with
object-oriented languages, the use of built-in mechanisms for automatic memory
management, responsible for managing the low-level lifecycle of objects in memory,
should be preferred. Whereas it is up to the programmer to determine the points where
objects are created in the program, mechanisms of memory management free the
programmer from the trouble of deleting the objects from memory at the points where
they become useless. Fortunately, most programming languages used in current
agent-based simulation toolkits are examples of high-level languages that have
built-in mechanisms for automatic memory management.
8.3.2.3 Software Reuse
We sometimes tend to prefer rewriting modules insofar as we believe that our

modules are better programmed than others. However, software modules that are
used and tested in a variety of different situations have fewer faults than modules
developed for the purposes of a single simulation. A good programmer should not
assume that all modules should be implemented especially for the simulation being
developed. Reusable modules that have been subjected to previous use and verifica-
tion should be preferred to those built from scratch. The more a simulation is based on
reusable modules the fewer modules need to be specified, implemented and verified.
Software reuse is used in social simulation as more and more developing
frameworks become available. Module reuse in simulation is basically a type of
model embedding. The use of agent-based toolkits as special-purpose extensions to
standard high-level languages based on Java, Objective C, or SmallTalk provides
developing standards and faster development, making simulations more compara-
ble to each other. Software reuse in computer science and engineering can be
considered at a number of different levels (Sommerville 1995, p. 397). Likewise
in social simulation we may identify different levels from the strongest to the
weakest degrees of dependency on the specifics of simulation platforms and
toolkits. In this sense, strong reuse means constraints on the simulation to existing
models and weak reuse lets the researcher develop models more freely.
Model Architecture Level
The control subsystem of a simulation, such as a discrete-time scheduling mecha-

nism, may be reused. In most agent-based simulation toolkits, scheduling consists
of setting up method calls on objects to occur at certain times. Scheduling
mechanisms are thus part of a reusable agent skeleton template, extended according
to the specifics of each agent model. Other examples include whole collections of
reusable objects and mechanisms, such as interaction environments that act as agent
containers and define the interactive relationship of agents relative to each other,
such as a network model or a torus grid. Yet other examples may include entire
canonical simulation models, which are used as components of another extended,
more sophisticated, simulation model.
Module or Object Level
Components representing a set of functions may be reused, such as class libraries

for generating sequence graphs; histograms or plots for visualization models;
statistical packages for data analysis; and widely tested pseudo-random number
generators that aggregate a set of different random number distributions into a
single class or package.
Programming Language API Level
Standard libraries and Application Programming Interfaces (APIs) conventionally

available in every implementation of high-level languages may and should be used.
Typically, abstract data structures representing such things as lists, queues and
trees, are coupled in a set of classes that implement reusable data structures, such as
the Java Collections Framework.
Generic-Type Level
The use of generic classes is a kind of reuse. Generics are a way of creating
parameterized general purpose templates for class types and subroutines (see e.g.
Sommerville 1995). If the same kind of abstract structure, such as a list, is used
for many different types of elements, like a list of agents and a list of events, this
avoids the need to develop a separate implementation of the list for each type of
element. As a result, programmers are less likely to make mistakes, thus
avoiding faults and improving confidence in the simulation. Generic program-
ming facilities are available in several object-oriented languages, such as Cþþ
and Java.
146 N. David
Routine Level
Independent components implementing a single procedure or function may be

reused, for example, a mathematical function used in a previously tested and
verified setting can be reused in another simulation.
8.3.3 Defensive Programming
The focus of good programming practices is on good code readability, flexible and
fast development as well as ease of debugging. Other focuses yet to be considered
are the actual testing and location of faults in the program. A typical test consists of
running test cases for which outputs are known and confirming these every time the
program is changed. Defensive approaches require that the programmer should
never assume that a module will work as expected but instead the programmer
should handle the appropriate testing for every module, possibly embedding the test
in the program itself. Rather than considering testing only after programming the
model, some of the testing should be planned and embedded as redundant code in
the program. In software engineering defensive programming stands for an
approach to fault tolerance for failure recovery. In social simulation the goal is
somewhat different and consists of including extra, redundant, code to test for
program failures and assist in the location of program faults.
Two techniques seem relevant to social simulation: contract-based programming
and unit testing. Although the latter is often referred to in the literature, it is not a
common practice in social simulation, probably because it requires a considerable
amount of redundant code to test programs.
The former stands for a methodology for checking the violation of specified
conditions in order to verify programs written with procedural languages on the
basis of declarative statements, called assertions, which are appropriate to the kind
of explorative programming used in simulation.
Testing and contract-based programming can be combined with the use of
assertions and exceptions. Assertions are predicates placed in the program to test
whether a specified condition is satisfied during testing. If the predicate evaluates to
false, the program should halt for further debugging. Exceptions are messages that
indicate the occurrence of exceptional conditions that were anticipated by the
programmer, which usually halt the program as well.
Contract-based programming is a methodology for designing class specifications
with checkable conditions at run time. The attributes of a class define the state space
for the objects of that class. The set of states considered valid act as constraints on
the range of values that objects of that class can assume. Constraints are limits
defined over the ranges of the attributes of a class and may define dependencies
among those attributes. If the valid state space of a class is violated in execution
time then an exceptional condition is launched and the program execution halts,
indicating information about the location and the type of violation occurred.
Constraints can also be specified at the routine level by specifying the range of
the parameter values and the dependencies among parameters. In fact, the specifi-
cation of a class may be understood as defining a contract between two
programmers, the client and the supplier of that class, resulting in two types of
constraints:
1. The methods’ pre-conditions specified by the supplier, to which the client should
abide;
2. The class invariant and the methods’ post-conditions to which the supplier
commits.
If the pre-conditions are not satisfied, then an exceptional condition occurs,
which must be dealt with by the client programmer who violated the contract of
that class. In contrast, if the class invariant or the post-conditions are not satisfied,
then the fault is due to the particular implementation of that class, resulting in an
assertion failure. This means that the supplier may not have programmed the class
appropriately.
As an example, consider a partial definition of a class named Agent that specifies
an agent in a culture dissemination model. Agents are distributed on a grid and the
culture of each actor is defined as a nonempty set of features with a positive number
of traits per feature. Each agent is aware of its position on the grid, which cannot be
outside the grid bounds. The attributes of that class include a list of features, the
number of traits per feature and the agent’s position. The invariant of such a class
would be defined as follows:
features 6¼ fg and number traits > 0 and position is legal
Suppose that agent interaction is specified with a bit-flipping kind of schema,

which occurs only between contiguous neighbors and if the agents have different
traits in at least one feature. Moreover, the agents must share the same trait in at
least one feature after the interaction. In order to guarantee that every state change
in objects of type Agent is valid, the invariant must be checked (1) right after the
calling of that method; (2) right before it returns to the caller; and (3) right before
the end of the class constructors. That is, whereas the bit-flipping pre-conditions
should be checked before the interaction takes place, the post-conditions should be
checked after the interaction. A partial definition of a Java class could be the
following (with redundant code in boldface):
public class Agent {
private SpaceGrid grid; // the agent´s interaction
environment
private int [] features;
private int number_of_traits;
private Position position; // the agent’s current
position in the grid
148 N. David
// class constructor
public Agent(final SpaceGrid grid, final int [] features,
final int number_of_traits, final Position
position) {
this.grid ¼ grid;
this.features ¼ features;
this.number_of_traits ¼ number_of_traits;
this.position ¼ position;
// check the class invariant after initialisation
assert checkInvariant();
}
// other methods here
// the bit-flipping operation
public void bitFlip(final Agent theOtherAgent) {
// check the invariant before every method that
change the agent’s state
assert checkInvariant();
// test pre-conditions for bit-flipping
if (theOtherAgent ¼¼ null)
throw new NullPointerException();
if (!grid.areNeigbourhs(this, theOtherAgent)) //
are the agents neighbours?
throw
new IllegalArgumentException(“Agents are not
neighbours, cannot bit-flip”);
// agents must have a different trait in at least one
feature
if (Arrays.dontMatch(features, theOtherAgent.
getFeatures()) < 1)
throw
new IllegalArgumentException(“Insufficient
number of features to bit-flip”);
// make bit-flipping
int[] other_features ¼ theOtherAgent.getFeatures();
int rand ¼ randomFeature(features);
while (features[rand] ¼¼ other_features[rand])
rand ¼ randomFeature(features);
other_features[rand] ¼ features[rand]; // bit-flip
// test post-conditions (agents must share at least

one feature after bit-flip)
assert Arrays.match(features, theOtherAgent.
getFeatures()) > ¼ 1;
// check the invariant after every method that
changed the agent’s state
assert checkInvariant() ;
}
private boolean checkInvariant() {
if (grid ¼¼ null || features ¼¼ null || position ¼¼
null ||
number_of_traits < 1 || grid.outOfBounds
(position))
return false;
return true;
}
} // end of class Agent
In the program, the invariant is checked at the end of the constructor, right after
initializing the agent’s attributes, and both before and after any state change in the
methods. Since the invariant concerns the testing on the part of the programmer of
that class (the supplier), rather than the user of that class (the client), the method
checkInvariant is set to private. Since the method bitFlip changes the agent’s state, the
method pre-conditions and post-conditions should be checked. If the pre-conditions
are not satisfied, an exception is thrown. This means that the error is on the caller of
that method. If the class invariant or the post-conditions are not satisfied, then an
assertion failure is launched, suggesting an error in the implementation of that class.
Contract-based programming helps programmers reason about the conceptual
model and its programs. Too often it is difficult to understand the details of an
implementation. Insofar as the contracts of a module are fully documented and are
regarded as a precise specification of the computerised model – akin to a declarative
programming language style – additional benefits are increased readability and the
fostering of replicability. During testing the programmer will typically run the
program with assertions enabled. For reasons of efficiency, the assertion mecha-
nism can either be set to on or off.
8.3.4 Replication for Model Alignment

Program testing is an essential activity for building simulations. However, testing
alone cannot distinguish between faults in the implementation of the pre-computerised
model and a surprising consequence of the model itself. Static methods are thus
unavoidable and the examination of the source code is important in all stages of the
150 N. David
Conceptual model A Computerized

model A compare
results
Conceptual model B ’ Computerized
model B ’
Computerized
Conceptual model B
model B
Fig. 8.3 Model alignment
Computerized
model A’
Published results compare
conceptual model A
results
and results
Computerized
model A’’
Fig. 8.4 Model replication
development process. In practice, static and dynamic methods are used together.
Most integrated development environments provide debugging tools and allow the
programmer to trace the execution of a program and observe the values of variables
statement by statement.
A slightly different technique, mostly used in numerical and technical simulation,
is called “structured walk-throughs.” This consists of having more than one person
reading and debugging a program. All members of the development team are given a
copy of a particular module to be debugged and the module developer goes through
the code but does not proceed from one statement to another until everyone is
convinced that a statement is correct (Law and Kelton 1991).
A different technique described by Sargent (1999) consists of reprogramming
critical modules to determine if the same results are obtained. If two members are
given the same specification for a module and they make two distinct
implementations and obtain the same results, then the confidence in the implemen-
tation is increased. A more general problem is the extent to which models can be
related to others so that their consequences and results are consistent with each
other. In its most general form, this concerns both to V&V. After Axtell et al.
(1996) it became known as the process of model alignment, which is used for
determining whether different published models describing the same class of social
phenomena produce the same results. Usually the alignment of two models A and B
requires modifying certain features of model B – for instance by turning off certain
features – in order to become equivalent to model A. This is represented in Fig. 8.3.
The term “model alignment” is frequently used synonymously for model repli-
cation. This assesses the extent to which building computerised models that draw
on the same conceptual, usually published, model give results compatible with the
ones reported in the published model. If the new results are similar to the published
results, then the confidence in the correspondence between the computerised and
the conceptual models is increased. Replication is represented in Fig. 8.4.
Computerized Alignment 1 (A)

Models Publication 1
Specification
Publication
Version 1 Alignment 2 (B)
Version 2 Alignment 3 Publication 3

...
Version 3 Alignment n Publication n
Replicating for model verification Replicating for model alignment
Fig. 8.5 N-version programming and replication for model alignment
In the vast majority of cases, replication processes are mostly reactive and not
proactive, in other words, the creator of the model publishes it and in the future
someone else replicates it. However, we should not take replication lightly and see it
as something that will possibly be done in the future. A verification technique, not so
frequently used in social simulation, is called N-version programming. This is similar
to reprogramming critical modules. In contrast to the practice of having other teams
replicating models after they have already been reviewed, accepted and published,
the effort of replication becomes focused on producing multiple computerised
versions of a conceptual model before submitting it for peer reviewing.
Both N-version programming and replication are depicted in Fig. 8.5. The right-
hand side of the diagram illustrates a perspective where a published conceptual model
is replicated. The left-hand side illustrates a perspective where a conceptual model
gives origin to multiple computerised versions of the model, implemented by differ-
ent persons before it is actually published. In any case, if all versions lead to the same
results, then there are reasonable grounds for trusting in the results obtained, as well
as in the correspondence between the conceptual and computerised models.
8.3.4.1 Types of Model Equivalence
The work of Axtell et al. (1996) is arguably the most-cited attempt to align two
distinct but similar models. Rather than re-implementing Axelrod’s culture dissem-
ination model, Axtell and colleagues focused on the general case of aligning two
models that reflected slightly distinctive mechanisms. For this purpose, Epstein and
Axtell’s Sugarscape model was progressively simplified in order to align with the
results obtained by Axelrod’s culture dissemination model.
Model alignment has been further investigated in a series of meetings called
model-to-model (M2M) workshops (see Rouchier et al. 2008). The M2M workshops
attract researchers interested in understanding and promoting the transferability of
knowledge between model users. The replication of Edmonds and Hales (2003) is
particularly informative on the problem of verification. They suggested that one
should not trust an unreplicated simulation, since its results are almost certainly
wrong in the sense that the computerised model differs from what was intended or
152 N. David
assumed in the conceptual model. In order to align with a published conceptual model
they had to rely on a double replication process insofar as the first replication did not
seem to result in alignment. It was only after implementing a second computerised
version of the published model that the authors were sure that the results reported in
the published model did not align with their own. They concluded that the original
implementation of the published model was producing results that could be
misleading. This was due to different interpretations of the correct implementation
of a mechanism of tournament selection for reproducing agents. Subtle differences in
the mechanism implied different conclusions about the functioning of the model.
But how do we determine whether or not two models produce equivalent results?
Axtell et al. (1996) defined three kinds of equivalence:
• Numerical identity: shows that the two models reproduce the results exactly.
• Relational equivalence: shows that the two models produce the same internal
relationship among their results, for instance that a particular variable is a
quadratic function of another.
• Distributional equivalence: shows that the two models produce distributions of
results that cannot be distinguished statistically.
Numerical identity is hardly attainable in social complexity, except in the
circumstances where models converge to some kind of final equilibrium, such as in
Axelrod’s culture dissemination model. Most simulations display several kinds of
global behaviours that do not converge to equilibrium, are in the medium term self-
reinforcing and make numerical identity unattainable. Among the three kinds of
equivalence, distributional equivalence is the most demanding: it is achieved when
the distributions of results cannot be distinguished statistically. What this shows
is that at conventional confidence probabilities the statistics from different
implementations may come from the same distribution, but it does not prove that
this is actually the case. In other words, it does not prove that two implementations are
algorithmically equivalent, but it allows us to disconfirm that they are. If this test
survives repeatedly we somehow increase our confidence in the equivalence but only
in the parameter range within which it has been tested.
8.3.4.2 Programming for Replication
An important aspect when programming a simulation is to guarantee that it may be

replicated by other researchers. Designing and programming for replicability
involves a number of aspects that should be considered. Simulations are often a
mix of conceptual descriptions and hard technical choices about implementation.
The author who reports a model should assume that an alignment may later be tried
and thus should be careful about providing detailed information for future use:
• Provide effective documentation about the conceptual and the computerised
models; provide information about those technical options where the translation
from the conceptual to the computerised model is neither straightforward nor
consensual. Even if the difference between two computerised models may seem
minor or irrelevant, small differences can affect the level of equivalence, even if
the overall character of the simulation does not seem to change significantly.
• Outline in pseudo-code the parts of the program that may be liable to
ambiguities; use modelling languages to represent implementation options,
like UML. This is a formalism used to describe not only the static structure of
the relations between classes but also different aspects of its dynamic behaviour,
for instance, with activity graphs and sequence diagrams.
• Make the source code available online and ready to install. If possible, use a
simulation platform to implement the model, hence fostering software reuse in
order to make simulations reliable and more comparable to each other. If possible,
make the simulation available to be run online, for instance, by using such
technologies as Applets, a technology that allows the embedding of an execution
model in Web pages, making it possible to be loaded and executed remotely.
Making the computerised model available is crucial for others to be able to run the
model with parameters settings that were not reported in your papers. Whereas for
a certain set of parameter settings two simulations may match, this may not
happen with other settings, suggesting the two models are not equivalent.
• Provide a detailed description about the results, statistical methods used, distri-
butional information and qualitative measures. Make the bulk outputs available
online or in appendices.
While programming for replicability is something to be considered on the part of
the team that programs the model, a number of aspects should be considered by the
team that decides to replicate a model. Often apparently irrelevant differences in two
implementations can be the cause of different statistics that do not match (Edmonds
and Hales 2003). This is particularly relevant if the description of the original model
is not detailed sufficiently, making it difficult to assess whether the computerised
model is implemented correctly and according to the conceptual model. When the
original implementation is available, high degrees of confidence in the new imple-
mentation requires matching results with runs based on different parameters from
those reported in the original model. The experiment of Edmonds and Hales provides
a good deal of informative techniques and tips on this kind of model alignment:
• If the simulation shows very dynamic self-reinforcing effects in the long run,
check the alignment of simulations in the first few cycles of their runs, averaged
over several runs, in order to test whether they are initialized in the same way.
• Use different parameter settings and for each setting make several runs of both
implementations over long-term periods. For each implementation collect several
statistics from each time period and average them over the number of runs. For
each such pair of sets of averages use statistical tests, such as the Kolmogorov-
Smirnov test for the goodness of fit of cumulative distribution functions.
• Use modularity to test features separately. If two simulations do not align, turn
off certain features of the implementations until they do so, and find the source
of errors in the different modules. Reprogramming whole critical modules may
also apply here.
• Use different toolkits or programming languages to re-implement simulations
and if possible have this done by different people.
154 N. David
8.3.5 Participative-Based Methods
Replication is feasible when models are simple. When the goal is modelling a specific
target domain, full of context, with significant amounts of rich detail, and use of
empirical data and stakeholder participation, such as with the Companion Modelling
approach, replication may not be feasible for verifying the computerised model. As in
any other simulation, good programming practices and defensive programming are
thus fundamental. In addition, insofar as some results may be due to software faults
and be unexpected for stakeholders, participative-based methods are also a form of
verification. This fact stresses the importance of involving domain experts and
stakeholders as much as possible in all stages of the development and implementation
process. Here, documentation and visualisation techniques can play a crucial role in
bridging between the stakeholders’ opinions and the intention of the programmer of
the simulation. This is discussed in more detail in Chap. 10 in this volume (Barreteau
et al. 2013).
8.4 Validation Approaches
We offered a conceptual definition of validation in Sect. 8.2.2. Had we given an

operational definition, things would have become somewhat problematical. Models
of social complexity are diverse and there is no definitive and guaranteed criterion
of validity. As Amblard et al. (2007) remarked, “validation suggests a reflection on
the intended use of the model in order to be valid, and the interpretation of the
results should be done in relation to that specific context.”
A specific use may be associated with different methodological perspectives for
building the model, with different strategies, types of validity tests, and techniques -
Fig. 8.6. Consider the kind of subjunctive, metaphorical, models such as
Schelling’s. In these models there is no salient validation step during the simulation
development process. Design and validation walk together, and the intended use is
not to show that the simulation is plausible against a specific context of social
reality but to propose abstract or schematic mechanisms as broad representations of
classes of social phenomena. In other different cases, the goal may be modelling a
specific target domain, full of context, with use of empirical data and significant
amounts of rich detail. Whereas in the former case a good practice could be
modelling with the greatest parsimony possible so as to have a computational
model sanctionable by human beings and comparable to other models, parsimony
can be in opposition to the goal of descriptive richness and thus inappropriate to the
latter case.
There are also different methodological motivations behind the use of a model,
such as those conceived to predict or explain and those merely conceived to
describe. Regardless of what method is used, the reproduction of characteristics
of the object domain is important, but this can be assessed through rather different
approaches during the model development process. If it is prediction you are
Purpose of Models
General goal of simulating social complexity
Basic methodological conceptions

(types of validity:
through prediction, retrodiction, structural similarity)
Validation Techniques
(diverse)
Relationship to modeling strategies

( subjunctive models, context-specific models)
Fig. 8.6 Validation implies considering the purpose of the model
seeking, validation consists of confronting simulated behaviour with the future

behaviour of the target system (however, attempting to establish numerical predic-
tion is not a normal goal in simulation). If it is explanation, validation consists of
building plausible mechanisms that are able to reproduce simulated behaviour
similar to real behaviour. If the goal is the more general aim of descriptiveness,
explanation may probably be a goal as well, and a creative integration of ways for
assessing the structure and results of the model, from quantitative to qualitative and
participatory approaches, will be applied.
In conclusion, one should bear in mind that there is no one special method for
validating a model. However, it is important to assess whether the simulation is
subjected to good practices during its conception, whether it fits the intended use of
the model builder and whether it is able to reproduce characteristics of the object
domain. Assessing whether the goals of the modellers are well stated and the
models themselves are well described in order to be understood and sanctioned
by other model builders are sine qua non conditions for good simulation modelling.
In the remainder of this section, we revise the issue of the purpose of validating
simulations along four dimensions: the general goal of simulation in social com-
plexity, three basic methodological conceptions of validity types, a set of usual
techniques applied in social simulation, and finally the relationship of validation to
different modelling strategies with respect to the level of descriptive detail embed-
ded in a simulation.
8.4.1 The Goal of Validation: Goodness of Description
If one is using a predictive model, then the purpose of the model is to predict either
past or future states of the target system. On the other hand, one may strive for a
156 N. David
model that is able to describe the target system with satisfactory accuracy in order
to become more knowledgeable about the functioning of the system, to exercise
future and past scenarios, and to explore alternative designs or inform policies.
The objective in this section is to define the purpose of validation in terms of the
purpose of simulating social complexity, which we will define as being of good
description. This position entails that there is no single method or technique for
validating a simulation. A diversity of methods for validating models is generally
applied.
In the rest of this chapter we adopt the multi-agent paradigm for modelling.
A conceptual understanding of validation, similar but more general than Moss and
Edmonds’ (2005), will be used:
The purpose of validation is to assess whether the design of micro-level mechanisms, put
forward as theories of social complexity validated to arbitrary levels, can be demonstrated
to represent aspects of social behaviour and interaction that are able to produce macro-level
effects either (i) broadly consistent with the subjacent theories; and/or (ii) qualitatively or
quantitatively similar to real data.
By broad consistency we mean the plausibility of both micro specification and

macro effects accounted as general representations of the target social reality. In its
most extreme expression, plausibility may be evaluated on a metaphorical basis. By
qualitative similarity to real data we mean a comparison with the model in terms of
categorical outcomes, accounted as qualitative features, such as the shape of the
outcomes, general stylized facts, or dynamical regimes. As for quantitative similar-
ity we mean the very unlikely case in which the identification of formal numerical
relationships between aggregate variables in the model and in the target – such as
enquiring as to whether both series may draw from the same statistical distribution –
proves to be possible.
Notice that this definition is general enough to consider both the micro-level
mechanisms and macro-level effects assessed on a participatory basis. It is also
general enough to consider two methodological practices – not necessarily incom-
patible – related to the extent to which models in social science simulation ought to
be constructed on the basis of formal theories or ought to be based on techniques
and approaches on the basis of the intuition of the model builders and stakeholders –
an issue that we will come back to later. These are omnipresent methodological
questions in the social simulation literature and are by no means irrelevant to the
purpose of simulation models.
Suppose that on the basis of a very abstract model, such as the Schelling model,
you were to evaluate the similarity of its outputs with empirical data. Then you will
probably not take issue with the fact that the goal of predicting future states of the
target would be out of the scope of simulation research for that kind of modelling.
However, despite the belief that other sorts of validation are needed, this does not
imply excluding the role of prediction, but emphasises the importance of description
as the goal of simulating social complexity. In truth, what could be more contentious
in assessing the Schelling model is the extreme simplicity used to describe the
domain of social segregation. The descriptive power of multi-agent models makes
them suited to model social complexity. Computational modelling corresponds to a
process of abstraction, in that it selects some aspects of a subject being modelled,

like entities, relations between entities and change of state, while ignoring those that
may be considered less relevant to the questions that are of interest to the model
builder. The expressiveness of multi-agent models allows the researcher to play with
intuitive representations of different aspects of the target, such as defining societies
with different kinds of agents, organizations, networks and environments, which
interact with each other and represent social heterogeneity. By selecting certain
aspects of social reality into a model, this process of demarcation makes multi-agent
modelling suited to represent sociality as perceived by researchers and often by the
stakeholders themselves.
The descriptive power of simulation is on par with the diversity of ways used for
informing the construction and validation of models, from theoretic approaches to the
use of empirical data or stakeholder involvement. At any rate, measuring the good-
ness of fit between the model and real data expressed with data series is neither the
unique nor a typical criterion for sanctioning a model. The very idea of using a
diversity of formal and informal methods is to assess the credibility of the
mechanisms of the model as good descriptions of social behaviour and interaction,
which must be shown to be resilient in the face of multiple tests and methods, in order
to provide robust knowledge claims and allow the model to be open to scrutiny.
8.4.2 Broad Types of Validity
When we speak about types of validity we mean three general methodological

perspectives for assessing whether a model is able to reproduce expected
characteristics of an object domain: validation through prediction, validation
through retrodiction and validation through structural similarity. Prediction refers
to validating a model by comparing the states of a model with future observations of
the target system; retrodiction compares the states of the model with past
observations of the target system; and structural similarity refers to assessing the
realism of the structure of the model in terms of empirical and/or theoretical
knowledge of the target system. In practice, all three approaches are interdependent
and no single approach is used alone.
8.4.2.1 Validation Through Prediction
Validation through prediction requires matching the model with aspects of the
target system before they were observed. The logic of predictive validity is the
following: If one is using a predictive model – in which the purpose of the model is
to predict future states of the target system – and the predictions prove satisfactory
in repeated tested events, it may be reasonable to expect the model outcomes to stay
reliable under similar conditions (Gross and Strand 2000). The purpose of predic-
tion is somewhat problematic in social simulation:
158 N. David
• Models of social complexity usually show nonlinear effects in which the global
behaviour of the model can become path-dependent and self-reinforcing, pro-
ducing high sensitivity to initial conditions, which limits the use of predictive
approaches.
• Many social systems show high volatility with unpredictable events, such as
turning points of macroeconomic trade cycles or of financial markets that are in
practice (and possibly in principle) impossible to predict; see (Moss and
Edmonds 2005) for a discussion on this.
• Many social systems are not amenable to direct observation, change too slowly,
and/or do not provide enough data to be able to compare model outcomes. Most
involve human beings and are too valuable to allow repeated intervention, which
hinders the acquisition of knowledge about its future behaviour. Policies based
on false predictions could have serious consequences, thus making the purpose
of prediction unusable (Gross and Strand 2000).
While quantitative prediction of the target system behaviour is rare or simply
unattainable, prediction in general is not able to validate per se the mechanisms of
the model as good representations of the target system. In the words of Troitzsch
(2004), “What simulations are useful to predict is only how a target system might
behave in the future qualitatively”. But a different model using different
mechanisms that could lead to the same qualitative prediction may always exist,
thus providing a different explanation for the same prediction. More often, the role
of predicting future states of the target system becomes the exploration of new
patterns of behaviour that were not identified before in the target system, whereby
simulation acquires a speculative character useful as a heuristic and learning tool.
What we are predicting is really new concepts that we had not realized as being
relevant just for looking into the target.
8.4.2.2 Validation Through Retrodiction
The difference from retrodiction to prediction is that in the former the intention is to
reproduce already observed aspects of the target system. Given the existence of a
historical record of facts from the target system, the rationale of retrodictive
validity for a predictive model is the following: If the model is able to reproduce
a historical record consistently and correctly, then the model may also be trusted for
the future (Gross and Strand 2000). However, as we have mentioned, predictive
models of social complexity are uncommon in simulation. Explanation rather than
prediction is the usual motive for retrodiction. The logic of retrodictive validity is
the following: If a model is able to consistently reproduce a record of past
behaviours of the target system, then the mechanisms that constitute the model
are eligible candidates for explaining the functioning of the target system. Never-
theless, retrodiction alone is not sufficient to assess the validity of the candidate
explanations:
• Underdetermination: Given a model able to explain a certain record of

behaviours or historical data, there will always be a different model yielding a
different explanation for the same record.
• Insufficient quality of data: In many cases it is impossible to obtain long
historical series of social facts in the target system. In the social sciences the
very notion of social facts or data is controversial, can be subjective, and is not
dissociable from effects introduced by the measurement process. Moreover,
even when data is available it may not be in a form suitable to be matched to
the bulk of data generated by simulation models.
Underdetermination and insufficient data suggest the crucial importance of
domain experts for validating the mechanisms specified in the model. A model is
only valid provided that both the generated outcomes and the mechanisms that
constitute the model are sanctioned by experts in the relevant domain. The impor-
tance of validating the mechanisms themselves leads us to the structural validity of
the model, which neither predictive nor retrodictive validity is able to assess alone.
8.4.2.3 Validation Through Structural Similarity
In practice, the evaluation of a simulation includes some kind of prediction and

retrodiction, based on expertise and experience. Given the implementation of
micro-level mechanisms in the simulation, classes of behaviour at the macroscopic
scale are identified in the model and compared to classes of behaviour identified in
the target. Similarly, known classes of behaviour in the target system are checked
for existence in the simulation. The former case is generally what we call the
“surprising” character of simulations in which models show something beyond
what we expect them to. However, only an assessment of the model at various
points of view, including its structure and properties on different grains and levels,
will truly determine whether the system reflects the way in which the target system
operates. For instance, do agents’ behaviour, the constituent parts and the structural
evolution of the model match the conception we have about the target system with
satisfactory accuracy? These are examples of the elements of realism between the
model and the system that the researcher strives to find, which requires expertise in
the domain on the part of the person who builds and/or validates the model.
8.4.3 Validation Techniques
In this section we describe validation techniques used in social simulation. Some

are used as common practices in the literature, and most of the terminology has
been inhered from simulation in engineering and computer science, particularly
from the reviews of validation and verification in engineering by Sargent (1999).
All techniques that we describe can be found in the literature, but it would be rare to
160 N. David
find a model in which only one technique was used, consistent with the fact that the
validation process should be diverse. Also, there are no standard names in the
literature and some techniques overlap with others.
8.4.3.1 Face Validity
Face validity is a general kind of test used both before and after the model is put to
use. During the model development process, the various intermediate models are
presented to persons who are knowledgeable about, or are relevant to the problem in
order to assess whether it is compatible with their knowledge and experience and
reasonable for its purpose (Sargent 1999). Face validity may be used for evaluating
the conceptual model, the components thereof, and the behaviour of the
computerised models in terms of categorical outcomes or direct input/output
relationships. This can be accomplished via documentation, graphing visualisation
models, and animation of the model as it moves through time. Insofar as this is a
general kind of test, it is used in several iterations of the model.
8.4.3.2 Turing Tests
People who are knowledgeable about the behaviour of the target system are asked if
they can discriminate between system and model outputs (Sargent 1999). The logic
of Turing tests is the following: If the outputs of a computerised model are
qualitatively or quantitatively indistinguishable from the observation of the target
system, a substantial level of validation has been achieved.
Note that the behaviour of the target system does not need to be observed
directly in the cases where a computerised representation is available. For example,
suppose that videos of car traffic are transformed into three-dimensional scenes,
whereby each object in the scene represents a car following the observed trajectory.
If an independent investigator is not able to distinguish the computerised reproduc-
tion from an agent-based simulation of car traffic, then a substantial level of
validation has been obtained for the set of behaviours represented in the simulation
model.
8.4.3.3 Historical Validity
Historical validity is a kind of retrodiction where the results of the model are
compared with the results of previously collected data. If only a portion of the
available historical data is used to design the model then a related concept is called
out-of-sample tests in which the remaining data are used to test the predicative
capacity of the model.
8.4.3.4 Event Validity
Event validity compares the occurrence of particular events in the model with the
occurrence of events in the source data. This can be assessed at the level of individual
trajectories of agents or at any aggregate level. Events are situations that should occur
according to pre-specified conditions, although not necessarily predictable. Some
events may occur at unpredictable points in time or circumstances. For instance, if the
target system data shows arbitrary periods of stable behaviours interwoven with
periods of volatility with unpredictable turning points, the simulation should produce
similar kinds of unpredictable turning events.
8.4.3.5 Extreme Condition Tests
Extreme conditions are used for both verifying and validating the validation. The
experimenter uses unlikely combinations of factors in the system, usually very high
or low values for the inputs and parameters in order to test whether the simulation
continues to make sense at the margins. For instance, if interaction among agents is
nearly suppressed the modeller should be surprised if such activities as trade or
culture dissemination continues in a population.
8.4.3.6 Sensitivity Analysis
As a precautious rule one should consider that a model is only valid for the range of
parameters that have been tested. Sensitivity analysis stands for tests in which
parameters or even the inter-relations of model components are systematically
varied in order to determine the effect on the behaviour of the model. It aims at
three sorts of related considerations:
• Understanding the basic conditions under which the model behaves as expected;
• Finding the conditions that maximize the agreement of the model behaviour with
the target system behaviour;
• Identifying the conditions that are sensitive, for instance, when changes in the
input yield outputs that do not remain within known intervals, even when
changes are carried out in a very controlled way.
Parameters that are sensitive should be made sufficiently accurate prior to using
the model. If the output remains unpredictable even with controlled changes, the
modeller should be concerned about making claims about the model.
Executing sensitivity tests is not a trivial task, and there are no special methods.
If one imagines sweeping three parameters with 10 distinct values each then 720
configurations of input sets will be defined. If for each configuration we carry out
five runs so as to obtain meaningful statistical figures, we can imagine 3,600 runs of
the model. Since it is likely that some of the parameters will interact they should be
162 N. David
swept in combinations – a fact which makes sensitivity tests already intractable for
only a relatively small number of parameters.
Experience, the availability of empirical data, and the use of sampling techniques
are the usual solution. A possible approach is constraining the range of values
according to empirical data by ignoring ranges of values that are known from the
start to be implausible in the target system. Yet, this might not be possible. The
correspondence between parameter ranges in the model and in the target must be
somehow known a priori, which requires that the model be subjected to some kind of
testing anyway.
Sampling the parameter space is the usual solution. Related techniques in social
simulation include learning algorithms for searching the parameter space, such as the
Active Nonlinear Test (Miller 1998). Genetic algorithms for exploring the parameters
of the model more efficiently are often used.
Besides sweeping parameters, any changes in the conditions of the model should
be tested. Two architectural levels in the model must be considered:
1. The conceptual level, which involves changing the internal mechanisms or sub-
models that constitute the larger model, such as changing the decision processes
of the agents, their learning mechanisms or their interaction topology.
2. The system level, which involves low-level elements of the model, such as the
effect of changing the agent activation regimes (e.g. uniform activation or
random activation).
If changing elements at the system level determines different behaviours of the
model that cannot be adequately interpreted, then the validity of the model can be
compromised. The case of changing elements at conceptual levels is more subtle,
and the validity of the results must be assessed by the researcher with reference to
the validity of the composing elements of the model. This is basically a kind of
cross-model or cross-element validation, as described below.
8.4.3.7 Cross-Sectional Validity
Cross-sectional validity refers to the examination of the fit of congruence of social

data to the results that simulation models produce in a specific point in time. This
may be accomplished by comparing empirical data, such as a cross-sectional
survey, with output generated by a model at a single time period. For example,
simulation models of withdrawal behaviours in organisations, based on fuzzy set
theory, have been used by Munson and Hulin (2000) for comparing correlations
among frequencies of withdrawal behaviours in a cross-section survey with
correlations among simulated behaviours after a certain number of iterations. The
model that generated the correlation matrix that fitted better with the empirical data
(evaluated by calculating root mean squared discrepancies) gained support as the
most useful model of the organisational context that was being analysed. Notwith-
standing, quantitative assessments between cross-sectional and simulated data are
rare. Moreover, the benefits of simulation for representing time suggest an obvious
disadvantage of cross-sectional approaches: they are unable to assess results from a
longitudinal point of view.
8.4.3.8 Comparison to Other Models
There are strong reasons to compare a model with other models. One reason is the
unavailability, or insufficient quality, of data. Another is that models are often
specified at a level of abstraction not compatible with available data. Moreover,
even if data were available, the goodness of fit between real and simulated data, albeit
reflecting evidence about the validity of the model as a data-generating process, does
not provide evidence on how it operates. The most important reason for comparing
models is intrinsic to the scientific practice. The very idea of validation is comparing
models with other descriptions of the problem modelled, and this may include other
simulation models that have been validated to some level.
If two models of the same puzzle lead to different conclusions, then they
motivate us to check the validity of the models. Or, as we mentioned in Sect. 8.2,
indicate that the computerised models have not been appropriately verified. The
bricolage of methods for relating models in social simulation has become more
mature after the aligning experience of Axtell et al. (1996), and is now a highly
cited area. The goal of relating models can have different origins. In Sect. 8.2.4 the
meaning of model alignment was described as well as the different methods for
replicating (computerised) models. These focused on the verification of algorithmic
equivalence of models through comparison of data. There are a number of
approaches for relating models focused on the validation perspective:
• Extending models or composing models in a larger model, where different
concepts, mechanisms or results are abstracted as components of the larger
model; software reuse, as described in Sect. 8.2.2, is a kind of model
composition.
• Docking data produced by a model to a set of data produced by another model;
this may require changing and calibrating the model, for instance, by turning off
features of the former in order to align with the latter.
• Varying systematically and controllably the internal structure of a model; in
other words, playing with models within models in order to assess the overall
validity of the larger model with reference to the validity of each one of the
composing models; this is a kind of sensitivity analysis that resembles cross-
element validation, which is mentioned below.
• Relating different computerised models that are based on different paradigms
and provide different scales and perspectives for the same class of target; for
instance, agent-based models and equation-based modelling.
164 N. David
8.4.3.9 Cross-Element Validity
Cross-element validation, rather than comparing whole models, compares the results
of a model whose architecture of the agents differs only in a few elements. The goal is
to assess the extent to which changing elements of the model architecture produces
results compatible with the expected results of the (larger) model; basically an
exercise of composing different models within a larger model, which resembles
structural sensitivity analysis. For instance, one may study the effects of using a
model with agents in a bargaining game employing either evolutionary learning or
reinforcement learning strategies, and assess which one of the strategies produces
results compatible with theoretical analysis in game theory (Takadama et al. 2003).
A difficult issue faced with cross-element validation is that different results
depend on the different elements used. The results obtained may only be a result of
sensitivity to the elements. How can different tendencies be compared resulting from
using different elements in the model? Arai and Watanabe (2008) have introduced a
promising approach for use with time-series data. A quantitative comparison method
based on the Fourier transform is used for measuring the distance between the results
of two models that use different elements (e.g. different learning mechanisms).
Another problem with abstract simulations, in which the context of the empirical
referent is vague, refers to the lack of real data for comparing which of the
simulation outcomes are more realistic. As a result, the outcomes may only be
assessed against a reference theoretical model. In contrast, when enough data from
the target system are available, cross-element validation becomes a particular case
of validation through retrodiction. For example, the cross-sectional validity tests
employed by Munson and Hulin (2000) are a kind of cross-element validation
through retrodiction, in which different theoretical models of individual withdrawal
behaviours were tested within the same organisational model, with post-model data
fit evaluation. Data generated for each theoretical model was compared with data
from structured interviews in order to determine which theoretical model provided
the best fit to the empirical data.
8.4.3.10 Participatory Approaches for Validation
Participatory approaches refer to the involvement of stakeholders both in the design

and the validation of a model. Such an approach, also known as Companion
Modelling (Barreteau et al. 2001), assumes that model development must be itself
considered in the process of social intervention, where dialogue among
stakeholders, including both informal and theoretical knowledge, is embedded in
the model development process. Rather than just considering the final shape of the
model, both the process and the model become instruments for negotiation and
decision making. It is particularly suited for policy or strategy development. This
topic is discussed in Chap. 10 “Participatory Approaches” (Barreteau et al. 2013).
8.4.4 Relationship to Modelling Strategies
Regarding the diversity of methodological conceptions for modelling, different

strategies may be adopted relating to the level of descriptive richness embedded
in the simulations.
Several taxonomies of modelling strategies have been described in the literature
(see David et al., 2004; Boero et al., 2005; Gibert 2008, pp. 42–44). Normally, the
adoption of these strategies does not depend on the class of the target to be modeled,
but in different ways to address it as the problem domain. For example, if a
simulation is intended to model a system for the purpose of designing policies,
this will imply representing more detail than a simulation intended for modeling
certain social mechanisms of the system in a metaphorical way. But this also means
that there is a trade-off between the effort that a modeler puts into verifying the
simulation and puts into validating it. As more context and richness are embedded
in a model, the more difficult it will be to verify it. Conversely, as one increases the
descriptive richness of a simulation, more ways will be available to assess its
validity. A tension that contrasts the tendency for constraining simulations by
formal-theoretical constructs – normally easier to verify – and constraining
simulations by theoretical-empirical descriptions – more amenable to validation
by empirical and participative-based methods. In the remaining sections of this
chapter, two contrasting modelling strategies are described.
8.4.4.1 Subjunctive Agent-Based Models
A popular strategy in social simulation consists of using models as a means for

expressing subjunctive moods to talk about possible worlds using what-if scenarios,
like ‘what would happen if something were the case’. The goal is building artificial
societies for modelling possible worlds that represent classes of social mechanisms,
while striving for maximal simplicity and strong generalisation power of the
representations used. Reasons for striving for simplicity include the computational
tractability of the model and to keep the data analysis as simple as possible.
Simplicity and generalization power are often seen as elements of elegance in a
model. However, making the model simpler in the social sciences does not neces-
sarily make the model more general. More often than not this kind of modelling
only makes it metaphorically general, or simply counterfactual (with false
assumptions). For example, ‘What would happen if world geography is regarded
as a two-dimensional space arranged on a 10 10 grid, where agents are thought
of as independent political units, such as nations, which have specific behaviours of
interaction according to simple rules?’ To assume that world geography is one-
dimensional, as Axelrod does in his Tribute Model, is clearly a false assumption.
Often these models are associated with a design slogan coined by Axelrod, called
the KISS approach – “Keep it Simple Stupid”. Despite their simplicity, these
models prove useful for concept formation and theoretical abstraction. The
166 N. David
emergence of macro regularities from micro-levels of interaction becomes

the source of concept formation and hypothesis illustration, with the power of
suggesting novel theoretical debates.
Given the preference for simplification and abstraction, mechanisms used in
these models are normally described in a formalized or mathematical way.
Axelrod’s models, such as the culture dissemination model, or Schelling’s residen-
tial segregation model, are canonical examples. Their simplicity and elegance have
been factors for popularity and dissemination that span numerous disciplines and
ease replication and verification.
However, whereas simplicity eases verification, the use of metaphorical models
also brings disadvantages. Consider a word composed of several attributes that
represents an agent’s culture, such as in Axelrod’s culture dissemination model. The
attributes do not have any specific meaning and are only distinguishable by their
relative position in the word and so can be interpreted according to a relatively
arbitrary number of situations or social contexts. However, such a representation
may also be considered too simplified to mean anything relevant for such a complex
concept as a cultural attribute. As a consequence, verification is hardly distinguish-
able from validation, insofar as the model does not represent a specific context of
social reality. In such a sense, the researcher is essentially verifying experimentally
whether the conceptions that he has in his mind are met by an operationalisation that
is computationally expressed (David et al. 2007). Nevertheless, given their simplic-
ity, subjunctive models can be easily linked and compared to other models,
extended with additional mechanisms, as well as modified for model alignment,
docking, or replication. Cross-element validation is a widely used technique.
At any rate, the fact that these models are easily replicable and comparable – but
hardly falsifiable by empirically acquired characteristics of social reality – stresses
their strong characteristic: when models based on strategies of maximal simplicity
become accepted by a scientific community, their influence seems to reach several
other disciplines and contexts. Perhaps for this reason, these kind of models are the
most popular in social simulation, and some models are able to reach a considerable
impact in many strains of social science.
8.4.4.2 Context-Specific Agent-Based Models
It would be simplistic to say that models in social simulation can be characterized

according to well-defined categories of validation strategies. Even so, the capacity
to describe social complexity, whether through simplicity or through rich detail and
context, is a determining factor for a catalogue of modelling strategies.
We cannot hope to model general social mechanisms that are valid in all
contexts. There are many models that are not designed to be markedly general or
metaphorically general, but to stress accurateness, diversity, and richness of
description. Instead of possible worlds representing very arbitrary contexts, models
are explicitly bounded to specific contexts. Constraints imposed on these models
can vary from models investigating properties of social mechanisms in a large band
of situations which share common characteristics, to models with the only ambition
of representing a single history, like Dean’s retrodiction of the patterns of settle-
ment of a specific population in the southwestern United States, household by
household (see Dean et al. 2000).
Constructing and validating a model of this kind requires the use of empirical
knowledge. They are, for this reason, often associated with the idea of “Empirical
Validation of Agent-Based Models.”
What is the meaning of empirical in this sense? If the goal is to discuss empirical
claims, then models should attempt to capture empirically enquired characteristics
of the target domain. Specifying the context of descriptions in the model will
typically provide more ways for enquiring quantitative and qualitative data in the
target, as well as using experimental and participative methods with stakeholders.
In this sense, empirical may be understood as a stronger link between the model and
a context-specific, well-circumscribed, problem domain.
The model of Dean et al. (2000), which attempted to retrodict the patterns of
settlement of the Anasazi in the Southwestern United States, household by house-
hold, is a well-known and oft-cited example of a highly contextualized model built
on the basis of numerous sources, from archaeological data to anthropological,
agricultural and ethnographic analyses, in a multidisciplinary context.
Given the higher specificity of the target domain, the higher diversity of ways for
enriching the model as well as the increased semantic specificity of the outputs
produced by the model, context-specific models may be more susceptible to be
compared with empirical results of other methods of social research. On the other
hand, comparison with other simulation models is complex and these models are
more difficult to replicate and verify.
8.4.4.3 Modus Operandi: Formal and Informal Approaches
The tension between simplicity and descriptive richness expresses two different
ways for approaching the construction and validation of a model. One can start with
a rich, complex, realistic description and only simplify it where this turns out to be
possible and irrelevant to the target system – known as the KIDS approach
(Edmonds and Moss 2005). Or one starts from the outset with the simplest possible
description and complexifies it only when it turns out to be necessary to make the
model more realistic, nevertheless keeping the model as simple as possible – known
as the KISS approach (Axelrod 1997b).
In practice, both trends are used for balancing trades-offs between the model’s
descriptive accuracy and the practicality of modelling, according to the purpose and
the context of the model. This raises yet another methodological question: the
extent to which models ought to be designed on the basis of formal theories, or
ought to be constrained by techniques and approaches just on the basis of the
intuition of the model builders and stakeholders. As we have seen, strong, subjunc-
tive, agent-based models with metaphorical purposes tend to adopt the simplicity
motto with extensive use of formal constructs, making the models more elegant
168 N. David
from a mathematical point of view, easier to verify, but less liable to validation
methods. Game theoretical models, with all their formal and theoretical apparatus,
are a canonical example. Results from these models are strongly constrained by the
formal theoretical framework used.
A similar problem is found when agent-based models make use of cognitive
architectures strongly constrained by logic-based formalisms, such as the kind of
formalisms used to specify BDI-type architectures. If the cognitive machinery of
the agents relies on heuristic approaches that have been claimed valid, many
researchers in the literature claim that cognitive agent-based models can be
validated in the empirical sense of context-specific models. Cited examples of
this kind usually point to agent-based models based on the Soar architecture.
At any rate, context-specific models are normally more eclectic and make use of
both formal and informal knowledge, often including stakeholder evidence in order
to build and validate the models. Model design tends to be less constrained a priori
by formal constructs. In principle, one starts with all aspects of the target domain
that are assumed to be relevant and then explores the behaviour of the model in
order to find out whether there are aspects that do not prove relevant for the purpose
of the model. The typical approach for modelling and validation can be summarized
in a cycle with the following iterative and overlapping steps:
A. Building and validating pre-computerised and computerised models: Several
descriptions and specifications are used to build a model, eventually in the form
of a computer program, which are micro-validated against a theoretical frame-
work and/or empirical knowledge, usually qualitatively. This may include the
individual agents’ interaction mechanisms (rules of behaviour for agents or
organisations of agents), their internal mechanisms (e.g. their cognitive machin-
ery), the kind of interaction topology or environment, and the passive entities
with which the agents interact. The model used should be as accurate as possible
for the context in consideration as well as flexible for testing how parameters
vary in particular circumstances. Empirical data – if available – should be
used to help configure the parameters. Both the descriptions of the model and
the parameters used should be validated for the specific context of the model.
For example, suppose empirical data are available for specifying the consumer
demand of products. If the demand varies from sector to sector, one may use
data to inform the distribution upon which the parameter could be based for each
specific sector.
B. Specifying expected behaviours of the computerised model: Micro and macro
characteristics that the model is designed to reproduce are established from the
outset based on theoretical and/or empirical knowledge. Any property, from
quantitative to qualitative measures, such as emergent key facts the model should
reproduce (stylized facts), the statistical characteristic or shape of time-data series
(statistical signatures) and individual agents’ behaviour along the simulation
(individual trajectories), can be assessed. This may be carried out in innumerable
ways, according to different levels of description or grain, and be more or less
general depending on the context of the model and the kind of empirical
knowledge available. For instance, in some systems it may be enough to predict

just a “weak” or “positive” measure on some particular output, such as a positive
and weak autocorrelation. Or we might look for the emergence of unpredictable
events, such as stable regimes interleaved with periods of strong volatility, and
check their statistical properties for various levels of granularity. Or the emer-
gence of different structures or patterns associated with particular kinds of agents,
such as groups of political agents with “extremist” or “moderate” agents.
C. Testing the computerised model and building and validating post-computerised
models: The computerised model is executed. Both individual and aggregate
characteristics are computed and tested for sensitivity analysis. These are micro-
validated and macro-validated against the expected characteristics of the model
established in step B according to a variety of validation techniques, as
described in Sects. 8.4.2 and 8.4.3. A whole process of building post-
computerised models takes place, possibly leading to the discovery of unex-
pected characteristics in the behaviour of the computerised model which should
be assessed with further theoretical or empirical knowledge about the problem
domain.
Further Reading
Good introductions to validation and verification of simulation models in general

are Sargent (1999) and Troitzsch (2004), the latter with a focus on social simula-
tion. Validation of agent-based models in particular is addressed by Amblard and
colleagues (Amblard et al. 2007).
For readers more interested in single aspects of V&V with regard to agent-based
models in the context of social simulation, the following papers provide highly
accessible starting points:
• Edmonds and Hales (2003) demonstrate the importance of model replication
(or model alignment) by means of a clear example.
• Boero and Squazzoni (2005) examine the use of empirical data for model
calibration and validation and argue that “the characteristics of the empirical
target” influence the choice of validation strategies.
• Moss and Edmonds (2005) discuss an approach for cross-validation that
combines the involvement of stakeholders to validate the model qualitatively
on the micro level with the application of statistical measures to numerical
outputs to validate the model quantitatively on the macro level.
Finally, for a more in-depth epistemological perspective on verification and
validation I would refer the inclined reader to a revised version of my EPOS
2006 paper (David 2009).
170 N. David
References
Amblard F, Bommel P, Rouchier J (2007) Assessment and validation of multi-agent models.

In: Phan D, Amblard F (eds) Agent-based modelling and simulation in the social and human
sciences. The Bardwell Press, Oxford, pp 93–114
Arai R, Watanabe S (2008) A quantitative comparison method for multi-agent based simulation in
the feature space. In: David N, Sichman JS (eds) Multi-agent-based simulation IX, MABS
2008, revised selected papers (Lecture notes in artificial intelligence), vol 5269. Springer,
Berlin, pp 154–166
Axelrod R (1997a) The dissemination of culture: a model with local convergence and global
polarization. J Conf Resolut 41(2):203–226
Axelrod R (1997b) Advancing the art of simulation in the social sciences. In: Conte R,
Hegselmann R, Terna P (eds) Simulating social phenomena. Springer, Berlin, pp 21–40
Axtell R, Axelrod R, Epstein J, Cohen M (1996) Aligning simulation models: a case study and
results. Comp Math Organiz Theory 1(2):123–141
Barreteau O, Bousquet F, Attonaty JM (2001) Role-playing games for opening the black box of
multi-agent systems: method and lessons of its application to Senegal River Valley irrigated
systems. J Artif Soc Soc Simul 4(2). http://jasss.soc.surrey.ac.uk/4/2/5.html
Boero R, Squazzoni F (2005) Does empirical embeddedness matter? Methodological issues on
agent-based models for analytical social science. J Artif Soc Soc Simul 8(4). http://jasss.soc.
David N (2009) Validation and verification in social simulation: patterns and clarification of
terminology. In: Squazzoni F (ed) Epistemological aspects of computer simulation in the social
sciences, second international workshop, EPOS 2006, Brescia, 5–6 Oct 2006, revised selected
and invited papers (Lecture notes in computer science), vol 5466. Springer, Berlin, pp 117–129
David N, Sichman J, Coelho H (2003) Towards an emergence-driven software process for agent-
based-simulation. In: Sichman JS, Bousquet F, Davidsson P (eds) Multi-agent-based simula-
tion II, third international workshop, MABS 2002, Bologna, 15–16 July 2002, revised papers
(Lecture notes in artificial intelligence), vol 2581. Springer, Berlin, pp 89–104
David N, Marietto M, Sichman J, Coelho H (2004) The structure and logic of interdisciplinary
research in agent-based social simulation. J Artif Soc Soc Simul 7(3). http://jasss.soc.surrey.ac.
uk/7/3/4.html
David N, Sichman J, Coelho H (2007) Simulation as formal and generative social science: the very
idea. In: Gershenson C, Aerts D, Edmonds B (eds) Worldviews, science and us: philosophy and
complexity. World Scientific, Singapore, pp 266–284
David N, Caldas JC, Coelho H (2010) Epistemological perspectives on simulation III: selected papers
of EPOS 2008. J Artif Soc Soc Simul 13(1), special section. http://jasss.soc.surrey.ac.uk/13/1/
Dean S et al (2000) Understanding Anasazi culture change through agent-based modeling.
In: Kohler T, Gumerman G (eds) Dynamics in human and primate societies: agent-based
modeling of social and spatial processes. Oxford University Press, New York/Oxford,
pp 179–205
Edmonds B, Hales D (2003) Replication, replication and replication: some hard lessons from
model alignment. J Artif Soc Soc Simul 6(4). http://jasss.soc.surrey.ac.uk/6/4/11.html
Edmonds B, Moss S (2005) From KISS to KIDS: an ‘anti-simplistic’ modelling approach.
In: Davidsson P, Logan B, Takadama K (eds) Multi-agent and multi-agent-based simulation
(Lecture notes in artificial intelligence), vol 3415. Springer, Berlin, pp 130–144
Frank U, Troitzsch KG (2005) Epistemological perspectives on simulation. J Artificial Soc Soc
Simul 8(4). http://jasss.soc.surrey.ac.uk/8/4/7.html
Galán JM et al (2013) Checking simulations: detecting and avoiding errors and artefacts. Chapter 6
in this volume
Gilbert N (2008) Agent-based models, vol 153, Quantitative applications in the social sciences.
Sage, London
Gross D, Strand R (2000) Can agent-based models assist decisions on large-scale practical
problems? A philosophical analysis. Complexity 5(6):26–33
Kleijnen JPC (1995) Verification and validation of simulation models. Euro J Operat Res 82:
145–162
Law A, Kelton D (1991) Simulation modelling and analysis, 2nd edn. McGraw Hill, New York
Miller JH (1998) Active nonlinear tests (ANTs) of complex simulation models. Manage Sci 44(6):
820–830
Moss S, Edmonds B (2005) Sociology and simulation: statistical and qualitative cross-validation.
Am J Sociol 110(4):1095–1131
Munson L, Hulin C (2000) Examining the fit between empirical data and theoretical simulations.
In: Ilgen DR, Hulin CL (eds) Computational modeling of behaviour in organisations: the third
scientific discipline. American Psychological Association, Washington, DC, pp 69–84
Rouchier J, Cioffi-Revilla C, Polhill JG, Takadama K (2008) Progress in model-to-model analysis.
J Artif Soc Soc Simul 11(2). http://jasss.soc.surrey.ac.uk/11/2/8.html
Sargent R (1999) Verification and validation of simulation models. In: Farrington PA, Nembhard
HB, Sturrock DT, Evans GW (eds) Proceedings of the 1999 winter simulation conference, Squaw
Peak, Phoenix, 5–8 Dec 1999, pp 39–48. http://www.informs-sim.org/wsc99papers/005.PDF
Sommerville I (1995) Software engineering, 5th edn. Addison Wesley, Boston
Squazzoni F (2009) Epistemological aspects of computer simulation in the social sciences,
second international workshop, EPOS 2006, Brescia, Revised selected and invited papers
(Lecture notes in computer science), vol 5466. Springer, Berlin, 5–6 Oct 2006
Takadama K, Suematsu Y, Sugimoto N, Nawa N, Shimohara K (2003) Cross-element validation in
multi-agent-based simulation: switching learning mechanisms in agents. J Artif Soc Soc Simul
6(4). http://jasss.soc.surrey.ac.uk/6/4/6.html
Troitzsch K (2004) Validating simulation models. In: Horton G (ed) Proceedings of the 18th European
simulation multiconference, ESM 2004, 13–16 June 2004, Magdeburg. SCS, Erlangen,
pp 265–270
Chapter 9
Understanding Simulation Results
Andrew Evans, Alison Heppenstall, and Mark Birkin
Why Read This Chapter? Overall, this chapter aims to help you understand the
results that a simulation model produces, by suggesting some ways to analyse and
visualise them. The chapter concentrates on the internal dynamics of the model
rather than its relationship to the outside world.
Abstract Simulation modelling is concerned with the abstract representation of
entities within systems and their inter-relationships; understanding and visualising
these results is often a significant challenge for the researcher. Within this chapter
we examine particular issues such as finding “important” patterns and interpreting
what they mean in terms of causality. We also discuss some of the problems with
using model results to enhance our understanding of the underlying social systems
which they represent, and we will assert that this is in large degree a problem of
isolating causal mechanisms within the model architecture. In particular, we high-
light the issues of identifiability and equifinality – that the same behaviour may be
induced within a simulation from a variety of different model representations or
parameter sets – and present recommendations for dealing with this problem. The
chapter ends with a discussion of avenues of future research.
9.1 Introduction
Simulation models may be constructed for a variety of purposes. Classically these

purposes tend to centre on either the capture of a set of knowledge or making
predictions. Knowledge capture has its own set of issues that are concerned
with structuring and verifying knowledge in the presence of contradiction and
uncertainty. The problems of prediction, closely associated with calibration and
A. Evans (*) • A. Heppenstall • M. Birkin

Centre for Spatial Analysis and Policy, University of Leeds, Leeds, Britain, LS2 9JT

174 A. Evans et al.
validation, centre around comparisons with real data, for which the methods
covered in Chap. 8 (David 2013) are appropriate. In this chapter, however, we
look at what our models tell us through their internal workings and logic; how we
might understand/interpret simulation results as results about an attempted simu-
lation of the real world, rather than as results we expect to compare directly with
the world. Here then, we tackle the third purpose of modelling: the exploration of
abstracted systems through simulation. In a sense, this is a purpose predicated
only on the limitations of the human mind. By common definition, simulation
modelling is concerned with abstract representations of entities within systems
and their interrelationships, and with the exploration of the ramifications of these
abstracted behaviours at different temporal and geographical scales. In a world in
which we had larger brains, models would not be required to reveal anything – we
would instantly see the ramifications of abstracted behaviours in our heads. To a
degree, therefore, models may be seen as replacing the hard joined-up thinking
that is required to make statements about the way the world works. This chapter
looks at what this simplifying process tells us about the systems we are trying to
replicate.
In part, the complications of simulation modelling are a product of the dimension-
ality of the systems with which we are dealing. Let us imagine that we are tackling a
system of some spatio-temporal complexity, for example, the prices in a retail market
selling items A, B, and C. Neighbouring retailers adjust their prices based on local
competition, but the price of raw materials keeps the price surface out of equilibrium.
In addition, customers will only buy one of the products at a time, creating a link
between the prices of the three items. Here, then, we have three interdependent
variables, each of which varies spatio-temporally, with strong auto- and cross-
correlations in both time and space. What kinds of techniques can be used to tease
apart such complex systems? In Sect. 9.2 of this chapter we will discuss some of the
available methodologies broken down by the dimensionality of the system in ques-
tion and the demands of the analysis. Since the range of such techniques is extremely
sizable, we shall detail a few traditional techniques that we believe might be helpful
in simplifying model data that shows the traits of complexity, and some of the newer
techniques of promise.
Until recently, most social science models represented social systems using
mathematical aggregations. We have over 2,500 years’ worth of techniques to
call upon that are founded on the notion that we need to simplify systems as rapidly
as we can to the point at which the abstractions can be manipulated within a single
human head. As is clear, not least from earlier contributions in this volume, it is
becoming increasingly accepted that social scientists might reveal more about
systems by representing them in a less aggregate manner. More specifically, the
main difference between mathematics and the new modelling paradigm is that we
now aspire to work at a scale at which the components under consideration can be
represented as having their own discrete histories; mathematics actually works in a
very similar fashion to modern models, but at all the other scales. Naturally there
are knock-ons from this in terms of the more explicit representation of objects,
states and events, but these issues are less important than the additional simulation
9 Understanding Simulation Results 175
and analytical power that having a history for each component of a system gives us.
Of course, such a “history” may just be the discrete position of an object at a single
historical moment, and plainly at this level of complication the boundary between
such models and, for example, Markov models is somewhat diffuse, however, as
the history of components becomes more involved, so the power of modern
modelling paradigms comes to the fore. What is lacking, however, are the
techniques that are predicated on these new architectures. Whilst models which
are specified at the level of individual entities or ‘agents’ may also be analysed
using conventional mathematical techniques, in Sect. 9.3 of the chapter we will
discuss some more novel approaches which are moving the direction of understand-
ing the outputs of these new, unaggregated, models on their own terms.
One of the reasons that simulation models are such a powerful methodology for
understanding complex systems is their ability to display aggregate behaviour
which goes beyond the simple extrapolation of the behaviour of the individual
component parts. In mathematical analysis, such as dynamical systems theory, this
behaviour tends to be linked to notions of equilibrium, oscillation, and catastrophe
or bifurcation. Individual and agent-based modelling approaches have veered more
strongly towards the notion of emergence, which can be defined as “an unforeseen
occurrence; a state of things unexpectedly arising” (OED 2010). The concept of
emergence is essentially a sign of our ignorance of the causal pathways within a
system. Nevertheless, emergence is our clearest hope for developing an understand-
ing of systems using models. We hope that emergence will give us a perceptual
shortcut to the most significant elements of a system’s behaviour. When it comes to
applications, however, emergence is a rather double-edged blade: emergence hap-
pily allows us to see the consequence of behaviours without us having to follow the
logic ourselves, however it is problematic in relying upon us to filter out which of
the ramifications are important to us. As emergence is essentially a sign of incom-
plete understanding, and therefore weakly relative, there is no objective definition
of what is “important”. One might imagine one day a classification of the kinds of
patterns that relate to different types of causal history, but there is no objective
manner of recognising a pattern as “important” as such. These two problems:
finding “important” patterns (in the absence of any objective way of defining
“important”) and then interpreting what they mean in terms of causality are the
issues standing between the researcher and perfect knowledge of a modelled
system. In the fourth section of this chapter, we will discuss some of the problems
with using model results to enhance our understanding of the underlying social
systems which they represent, and we will assert that this is in large degree a
problem of isolating causal mechanisms within the model architecture. In particu-
lar, we highlight the issues of equifinality and identifiability – that the same
behaviour may be induced within a simulation from a variety of different model
representations or parameter sets – and present recommendations for dealing with
this problem. Since recognising emergence and combating the problems of
identifiability and equifinality are amongst the most urgent challenges to effective
modelling of complex systems, this leads naturally to a discussion of future
directions in the final section of the chapter.
176 A. Evans et al.
9.2 Aggregate Patterns and Conventional Representations

of Model Dynamics
Whether a model is based on deductive premises or inferred behaviours, any new

understanding of a given modelled system tends to be developed inductively.
Modellers examine model outputs, simplify them, and then try to work out the
cause utilising a combination of hypothesis dismissal, refinement and experimenta-
tion. For example, a modeller of a crowd of people might take all the responses of
each person over time and generate a single simple mean statistic for each person;
these might then be correlated against other model variables. If the correlation
represents a real causal connection, then varying the variables should vary the
statistic. Proving such causal relationships is not something we often have the
ability to do in the real world. During such an analysis the simplification process
is key: it is this that reveals the patterns in our data. The questions are: how do we
decide what needs simplifying and, indeed, how simple to make it?
We can classify model results by the dimensionality of the outputs. A general
classification for social systems would be:
• Single statistical aggregations (1D)
• Time series of variables (2D)
• The spatial distributions of invariants (2D) or variables (3D)
• Spatio-temporal locations of invariants (3D) or variables (4D)
• Other behaviours in multi-dimensional variable space (nD)
For simplicity, this assumes that geographical spaces are essentially two-
dimensional (while recognising that physical space might also be represented
along linear features such as a high street, across networks, or within a three
dimensional topographical space for landforms or buildings). It should also be
plain that in the time dimension, models do not necessarily produce just a stream
of data, but that the data can have complex patternation. By their very nature,
individual-level models, predicated as they are on a life-cycle, will never stabilise
in the way a mathematical model might (Grimm 1999); instead models may run
away or oscillate, either periodically or chaotically.
Methods for aiding pattern recognition in data break down, again, by the
dimensionality of the data, but also by the dimensionality of their outputs. It is
quite possible to generate a one-number statistic for a 4D spatio-temporal distribution.
In some cases, the reduction of dimensionality is the explicit purpose of the technique,
and the aim is that patterns in one set of dimensions should be represented as closely
as possible in a smaller set of dimensions so they are easier to understand. Table 9.1
below presents a suite of techniques that cross this range (this is not meant to be an
exhaustive list; after all, pattern recognition is a research discipline of its own with a
whole body of literature including several dedicated journals).
To begin with, let us consider some examples which produce outputs in a single
dimension. In other words, techniques for generating global and regional statistics
describing the distribution of variables across space, either a physical space, or a
Table 9.1 Pattern recognition techniques for different input and output data dimensions
1D output 2D output 3D output 4D output ND
1D input
2D input Exploratory Cluster
statistics locating
Fourier/wavelet
transforms
3D input Entropy Phase diagrams
statistics Fourier/wavelet
transforms
4D input Diffusion Time slices Recurrence
statistics plots
nD Network Eigenvector Sammon Animations Heuristic
statistics analysis mapping techniques
variable space. Such statistics generally tend to be single time-slice, but can be
generated for multiple time-slices to gauge overall changes in the system dynamics.
Plainly, standard aggregating statistics used to compare two distributions, such as
the variable variance, will lose much of interest, both spatially and temporally. If we
wish to capture the distribution of invariants, basic statistics like Nearest-Neighbour
(Clark and Evans 1954) or the more complex patch shape, fragmentation and
connectivity indices of modern ecology (for a review and software see McGarigal
2002) provide a good starting point. Networks can be described using a wide variety
of statistics covering everything from shortest paths across a network, to the quantity
of connections at nodes (for a review of the various statistics and techniques
associated with networks, see Boccaletti et al. 2006; Evans 2010). However, we
normally wish to assess the distribution of a variable across a surface – for example a
price surface or a surface of predicted retail profitability. One good set of global
measures for such distributions are entropy statistics. Suppose we have a situation in
which a model is trying to predict the number of individuals that buy product A in one
of four regions. The model is driven by a parameter, beta. In two simulations we get
the following results: simulation one (low beta): 480, 550, 520, 450; simulation two:
(high beta) 300, 700, 500, 400. Intuitively the first simulation has less dispersal or
variability than the second simulation. An appropriate way to measure this variability
would be through the use of entropy statistics. The concept of entropy originates in
thermodynamics, where gases in a high entropy state contain dispersed molecules.
Thus high entropy equates to high levels of variability. Entropy statistics are closely
related to information statistics where a high entropy state corresponds to a high
information state. In the example above, simulation two is said to contain more
‘information’ than simulation one, because if we approximate the outcome using
no information we would have a flat average – 500, 500, 500, 500 – and this is closer
178 A. Evans et al.
to simulation one than simulation two. Examples of entropy and information statistics
include Kolmogorov-Chaitin, mutual information statistics and the Shannon infor-
mation statistic. Most applications in the literature use customised code for the
computation of entropy statistics, although the computation of a limited range of
Generalised Entropy indices is possible within Stata.1 Entropy statistics can also be
used to describe flows across networks. In this sense they provide a valuable addition
to networks statistics: most network statistics concentrate on structure rather than the
variable values across them. Unless they are looking specifically at the formation of
networks over time, or the relationship between some other variable and network
structure, modellers are relatively bereft of techniques to look at variation on a
network.
In the case where variability is caused and constrained by neighbourhood effects,
we would expect the variation to be smoother across a region. We generally expect
objects in space under neighbourhood effects to obey Tobler’s first law of geography
(1970) that everything is related, but closer things are related more. This leads to
spatial auto- or cross-correlation, in which the values of variables at a point reflect
those of their neighbours. Statistics for quantifying such spatial auto- or cross-
correlation at the global level, or for smaller regions, such as Moran’s I and Geary’s
C, are well-established in the geography literature (e.g. Haining 1990); a useful
summary can be found in Getis (2007).
Such global statistics can be improved on by giving some notion of the direction
of change of the auto- or cross-correlation. Classically this is achieved through
semi-variograms, which map out the intensity of correlation in each direction
traversed across a surface (for details, see Isaaks and Srivastava 1990). In the
case where it is believed that local relationships hold between variables, local linear
correlations can be determined, for example using Geographically Weighted
Regression (GWR: for details, see Fotheringham et al. 2002). GWR is a technique
which allows the mapping of R2s calculated within moving windows across a
multivariate surface and, indeed, mapping of the regression parameter weights.
For example, it would be possible in our retail results to produce a map of the
varying relationship between the amount of A purchased by customers and the
population density, if we believed these were related. GWR would not just allow a
global relationship to be determined, but also how this relationship changed across
a country. One important but somewhat overlooked capability of GWR is its ability
to assess how the strength of correlations varies with scale by varying the window
size. This can be used to calculate the key scales at which there is sufficient overlap
between the geography of variables to generate strong relationships (though some
care is needed in interpreting such correlations, as correlation strength generally
increases with scale: Robinson 1950; Gehlke and Biehl 1934). Plainly, identifying
the key scale at which the correlations between variables improve gives us some
1
Confusingly, ‘generalised entropy’ methods are also widely used in econometrics for the
estimation of missing data. Routines which provide this capability, e.g. in SAS, are not helpful
in the description of simulation model outputs!
ability to recognise key distance-scales at which causality plays out. In our example,
we may be able to see that the scale at which there is a strong relationship
between sales of A and the local population density increases as the population
density decreases, suggesting rural consumers have to travel further and a concom-
itant non-linearity in the model components directing competition.
If, on the other hand, we believe the relationships do not vary smoothly across a
modelled surface, we instead need to find unusual clusters of activity. The ability to
represent spatial clustering is of fundamental importance, for example, within
Schelling’s well-known model of segregation in the housing market (Schelling
1969). However clustering is often not so easy to demonstrate within both real
data and complex simulation outputs. The most recent techniques use, for example,
wavelets to represent the regional surfaces, and these can then be interpreted for
cluster-like properties. However, for socio-economic work amongst the best soft-
ware for cluster detection is the Geographical Analysis Machine (GAM), which not
only assesses clustering across multiple scales, but, also allows assessment of
clustering in the face of variations in the density of the population at risk. For
example, it could tell us where transport network nodes were causing an increase in
sales of A, by removing regions with high sales caused by high population density
(the population “at risk” of buying A). Clusters can be mapped and their signifi-
cance assessed (Openshaw et al. 1988).
Often, simulations will be concerned with variations in the behaviour of systems,
or their constituent agents, over time. In common with physical systems, social and
economic systems are often characterised by periodic behaviour, in which similar
states recur, although typically this recurrence is much less regular than in many
physical systems. For example, economic markets appear to be characterised by
irregular cycles of prosperity and depression. Teasing apart a model can provide non-
intuitive insights into such cycles. For example, Heppenstall et al. (2006) considered
a regional network of petrol stations and showed within an agent simulation how
asymmetric cyclical variations in pricing (fast rises and slow falls), previously
thought to be entirely due to a desire across the industry to maintain artificially
high profits, could in fact be generated from more competitive profit-maximisation in
combination with local monitoring of network activity. While it is, of course, not
certain these simpler processes cause the pattern in real life, the model exploration
does give researchers a new explanation for the cycles and one that can be
investigated in real petrol stations.
In trying to detect periodic behaviour, wavelets are rapidly growing in popularity
(Graps 2004). In general, one would assume that the state of the simulation can be
represented as a single variable which varies over time (let’s say the average price
of A). A wavelet analysis of either observed or model data would decompose this
trend into chunks of time at varying intervals, and in each interval the technique
identifies both a long-term trend and a short-term fluctuation. Wavelets are there-
fore particularly suitable for identifying cycles within data. They are also useful as
filters for the removal of noise from data, and so may be particularly helpful in
trying to compare the results from a stylised simulation model with observed data
which would typically be messy, incomplete or subject to random bias. It has been
180 A. Evans et al.
argued that such decompositions are fundamentally helpful in establishing a basis

for forecasting (Ramsey 2002). A software environment for wavelet analysis is
provided by Wavelab, using procedures derived from the Matlab statistical
package.
Wavelets are equally applicable in both two and three dimensions. For example,
they may be useful in determining the diffusion of waves across a two-dimensional
space and over time, and can be used to analyse, for example, the relationship
between wave amplitude and propagation distance. Viboud et al. (2006) provide a
particularly nice example of such a use, looking at the strength of the propagation of
influenza epidemics as influenced by city size and average human travel distances
in the USA. Other, more traditional statistics, such as the Rayleigh statistic (Fisher
et al. 1987; Korie et al. 1998) can also be used to assess the significance of diffusion
from point sources.
In addition to global and regional aggregate statistics of single variables or cross-
correlations, it may be that there is simply too great a dimensionality to recognise
patterns in outputs and relate them to model inputs. At this point it is necessary to
engage in multi-dimensional scaling. If an individual has more than four
characteristics, then multi-dimensional scaling methods can be used to represent the
individuals in two or three dimensions. In essence, the problem is to represent the
relation between individuals such that those which are most similar in n-dimensions
still appear to be closest in a lower dimensional space which can be visualised more
easily. The most popular technique is Sammon mapping. This method relies on the
ability to optimise an error function which relates original values in high dimensional
space to the transformed values. This can be achieved using standard optimisation
methods within packages such as Matlab. Multidimensional scaling can be useful in
visualising the relative position of different individuals within a search space, for
exploring variations in a multi-criteria objective function within a parameter space, or
for comparing individual search paths within different simulations (Pohlheim 2006).
Eigenvector methods are another form of multi-dimensional scaling. Any multi-
dimensional representation of data in n-dimensional space can be transformed into
an equivalent space governed by n orthogonal eigenvectors. The main significance
of this observation is that the principal eigenvector constitutes the most efficient
way to represent a multi-dimensional space within a single value. For example,
Moon, Schneider and Carley (2006) use the concept of ‘eigenvector centrality’
within a social network to compute a univariate measure of relative position based
on a number of constituent factors.
Clustering techniques collapse multi-dimensional data so that individual cases
are members of a single group or cluster, classified on the basis of a similarity
metric. The method may therefore be appropriate if the modeller wishes to under-
stand the distribution of an output variable in relation to the combination of several
input variables. For example, Schreinemachers and Berger (2006) present a model
of land-use in relation to farm characteristics such as soil quality, agricultural
intensity, market orientation, water use, social capital and migration. In order to
understand land use outcomes (for example, preference for crops versus livestock)
farms were clustered into nine groups based on their shared characteristics. Cluster
analysis is easy to implement in all the major statistics packages (SAS, SPSS,
BMDP). The technique is likely to be most useful in empirical applications with a
relatively large number of agent characteristics (i.e. six or more) than in idealised
simulations with simple agent rules. One advantage of this technique over multidi-
mensional scaling is that it is possible to represent statistical variation within the
cluster space.
9.3 Individual Patterns, Novel Approaches and Visualisation
Plainly aggregate statistics like those above are a useful way of simplifying
individual-level data, both in terms of complexity and dimensionality. However,
they are the result of over 2,500 years of mathematical development in a research
environment unsuited to the mass of detail associated with individual-level data.
Now, computers place us in the position of being able to cope with populations of
individual-level data at a much smaller scale. We still tend to place our own
understanding at the end of an analytical trail, constraining the trail to pass through
some kind of simplification and higher level of aggregation for the purposes of
model analysis. Despite this, it is increasingly true that individual-level data is dealt
with at the individual-level for the body of the analysis and this is especially true in
the case of individual-level modelling, in which experimentation is almost always
enacted at the individual-level. Whether it is really necessary to simplify for human
understanding at the end of an analysis is not especially clear. It may well be that
better techniques might be developed to do this than those built on an assumption of
the necessity of aggregation.
At the individual-level, we are interested in recognising patterns in space and time,
seeing how patterns at different scales affect each other, and then using this to say
something about the behaviour of the system/individuals. Patterns are often indicators
of the attractors to which individuals are drawn in any given system, and present a
shortcut to understanding the mass of system interactions. However, it is almost as
problematic to go through this process to understand a model as it is, for example, to
derive individual-level behaviours from real large-size spatio-temporal datasets of
socio-economic attributes. The one advantage we have in understanding a model is
that we do have some grip on the foundation rules at the individual-scale. Nonethe-
less, understanding a rule, and determining how it plays out in a system of multiple
interactions are very different things. Table 9.2 outlines some of the problems.
Our chief tool for individual-level understanding without aggregation is, and
always has been, the human ability to recognise patterns in masses of data.
Visualisation, for all its subjectivity and faults, remains a key element of the
research process. The standard process is to present one or more attributes of the
individuals in a map in physical or variable space. Such spaces can then be evolved
in movies or sliced in either time or space (Table 9.3 shows some examples). In
general, we cannot test the significance of a pattern without first recognising it
exists, and to that extent significance testing is tainted by the requirement that it test
182 A. Evans et al.
Table 9.2 Issues related to understanding a model at different levels of complexity

Complexity Issues
Spatial What is the impact of space (with whom do individuals initiate transactions
and to what degree)?
Temporal How does the system evolve?
Individuals How do we recognise which individual behaviours are playing out in the
morass of interactions?
Relationships How do we recognise and track relationships?
Scale How can we reveal the manner in which individual actions affect the
large scale system and vice versa?
our competency in recognising the correct pattern as much as that the proposed
pattern represents a real feature of the distribution of our data. Visualisation is also a
vital tool in communicating results within the scientific community and to the wider
public. The former is not just important for the transmission of knowledge, but
because it allows others to validate the work. Indeed, the encapsulation of good
visualisation techniques within a model framework allows others to gain deeper
understanding of one’s model, and to experiment at the limits of the model – what
Grimm (2002) calls “Visual Debugging”. Good model design starts like the design
of any good application, with an outline of what can be done to make it easy to use,
trustworthy, and simple to understand. Traditionally, user interface design and
visualisation have been low on the academic agenda, to the considerable detriment
of both the science and the engagement of taxpayers. Fortunately, in the years since
the turn of the millennium there has been an increasing realisation that good design
engages the public, and that there is a good deal of social-science research that can
be built on that engagement. Orford et al. (1999) identify computer graphics,
multimedia, the World Wide Web, and virtual reality as four visualisation
technologies that have recently seen a considerable evolution within the social
sciences. There is an ever increasing array of visualisation techniques at our
disposal: Table 9.3 presents a classification scheme of commonly used and more
novel visualisation methods based on the dimensionality and type of data that is
being explored.
Another classification scheme of these techniques that is potentially very useful
comes from Andrienko et al. (2003). This classification categorises techniques
based on their applicability to different types of data.
• “Universal” techniques that can be applied whatever the data, e.g. querying and
animation.
• Techniques revealing existential change, e.g. time labels, colouring by age,
event-lists and space-time cubes.
• Techniques about moving objects, e.g. trajectories, space-time cubes, snapshots
in time.
• Techniques centered on thematic/numeric change, e.g. change maps, time-series
and aggregations of attribute values.
Table 9.3 Classification of visualisation methods according to dimensionality and type of data
Method Pro Con
Spatial 1D/2D Map: overlay; View of whole Cannot analyse trajectory
animated trajectory of movement. If several
trajectory of an object objects cross paths,
representation cannot tell whether
(e.g. arrows); objects met at crossing
snapshots point or visited points at
different times
Spatial distribution: Gives a snapshot Cannot see how a system
e.g. choropleth of an area evolves through time.
maps Aggregate view of area.
Only represents one
variable; hard to
distinguish relationships
Temporal 1D Time-series graphs/ Show how the system No spatial element. Hard to
linear and (or parameters) correlate relationships
cyclical graphs change over time between multivariate
variables
Rank clocks (e.g. Good for visualising No spatial element
Batty 2006) change over time
in ranked order of
any set of objects
Rose diagrams Good for No spatial element
(e.g. Parry 2005) representation of
circular data e.g.
wind speed and
direction
Phase diagram Excellent for No spatial element. Gets
examining system confusing quickly with
behaviour over more than two variables
time for one or
two variables
Spatio-temporal Map animation Can see system Hard to quantify or see
3D/4D (e.g. Patel and evolving spatially impacts of individual
Hudson-Smith and temporally behaviour, i.e. isolated
2012) effects
Space-time cube Can contain space- Potentially difficult to
(Andrienko time paths for interpret
et al. 2003) individuals
Recurrence plot Reveals hidden Computationally intensive.
structures over Methods difficult to
time and in space apply. Have to generate
multiple snapshots and
run as an animation
Vector plotting/ Ability to visualise Hard to quantify
Contour slicing 2D or 3D data individual effects
(Ross and Vosper and multiple
2003) dimensional
data set
184 A. Evans et al.
For information on other visualisation schemes, see Cleveland (1983),

Hinneburg et al. (1999) and Gahegan (2001).
In each case, the techniques aim to exploit the ease with which humans recognise
patterns (Muller and Schumann 2003). Pattern recognition is, at its heart, a human
attribute, and one which we utilise to understand models, no matter how we process
the data. The fact that most model understanding is founded on a human recognition
of a “significant” pattern is somewhat unfortunate, as we will bring our own biases
to the process. At worst we only pay attention to those patterns that confirm our
current prejudices: what Wyszomirski et al. (1999) call the WYWIWYG – What You
Want is What You Get – fallacy. At best, we will only recognise those patterns that
match the wiring of the human visual system and our cultural experiences. The
existence of visualisation techniques generally points up the fact that humans are
better at perceiving some patterns than others, and in some media than others – it is
easier to see an event as a movie and not a binary representation of the movie file
displayed as text. However, in addition to standard physiological and psychological
restrictions on pattern recognition consistent to all people, it is also increasing
apparent there are cultural differences in perceptions. Whether there is some
underlying biological norm for the perception of time and space is still moot
(Nisbett and Masuda 2003; Boroditsky 2001) but it is clear that some elements of
pattern recognition vary by either culture or genetics (Nisbett and Masuda 2003;
Chua et al. 2005). Even when one looks at the representation of patterns and
elements like colour, there are clear arguments for a social influence on the
interpretation of even very basic stimuli into perceptions (Roberson et al. 2004).
Indeed, while there is a clear and early ability of humans to perceive moving objects
in a scene as associated in a pattern (e.g. Baird et al. 2002), there are cultural traits
associated with the age at which even relatively universal patterns are appreciated
(Clement et al. 1970). The more we can objectify the process, therefore, the less our
biases will impinge on our understanding. In many respects it is easier to remove
human agents from data comparison and knowledge development than pattern
hunting, as patterns are not something machines deal with easily. The unsupervised
recognition of even static patterns repeated in different contexts is far from com-
putationally solved (Bouvrie and Sinha 2007). Most pattern-hunting algorithms try
to replicate the process found in humans and in that sense one suspects we would do
better to skip the pattern hunting and concentrate on data consistency and the
comparison of full datasets directly. At best we might say that an automated
“pattern” hunter that wasn’t trying to reproduce the human ability would instead
seek to identify attractors within the data.
Figure 9.1 presents several visualisation methods that are commonly found in
the literature, ranging from 1D time-series representation Fig. 9.1a to contour plots
Fig. 9.1d that could be potentially used for 4D representation.
Visualisations are plainly extremely useful. Here we’ll look at a couple of
techniques that are of use in deciphering individual-level data: phase maps and
recurrence plots. Both techniques focus on the representation of individual level
states and the relationships between stated individuals.
a b
Income (pounds) 25000
400
20000
25000
15000
300 20000
15000 10000
10000 5000
Price
200
5000 0
0
100 450
445
440
435
390 430
0 400 425
410 420
0 5 10 15 420 415 Northing (km)
430 410
Time Easting (km) 440
450 405
N
c 15000 d U and streamlines
30
4.50
10000
4.00
3.50
20 3.00
5000
2.50
z (m)
2.00
W 0 E
15000 10000 5000 0 5000 10000 15000
1.50
10
5000
1.00
0.500
10000 0.00
0 –0.500
–100 0 100
15000 X (m)
S
Fig. 9.1 Examples of different visualisation methods. (a) 1D time-series graph (idealised data).
(b) 3D interpolated map (idealised data). (c) Rose diagram. (d) Contour plot
9.3.1 Phase Maps
Phase space maps are commonly used by physicists to study the behaviour of
physical systems. In any graphical representation, a phase space map represents
an abstract view of the behaviour of one or more of the system components. These
can be particularly useful to us as we can plot the behaviour of our system over
time. This allows us to understand how the system is evolving and whether it is
chaotic, random, cyclical or stable (Fig. 9.2).
Each of the graphs produced in Fig. 9.2 are a representation of the coincident
developments in two real neighbouring city centre petrol stations in Leeds (UK)
over a 30 day period (sampled every other day). Figure 9.2a represents a stable
system. Here, neither of the stations is changing in price and thus, a fixed point is
produced. However, this behaviour could easily change if one or both of the stations
alters it price. This behaviour is seen in Fig. 9.2b. Both stations are changing their
prices each day (from 75.1p to 75.2p to 75.1p), this creates a looping effect; the
stations are cycling through a pattern of behaviour before returning to their starting
186 A. Evans et al.
a 75.5 b 75.22
75.45
Price at Station 2
Price at Station 2
75.2
75.4
75.35 75.18
75.3
75.16
75.25
75.2 75.14
75.15 75.12
75.1
75.1
75.05
75 75.08
75 75.05 75.1 75.15 75.2 75.25 75.3 75.08 75.1 75.12 75.14 75.16 75.18 75.2 75.22
Price at Station 1 Price at Station 1
c 90 d
88 85
Price at Station 2
Price (p) Station 2
86 84
84
83
82
80 82
78 81
76
80
74
72 79
70 78
69 70 71 72 73 74 75 76 77 78 73 74 75 76 77 78 79
Price (p) Station 1 Price (p) Station 1
Fig. 9.2 Examples of different types of behaviour found in urban petrol stations (Leeds).
(a) Stable (b) Looping (c) Two types of behaviour (d) Chaotic
point. Note that the graph appears to reveal a causative link between the two
stations as they are never simultaneously low. Figure 9.2c and d show a more
varied pattern of behaviour between the stations. In Fig. 9.2c, one point is rising in
price whilst the other is oscillating. In Fig. 9.2d, there is no apparent pattern in the
displayed behaviour. Simply knowing about these relationships is valuable infor-
mation and allows us a greater understanding of this system, its behaviour and its
structure. For example, it may be that the only difference between the graphs is one
of distance between the stations, but we would never see this unless the graphs
allowed us to compare at a detailed level the behaviours of stations that potentially
influence each other.
9.3.2 Recurrence Plots
Recurrence plots (RPs) are a relatively new technique for the analysis of time series
data that allows both visualisation and quantification of structures hidden within
data or exploration of the trajectory of a dynamical system in phase space
(Eckmann et al. 1987). They are particularly useful for graphically detecting hidden
patterns and structural changes in data as well as examining similarities in patterns
across a time series data set (where there are multiple readings at one point). RPs
40 60 80 100 80 85 90 95 100
a b
1.00 1.00
100 0.900 100 0.900
0.800 0.800
0.700 0.700
95
0.600
Day 2
Day 2
80 0.600
0.500 0.500
90 0.400
0.400
0.300 0.300
60
0.200 85 0.200
0.100 0.100
0.00 80 0.00
40
40 60 80 100 80 85 90 95 100
Day 1 Day 1
Fig. 9.3 Example of recurrence plots
can be also used to study the nonstationarity of a time series as well as to indicate its
degree of aperiodicity (Casadagli 1997; Kantz and Schreiber 1997). These features
make RPs a very valuable technique for characterising complex dynamics in the
time domain (Vasconcelos et al. 2006), a factor reflected in the variety of
applications that RPs can now found in, ranging from climate variation (Marwan
and Kruths 2002), and music (Foote and Cooper 2001) to heart rate variability
(Marwan et al. 2002).
Essentially a RP is constructed via a matrix where values at a pair of time-steps
are compared against each other. If the system at the two snapshots is completely
different, the result is 1.0 (black), while completely similar periods are attributed
the value 0.0 (represented as white). Through this, a picture of the structure of the
data is built up. Figure 9.3a shows the RP of the change in price at a retail outlet
over 100 days. Above the RP is a time-series graph diagrammatically representing
the change in price. Changes in price, either increases, decreases, or oscillations,
can be clearly seen in the RP. Figure 9.3b illustrates how oscillations in the change
in the price data are represented in the RP.
Early work on this area has shown that there is considerable potential in the
development and adaptation of this technique. Current research is focused on the
development of cross-reference RPs (consideration of the phase space trajectories
of two different systems in the same phase space) and spatial recurrence plots.
9.4 Explanation, Understanding and Causality
Once patterns are recognised, “understanding” our models involves finding

explanations highlighting the mechanisms within the models which give rise to
these patterns. The process of explanation may be driven with reference to current
theory or developing new theory. This is usually achieved through:
188 A. Evans et al.
1. Correlating patterns visually or statistically with other parts of the model, such as
different geographical locations, or with simulations with different starting
values;
2. Experimentally adjusting the model inputs to see what happens to the outputs;
3. Tracking the causal processes through the model.
It may seem obvious, and yet it is worth pointing out that model outputs can only
causally relate to model inputs, not additional data in the real world. Plainly insights
into the system can come from comparison with external data that is correlated or
mis-correlated with model outputs, but this is not the same as understanding your
model and the way it currently represents the system. One would imagine that this
means that understanding of a model cannot be facilitated by comparing it with
other, external, data, and yet it can often be worth:
4. Comparing model results with real world data,
because the relationships between real data and both model inputs and model
outputs may be clearer than the relationships between these two things within the
model.
Let’s imagine, for example, a model that predicts the location of burglaries
across a day in a city region where police corruption is rife. The model inputs are
known offenders’ homes, potential target locations and attractiveness, the position
of the owners of these targets and the police, who prefer to serve the wealthy. We
may be able to recognise a pattern of burglaries that moves, over the course of the
day, from the suburbs to the city centre. Although we have built into our model the
fact that police respond faster to richer people, we may find, using (1) that our
model doesn’t show less burglaries in rich areas, because the rich areas are so
spatially distributed that the police response times are stretched between them. We
can then alter the weighting of the bias away from the wealthy (2) to see if it
actually reduces the burglary rate in the rich areas by placing police nearer these
neighbourhoods as an ancillary effect of responding to poor people more. We may
be able to fully understand this aspect of the model and how it arises (3), but still
have a higher than expected burglary rate in wealthy areas. Finally, it may turn out
(4) that there is a strong relationships between these burglaries and real data on
petrol sales, for no other reason than both are high at transition times in this social
system, when the police would be most stretched between regions – suggesting in
turn that the change in police locations over time is as important as their positions at
any one time.
Let us look at each of these methodologies for developing understanding in turn.
Correlation: Most social scientists will be familiar with linear regression as a
means for describing data or testing for a relationship between two variables; there
is a long scientific tradition of correlating data between models and external
variables, and this tradition is equally applicable to intra-model comparisons.
Correlating datasets is one of the areas where automation can be applied. As an
exploratory tool, regression modelling has its attractions, not least its simplicity in
both concept and execution. Simple regressions can be achieved in desktop
applications like Microsoft Excel, as well as all the major statistical packages
(SAS, SPSS, BMDP, GLIM etc.). Standard methodologies are well known for
cross-correlation of both continuous normal data and time series. However even
for simple analyses with a single input and single output variable, linear regression
is not always an appropriate technique. For example, logistic regression models will
be more appropriate for binary response data, Poisson models will be superior when
values in the dependent tend to be highly clustered, while binomial models may be
the most effective when observations are highly dispersed around the mean. An
interesting example is Fleming and Sorenson (2002) in which binomial estimates of
technological innovation are compared to the complexity of the invention measured
by both the number of components and the interdependence between those
components. In behavioural space, methodologies such as Association Rule
Making (e.g. Hipp et al. 2002) allow the Bayesian association of behavioural
attributes. It is worth noting that where models involve a distribution in physical
space this can introduce problems, in particular where the model includes
neighbourhood-based behaviours and therefore the potential to develop spatial
auto and cross-correlations. These alter the sampling strategies necessary to prove
relationships – a full review of the issues and methodologies to deal with them can
be found in (Wagner and Fortin 2005).
Experimentation: In terms of experimentation, we can make the rather artificial
distinction between sensitivity testing and “what if?” analyses – the distinction is
more one of intent than anything. In sensitivity analysis one perturbs model inputs
slightly to determine the stability of the outputs, under the presumption that models
should match the real world in being insensitive to minor changes (a presumption not
always well founded). In “what if?” analyses, one alters the model inputs to see what
would happen under different scenarios. In addition to looking at the output values at
a particular time-slice, the stability or otherwise of the model, and the conditions
under which this varies, also give information about the system (Grimm 1999).
Tracking Causality: Since individual-based models are a relatively recent
development, there is far less literature dealing with the tracking of causality
through models. It helps a little that the causality we deal with in models, which
is essentially a mechanistic one, is far more concrete than the causality perceived by
humans, which is largely a matter of the repeated co-incidence of events. Never-
theless, backtracking through a model to mark a causality path is extremely hard,
primarily for two reasons. The first is what we might call the “find the lady
problem” – that the sheer number of interactions involved in social processes
tends to be so large we don’t have the facilities to do the tracking. The second
issue, which we might call the “drop in the ocean problem”, is more fundamental as
it relates to a flaw in the mathematical representation of objects, that is, that
numbers represent aggregated quantities, not individuals. When transacted objects
in a system are represented with numbers greater than one, it is instantly impossible
to reliably determine the path taken by a specific object through that system. For
objects representing concepts, either numerical (for example, money) or non-
numerical (for example, a meme), this isn’t a problem (one dollar is much like
any other; there is only one YouTube to know). However, for most objects such
aggregations place ambiguous nodes between what would otherwise be discrete
190 A. Evans et al.
causal pathways. Fortunately, we tend to use numbers in agent-models as a meth-

odology to cope with our ignorance (for example, in the case of calibrated
parameters), or the lack of the computing power we’d need to deal with individual
objects and their transactional histories (for example, in the case of a variable like
“number of children”). As it happens, every day brings improvements to both. It
addition, the last 10 years or so has seen considerable theoretical advances in the
determination of the probabilities of causation (for example, Pearl and Verma 1991;
Greenland and Pearl 2006). For now, however, the tracking of causality is much
easier if the models build in appropriate structures from the start. While they are in
their infancy, techniques like Process Calculi (Worboys 2005) and Petri Nets show
the potential of this area.
The inability to track causality leads to the perennial problem of identifiability,
that is, that a single model outcome may have more than one history of model
parameters that leads to it. Identifiability is part of a larger set of issues with
confirming that the model in the computer accurately reflects the system in the real
world – the so-called equifinality issue. These are issues that play out strongly during
model construction from real data and when validating a model against real data, and
a review of techniques to examine these problems, including using model variation to
determine the suitability of variables and parameters, can be found in Evans (2012).
At the model stage we are interested in, however, we at least have the advantage that
there is only one potential model that may have created the output – the one running.
Nevertheless, the identifiability of the parameters in a running model still makes it
hard to definitively say when model behaviour is reasonable. For those modelling for
prediction, this is of little consequence – as long as the model gives consistently good
predictions it may as well be a black-box. However, if we wish to tease the model
apart and look at how results have emerged, these issues become more problematic.
The mechanisms for dealing with these problems are pragmatic.
1. Examine the stability of the calibration process and/or the state of internal
variables that weren’t inputs or outputs across multiple runs.
2. Validate internal variables that weren’t inputs or outputs used in any calibration
against real data.
3. Run the model in a predictive mode with as many different datasets as possible –
the more the system can replicate reality at output, the more likely it is to
replicate reality internally. If necessary engage in inverse modelling: initialize
parameters randomly then adjust them over multiple runs until they match all
known outputs.
Of these, by far the easiest, but the least engaged with, is checking the stability of
the model in parameter space (see Evans 2012 for a review). Various AI techniques
have been applied to the problem of optimizing parameters to fit model output
distributions to some pre-determined pattern (such as a ‘real world’ distribution).
However, the stability of these parameterizations and the paths AIs take to generate
them are rarely used to examine the degree to which the model fluctuates between
different states, let alone to reflect on the nature of the system. The assumption of
equifinality is that the more parameterized a model, the more likely it is a set of
parameter values can be derived which fit the data but don’t represent the true
values. However, in practice the limits on the range of parameter values within any
given model allows us an alternative viewpoint: that the more parameterized rules
in a model, the more the system is constrained by the potential range of the elements
in its structure and the interaction of these ranges. For example, a simple model
a ¼ b has no constraints, but a ¼ b/c, where c ¼ distance between a and b, adds an
additional constraint even though there are more parameters. As such rules build up
in complex systems, it is possible that parameter values become highly constrained,
even though, taken individually, any given element of the model seems reasonably
free. This may mean that if a system is well modelled, exploration of the model’s
parameter space by an AI might reveal the limits of parameters within the
constraints of the real complex system. For example, Heppenstall et al. (2007)
use a Genetic Algorithm to explore the parameterisation of a petrol retail model/
market, and find that while some GA-derived parameters have a wide range, others
consistently fall around specific values that match those derived from expert
knowledge of the real system.
The same issues as hold for causality hold for data uncertainty and error. We
have little in the way of techniques for coping with the propagation of either
through models (see Evans 2012 for a review). It is plain that most real systems
can be perturbed slightly and maintain the same outcomes, and this gives us some
hope that errors at least can be suppressed, however we still remain very ignorant as
to how such homeostatic forces work in real systems, and how we might recognise
or replicate them in our models. Data and model errors can breed patterns in our
model outputs. An important component of understanding a model is understanding
when this is the case. If we are to use a model to understand the dynamics of a real
system and its emergent properties, then we need to be able to recognise novelty in
the system. Patterns that result from errors may appear to be novel (if we are lucky),
but as yet there is little in the way of toolkits to separate out such patterns from truly
interesting and new patterns produced intrinsically.
Currently our best option for understanding model artifacts is model-to-model
comparisons. These can be achieved by varying one of the following contexts while
holding the others the same: the model code (the model, libraries, and platform), the
computer the model runs on, or the data it runs with (including internal random
number sequences). Varying the model code (for instance from Java to C++, or
from an object-orientated architecture to a procedural one) is a useful step in that it
ensures the underlying theory is not erroneously dependent on its representation.
Varying the computer indicates the level of errors associated with issues like
rounding and number storage mechanisms, while varying the data shows the degree
to which model and theory are robust to changes in the input conditions. In each
case, a version of the model that can be transferred between users, translated onto
other platforms and run on different data warehouses would be useful. Unfortu-
nately, however, there is no universally recognised mechanism for representing
models abstracted from programming languages. Mathematics, UML, and natural
languages can obviously fill this gap to a degree, but not in a manner that allows for
complete automatic translation. Even the automatic translation of computer
192 A. Evans et al.
languages is far from satisfactory when there is a requirement that the results be
understood by humans so errors in knowledge representation can be checked. In
addition, many such translations work by producing the same binary executable.
We also need standard ways of comparing the results of models, and these are no
more forthcoming. Practitioners are only really at the stage where we can start to
talk about model results in the same way (see, for example, Grimm et al. 2006).
Consistency in comparison is still a long way off, in part because statistics for
model outputs and validity are still evolving, and in part because we still don’t have
much idea which statistics are best applied and when (for one example bucking this
trend see Knudsen and Fotheringham 1986).
9.5 Future Directions
Recognising patterns in our modelled data allows us to:

1. Compare it with reality for validation.
2. Discover new information about the emergent properties of the system.
3. Make predictions.
Of these, discovering new information about the system is undoubtedly the
hardest, as it is much easier to spot patterns you are expecting. Despite the above
advances, there are key areas where current techniques do not match our
requirements. In particular, these include:
1. Mechanisms to determine when we do not have all the variables we need to
model a system and which variables to use.
2. Mechanisms to determine which minor variables may be important in making
emergent patterns through non-linearities.
3. The tracking of emergent properties through models.
4. The ability to recognise all but the most basic patterns in space over time.
5. The ability to recognise action across distant spaces over space and time.
6. The tracking of errors, error acceleration, and homeostatic forces in models.
While we have components of some of these areas, what we have is but a drop in
the ocean of techniques we need. In addition, the vast majority of our techniques are
built on the 2,500 years of mathematics that resolved to simplify systems that were
collections of individuals because we lacked the ability (either processing power or
memory) to cope with the individuals as individuals. Modern computers have given
us this power for the first time and, as of yet, the ways we describe such systems
have not caught up, even if we accept that some reduction in dimensionality and
detail is necessary for a human to understand our models. Indeed in the long-run it
might be questioned whether the whole process of model understanding and
interpretation might be divorced from humans, and delegated instead to an artifi-
cially intelligent computational agency that can better cope with the complexities
directly.
Further Reading
Statistical techniques for spatial data are reviewed by McGarigal (2002) while for
network statistics good starting points are (Newman 2003) and (Boccaletti et al.
2006), with more recent work reviewed by Evans (2010). For information on
coping with auto/cross-correlation in spatial data, see (Wagner and Fortin 2005).
Patel and Hudson-Smith (2012) provide an overview of the types of simulation tool
(virtual worlds and virtual reality) available for visualising the outputs of spatially
explicit agent-based models. Evans (2012) provides a review of techniques for
analysing error and uncertainty in models, including both environmental/climate
models and what they can bring to the agent-based field. He also reviews techniques
for identifying the appropriate model form and parameter sets.
References
Andrienko N, Andrienko G, Gatalsky P (2003) Exploratory spatio-temporal visualisation: an

analytical review. J Vis Lang Comput 14(6):503–541
Baird AA et al (2002) Frontal lobe activation during object permanence: data from near-infrared
spectroscopy. Neuroimage 16:1120–1126
Batty, M (2006) Rank Clocks. Nature, 444: 592–596
Boccaletti S, Latora V, Moreno Y, Chavez M, Hwang D-U (2006) Complex networks: structure
and dynamics. Phys Rep 424(4–5):175–308
Boroditsky L (2001) Does language shape thought? Mandarin and English Speakers’ conceptions
of time. Cogn Psychol 43:1–22
Bouvrie JV, Sinha P (2007) Visual object concept discovery: observations in congenitally blind
children, and a computational approach. Neurocomputing 70(13–15):2218–2233
Casdagli M (1997) Recurrence plots revisited. Phys D 108(1–2):12–44
Chua HF, Boland JE, Nisbett RE (2005) Cultural variation in eye movements during scene
perception. Proc Natl Acad Sci USA 102(35):12629–12633
Clark PJ, Evans FC (1954) Distance to nearest neighbor as a measure of spatial relationships in
populations. Ecology 35(4):445–453
Clement DE, Sistrunk F, Guenther ZC (1970) Pattern perception among Brazilians as a function of
pattern uncertainty and age. J Cross Cult Psychol 1(4):305–313
Cleveland WS (1983) Visualising data. Hobart Press, Summit
Eckmann JP, Kamphorst SO, Reulle D (1987) Recurrence plots of dynamical systems.
Europhys Lett 4(9):973–977
Evans AJ (2010) Complex spatial networks in application. Complexity 16(2):11–19
Evans AJ (2012) Uncertainty and error. In: Heppenstall AJ, Crooks AT, See LM, Batty M (eds)
Agent-based models of geographical systems. Springer, Berlin, chapter 15
Fisher N, Lewis T, Embleton B (1987) Statistical analysis of spherical data. Cambridge University
Press, Cambridge
Fleming L, Sorenson O (2001) Technology as a complex adaptive system: evidence from patent
data. Research Policy 30: 1019–1039
Foote J, Cooper M (2001) Visualising music structure and rhythm via self-similarity.
In: Proceedings of the international computer music conference, ICMC’01, Havana. ICMA,
San Francisco, pp 419–422
194 A. Evans et al.
Fotheringham AS, Brunsdon C, Charlton M (2002) Geographically weighted regression:

the analysis of spatially varying relationships. Wiley, Chichester
Gahegan M (2001) Visual exploration in geography: analysis with light. In: Miller HJ, Han J (eds)
Geographic data mining and knowledge discovery. Taylor & Francis, London, pp 260–287
Gehlke CE, Biehl H (1934) Certain effects of grouping upon the size of correlation coefficients in
census tract material. J Am Stat Assoc 29(2):169–170
Getis A (2007) Reflections on spatial autocorrelation. Reg Sci Urban Econ 37(4):491–496
Graps A (2004) Amara’s wavelet page. Access date 8 Jun 2011. http://web.archive.org/web/
20110608005544/http://www.amara.com/current/wavelet.html
Greenland S, Pearl J (2006) Causal diagrams (Technical report, R-332). UCLA Cognitive Systems
Laboratory, Los Angeles. Access date 8 June 2011 http://ftp.cs.ucla.edu/pub/stat_ser/r332.pdf
Grimm V (1999) Ten years of individual-based modelling in ecology: what have we learned and
what could we learn in the future? Ecol Model 115(2):129–148
Grimm V (2002) Visual debugging: a way of analyzing, understanding, and communicating
bottom-up simulation models in ecology. Nat Resour Model 15:23–38
models. Ecol Model 198(1–2):115–126
Haining R (1990) Spatial data analysis in the social and environmental sciences. Cambridge University
Press, Cambridge
Heppenstall AJ, Evans AJ, Birkin MH (2006) Using hybrid agent-based systems to model
spatially-influenced retail markets. J Artif Soc Soc Simul 9(3). http://jasss.soc.surrey.ac.uk/9/
3/2.html
Heppenstall AJ, Evans AJ, Birkin MH (2007) Genetic algorithm optimisation of a multi-agent
system for simulating a retail market. Environ Plann B 34(6):1051–1070
Hinneburg A, Keim DA, Wawryniuk M (1999) HD-eye: visual mining of high-dimensional data.
IEEE Comput Graph Appl 19(5):22–31
Hipp J, Güntzer U, Nakhaeizadeh G (2002) Data mining of association rules and the process of
knowledge discovery in databases. In: Perner P (ed) Advances in data mining, vol 2394,
Lecture notes in computer science. Springer, Berlin, pp 207–226
Isaaks EH, Srivastava RM (1990) Applied geostatistics. Oxford University Press, Cary
Kantz H, Schreiber T (1997) Non-linear time series analysis. Cambridge University Press, Cambridge
Knudsen DC, Fotheringham AS (1986) Matrix comparison, goodness-of-fit, and spatial interaction
modelling. Int Reg Sci Rev 10:127–147
Korie S et al (1998) Analysing maps of dispersal around a single focus. Environ Ecol Stat 5(4):
317–344
Marwan N, Kruths J (2002) Nonlinear analysis of bivariate data with cross recurrence plots.
Phys Lett A 302(5–6):299–307
Marwan N, Wessel N, Meyerfeldt U, Schirdewan A, Kurths J (2002) Recurrence-plot-based measures
of complexity and their application to heart-rate-variability data. Phys Rev E 66(2):026702
McGarigal K (2002) Landscape pattern metrics. In: El-Shaarawi AH, Piegorsch WW (eds)
Encyclopedia of environmentrics, vol 2. Wiley, Chichester, pp 1135–1142
Moon I-C, Schneider M, Carley K (2006) Evolution of player skill in the America’s army game.
Simulation 82(11):703–718
Müller W, Schumann HS (2003) Visualisation methods for time-dependent data: an overview.
In: Chick S, Sánchez PJ, Ferrin D, Morrice DJ (eds) Proceedings of winter simulation 2003,
New Orleans, 7–10 Dec 2003. http://informs-sim.org/wsc03papers/090.pdf
Newman MEJ (2003) The structure and function of complex networks. SIAM Rev 45:167–256
Nisbett RE, Masuda T (2003) Culture and point of view. Proc Natl Acad Sci USA 100(19):
11163–11170
OED (2010) Oxford English dictionary. Access date 21 Dec 2010 http://www.oed.com/
Openshaw S, Craft AW, Charlton M, Birch JM (1988) Investigation of leukaemia clusters by use
of a geographical analysis machine. Lancet 331(8580):272–273
Orford S, Harris R, Dorling D (1999) Geography: information visualisation in the social sciences.
Soc Sci Comput Rev 17(3):289–304
Parry H (2005) Effects of Land Management upon Species Population Dynamics: A Spatially
Explicit, Individual-based Model. Unpublished PhD thesis, University of Leeds. pp. 281
Patel A, Hudson-Smith A (2012) Agent tools, techniques and methods for macro and microscopic
simulation. In: Heppenstall AJ, Crooks AT, See LM, Batty M (eds) Agent-based models of
geographical systems. Springer, Berlin, chapter 18
Pearl J, Verma TS (1991) A theory of inferred causation. In: Allen JA, Fikes R, Sandewall E (eds)
Proceedings of the 2nd international conference on principles of knowledge representation and
reasoning (KR’91), Cambridge, MA, 22–25 Apr 1991. Morgan Kaufmann, San Mateo,
pp 441–452
Pohlheim H (2006) Multidimensional scaling for evolutionary algorithms: visualisation of the path
through search space and solution space using Sammon mapping. Artif Life 12(2):203–209
Ramsey JB (2002) Wavelets in economics and finance: past and future. Stud Nonlinear Dynam
Econ 6(3):1–27
Roberson D, Davidoff J, Davies IRL, Shapiro LR (2004) The development of color categories in
two languages: a longitudinal study. J Exp Psychol Gen 133(4):554–571
Robinson WS (1950) Ecological correlations and the behaviour of individuals. Am Sociol Rev 15:
351–357
Ross AN, Vosper SB (2003) Numerical simulations of stably stratified flow through a mountain
pass. Q J R Meteor Soc 129:97–115
Schelling TC (1969) Models of segregation. Am Econ Rev 59(2):88–493
Schreinemachers P, Berger T (2006) Land-use decisions in developing countries and their
representation in multi-agent systems. J Land Use Sci 1(1):29–44
Tobler WR (1970) A computer model simulation of urban growth in the Detroit region.
Econ Geogr 46(2):234–240
Vasconcelos DB, Lopes SR, Kurths J, Viana RL (2006) Spatial recurrence plots. Phys Rev E 73(5):
056207
Viboud C et al (2006) Synchrony, waves, and spatial hierarchies in the spread of influenza.
Science 312(5772):447–451
Wagner HH, Fortin M-J (2005) Spatial analysis of landscapes: concepts and statistics.
Ecology 86:1975–1987
Worboys MF (2005) Event-oriented approaches to geographic phenomena. Int J Geogr Info Sci
19(1):1–28
Wyszomirski T, Wyszomirska I, Jarzyna I (1999) Simple mechanisms of size distribution dynamics
in crowded and uncrowded virtual monocultures. Ecol Model 115(2–3):253–273
Chapter 10
Participatory Approaches
Olivier Barreteau, Pieter Bots, Katherine Daniell, Michel Etienne,

Pascal Perez, Cécile Barnaud, Didier Bazile, Nicolas Becu,
Jean-Christophe Castella, William’s Daré, and Guy Trebuil
O. Barreteau (*)
IRSTEA, UMR G-EAU, 361 rue Jean-François Breton, BP 5095, Montpellier 34196, France
P. Bots
Delft University of Technology, Faculty of Technology, Policy and Management,
Delft 50152600 GA, The Netherlands
K. Daniell
Centre for Policy Innovation, The Australian National University, Canberra ACT 0200, Australia
M. Etienne
Ecodevelopment UnitDomaine St-Paul, site Agroparc, Avignon cedex 9 84914, France
P. Perez
SMART, University of Wollongong, Northfields Ave, Wollongong NSW 2522, Australia
C. Barnaud
INRA, Centre INRA de Toulouse UMR Dynafor Chemin de Borde Rouge,
Castanet Tolosan Cedex BP 5262731326, France
D. Bazile • W. Daré • G. Trebuil
Cirad GREEN, TA C-47/F. Campus international de Baillarguet, Montpellier Cedex 5 34398,
France
N. Becu
CNRS, Laboratoire de géographie PRODIG 2, rue Valette, Paris 75005, France
J.-C. Castella
Institut de Recherche pour le Développement (IRD), UMR 220 GRED (IRD UPV Montpellier 3),
PO Box 5992, Vientiane, Laos

198 O. Barreteau et al.
Why Read This Chapter? To help you understand how one might involve
stakeholders in all stages of the modelling process. This approach allows for
including stakeholders’ expertise as well as giving them more control over the
process.
Abstract This chapter aims to describe the diversity of participatory approaches in
relation to social simulations, with a focus on the interactions between the tools and
participants. We consider potential interactions at all stages of the modelling
process: conceptual design; implementation; use; and simulation outcome analysis.
After reviewing and classifying existing approaches and techniques, we describe
two case studies with a focus on the integration of various techniques. The first case
study deals with fire hazard prevention in southern France, and the second one with
groundwater management on the Atoll of Kiribati. The chapter concludes with a
discussion of the advantages and limitations of participatory approaches.
10.1 Introduction
In this chapter, social simulation is cross-examined with a currently very active

trend in policy making: participation or stakeholder involvement. This cross-
examination has two main outputs: the development of tools and methods to
improve or facilitate participation; and the development of more grounded simula-
tion models through participatory modelling. Technological development provides
new devices to facilitate interaction around simulation models: from the phase of
conceptual design to that of practical use. In many fields there is a growing
requirement from stakeholders and the public to become more actively involved
in policy making and to be aware of probable changing trends due to global policy
decisions. New tools and methods related to social simulation have started to be
made available for this purpose such as many Group Decision Support Systems
which use computer simulation, including potentially social items components, to
facilitate communication to formulate and solve problems collectively (DeSanctis
and Gallupe 1987; Shakun 1996; Whitworth et al. 2000). In addition, simulation of
social complexity occurs in models whose validation and suitability depend on their
close fit to society, as well as on their acceptability by it. These issues are tackled
through the use of participatory modelling, such as group model building (Vennix
1996) or participatory agent based simulations (Bousquet et al. 1999; Guyot and
Honiden 2006; Moss et al. 2000; Pahl-Wostl and Hare 2004; Ramanath and Gilbert
2004). The topic is also related to participatory design as it is a mean of involving
end-users of computer systems in their design, including social simulations
focussed ones (Schuler and Namioka 1993).
Group Decision Support as well as Participatory Modelling stem from the
interactions between simulation models and participants. There is a diversity of
ways though which these interactions might take place. They are related to the
diversity of approaches to simulate society or to organise participation. It is important
to make the choices made for these interactions explicit: for distinction between
approaches to be possible; to provide the opportunity for stakeholders to discuss the
10 Participatory Approaches 199
process; and for them to be prepared to be involved in. There is a need to go further
than the development of tools as they are liable to create filters that reshape the
understanding of social complexity. Description of the mechanisms behind
interactions is a way to qualify the potential effects of these interactions.
This chapter aims to describe the diversity of participatory approaches in
relation to social simulations, with a focus on the interactions between the tools
and participants. This overview is limited to simulation models. Model is consid-
ered here as a representation of shared knowledge, which means the gathering of
pieces of knowledge and assumptions about a system, written altogether in a model
so that they might play or work together. We limit this scope further to simulation
model, hence models including the representation of dynamics. We consider here
potential interactions among participatory and modelling processes at all stages of
the modelling process: conceptual design; implementation; use; and simulation
outcome analysis.
The first section of this chapter outlines a number of factors which have paved
the way for development of the association between social simulation and partici-
pation. There is a large body of literature in which authors have developed their
own participatory modelling approaches, justified by some specific expectations on
participation for modelling or vice-versa. This first section makes a synthesis of
these expectations and draws out some principles on which various participatory
modelling settings should be assessed. The second section describes some existing
techniques and approaches. The third section proposes a classification of these
participatory approaches according to three dimensions: the level of involvement in
the process; the timeliness of involvement; and the heterogeneity of population
involved. The fourth section describes two case studies with a focus on the
integration of various techniques. We discuss the advantages of these approaches
but also some limits, according to the expectations and in comparison with more
traditional techniques in the fifth section.
10.2 Expectations of Using Participatory Approaches

with Simulation of Social Complexity
Joint use of participatory approaches with social simulations is based upon three
categories of expectations. They vary according to the target of the expected
benefits of the association:
1. Quality of the simulation model per se;
2. Suitability of the simulation model for a given use; and
3. Participation support.
These three targets are linked to three different components of a modelling
process. Target one is linked to the output, target three to the source system, and
target two to the relation between both the output and source system. In this section
we further develop these three categories.
10.2.1 Increasing Quality of Simulation Models of Social

Complexity
The objective here is to produce a good quality model to simulate social complexity.
Participation is then pragmatically assumed to be a means for improving this
quality. There is no normative belief which would value participation by itself in
this category of expectations.
Quality of the simulation model is understood here rather classically with the
following indicators:
– Realism: is the simulation model able to tackle key features of the social
complexity it aims to represent?
– Efficiency: is the simulation model representing its target system with a mini-
mum of assumptions and minimal simulation run-times?
Quality of the representation according to its use is another classical indicator of
a simulation model’s quality. It is specifically tackled in the following subsection.
10.2.1.1 Taking Social Diversity and Capacity to Evolve into Account
One of the key features to be taken into account when representing a social system is
to deal with its diversity. This diversity is related not only to individual
characteristics, but also to viewpoints, expectations towards the system, and positions
in the decision making processes. Dealing with diversity in simulation of social
complexity involves embracing it as well as profiting by its existence.
Classically, dealing with diversity is a process of aggregation or selection.
Aggregation consists of the identification of classes of individuals and
representatives for them. Selection consists of choosing a few cases with all of
their characteristics. This may lead to very simple simulation models with a generic
diversity. Aggregation is rather greedy on data and modelling time and is still
dependent on the viewpoint of the observers who provide the information leading to
the categorisation. Selection is weak to cope with relations among various sources
of diversity.
Involvement of stakeholders in the modelling process allows them to bring their
own diversity. Concerns over representation are then transferred onto the constitu-
tion of the sample of participants. Fischer and colleagues have shown through
development of situations to support creativity in various fields, such as art, open
source development and urban planning, that diversity, as well as complexity, is
important to enhance creativity (Fischer et al. 2005). This creativity is expected to
pave the way for surprises in the simulation model.
Involvement of stakeholders in the modelling process is a way to externalise part
of this diversity outside the model towards a group of stakeholders. The issue is then
to work on the relation between the model and a number of stakeholders to allow a
transfer of knowledge and ideas.
Social systems are open and evolving. Their definition depends on the viewpoint
of the analyst. As far as simulation is concerned, this means depending on the
viewpoint of the model designer(s). This choice means framing: cutting a number
of links around the boundaries of the system studied, as well as around the
interpretation which might occur based on the simulation outcomes (Dewulf et al.
2006). Firstly, participation provides the opportunity to consider problem
boundaries which would be plurally defined, increasing the potential coherence of
the model. However, it is still an operation of cutting links out of the real world
situation, even though these chosen cuttings are more grounded and discussed.
Secondly, interactive use of a simulation model is a means to keep some of these
links open and active, with participants as driving belts. Stakeholders are embedded
in social networks which cross the boundaries into the physical and environmental
networks. They make the links come alive, which allows them to function and be
updated.
There is thus a need to question the boundaries set in the interactive setting:
actors in the neighbourhood; concerns of actors connected to those tackled by the
(simulation) model; and how these relations are to be mobilised in the interaction.
10.2.1.2 Distribution of Control
A key characteristic of social systems which is to be addressed through social

simulation is their complexity. This complexity leads to various consequences, such
as the emergence of phenomena, delay effects or discontinuities in some trends,
which are present in social systems as in any complex systems. These are usually
the effects which one likes to discover or better understand when experimenting
with social simulations. From the internal point of view of simulations, Schelling
has shown experimentally that reproducing settings with multiple decision centres
improves the quality of representation of complexity (Schelling 1961). He could
generate complexity through experimental games because of the presence of
independent decision centres, the players. This result has also been shown with
simulations used for forecasting (Green 2002). Green compared the capacity of
forecasting the outcome of past social conflicts with: a role playing game with
students; game theorists; and a group of experts. He compared the simulated
outcomes with those from the real negotiations and found that the role playing
game setting produced the best results. This was the one with the main distribution
of decisions among autonomous centres.
The purpose of associating participatory processes and social simulation here is
then to increase the complexity through interactive use or implementation of a
social model. Unless computational agents are effectively used, which is rare
(Drogoul et al. 2003), formal theories of complex systems that are completely
embedded in a simulation model do not simulate complex patterns but implement
an explanation of a complex pattern. In other words, they should be implemented in
a distributed setting with autonomous entities. Participatory approaches provide
such settings. There is then an issue of a deep connection between a simulation

model and participants in a participatory modelling setting.
10.2.2 Improving Suitability of Simulation Model’s Use
Quality of a model is also assessed according to its suitability for its intended use. In
this subsection, two cases of use are considered: knowledge increase; and policy
making. In both cases, it is expected that involvement of stakeholders at any stage
of a modelling process will aid better tuning of the model with its intended use:
either through interactions with people represented in the model, or with potential
users. Both cases have a major concern with making viewpoints explicit.
10.2.2.1 Case of Increasing Knowledge
The case of use for knowledge increase builds upon the previous subsection. The
key element treated here deals with the uncertainty of social systems. The involve-
ment of stakeholders represented in the simulation model is a way to improve its
validation or calibration. Participants may bring their knowledge to reduce or better
qualify some uncertainties. The simulation model is then expected to give back to
the participants simulation outputs based on the interactions between their pieces of
knowledge. On the other hand, this feedback is sometimes difficult to validate
(Manson 2002). Its presentation and discussion with stakeholders represented in the
simulation model is a way to cope with this issue. This approach has been explored
by Barreteau and colleagues to improve the validation of an Agent Based Model of
irrigated systems in Senegal River valley (Barreteau and Bousquet 1999). The
format of this feedback, information provided and medium of communication,
might make the model really open to discussion.
This joins another expectation which is probably the most common in work that
has so far implemented such participatory approaches with a social simulation
model: making each participant’s assumptions explicit, included the modellers
(Fischer et al. 2005; Moss et al. 2000; Pahl-Wostl and Hare 2004). This is a
requirement from the simulation modelling community: making stakeholders’
beliefs, points of view and tacit knowledge explicit (Barreteau et al. 2001; Cockes
and Ive 1996; D’Aquino et al. 2003; McKinnon 2005). Moreover, so that
participants might become part of the model, the assumptions behind the model
should be made explicit in order to be discussed, as should the outputs of the
simulations so that they can also be discussed, transferred and translated in new
knowledge. This is to overcome one major pitfall identified with the development of
models which is the under-use of decision support models because of their opacity
(Loucks et al. 1985; Reitsma et al. 1996). This concern of making explicit
assumptions in the modelling process is also at the heart of the participatory
approach community. One aim of gathering people together and making them
collectively discuss their situation in a participatory setting is to make them aware

of others’ viewpoints and interests. This process involves and stimulates some
explanation of tacit positions.
This means that the interactive setting should allow a bi-directional transfer of
knowledge between stakeholders and the simulation model: knowledge elicitation
in one direction and validation and explanation of simulation outputs in the other
direction.
10.2.2.2 Case of Policy Making
In the case of simulation focusing on policy issues, there is a pragmatic, moral, and
now sometimes legal need to involve stakeholders, which may lead to open the
black box of models of social complexity used in policy making. Post-normal
approaches aim at making the decision process and its tools explicit so that
stakeholders can better discuss it and appropriate its outcomes. When this decision
process involves the use of decision support tools, which might include social
simulation models, this means that the models themselves should be opened to
stakeholders (Funtowicz et al. 1999). A simulation model is then expected to be
explicit enough so that stakeholders who might be concerned by the implementa-
tion of the policy at stake could discuss it. This legitimisation is socially based,
while validation, as mentioned with the previous case of use, is scientifically based
(Landry et al. 1996). Even though validation is still required in this case of use,
because it is the mode of evaluation for some participants, it is rather the
legitimisation of the model by the stakeholders which is to be worked out.
Participatory approaches may be a means for opening these models to
stakeholders, provided that formats of communication of models’ assumptions and
structure can be genuinely discussed. Involvement of stakeholders is expected to raise
their awareness of the assumptions of the model and potentially able to discuss these
and modify them. This includes the evolution of underlying values and choices made
in the design of model.
10.2.3 Simulation as a Means to Support Participation
Social simulation might also benefit to participation. While the previous subsection
was dedicated to appropriateness between the model and its use as a group decision
support tool, we focus here on participation which might be a component of a
decision making process.
Social simulation is seen here as an opportunity to foster participation and cope
with some of its pitfalls (Eversole 2003). Use of simulation models may lead to
some outcomes such as community building or social learning.
10.2.3.1 Dynamics and Uncertainties
Social systems have to deal with uncertainties just as social simulation models do.
This might hamper participatory processes: in wicked problems (Rittel and Webber
1973), encountered in many situations where participatory processes are organised,
stakeholders always maintain the opportunity related to these uncertainties to
challenge others’ viewpoints or observations. As an example: origin, flow and
consequences of non point source pollution are uncertain. This leads some farmers
to challenge the accusation, made by domestic water companies downstream of
their fields, that they are polluting their sources. Sometimes, disparate viewpoints
do not conflict. The gathering of these disparate pieces of knowledge is a way to
reduce uncertainty and allow the group of stakeholders involved in a participatory
process to progress; provided that they can work together.
Another characteristic of any social system which might hamper participation is
its dynamicity. Socio-ecological systems exhibit a range of dynamics; not only
social, but also natural, which evolve at various paces. In the application developed
by Etienne and colleagues in Causse Mejan, pine tree diffusion has a typical time
step of 20 years which is long according to the typical time steps of land use choices
and assessment (Étienne et al. 2003). In a participatory process it might be difficult
to put these dynamics on the agenda. Simulation models are known to be good tools
to deal with dynamic systems.
Simulation models are therefore a means to gather distributed pieces of knowl-
edge among stakeholders and to cope with scenarios in the face of uncertainties.
They can also help make the participants aware of potential changes or regime
shifts generated by their interactions (Kinzig et al. 2006).
10.2.3.2 Towards Social Learning
Participation is often linked with the concept of social learning (Webler et al. 1995).
However, for social learning to occur, participants should have a good understand-
ing of their interdependencies as well as of the system’s complexity. Social
simulation can provide these bases, provided that the communication is well
developed (Pahl-Wostl and Hare 2004).
This learning comes from exchanges among stakeholders involved in the par-
ticipatory process but also from new knowledge which emerges in the interaction.
Externalisation of tacit knowledge in boundary objects (Star and Griesemer 1989)
is useful for both: it facilitates communication in giving a joint framework to make
one’s knowledge explicit; and it enhances individual, as well as social, creativity
(Fischer et al. 2005).
Simulation models are good candidates to become such boundary objects. Agent
based models have long been considered as blackboards upon which various
disciplines could cooperate (Hochman et al. 1995). Through simulation outputs,
they provide the necessary feedback for reflexivity, be it individual or collective.
The question then remains whether such models constrain the format of knowl-
edge which might be externalised.
10.2.4 Synthesis: A Key Role of the Interaction Pattern

Between Model and Stakeholders
These three categories of expectations have led to specific requests for the devel-
opment of participation in relation to social simulation models. In the following
section, we provide an overview of these techniques. On the basis of the previous
requests, these techniques and methods have to be analysed according to the
following dimensions:
– Set of connections between the participation arena and simulation model: its
structure, its content, and organisation of its mobilisation;
– Control of the process; and
– Format of information which can travel from one pole to another: openness and
suitability to the diversity of stakeholders’ competencies.
10.3 A Diversity of Settings
In this section, we describe some examples of participatory techniques and

approaches associated with social simulation models. Settings described in this
overview stem from various fields and disciplines. Most of these have already
produced some reviews on participatory approaches. For the purpose of the discus-
sion in relation with social simulation, a synthesis of these reviews is provided here
with a focus on the requests identified in the previous section.
10.3.1 From System Science and Cybernetics
Cybernetics and system sciences have produced a first category of simulation models
of social complexity (Gilbert and Troitzsch 1999). These models are based on tools
originating from system dynamics, using specific software. They focus on flows of
resources and information between stocks which can be controlled.
Two main types of interactions between these models and stakeholders have so
far emerged: group model building (Vennix 1996); and management flight
simulators or microworlds (Maier and Grössler 2000).
Group Model Building experiments focus on the interaction with stakeholders in
the design stage of a modelling process. It associates techniques of system dynam-
ics modelling with brainstorming tools and other techniques of group work, mainly
based on workshops and meetings. This trend consists of integrating future users of
the model in the design stage. The participants are supposed to be the clients of the
modelling process. Rouwette and colleagues analysed 107 cases of such
experiments and proposed a number of guidelines to facilitate consistent reporting
on participatory modelling exercises. These guidelines focus on three categories:
context, mechanisms and results (Rouwette et al. 2002). The second category
focuses predominately on preparation activities and description of meetings,
along with factual elements and the modelling process.
This category of participatory modelling deals with the expectations identified in
the first section in the following manner:
– The participation arena is constituted of a rather small or medium size well
identified group. The structure of the interaction is rather global: debates tackle
the whole model, and participants are supposed to be concerned by the model
entity as a whole. The connections may convey information on the tacit knowl-
edge of stakeholders, as well as on their purposes. This is still very diverse
among the experiments. The group of stakeholders is mobilised within specific
events, workshops, which might be repeated. The aim is to feed the model but
also to increase the probability of use of the models produced.
– The process is predominately controlled by the modellers; and
– The format of information is generally not well formalised, even though
techniques, such as hexagons brainstorming or causal diagrams (Akkermans
1995), appear to organise the knowledge brought by stakeholders. This low
formalisation allows the issues related to stakeholder diversity to be tackled
and alleviated in the problem framing phase, but it leaves a large place to the
modellers’ interpretation.
Management flight simulators or microworlds constitute a complementary tech-
nique, which focuses more on the stages of use and simulation outcomes analysis,
even though this technique may also be used in a design stage to elicit tacit
knowledge. A key characteristic of this type of technique is to encourage learning
by doing. Participants, who might be the clients or other concerned people without
any formal relation to the modelling team, have to play through a simulation of the
model. Martin and colleagues have used this technique to validate a system
dynamics model on the hen industry (Martin et al. 2007). Participants were asked
to play with some parameters of the model.
When used to elicit knowledge, microworlds attempt to provide events that are
similar to those that participants already face or are likely to face in their activities
related to the issue at stake in the model. Le Bars and colleagues have thus
developed a game setting to lead farmers to understand the dynamics of their
territory with regard to water use and changes in EU Common Agricultural Policy
(Le Bars et al. 2004). In flight simulator experiments, interaction between
stakeholders and the simulation model is structured around future users of the
model or people whose stakes are represented in the model, with a slightly deeper
connection than with previous group modelling building approaches. Participants
are asked to deal with parameters of the model and are framed in the categories used
in the model. There is no a priori differentiation among participants. The

connections convey information about the object from the model to participants.
It also conveys the participants’ reactions to this object, and some behavioural
patterns observed that can provide new information for the modellers. This connec-
tion is activated by the participants working through specific events and focus on
the use of the tool. Control is still on the side of modellers, who frame the
interactions. The format of information is largely formalised from model to
stakeholders. It is not formalised from stakeholders to model.
10.3.2 Knowledge Engineering: Between Artificial Intelligence

and Social Psychology
Knowledge engineering focuses on a specific time of the interaction between

stakeholders and a simulation model in the design stage: the process of translating
tacit knowledge into conceptual or sometimes computational models. Many knowl-
edge elicitation techniques are useful in transforming written or oral text into pieces
of simulation models. The purpose of these techniques is to separate the
contributions made directly to the model from the design of the model itself.
Knowledge engineering aims to provide interfaces for this gap.
To deal with this interface, techniques have been developed, grounded in artifi-
cial intelligence, (social) psychology and cognitive science. Behavioural patterns in
social simulation models are often borrowed in simplified versions from these fields
(Moss et al. 2000; Pahl-Wostl ND Hare 2004). This cross-pollination of disciplines
can be potentially fruitful for model design. As an example, Abel and colleagues
have built upon the concept of a mental model. They assume that individuals have
representations of their world which may be formalised in causal rules. Working in
the Australian bush, they have designed specific individual interview protocols and
analysis frameworks to elicit these mental models (Abel et al. 1998). In this case,
interaction with the model occurs through the interviewer who in this case was also
the modeller. There was no collective interaction. Researchers dealing with the
interviews and the corresponding model design clearly guide the process. The
format of information is speech (in the form of a transcribed text), which is
transformed into a modelling language in this elicitation process.
Building upon Abel’s work, Becu has further minimised the involvement of the
modeller, still using individual interviews. This has led him to collaborate with an
anthropologist and to use ethnographic data as a benchmark. Individual interviews,
with the interviewee in the environment suitable to the purpose of the interview, led
him to identify objects and relations among these objects. These constitute the
initial basis for an exercise, labelled as playable stories: stakeholders, in his case
farmers from Northern Thailand, are asked to choose the key elements to describe
their world from their own viewpoints (with the possibility of adding new
elements), then to draw relations among them and to tell a story with this support
(Becu 2006; Becu et al. 2006). In this case, interaction between stakeholders and the
simulation model is still on an individual basis. The format of conveyed informa-
tion is finally less formal, but the work of translation is less important. However,
control of the process still remains largely in the hand of the modeller, but to a lesser
degree than in previous examples. This technique was further associated with semi-
automatic ontology building procedures by Dray and colleagues in order to gener-
ate collective representations of water management in the atoll of Tarawa
(Dray et al. 2006a).
With inspiration coming similarly from the domain of ethnography, Bharwani
and colleagues have developed the KNeTS method to elicit knowledge. Apart from
a first stage with a focus group, this method is also based on individual interviews.
As in Becu’s work, interaction occurs in two phases: elicitation through
questionnaires and involvement in the model design at the validation stage, which
is also considered as a learning phase for stakeholders. These authors used an
interactive decision tree to check with stakeholders whether the output of simula-
tion would fit their points of view (Bharwani 2006). Control of this process is on the
modeller’s side. The stakeholders’ interaction is marginally deeper in the model
than in previous examples, since there is a direct interaction with the model as in
management flight simulator. On the other hand, the ontology which is manipulated
seems to be poorer, since the categories of choices open in the interaction are rather
reduced. The format of information is open in the first phase and very structured in
the decision tree in the second phase. The structuration process used in the
modelling process occurs outside of the field of interaction with the stakeholders.
On its side, Group Decision Support System design domain is based on a
collective interaction with stakeholders as early as the design stage. These systems
tend to be used to address higher level stakeholders. In the method he developed,
ACKA, Hamel organised a simulation exercise with the stakeholders of a poultry
company. In this exercise, the participants were requested to play their own roles in
the company. He constrained the exchanges taken place during the exercise through
the use of an electronic communication medium so that he could analyse them and
keep track of them later. All of the participants’ communication was transformed
into graphs and dynamic diagrams (Hamel and Pinson 2005). In this case, the
format of information was quite structured.
10.3.3 From Software Engineering
Close to the artificial intelligence trend, working like Hamel and Pinson on the
design of Agent Based Models, there is an emerging trend in computing science
based on agent based participatory simulations (Guyot and Honiden 2006) or
participatory agent based design (Ramanath and Gilbert 2004). This trend focuses
on the development of computer tools, multi-agent systems, which originate from
software engineering. Guyot proposes the implementation of hybrid agents, with
agents in the software controlled by real agents, as avatars (Guyot 2006). These
avatars help the players’ understanding the system (Guyot and Honiden 2006).
They can be thought as learning agents: they learn from choices of their associated
player and are progressively designed (Rouchier 2003). The approaches working on
hybrid agents implement a deep connection between participants and the social
simulation model. Information conveyed in the interaction is relative to the model
assumptions, as well as to the model content.
Ramanath and Gilbert have reviewed a number of software engineering techniques
which may be coupled to participatory approaches (Ramanath and Gilbert 2004). This
union between software design and participatory approaches is based on joint pro-
duction not only between developers but also with end-users. Not only interaction
with stakeholders contributes to better software ergonomics – the Computer
Supported Cooperative Work (CSCW) workshops series being an example – but
their participation tends to improve their acceptation and further appropriation of the
mode.
The implementation of interactive techniques may take place at all stages of a
software development process. In early stages, joint application design (Wood and
Silver 1995) allows issues raised to be dealt with during the software development
phase, attributing a champion to each issue. It is also concerned with technical
issues. This protocol might involve other developers, as well as potential users. It
may also increase the computing literacy of the participants involved in the process.
This process is based on the implementation of rather well framed workshops.
Joint application design is supported by using prototypes. It is here we find a link
with a second technique: prototyping. This technique can be used all the way
through a software development cycle. It is based around providing rough versions
or parts of the targeted product. For example, it allows the pre-product to be
criticised, respecified, or the interface improved. Quite close to prototyping, in
the final stages of the process, user panels can be used to involve end-users in
assessment of the product. These panels are based on a demonstration or a test of the
targeted product.
In these cases, control of the process is dependent on the hiring of a skilful
facilitator. Otherwise, control of the process may become rather implicit. The
content of the interaction is rather technical, which makes it potentially unbalanced
according to participants’ literacy in computer science. An assessment of 37 joint
application design experiments has shown that the participation of users during the
process is actually rather poor, notably due to the technical nature of debates, which
is hardly compatible with the time allocated to a joint application design process by
users, compared to the time allocated by developers (Davidson 1999). Interaction is
rather superficial and needs translation. However, identification of a champion of
specific tasks gives a little bit more control to participants, as does involvement in
the content of pieces of the tool being developed.
Besides these approaches originating from software engineering, people work-
ing in thematic fields such as the environmental sciences propose co-design
workshops that focus on the development of simulation models. Such workshops
are a type of focus group, organised around the identification of actors, resources,
dynamics and interactions, suitable for a set of stakeholders to represent from a
socio-ecological system on which they express their own point of view (Étienne
2006). This approach, which occurs at the design stage of the modelling process is
supposed to lead participants to design the simulation model by themselves, by
formalising the conceptual model through a series of diagrams and a set of logical
sentences. The final interaction diagram and the attached logical sentences are then
translated by the modeller in computer code. It is in this type of process that a deep
interaction can occur between participants and the model. This interaction conveys
information on the model content, which is attached to the representations and
knowledge of each participant.
10.3.4 From Statistical Modelling
Bayesian Belief Networks have been developed to include in the computation of

probabilities, their dependence on the occurrence of any event. They can be useful
to represent complex systems and increasingly used in participatory settings
because their graphical nature facilitates discussion (Henriksen et al. 2004). A
group of participants can be asked individually or collectively to generate relations
between events and possibly probabilities as well. Henriksen and his colleagues
propose a method in seven stages which alternates between individual and collec-
tive assessment and revision of an existent Bayesian Belief Network diagram.
This approach is reported to still present some difficulties in encouraging strong
participant involvement due to the mathematical functions behind the network
structure. However, other researchers and practitioners have improved their com-
munication and facilitation of the technique with their own Bayesian Belief Net-
work processes and are receiving positive stakeholder engagement in the modelling
processes (Ticehurst et al. 2005). In the example of Henriksen and colleagues, the
process is controlled by the modeller and includes only a rather superficial coupling
between participants and the model. The translation of participant-provided infor-
mation into probabilities is mediated by the modeller and is rather opaque, as in
many participatory modelling approaches.
10.3.5 From the Social Sciences
The association of participatory approaches and social simulation modelling also

originates from disciplines not focussing on the production of tools but on under-
standing social systems. Social psychology, economics, management and policy
sciences have all developed their own interactive protocols to involve stakeholders
in the design and/or use of their models. Sociology is still at the beginning of this
process (Nancarrow 2005). These protocols propose a variety of structures of
experimental settings, from laboratory to in vivo experiments through interactive
platforms (Callon and Muniesa 2006). These three categories vary according to
their openness to the influence given to participants. The in vivo category is beyond
the scope of this paper since it does not involve modelling: the society in which the
experiment is embedded provides its own model (Callon and Muniesa 2006).
Laboratory settings are very controlled experiments, involving human subjects.
This is the case for most economic experiments. Participants are encouraged to
behave with a given rationality through instructions and payments at the end of the
session. In canonical experiments, analysis of the experiments is performed by the
scientist. The focus of the analysis is to understand the individual and collective
behavioural patterns generated by these settings. The purpose of these experiments
is either: the testing of theories and models; new knowledge on human behavioural
patterns in given situations; or the testing of new institutional configurations
(Friedman and Sunder 1994). These experiments are particularly efficient for
situations with strong communication issues or with important inter-individual
interactivity (Ostrom et al. 1994). The issue of simulating a real situation is not
considered, but rather the testing of a theoretical model. This field is currently very
active and evolve with the emergence of field experiments involving stakeholders
concerned by the issues idealised in the model tested, asking them to play in their
environment (Cardenas et al. 2000). With this configuration, interactions are rather
deep since participants act as parts of the model. The participants convey action
choices. However, the experimentalist strongly controls the process.
A platform is an intermediary setting more open to compromise and hybridisation
than the laboratory. Heterogeneity of participants is also more welcome, since the
setting is designed to enhance sharing interests. Through experimentation, a platform
is supposed to bridge through experimentation the gap between the world of the
model and that of the stakeholders (Callon and Muniesa 2006). Policy exercises and
role playing games, as developed in the companion modelling approach, are kinds of
these platforms (Richard and Barreteau 2006). Policy exercises embed stakeholders
in potential situations they might have to face in the future (Toth 1988). They stem
from war games that have been developed since the time of Ancient China and are
now used in public policy assessment (Duke and Geurts 2004) or environmental
foresighting (Mermet 1993). They are actually quite similar to the business games and
the system dynamics trend explained previously in subsection 10.2.1. However, the
underlying social simulation model is rather implicit; though it exists to create the
potential situation and to help identify the participants relevant to the exercise.
Association with a computer tool tends to be with a simulation model of the environ-
ment, that does not necessary involve a social component. The interaction between
participants and the social model is rather deep since they are pieces of the model and
connect with the model of their environment. Control of the process is rather diffuse.
There might be a genuine empowerment of participants since they have the possibility
of bringing their own parts of the social model to the process, and can adapt it in ways
different to what the designers expected. Alike with laboratory settings, platforms
provide information to the modeller about behavioural patterns of the participants.
Reaction to taboos or innovative behaviours in situations new to the participants, tacit
routines, and collective behavioural patterns can be elicited using these platforms,
while it is difficult with classical interviewing techniques.
Between experimental laboratory settings and policy exercises, the companion

modelling approach proposes an association of role playing games and agent based
simulations (Bousquet et al. 2002). Even though authors in this approach claim not
to limit themselves to these two categories of tools, they predominately rest in the
trend of participatory agent based simulations, and are thus close to the software
design and artificial intelligence trends presented above. This approach makes a full
use of similarities in architecture between role playing games and agent based
simulations (Barreteau 2003). Both implement autonomous agents that interact
within a shared dynamic environment. Joint use of both agent based simulation
and role playing games builds upon these similarities to express the same concep-
tual model. Authors in this approach use this to reinforce a principle of making all
the assumptions underlying a model used or design interactively with stakeholders
explicit and understood. At the design stage, this approach aims to incorporate
stakeholders’ viewpoints in the model. At the model use stage, it aims to improve
the appropriation of the tool produced, as well as to increase its legitimacy for
further operational use. However, this appropriation is still under discussion and
might be rather heterogeneous (Barreteau et al. 2005).
10.4 Participation in the Modelling Process:

Diversity of Phases and Intensity
While many authors claim to use participatory approaches for the simulation of social
complexity, there remains a large diversity of actual involvement of stakeholders and
of activities hidden behind this involvement. Associations of participatory methods
with social simulation models are rather heterogeneous. It is thus important to qualify
the actual involvement of stakeholders in these processes. This level of participation
can range from mere information received by concerned parties related to the output
of a process to the full involvement of a wide range of stakeholders at all stages of a
process. There are also many intermediary situations imaginable. Participation should
not be thought of as just talking, and diversity should be made explicit so that
criticisms towards participation as a global category (Irvin and Stansbury 2004) can
focus on specific implementations. This section explores the potential consequences
of this diversity in three dimensions: stage in the modelling process, degree of
involvement and heterogeneity of stakeholders involved.
10.4.1 Stages in the Modelling Process
The modelling process can be subdivided into the following stages, with the
possibility of iterating along them:
– Preliminary synthesis/diagnosis (through previously available data). This
includes making explicit the goal of the modelling process
– Data collection (specific to the modelling purpose)
– Conceptual model design

– Implementation
– Calibration and verification
– Simulation process (might be running a computer simulation model, playing a
game session, etc.)
– Validation
– Discussion of results
Involvement of stakeholders in each of the different stages of the modelling
process does not generate the same level of empowerment or learning, even if we
assume that this involvement is sincere. Preliminary synthesis, conceptual model
design, validation and, to some extent, discussion of results are framing stages;
stakeholder involvement at these levels gives power to stakeholders to orientate the
process. In the preliminary synthesis/diagnosis, stakeholders have the opportunity
to play a part in setting the agenda. This is the stage of problem structuring which is
identified as a key one in all participatory processes (Daniell et al. 2006). Even if the
agenda developed with stakeholder involvement might further evolve, its
initialisation generates a strong irreversibility in the process: data collection,
participants selection and (partially) modelling choices (architecture, platform)
are related to this agenda and are costly, either directly or through the necessity
of re-programming. The modelling process is a sequential decision process, and as
shown in theory of sequential decisions: initial decisions are often at the source of
more consequences than envisaged (Henry 1974; Richard and Trometter 2001).
Conceptual model design constitutes a landmark in the process. It is the
crystallisation of viewpoints that serves as a reference in further stages. Validation
is the compulsory stage where stakeholders will have the opportunity to check the
effectiveness of the computer model in representing correctly their behaviours and
ways of acting. Discussion of results may also constitute a framing phase, according
to the purpose of the discussion. If dimensions of discussion are to be defined and
model is open to be modified, there is some place for participants to (re-)orientate
the modelling process. Otherwise, if the discussion of results aims to choose from a
few scenarios for example, the choice is very narrow and might be completely
manipulated. In this regard, it has been shown that for any vote among composite
baskets, it is possible to maintain that one item always selected according to the way
the baskets are constituted (Marengo and Pasquali 2003). A scenario in this case is a
kind of composite basket.
In other stages of a modelling process, the influence of stakeholder involvement
on the overall process is less important. When data collection, or calibration and
verification involve participants, stakeholders tend to take the role of informants.
Among the various levels proposed in the classical ladder of participation explained
in the following subsection, these stages deal predominately with consultation.
Their involvement is framed by the format of information which is expected, and
on the parts of the model which are to be calibrated or validated. If the process is
open to modification in these frames, the level of participation might be higher, but
still with a limited scope.
Implementation stage is another mean to empower participants. It is often

implicitly framing. But empowerment through involving stakeholders in this tech-
nical activity is rather to raise their literacy in this part and raise the probability of
their appropriation of the model. Simulation stage is basically providing informa-
tion to stakeholders on what is being done. This is a technical stage (running the
simulation) which keeps a part of strategic choices (design of scenarios and
indicators to track the simulation progress). Involvement of stakeholders in the
technical part, such as through role playing games, increases their knowledge of the
model from inside, provided stakeholders have the literacy for that. Involvement in
strategic part is connected to the initial stage which has set the agenda. The further
this initialisation has gone in formalising the questions, the less empowering is this
involvement.
10.4.2 Level of Involvement
Level of involvement is a more classical dimension. It is inspired by the classical

hierarchy of participation levels proposed first by Arnstein (1969). Several reviews
and adaptations have been made since then, with the same focus on power issues
(Mostert 2006; Asselt et al. 2001). These works focus on what participation means
in decision making terms (the bases of many political or democratic theories), with
Democracy Cube (Fung 2006) or the work of Pateman (1990) and Rocha (1997). In
most of these examples, the emphasis is placed on who (citizens, managers or
policy makers) has the balance of power for final decision-making (i.e. the choice
phase of a decision process (Simon 1977)) but other issues of process are not
specifically mentioned. Such participation classifications, although useful in a
very general sense for the question of participation in modelling processes, do not
explicitly treat the issue of the place of a modeller or researchers with expert
knowledge (Daniell et al. 2006).
On these bases, we consider here the five following levels in which there are at
least some interactions between a group of citizens and a group of decision makers:
– Information supply: citizens are provided access to information. This is not
genuine participation since it is a one way interaction;
– Consultation: solicitation of citizens’ views;
– Co-thinking: real discussions between both groups;
– Co-design: citizens have an active contribution in policy design; and
– Co-decision making: decisions are taken jointly by members of both groups.
Since a modelling process is a kind of decision process, this hierarchy might apply
to modelling process as well. This is a little bit more complicated because two
processes are behind the modelling process and the network of interactions cannot
be represented with a group of citizens and a group of decision or policy makers only.
A modelling process with the purpose of simulation has two dimensions along
which these scales might be assessed: model content and building on one hand; and
control over model use on the other. Though these two dimensions are related, it is
useful to consider them separately as they provide power and knowledge: either
within the process; or in the system in which the process takes place. Each of these
dimensions is more closely related to specific stages in the modelling process
presented in the previous subsection. However, some stages, such as model design
or implementation, contribute to both dimensions.
Therefore we consider the following categories:
– Information on a model’s content and no control over model use;
– Consultation and no control over model use;
– Dialogue with modellers and no control over model use;
– Dialogue with modellers and control over model use;
– Co-building of a model and no control over model use; and
– Co-building of a model and control over model use;
Each category is described in the following sub-section by a flow of interactions
within an interaction network based on four poles: A, R, M, and P. A stands for all
people who are involved in and/or concerned by the social complexity at stake in
the modelling process. This includes policy makers and citizens. R stands for
researchers involved in the modelling process. M stands for the model. P stands
for policy makers. P is a subset of A, which gathers the actors who might use
the model and its output for the design of new regulations or policies concerning
the system as a whole. We chose to gather citizens and policy makers in A, as in the
modelling process they are rather equivalent in their interactions with the
researchers about the model. Their distinction is useful for the second dimension:
model use and dissemination. We assume that the default situation is an access of P
members to the output of the modelling process.
10.4.2.1 Information and No Control
A R
P M
Participants are informed about the model’s content and the simulation by
researchers, who are the only designers. No control over the model’s use or
dissemination is deputed to participants as such. Whatever the use of the model
may be afterwards, citizens become only better aware of the basis on which this
model has been built. However, the model exists and can be used by members of P.
This is the classical situation with simulation demonstration and explanation of a
model’s assumptions. This explanation might be achieved by more active means,
such as a role playing game. A switch to the following category occurs when this
explanation leads to a debate that makes the model open to modifications. Other-
wise, it remains mere information.
10.4.2.2 Consultation and No Control
A R
P M
Participants are consulted about the model’s content and its simulation that is by the
researchers, who are the only designers. They provide information and solicit
comments on the model. Mere data collection through a survey does not fall in
this category because it assumes active involvement from participants in providing
information to the modellers. Some knowledge elicitation techniques, such as BBN
design, tend to fall mostly in this category. Translation of the inputs originating
from participants into pieces of a model is performed only by researchers. This
translation is not necessary transparent. No control over use or dissemination of the
model is deputed to participants as such. Compared to previous category,
participants have the ability to frame marginally more of what is performed by
the model through their inputs to the model’s content However, the extent of this
ability depends on the participants’ skills to identify potential uses of a model. As in
any participatory process, when there is an unbalanced power relation between
parties, the process is also a way for policy makers to gain information from
stakeholders; information that could be used for strategic purposes. This bias can
be alleviated if the involvement of A includes all members of A, including the
subset P. The constructed model in this case may be used by the members of P.
10.4.2.3 Dialogue with Modellers and No Control
A R
P M
In this category, iterative and genuinely interactive processes between stakeholders

and modellers start to appear. There is still a translation of inputs from participants
into the model through the researchers, but there is feedback about these
developments to the stakeholders. This leads to discussion about the model. Conver-
gence of the discussion remains on the researchers’ side. Group Model Building
experiments predominately fall into this category. In this case, stakeholders may
increase their influence on the framing of the model with better prior assessment of
the scope of simulations to be examined. Biases related to strategic information being
revealed in the dialogue process are still present if there is unbalanced involvement of
Members of A, and notably if members of P are less active, but still present. However,
this category still represents indirect control and no specification of model use is left
open to the stakeholders. At the end of the process, the created model can be used by
members of P without any control or any roadmap set by other members of A.
10.4.2.4 Dialogue with Modellers and Control
A R
P M
This category is the same as the previous one with translation of stakeholders’
inputs and feedback from the researchers about them. However, the output of the
discussion, the model, is appropriated by stakeholders. They have control over its
use and dissemination of models which may have been produced through the
modelling process: who might use them; with which protocol; and what is the
value of their outputs. They can decide whether the model and simulations are
legitimate to be used for the design of policies that may concern them. However,
this appropriation raises issues of dialogue between researchers and stakeholders
about the suitability of model for various uses. Comparison of several participatory
agent based simulations has shown that there is a need for dialogue about not only a
model’s content but also on its domain of validity (Barreteau et al. 2005).
10.4.2.5 Co-building of a Model and No Control
A R
P M
A further stage of empowerment of stakeholders through participation in a modelling

process is their co-building of the model. The design and/or implementation of such a
model are joint activities between the researchers and stakeholders. Co-design
workshops or joint application development fall into this category, provided that
there is genuinely no translation of stakeholders’ inputs by the researchers.
Techniques originating from Artificial Intelligence and knowledge engineering, as
presented above, aim to reach this level, either through the implementation of virtual
agents extending stakeholders, or through constraining the interactions between
actors through a computer network. This involvement increases the fidelity of the
model to match stakeholders’ viewpoints and behavioural patterns. However, at the
end of the process, the created model can still be used by members of P without any
control or any roadmap set by other members of A.
10.4.2.6 Co-building of a Model and Control
A R
P M
This category is the same as the previous one, but actors now have control over use
and dissemination of models which may be produced through the process. This
leads to possible stakeholder appropriation of the models, raising the same issues as
in Sect. 3.2.4.
10.4.3 Heterogeneity of Actors
Eversole points out the need for participatory processes to take into account the
complexity of the society involved including: power relations; institutions; and the
diversity of viewpoints (Eversole 2003). This is all the more true when applied to
the participatory process of social simulation modelling. Most settings presented in
Sect. 10.2 have a limited capacity to involve a large numbers of people in
interactions with a given version of a model. When interactions convey viewpoints
or behavioural patterns, heterogeneity may not appear if no attention is paid to it.
Due to limits in terms of number of participants, participatory approaches that deal
with social simulation modelling involve usually representatives or spokespeople.
The issue of their statistical representativeness is left aside here, as the aim is to
comprehend the diversity of possible viewpoints and behavioural patterns. There is
still an issue of their representativeness through their legitimacy to speak for the
group they represent, as well as their competency to do so. The feedback of these
spokespersons to their group should also be questioned. When issues of empower-
ment are brought to the fore, the potential for framing or controlling the process is
dedicated to the participants. This might induce echoes in power relations within
the group, notably due to training that may be induced.
Van Daalen and Bots have proposed a categorisation of participatory modelling
according to this dimension with three scales: individual involvement; a group
considered as homogeneous, and a heterogeneous group (Daalen and Bots 2006).
Table 10.1 provides examples of each level according to the two processes involved
that were explained in previous subsection.
These three categories are represented in the diagrams below, as expansions of
the relation between A and (M [ R) in the previous subsection. The third category
corresponds to the deep connection mentioned in the first section (Figs. 10.1, 10.2,
and 10.3).
Some other ways are currently explored with hybrid agents to technically over-
come the difficulty of dealing with representatives: by involving them all in large
systems. The internet or mobile phone networks provide the technical substrate for
Table 10.1 Categories of participation according to level of heterogeneity embraced (From

Daalen and Bots 2006)
Level Model construction Model use
Computer model Gaming simulation
1. Individual Knowledge elicitation Model can be executed Individual can ‘play’
stakeholders involving one or and individual an actor in a flight
more individuals stakeholders are simulator setting
separately; informed of the (e.g. Maier and
depending on the result (e.g. Dudley Grössler 2000;
modelling method 2003) Sterman 1992)
this may consist of
interviews about
(perceptions on) a
system or
questionnaires
related to the aspects
being modelled
(e.g. Molin 2005)
2. Homogenous Same as 1, but group Use of a model in a Multi-player gaming
group model building homogenous group simulation can be
includes interaction means that the model conducted, the
between can be run in a game is followed by
stakeholders (e.g. workshop setting and a debriefing (e.g.
Castella et al. 2005) model results are Mayer et al. 2005)
discussed (e.g.
Daalen et al. 1998)
3. Heterogeneous Same as 2, but group Same as 2, but results Same as 2, but full
group model building discussed with stakeholder group
interaction between stakeholders with involved (e.g.
stakeholders with different Étienne et al. 2003)
different perceptions/beliefs
perceptions/beliefs
(e.g. Van den Belt
2004)
such interaction. A large number of participants have a virtual component in a large

system, interacting with other components, possibly with the purpose of building a
model (Klopfer et al. 2004). However, in this case it is rather an individual interaction
of these participants with the system, than genuine interactions amongst the
participants.
10.4.4 Which Configurations Can Meet the Expectations

of the First Section?
In this subsection we revisit the expectations towards the joint use of participatory
approaches and social simulation presented in the first section, through the
Fig. 10.1 Individual A1

involvement
A2
A3
A4 M
A5
A6
A7
Fig. 10.2 Homogenous

group involvement A1
A2
A3
A4
M
A5
A6
A7
Fig. 10.3 Heterogenous

group involvement A1
A2
A3
A4 M
A5
A6
A7
Table 10.2 Matching expectations on joint use of participatory approaches and social simulation
modelling with categories of participation
Key stage(s) Minimum level Level of
Expectation for participation of empowerment heterogeneity
Increase model’s quality Simulation Information and Heterogeneous
with social diversity no control group
and capacity to evolve
Increase model’s quality Simulation Information and Heterogeneous
through distribution no control group
of control
Improve suitability of Design Dialogue and Individual
simulation model’s no control
use for increasing
knowledge
Improve suitability of Design and Dialogue and Homogeneous
simulation model’s discussion control group
use for policy making of results
Simulation as a means to Discussion of Consultation and Homogeneous
support participation results no control group
to deal with dynamics (depend on
and uncertainties participatory
process to be
supported)
Simulation as a means to Preliminary diagnosis, Co-building Heterogeneous
support participation design and discussion and control group
through social learning of results (to be preferred)
categorisations above. This is a tentative mapping of participatory approach

categorisation with model expectations. Table 10.2 below synthesises this mapping.
The two expectations dealing with increasing a model’s quality often actually
use participants as (sometimes cheap) resources in the simulation modelling pro-
cess. The most important stage is simulation, because participants are supposed to
bring missing information to the simulation, as well as the missing complexity. The
minimum level of empowerment is rather low. These processes are hardly partici-
patory in that sense, because participants are not supposed to benefit from the
process, except a potential payment. A higher level of empowerment might increase
the quality of participants’ involvement in the process through a deeper concern in
the outcome of the simulation. Finally, the heterogeneous group level is obviously
to be respected because it can instil a deep connection between stakeholders and the
model still and concurrently profit from their interactions with each other.
To make simulation models match their intended use, the key stage is the design
process. Stakeholders are supposed to aid the building of an appropriate model. The
main difference between targets of simulation model’s use is in the necessity to give
control over the process to stakeholders in case of policy making. New knowledge is
of individual benefit to all participants and the emergence of fruitful interactions can
also become an individual benefit. There are few direct consequences of this new
knowledge. Therefore, control over the process in this case is useless and involvement
might be individual, as with knowledge elicitation techniques. However, higher level

of stakeholder heterogeneity might raise the knowledge acquired in the process.
When simulation is used to support participation, discussion of results is a key
stage. Previous stages aid in the problem framing and literacy increase of participants
that allow them to reach more solid interpretations. The empowerment level is rather
dependent on the participatory process that is being supported. However, consultation
in the modelling process should be a minimum requirement so that uncertainties and
dynamics tackled by the simulations are relevant to the stakeholders. When focusing
on social learning, co-building and control should be preferred because this category
increases the potential for exploration and creativity. However, some social learning
might take place in lesser levels, provided that group heterogeneity is well
encouraged in the process.
10.5 Combining Approaches and Techniques at Work
We present in this section two case studies implementing various methods for
joining social simulation modelling and participatory approaches. The first deals
with fire hazard prevention in southern France, and the second one with groundwa-
ter management in the Atoll of Kiribati.
10.5.1 The Fire Hazard Case Study
In December 2005, the Forest Service of the Gard Department of Agriculture

(DDAF), decided to start a fire prevention campaign focused on fire hazard at the
interface between urban and forest areas. Interested in the participatory approaches
developed by INRA researchers on fire prevention and forest management planning
(Etienne 2003), they ask for an adaptation of the SylvoPast model to the periurban
context in order to make local politicians aware of the increasing fire hazard
problem. The District of Nı̂mes City (NM) who was already interested in the use
of role playing games for empowering stakeholders and decision makers, asked the
Ecodevelopment Unit of the INRA of Avignon to develop a companion modelling
approach based on social simulations and a participatory involvement of all the
mayors of the district.
The modelling process was subdivided into seven stages:
1. Collection and connection on a GIS of relevant cartographic data on forests,
land-use and urbanization, and individual interviews with local extensionists on
farmers, foresters and property developers practices.
2. Co-construction with DDAF and NM of a virtual but implicit map representing
three villages typical from the northern area of Nı̂mes city and validation of the
map (shape, land-use attributes and scale) by a group of experts (EX) covering the
main activities of the territory (agriculture and livestock extensionists, forest

managers, hunting manager, land tenure regulator, fire brigade captain and town
planner).
3. Co-construction, with NM, DDAF and experts of a conceptual model accounting
for the current functioning of the territory and the probable dynamic trends to
occur during the next 15 years. This participatory process followed the ARDI
methodology mentioned in Sect. 10.2.3 (Étienne 2006).
4. Implementation of the NimetFeu model on Cormas multi-agent platform by
INRA researchers and validation of the model by simulating with the co-
construction group, the current situation and its consequences on fire hazard
and landscape dynamics for the 15 following years.
5. Co-construction and test of a role-playing game (RPG) using the NimetFeu
model as a way to simulate automatically natural processes and some social
decisions (vineyard abandonment, horse herding, fire fighting). The other social
decisions were programmed to be taken directly by the players and used as an
input to the model.
6. Use of the RPG during several sessions gathering 6 players (3 mayors,
1 developer, 1 NM representative, 1 DDAF technician) until the 14 villages
involved in the project did participate to a session.
7. Adaptation of NimetPasLeFeu to other ecological conditions, and decision of the
Gard Department to become autonomous in running RPG sessions. A facilitator
and a data manager were trained and tested during sessions organized in the
framework of an INTEREG project with mayors and fire prevention experts
from France, Spain, Italy and Portugal
The approach is based on a mutual comprehension of the elements of the territory
that make sense with the question asked. This sharing of representations is done by
means of a series of collective workshops during which Actors, Resources, Dynamics
and Interactions (ARDI) which make the stakes of the territory are identified and
elicited. To facilitate this sharing, the answers to the questions are formalized into
easily comprehensible diagrams, with a minimum of coding making it possible to
classify the provided information. The role of the facilitator only consists in calling
upon each participant, writing down the proposals in a standard way, and asking for
reformulating when the proposal is too generic, enounced with a polysemous word or
can lend to confusion.
In both models, the environment is divided into three neighbouring villages
covering the gradient of urbanization and agricultural land/woodland ratio currently
observed around Nı̂mes city. It is visualized by means of a cellular automaton through
a spatial grid representing 18 land-use types that can change according to natural
transitions or human activities.
Four categories of social entities are identified: property developers, mayors,
farmers and fire prevention managers. The developers propose new urban
developments according to social demand and land prices. They have to respect the
government regulations (flood hazard, protected areas, urban zoning). Mayors select
an urbanization strategy (to densify, to develop on fallow land, olive groves or
forests), update their urban zoning according to urban land availability and social
demand and make agreements with the developers. When updating the urban zoning,
they can create new roads. Farmers crop their fields using or not current practices that
impact fire hazard (vineyards weeding, stubble ploughing) or adapt to the economic
crisis of certain commodities by uprooting and setting aside lowland vineyards or
olive groves near to urban zones. The fire prevention manager establishes a fuel-break
in a strategic place, selected according to fire hazard ranges in the forest and the
possible connections with croplands, as well as available funds and forest cleaning
costs.
Four biophysical models issued from previous researches and adjusted to the local
conditions are integrated to the MAS to account for fallow development, shrub
encroachment, pine overspreading and fire propagation. The model is run at a
1-year time step, the state represented on the map corresponding to the land cover
at the end of June (beginning of the wildfire period). Each participant was invited to
propose a set of key indicators that permit them to monitor key changes on ecological
or socio-economic aspects. A common agreement was made on what to measure, on
which entity, with which unit, and on the way to represent the corresponding
qualitative or quantitative value (visualizing probes on graphs or viewpoints on
maps). They were encouraged to elaborate simple legends, in order to be able to
share their point of view with the other participants while running the model.
The first MAS was exclusively used to support the collective thinking on which
procedures and agents will be affected to players, and which ones would be
automatically simulated by the computer. In the RPG model, the playing board
was strongly simplified with only four types of land cover. Running the game gives
participants the opportunity to play individually or collectively by turns, according
to a precisely defined sequence. While the mayors players draw the limits of the
urban zone and rank the price of constructible land according to its current land-use,
the developer player sorts randomly a development demand and elaborates a
strategy (village, density, livelihood). Then begin a series of negotiations between
the developer and the three mayors in order to decide where to build, at which
density and with which type of fire prevention equipment. All the players’ decisions
are input into the computer and landscape dynamics is simulated by running the
model. Players get different types of output from the simulation run: budget
updating, new land-use mapping, popularity scoring. Each round corresponds to a
3-year lapse and is repeated three to four times according to players’ availability.
A specific effort is made in the RPG design to account for physical remoteness
and territory identity among participant: the playing room is set up into three
neighbouring but distinct boxes for the three mayors (each box represents one
village), one isolated small table for the developer, and another game place with
two tables, one small for the DDAF and a huge one for NM. Lastly, in a corner, the
computer stuff is placed with an interactive board than can both be used as a screen
to project different viewpoints on the map or as an interactive town plan to identify
the parcels’ number.
At the end of the game, all the participants are gathered in the computer room
and discuss collectively, with the support of fast replays of the game played.
Table 10.3 Classification of type of participation in various stages of the NimetPasleFeu

experiment
Involvement Heterogeneity nb
Preliminary diagnosis Consultation Individuals 10
Data collecting Consultation Individuals 3
Conceptual model designing Co-design Heterogeneous group 14
Implementing Information Individuals 2
Calibrating and validating Co-thinking Heterogeneous group 14
Role-playing game designing Co-design Heterogeneous group 14
RPG playing and debriefing Co-decision making Heterogeneous group 30
Getting self sufficient Information Individuals 3
Different topics are tackled related to ecological processes (effect of fire, main
dynamics observed), attitudes (main concerns, changes in practices), and social
behaviours (negotiations, alliances, strategies).
Along these various stages, this experiment features a diversity of involvement
as well as of structure of interactions. This is synthesised in the Table 10.3 above.
10.5.2 The AtollGame Experiment
This study is carried out in the Republic of Kiribati, on the low-lying atoll of
Tarawa. The water resources are predominantly located in freshwater lenses on the
largest islands of the atoll. South Tarawa is the capital and main population centre
of the Republic. The water supply for the urban area of South Tarawa is pumped
from horizontal infiltration galleries in groundwater protection zones. These cur-
rently supply about 60 % of the needs of South Tarawa’s communities. The
government’s declaration of water reserves over privately owned land has lead to
conflicts, illegal settlements and vandalism of public assets (Perez et al. 2004).
The AtollGame experiment aims at providing the relevant information to the local
actors, including institutional and local community representatives, in order to facili-
tate dialogue and to help devise together sustainable and equitable water management
practices. Knowledge elicitation techniques as well as Multi Agent-Based
Simulations (MABS) coupled with a role-playing game have been implemented to
fulfil this aim. In order to collect, understand and merge viewpoints coming from
different stakeholders, the following five-stage methodology is applied: (1) collecting
local and expert knowledge; (2) blending the different viewpoints into a game-based
model; (3) playing the game with the different stakeholders; (4) formalising the
different scenarios investigated in computer simulations; and (5) exploring the
simulated outcomes with the different stakeholders (Dray et al. 2006b).
Initial knowledge elicitation (Stages 1 and 2) relies on three successive methods.
First, a Global Targeted Appraisal focuses on social group leaders in order to collect
different standpoints and their articulated mental models. These collective models
are partly validated through Individual Activities Surveys focusing on behavioural
patterns of individual islanders. Then, these individual representations are merged

into one collective model using qualitative analysis techniques. This conceptual
model is further simplified in order to create a computer-assisted role-playing game
(AtollGame). The range of contrasted viewpoints confirms the need for an effective
consultation, and engagement of the local population in the design of future water
management schemes in order to warrant the long-term sustainability of the system.
Clear evidence of the inherent duality between land and water use rights on one
hand, and between water exploitation and distribution on the other hand, provides
essential features to frame the computer-assisted Role Playing Game.
The assistance of a computer is needed as far as interactions between groundwa-
ter dynamics and surface water balance involve complex spatial and time-
dependent interactions (Perez et al. 2004). The use of Agent-Based Modelling
(ABM) enables us to take full advantage of the structure of the conceptual model.
We developed the AtollGame simulator with the CORMAS# platform (Bousquet
et al. 1998).1
A board game version reproduces the main features of the AtollGame simulator
(Dray et al. 2006a). 16 players – 8 on each island – are able to interact according to a
set of pre-defined rules. Their choices and actions are directly incorporated into the
simulator at the end of each round of the game. During the game, players can ask for
more information from the simulator or discuss the results provided by the simula-
tor (salinity index, global demand). Landowners, traditional or new buyers, are the
essential actors in the negotiations with the government. The connection between
land tenure issues and water management is essential. It drives the land use
restrictions and land leases discussions. The population increase, mainly through
immigration, is perceived as a threat in terms of water consumption, pollution
generation and pressure on the land. Financial issues linked with water management
usually deal with land leases, equipment investment and, seldom, with water
pricing. Hence, the model features:
– Agents/players becoming a local landowners;
– Land and water allocation conflicting rules, and various sources of incomes;
– An increasing number of new settlers on agents/players’ land.
The individual objective of the players is to minimize the number of angry or
sick people in their house. People may become ANGRY because they didn’t have
enough water to drink during the round. People may become SICK if they drank
unhealthy (polluted or salty) water during the round. POLLUTION depends on the
number of people living on the island and contaminating the freshwater lens.
SALTY WATER depends on the recharge rate of the fresh water lens and the
location of the people on the island.
At first, representatives from the different islands displayed different viewpoints
about the Water Reserves. Hence, the group meetings organized in the villages prior
1
More details about the AtollGame can be found online at http://cormas.cirad.fr/en/applica/
atollGame.htm.
Fig. 10.4 Flowchart of financial, technical, and social solutions agreed on by the participants of
the AtollGame experiment
the workshop allowed for a really open debate. On the institutional side, the position
of the different officers attending the workshop demonstrated a clear commitment to
the project. All the participants showed the same level of motivation either to
express their views on the issue or to genuinely try to understand other viewpoints.
Participants also accepted to follow the rules proposed by the project team, espe-
cially the necessity to look at the problem from a broader perspective. During the
first rounds, the players quickly handled the game and entered into interpersonal
discussions and comparisons. The atmosphere was good and the game seemed
playful enough to maintain the participants’ interest alive. The second day, the
introduction of a Water Management Agency and the selection of its (virtual)
Director created a little tension among the participants. But, after a while, the
players accepted the new situation as a gaming scenario and started to interact
with the newly created institution. At this stage, players started to mix arguments
based on the game with other ones coming directly from the reality. On Island 1,
players entered direct negotiations with the (virtual) Director of the Water Manage-
ment Agency. On Island 2, discussions opposed players willing or not to pay the fee.
Finally, the project team introduced the fact that the Water Management Agency
was no longer able to maintain the reticulated system due to a poor recovery of the
service fees. It had for immediate consequence a sharp decrease of the water
quantity offered on Island 2.
Then, players from both tables were asked to list solutions to improve the
situation on their island. When the two lists were completed, the project team and
the participants built a flowchart of financial, technical and social solutions, taking
into account issues from both islands (Fig. 10.4).
A collective analysis of the flowchart concluded that the actual situation was
largely unsustainable either from a financial or social viewpoint. The flowchart
above provides a set of inter-dependent solutions that should be explored in order to
gradually address the present situation.
10.6 Discussion: Relations Between Participants and Models
The diverse categories of joint implementation of participatory approaches and

social simulation modelling feature a diversity of relations between a set of people,
participants, and a model.
Classical social simulation models do not feature any participant. People are
represented in the model, sometimes from assumed or theoretical behavioural
patterns. This entails exploring potential emergent phenomena from interactions
among these behavioural patterns. Some participatory approaches involve only an
implicit social model. Within this scope, there is a large diversity of relations. This
diversity is based on the role undertaken by stakeholders, their actual involvement
and issues tackled by the model.
In all the processes allying social simulation models and participation,
stakeholders take on various roles: pieces in simulation, interfaces for coupling
various sources of knowledge, beneficiaries of the process, key informants. . . As
pointed out by Ryan, managers are overwhelmed by the complexity to be managed.
Participation is a way to share this burden (Ryan 2000). Stakeholders provide the
missing interactions and add missing pieces of knowledge, such as tacit knowledge
(Johannessen et al. 2001). If involvement of stakeholders is useful for principal agents
such as managers, we propose it as a rule that they should gain some empowerment in
the process.
Stakeholders can be key pieces of the modelling process itself as well. In the
simulation they are an alternative to computer code to provide the engine
(Hanneman 1995). They provide an answer to issues of coupling several viewpoints
(Robinson 1991).
However actual involvement of people in a participatory modelling process
might largely differ from formal involvement planned. Leaving aside cases of
manipulation and announce effect, people have also to find their place in the
participatory process. Suitability of participatory approaches in a specific society
has to be taken into account: context (including social) is a key driver for success in
stakeholder involvement (Kujala 2003), and practice of interactive policy making
processes depend on local culture (Driessen et al. 2001). Representation
mechanisms have already be pointed out as a major factor. It has to be tuned to
this local social and cultural context. At a finer grain, facilitator has a key role to
lead people towards the level of involvement they are invited (Akkermans and
Vennix 1997).
10.7 Conclusion
This chapter provides a review of the diversity of association of participatory

approaches and social simulation, for their mutual benefit. This diversity of
approaches allows tackling expectations about increasing model’s quality, model’s
suitability to its intended use and improving participation. Their diversity is built
upon ingredients coming from various disciplines from social sciences to computer
sciences and management. It is expressed according to the implementation of
interactions between the participants and the simulation model, the control of the
process and the format of information. This leads to expand the classical ladder of
participation towards categorization according to the stage in the modelling process
when participation takes place and the structure of the interaction to cope with the
heterogeneity of stakeholders.
This diversity requires a cautious description of each implementation in situa-
tion, so that any evaluation is specific to the implementation of a given association
in its context. Generalisation can then be done only on the relation of this practice of
participatory simulation and its suitability to its context and purpose. Efficiency to
induce changes in practice or knowledge depends on the respect of a triple contin-
gency of collective decision processes: time, people, means (Miettinen and
Virkkunen 2005). This means to respect and take into account the own dynamics
within the social system at stake, to allow the participation of people with their
whole essence (including tacit knowledge, networks, relations to the world), and to
be adaptive to means and competences present within the system (Barreteau 2007).
Another dimension of evaluation should be democracy, since it is often put to the
front. This raises the issue of the existence of a control of the process. Does it rely
only on modellers or is it more shared? Finally there is a necessity of being more
explicit on the kind of PA which is used because of the potential deconsideration of
the whole family if expectations are deceived.
Further Reading
Participatory modelling is increasingly present in special sessions of conferences or

special features of scientific journals. A first source of further readings consists in
case studies. Among others, Environmental Modelling and Software had a special
issue on modelling with stakeholders (Bousquet and Voinov 2010), where readers
will find a whole set of well described case studies using various methods. The
biennial international Environmental Modelling and Software conferences have
also specific tracks for participatory modelling; proceedings are available online
(see http://www.iemss.org/society/ under publications). For specific tools, refer to
the papers of a symposium on simulation and gaming in natural resource manage-
ment, published as a special issue of Simulation and Gaming (volume 38, issues
2 and 3). The introductory paper giving an overview is (Barreteau et al. 2007).
Reflexivity is crucial for practitioners of participatory processes, as part of the

need for more cautious evaluation of participatory processes as pointed out by
Rowe and Frewer (2004). Another direction for reading consists in methods for
evaluation and assessment of stakeholder involvement in modelling processes.
Etienne edited a whole book aiming at assessing consequences of a specific
approach, so-called companion modelling (Étienne 2011).
Readers who are more interested in stakeholder involvement in modelling at a
more technical level should go for the review paper of Ramanath and Gilbert (2004)
which provides a nice overview of this point of view.
References
Abel N, Ross H, Walker P (1998) Mental models in rangeland research, communication and
management. Rangel J 20:77–91
Akkermans HA (1995) Developing a logistics strategy through participative business modelling.
Int J Oper Prod Manag 15:100–112
Akkermans HA, Vennix JAM (1997) Clients’ opinions on group model building: an exploratory
study. Syst Dyn Rev 13:3–31
Arnstein S (1969) A ladder of citizen participation. J Am Plann Assoc 35:216–224
Barreteau O (2003) The joint use of role-playing games and models regarding negotiation
processes: characterization of associations. J Artif Soc Soc Simul 6(2), http://jasss.soc.surrey.
ac.uk/6/2/3.html
Barreteau O (2007) Modèles et processus de décision collective: entre compréhension et facilita-
tion de la gestion concertée de la ressource en eau (HDR Thesis). Paris Dauphine University,
Paris
Barreteau O, Bousquet F (1999) Jeux de rôles et validation de systèmes multi-agents. In: Gleizes
M-P, Marcenac P (eds) Ingénierie des systèmes multi-agents, actes des 7èmes JFIADSMA.
Hermès, Paris, pp 67–80
Barreteau O, Bousquet F, Attonaty J-M (2001) Role-playing games for opening the black box of
multi-agent systems: method and teachings of its application to Senegal River Valley irrigated
systems. J Artif Soc Soc Simul 4(2), http://jasss.soc.surrey.ac.uk/4/2/5.html
Barreteau O, Hare M, Krywkow J, Boutet A (2005) Model designed through participatory
processes: whose model is it? In: Ferrand N, Perez P, Batten D (eds) Joint conference on
multiagent modelling for environmental management, CABM-HEMA-SMAGET 2005. Bourg
St Maurice – Les Arcs, 21–25 Mar 2005
Barreteau O, Le Page C, Perez P (2007) Contribution of simulation and gaming to natural resource
management issues: An introduction. Simul Gaming 38(2):185–194
Becu N (2006) Identification et modélisation des représentations des acteurs locaux pour la gestion
des bassins versants. PhD thesis, Sciences de l’eau, Université Montpellier 2, Montpellier,
France
Becu N, Barreteau O, Perez P, Saising J, Sungted S (2006) A methodology for identifying and
formalizing farmers’ representations of watershed management: a case study from Northern
Thailand. In: Bousquet F, Trebuil G, Hardy B (eds) Companion modeling and multi-agent
systems for integrated natural resource management in Asia. IRRI, Los Baños, pp 41–62
Bharwani S (2006) Understanding complex behavior and decision making using ethnographic
knowledge elicitation tools (KnETs). Soc Sci Comput Rev 24:78–105
Bousquet F, Voinov A (eds) (2010) Thematic issue – modelling with stakeholders. Environ Modell
Softw 25(11):1267–1488
Bousquet F, Bakam I, Proton H,Le Page C (1998) Cormas: common-pool resources and multi-
agent systems. In: Lecture notes in artificial intelligence 1416, Springer, Berlin pp 826–838
Bousquet F, Barreteau O, Le Page C, Mullon C, Weber J (1999) An environmental modelling
approach: the use of multi-agent simulations. In: Blasco F, Weill A (eds) Advances in
environmental and ecological modelling. Elsevier, Amsterdam, pp 113–122
Bousquet F et al (2002) Multi-agent systems and role games: an approach for ecosystem co-
management. In: Janssen M (ed) Complexity and ecosystem management: the theory and
practice of multi-agent approaches. Edward Elgar, Northampton, pp 248–285
Callon M, Muniesa F (2006) Economic experiments and the construction of markets.
In: MacKenzie D, Muniesa F, Siu L (eds) Do economists make markets? On the performativity
of economics. Princeton University Press, Princeton, pp 163–189
Cardenas J-C, Stranlund J, Willis C (2000) Local environmental control and institutional
crowding-out. World Dev 28:1719–1733
Castella JC, Tran Ngoc T, Boissau S (2005) Participatory simulation of land-use changes in the
northern mountains of Vietnam: the combined use of an agent-based model, a role-playing
game, and a geographic information system. Ecol Soc 10(1):27, http://www.ecologyand
society.org/vol10/iss1/art27/
Cockes D, Ive J (1996) Mediation support for forest land allocation: the SIRO-MED system.
Environ Manage 20(1):41–52
Daalen CE van, Bots PWG (2006) Participatory model construction and model use in natural
resource management. In: Proceedings of the workshop on formalised and non-formalised
methods in resource management – knowledge and learning in participatory processes,
Osnabrück, 21–22 Sept 2006, http://www.partizipa.uni-osnabrueck.de/wissAbschluss.html
Daniell KA, Ferrand N, Tsoukias A (2006) Investigating participatory modelling processes for
group decision aiding in water planning and management. In: Seifert S, Weinhardt C (eds)
Proceedings of group decision and negotiation (GDN) 2006, International conference,
Karlsruhe, 25–28 June 2006, Universitätsverlag Karlsruhe, Karlsruhe, pp 207–210
D’Aquino P, Le Page C, Bousquet F, Bah A (2003) Using self-designed role-playing games and a
multi-agent system to empower a local decision-making process for land use management: the
SelfCormas experiment in Senegal. J Artif Soc Soc Simul 6(3), http://jasss.soc.surrey.ac.uk/6/
3/5.html
Davidson EJ (1999) Joint application design (JAD) in practice. J Syst Softw 45:215–223
DeSanctis G, Gallupe RB (1987) A foundation for the study of group decision support systems.
Manag Sci 33:589–609
Dewulf A, Bouwen R, Tailleu T (2006) The multi-actor simulation ‘Podocarpus National Park’ as
a tool for teaching and researching issue framing. In: Proceedings of IACM 2006 Montreal
Meetings. http://ssrn.com/abstract¼915943
Dray A et al (2006a) The AtollGame experience: from knowledge Engineering to a computer-
assisted role playing game. J Artif Soc Soc Simul 9(1), http://jasss.soc.surrey.ac.uk/9/1/6.html
Dray A, Perez P, Le Page C, D’Aquino P, White I (2006b) AtollGame: a companion modelling
experience in the pacific. In: Perez P, Batten D (eds) Complex science for a complex world:
exploring human ecosystems with agents. ANU E Press, Canberra, pp 255–280
Driessen PPJ, Glasbergen P, Verdaas C (2001) Interactive policy making: a model of management
for public works. Eur J Oper Res 128:322–337
Drogoul A, Vanbergue D, Meurisse T (2003) Multi-agent based simulation: where are the agents?
In: Sichman JS, Bousquet F, Davidsson P (eds) Multi-agent-based simulation II: third interna-
tional workshop, MABS 2002, Bologna, Revised papers (Lecture notes in computer science,
2581), Springer, Berlin/Heidelberg, 15–16 July 2002, pp 1–15
Dudley RG (2003) Modeling the effects of a log export ban in Indonesia. Syst Dyn Rev 20:99–116
Duke RD, Geurts JLA (2004) Policy games for strategic management. Dutch University Press,
Amsterdam
Étienne M (2006) Companion modelling: a tool for dialogue and concertation in biosphere
reserves. In: Bouamrane M (ed) Biodiversity and stakeholders: concertation itineraries, Bio-
sphere reserves – technical notes 1, Unesco, Paris, pp 44–52
Étienne M (ed) (2011) Companion modeling: a participatory approach to support sustainable

development. QUAE, Versailles
Etienne M (2003) SYLVOPAST a multiple target role-playing game to assess negotiation
processes in sylvopastoral management planning. Journal of Artificial Societies and Social
Simulations 6, http://jasss.soc.surrey.ac.uk/6/2/5.html
Étienne M, Le Page C, Cohen M (2003) A step by step approach to build up land management
scenarios based on multiple viewpoints on multi-agent systems simulations. J Artif Soc Soc
Simul 6(2), http://jasss.soc.surrey.ac.uk/6/2/2.html
Eversole R (2003) Managing the pitfalls of participatory development: some insight from
Australia. World Dev 31:781–795
Fischer G, Giaccardi E, Eden H, Sugimoti M, Ye Y (2005) Beyond binary choices: integrating
individual and social creativity. Int J Hum-Comput Stud 63:482–512
Friedman D, Sunder S (1994) Experimental methods, a primer for economists. Cambridge
University Press, Cambridge
Fung A (2006) Varieties of participation in complex governance. Public Adm Rev 66:66–75
Funtowicz SO, Martinez-Alier J, Munda G, Ravetz JR (1999) Information tools for environmental
policy under conditions of complexity, vol 9, Environmental issues series. European Environ-
ment Agency, Copenhagen
Buckingham
Green KC (2002) Forecasting decisions in conflict situations: a comparison of game theory, role-
playing and unaided judgment. Int J Forecast 18:321–344
Guyot P (2006) Simulations multi-agents participatives. PhD thesis, Informatique, Université
Paris VI, Paris
Guyot P, Honiden S (2006) Agent-based participatory simulations: merging multi-agent systems
and role-playing games. J Artif Soc Soc Simul 9(4), http://jasss.soc.surrey.ac.uk/9/4/8.html
Hamel A, Pinson S (2005) Conception participative de simulations multi-agents basée sur une
approche d’analyse multi-acteurs. In: Drogoul A, Ramat E (eds) Systèmes multi-agents: Vers
la conception de systèmes artificiels socio-mimétiques (JFSMA 2005). Hermès, Paris, pp 1–15
Hanneman RA (1995) Simulation modeling and theoretical analysis in sociology. Sociol Perspect
38:457–462
Henriksen HJ, Rasmussen P, Brandt G, von Bülow D, Jensen FV (2004) Engaging stakeholders in
construction and validation of Bayesian belief networks for groundwater protection.
In: Proceedings of IFAC workshop on modelling and control for participatory planning and
managing water systems, Venice, 29 Sept–1 Oct 2004
Henry C (1974) Investment decisions under uncertainty: the irreversibility effect. Am Econ Rev
64:1006–1012
Hochman Z, Hearnshaw H, Barlow R, Ayres JF, Pearson CJ (1995) X-breed: a multiple domain
knowledge based system integrated through a blackboard architecture. Agr Syst 48:243–270
Irvin RA, Stansbury J (2004) Citizen participation in decision making: is it worth the effort. Public
Adm Rev 64:55–65
Johannessen J-A, Olaisen J, Olsen B (2001) Mismanagement of tacit knowledge: the importance
of tacit knowledge, the danger of information technology, and what to do about it. Int J Inf
Manag 21:3–20
Kinzig A, Ryan P, Etienne M, Allyson H, Elmqvist T, Walker B (2006) Resilience and regime
shifts: assessing cascading effects. Ecol Soc 11(1):20, http://www.ecologyandsociety.org/
vol11/iss1/art20
Klopfer E, Yoon S, Rivas L (2004) Comparative analysis of palm and wearable computers for
participatory simulations. J Comput Assist Lear 20:347–359
Kujala S (2003) User involvement: a review of the benefits and challenges. Behav Inf Technol
22:1–16
Landry M, Banville C, Oral M (1996) Model legitimation in operational research. Eur J Oper Res
92:443–457
Le Bars M, Le Grusse P, Allaya M, Attonaty J-M, Mahjoubi R (2004) NECC: Un jeu de simulation
pour l’aide à la décision collective; Application à une région méditerranéenne virtuelle.
In: Projet INCO-WADEMED, Séminaire Modernisation de l’Agriculture Irriguée, Rabat
Loucks DP, Kindler J, Fedra K (1985) Interactive water resources modeling and model use: an
overview. Water Resour Res 21:95–102
Maier FH, Grössler A (2000) What are we talking about? A taxonomy of computer simulations to
support learning. Syst Dynam Rev 16:135–148
Manson SM (2002) Validation and verification of multi-agent systems. In: Janssen M (ed)
Complexity and ecosystem management: the theory and practice of multi-agent approaches.
Edward Elgar, Northampton, pp 63–74
Marengo L, Pasquali C (2003) How to construct and share a meaning for social interactions?
In Conventions et Institutions: Approfondissements théoriques et Contributions au Débat
Politique, Paris
Martin L et al (2007) Microworld gaming of a local agricultural production chain in Poland. Simul
Gaming 38(2):211–232
Mayer IS, van Bueren EM, Bots PWG, van der Voort HG, Seijdel RR (2005) Collaborative
decision-making for sustainable urban renewal projects: a simulation-gaming approach. Envi-
ron Plan B Plan Des 32:403–423
McKinnon J (2005) Mobile interactive GIS: bringing indigenous knowledge and scientific infor-
mation together; a narrative account. In: Neef A (ed) Participatory approaches for sustainable
land use in Southeast Asia. White Lotus, Bangkok, pp 217–231
Mermet L (1993) Une méthode de prospective: Les exercices de simulation de politiques. Nature
Sciences Sociétés 1:34–46
Miettinen R, Virkkunen J (2005) Epistemic objects, artefacts and organizational change. Organi-
zation 12:437–456
Molin E (2005) A causal analysis of hydrogen acceptance. Transport Res Rec 1941:115–121
Moss S, Downing T, Rouchier J (2000) Demonstrating the role of stakeholder participation: an
agent based social simulation model of water demand policy and response (CPM report,
00–76). Centre for Policy Modelling, Manchester Metropolitan University, Manchester,
http://cfpm.org/cpmrep76.html
Mostert E (2006) Participation for sustainable water management. In: Giupponi C, Jakeman AJ,
Karssenberg D, Hare MP (eds) Sustainable management of water resources. Edward Elgar,
Northampton, pp 153–176
Nancarrow B (2005) When the modeller meets the social scientist or vice-versa. In: Zerger A,
Argent RM (eds) MODSIM 2005 international congress on modelling and simulation.
Modelling and Simulation Society of Australia and New Zealand, Melbourne, pp 38–44
Ostrom E, Gardner R, Walker J (1994) Rules, games and common-pool resources. The University
of Michigan Press, Ann Arbor
Pahl-Wostl C, Hare M (2004) Processes of social learning in integrated resources management.
J Community Appl Soc Psychol 14:193–206
Pateman C (1990) Participation and democratic theory. Cambridge University Press, Cambridge
Perez P, Dray A, Le Page C, D’Aquino P, White I (2004) Lagoon, agents and kava: a companion
modelling experience in the Pacific. In: van Dijkum C, Blasius J, Durand C (eds) Recent
developments and applications in social research methodology: Proceedings of RC33 sixth
international conference on social science methodology, Amsterdam Barbara Budrich
Publishers, Opladen, p 282
Ramanath AM, Gilbert N (2004) The design of participatory agent based simulations. J Artif Soc
Soc Simul 7(4), http://jasss.soc.surrey.ac.uk/7/4/1.html
Reitsma R, Zigurs I, Lewis C, Wilson V, Sloane A (1996) Experiment with simulation models in
water-resources negotiations. J Water Resour Plan Manag 122:64–70
Richard A, Barreteau O (2006) Concert’eau: un outil de sociologie expérimentale pour l’étude de
dispositifs de gestion locale et concertée de l’eau. In: Proceedings of 2e Congrès de l’Associa-
tion Française de Sociologie, Bordeaux, 5–8 Sept 2006
Richard A, Trometter M (2001) Les caractéristiques d’une décision séquentielle: Effet

irréversibilité et endogénéisation de l’environnement. Revue Economique 52:739–752
Rittel HWJ, Webber MM (1973) Dilemmas in a general theory of planning. Policy Sci 4:155–169
Robinson JB (1991) Modelling the interactions between human and natural systems. Int Soc Sci J
130:629–647
Rocha EM (1997) A ladder of empowerment. J Plan Edu Res 17:31–44
Rouchier J (2003) Re-implementation of a multi-agent model aimed at sustaining experimental
economic research: the case of simulations with emerging speculation. J Artif Soc Soc Simul
6(4), http://jasss.soc.surrey.ac.uk/6/4/7.html
Rouwette EAJA, Vennix JAM, van Mullekorn T (2002) Group model building effectiveness:
a review of assessment studies. Syst Dyn Rev 18:5–45
Rowe G, Frewer LJ (2004) Evaluating public-participation exercises: a research agenda.
Sci Technol Hum Val 29(4):512–556
Ryan T (2000) The role of simulation gaming in policy making. Syst Res Behav Sci 17:359–364
Schelling TC (1961) Experimental games and bargaining theory. World Polit 14:47–68
Schuler D, Namioka A (eds) (1993) Participatory design: principles and practices. Lawrence
Erlbaum, Hillsdale
Shakun ME (1996) Modeling and supporting task-oriented group processes: purposeful complex
adaptive systems and evolutionary systems design. Group Decis Negot 5:305–317
Simon H (1977) The new science of management decision. Prentice-Hall, Englewood Cliffs
Star SL, Griesemer JR (1989) Institutional ecology, ‘translations’ and boundary objects: amateurs
and professionals in Berleley’s museum of vertebrate zoology, 1907–39. Soc Stud Sci
19:387–420
Sterman JD (1992) Teaching takes off – flight simulators for management education. OR/MS
Today, pp 40–44
Ticehurst J, Rissik D, Letcher RA, Newham LTH, Jakeman AJ (2005) Development of decision
support tools to assess the sustainability of coastal lakes. In: Zerger A, Argent RM (eds)
MODSIM 2005 international congress on modelling and simulation, Modelling and Simula-
tion Society of Australia and New Zealand, Melbourne, pp 2414–2420, http://www.mssanz.
org.au/modsim05/papers/ticehurst.pdf
Toth FL (1988) Policy exercises: objectives and design elements. Simul Games 19:235–255
van Asselt M et al (2001) Building blocks for participation in integrated assessment: a review of
participatory methods. ICIS, Maastricht
van Daalen CE, Thissen WAH, Berk MM (1998) The Delft process: experiences with a dialogue
between policy makers and global modellers. In: Alcamo J, Leemans R, Kreileman E (eds)
Global change scenarios of the 21st century: results from the IMAGE 2.1 model. Elsevier,
London, pp 267–285
Van den Belt M (2004) Mediated modeling: a system dynamics approach to environmental
consensus building. Island Press, Washington, DC
Vennix JAM (1996) Group model building, facilitating team learning using system dynamics.
Wiley, Chichester
Webler T, Kastenholz H, Renn O (1995) Public participation in impact assessment: a social
learning perspective. Environ Impact Assess Rev 15:443–463
Whitworth B, Gallupe B, McQueen R (2000) A cognitive three-process model of computer
mediated group interaction. Gr Decis Negot 9:431–456
Wood J, Silver D (1995) Joint application development. Wiley, New York
Chapter 11
Combining Mathematical and Simulation
Approaches to Understand the Dynamics
of Computer Models
Luis R. Izquierdo, Segismundo S. Izquierdo, José M. Galán,

and José I. Santos
Why Read This Chapter? To learn how to better understand the dynamics of
computer models using both simulation and mathematical analysis. Our starting
point is a computer model which is already implemented and ready to be run; the
objective is to gain a thorough understanding of its dynamics. Combining computer
simulation with mathematical analysis can help to provide a picture of the model
dynamics that could not be drawn by using only one of the two techniques.
Abstract This chapter shows how computer simulation and mathematical analysis
can be used together to understand the dynamics of computer models. For this
purpose, we show that it is useful to see the computer model as a particular
implementation of a formal model in a certain programming language. This formal
model is the abstract entity which is defined by the input-output relation that the
computer model executes, and can be seen as a function that transforms probability
distributions over the set of possible inputs into probability distributions over the set
of possible outputs.
It is shown here that both computer simulation and mathematical analysis are
extremely useful tools to analyse this formal model, and they are certainly comple-
mentary in the sense that they can provide fundamentally different insights on the
same model. Even more importantly, this chapter shows that there are plenty of
synergies to be exploited by using the two techniques together.
The mathematical approach to analyse formal models consists in examining the rules
that define the model directly. Its aim is to deduce the logical implications of these rules
for any particular instance to which they can be applied. Our analysis of mathematical
L.R. Izquierdo (*) • J.M. Galán • J.I. Santos

Departamento de Ingenierı́a Civil, Universidad de Burgos, E-09001, Burgos, Spain
S.S. Izquierdo
Departamento de Organización de Empresas y C.I.M., Universidad de Valladolid, E-47011,
Valladolid, Spain

236 L.R. Izquierdo et al.
techniques to study formal models is focused on the theory of Markov Chains, which is
particularly useful to characterise the dynamics of computer models.
In contrast with mathematical analysis, the computer simulation approach does
not look at the rules that define the formal model directly, but instead tries to infer
general properties of these rules by examining the outputs they produce when
applied to particular instances of the input space. Thus, conclusions obtained with
this approach may not be general. On a more positive note, computer simulation
enables us to explore formal models beyond mathematical tractability, and we can
achieve any arbitrary level of accuracy in our computational approximations by
running the model sufficiently many times.
Bearing in mind the relative strengths and limitations of both approaches, this
chapter explains three different ways in which mathematical analysis and computer
simulation can be usefully combined to produce a better understanding of the
dynamics of computer models. In doing so, it becomes clear that mathematical
analysis and computer simulation should not be regarded as alternative – or even
opposed – approaches to the formal study of social systems, but as complementary.
Not only can they provide fundamentally different insights on the same model, but
they can also produce hints for solutions for each other. In short, there are plenty of
synergies to be exploited by using the two techniques together, so the full potential
of each technique cannot be reached unless they are used in conjunction.
11.1 Introduction
This chapter is about how to better understand the dynamics of computer models
using both simulation and mathematical analysis. Our starting point is a computer
model which is already implemented and ready to be run; our objective is to gain a
thorough understanding of its dynamics. Thus, this chapter is not about how to
design, implement, verify, or validate a model; this chapter is about how to better
understand its behaviour.
Naturally, we start by clearly defining our object of study: a computer model.
The term ‘computer model’ can be understood in many different ways – i.e. seen
from many different perspectives –, and not all of them are equally useful for every
possible purpose. Thus, we start by interpreting the term ‘computer model’ in a way
that will prove useful for our objective: to characterise and understand its
behaviour. Once our object of study has been clearly defined, we then describe
two techniques that are particularly useful to understand the dynamics of computer
models: mathematical analysis and computer simulation.
In particular, this chapter will show that mathematical analysis and computer
simulation should not be regarded as alternative – or even opposed – approaches to
the formal study of social systems, but as complementary (Gotts et al. 2003a, b).
They are both extremely useful tools to analyse formal models, and they are
certainly complementary in the sense that they can provide fundamentally different
insights on the same model. Even more importantly, this chapter will show that
11 Combining Mathematical and Simulation Approaches to Understand the. . . 237
there are plenty of synergies to be exploited by using the two techniques together,
i.e. the full potential of each technique will not be reached until they are used in
conjunction. The remaining of this introduction outlines the structure of the chapter.
Sections 11.2, 11.3 and 11.4 are devoted to explaining in detail what we
understand by ‘computer model’, and they therefore provide the basic framework
for the rest of the chapter. In particular, Sect. 11.2 shows that a computer model can
be seen as an implementation – i.e. an explicit representation – of a certain
deterministic input-output function in a particular programming language. This
interpretation is very useful since, in particular, it will allow us to abstract from
the details of the modelling platform where the computer model has been
programmed, and focus on analysing the formal model that the computer model
implements. This is clarified in Sect. 11.3, which explains that any computer model
can be re-implemented in many different formalisms (in particular, in any sophisti-
cated enough programming language), leading to alternative representations of the
same input-output relation.
Most computer models in the Social Simulation literature make use of pseudo-
random number generators. Section 11.4 explains that – for these cases and given
our purposes – it is useful to abstract from the details of how pseudo-random
numbers are generated, and look at the computer model as an implementation of
a stochastic process. In a stochastic process, a certain input does not necessarily
lead to one certain output only; instead, there are many different paths that the
process may take with potentially different probabilities. Thus, in a stochastic
process a certain input will generally lead to a particular probability distribution
over the range of possible outputs, rather than to a single output only. Stochastic
processes are used to formally describe how a system subjected to random events
evolves through time.
Having explained our interpretation of the term ‘computer model’, Sect. 11.5
introduces and compares the two techniques to analyse formal models that are
assessed in this chapter: computer simulation and mathematical analysis. The
following two sections sketch possible ways in which each of these two techniques
can be used to obtain useful insights about the dynamics of a model. Section 11.8 is
then focused on the joint use of computer simulation and mathematical analysis.
It is shown here that the two techniques can be used together to provide a picture of
the dynamics of the model that could not be drawn by using one of the two
techniques only. Finally, our conclusions are summarised in Sect. 11.9.
11.2 Computer Models as Input-Output Functions
At the most elementary level, a computer model can be seen as an implementation –

i.e. an explicit representation – of a certain deterministic input-output function in a
particular programming language. The word ‘function’ is useful because it
correctly conveys the point that any particular input given to the computer model
will lead to one and only one output.1 (Obviously, different inputs may lead to the
same output.) Admittedly, however, the word ‘function’ may also mislead the
reader into thinking that a computer model is necessarily simple. The computer
model may be as complex and sophisticated as the programmer wants it to be but,
ultimately, it is just an entity that associates a specific output to any given input,
i.e. a function. In any case, to avoid confusion, we will use the term ‘formal model’
to denote the function that a certain computer model implements.2 To be sure, the
‘formal model’ that a particular computer model implements is the abstract entity
which is defined by the input-output relation that the computer model executes.3
Thus, running a computer model is just finding out the logical implications of
applying a set of unambiguously defined formal rules (which are coded in the
program and define the input-output function or formal model) to a set of inputs
(Balzer et al. 2001). As an example, one could write the computer program
“y ¼ 4x” and apply it to the input “x ¼ 2” to obtain the output “y ¼ 8”. The
output (y ¼ 8), which is fully and unequivocally determined by the input (x ¼ 2)
and the set of rules coded in the program (y ¼ 4x), can be seen as a theorem
obtained by pure deduction ({x ¼ 2; y ¼ 4x} ) y ¼ 8). Naturally, there is no
reason why the inputs or the outputs should be numbers4; they could equally well be
e.g. strings of characters. In the general case, a computer run is a logical theorem
that reads: the output obtained from running the computer simulation follows (with
logical necessity) from applying to the input the algorithmic rules that define the
model. Thus, regardless of its inherent complexity, a computer run constitutes a
perfectly valid sufficiency theorem (see e.g. Axtell 2000).
It is useful to realise that we could always apply the same inference rules
ourselves to obtain – by logical deduction – the same output from the given
input. While useful as a thought, when it comes to actually doing the job, it is
much more convenient, efficient and less prone to errors to let computers derive the
output for us. Computers are inference engines that are able to conduct many
algorithmic processes at a speed that the human brain cannot achieve.
1
Note that simulations of stochastic models are actually using pseudo-random number generators,
which are deterministic algorithms that require a seed as an input.
2
A formal model is a model expressed in a formal system (Cutland 1980). A formal system
consists of a formal language and a deductive apparatus (a set of axioms and inference rules).
Formal systems are used to derive new expressions by applying the inference rules to the axioms
and/or previously derived expressions in the same system.
3
The mere fact that the model has been implemented and can be run in a computer is a proof that
the model is formal (Suber 2002).
4
As a matter of fact, strictly speaking, inputs and outputs in a computer model are never numbers.
We may interpret strings of bits as numbers, but we could equally well interpret the same strings of
bits as e.g. letters. More importantly, a bit itself is already an abstraction, an interpretation we
make of an electrical pulse that can be above or below a critical voltage threshold.
11.3 Different Ways of Representing the Same Formal Model
A somewhat controversial issue in the Social Simulation literature refers to the

allegedly unique features of some modelling platforms. It is important to realise that
any formal model implemented in a computer model can be re-implemented in many
different programming languages, leading to exactly the same input-output relation.
Different implementations are just different ways of representing one same formal
model, much in the same way that one can say ‘Spain’ or ‘España’ to express the
same concept in different languages: same thing, different representation, that’s all.
Thus, when analysing the dynamics of a computer model, it is useful to abstract
from the details of the modelling platform that has been used to implement the
computer model, and focus strictly on the formal model it represents, which could
be re-implemented in any sophisticated enough5 modelling platform. To be clear,
let us emphasise that any computer model implemented in Objective-C (e.g. using
Swarm) can be re-implemented in Java (e.g. using RePast or Mason), NetLogo,
SDML, Mathematica# or Matlab#. Similarly, any computer model can be
expressed as a well-defined mathematical function (Epstein 2006; Leombruni and
Richiardi 2005; Richiardi et al. 2006).
Naturally, the implementation of a particular formal model may be more
straightforward in some programming languages than in others. Programming
languages differ in where they position themselves in the well-known trade-offs
between ease of programming, functionality and performance; thus, different
programming languages lead to more or less natural and more or less efficient
implementations of any given formal model. Nonetheless, the important point is
this: whilst we may have different implementations of the same formal model, and
whilst each of these implementations may have different characteristics (in terms of
e.g. code readability), ultimately they are all just different representations of the
same formal model, and they will therefore return the same output when given the
same input.
In the same way that using one or another formalism to represent a particular
formal model will lead to more or less natural implementations, different formalisms
also make more or less apparent certain properties of the formal model they
implement. For example, we will see in this chapter that representing a computer
model as a Markov chain, i.e. looking at the formal model implemented in a
5
A sufficient condition for a programming language to be “sophisticated enough” is to allow for
the implementation of the following three control structures:
• Sequence (i.e. executing one subprogram, and then another subprogram),
• Selection (i.e. executing one of two subprograms according to the value of a boolean variable,
e.g. IF[boolean ¼¼ true]-THEN[subprogram1]-ELSE[subprogram2]), and
• Iteration (i.e. executing a subprogram until a boolean variable becomes false, e.g. WHILE
[boolean ¼¼ true]-DO[subprogram]).
Any programming language that can combine subprograms in these three ways can implement
any computable function; this statement is known as the “structured program theorem” (Böhm and
Jacopini 1966; Harel 1980; Wikipedia 2007).
computer model through Markov’s glasses, can make apparent various features of
the computer model that may not be so evident without such glasses. In particular, as
we will show later, Markov theory can be used to find out whether the initial
conditions of a model determine its asymptotic dynamics or whether they are
actually irrelevant in the long term. Also, the theory can reveal whether the model
will sooner or later be trapped in an absorbing state.
11.4 ‘Stochastic’ Computer Models as Stochastic Processes
Most computer models in the Social Simulation literature contain stochastic

components. This section argues that, for these cases and given our purposes, it is
convenient to revise our interpretation of computer models as deterministic input-
output relations, abstract from the (deterministic) details of how pseudo-random
numbers are generated, and reinterpret the term ‘computer model’ as an implemen-
tation of a stochastic process. This interpretation will prove useful in most cases
and, importantly, does not imply any loss of generality: even if the computer model
to be analysed does not contain any stochastic components, our interpretation will
still be valid.
In the general case, the computer model to be analysed will make use of (what
are meant to be) random numbers, i.e. the model will be stochastic. The word
‘stochastic’ requires some clarification. Strictly speaking, there does not exist a
truly stochastic computer model, but one can approximate randomness to a very
satisfactory extent by using pseudo-random number generators. The pseudo-
random number generator is a deterministic algorithm that takes as input a value
called the random seed, and generates a sequence of numbers that approximates the
properties of random numbers. The sequence is not truly random in that it is
completely determined by the value used to initialise the algorithm, i.e. the random
seed. Therefore, if given the same random seed, the pseudo-random number
generator will produce exactly the same sequence of (pseudo-random) numbers.
(This fact is what made us define a computer model as an implementation of a
certain deterministic input-output function in Sect. 11.2.)
Fortunately, the sequences of numbers provided by current off-the-shelf pseudo-
random number generators approximate randomness remarkably well. This basically
means that, for most intents and purposes in this discipline, it seems safe to assume
that the pseudo-random numbers generated in one simulation run will follow the
intended probability distributions to a satisfactory degree. The only problem we
might encounter appears when running several simulations which we would like to
be statistically independent. As mentioned above, if we used the same random seed
for every run, we would obtain the same sequence of pseudo-random numbers, i.e. we
would obtain exactly the same results. How can we truly randomly select a random
seed? Fortunately, for most applications in this discipline, the state of the computer
system at the time of starting a new run can be considered a truly random variable;
and, conveniently, if no seed is explicitly provided to the pseudo-random number
INPUT OUTPUT
SPACE SPACE
COMPUTER
MODEL
Pseudorandom
number
generator
Fig. 11.1 A computer model can be usefully seen as the implementation of a function that
transforms any given input into a certain probability distribution over the set of possible outputs
generator, most platforms generate a seed from the state of the computer system
(e.g. using the time). When this is done, the sequences of numbers obtained with
readily available pseudo-random number generators approximate statistical
randomness and independence remarkably well.
Given that – for most intents and purposes in this discipline – we can safely
assume that pseudo-random numbers are random and independent enough, we
dispense with the qualifier ‘pseudo’ from now on for convenience. Since every
random variable in the model follows a specific probability distribution, the
computer model will indeed generate a particular probability distribution over the
range of possible outputs. Thus, to summarise, a computer model can be usefully
seen as the implementation of a stochastic process, i.e. a function that transforms
any given input into a certain probability distribution over the set of possible
outputs (Fig. 11.1).
Seeing that we can satisfactorily simulate random variables, note that studying
the behaviour of a model that has been parameterised stochastically does not
introduce any conceptual difficulties. In other words, we can study the behaviour
of a model that has been parameterised with probability distributions rather than
certain values. An example would be a model where agents start at a random initial
location.
To conclude this section, let us emphasise an important corollary of the previous
paragraphs: any statistic that we extract from a parameterised computer model
follows a specific probability distribution (even if the values of the input
parameters have been expressed as probability distributions).6 Thus, a computer
model can be seen as the implementation of a function that transforms probability
distributions over the set of possible inputs into probability distributions over the set
of possible outputs (Fig. 11.2). The rest of the chapter is devoted to characterising
this function.
6
Note that statistics extracted from the model can be of any nature, as long as they are unambigu-
ously defined. For example, they can refer to various time-steps, and only to certain agents (e.g.
“average wealth of female agents in odd time-steps from 1 to 99”).
INPUT OUTPUT
SPACE SPACE
COMPUTER
MODEL
Pseudorandom
number
generator
Fig. 11.2 A computer model can be seen as the implementation of a function that transforms
probability distributions over the set of possible inputs into probability distributions over the set of
possible outputs
11.5 Tools to Understand the Behaviour of Formal Models
Once it is settled that a computer model can be seen as a particular implementation of a

(potentially stochastic) function in a certain programming language, let us refer to such
a function as the ‘formal model’ that the computer model implements. As mentioned
before, this formal model can be expressed in many different formalisms – in particular,
it can always be expressed as a set of well defined mathematical equations (Leombruni
and Richiardi 2005) – and our objective consists in understanding its behaviour. To do
that, we count with two very useful tools: mathematical analysis7 and computer
simulation.
The advantages and limitations of these two tools to formally study social
systems have been discussed at length in the literature (see e.g. Axtell 2000; Axtell
and Epstein 1994; Edmonds 2005; Gilbert 1999; Gilbert and Troitzsch 1999; Gotts
et al. 2003a; Holland and Miller 1991; Ostrom 1988). Here we only highlight the
most prominent differences between these two techniques (see Fig. 11.3).
In broad terms, when using mathematical analysis, one examines the rules that
define the formal model directly, and tries to draw general conclusions about these
rules. These conclusions are obtained by using logical deduction; hence they follow
with logical necessity from the premises of the formal model (and the axioms of the
mathematics employed). The aim when using mathematical analysis is usually to
“solve” the formal system (or, most often, certain aspects of it) by producing
general closed-form solutions that can be applied to any instance of the whole
input set (or, at least, to large portions of the input set). Since the inferences
obtained with mathematical analysis pertain to the rules themselves, such
inferences can be safely particularised to any specific parameterisation of the
model, even if such a parameterisation was never explicitly contemplated when
analysing the model mathematically. This greatly facilitates conducting sensitivity
analyses and assessing the robustness of the model.
Computer simulation is a rather different approach to the characterisation of the
formal model (Epstein 2006; Axelrod 1997a). When using computer simulation,
7
We use the term “mathematical analysis” in its broadest sense, i.e. we do not refer to any
particular branch of mathematics, but to the general use of (any type of) mathematical technique
to analyse a system.
Mathematical
Analysis
INPUT OUTPUT
SPACE SPACE
COMPUTER
MODEL
Pseudorandom
number
… … ……
generator
Computer
Simulation
Fig. 11.3 In general terms, mathematical analysis tends to examine the rules that define the
formal model directly. In contrast, computer simulation tries to infer general properties of such
rules by looking at the outputs they produce when applied to particular instances of the input space
one often treats the formal model as a black box, i.e. a somewhat obscure abstract
entity that returns certain outputs when provided with inputs. Thus, the path to
understand the behaviour of the model consists in obtaining many input-output
pairs and – using generalisation by induction – inferring general patterns about how
the rules transform the inputs into the outputs (i.e. how the formal model works).
Importantly, the execution of a simulation run, i.e. the logical process that
transforms any (potentially stochastic) given input into its corresponding (poten-
tially stochastic) output is pure deduction (i.e. strict application of the formal rules
that define the model). Thus, running the model in a computer provides a formal
proof that a particular input (together with the set of rules that define the model) is
sufficient to generate the output that is observed during the simulation. This first part
of the computer simulation approach is therefore, in a way, very “mathematical”:
outputs obtained follow with logical necessity from applying to the inputs the
algorithmic rules that define the model.
In contrast, the second part of the computer simulation approach, i.e. inferring
general patterns from particular instances of input-output pairs, can only lead to
probable – rather than necessarily true – conclusions.8 The following section
explains how to rigorously assess the confidence we can place on the conclusions
obtained using computer simulation, but the simple truth is irrefutable: inferences
obtained using generalisation by induction can potentially fail when applied to
instances that were not used to infer the general pattern. This is the domain of
statistical extrapolation.
8
Unless, of course, all possible particular instances are explored.
So why bother with computer simulation at all? The answer is clear: computer
simulation enables us to study formal systems in ways that go beyond mathematical
tractability. This role should not be underestimated: most models in the Social
Simulation literature are mathematically intractable, and in such cases computer
simulation is our only chance to move things forward. As a matter of fact, the
formal models that many computer programs implement are often so complicated
and cumbersome that the computer code itself is not that far from being one of the
best descriptions of the formal model that can be provided.
Computer simulation can be very useful even when dealing with formal models
that are mathematically tractable. Valuable uses of computer simulation in these
cases include conducting insightful initial explorations of the model and presenting
dynamic illustrations of its results.
And there is yet another important use of computer simulation. Note that
understanding a formal model in depth requires identifying the parts of the model
(i.e. the subset of rules) that are responsible for generating particular (sub)sets of
results or properties of results. Investigating this in detail often involves changing
certain subsets of rules in the model, so one can pinpoint which subsets of rules are
necessary or sufficient to produce certain results. Importantly, changing subsets
of rules can make the original model mathematically intractable and in such
(common) cases, computer simulation is, again, our only hope. In this context,
computer simulation can be very useful to produce counter-examples. This
approach is very common in the literature of e.g. evolutionary game theory,
where several authors (see e.g. Hauert and Doebeli 2004; Imhof et al. 2005;
Izquierdo and Izquierdo 2006; Lieberman et al. 2009; Nowak and May 1992;
Nowak and Sigmund 1992, 1993; Santos et al. 2006; Traulsen et al. 2006) resort
to computer simulations to assess the implications of assumptions made in mathe-
matically tractable models (e.g. the assumptions of “infinite populations” and
“random encounters”).
It is important to note that the fundamental distinction between mathematical
analysis and computer simulation as presented here is not about whether one uses
pen and paper or computers to analyse formal models. We can follow either approach
with or without computers, and it is increasingly popular to do mathematical analysis
with computers. Recent advancements in symbolic computation have opened up
a new world of possibilities to conduct mathematical analyses (using e.g.
Mathematica#). In other words, nowadays it is perfectly possible to use computers
to directly examine the rules that define a formal model (see Fig. 11.3).
Finally, as so often in life, things are not black or white, but involve some shade
of grey. Similarly, most models are not tractable or intractable in mathematical
terms; most often they are partially tractable. It is in these cases where an adequate
combination of mathematical analysis and computer simulation is particularly
useful. We illustrate this fact in Sect. 11.8, but first let us look at each technique
separately. The following two sections provide some guidelines on how computer
simulation (Sect. 11.6) and mathematical analysis (Sect. 11.7) can be usefully
employed to analyse formal models.
11.6 Computer Simulation: Approximating the Exact

Probability Distribution by Running the Model
The previous sections have argued that any statistic obtained from a (stochastically or
deterministically) parameterised model follows a specific probability distribution.
The statistic could be anything as long as it is unambiguously defined; in particular, it
could refer to one or several time-steps, and to one or various subcomponents of the
model. Ideally, one would like to calculate the exact probability distribution for the
statistic using mathematical analysis, but this will not always be possible. In contrast,
using computer simulation we will always be able to approximate this probability
distribution to any arbitrary level of accuracy; this section provides basic guidelines
on how to do that.
The output probability distribution – which is fully and unequivocally deter-
mined by the input distribution – can be approximated to any degree of accuracy by
running enough simulation runs. Note that any specific simulation run will be
conducted with a particular certain value for every parameter (e.g. a particular
initial location for every agent), and will produce one and only one particular
certain output (see Fig. 11.3). Thus, in order to infer the probability distribution
over the set of outputs that a particular probability distribution over the set of inputs
leads to, there will be a need to run the model many times (with different random
seeds); this is the so-called Monte Carlo method.
The method is straightforward: obtain as many random samples as possible (i.e.
run as many independent simulations as possible), since this will get us closer and
closer to the exact distribution (by the law of large numbers). Having conducted a
large number of simulation runs, the question that naturally comes to mind is: How
close to the exact distribution is the one obtained by simulation?
To illustrate how to assess the quality of the approximation obtained by simulation,
we use CoolWorld, a purpose-built agent-based model (Gilbert 2007) implemented in
NetLogo 4.0 (Wilensky 1999). A full description of the model, an applet and the
source code can be found at the dedicated model webpage http://luis.izquierdo.name/
models/coolworld. For our purposes, it suffices to say that in CoolWorld there is
a population of agents called walkers, who wander around a 2-dimensional grid
made of square patches; some of the patches are empty whilst others contain a
house (see Fig. 11.4). Patches are at a certain predefined temperature, and walkers
tend to walk towards warmer patches, staying for a while at the houses they encounter
in their journey.
Let us assume that we are interested in studying the number of CoolWorld
walkers staying in a house in time-step 50. Initial conditions (which involve 100
walkers placed at a random location) are unambiguously defined at the model
webpage and can be set in the implementation of CoolWorld provided by clicking
on the button “Special conditions”. Figure 11.4 shows a snapshot of CoolWorld
after having clicked on that button.
As argued before, given that the (stochastic) initial conditions are unambiguously
defined, the number of CoolWorld walkers in a house after 50 time-steps will follow
Fig. 11.4 Snapshot of CoolWorld. Patches are coloured according to their temperature: the higher
the temperature, the darker the shade of grey. Houses are coloured in black, and form a circle
around the central patch. Walkers are coloured in grey, and represented as a person if standing on a
patch without a house, and as a smiling face if standing on a patch with a house. In the latter case,
the white label indicates the number of walkers in the same house
a specific probability distribution that we are aiming to approximate. For that, let us
assume that we run 200 runs, and plot the relative frequency of the number of
walkers in a patch with a house after 50 time-steps (see Fig. 11.5).
Figure 11.5 does not provide all the information that can be extracted from the
data gathered. In particular, we can plot error bars showing the standard error for
each calculated frequency without hardly any effort.9 Standard errors give us
information about the error we may be incurring when estimating the exact
probabilities with the empirical frequencies. Another simple task that can be
conducted consists in partitioning the set of runs into two batteries of approximately
9
The frequency of the event “there are i walkers in a patch with a house” calculated over n
simulation runs can be seen as the mean of a sample of n i.i.d. Bernouilli random variables where
success denotes that the event occurred and failure denotes that it did not. Thus, the frequency
f is the maximum likelihood (unbiased) estimator of the exact probability with which the event
occurs. The standard error of the calculated frequency f is the standard deviation of the sample
divided by the square root of the sample size. In this particular case, the formula reads:
Std: error ð f ; nÞ ¼ ð f ð1 f Þ=ðn 1ÞÞ1=2
Where f is the frequency of the event, n is the number of samples, and the standard deviation of
the sample has been calculated dividing by (n – 1).
0.14
200 simulation runs
0.12
Empirical relative
Relative frequency frequency distribution
0.10
0.08
0.06
0.04
0.02
0.00
75 80 85 90 95 100
Number of walkers in a house after 50 time-steps
Fig. 11.5 Relative frequency distribution of the number of walkers in a house after 50 time-steps,
obtained by running CoolWorld 200 times, with initial conditions set by clicking on “Special
conditions”
0.20
Empirical relative 100 simulation runs
frequency distribution Battery A
Error bars
0.15
(std. error)
Relative frequency
Exact
0.10
prob. function
0.05
75 80 85 90 95 100
Fig. 11.6 In darker grey: Relative frequency distribution of the number of walkers in a house after
50 time-steps, obtained by running CoolWorld 100 times (Battery A), with initial conditions set by
clicking on “Special conditions”. In lighter grey: Exact probability distribution (Calculated using
Markov chain analysis)
equal size and comparing the two distributions. If the two distributions are
not similar, then there is no point in proceeding: we are not close to the exact
distribution, so there is a need to run more simulations.
Figures 11.6 and 11.7 show the data displayed in Fig. 11.5 partitioned in two
batteries of 100 simulation runs, including the standard errors. Figures 11.6 and 11.7
also show the exact probability distribution we are trying to approximate, which has
been calculated using mathematical methods that are explained later in this chapter.
0.20
Empirical relative 100 simulation runs
frequency distribution Battery B
0.15 Error bars

(std. error)
Relative frequency
Exact
0.10 prob. function
0.05
75 80 85 90 95 100
50 time-steps, obtained by running CoolWorld 100 times (Battery B), with initial conditions set by
clicking on “Special conditions”. In lighter grey: Exact probability distribution (Calculated using
Markov chain analysis)
0.20
Empirical relative 50 000 simulation runs
frequency distribution Battery A
0.15
Error bars
(std. error)
Relative frequency
Exact
0.10 prob. function
0.05
75 80 85 90 95 100
50 time-steps, obtained by running CoolWorld 50,000 times (Battery A), with initial conditions set
by clicking on “Special conditions”. In lighter grey: Exact probability distribution (Calculated
using Markov chain analysis)
Figures 11.6 and 11.7 indicate that 100 simulation runs may not be enough to
obtain a satisfactory approximation to the exact probability distribution. On the
other hand, Figs. 11.8 and 11.9 show that running the model 50,000 times does
seem to get us close to the exact probability distribution. The standard error, which
0.20

frequency distribution Battery B
0.15 Error bars
Relative frequency
(std. error)
Exact
0.10 prob. function
0.05
75 80 85 90 95 100
50 time-steps, obtained by running CoolWorld 50,000 times (Battery B), with initial conditions set
by clicking on “Special conditions”. In lighter grey: Exact probability distribution (Calculated
using Markov chain analysis)
is inversely proportional to the square root of the sample size (i.e. the number of
runs), is naturally much lower in these latter cases.
When, like in this example, the space of all possible outcomes in the distribution
under analysis is finite (the number of walkers in a house must be an integer
between 0 and 100), one can go further and calculate confidence intervals for the
obtained frequencies. This is easily conducted when one realises that the exact
probability distribution is a multinomial. Genz and Kwong (2000) show how to
calculate these confidence intervals.
To conclude this section, let us emphasise that all that has been written here
applies to any statistic obtained from any computer model. In particular, the statistic
may refer to predefined regimes (e.g. “number of time-steps between 0 and 100
where there are more than 20 walkers in a house”) or to various time-steps (e.g.
“total number of walkers in a house in odd time-steps in between time-steps 50 and
200”). These statistics, like any other one, follow a specific probability distribution
that can be approximated to any degree of accuracy by running the computer model.
11.7 Mathematical Analysis: Time-Homogenous

Markov Chains
The whole range of mathematical techniques that can be used to analyse formal
systems is too broad to be reviewed here. Instead, we focus on one specific
technique that seems to us particularly useful to analyse Social Simulation models:
Markov chain analysis. Besides, there are multiple synergies to be exploited by

using Markov chain analysis and computer simulation together, as we will see in the
next section.
Our first objective is to learn how to represent a particular computer model as a
time-homogeneous Markov chain. This alternative representation of the model will
allow us to use several simple mathematical results that will prove useful to
understand the dynamics of the model. We therefore start by describing time-
homogeneous Markov chains.
11.7.1 What Is a Time-Homogeneous Markov Chain?
Consider a system that in time-step n ¼ {1,2,3,. . .} may be in one of a finite

number of possible states S ¼ {s1, s2,. . ., sM}. The set S is called the state space;
in this chapter we only consider finite state spaces.10 Let the sequence of random
variables Xn 2 S represent the state of the system in time-step n. As an example,
X3 ¼ s9 means that at time n ¼ 3 the system is in state s9. The system starts at a
certain initial state X0 and moves from one state to another. The system is stochastic
in that, given the present state, the system may move to one or another state with a
certain probability (see Fig. 11.10). The probability that the system moves from
state i to state j in one time-step, P(Xn+1 ¼ j | Xn ¼ i), is denoted by pi,j. As an
example, in the Markov chain represented in Fig. 11.10, p4,6 equals 0 since the
system cannot go from state 4 to state 6 in one single time-step. The system may
also stay in the same state i, and this occurs with probability pi,i. The probabilities
pi,j are called transition probabilities and they are often arranged in a matrix,
namely the transition matrix P.
Implicitly, our definition of transition probabilities assumes two important
properties about the system:
(a) The system has the Markov property. This means that the present state
contains all the information about the future evolution of the system that can
be obtained from its past, i.e. given the present state of the system, knowing the
past history about how the system reached the present state does not provide
any additional information about the future evolution of the system. Formally,
PðXnþ1 ¼ xnþ1 j Xn ¼ xn ; Xn1 ¼ xn1 ; . . . ; X0 ¼ x0 Þ ¼ PðXnþ1 ¼ xnþ1 j Xn ¼ xn Þ
(b) In this chapter we focus on time-homogeneous Markov chains, i.e. Markov

chains with time-homogeneous transition probabilities. This basically means
that transition probabilities pi,j are independent of time, i.e. the one-step transi-
tion probability pi,j depends on i and j but is the same at all times n. Formally,
10
The term ‘Markov chain’ allows for countably infinite state spaces too (Karr 1990).
Fig. 11.10 Schematic Initial

transition diagram of a state
Markov chain. Circles denote
s9 s12
states and directed arrows s7 s11
indicate possible transitions s8
s3 s5
between states. In this figure,
thicker circles and arrows s1 s4 s6 s2
represent one possible path s10
where the initial state X0 is s8
Final
and the final state is s2
state
PðXnþ1 ¼ j j Xn ¼ iÞ ¼ PðXn ¼ j j Xn1 ¼ iÞ ¼ pi ;j
The crucial step in the process of representing a computer model as a time-

homogeneous Markov chain (THMC) consists in identifying an appropriate set of
state variables. A particular combination of specific values for these state variables
will define one particular state of the system. Thus, the challenge consists in
choosing the set of state variables in such a way that the computer model can be
represented as a THMC. In other words, the set of state variables must be such that
one can see the computer model as a transition matrix that unambiguously
determines the probability of going from any state to any other state.
Example: A Simple Random Walk
Let us consider a model of a simple 1-dimensional random walk and try to see it as a
THMC. In this model – which can be run and downloaded at the dedicated model
webpage http://luis.izquierdo.name/models/randomwalk – there are 17 patches in
line, labelled with the integers between 1 and 17. A random walker is initially
placed on one of the patches. From then onwards, the random walker will move
randomly to one of the spatially contiguous patches in every time-step (staying still
is not an option). Space does not wrap around, i.e. patch 1’s only neighbour is
patch 2 (Fig. 11.11).
This model can be easily represented as a THMC by choosing the agent’s
position (e.g. the number of the patch she is standing on) as the only state variable.
To be sure, note that defining the state of the system in this way, it is true that there
is a fixed probability of going from any state to any other state, independent of time.
The transition matrix P ¼ [pi,j] corresponding to the model is:
2 3
0 1 0 0
6 .. 7
6 0:5 0 0:5 0 . 7
6 7
6 0 0:5 0 0:5 0 7
6 7
6 0 0:5 0 0:5 0 7
6
P ¼ ½pi;j ¼ 6 .. .. .. .. .. 7 (11.1)
. . . . . 7
6 7
6 0 0:5 0 0:5 0 7
6 7
6 .. 7
4 . 0 0:5 0 0:5 5
0 0 1 0
Fig. 11.11 Snapshot of the 1-dimensional random walk applet. Patches are arranged in a
horizontal line on the top right corner of the figure; they are labelled with integers, and coloured
in shades of grey according to the number of times that the random walker has visited them: the
higher the number of visits, the darker the shade of blue. The plot beneath the patches shows the
time series of the random walker’s position
Where, as explained above, pi,j is the probability P(Xn+1 ¼ j j Xn ¼ i) that the

system will be in state j in the following time-step, knowing that it is currently in
state i.
11.7.2 Transient Distributions of Finite THMCs
The analysis of the dynamics of THMCs is usually divided into two parts: transient
dynamics (finite time) and asymptotic dynamics (infinite time). The transient
behaviour is characterised by the distribution of the state of the system Xn for a
fixed time-step n 0. The asymptotic behaviour (see Sects. 11.7.3 and 11.7.4) is
characterised by the limit of the distribution of Xn as n goes to infinity, when this
limit exists.
This section explains how to calculate the transient distribution of a certain
THMC, i.e. the distribution of Xn for a fixed n 0. In simple words, we are after a
vector a(n) containing the probability of finding the process in each possible state in
time-step n. Formally, a(n) ¼ [a1(n), . . ., aM(n)], where ai(n) ¼ P(Xn ¼ i), denotes
the distribution of Xn for a THMC with M possible states. In particular, a(0) denotes
the initial distribution over the state space, i.e. ai(0) ¼ P(X0 ¼ i). Note that there is
no problem in having uncertain initial conditions, i.e. probability functions over the
space of possible inputs to the model.
0.12
0.10
frequency distribution
Exact
Error bars prob. function
Relative frequency
0.08 (std. error)
0.06
0.04
0.02
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Location of random walker after 100 time-steps
Fig. 11.12 Probability function of the position of the 1-dimensional random walker in time-step
100, starting at an initial random location
It can be shown that one can easily calculate the transient distribution in time-
step n, simply by multiplying the initial conditions by the n-th power of the
transition matrix P.
Proposition 1. a(n) ¼ a(0) · Pn
Thus, the elements p(n)i,j of Pn represent the probability that the system is in state j
after n time-steps having started in state i, i.e. p(n)i,j ¼ P(Xn ¼ j j X0 ¼ i). A
straightforward corollary of Proposition 1 is that a(n+m) ¼ a(n) · Pm.
As an example, let us consider the 1-dimensional random walk again. Imagine
that the random walker starts at an initial random location, i.e. a(0) ¼ [1/17, . . .,
1/17]. The exact distribution of the walker’s position in time-step 100 would then be
a(100) ¼ a(0) · P100. This distribution is represented in Fig. 11.10, together with an
empirical distribution obtained by running the model 50,000 times (Fig. 11.12).
Having obtained the probability function over the states of the system for any
fixed n, namely the probability mass function of Xn, it is then straightforward to
calculate the distribution of any statistic that can be extracted from the model. As
argued in the previous sections, the state of the system fully characterises it, so any
statistic that we obtain about the computer model in time-step n must be, ultimately,
a function of {X0, X1, . . ., Xn}.
Admittedly, the transition matrix of most computer models cannot be easily
derived, or it is unfeasible to operate with it. Nonetheless, this apparent drawback is
not as important as one might expect. As we shall see below, it is often possible to
infer many properties of a THMC even without knowing the exact values of its
transition matrix, and these properties can yield useful insights about the dynamics
of the associated process. Knowing the exact values of the transition matrix allows
us to calculate the exact transient distributions using Proposition 1; this is desirable
but not critical, since we can always approximate these distributions by conducting
many simulation runs, as explained in Sect. 11.6.
11.7.3 Important Concepts
This section presents some basic concepts that will prove useful to analyse the
dynamics of computer models. The notation used here follows the excellent book
on stochastic processes written by Kulkarni (1995).
Definition 1: Accessibility
A state j is said to be accessible from state i if starting at state i there is a chance that
the system may visit state j at some point in the future. By convention, every state is
accessible from itself. Formally, a state j is said to be accessible from state i if for
some n 0, p(n)i,j > 0.
Note that j is accessible from i 6¼ j if and only if there is a directed path from i to
j in the transition diagram. In that case, we write i ! j. If i ! j we also say that i
leads to j. As an example, in the THMC represented in Fig. 11.10, s2 is accessible
from s12 but not from s5. Note that the definition of accessibility does not depend on
the actual magnitude of p(n)i,j, only on whether it is exactly zero or strictly positive.
Definition 2: Communication
A state i is said to communicate with state j if i ! j and j ! i.
If i communicates with j we also say that i and j communicate and write i$j. As
an example, note that in the simple random walk presented in Sect. 11.7.1, every
state communicates with every other state. It is worth noting that the relation
“communication” is transitive, i.e.
i $ j; j $ k ) i $ k:
Definition 3: Communicating Class

A set of states C S is said to be a communicating class if:
• Any two states in the communicating class communicate with each other.
Formally,
i 2 C; j 2 C ) i $ j
• The set C is maximal, i.e. no strict superset of a communicating class can be a

communicating class. Formally,
i 2 C; i $ j ) j 2 C
As an example, note that in the simple random walk presented in Sect. 11.7.1
there is one single communicating class that contains all the states. In the THMC
represented in Fig. 11.10 there are four communicating classes: {s2}, {s5}, {s10},
{s1, s3, s4, s6, s7, s8, s9, s11, s12}.
Definition 4: Closed Communicating Class (i.e. Absorbing Class). Absorbing
State
A communicating class C is said to be closed if no state within C leads to any state
outside C. Formally, a communicating class C is said to be closed if i 2 C and j 2
=C
implies that j is not accessible from i.
Note that once a Markov chain visits a closed communicating class, it cannot
leave it. Hence we will sometimes refer to closed communicating classes as
“absorbing classes”. This latter term is not standard in the literature, but we find
it useful here for explanatory purposes. Note that if a Markov chain has one single
communicating class, it must be closed.
As an example, note that the communicating classes {s10} and {s1, s3, s4, s6, s7, s8,
s9, s11, s12} in the THMC represented in Fig. 11.10 are not closed, as they can be
abandoned. On the other hand, the communicating classes {s2} and {s5} are indeed
closed, since they cannot be abandoned. When a closed communicating class consists
of one single state, this state is called absorbing. Thus, s2 and s5 are absorbing states.
Formally, state i is absorbing if and only if pi,i ¼ 1 and pi,j ¼ 0 for i ¼
6 j.
Proposition 2. Decomposition Theorem (Chung 1960)
The state space S of any Markov chain can be uniquely partitioned as follows:
S ¼ C1 [ C2 [ . . . [ Ck [ T
where C1, C2, . . ., Ck are closed communicating classes, and T is the union of all
other communicating classes.
Note that we do not distinguish between non-closed communicating classes: we
lump them all together into T. Thus, the unique partition of the THMC represented in
Fig. 11.10 is S ¼ {s2} [ {s5} [ {s1, s3, s4, s6, s7, s8, s9, s10, s11, s12}. The simple
random walk model presented in Sect. 11.7.1 has one single (closed) communicating
class C1 containing all the possible states, i.e. S C1.
Definition 5: Irreducibility
A Markov chain is said to be irreducible if all its states belong to a single closed
communicating class; otherwise it is called reducible. Thus, the simple random walk
presented in Sect. 11.7.1 is irreducible, but the THMC represented in Fig. 11.10 is reducible.
Definition 6: Transient and Recurrent States
A state i is said to be transient if, given that we start in state i, there is a non-zero
probability that we will never return back to i. Otherwise, the state is called
recurrent. A Markov chain starting from a recurrent state will revisit it with
probability 1, and hence revisit it infinitely often. On the other hand, a Markov
chain starting from a transient state has a strictly positive probability of never
coming back to it. Thus, a Markov chain will visit any transient state only finitely
many times; eventually, transient states will not be revisited anymore.
Definition 7: Periodic and Aperiodic States. Periodic and Aperiodic Commu-
nicating Classes
A state i has period d if any return to state i must occur in multiples of d time-steps.
If d ¼ 1, then the state is said to be aperiodic; otherwise (d > 1), the state is said to
be periodic with period d. Formally, state i’s period d is the greatest common
divisor of the set of integers n > 0 such that p(n)i,i > 0. For our purposes, the
concept of periodicity is only relevant for recurrent states. As an example, note that
every state in the simple random walk presented in Sect. 11.7.1 is periodic with
period 2.
An interesting and useful fact is that if i$j, then i and j must have the same
period (see Theorem 5.2. in Kulkarni (1995)). In particular, note that if pi,i > 0 for
any i, then the communicating class to which i belongs must be aperiodic. Thus, it
makes sense to qualify communicating classes as periodic with period d, or
aperiodic. A closed communicating class with period d can return to its starting
state only at times d, 2d, 3d, . . .
The concepts presented in this section will allow us to analyse the dynamics of
any finite Markov chain. In particular, we will show that, given enough time, any
finite Markov chain will necessarily end up in one of its closed communicating
classes (i.e. absorbing classes).
11.7.4 Limiting Behaviour of Finite THMCs
This section is devoted to characterising the limiting behaviour of a THMC, i.e.

studying the convergence (in distribution) of Xn as n tends to infinity. Specifically,
we aim to study the behaviour of ai(n) ¼ P(Xn ¼ i) as n tends to infinity. From
Proposition 1 it is clear that analysing the limiting behaviour of Pn would enable us
to characterise ai(n). There are many introductory books in stochastic processes that
offer clear and simple methods to analyse the limiting behaviour of THMCs when
the transition matrix P is tractable (see e.g. Chap. 5 in (Kulkarni 1999), Chaps. 2–4
in (Kulkarni 1995), Chap. 3 in (Janssen and Manca 2006) or the book chapter
written by Karr (1990)). Nonetheless, we focus here on the general case, where
operating with the transition matrix P may be computationally unfeasible.
11.7.4.1 General Dynamics
The first step in the analysis of any THMC consists in identifying all the closed
communicating classes, so we can partition the state space S as indicated by the
decomposition theorem (see proposition 2). The following proposition (Theorems
3.7 and 3.8 in Kulkarni (1995)) reveals the significance of this partition:
Proposition 3. General Dynamics of Finite THMCs
Consider a finite THMC that has been partitioned as indicated in proposition 2. Then:
1. All states in T (i.e. not belonging to a closed communicating class) are transient.
2. All states in Cv (i.e. in any closed communicating class) are recurrent; v 2 {1, 2, . . ., k}.
Proposition 3 states that sooner or later the THMC will enter one of
the absorbing classes and stay in it forever. Formally, for all i 2 S and all j 2 T:
ðnÞ
lim p ¼ 0, i.e. the probability of finding the process in a state belonging to a non-
n!1 i;j
closed communicating class goes to zero as n goes to infinity. Naturally, if the
initial state already belongs to an absorbing class Cv, then the chain will never
abandon such a class. Formally, for all i 2 Cv and all j 2 = Cv: p(n)i,j ¼ 0 for all n 0.
As an example of the usefulness of Proposition 3, consider the THMC
represented in Fig. 11.10. This THMC has only two absorbing classes: {s2} and
{s5}. Thus, the partition of the state space is: S ¼ {s2} [ {s5} [ {s1, s3, s4, s6, s7, s8,
s9, s10, s11, s12}. Hence, applying Proposition 3 we can state that the process will
eventually end up in one of the two absorbing states, s2 or s5. The probability of
ending up in one or the other absorbing state depends on the initial conditions a(0)
(and on the actual numbers pi,j in the transition matrix, of course). Slightly more
formally, the limiting distribution of Xn exists, but it is not unique, i.e. it depends on
the initial conditions.
11.7.4.2 Dynamics Within Absorbing Classes
The previous section has explained that any simulation run will necessarily end up in
a certain absorbing class; this section characterises the dynamics of a THMC that is
already “trapped” in an absorbing class. This is precisely the analysis of irreducible
Markov chains, since irreducible Markov chains are, by definition, Markov chains
with one single closed communicating class (see definition 5). In other words, one
can see any THMC as a set of transient states T plus a finite number of irreducible
Markov sub-chains.
Irreducible THMCs behave significantly different depending on whether they
are periodic or not. The following sections characterise these two cases.
Irreducible and Aperiodic THMCs
Irreducible and aperiodic THMCs are often called ergodic. In these processes the
probability function of Xn approaches a limit as n tends to infinity. This limit is
called the limiting distribution, and is denoted here by π. Formally, the following
limit exists and is unique (i.e. independent of the initial conditions ai(0)):
ðnÞ
lim ai ¼ πi > 0 i2S
n!1
Thus, in ergodic THMCs the probability of finding the system in each of its
states in the long run is strictly positive and independent of the initial conditions
(Theorems 3.7 and 3.15 in Kulkarni (1995)). As previously mentioned, calculating
such probabilities may be unfeasible, but we can estimate them sampling many
simulation runs at a sufficiently large time-step.
Importantly, in ergodic THMCs the limiting distribution π coincides with the

occupancy distribution π*, which is the long-run fraction of the time that the
THMC spends in each state.11 Naturally, the occupancy distribution π* is also
independent of the initial conditions. Thus, in ergodic THMCs, running just one
simulation for long enough (which enables us to estimate π*) will serve to estimate
π just as well.
The question that comes to mind then is: How long is long enough? I.e. when
will I know that the empirical distribution obtained by simulation resembles the
limiting distribution π? Unfortunately there is no answer for that. The silver lining is
that knowing that the limiting and the occupancy distribution coincide, that they
must be stable in time, and that they are independent of the initial conditions,
enables us to conduct a wide range of tests that may tell us when it is certainly not
long enough. For example, we can run a battery of simulations and study the
empirical distribution over the states of the system across samples as time goes
by. If the distribution is not stable, then we have not run the model for long enough.
Similarly, since the occupancy distribution is independent of the initial conditions,
one can run several simulations with widely different initial conditions, and
compare the obtained occupancy distributions. If the empirical occupancy
distributions are not similar, then we have not run the model for long enough.
Many more checks can be conducted.
Admittedly, when analysing a computer model one is often interested not so
much in the distribution over the possible states of the system, but rather in the
distribution of a certain statistic. The crucial point is to realise that if the statistic is a
function of the state of the system (and all statistics that can be extracted from the
model are), then the limiting and the occupancy distributions of the statistic exist,
coincide and are independent of the initial conditions.
Irreducible and Periodic THMCs
In contrast with aperiodic THMCs, the probability distribution of Xn in periodic

THMCs does not approach a limit as n tends to infinity. Instead, in an irreducible
THMC with period d, as n tends to infinity, Xn will in general cycle through d
probability functions depending on the initial distribution.
As an example, consider the simple random walk again (which is irreducible
and periodic, with period 2), and assume that the random walker starts at patch
number 1 (i.e. X0 ¼ 1). Given these settings, it can be shown that
11
Formally, the occupancy of state i is defined as:
EðNi ðnÞÞ
π i ¼ lim
n!1 nþ1
where Ni(n) denotes the number of times that the THMC visits state i over the time span {0, 1,. . ., n}.

ð2nÞ 1 1 1 1 1 1 1 1 1
lim a ¼ ; 0; ; 0; ; 0; 0; ; 0; ; 0; ; 0; ; 0;
n!1 i 16 8 8 8 8 8 8 8 16

ð2nþ1Þ 1 1 1 1 1 1 1 1
lim ai ¼ 0; ; 0; ; 0; ; 0; ; 0; ; 0; ; 0; ; 0; ; 0
n!1 8 8 8 8 8 8 8 8
In particular, the limits above show that the random walker cannot be at a patch
with an even number in any even time-step, and he cannot be at a patch with an odd
number in any odd time-step. In contrast, if the random walker started at patch
number 2 (i.e. X0 ¼ 2), then the limits above would be interchanged.
Fortunately, every irreducible (periodic or aperiodic) THMC does have a unique
occupancy distribution π*, independent of the initial conditions (see Theorem 5.19
in Kulkarni (1999)). In our particular example, this is:

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
π ¼ ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ;
32 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 32
Thus, the long-run fraction of time that the system spends in each state in any
irreducible THMC is unique (i.e. independent of the initial conditions). This is a
very useful result, since any statistic which is a function of the state of the system
will also have a unique occupancy distribution independent of the initial conditions.
As explained before, this occupancy distribution can be approximated with one
single simulation run, assuming it runs for long enough.
11.8 Synergies Between Mathematical Analysis

and Computer Simulation
In this section we present various ways in which mathematical analysis and

computer simulation can be combined to produce a better understanding of the
dynamics of a model. Note that a full understanding of the dynamics of a model
involves not only characterising (i.e. describing) them, but also finding out why
such dynamics are being observed, i.e. identifying the subsets of rules that are
necessary or sufficient to generate certain aspects of the observed dynamics. To do
this, one often has to make changes in the model, i.e. build supporting models that
differ only slightly from the original one and may yield useful insights about its
dynamics. These supporting models will sometimes be more tractable (e.g. if
heterogeneity or stochasticity are averaged out) and sometimes more complex
(e.g. if interactions that were assumed to be global in the original model may
only take place locally in the supporting model). Thus, for clarity, we distinguish
three different cases and deal with them in turn (see Fig. 11.13):
Fig. 11.13 To fully understand the dynamics of a model, one often has to study supporting models
that differ only slightly from the original one. Some of these supporting models may be more
tractable whilst others may be more complex
1. Characterisation of the dynamics of a model.

2. Moves towards greater mathematical tractability. This involves creating and
studying supporting models that are simpler than the original one.
3. Moves towards greater mathematical complexity. This involves creating and
studying supporting models that are less tractable than the original one.
11.8.1 Characterising the Dynamics of the Model
There are many types of mathematical techniques that can be usefully combined
with computer simulation to characterise the dynamics of a model (e.g. Stochastic
Approximation Theory (Benveniste et al. 1990; Kushner and Yin 1997)), but for
limitations of space we focus here on Markov chain analysis only.
When using Markov chain analysis to characterise the dynamics of a model it
may happen that the transition matrix can be easily computed and we can operate
with it, or it may not. In the former case – which is quite rare in Social Simulation
models –, one can provide a full characterisation of the dynamics of the model just
by operating with the transition matrix (see Proposition 1 and the beginning of
Sect. 11.7.4 for references). In general, however, deriving and operating with the
transition matrix may be unfeasible, and it is in this common case where there is a
lot to gain in using Markov chain analysis and computer simulation together. The
overall method goes as follows:
• Use Markov chain analysis to assess the relevance of initial conditions and to
identify the different regimes in which the dynamics of the model may end up
trapped.
• Use the knowledge acquired in the previous point to design suitable computa-
tional experiments aimed at estimating the exact probability distributions for the
relevant statistics (which potentially depend on the initial conditions).
The following describes this overall process in greater detail. Naturally, the first
step consists in finding an appropriate definition of the state of the system, as
explained in Sect. 11.7.1. The next step is to identify all the closed communicating
(i.e. absorbing) classes in the model Cv (v 2 {1, 2, . . ., k}). This allows us to
partition the state space of the Markov chain as the union of all the closed
communicating classes C1, C2, . . ., Ck in the model plus another class T containing
all the states that belong to non-closed communicating classes. Izquierdo et al.
(2009) illustrate how to do this in ten well-known models in the Social Simulation
literature.
In most cases, conducting the partition of the state space is not as difficult as it
may seem at first. In particular, the following proposition provides some simple
sufficient conditions that guarantee that the computer model contains one single
aperiodic absorbing class, i.e. the finite THMC that the computer model implements
is irreducible and aperiodic (i.e. ergodic).
Proposition 4. Sufficient Conditions for Irreducibility and Aperiodicity
1. If it is possible to go from any state to any other state in one single time-step
(pi,j > 0 for all i 6¼ j) and there are more than two states, then the THMC is
irreducible and aperiodic.
2. If it is possible to go from any state to any other state in a finite number of time-
steps (i$j for all i 6¼ j), and there is at least one state in which the system may
stay for two consecutive time-steps (pi,i > 0 for some i), then the THMC is
irreducible and aperiodic.
3. If there exists a positive integer n such that p(n)i,j > 0 for all i and j, then the
THMC is irreducible and aperiodic (Janssen and Manca 2006, p. 107).
If one sees the transition diagram of the Markov chain as a (directed) network,
the conditions above can be rewritten as:
1. The network contains more than two nodes and there is a directed link from
every node to every other node.
2. The network is strongly connected and there is at least one loop.
3. There exists a positive integer n such that there is at least one walk of length n
from any node to every node (including itself).
Izquierdo et al. (2009) show that many models in the Social Simulation literature
satisfy one of these sufficient conditions (e.g. Epstein and Axtell’s (1996)
Sugarscape, Axelrod’s (1986) metanorms models, Takahashi’s (2000) model of
generalized exchange, and Miller and Page’s (2004) standing ovation model with
noise). This is important since, as explained in Sect. 11.7.4.2, in ergodic THMCs
the limiting and the occupancy distributions of any statistic exist, coincide and are
independent of the initial conditions (so running just one simulation for long
enough, which enables us to estimate the occupancy distribution, will serve to
estimate the limiting distribution just as well).
Let us return to the general case. Having partitioned the state space, the analysis
of the dynamics of the model is straightforward: all states in T (i.e. in any finite
communicating class that is not closed) are transient, whereas all states in Cv (i.e. in
any finite closed communicating class) are recurrent. In other words, sooner or later
any simulation run will enter one of the absorbing classes Cv and stay in it forever.
Here computer simulation can play a crucial role again, since it allows us to
estimate the probability of ending up in each of the absorbing classes for any
(stochastic or deterministic) initial condition we may be interested in. A case-in-
point would be a model that has only a few absorbing states, or where various
absorbing states are put together into only a few groups. Izquierdo et al. (2009)
analyse models that follow that pattern: Axelrod’s (1997b) model of dissemination
of culture, Arthur’s (1989) model of competing technologies, and Axelrod and
Bennett’s (1993) model of competing bimodal coalitions. CharityWorld (Polhill
et al. 2006; Izquierdo and Polhill 2006) is an example of a model with a unique
absorbing state.
The following step consists in characterising the dynamics of the system within
each of the absorbing classes. Once the system has entered a certain absorbing class
Cv, it will remain in it forever exhibiting a unique conditional12 occupancy distri-
bution πv* over the set of states that compose Cv. Naturally, the same applies to any
statistic we may want to study, since all statistics that can be extracted from the
model are a function of the state of the system.
The conditional occupancy distribution πv* denotes the (strictly positive) long-
run fraction of the time that the system spends in each state of Cv given that the
system has entered Cv. Importantly, the conditional occupancy distribution πv* is
the same regardless of the specific state through which the system entered Cv. The
role of simulation here is to estimate these conditional occupancy distributions for
the relevant statistics by running the model for long enough.
Finally, recall that some absorbing classes are periodic and some are aperiodic.
Aperiodic absorbing classes have a unique conditional limiting distribution πv
denoting the long-run (strictly positive) probability of finding the system in each
of the states that compose Cv given that the system has entered Cv. This conditional
limiting distribution πv coincides with the conditional occupancy distribution πv*
and, naturally, is also independent of the specific state through which the system
entered Cv. (Again, note that this also applies to the distribution of any statistic, as
they are all functions of the state of the system, necessarily.)
In contrast with aperiodic absorbing classes, periodic absorbing classes do
not generally have a unique limiting distribution; instead, they cycle through d
probability functions depending on the specific state through which the system
entered Cv (where d denotes the period of the periodic absorbing class). This is
knowledge that one must take into account at the time of estimating the relevant
probability distributions using computer simulation.
Thus, it is clear that Markov chain analysis and computer simulation greatly
complement each other. Markov chain analysis provides the overall picture of the
dynamics of the model by categorising its different dynamic regimes and
12
Given that the system has entered the absorbing class Cv.
identifying when and how initial conditions are relevant. Computer simulation uses
this information to design appropriate computational experiments that allow us to
quantify the probability distributions of the statistics we are interested in. As
explained above, these probability distributions can always be approximated with
any degree of accuracy by running the computer model several times.
There are several examples of this type of synergetic combination of Markov
chain analysis and computer simulation in the literature. Galán and Izquierdo
(2005) analysed Axelrod’s (1986) agent-based model as a Markov chain, concluded
that the long-run behaviour of that model was independent of the initial conditions,
in contrast to the initial conclusions of the original analysis. Galán and and
Izquierdo (2005) also used computer simulation to estimate various probability
distributions. Ehrentreich (2002, 2006) used Markov chain analysis on the Artificial
Stock Market (Arthur et al. 1997; LeBaron et al. 1999) to demonstrate that the
mutation operator implemented in the model is not neutral to the learning rate, but
introduces an upward bias.13 A more positive example is provided by Izquierdo
et al. (2007, 2008b), who used Markov chain analysis and computer simulation to
confirm and advance various insights on reinforcement learning put forward by
Macy and Flache (2002) and Flache and Macy (2002).
11.8.2 Moves Towards Greater Mathematical Tractability:

Simplifications
There are at least two types of simplifications that can help us to better understand
the dynamics of a model. One consists in studying specific parameterisations of the
original model that are thought to lead to particularly simple dynamics, or to more
tractable situations (Gilbert and Terna 2000; Gilbert 2007). Examples of this type of
activity would be to run simulations without agents or with very few agents, explore
the behaviour of the model using extreme parameter values, model very simple
environments, etc. This activity is common practice in the field (see e.g. Gotts et al.
2003c, d).
A second type of simplification consists in creating an abstraction of the original
model (i.e. a model of the model) which is mathematically tractable. An example of
one possible abstraction would be to study the expected motion of the dynamic
system (see the studies conducted by Galán and Izquierdo (2005), Edwards et al.
(2003), Castellano et al. (2000), Huet et al. (2007), Mabrouk et al. (2007), Vilà
(2008) and Izquierdo et al. (2007, 2008b) for illustrations of mean-field
approximations). Since these mathematical abstractions do not correspond in a
one-to-one way with the specifications of the formal model, any results obtained
with them will not be conclusive in general, but they may give us insights
13
This finding does not refute some of the most important conclusions obtained by the authors of
the original model.
suggesting areas of stability and basins of attraction, clarifying assumptions,

assessing sensitivity to parameters, or simply giving the option to illustrate
graphically the expected dynamics of the original model. This approach can also
be used as a verification technique to detect potential errors and artefacts (Galán
et al. 2009).
11.8.3 Moves Towards Greater Mathematical Complexity:

Extensions
As argued before, understanding the dynamics of a model implies identifying the

set of assumptions that are responsible for particular aspects of the obtained results.
Naturally, to assess the relevance of any assumption in a model, it is useful to
replace it with other alternatives, and this often leads to greater mathematical
complexity.14
Ideally, the evaluation of the significance of an assumption is conducted by
generalisation, i.e. by building a more general model that allows for a wide range of
alternative competing assumptions, and contains the original assumption as a
particular case. An example would be the introduction of arbitrary social networks
of interaction in a model where every agent necessarily interacts with every other
agent. In this case, the general model with arbitrary networks of interaction would
correspond with the original model if the network is assumed to be complete, but
any other network could also be studied within the same common framework.
Another example is the introduction of noise in deterministic models.
Building models by generalisation is useful because it allows for a transparent,
structured and systematic way of exploring the impact of various alternative
assumptions that perform the same role in the model, but it often implies a loss in
mathematical tractability (see e.g. Izquierdo and Izquierdo 2006). Thus, it is often
the case that a rigorous study of the impact of alternative assumptions in a model
requires being prepared to slide up and down the tractability continuum depicted in
Fig. 11.13 (Gotts et al. 2003b). In fact, all the cases that are mentioned in the rest of
this section involved greater complexity than the original models they considered,
and computer simulation had to be employed to understand their dynamics.
In the literature there are many examples of the type of activity explained in this
section. For example, Klemm et al. studied the relevance of various assumptions in
Axelrod’s model of dissemination of culture (1997b) by changing the network
topology (Klemm et al. 2003a), investigating the role of dimensionality (Klemm
14
This is so because many assumptions we make in our models are, to some extent, for the sake of
simplicity. As a matter of fact, in most cases the whole purpose of modelling is to build an
abstraction of the world which is simpler than the world itself, so we can make inferences about the
model that we cannot make directly from the real world (Edmonds 2001; Galán et al. 2009;
Izquierdo et al. 2008a).
et al. 2003b, 2005), and introducing noise (Klemm et al. 2003c). Another example
is given by Izquierdo and Izquierdo (2007), who analysed the impact of using
different structures of social networks in the efficiency of a market with quality
variability.
In the context of decision-making and learning, Flache and Hegselmann (1999)
and Hegselmann and Flache (2000) compared two different decision-making
algorithms that a set of players can use when confronting various types of social
dilemmas. Similarly, Takadama et al. (2003) analysed the effect of three different
learning algorithms within the same model.
Several authors, particularly in the literature of Game Theory, have investigated
the effect of introducing noise in the decision-making of agents. This is useful not
only to investigate the general effect of potential mistakes or experimentation, but
also to identify the stochastic stability of different outcomes (see Sect. 10 in
Izquierdo et al. (2009)). An illustrative example is given by Izquierdo et al.
(2008b), who investigate the reinforcement learning algorithm proposed by Bush
and Mosteller (1955) using both mathematical analysis and simulation, and find that
the inclusion of small quantities of randomness in players’ decisions can change the
dynamics of the model dramatically.
Another assumption investigated in the literature is the effect of different spatial
topologies (see e.g. Flache and Hegselmann (2001), who generalised two of their
cellular automata models by changing their – originally regular – grid structure).
Finally, as mentioned in Sect. 11.5, it is increasingly common in the field
of evolutionary game theory to assess the impact of various assumptions using
computer simulation (see e.g. Galán and Izquierdo 2005; Santos et al. 2006;
Traulsen et al. 2006; Izquierdo and Izquierdo 2006).
11.9 Summary
In this chapter we have provided a set of guidelines to understand the dynamics of

computer models using both simulation and mathematical analysis. In doing so, it
has become clear that mathematical analysis and computer simulation should not be
regarded as alternative – or even opposed – approaches to the formal study of social
systems, but as complementary (Gotts et al. 2003a, b). Not only can they provide
fundamentally different insights on the same model, but they can also produce hints
for solutions for each other. In short, there are plenty of synergies to be exploited by
using the two techniques together, so the full potential of each technique cannot be
reached unless they are used in conjunction.
To understand the dynamics of any particular computer model, we have seen that
it is useful to see the computer model as the implementation of a function that
transforms probability distributions over the set of possible inputs into probability
distributions over the set of possible outputs. We refer to this function as the formal
model that the computer model implements.
The mathematical approach to analyse formal models consists in examining the

rules that define the model directly; the aim is to deduce the logical implications of
these rules for any particular instance to which they can be applied. Our analysis
of mathematical techniques to study formal models has been focused on the theory
of Markov Chains. This theory is particularly useful for our purposes since
many computer models can be meaningfully represented as time-homogenous
Markov chains.
In contrast with mathematical analysis, the computer simulation approach does
not look at the rules that define the formal model directly, but instead tries to infer
general properties of these rules by examining the outputs they produce when
applied to particular instances of the input space. Thus, in the simulation approach,
the data is produced by the computer using strict logical deduction, but the general
patterns about how the rules transform the inputs into the outputs are inferred using
generalisation by induction. Thus, in the general case – and in contrast with
mathematical analysis –, the inferences obtained using computer simulation will
not be necessarily correct in a strict logical sense; but, on the other hand, computer
simulation enables us to explore formal models beyond mathematical tractability,
and the confidence we can place on the conclusions obtained with this approach can
be rigorously assessed in statistical terms. Furthermore, as shown in this chapter, we
can achieve any arbitrary level of accuracy in our computational approximations by
running the model sufficiently many times.
Bearing in mind the relative strengths and limitations of both approaches, we
have identified at least three different ways in which mathematical analysis and
computer simulation can be usefully combined to produce a better understanding of
the dynamics of computer models.
The first synergy appears at the time of characterising the dynamics of the formal
model under study. To do that, we have shown how Markov chain analysis can be
used to provide an overall picture of the dynamics of the model by categorising its
different dynamic regimes and identifying when and how initial conditions are
relevant. Having conducted such an analysis, one can then use computer simulation
to design appropriate computational experiments with the aim of quantifying the
probability distributions of the variables we are interested in. These probability
distributions can always be approximated with any degree of accuracy by running
the computer model several times.
The two other ways in which mathematical analysis and computer simulation can
be combined derive from the fact that understanding the dynamics of a model
involves not only characterising (i.e. describing) them, but also finding out why
such dynamics are being observed (i.e. discovering causality). This often implies
building supporting models that can be simpler or more complex than the original
one. The rationale to move towards simplicity is to achieve greater mathematical
tractability, and this often involves studying particularly simple parameterisations of
the original model, and creating abstractions which are amenable to mathematical
analysis. The rationale to move towards complexity is to assess the relevance of
specific assumptions, and it often involves building generalisations of the original
model to explore the impact of competing assumptions that can perform the same
role in the model but may lead to different results.
Let us conclude by encouraging the reader to put both mathematical analysis and
computer simulation in their backpack, and be happy to glide up and down the
tractability spectrum where both simple and complex models lie. The benefits are
out there.
Acknowledgements The authors have benefited from the financial support of the Spanish
Ministry of Education and Science (projects DPI2004-06590, DPI2005-05676 and TIN2008-
06464-C03-02), the Spanish Ministry for Science and Innovation (CSD2010-00034), and the
JCyL (projects VA006B09, BU034A08 and GREX251-2009). We are also very grateful to Nick
Gotts, Bruce Edmonds, Gary Polhill, and Cesáreo Hernández for many extremely useful
discussions.
Further Reading
Firstly we suggest three things to read to learn more about Markov Chain models.
Grinstead and Snell (1997) provides an excellent introduction to the theory of finite
Markov Chains, with many examples and exercises. Häggström (2002) gives a clear
and concise introduction to Probability theory and Markov Chain theory, and then
illustrates the usefulness of these theories by studying a range of stochastic
algorithms with important applications in optimisation and other problems in
computing. One of the algorithms covered is the Markov chain Monte Carlo
method. Finally, Kulkarni (1995) provides a rigorous analysis of many types of
useful stochastic processes, e.g. discrete and continuous time Markov Chains,
renewal processes, regenerative processes, and Markov regenerative processes.
The reader may find three other papers helpful. Izquierdo et al. (2009) analyses the
dynamics of ten well-known models in the social simulation literature using the theory
of Markov Chains, and is thus a good illustration of the approach in practice within the
context of social simulation.15 Epstein (2006) is a more general discussion, treating a
variety of foundational and epistemological issues surrounding generative explanation
in the social sciences, and discussing the role of agent-based computational models in
generative social science. Finally, Leombruni and Richiardi (2005) usefully discusses
several issues surrounding the interpretation of simulation dynamics and the
generalisation of the simulation results. For a different approach to analysing the
dynamics of a simulation model we refer the interested reader to Chap. 9 in this
volume (Evans et al. 2013).
15
This comment was added by the editors as the authors are too modest to so describe their own
work.
References
Arthur WB (1989) Competing technologies, increasing returns, and lock-in by historical events.
Econ J 99(394):116–131
Arthur WB, Holland JH, LeBaron B, Palmer R, Tayler P (1997) Asset pricing under endogenous
expectations in an artificial stock market. In: Arthur WB, Durlauf S, Lane D (eds) The
economy as an evolving complex system II. Addison-Wesley Longman, Reading, pp 15–44
Axelrod RM (1986) An evolutionary approach to norms. Am Polit Sci Rev 80(4):1095–1111
Axelrod RM (1997a) Advancing the art of simulation in the social sciences. In: Conte R,
Hegselmann R, Terna P (eds) Simulating social phenomena (Lecture notes in economics and
mathematical systems, 456). Springer, Berlin, pp 21–40
Axelrod RM (1997b) The dissemination of culture: a model with local convergence and global
polarization. J Confl Resolut 41(2):203–226
Axelrod RM, Bennett DS (1993) A landscape theory of aggregation. Br J Polit Sci 23(2):211–233
Axtell RL (2000) Why agents? On the varied motivations for agent computing in the social
sciences. In: Macal VM, Sallach D (eds) Proceedings of the workshop on agent simulation:
applications, models, and tools, Argonne National Laboratory, Argonne, pp 3–24
Axtell RL, Epstein JM (1994) Agent based modeling: understanding our creations. The Bulletin of
the Santa Fe Institute, Winter 1994, pp 28–32
Balzer W, Brendel KR, Hofmann S (2001) Bad arguments in the comparison of game theory and
simulation in social studies. J Artif Soc Soc Simul 4(2), http://jasss.soc.surrey.ac.uk/4/2/1.html
Benveniste A, Métivier M, Priouret P (1990) Adaptive algorithms and stochastic approximations.
Springer, Berlin
Böhm C, Jacopini G (1966) Flow diagrams, turing machines and languages with only two
formation rules. Commun ACM 9(5):366–371
Bush RR, Mosteller F (1955) Stochastic models for learning. Wiley, New York
Castellano C, Marsili M, Vespignani A (2000) Nonequilibrium phase transition in a model for
social influence. Phys Rev Lett 85(16):3536–3539
Cutland N (1980) Computability: an introduction to recursive function theory. Cambridge
Edmonds B (2001) The use of models: making MABS actually work. In: Moss S, Davidsson P
(eds) Multi-agent-based simulation (Lecture notes in artificial intelligence, 1979). Springer,
Berlin, pp 15–32
Edmonds B (2005) Simulation and complexity: how they can relate. In: Feldmann V, Mühlfeld K
(eds) Virtual worlds of precision: computer-based simulations in the sciences and social
sciences. Lit-Verlag, Münster, pp 5–32
Edwards M, Huet S, Goreaud F, Deffuant G (2003) Comparing an individual-based model of
behaviour diffusion with its mean field aggregate approximation. J Artif Soc Soc Simul 6(4),
Ehrentreich N (2002) The Santa Fe Artificial Stock Market re-examined: suggested corrections
(Betriebswirtschaftliche Diskussionsbeiträge Nr. 45/02). Wirtschaftswissenschaftliche Fakultät,
Martin-Luther-Universität, Halle/Saale, http://econwpa.wustl.edu:80/eps/comp/papers/0209/
0209001.pdf
Ehrentreich N (2006) Technical trading in the Santa Fe Institute Artificial Stock Market revisited.
J Econ Behav Organ 61(4):599–616
Epstein JM (2006) Remarks on the foundations of agent-based generative social science. In: Judd KL,
Tesfatsion L (eds) Handbook of computational economics, vol. 2: agent-based computational
economics. North-Holland, Amsterdam, pp 1585–1604
Brookings Institution Press/MIT Press, Cambridge
Evans A, Heppenstall A, Birkin M (2013) Understanding simulation results. Chapter 9 in this
volume
Flache A, Hegselmann R (1999) Rationality vs. learning in the evolution of solidarity networks: a
theoretical comparison. Comput Math Organ Theory 5(2):97–127
Flache A, Hegselmann R (2001) Do irregular grids make a difference? Relaxing the spatial
regularity assumption in cellular models of social dynamics. J Artif Soc Soc Simul 4(4),
Flache A, Macy MW (2002) Stochastic collusion and the power law of learning. J Confl Resolut
46(5):629–653
Galán JM, Izquierdo LR (2005) Appearances can be deceiving: lessons learned re-implementing
Axelrod’s ‘Evolutionary approach to norms’. J Artif Soc Soc Simul 8(3), http://jasss.soc.
Galán JM et al (2009) Errors and artefacts in agent-based modelling. J Artif Soc Soc Simul 12(1),
Genz A, Kwong KS (2000) Numerical evaluation of singular multivariate normal distributions.
J Stat Comput Sim 68(1):1–21
Gilbert N (1999) Simulation: a new way of doing social science. Am Behav Sci 42(10):1485–1487
Sage, London
Gilbert N, Terna P (2000) How to build and use agent-based models in social science. Mind Soc
1(1):57–72
Buckingham
Gotts NM, Polhill JG, Law ANR (2003a) Agent-based simulation in the study of social dilemmas.
Artif Intell Rev 19(1):3–92
Gotts NM, Polhill JG, Adam WJ (2003b) Simulation and analysis in agent-based modelling of land
use change. In: Online proceedings of the first conference of the European social simulation
association, Groningen, The Netherlands, 18–21 Sept 2003, http://www.uni-koblenz.de/~essa/
ESSA2003/proceedings.htm
Gotts NM, Polhill JG, Law ANR (2003c) Aspiration levels in a land-use simulation. Cybern Syst
34(8):663–683
Gotts NM, Polhill JG, Law ANR, Izquierdo LR (2003d) Dynamics of imitation in a land use
simulation. In: Dautenhahn K, Nehaniv C (eds) Proceedings of the second international
symposium on imitation in animals and artefacts, University of Wales, Aberystwyth, 7–11
April 2003, pp 39–46
Grinstead CM, Snell JL (1997) Chapter 11: Markov chains. In: Grinstead CM, Snell JL (eds)
Introduction to probability (Second revised edition). American Mathematical Society,
Providence, pp 405–470, http://www.dartmouth.edu/~chance/teaching_aids/books_articles/
probability_book/book.html
Häggström O (2002) Finite Markov chains and algorithmic applications. Cambridge University Press,
Cambridge
Harel D (1980) On folk theorems. Commun ACM 23(7):379–389
Hauert C, Doebeli M (2004) Spatial structure often inhibits the evolution of cooperation in the
snowdrift game. Nature 428(6983):643–646
Hegselmann R, Flache A (2000) Rational and adaptive playing. Anal Krit 22(1):75–97
Holland JH, Miller JH (1991) Artificial adaptive agents in economic theory. Am Econ Rev 81(2):
365–370
Huet S, Edwards M, Deffuant G (2007) Taking into account the variations of neighbourhood sizes
in the mean-field approximation of the threshold model on a random network. J Artif Soc Soc
Imhof LA, Fudenberg D, Nowak MA (2005) Evolutionary cycles of cooperation and defection.
Proc Natl Acad Sci USA 102(31):10797–10800
Izquierdo SS, Izquierdo LR (2006) On the structural robustness of evolutionary models of
cooperation. In: Corchado E, Yin H, Botti VJ, Fyfe C (eds) Intelligent data engineering and
automated learning – IDEAL 2006 (Lecture notes in computer science, 4224). Springer,
Berlin/Heidelberg, pp 172–182
Izquierdo SS, Izquierdo LR (2007) The impact on market efficiency of quality uncertainty without
asymmetric information. J Bus Res 60(8):858–867
Izquierdo LR, Polhill JG (2006) Is your model susceptible to floating point errors? J Artif Soc Soc
Izquierdo LR, Izquierdo SS, Gotts NM, Polhill JG (2007) Transient and asymptotic dynamics of
reinforcement learning in games. Game Econ Behav 61(2):259–276
Izquierdo LR, Galán JM, Santos JI, Olmo R (2008a) Modelado de sistemas complejos mediante
simulación basada en agentes y mediante dinámica de sistemas. Empiria 16:85–112
Izquierdo SS, Izquierdo LR, Gotts NM (2008b) Reinforcement learning dynamics in social
dilemmas. J Artif Soc Soc Simul 11(2), http://jasss.soc.surrey.ac.uk/11/2/1.html
Izquierdo LR, Izquierdo SS, Galán JM, Santos JI (2009) Techniques to understand computer
simulations: Markov chain analysis. J Artif Soc Soc Simul 12(1), http://jasss.soc.surrey.ac.uk/
12/1/6.html
Janssen J, Manca R (2006) Applied semi-Markov processes. Springer, New York
Karr AF (1990) Markov processes. In: Heyman DP, Sobel MJ (eds) Stochastic models, vol 2,
Handbooks in operations research and management science. Elsevier, Amsterdam, pp 95–123
Klemm K, Eguı́luz VM, Toral R, San Miguel M (2003a) Nonequilibrium transitions in complex
networks: a model of social interaction. Phys Rev E 67(2):026120
Klemm K, Eguı́luz VM, Toral R, San Miguel M (2003b) Role of dimensionality in Axelrod’s
model for the dissemination of culture. Phys A 327(1–2):1–5
Klemm K, Eguı́luz VM, Toral R, San Miguel M (2003c) Global culture: a noise-induced transition
in finite systems. Phys Rev E 67(4):045101
Klemm K, Eguı́luz VM, Toral R, San Miguel M (2005) Globalization, polarization and cultural
drift. J Econ Dyn Control 29(1–2):321–334
Kulkarni VG (1995) Modeling and analysis of stochastic systems. Chapman & Hall/CRC,
Boca Raton
Kulkarni VG (1999) Modeling, analysis, design, and control of stochastic systems. Springer,
New York
Kushner HJ, Yin GG (1997) Stochastic approximation algorithms and applications. Springer,
New York
Lebaron B, Arthur WB, Palmer R (1999) Time series properties of an artificial stock market.
J Econ Dyn Control 23(9–10):1487–1516
Leombruni R, Richiardi M (2005) Why are economists sceptical about agent-based simulations?
Phys A 355(1):103–109
Lieberman E, Havlin S, Nowak MA (2009) Evolutionary dynamics on graphs. Nature 433(7023):
312–316
Mabrouk N, Deffuant G, Lobry C (2007) Confronting macro, meso and micro scale modelling of
bacteria dynamics. In: M2M 2007: Third international model-to-model workshop, Marseille,
France, 15–16 Mar 2007, http://m2m2007.macaulay.ac.uk/M2M2007-Mabrouk.pdf
Macy MW, Flache A (2002) Learning dynamics in social dilemmas. Proc Natl Acad Sci USA
99(3):7229–7236
Miller JH, Page SE (2004) The standing ovation problem. Complexity 9(5):8–16
Nowak MA, May RM (1992) Evolutionary games and spatial chaos. Nature 359(6398):826–829
Nowak MA, Sigmund K (1992) Tit for tat in heterogeneous populations. Nature 355(6357):
250–253
Nowak MA, Sigmund K (1993) A strategy of win-stay, lose-shift that outperforms tit for tat in the
Prisoner’s Dilemma game. Nature 364(6432):56–58
Ostrom T (1988) Computer simulation: the third symbol system. J Exp Soc Psychol 24(5):381–392
Polhill JG, Izquierdo LR, Gotts NM (2006) What every agent based modeller should know about
floating point arithmetic. Environ Model Software 21(3):283–309
Richiardi M, Leombruni R, Saam NJ, Sonnessa M (2006) A common protocol for agent-based
social simulation. J Artif Soc Soc Simul 9(1), http://jasss.soc.surrey.ac.uk/9/1/15.html
Santos FC, Pacheco JM, Lenaerts T (2006) Evolutionary dynamics of social dilemmas in
structured heterogeneous populations. Proc Natl Acad Sci USA 103(9):3490–3494
Suber P (2002) Formal systems and machines: an isomorphism (Electronic hand-out for the course
“Logical Systems”, Earlham College, Richmond, IN) http://www.earlham.edu/~peters/courses/
logsys/machines.htm
Takadama K, Suematsu YL, Sugimoto N, Nawa NE, Shimohara K (2003) Cross-element
validation in multiagent-based simulation: switching learning mechanisms in agents. J Artif
Soc Soc Simul 6(4), http://jasss.soc.surrey.ac.uk/6/4/6.html
Takahashi N (2000) The emergence of generalized exchange. Am J Sociol 10(4):1105–1134
Traulsen A, Nowak MA, Pacheco JM (2006) Stochastic dynamics of invasion and fixation.
Phys Rev E 74(1):011909
Vilà X (2008) A model-to-model analysis of Bertrand competition. J Artif Soc Soc Simul 11(2),
Wikipedia (2007) Structured program theorem, http://en.wikipedia.org/w/index.php?title¼
Structured_program_theorem&oldid¼112885072
Wilensky U (1999) NetLogo, http://ccl.northwestern.edu/netlogo
Chapter 12
Interpreting and Understanding Simulations:
The Philosophy of Social Simulation
R. Keith Sawyer
Why Read This Chapter? To gain an overview of some key philosophical issues
that underlie social simulation. Providing an awareness of them may help avoid the
risk of presenting a very limited perspective on the social world in any simulations
you develop.
Abstract Simulations are usually directed at some version of the question: What is
the relationship between the individual actor and the collective community?
Among social scientists, these questions generally fall under the topic of emer-
gence. Sociological theorists and philosophers of science have developed sophisti-
cated approaches to emergence, including the critical question: to what extent can
emergent phenomena be reduced to explanations in terms of their components?
Modelers often proceed without considering these issues; the risk is that one might
develop a simulation that does not accurately reflect the observed empirical facts, or
one that implicitly sides with one side of a theoretical debate that remains unre-
solved. In this chapter, I provide some tips for those developing simulations, by
drawing on a strong recent tradition of analyzing scientific explanation that is found
primarily in the philosophy of science but also to some extent in sociology.
12.1 Introduction
Researchers who develop multi-agent simulations often proceed without thinking

about what the results will mean or how the results of the simulation might be used.
After the simulation is completed, it is too late to begin to think about interpretation
and understanding, because poorly designed simulations often turn out to be uninter-
pretable. An analysis of interpretation and understanding has to precede and inform
the design. In this chapter, I provide some tips for those developing simulations, by
R.K. Sawyer (*)

Washington University, St. Louis, MO, USA

274 R.K. Sawyer
drawing on a strong recent tradition of analyzing scientific explanation that is found

primarily in the philosophy of science but also to some extent in sociology. My hope
is that this exploration can help modelers by identifying hidden assumptions that
often underlie simulation efforts, assumptions that sometimes can limit the explana-
tory power of the resulting simulation.
Modelers often proceed without considering these issues, in part, because of the
historical roots of the approach. During the 1990s, two strands of work in computer
science began to merge: intelligent agents, a tradition of studying autonomous
agents that emerged from the artificial intelligence community, and artificial life,
typically a graphically-represented two-dimensional grid in which each cell’s
behavior is determined by its nearby neighbors. In the 1990s, a few members of
both of these communities began to use the new technologies to simulate social
phenomena. Several intelligent agent researchers began to explore systems with
multiple agents in communication, which became known as distributed artificial
intelligence, and artificial life researchers began to draw parallels between their
two-dimensional grids and various real-world phenomena, such as ant colonies and
urban traffic flows. The new field that emerged within computer science is generally
known as multi agent systems; the term refers to engineering computer systems with
many independent and autonomous agents, which communicate with each other
using a well-specified set of message types. When multi agent systems are used
specifically to simulate social phenomena, the simulations are known as multi-
agent based simulations (MABS), artificial societies, or social simulations.
Computer scientists had been working with multiple-processor systems since the
1980s, when computer scientists began to experiment with breaking up a computa-
tional task into sub-tasks and then assigning the subtasks to separate, standalone
computers. During this time, a specialized massively parallel computer called the
thinking machine was famously built by the Thinking Machines Corporation, and
computer scientists began developing formalisms to represent distributed compu-
tational algorithms. These early efforts at parallel computation were centrally
managed and controlled, with the distribution of computation being used to speed
up a task or to make it more efficient.
The multi agent systems of the 1990s represented a significant shift from these
earlier efforts, in that each computational agent was autonomous: capable of making
its own decisions and choosing its own course of action. The shift to autonomous
agents raised several interesting issues among computer scientists: if an agent were
asked to execute a task, when and why would the agent agree to do it? Perhaps an agent
would have a different understanding of how to execute a task than the requesting
agent; how could two agents negotiate these understandings? Such questions would be
unlikely to arise if one central agent had control over all of the distributed agents. But
the rapid growth of the Internet – where there are, in fact, many autonomous
computers and systems that communicate with each other every day – resulted in a
real-world situation in which independent computational agents might choose not to
respond to a request, or might respond differently than the requester expected.
As a result of these historical developments, computer scientists found them-
selves grappling with questions that have long been central to sociology and
12 Interpreting and Understanding Simulations: The Philosophy of Social Simulation 275
economics. Why should a person take an action on behalf of the collective good?
How to prevent free-riding and social loafing – situations where agents benefit from
collective action but without contributing very much? What is the relationship
between the individual actor and the collective community? What configurations
of social network are best suited to different sorts of collective tasks?
Among social scientists, these questions generally fall under the topic of emer-
gence (Sawyer 2005). Emergence refers to a certain kind of relation between
system-level phenomena or properties and properties of the system components.
The traditional method of understanding complex system phenomena has been to
break the system up into its components, to analyze each of the components and the
interactions among them. In this reductionist approach, the assumption is that at the
end of this process, the complex system will be fully explained. In contrast, scholars
of emergence are unified in arguing that this reductionist approach does not work
for a certain class of complex system phenomena. There are different varieties of
this argument; some argue that the reduction is not possible for epistemological
reasons (it is simply too hard to explain using reduction although the whole is really
nothing more than the sum of the parts), while others argue that it is not possible for
ontological reasons (the whole is something more and different than the sum of the
parts). Emergent phenomena have also been described as novel – not observed in
any of the components – and unpredictable, even if one has already developed a full
and complete explanation of the components and of their interactions.
Among sociologists, this debate is generally known as the individualism-
collectivism debate. Methodological individualists are the reductionists, those who
argue that all social phenomena can be fully explained by completely explaining the
participating individuals and their interactions. Collectivists argue, in contrast, that
some social phenomena cannot be explained by reduction to the analysis of
individuals and their interactions. Several scholars have recently noted that agent-
based simulations are an appropriate tool to explore these issues (Neumann 2006;
Sawyer 2005; Schmid 2006).
In the mid 1990s, a few sociologists who were interested in the potential of
computer simulation to address these questions began to join with computer
scientists who were fascinated with the more theoretical dimensions of these very
practical questions, and the field of multi-agent based simulation (MABS) was
born. Since that time, technology has rapidly advanced, and now there are several
computer tools, relatively easy to use, that allow social scientists to develop multi-
agent simulations of social phenomena.
After almost 10 years, this handbook provides an opportunity to look reflexively
at this work, and to ask: what do these simulations mean? How should scientists
interpret them? Such questions have traditionally been associated with the philoso-
phy of science. For over a century, philosophers of science have been exploring
topics that are fundamental to science: understanding, explanation, perception,
validation, interpretation. In the following, I draw on concepts and arguments from
within the philosophy of science to ask a question that is critical to scientists: What
do these multi-agent based simulations mean? How should we interpret simulations?
276 R.K. Sawyer
In spite of the occasional participation of sociologists in social simulation

projects, most modelers continue to proceed without an awareness of these impor-
tant foundational questions. These are of more than simple theoretical importance;
without an awareness of these somewhat complex issues of interpretation and
understanding, the risk is that one might develop a simulation that does not
accurately reflect the observed empirical facts, or one that implicitly sides with
one side of a theoretical debate that remains unresolved (thus clouding the interpre-
tation of the simulation’s results).
I address these questions by delving into the nature of explanation. Many
developers of MABS believe that the simulations provide explanations of real
phenomena – that the concrete specifics of what is going on in the world are
revealed by examining the simulation. Interpreting and understanding the simula-
tion thus results in an explanation of the target phenomenon.
Although simulation has many potential benefits to offer scientists, I argue that
the meaning of a simulation is rarely obvious and unequivocal. For example, there
are many different ways to understand and interpret a given simulation of a social
phenomenon. These run roughly along a spectrum from: a very narrowly focused
and ungeneralizable simulation of a very specific instance of a social phenomenon,
to a grand-theoretical type of simulation that explains a very broad range of social
emergence phenomena. The specific end of the spectrum results in better under-
standing of a single phenomenon, but not in any lawful regularities nor in any
general knowledge about social life. The general end of the spectrum is something
like the tradition of grand theory in sociology, with generalizable laws that explain
a wide range of social phenomena. The center of the spectrum is associated with
what Merton famously called “middle range theories”; this is what most main-
stream sociologists today believe is the appropriate task of sociology, as the field
has turned away from grand theorizing in recent decades.
All of these are valid forms of sociological explanation, and each has the
potential to increase our understanding of social life. My concern is with the
specific end of the spectrum: in some cases, it could be that a simulation explains
only a very narrow single case, with no generalizability. This would be of limited
usefulness to our understanding.
12.2 Interpreting Multi-agent Simulations
How should scientists outside of the simulation community interpret a multi-agent

simulation of a real phenomenon? Although there has been almost no philosophical
attention to these simulations, simulation developers themselves have often engaged
in discussions of the scientific status of their simulations. Within this community,
there is disagreement about the scientific status of the simulations. Opinions fall into
two camps: the “simulation as theory” camp, and the “simulation as experiment”
camp.
Representing the first group, many of those developing computer simulations

believe that in building them, they are engaged in a form of theory construction
(Conte et al. 2001; Markovsky 1997; also see Ostrom 1988). They argue that social
simulation, for example, is a more sophisticated and advanced form of social theory,
because concepts and axioms must be rigorously specified to be implemented in a
computer program, unlike the “discursive theorizing” of many sociologists, which
is relatively vague and hard to empirically test (Conte et al. 2001; Turner 1993).
As Markovsky (1997) noted, turning a (discursive) sociological theory into a
simulation is not a transparent translation. A variable in the theory may turn out to
be central to the simulation, or it may turn out not to matter very much; one cannot
know which without going through the exercise of programming the simulation.
Developing a simulation almost always reveals logical gaps in a theory, and these
must be filled in before the simulation will work. As a result, simulations often
introduce logical relationships that the original theory did not specify, and they
contain gap-filling assumptions that the theory never made.
Representing the second group, other modelers have argued that a simulation is a
virtual experiment (Carley and Gasser 1999). From this perspective, simulations
cannot explain in and of themselves, but can only serve as tests of a theory – and the
theory is what ultimately does the explaining. In a virtual experiment, a model is
developed that simulates some real-world social phenomenon, but with one or more
features modified to create experimental conditions that can be contrasted. For
example, the same business organization could be modeled multiple times, but
with a strong authority figure in one simulation, and a diffuse authority structure in
another (Carley and Gasser 1999). Whereas it would probably be impossible to
implement such an experiment with real-world societies, a computer model readily
allows such manipulation. When the model is started, the simulations that result
behave in ways that are argued to be analogous to how the real-world organization
would have behaved, in each of the different conditions. In this view, because the
simulation plays the role of a data-generating experiment, it doesn’t provide an
explanation; rather, it provides raw data to aid in theorizing, and the theory ulti-
mately provides the explanation.
12.3 Scientific Explanation
Explanations are attempts to account for why things happen – singular events or
regular, repeatable patterns. In the philosophy of science, there is a long history of
discussion surrounding scientific explanation, including the deductive-nomological
(D-N) or covering law approach (Hempel 1965), the statistical relevance approach
(Salmon 1971), and mechanistic approaches (Bechtel and Richardson 1993;
Salmon 1984). Here I limit the term “explanation” to causal explanation (cf. Little
1998; Woodward 2003). The relation between causation and explanation is
complex; some philosophers of science hold that all explanation must be causal,
whereas others deny this. For example, in the deductive-nomological tradition of
278 R.K. Sawyer
logical empiricism, laws are said to provide explanations, even though the status
of causation is questionable – causation is thought to be nothing more than an
observed regularity as captured by a covering law. In the more recent mechanistic
approach, in contrast, causation is central to explanation. I take the mechanistic
position that causal mechanism is central to explanation, but I first briefly summa-
rize the covering law notion of explanation.
In the covering law approach, a phenomenon is said to be explained when salient
properties of the event are shown to be consequents of general laws, where
the antecedents can also be identified. The phenomenon is said to be explained by
the combination of the antecedent conditions and the laws that then result in the
phenomenon. A strength of the covering-law approach is that laws both explain
and predict; once a law is discovered, it can be used both to explain past phenomena
and also to predict when similar phenomena will occur in the future.
Covering law models have always been problematic in the social sciences,
primarily because of difficulty translating the notion of “law” to social reality.
After all, advocates of the covering law model have had trouble adequately defining
“law” even in the physical world (Hempel 1965). Candidates for social laws always
have exceptions, and laws with exceptions are problematic in the DN approach.
There is a history of debate concerning whether social laws exist at all, with
prominent social theorists such as Anthony Giddens arguing that there are no social
laws (Giddens 1984), and other prominent social theorists arguing that there are
(e.g. Peter Blau (1977, 1983)). Philosophers of social science have taken various
positions on the status of social laws (Beed and Beed 2000; Kincaid 1990; Little
1993; McIntyre 1996). Much of this discussion centers on what constitutes a law:
must it be invariant and universal (Davidson’s (1980) “strict law”), or can it admit
of some exceptions? Even the strongest advocates of lawful explanation admit that
there are no strict laws in the social sciences; these laws will typical have
exceptions, and the law cannot explain those exceptions.
In the last decade or so, philosophers of biology (Bechtel 2001; Bechtel and
Richardson 1993; Craver 2001, 2002; Glennan 1996; Machamer et al. 2000) and
philosophers of social science (Elster 1989; Hedström 2005; Hedström and Swedberg
1998; Little 1991, 1998; Stinchcombe 1991) have begun to develop a different
approach to explanation, one based on causal mechanisms rather than laws. In the
mechanism approach, a phenomenon is said to be explained when the realizing
mechanism that gave rise to the phenomenon is sufficiently described. Mechanistic
accounts of explanation are centrally concerned with causation. For example,
Salmon’s (1984, 1994, 1997) causal mechanical model focuses on causal processes
and their causal interactions; an explanation of an event traces the causal processes
and interactions leading up to that event, and also describes the processes and
interactions that make up the event.
Hedström (2005) presented an account of how social simulations correspond to
mechanistic explanations. Mechanistic explanations differ from covering-law
explanations by “specifying mechanisms that show how phenomena are brought
about” (p. 24). Of course, how one defines “mechanism” is the crux of the approach.
Some theorists believe that mechanisms provide causal explanations, whereas
others do not. But what the approaches share is that “a mechanism explicates the
details of how the regularities were brought about” (p. 24). Rather than explanation
in terms of laws and regularities, a mechanism approach provides explanations by
postulating the processes constituted by the operation of mechanisms that generate
the observed phenomenon. For Hedström, “A social mechanism. . .describes a
constellation of entities and activities that are organized such that they regularly
bring about a particular type of outcome” (p. 25). The explanation is provided by
the specification of often unobservable causal mechanisms, and the identification of
the processes in which they are embedded. In other words, mechanists are willing to
grant that macro-level regularities are observed; the covering-law approach to
sociology has, after all, resulted in the identification of law-like regularities that
are empirically supported. But for a mechanist, a covering law does not explain:
“correlations and constant conjunctions do not explain but require explanation by
reference to the entities and activities that brought them into existence” (p. 26).
12.4 MABS: Explaining by Simulating
Using MABS, researchers have begun to model the mechanisms whereby macro
social properties emerge from interacting networked agents. A MABS contains
many autonomous computational agents that negotiate and collaborate with each
other, in a distributed, self-organizing fashion. The parallels with causal mechanism
approaches in the philosophy of science are striking (Sawyer 2004).
Hedström (2005) refers to his social simulation method as empirically calibrated
agent-based models (ECA) to emphasize that the models should be grounded in
quantitative empirical data. His recommended method is to (1) develop a stylized
agent-based model “that explicates the logic of the mechanism assumed to be
operative” (p. 143); (2) use relevant data to verify the mechanism actually works
this way; (3) run the model and modify it until it best matches relevant data. “Only
when our explanatory account has passed all of these three stages can we claim to
have an empirically verified mechanism-based explanation of a social outcome”
(p. 144). Hedström provides an extended demonstration of the ECA method by
modeling how social interactions might have given rise to the increase in youth
unemployment in Stockholm in the 1990s. His model includes exactly as many
computational agents as there were unemployed 20–24-year-olds in Stockholm
during 1993–1999: 87,924. The demographic characteristics of these computational
agents were an accurate reflection of their real-world counterparts. He then created
a variety of different simulations as “virtual experiments” to see which simulation
resulted in the best match to the observed empirical data.
The Stockholm simulation falls at the more specific end of the explanatory
spectrum; even if it successfully simulates unemployment in Stockholm, it may
not be helpful at understanding unemployment in any other city. Other MABS are
designed to provide more general explanations. For example, many MABS have
explored one of the most fundamental economic and sociological questions: What is
280 R.K. Sawyer
the origin of social norms? For example, how do norms of cooperation and trust
emerge? If autonomous agents seek to maximize personal utility, then under what
conditions will agents cooperate with other agents? In game theory terms, this is a
prisoner’s dilemma problem. Many studies of cooperation in artificial societies have
been implementations of the iterated prisoner’s dilemma (IPD), where agents
interact in repeated trials of the game, and agents can remember what other agents
have done in the past (Axelrod 1997).
The sociologists Macy and Skvoretz (1998) developed an artificial society to
explore the evolution of trust and cooperation between strangers. In prior
simulations of the prisoner’s dilemma, trust emerged in the iterated game with
familiar neighbors, but trust did not emerge with strangers. Macy and Skvoretz
hypothesized that if the agents were grouped into neighborhoods, norms of trust
would emerge among neighbors within each neighborhood, and that these norms
would then extend to strangers. Their simulation contained 1,000 agents that played
the prisoner’s dilemma game with both familiar neighbors and with strangers. To
explore the effects of community on the evolution of PD strategy, the simulation
defined neighborhoods that contained varying numbers of agents – from nine agents
per neighborhood to 50. Different runs of the simulation varied the embeddedness
of interaction: the probability that in a given iteration, a player would be interacting
with a neighbor or a stranger. These simulations showed that conventions for
trusting strangers evolved in neighborhoods of all sizes, as long as agents interacted
more with neighbors than strangers (embeddedness greater than 0.5). The rate of
cooperation among strangers increased linearly as embeddedness was raised from
0.5 to 0.9. Simulations with smaller neighborhoods resulted in a higher rate of
cooperation between strangers: at 0.9 embeddedness, the rate of cooperation
between strangers was 0.62 in the 10-member neighborhood simulation, and 0.45
in the 50-member neighborhood simulation (p. 655).
Macy and Skvoretz concluded that these neighborhoods – characterized by
relatively dense interactions – allow conventions for trusting strangers to emerge
and become stable and then diffuse to other neighborhoods via weak ties. If an
epidemic of distrusting behavior evolves in one segment of the society, the large
number of small neighborhoods facilitates the restoration of order (p. 657). This
simulation demonstrates how social structure can influence micro-to-macro emer-
gence processes; cooperation with strangers emerges when agents are grouped into
neighborhoods, but not when they are ungrouped.
An advocate of the causal mechanist approach to explanation would argue that
the Macy and Skvoretz simulation provides candidate explanations of several social
phenomena. First, the simulation explains how norms of cooperation could emerge
among friends in small communities – because exchanges are iterated, and agents
can remember their past exchanges with each other, they learn that cooperation
works to everyone’s advantage. Second, the simulation explains how norms of
cooperation with strangers could emerge – as local conventions diffuse through
weak ties. And in addition, the simulation explains how several variables contribute
to these effects – variables like the size of the neighborhood and the embeddedness
of each agent.
Advocates of a covering-law approach to explanation might prefer to think in

terms of lawful generalizations. The above simulation suggests at least two: first,
cooperation among strangers is greater when the neighborhoods are smaller, and
second, cooperation among strangers increases linearly with embeddedness. In a
D-N empiricist approach, such laws could be hypothesized and then tested through
empirical study of existing human societies, and no understanding of the causal
mechanism would be necessary. A mechanist like Hedström (2005) would counter
that the identification of empirically supported lawful relations does not constitute
an explanation. One has not identified a causal explanation until one has identified
the underlying social mechanisms that realize the regularities captured by the law.
The Macy and Skvoretz simulation helps to provide this form of causal explanation.
12.5 Potential Limitations of Simulations as Explanations
I am sympathetic to the mechanism approach. As Hedström points out, “It tends to

produce more precise and intelligible explanations” (Hedström 2005, p. 28); this is
desirable for sociology, which I believe must work toward being an empirically
grounded and theoretically rigorous science. It reduces theoretical fragmentation,
because a single mechanistic account might potentially explain many different
observed phenomena, from crime to social movements; think of Gladwell’s best
seller The Tipping Point (Gladwell 2000). And finally, knowing the underlying
mechanism allows you to make a certain kind of causal claim, whereas covering
law approaches give you essentially only correlations.
However, causal mechanist accounts of scientific explanation can be epistemi-
cally demanding. For example, many behaviors of a volume of gas can be explained
by knowing a single number, its pressure; yet a mechanist account requires the
identification of the locations and movements of all of the contained molecules.
A strict focus on mechanistic explanation would hold that the ideal gas law does not
explain the behavior of a volume of gas; only an account in terms of the individual
trajectories of individual molecules would be explanatory. And even that would be
an incomplete explanation, because the gas would manifest the same macroscopic
behavior even if the individual molecules had each taken a different trajectory;
certainly, an explanation should be able to account for these multiple realizations.
Many advocates of mechanism are unwilling to accept any place for the covering
law approach: they argue that mechanisms are the only proper form of sociological
explanation. In several publications, Sawyer (2003a, 2004, 2005) has described a
class of social phenomena in which mechanism-based explanations would be of
limited usefulness: when macro-level regularities are multiply realized in a wide
range of radically different underlying mechanisms. In such a situation, an account
of the mechanism underlying one realization of the regularity would not provide an
explanation of the other instances of that regularity. Hedström’s Stockholm simu-
lation does not necessarily explain the rise in youth employment in any other city,
for example; other cities might have different realizing mechanisms. But even
282 R.K. Sawyer
though it might not be possible to develop a single simulation that accurately

represents unemployment processes in all large cities, it might nonetheless be
possible to develop a law (or set of laws) that was capable of explaining (in the
deductive-nomological sense of the term) unemployment in a large number of
cities.
A social mechanist account often requires information that is unavailable, or that
science is unable to provide. The Stockholm simulation was only possible because
of the availability of detailed data gathered by the Swedish government. But
covering law explanations can be developed with much less data, or with data at
a much larger grain size. For example, many behaviors of a society can be
explained by knowing whether it is individualist or collectivist (Markus and
Kitayama 1991; Triandis 1995). Such properties figure in lawful generalizations
like “individualist societies are more likely to be concerned with ownership of
creative products” (Sawyer 2006) and “collectivist societies are more likely to
practice co-sleeping” (Morelli et al. 1992). In contrast to such simple and easy-
to-understand regularities, a mechanist explanation of the same patterns requires
quite a bit of knowledge about each participant in that society, and their interactions
with each other.
Even if a very good social simulation were developed, it might be very difficult to
use that simulation to communicate to a broad, non-technical audience what mean-
ing or understanding to attribute to the phenomenon (or the simulation). And in
extremely complex systems like human societies, it may be impossible to develop an
explanation of macro phenomena in terms of individual actions and interactions,
even though we may all agree that such processes nonetheless must exist at the
individual level. The issue here is identifying the right level of description, and the
mechanistic or realizing level is often too detailed to provide us with understanding.
There are many cases in science where it seems that reduction is not the best strategy
for scientific explanation. For example, higher-level events like mental events
supervene on physical processes but do not seem to be reducible to a unique set of
causal relationships in terms of them.
The most accurate simulation would come very close to replicating the natural
phenomenon in all its particulars. After such a simulation has been successfully
developed, the task remains to explain the simulation; and for a sufficiently detailed
simulation, that could be just as difficult as the original task of explaining the data
(Cilliers 1998). Computer programmers often have difficulty explaining exactly
why their creations behave as they do, and artificial society developers are no
different. Mechanistic accounts of explanation need to more directly address issues
surrounding levels of explanation and epistemic and computational limits to human
explanation and understanding (see Sawyer 2003a, 2004).
Social simulation unavoidably touches on the unresolved sociological issue of
how explanation should proceed. Social simulations represent only individual agents
and their interactions, and in this they are methodologically individualist (Conte
et al. 2001; also see Drennan 2005). Methodological individualism is a sociological
position that has its roots in the nineteenth century origins of sociology; it argues that
sociology should proceed by analyzing the individual participants in the social
system, then their relations and the behaviors of bigger system components, and
all the way up until we have an explanation of the social system. But if there are
real emergent social properties, with downward causal powers over component
individuals, then methodologically individualist simulation will fail to provide
explanations of those social phenomena – for essentially the same reasons that
philosophers of mind now believe that physicalism is inadequate to explain mental
phenomena (see Sawyer 2002). Some social properties – such as the property of
“being a collectivist society” or “being a church” – are multiply realized in widely
different social systems. A simulation of a realizing mechanism of one instance of
“being a church” would explain only one token instance, but would fail to broadly
explain the full range of mechanisms that could realize the social property. To return
to the Macy and Skvoretz simulation, the norm of cooperation could emerge in many
other realizing social systems, yet the norm might have the same downward causal
effects regardless of its realizing mechanism. If so, then a simulation of one realiza-
tion is only a partial explanation of a more general social phenomenon; it does not
explain the other ways that human cooperative behavior could be realized.
Social simulations which contain only individual agents deny a sociological
realism that accepts social properties as real. If macro social properties are real, then
they have an ontological status distinct from their realizing mechanisms, and may
participate in causal relations (this point continues to be actively debated and the
arguments are complex; see Sawyer 2003b). An accurate simulation of a social
system that contains multiply realized macro social properties would have to
represent not only individuals in interaction, but also these higher-level system
properties and entities (Sawyer 2003a).
The problem is that although a social simulation may provide a plausible account
of how individual actions and interactions give rise to an emergent macro pattern,
it is hard to know (1) whether or not that social simulation in fact captures the
empirical reality of the emergence; and more critically, (2) even if all agree that the
social simulation accurately captures an instance of the emergence of a macro
pattern from a network of individual agents, there may be other networks and other
emergence processes that could also give rise to the same macro pattern.
Issue (1) is the issue of validation and it is addressed in another chapter in this
same volume (David 2013). My concern here is with issue (2), which Sawyer
(2005) called multiple realizability: a social simulation may accurately represent
the social mechanisms by which individual actions together give rise to an emer-
gent macro phenomenon. But for a given macro social phenomenon, there could
potentially be many different networks of people, acting in different combinations,
that result in different emergence processes that lead to the same macro social
phenomenon. If so, the social simulation would not provide a complete explanation
of the macro social phenomenon, but instead would only provide a limited and
partial explanation.
The usual response to the multiple realizability issue is to argue that in many
cases, the alternate realizations of the macro phenomena are not significantly
different from each other. After all, every basketball team has five different players,
but the fact that the five positions are occupied by different human beings does not
284 R.K. Sawyer
substantially change the possible ways that the five can interact, and the possible
ways that plays emerge from the interactions of the five individuals. Some pairs of
realization are quite similar to each other, so similar that understanding one reali-
zation is tantamount to understanding the other one of the pair – without necessarily
developing an entirely distinct social simulation.
The problem with this response is that it fails in the face of wild disjunction
(Fodor 1974): when a given macro phenomenon has multiple realizations, and those
realizations have no lawful relations with one another. If the multiple realizations
are wildly disjunctive, then a social simulation of one of the realizations does not
provide us with any understanding beyond that one realization. We are still left
needing explanations of all of the other realizations. In many cases, wild disjunction
is related to functionalist arguments (and the argument originated from functionalist
perspectives in the philosophy of mind: Fodor 1974). “Being a church” is more
likely to be multiply realized in wildly disjunctive fashion, if “church” is defined in
terms of the functional needs it satisfies for its society rather than in terms of
structural features internal to the institution.
The mechanist could respond to wild disjunction concerns by empirically
identifying all of the different ways that the macro phenomenon in question might
emerge, and then developing a suite of social simulations, each one of which would
represent one of the realizing mechanisms. Then, could we say that the suite of social
simulations, together, constituted a complete explanation of the macro phenomenon?
I think so, although I would prefer to speak of a suite of “explanations” rather than to
call the set a single explanation. The Stockholm unemployment simulation might
only work for societies with generous social welfare systems; but then, another
simulation could be developed for stingier governments, and two simulations
together is not oppressively large given the outcome that unemployment everywhere
is now fully explained.
The suite-of-simulations approach works fine as long as the number of realizing
social simulations is manageable. But at some point, a suite of simulations would
become so large that most sociologists would agree that it provided limited under-
standing of a phenomenon. Is 200 too many to be a meaningful explanation? Could
as few as 20 still be too many? Even if it is computationally plausible and the
number of person-hours required to develop all of the simulations is not excessive,
it might nonetheless be of questionable value to undertake that work: because
another path toward social explanation is available: the covering law model.
Many philosophical advocates of mechanism believe that mechanistic explanation
is compatible with the existence of higher-level laws. Mechanisms are said to explain
laws (Beed and Beed 2000; Bunge 2004; Elster 1998). Bunge (2004) and Little
(1998) argued that causal mechanistic accounts are fully compatible with covering
law explanations; the mechanisms do the explanatory work, and the covering laws
provide a convenient shorthand that is often useful in scientific practice. However,
it is possible that social laws may exist that are difficult to explain by identifying
realizing mechanisms – in those cases where the laws relate wildly disjunctive,
multiply realized social properties. If so, the scope of mechanistic explanation
would be limited.
Many sociological theorists use the philosophical notion of emergence to argue

that collective phenomena are collaboratively created by individuals, yet are not
reducible to individual action (Sawyer 2005). In the social sciences, emergence
refers to processes and mechanisms of the micro-to-macro transition. Many of these
accounts argue that although only individuals exist, collectives possess emergent
properties that are irreducibly complex and thus cannot be reduced to individual
properties. Thus they reject sociological realism and are methodologically collec-
tivist. Other accounts argue that emergent properties are real.
The resolution to the apparent contradiction between mechanistic explanation
and social emergence is to develop a sufficiently robust account of emergence so that
mechanistic explanation and lawful explanation can be reconciled. Sawyer (2002,
2003b) proposed a version of emergence that he called nonreductive individualism
(NRI). Some emergent social properties may be real, and may have autonomous
causal powers, just like real properties at any other level of analysis. Nonreductive
individualism argues that this is the case for social properties that are multiply
realized in wildly disjunctive mechanisms. To the extent that social properties are
multiply realized, artificial society simulations may be limited to the explanation of
individual cases that do not generalize widely, resulting in a case study approach
rather than a science of generalizable laws and theories. The emergentist nature of
NRI is compatible with a more limited form of mechanism, but one that is elaborated
in a sociologically realist direction – with the mechanisms containing explicit
models of social properties at levels of analysis above the individual.
If a social property is multiply realized in many different (methodologically
individualist) mechanisms, a mechanistic explanation of any one realizing instance
will have limited explanatory power – particularly if the social property participates
in causal relations across its multiple realizations. A covering law approach might be
necessary to capture generalizations of higher-level phenomena across different
realizing mechanisms. Alternately, a mechanism could be proposed which explicitly
models emergent social properties, in addition to individuals and their interactions.
Although almost all artificial societies are currently individualist – with no repre-
sentation of higher-level social properties – there is no reason why computer
simulations could not be extended to model both individuals and macrosocial
phenomena, apart from an implicit commitment to methodological individualism.
12.6 Conclusion
Social simulations are almost all methodologically individualist, in that they repre-
sent agents and their interactions, but not higher level entities or properties. In other
words, social simulations are representations of the realizing mechanisms of higher-
level social properties. More generally, almost all agent-based simulations are
representations of a realizing mechanism of some system-level phenomenon.
Whether or not a complex system can be explained at the level of its realizing
mechanisms, or requires explanation at the level of emergent macro properties,
286 R.K. Sawyer
is an empirical question (Sawyer 2005). For example, it cannot be known a priori

whether or not a given social property can be given a useful mechanistic explana-
tion in terms of individuals – nor whether a given social property can be adequately
simulated by representing only individuals and their interactions in the model.
If this “realizing mechanism” approach begins to seem limiting, then modelers
could respond by incorporating system-level entities into their simulations. For
example, sociologists could respond by developing simulations that contain the
terms and properties of macro sociology, in addition to individual properties and
relations. If macrosocial properties are indeed real, and have autonomous causal
powers, then to be empirically accurate any model would have to incorporate those
properties. Although the social mechanism approach is commonly associated with
methodological individualism – because its advocates assume that a social mecha-
nism must be described in terms of individuals’ intentional states and relations
(e.g. Elster 1998; Hedström 2005) – there is no reason why social simulations
cannot include systems and mechanisms at higher levels of analysis. The system
dynamics models of an earlier era focused on macrosocial properties; but with the
availability of multi agent technology, new hybrid simulations could be developed
that contain both societies of autonomous agents, and explicit simulations of
emergent macro social properties.
To explore how we should interpret and understand simulations, I have drawn
on contemporary philosophical accounts of explanation and of causal mechanism.
I conclude by cautioning against being overly confident that agent-based
simulations, at least those that are based on mechanistic assumptions, provide
complete explanations of a given system-level phenomenon. The explanation may
not be complete even in those cases where the simulation is well-conceived, is
grounded in empirical observation and theory, and generates emergent processes
that lead to empirically observed system outcomes. These successful simulations
should be considered to be explanations of a given realizing instance of an
emergence process, but not necessarily considered to be complete explanations
of the target system phenomenon.
The question for sociologists is ultimately: What path should sociology take? All
sociologists define their goal to be the explanation of macro social phenomena – of
groups and organizations, rather than of single individuals. Many sociologists
believe that they can explain macro social phenomena without attending to the
specific realizing mechanisms at the level of individuals and their interactions.
These sociologists define sociology as the science of an autonomous social level of
analysis. The social mechanists believe that this is the wrong approach; instead, the
goal of sociology should be to identify and characterize these individual realizing
mechanisms. Mechanists believe that there can be no autonomous science at the
macro social level of analysis. These are the issues to be faced as we attempt to
interpret and understand social simulations.
The question for complexity researchers more generally is the same: Can a
mechanistic approach provide a complete explanation, or will scientific explanation
need to incorporate some higher-level properties and entities? The answer to this
question has direct implications for modelers. If the mechanistic approach is
capable of providing a complete explanation of a given phenomenon, then a strict

agent-based approach is appropriate. But if it is necessary to incorporate higher-
level properties or entities, then agent-based simulations will need to include model
entities that represent higher levels of organization.
Further Reading
(Bechtel and Richardson 1993) provides a discussion of a range of philosophical

issues related to the likely success of reductionist strategies in understanding and
explaining complex systems, inspired by connectionist accounts of cognition, but
relevant to complex systems at any level of analysis. (Hedström 2005) makes a
strong case for reductionist explanation of social systems, using mechanistic expla-
nation and specifically, multi-agent based simulation in connection with empirical
study.
For an examination of the philosophical accounts of mechanistic explanation
and theories of emergence in sociology and philosophy see (Sawyer 2004). For an
extensive review of historical and contemporary theories of emergence in the social
sciences, primarily psychology and sociology see (Sawyer 2005). This advocates
that sociology should be the science of social emergence. (Conte et al. 2001) is a
discussion between four different viewpoints specifically as they concern social
simulation.
References
Axelrod R (1997) The complexity of cooperation: agent-based models of competition and collab-
oration. Princeton University Press, Princeton
Bechtel W (2001) The compatibility of complex systems and reduction: a case analysis of memory
research. Minds Machines 11:483–502
Bechtel W, Richardson RC (1993) Discovering complexity: decomposition and localization as
strategies in scientific research. Princeton University Press, Princeton
Beed C, Beed C (2000) Is the case for social science laws strengthening? J Theory Soc Behav
30(2):131–153
Blau PM (1977) A macrosociological theory of social structure. Am J Sociol 83(1):26–54
Blau PM (1983) Comments on the prospects for a nomothetic theory of social structure. J Theory
Soc Behav 13(3):265–271
Bunge M (2004) How does it work?: the search for explanatory mechanisms. Philos Soc Sci
34(2):182–210
Carley KM, Gasser L (1999) Computational organization theory. In: Weiss G (ed) Multiagent
systems: a modern approach to distributed artificial intelligence. MIT Press, Cambridge,
pp 299–330
Cilliers P (1998) Complexity and postmodernism: understanding complex systems. Routledge,
New York
Conte R, Edmonds B, Moss S, Sawyer RK (2001) Sociology and social theory in agent based
social simulation: a symposium. Comput Math Organ Theory 7(3):183–205
288 R.K. Sawyer
Craver C (2001) Role functions, mechanisms and hierarchy. Philos Sci 68:31–55
Craver CF (2002) Interlevel experiments and multilevel mechanisms in the neuroscience of
memory. Philos Sci 69(Supplement):S83–S97
Davidson D (1980) Essays on actions and events. Oxford University Press, New York
Drennan M (2005) The human science of simulation: a robust hermeneutics for artificial societies.
J Artif Soc Soc Simulat 8(1). http://jasss.soc.surrey.ac.uk/8/1/3.html
Elster J (1989) Nuts and bolts for the social sciences. Cambridge University Press, New York
Elster J (1998) A plea for mechanism. In: Hedström P, Swedberg R (eds) Social mechanisms: an
analytical approach to social theory. Cambridge University Press, Cambridge, pp 45–73
Fodor JA (1974) Special sciences (Or: the disunity of science as a working hypothesis). Synthese
28:97–115
Giddens A (1984) The constitution of society: outline of the theory of structuration. University of
California Press, Berkeley
Gladwell M (2000) The tipping point. Little, Brown, New York
Glennan SS (1996) Mechanisms and the nature of causation. Erkenntnis 44:49–71
Hedström P (2005) Dissecting the social: on the principles of analytic sociology. Cambridge
Hedström P, Swedberg R (eds) (1998) Social mechanisms: an analytical approach to social theory.
Cambridge University Press, Cambridge
Hempel CG (1965) Aspects of scientific explanation and other essays in the philosophy of science.
Free Press, New York
Kincaid H (1990) Defending laws in the social sciences. Philos Soc Sci 20(1):56–83
Little D (1991) Varieties of social explanation: an introduction to the philosophy of science.
Westview Press, Boulder
Little D (1993) On the scope and limits of generalization in the social sciences. Synthese
97(2):183–207
Little D (1998) Microfoundations, method and causation: on the philosophy of the social sciences.
Transaction Publishers, New Brunswick
Machamer P, Darden L, Craver CF (2000) Thinking about mechanisms. Philos Sci 67:1–25
Macy MW, Skvoretz J (1998) The evolution of trust and cooperation between strangers:
a computational model. Am Sociol Rev 63:638–660
Markovsky B (1997) Building and testing multilevel theories. In: Szmatka J, Skvoretz J, Berger J
(eds) Status, network, and structure: theory development in group processes. Stanford, Palo
Alto, pp 13–28
Markus HR, Kitayama S (1991) Culture and the self: implications for cognition, emotion, and
motivation. Psychol Rev 98(2):224–253
McIntyre LC (1996) Laws and explanation in the social sciences: defending a science of human
behavior. Westview, Boulder
Morelli GA, Rogoff B, Oppenheim D, Goldsmith D (1992) Cultural variation in infants’ sleeping
arrangements: questions of independence. Dev Psychol 28(4):604–613
Neumann M (2006) Emergence as an explanatory principle in artificial societies: reflections on the
bottom-up approach to social theory. In: Squazzoni F (ed) Epistemological aspects of computer
simulation in the social sciences: second international workshop, EPOS 2006, Brescia, 5–6 Oct
2006, revised selected and invited papers (Lecture notes in computer science), vol 5466.
Ostrom T (1988) Computer simulation: the third symbol system. J Exp Soc Psychol 24:381–392
Salmon W (1971) Statistical explanation. In: Salmon W (ed) Statistical explanation and statistical
relevance. University of Pittsburgh Press, Pittsburgh, pp 29–87
Salmon W (1984) Scientific explanation and the causal structure of the world. Princeton Univer-
sity Press, Princeton
Salmon W (1994) Causality without counterfactuals. Philos Sci 61:297–312
Salmon W (1997) Causality and explanation: a reply to two critiques. Philos Sci 64:461–477
Sawyer RK (2002) Nonreductive individualism, part 1: supervenience and wild disjunction. Philos
Soc Sci 32(4):537–559
Sawyer RK (2003a) Artificial societies: multi agent systems and the micro–macro link in socio-
logical theory. Sociol Methods Res 31(3):37–75
Sawyer RK (2003b) Nonreductive individualism, part 2: social causation. Philos Soc Sci
33(2):203–224
Sawyer RK (2004) The mechanisms of emergence. Philos Soc Sci 34(2):260–282
Sawyer RK (2005) Social emergence: societies as complex systems. Cambridge University Press,
New York
Sawyer RK (2006) Explaining creativity: the science of human innovation. Oxford University
Press, New York
Schmid A (2006) What does emerge in computer simulations? Simulation between epistemologi-
cal and ontological emergence. In: Squazzoni F (ed) Epistemological aspects of computer
simulation in the social sciences: second international workshop, EPOS 2006, Brescia, 5–6 Oct
2006, revised selected and invited papers (Lecture notes in computer science), vol 5466.
Stinchcombe AL (1991) The conditions of fruitfulness of theorizing about mechanisms in social
science. Philos Soc Sci 21(3):367–388
Triandis HC (1995) Individualism and collectivism. Westview Press, Boulder
Turner JH (1993) Classical sociological theory: a positivist perspective. Nelson-Hall, Chicago
Woodward J (2003) Making things happen: a theory of causal explanation. Oxford University
Press, New York
Part III
Mechanisms
Chapter 13
Utility, Games, and Narratives
Guido Fioretti
Why Read This Chapter? To appreciate how decision making can be modelled
in terms of utility maximisation and game theory. To understand some of the
paradoxes, limitations and major criticism of this approach and some of the
alternatives.
Abstract This chapter provides a general overview of theories and tools to model
decision-making. In particular, utility maximization and its application to collective
decision-making, i.e. Game Theory, are discussed in detail. The most important
exemplary games are presented, including the Prisoner’s Dilemma, the Game of
Chicken and the Minority Game, also known as the El Farol Bar Problem. After
discussing the paradoxes and pitfalls of utility maximisation an alternative
approach is introduced, which is based on appropriateness of decisions rather
than consequences. An assessment of the pros and cons of competing approaches
to modelling decision making concludes the chapter.
13.1 Introduction
This chapter provides a general overview of theories and tools to model individual
and collective decision-making. In particular, stress is laid on the interaction of
several decision-makers.
A substantial part of this chapter is devoted to utility maximization and its
application to collective decision-making, known as Game Theory. However, the
pitfalls of utility maximization are thoroughly discussed, and the radically alternative
approach of viewing decision-making as constructing narratives is presented with its
emerging computational tools. In detail, the chapter is structured as follows.
G. Fioretti (*)
University of Bologna, Department of Management Science, Via Capo di Lucca 34,
40126 Bologna, Italy

294 G. Fioretti
Section 13.2 presents utility maximization and Game Theory with its Nash
equilibria. The most important prototypical games are expounded in this section.
Section 13.3 presents games that are not concerned with Nash equilibria. Sec-
tion 13.4 illustrates the main paradoxes of utility maximization, as well as the
patches that have been proposed to overcome them. Section 13.5 expounds the
vision of decision-making as constructing a narrative, supported by an empirical
case-study. Section 13.6 aims at providing computational tools for this otherwise
literary vision of decision-making. Finally, Sect. 13.7 concludes by assessing the
pros and cons of competing approaches.
This chapter touches so many issues that a complete list of references to the
relevant literature would possibly be longer than the chapter itself. Instead of
references, a guide to the most relevant bibliography is provided at the end of the
chapter.
13.2 Utility and Games
Let fa1 ; a2 ; . . . am g be a set of alternatives. Let ai denote a generic alternative,

henceforth called the i-th alternative where i ¼ 1,2,. . .m.
By selecting an alternative, a decision-maker obtains one out of several possible
consequences. Let fci1 ; ci2 ; . . . cin g be the set of possible consequences of alterna-
tive ai . Let cij denote a consequence of ai , where i ¼ 1,2,. . .m and j ¼ 1; 2; . . . ni .
The expected utility of alternative ai is:
X
ni
uðai Þ ¼ pðcij Þuðcij Þ (13.1)
j¼1
where pðcij Þ is the probability of obtaining consequence cij and uðcij Þ is the utility of
consequence cij .
It is suggested that the one alternative should be chosen, that maximizes expected
utility. Frank Ramsey, Bruno De Finetti and Leonard Savage demonstrated that this is
the only choice coherent with a set of postulates that they presented as self-evident.
Among these postulates, the following ones have the strongest intuitive appeal:
Transitivity Transitivity of preferences means that if ai aj and aj ak , then
ai ak .
Independence Independence of irrelevant alternatives means that ai aj iff ai [ ak
aj [ ak ; 8ak .
Completeness Completeness means that 8ðai ; aj Þ , a preference relation is
defined.
Utility maximization is neither concerned with conceiving alternatives, nor with
the formation of preferences, which are assumed to be given and subsumed by the
utility function. Probabilities may eventually be updated by means of frequency
13 Utility, Games, and Narratives 295
measurement, but at least their initial values are supposed to be given as well. Thus,
utility maximization takes as solved many of the problems with which its critics are
concerned.
Utility maximization takes a gambler playing dice or roulette as its prototypical
setting. In fact, in this setting the set of alternatives is given, utilities coincide with
monetary prizes and probabilities can be assessed independently of utilities. By
contrast, for some critics of utility maximization, gambling is not an adequate
prototype of most real-life situations.
The interaction of several utility-maximizing decision-makers is covered by
Game Theory. Game Theory assumes that collective decision-making is the
combination of several individual decision processes, where each individual
maximizes his utility depending on the alternatives selected by the other
individuals. Since selecting an alternative implies considering what alternatives
other players may select, alternatives are generally called strategies in this
context.
Utility is called payoff in Game Theory. Games in which one player does better
at another’s expense are called zero-sum games. Games may be played once, or they
may be repeated.
The bulk of Game Theory is concerned with equilibria. If each player knows the
set of available strategies and no player can benefit by changing his or her strategy
while the other players keep theirs unchanged, then the current choice of strategies
and the corresponding payoffs constitute a Nash equilibrium. Since this implies
stepping in another player’s shoes in order to figure out what she would do if one
selects a particular strategy, Nash-equilibria are fixed points in self-referential loops
of the kind “I think that you think that I think . . .”.
Note that being at a Nash equilibrium neither implies that each player reaches
the highest possible payoff that she can attain, nor that the sum of all payoffs of all
players is the highest that can be attained. This is eventually a concern for
economics, for it implies that individual interests may not produce the common
good.
If a game is repeated, a Nash equilibrium may be either realized with pure
strategies, meaning that the players choose consistently one single alternative, or
mixed strategies, meaning that the players select one out of a set of available
strategies according to a probability distribution. Accepting the idea of mixed
strategies often allows to find Nash equilibria where there would be none if only
pure strategies are allowed. However, the realism of random decision-makers
choosing strategies according to a probability distribution, is at least questionable.
Most of the games analysed by Game Theory involve two, or in any case a very
limited number of players. On the contrary, evolutionary games (Weibull 1997)
concern large populations of players playing different strategies, that are subject to
an evolutionary dynamics regulated by replicator equations. Successful strategies
replicate and diffuse, unsuccessful strategies go extinct. Occasionally, new
strategies may originate by random mutation.
The equilibrium concept of evolutionary games is that of evolutionarily stable
strategies. An evolutionary stable strategy is such that, if almost every member of the
296 G. Fioretti
Opera Football
Opera 3, 2 0, 0 M, L 0, 0
Football 0, 0 2, 3 0, 0 L, M
Fig. 13.1 A payoff matrix for the Battle of the Sexes (left) and its generic representation (right).
The left number is the payoff of the row player (wife), the right number is the payoff of the column
player (husband). In this generic representation, L is the payoff of the least preferred alternative
whereas M is the payoff of the most preferred alternative
population follows it, no mutant can successfully invade. Alternatively, evolutionary

games may be played in order to observe typical dynamics, in which case they
become akin to the influence games that will be handled in Sect. 13.3.
The following games propose prototypical modes of human interaction. Games
used by experimental economics in order to evince human attitudes do not pertain
to this list.
13.2.1 The Battle of Sexes
Imagine a couple where the husband would like to go to the football game, whereas
the wife would like to go to the opera. Both would prefer to go to the same place
rather than different ones. They cannot communicate, so they cannot agree, e.g., to
go alternatively to the opera and to the football game.
The payoff matrix in Fig. 13.1 is an example of the Battle of Sexes, where the
wife chooses a row and the husband chooses a column. Aside, a generic represen-
tation of the game where L < M.
This representation does not account for the additional harm that might come
from going to different locations and going to the wrong one, i.e., the husband goes
to the opera while the wife goes to the football game, satisfying neither. Taking
account of this effect, this game would bear some similarity to the Game of Chicken
of Sect. 13.2.7.
This game has two pure-strategy Nash-equilibria, one where both go to the opera
and another where both go to the football game. Furthermore, there is a Nash
equilibrium in mixed strategies, where the players go to their preferred event more
often than to the other one.
None of these equilibria is satisfactory. One possible resolution involves a
commonly observed randomizing device, e.g., the couple may agree to flip a coin
in order to decide where to go.
Stag Hare
Stag 3, 3 0, 1 C, C S, B
Hare 1, 0 1, 1 B, S D, D
Fig. 13.2 A payoff matrix for the Stag Hunt (left) and its generic representation (right). The left
number is the payoff of the row player, the right number is the payoff of the column player. In this
generic representation, C is the payoff that accrues to both players if they cooperate, D is the
payoff that accrues to both players if they defect from their agreement, S is the sucker’s payoff and
B is the betrayer’s payoff
13.2.2 The Stag Hunt
Rousseau described a situation where two individuals agree to hunt a stag, which
none of them would be able to hunt alone. Each hunter may eventually notice a hare
and shoot at it. This would destroy the stag hunt, so the other hunter would get
nothing.
An example of the payoff matrix for the stag hunt is pictured in Fig. 13.2, along
with its generic representation. The stag hunt requires that C > B D > S.
This game has two pure-strategy Nash-equilibria, one where both hunters hunt
the stag, the other one where both hunters hunt a hare. The first equilibrium
maximizes payoff, but the second equilibrium minimizes risk. There exists also a
mixed-strategy Nash-equilibrium, but no payoff matrix can make the hunters play
“stag” with a probability higher than 1/2.
The Stag Hunt exemplifies the idea of society originating out of contracts between
individuals. The examples of “social contract” provided by Hume are Stag Hunts:
• Two individuals must row a boat. If both choose to row they can successfully
move the boat, but if one does not, the other wastes his effort.
• Two neighbours wish to drain a meadow. If they both work to drain it they will
be successful, but if either fails to do his part the meadow will not be drained.
Several animal behaviours have been described as stag hunts. For instance, orcas
corral large schools of fish to the surface and stun them by hitting them with their
tails. This works only if fishes do not have ways to escape, so it requires that all
orcas collaborate to kill all fishes they caught rather than catching a few of them.
13.2.3 The Prisoner’s Dilemma
The Prisoner’s Dilemma is a central subject in economics, for it apparently contradicts

its basic assumption that common good arises out of self-interested individuals. This
difficulty is eventually overcome by repeating the game.
298 G. Fioretti
Cooperate Defect
Cooperate 3, 3 0, 5 C, C S, B
Defect 5, 0 1, 1 B, S D, D
Fig. 13.3 A payoff matrix for the Prisoner’s Dilemma (left) and its generic representation (right).
The left number is the payoff of the row player, right number is the payoff of the column player. In
this generic representation, C is the payoff if both players cooperate, D is the payoff if both defect
from their agreement, S is the sucker’s payoff, B is the betrayer’s payoff
The basic formulation of the Prisoner’s Dilemma is as follows. Two suspects,

A and B, are arrested by the police. Having insufficient evidence for a conviction,
the police visits them separately, offering each prisoner the following deal: if one
testifies for the prosecution against the other and the other remains silent, the
betrayer goes free and the silent accomplice receives the full 10-year sentence.
If both stay silent, both prisoners are sentenced to only 6 months in jail for
a minor charge. If each betrays the other, each receives a 5-year sentence. Each
prisoner must make the choice of whether to betray the other or to remain silent;
unfortunately, neither prisoner knows what choice the other prisoner made.
The Prisoner’s Dilemma describes any situation where individuals have an
interest to be selfish, though if everyone cooperates a better state would be attained.
Examples may include unionising, paying taxes, not polluting the environment, or
else. Figure 13.3 illustrates a payoff matrix for the Prisoner’s Dilemma, as well as
its generic representation. The Prisoner’s Dilemma requires that B > C > D > S.
The Prisoner’s Dilemma has only one Nash equilibrium at (D, D). Notably, all
individual incentives push towards this equilibrium. Nevertheless, this equilibrium
is not socially optimal.
Eventually, the difficulty raised by the Prisoner’s Dilemma can be overcome if
players can repeat the game (which requires 2C > B + S). In particular, by playing
the Prisoner’s Dilemma as an evolutionary game with large numbers of players and
strategies it is possible that islands of cooperation sustain themselves in a sea of
selfish choices. One possibility for islands of cooperation to emerge is to allow
reciprocity, e.g., with a “tit-for-tat” strategy: start with cooperating whenever you
meet a new player, but defect if the other does. Another possibility is that players
cooperate when they meet players that exhibit a randomly selected tag – e.g., a tie
may be worn in order to inspire confidence – so that islands of cooperation emerge
even if agents have no memory.
13.2.4 The Traveller’s Dilemma
The Traveller’s dilemma is a non-zero-sum game in which two players attempt to

maximize their own payoff, without any concern for the other player’s payoff. It is a
game that aims at highlighting a paradox of rationality. It is a thought experiment on

the following problem.
An airline loses two suitcases belonging to two different travellers. The suitcases
contain identical antiques. An airline manager tasked to settle the claims of both
travellers explains that the airline is liable for a maximum of $100 per suitcase, and
in order to determine a honest appraised value of the antiques the manager separates
both travellers and asks each of them to write down the amount of their value at no
less than $2 and no more than $100. He also tells them that if both write down the
same number, he will treat that number as the true value of both suitcases and
reimburse both travellers that amount. However, if one writes down a smaller
number than the other, this smaller number will be taken as the true value, and
both travellers will receive that amount along with a bonus/malus: $2 extra will be
paid to the traveller who wrote down the lower value and a $2 deduction will be
taken from the person who wrote down the higher amount. The challenge is: what
strategy should both travellers follow in order to decide what value they should
write down?
If this game is actually played, nearly all the times everyone chooses $100 and
gets it. However, rational players should behave differently.
Rational players should value the antique slightly less than their fellow traveller,
in order to get the the bonus of $2. For instance, by pricing at $99 one would get
$101, whereas the opponent would get $97. However, this triggers an infinite
regression such that $2 is the only Nash-equilibrium of this game. Thus, being
rational does not pay.
The Traveller’s Dilemma suggests that in reality people may coordinate and
collaborate because of their bounded rationality, rather than in spite of it. If they
would be smarter than they are, they would obtain less.
13.2.5 The Dollar Auction
The dollar auction is a non-zero sum sequential game designed to illustrate a

paradox brought about by rational choice theory. In this game, players with perfect
information are compelled to make an ultimately irrational decision based on a
sequence of rational choices.
The game involves an auctioneer who offers a one-dollar bill with the following
rule: the dollar goes to the highest bidder, who pays the amount he bids. The
second-highest bidder must also pay the highest amount that he bids, but gets
nothing in return.
Suppose that the game begins with one of the players bidding 1 cent, hoping to
make a $0.99 profit. He will be quickly outbid by another player bidding 2 cents, as
a $0.98 profit is still desirable. Similarly, another bidder may bid 3 cents, making a
$0.97 profit. At this point the first bidder may attempt to convert their loss of 1 cent
into a gain of $0.97 by also bidding 3 cents. In this way, a series of bids is
maintained.
300 G. Fioretti
One may expect that the bidders end up with offering $1.00 for a one-dollar bill,
which is what the auction is for. However, a problem becomes evident as soon as
the bidding reaches 99 cents. Suppose that one player had bid 98 cents. The other
players now have the choice of losing 98 cents or bidding a dollar even, which
would make their profit zero. After that, the original player has a choice of either
losing 99 cents or bidding $1.01, losing only 1 cent. After this point these rational
players continue to bid the value up well beyond the dollar, and neither makes a
profit.
13.2.6 Pure Coordination Games
Pure coordination games are an empirical puzzle for Game Theory. Pure coordina-
tion games are one-shot games where players face a set of alternatives knowing that
a positive payoff will only accrue to them if they coordinate on the same choice.
For instance, two subjects may be shown a city map and asked, independently of
one another, to select a meeting point. Or, subjects may be asked to select a positive
integer. In the first case they obtain a positive payoff if they select the same meeting
point; in the second case, if they select the same integer.
The difficulty of pure coordination games derives from the fact that players
cannot communicate and that the game is not repeated. The astonishing fact about
pure coordination games is that, if they are actually played, players reach an
agreement much more often than they would if they would play randomly.
The commonly held explanation is that pure coordination games generally entail
cues that single out one choice as more “salient” than others. For instance, subjects
asked to select a meeting point generally end up with the railway station, whereas
the majority of those asked to name a positive integer select the number 1.
However, this suggests that coordination may eventually be attained because of
conventions, habits or values that do not enter the description of decision settings.
People may not even be aware of what makes them coordinate with one another.
13.2.7 The Game of Chicken
The game of Chicken models two drivers, both headed for a single lane bridge from
opposite directions. One must swerve, or both will die in the crash. However, if one
driver swerves but the other does not, he will be called a “chicken”. Figure 13.4
depicts a typical payoff matrix for the Chicken Game, as well as its generic form.
Chicken is an anti-coordination game with two pure-strategy Nash-equilibria
where each player does the opposite of what the other does. Which equilibrium is
selected depends very much on the effectiveness in signaling pre-commitment
before the game is played. For instance, a driver who disables the brakes and the
Swerve Straight
Swerve 0, 0 −1, +1 V/2, V/2 0, V
Straight +1, −1 −10, −10 V, 0 (V−C), (V−C)

2 2
Fig. 13.4 A payoff matrix for the Game of Chicken (left) and its generic representation (right).
The left number is the payoff of the row player, the right number is the payoff of the column
player. In this generic representation, V is the value of power, prestige, or of the available resource
to be obtained, C is the cost if both players choose “straight”
steering wheel of his car and makes it known to the other driver, may induce him to
swerve.
Bertrand Russell remarked that the nuclear stalemate was much like the Game of
Chicken1:
As played by irresponsible boys, this game is considered decadent and immoral, though
only the lives of the players are risked. But when the game is played by eminent statesmen,
who risk not only their own lives but those of many hundreds of millions of human beings,
it is thought on both sides that the statesmen on one side are displaying a high degree of
wisdom and courage, and only the statesmen on the other side are reprehensible. This, of
course, is absurd. Both are to blame for playing such an incredibly dangerous game. The
game may be played without misfortune a few times, but sooner or later it will come to be
felt that loss of face is more dreadful than nuclear annihilation. The moment will come
when neither side can face the derisive cry of ’Chicken! ’ from the other side. When that
moment is come, the statesmen of both sides will plunge the world into destruction.
The Game of Chicken has been re-interpreted in the context of animal behaviour.
It is known as Hawk-Dove Game among ethologists, where the Hawk-Dove game has
the same payoff matrix as in Fig. 13.4. In the Hawk-Dove game, “swerve” and
“straight” correspond to the following strategies, respectively:
Dove Retreat immediately if one’s opponent initiates aggressive behaviour;
Hawk Initiate aggressive behaviour, not stopping until injured or until the opponent
backs down.
Whilst the original Game of Chicken assumes C > V and cannot be repeated, the
Hawk-Dove game lacks this requirement and is generally conceived as an evolu-
tionary game.
The strategy “Dove” is not evolutionary stable, because it can be invaded by a
“Hawk” mutant. If V > C, then the strategy “Hawk” is evolutionarily stable. If
V < C there is no evolutionarily stable strategy if individuals are restricted to
following pure strategies, although there exists an evolutionarily stable strategy if
players may use mixed strategies.
1
Russell BW (1959) Common Sense and Nuclear Warfare. George Allen/Unwin, London.
302 G. Fioretti
13.2.8 The War of Attrition
The war of attrition is a game of aggression where two contestants compete for a
resource of value V by persisting with their intentions while constantly
accumulating costs. Equivalently, this game can be seen as an auction in which
the prize goes to the player with the highest bid Bh, and each player pays the loser’s
low bid Bl .
The war of attrition cannot be properly solved using its payoff matrix. In fact, the
players’ available resources are the only limit to the maximum value of bids. Since
bids can be any number, if available resources are ignored, then the payoff matrix
has infinite size. Nevertheless, its logic can be analysed.
Since players may bid any number, they may even exceed the value V that is
contested over. In fact, if both players bid higher than V, the high bidder does not so
much win as lose less, in the sense that Bl < V Bh < 0 – a Pyrrhic victory.
Since there is no value to bid which is beneficial in all cases, there is no dominant
strategy. However, this does not preclude the existence of Nash-equilibria. Any pair
of strategies such that one player bids zero and the other player bids any value equal
to V or higher, or mixes among any values V or higher, is a Nash-equilibrium.
The War of Attrition is akin to a Chicken or Hawk-Dove game – see Sect. 13.2.7 –
where if both players choose “swerve”/“Dove” they obtain 0 instead of V/2 as in
Fig. 13.4.
The evolutionarily stable strategy when playing it as an evolutionary game is a
probability density of random persistence times which cannot be predicted by the
opponent in any particular contest. This result has led to the conclusion that, in this
game, the optimal strategy is to behave in a completely unpredictable manner.
13.3 Influence Games
Contrary to those of Sect. 13.2, the games in this section are not concerned with
Nash-equilibria. Players are not assumed to figure out which alternatives the other
players might choose, originating infinite regressions that can only stop at equilib-
rium points.
Rather, boundedly rational players are assumed to follow certain rules, that may
be quite simple but need not be necessarily so. The game then concerns what
collective behaviours emerges out of mutual influence.
However, the games in this section are not so different from those of Sect. 13.2
when they are played as evolutionary games. Such is the case, for instance, of
simulations where a large number of players iterate the Prisoner’s Dilemma.
Two prototypical games will be expounded in this section. The Ising model
(originally developed in physics, where it is also known as spin glass model) is
concerned with imitation. The minority game, also known as the El Farol Bar
Problem, is concerned with the contrary of imitation. It is about doing the opposite
of what others do.
13.3.1 The Ising Model
The Ising model was originally developed in physics in order to study the interac-
tion between atoms in a ferromagnetic material. For this reason its agents can only
take two states, or opinions in social applications, and are fixed in space.
The Ising model is an exceedingly stylized model of imitation dynamics. Clearly,
many imitation models are more complex and more realistic than the Ising model.
However, the closed-form solutions of the Ising model may guide the builder of more
complex models in the process of understanding their behaviour.
In general, the Ising model is not presented as a game. It is done here in order to
stress its symmetry with the minority game.
Let N players be denoted by means of an index i ¼ 1,2,. . .N. Players must
choose between an alternative A ¼ 1 and an alternative A ¼ 1.
The payoff of a player does not only depend on the alternative that she has
chosen, but also on the average of the alternatives chosen by the other players. Let
m denote this average.
Since we want to reproduce situations where the individual follows the herd,
the effect of m should be the stronger, the more homogeneous the group. Since
A2{1;1} and consequently m2{1;1}, we can reach this goal by requiring that
the payoff depends on a term A m. This term may eventually be multiplied by a
coefficient J > 0.
A stochastic term ε is necessary in order to understand our game as a system
jumping between many equilibria. This term will disappear when expected values
will be taken.
In the end, the following functional form is chosen for the payoff of a player:
uðAÞ ¼ vðAÞ þ J A m þ ε (13.2)
where u(A) is the total payoff of a player and v(A) is its individual component.
Furthermore, let us assume that this individual component takes the following
form:
(
h if A ¼ 1
vðAÞ ¼ (13.3)
h if A ¼ 1:
where h 2 <; h > 0.

By assuming that the stochastic terms εi are Gumbel-distributed, we can apply
the logit model. By combining Eqs. 13.2 and 13.3 we derive the following
expressions for the probability that a player selects one of the two alternatives:
304 G. Fioretti
eμðhJmþεÞ
pfA ¼ 1g ¼ (13.4)
eμðhJmþεÞ þ eμðhþJmþεÞ
eμðhþJmþεÞ
pfA ¼ 1g ¼ (13.5)
eμðhJmþεÞ þ eμðhþJmþεÞ
The expected value of the selected alternative is E{A} ¼ 1p{A ¼ 1} þ

1p{A ¼ 1}. Since it is also E{A} ¼ m we obtain the following expression:
m ¼ tanhðμh þ μ JmÞ (13.6)
where tanhðxÞ ¼ ðex ex Þ=ðex þ ex Þ is the hyperbolic tangent.

Equation 13.6 provides an analytic description of a game with herd behaviour on
two alternatives described by means of a mean-field approximation. It admits a
closed form solution that provides the following findings:
• If μ J < 1 and h ¼ 0 there exists one single solution at m ¼ 0. Consider that this
is a discrete-time system, so its attractors are stable if all eigenvalues of the state
transition function are in (1,1). Intuitively, μ J < 1 means that this system is
globally stable. Furthermore, h ¼ 0 means that the individual component of the
payoff is zero so the players have no incentive to choose one of the two
alternatives. Consequently, the stochastic term makes m ¼ 0 the only solution.
• If μ J < 1 and h 6¼ 0 there exists one single solution with the same sign as h. In
fact, as in the previous case the system is globally stable so it admits one single
solution. However, since in this case the players’ payoff includes an individual
component, it is this component that determines what equilibriun is reached. If
most players prefer A ¼ 1 the equilibrium will be m 1; likewise, if most
players prefer A ¼ 1 then the equilibrium will be m 1.
• If μ J 1 and h ¼ 0 there exist three solutions: m ¼ 0 and m ¼ m(μ J).
In fact, the system is globally unstable but locally stable equilibria may exist.
Since the individual component of the payoff is zero, the system may either tend
towards m ¼ 0, or m 1, or m 1.
• If μ J 1 and h 6¼ 0, the following subcases must be distinguished:
– If, for any given μ and J, there exists a threshold H(h) > 0 such that |h| H,
then three solutions exist, one with the same sign as h and the other two with
opposite sign. Condition |h| < H means that the individual component of the
payoff is limited even if not zero. Therefore, results are similar to the previous
case.
– If, for any given μ and J, there exists a threshold H(h) > 0 such that |h| > H,
then there exists one single solution with the same sign as h. In fact, if the
individual component of the payoff can take any value, then the whole system
is forced into its direction.
In the Ising model, each player observes the average behaviour of all other
players. If each player observes only the behaviour of his neighbours, one obtains
Schelling’s model of racial segregation (According to Schelling’s model of racial
segregation, a city where Blacks and Whites are randomly distributed turns into a
chessboard of homogeneous quarters if its inhabitants, although absolutely ready to
accept the presence of the other colour, do not want to be a small minority in their
own neighbourhoods).
13.3.2 The Minority Game
The minority game originates from a consideration inspired by the El Farol bar in
Santa Fe, New Mexico (USA). The economist Brian Arthur remarked that people
go to the bar in order to meet other people, but they do not want to go when all other
people go, because the bar is too crowded on such occasions. Thus, they want to do
the opposite of what most people do –go to the bar when most people stay at home,
stay at home when most people go to the bar. This is interesting, because the
“El Farol Bar Problem” cannot have a stable equilibrium. In fact, once the majority
observed what the minority did, it wants to imitate it, which turns the minority into
majority, and so on endlessly.
Physicists Damien Challet and Yi-Cheng Zhang remarked that this is the essence
of stock market dynamics. In fact, in the stock market those traders gain, who buy
when prices are low (because most traders are selling) and sell when prices are high
(because most traders are buying). So all traders want to belong to the minority,
which is clearly impossible, hence the inherent instability of this game. Among the
physicists, the “El Farol Bar Problem” became the “Minority Game”.2
Let us consider N players who either belong to a group denoted 0 or a group
denoted 1. Players belonging to the minority group receive a positive payoff.
Players belonging to the majority receive zero.
Strategies are functions that predict which will be the minority group in the next
step given the minority group in the m previous steps. Thus, a strategy is a matrix
with 2m rows (dispositions with repetition of two elements of class m) and two
columns. The first column entails all possible series of minority groups in the
previous m steps, henceforth histories. The second column entails the group
suggested to be minority in the next step. As an example, Fig. 13.5 illustrates a
strategy with m ¼ 2.
Each player owns s strategies. If s ¼ 1, the game is trivial because the time
series of the minority group is periodical.
If s > 1, players choose the strategy that cumulated the greatest amount of
payoffs. Thus, a number of feedbacks may arise between what strategies are chosen
2
The rest of this section has been extensively drawn from E. Moro, The Minority Game:
An Introductory Guide, working paper available online.
306 G. Fioretti
History Prediction
00 0
01 1
10 1
11 1
Fig. 13.5 An example of a strategy based on the two previous steps of the minority game. The first
column lists all possible stories. The second column, depending on past history, makes a prediction
and their capability to predict the minority. In fact, in this game players must adapt
to an environment that they themselves create.
An important magnitude in this game is the variance of the time series of the
number of players belonging to group 1 (or, equivalently, group 0). Henceforth, this
magnitude will be denoted by σ 2 .
The average of the number of players belonging to each group is generally close
to N/2. If σ 2 is small, then the distribution of the number of players belonging to
group 1 is concentrated around N/2. This implies that the minority is large,
eventually close to its maximum (N/2 1). On the contrary, if σ 2 is large the
number of players belonging to group 1 tends to be either much smaller or much
larger than N/2, implying that the minority is often very small.
Let us consider σ 2 =N in order to normalize to the number of players. Let us define
the efficiency of coordination ec ¼ N=σ 2 as the reciprocal of the extent to which
players behave differently from one another.
Figure 13.6 depicts numerical simulations of ec as a function of the number of
histories in a strategy 2m =N . Graphs are shown for different values of s. The
horizontal line marks the value that ec attains if players would make a random
choice among the strategies available to them.
With low m the efficiency of coordination is low. This happens because if
memory is short, then players have greater difficulties to adapt to the changing
features of the game.
If only few strategies are available (s ¼ 2, s ¼ 3, s ¼ 4), at intermediate values
of m many players guess the correct strategy so ec increases above the level that can
be attained if strategies are chosen randomly. This threshold is marked by the
dashed vertical line. However, this effect disappears if many strategies are available
(s ¼ 8, s ¼ 16). In this case the decision process becomes similar to a random
choice so even at intermediate values of m the efficiency of coordination is close to
the level attained when strategies are chosen randomly.
Independently of the number of available strategies, with increasing m the value
of ec tends to the level attained when strategies are chosen randomly. In fact, a
history of length m occurs again after 2m steps on average, so a strategy that is
successful with a particular history needs 2m steps in order to be successful again.
With very high values of m, no strategy can present itself as particularly successful;
therefore, a nearly-random dynamics ensues.
102
ec
101
ec
100
s=2
10–1 3
4
8
16
10–2
10–3 10–2 10–1 100 101 102 103
2m/N
Fig. 13.6 Efficiency of coordination ec as a function of the number of histories in a strategy 2m =N,
for different values of the number of available strategies s. The horizontal line at ec ¼ ec marks the
efficiency level when player select a strategy at random. The vertical dashed line marks the point
where ec can be greater than ec
Let us consider what information is available to players. The only information

available to them is what group was the minority in previous time steps. Let this
information be carried by a variable Wt, where Wt ¼ 0 means that at time t the group
0 has been minority, Wt ¼ 1 otherwise. The issue is whether this information is used
efficiently; if it is not, there may exist exist arbitrage possibilities for players who
utilize information more efficiently than their peers.
Let us consider Wt and Wtþ1 as distinct signals. Let us compute their mean mutual
information IðWt ; Wtþ1 Þ.3
Mean mutual information measures whether the information entailed in the
outcomes of two steps of the game, taken together, is greater than the sum of
the information entailed in the outcomes of the two steps taken independently of
one another. Thus, mean mutual information says whether a player, by observing
the time series of the outcome of the game, could do better than his peers.
3
Given a source of binary symbols fa1 ; a2 ; . . . aM g issued with probabilities p1 ; p2 ; . . . pM , the
P
M
average information that they convey is defined as HðAÞ ¼ pðai Þlg2 1=pðai Þ and it is called
i¼1
information entropy. Suppose that there is a second source issuing symbols fb1 ; b2 ; . . . bN g with
information entropy H(B). Let H(A,B) denote the information entropy of the whole system. Mean
mutual information H(A) + H(B) H(A,B) measures to what extent the two sources interact to
correlate their messages. Mean mutual information is zero if the two sources are independent of
one another.
308 G. Fioretti
Fig. 13.7 Efficiency of information exploitation as a function of the number of stories in a

strategy, normalized to the number of players
Recalling the analogy with the stock market, IðWt ; Wtþ1 Þ > 0 means that a trader
could gain from arbitrage.
Let us introduce information efficiency ei ¼ 1=IðWt ; Wtþ1 Þ. Being the reciprocal
of mean mutual information, information efficiency is high when mean mutual
information is low, i.e., when information is efficiently exploited by the player so
there is little room for arbitrage.
Figure 13.7 depicts numerical simulations of ei as a function of the number of
stories in a strategy 2m =N. Graphs are shown for different values of s.
One may observe in Fig. 13.7 a sudden drop of ei in the [0.3,1] interval. This
interval is entailed in the interval [0.1,1] where ec was observed to rise above the
level corresponding to random choice in Fig. 13.6. Thus, we may subsume the
behaviour of the minority game in the following Table 13.1:
Table 13.1 shows that the minority game has two large behaviour modes, one
inefficient in coordination but efficient in the exploitation of information, the other
one efficient in coordination but inefficient in the exploitation of information. In
between, a tiny space where the efficiency of coordination and the efficiency of
information exploitation may change dramatically depending on s and m.
Since the minority game is a stylised representation of stock markets, we may
ask in which region stock markets operate. It is well known that very many traders
operate in stock markets, so we may assume that N is very large. Human bounded
rationality suggests that traders do not make use of complicated algorithms that take
account of events far back in the past, so m should be in the order of a few units.
Consequently, 2m =N is likely to be very small.
This suggests that financial markets are characterised by low coordination, which
implies irregular oscillations where large majorities and small minorities may appear.
At the same time, financial markets are efficient in exploiting information. Thus, the
observation of its time series offers few possibilities to extrapolate future courses.
Table 13.1 Efficiency of coordination and efficiency of information exploitation in the minority
game
2m =N < 0:1 2m =N > 1
Inefficient coordination Efficient coordination
Low ec High ec
Efficient information exploitation Inefficient information exploitation
High ei Low ei
13.4 Some Pitfalls of Utility Maximization
Utility maximization strikes its adepts for its elegance, simplicity and beauty. Unfor-
tunately, empirical tests have have shown that in many situations decision-makers do
not follow its prescriptions. Furthermore, there are cases where maximizing utility
leads to paradoxical decisions.
Some of these paradoxes can be reduced to utility maximization by means of
special additions to the basic theory. Others cannot, thereby suggesting that utility
maximization, besides poor descriptive strength, may have poor normative value as
well. In this section the main paradoxes will be discussed, together with their
eventual resolution within the utility maximization framework.
13.4.1 Ellsberg’s Paradox and Sub-additive Probabilities
Suppose that a decision-maker is placed in front of two urns, henceforth denoted

A and B. The decision-maker is informed that urn A entails white and black
balls in equal proportion, e.g., urn A may contain 10 white balls and 10 black
balls. Regarding urn B, the decision-maker knows only that it entails white and
black balls. Suppose to ask the decision-maker to evaluate the probability to
extract a white ball from urn A and the probability to extract a white ball from
urn B.
Since urn A entails white and black balls in equal proportions, the probability to
extract a white ball from urn A is 0.5. On the contrary, nothing is known regarding
the proportion of white to black balls in urn B. In cases like this, the so-called
“principle of insufficient reason” – i.e., the fact that there is no reason to think
otherwise – suggests to imagine that also urn B entails white and black balls in
equal proportions. Thus, also in this case the probability to extract a white ball is
assessed at 0.5. And yet, something is not in order: intuitively, urn B should be
characterized by a greater uncertainty than urn A!
Ellsberg’s paradox actually deals with the size of the sample on which probabilities
are evaluated. In fact, Ellsberg’s paradox places two extreme situations aside.
310 G. Fioretti
In the case of urn A, since we know that it entails white and black balls in equal
proportions we are able to compute probability with infinite precision. It is just like
extracting a ball (and replacing it afterwards) infinite times. We are measuring
probability on a sample of infinite size.
In the case of urn B, lack of knowledge on the proportion of white to black balls
is equivalent to estimating the probability of extracting a white ball prior to any
extraction. It means that the probability must be measured on a sample of size zero.
We guess its value at 0.5, but the reliability of our estimate is very low.
One possibility for overcoming Ellsberg’s paradox is that of representing
uncertainty by means of two magnitudes. The first one is probability, whereas
the second one is sample size. In statistics, sample size is expressed by precision
indicators.
Another possibility is to resort to the theory of sub-additive probabilities. While
according to classical probability theory the sum of the probabilities of an exhaus-
tive set of events must be equal to 1, according to the theory of sub-additive
probabilities this holds only if probabilities are measured on a sample of infinite
size. In all other cases, probabilities take values such that their sum is smaller
than 1.
Let us consider the following example: We are playing dice in a clandestine
gambling room. Since we fear that we are playing with an unfair die, we may not
assign probability 1/6 to each face, but rather less, e.g. 1/8. Thus, the sum of the
probabilities of all faces is 6 1/8 ¼ 3/4, which is smaller than 1. Subsequently, if
we have a possibility to throw the die many times –i.e. if we can increase the size of
our sample –we may find out that the die is unfair in the sense that, e.g., face “2”
comes out with probability 1/3 while the other faces come out with probability 2/15.
The sum of all these probabilities is 5 2/15 + 1/3 ¼ 2/3 + 1/3 ¼ 1.
Let us return to Ellsberg’s paradox. In the case of urn A, the probability to extract
a white ball is 0.5 and the probability to draw a black ball is 0.5. The sum of these
probabilities is 1. In the case of urn B, the decision-maker may judge that the
probability to extract a white ball is, for instance, 0.4, and that the probability of
extracting a black ball is also 0.4. The sum of these probabilities is 0.8, but this does
not constitute a problem for the theory.
By employing sub-additive probabilities, utility maximization can be meaning-
fully applied to situations where probabilities are less-then perfectly known. Thus,
utility maximization can safely deal with the difficulty raised by Ellsberg’s paradox.
13.4.2 Allais’ Paradox and Prospect Theory
The following experiment was proposed by Maurice Allais. Subjects are asked to
choose between the alternatives A and B reported on the rows of Table 13.2. It is
empirically observed that most people choose alternative B.
Table 13.2 The first choice in Allais’ experiment

Consequence 1 Consequence 2 Consequence 3
Alternative A Receive $2,500 Receive $2,400 Receive nothing
with prob. 0.33 with prob. 0.66 with prob. 0.01
Alternative B Receive $2,400
with prob. 1.00
Table 13.3 The second choice in Allais’ experiment

Consequence 1 Consequence 2
Alternative C Receive $2,500 with probability 0.33 Receive nothing with probability 0.67
Alternative D Receive $2,400 with probability 0.34 Receive nothing with probability 0.66
Subsequently, the same subjects are confronted with the alternatives C and D
reported on the rows of Table 13.3. It is empirically observed that most people
choose alternative C.
Let us now examine the expected utilities of these two pairs of alternatives,
(A,B) and (C,D). Preferring (B) to (A) means that u(2,400) > 0.33 u(2,500) þ
0.66 u(2,400), which can be written as 0.34 u(2,400) > 0.33 u(2,500).
Unfortunately, preferring (C) to (D) implies the opposite, i.e. that 0.33 u(2,500) >
0.34 u(2,400). So it turns out that most people either do not behave rationally, or do
not maximise utility.
Allais’ paradox is due to the presence of a tiny probability of not obtaining
anything in alternative (A). Thus, it is due to aversion to risk.
Daniel Kahneman and Amos Tversky introduced non-linear transformations of
utilities and probabilities in order to balance risk aversion. The transformed utilities
and probabilities can describe the observed behaviour as expected utility maximi-
zation. This is called Prospect Theory.
A prospect is a set of pairs fðc1 ; p1 Þ; ðc2 ; p2 Þ; . . .g, where cj is a consequence that
will obtain with probability pj . As a preliminary step, prospects with identical
consequences are summed, dominated prospects are eliminated and riskless
components are ignored.
Prospects Theory prescribes that the utilities and the probabilities of the above
prospects be transformed according to the following rules:
1. Utility is transformed by means of a non-linear function v ¼ f(u) such that
f 0 (u) > 0 and f 00 (u) < 0 for u > 0, f 0 (u) > 0 and f 00 (u) > 0 for u < 0, with
j f 00 ðuÞju<0 > j f 00 ðuÞju>0 .
2. Probabilities p are transformed into “weights” w by means of a non-linear
function w ¼ g(p) such that g(0) ¼ 0 and g(1) ¼ 1 but 9 p 2 ð0; 1Þ such that
8p < p it is g(p) p and 8p > p it is g(p) p.
312 G. Fioretti
3. Weights w are transformed into coefficients q by means of the following rules:
q
h ¼ w ðph Þ for j ¼ h
q
i ¼ w ðph þ . . . þ pi Þ w ðph þ . . .þ pi1 Þ for h < j 0
qþ
i ¼ wþ ðpi þ . . . þ pk Þ wþ ðpiþ1 þ . . .þ pk Þ for 0j<k
qþ
k ¼ wþ ðpk Þ for j¼k
where w and q refer to prospects with negative utility, denoted with an index
j2[h, 0], whereas wþ and qþ refer to prospects with positive utility, denoted
with an index j2[0, k].
The v and q obtained at the end of this procedure can be used much like utilities
and probabilities, respectively. Prospect Theory succeeds to eliminate the
inconsistencies highlighted by Allais’ paradox, but it does not explain why it
works. It should be called a heuristic, rather than a theory. In the end, utility
maximisation does succeed to cope with Allais’ paradox, but at the cost of a patch
that has a flavour akin to the epicycles that had to be added to the Ptolemaic system in
order to support the idea that it was the Sun that was turning around the Earth.
13.4.3 Preference Reversal in Slovic’s Paradox
Let us consider a series of bets with different characteristics. For instance, a series
of bets on different horses, or playing on a series of different slot machines, or a
series of unfair dice different from one another. The game consists of choosing to
bet on a specific horse, choosing to play on a specific slot machine or selecting a
specific die to throw. In other words, the game consists of choosing one bet out of a
series of bets.
In order to simplify matters, let us consider series composed by two bets. More
specifically, let us consider the four pairs of bets in Table 13.4.
For any pair of bets, subjects are asked to select either bet A or bet B. On
average, the number of subjects who prefer A to B is slightly greater than the
number of subjects who prefer B to A.
At this point, a different game is played. Subjects are asked to imagine that they
own a lottery ticket for each bet, and that they have a possibility to sell it. That is,
they can either wait for the outcome of each bet, where they may win or loose with a
certain probability, or they can sell the ticket. In order to compare the willingness to
play to the willingness to sell the ticket, subjects are asked to fix a minimum selling
price for each bet.
In general, it is empirically observed that most people ask a higher price for bets
B than for bets A.
However, for each pair of bets, bet A has the same expected (utility) value than
bet B. Thus, utility maximizers should be indifferent between A and B. On the
contrary, most subjects have a slight preference for A if they are asked to play one
of the two bets but they definitely prefer B if they are asked to fix a selling price.
Table 13.4 Slovic’s Pair of bets I

experiment
Bet AI Win $4.00 Loose $1.00
With probability 0.99 With probability 0.01
Bet BI Win $16.00 Loose $2.00
Pair of bets II
Bet AII Win $3.00 Loose $2.00
Win $6.50 Loose $1.00
Pair of bets III

Bet AIII Win $2.00 Loose $1.00
Bet BIII Win $9.00 Loose $0.50
Pair of bets IV
Bet AIV Win $4.00 Loose $0.50
Bet BIV Win $40.00 Loose $1.00
The distinguishing feature of bets A is that the first consequence has a much
higher probability than the second one. Thus, one may assume that it is this
difference of probability values that orientates decision-making.
The distinguishing feature of bets B is that the first consequence concerns a
much larger amount of money than the second one. Probabilities, on the contrary,
are sometimes very similar, sometimes very different from one another. Thus, one
may assume that it is this difference of money values that orientates decision-
making.
If subjects are asked to bet their attention is caught by probabilities, so either
they are indifferent or they prefer A. If subjects are asked to sell lottery tickets their
attention is caught by money values, so they prefer B.
Slovic’s paradox shows that preferences change if the decision-maker focuses on
the probability of a consequence or, rather, on its utility (here, money value). This
means that human beings are unable to evaluate probabilities and utilities indepen-
dently of one another.
Slovic’s paradox – often known as “preference reversal” – is destructive for
utility maximization. In fact, it undermines the assumption that a utility function
and a probability function can be defined, independently of one another. Ultimately,
Slovic’s paradox suggests that uncertain belief cannot be split into utilities and
probabilities.
314 G. Fioretti
Obviously, several attempts to reconcile preference reversal with the theory of

rational choice have been made. Preference reversal can be accommodated with the
theory of rational choice if either violations of transitivity, or of independence, or of
completeness of preferences are accepted. While the attempts to reconcile prefer-
ence reversal with the theory of rational decision by relaxing transitivity or inde-
pendence of preferences did not receive much attention because these properties are
essential for our idea of rationality – see Sect. 13.2, the more recent idea of dropping
completeness deserves some discussion. In fact, allowing preferences to be incom-
plete amounts to accept the idea that a utility function can be defined, at most, for
only some alternatives. Possibly, just the simplest and most repetitive ones.
13.4.4 Arrow’s Paradox
The following paradox of social choice is due to Kenneth Arrow. Let A, B and C
denote three alternatives, and let 1, 2 and 3 denote three individuals. Let us assume
that:
• Individual 1 prefers alternative A to alternative B and alternative B to alternative
C. Thus, he prefers alternative A to alternative C.
• Individual 2 prefers alternative B to alternative C and alternative C to alternative
A. Thus, he prefers alternative B to alternative A.
• Individual 3 prefers alternative C to alternative A and alternative A to alternative
B. Thus, he prefers alternative C to alternative B.
If these three individuals constitute a democratic community with a majority
rule, then this community prefers A to B (individuals 1 and 3) and alternative B to
alternative C (individuals 1 and 2). Thus, if the community wants to have transitive
preferences, it must prefer A to C. But, the majority of its members (individuals
2 and 3) prefers C to A!
Arrow’s paradox shows that there are conditions such that the aggregate
outcome contradicts a basic assumption of utility maximization, even if
individuals do not. It is not destructive for utility maximization as a theory of
individual decision-making, but it impairs its extension to group or organizational
decision-making.
Several proposals have been made in order to overcome Arrow’s paradox. The
most common way out is to allow individuals to have different preferences if all
alternatives are presented to them, instead of being presented with pairs of
alternatives. Or, one may limit voters to two alternatives presented in tournaments.
In this way, Arrow’s paradox would disappear, yet the final choice is not necessarily
the one that would be preferred by the largest possible majority.
13.5 Logic of Consequence and Logic of Appropriateness
As we have seen in Sect. 4, utility maximization is not a good descriptor of decision

processes. Its proponents have objected that utility maximization is not meant to be
a faithful description of what people actually do, but rather a prescription of what
they should do. It pretends to be a normative theory, although it is not a descriptive
theory.
However, the preference reversals highlighted by Slovic point to such a huge
distance between theory and reality, that the normativeness of utility maximization
might be questioned. If utilities do not exist prior to decision-making, it may make
little sense to tell decision-makers that they should maximize them. Furthermore, if
evolution shaped human reasoning along patterns that are different from utility
maximization, we ought to be careful to declare these patterns “illogical”, or
“irrational”. Rather, it may make sense to observe how human beings actually
make their decisions, understand their rationales, and eventually revise our theories.
Sociology knows a distinction between the “logic of consequences” that underlies
utility maximization, and a “logic of appropriateness” where human beings behave
according to what they deem appropriate depending on past experiences and social
pressures to conformity in specific settings. Note that the logic of appropriateness
does not separate an individualistic step (Utility) from social interaction (Game
Theory).
According to the logic of appropriateness, human minds are viewed as
coherence-seeking machines that make use of available information in order to
construct a plausible interpretation of reality, be it social roles, scientific theories, or
else. By drawing causal relationships and eliminating inconsistencies a decision-
maker tells herself a story that explains why certain facts are the way they are and
why certain people did what they did. This story, a founding story that suggests a
decision-maker what it is appropriate to do, is called a narrative.
The construction of a narrative may require that issues that do not fit into the
picture are ignored, downplayed or forgotten. It may require that opinions are
changed even dramatically, and yet their purporters candidly claim that they have
always been coherent throughout their lives, or that they have been coherent in spite
of having changed their opinion, if their story is seen from a particular point of view.
Albeit disturbing for our idea of rationality, the extent and easiness with which
human beings distort previous experiences is proven by a number of experiments in
psychology. It is easy to induce the subjects of experiments to change opinion while
they are still convinced to have been coherent throughout the whole experiment.
Experiments show that people remember past events to the extent that they fit their
narratives, and that they are ready to change their interpretation of the past if new
evidence must be accommodated. On the whole, experimental evidence tells us that
human beings are ready to lie to themselves in order to build coherent narratives.
This attitude is puzzling, because distorting reality in order to construct a
coherent narrative is at odds with our idea of rationality. So either human nature
is inherently irrational, or our idea of rationality is incorrect.
316 G. Fioretti
According to James March, re-inventing the past is a crucial ability that enables
decision-makers to conceive new goals and figure out a strategy in an uncertain
future. Later, a similar argument has been made by Karl Weick under the label of
“sensemaking”. Essentially, these authors suggest that in order to make decisions in
the face of an uncertain future it is good to have a narrative that explains the past as
if previous decisions had been made along a coherent line. This line guides the
decision-maker into the future, providing a rationale for action even if certainties
are very few.
So here comes a straightforward argument for normativeness. If seeking coher-
ence has the purpose of constructing a narrative, and if narratives are useful, then a
decision theory based on constructing narratives should be regarded as rational, and
openly prescribed.
In business, politics and other fields, narratives may constitute the bulk of
strategies. David Lane and Robert Maxfield have made a years-long field observa-
tion of the elaboration and modification of the narrative of a Silicon Valley firm.
This study is worth reporting, because it is very clear in making us understand that
narratives are useful precisely because they provide a guidance in the face of an
uncertain future, and that their usefulness is not impaired by the fact that their
coherence is based on an arbitrary interpretation of reality.
13.5.1 A Real Story
In 1990, Echelon, a Silicon Valley company, launched LonWorks, an innovative

technology for distributed control. Previously, control was centralized by one main
processing unit that commanded several peripheral devices. With LonWorks, each
device is endowed with a microprocessor and can communicate with all others, so the
peripheral electromechanical devices control each other. Distributed control is more
resilient than centralized control, and easily implements modular architectures to
which additional devices can be added.
Distributed control is particularly suited to the automation of office spaces in
large buildings, post-Fordist productive plants, as well as any setting where a large
number of heterogeneous devices must coordinate their operations while retaining
some flexibility. Thus, in its early days Echelon focused on partnerships with large
producers of the devices to be automated, e.g., a producer in the field of heating and
air conditioning was offered a possibility to integrate a microchip in their devices,
as well on lifts, doors and windows in order to integrate all controls in a large
building, from lighting to heating to theft protection.
With some disappointment, Echelon had to recognize that the LonWorks tech-
nology was not exploited in its full potentialities. In fact, each large producer was
specialized in one tiny sector so it had neither the power nor the capability to
implement LonWorks on all devices. For instance, a producer in the field of heating
and air conditioning found it difficult to install LonWorks on doors, windows, lights
and lifts, for the production of these devices was covered by other firms. Indeed, the
difficulty was that Echelon was attempting to create a single market for automation
where the marketplace was covered by producers of several devices.
Echelon was conscious of the enormous difficulties connected with the creation
of a new market. Nevertheless, it deemed that long-term relations with a few
specialized producers would pay in the long run. Echelon had a narrative, saying
that large specialized producers would slowly but persistently adopt and impose
LonWorks. Consequently, it invested all of its resources in these relations.
By 1994, Echelon was loosing confidence in this narrative. Echelon started to
approach large system integrators of ICT, such as Olivetti and Ameritech. However,
the crucial move was that of hiring a person for this job, who did not come from
Silicon Valley as all other executives did. Through this employee Echelon
approached smaller companies, that integrated devices they bought from different
producers. Some people at Echelon conceived the idea of embedding LonWorks in a
box that could be attached to any electromechanical device, of whatever producer.
Scholars of technological innovation know how difficult it is for visionary
employees to convince their boss of the value of their idea. In the case of Echelon
the CEO embraced enthusiastically the new idea, because it appeared to fit with his
previous experience.
Echelon’s CEO had been the successful entrepreneur of a small firm that
exploited digital technologies to produce private branch exchange systems (PBX)
with innovative features. This firm had been able to displace giants such as AT&T
by providing small independent installers with a superior product. When this CEO
met small independent integrators of electromechanical devices, he mapped the
idea proposed by his new employee onto his previous experience.
In 1996, and within a few months, Echelon changed its narrative. Echelon
presented itself as a provider of an innovative microchip for independent system
integrators, a microchip that could be installed on any electromechanical device, of
whatever producer.
Most importantly, Echelon told itself that it had always pursued this strategy.
Nobody in the firm seemed to be aware that the firm’s strategy had changed.
According to the narrative that they had developed, they had always done what
they were doing.
Moreover, when faced with evidence that the firm did change its strategy, man-
agement wished that the final publication would not stress this aspect (Lane: personal
communication). This makes sense, for according to our idea of rationality narratives
should reflect “objective information”, and decision-makers should stick to it. Thus,
management did not want to appear irrational according to common wisdom.
However, the case of Echelon highlights that constructing a narrative by re-
interpreting the past may be good and useful for decision-makers. In fact, the
reported case reveals that precisely by re-interpreting its mission Echelon was
able to direct its investments. If the future is uncertain, as it is often the case,
interpreting the past in order to find a direction for the future is a sensible activity.
So the trouble may rather lie with our idea of rationality. Since re-interpreting
the past is regarded as irrational, then it must be done in secrecy. However, if re-
interpreting the past may have positive effects, then it should be prescribed.
318 G. Fioretti
13.6 Tools for the Logic of Appropriateness
Although the logic of appropriateness cannot propose itself with a ready-made and
ready-to-use formula such as utility maximization, there exist some tools that can
be used to reproduce its building blocks. These are essentially classification tools,
i.e., tools that form concepts out of information, and coherence tools, i.e., tools that
arrange concepts into coherent stories.
In particular, the following tools will be reviewed in this section:
1. Unsupervised neural networks;
2. Constraint Satisfaction Networks;
3. Evidence Theory, also called Belief Functions Theory.
Unsupervised neural networks reproduce the formation of mental categories out
of a flow of information. Constrain Satisfaction Networks arrange concepts into
coherent explanations. Finally, Evidence Theory assumes that an actor receives
information on possibilities and arranges them into coherent hypotheses. Although
these tools have not been integrated into one another, they all concern the process of
selecting some items from the flow of experiences, arranging them in a coherent
narrative, and deciding accordingly.
The logic of consequence makes sense in the restricted realm of games of
chance, where it is possible to overview an exhaustive set of possibilities and enlist
all of the consequences of any alternative. On the contrary, the logic of appropri-
ateness makes sense precisely because, quite often, such conditions do not hold.
Thus, this review does not cover tools concerned with classification in a given set of
categories, such as Case-Based Decision Theory and supervised neural networks.
13.6.1 Unsupervised Neural Networks
Human mental categories are not defined by pre-specified similarity criteria that the
objects to be classified should fulfil. Rather, mental categories are continuously
constructed and modified according to the similarity of a just-received piece of
information to the pieces of information that have already been stored in existing
categories. For instance, a child observing house chairs may start with an idea of
“chair” as an object having four legs, then observe an office chair with only one leg
and yet sufficiently similar to house chairs to be added to their category, which from
this time on does not have the number of legs as a common property of all the
objects that it entails.
This is a radically different process from that of defining criteria in order to
classify objects into given categories. In the case of mental categories, categories do
not exist prior to the beginning of the classification process. Categories form out of
similarity of certain input information to some information that has been previously
received, so these items are stored in the same “place” and concur to build up a
mental category. There exist no criteria defined ex-ante that control the classifica-
tion of input information; since the only rule is similarity of information items,
categories form depending on what items are received. For instance, the aforemen-
tioned mental category “chair” may depend on what chairs have been observed in
the course of a life spent in Manhattan, rural China or tropical Africa, in modern
times or the Middle Age.
By contrast, logical reasoning works with objects that have been clearly defined.
The point here is that these definitions are made once mental categories have been
formed. Eventually, a mental category formed around similarity judgements may
suggest a definition for all the objects that it entails if they share some common
feature.
However, this is not necessarily the case. There are instances where definitions are
not possible, simply because a mental category entails objects that do not have any
common feature. For instance, the mental category expressed by the word “game”
refers to children amusing themselves with toys, adults involved in a serious compe-
tition on a chess board and as well as a set of wild animals. One may speculate that
man transposed the emotions involved in hunting into the more intellectual context of
chess, and that the fact that chess was an amusement suggested some similarity to
what children were supposed to do. So pairwise intersections of the meanings of the
word “game” do exist, but this does not imply that all of their meanings have a
common intersection. Nevertheless, human beings are perfectly at ease with this as
well as many other concepts that cannot be defined.
Unsupervised neural networks (UNN) are able to reproduce the idea that mental
categories arise out of adding examples. In fact, these networks construct categories
around the most frequent input patterns, mimicking the idea that a child creates a
category “chair” upon observation of many objects similar to one another.
It is important to stress once again that the ensuing account deals with unsuper-
vised neural networks (UNN) only. Supervised neural networks (SNN), with their
neurons arranged in layers and training phases to teach them in what categories they
must classify input information, do not fit into the present account. Nor does Case-
Based Decision Theory which, similarly to SNN, is concerned with classifying
information into previously defined categories. Directions to these tools will be
provided in the bibliography, but they do not pertain to this chapter.
Neural networks, of whatever kind, are composed by a set of neurons which
produce an output y 2 < by summing inputs x1 ; x2 ; . . . xN 2 < by means of
coefficients a1 ; a2 ; . . . aN :
X
N
y¼ ai xi (13.7)
i¼1
For any set of coefficients ai , this simple device is able to classify inputs in a
category by yielding the same output y for several input vectors x. In fact, there
exist several vectors x whose weighted sum yields the same y. For instance, if 8i it
is ai ¼ 1, then e.g. y ¼ 10 can arise out of x0 ¼ [9 1], x00 ¼ [2.5 7.5], as well as
320 G. Fioretti
many other vectors. In this sense, the neuron classifies the input vectors [9 1] and
[2.5 7.5] in the same category.
Note that a neuron has no difficulty to classify input vectors that do not perfectly
fit its categories. For instance, if there is a category y ¼ 10 and a category y ¼ 11,
an input vector x000 ¼ [2.1 8] is classified in the category y ¼ 10 just as x0 and x00 .
The shape of the categories implemented by a neuron depends on the coefficients
ai . For instance, if a1 ¼ 0:5 and a2 ¼ 20 the input vector x0 ¼ [9 1] yields
y ¼ 24.5 and may not lie in the same category as x00 ¼ [2.5 7.5], which yields
y ¼ 151.25.
The coefficients ai may be chosen by the user of the network during a training
phase, in which case we are dealing with a SNN. Alternatively, the coefficients ai
may be initialised at random and subsequently changed by the network itself
according to some endogenous mechanism. In this case we have a UNN, of
which Kohonen networks are the most common instance.
In UNN, the ability of a neuron to change its categories stems from a feed-back
from output y and a feed-forward from input x, towards coefficients ai :
dai
¼ φða; yÞxi γða; yÞai 8i (13.8)
dt
where φ(a,y) and γ(a,y) may be linear or non-linear functions.

In Eq. 13.8, the term φða; yÞxi enables the neuron to learn input patterns. It entails
both a feed-back (from y) and a feed-forward (from xi). This learning term makes ai
increase when both y and xi take high values, thereby enhancing those coefficients
that happened to yield a high y when a particular xi was high. Thus, the structure of
coefficients vector a ultimately depends on which vectors x appeared most often as
inputs.
The learning term is such that the neuron learns the patterns that it receives most
often. This is sufficient to make the network work, but makes it unable to construct
different categories if different patterns appear. Furthermore, since the learning
term works by multiplying inputs and outputs, it may produce an explosive output.
This should be curbed in order to use the network.
For both reasons, a forgetting term that makes the coefficients ai decay towards
zero is in order. In Eq. 13.8 the forgetting term is γða; yÞai. It depends on a feed-back
from output y and, most importantly, on coefficient ai itself.
Figure 13.8 illustrates the feed-backs and -forwards within a neuron of a UNN.
Simple, but non trivial examples of Eq. 13.8 are: a_ ¼ μyx νa, a_ ¼ μx νya,
a_ ¼ μyx νya, a_ ¼ μyx νy2 a, where μ and ν are constants. Each functional form
corresponds to different strengths of the learning and forgetting terms.
In general, a network of neurons is able to discriminate input information
according to much finer categories than a single neuron can do. As a rule, the greater
the number of neurons, the finer the categories that a network constructs. However, a
neural network is useful precisely because it is able to classify a huge amount of
information into a few broad categories. If categories are so fine that they track input
information exactly, then a neural net becomes useless. Thus, the number of neurons
x
Σ y
a
φ, γ
Fig. 13.8 The neuron of a UNN. The feed-backs and -forwards are responsible for the most
notable properties of UNN, including the absence of a training phase. In a sense, the “training
phase” of SNN may be seen as a feed-back and -forward passing through a human operator
that a network should possess depends on the variability of the input as well as on user
needs.
However, the behaviour of a neural network does not only depend on the number
of its neurons, but also on the structure of the connections between them. In fact,
just like the capabilities of neurons depend on feed-backs and -forwards, the
capabilities of a neural network depend on linkages that eventually enable informa-
tion to circulate in loops. If information can circulate within the network, then the
whole network acquires a memory.
It is a distributed memory, fundamentally different in nature from the more usual
localised memories. Localised memories such as books, disks, tapes etc., store
information at a particular point in space. This information can only be retrieved if
one knows where its support is (e.g. the position of a book in a library, or the
address of a memory cell on a hard disk).
On the contrary, in a neural network each neuron may be part of a number of
information circuits where information is “memorised” as long as it does not stop to
circulate. Although this is a memory, one cannot say that information is stored at
any particular place. Hence the name.
For obvious reasons, the information stored in a distributed memory cannot be
retrieved by means of an address. However, a piece of information flowing in a
particular loop can be retrieved by some other piece of information that is flowing
close enough to it. Thus, in a distributed memory information can be retrieved by
means of associations of concepts, with a procedure that reminds of human “intui-
tion”. Intuition, according to this interpretation, would consist of associations
between concepts that occur when information flowing in different but neighbouring
circuits comes in touch.
13.6.2 Constraint Satisfaction Networks
Parallel constraint satisfaction networks (CSN) arrange concepts into coherent

theories. Although they belong to the larger family of neural networks, they do
not carry out any classification process.
322 G. Fioretti
Constraint satisfaction networks are characterized by:

• Excitatory and inhibitory connections;
• Feedbacks between neurons.
Neurons represent possibilities, or concepts, or propositions. Connections
represent inferences: an excitatory connection from neuron A to neuron B
means “A implies B”, whereas an inhibitory connection from neuron A to neuron
B means “A implies ¬ B”.
Let ai denote the activation (the output) of neuron i, with ai 2 <. Let wij 2 <
denote the weight by which neuron i multiplies the input arriving from neuron j.
The net excitatory input to neuron i is:
X
ene ti ¼ wij aj if wij aj 0 (13.9)
j
The net inhibitory input to neuron i is:

X
ene ti ¼ wij aj if wij aj < 0 (13.10)
j
At each time step, the activation of neuron i is increased by its excitatory inputs
and decreased by its inhibitory inputs:
Δai ¼ eneti ðamax ai Þ þ ineti ðai amin Þ (13.11)
where, in general, amax ¼ 1 and amin ¼ 1.

Feedbacks between neurons make the network maximize consonance:
XX
C¼ wij ai aj (13.12)
i j
or, equivalently, minimize energy E ¼ C.

Consonance maximization means that those neurons are strengthened, that
represent possibilities, concepts or propositions that are coherent with one another.
Thus, constraint satisfaction networks can be used to model any cognitive process
characterized by a search for coherence. In particular, researchers have emphasized
the ability of CSN to construct narratives, much like humans actually do.
Notable applications of CSN are the elaboration of scientific theories, which
amounts to arrange empirical findings in a network of coherent causal relations, as
well as the evaluation of guilt or innocence in a trial, which amounts to fitting
testimonies in a coherent frame. Under this respect, CSN share a common concern
with Evidence Theory or, to be more precise, with its Dempster-Shafer’s combina-
tion rule reported in Eq. 13.16 of Sect. 13.6.3.
When modelling decision-making, CSN can be used to model the process of

emphasizing the positive aspects of one alternative and the negative aspects of its
competiting alternatives until a coherent frame is available and a decision can be
made. This oscillation between competing explanations reproduce at least one
important aspect of Gestalt theories, namely, the idea that the human mind may
shift among alternative interpretations of reality, as exemplified by Rubin’s vase
and other images where at least two interpretations are possible.4 Many cues
suggest that this is a fundamental pattern in decision-making.
A clear limitation of CSN is that they work with given possibilities, concepts, or
propositions. In other words, CSN can reproduce the arrangement of possibilities
and concepts into narratives, but not their arousal. By contrast, UNN reproduce the
arousal of concepts out of empirical experiences. Possibly, future models will be
able to couple UNN to CSN in order to model both processes at a time.
13.6.3 Evidence Theory
Evidence Theory is a branch of the mathematics of uncertain reasoning that, unlike

Probability Theory, does not assume that a decision-maker knows the set of all
possible events. Rather than defining a “residual event” for anything that cannot be
clearly expressed, Evidence Theory leaves a decision-maker’s possibility set open
to novelties.
Evidence Theory makes use of a particular class of monotone uncertainty
measures, Choquet capacities of infinite order. Furthermore, it assumes that no
operation is attached to the possibility set, which frees a decision-maker to define a
“residual event” by complementation. Novel possibilities can appear in the possi-
bility set in the course of the calculations, and the possibility set is called frame of
discernment in order to stress its cognitive nature.
Evidence Theory does not take a gambler as its prototypical subject, but a judge
or a detective. The reason is that a gambler playing with dice or throwing a coin
knows what possibilities can occur. On the contrary, judges and detectives know
that unexpected proves and testimonies may open up novel possibilities. Possibly,
managers making investments, politicians steering their countries, or just anyone in
the important choices of her daily life is more akin to a judge or a detective looking
for cues than a gambler looking for luck.
Let us consider a frame of discernment Θ. Let us suppose that a person receives
testimonies, or bodies of evidence, as numbers that to various extents support a
set of possibilities A1 ; A2 ; . . . AN , where A1
Θ, A2
Θ . . . AN
Θ and where the
4
The simplest picture of this kind is a cube depicted by its edges: it is up to the observer to choose
which face stays in the front and which face stays in the rear. Rubin’s vase is white and stands
against a black background. The observer may see a white vase, or two black profiles in front of
one another.
324 G. Fioretti
Ai s are not necessarily disjoint sets.5 Let us denote these numbers fmðA1 Þ; mðA2 Þ;
. . . mðΘÞg, where mðAi Þ measures the amount of empirical evidence that supports
the possibility Ai .
Numbers m are exogenous to the person (the judge, the detective) who owns the
frame of discernment. They are not subjective measures for this person, though they
may be subjective evaluations of those who provide the testimonies. Numbers m are
cardinal measures of the amount of empirical evidence supporting each possibility.
Since no operation is defined on the frame of discernment, the number m that has
been assigned to Θ does not concern any specific possibility. Rather, it indicates
how small the evidence is, that supports the possibilities envisaged in the testimony,
or, in other words, how strongly a person fears that the possibilities that she is
envisaging are not exhaustive. The greater the ignorance of a person on which
possibilities exist, the greater m(Θ).
Note that m(Θ) can be smaller than any m of the Ai s that it entails. Indeed, this
applies to the Ai s as well: if Ai Aj , this does not imply that mðAi Þ > mðAj Þ.
Although it is not strictly essential for Evidence Theory, numbers m are gener-
ally normalised by requiring that:
X
N
mðAi Þ þ mðΘÞ ¼ 1 (13.13)
i¼1
For instance, if the original format of the testimony is:
f5; 32; 12; 3g
by applying Eq. 13.13 we obtain:
f0:096; 0:615; 0:231; 0:058g
whose numbers sum up to one.

Let us suppose that the decision-maker wants to evaluate to what extent the
available empirical evidence supports certain hypotheses that she is entertaining in
her mind. Since a hypothesis concerns the truth of a possibility or a set of
possibilities, hypotheses are subsets of the frame of discernment just as possibilities
are. A body of evidence fmðA1 Þ; mðA2 Þ; . . . mðΘÞg supports a hypothesis H to the
extent that some Ai s are included or at least intersect H.
Note that, whilst the possibilities Ai entailed in the testimonies cannot be combined
with one another (intersected, complemented, etc.) to form novel possibilities, a
hypothesis H represents a construct of the owner of a frame of discernment
(the judge, the detective, etc.). This person is absolutely free to conceive any
5
For simplicity, the theory is expounded with respect to a finite number of possibilities. No
substantial change is needed if an infinite number of possibilities is considered.
hypothesis, as well as its opposite. Thus, although Ai s are forbidden, H can be safely
considered.
Given a testimony fmðA1 Þ; mðA2 Þ; . . . mðΘÞg , the belief in hypothesis H is
expressed by the following belief function:
X
BelðHÞ ¼ mðAi Þ (13.14)
Ai
H
By definition, Bel(Ø) ¼ 0 and Bel(Θ) ¼ 1. However, this last condition does not
imply that any of the possibilities included in the frame of discernment must
necessarily realise. It simply means that any possibility must be conceived within
the frame of discernment, independently of what possibilities are envisaged at a
certain point in time.
The belief function takes account of all evidence included in H. The plausibility
function takes account of all evidence that intersects H:
X
PlðHÞ ¼ mðAi Þ (13.15)
H\Ai 6¼
It can be shown that belief and plausibility are linked by the relation PlðHÞ ¼
1 BelðHÞ , where H denotes the hypothesis opposite to H. If m(Θ) > 0 these
two measures are not equivalent, so both of them need to be considered. In general,
Bel(H) Pl(H).
Let us suppose that some unexpected facts occur, that are told by a new
testimony. The new testimony must be combined with previous knowledge,
confirming it to the extent that it is coherent with it. On the contrary, previous
beliefs must be weakened if the new evidence disconfirms them.
Let fmðB1 Þ; mðB2 Þ; . . . mðΘÞg be the new testimony, which must be combined
with fmðA1 Þ; mðA2 Þ; . . . mðΘÞg. The new testimony may entail possibilities that are
coherent with those of the previous testimony, possibilities that contradict those of
the previous testimony, and possibilities that partially support, partially contradict
the previous testimony. Figure 13.9 illustrates contradictory, coherent and partially
coherent/contradictory possibilities on the frame of discernment. Contradictory
possibilities appear as disjoint sets. A possibility is coherent with another if it is
included in it. Finally, two possibilities that are partially coherent, partially contra-
dictory, intersect one another.
Let us suppose that two testimonies
fmðA1 Þ; mðA2 Þ; . . . mðΘÞg
and
fmðB1 Þ; mðB2 Þ; . . . mðΘÞg

326 G. Fioretti
Ai Bj Ai
Ai
Bj Bj
Θ Θ Θ
Contradictory Possibilities Coherent Possibilities Partially contradictory-

-coherent possibilities
Fig. 13.9 Left, two contradictory possibilities. Centre, two coherent possibilities. Right, two
partially coherent, partially contradictory possibilities
that satisfy Eq. 13.13, must be combined into a testimony
fmðC1 Þ; mðC2 Þ; . . . mðΘÞg
that also satisfies Eq. 13.13. Dempster-Shafer’s combination rule yields a combined
testimony fmðCk Þg where the coherent possibilities between fmðAi Þg and fmðBj Þg
have been stressed.
According to Dempster-Shafer combination rule, possibilities fCk g are defined
by all intersections of each possibility in fA1 ; A2 ; . . . Θg with each possibility in
fB1 ; B2 ; . . . Θg. For any possibility Ck , the amount of empirical evidence is:
P
mðAi ÞmðBj Þ
Ai \Bj ¼Ck
mðCk Þ ¼ P (13.16)
1 mðAi ÞmðBj Þ
Ai \Bj ¼
The numerator of Eq. 13.16 measures the extent to which both the first and the
second testimony support the possibility Ck . In fact, for each possible Ck the sum
extends to all pairs of possibilities from the two testimonies that are coherent on Ck
(see Fig. 13.9). The more the intersections between the Ai s and the Bj s that give rise
to Ck , and the greater their amounts of evidence, the larger the numerator.
The denominator is the complement to one of those elements of the second
testimony that contradict the first one. In fact, the complement to one is made on
those Ai s and Bj s that are disjoint sets (see Fig. 13.9). The denominator represents a
measure of the extent to which the two testimonies are coherent, in the sense that all
evidence that supports contradictory possibilities is excluded.
Essentially, Dempster-Shafer combination rule says that the evidence suppor-
ting possibility Ck is a fraction of the coherent evidence between fmðA1 Þ; mðA2 Þ;
. . . mðΘÞg and fmðB1 Þ; mðB2 Þ; . . . mðΘÞg. The amount of this fraction depends on
the sum of all elements of the testimonies that support Ck .
Dempster-Shafer’s rule can be iterated to combine any number of testimonies.

The outcome of Dempster-Shafer combination rule is independent of the order in
which two testimonies are combined.
The above description made clear that Evidence Theory provides an algorithm
for handling an exogenous flow of new, unexpected possibilities. Indeed, the
decision-maker of Evidence Theory is not supposed to conceive possibilities. She
merely listens to exogenous testimonies that consist of possibilities and degrees of
evidence supporting them, and combines these testimonies into a coherent whole by
means of Dempster-Shafer theory. She does not conceive novel possibilities out of a
creative effort. Rather, novel possibilities – the fCk g – arise out of combination of
exogenous inputs.
Under this respect Evidence Theory bears some similarly to UNN. In fact, both
tools are open to novel arrangements of input information. These dynamical, open
arrangements may be variably called “possibilities” in Evidence Theory or
“categories” in UNN, but they express a similar reasoning.
At the same time, Evidence Theory bears some similarity to CSN when it comes
to the combination of evidence. In fact bot CSN and Dempster-Shafer’s combina-
tion rule (13.16) seek to improve the coherence of available information.
While UNN and CSN take on the complementary tasks of modelling the arousal
of new concepts and, respectively, their combination into coherent narratives,
Evidence Theory is a mathematical theory that may be developed into a framework
able to encompass both CSN and UNN as simulation tools. This is, at present, just
pure speculation; however, it is also a possible narrative that may direct future
research efforts.
13.7 Conclusions
This review presented tools to model decision-making according to two opposing

paradigms, namely, the logic of consequence and the logic of appropriateness. The
reader may feel unease because scientists do not provide a univocal answer to the
demands of the modeller.
However, a pragmatic attitude may suggest that tools should be used depending
on conditions. Utility maximization and Game Theory require that all available
alternatives and all of their possible consequences can be listed. Thus, it may be
sensible to make use of these tools when one such exhaustive list is available,
eventually releasing the requirement of perfect rationality and the pursuit of Nash-
equilibria while assuming some form of bounded rationality as influence games do.
Unsupervised neural networks, constraint satisfaction networks and Evidence The-
ory, on the contrary, may be used when more challenging decision settings must be
modelled. The modeller should remember that constructing narratives makes sense
because decision-makers may be uncertain regarding what possibilities exist, so
these tools become necessary only in this kind of decision settings.
328 G. Fioretti
The trouble, in this last case, is that the tools mentioned above have not been
integrated into a unified framework. No simple formula is available to be readily
used, so the modeller must resort to a higher degree of creativity and intuition. On
the other hand, here is an exciting opportunity for modellers to participate to theory
development.
Further Reading
Game Theory is a huge subject. Relevant handbooks are Aumann and Hart (1992,
1994, 2002) and, at a more introductory level, Rasmusen (2007). However, agent-
based modelers should keep in mind that a substantial part of Game Theory has
been developed around equilibrium states, which are generally not a main concern
for agent-based modelers. Evolutionary games, thoroughly discussed in the above
handbooks, are possibly closest to agent-based modeling. For other evolutionary
mechanism, see Chap. 18 in this volume (Chattoe-Brown and Edmonds 2013).
Neural networks are a huge subject as well. This field is currently split in two
streams: On the one hand, research on neural networks as a model of cognitive
processes in the brain. On the other hand, research on neural networks as an
engineering tool for signal processing. A handbook oriented towards cognitive
problems is Arbib (2002). Handbooks oriented towards engineering problems are
Hu and Hwang (2002) and Graupe (2007). Specifically, unsupervised neural
networks are often employed in pattern recognition. A comprehensive treatment
of pattern recognition techniques is Ripley (1996).
All other tools and issues discussed in this chapter are in their infancy, so no
generic reading can be mentioned. Interested scholars are better advised to start
with the original papers mentioned in the bibliography, tracking developments on
recent publications and working papers.
Reasoned Bibliography
This chapter covered too many topics to be able to provide detailed references.
Henceforth, a few basic publications will be listed, that may be used by interested
readers as a first orientation to each of the topics mentioned in this chapter.
Utility and Games
Utility maximization was pioneered by Frank Ramsay and Bruno De Finetti in the
1930s, and subsequently refined by Leonard Savage in the 1950s. Savage still
provides the most comprehensive explanation of this approach to uncertain
reasoning.
Game Theory was initiated by John Von Neumann and Oskar Morgenstern in the
1940s. It subsequently developed into a huge research field within economics, with
several specialized journals. Today, game theory is a field characterized by extreme
mathematical sophistication and intricate conceptual constructions.
This chapter did not focus on the assumptions and methods of Game Theory, but
rather aimed at presenting the main prototypical games that have been devised
hitherto. A classical treatise by Duncan Luce and Howard Raiffa may introduce the
subject more easily than Von Neumann and Morgenstern did. Luce and Raiffa were
first to present the Battle of the Sexes as well as the Prisoner’s Dilemma, which they
ascribed to Albert Tucker anyway.
Readers interested in evolutionary games may rather read the treatises written by
Jörgen Weibull and Herbert Gintis, respectively. The former is more specific on
evolutionary games, but also more technical than the second one.
Robert Axelrod is the main reference so far it regards simulations of the iterated
Prisoner’s Dilemma with retaliation strategies. The idea that the iterated Prisoner’s
Dilemma could yield cooperation simply relying on tags is due to Rick Riolo.
The Stag Hunt and the Game of Chicken are classical, somehow commonsensi-
cal games. The Game of Chicken has been turned into the Hawk-Dove Game by
Maynard Smith and George Price. The Hawk-Dove game is not terribly different
from The War of Attrition, conceived by Maynard Smith and improved by Timothy
Bishop and Chris Cannings.
The Traveller’s Dilemma and The Dollar Auction are recent games invented by
Kaushik Basu and Martin Shubik, respectively. Pure Coordination games have been
discovered by Thomas Schelling.
Axelrod RM (1984) The evolution of cooperation. Basic Books, New York
Basu K (1994) The traveller’s dilemma: paradoxes of rationality in game theory.
Am Econ Rev 84:391–395
Bishop DT, Cannings C, Smith JM (1978) The war of attrition with random
rewards. J Theor Biol 74:377–389
Gintis H (2000) Game theory evolving. Princeton University Press, Princeton
Luce RD, Raiffa H (1957) Games and decision: introduction and critical survey.
Wiley, New York
Riolo RL, Cohen MD, Axelrod RM (2001) Evolution of cooperation without
reciprocity. Nature 414:441–443
Savage L (1954) The foundations of statistics. Wiley, New York
Schelling TC (1960) The strategy of conflict. Harvard University Press,
Cambridge, MA
Shubik M (1971) The dollar auction game: a paradox in noncooperative behav-
ior and escalation. J Conflict Resolut 15:109–111
Smith JM, Price GR (1973) The logic of animal conflict. Nature 246:15–18
Weibull J (1997) Evolutionary game theory. The MIT Press
330 G. Fioretti
Influence Games
Ernst Ising introduced his model in the 1920s. Since then, a huge literature
appeared.
The Ising model is taught in most Physics courses around the world, so a number
of good introductions are available on the Internet. A printed introduction by Barry
Cipra is mentioned here for completeness.
Schelling’s model of racial segregation was developed independently of the
Ising model. However, it may be considered a variation of it.
The El Farol Bar Problem was conceived by Brian Arthur. Renamed The
Minority Game and properly formalized, it was introduced to physicists by Damien
Challet and Yi-Cheng Zhang.
A huge literature on the Minority Game has appeared on Physics journals. Good
introductions have been proposed, among others, by Esteban Moro and Chi-Ho
Yeung and Yi-Cheng Zhang.
Arthur WB (1994) Inductive reasoning and bounded rationality. Am Econ Rev
84:406–411
Challet D, Zhang YC (1997) Emergence of cooperation and organization in an
evolutionary game. Phys A 246:407–418
Cipra BA (1987) An introduction to the Ising model. The American Mathemati-
cal Monthly 94:937–959
Moro E (2004) The minority game: an introductory guide. In: Korutcheva E,
Cuerno R (eds) Advances in condensed matter and statistical physics. Nova Science
Publishers, New York, pp 263–286. (Also available online at http://arxiv.org/abs/
cond-mat/0402651v1)
Yeung CH, Zhang YC (2009) Minority games. In: Meyers RA (ed) Encyclope-
dia of complexity and systems science. Springer, Berlin. (Also available online at
http://arxiv.org/abs/0811.1479v2)
Some Pitfalls of Utility Maximization
The idea that probabilities measured on samples of size zero are somewhat awk-
ward is quite old, and evidently linked to the frequentist view of probabilities.
Daniel Ellsberg circulated this idea among economists, where in the meantime the
subjectivist view of probability judgements had become dominant. Sub-additive
probabilities were conceived by Bernard Koopman in the 1940s and popularized
among economists by David Schmeidler in the 1980s.
Maurice Allais submitted his decision problem to Leonard Savage, who did not
behave according to his own axioms of rational choice. Since then, Savage
presented utility maximization as a normative, not as a descriptive theory. Prospect
Theory was advanced by Daniel Kahneman and Amos Tversky; it comes in a first
version (1953), and a second version (1992).
The preference reversals highlighted by Paul Slovic have triggered a huge
literature. A recent book edited by Sarah Lichtenstein and Paul Slovic gathers the
most important contributions.
Kenneth Arrow originally devised his paradox as a logical difficulty to the idea
of a Welfare State that would move the economy towards a socially desirable
equilibrium. However, it may concern any form of group decision-making.
Michael Mandler is the main reference for a possible conciliation of Slovic’s and
Arrow’s paradoxes with utility maximization, provided that preferences are
incomplete.
Allais M (1953) Le comportement de l’homme rationnel devant le risque:
critique des postulats et axiomes de l’école americaine. Econometrica 21:503–546
Arrow KJ (1950) A difficulty in the concept of social welfare. J Polit Econ
58:328–346. (Reprinted in The collected papers of Kenneth J. Arrow. Blackwell,
Oxford, 1984)
Ellsberg D (1961) Risk, ambiguity, and the savage axioms. Quart J Econ
75:643–669
Kahneman D, Tversky A (1953) Prospect theory: an analysis of decision under
risk. Econometrica 21:503–546
Koopman BO (1940) The axioms and algebra of intuitive probability. Annal
Math 41:269–292
Lichtenstein S, Slovic P (2006) (eds) The construction of preference. Cambridge
Mandler M (2005) Incomplete preferences and rational intransitivity of choice.
Games Econ Behav 50:255–277
Savage L (1967) Difficulties in the theory of personal probability. Philos Sci
34:305–310
Schmeidler D (1989) Subjective probability and expected utility without addi-
tivity. Econometrica 57:571–587
Tversky A, Kahneman D (1992) Advances in prospect theory: cumulative
representation of uncertainty. J Risk Uncertainty 5:297–323
Logic of Consequence and Logic of Appropriateness
An introduction and thorough discussion of the differences between the logic of

consequence and the logic of appropriateness can be found in a book by James
March, A Primer on Decision Making. The same Author, in 1974 and 1976, has
been the first to point to the fact that human beings distort their memories of the past
in order to construct coherent stories that guide them into the future. This point has
been also made by Karl Weick, who wrote a lengthy treatise on this subject a few
decades later.
332 G. Fioretti
A number of psychological experiments confirm this idea. Interested readers

may start with the works of Daryl Bem, Michael Conway, Michael Ross and a book
edited by Ulric Neisser and Robyn Fivush.
However, the trouble with the idea of human beings reconstructing the past is
that they are not willing to concede that they do so. Thus it is extremely difficult to
find case-studies. The one by David Lane and Robert Maxfield is possibly the only
exception, though they were not allowed to published all the material they gathered
during their investigation (Lane, personal communication).
A final remark on lack of communication in this stream of research. James
March, Karl Weick and David Lane worked independently, possibly unaware of
one another, focusing on the same issue but employing different expressions.
Bem DJ (1966) Inducing belief in false confessions. J Pers Soc Psychol
3: 707–710
Bem DJ (1967) Self-perception: an alternative interpretation of cognitive disso-
nance phenomena. Psychol Rev 74:183–200
Cohen MD, March JG (1974) Leadership and ambiguity: the American College
President. McGraw-Hill, New York
Conway M, Ross M (1984) Getting what you want by revising what you had.
J Pers Soc Psychol 47:738–748
Greenwald A (1980) The totalitarian ego: fabrication and revision of personal
history. Am Psychol 35:603–618
Lane DA, Maxfield RR (2005) Ontological uncertainty and innovation. J Evolut
Econ 15:3–50
March JG (1994) A primer on decision making. The Free Press, New York
March JG, Olsen JP (1976) Organizational learning and the ambiguity of the
past. In: March JG, Olsen JP (eds) Ambiguity and choice in organizations.
Universitetsforlaget, Bergen
Neisser U, Fivush R (1994) (eds) The remembering self: construction and
accuracy in the self-narrative. Cambridge University Press, Cambridge
Ross M, Newby-Clark IR (1998) Costructing the past and future. Soc Cognit
16:133–150
Weick KE (1979) The social psychology of organizing. Random House,
New York
Weick KE (1995) Sensemaking in organizations. Sage, Thousand Oaks
Tools for the Logic of Appropriateness
This chapter did not deal with tools where categories pre-exist to the information
that is being received, namely, supervised neural networks and Case-Based Deci-
sion Theory. Readers interested in supervised neural networks may start with the
classical handbook by Rumelhart, McClelland and the PDP Research Group.
Readers interested in Case-Based Decision Theory may refer to a series of articles
by Izhak Gilboa and David Schmeidler.
The earliest intuitions on the nature of mental categories date back to Ludwig
Wittgenstein. A good explanation of the main features of mental categories, and
why they are so different from our common idea of what a “category” is, is provided
by George Lakoff in his Women, Fire, and Dangerous Things.
So far it regards unsupervised neural networks, the classic book by Teuvo
Kohonen is still unrivaled for its combination of mathematical rigour and philo-
sophical insight. Having been written at an early stage, it still keeps a strong link
between artificial neural networks and the human brain.
Paul Thagard is the basic reference for constraint satisfaction networks. Con-
straint satisfaction networks appear in several contributions to the book The Con-
struction of Preference, edited by Sarah Lichtenstein and Paul Slovic, mentioned in
the section “Some Pitfalls of Utility Maximization”. Regarding the importance of
focussing on two alternatives in order to arrive at a decision, see a working paper by
Guido Fioretti.
Evidence Theory started with a book by Glann Shafer in 1976, and triggered a
small but continuous flow of mathematical works since then. An article by Guido
Fioretti explains it to social scientists, along with examples of applications to
decision problems.
Fioretti G (2009) Evidence theory as a procedure for handling novel events.
Metroeconomica 60:283–301
Fioretti G (2011) Either, or: exploration of an emerging decision theory (Work-
ing paper). http://ssrn.com/abstract¼1314352
Gilboa I, Schmeidler D (1995) Case based decision theory. Quart J Econ
110:605–639
Kohonen T (1989) Self-organization and associative memory. Springer, Berlin
Lakoff G (1987) Women, fire, and dangerous things. University of Chicago
Press, Chicago
Rumelhart DE, McClelland JL, The PDP Research Group (1986) Parallel
distributed processing: explorations in the microstructure of cognition. MIT
Press, Cambridge, MA
Shafer G (1976) A mathematical theory of evidence. Princeton University Press,
Princeton
Thagard P (2000) Coherence in thought and action. MIT Press, Cambridge, MA
References
Arbib MA (ed) (2002) The handbook of brain theory and neural networks, 2nd edn. MIT Press,
Cambridge, MA
Aumann R, Hart S (eds) (1992) Handbook of game theory with economic applications, vol 1.
North-Holland, Amsterdam
334 G. Fioretti
Chattoe-Brown E, Edmonds B (2013) Evolutionary mechanisms. Chapter 18 in this volume

Graupe D (2007) Principles of artficial neural networks. World Scientific, Singapore
Hu YH, Hwang JN (eds) (2002) Handbook of neural network signal processing. CRC Press,
Boca Raton
Rasmusen E (2007) Games and information: an introduction to game theory, 4th edn. Blackwell,
Malden
Ripley BD (1996) Pattern recognition and neural networks. Cambridge University Press,
Cambridge
Weibull JW (1997) Evolutionary game theory. MIT Press, Cambridge, MA
Chapter 14
Social Constraint
Martin Neumann
Why Read This Chapter? To understand social norms and their complexities,
including: how they can operate, how they can effectively constrain action and how
such processes have been represented within simulations. The chapter also helps the
reader to acquire an integrated view of norms and become aware of some of the
relevant work simulating them using this framework.
Abstract This chapter examines how a specific type of social constraint operates
in Artificial Societies. The investigation concentrates on bottom-up behaviour
regulation. Freedom of individual action selection is constraint by some kind of
obligations that become operative in the individual decision making process. This is
the concept of norms. The two-way dynamics of norms is investigated in two main
sections of the chapter: the effect of norms on a social macro-scale and the
operation of social constraints in the individual agent. While normative modelling
is becoming useful for a number of practical purposes, this chapter specifically
addresses the benefits of this expanding research field to understand the dynamics
of human societies. For this reason, both sections begin with an elaboration of the
problem situation, derived from the empirical sciences. This enables to specify
questions to agent-based modelling. Both sections then proceed with an evaluation
of the state of the art in agent-based modelling. In the first case, sociology is
consulted. Agent-based modelling promises an integrated view on the conception
of norms in role theoretic and individualistic theories of society. A sample of
existing models is examined. In the second case, socialisation research is consulted.
In the process of socialisation the obligatory force of norms become internalised by
the individuals. A simulation of the feedback loop back into the mind of agents is
only in the beginning. Research is predominantly on the level of the development of
architectures. For this reason, a sample of architectures is evaluated.
M. Neumann (*)
University of Koblenz-Landau, Institute of information systems in Business and Public
Administration, Universitaetsstr. 1, 56070 Koblenz, Germany

336 M. Neumann
14.1 Introduction
Some kind of mechanism for action selection has to be implemented in the agents.
The decision making process of isolated agents may be governed by BDI
architectures. Developed by the philosopher Michael Bratman as a model of
rational decision making – in particular to clarify the role of intentions in practical
reasoning according to the norms of rationality – (Bratman 1987), the BDI frame-
work has been adopted by Rao and Georgeff for the development of software
technology (Rao and Georgeff 1991). BDI agents are data structures that represent
beliefs about the environment and desires of the agents. Desires enable the goal-
directed behaviour of an agent. Moreover, the decision-making process has to be
managed, which is accomplished in two stages. To achieve a goal, a certain plan has
to be selected, denoted as the intention of the agent. Secondly, the agent undertakes
means-ends calculations, to estimate which actions are necessary to reach its goals.
Intentions are crucial in mediating between the agent’s beliefs and desires and the
environment.
In groups, however, behaviour is more effective if agents orient their actions
around other agents. The simplest example is of two robots moving towards each
other. They have to decide how to pass. Hence, individual action selection has to be
restricted by social constraints. In principle there exist two options to model social
constraints on the individual agent’s action selection: top down regulation by a
central authority or bottom-up regulation of action selection. Top-down regulation
can be exhibited computationally by a central processor, or – in human societies –
by coercion of a central authority. An evaluation of modelling approaches to the
former kind of social regulation can be found in the chapter on Power and
Authority. In bottom-up approaches, freedom of individual action selection is
constraint by some kind of obligatory forces that become operative in the individual
decision making process even without a controlling authority. This is denoted by
the concept of norms. Normative behaviour regulation can be enforced by some
kind of generalised social influence such as sanctions and encouragement. Crucial
for the normative behaviour regulation is that social constraints are internalised in
the individual agent’s decision making process. This can range from more or less
automatically executed habits to processes of normative reasoning and balancing
competing goals. In contrast to top-down regulation it is not based purely on
coercion. Since bottom-up regulation is more flexible than pre-determined
constraints by a central authority, the past decade(s) have witnessed a growing
interest in the inclusion of norms in multi-agent simulation models. Numerous
factors are responsible for attention being paid to norms in artificial societies,
ranging from technical problems in the co-ordination of multi-agent system, e.g.
with moving robots (Shoham and Tennenholtz 1992; Boman 1999), or practical
problems such as e-commerce or electronic institutions (Lopez and Marquez 2004,
Vazquez-Salceda et al. 2005) to philosophical interest in the foundation of morality
(Axelrod 1986; Skyrms 1996, 2004) and the investigation of the wheels of social
order (Conte and Castelfranchi 2001). The contribution of Artificial Societies to the
14 Social Constraint 337
latter problem is in the focus of this chapter: what is the potential contribution of
Artificial Societies to the investigation of social order in human societies?
However, in the literature – not only, but also in the simulation literature – a
great variety of concepts of norms exist. Typically, these differences are not much
discussed in the AI literature. Nevertheless, some decisions are made, consciously
or unconsciously. For this reason, this chapter aims to introduce the reader to the
different concepts that are often only implicit in the different approaches to
normative simulation models. This requires some background information about
the different concepts of norms in the different empirical sciences. However, one
reason for the variety of concepts of norms is that their investigation is scattered
over a vast variety of different disciplines. For this reason also this chapter has to
concentrate on some empirical disciplines that are of particular importance for
agent-based modelling. A first restriction is motivated by the decision to concen-
trate on bottom-up behaviour regulation in this chapter. This suggests to exclude the
literature on norms that can be found in the political sciences or the theory of law,
since these norm concepts are more closely related to top-down approaches.
Secondly, agent-based modelling is of particular relevance for the study of social
mechanisms. These mechanisms, however, are of some kind of generality. Typi-
cally, Artificial Societies investigate stylised facts. This suggests focussing the
examination by excluding historical narratives that can be found in anthropological
studies or historical sociology such as the work of Michel Foucault or Norbert Elias.
Instead, the survey of the empirical sciences concentrates on the literature that
investigates the theoretical foundation of the specific dynamics between the micro
and the macro level that is the particular focus of agent-based modelling. Namely,
how norms influence the operations and effects of social systems and how such a
normative structure of the social system recursively affects the generating agent
level. This calls for a survey of sociology and socialisation research in social-
psychology. They are most relevant sciences for an investigation of the contribution
of norms to the wheels of social order.
Beside this introduction, the chapter contains three main sections. Sections 14.3
and 14.4 consist of two parts: one providing a broad overview of the empirical
counterpart and a subsequent one about modelling. The sections can be read
independently. A reader with a background in the empirical sciences who wants
to get informed about simulation might concentrate on the parts evaluating simula-
tion models. The sections provide the following information:
Section 14.2: First, a brief exposition of the core concepts of norms in the
empirical sciences is provided. It is suggested to have a look at it, because here the
core problems are exposed that are investigated in the following sections. These are
the two main parts.
Section 14.3: an investigation of the dynamics of norm spreading from
individuals to a social group and how this effects the operations of the social system
is undertaken in this part. This refers to the sociological question of the operations
and effects of normative behaviour regulation on the social macro-level (the
emergence of a behaviour regularity). The section is divided into two parts: first
the sociological questions to agent-based modelling are developed. Then a sample
338 M. Neumann
of models is examined. The sample is divided into two categories of models:

models inspired by a game theoretic problem description and cognitive models in
an AI tradition. This section is particular relevant for modelling normative regula-
tion in Artificial Societies.
Section 14.4: in this part the recursive feedback loop is investigated. It is
examined, how social behaviour regulation is executed in the individual agent.
This refers to the socio-psychological problem of norm internalisation. To represent
this process, cognitively rich agents are necessary. Also this section is divided into
two parts: first, a problem exposition will be provided. This refers to theories of
socialisation. Then a sample of architectures is examined with regard to the
question how norms become effective within the agent. This section might be the
most interesting one for readers interested cognitive modelling.
Finally, the main findings of the chapter are summarised in concluding remarks
and further reading suggested for those who want to investigate further.
14.2 The Concept of Norms
Before turning the attention to an examination of existing modelling approaches, a

brief exposition of the core concepts of norms will be provided. In some way, some
of these aspects have to be represented in a normative simulation model. Dependent
on the research question, the developer of a normative simulation model may
concentrate on only some aspects. In fact no model includes all aspects. However,
it might be useful to be aware of a more comprehensive perspective. Operationally,
norms can be described as correlated structures. Agent behaviour exhibits
regularities that can be found – to a certain degree – in a entire population. This
can be described as a social constraint. From a social science perspective, norms are
the most important concept of social constraints on behaviour regulation. Norms
belong to the fundamental concepts in sociology to explain human behaviour.
Summarising the individual and social aspects of norms, Gibbs provided the
definition that “a norm is a belief shared to some extent by members of a social
unit as to what conduct ought to be in particular situations or circumstances” (Gibbs
1981, p.7). Unfolding this concise definition reveals three essential components:
• An individual component: a belief
• A social component: the belief is shared by other members of a social unit
• A deontic: a conduct is obliged
1. The capability to understand the meaning of a deontic language game is acquired
in the childhood. The deontic prescribes individual behaviour. The capacity to
play language games that are centred around words such as ‘you ought’ is the
precondition for a normative orientation. It allows for the internalisation of
norms. Although in the past decades consideration has been given to understand-
ing processes of internalisation over an individual’s entire life span, the central
processes arguably take place during childhood. Indeed, the transmission of
cultural values has even been denoted as a ‘second birth’ (Claessens 1972). An
individual may become accepted into a wider society through a variety of means.
Norm internalisation is argued to represent one of the stronger mechanisms by
which this process occurs. In contrast to compliance under coercion or to simply
copying other behaviour patterns, internalisation, it is argued, is coupled with an
individual’s intrinsic motivation and sense of identity. The mechanisms by
which this occurs are the focus of socialisation research. Socialisation is the
bridge between the individual and society (Krappmann 2006). Hence,
socialisation research is at the border of psychology and sociology, and
contributions from both disciplines can be found in the literature.
2. It is a central element of the concept of norms that it implies both a psychological
component, namely the mental state of a belief, as well as a social component.
This becomes apparent by the fact that this belief is shared by a number of
individuals. Norms are thus essential for a comprehension of the relation
between structure and agency. This is often denoted as the micro–macro link.
While the aspect of agency can be found in the psychological aspect of norms,
structure is entailed in the social prescription. This makes norms a fundamental
building block of the wheels of social order. An understanding of norms is thus
crucial for an understanding of the social world. They are a central mechanism in
the two way dynamics creating of social order. On the one hand, individual
interactions might produce social macro-patterns, namely a normative order. On
the other hand, once the agents on the individual level recognise patterns on the
resultant social level, these patterns might have an effect on the mind of the
individual agent, namely the recognition of a deontic prescription. This complex
dynamics can be embraced in the following schema:
This schema can be illustrated by an example: interactions of individual actors
might lead to a certain aggregate result, for instance, social stratification. This
macro-property, however, might have an effect on how the actor cognitively
structures the perception of the world. This is an intra-agent process. For instance,
actors might regard themselves as helpless losers or they might think that you can
get what you want. This world perception, in turn, has an effect on how these actors
behave in the interactions between the agents. Now the feedback loop is closed: the
interactions result again in an aggregated macro-structure. The reader may think of
other examples appropriate for his or her purpose. Presumably, it will turn out that
the dynamics of the particular models can be described in terms of such a two way
dynamics between inter – and intra agent processes (Fig. 14.1).
14.2.1 The Sociological Perspective
In this section process (1) is considered: the emergence and the properties of a
normative macro structure. Hence, this section will concentrate mainly on the social
aspect of norms, i.e. the one-way dynamics of inter agent processes. Before turning
the attention to existing modelling approaches, the questions will be developed that
340 M. Neumann
Fig. 14.1 The two-way

dynamics of norms
Structure
Inter – agent Intra – agent

process process
Agency
need to be posed to normative multi-agent models for a comprehension of the

effects and operations of norms in a social system. For this purpose sociological
theory will be utilised. Readers that are either familiar with this subject or with a
purely practical interest might go on to Sect. 14.3. (2). However, a more or less
implicit reference to the problems posed by social theory can be discerned in most
of the simulation models.
The function of normative behaviour regulation for social integration (i.e. a
social structure) has been particularly emphasised by the classical role theoretic
account within the sociological theory (Neumann 2008b). It can be traced back to
the functionalist sociology of Durkheim and Parsons. This theoretical account was
decisively influenced by anthropological research of the late 1920s when cultural
anthropologists discovered the variety of the organisation of the social life and the
correlation of the structure of personality and society (Malinowski 1927; Mead
1928). Role theory claims that action is guided by a normative orientation, insofar
as social roles are described by social norms. The concept of norms is introduced in
this account as a structural constraint of individual actors. This theory of action is
often paraphrased as the ‘homo sociologicus’. For a long time this remained the
dominant stream of sociological theory.
To specify where norms are placed in the theoretical architecture, let us consider
the example of the famous Mr Smith, introduced by Ralf Dahrendorf (1956) in his
analysis of the ‘homo sociologicus’ to characterise key elements of sociological
role theory. We first meet him at a cocktail party (in the early 1950s) and want to
learn more about him. What is there to find out?
Mr Smith is an adult male, ca. 35 years old. He holds a PhD, and is an academic.
Since he wears a wedding ring, we know that he is married. He lives in a middle-
sized town in Germany and is a German citizen. Moreover, we discover that he is
Protestant and that he arrived as a refugee after the second World War in a town
populated mostly by Catholics. We are told that this situation caused some
difficulties for him. He is a Lecturer by profession and he has two kids. Finally,
we learn that he is the third chairmen of the local section of a political party, Y,
a passionate and skilful card player and a similarly passionate though not very good
driver. This approximates to what his friends would tell us.
In fact, we may have the feeling that we know him rather better now. After all,
we have some expectations as to how a lecturer is likely to behave. As a lecturer he
stands in certain relations to colleagues and pupils. As a father he will love and care
for his children, and card playing is also typically associated with certain habits. If
we know the party Y, we will know a lot more about his political values. However,
all that we have found out represent social facts. There are a lot more lecturers,
fathers and German citizens beside Mr Smith. In fact, none of this information tell
us anything about the unique identity of Mr Smith. We simply discovered informa-
tion about social positions, which can, of course, be occupied by varying persons.
However, social positions are associated with specific social roles. Roles are
defined by the specific attributes, behaviour and social relations required of them.
Demands of society determine—to a certain degree—individual behaviour.
Individuals are faced with obligations and expectations. This social demand is
transmitted to the individual by norms. Norms are the ‘casting mould’ (Durkheim
1895) of individual action. They regulate how lecturers, fathers and members of
political parties ought to act to fulfil the role expectations of society. However
nowadays, it is improbable that we will be told anything about driving competence.
Thus, we have learned another lesson: Norms may change over the course of time.
In particular, Talcott Parsons (1937) emphasised that the ends of individual actions
are not arbitrary, but rather are prescribed by social norms. Thus, norms represent a
key concept within sociological role theory.
To examine the explanatory account of this theoretical approach, the present
investigation will abstract from a description of the content of concrete norms. We
concentrate on the methodological characteristics of norms in general, rather than
the content of specific norms. On closer inspection of this example, we find out that
the concept of social norms is characterised by three key elements:
First, norms show some degree of generality. They are regarded as the ‘casting mould’ of
individual action (Durkheim 1895). The very idea of role theory is that a social role must not
be restricted to a unique individual. For instance, the roles of lecturer or chairman of a
political party can be performed by different individuals. It might not, of course, be arbitrary
as to who will play this role. In fact, it is a major focus of the empirical counterpart of role
theory, statistical analysis of variables, to investigate the distribution of roles. For instance,
monetary background might determine an individual’s chances of securing an academic
position. Nevertheless, academic positions are a feature of society, not the individual. In the
classical account, the generality of norms is simply a given. However, agent-based models
start from individual agents. Thus, in the individualistic approach norms have to spread in
some way from one agent to another to gain generality. The explanation of norm spreading
is essential for a reconstruction of social norms in terms of individual actors.
Secondly, the role set of father or lecturer encompasses a huge action repertoire. The
choice of a concrete action cannot be determined only solely by an external force. The ends
of an action have to be determined internally by the individual actor executing a specific
role. This means that the ends of individual actions are (to a certain degree) determined by
the society. For instance, it is a social norm how a caring father would look like and what
kind of actions are to be undertaken. In fact, this varies between societies. This knowledge
is often denoted as internalisation, even though the psychological mechanisms are not in
the focus of sociological theory. Thus, already a comprehension of the inter agent processes
that constitute a social role calls for a related subjective element at the level of the
individual agent.
342 M. Neumann
Thirdly, this approach is characterised by a certain type of analysis: the normative

integration of society (Parsons 1937; Davies and Moore 1945; Merton 1957). Hence, the
question is to a lesser extent concerned with the origin of norms than with the function of
norms for a society. For instance, the role of the father is to educate his child. The role of the
lecturer is crucial for the socialisation of pupils. Thus, both roles are functionally relevant
for the reproduction of the society.
However, in the past decades, this paradigm has been severely criticised: Firstly,
role theory has been criticised for sketching an oversocialised picture of man
(Wrong 1961). Already in the 1960s, Homans (1964) claimed to ‘bring man back
in’. In fact, the individual actors of the role theory have often been regarded as more
or less social automata. If they have properly internalised the norms, they execute
the program prescribed by their roles. Secondly, role theory is built on the claim
that social phenomena should be explained with social factors (Durkheim 1895).
Roles are a pre-given element in this theoretical architecture. Roles (and thus:
norms) emanate from society. However, the origin of both roles and society is
left unexplained. In this so-called functionalist methodology of the role theoretic
account, it is argued that to perform a social role is to perform a social function and
this social function is argued to be the source of the norm. However, this perspec-
tive lacks a description of a mechanism by which norms become effectively
established. This deficit suggests to ‘bring man back in’.
In fact, the explanatory deficit engendered an alternative. Others have suggested
building sociology on the foundations of individual actors. This built on views
advocated by John Stuart Mill. This is the programme of the so-called methodolog-
ical individualism (Boudon 1981; Raub and Voss 1981; Coleman 1990; Esser
1993). A shift ‘from factors to actors’ (Macy and Willer 2002) can be observed in
the foundations of sociology in the recent decades. This approach regards the
appearance of norms as an aggregate product of a sum of individual actions. This
approach to norms has been particularly stressed by Rational Choice Theories. The
deficit of this approach, however, is a lack of a cognitive mechanism that could
explain the deontic component of norms, i.e. how the social level may exhibit an
obligatory force on the individual by norms. Agent-based modelling promises to
overcome the complementary deficits of both theoretical accounts:
On the one hand, agent-based models contribute to the individualist theory
building strategy: obviously, agents are the fundamental building block of agent-
based models. In agent-based simulation models (Artificial Societies), structures
emerge from individual interaction. On the other hand, however, the great advan-
tage of this methodology is that it allows to explicitly consider the complexity
generated by individual interactions. Compared to purely analytical models it is a
great advantage of simulation approaches that they are not restricted to single or
representative actors. This complexity generated by this particular feature enables
the investigation of the feedback loop between individual interaction and collective
dynamics. This has led to a growing awareness of its potential for investigating the
building-blocks of social structure. For this reason, it is even claimed that agent-
based simulation allows us to “discover the language in which the great book of
social reality is written” (Deffuant et al. 2006), by constituting the promise to
understand “how actors produce, and are at the same time a product of social
reality” (Deffuant et al. 2006). Since structures on the macro level are a product
of individual interaction, a causal understanding of the processes at work seems
possible that might fill the gap between individual action and social structure with
agent-based models.
This feature of agent-based modelling suggests that this methodology enables to
bring together the role theoretic perspective that norms are a structural constraints
and the individualistic perspective of norms as an aggregated product of a sum of
individual actions. Hence, it might provide both a causal mechanism of how
normative action constraints get established as well as how social forces exhibit
cognitive power. For this reason the focus will now shift to an evaluation of
simulation models.
14.2.2 Simulation Models of Norms
After the theoretical problem exposition, the particular focus of this section is an
investigation of the contribution of simulation experiments to sociological theory.
The recourse to sociological theory in Sect. 14.3.(1) revealed the following
questions:
• Can they provide insights into the normative regulation of society; that is, do
they also reproduce the findings of a functional analysis of the effects and
operations of norms in a society (focus of contribution)?
• Moreover, do they allow for a causal reconstruction of the mechanisms that
generate the functional interconnectedness on the social level? This implies that
two further questions have to be addressed:
• What transforms the agents in such a way that they factually follow norms? That
is, what are the causal mechanisms at work that enable an internalisation of
norms (transformation problem)?
• By what mechanisms in the model can norm-abiding behaviour spread to or
decay from one agent to another (transmission problem)?
These questions will be examined in this section. It has to be emphasised that the
investigation concentrates on methodology, not on the contents of norms governing
concrete roles such as father or lecturer. However, existing models are clustered
around various intuitions about norms, conventions or standards of behaviour. The
concrete research question differs from model to model. Some models concentrate
on the emergence or spreading of norms. Others concentrate on functional aspects
or the feedback of norms on individual agent’s behaviour. A multiplicity of
concepts is at hand. Hence, a comprehensive review of all models and accounts
that may be in some way related to the study of norms would go beyond the scope of
this investigation.
The overwhelming mass of models, however, can be traced back to (or is at least
influenced by) two traditions in particular: first, game theory and secondly an
344 M. Neumann
architecture of cognitive agents with some roots in Artificial Intelligence. Tradition

and theoretical background has a direct impact on the terminology used. Depending
on their background, the models tend to be communicated in different scientific
communities. Additionally, references in articles tend to depend on their authors’
background. Under the perspective of content, the models in the AI tradition
typically contain references to conceptual articles relating to agent architectures.
Articles with models in a more game theoretical tradition typically refer to game
theoretic literature for the characterisation of the interaction structures in which the
authors are interested. Of course, this tradition-influenced framing, publishing and
referencing is a tendency. It does not constitute a clear-cut disjunction without any
intersection. It has to be emphasised that this is neither a very precise nor a
disjunctive categorisation. To some degree, the distinction between game theory
and DAI is a distinction in the mode of speech employed by the authors. Some
problems of game theoretic models could also be formulated in a DAI language and
vice versa. The categorisation of models as following the DAI tradition shall only
indicate that the agents employed by these models are in some way cognitively
richer than those in the so-called game theoretic models.
Nevertheless, this distinction gives a rough sketch of the line of thought followed
by the models, and also of the kind of problems, the concepts for their investigation,
and the mode of speech in which the paper is presented. Moreover, this
categorisation provide hints to other areas of research that are closely related to
the models considered in this article: For instance, simulation is only a small sub-
discipline of game theory in general and the distinction between analytical and
simulation results is only gradual. Simulation models might describe problems in
game theoretic terms, but the method of resolution is not that of analytical game
theory (Binmore 1998). In fact, investigating norms with the means of analytical
game theory is a highly active research field.
14.2.2.1 Models Using a Game Theoretic Problem Description
First a closer examination of a sample of models applying a game theoretic problem

description will be undertaken. The models investigated here build on the frame-
work described in the chapter on games and utility in this handbook (Fioretti 2013).
Axelrod (1986) studies the evolution of a standard not to cheat via a “norms game” and a
“meta-norms game”. In the “norms game” defectors may be punished by observers. In the
“meta-norms game” it is also the case that observers of a defection that do not punish the
defector may be punished. Only the latter game leads to a widespread standard of not
defecting.
Coleman (1987) investigates the effect of interaction structures on the evolution of co-
operation in a prisoner’s dilemma situation. Only small groups can prevent the exploitation
of strangers.
Macy and Sato (2002) examine the effect of mobility on the emergence of trust among
strangers in a trust game. While agents with low mobility trust only their neighbours, high
mobility supports the evolution of trust among strangers.
Vieth (2003) investigates the evolution of fair division of a commodity in an ultimatum

game. Including the ability to signal emotions leads to a perfectly fair share. If detection of
emotions is costly the proposals even exceed fair share.
Bicchieri et al. (2003) present a model of a trust game. It demonstrates how a trust and
reciprocate norm emerges in interactions among strangers. This is realised by several
different conditional strategies.
Savarimuthu and colleagues (2007) study the convergence of different norms in the
interactions of two different societies. Both societies play an ultimatum game against each
other. Two mechanisms are examined: a normative advisor and a role model agent.
In the model by Sen and Airiau (2007) a co-ordination and a social dilemma game are
examined. Agents learn norms in repeated interactions with different agents. This is
denoted as social learning to distinguish this interaction type from repeated games with
the same player. The whole population converges to a consistent norm.
Obviously, all models have been developed for differing concrete purposes. To
examine the extend to which these models capture the explanatory problems of the
contribution problem, transformation problem and transmission problem, the vari-
ous accounts of the different models will be outlined in a table. Moreover, a short
hint to the concrete implementation is provided. This will enable an evaluation
inasmuch normative agent-based models have so far reached the goal to discover
‘the language in which social reality is written’.
Contribution Transformation Transmission Implementation

Axelrod Norm conceived!) Sanctions Social learning;
dynamics replicator
(norms dynamics
broadly
Dynamical
propensities
Coleman Norm Punishment by defections (a) Group size Conditional
dynamics (memory restrictions (acquaintance) strategies
for identifying (b) Additionally:
defections as sanctions) replicator
dynamics
Macy & Sato Norm Losses by exclusion from Social learning Dynamical
dynamics interaction propensities
Vieth Norm Losses by rejection Social learning; Dynamical
dynamics replicator propensities
dynamics
Bicchieri et al. Norm Sanctions by retaliating Strategy Conditional
dynamics super game strategies evolution; strategies
replicator
dynamics
Savarimuthu Norm Losses by rejection; advice Advice updating Dynamical
et al. dynamics; based on propensities
functional collective
analysis experience
Sen & Airiau Norm Experience Social learning Dynamical
dynamics guiding propensities
behaviour
convergence
346 M. Neumann
Lessons
The classical model employing a game theoretical approach for the problem descrip-
tion is Axelrod’s model. The main contribution of this approach is a clear under-
standing of the emergence of (commonly shared) normative behaviour constraints.
The starting point of the models is a dilemma situation. This is a consequence of the
game theoretic problem description. Simulation allows for an evolutionary perspec-
tive by analysing repeated games. Typically, in the long run and under specific
conditions (which vary from model to model) it is possible that behaviour
conventions emerge, that are in the benefit of all agents or that represent some
intuitions about fairness. The diffusion of a behavioural regularity is then regarded
as a norms. The subjective side of an obligatory force is not in the focus of this
approach. Hence, what lessons can be gained from the investigation of this concep-
tion to model normative action constraints with regard to the specific questions?
1. Contribution
Already Axelrod’s classical model provides a causal explanation for norm
spreading. This includes a designation of mechanisms of norm transmission
and normative transformation. An investigation of the functional effect of
norms on the society is left aside. This orientation remained predominant in
this line of research. Typically models with a game theoretic background con-
centrate on the question of norm dynamics. They ask how a behaviour regularity
emerges in an agent population. This is the problem of the Rational Choice
tradition in sociological theory, namely the perspective of norms as an aggregated
product of a sum of individual actions.
2. Transformation
It is striking that, except for the model by Sen and Airiau, the transformation of
individual behaviour in all models is driven by some kind of sanctions. However,
also in the Sen & Airiau model agents react to losses of utility values. This is the
causal mechanism of norm spreading. The great advantage of this account is to
shed light on the process of norm change. As it has become apparent in discussing
Mr Smith, this process can also be observed in human societies. However, norm
change is only barely captured by the functional account of role theory. On the
other hand, the models of this tradition only include a restricted functional
perspective: On an individual level, the agents’ choice of action is guided by
the functional consideration of calculating the expected utility. However, a
corresponding analysis on the social macro-level can be found only in
Savarimuthu et al.’s model.
3. Transmission
With regard to the transmission of norms it is striking that social learning is
implemented in many game theoretic models by a replicator dynamics. Typi-
cally this is interpreted as social learning by imitation. If applied in a context
where no real natural selection, rather than some kind of learning is at work, then
using a replicator dynamics amounts to saying: Somehow the individuals learn in
a way that – measured by the relative overall success of their type of behaviour –
more successful types of behaviour become more frequent. As an effect this may
be true. However, no mechanism is indicated. In this dimension, the models

struggle with the same kind of problem as functional analysis which the individ-
ualistic program tries to resolve; namely, the lack of a causal explanation.
4. Implementation
From the perspective of the role theory of action, a weakness of this approach
becomes apparent, one immediately related to the game theoretic problem descrip-
tion. Agents are faced with a strategic (binary) decision situation. Thus, they have a
fixed set of behaviour (Moss 2001). For this reason behaviour change can be
implemented by dynamical propensities (i.e. the propensity to defect is
dynamically updated). Faced with this situation, agents choose the alternative
that maximises their expected utility. However, behaviour change goes not along
with goal change. Agents can do no more than react to different environmental
conditions. The agents’ behaviour is guided strategic adaptation. An active ele-
ment of normative orientation in the choice relating to the ends of action cannot be
found in a game theoretic approach. This is simply due to the fact that agents do not
possess any internal mechanism to reflect and eventually change their behaviour,
other than the desire to maximise utility. This point has already been highlighted in
Parsons’ critique of ‘utilitarian theories’ of action (Parsons 1937): namely, that the
ends of individual actions are in some way arbitrary. Even though the modelling of
behaviour transformation is the strength of this kind of models, the ends of the
action remain unchanged: the goal is to maximise utility. In this respect, the relation
between the action and the ends of the action remains arbitrary.
However, the very idea of role theory is to provide an answer to the question:
Where do ends come from? Parsons’ (and Durkheim’s) answer was the
internalisation of norms. A corresponding answer to this problem is not supplied
in game theoretical models. This is due to the fact that agents do not act because
they want to obey (or deviate from) a norm. They do not even ‘know’ norms.
Even though the model provides a mechanism for the transformation of the
agents, this is not identical with norm internalisation. This remains beyond the
scope of this account. The agents’ behaviour can only be interpreted as norma-
tive from the perspective of an external observer. Thus, transformation is not
identical with internalisation. While the model provides a mechanism for
behaviour transformation, it cannot capture the process of internalisation. Com-
pared to the classical role theory, this is a principle limitation of a game
theoretical description of the problem situation.
14.2.2.2 Models Utilising Cognitive Agents
This shortcoming calls for cognitively richer agents. For this reason, a sample of
models in the AI tradition will be examined more closely.
Conte and Castelfranchi (1995b) investigate three different populations of food gathering
agents: aggressive, strategic and normative agent populations. Aggressive agents attack
‘eating’ agents, strategic agents attack only weaker agents and normative agents obey a
finder-keeper norm. The aggregated performance of the normative population is the best
with regard to the degree of aggression, welfare and equality.
348 M. Neumann
In an extension of the above model, Castelfranchi et al. (1998) study the interaction of
the different agent populations. Interaction leads to a breakdown of the beneficent effects of
norms, which can only be preserved with the introduction of normative reputation and
communication among agents.1
Saam and Harrer (1999) present a further extension of Conte and Castelfranchi’s model.
They investigate the influence of social inequality and power relations on the effectiveness
of a ‘finder-keeper’ norm.
Epstein (2000) examines the effect of norms on both the social macro and the individual
micro level. On the macro level, the model generates patterns of local conformity and
global diversity. At the level of the individual agents, norms have the effect of relieving
agents from individual thinking.
Flentge et al. (2001) study the emergence and effects of a possession norm by processes
of memetic contagion. The norm is beneficent for the society, but has short-term
disadvantages for individual agents. Hence, the norm can only be retained in the presence
of a sanctioning norm.
Verhagen (2001) tries to obtain predictability of social systems while preserving
autonomy on the agent level through the introduction of norms. In the model, the degree
of norm spreading and internalisation is studied.
Hales (2002) extends the Conte/Castelfranchi model by introducing stereotyping
agents. Reputation is projected not on individual agents but on whole groups. This works
effectively only when stereotyping is based on correct information. Even slight noise
causes the norms to breakdown.
Burke and colleagues (2006) investigate the emergence of a spatial distribution of a
binary norm. Patterns of local conformity and global diversity are generated by a decision
process which is dependent on the local interactions with neighbouring agents.
The contribution of these models to the questions specified above can be

summarised in the following table, which also includes a brief remark on the
implementation specification.

Conte & Functional /b /b Conditional
Castelfranchi analysis strategies
Castelfranchi Functional Updating Updating knowledge Conditional
et al. analysis conditionals by experience strategies
(of strategies) (and
through communicationa)
knowledge
Saam & Harrer Functional (a) /b (a) /b Conditional
analysis (b) Internalisation (b) Obligationa
a
strategies
Epstein Norm Observation Social learning Dynamical
dynamics; updating
functional
analysis
Flentge et al. Functional Memetic Contact Conditional
analysis contagion strategies
(continued)
1
For a more in-depth discussion of this model the interested reader is referred to the chapter on
reputation (Giardini et al. 2013).

Verhagen Norm Internalisation Communication Decision tree
dynamics
Hales Functional Updating Updating knowledge Conditional
analysis conditionals by experience strategies
(of strategies) (and
through communicationa)
knowledge
Burke et al. Norm Signals Social learning Dynamical
dynamics guiding propensities
behaviour (Threshold)
convergence
a
Only in a second experiment
b
The agents are/are not already moral agents
Lessons
The classical model of this kind of models is the one developed by Conte and
Castelfranchi in (1995b). It was the starting point for several extensions. While the
scope of these models has been significantly extended in the past decade, the most
significant contribution of this approach still can be regarded as to enhance the
understanding of the operation and effects of normative behaviour constraints.
Epstein, Burke et al. and Verhagen do not study specific norms but examine
mechanisms related to the operations of norms. In particular, the spreading of
norms is studied by these authors. In this respect, they recover and refine (by the
notion of local conformity and global diversity, a pattern that cannot be found in
game theoretic models) the findings of game theoretic models with the means of
cognitive agents. The other models concentrate mainly on studying the effects and
operations of specific norms in the society. Analysed are questions such as the
effect of possession norms, for instance under the condition of social inequality and
power relations. This is strongly influenced by Conte and Castelfranchi’s (1995b)
problem exposition.
What lessons can be gain from the investigation of this conception to model
normative action constraints with regard to the specific questions?
1. Contribution
A striking feature of these models is that they demonstrate how agent-based
models are able to contribute to a functional analysis of norms. However, in
contrast to the classical scheme of functional explanations in the social sciences,
this result is reached by interactions of individual agents. Moreover, in these
model a much stronger notion of norms is deployed than typically in game
theoretic models. Norms are not just reached by mutual agreement, but are an
explicitly prescribed action routine. The concept of norms in these models is in
line with the role theoretical conception of norms as a structural constraint of
individual actors. This conception of norms allows for a wider field of
applications that could cover the role theoretic norm conception: these can be
interpreted as internalised properties of the agents.
350 M. Neumann
2. Transformation
There exist a wide range of varieties how agents change their behaviour.
Behaviour transformation is not as straightforward as in game theoretic models.
However, it has to be emphasised that the very first model of Conte and
Castelfranchi (1995b) did not include any behaviour transformation at all.
Agents have no individual freedom in this model. As critics accuse the role
theory, the action repertoire is also (depending on conditions) deterministic.
Thus, even though the authors succeed in ‘bringing man back in’, the agents in
the model are merely normative automata. Insofar as the norms are a pre-given
element in the model, the approach can also be regarded as an ‘over-socialised’
conception of man.
This limitation has been overcome by the subsequent developments. With
regard to the transformation problem, the agents have become more flexible than
in the very first model. However, a key difference to game theoretic models still
remains: while game theoretic models mostly concentrate on sanctioning, in
models of cognitive agents sanctions are only employed by Flentge et al. as the
transformation mechanism.
However, while the norms in these models can be interpreted as internalised
properties of the agents, an investigation of the process of internalisation is only
in the beginning. So far no commonly accepted mechanism of internalisation has
been identified. Memetic contagion is a candidate. In Verhagen’s model a quite
sophisticated account is undertaken, including a self-model, a group model and a
degree of autonomy. It is highly advanced in constructing a feedback loop
between individual and collective dynamics. By the combination of a self- and
a group-model a representation of the (presumed) beliefs held in the society is
integrated in the belief system of individual agents. Conceptually, this is quite
close to Mr Smith. However, it might be doubted whether the mechanisms
applied are a theoretically valid representation of real processes.
3. Transmission
Complementary to the wide range of different mechanisms of agent transforma-
tion, also a variety of different transmission mechanisms are applied. Basically,
agents apply some kind of knowledge updating process, if agent transformation
takes place at all. Up to date the transmission problem is no longer a blind spot of
cognitive agents as it was the case in the Conte and Castelfranchi (1995b) model.
By comparison, communication plays a much more important role than in game
theoretic models and is much more explicitly modelled in models within the AI
tradition. The processes utilised are more realistic mechanisms than the
replicator dynamics of game theoretic models. However, no consensus has
been reached, what an appropriate mechanism would be. This is also due to
the fact, that a modelling of agent transformation and norm transmission is
computationally more demanding than in game theoretic models. It has to be
emphasised, however, that with regard to the transformation and the transmis-
sion problem, the borderlines of both approaches are no longer clear-cut. The
models of Verhagen and Savarimuthu et al. include elements of the other line of
thought.
4. Implementation
Since actions performed by cognitive agents cannot be reduced to the binary
decision to cooperate or defect, more complex behaviour rules than dynamic
propensities have to be applied. The dominant approach for the implementation
of normative behaviour constraints in cognitive agents is based on a straightfor-
ward intuition, namely to apply conditional strategies that are conditionally
based on the agent’s knowledge base. The strategy repertoire, however, depends
on the concrete model.
The overview of existing models has revealed that the focus of their contri-
bution is mainly on the dynamics on the macro level. The questions of norm
transmission and the focus of contribution concentrate on the social macro-scale.
The analysis is focused on the one-way dynamics of effects and operations of
norms on the social level. Hence, the analysis is focused on the emergence of
social structure out of individual interaction rather than on the relation between
structure and agency. This is one aspect of the full dynamics, namely the
following process:
inter-agent processes : interaction ) macro-property : structure
14.3 Socialisation: Norm Internalisation
It has been outlined, however, that the definition of a norm posses social and
psychological components. Norms are essential for a comprehension of the relation
of structure and agency. While processes of emergence of aggregated behaviour
standards from interaction among individual agents has been extensively studied, a
comprehension of the reverse process how the aggregate level gets back into
agents’ minds is not as yet fully reached. A full comprehension of normative
behaviour regulation, however, has also to include the reverse dynamics of the
effect of social structure on the individual agency. Already the problem of agent
transformation refers to the effect of structure on the level of the individual agent.
This is the most problematic aspect in the agents’ design. It would include the
following dynamics:
structure ) Intra-agent processes : agency
This would be a step towards closing the feedback loop of the two-way dynam-
ics. Obviously, intra agent processes are closely related to learning. The reader can
find a more comprehensive examination of the state of the art in the chapter on
evolution and learning. In particular, the effect of structure on (a transformation of)
individual agency is particular relevant for studying the effects of norms, if agency
is not restricted to a representative agent.
To represent such intra-agent processes, in particular the concept of social
learning is well know in agent-based models. It is applied in a number of game
352 M. Neumann
theoretic models and can also be found in models of the AI tradition. The concepts
of social learning, but also knowledge updating, can be traced back to
behaviouristic psychological theories. Behaviourism is a theory of learning devel-
oped principally through experiments with animals. For instance, the conditioning
experiments of Ivan Pavlov are well-known: he demonstrated that dogs can be
trained to exhibit a specific reaction such as salivation by presenting a specific
stimulus such as the sound of a bell together with food (Pavlov 1927). Bandura
(1962, 1969) extended the behaviouristic approach with a social dimension by
developing a theory of social learning through imitation. From a behaviouristic
perspective, norms constitute a learned behaviour and thus have to be explained
using these theories. The dynamical propensities of models inspired by game
theoretical concepts are a straightforward implementation of such a view on intra
agent processes. The propensity to co-operate or defect is updated in proportion to
the propensity of sanctions. The propensity of sanctions, however, is a structural
component resulting from inter-agent processes. Hence, agents learn to modify
their behaviour according to structural conditions.
Here we find the feedback loop between social and individual components that
are in fact essential for the concept of norms. However, the third component is
missing: this approach does not include a concept of obligations. Deontics are out of
the scope of this approach. This shortcoming can be traced back to the psychologi-
cal theory that is represented in the agents: behaviourism is not capable of capturing
mental processes. Indeed, it specifically avoids commenting on the mental pro-
cesses involved. Under the influence of positivism, reference to unobservable
entities such as the ‘mind’ has been regarded as not scientifically valid. Obligations
are such unobservable entities. Hence, they cannot be represented by the means of
behaviouristic learning theories that are applied in agent models.
In socialisation research, the complex cognitive processes necessary for grasping
the meaning of an obligation is denoted as internalisation. It has already been
shown that agent transformation is not the same as the internalisation of norms. This
is also behaviourally important because normative behaviour, guided by deontics,
need not be a statistical regularity, guided by propensities. In particular if moral
reasoning is involved, deviant behaviour is not explained by chance variation,
leading to some kind of normal distribution (where the mean value might be
updated). There is a difference between norms and the normal.
To represent a complex cognitive concept such as norm internalisation calls for
the cognitively rich agents of the AI tradition. However, the examination of current
models has revealed that a comprehension of the cognitive mechanisms by which
social behaviour regulation becomes effective in the individual mind is still in its
fledgling stages. It has been shown that a multiplicity of concepts is at hand: While
in the very beginning the agents where merely normative automata there exist
conceptualisations of normative agent transformation ranging from updating
conditionals (of strategies) through knowledge to signalling and memetic conta-
gion. However, no consensus is reached what are the most relevant mechanisms. It
can be suspected that they remain effect generating rather than process representa-
tion mechanisms. As an agenda for the next decade, a closer examination of the
processes by which normative obligation become accepted by humans might be

useful. For this purpose it is necessary to recall socialisation theory. This will help
to clarify the problem situation. However, the results of socialisation research are
not unequivocal. Hence, in the development of a normative architecture some
fundamental decisions have to be made.
14.3.1 The Perspective of Socialisation Research on Norms
To grasp an understanding of the decisions that have to be made in the cognitive

design of an agent, first some fundamental aspects of theories of socialisation are
briefly highlighted. Subsequently, current architectures of normative agents are
evaluated with regard to the questions posed by the empirical science.
Broadly speaking a conceptual dichotomy of two main approaches can be
identified Geulen (1991) in socialisation research concerning the relation between
the individual and society: One position assumes a harmony between, or identity of,
the individual and society. Philosophical precursors of this approach are Aristotle,
Leibniz and Hegel. The second position stands in contrast and postulates an
antagonism between the individual and society. Within this position two further
standpoints can be distinguished already in the philosophical tradition: Hobbes, for
example, is representative of the argument that society should tame the individual.
By contrast, the position of Rousseau is paradigmatic of an approach that advocates
the need for releasing the individual from society. Both philosophers share the
assumption that an antagonism exists between the individual and society, although
they disagree about the implications.
As it has been outlined, socialisation research sits at the border of psychology
and sociology, and contributions from both disciplines can be found in the litera-
ture. From a sociological perspective, the beginning of investigating socialisation
processes cumulated in the work of Emil Durkheim, founding father of sociology
and professor of pedagogy. Starting from a clinical and psychological perspective,
Sigmund Freud developed a theory of socialisation, which in many aspects is
surprisingly akin to Durkheim’s approach.
The early theories of Freud and Durkheim agree in that they assume an antago-
nism between individual and society. Durkheim asserted that the individual consists
of two parts: first, a private domain that is egoistic and guided purely by basic
drives. The egoistic domain corresponds to that of the newborn child. The original
human is a ‘tabula rasa’ in which social norms have to be implemented. Only
through the process of socialisation do humans become socially and morally
responsible persons. This is the second ‘part’ of the individual. Durkheim claimed
that the best of us is of a social nature. Society, however, is coercive (Durkheim
1895), and can even compel individuals to commit suicide (Durkheim 1897).
Norms are finally internalised once the individual no longer perceives this coercion
(Durkheim 1907). Yet for Durkheim coercion none the less remains. As with
Durkheim, Freud assumed the existence of an antagonism between individuals
354 M. Neumann
and society. This assumption can be discerned in his distinction between Ego, Id
and Super-Ego. The Id represents the drives of the child-like portion of the person.
It is highly impulsive and takes into account only what it wants. It exclusively
follows the pleasure principle (Freud 1932). The Super-Ego enables control of the
primary drives: it represents the moral code of a society and involves feelings of
shame and guilt (Freud 1955). It is the place where social norms can be found. It has
been argued that the degree to which feelings of guilt are experienced is indicative
of the degree of norm internalisation (Kohlberg 1996). Finally, the Ego is the
controlling instance: it coordinates the demands of Id, Super-Ego and outer
world. According to Freud, the Super-Ego is the mirror of the society. Freud’s
theory of Ego, Super-Ego and Id, then, parallels Durkheim’s assumption that the
internalisation of norms involves social coercion (Geulen 1991). From the perspec-
tive of both, society is in radical conflict with human nature. Norms are given as an
external fact. Both Durkheim and Freud regard the individual as passive and
internalisation as a unidirectional process.
Building on G.H. Mead (1934), and the theories of cognitive and moral devel-
opment of Piaget (1932; 1947) and Kohlberg (1996), in recent times identity
theories have become influential in socialisation research. In contrast to an orienta-
tion based solely either on the individual subject or the society, identity theories
emphasise the interaction of culture and individuals in the development of a
morally responsible person (Bosma and Kunner 2001; Fuhrer and Trautner 2005;
Keupp 1999).
Mead developed a concept of identity that in contrast to Durkheim did not
reduce the individual to a private subject guided purely by drives. Instead, he
developed a theory of the social constitution of individual identity. A crucial
mechanism in the development of personality is the capability of role taking: to
regard oneself from the perspective of the other. This ability enables individuals to
anticipate the perspectives and expectations of others and thereby to come to accept
social norms. In the process of role taking, the individual develops a consciousness
whereby the individual is itself a stimulus for the reaction of the other in situations
of social interaction. This is the distinction between the spontaneous ‘I’ and self-
reflected ‘me’. Together they form what Mead denoted as identity: in other words,
the ‘self’. An abstraction of this process leads to the notion of the ‘generalised
other’. This is not a specific interaction partner but a placeholder for anybody. The
notion of the ‘generalised other’ is the representation of society.
Identity theories follow Mead in seeing individual identity as the key link
between person and culture. In contrast to the perspective to regard the social as
constraining the individual, identity theories argue that socially embedded identity
enables action selection. Action determination can be intrinsically or extrinsically
motivated. The identity of individuals contributes to the development of their
intrinsic motivation. There exist clear empirical evidence that sanctions and even
incentives undermine intrinsic motivation (Deci and Ryan 2000). Norms, however,
constitute a socially determined pattern of behaviour. Thus norm obedience is
always extrinsically motivated. However, at this point internalisation comes into
play. Extrinsic motivation can be internalised to different degrees, ranging from
purely extrinsic behavioural regulation (e.g. sanctions) to motivations that are

integrated into the self. Integration is attained when external guidelines have
become part of personal identity. This is the highest degree of a transformation of
external regulation into self-determination and is denoted as self-determined extrin-
sic motivation (Deci and Ryan 2000). In this case, a person is in line with itself if he
or she orients behaviour around social norms. Integrated behaviour regulation is
highly salient. Since norms are in full accordance with the personal values, action is
regarded as autonomously motivated. Hence, the scale of internalisation from
external regulation to integration is regarded as the scale from external control to
autonomy. Internalisation of extrinsic motivation represents the bridge between
psychological integrity and social cohesion. While intrinsic motivation is the
paradigm of autonomous motivation, in the case of social behaviour regulation
autonomy can only be reached through a process of internalisation.
How is the concept of identity described? With reference to William James
(1890), criteria for identity are formulated as consistency, continuity and effective-
ness. Identity consists of an inner and outer perspective. Moreover personal and
social identities are differentiated (Tajfel 1970; Turner 1982; Turner and Onorato
1999). While the inner perspective is grounded on individual decisions, the outer
perspective is based on ascription of others. Examples are ethnic or gender identity.
However, the individual might decide to identify with these ascriptions. Then the
ascription becomes part of the inner perspective. Examples can be found throughout
the history. For instance, (beside other factors) elements of this psychological
mechanism can be revealed in the black power movement in the 1960s, or the
raise of ethnic conflicts since the 1990s. Personal identity is the self-construction of
a personal biography. Social identity is determined by peer and reference groups.
This refers to social networks. While peers are the group to which the individual
factually belongs, the individual need not belong to the reference group. It is
sufficient to identify with the values of this group. For instance, this identification
might constitute sympathy for a political party. The social identity is decisively
responsible for the process by which social norms and values become part of
individual goals. This is particularly dependent on the salience of group member-
ship. Norm internalisation, however, is not a unidirectional process of the transmis-
sion of a given norm. While embedded in a social environment, the individual has
an active role in the social group.
14.3.2 Normative Architectures
The brief overview of socialisation research suggests that for the design of norma-
tive agents in particular two main decisions have to be made:
Is an antagonism or an identity (respective harmony) between individual and society
presumed? Hence, does the Artificial Society represent the theories of Durkheim and
Freud, or identity theories that follow G. H. Mead?
356 M. Neumann
How is the effect of normative behaviour regulation on the individual agent

represented? Does the individual agent play an active or a passive role, i.e. has the
individual agent something comparable to a personal identity?
The second question leads to a follow-up question, namely:
If agents play an active role, if and how can this represent a process of identity formation?
In particular, are agents embedded in social networks of peer or reference groups?
It will now be examined how current architectures can be assessed from an

empirical perspective. As the examination of current simulation models has
revealed, a comprehension of the effects of the social level on individual agents,
is far from being sufficient so far. To provide an outlook of possible future
modelling approaches of the effects of social norms on individual agents, a brief
sample of normative agent architectures will be provided. In fact, the number of
conceptually oriented articles on the architecture of normative agents exceeds the
number of existing models. These architectures study how these processes could be
modelled in principle. Typically, norms in concrete models are less sophisticated
than concepts proposed in formal architectures Conte and Dignum (2001). The
development of architectures is a kind of requirement analysis: it specifies the
essential components of normative agents. It can be expected that future
implementations will be guided by deliberations that can be found in these
architectures. For this reason a sample of cases is selected (Neumann 2008a) for
a closer examination with regard to the question of what decision are made how to
represent effects of norms on individual agents.
Andrighetto et al. (2007) investigate the process of norm innovation. The behaviour of an
agent may be interpreted by an observing agent as normative if it is marked as salient in the
observer’s normative board. Thus, norm instantiation is regarded as an inter-agent process.
Boella and van der Torre (2003) differentiate between three types of agents: agents who
are the subject of norms, so-called defender agents, who are responsible for norm control,
and a normative authority that has legislative power and that monitors defender agents.
Boella and van der Torre (2006) rely on John Searle’s notion of institutional facts (so-
called ‘counts-as’ conditionals) to represent social reality in the agent architecture. A
normative base and a ‘counts-as’ component transforms brute facts into obligations and
permissions.
The Belief-Obligation-Intentions-Desire (BOID) Architecture (Broersen et al. 2001) is
the classical approach to represent norms in agent architectures. Obligations are added to
the BDI Architecture to represent social norms while preserving the agent’s autonomy.
Principles of the resolution of conflicts between the different components are investigated
in the paper.
Boman (1999) proposes the use of super-soft decision theory to characterise real-time
decision-making in the presence of risk and uncertainty. Moreover, agents can communi-
cate with a normative decision module to act in accordance with social demands. Norms act
as global constraints on individual behaviour.
Castelfranchi et al. (2000) explore the principles of deliberative normative reasoning.
Agents are able to receive information about norms and society. The data is processed in a
multi-level cognitive architecture. On this basis, norms can be adopted and used as meta-
goals in the agent decision process.
Conte and Castelfranchi (1999) distinguish between a conventionalist (in rational

philosophy) and a prescriptive (in philosophy of law) perspective on norms. A logical
framework is introduced to preserve a weak intuition of the prescriptive perspective which
is capable of integrating the conventionalist intuition.
Conte and Dignum (2001) argue that imitation is not sufficient to establish a cognitive
representation of norms in an agent. Agents infer abstract standards from observed
behaviour. This allows for normative reasoning and normative influence in accepting (or
defeating) and defending norms.
Dignum et al. (2002) investigate the relations and possible conflicts between different
components in an agent’s decision process. The decision-making process of so-called ‘B-
doing agents’ is designed as a two-stage process, including norms as desires of society. The
authors differentiate between abstract norms and concrete obligations.
Garcia-Camino et al. (2006) introduce norms as constraints for regulating the rules of
interaction between agents in situations such as a Dutch auction protocol. These norms are
regulated by an electronic institution (a virtual auctioneer) with an explicitly represented
normative layer.
Lopez and Marquez (2004) explore the process of adopting or rejecting a normative
goal in the BDI framework. Agents must recognise themselves as addressees of norms and
must evaluate whether a normative goal has a higher or lower priority than those hindered
by punishment for violating the norm.
Sadri et al. (2006) extend their concept of knowledge, goals and plan (KGP) agents by
including norms based on the roles played by the agents. For this reason, the knowledge
base KB of agents is upgraded by KBsoc, which caters for normative reasoning, and KBrev,
which resolves conflicts between personal and social goals.
Shoham and Tennenholtz (1992) propose building social laws into the action represen-
tation to guarantee the successful coexistence of multiple programs (i.e. agents) and
programmers. Norms are constraints on individual freedom. The authors investigate the
problem of automatically deriving social laws that enable the execution of each agent’s
action plans in the agent system.
Vazquez-Salceda et al. (2005) provide a framework for the normative regulation of
electronic institutions. Norms are instantiated and controlled by a central institution, which
must consist of a means to detect norm violation and a means to sanction norm violators and
repair the system.
How can these examples be evaluated with regard to the design decision
suggested by socialisation research? The existing approaches can be regarded as a
hierarchy of increasingly sophisticated accounts, ranging from mere constraints to
abstract concepts. Broadly speaking three concepts of norms can be differentiated:
norms as constraints (the simplest choice), as obligations, or as abstract concepts
(the most sophisticated choice). This can the summarised in the following table:
Constraints Obligations Abstract concepts

Garcia-Camino et al. Sadri et al. Dignum et al.
Boman Broersen et al. Andrighetto et al.
Shoham & Tennenholtz Boella and van der Torre (2006) Conte & Dignum
Lopez & Marquez
Boella and van der Torre (2003)
Conte & Castelfranchi
Vazquez-Salceda et al.
Castelfranchi et al.
358 M. Neumann
(a) Constraints
The simplest and most straightforward way is to regard norms as mere
constraints on the behaviour of individual agents. For example, the norm to
drive on the right-hand side of the road restricts individual freedom. In this case,
norms need not necessarily be recognised as such. They can be implemented
off-line or can emerge in interaction processes. This may be sufficient for
practical purposes. However, it follows that it is not possible to distinguish
the norm from the normal. Hence, even though norms cannot be in contrast to
individual desires in this account, the agents have no concept of obligations.
They do not ‘know’ norms. Agents have a purely passive role. Since no
decisions are possible they remain merely normative automata.
(b) Obligations
More sophisticated accounts treat norms as mental objects (Castelfranchi et al.
2000, Conte and Castelfranchi 1995b). This allows for deliberation about norms
and, in particular, for the conscious violation of norms. Norms intervene in the
process of goal generation, which might – or might not – lead to the revision of
existing personal goals and the formation of normative goals. A number of
accounts (such as the BOID architecture) rely on the notion of obligations.
Obligations are explicit prescriptions that are always conditional to specific
circumstances. One example of an obligation is not being permitted to smoke in
restaurants. The rationale for including a separate obligation component next to
a component of individual desires is geared towards ensuring an agent’s
autonomy: by explicitly separating individual and social desires, it is possible
that the agent can deliberate over which component has priority. Conflicts may
arise between different components. Compared to the literature on
socialisation, a partial convergence with older theories can be observed. In
particular, it is striking that Freud’s architecture of the human psyche has some
parallels to BOID agents: the Id, guided by egoistic drives taking into account
only what it wants, can be found in the ‘desires’ component. Moreover, there is
an obvious temptation to identify Freud’s Super-Ego with the ‘obligations’
component. In fact, ‘obligations’ have been explicitly described as the desires
of a society (Dignum et al. 2002). This conception is well supported by Freud’s
theory. With regard to current identity theories and the theory of self-
determination, the situation is different: these theories emphasise that a full
internalisation of norms is only realised when they have become part of one’s
identity. Thus, internalised norms form part of the person’s own goals.
According to identity theories, agents of this kind of architecture have not yet
fully internalised norms. Norms, implemented in an ‘obligations’ component
do not represent complete external regulation, but by the same token are
not part of the agent’s own desires. In fact, the dichotomy between obligations
and desires becomes only effective once conflicts between both components
arise. This is explicitly wanted: the ‘obligations’ component is added to the
architecture to enable norm compliance as well as violation. It is claimed that
this process preserves the agent’s autonomy. Hence, the dichotomy of obliga-
tions and desires refers to an antagonism between an individual and society.
To represent the process of norm internalisation as described by modern

theories, a dynamic relation between the components ‘obligations’ and ‘desires’
would be required: contingent on the salience of a norm, elements of the
‘obligations’ component should be imported to the ‘desires’ component.
(c) Abstract Concepts
Agents may face several obligations that may contradict one another. For this
reason, some authors differentiate between norms and obligations. Norms are
regarded as more stable and abstract concepts than mere obligations (Dignum
et al. 2002; Conte and Dignum 2001). One example of such an abstract norm is
‘being altruistic’: further inference processes are needed for the formation of
concrete action goals from this abstract norm. The striking feature of this
approach is to allow for normative reasoning. This calls for an active role of
the agent. This conception of norms is a precondition for a modelling approach
of social behaviour regulation based on identity conceptions.
In particular the cognitive capacity of role taking constitutes a crucially
important step in the development of goals from abstract concepts: that is, the
ability to regard oneself from another’s perspective. Interestingly, steps in this
direction can be found in the AI literature. In Boella and van der Torre’s
architecture of a ‘norm governed system’ (2003), the agent’s decision-making
process is governed by the belief that they are observed by other agents and by
the belief that the other agents have expectancies with regard to how they ought
to behave. This can be regarded as a first step in simulating identity theory.
However, from the perspective of socio-psychological identity theories, it is a
shortcoming of this architecture that the agents regard themselves only in terms
of the question concerning whether they fulfil their – externally given – social
role. Identity consists of an inner and outer perspective. The inner perspective is
dependent on one’s personal decisions. This is not the case in this architecture,
which consists solely of an outer perspective. It can be questioned if and how an
inner perspective can be modelled: among other things, the development of an
inner perspective is correlated to a social identity. This social identity, however,
is correlated to peer groups and reference groups. Hence, it refers to social
networks, which can be simulated. In principle, a propensity to take over group
norms could be simulated, dependent on the salience of group membership. To
model such cognitive development the agents thus need to be embedded in
micro-social structures.
In conclusion, the concepts of norms as obligations and as an abstract
concept are more closely related to concepts in empirical sciences than a
mere constraint, which might be perfectly sufficient for practical purposes. It
has to be noted, however, that they refer to different theories: the obligations
concept of norms presumes an antagonism between the individual and the
society, which is in line with Durkheim and Freud. The idea of norms as an
abstract concept demands for a more active role of the agents. This is a
precondition for modelling identity. There exist very first attempts that can be
regarded as a modelling approach towards identity formation. Yet, it has to be
emphasised that these are very first steps, and much is still not realised, such as,
360 M. Neumann
for instance, to implement a correlation between network structures and

salience of normative orientation. However, one principle deficiency of current
models and architectures in attempting to represent the process of norm
internalisation remains; namely, that agents do not have a childhood (Guerin
2008). However, socialisation theory describes childhood as the most important
site for the internalisation of norms. Since agents have no childhood, the
process of human cognitive development cannot be represented.
14.4 Conclusion
In conclusion, it can be retained that the interaction processes, resulting in macro-

structural constraints are quite well understood. In particular, the perspective to
regard norms as an aggregated product of individual interactions is considerably
elaborated. This is the view of sociological Rational Choice theories. In particular,
the game theoretic paradigm has proved to be an effective means to study the
dynamics of collective behaviour regularities. However, it lacks of an active
element of normative orientation in the choice of the ends of action. The agents
do not ‘know’ norms. Thus, these models do not capture the process of norm
internalisation. Behaviour is merely guided by adaptation of agents to changing
environmental conditions.
The role theoretic tradition emphasises that norms are structural constraints of
individual behaviour. While models of cognitive agents in the AI tradition also have
reached a substantial insights into norm dynamics, this aspect has been particularly
studied these models. They have provided considerable insights into the effects of
such structural constraints on a social macro-level. Hence, the inter-agent processes
of interaction, leading to a macro-property of some kind of normatively structured
social macro level, are relatively good understood. There is, however, still a lot to
do with regard to achieving a comprehensive understanding of how actors produce,
and are at the same time a product of social reality. While agent-based modelling
has reached a substantial understanding of inter-agent processes, an investigation
of the recursive impact on intra-agent processes is still in its fledgling stages.
This becomes apparent when considering socialisation theories of norm
internalisation. An investigation of the effects of social behaviour regulation on
individual agents is mostly at the level of conceptual considerations in the devel-
opment of architectures. Here the development of new models has to be aware of
the decisions to be made. The philosophical orientation that implicitly underlies the
BOID architecture is inspired by the classical accounts of Durkheim and Freud: by
opposing obligations and desires, an antagonism between individuals and society is
assumed. However, it has to be (and – implicitly – is) decided whether an antago-
nism individual and society is assumed or not. This is the question, whether the
social macro-level is perceived as action constraint or as enabling action selection.
Empirical research suggests that a social embedding in networks of peer- and
reference groups has a substantial impact on normative reasoning and thereby on
the process of action selection (i.e. the agents’ desires).
A comprehension of the two-way dynamics of the operations and effects of

social behaviour regulation on a psychological as well as on the social level calls for
interdisciplinary research. Agent-based modelling is an appropriate methodological
tool for this effort. However, it has to be emphasised, that developmental processes
in the socialisation process are only barely captured by current simulations.
Further Reading
Even though they are quite old and some of their findings are out of date by now, it
is still a good start (and not too much effort) to study the following two models to
become familiar with the research field of normative agent-based models: Axelrod’s
(1986) evolutionary approach to norms, and Conte and Castelfranchi’s (1995b) paper
on understanding the functions of norms in social groups (using simulation).
As an introduction into the design and logical foundations of normative
architectures the following anthologies are suggested: (Boella et al. 2005; Boella
et al. 2007).
The relation of modelling and theory is particularly highlighted in the two
anthologies (Conte and Dellarocas 2001; Lindemann et al. 2004). Here the reader
will also find hints for further readings about the empirical and theoretical
background.
For an overview of the theoretical background and developments in theorising
norms it is suggested to refer to (Conte and Castelfranchi 1995a; Therborn 2002).
References
Andrighetto G, Campennı̀ M, Conte R, Paolucci M (2007) On the immergence of norms:

a normative agent architecture. In: Trajkovski GP, Collins SG (eds) Emergent agents and
socialities: social and organizational aspects of intelligence; proceedings of the 2007 fall
AAAI symposium. AAAI Press, Menlo Park, pp 11–18
Axelrod R (1986) An evolutionary approach to norms. Am Polit Sci Rev 80:1095–1111
Bandura A (1962) Social learning through imitation. In: Jones M (ed) Nebraska symposium on
motivation. University of Nebraska Press, Lincoln
Bandura A (1969) Principles of behaviour modification. Holt, New York
Bicchieri C, Duffy J, Tolle G (2003) Trust among strangers. Philos Sci 71:286–319
Binmore K (1998) Review of the book “The complexity of cooperation: agent-based models of
competition and collaboration” by Axelrod R, Princeton, Princeton University Press. J Artif
Soc Soc Simulat 1(1), http://jasss.soc.surrey.ac.uk/1/1/review1.html
Boella G, van der Torre L (2003) Norm Governed multiagent systems: the delegation of control to
autonomous agents. In proceedings of the IEEE/WIC international conference on intelligent
agent technology 2003 (IAT 2003). IEEE Press, pp 329–335
Boella G, van der Torre L (2006) An architecture of a normative system: counts-as conditionals,
obligations, and permissions. In: AAMAS ’06: Proceedings of the fifth international joint
conference on autonomous agents and multiagent systems. ACM Press, New York, pp 229–231
362 M. Neumann
Boella G, van der Torre L, Verhagen H (eds) Proceedings of the symposium on Normative Multi-
Agent Systems (NORMAS 2005), AISB’05 convention: social intelligence and interaction in
animals, robots and agents, 12–15 Apr 2005, University of Hertfordshire, Hatfield. Hatfield:
society for the Study of Artificial Intelligence and the Simulation of Behaviour (SSAISB)
Boella G, van der Torre L, Verhagen H (2007) Introduction to normative multiagent systems. In:
Boella G, van der Torre L, Verhagen H (eds) Normative multi-agent systems, Dagstuhl
seminar proceedings 07122, vol I. IBFI, Schloss Dagstuhl, pp 19–25
Boman M (1999) Norms in artificial decision making. Artif Intell Law 7:17–35
Bosma H, Kunner E (2001) Determinants and mechanisms in ego development: a review and
synthesis. Dev Rev 21(1):39–66
Boudon R (1981) The logic of social action. Routledge, London
Bratman M (1987) Intentions, plans and practical reasoning. CSLI Publications, Stanford
Broersen J, Dastani M, Huang Z, van der Torre L (2001) The BOID architecture: conflicts between
beliefs, obligations, intentions, and desires. In: André E, Sen S, Frasson C, Müller JP (eds)
Proceedings of the 5th international conference on autonomous agents. ACM Press, New York,
pp 9–16
Burke M, Fournier G, Prasad K (2006) The emergence of local norms in networks. Complexity
11:65–83
Castelfranchi C, Conte R, Paolucci M (1998) Normative reputation and the costs of compliance.
J Artif Soc Soc Simulat 1(3), http://www.soc.surrey.ac.uk/JASSS/1/3/3.html
Castelfranchi C, Dignum F,Treur J (2000) Deliberative normative agents: principles and architec-
ture. In: Jennings NR, Lesperance Y (eds) Intelligent agents VI, agent theories, architectures,
and languages (ATAL), 6th international workshop, ATAL ’99, Orlando, July 15–17, 1999,
proceedings (LNCS, 1757). Springer, Berlin, pp 364–378
Claessens D (1972) Familie und wertsystem: eine studie zur zweiten sozio-kulturellen geburt des
menschen. Duncker & Humblot, Berlin
Coleman J (1987) The emergence of norms in varying social structures. Angew Sozialforschung
14:17–30
Coleman J (1990) Foundations of social theory. Harvard, Belknap
Conte R, Castelfranchi C (1995a) Cognitive and social action. UCL Press, London
Conte R, Castelfranchi C (1995b) Understanding the functions of norms in social groups through
simulation. In: Gilbert N, Conte R (eds) Artificial societies: the computer simulation of social
life. UCL Press, London, pp 252–267
Conte R, Castelfranchi C (1999) From conventions to prescriptions: towards an integrated view of
norms. Artif Intell Law 7:323–340
Conte R, Castelfranchi C (2001) Are incentives good enough to achieve (info) social order? In:
Conte R, Dellarocas C (eds) Social order in multiagent systems. Kluwer, Dordrecht
Conte R, Dellarocas C (eds) (2001) Social order in multiagent systems. Kluwer, Dordrecht
Conte R, Dignum F (2001) From social monitoring to normative influence. J Artif Soc Soc Simulat
4(2), http://www.soc.surrey.ac.uk/JASSS/4/2/7.html
Dahrendorf R (1956) Homo sociologicus. Ein Versuch zu Geschichte, Bedeutung und Kritik der
Kategorie der sozialen Rolle. Westdeutscher Verlag, Opladen
Davies K, Moore W (1945) Some principles of stratification. Am Sociol Rev 10:242–249
Deci E, Ryan R (2000) The “what” and “why” of goal pursuits: human needs and the self-
determination of behavior. Psychol Inq 11(4):227–268
Deffuant G, Moss S, Jager W (2006) Dialogues concerning a (Possibly) new science. J Artif Soc
Soc Simulat 9(1), http://jasss.soc.surrey.ac.uk/9/1/1.html
Dignum F, Kinny D, Sonenberg L (2002) From desires, obligations and norms to goals. Cognitive
science quarterly, 2, http://www.cs.uu.nl/people/dignum/papers/CSQ.pdf
Durkheim E ([1895] 1970) Regeln der Soziologischen Methode. Neuwied: Luchterhand
Durkheim E ([1897] 2006) Der Selbstmord. Frankfurt a.M.: Suhrkamp
Durkheim E ([1907] 1972) Erziehung und Soziologie. Düsseldorf: Schwann
Epstein J (2000) Learning to be thoughtless: social norms and individual computation (working
paper, no. 6). Center on social and economic dynamics, http://www.santafe.edu/media/
workingpapers/00-03-022.pdf
Esser H (1993) Soziologie: allgemeine grundlagen. Campus, Frankfurt a. M

Fioretti G (2013) Utility, games, and narratives. Chapter 13 in this volume
Flentge F, Polani D,Uthmann T (2001) Modelling the emergence of possession norms using
memes. J Artif Soc Soc Simulat 4(4), http://jasss.soc.surrey.ac.uk/4/4/3.html
Freud S (1932) Neue Vorlesungen zur Einführung in die Psychoanalyse. In: Freud S, Gesammelte
Werke, vol. 15. London: Imago
Freud S (1955) Abriss der psychoanalyse. Fischer, Frankfurt
Fuhrer U, Trautner N (2005) Entwicklung und identität. In: Asendorpf J (ed) Enzyklopädie der
pschologie – entwicklungspsychologie, vol 3. Göttingen, Hofgrebe
Garcia-Camino A, Rodriguez-Aguilar JA, Sierra C, Vasconcelos W (2006) Norm-oriented pro-
gramming of electronic institutions: a rule-based approach. In: Nakashima H, Wellman M,
Weiss G, Stone P, Proceedings of the 5th international joint conference on autonomous agents
and multi-agent systems 2006, Hakodate, 08–12 May 2006. ACM Press, New York, pp
670–672
Geulen D (1991) Die historische entwicklung sozialisationstheoretischer ansätze. In: Hurrelmann
K, Ulich D (eds) Neues handbuch der sozialisationsforschung. Beltz, Weinheim
Giardini F, Conte R, Paolucci M (2013) Reputation. Chapter 15 in this volume
Gibbs JP (1981) Norms, deviance and social control: conceptual matters. Elsevier, New York
Guerin F (2008) Constructivism in AI: prospects, progress and challenges. In proceedings of the
AISB convention 2008, vol 12: computing and philosophy, Aberdeen, Scotland, 1–4 Apr 2008.
London: The society for the study of artificial intelligence and simulation of behaviour,
pp 20–27
Hales D (2002) Group reputation supports beneficent norms. J Artif Soc Soc Simulat 5(4), http://
Homans G (1964) Bringing man back in. Am Sociol Rev 29:809–818
James W (1890) The principles of psychology. Holt, New York
Keupp H (1999) Identitätskontruktionen. Rowohlt, Hamburg
Kohlberg L (1996) Die psychologie der moralentwicklung. Suhrkamp, Frankfurt
Krappmann L (2006) Sozialisationsforschung im spannungsfeld zwischen gesellschaftlicher
reproduktion und entstehender handlungsfähigkeit. In: Schneider W, Wilkening F (eds)
Enzyklopädie der psychologie – entwicklungspsychologie, vol 1. Göttingen, Hofgrebe
Lindemann G, Moldt D, Paolucci M (eds) (2004) Regulated agent-based social systems: first
international workshop, RASTA 2002, vol 2934, Lecture notes in artificial intelligence.
Springer, Berlin
Lopez F, Marquez A (2004) An architecture for autonomous normative agents. In: proceedings of
the 5th Mexican international conference in computer science, ENC’04. IEEE computer
society, Los Alamitos, pp 96–103
Macy M, Sato Y (2002) Trust, cooperation, and market formation in the U.S. and Japan. PNAS, 99,
pp 7214–7220
Macy M, Willer R (2002) From factors to actors: computational sociology and agent-based
modelling. Am Rev Sociol 28:143–166
Malinowski B (1927) Sex and repression in primitive society. Humanities Press, New York
Mead M (1928) Coming of age in Samoa. William Morrow & Co., New York
Mead GH ([1934] 1968) Geist, Identität und Gesellschaft. Suhrkamp, Franfurt
Merton R (1957) Social theory and social structure, 2nd edn. The Free Press, Glencoe
Moss S (2001) Game theory: limitations and an alternative. J Artif Soc Soc Simulat 4(2), http://
www.soc.surrey.ac.uk/JASSS/4/2/2.html
Neumann M (2008a) A classification of normative architectures. In: Deffuant G, Cioffi-Revilla C
(eds) Proceedings of the second world congress of social simulation (WCSS’08), Fairfax,
14–17 July 2008
Neumann M (2008b) Homo socionicus: a case study of normative agents. J Artif Soc Soc Simulat
Parsons T ([1937] 1968) The structure of social action: a study in social theory with special
reference to a group of recent European writers. New York: Free Press
364 M. Neumann
Pawlow IP (1927) Conditioned reflexes: an investigation of the physiological activity of the

cerebral cortex. Oxford University Press, London
Piaget J ([1932] 1983) Das moralische Urteil beim Kinde. Stuttgart: Klett-Cotta
Piaget J ([1947] 1955) Psychologie der Intelligenz. Rascher, Zürich
Rao AM, Georgeff M (1991) Modeling rational agents within a BDI architecture. In: Allen JF,
Fikes R, Sandewall E (eds) Proceedings of the 2nd international conference on principles of
knowledge representation and reasoning (KR’91), Cambridge, 22–25 Apr 1991. Morgan
Kaufmann, San Francisco, pp 473–484
Raub W, Voss T (1981) Individuelles handeln und gesellschaftliche folgen: das individualistische
programm in den sozialwissenschaften. Luchterhand, Darmstadt
Saam N, Harrer A (1999) Simulating norms, social inequality, and functional change in artificial
societies. J Artif Soc Soc Simulat 2(1), http://jasss.soc.surrey.ac.uk/2/1/2.html
Sadri F, Stathis K, Toni F (2006) Normative KGP agents. Comput Math Organ Theor 12:101–126
Savarimuthu B, Purvis M, Cranefield S, Purvis MA (2007) How do norms emerge in multi-agent
societies? Mechanism design (The information science discussion paper series, 2007/01).
University of Otago, Dunedin
Sen S, Airiau S (2007) Emergence of norms through social learning. In: Veloso M (ed)
Proceedings of the twentieth joint conference on artificial intelligence (IJCAI-07), Hyderabad,
6–12 Jan 2007. pp 1507–1512
Shoham Y, Tennenholtz M (1992) On the synthesis of useful social laws for artificial agent
societies (preliminary report). In: Swartout WR (ed) Proceedings of the 10th National confer-
ence on artificial intelligence (AAAI’92), 12–16 July 1992 AAAI Press/The MIT Press, San
Jose, pp 276–281
Skyrms B (1996) Evolution of the social contract. Cambridge University Press, Cambridge
Skyrms B (2004) The stage hunt and the evolution of social structure. Cambridge University Press,
Cambridge
Tajfel H (1970) Experiments in intergroup discrimination. Sci Am 223(5):96–102
Therborn G (2002) Back to norms! on the scope and dynamics of normative action. Curr Sociol 50
(6):863–880
Turner JC (1982) Towards a cognitive redefinition of the social group. In: Tajfel H (ed) Social
identity and intergroup relations. Cambridge University Press, Cambridge
Turner JC, Onorato RS (1999) Social identity, personality, and the self-concept: a self categorising
perspective. In: Tylor T et al (eds) The psychology of the social group. Erlbaum, Mahwah
Vazquez-Salceda J, Aldewereld H, Dignum F (2005) Norms in multiagent systems: from theory to
practice. Int J Comput Syst Engin 20:225–236
Verhagen H (2001) Simulation of the learning of norms. Soc Sci Comput Rev 19:296–306
Vieth M (2003) Die evolution von fairnessnormen im ultimatumspiel: eine spieltheoretische
modellierung. Z Soziol 32:346–367
Wrong D (1961) The oversocialised conception of man. Am Sociol Rev 26:183–193
Chapter 15
Reputation
Francesca Giardini, Rosaria Conte, and Mario Paolucci
Why Read This Chapter? To understand the different conceptions underlying

reputation in simulations up to the current time and to get to know some of the
approaches to implementing reputation mechanisms, which are more cognitively
sophisticated.
Abstract In this chapter, the role of reputation as a distributed instrument for social
order is addressed. A short review of the state of the art will show the role of
reputation in promoting (a) social control in cooperative contexts – like social groups
and subgroups – and (b) partner selection in competitive contexts, like (e-) markets
and industrial districts. In the initial section, current mechanisms of reputation – be
they applied to electronic markets or MAS – will be shown to have poor theoretical
backgrounds, missing almost completely the cognitive and social properties of the
phenomenon under study. In the rest of the chapter a social cognitive model of
reputation developed in the last decade by some of the authors will be presented.
Its simulation-based applications to the theoretical study of norm-abiding behaviour,
partner selection and to the refinement and improvement of current reputation
mechanisms will be discussed. Final remarks and ideas for future research will
conclude the chapter.
15.1 Reputation in Social Systems: A General Introduction
Ever since hominid settlements started to grow, human societies needed to cope with the
problem of social order (Axelrod, 1984): how to avoid fraud and cheating in wider,
unfamiliar groups? How to choose trustworthy partners when the likelihood of re-encounter
is low? How to isolate cheaters and establish worthwhile alliances with the good guys?
F. Giardini (*) • R. Conte • M. Paolucci

Laboratory for Agent-Based Social Simulation (LABSS), Institute of Cognitive Sciences and
Technologies, National Research Council, Via Palestro, 32 – 00185, Rome

366 F. Giardini et al.
Social knowledge like reputation and its transmission, i.e. gossip, plays a
fundamental role in social order, adding at the same time cohesiveness to social
groups and allowing for distributed social control and sanctioning (plus a number of
other functionalities, see Boehm 1999). Reputation is a property that unwilling and
unaware individuals derive from the generation, transmission and manipulation of a
special type of social beliefs, namely social evaluations, and that contributes to
regulate natural societies from the morning of mankind (Dunbar 1996). People use
reputational information to make decisions about possible interactions, to evaluate
candidate partners, to understand and predict their behaviours, and so on.
That reputation is a fundamental generator, vehicle and manipulator of social
knowledge for enforcing reciprocity and other social norms is known since long
(see a review in Conte and Paolucci 2002). In particular, in the study of cooperation
and social dilemmas, the role of reputation as a partner selection mechanism started
to be appreciated in the early 1980s (Kreps and Wilson 1982). However, little
understanding of its dynamic and cognitive underpinnings was achieved at that
stage. Despite its critical role in the enforcement of altruism, cooperation and social
exchange, the social cognitive study of reputation is relatively new. Hence, it has
not yet been fully clarified how this critical type of knowledge is manipulated in the
minds of agents, how social structures and infrastructures generate, transmit and
transform it, and consequently how it affects agents’ behaviour.
The aim of this chapter is to guide the reader through the multiplicity of computa-
tional approaches concerned with the impact of reputation and its dynamics. Reputa-
tion is a complex social phenomenon that cannot be treated as a static attribute of
agenthood, with no regard for the underlying process of transmission. We claim that
reputation is both the process and the effect of transmitting information, and that
further specifications about the process and its mechanisms are needed. Following
these premises, we will first review some applications of reputation in computational
simulation, highlighting problems and open questions, and then we will propose a
theoretical social cognitive model of reputation. Moreover, we will present three
different ways of applying the cognitive theory of reputation to model social phenom-
ena: the Sim-Norm model, the SOCRATE framework and the REPAGE architecture.
This brief introduction will be followed by an outline of reputation research in
different domains (social psychology, management and experimental economics,
agent-based simulation), in order to show how many different viewpoints can be
used to describe and explore this complex phenomenon. We will then focus on
some of the results in electronic markets and multi-agent simulations. Electronic
markets are a typical example of a complex environment where centralized control
is not possible and decentralized solutions are far from being effective. In recent
years, the Internet contributed to a growth of auction sites facilitating the exchange
of goods between individual consumers, without guaranteeing transparency and
safety of transactions. On the other hand, multi-agent applications are concerned
with the problem of assessing the reliability of single agents and of social networks.
In Sect. 15.2 we will propose a cognitive model of reputation, which aims to solve
15 Reputation 367
some of the problems left open by existing systems, moving from a theoretical
analysis of cognitive underpinnings of reputation formation and spreading. This
model will be tested in the following section, where a description of three different
implementations will be provided. Finally, we will draw some conclusions about
future work and directions.
15.1.1 State of the Art
According to Frith and Frith (2006), there are three ways to learn about other people:
through direct experience, through observation and through “cultural information”.
When the first two modalities are not available, reputational information becomes
essential in order to obtain some knowledge about one’s potential partner(s) in an
interaction, and thus to predict their behaviour. Reputation allows people to predict,
at least partially or approximately, what kind of social interaction they can expect and
how that interaction may possibly evolve. Reputation is therefore a coordination
device whose predictive power is essential in social interactions (Paolucci and Conte
2009). Furthermore, reputation has a strategic value and can be used to pursuit self-
interest (Paine 1967; Noon and Delbridge 1993).
Reputation and its transmission (gossip) has an extraordinary preventive power:
it substitutes personal experience in (a) identifying cheaters and isolating them, and
in (b) easily finding trustful partners. It makes available all the benefits of
evaluating someone, without implying the costs of direct interaction.
Furthermore, in human societies gossip facilitates the formation of groups
(Gluckman 1963): gossipers share and transmit relevant social information about
group members within the group (Barkow 1996), at the same time isolating out-
groups. Besides, gossip contributes to stratification and social control, since it works
as a tool for sanctioning deviant behaviours and for promoting, even through
learning, those behaviours that are functional with respect to the group’s goals and
objectives. Reputation is also considered as a means for sustaining and promoting the
diffusion of norms and norm conformity (Wilson et al. 2000).
The theories of indirect reciprocity and costly signals show how cooperation
in large groups can emerge when the agents are endowed with, or can build, a
reputation (Nowak and Sigmund 1998a; 1998b; Gintis et al. 2001). As Alexander
(1987) pointed out, “indirect reciprocity involves reputation and status, and results
in everyone in the group continually being assessed and reassessed”. According to
this theory large-scale human cooperation can be explained by individuals helping
others in order to uphold a reputation and thus be included in future cooperation
(Panchanathan and Boyd 2004).
Despite important advances in the study of reputation as a means to support
cooperation (Sommerfeld et al. 2008), no explicit theory of the cognitive
ingredients and processes which reputation is made of was provided. More recently,
reputation and gossip has started to become crucial in other fields of the social
sciences like management and organisation science, governance, or business ethics,
where the importance of branding became apparent. The economic interest in the
subject matter implied an extension of reputation to the super-individual level:
Corporate reputation is considered as an external and intangible asset tied to the
history of a firm and coming from stakeholders’ and consumers’ perceptions. Rose
and Thomsen (2004) claim that good reputation and financial performance are
mutually dependent, hence a good reputation may influence the financial asset of
a firm and vice versa. Several researchers have tried to create a corporate reputation
index containing the most relevant dimensions to take into account when dealing
with corporate reputation. Cravens et al. (2003) interviewed 650 CEO in order to
create a reliable index, but their index has so many entries, ranging from global
strategy to employees’ attributes, that it is not easy to foresee how such a tool could
be used. Gray and Balmer (1998) distinguish between corporate image and corpo-
rate reputation. Corporate image is the mental picture consumers hold about a firm,
therefore is similar to an individual perception, whereas the reputation results from
the firm’s communication and long-term strategy. Generally speaking, corporate
reputation is treated as an aggregate evaluation stakeholders, consumers, managers,
employees, and institutions form about a firm, but the mechanisms leading to the
final result are still only vaguely defined.
Over the last 10 years several studies in experimental economics have investigated
reputational dynamics through standard experimental settings, such as trust games,
public good games, and moral hazard problems (see Fehr and Gachter 2000, for an
introduction). The aim of these studies is to explain mechanisms underlying reciproc-
ity, altruism and cooperation in humans, in order to answer the puzzling question:
“Why do humans cooperate?” Axelrod (1984). According to Nowak and Sigmund
(1998a, b), reputation sustains the emergence of indirect reciprocity, which gets
people to cooperate in order to receive cooperation even from strangers. Following
this hypothesis, holding a good reputation in your social group makes it more
probable that someone else would help you when you will need help.
If social order is a constant of human evolution, it becomes particularly crucial in
the e-society where the boundaries of interaction are extensively widened. The
portentous development pace of ICT technologies dramatically enlarges the range
of interaction among users, generating new types of aggregation, from civic
communities to electronic markets, from professional networking to e-citizenship,
etc. What is the effect of this widening of social boundaries? Communication and
interaction technologies modify the range, structures and modalities of interaction,
with consequences that are only partially explored, often only to resume the stereo-
type of technological unfriendliness (see the negative impact of computer terminals,
as opposed to face-to-face interaction, on subjects’ cooperativeness in experimental
studies of collective and social dilemmas, Sell and Wilson 1991; Rocco and
Warglien 1995). A detailed approach to the effects of technological infrastructures
on interaction styles and modes has never been adopted. Perhaps, an exception to
15 Reputation 369
this is represented by the research on the effects of asymmetry of information on the

markets. Known to encourage fraud and low-quality production, this phenomenon
appears as an intrinsic feature of e-markets, but in fact it goes back to eleventh
century Maghribi traders moving along the coast of the Mediterranean sea
(Greif 1993). As exemplified by Akerlof (1970), asymmetry of information drives
honest traders and high quality goods out of the market. The result is a market where
only “lemons”, or fraudulent commodities, are available – often to the detriment of
both sellers and buyers. The classical example of a lemon market is the used car
market, where only sellers have information about problems with the cars they are
selling, and most consumers are incapable of discerning these problems. Contempo-
rary online traders such as users of Internet auction sites face the same problem:
online buyers can learn about the quality (or condition) of the good only once they
have already paid for it.
Auction sites may be very generic concerning the products being offered and
operate on a global scale (e.g., e-Bay), or may focus on specific products on a
national scale (many car auction sites). At the moment virtually all consumer
products are being auctioned on the Internet, ranging from used toys and CDs to
cars and houses Utz et al. (2009). Compared to buying through online retailers,
buying through auction sites is even less controllable, as the sellers are not visible
and have not made major investments. Consumers who purchase through auction
sites must rely on the accuracy and reliability of the seller (Melnik and Alm 2002).
Sellers on the Internet may actively try to communicate their reputation to potential
buyers, inflating common expectations on the effects of reputation. Melnik and Alm
(2002) investigated whether an e-seller’s reputation matters. Their results indicated
that reputation had a positive – albeit relatively small – impact on the price levels
consumers were willing to pay. Moreover, Yamagishi et al. (2009) show that
reputation has significant positive effects on the quality of products.
Despite the role of reputation in economic transactions, online reputation
reporting systems are only moderately efficient (Resnick and Zeckhauser 2002;
Bolton et al. 2002). eBay, one of the best (if not the best) known examples, is also
one of the world’s largest online marketplaces with a community of over 50 million
users registered. On eBay, most items are sold through English-type auctions, and
the reputation mechanism used is based on the ratings that users perform after the
completion of transactions. The user can give three possible values: positive (1),
negative (1) or neutral (0). The “reputation” value is the sum of the last 6 months’
ratings, weighted by the relevance of the transaction.
In all of these models, the notion of reputation is weak and essentially reduced to
centralized image: no direct exchange of information takes place among participants
but only reports to a central authority, which calculates the final reputation value.
The actual utility of this mechanism is debatable. For example, when forums are
available, real reputation exchanges are performed in parallel, ignoring the centrally
calculated reputation rating. Moreover, people are likely to not provide reputational
feedback (under-provision) and if they do, they lean on providing only good reports
(over-scoring). In sum, the reputation system does not seem to perform efficiently.
However, from an economic point of view, eBay prospers. How is this possible?
What would happen if reputation worked more efficiently? Which type of reputation
system should be used for which objective? As current online reputation systems are
not theory-driven instruments, based upon an understanding of what reputation is,
how it is generated and distributed, and how it works, all of these questions are
still open.
However, the effects of reputation are at least partially known, generally restricted
to partner selection. Indeed, agent-based social simulation has taught us some
lessons: (1) what matters about reputation is its transmission (Castelfranchi et al.
1998), since by this means agents acquire zero-cost relevant information; (2) reputa-
tion has more impact than image: if agents transmitted only their own evaluations
about one another (image), the circulation of social knowledge would stop soon
(Pinyol et al. 2008). To exchange information about reputation, agents need to
participate in circulating reputation whether they believe it or not (gossip) and, to
preserve their autonomy, they must decide how, when and about whom to gossip.
What is missing in the study of reputation is the merging of these separate directions
in an interdisciplinary integrated approach, which accounts for both its social
cognitive mechanisms and structures.
15.1.2 Simulating Reputation: Current Systems
So far, the simulation-based study of reputation has been undertaken for the sake of
social theory, namely in the account of pro-social behaviour – be it cooperative,
altruistic, or norm-abiding – among autonomous, i.e. self-interested agents.
In this chapter, we will concentrate instead on two specific functionalities of
reputation:
1. To promote norm-abiding behaviour in cooperative settings (e.g. social groups).
2. To favour partner selection in electronic markets and agentized environments.
The second aspect has been dealt with in the literature to some extent and the
remainder of this section will give an overview of existing approaches; for a
treatment of the first aspect we refer the reader to Sect. 15.2, where we mainly
discuss our own models and studies of reputation.
Several attempts have been made to model and use reputation in artificial
societies, especially in two sub-fields of information technologies: computerized
interaction (with a special reference to electronic marketplaces) and agent-mediated
interaction. It is worth emphasizing that in these domains trust and reputation are
actually treated as the same phenomenon, and often the fundamentals of reputation
mechanisms are derived from trust algorithms. Moreover, several authors (Moukas
et al. 1999; Zacharia 1999; Zacharia et al. 1999) explain reputation in terms of trust
and vice versa, continuously mixing up these two phenomena. We will review some
15 Reputation 371
of the main contributions in online reputation reporting systems and in multi-agent

systems, in order to achieve a better understanding of the complex issue of
implementing and effectively using reputation in artificial societies.
15.1.2.1 Online Reputation Reporting Systems
The continuously growing volume of transactions on the World Wide Web and the
growing number of frauds this appears to entail1 led scholars from different
disciplines to develop new online reputation reporting systems. These systems are
meant to provide a reliable way to deal with reputation scores or feedbacks, allowing
agents to find cooperative partners and avoid cheaters.
The existing systems can be roughly divided into two sub-sets: Agent-oriented
individual approaches and agent-oriented social approaches, depending on how
agents acquire reputational information about other agents.
The agent-oriented individual approach has been dominated by Marsh’s ideas on
trust (Marsh 1992, 1994a, b), on which many further developments and algorithms
are based. This kind of approach is characterized by two attributes: (1) any agent
may seek potential cooperation partners, and (2) the agent only relies on its
experiences from earlier transactions. When a potential partner proposes a transac-
tion, the recipient calculates the “situational reputation” by weighing the reputation
of his potential trading partner with further factors, such as potential output and the
importance of the transaction. If the resulting value is higher than a certain “cooper-
ation threshold”, the transaction takes place and the agent updates the reputation
value according to the outcomes of the transaction. If the threshold is not reached,
the agent rejects the transaction offer, which may be punished by a “reputation
decline”. These individual-based models (Bachmann 1998; Marsh 1994a; Ripperger
1998) differ with regard to the length of memory span they apply. Agents may forget
their experiences slowly, fast, or never.
In agent-oriented social approaches agents not only rely on their direct experi-
ence, but are also allowed to consider third-party information (Abdul-Rahman and
Hailes 1997a; Rasmusson 1996; Rasmusson and Janson 1996; Schillo 1999; Yu and
Singh 2000). Although these approaches share the same basic idea, i.e. experiences
of other agents in the network can be used when searching for the right transaction
partner, they rely upon different solutions when it comes to weigh the third-party
information and to deal with “friends of friends”. Thus the question arises: how to
react to information from agents who do not seem to be very trustworthy?
Another problem lies in the storage and distribution of information. To form a
complete picture of its potential trading partners, each agent needs both direct
(its own) and indirect (third-party) evaluations in order to be able to estimate the
validity and the informational content of such a picture.
1
The US-based Internet Crime Complaint Center (IC3) received 231,493 complaints for the year
2005, 62.7 % of which were related to electronic auctioning (IC3 2005).
Regan and Cohen (2005) propose a system for computing indirect and direct
reputation in a computer mediated market. Buyers rely on reputation information
about sellers when choosing from whom to buy a product. If they do not have direct
experience from previous transactions with a particular seller they take indirect
reputation into account by asking other buyers for their evaluations of the potential
sellers. The received information is then combined to mitigate effects of deception.
The objective of this system is to propose a mechanism which reduces the “unde-
sirable practices” on actual reputation in online applications, especially on the part
of sellers, and to prevent the market from turning into a “lemons market” where
only low quality goods are listed for sale.
One serious problem with this and similar models concerns the reputation
transmission. Agents only react to reputation requests, while proactive, spontane-
ous delivery of reputation information to selected recipients is not considered.
On the other hand, despite its simplicity, this type of model tackles the problem
of collusion between rating agents by keeping secret the evaluation of sellers
amongst buyers, i.e. not disclosing it to the sellers.
As to electronic marketplaces, classic systems like eBay show a characteristic
bias to positive evaluations (Resnick and Zeckhauser 2002), suggesting that factual
cooperation among users at the information level may lead to a “courtesy” equilib-
rium (Conte and Paolucci 2003). As Cabral and Hortaçsu (2006) formally prove,
initial negative feedbacks trigger a decline in sale price that drives the targeted
sellers out of the market. Good sellers, however, can gain from ‘buying a reputa-
tion’ by building up a record of favourable feedback through purchases rather than
sales. Thus those who suffer a bad reputation stay out – at least until they decide to
change identity – while those who stay in can take advantage of a good reputation:
after a good start, they will hardly receive negative feedback and even if they do,
negative feedbacks will not get to the point of spoiling their good name. Under such
conditions, even good sellers may have an incentive to sell lemons.
Intuitively, the courtesy equilibrium reduces the deterrent effect of reputation. If
a reputation system is meant to impede frauds and improve the quality of products,
it needs to be constructed in such a way as to avoid the emergence of a courtesy
equilibrium. It is not by chance that among the possible remedies to ameliorate
eBay, Dellarocas (2003) suggested a short-memory system, erasing all feedbacks
but the very last one. While this might work for eBay, we believe that such remedies
based on fragmented models and tailored to a particular application are not the way
forward. Instead, a general theory of how reputation and its transmission work
needs to be developed. On top of such a theory, different systems for different
objectives can then be constructed. We will pursue this further in Section 15.3.
15.1.2.2 MAS Applications
Models of trust and reputation for multi agent systems applications (e.g. Yu and Singh
2000; Carbo et al. 2002; Sabater and Sierra 2002; Schillo et al. 2000; Huynh et al.
2004; for exhaustive reviews see Ramchurn et al. 2004a; Sabater and Sierra 2005)
15 Reputation 373
present interesting new ideas and advances over conventional online reputation
systems, with their notion of centralized global reputation.
Yu and Singh (2000) proposed an agent-oriented model for social reputation and
trust management which focuses on electronic societies and MAS. Their model
introduces a gossip mechanism for informing neighbours of defective transaction
partners, in which the gossip is transferred incrementally through the network of
agents. It also provides a mechanism to include other agents’ testimonies in an
agent’s reputation calculation. Agents store information about the outcome of every
transaction they had and recall this information in case they are planning to bargain
with the same agent again (direct evaluation). If the agent meets an agent it has not
traded with before, the reputation mechanism comes into play. In this mechanism,
so-called referral chains are generated that can make third-party information avail-
able across several intermediate stations. An agent is thus able to gain reputation
information with the help of other agents in the network. Since a referral chain
represents only a small part of the whole network, the information delivered will
most likely be partial instead of global as in centralized systems like eBay.
In the context of several extensive experiments, Yu and Singh showed that the
implementation of their mechanism results in a stable system, in which the reputa-
tion of cheaters decreases rapidly while the cooperating agents experienced a slow,
almost linear increase in reputation. Still, some problems remain. The model does
not combine direct and indirect reputation, i.e. if an agent has already traded with
another agent he has to rely on his own experience and cannot use the network
information anymore. Thus it might take unnecessarily long to react to a suddenly
defecting agent that cooperated before. In addition, Singh and Yu do not give an
explanation of how their agent-centred storage of social knowledge (for example
the referral chains) is supposed to be organized. Consequently, no analysis of
network load and storage capacity can be done.
As this example shows, the “agentized environment” is likely to produce
interesting solutions that may apply also to online communities. This is so for
two main reasons. First, in this environment two problems of order arise: to meet
the users’ expectations (external efficiency), and to control agents’ performance
(internal efficiency). Internal efficiency is instrumental to the external one, but it
re-proposes the problem of social control at the level of the agent system. In order to
promote the former, agents must control, evaluate, and act upon each other.
Reliability of agents is a proxy for reliability of users. Secondly, and consequently,
the agent system plays a double role: it is both a tool and a simulator. In it one can
perceive the consequences of given premises, which may be transferred to the level
of user interactions. In a sense, implemented agent systems for agent-mediated
interaction represent both parallel and nested sub-communities.
As a consequence, solutions applied to the problems encountered in this envi-
ronment are validated more strictly, against both external and internal criteria.
Their effects are observable at the level of the virtual community, with a procedure
essentially equivalent to agent-based simulation and with the related advantages.
Moreover, solutions may not (only) be implemented between agents, but (also)
within agents, which greatly expands the space for modelling. So far, however,
these potentials have not been fully exploited. Models have mainly been aimed at
ameliorating existing tools implemented for computerized markets. We suggest that
agent systems can do much more than this: they can be applied to answer the
question as to (a) what type of agent, (b) what type of beliefs, and (c) what type of
processes among agents are required to achieve useful social control. More specifi-
cally, what type of agents and processes are needed for which desirable result:
better efficiency, encouraging equity and hence users’ trust, discouraging either
positive or negative discrimination (or both), foster collaboration at the information
level or at the object level (or both), etc.
15.1.3 Concluding Remarks
This review has given an overview of how reputation have been discussed and
modelled in studies regarding online markets and multi-agent systems.
In the Internet-based models, the notion of reputation is weak and essentially
reduced to centralized image: participants do not exchange information directly but
only report their evaluations to a central authority, which calculates the final
reputation value. The actual utility of this mechanism is debatable.
The solutions proposed for MAS systems are interesting but so far insufficient to
meet the problems left open by online systems. There is a tendency to consider
reputation as an external attribute of agents without taking into account the process
of creation and transmission of that reputation.
We argue that a more theory-driven approach is needed, based upon a conceptual
analysis of the differences and analogies among notions concerning social evalua-
tion. In the next section we will therefore introduce our social cognitive approach
towards reputation.
15.2 An Alternative Approach: The Social Cognitive Theory

of Reputation
In this section we will present a social cognitive model of reputation, we will define
the difference between image and reputation, introduce the roles different agents
play when evaluating someone and transmitting this evaluation and, finally, we will
explain the decision processes underlying reputation.
Let us first clarify the term “social cognitive”. A cognitive process involves
symbolic mental representations (such as goals and beliefs) and is effectuated by
means of the mental operations that agents perform upon these representations
(reasoning, decision-making, etc.). A social cognitive process is a process that
involves social beliefs and goals, and that is effectuated by means of the operations
that agents perform upon social beliefs and goals (e.g. social reasoning). A belief or a
15 Reputation 375
goal is social when it mentions another agent and possibly one or more of his or her
mental states (for an in-depth discussion of these notions, see Conte and Castelfranchi
1995; Conte 1999).
Social cognitive processes are receiving growing attention within several
subfields of the Sciences of the Artificial, in particular intelligent software agents,
multi-agent systems, and artificial societies. Unlike the theory of mind (cf. Leslie
1992), which focuses upon one, albeit important aspect of social agency, namely
social beliefs, this approach aims at modelling and possibly implementing systems
acting in a social – be it natural or artificial – environment. Thus, it is aimed at
modelling the variety of mental states (including social goals, motivations,
obligations) and operations (such as social reasoning and decision-making) neces-
sary for an intelligent social system to act in some domain and influence other
agents (social learning, influence, and control).
15.2.1 Image and Reputation
The social cognitive model presented here is a dynamic approach that considers
reputation as the output of a social process of transmission of information. The input
to this process is the evaluation that agents directly form about a given agent during
interaction or observation. We will call this evaluation the social image of the agent.
An agent’s reputation is argued to be distinct from, although strictly interrelated
with, its image. Image consists of a set of evaluative beliefs (Miceli and
Castelfranchi 2000) about the characteristics of the target, i.e. it is an assessment
of her positive or negative qualities with regard to a norm, a competence, etc.
Reputation is both the process and the effect of transmission of a target image.
The image relevant for social reputation may concern a subset of the target’s
characteristics, e.g. its willingness to comply with socially accepted norms and
customs. More precisely, we define reputation to consist of three distinct but
interrelated objects: (1) a cognitive representation, i.e. a believed evaluation; (2) a
population object, i.e. a propagating believed evaluation; (3) an objective emergent
property at the agent level, i.e. what the agent is believed to be.
Reputation is a highly dynamic phenomenon in two distinct senses: it is subject
to change, especially as an effect of corruption, errors, deception, etc.; and it
emerges as an effect of a multi-level bidirectional process (Conte and Paolucci
2002). In particular, it proceeds from the level of individual cognition to the level of
social propagation (population level) and from there back to individual cognition.
What is even more interesting, once it reaches the population level it gives rise to an
additional property at the agent level. From the very moment an agent is targeted by
the community, their life will change whether they want or believe it or not.
Reputation has become the immaterial, more powerful equivalent of a scarlet letter
sewn to one’s clothes. It is more powerful because it may not even be perceived by
the individual to whom it is attached, and therefore it is not in the individual’s
power to control and manipulate. Reputation is an objective social property that
emerges from a propagating cognitive representation. This lack of an identified

source, i.e. its impersonality, is the distinctive feature of reputation, whereas image
always requires identifying the individual who made the evaluation.
To formalise these concepts we will begin by defining the building block of
image. An agent has made an evaluation when he or she believes that a given entity
is good for, or can achieve a given goal. An agent has made a social evaluation
when his or her belief concerns another agent as a means for achieving this goal.
A given social evaluation includes three sets of agents:
1. A nonempty set E of agents who share the evaluation (evaluators)
2. A nonempty set T of evaluation targets
3. A nonempty set B of beneficiaries, i.e., the agents sharing the goal with regard to
which the elements of T are evaluated.
Often, the sets of evaluators and beneficiaries overlap but this is not necessarily
the case. A given agent t is a target of a social evaluation when t is believed to be a
good/bad means for a given goal of B, which may or may not include the evaluator.
Evaluations may concern physical, mental, and social properties of targets; agents
may evaluate a target with regard to both capacity and willingness to achieve a
shared goal. The latter, willingness to achieve a goal or interest, is particular to
social evaluations. Formally, e (with e2E) may evaluate t (t2T) with regard to a
state of the world that is in b’s (b2B) interest, but of which b may not be aware.
The interest/goal with regard to which t is evaluated may be a distributed or
collective advantage. It is an advantage for the individual members of B, or it may
favour a supra-individual entity, that results from the interactions among the
members of B (for example, if B’s members form a team).
To make this analysis more concrete, we will start with an example drawn from
the Sim-Norm model, which will be described in more detail in Section 15.3.1. Let
us consider a classical multi-agent situation, a set of agents fighting for access to a
scarce resource (food). Assume that a norm of precedence (a proscription against
attacking agents who are consuming their “own” resources) is applied to reduce
conflicts. The norm is disadvantageous for the norm follower in the short run, but is
advantageous for the community, and thus eventually for the individual followers.
We will call N the set of norm followers, or normative agents, and C the set of
cheaters, or violators of the norm. With regard to social evaluations (image), the
targets coincide with the set of all agents; T ¼ N[ C (all are evaluated). The agents
carrying out the evaluation are restricted to the norm followers: E ¼ N, because
only the normative find purpose in evaluating. Finally, B ¼ N[ C: indeed, if
normative agents benefit globally from the presence of the norm, cheaters in this
simple setting benefit even more by exploiting the norm; they can attack the weaker
while they themselves are safe from attacks by the gullible normative.
It is very easy to find examples where all three sets coincide. General behaviour
norms, such as “Do not commit murder” apply to, benefit, and are evaluated by the
whole universe of agents. There are situations in which beneficiaries, targets, and
evaluators are separate, for example, when norms safeguard the interests of a subset
of the population. Consider the quality of TV programs for children broadcast in the
15 Reputation 377
afternoon. Here, we can identify three more or less distinct sets. Children are the
beneficiaries, while adults entrusted with taking care of children are the evaluators.
It could be argued that B and E still overlap, since E may be said to adopt B’s
interests. The targets of evaluation are the writers of programs and the decision-
makers at the broadcast stations. There may be a non-empty intersection between E
and T but no full overlap. In case the target of evaluation is the broadcaster itself, a
supra-individual entity, the intersection can be considered to be empty: E \ T ¼ Ø.
Extending this formalisation to include reputation, we have to differentiate
further. To assume that a target t is assigned a given reputation implies assuming
that t is believed to be “good” or “bad,” but it does not imply sharing either
evaluation. Reputation therefore involves four sets of agents:
1. A nonempty set E of agents who share the evaluation;
2. A nonempty set T of evaluation targets;
3. A nonempty set B of beneficiaries, i.e. the agents sharing the goal with regard to
which the elements of T are evaluated;
4. A nonempty set M of agents who share the meta-belief that members of E share
the evaluation; this is the set of all agents aware of the effect of reputation (as
stated above, effect is only one component of it; awareness of the process is not
implied).
Often, E can be taken as a subset of M; the evaluators are aware of the effect of
evaluation. In most situations, the intersection between the two sets is at least
nonempty, but exceptions exist. M in substance is the set of reputation transmitters,
or third parties. Third parties share a meta-belief about a given target, whether they
share the underlying belief or not.
15.2.2 Reputational Roles
Agents may play more than one role simultaneously: evaluator, beneficiary, target,
and third party. In the following, we will examine the characteristics of the four
roles in more detail.
15.2.2.1 Evaluator
Any autonomous agent is a potential evaluator Conte et al. 1998. Social agents are
likely to form evaluative beliefs about one another (see Castelfranchi 1998) as an
effect of interaction and social perception. These may be positive or negative,
depending on an agent’s experiences. When agents evaluate one another with
regard to their individual goals they obtain social evaluations. Image, the result of
such social evaluations, serves to identify friends and partners and to avoid
enemies. Who should the agent resort to for help? Who should he or she cooperate
with? And who should be avoided due to being dangerous or ill willed?
Furthermore, agents may not only evaluate one another with regard to their
own goals, but with regard to the goals or interests of a given set of agents (the
beneficiaries), to which the evaluators may belong. A negative evaluation may be
formed about agents violating others’ rights or behaving in an apparently malevolent
and hostile manner, whether or not the evaluators consider themselves potential
victims of such actions. Information thus obtained may be used to infer that the
target could violate other rights in the future, namely, those of the evaluator. In
addition, evaluators may be concerned with one another’s power to achieve the goals
or interests of abstract social entities or institutions, as when we judge others’
attitudes towards norms, the church, the government or a political party.
To sum up, agents evaluate one another with regard to their own goals and the
goals they adopt from either other individual agents (e.g. their children) or supra-
individual agents, such as groups, organisations, or abstract social entities.
15.2.2.2 Beneficiary
A beneficiary is the entity that benefits from the action with regard to which targets
are evaluated. Beneficiaries can either be individual agents, groups and organisations
or even abstract social entities like social values and institutions. Beneficiaries may
be aware of their goals and interests, and of the evaluations, but this is not necessarily
the case. In principle, their goals might simply be adopted by the evaluators – as it
happens, for example, when members of the majority support norms protecting
minorities. Evaluators often are a subset of the beneficiaries.
Beneficiaries may be implicit in the evaluation. This is particularly the case
when it refers to a social value (honesty, altruism, etc.); the benefit itself and those
who take advantage of it are left implicit, and may coincide with the whole society.
The beneficiary of the behaviour under evaluation is also a beneficiary of this
evaluation: the more an (accurate) evaluation spreads, the likelier the execution
of the positively evaluated behaviour.
15.2.2.3 Target
The target of any social evaluation is the evaluated entity. While targets may even
be objects or artefacts to be used by others when the evaluation pertains to image,
for reputation mental and moral components are necessarily involved. Holders of
reputation (targets) are endowed with the following important characteristics:
• Agency, in particular autonomous agency and sociality (the target is evaluated
with regard to a given behaviour)
• Mental states, specifically willingness to perform the respective behaviour
• Decision making or deliberative capacity, i.e. the ability to choose a desirable
behaviour from a set of options
• Social responsibility, i.e. the power to prevent social harm and possibly to
respond for it, in case any damage occurs.
15 Reputation 379
Intuitively, targets are complementary to beneficiaries, but this is not necessarily

the case. Targets may be evaluated for their willingness to exhibit a behaviour from
which they are expected to benefit themselves.
Other than beneficiaries, targets are always explicit. They may be individual
entities or supra-individual like a group, a collective, an abstract entity, or a social
artefact, such as an institution, provided this can be attributed the capacity to make
decisions, achieve goals, and perform actions. A further distinguishing characteris-
tic of targets is that they may inherit the reputation of the set they belong to.
Offspring may inherit their parents’ reputation; employees may suffer from the
bad reputation of the firm they work for; members may inherit the reputation of
their social groups, whether they have had a chance to decide to join those groups or
not. In the latter case, the attitude under evaluation is considered to be a built-in
tendency shared by the targets.
Two dynamic processes characterise reputation. The first concerns the propaga-
tion of a social cognitive representation from one agent to another. We call this
transmission of reputation (or gossip). As we have seen so far, it is intrinsic to
reputation spreading. The second process concerns the extension of a given agent’s
reputation to other agents who are related to the former by affiliation, proximity,
similarity, etc. We will call it contagion of reputation (or inheritance). This
dynamic is not intrinsic to reputation, although it may empirically co-occur with
reputation spreading. It corresponds to what is usually called prejudice.
The transmission process spreads reputation as a cognitive representation,
whereas the contagious process spreads reputation as a property. Both may proceed
vertically and horizontally, within social groups and hierarchies and from one agent
to his or her neighbours in social space. In this sense, reputation may be perceived
as a highly fertile phenomenon, whose propagation is favoured by social member-
ship and proximity.
15.2.2.4 Third Party or Gossiper
An agent is a (potential) third party if she transmits (is in position to transmit)

reputation information about a target to another agent or set of agents. Although
sharing awareness of a given target reputation, third parties do not necessarily share
the corresponding image (social evaluation) of the target. That is, they do not
necessarily believe it to be true. Actually, an agent could share a given evaluation
with others without being aware of others’ existence. This would put him in a poor
position to transmit reputation. If there is no belief about E, it is very difficult to
justify (that is, to assign a goal to) the act of reputation transmission. Instead, the
meta-belief about the sharing of the evaluation is a powerful tool and paves the way
to complex (and potentially deceiving) new behaviours.
Third parties (if they are also targets) may deserve a negative evaluation; they
may actually deceive their fellow third parties, the beneficiaries, the targets, or the
society as a whole, by conveying information that they hold to be false. A third
party may be bluffing: he or she may pretend to be benevolent with regard to
beneficiaries, in order to (1) enjoy the advantages of sharing reputation information,

(2) be considered as part of the in-group by other evaluators, and therefore (3) gain a
good reputation without sustaining the costs of its acquisition (as would be implied
by performing the socially desirable behaviour), and (4) avoid the consequences of
a bad reputation. Agents may also spread a false reputation, i.e., pretend that a target
has a given reputation when this is not the case. Agents do this in order to achieve
the aforementioned benefits without taking responsibility for spreading a given
social evaluation.
Agents may have incomplete and inaccurate information – a situation that can
result in ill-reputed agents being taken as well-reputed and vice versa. One reason
for incomplete or inaccurate information is the lack of personal experience and
familiarity with the target. In such a case, an agent might have received information
from other third parties or from agents who have had a direct contact with the target.
To demonstrate how the social categories so far identified may help in the analysis
of reputation spreading we will return to the example of the quality of TV programs
for children. Transmission of evaluations involves two distinct sets of agents,
children (beneficiaries) and adults with the children’s welfare in mind (evaluators),
while the targets – operators and decision-makers at broadcast stations – are a subset
of the evaluators. The set of third parties comprises the whole universe of adults,
thereby including evaluators and targets but not the beneficiaries. The targets (broad-
caster) offer an interesting example of the interplay between institutional and per-
sonal roles. While no assumption of benevolence towards B can be made on the part
of the institution (broadcast stations) itself, whose purpose is profit, things are more
intriguing in the case of their employees. These belong to E and are thus benevolent
with regard to B, but are also members of T and thus committed to their employer.
15.2.3 Reputation-Based Decisions
After identifying the different roles agents can take with regard to image, reputation
and its transmission let us now examine the relevant decision-making processes. To
understand the difference between image and reputation, we will categorise the
mental decisions based upon them into three levels:
1. The epistemic level is concerned with decisions about accepting or rejecting an
evaluative belief.
2. The pragmatic–strategic level concerns decisions about interacting with another
agent (target).
3. The memetic level concerns decisions about transmitting evaluative beliefs about
a given target to others
15 Reputation 381
15.2.3.1 Epistemic Level
When an agent decides whether or not to accept a particular belief that forms either
a given image or acknowledges a given reputation he makes an epistemic decision.
This involves direct evaluation of the belief in question. To acknowledge a given
reputation implies two more specific beliefs: (a) that a nonempty set E of agents
within the same population shares the given image about the target (reputation
effect), and/or (b) that information about the target’s image has been circulated
(reputation process) among those agents.
Image and reputation about one and the same agent do not necessarily overlap.
From an agent X’s meta-belief that agent Y is said to be e.g. a womaniser we are not
entitled to derive X’s belief that Y is in fact a womaniser (although X may draw such
a more or less arbitrary conclusion). Image and reputation about the same Y may be
consistent (for example, X believes that Y is and is believed to be a womaniser) or
inconsistent (X believes that Y either suffers from an undeserved bad reputation or
enjoys an unworthy good one).
To accept one does not imply acceptance of the other. The two processes of
acceptance are different not only in their respective outputs (evaluations and meta-
beliefs), but also in the operations involved. To accept a given image implies
coming to share it. The acceptance may be based, for example, upon supporting
evidence and first-hand experience with the image target, consistent pre-existing
evaluations (concerning, for example, the class of objects to which the target
belongs), or trust in the source of the given evaluative belief.
Acknowledging a given reputation, on the other hand, does not necessarily lead
to sharing others’ evaluations but rather to believing that these evaluations are held
or at least circulated by others. To assess the value of such a meta-belief is a rather
straightforward operation. For the recipient to be relatively confident about this
meta-belief, it is probably sufficient for him or her to hear some rumours.
15.2.3.2 Pragmatic-Strategic Level
Agents resort to their evaluative beliefs in order to achieve their goals (Miceli and
Castelfranchi 2000). In general, evaluations are guidelines for planning; therefore
social evaluations (evaluations about other agents) are guidelines for social action
and social planning. The image an agent X has about a target agent Y will guide his
or her action with regard to Y, will suggest whether it is convenient to interact with
Y or not, and also will suggest what type of interaction to establish with Y. Image
may also be conveyed to others in order to guide their actions towards the target in a
positive or negative sense. To do so, the agent must (pretend to) be committed to
their evaluation and take responsibility for its truth value.
While image dominates direct pragmatic-strategic decisions reputation may be
used when it is consistent with image or when no image of the target has been
formed. To influence others’ social decisions, however, agents tend to transmit

information about the target’s reputation rather than their image of the target. Two
main reasons explain this inverse pattern:
Agents expect that a general opinion is more credible and acceptable than an
individual one.
Agents reporting on reputation do not need to commit to its truth value, and do not
have to take responsibility for it; consequently, they may influence others at a
lower personal cost.
15.2.3.3 Memetic Level
A memetic decision can be roughly described as the decision to spread reputation.

Consequently, communication about reputation is communication about a meta-
belief, i.e. about others’ mental attitudes. To spread news about someone’s reputa-
tion does not bind the speaker to commit himself to the truth value of the evaluation
conveyed but only to the existence of rumours about it. In other words, communi-
cation about reputation does neither imply any personal commitment of the speaker
with regard to the content of the information delivered – if an agent X reports on Y’s
bad reputation he is by no means stating that Y deserved it – nor any responsibility
with regard to the credibility of (the source of) information (“I was told that Y is a
bad guy”). This is due to the source of the meta-belief being implicit (“I was
told. . .”) and the set of agents to whom the belief is attributed being undefined
(“Y is ill/well reputed”).
Communication about reputation is not always sincere. On the contrary, the lack
of commitment and responsibility can (and often does) lead to deceptive reputation
transmission. To deceive other agents about Y’s reputation agent X only needs to
report it as a rumour independent of or even opposite to his own beliefs.
Several factors may affect an agent’s decision to transmit reputation information
(see Conte and Paolucci 2002):
• Certainty and acceptance of the evaluation,
• Reputation of the source from which information was received,
• Responsibility and accountability for the effects of distributing this evaluation to
others,
• Benevolence toward the beneficiary as opposed to benevolence toward the
target, or no benevolence at all.
Social responsibility can be defined as the power attributed to intelligent autono-
mous social systems to predict and prevent harm to others and/or themselves.
A memetic agent may be said to have harmed another agent by transmitting bad
reputation information about her. Responsibility comes from accountability, i.e. the
power and obligation to respond and repair harm that one has been found account-
able for. A memetic agent may be asked to take responsibility and possibly repair the
15 Reputation 383
harm. However, responsibility is much less crucial in a memetic than in a pragmatic

decision concerning reputation. The decision to transmit reputation will thus depend
on the extent to which the memetic agent is aware of potentially serious effects of
reputation transmission on the target and his perceived contribution to these effects.
The latter is influenced by the perceived “distance” from the target. It is easier
to convey potentially dangerous evaluations when the target is far away, absent,
unknown, etc., not only because the memetic agent is less likely to perceive the
effective occurrence of harm but also because he or she will not be a direct cause of
it. Harm will actually depend on a number of intermediate events, others’ decisions,
etc. The greater the distance and the smaller the memetic agent’s (perceived)
responsibility, the less cautious the memetic agent is likely to be.
Benevolence means whether and to what extent a candidate third party adopts
the goals or interests of either the target or the beneficiary of reputation transmis-
sion. When circulating reputation, memetic agents may follow different strategies,
according to the direction of their benevolence.
In case of benevolence towards the set of beneficiaries B, gossiping may follow
some prudence rule like “pass on negative evaluation even if uncertain, pass on
positive evaluation only if certain”. This may give way to circulation of a reputation
which is overly critical and worse than the real characteristics of the target.
When the benevolence is target-oriented (set T), it is possible to expect the
application of some courtesy rule like “pass on positive evaluation even if uncer-
tain, negative evaluation only if certain”. This may lead to a courtesy equilibrium
where no-one expresses critique anymore, especially fearing retaliation.
If memetic agents are not benevolent towards any of the two sets B and T, Conte
and Paolucci (2002) predicts scarcity of reputation transmission. In this case
production of information may be induced by institutional reinforcement of it,
e.g. a university requesting students to academically evaluate their lecturers.
Systematic application of a courtesy or prudence rule in reputation spreading
may result in selective transmission of either positive or negative evaluations. We
expect general adoption of such rules as a consequence of self-interested decisions
of single memetic agents. We may reasonably suppose that a lower responsibility of
memetic agents increments the quantity of circulating information, while benevo-
lence direction is a consequence – all other factors being equal – of the overlapping
of reputational roles of agents. Role overlapping means that the same people
(or people having the same norm-related goals) are involved in more than one
reputational role. Such agents then may consider it useful to be prudent or generous
in their information spreading, or they may have not enough motivation to circulate
information of any sort.
As we have seen in this and the previous section, operationalization of cognitive
properties of reputation and dynamics of reputational groups allows us to express
testable hypotheses, which can be investigated both by experimenting with human
subjects and by designing software reputation agents.
15.3 Simulating Reputation
Social simulation models have been successfully employed to investigate the

effects of reputation in different contexts. In this section we will present an
overview of three different ways of designing agent-based systems based on the
cognitive theory of reputation presented previously. These models can be distin-
guished according to:
• Different levels of cognitive complexity of the agents,
• The kind of setting (purely cooperative, competitive or both)
• The scenario.
The first, Sim-Norm, based on a very simple concept of reputation, has been
applied to observe the impact of reputation in social control. The second, REPAGE,
based on a more complex agent architecture and on the concepts of reputation and
image introduced in the previous section, has been applied to explore the impact of
social evaluation on the market. Finally, in the third model, called SOCRATE,
relatively simple agents interact in a complex market in which they exchange both
goods and information.
15.3.1 Sim-Norm
This model was developed to examine the effect of reputation on the efficiency of a
norm of precedence (Castelfranchi, Conte, Paolucci 1998; Conte and Paolucci
1999; Paolucci 2000) in reducing aggression, measured both at the global (i.e.
societal) and local (i.e. individual) level. In particular, Sim-Norm was designed to
explore why self-interested agents exercise social control. Albeit far from reaching
a final conclusion on this issue, the studies based on Sim-Norm confirmed a positive
impact of reputation on social control.
More precisely, while individually acquired evaluation of other agents gave
norm executors no significant advantage, the transmission of these evaluations
among norm executors proved decisive in levelling the outcomes of norm-abiders
and cheaters (if numerically balanced).
15.3.1.1 Hypotheses
Sim-Norm revolved around the question of which ingredients are necessary for
social order to be established in a society of agents. The role of norms as aggression
controllers in artificial populations living under conditions of resource scarcity was
addressed. We set out to explore two hypotheses:
Norm-based social order can be maintained and its costs reduced via distributed
social control.
15 Reputation 385
Social cognitive mechanisms are needed to account for distributed social control. In
particular, the propagation of social beliefs plays a decisive role in distributing
social control at low or zero individual costs and high global benefit.
15.3.1.2 Model Description and Experimental Conditions
The model defines agents as objects moving in a two-dimensional environment

(a 10 10 grid) with randomly scattered food. At the beginning of a run, agents
and food items are assigned locations at random. A location is a cell in the grid. The
same cell cannot contain more than one object at a time (except when an agent is
eating). The agents move through the grid in search of food, stopping to eat to build
up their strength when they find it. The agents can be attacked only when eating;
no other type of aggression is allowed.
At the beginning of each step, every agent selects an action from the six
available routines: eat, move-to-food-seen, move-to-food-smelled, attack, move-
random, and pause. Actions are supposed to be simultaneous and time consuming.
To investigate the role of norms in the control of aggression, we compared
scenarios in which agents follow a norm – implemented as a restriction on attacks –
with identical scenarios, in which they follow utilitarian rules. In all scenarios, each
agent can perform only one of three strategies:
Blind aggression, or control condition, in which aggression is not constrained. If the
agent can perform no better move (eating, moving to food seen or smelled), then
it will attack without further considerations. Blind agents have access to neither
their own strength nor the eater’s strength; these parameters never enter their
decision-making process.
Utilitarian, in which aggression is constrained by strategic reasoning. Agents will
only attack those eaters whose strength is lower than their own. An eater’s
strength is “visible” one step away from the agent’s current location. While
blind agents observe no rule at all, utilitarian agents observe a rule of personal
utility, which does not qualify as a norm.
Normative (N), in which aggression is constrained by a norm. We introduced a
finder-keeper precept, assigning a “moral right” to food items to finders, who
become possessors of the food. Possession of food is ascribed to an agent on the
grounds of spatial vicinity; food owned is flagged, and every player knows to
whom it belongs. Each food unit may have up to five owners, decided on the
basis of proximity at the time of creation. The norm then prescribes that agents
cannot attack other agents who are eating their own food.
The strategies can also be characterised by the kind of agents they allow to
attack: while blind agents attack anybody, the utilitarian agents attack only the
weaker, and the normative agents, respecting a norm of private property, will not
attack agents who are eating their own food.
These strategies were compared (Castelfranchi et al. 1998) using an efficiency

measure – the average strength of the population after n periods of simulation – and
a fairness measure, the individual deviation from the average strength.
15.3.1.3 Findings
The first two series of experiments showed that normative agents perform less well
than non-normative agents in mixed populations, as they alone bear the costs of
social control and are exploited by utilitarian agents. In a following series of
experiments, “image” was added to the preceding experimental picture. Now,
each normative agent collects information about the behaviour of other agents in
an image vector. This information is binary and discriminates between the
“respectful,” who will abide with the norm and “cheaters,” who will not respect
the principle of finders–keepers. The vector is initialised to “all respectful” (pre-
sumption of innocence), but every time a normative agent is attacked while eating
its own food, the attacker is recorded as a cheater. Moreover, the normative
algorithm is modified so that the agents respect the norm only when facing agents
known as respectful, while they behave with known cheaters according to one of the
retaliation strategies listed above.
The results from another set of experimental runs on a mixed population equally
composed of normative and utilitarian agents show that in this model useful
knowledge can be drawn from personal experience, but therefore still at one’s
own cost. To reduce cost differences among subpopulations, image is insufficient.
Henceforth, we provided the respectful agents with the capacity to exchange
with their (believed-to-be) respectful neighbours at distance one images of other
agents. With the implementation of a mechanism of transmission of information,
we can speak of a reputation system. We ran the experiments again with normative
agents exchanging information about cheaters. The results suggest that circulating
knowledge about others’ behaviours significantly improves normative agents’
outcomes in a mixed population.
The spreading of reputation can then be interpreted as a mechanism of cost
redistribution for the normative population. Communication allows compliant
agents to easily acquire preventive information, sparing them the costs of direct
confrontations with cheaters. By spreading the news that some guys cheat, the good
guys (1) protect themselves, (2) at the same time punish the cheaters, and possibly
(3) exercise an indirect influence on the bad guys to obey the norm. Social control is
therefore explained as an indirect effect of a “reciprocal altruism” of knowledge.
15.3.1.4 Concluding Remarks
The Sim-Norm model presented in this section was studied with the purpose of
clarifying the role of norms controlling aggression in simple multi-agent systems.
The model shows that a simple norm of precedence is not as efficient as a utilitarian
15 Reputation 387
rule when utilitarian agents are present; the normative agents must resort to retalia-
tion against cheaters. The addition of image is not enough to defend the norm, but
image coupled with a mechanism of information transmission is. The necessity of
information transmission points out the relevance of our distinction between image
and reputation.
The model inspired further research in the social simulation community: Saam
and Harrer (1999) used the same model to explore the interaction between norma-
tive control and power, whereas Hales (2002) applied an extended version of Sim-
Norm to investigate the effects of group reputation. In his model agents are given
the cognitive capacity to categorise other agents as members of a group and project
reputation onto whole groups instead of individual agents (a form of stereotyping).
15.3.2 The REPAGE Cognitive Architecture
REPAGE is a computational system for partner selection in a competitive setting

(marketplace), although in principle it can also be used in cooperative contexts
(organizations). Based on a model of REPutation, imAGE and their interplay,
REPAGE provides evaluations of potential partners and is fed with information
from others plus outcomes from direct experience. This is fundamental to account
for (and to design) limited autonomous agents as exchange partners. To select good
partners, agents need to form and update own social evaluations; hence, they must
exchange evaluations with one another.
But in order to preserve their autonomy, agents need to decide whether or not to
share others’ evaluations of a given target. If agents would automatically accept
reported evaluations and transmit them as their own, they would not be autonomous
anymore. In addition, in order to exchange information about reputation, agents
need to participate in circulating it, whether they believe it or not (gossip); but again
to preserve their autonomy, they must decide how, when and about whom to gossip.
In sum, the distinction between image and reputation suggests a way out from the
paradox of sociality, i.e. the trade-off between agents’ autonomy and their need to
adapt to social environment. On one hand, agents are autonomous if they select
partners based on their social evaluations (images). On the other, they need to update
evaluations by taking into account others’ evaluations. Hence, social evaluations
must circulate and be represented as “reported evaluations” (reputation), before and
in order for agents to decide whether to accept them or not. To represent this level of
cognitive detail in artificial agents’ design, there is a need for a specialised subsys-
tem. This is what REPAGE provides.
In the following we briefly describe the architecture of the REPAGE model and
how it is integrated with the other elements that compose a typical deliberative
agent (cf. Sabater et al. 2006 for a more detailed description). REPAGE is com-
posed of three main elements: a memory, a set of components called detectors and
the analyzer. An implementation of REPAGE in Java has been published as a
Sourceforge project.
Repage
Analyzer Repage memory Detectors
Planner Agent main memory Repage
Main
Planner
mem
Comm
module Comm
module
a b
Fig. 15.1 REPAGE and its environment. For the sake of clarity only agent components that
interact with REPAGE are depicted
15.3.2.1 Memory
In the implementation, to support the specialized nature of REPAGE, memory is

actually composed by a set of references to the predicates in the agent general-
purpose memory (see section A of Fig. 15.1). Only those predicates that are relevant
for dealing with image and reputation are considered. Therefore, of all the
predicates in the main memory, only a subset is also part of the REPAGE memory.
A change in a predicate is immediately visible in both memories.
To mirror their dependence connections, in the REPAGE memory predicates are
conceptually organized in different levels and inter-connected. Each predicate
reference is wrapped by a component that adds connection capabilities to it.
Predicates contain a fuzzy evaluation, consisting of three aspects: the type of the
evaluation (e.g. personal experience, image, third party image), the role of the target
(e.g. informant or seller) and the actual content. To store the content, a simple
number is used in e-Bay and in most reputation systems. This sharp representation,
however, is quite implausible in inter-agent communication, which is one of the
central aspects of REPAGE; in real life no-one is told that people are saying Jane is
0.234 good. To capture the lack of precision coming (a) from vague utterances, e.g.
“I believe that agent X is good, I mean, very good – good, that is”, and (b) from noise
in the communication or in the recollection from memory, we decided to model the
actual value of an evaluation with a tuple of positive real values that sum to one.
Aggregation of these evaluations was detailed in Sabater and Paolucci (2007).
Finally, each predicate has a strength value associated to it. This value is a
function of (1) the strength of its antecedents and of (2) some special characteristics
intrinsic to that type of predicate. The network of dependencies specifies which
15 Reputation 389
predicates contribute to the values of other predicates. Each predicate in the

REPAGE memory has a set of antecedents and a set of consequents. If an anteced-
ent changes its value or is removed, the predicate is notified. Then the predicate
recalculates its value and notifies the change to its consequents.
15.3.2.2 Detectors
The detectors are inference units specialized in certain predicates. They populate
the REPAGE memory (and consequently the main memory of the agent) with new
predicates inferred from those already in the memory. They are also responsible for
removing predicates that are no longer useful and, more importantly, for creating
the network of dependencies among the predicates.
Each time a new predicate is added to or removed from the main memory (either
by the action of another agent module – planner, communication module, etc. – or
by the action of a detector) the REPAGE memory notifies all detectors ‘interested’
in that type of predicate. This starts a cascading process where several detectors
are activated one after the other. At the same time, the dependency network ensures
the correct update of the predicate values according to the new additions and
subtractions.
15.3.2.3 Analyzer
The main task of the analyzer is to propose actions that (1) can improve the
accuracy of the predicates in the REPAGE memory and (2) can solve cognitive
dissonances trying to produce a situation of certainty. The analyzer can propose one
or more suggestions to the planner, which then decides whether to execute them
or not.
15.3.2.4 Integration with Deliberative Agent Architecture
One of the key points of the REPAGE design is its easy integration with the other
elements that compose a deliberative agent. REPAGE is not only a passive module
the agent can query to obtain information about image and reputation of another
agent. The aim of REPAGE is also to provide the agent (or more specifically, the
planner module of the agent) with a set of possibilities that can be followed to
improve the reliability of the provided information.
The communication module connects the agent with the rest of the world. After a
possible process of filtering and/or transformation of the received information, new
predicates are added to the main memory.
The REPAGE memory contains references to those predicates in the main
memory of the agent that are relevant to deal with image and reputation. The
actions of the detectors over the REPAGE memory result in addition/removal of
predicates as well as the creation of the dependence network. While the addition or
removal of predicates has again an immediate effect on the main memory, the
dependence network is present only in the REPAGE memory.
The planner uses the information in the main memory to produce plans. This
information includes the information generated by REPAGE. By means of the
analyzer, REPAGE always suggests new actions to the planner in order to improve
the accuracy of existing images and reputations. It is a task of the planner to decide
which actions are worth being performed. These actions (usually asking informers
or interacting with other agents) will hopefully provide new information that will
feed REPAGE and improve its accuracy. This cycle is illustrated in Fig. 15.1(B).
15.3.2.5 Demonstration
To illustrate some of the main points in the behaviour of REPAGE let us consider
two particular situations that are quite common in the application area of markets.
The general scenario is the following: Agent X is a buyer who knows that agent
Y sells what he needs but knows nothing about the quality of agent Y (the target of
the evaluations) as a seller. Therefore, he turns to other agents in search for
information – the kind of behaviour that can be found, for example, in Internet
fora, auctions, and in most agent systems.
In the first situation, agent X receives a communication from agent Z saying that
his image of agent Y as a seller is very good. Since agent X does not yet have an
image about agent Z as informer he resorts to a default image that is usually quite
low. The uncertain image as an informer adds uncertainty to the value of the
communication.
Later on, agent X has received six communications from different agents
containing their image of agent Z as an informer. Three of them give a good report
and three a bad one. This information is enough for agent X now to build an image
about agent Z as an informer so this new image substitutes the default candidate
image that was used so far. However, the newly formed image is insufficient to take
any strategic decision – the target seems to show an irregular behaviour.
At this point, agent X decides to try a direct interaction with agent Y. Because he
is not sure about agent Y he resorts to a low risk interaction. The result of this
interaction is completely satisfactory and has important effects in the REPAGE
memory. The candidate image about agent Y as a seller becomes a full image, in this
case a positive one.
Moreover, this positive image is compared (via a fuzzy metric) with the infor-
mation provided by agent Z (which was a positive evaluation of agent Y as a seller);
since the comparison shows that the evaluations are similar, a positive confirmation
of the image of agent Z as an informer is generated. This reinforcement of the image
of agent Z as a good informer at the same time reinforces the image of agent Y as a
good seller. As a consequence, there is a positive feedback between the image of
agent Y as a good seller and the image of agent Z as a good informer. This feedback
is a necessary and relevant part of the REPAGE model.
15 Reputation 391
The purpose of the second situation is to show how REPAGE differentiates

between image and reputation. In this case agent X, after a couple of successful
interactions with agent Y, receives four communications from different informants.
Each informant communicates the reputation of agent Y as a seller, which happens
to be negative. This contradicts agent X’s own positive evaluations of Y stemming
from their direct interactions. However, it is not a problem in REPAGE because
there is a clear distinction between image and reputation. In addition, unlike
communicated images (see first situation) communications about reputation do
not generate confirmations that reinforce or weaken the image of the informant.
REPAGE is a cognitive architecture to be used by artificial autonomous agents in

different scenarios, be they MAS applications, artificial markets or teamwork. Its
main objective is to give agents access to a fundamental module for social
reasoning and decision-making on the first two levels described in Sect. 15.2.3.1
and 15.2.3.2, epistemic and strategic. Thus, the main outputs of REPAGE consist of
“advice” about what to think of a given target and how to interact with it.
However, there are still some aspects of reputation-based decisions, which need
further theoretical elaboration and inclusion into REPAGE. First, the interplay
between image and reputation in the epistemic decision: which received evaluations
will be transformed into own evaluations of a given target and why? Secondly, the
grounds on which memetic decisions are made: what is said to whom, and moreover,
why? Both these issues will prove particularly relevant in testing this architecture by
means of simulation.
15.3.3 SOCRATE
SOCRATE is an attempt to test the cognitive theory of reputation in a ideal-typical

economic setting, modeled after an industrial district in which firms exchange
goods and information (Giardini et al. 2008; Di Tosto et al. 2010). In this model,
the focus is on social relationships among agents and their impact on the focusing
on social links and on the resulting social structure, usually informal, which is a
defining feature of industrial clusters (Porter 1998; Fioretti 2005; Squazzoni and
Boero 2002). Social evaluations are the building blocks of social and economic
relationships inside the cluster; they are used to select trustworthy partners, to
create and enlarge the social network (Giardini and Cecconi 2010), and to exert
social control on cheaters.
15.3.3.1 Agents and Environment
We implemented an artificial environment in which agents can choose among

several potential suppliers by relying either on their own evaluations, or on other
agents’ evaluations. In the latter case, the availability of truthful information could
help agents to find reliable partners without bearing the costs of potentially nega-
tive, i.e. harmful, interactions with bad suppliers. Moreover, evaluations can be
transmitted either as image (with an explicit source and the consequent risk of
retaliation) or as reputation.
We tried to answer the following questions:
• How relevant is image when firms need to select suppliers, service providers and
so on?
• Does transmission of image promote the improvement of quality in a cluster?
• How does false information affect the quality of the cluster?
• What are the effects of image and reputation, respectively, on the economic
performance of firms?
Our model is characterized by the existence of two different kinds of interactions
among agents: material exchange and evaluation exchange. The former refers to the
exchange of products between leader firms and their suppliers, and it leads to the
creation of a supply chain network. On the other hand, the flows of social
evaluations among the firms create a social network. In this setting, agents can
transmit true or false evaluations in order to either help or hamper their fellows
searching for a good partner.
Agents are firms organized into different layers, in line with their role in the
production cycle. The number of layers can vary according to the characteristics of
the cluster, but a minimum of two layers is required. Here, we have three layers, but
n possible layers can be added, in order to develop a more complex production
process:
Layer 0 (L0) is represented by leader firms that supply the final product
Layer 1 (L1) is represented by suppliers of L0
Layer 2 (L2) are firms providing raw material to firms in L1.
When image transmission is allowed, both leader firms and suppliers exchange
information with their fellows, thus creating and taking part in a social network.
This process works only horizontally: L0 and L1 are not allowed to talk each other.
Agents in both layers can play two possible roles:
Questioner – asks an Informer, i.e. another firm of the same layer, to suggest a good
supplier;
Informer – provides her own image of a good supplier. Honest informers suggest
their best rated supplier, whereas cheaters transmit the image of their worse
supplier (as if it was a good one).
15 Reputation 393
Agents in L0 have to select suppliers that produce with a quality above the
average among all L1 agents. Suppliers can be directly tested or they can be chosen
thanks to the information received by other L0 firms acting as Informers. Buying
products from L1 and asking for information to L0 fellows are competing activities
that can not be performed contemporaneously. In turn, once received an order for a
product, L1 firms should select a good supplier (above the average quality) among
those in L2. After each interaction with a supplier, both L0 and L1 agents create an
evaluation, i.e. an image, of it, comparing the quality of the product they bought
with the quality threshold value. Agents are endowed with an “Image Table” in
which all the values of the tested partners are recorded and stored for future
selections. In the Reputation condition, evaluations are exchanged without reveal-
ing their source, thus injecting the cluster with untested information. In this
condition, retaliation against untrustful informers is unattainable.
At each simulation cycle, firms attempt to interact with the best known suppliers.
Every time the best known supplier is unavailable they query their fellows about
other high quality suppliers, which will be tested and integrated into the Image-
Table.
15.3.3.2 Results
The exchange of true and valuable information led to a significant increase in

quality of exchanged goods, making the cluster exploration faster and permitting
firms to obtain higher profits. However, in the Image condition average quality was
negatively affected by the presence of cheaters, because false information triggered
a mechanism of reciprocal retaliation with detrimental effects on the cluster as a
whole. In the Reputation Condition, the cluster could absorb relatively high
percentages of cheaters, without compromising its economic performance.
SOCRATE results provided further support to the hypotheses about the impor-
tance of reputation for social control, showing again that social evaluations and
their features have consequences also in economic terms.
Given the assumption that in this “small-world”, as in the real world, evaluations
are crucial to selecting trustworthy partners and to isolating cheaters, we tried to
demonstrate how useful this exchange is, especially in terms of global cluster
quality and profits. Firms receiving reliable information about potential partners
found good suppliers in a faster and more efficient way, compared to firms that were
systematically cheated by their fellows. More interesting results are expected after
we include an enriched economic structure and the implementation of reputation.
15.4 Conclusion and Future Work
In the last decade there has been a significant increase in research on reputation and
gossip. There is growing evidence on the fact that the presence of reputation
strongly promotes cooperation and represents an effective way to maintain social
control. Exercising social control roughly means to isolate and punish the cheaters.
However, punishment is costly and it inevitably implies the problem of second-
order cooperation.
In this chapter, we discussed current studies of reputation as a distributed
instrument for social order. After a critical review of current technologies of
reputation in electronic institutions and agentized environments, a theory of repu-
tation as a social cognitive artefact was presented. In this view, reputation allows
agents to cooperate at a social meta-level, exchanging information (a) for partner
selection in competitive settings like markets and (b) for cheater isolation and
punishment in cooperative settings like teamwork and grouping.
To exemplify both functionalities, we introduced three simulation models of
reputation in artificial societies developed within our research group during the last
decade. Both have been used mainly as a theory-building tool.
The first, Sim-Norm, is a reputation-based model for norm compliance. The
main findings from simulations show that, if circulated among norm-abiders only,
reputation allows for the costs of compliance to be redistributed between two
balanced subpopulations of norm-abiders and cheaters. In such a way, it contributes
to the fitness of the former, neutralising the advantage of cheaters. However, results
also show that as soon as the latter start to bluff and optimistic errors begin to spread
in the population, things worsen for norm-abiders, to the point that the advantage
produced by reputation is nullified.
REPAGE, a much more complex computational model than SimNorm, was
developed to test the impact of image, reputation and their interaction on the
market. Based on our social cognitive theory, it allows the distinction between
image and reputation to be made, and the trade-off between agents’ autonomy and
their liability to social influence to be coped with. REPAGE allows the circulation
of reputation whether or not third parties accept it as true.
Finally, SOCRATE is an attempt to combine fairly complex agents (endowed
with a memory and able to manage different kinds of evaluations) with a market in
which agents must protect themselves from both informational and material
cheating. In this context, reputation has been proven to be useful to punish cheaters
but it also prevented the social network from collapse.
These results clearly show that differentiating image from reputation provides a
means for coping with informational cheating and that further work is needed to
achieve a better understanding of this complex phenomenon. The long term results of
these studies are expected to (a) answer the question on how to cope with informa-
tional cheating (by testing the above hypothesis), (b) provide guidelines about how to
realize technologies of reputation that achieve specified objectives (e.g. promoting
15 Reputation 395
respect of contracts vs. increasing volume of transactions), and finally (c) show the
impact of reputation on the competitiveness of firms within and between districts.
Acknowledgements The authors would like to thank Jordi Sabater and Samuele Marmo for their
helpful collaboration. This work was partially supported by the Italian Ministry of University and
Scientific Research under the Firb programme (SOCRATE project, contract number RBNE03Y338),
by the European Community under the FP6 programme (eRep project, contract number CIT5-
028575; EMIL project, contract number IST-FP6-33841).
Further Reading
For a more in-depth treatment of the contents of this chapter we refer the reader to
the monograph Reputation in Artificial Societies (Conte and Paolucci 2002).
A review of this book was published in JASSS (Squazzoni 2004). For more on
the same line of research, with an easier presentation aimed to dissemination, we
suggest the booklet published as the result of the eRep project (Paolucci et al. 2009).
A conference aiming to propose a scientific approach to Reputation has been
organized in 2009: the first International Conference on Reputation, ICORE 2009.
Its proceedings (Paolucci 2009), available online, contain a collection of papers that
give an idea of the range of approaches and ideas on Reputation from several
academic disciplines.
Due to the focus on the theoretical background of reputation only a narrow
selection of simulation models of reputation could be discussed in this chapter.
Sabater and Sierra (2005) give a detailed and well-informed overview of current
models of trust and reputation using a variety of mechanisms. Another good starting
point for the reader interested in different models and mechanisms is the review by
Ramchurn and colleagues (Ramchurn et al. 2004a).
Further advanced issues for specialised reputation subfields can be found in
(Jøsang et al. 2007), a review of online trust and reputation systems, and in (Koenig
et al. 2008), regarding the Internet of Services approach to Grid Computing.
References
Abdul-Rahman A, Hailes S (1997a) A distributed trust model. In: Proceedings of NSPW97,

new security paradigms workshop VI, Langdale, Cumbria, 23–26 Sept 1997. ACM Press,
New York, pp 48–60
Akerlof G (1970) The market for lemons: quality uncertainty and the market mechanisms. Q J
Econ 84:488–500
Alexander R (1987) The biology of moral systems. De Gruyter, New York
Bachmann R (1998) Kooperation, vertrauen und macht in systemen verteilter künstlicher
intelligenz: eine vorstudie zum verhältnis von soziologischer theorie und technischer
modellierung. In: Malsch T (ed) Sozionik. Edition Sigma, Berlin, pp 197–234
Barkow JH (1996) Beneath new culture is old psychology: gossip and social stratification. In:
Barkow JH, Cosmides L, Tooby J (eds) The adapted mind: evolutionary psychology and the
generation of culture. Oxford University Press, New York, pp 627–638
Bohem C (1999) Hierarchy in the forest: the evolution of egalitarian behavior. Harvard University Press,
Cambridge, MA
Bolton G, Katok E, Ockenfels A (2002) How effective are online reputation mechanisms? An
experimental investigation (Working paper). Max Planck Institute of Economics, Jena
Cabral L, Hortaçsu A (2006) The dynamics of seller reputation: theory and evidence from eBay
(Discussion papers, 4345). Centre for Economic Policy Research, London
Carbo J, Molina JM, Davila J (2002) Comparing predictions of SPORAS vs. a fuzzy reputation
agent system. In: Grmela A, Mastorakis N (eds) Proceedings of the 3rd WSEAS international
conference on Fuzzy Sets and Fuzzy Systems (FSFS ’02), Interlaken, 11–15 Feb 2002.
WSEAS, pp 147–153
Castelfranchi C (1998) Modelling social action for AI agents. Artif Intell 103(1–2):157–182
Castelfranchi C, Conte R, Paolucci M (1998) Normative reputation and the costs of compliance.
Conte R (1999) Social intelligence among autonomous agents. Comput Math Organ Theory 5:
202–228
Conte R, Castelfranchi C (1995) Cognitive and social action. UCL Press, London
Conte R, Paolucci M (1999) Reproduction of normative agents: a simulation study. Adapt Behav 7
(3/4):301–322
Conte R, Paolucci M (2002) Reputation in artificial societies: social beliefs for social order.
Kluwer, Dordrecht
Conte R, Paolucci M (2003) Social cognitive factors of unfair ratings in reputation reporting
systems. In: Liu J, Liu C, Klush M, Zhong N, Cercone N (eds) Proceedings of the IEEE/WIC
international conference on web intelligence – WI 2003. IEEE Computer Press, Halifax,
pp 316–322
Conte R, Castelfranchi C, Dignum F (1998) Autonomous norm-acceptance. In: Müller JP, Singh
MP, Rao AS (eds) ATAL ’98: proceedings of the 5th international workshop on intelligent
agents V, agent theories, architectures, and languages. Springer, London, pp 99–112
Cravens K, Goad Oliver E, Ramamoorti S (2003) The reputation index: measuring and managing
corporate reputation. Eur Manage J 21(2):201–212
Dellarocas CN (2003) The digitalization of word-of-mouth: promise and challenges of online
feedback mechanisms (MIT Sloan working paper, 4296–03). MIT Sloan School of Manage-
ment, Cambridge, MA
Di Tosto G, Giardini F, Conte R (2010) Reputation and economic performance in industrial
districts: modelling social complexity through multi-agent systems. In: Takadama K, Cioffi-
Revilla C, Deffuant G (eds) Simulating interacting agents and social phenomena: the second
world congress, vol 7, ABSS. Springer, Tokyo, pp 165–176
Dunbar R (1996) Grooming, gossip, and the evolution of language. Faber and Faber, London
Fehr E, Gachter S (2000) Fairness and retaliation: the economics of reciprocity. J Econ Perspect
14:159–181
Fioretti G (2005) Agent based models of industrial clusters and districts (Working paper, series
urban/regional, 0504009). EconWPA http://ideas.repec.org/p/wpa/wuwpur/0504009.html
Frith CD, Frith U (2006) How we predict what other people are going to do. Brain Res 1079
(1):36–46
Giardini F, Cecconi F (2010) Social evaluations, innovation and networks. In: Ahrweiler P (ed)
Innovation in complex social systems. Routledge, London, pp 277–289
Giardini F, Di Tosto G, Conte R (2008) A model for simulating reputation dynamics in industrial
districts. Simul Model Pract Theory 16:231–241
Gintis H, Smith EA, Bowles S (2001) Costly signaling and cooperation. J Theor Biol 213:103–119
Gluckman M (1963) Gossip and scandal. Curr Anthropol 4:307–316
15 Reputation 397
Gray ER, Balmer JMT (1998) Managing corporate image and corporate reputation. Long Rage
Plann 31(5):695–702
Greif A (1993) Contract enforceability and economic institutions in early trade: the maghribi
traders’ coalition. Am Econ Rev 83(3):525–548
Hales D (2002) Group reputation supports beneficent norms. J Artif Soc Soc Simul 5(4). http://
Huynh D, Jennings NR, Shadbolt NR (2004) Developing an integrated trust and reputation model
for open multi-agent systems. In: Falcone R, Barber S, Sabater-Mir J, Singh M (eds)
Proceedings of the 7th international workshop on trust in agent societies, New York, 19 July
2004, pp 65–74
IC3 (2005): IC3 2005 Internet crime report http://www.ic3.gov/media/annualreport/2005_
IC3Report.pdf
Jøsang A, Ismail R, Boyd C (2007) A survey of trust and reputation systems for online service
provision. Decis Support Syst 43(2):618–644
Koenig S, Hudert S, Eymann T, Paolucci M (2008) Towards reputation enhanced electronic
negotiations for service oriented computing. In: Falcone R, Barber S, Sabater-Mir J, Singh M
(eds) Trust in agent societies: 11th international workshop, TRUST 2008, Estoril, 12–13 May
2008, revised selected and invited papers. Springer, Berlin, pp 273–291
Kreps D, Wilson R (1982) Reputation and imperfect information. J Econ Theory 27:253–279
Leslie AM (1992) Autism and the ‘theory of mind’ module. Curr Dir Psychol Sci 1:18–21
Marsh S (1992) Trust in distributed artificial intelligence. In: Castelfranchi C, Werner E (eds)
Artificial social systems, 4th European workshop on modelling autonomous agents in a multi-
agent world, MAAMAW ’92, S. Martino al Cimino, 29–31 July 1992, selected papers.
Marsh S (1994a) Formalising trust as a computational concept. PhD Thesis, Department of
Computing Science and Mathematics, University of Stirling, Stirling. http://www.cs.stir.ac.
uk/research/publications/techreps/pdf/TR133.pdf
Marsh S (1994b) Optimism and pessimism in trust. In: Proceedings of the Ibero-American
conference on artificial intelligence (IBERAMIA-94), Caracas, 25–28 Oct 1994
Melnik MI, Alm J (2002) Does a seller’s eCommerce reputation matter? Evidence from eBay
auctions. J Ind Econ 50(3):337–349
Miceli M, Castelfranchi C (2000) The role of evaluation in cognition and social interaction. In:
Dautenhahn K (ed) Human cognition and social agent technology. Benjamins, Amsterdam,
pp 225–262, chapter 9
Moukas A, Zacharia G, Maes P (1999) Amalthaea and histos: multiAgent systems for WWW sites
and reputation recommendations. In: Klusch M (ed) Intelligent information agents: agent-
based information discovery and management on the internet. Springer, Berlin, pp 292–322
Noon M, Delbridge R (1993) News from behind my hand: gossip in organizations. Organ Stud 14:
23–36
Nowak MA, Sigmund K (1998a) Evolution of indirect reciprocity by image scoring. Nature 393:
573–577
Nowak MA, Sigmund K (1998b) The dynamics of indirect reciprocity. J Theor Biol 194:561–574
Paine R (1967) What is gossip about? An alternative hypothesis. Man 2(2):278–285
Panchanathan K, Boyd R (2004) Indirect reciprocity can stabilize cooperation without the second-
order free rider problem. Nature 432:499–502
Paolucci M (2000) False reputation in social control. Adv Complex Syst 3(1–4):39–51
Paolucci M (ed) (2009) Proceedings of the first international conference on reputation: theory and
technology – ICORE 09, Gargonza. Institute of Science and Technology of Cognition,
National Research Centre (ISTC-CNR), Rome. http://pagesperso-systeme.lip6.fr/Erika.Rosas/
pdf/ICORE09.pdf
Paolucci M, Conte R (2009) Reputation: social transmission for partner selection. In: Trajkovski GP,
Collins SG (eds) Handbook of research on agent-based societies: social and cultural interactions.
IGI, Hershey, pp 243–260
Paolucci M et al (2009) Theory and technology of reputation (Technical report, FP6 Research
Project “Social knowledge for e-Governance” (eRep)). Institute of Science and Technology of
Cognition, National Research Centre (ISTC-CNR), Rome. http://issuu.com/mario.paolucci/
docs/erep_booklet
Pinyol I, Paolucci M, Sabater-Mir J, Conte R (2008) Beyond accuracy: reputation for partner
selection with lies and retaliation. In: Antunes L, Paolucci M, Norling E (eds) Multi-agent-
based simulation VIII: international workshop, MABS 2007, Honolulu, 15 May 2007, revised
and invited papers (Lecture notes in computer science 5003). Springer, Berlin, pp 128–140
Porter M (1998) Clusters and the new economics of competition. Harv Bus Rev 76:77–90
Ramchurn SD, Huynh D, Jennings NR (2004a) Trust in multiagent systems. Knowl Eng Rev 19(1):1–25
Ramchurn SD, Sierra C, Godo L, Jennings NR (2004b) Devising a trust model for multi-agent
interactions using confidence and reputation. Int J Appl Artif Intell 18:833–852
Rasmusson L (1996) Socially controlled global agent systems (Working paper). Department of
Computer and Systems Science, Royal Institute of Technology, Stockholm
Rasmusson L, Janson S (1996) Simulated social control for secure internet commerce. In:
Proceedings of the 1996 workshop on new security Paradigms, Lake Arrowhead. ACM
Press, New York, pp 18–25
Regan K, Cohen R (2005) Indirect reputation assessment for adaptive buying agents in electronic
markets. In: Proceedings of workshop business agents and the semantic web (BASeWEB’05),
8 May 2005, Victoria, British Columbia
Resnick P, Zeckhauser R (2002) Trust among strangers in internet transactions: empirical analysis
of eBay’s reputation system. In: Baye MR (ed) The economics of the internet and e-commerce,
vol 11, Advances in applied microeconomics. Elsevier Science, Amsterdam, pp 127–157
Ripperger T (1998) Ökonomik des Vertrauens: Analyse eines Organisationsprinzips. Mohr Siebeck,
Tübingen
Rocco E, Warglien M (1995) La comunicazione mediata da computer e l’emergere
dell’opportunismo elettronico. Sistemi Intelligenti 7(3):393–420
Rose C, Thomsen S (2004) The impact of corporate reputation on performance: some Danish
evidence. Eur Manage J 22(2):201–210
Saam N, Harrer A (1999) Simulating norms, social inequality, and functional change in artificial
societies. J Artif Soc Soc Simul 2(1), http://jasss.soc.surrey.ac.uk/2/1/2.html
Sabater J, Paolucci M (2007) On representation and aggregation of social evaluations in computa-
tional trust and reputation models. Int J Approx Reason 46(3):458–483
Sabater J, Sierra C (2002) Reputation and social network analysis in multi-agent systems. In:
Proceedings of the first international joint conference on autonomous agents and multiagent
systems (AAMAS 2002), 15–19 July 2002, Bologna. ACM Press, New York, pp 475–482
Sabater J, Sierra C (2005) Review on computational trust and reputation models. Artif Intell Rev
24(1):33–60
Sabater J, Paolucci M, Conte R (2006) REPAGE: REPutation and ImAGE among limited
autonomous partners. J Artif Soc Soc Simul 9(2), http://jasss.soc.surrey.ac.uk/9/2/3.html
Schillo M (1999) Vertrauen: Ein Mechanismus zur sicheren Koalitionsbildung in künstlichen
Gesellschaften (Trust: a mechanism for reliable coalition formation in artificial societies).
Master’s thesis, Department of Computer Science, Saarland University, Saarbrücken,
Germany.
Schillo M, Funk P, Rovatsos M (2000) Using trust for detecting deceitful agents in artificial
societies. Appl Artif Intell 14:825–848
Sell J, Wilson R (1991) Levels of information and contributions to public goods. Soc Forces 70:
107–124
Sommerfeld RD, Krambeck H, Milinski M (2008) Multiple gossip statements and their effects on
reputation and trustworthiness. Proc R Soc B 275:2529–2536
Squazzoni F, Boero R (2002) Economic performance, inter-firm relations and local institutional
engineering in a computational prototype of industrial districts. J Artif Soc Soc Simul 5(1).
15 Reputation 399
Utz S, Matzat U, Snijders CCP (2009) On-line reputation systems : the effects of feedback
comments and reactions on building and rebuilding trust in on-line auctions. International
Journal of Electronic Commerce 13(3):95–118
Wilson DS, Wilczynski C, Wells A, Weiser L (2000) Gossip and other aspects of language as
group-level adaptations. In: Heyes C, Huber L (eds) The evolution of cognition. MIT Press,
Cambridge, pp 347–366
Yamagishi T, Matsuda M, Yoshikai N, Takahashi H, Usui Y (2009) Solving the lemons problem
with reputation. An experimental study of online trading. In: Cook KS, Snijders C, Vincent B,
Cheshire C (eds) eTrust: forming relationships in the online world. Russel Sage Foundation,
Yu B, Singh MP (2000) A social mechanism of reputation management in electronic communities.
In: Klusch M, Kerschberg L (eds) Cooperative information agents IV, the future of information
agents in cyberspace, 4th international workshop, CIA 2000, Boston, 7–9 July 2000,
proceedings (Lecture notes in computer science, 1860). Springer, Berlin, pp 154–165
Zacharia G (1999) Trust management through reputation mechanisms. In: Castelfranchi C,
Falcone R, Firozabadi BS (eds) Proceedings of the workshop on deception, fraud and trust in
agent societies at autonomous agents’99, Seattle. pp 163–167
Zacharia G, Moukas A, Maes P (1999) Collaborative reputation mechanisms in electronic
marketplaces. In: Proceedings of the 32nd annual Hawaii international conference on system
sciences (HICSS-32), 5–8 Jan 1999, Maui, Track 8: software technology. IEEE computer society,
Los Alamitos
Chapter 16
Social Networks and Spatial Distribution
Frédéric Amblard and Walter Quattrociocchi
Why Read This Chapter? To learn about interaction topologies for agents,
from social networks to structures representing geographical space, and the main
questions and options an agent-based modeller has to face when developing and
initialising a model.
Abstract In most agent-based social simulation models, the issue of the
organisation of the agents’ population matters. The topology, in which agents
interact, – be it spatially structured or a social network – can have important impacts
on the obtained results in social simulation. Unfortunately, the necessary data about
the target system is often lacking, therefore you have to use models in order to
reproduce realistic spatial distributions of the population and/or realistic social
networks among the agents. In this chapter we identify the main issues concerning
this point and describe several models of social networks or of spatial distribution
that can be integrated in agent-based simulation to go a step forward from the use of
a purely random model. In each case we identify several output measures that allow
quantifying their impacts.
16.1 Introduction
Independent of the methodology followed to build a model, any agent-based

modeller has to face not only the design of the agents and their behaviour, but
also the design of the topology in which the agents interact. This can be a spatial
structure, so that the agents are distributed within a representation of geographical
space, or a social structure, linking agents as nodes in a network, or even both.
F. Amblard (*)
IRIT, Université Toulouse 1 Capitole, Toulouse, France
W. Quattrociocchi
LABSS, University of Siena, Italy

402 F. Amblard and W. Quattrociocchi
Interaction topologies can be determined explicitly or implicitly. They are

explicit when they are specified as modelling hypotheses and thus clearly defined
within the model. Conversely, topologies are implicit when they are inferred from
other processes and thus not a definite part of the model. We will come back to the
consequences of such a classification in the next section.
There are three major issues to solve when dealing with social and spatial
structures:
Implementation: How to represent the structure in the model, e.g. continuous vs.
discrete representation of space or which data structure to choose for the social
network;
Initialisation: How to initialise the chosen structure, e.g. which initial shape of the
network should be chosen and how should the population of agents be
distributed on this network;
Observation: How to characterise a given structure and/or its evolution, potentially
taking into account agents’ states related to their place in the structure. This
latter point raises the question of the indicators to observe during the simulation
in order to follow changes in either spatial or social structures.
Since the answers to these questions differ quite substantially for spatial and
social structures, we will discuss them separately in Sects. 16.3 and 16.4. The
following section, while focussing on the topic of explicitly versus implicitly defined
interaction topologies, will also discuss the situation where both social and spatial
structures have to be taken into account at the same time; leading to either spatially-
embedded networks or graph-like representations of spatial structures.
16.2 Explicit and Implicit Structures
As mentioned in the introduction, there is a difference between explicit and implicit

structures. We define explicit structures as clearly implemented modelling
hypotheses, which can therefore be identified in the model. Implicit structures, on
the other hand, are not directly defined in the model and are rather determined as a
result of other processes and/or hypotheses in the model. We will demonstrate
the difference with the help of two examples and then go on to explore the
consequences of ex- versus implicitness with regard to the three issues of imple-
mentation, initialisation and observation introduced above.
16.2.1 Example 1: Schelling’s Segregation Model
To present it briefly, the segregation model of Thomas Schelling (1971) is

composed of a set of agents, some red, the others green. Each agent is positioned
on an empty square of a chessboard (representing the environment). If the
16 Social Networks and Spatial Distribution 403
Fig. 16.1 Schelling’s segregation model in NetLogo, the spatial structure is represented on the left
and the inferred social structure represented as a graph on the right figure
proportion of neighbours of the same colour falls below the value of the agent’s
tolerance variable, the agent moves to a randomly chosen empty square (originally:
the nearest empty square at which it would be satisfied with the neighbourhood);
otherwise, it stays where it is.
The spatial structure of this model is explicit and is represented explicitly as a
grid. As such, it is discrete, regular and static (the distribution of the agents on this
structure evolves, not the structure itself). The social structure, on the other hand, is
implicit in this model. During the simulation the agents take into account the type
(green or red) of their neighbours on the grid, but the corresponding social structure
is not defined as such and is inferred from the spatial distribution of the agents
(cf. Fig. 16.1). As discussed later, it can be interesting in such a case to characterise
the implicit social structure of the model, as this is the one that drives the model,
whereas the spatial structure merely acts as a constraint (in particular concerning
the maximal number of neighbours) on the evolution of the social structure.
16.2.2 Example 2: IMAGES Innovation Dynamics Model
In the innovation dynamics model developed in the FAIR-IMAGES project

(IMAGES 2004) agents, representing farmers, are linked via a social network.
This defines the paths, over which information is diffused, and controls which agents
influence each other. The social network is determined, at least in part, from the
geographical locations of the agents to account for the fact that geographical
neighbours tend to know each other. In this example the social structure is explicit
as it is built into the model by the hypothesis (Fig. 16.2).
Fig. 16.2 Reconstruction of the social network among agents incorporating the geographical
distance (IMAGES 2004)
16.2.3 Questions Linked to the Implicit/Explicit Property

of the Structures
The property of being explicit or implicit enables us to narrow down the range of
possible answers to the three questions raised in the introduction. To begin with, the
question of implementation can only be asked when dealing with explicit
structures; the same is true for the question of initialisation. However,
characterising the implicit social structure in spatial models, i.e. identifying at a
given time step the whole set of interactions among agents that could or do take
place, can give useful hints for understanding the underlying dynamics. Identifying,
for instance, separate components in the implicit social network inferred from a
spatial model is more informative than solely identifying spatial clusters as it
confirms that there is effectively no connection among the different groups.
In the next two sections we will discuss the conditions in which social and spatial
structures can be considered as independent features of social analysis and can
therefore be presented independently. This is generally the case but, as we will
detail in the last section, there are some exceptions.
16.3 Social Networks
Social networks have been analysed extensively during the last decade. From the
social network of scientists (co-authorship network or co-citation network)
(Newman 2001, 2004; Jeong et al. 2002; Meyer et al. 2009) to the social network
of dolphins (Lusseau 2003), many empirical studies on large graphs popularised
this fascinating subject: ‘social’ links between social beings. Neither empirical
analysis nor theoretical modelling is new in this field. From the formalisation of
graphs by Euler in the eighteenth century in order to represent paths in Königsberg,
which led to the now well established graph theory, to social network analysis in
sociology, originating from the use of the socio-matrix of Moreno, social networks
are now quite commonly used in agent-based simulation to explicitly represent the
topology of interactions among a population of agents.
In order to clarify the kind of modelling issues you are dealing with, we can
divide modelling of social networks into three categories: (a) static networks,
(b) dynamic networks with the dynamics independent of the agents’ states (for
instance random rewiring process), (c) dynamic networks evolving dependent on
the agents’ states. In this chapter we will concentrate on the first case since it is the
most common; although the use of the second case has recently started to grow
rapidly, while the third case is still in its incipient stage.
In each case, the same three questions arise with regard to implementation,
initialisation and observation:
• Which data structure is best suited to represent the network?
• Which initial (and in the static case, only) network configuration to use?
• How to identify something interesting from my simulation including the network
(e.g. a social network effect)?
16.3.1 Which Data Structure to Use?
Although it could seem trivial, especially when you use a high-level modelling
platform such as NetLogo or Repast, this issue is important concerning the execu-
tion efficiency of your model. More important even, depending on your choice,
biases are linked to some data structures when using particular classes of networks
such as scale-free networks.
Basically using an object-oriented approach, you have two choices: either to
embed social links within the agent as pointers to other agents or to externalise the
whole set of links as a global collection (called Social Network, for instance). The
former is more practical when having 1 n interactions rather than 1–1
interactions, i.e. taking into account all neighbours’ states to determine the new
state of the agent rather than picking one agent at random in the neighbourhood.
The difference between the two solutions is mainly related to the scheduling you
will use. You can choose either to first schedule the agents, picking an agent at
random from the population and then selecting one (or more) of its social links. Or
you can choose to pick a random link from the global collection and then execute
the corresponding interaction. While this choice will depend a lot on the kind of
model you are implementing, it is crucial when using scale-free networks since both
options may produce a bias and you will have to choose the solution that is more
relevant for the purpose of your model.
To be able to explain this further, we need to quickly introduce scale-free

networks (they will be presented in more detail in Sect. 16.3.2.4). Their main
property is that the distribution of the number of links per agent (node) follows a
power law, meaning that very few agents – the so-called “hubs” – have a lot of links
while the majority of agents have only a few.
If you now choose to schedule the agents first, you effectively apply an egalitarian
rule on the individuals, which however results in the links involving the hubs being
less frequently scheduled than the links involving agents with few social relations.
Take for example a population of 100 agents where one agent has 30 links and
another has only one link. Each of these two agents has the same probability to be
scheduled (0.01), but if you then proceed to select a random link from the scheduled
agent, the links of the hub agent each have a 0.01/30 probability to be chosen while
the one link of the other agent still has a 0.01 probability to be picked.
Conversely, if you schedule the collection of links first, i.e. apply an egalitarian
rule on the individual links, the influence of the hubs in the global dynamics will be
strengthened, as they are involved in more links. Therefore, the initial states of the
hub agents are also interesting with respect to the whole dynamic of the model.
16.3.2 Which Initial Network Configuration to Use?
This question mainly arises when choosing which (initial) social structure to
implement. There is a large choice of models – each with some advantages and
some drawbacks – that can be distinguished into four categories: (a) regular graphs,
lattices or grids being the most longstanding structure used in social simulation
(inherited from the cellular automata approaches), (b) random graphs, (c) small-
world networks, and (d) scale-free networks. Concerning the latter three categories,
Newman et al. (2006) regroup an important set of articles that can be useful as
advanced material on this point.
16.3.2.1 Lattices
The field of social modelling inherited many tools from mathematics and physics
and in particular cellular automata (Wolfram 1986). The corresponding underlying
interaction structure is then in general a grid and in many cases a torus. The cells of
the automata represent the agents and their social neighbourhood is defined from
the regular grid with a von Neumann or a Moore neighbourhood. The von Neumann
neighbourhood links a cell to its four adjacent cells (North, East, South, West) while
a Moore neighbourhood adds four more neighbours (NE, SE, SW, NW; cf.
Fig. 16.3).
The main advantage of using regular grids stems from visualisation, regular
grids enabling very efficient visualisations of diffusion processes or clustering
process (cf. Fig. 16.4).
Fig. 16.3 von Neumann (a) and 3 3 Moore neighbourhood (b) on a regular grid; the illustration
on the right shows a torus, i.e. the result of linking the borders (north with south, east with west) of
a grid (Flache and Hegselmann 2001)
Fig. 16.4 Opinion dynamics model on a regular grid (Jager and Amblard 2005)
Fig. 16.5 Regular 1D

structure with
connectivity ¼ 4
The first important point concerning the regular structures deals with connectivity.
In contrast to other kinds of networks (random ones for instance, see next section),
using regular networks makes it difficult to change the connectivity of the structure,
i.e. the number of links per agent. The exploration of connectivity effects on the model
behaviour is limited in this case to specific values (4, 8, 12, 24. . . in the case of a
two-dimensional regular grid).
The second point deals with the dimension of the regular structure. A one-
dimensional lattice corresponds to a circle (see Fig. 16.5), two-dimensional
structures to grids (chessboard), three-dimensional structures to cubic graphs.
However, we have to notice that only 2D regular structures benefit from a visuali-
zation advantage, higher dimensions suffering from the classical disadvantage
associated with the visualization of dynamics on complex graphs.
The presence or absence of borders is important in regular graphs. Classic
example is the 2D grid, which – if not implemented as a torus – is not a regular
structure anymore, since agents localised at the borders have fewer connections.
Moreover, these agents being linked to each other creates a bias in the simulation
(Chopard and Droz 1998). This bias is sometimes needed for any dynamics to
happen and could either correspond to a modelling hypothesis or to an unwanted
artefact. This point is probably far clearer on one-dimensional graphs, where if you
do not close the circle the diameter1 is approximately the size of the population,
whereas if you close it, the diameter is half the population size. This issue
corresponds to border effects identified on cellular automata.
1
The diameter of a graph is defined as the length of the longest shortest-path in the graph.
Fig. 16.6 The β-model of Watts (1999) enables to go from regular graphs (on the left) to random
graphs (on the right) using rewiring of edges
16.3.2.2 Random Graphs
Another kind of model that can be used to generate social structures is the random
graph model (Solomonoff and Rapoport 1951; Erdös and Renyi 1960); see Fig. 16.6
on the right for an illustration. As told by Newman et al. (2006), there are two ways
to build random graphs containing n vertices: one (denoted as Gn,m) is to specify the
number m of edges between vertex pairs chosen at random, the other (denoted as
Gn,p) is to specify a probability p for an edge to link any two vertices. Both of them
correspond to graphs that have on average the same properties when they are big
enough. The only difference is that in Gn,m the number of edges is fixed, while in
Gn,p it may fluctuate from one instance to the other; but on average it is also fixed.
The first important property of random graphs is that they show a phase transition
when the average degree of a vertex is 1. Below this transition we obtain a number of
small components, while above this threshold the model exhibits a giant component
with some isolated nodes. The giant component is a subset of the graph vertices,
each of which is reachable from any of the others along some path(s). The random
graphs that are used most often in agent-based social simulation are slightly above
the threshold, in the giant-component phase with some isolated nodes.
Another important point concerns the degree distribution. The properties and
behaviour of a network are affected in many ways by its degree distribution (Albert
et al. 2000; Cohen et al. 2000; Callaway et al. 2000). Gn,p has a binomial degree
distribution (Poisson distribution for large n), which is sharply peaked and has a tail
that decays quicker than any exponential distribution.
It is possible to define random graphs with any desired degree distribution
(Bender and Canfield 1978; Luczak 1992; Molloy and Reed 1995). In this case,
one considers graphs with a given degree sequence rather than with a degree
distribution. A degree sequence is a set of degrees k1, k2, k3. . . for each of the
corresponding vertices 1, 2, 3. . .. Molloy and Reed (1995) suggest the following
algorithm:
• Create a list in which the label i of each vertex appears exactly ki times.
• Pair up elements from this list uniformly at random until none remain.
• Add an edge to the graph joining the two vertices of each pair.
According to Molloy and Reed (1995), these graphs possess a phase transition at
which a giant component appears, just as in the standard Poisson random graph.
In the context of agent-based social simulation, a great advantage of random
graphs over regular graphs is that you can easily change and precisely tune the
average connectivity of the graph and – applying Molloy and Reed’s algorithm –
the distribution of the edges among vertices.
Replacing regular graphs with random graphs, several scientists experienced a
“social network” effect, i.e. models having different macroscopic behaviours
depending on the chosen interaction structure (Stocker et al. 2001, 2002; Holme
and Grönlund 2005; Huet et al. 2007; Deffuant 2006; Gong and Xiao 2007;
Kottonau and Pahl-Wostl 2004; Pujol et al. 2005). The fact is that these two classes
of networks have very different characteristics. In terms of clustering, regular
graphs exhibit more clustering or local redundancy than the random graphs. On
the other hand, random graphs lead to a shorter diameter and average path length
among the pairs of individuals than a regular graph. Mean path length for a random
graph scales logarithmically with graph size. For more details concerning random
graphs models, we refer the interested reader to (Bollobas 2001).
16.3.2.3 Small World Networks
The question arising at this stage is: are there classes of graphs between these two
extremes (random and regular graphs) that may have other characteristics? Watts
and Strogatz (1998) introduced such a model, motivated by the observation that
many real world graphs share two main properties:
• The small-world effect, i.e. most vertices are connected via a short path in the
network.2
• High clustering, corresponding to the phenomenon that the more neighbours two
individuals have in common, the more likely they are to be connected
themselves.
Watts and Strogatz defined a network to be a small-world network if it exhibits
both of these properties, that is, if the mean vertex-vertex distance l is comparable
to that of a random graph and the clustering coefficient is comparable to that of a
regular lattice.
To construct such a network, Watts and Strogatz found the following algorithm.
Starting from a regular lattice with the desired degree, each link has a probability
p to be rewired, i.e. to be disconnected from one of its vertices and reconnected with
another vertex chosen uniformly at random. The result is the creation of shortcuts in
the regular structure. Watts and Strogatz (1998) imposed additional constraints on
the rewiring process: a vertex cannot be linked to itself, and any two vertices cannot
2
Short path being defined by Watts and Strogatz as comparable to those found in random graphs of
the same size and average degree.
Fig. 16.7 Average path

length (dark grey), and
clustering coefficient (light
grey) represented as the
probability of rewiring
evolves
be linked by more than one edge. Moreover, the rewiring process only rewired one
end and not both ends of the link. These added conditions prevent the resulting
network from being a random graph even in the limit p ¼ 1.
A simplified version of this model was proposed by Newman and Watts (1999).
Starting with a lattice and taking each link one by one, add another link between a
pair of vertices chosen at random with the probability p without removing the
existing one. This corresponds to the addition of Lkp new links on average to the
starting lattice, Lk being the initial number of links in the graph (Fig. 16.7).
A number of other models have been proposed to achieve the combination of
short path length with high clustering coefficient. The oldest one is the random
biased net of Rapoport (1957), in which clustering is added to a random graph by
triadic closure, i.e. the deliberate completion of connected triples of vertices to
form triangles in the network, thereby increasing the clustering coefficient. Another
attempt to model clustering was made in the 1980s by Holland and Leinhardt
(1981), using the class of network models known as exponential random graphs.
Another method for generating clustering in networks could be membership of
individuals in groups. This has been investigated by Newman et al. (2001; Newman
2003b) using so-called bipartite graph models.
16.3.2.4 Scale-Free Networks
Even with the small-world effect, the hypothesis that complex systems such as cells
or social systems are based upon components – i.e. molecules or individuals –
randomly wired together, has proven to be incomplete.
In fact, several empirical data analysis of real networks found out that for many
systems, including citation networks, the World Wide Web, the Internet, and
metabolic networks, the degree distribution approximates a power law (Price
1965; Albert et al. 1999; Faloustos et al. 1999; Broder et al. 2000).
This corresponds to a new class of network since neither of the previously
discussed networks such as random graphs and small-world models have a
power-law degree distribution. Barabási and Albert (1999) called these graphs
scale-free networks and proposed that power laws could potentially be a generic
property of many networks and that the properties of these networks can be
explained by having the graph grow dynamically rather than being static. Their
paper proposes a specific model of a growing network that generates power-law
degree distributions similar to those seen in the World Wide Web and other
networks.
Their suggested mechanism has two components: (1) the network is growing, i.e.
vertices are added continuously to it, and (2) vertices gain new edges in proportion
to the number they already have, a process that Barabási and Albert call preferential
attachment.3 Therefore, the network grows by addition of a single new vertex at
each time step, with m edges connected to it. The other end of the edge is chosen at
random with probability proportional to degree:
ki
Pð k i Þ ¼ P (16.1)
j kj
An extension to this model proposed by the same authors (Albert and Barabási
2000) follows another way of building the graph. In this model one of three events
occurs at each time step:
• With probability p, m new edges are added to the network. One end of each new
edge is connected to a node selected uniformly at random from the network and
the other to a node chosen using the preferential attachment process according to
the probability given just before.
• With probability q, m edges are rewired, meaning that a vertex i is selected at
random and one of its edges chosen at random is removed and replaced with a
new edge whose other end is connected to a vertex chosen again according to the
preferential attachment process.
• With probability 1 – p – q, a new node is added to the network. The new node
has m new edges that are connected to nodes already present in the system via
preferential attachment in the normal fashion
This model produces a degree distribution that again has a power-law tail, with
an exponent γ that depends on the parameters p, q and m, and can vary anywhere in
the range from 2 to 1.
Dorogovtsev et al. (2000) consider a variation on the preceding model applied to
directed graphs, in which preferential attachment takes place with respect only to
the incoming edges at each vertex. Each vertex has m outgoing edges, which attach
to other pre-existing vertices with attachment probability proportional only to those
vertices’ in-degree.
3
The preferential attachment mechanism has appeared in several different fields under different
names. In information science it is known as cumulative advantage (Price 1976), in sociology as
the Matthew effect (Merton 1968), and in economics as the Gibrat principle (Simon 1955).
Fig. 16.8 Scale-free network generated using the preferential attachment model implemented in
NetLogo (left) and the distribution of number of nodes per degree (right), which follows a scale-
free distribution
Another extension proposed by Krapivsky et al. (2000) explores what happens

when Π(k), the probability distribution of connectivity per node, is not linear in k.
They studied the case in which Π(k) takes the power-law form ΠðkÞ / kα
In a subsequent model, Bianconi and Barabási (2001), motivated by the Google
effect (emergence of a new hub from an existing scale-free network), introduced the
idea that some nodes are intrinsically “better” or faster growing than others. In the
model, each vertex in the network has a fixed fitness value ηi, chosen from some
distribution ρ(η), that corresponds to an intrinsic ability to compete for edges at the
expense of other vertices. Each such vertex connects to m others, the probability
of connecting to vertex i being proportional to the product of the degree and the
fitness of i:
η ki
Π ¼ Pi (16.2)
η j kj
j
The use of models like preferential attachment in an agent-based social simula-

tion context follows two motivations that are relatively distinct. On the one hand,
such models are used as initial configurations for the simulation: therefore the
construction of the scale-free network, even if it could be considered as a model
of a growing network, should rather be envisaged as an algorithm to build the initial
state of the simulation. On the other hand, it can also be seen as a subject of
research, the focus of a model, for instance in the field of modelling social network
dynamics. In this case, preferential attachment mechanisms could be included in the
model of social network evolution (if such a mechanism is relevant, of course)
(Fig. 16.8).
As the structure and the evolution of a scale-free network cannot be considered
as separate concepts, the same holds true for the topology and the overlying
interactions among components. For instance, Pastor-Satorras and Vespignani
(2001) discovered that the epidemic threshold on top of scale-free networks
converges towards 0, meaning that in scale-free topologies even weak viruses
diffuse fast within the network. In particular the dynamic aspects of real phenomena
occurring on top of scale-free networks would allow investigating several

properties of real social networks, for instance how changes in the network structure
trigger emergent phenomena by local interaction.
16.3.3 How to Distribute Agents on the Network?
This question deals with the efficiency of the model, because each link between
nodes (agents) represents an interaction, the shape of the topology, i.e. the interac-
tion space that we are interested in, characterising its emergent properties. To
exemplify, let us take the example of an opinion dynamics model including
extremist agents that have a significant impact on the global dynamics. What should
the initial state of the population be? Are the extremists homogeneously distributed
in the population, i.e. over the social network, or are they organised in dense
communities, or any case in between? This question could have an important
impact on the emerging phenomenon and is the consequence of two choices: a
modelling choice, i.e. what is the initial state of your model, and a more technical
choice, i.e. the possibility to generate the distribution wanted. With regard to the
latter, you can always generate an ad-hoc situation either using real data or building
them artificially; however, it is often the case that you wish to use a generic way to
distribute agents among a network. Usually, the starting point would be to use a
random distribution. You first generate you network and you distribute the different
agents uniformly over this network assuming that each agent’s state is independent
from its location in the network. However, you do not have to believe strongly in
any social determinism to question this independence and to wish to explore other
kinds of distributions. Two cases occur then:
You have modelling hypotheses concerning the relation between the existence
of a link between two agents and their actual state. In this case you would probably
proceed using a stochastic algorithm that enables some flexibility applying this
relation.
You are only able to describe the global state you want to attain. In this case, you
probably have to operate iterative permutations among the agents. For each permu-
tation, you have to compute the result. If it is better (i.e. closer to the state wanted)
then you keep the permutation if not you reject this permutation and keep the
previous state.
Note that the latter solution can become very expensive with regard to the
evaluation of a given state and is not guaranteed to obtain the wanted final state
in a finite time. Therefore the most prudent approach would be to take the agents’
characteristics into account when deciding on the presence of a link between two
agents. Such a solution has been proposed by Thiriot and Kant (2008) explicitly
drawing on agents’ attributes and link types for the generation of interaction
networks. In this perspective, the generated social network is a part of the modelling
hypotheses and it has been shown that the corresponding graphs differ a lot from the
ones obtained with abstract models like the ones described in this section.
16.3.4 Which Measures to Use?
Once you have your model running, especially when it deals with agents changing
their state over a (possibly evolving, irregular) social network, the first image is
usually a messy network with overlapping links and nodes changing colours over
time. Without specific efforts dedicated to the network layout (which is not a trivial
task) the global picture is not understandable at first glance. Therefore you would
want to add some indicators enabling you to better understand what happens in the
model. There are two main categories: the first one – which is currently under
development – concerns indicators linking the states of nodes with their position in
the network, while the more classical indicators of the second category help to
characterise the structure of the network only.
Some important properties associated with graphs influence the characteristics
of the dynamics taking place on the graph. This is mainly the case for the diameter,
especially dealing with diffusion processes, and the clustering coefficient, dealing
for instance with influence processes. In the particular case of a regular network,
where all nodes are equal, it can be determined depending on the connectivity and
the dimension of the graph. The average path length gives information that is
equivalent. The clustering coefficient is defined locally as, for a given node, the
rate of existing links among its neighbours compared to the number of possible
links. Local redundancy of links is therefore important as an inertial process that
can reinforce or go against an influence dynamics. In the following paragraphs we
will briefly describe the main indicators you can use to gain some insight about the
phenomena emerging in the model.
A first question you could ask is, looking at a particular dimension of the agents’
state vector, do they have a tendency to regroup or cluster according to this state or
not. Or phrased in a different way: do connected agents tend to become similar? A
useful indicator for this would be an averaged similarity measure over the network,
calculating the mean distance among all connected agents of the population.
In order to characterise the two main features of small-world networks (i.e. small
world effect and high clustering) several indicators are used. The small-world effect
is measured simply by averaging the distance among any pair of vertices in the
graph. This results in the average path length index.
Concerning the clustering, a number of indicators have been proposed. Watts
and Strogatz (1998) suggested the clustering coefficient, used classically in social
network analysis and consisting in averaging a local clustering coefficient over all
vertices of the network. The local clustering coefficient Ci of the vertex i is defined
as the number of existing links among the neighbours of the vertex i, divided by the
number of possible links among these neighbours. This quantity is 1 for a fully
connected graph but tends towards 0 for a random graph as the graph becomes
large.
The problem is that this indicator is heavily biased in favour of vertices with a
low degree due to the small number of possible links (denominator) they have.
When averaging the local clustering coefficient over all nodes without additional
weighting, this could make a huge difference in the value of C. A better way to
calculate the average probability of a pair of neighbours being connected is to count

the total number of pairs of vertices on the entire graph that have a common
neighbour and the total number of such pairs that are themselves connected, and
divide the one by the other. Newman et al. (2001) have expressed this, i.e. the
fraction of transitive triples, as:
C ¼ 3
ðnumber of triangles on the graphÞ=ðnumberof connected triples of verticesÞ
Another definition for clustering coefficient proposed by Newman (2003a)

consists in calculating the probability that the friend of your friend is also your
friend, resulting in:
C ¼ 6 ðnumber of triangles on the graphÞ=ðnumber of paths of length 2Þ
Dealing with small-world networks, Watts and Strogatz (1998) defined a network
to have high clustering if C > > Crg (this latter being the clustering coefficient for
random graphs).
Watts and Strogatz defined a network to be a small-world network if it shows
both of those properties, that is, if the mean vertex-vertex distance l is comparable
with that on a random graph (lrg) and the clustering coefficient is much greater than
that for a random graph. Walsh (1999) used this idea to define the proximity ratio:
C=Crg
μ¼ (16.3)
l=lrg
which is of order 1 on a random graph, but much greater than 1 on a network

obeying the Watts-Strogatz definition of a small-world network.
One of the most important properties of social networks is the so-called notion of
power. As a shared definition of power is still object of debate, the design of metrics
able to characterise its causes and consequences is a pressing challenge. In particular
the social network approach emphasises the concept of power as inherently relational,
i.e., determined by the network topology. Hence, the focus must be put on the relative
positions of nodes. In order to characterise such a property the concept of centrality
has emerged. The simplest centrality metric, namely the degree centrality, measures
the number of edges that connect a node to other nodes in a network. Over the years
many more complex centrality metrics have been proposed and studied, including
status score (Katz 1953), α-centrality (Bonacich and Lloyd 2001), betweenness
centrality (Freeman 1979), and several others based on the random walk, the most
famous of which is the eigenvector centrality used by Google’s PageRank algorithm
(Page et al. 1999). The temporal declination of these concepts is meaningful, and
Kossinets et al. (2008) have shown that nodes that are topologically more central are
not necessarily central from a temporal point of view, hence the concept of temporal
centrality. The temporalization of network metrics is currently a pressing scientific
challenge (Casteigts et al. 2010; Santoro et al. 2011).
16.3.5 What Kind of Network Effect Can You Anticipate?
As mentioned by Newman et al. (2006), the behaviour of dynamical systems on

networks is one of the most challenging areas of research. At the end of their paper,
Watts and Strogatz (1998) present one of the most tractable problems in the field
that is the spread of disease using the SIR (susceptible/infective/removed) epidemic
model. They measure the position of the epidemic threshold – the critical value of
the probability r of infection of a susceptible individual by an infective one at which
the disease first becomes an epidemic, spreading through the population rather than
dying out. They found a clear decline in the critical r with increasing p, the
probability of links rewiring, indicating that the small-world topology makes it
easier for the disease to spread. They also found that the typical time for the disease
to spread through the population decreases while increasing p.
One of the simplest models of the spread of non-conserved information on a
network is one in which an idea or a disease starts at a single vertex and spreads first
to all its neighbouring vertices. From these it spreads to all of their neighbours, and
so forth, until there are no accessible vertices left that have not yet been infected.
This process is known in computer science as breadth-first search. Depending on
the type of network model chosen, it could happen (this is mostly the case on real
networks) that you obtain a large number of small components plus, optionally, a
single giant component. If we simulate the spread of a rumour or disease on such
a network using breadth-first search, either we will see a small outbreak
corresponding to one of the small components, or, if there is a giant component
and we happen to start a breadth-first search within it, we will see a giant outbreak
that fills a large portion of the system. The latter is precisely what we call an
epidemic, when we are talking about disease, and the phase transition at which the
giant component appears in a graph is also a phase transition between a regime in
which epidemics are not possible and a regime in which they are. In fact, if the
disease starts its breadth-first search at a randomly chosen vertex in the network,
then the probability of seeing an epidemic is precisely equal to the fraction of the
graph occupied by the giant component. By studying, either analytically or numeri-
cally, the behaviour of various epidemiological models on networks, the authors
hope to get a better idea of how real diseases will behave, and in some cases they
have found entirely new behaviours that had not previously been observed in
epidemiological models.
Ball et al. (1997) consider networks with two levels of mixing, meaning that
each vertex in the network belongs both to the network as a whole and to one of a
specified set of subgroups (e.g. family) with different properties of spread within
the network. Disease spreading is again modelled using the SIR model. In the
model, people can be in one of three states: susceptible (S), meaning they can catch
the disease but haven’t yet; infective (I), meaning they have caught the disease and
can pass it on to others; and removed (R), meaning they have recovered from the
disease and can neither pass it on nor catch it again, or they have died. Ball et al.
(1997) found that the rapid spread of a disease within groups such as families can
lead to epidemic outbreaks in the population as a whole, even when the probability
of inter-family communication of the disease is low enough that epidemic
outbreaks normally would not be possible. The reason for this is the following. If
transmission between family members takes place readily enough that most
members of a family will contract the disease once one of them does, then we
can regard the disease as spreading on a super network in which vertices are the
families, not individuals. Roughly speaking, spread of the disease between families
will take place with n2 times the normal person-to-person probability, where n is the
number of people in a family.
An alternative approach to calculating the effect of clustering on SIR epidemics
has been presented by Keeling (1999). What Keeling’s method does is to include, in
approximate form, the effect of the short-scale structure of the network – the
clustering – but treat everything else using a standard fully mixed approximation.
Thus things like the effect of degree distributions are absent from his calculations.
But the effect of clustering is made very clear. Keeling finds that a lower fraction of
the population need to be vaccinated against a disease to prevent an epidemic if the
clustering coefficient is high.
In another paper, Pastor-Satorras and Vespignani (2001) address the behaviour of
an endemic disease model on networks with scale-free degree distributions. Their
motivation for the work was an interest in the dynamics of computer virus infections,
which is why they look at scale-free networks; computer viruses spread over the
Internet and the Internet has a scale-free form, as demonstrated by Faloutsos et al.
(1999). They used a derivation of the SIR model, the SIS (susceptible/infected/
susceptible) model, which simply considers that individuals recover with no immu-
nity to the disease and are thus immediately susceptible once they have recovered. In
their work, Pastor-Satorras and Vespignani grow networks according to the scale-free
model of Barabási and Albert (1999) and then simulate the SIS model on this network
starting with some fixed initial number of infective computers. Pastor-Satorras and
Vespignani do not find oscillations in the number of infected individuals for any
value of the independent parameter of the model. No matter what value the parameter
takes, the system is always in the epidemic regime; there is no epidemic threshold in
this system. No matter how short a time computers spend in the infective state or how
little they pass on the virus, the virus remains endemic. Moreover, the average
fraction of the population infected at any one time decreases exponentially with the
infectiousness of the disease.
Watts (2002) has looked at the behaviour of cascading processes. Unlike disease,
the spread of some kinds of information, such as rumours, fashions, or opinion,
depends not only on susceptible individuals having contacts with infective ones, but
on their having contact with such individuals in sufficient numbers to persuade
them to change their position or beliefs on an issue. People have a threshold for
adoption of trends. Each individual has a threshold t for adoption of the trend being
modelled, which is chosen at random from a specified distribution. When the
proportion of a person’s contacts that have already adopted the trend rises above
this threshold, the person will also adopt it. This model is similar to the rioting
model of Granovetter (1978). Watts gives an exact solution for his model on
random graphs for the case where initially a low density of vertices has adopted the
trend. The solution depends crucially on the presence of individuals who have very
low thresholds t. In particular, there must exist a sufficient density of individuals in
the network whose thresholds are so low that they will adopt the trend if only a
single one of their neighbours does. As Watts argues, the trend will only propagate
and cause a cascade if the density of these individuals is high enough to form a
percolating sub-graph in the network. The fundamental result of this analysis is that,
as a function of the average degree z of a vertex in the graph, there are two different
thresholds for cascading spread of information. Below z ¼ 1, no cascades happen
because the network itself has no giant component. Cascades also cease occurring
when z is quite large, the exact value depending on the chosen distribution of t. The
reason for this upper threshold is that as z becomes large, the value of t for vertices
that adopt the trend when only a single neighbour does so becomes small, and hence
there are fewer such vertices. For large enough z these vertices fail to percolate and
so cascades stop happening.
Another field of application deals with the robustness of networks. If we have
information or disease propagating through a network, how robust is the propaga-
tion to failure or removal of vertices? The Internet, for example, is a highly robust
network because there are many different paths by which information can get from
any vertex to any other. The question can also be rephrased in terms of disease. If a
certain fraction of all the people in a network are removed in some way from a
network – by immunization against disease, for instance – what effect will this have
on the spread of the disease?
Albert et al. (2000) discuss network resilience for two specific types of model
networks, random graphs and scale-free networks. The principal conclusion of the
paper is that scale-free networks are substantially more robust to the random
deletion of vertices than Erdös-Rényi random graphs, but substantially less robust
to deletion specifically targeting the vertices with the highest degrees. The mean
vertex-vertex distance in the scale-free network increases as vertices are deleted,
but it does so much more slowly than in the random graph. Similarly, the size of the
largest component goes down substantially more slowly as vertices are deleted in
the scale-free network than in the random graph. By contrast, the random graph’s
behaviour is almost identical whether one deletes the highest-degree vertices or
vertices chosen at random. Thus scale-free networks are highly robust to random
failure, but highly fragile to targeted attacks. More recent work has shown that there
are networks with even higher resilience than scale-free networks (Costa 2004;
Rozenfeld and ben-Avraham 2004; Tanizawa et al. 2005).
As a conclusion on networks effects, the existence of isolated components and
the average degree of the graph are the first important factors that play a role in the
dynamics occurring in agent-based social simulation. After that, depending on the
kind of phenomenon studied, social influence or diffusion dynamics for instance,
clustering coefficient and average path length should be considered.
16.4 Spatial Structure
16.4.1 Which Spatial Structure to Use?
Despite the increased use of social networks in agent-based simulation, in many

cases one has to take into account the spatial dimension. In this section, we will first
present some issues dealing with this point, then some models that allow the
distribution of a population of agents in a geographical space, as well as some
measures to try and qualify such distributions and their effects on the simulation.
16.4.1.1 Torus or Not?
There may also be arbitrary effects introduced by the spatial bounds or limits placed
on the phenomenon or study area. This occurs since spatial phenomena may be
unbounded or have ambiguous transition zones in reality. In the model, ignoring
spatial dependency or interaction outside the study area may create edge effects.
For example, in Schelling’s segregation model there is a higher probability to
stabilise when you have fewer neighbours (e.g. three in the corners or five on a
border) for a particular density.
The choice of spatial bounds also imposes artificial shapes on the study area that
can affect apparent spatial patterns such as the degree of clustering. A possible
solution is similar to the sensitivity analysis strategy for the modifiable areal unit
problem, or MAUP: change the limits of the study area and compare the results of
the analysis under each realization. Another possible solution is to over-bound the
study area, i.e. to deliberately model an area that encompasses the actual study area.
It is also feasible to eliminate edge effects in spatial modelling and simulation by
mapping the region to a boundless object such as a torus or sphere.
16.4.2 How to Distribute Agents on the Space?
To begin with the simplest case, let us present the Poisson process in a 2D
continuous space. Such a distribution can be used in particular to define the null
hypothesis of a totally random structure for spatial distribution that will enable us to
compare other kinds of spatial structures. In order to simulate a Poisson process of
intensity λ on a domain of surface D, we first define the number of points N to be
distributed, picking it at random from a Poisson law of parameter λD. For each point
Ai the coordinates xi and yi are therefore taken at random from a uniform law. For a
rectangular domain, it is sufficient to bound the values of x and y depending on the
studied domain. For more complex zones, the method can be the same, deleting the
points in a larger rectangle that are not part of the specific zone. Studying a
population where the number of agents N is known, we can use the same method
Fig. 16.9 Two different spatial distributions, random (left) and aggregated (right) (From Goreaud
2000)
without picking the number N at random and using the known value instead. The
corresponding process is called a binomial process (Tomppo 1986). Moreover, note
that because in a Poisson process random points can be infinitely close, the space
occupied by the represented agents is not taken into account. Some processes have
been developed to deal with this (Ripley 1977; Cressie 1993), corresponding to a
random repartition of the points under the constraint that two points have to be at
least 2R apart, where R is the specified radius of an agent. It can also be interpreted
as the random repartition of non-overlapping disks of radius R.
Practically, spatial structures are rarely totally random and it is quite frequent to
have more aggregated structures. The Neyman-Scott cluster process can be used to
simulate such more elaborated structures (Ripley 1977; Cressie 1993). The
Neyman-Scott process is built from a master Poisson process whose Nag (number
of aggregates) points are used as centres for the aggregates. Each aggregate is then
composed of a random number of points whose positions are independent and
distributed within a radius R around the centre of the aggregate (Fig. 16.9).
In order to simulate complex interaction, the Gibbs process can be useful. The
general class of Gibbs processes allows obtaining various complex structures
(Tomppo 1986). The main idea of the Gibbs process is to distribute the points of
the pattern according to attraction/repulsion relations with a range. Such a process
can be defined using a cost function f(r) which represents the cost associated to the
presence of two points in the pattern separated by a distance r. For a given point
pattern we can calculate a total cost, equal to the sum of the costs associated to each
couple of points (see Eq. 16.4 below). By definition, the lower the total cost
obtained on a pattern, the stronger is the probability for this pattern to be the result
of the Gibbs process considered. For a given r, if the cost function f(r) is positive,
the probability is low to have two points at a distance r (we could say that there is
repulsion at distance r). Conversely, if the cost function f(r) is negative, it is highly
probable to find a couple of points separated by a distance r (we could say there is
attraction at a distance r).
X
totalCost ¼ f ðdðAi ; Aj ÞÞ (16.4)
i;j
Where d is the distance between points Ai and Aj.

To obtain the most probable realization of the process, we aim to simulate
patterns whose total costs are low enough, i.e. get closer to a minimum of the
cost function. Such patterns can be obtained using an algorithm of depletions and
replacements: starting for instance from a Poisson pattern, we iterate a big number
of times a modification of this pattern, which consists in moving randomly an
arbitrarily chosen point. If this move increases the total cost, the previous configu-
ration is kept unchanged, else the new position is kept. Then, the attractions and
repulsions among points lead to a progressive reorganization of the original pattern
toward a pattern that is more aggregated or more regular. Following Tomppo
(1986), this algorithm converges quickly enough towards a realization of the
process.
In the preceding paragraphs, we were discussing the spatial distribution of
homogeneous points. What if the points are heterogeneous? In the case of hetero-
geneous patterns that do not correspond to the superposition of independent
distributions of different homogeneous groups, the preceding methodology can be
used independently for each class of points.
In order to consider a more difficult case, let us consider that now each point can
have a radius as well as a colour. More generally speaking, the variables associated
to the points are called “marks” and the following approach can be applied either to
qualitative (the colour) or quantitative (the radius) marks. We can then define
stochastic processes whose realizations are marked point patterns. These are
marked punctual processes, some mathematical objects that can generate an infinite
number of marked point patterns, all different, but sharing some common
properties, in particular the spatial structure of points and marks (Stoyan 1987;
Goulard et al. 1995).
For a marked pattern, one can consider the position of points (the pattern as such)
and the attribution of marks separately. The positioning of points can be done using
tools and methods from the preceding part. Rather than to consider probabilities of
having points separated by a distance r, we now have to reason about the joint
probability P (see Eq. 16.5) that two points at a distance r have marks m and m0 .
This function expresses neighbourhood relationships among values of a mark. The
function gM(r,m,m0 ) can moreover be linked to the number of neighbours having a
mark m0 at a distance r of a point marked m.
PðMðA1 Þ ¼ m and MðA2 Þ ¼ m0 Þ ¼ λM ðmÞ λM ðm0 ÞgM ðr; m; m0 Þ (16.5)
The hypothesis of a totally random distribution of marks can be used as null

hypothesis for marked stochastic processes. Then if, for a given case, the points of a
subpopulation SP2 are less often near the points of a subpopulation SP1 than in the
null hypothesis, we can speak about repulsion between the two subpopulations
considered. Conversely, if the points from SP2 are more numerous in the vicinity of
SP1 than in the null hypothesis we can talk about attraction of SP2 by SP1.
One can easily use a “depletions and replacements” algorithm with an associated
cost function in this particular case to generate marked point patterns of the wanted
properties with regard to inter- and intra-attraction/repulsion among the elements of
the different groups.
This situation can be seen as two different cases that could be treated using
nearly the same methods: the case where mark values are drawn at random from a
known statistical law and the case where mark values are fixed from an existing
distribution (histogram for instance) and where we have to distribute these mark
values spatially. This corresponds to a random distribution of marks and to a list of
given marks, respectively.
16.4.3 Which Measures to Use?
To analyse a spatial structure, there are several established methods from spatial
statistics to characterise the spatial structure of point patterns (Ripley 1981; Diggle
1983; Cressie 1993).
One can distinguish between methods based on quadrants, the required data for
which are the number of individuals in the quadrants, having variable position and
size (Chessel 1978), and methods based on distance, for which the distances among
points or individuals or positions are used as input. Indicators à la “Clark and
Evans” (Clark and Evans 1954) are classical examples of methods based on
distance. They calculate for each point an index of spatial structure considering
the n (fixed) closest neighbours. The average value of this index can then be
compared to the theoretical value calculated for a null hypothesis (for instance a
Poisson process) in order to characterise the spatial structure with regard to
aggregates or regularity, attraction or repulsion, etc. (Pretzsch 1993; Füldner 1995).
For a homogeneous punctual process having a density λ, Ripley (1976, 1977)
showed that it is possible to characterise the second order propriety using a function
K(r) such that λK(r) is the expected value of the number of neighbours at a distance
r of a point chosen at random from the pattern. This function is linked directly to the
function of density of a couple of points g(r) defined previously.
λKðrÞ ¼ Eðnumber of neighbours at distance <¼ r from Ai Þ
ðr
KðrÞ ¼ gðsÞ2π sds (16.6)
s¼0
For the Poisson process, which serves as null hypothesis, the expected value of
the number of neighbours at distance r of a given point from the pattern is λπr2, and
then K(r) ¼ πr2. For an aggregated process, the points have on average more
Fig. 16.10 Estimation of the

K(r) function for different
repartitions (aggregated,
poisson and regular) (From
Goreaud 2000)
neighbours than in the null hypothesis and thus K(r) > πr2. Conversely, for a
regular process, the points have on average fewer neighbours than in the null
hypothesis and K(r) < πr2 (Fig. 16.10).
In the most frequent case where we do not know the process, the function K(r)
has to be estimated with the unique known realization: the real point pattern. We
then approach the expected value of the number of neighbours around a given point
using its average on the whole set of individuals belonging to the pattern. We then
^ defined as follows:
obtain a first approximated indicator of K(r), noted as KðrÞ
XN X
^ ¼1 1
KðrÞ kij (16.7)
λ N i¼1 j6¼i
N is the number of points in the studied domain of surface D; λ ¼ N=D is the

estimator of the process density and kij takes the value 1 if the distance between
points i and j is below r and 0 otherwise.
This first estimator is unfortunately biased as it underestimates the values of K(r)
because of the points situated close to the border of the studied domain having
fewer neighbours than a given point of the process. This general problem, known as
border or edge effect and discussed in Sect. 16.3.2.1, is met in a lot of methods for
the analysis of spatial structure.
To correct the border effect, two methods are known from the literature:
• Ohser and Stoyan (Stoyan et al. 1995) propose a global correction method,
which is based on the geometrical properties of the studied domain. This method
has the advantage of being quite simple, so quick enough to calculate.
• Ripley (1977) proposes a local correction method that is based on the contribu-
tion of each point situated nearby the border. This local correction has the
advantage of being equally usable to calculate the individual indices relative
to each point of the pattern, just as those proposed by Getis and Franklin (1987).
Some works seem to show that this class of estimators is more robust than the
estimators of the Ohser and Stoyan type (Kiêu and Mora 1999).
Fig. 16.11 Estimation of the

L(r) function for different
repartitions (aggregated,
poisson and regular) (From
Goreaud 2000)
This latter solution consists in replacing the coefficient kij with the inverse of the
proportion of the perimeter of the circle Cij (centred on Ai and passing through Aj)
for the points situated near the border of the domain studied. It corresponds to an
estimation of the number of points situated at the same distance that would be
outside of this domain. Ripley (1977) shows this estimator is not biased.
totalPerimeter 2πr
kij ¼ ¼ 1 (16.8)
PerimeterInsideTheZone Cinside
An alternative to the function K(r) is the function L(r) proposed by Besag (1977)
which is easier to interpret. For a Poisson process, for any distance r, L(r) ¼ 0.
Aggregated processes and regular ones are situated under and above the x-axis,
respectively.
rffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffi
KðrÞ ^
LðrÞ ¼ ^ ¼ KðrÞ r
r estimated by LðrÞ (16.9)
π π
However appropriate such methods of punctual processes are to deal with agent-
based simulations and allow initialising and measuring and therefore characterising
such systems, agent-based simulations have an important aspect that is not
addressed by this literature: the dynamics. In many cases, the agents in such
systems will move in the environment and the spatial properties of the system is
represented more accurately considering a set of trajectories of the agents rather
than a succession of static spatial repartition. Even though punctual processes
enable to capture spatial clustering effects, when considering for example the
Boids model (Reynolds 1987) the dynamics of the flock will be overlooked.
Therefore there is a need for statistical tools that enable to characterise the
interrelations between sets of trajectories to analyze some models properly
(Fig. 16.11).
16.4.4 What Effects Can You Anticipate?
In spatial agent-based simulation, and without aiming at being exhaustive, the main
phenomena you can look at and therefore characterise depend greatly on the
population of agents in the model. Dealing with a homogeneous population of
agents, you may observe spatial clustering, i.e. the formation of groups in space,
that could be characterised by the methods presented beforehand. Introducing
heterogeneity into the population, the main question deals with the spatial mixing
of subpopulations, as present for instance in Schelling’s segregation model. Dealing
with a population of agents which may change their states, methods to characterise
diffusion processes could be used. However, one of the most interesting points in
this case does not really deal with the efficiency of the diffusion (evolution of the
number of infected agents in an epidemiological model for instance), but rather
with the characterization of the shape of such a diffusion. In this case, considering a
diffusion process that takes place, you aim at characterising the shape of the
interface rather than the global phenomenon that takes place. To this aim, fractal
analysis gives interesting hints but is not appropriate for many phenomena,
especially dynamical ones.
Finally, an important phenomenon that can occur at the macro-level of any
spatial agent-based simulation concerns the density-dependence of the results.
For an identified phenomenon (if we take for instance the segregation model of
Schelling), there will be an impact of the density of agents (either locally or
globally) on the appearance of the phenomenon. This is exactly the case of
segregation in Schelling’s model, where a very low number of empty places
(thus, a high density of agents) can freeze the system.
16.5 Conclusion
This chapter aims at presenting ways to deal with the distribution of agents in social
simulation. The two kinds of distributions considered are distribution over a graph
or social network on the one hand or spatial distribution on the other hand. While
these two cases are very common in social simulation too little effort is spent on
either the characterization or the investigation of the impact of the distribution on
the final results. The methods presented in this chapter are certainly not exhaustive
and not even pertinent to all cases but they do present a first step towards pointing
the curiosity of agent-based modellers at techniques that would definitely be useful
for social simulation.
Further Reading
The literature on dynamic aspects of social networks is rapidly developing.

For social network models and their analysis we currently recommend (Newman
et al. 2006). For spatial aspects we recommend (Diggle 1983). For more details
concerning random graphs models, we refer the interested reader to (Bollobas
2001).
References
Albert R, Barabási A-L (2000) Topology of evolving networks: local events and universality.
Phys Rev Lett 85:5234–5237
Albert R, Jeong H, Barabási A-L (1999) Diameter of the World-Wide Web. Nature 401:130–131
Albert R, Jeong H, Barabási A-L (2000) Error and attack tolerance of complex networks.
Nature 406:379–381
Ball F, Mollisson D, Scalia-Tomba G (1997) Epidemics with two levels of mixing. Ann Appl
Probab 7:46–89
Barabási A-L, Albert R (1999) Emergence of scaling in random networks. Science 286:509–512
Bender EA, Canfield ER (1978) The asymptotic number of labeled graphs with given degree
sequences. J Comb Theory A 24:296–307
Besag J (1977) Contribution to the discussion of Dr Ripley’s paper. J Roy Stat Soc B 39:193–195
Bianconi G, Barabási A-L (2001) Competition and multiscaling in evolving networks. Europhys
Lett 90:436–442
Bollobas B (2001) Random graphs, 2nd edn. Academic, New York
Bonacich P, Lloyd P (2001) Eigenvector-like measures of centrality for asymmetric relations.
Soc Networks 23(3):191–201
Broder A et al (2000) Graph structure in the web. Comput Networks 33:309–320
Callaway DS, Newman MEJ, Strogatz SH, Watts DJ (2000) Network robustness and fragility:
percolation on random graphs. Phys Rev Lett 85:5468–5471
Casteigts A, Flocchini P, Quattrociocchi W, Santoro N (2010) Time-varying graphs and dynamic
networks. arXiv: 1012.0009v3
Chessel D (1978) Description non paramétrique de la dispersion spatiale des individus d’une
espèce. In: Legay JM, Tomassone R (eds) Biométrie et écologie. Société française de
biométrie, pp 45–133
Chopard B, Droz M (1998) Cellular automata modeling of physical systems. Cambridge University Press,
Cambridge
Clark PJ, Evans FC (1954) Distance to nearest neighbor as a measure of spatial relationships in
populations. Ecology 35(4):445–453
Cohen R, Erez K, Ben-Avraham D, Havlin S (2000) Resilience of the internet to random
breakdowns. Phys Rev Lett 85:4626–4628
Costa LD (2004) Reinforcing the resilience of complex networks. Phys Rev E 69, 066127
Cressie NAC (1993) Statistics for spatial data. Wiley, New York
de Price DJS (1965) Networks of scientific papers. Science 149:510–515
de Price DJS (1976) A general theory of bibliometric and other cumulative advantage processes.
J Am Soc Inform Sci 27:292–306
Deffuant G (2006) Comparing extremism propagation patterns in continuous opinion models.
Diggle PJ (1983) Statistical analysis of spatial point patterns. Academic, New York
Dorogovtsev SN, Mendes JFF, Samukhin AN (2000) Structure of growing networks with prefer-
ential linking. Phys Rev Lett 85:4633–4636
Erdös P, Rényi A (1960) On the evolution of random graphs. Publ Math Inst Hung Acad Sci
5:17–61
Faloustos M, Faloustos O, Faloustos C (1999) On power-law relationships of the internet topology.
Comput Commun Rev 29:251–262
Flache A, Hegselmann R (2001) Do irregular grids make a difference? Relaxing the spatial
regularity assumption in cellular models of social dynamics. J Artif Soc Soc Simul 4(4).
http://www.soc.surrey.ac.uk/JASSS/4/4/6.html
Freeman LC (1979) Centrality in social networks: conceptual clarification. Soc Networks
1:215–239
Fuldner K (1995) Strukturbeschreibung von buchen-edellaubholz-Mischwäldern. Cuvillier Verlag,

Göttingen
Getis A, Franklin J (1987) Second order neighborhood analysis of mapped point patterns.
Ecology 68(3):473–477
Gong X, Xiao R (2007) Research on multi-agent simulation of epidemic news spread
characteristics. J Artif Soc Soc Simul 10(3). http://jasss.soc.surrey.ac.uk/10/3/1.html
Goreaud F (2000) Apports de l’analyse de la structure spatiale en forêt tempérée à l’étude et la
modélisation des peuplements complexes (Diploma Thesis). ENGREF (École nationale du
génie rural des eaux et forêts), Nancy
Goulard M, Pages L, Cabanettes A (1995) Marked point process: using correlation functions to
explore a spatial data set. Biom J 37(7):837–853
Granovetter M (1978) Threshold models of collective behaviour. Am J Sociol 83(6):1420–1443
Holland PW, Leinhardt S (1981) An exponential family of probability distributions for directed
graphs. J Am Stat Assoc 76:33–65
Holme P, Grönlund A (2005) Modelling the dynamics of youth subcultures. J Artif Soc Soc Simul
Huet S, Edwards M, Deffuant G (2007) Taking into account the variations of neighbourhood sizes
in the mean-field approximation of the threshold model on a random network. J Artif Soc Soc
IMAGES (2004) Improving agri-environmental policies: a simulation approach to the role of the
cognitive properties of farmers and institutions (Final report of the FAIR3 CT 2092 – IMAGES
project), http://wwwlisc.clermont.cemagref.fr/ImagesProject/FinalReport/final%20report%
20V2last.pdf
Jager W, Amblard F (2005) Uniformity, bipolarisation and pluriformity captured as generic
stylised behaviour with an agent-based simulation model of attitude change. Comput Math
Organ Theory 10:295–303
Jeong H et al (2002) Evolution of the social network of scientific collaborations. Physica A 311:
590–614
Katz L (1953) A new status derived from sociometric analysis. Psychometrika 18:39–43
Keeling MJ (1999) The effects of local spatial structure on epidemiological invasion. Proc R Soc
London B 266:859–867
Kieu K, Mora M (1999) Estimating the reduced moments of a random measure. Adv Appl Probab
Sect Stoch Geom Stat Appl 31(1):48–62
Kossinets G, Kleinberg J, Watts D (2008) The structure of information pathways in a social
communication network. In: Li Y, Liu B, Sarawagi S (eds) proceedings of the 14th ACM
SIGKDD international conference on knowledge discovery and data mining (KDD’08). ACM,
Kottonau J, Pahl-Wostl C (2004) Simulating political attitudes and voting behaviour. J Artif Soc
Soc Simul 7(4). http://jasss.soc.surrey.ac.uk/7/4/6.html
Krapivsky PL, Redner S, Leyvraz F (2000) Connectivity of growing random networks. Phys Rev
Lett 85:4629–4632
Luczak T (1992) Sparse random graphs with a given degree sequence. In: Frieze AM, Luczak T
(eds) Proceedings of the symposium on random graphs, Poznan 1989. Wiley, New York,
pp 165–182
Lusseau D (2003) The emergent properties of a dolphin social network. Proc R Soc London B 270(2):
S186–S188
Merton RK (1968) The Matthew effect in science. Science 159:56–63
Meyer M, Lorscheid I, Troitzsch K (2009) The development of social simulation as reflected in the
first ten years of JASSS: a citation and co-citation analysis. J Artif Soc Soc Simul 12(4). http://
Molloy M, Reed B (1995) A critical point for random graphs with a given degree sequence.
Random Struct Algor 6:161–179
Newman MEJ (2001) The structure of scientific collaboration networks. PNAS 98(2):404–409
Newman MEJ (2003a) Ego-centered networks and the ripple effect. Soc Networks 25:83–95
Newman MEJ (2003b) Properties of highly clustered networks. Phys Rev E 68, 016128
Newman MEJ (2004) Co-authorship networks and patterns of scientific collaboration. PNAS
101:5200–5205
Newman MEJ, Watts DJ (1999) Scaling and percolation in the small-world network model.
Phys Rev E 60:7332–7342
Newman MEJ, Strogatz SH, Watts DJ (2001) Random graphs with arbitrary degree distributions
and their applications. Phys Rev E 64, 026118
Newman MEJ, Barabási A-L, Watts DJ (2006) The structure and dynamics of networks.
Page L, Brin S, Motwani R, Winograd T (1999) The pageRank citation ranking: bringing order to
the web (Technical report, 1999–66). Stanford InfoLab, Stanford University, http://ilpubs.
stanford.edu:8090/422/1/1999-66.pdf
Pastor-Satorras R, Vespignani A (2001) Epidemic spreading in scale-free networks. Phys Rev Lett
86:3200–3203
Pretzsch H (1993) Analyse und reproduktion räumlicher bestandesstrukturen: versuche mit dem
strukturgenerator STRUGEN Writings from the forestry faculty of the University of Gottingen
954 and NFV, 114
Pujol JM, Flache A, Delgado J, Sanguesa R (2005) How can social networks ever become
complex? Modelling the emergence of complex networks from local social exchanges.
Rapoport A (1957) Contribution to the theory of random and biased nets. Bull Math Biophys
19:257–277
Reynolds CW (1987) Flocks, herds, and schools: a distributed behavioral model. Comput Graph
21(4):25–34
Ripley BD (1976) The second order analysis of stationary point process. J Appl Probab
13:255–266
Ripley BD (1977) Modelling spatial patterns. J R Stat Soc B 39:172–212
Ripley BD (1981) Spatial statistics. Wiley, New York
Rozenfeld HD, ben-Avraham D (2004) Designer nets from local strategies. Phys Rev E 70, 056107
Santoro N, Quattrociocchi W, Flocchini P, Casteigts A, Amblard F (2011) Time-varying graphs
and social network analysis: temporal indicators and metrics. In: Proceedings of SNAMAS
2011, pp 33–38
Schelling T (1971) Dynamic models of segregation. J Math Sociol 1(1):143–186
Simon HA (1955) On a class of skew distribution functions. Biometrika 42:425–440
Solomonoff R, Rapoport A (1951) Connectivity of random nets. Bull Math Biophys 13:107–117
Stocker R, Green DG, Newth D (2001) Consensus and cohesion in simulated social networks.
Stocker R, Cornforth D, Bossomaier TRJ (2002) Network structures and agreement in social
network simulations. J Artif Soc Soc Simul 5(4). http://jasss.soc.surrey.ac.uk/5/4/3.html
Stoyan D (1987) Statistical analysis of spatial point process: a soft-core model and cross
correlations of marks. Biom J 29:971–980
Stoyan D, Kendall WS, Mecke J (1995) Stochastic geometry and its applications. Wiley,
New York
Tanizawa T, Paul G, Cohen R, Havlin S, Stanley HE (2005) Optimization of network robustness to
waves of targeted and random attacks. Phys Rev E 71, 047101
Thiriot S, Kant J-D (2008) Generate country-scale networks of interaction from scattered statistics.
In: Squazzioni F (ed) Proceedings of the 5th conference of the European social simulation
association, ESSA ‘08, Brescia
Tomppo E (1986) Models and methods for analysing spatial patterns of trees, vol 138,
Communicationes instituti forestalis fenniae. The Finnish Forest Research Institute, Helsinki
Walsh T (1999) Search in a small world. In: Dean T (ed) Proceedings of the 16th international joint
conference on artificial intelligence, IJCAI ’99. Morgan Kauffman, San Francisco,
pp 1172–1177
Watts DJ (1999) Small worlds: the dynamics of networks between order and randomness.
Watts DJ (2002) A simple model of global cascades on random networks. PNAS 99:5766–5771
Watts DJ, Strogatz SH (1998) Collective dynamics of small world networks. Nature 393:440–442
Wolfram S (1986) Theory and applications of cellular automata. World Scientific, Singapore
Chapter 17
Learning
Michael W. Macy, Stephen Benard, and Andreas Flache
Why Read This Chapter? To understand the properties of various individual or

collective learning algorithms and be able to implement them within an agent
(where evolution is a particular kind of collective learning).
Abstract Learning and evolution are adaptive or “backwards-looking” models of
social and biological systems. Learning changes the probability distribution of traits
within an individual through direct and vicarious reinforcement, while evolution
changes the probability distribution of traits within a population through reproduc-
tion and selection. Compared to forward-looking models of rational calculation that
identify equilibrium outcomes, adaptive models pose fewer cognitive requirements
and reveal both equilibrium and out-of-equilibrium dynamics. However, they are
also less general than analytical models and require relatively stable environments.
In this chapter, we review the conceptual and practical foundations of several
approaches to models of learning that offer powerful tools for modeling social
processes. These include the Bush-Mosteller stochastic learning model, the Roth-
Erev matching model, feed-forward and attractor neural networks, and belief
learning. Evolutionary approaches include replicator dynamics and genetic
algorithms. A unifying theme is showing how complex patterns can arise from
relatively simple adaptive rules.
M.W. Macy (*)

Department of Information Science, Cornell University, Ithaca NY 14853, USA
S. Benard
Department of Sociology, Indiana University, 744 Ballantine Hall, Bloomington IN 47405
A. Flache
Department of Sociology, ICS, University of Groningen, Faculty of Behavioral and Social
Sciences, NL-9712 TG Groningen, The Netherlands

432 M.W. Macy et al.
17.1 Introduction
Evolution and learning are basic explanatory mechanisms for consequentialist

theories of adaptive self-organization in complex systems.1 These theories are
consequentialist in that behavioral traits are selected by their outcomes. Positive
outcomes increase the probability that the associated trait will be repeated (in
learning theory) or reproduced (in evolutionary theory), while negative outcomes
reduce it. Explanatory outcomes might be rewards and punishments (in learning
theory), survival and reproduction (in evolutionary models), systemic requisites (in
functionalism), equilibrium payoffs (in game theory), or the interests of a dominant
class (in conflict theory).
An obvious problem in consequentialist models is that the explanatory logic runs
in the opposite direction from the temporal ordering of events. Behavioral traits are
the explanandum and their outcomes the explanans. This explanatory strategy
collapses into teleology unless mechanisms can be identified that bridge the tem-
poral gap. While expected utility theory and game theory posit a forward-looking
and analytic causal mechanism, learning and evolution provide a backward-looking
and experiential link. In everyday life, decisions are often highly routine, with little
conscious deliberation. These routines can take the form of social norms, protocols,
habits, traditions, and rituals. Learning and evolution explain how these routines
emerge, proliferate, and change in the course of consequential social interaction,
based on experience instead of calculation. In these models, repetition, not predic-
tion, brings the future to bear on the present, by recycling the lessons of the past.
Through repeated exposure to a recurrent problem, the consequences of alternative
courses of action can be iteratively explored, by the individual actor (learning) or by
a population (evolution).
Backward-looking rationality is based on rules rather than choices (Vanberg
1994). A choice is an instrumental, case-specific comparison of alternative courses
of action, while rules are behavioral routines that provide standard solutions to
recurrent problems. Rules can take the form of strategies, norms, customs, habits,
morals, conventions, traditions, rituals, or heuristics. Rule-based decision-making
is backward-looking in that the link between outcomes and the actions that produce
them runs backward in time. The outcomes that explain the actions are not those the
action will produce in the future, they are the outcomes that were previously
experienced when the rule was followed in the past.
Learning alters the probability distribution of behavioral traits within a given
individual, through processes of direct and vicarious reinforcement. Evolution
alters the frequency distribution of traits within a population, through processes
of reproduction and selection. Whether selection operates at the individual or
population level, the units of adaptation need not be limited to human actors but
may include larger entities such as firms or organizations that adapt their behavior
1
Much of the material in this chapter has been previously published in (Macy 1996, 1997, 1998,
2004; Macy and Flache 2002; Flache and Macy 2002).
17 Learning 433
in response to environmental feedback. Nor is evolutionary adaptation limited to

genetic propagation. In cultural evolution, norms, customs, conventions, and rituals
propagate via role modeling, occupational training, social influence, imitation, and
persuasion. For example, a firm’s problem-solving strategies improve over time
through exposure to recurrent choices, under the relentless selection pressure of
market competition. Sub-optimal routines are removed from the repertoires of
actors by learning and imitation, and any residuals are removed from the population
by bankruptcy and takeover. The outcomes may not be optimal, but we are often
left with well-crafted routines that make their bearers look much more calculating
than they really are (or need to be), like a veteran outfielder who catches a fly ball as
if she had computed its trajectory.
17.2 Evolution
Selection pressures influence the probability that particular traits will be replicated,
in the course of competition for scarce resources (ecological selection) or competi-
tion for a mate (sexual selection). Although evolution is often equated with ecolog-
ical selection, sexual selection is at least as important. By building on partial
solutions rather than discarding them, genetic crossover in sexual reproduction
can exponentially increase the rate at which a species can explore an adaptive
landscape, compared to reliance on trial and error. Paradoxically, sexual selection
can sometimes inhibit ecological adaptation, especially among males. Gender
differences in parental investment cause females to be choosier about mates and
thus sexual selection to be more pronounced in males. An example is the peacock’s
large and cumbersome tail, which attracts the attention of peahens (who are
relatively drab) as well as predators. Sexually selected traits tend to become
exaggerated as males trap one another in an arms race to see who can have the
largest antlers or to be bravest in battle.
Selection pressures can operate at multiple levels in a nested hierarchy, from
groups of individuals with similar traits, down to individual carriers of those traits,
down to the traits themselves. Evolution through group selection was advanced
by Wynne-Edwards (1962, 1986) as a solution to one of evolution’s persistent
puzzles – the viability of altruism in the face of egoistic ecological counter
pressures. Pro-social in-group behavior confers a collective advantage over rival
groups of rugged individualists. However, the theory was later dismissed by
Williams in Adaptation and Natural Selection (1966), which showed that
between-group variation gets swamped by within-group variation as group size
increases. Moreover, group selection can depend on differential rates of
extinction, with no plausible mechanism for the whole-cloth replication of
successful groups.
Sexual selection suggests a more plausible explanation for the persistence of
altruistic behaviors that reduce the chances of ecological selection. Contrary to
Herbert Spencer’s infamous view of evolution as “survival of the fittest,” generosity
can flourish even when these traits are ecologically disadvantageous, by attracting
females who have evolved a preference for “romantic” males who are ready to
sacrifice for their partner. Traits that reduce the ecological fitness of an individual
carrier can also flourish if the trait increases the selection chances of other
individuals with that trait. Hamilton (1964) introduced this gene-centric theory of
kin altruism, later popularized by Dawkins’ in the Selfish Gene (1976).
Allison (1992) extended the theory to benevolence based on cultural relatedness,
such as geographical proximity or a shared cultural marker. This may explain why
gene-culture co-evolution seems to favor a tendency to associate with those who are
similar, to differentiate from “outsiders”, and to defend the in-group against social
trespass with the emotional ferocity of parents defending their offspring.
This model also shows how evolutionary principles initially developed to
explain biological adaptation can be extended to explain social and cultural change.
Prominent examples include the evolution of languages, religions, laws, organi-
zations, and institutions. This approach has a long and checkered history. Social
Darwinism is a discredited nineteenth century theory that used biological principles
as analogs for social processes such as market competition and colonial domination.
Many sociologists still reject all theories of social or cultural evolution, along with
biological explanations of human behavior, which they associate with racist and
elitist theories of “survival of the fittest”. Others, like the socio-biologist E. O.
Wilson (1988, p. 167), believe “genes hold culture on a leash”, leaving little room
for cultural evolution to modify the products of natural selection. Similarly, evolu-
tionary psychologists like Cosmides and Tooby search for the historical origins of
human behavior as the product of ancestral natural selection rather than on-going
social or cultural evolution.
In contrast, a growing number of sociologists and economists are exploring the
possibility that human behaviors and institutions may be heavily influenced by
processes of social and cultural selection that are independent of biological
imperatives. These include DiMaggio and Powell (the new institutional sociology),
Nelson and Winter (evolutionary economics), and Hannan and Freeman (organiza-
tional ecology).
One particularly compelling application is the explanation of cultural diversity.
In biological evolution, speciation occurs when geographic separation allows
populations to evolve in different directions to the point that individuals from
each group can no longer mate. Speciation implies that all life has evolved from a
very small number of common ancestors, perhaps only one. The theory has been
applied to the evolution of myriad Indo-European languages that are mutually
incomprehensible despite having a common ancestor. In socio-cultural models,
speciation operates through homophily (attraction to those who are similar), xeno-
phobia (aversion to those who are different), and influence (the tendency to become
more similar to those to whom we are attracted and to differentiate from those we
despise).
Critics counter that socio-cultural evolutionists have failed to identify any under-
lying replicative device equivalent to the gene. Dawkins has proposed the “meme”
as the unit of cultural evolution but there is as yet no evidence that these exist. Yet
17 Learning 435
Charles Darwin developed the theory of natural selection without knowing that
phenotypes are coded genetically in DNA. Perhaps the secrets of cultural evolution
are waiting to be unlocked by impending breakthroughs in cognitive psychology.
The boundary between learning and evolution becomes muddied by a hybrid
mechanism, often characterized as “cultural evolution.” In cultural evolution,
norms, customs, conventions, and rituals propagate via role modeling, occupational
training, social influence, and imitation. Cultural evolution resembles learning in
that the rules are soft-wired and can therefore be changed without replacing the
carrier. Cultural evolution also resembles biological evolution in that rules used by
successful carriers are more likely to propagate to other members of the population.
For example, norms can jump from one organism to another by imitation (Dawkins
1976; Durham 1992; Boyd and Richerson 1985; Lopreato 1990). A successful norm
is one that can cause its carrier to act in ways that increase the chances that the norm
will be adopted by others. Cultural evolution can also be driven by social learning
(Bandura 1977) in which individuals respond to the effects of vicarious reinforce-
ment. Social learning and role-modeling can provide an efficient shortcut past the
hard lessons of direct experience.
Imitation of successful role models is the principal rationale for modeling cultural
evolution as an analog of natural selection (Boyd and Richerson 1985; Dawkins
1976). However, social influence differs decisively from sociobiological adaptation.
Softwired rules can spread without replacement of their carriers, which means that
reproductive fitness loses its privileged position as the criteria for replication. While
“imitation of the fittest” is a reasonable specification of cultural selection pressures,
it is clearly not the only possibility. Replication of hardwired rules may be a
misleading model for cultural evolution, and researchers need to be cautious in
using Darwinian analogs as templates for modeling the diffusion of cultural rules. In
cultural models of “imitation of the fittest,” actors must not only know which actor is
most successful, they must also know the underlying strategy that is responsible for
that success. Yet successful actors may not be willing to share this information. For
very simple strategies, it may be sufficient to observe successful behaviors. How-
ever, conditional strategies based on “if-then” rules cannot always be deduced from
systematic observation. Researchers should therefore exercise caution in using
biological models based on Darwinian principles to model cultural evolution,
which is a hybrid of the ideal types of evolution and learning.
17.3 Learning
The most elementary principle of learning is simple reinforcement. Thorndike

(1898) first formulated the theory of reinforcement as the “Law of Effect,” based
on the principle that “pleasure stamps in, pain stamps out.” If a behavioral response
has a favorable outcome, the neural pathways that triggered the behavior are
strengthened, which “loads the dice in favor of those of its performances which
make for the most permanent interests of the brain’s owner” (James 1981,
p. 143). This connectionist theory anticipates the error back-propagation used in

contemporary neural networks (Rumelhart and McClelland 1988). These models
show how highly complex behavioral responses can be acquired through repeated
exposure to a problem.
Reinforcement theory relaxes three key behavioral assumptions in models of
forward-looking rationality:
1. Propinquity replaces causality as the link between choices and payoffs.
2. Reward and punishment replace utility as the motivation for choice.
3. Melioration replaces optimization as the basis for the distribution of choices over
time.
We consider each of these in turn.
1. Propinquity, not causality. Compared to forward-looking calculation, the Law
of Effect imposes a lighter cognitive load on decision makers by assuming
experiential induction rather than logical deduction. Players explore the likely
consequences of alternative choices and develop preferences for those
associated with better outcomes, even though the association may be coincident,
“superstitious,” or causally spurious. The outcomes that matter are those that
have already occurred, not those that an analytical actor might predict in the
future. Anticipated outcomes are but the consciously projected distillations of
prior exposure to a recurring problem. Research using fMRI supports the view
that purposive assessment of means and ends can take place after decisions are
made, suggesting that “rational choice” may be not so much a theory of decision
but a theory of how decisions are rationalized to self and others.
Reinforcement learning applies to both intended and unintended consequences
of action. Because repetition, not foresight, links payoffs back to the choices that
produce them, learning models need not assume that the payoffs are the intended
consequences of action. Thus, the models can be applied to expressive behaviors
that lack a deliberate or instrumental motive. Frank’s (1988) evolutionary model
of trust and commitment formalizes the backward-looking rationality of
emotions like vengeance and sympathy. An angry or frightened actor may not
be capable of deliberate and sober optimization of self-interest, yet the response
to the stimulus has consequences for the individual, and these in turn can modify
the probability that the associated behavior will be repeated.
2. Reward and punishment, not utility. Learning theory differs from expected
utility theory in positing two distinct cognitive mechanisms that guide decisions
toward better outcomes, approach (driven by reward) and avoidance (driven by
punishment). The distinction means that aspiration levels are very important for
learning theory. The effect of an outcome depends decisively on whether it is
coded as gain or loss, satisfactory or unsatisfactory, pleasant or aversive.
3. Melioration, not optimization. Melioration refers to suboptimal gradient climbing
when confronted with what Herrnstein and Drazin (1991) call “distributed
choice” across recurrent decisions. A good example of distributed choice is the
decision whether to cooperate in an iterated Prisoner’s Dilemma game. Suppose
17 Learning 437
each side is satisfied when the partner cooperates and dissatisfied when the
partner defects. Melioration implies a tendency to repeat choices with satisfactory
outcomes even if other choices have higher utility, a behavioral tendency March
and Simon (1958) call “satisficing.” In contrast, unsatisfactory outcomes induce
searching for alternative outcomes, including a tendency to revisit alternative
choices whose outcomes are even worse, a pattern we call “dissatisficing.” While
satisficing is suboptimal when judged by conventional game-theoretic criteria, it
may be more effective in leading actors out of a suboptimal equilibrium than if
they were to use more sophisticated decision rules, such as “testing the waters” to
see if they could occasionally get away with cheating. Gradient search is highly
path dependent and not very good at backing out of evolutionary cul de sacs.
Course correction can sometimes steer adaptive individuals to globally optimal
solutions, making simple gradient climbers look much smarter than they need to
be. Often, however, adaptive actors get stuck in local optima. Both reinforcement
and reproduction are biased toward better strategies, but they carry no guarantee
of finding the highest peak on the adaptive landscape, however relentless the
search. Thus, learning theory can be usefully applied to the equilibrium selection
problem in game theory. In repeated games (such as an on-going Prisoner’s
Dilemma), there is often an indefinitely large number of analytic equilibria.
However, not all these equilibria are learnable, either by individuals (via rein-
forcement) or by populations (via evolution). Learning theory has also been
used to identify a fundamental solution concept for these games stochastic
collusion based on a random walk from a self-limiting non-cooperative
equilibrium into a self-reinforcing cooperative equilibrium (Macy and Flache
2002).
17.4 Modeling Evolution
Replicator dynamics are the most widely used model of evolutionary selection
(Taylor and Jonker 1978). In these models, the frequency of a strategy changes from
generation to generation as a monotonic function of its “payoff advantage,” defined
in terms of the difference between the average payoff of that strategy and the
average payoff in the population as a whole. The more successful a strategy is on
average, the more frequent it tends to be in the next generation.
Replicator dynamics typically assume that in every generation every population
member encounters every other member exactly once, and replication is based on
the outcome of this interaction relative to the payoff earned by all other members of
the population. However, in natural settings, actors are unlikely to interact with or
have information about the relative success of every member of a large population.
The mechanism can also be implemented based on local interaction (limited to
network “neighbors”) and local replication (neighbors compete only with one
another for offspring).
The outcomes of replicator dynamics depend on the initial distribution of

strategies, since the performance of any given strategy will depend on its effective-
ness in interaction with other strategies. For example, aggressive strategies perform
much better in a population that is accommodating than one that is equally
aggressive. It is also not possible for replicator dynamics to invent new strategies
that were not present at the outset.
These limitations are minimized by using genetic algorithms. The genetic
algorithm was proposed by Holland (1975) as a problem-solving device, modeled
after the recursive system in natural ecologies. The algorithm provides a simple but
elegant way to write a computer program that can improve through experience. The
program consists of a string of symbols that carry machine instructions. The
symbols are often binary digits called “bits” with values of 0 and 1. The string is
analogous to a chromosome containing multiple genes. The analog of the gene is a
bit or combination of bits that comprises a specific instruction. The values of the
bits and bit-combinations are analogous to the alleles of the gene. A one-bit gene
has two alleles (0 and 1), a two-bit gene has four alleles (00, 01, 10, and 11), and so
on. The number of bits in a gene depends on the instruction. An instruction to go left
or right requires only a single bit. However, an instruction to go left, right, up, or
down requires two bits. When the gene’s instructions are followed, there is some
performance evaluation that measures the program’s reproductive fitness relative to
other programs in a computational ecology. Relative fitness determines the proba-
bility that each strategy will propagate. Propagation occurs when two mated
programs recombine through processes like “crossover” and “inversion.” In cross-
over, the mated programs (or strings) are randomly split and the “left” half of one
string is combined with the “right” half of the other, and vice versa, creating two
new strings. If two different protocols are each effective, but in different ways,
crossover allows them to create an entirely new strategy that may combine the best
abilities of each parent, making it superior to either. If so, then the new rule may go
on to eventually displace both parent rules in the population of strategies.
In addition, the new strings contain random copying errors. These mutations
continually refresh the heterogeneity of the population, in the face of selection
pressures that tend to reduce it.
To illustrate, consider the eight-bit string 10011010 mated with 11000101. (The
typefaces might represent gender, although the algorithm does not require sexual
reproduction.) Each bit could be a specific gene, such as whether to trust a partner
under eight different conditions (Macy and Skvoretz 1998). In mating, the two
parent strings are randomly broken, say after the third gene. The two offspring
would then be 10000101 and 11011010. However, a chance copying error on the
last gene might make the second child a mutant, with 11011011.
At the end of each generation, each individual’s probability of mating is a
monotonic (often linear) function of relative performance during that generation,
based on stochastic sampling (Goldberg 1989).
17 Learning 439
Pij ¼ P
N
Fi
for j ¼ 1 to N; j 6¼ i (17.1)
Fn
n¼1
where Pij is the probability that j is mated with i, Fi is i’s “fitness” (or cumulative
payoff over all previous rounds in that generation), and N is the size of the
population. If the best strategy had only a small performance edge over the worst,
it had only a small edge in the race to reproduce. With stochastic sampling, each
individual, even the least fit, selects a mate from the fitness-weighted pool of
eligibles. In each pairing, the two parents combined their chromosomes to create
a single offspring that replaces the less-fit parent. The two chromosomes are
combined through crossover.
17.5 Learning Models
The need for a cognitive alternative to evolutionary models is reflected in a growing

number of formal learning-theoretic models of behavior (Macy 1991; Roth and
Erev 1995; Fudenberg and Levine 1998; Young 1998; Cohen et. al. 2001). In
general form, learning models consist of a probabilistic decision rule and a learning
algorithm in which outcomes are evaluated relative to an aspiration level, and the
corresponding decision rules are updated accordingly.
All stochastic learning models share two important principles, the law of effect
and probabilistic decision-making (Macy 1989, 1991; Börgers and Sarin 1996;
1997; Roth and Erev 1995; Erev and Roth 1998; Erev et al. 1999; for more
references cf. Erev et al. 1999). The law of effect implies that the propensity of
an action increases if it is associated with a positively evaluated outcome and it
declines if the outcome is negatively evaluated. Probabilistic choice means that
actors hold a propensity qX for every action X. The probability pX to choose action
X then increases in the magnitude of the propensity for X relative to the propensities
for the other actions.
Whether an outcome is evaluated as positive or negative depends on the evalua-
tion function. An outcome is positive if it exceeds the actor’s aspirations. There are
basically three substantively different approaches for modeling the aspiration level,
fixed interior aspiration, fixed zero aspiration and moving average aspiration. Fixed
interior aspiration assumes that some payoffs are below the aspiration level and are
evaluated negatively, while other payoffs are above the aspiration level and are
evaluated positively (e.g. Macy 1989, 1991; Fudenberg and Levine 1998). The
fixed zero aspiration approach also fixes the aspiration level, but it does so at the
minimum possible payoff (Roth and Erev 1995; Börgers and Sarin 1997; Erev and
Roth 1998). In other words: in the fixed zero aspiration model every payoff
is deemed “good enough” to increase or at least not reduce the corresponding
propensity, but higher payoffs increase propensities more than lower payoffs do.
Finally, moving average aspiration models assume that the aspiration level
approaches the average of the payoffs experienced recently, so that players get used
to whatever outcome they may experience often enough (Macy and Flache 2002;
Börgers and Sarin 1997; Erev and Rapoport 1998; Erev et al. 1999). Clearly, these
assumptions have profound effects on model dynamics.
17.5.1 Bush-Mosteller Stochastic Learning Model
One of the simplest models of reinforcement learning is the Bush-Mosteller model

(Bush and Mosteller 1950). The Bush-Mosteller stochastic learning algorithm
updates probabilities following an action a as follows:
8
< pa;t þ ð1 pa;t Þ l sa;t , sa;t 0
pa;tþ1 ¼ ; (17.2)
:
pa;t þ pa;t l sa;t , sa;t < 0
In Eq. 17.2, pa,t is the probability of action a at time t and sa;t is a positive or
negative stimulus (0 jsj 1). The change in the probability for the action
not taken, b, obtains from the constraint that probabilities always sum to one, i.e.
pb;tþ1 ¼ 1 pa;tþ1 . The parameter l is a constant (0 < l < 1) that scales the
learning rate. With l 0, learning is very slow, and with l 1, the model
approximates a “win-stay, lose-shift” strategy (Catania 1992).
For any value of l, Eq. 17.2 implies a decreasing effect of reward as the
associated propensity approaches unity, but an increasing effect of punishment.
Similarly, as the propensity approaches zero, there is a decreasing effect of punish-
ment and a growing effect of reward. This constrains probabilities to approach
asymptotically their natural limits.
17.5.2 The Roth-Erev Matching Model
Roth and Erev (Roth and Erev 1995; Erev and Roth 1998; Erev et al. 1999) have
proposed a learning-theoretic alternative to the Bush-Mosteller formulation. Their
model draws on the “matching law” which holds that adaptive actors will choose
between alternatives in a ratio that matches the ratio of reward. Like the Bush-
Mosteller model, the Roth-Erev payoff matching model implements the three basic
principles that distinguish learning from utility theory – experiential induction
(vs. logical deduction), reward and punishment (vs. utility), and melioration
(vs. optimization). The similarity in substantive assumptions makes it tempting to
assume that the two models are mathematically equivalent, or if not, that they
nevertheless give equivalent solutions.
17 Learning 441
Flache and Macy (2002) show that Bush-Mosteller and Roth-Erev are special
cases of a more general learning model and identify important differences between
the two specifications. Each specification implements reinforcement learning in
different ways, and with different results. Roth and Erev (1995; Erev and Roth
1998) propose a baseline model of reinforcement learning with a fixed zero
reference point. The law of effect is implemented such that the propensity
for action X is simply the sum of all payoffs a player ever experienced when
playing X. The probability to choose action X at time t is then the propensity for
X divided by the sum of all action propensities at time t. The sum of the propensities
increases over time, such that payoffs have decreasing effects on choice
probabilities. However, this also undermines the law of effect. Suppose, after
some time a new action is carried out and yields a higher payoff than every other
action experienced before. The probability of repetition of this action will never-
theless be negligible, because its recent payoff is small in comparison with the
accumulated payoffs stored in the propensities for the other actions. As a conse-
quence, the baseline model of Roth and Erev (1995) fails to identify particular
results, because it has the tendency to lock the learning dynamics into any outcome
that occurs sufficiently often early on. Roth and Erev amend this problem by
introducing a “forgetting parameter” that keeps propensities low relative to recent
payoffs. With this, they increase the sensitivity of the model to recent
reinforcement.
Roth and Erev used a variant of this baseline model to estimate globally
applicable parameters from data collected across a variety of human subject
experiments. They concluded that “low rationality” models of reinforcement
learning may often provide a more accurate prediction than forward looking
models. Like the Bush-Mosteller, the Roth-Erev model is stochastic, but the
probabilities are not equivalent to propensities. The propensity q for action a at
time T is the sum of all stimuli sa a player has ever received when playing a:
X
T
qa;T ¼ sa;t ; (17.3)
t¼1
Roth and Erev then use a “probabilistic choice rule” to translate propensities into
probabilities. The probability pa of action a at time t þ 1 is the propensity for a
divided by the sum of the propensities at time t:
qa;t
pa;tþ1 ¼ ; a 6¼ b (17.4)
qa;t þ qb;t
where a and b represent binary choices. Following action a, the associated propen-
sity qa increases if the payoff is positive relative to aspirations (by increasing the
numerator in Eq. 17.4) and decreases if negative. The propensity for b remains
constant, but the probability of b declines (by increasing the denominator in the
equivalent expression for pb,tþ1).
Fig. 17.1 An example feed-forward network
17.5.3 Artificial Neural Networks
Bush-Mosteller and Roth-Erev are very simple learning models that allow an actor
to identify strategies that generally have more satisfactory outcomes. However, the
actor cannot learn the conditions in which a strategy is more or less effective.
Artificial neural nets add perception to reinforcement, so that actors can learn
conditional strategies.
An artificial neural network is a simple type of self-programmable learning
device based on parallel distributed processing (Rumelhart and McClelland
1988). Like genetic algorithms, neural nets have a biological analog, in this case,
the nerve systems of living organisms. In elementary form, the device consists of a
web of neuron-like nodes (or neurodes) that fire when triggered by impulses of
sufficient strength, and in turn stimulate other nodes when fired. The magnitude of
an impulse depends on the strength of the connection (or “synapses”) between the
two neurodes. The network learns by modifying these path coefficients, usually in
response to environmental feedback about its performance.
There are two broad classes of neural networks that are most relevant to social
scientists, feed-forward networks and attractor networks. Feed-forward networks
consist of four types of nodes, usually arranged in layers, as illustrated in Fig. 17.1.
The most straightforward are input (sensory) and output (response) nodes. Input
nodes are triggered by stimuli from the environment. Output nodes, in turn, trigger
action by the organism on the environment.
The other two types of nodes are less intuitive. Intermediate (or “hidden”) nodes
link sensory and response nodes so as to increase the combinations of multiple
stimuli that can be differentiated. Figure 17.1 shows a network with a single layer
17 Learning 443
containing two hidden nodes, H1 and H2. The number of hidden layers and the
number of hidden nodes in each layer vary with the complexity of the stimulus
patterns the network must learn to recognize and the complexity of the responses
the network must learn to perform. Unlike sensory and response nodes, hidden
nodes have no direct contact with the environment, hence their name.
Bias nodes are a type of hidden node that has no inputs. Instead, a bias node
continuously fires, creating a predisposition toward excitation or inhibition in the
nodes it stimulates, depending on the size and sign of the weights on the pathway
from the bias node to the nodes it influences. If the path weight is positive, the bias
is toward excitation, and if the path is negative, the bias is toward inhibition. The
weighted paths from the bias node to other hidden and output nodes correspond to
the activation thresholds for these nodes.
A feed-forward network involves two processes – action (firing of the output
nodes) and learning (error correction). The action phase consists of the forward
propagation of influence (either excitation or inhibition) from the input nodes to the
hidden nodes to the output nodes. The influence of a node depends on the state of
the node and the weight of the neural pathway to a node further forward in the
network. The learning phase consists of the backward propagation of error from the
output nodes to the hidden nodes to the input nodes, followed by the adjustment of
the weights so as to reduce the error in the output nodes.
The action phase begins with the input nodes. Each node has a “state” which can
be binary (e.g. 0 or 1 to indicate whether the node “fires”) or continuous (e.g. .9 to
indicate that the node did not fire as strongly as one whose state is 1.0). The states of
the input nodes are controlled entirely by the environment and correspond to a
pattern that the network perceives. The input nodes can influence the output nodes
directly as well as via hidden nodes that are connected to the input nodes.
To illustrate, consider a neural network in which the input nodes are numbered
from i ¼ 1 to I. The ith input is selected and the input register to all nodes j
influenced by i are then updated by multiplying the state of i times the weight wij
on the ij path. This updating is repeated for each input node. The pathways that link
the nodes are weighted with values that determine the strength of the signals
moving along the path. A low absolute value means that an input has little influence
on the output. A large positive weight makes the input operate as an excitor. When
it fires, the input excites an otherwise inhibited output. A large negative path weight
makes the input operate as an inhibitor. When it fires, the input inhibits an otherwise
excited output.
Next, the nodes whose input registers have been fully updated (e.g. the nodes in
the first layer of hidden nodes) must update their states. The state of a node is
updated by aggregating the values in the node’s input register, including the input
from its bias node (which always fires, e.g. Bi ¼ 1).
Updating states is based on the activation function. Three activation functions
are commonly used. Hard-limit functions fire the node if the aggregate input
exceeds zero. Sigmoid stochastic functions fire the node with a probability given
by the aggregate input. Sigmoid deterministic functions fire the node with a
magnitude given by the aggregate input.2 For example, if node k is influenced by
two input nodes, i and j and a bias node b, then k sums the influence ik ¼ i*wij þ
k*wik þ b*wib. Positive weights cause i, j, and b to activate k and negative weights
inhibit. If j is hard-limited, then if ik > 0, k ¼ 1, else k ¼ 0. If the activation
function is stochastic, k is activated with a probability p ¼ 1/1 þ eik. If the
sigmoid function is deterministic, k is activated with magnitude p.
Once the states have been activated for all nodes whose input registers have been
fully updated (e.g., the first layer of hidden nodes and/or one or more output nodes
that are directly connected to an input node), these nodes in turn update the input
registers of nodes they influence going forward (e.g. the second layer of hidden
nodes and/or some or all of the output nodes). Once this updating is complete, all
nodes whose input registers have been fully updated aggregate across their inputs
and update their states, and so on until the states of the output nodes have been
updated. This completes the action phase.
The network learns by modifying the path weights linking the neurodes.
Learning only occurs when the response to a given sensory input pattern is
unsatisfactory. The paths are then adjusted so as to reduce the probability of
repeating the mistake the next time this pattern is encountered. In many
applications, neural nets are trained to recognize certain patterns or combinations
of inputs. For example, suppose we want to train a neural net to predict stock prices
from a set of market indicators. We first train the net to correctly predict known
prices. The net begins with random path coefficients. These generate a prediction.
The error is then used to adjust the weights in an iterative process that improves the
predictions. These weights are analogous to those in a linear regression, and like a
regression, the weights can then be applied to new data to predict the unknown.
17.5.3.1 Back Propagation of Error
A feed-forward neural network is trained by adjusting the weights on the paths

between the nodes. These weights correspond to the influence that a node will have
in causing nodes further forward in the network to fire. When those nodes fire
incorrectly, the weights responsible for the error must be adjusted accordingly. Just
as the influence process propagates forward, from the input nodes to the hidden
layer to the output nodes, the attribution of responsibility for errors propagates
backward, from the output nodes to the hidden layers.
The back propagation of error begins with the output nodes and works back
through the network to the input nodes, in the opposite direction from the influence
process that “feeds forward” from input nodes to hidden nodes to output nodes.
First, the output error is calculated for each output node. This is simply the
2
A multilayer neural net requires a non-linear activation function (such as a sigmoid). If the
functions are linear, the multilayer net reduces to a single-layer I-O network.
17 Learning 445
difference between the expected state of the ith output node (Sî ) and the state that
was observed (Si). For an output node, error refers to the difference between the
expected output for a given pattern of inputs and the observed output.
eo ¼ so ð1 so Þð^
so so Þ (17.5)
where the term so(1 so) limits the error. If the initial weights were randomly
assigned, it is unlikely that the output will be correct. For example, if we observe an
output value of 0.37 and we expected 1, then the error is 0.53.
Once the error has been updated for each output node, these errors are used to
update the error for the nodes that influenced the output nodes. These nodes are
usually located in the last layer of hidden nodes but they can be anywhere and can
even include input nodes that are wired directly to an output.3 Then the errors of the
nodes in the last hidden layer are used to update error back further still to the hidden
nodes that influenced the last layer of hidden nodes, and so on, back to the first layer
of hidden nodes, until the errors for all hidden nodes have been updated. Input
nodes cannot have error, since they simply represent an exogenous pattern that the
network is asked to learn.
Back propagation means that the error observed in an output node o ð^ so so Þ is
allocated not only to o but to all the hidden nodes that influenced o, based on the
strength of their influence. The total error of a hidden node h is then simply the
summation over all nh allocated errors from the n nodes i that h influenced,
including output nodes as well as other hidden nodes:
X
n
eh ¼ sh ð1 sh Þ whi ei (17.6)
i¼1
Once the errors for all hidden, bias, and output nodes have been back
propagated, the weight on the path from i to j is updated:
w0 ij ¼ wij þ λsi ej (17.7)
where λ is a fractional learning rate. The learning algorithm means that the
influence of i on j increases if j’s error was positive (i.e. the expected output
exceeded the observed) and decreases if j’s influence was negative.
Note that the Bush-Mosteller model is equivalent to a neural net with only a
single bias unit and an output, but with no sensory inputs or hidden units. Such a
device is capable of learning how often to act, but not when to act, that is, it is
incapable of learning conditional strategies. In contrast, a feed-forward network can
3
However, if an input node is wired to hidden nodes as well as output nodes, the error for this node
cannot be updated until the errors for all hidden nodes that it influenced have been updated.
learn to differentiate environmental cues and respond using more sophisticated

protocols for contingent strategies.
17.5.3.2 Attractor Neural Network
Feed-forward networks are the most widely used but not the only type of artificial
neural network. An alternative design is the attractor neural network (Churchland
and Sejnowski 1994; Quinlan 1991), originally developed and investigated by
Hopfield (1982; Hopfield and Tank 1985). In a recent article, Nowak and Vallacher
(1998) note the potential of these computational networks for modeling group
dynamics. This approach promises to provide a fertile expansion to social network
analysis, which has often assumed that social ties are binary and static. A neural
network provides a way to dynamically model a social network in which learning
occurs at both the individual and structural levels, as relations evolve in response to
the behaviors they constrain.
Unlike feed-forward neural networks (Rumelhart and McClelland 1988), which
are organized into hierarchies of input, hidden, and output nodes, attractor (or
“feed-lateral”) networks are internally undifferentiated. Nodes differ only in their
states and in their relational alignments, but they are functionally identical. Without
input units to receive directed feedback from the environment, these models are
“unsupervised” and thus have no centralized mechanism to coordinate learning of
efficient solutions. In the absence of formal training, each node operates using a set
of behavioral rules or functions that compute changes of state (“decisions”) in light
of available information. Zeggelink (1994) calls these “object-oriented models”,
where each agent receives input from other agents and may transform these inputs
into a change of state, which in turn serves as input for other agents.
An important agent-level rule that characterizes attractor networks is that indi-
vidual nodes seek to minimize “energy” (also “stress” or “dissonance”) across all
relations with other nodes. As with feed-forward networks, this adaptation occurs in
two discrete stages. In the action phase, nodes change their states to maximize
similarity with nodes to which they are strongly connected. In the learning phase,
they update their weights to strengthen ties to similar nodes. Thus, beginning with
some (perhaps random) configuration, the network proceeds to search over an
optimization landscape as nodes repeatedly cycle through these changes of weights
and states.
In addition to variation in path strength, artificial neural networks typically have
paths that inhibit as well as excite. That is, nodes may be connected with negative as
well as positive weights. In a social network application, agents connected by
negative ties might correspond to “negative referents” (Schwartz and Ames
1977), who provoke differentiation rather than imitation.
Ultimately, these systems are able to locate stable configurations (called
“attractors”), for which any change of state or weight would result in a net increase
in stress for the affected nodes. Hopfield (1982) compares these attractors to
memories, and shows that these systems of undifferentiated nodes can learn to
17 Learning 447
implement higher-order cognitive functions. However, although the system may

converge at a stable equilibrium that allows all nodes to be locally satisfied (i.e., a
“local optimum”), this does not guarantee that the converged pattern will minimize
overall dissonance (a “global optimum”).
This class of models generally uses complete networks, with each node
characterized by one or more binary or continuous states and linked to other nodes
through endogenous weights. Like other neural networks, attractor networks learn
stable configurations by iteratively adjusting the weights between individual nodes,
without any global coordination. In this case, the weights change over time through a
Hebbian learning rule: the weight wij is a function of the correlation of states for nodes
i and j over time. Specifically, Hebbian learning implies the following rules:
• To the extent that nodes i and j adopt the same states at the same time, the weight
of their common tie will increase until it approaches some upper limit (e.g., 1.0).
• To the extent that nodes i and j simultaneously adopt different states, the weight
of their common tie will decrease until it approaches some lower limit (e.g., 0.0).
Although Hebbian learning was developed to study memory in cognitive
systems, it corresponds to the homophily principle in social psychology (Homans
1951) and social network theory (McPherson and Smith-Lovin 1987), which holds
that agents tend to be attracted to those whom they more closely resemble. This
hypothesis is also consistent with structural balance theory (Cartwright and Harary
1956; Heider 1958) and has been widely supported in studies of interpersonal
attraction and interaction, where it has been called “the Law of Attraction”
(Byrne 1971; Byrne and Griffit 1966).
An important property of attractor networks is that individual nodes seek to
minimize “energy” (or dissonance) across all relations with other nodes – a process
that parallels but differs from the pursuit of balanced relations in structural balance
theory. These networks also posit self-reinforcing dynamics of attraction and
influence as well as repulsion and differentiation.
Following Nowak and Vallacher (1998), Macy et al. (2003) apply the Hopfield
model of dynamic attraction to the study of polarization in social networks. In this
application, observed similarity/difference between states determines the strength
and valence of the tie to a given referent. This attraction and repulsion may be
described anecdotally in terms of liking, respect, or credibility and their opposites.
In their application of the Hopfield model, each node has N 1 undirected ties
to other nodes. These ties include weights, which determine the strength and
valence of influence between agents. Formally, social pressure on agent i to adopt
a binary state s (where s ¼ 1) is the sum of the states of all other agents j, where
influence from each agent is conditioned by the weight (wij) of the dyadic tie
between i and j (1.0 < wij < 1.0):
P
N
wij sj
j¼1
Pis ¼ ; j 6¼ i (17.8)
N1
Thus, social pressure (1 < Pis < 1) to adopt s becomes increasingly positive as
i’s “friends” adopt s (s ¼ 1) and i’s “enemies” reject s (s ¼ 1). The pressure can
also become negative in the opposite circumstances. The model extends to multiple
states in a straightforward way, where Eq. 17.8 independently determines the
pressure on agent i for each binary state s.
Strong positive or negative social pressure does not guarantee that an agent will
accommodate, however. It is effective only if i is willing and able to respond to peer
influence. If i is closed-minded or if a given trait is not under i’s control (e.g.,
ethnicity or gender), then no change to s will occur. The probability π that agent i
will change state s is a cumulative logistic function of social pressure:
1
π is ¼ (17.9)
1 þ e10Pis
Agent i adopts s if π > C ε, where C is the inflection point of the sigmoid, and
ε is an exogenous error parameter (0 ε 1). At one extreme, ε ¼ 0 produces
highly deterministic behavior, such that any social pressure above the trigger value
always leads to conformity and pressures below the trigger value entail differentia-
tion. Following Harsanyi (1973), ε > 0 allows for a “smoothed best reply” in which
pressure levels near the trigger point leave the agent relatively indifferent and thus
likely to explore behaviors on either side of the threshold.
In the Hopfield model, the path weight wij changes as a function of similarity in
the states of node i and j. Weights begin with uniformly distributed random values,
subject to the constraints that weights are symmetric (wij ¼ wji). Across a vector of
K distinct states sik (or the position of agent i on issue k), agent i compares its own
states to the observed states of another agent j and adjusts the weight upward or
downward corresponding to their aggregated level of agreement or disagreement.
Based on the correspondence of states for agents i and j, their weight will change at
each discrete time point t in proportion to a parameter λ, which defines the rate of
structural learning (0 < λ < 1):
λ XK
wij;tþ1 ¼ wijt ð1 λÞ þ sjkt sikt ; j 6¼ i (17.10)
K k¼1
As correspondence of states can be positive (agreement) or negative (disagree-

ment), ties can grow positive or negative over time, with weights between any two
agents always symmetric.
Note one significant departure from structural balance theory. Although the
agents in this model are clearly designed to maintain balance in their behaviors
with both positive and negative referents, this assumption is not “wired in” to the
relations themselves. That is, two agents i and j feel no direct need for consistency
in their relations with a third agent h. Indeed, i has no knowledge of the jh
relationship and thus no ability to adjust the ij relation so as to balance the triad.
17 Learning 449
Given an initially random configuration of states and weights, these agents will
search for a profile that minimizes dissonance across their relations. Structural
balance theory predicts that system-level stability can only occur when the group
either has become uniform or has polarized into two (Cartwright and Harary 1956)
or more (Davis 1967) internally cohesive and mutually antipathetic cliques. How-
ever, there is no guarantee in this model that they will achieve a globally optimal
state in structural balance.
17.5.4 Belief Learning
Actors learn not only what is useful for obtaining rewards and avoiding
punishments, they also update their beliefs about what is true and what is false.
There are two main models of belief learning in the literature, Bayesian belief
learning and fictitious play4 (cf. Offerman 1997; Fudenberg and Levine 1998).
These models differ in their assumptions about how players learn from
observations. Both models assume that players believe that something is true with
some fixed unknown probability p. In Bayesian learning, players then use Bayes’
learning rule to rationally update over time their beliefs about p. In a nutshell,
Bayes’ learning rule implies that actors’ assessment of the true value of p converges
in the long run on the relative frequency of events that they observe. However, in
the short and medium term, Bayesian learners remain suspicious in the sense that
they take into account that observed events are an imperfect indication of
p (Offerman 1997).
Fudenberg and Levine note that fictitious play is a special case of Bayesian
learning. Fictitious play is a form of Bayesian learning that always puts full weight
on the belief that the true value of p corresponds to the relative frequency observed
in past events. Fudenberg and Levine (1998) note that it is an implausible property
of fictitious play that a slight change in beliefs may radically alter behavior. The
reason is that the best reply function always is a step function. As a remedy,
Fudenberg and Levine introduce a smooth reply curve. The reply curve assigns a
probability distribution that corresponds to the relative frequency of events. With
strict best reply, the reply curve is a step function. Instead, a smooth reply curve
assigns some probability to the action that is not strict best reply. This probability
decreases in the difference between the expected payoffs. Specifically, when
expected payoffs are equal, actors choose with equal probability, whereas their
choice probabilities converge on pure strategies when the difference in expected
payoffs approaches the maximum value.
4
The Cournot rule may be considered as a third degenerate model of belief learning. According to
the Cournot rule, players assume that the behavior of the opponent in the previous round will
always occur again in the present round.
The strict best reply function corresponds to the rule, “play X if the expected
payoff for X is better than the expected payoff for Y, given your belief p. Otherwise
play Y”. Smooth best reply is then introduced with the modification to play the
strict best reply strategy only with a probability of 1 η, whereas the alternative is
played with probability η. The probability η, in turn, decreases in the absolute
difference between expected payoffs where η ¼ 0.5 if players are indifferent.
Belief learning generally converges with the predictions of evolutionary selec-
tion models. The approaches are also similar in the predicted effects of initial
conditions on end results. Broadly, the initial distribution of strategies in indepen-
dent populations in evolutionary selection corresponds to the initial beliefs
players’ hold about their opponent. For example, when two pure strategy Nash
equilibria are feasible, then the one tends to be selected towards which the initial
strategy distribution in evolutionary selection or initial beliefs in belief learning
are biased.
17.6 Conclusion
Evolution and learning are examples of backward-looking consequentialist models,

in which outcomes of agents’ past actions influence their future choices, either
through selection (in the case of evolution) or reinforcement (in the case of
learning). Backward-looking models make weaker assumptions about agents’ cog-
nitive capacities than forward looking models, and thus may be appropriate for
settings in which agents lack the ability, resources, information, and motivation to
engage in intensive cognitive processing, as in most everyday instances of collec-
tive action. Backward-looking models may also be useful for understanding behav-
ior driven by affect, rather than calculation. Forward looking models may be more
appropriate in applications such as investment decisions, international diplomacy,
or military strategy, where the stakes are high enough to warrant collection of all
relevant information and the actors are highly skilled strategists. However, even
where the cognitive assumptions of the models are plausible, forward-looking
models are generally limited to the identification of static equilibria but not
necessarily whether and how agents will reach those equilibria.
When implemented computationally, backward-looking models can show how
likely agents are to reach particular equilibria, as well as the paths by which those
equilibria may be reached. However, computational models are also less general
than analytical models. Furthermore, backward-looking models will be of little help
if change in the environment outpaces the rate of adaptation. These limitations
underscore the importance of robustness testing over a range of parameter values.
Evolutionary models are most appropriate for theoretical questions in which
adaptation takes place at the population level, through processes of selection.
Biological evolution is the most obvious analog, but social and cultural evolution
are likely to be more important for social scientists. However, as we note above,
17 Learning 451
researchers must be cautious about drawing analogies from the biological to social/
cultural dynamics.
Learning models based on reinforcement and Bayesian updating are useful in
applications that do not require conditional strategies based on pattern recognition.
When agents must learn more complex conditional strategies, feed-forward neural
networks may be employed. Furthermore, attractor neural networks are useful for
modeling structural influence, such as conformity pressures from peers.
Models of evolution and learning are powerful tools for modeling social pro-
cesses. Both show how complex patterns can arise when agents rely on relatively
simple, experience-driven decision rules. This chapter seeks to provide researchers
with an overview of promising research in this area, and the tools necessary to
further develop this research.
Further Reading
We refer readers interested in particular learning models and their application in

agent-based simulation to (Macy and Flache 2002), which gives a brief introduction
into principles of reinforcement learning and discusses by means of simulation
models how reinforcement learning affects behavior in social dilemma situations,
whereas (Macy 1996) compares two different approaches of modeling learning
behavior by means of computer simulations. (Fudenberg and Levine 1998) gives a
very good overview on how various learning rules relate to game theoretic rational-
ity and equilibrium concepts.
For some wider background reading, we recommend (Macy 2004), which
introduces the basic principles of learning theory applied to social behavior,
(Holland et al. 1986), which presents a framework in terms of rule-based mental
models for understanding inductive reasoning and learning, and Sun (2008), which
is a handbook of computational cognitive modeling.
References
Allison P (1992) The cultural evolution of beneficent norms. Soc Forces 71:279–301
Bandura A (1977) Social learning theory. Prentice Hall, Englewood Cliffs
Börgers T, Sarin R (1996) Naive reinforcement learning with endogenous aspirations. Interna-
tional Economic Review 4:921–950
Börgers T, Sarin R (1997) Learning through reinforcement and replicator dynamics. J Econ
Theory 77:1–14
Boyd R, Richerson PJ (1985) Culture and the evolutionary process. University of Chicago Press,
Chicago
Bush RR, Mosteller F (1950) Stochastic models for learning. Wiley, New York
Byrne DE (1971) The attraction paradigm. Academic, New York
Byrne DE, Griffith D (1966) A development investigation of the law of attraction. J Pers Soc
Psychol 4:699–702
Cartwright D, Harary F (1956) Structural balance: a generalization of Heider’s theory. Psychol

Rev 63:277–293
Catania AC (1992) Learning, 3rd edn. Prentice Hall, Englewood Cliffs
Churchland PS, Sejnowski TJ (1994) The computational brain. MIT Press, Cambridge, MA
Cohen MD, Riolo R, Axelrod R (2001) The role of social structure in the maintenance of
cooperative regimes. Ration Soc 13(1):5–32
Davis JA (1967) Clustering and structural balance in graphs. Hum Relat 20:181–187
Dawkins R (1976) The selfish gene. Oxford University Press, Oxford
Durham WH (1992) Coevolution: genes, culture and human diversity. Stanford University Press,
Stanford
Erev I, Rapoport A (1998) Coordination, “magic”, and reinforcement learning in a market entry
game. Game Econ Behav 23:146–175
Erev I, Roth AE (1998) Predicting how people play games: reinforcement learning in experimental
games with unique, mixed strategy equilibria. Am Econ Rev 88(4):848–879
Erev I, Bereby-Meyer Y, Roth AE (1999) The effect of adding a constant to all payoffs: experi-
mental investigation, and implications for reinforcement learning models. J Econ Behav Organ
39(1):111–128
Flache A, Macy MW (2002) Stochastic collusion and the power Law of learning: a general
reinforcement learning model of cooperation. J Confl Resolut 46(5):629–653
Frank R (1988) Passions within reason: the strategic role of the emotions. Norton, New York
Fudenberg D, Levine D (1998) The theory of learning in games. MIT Press, Boston
Goldberg D (1989) Genetic algorithms in search, optimization, and machine learning. Addison-
Wesley, New York
Hamilton W (1964) The genetic evolution of social behaviour. J Theor Biol 17:1–54
Harsanyi J (1973) Games with randomly disturbed payoffs. Int J Game Theory 2:1–23
Heider F (1958) The psychology of interpersonal relations. Wiley, New York
Herrnstein RJ, Drazin P (1991) Meliorization: a theory of distributed choice. J Econ Perspect 5
(3):137–156
Holland J (1975) Adaptation in natural and artificial systems. University of Michigan Press, Ann
Arbor
Holland JH, Holyoak KJ, Nisbett RE, Thagard PR (1986) Induction: processes of inference,
learning, and discovery. MIT Press, Cambridge, MA
Homans GC (1951) The human group. Harcourt Brace, New York
Hopfield JJ (1982) Neural networks and physical systems with emergent collective computational
abilities. Proc Am Acad Sci 79:2554–2558
Hopfield JJ, Tank DW (1985) “Neural” computation of decisions in optimization problems. Bio
Cybernetics 52:141–152
James W (1981) Principles of psychology. Harvard University Press, Cambridge, MA
Lopreato J (1990) From social evolutionism to biocultural evolutionism. Sociol Forum 5:187–212
Macy M (1989) Walking Out of social traps: a stochastic learning model for the Prisoner’s
dilemma. Ration Soc 2:197–219
Macy M (1991) Learning to cooperate: stochastic and tacit collusion in social exchange.
Am J Sociol 97:808–843
Macy M (1996) Natural selection and social learning in prisoner’s dilemma: co-adaptation with
genetic algorithms and artificial neural networks. Sociol Method Res 25:103–137
Macy M (1997) Identity, interest, and emergent rationality. Ration Soc 9:427–448
Macy M (1998) Social order in an artificial world. J Artif Soc Soc Simul 1(1). http://jasss.soc.
Macy M (2004) Learning theory. In: Ritzer G (ed) Encyclopedia of social theory. Sage, Thousand
Oaks
Macy M, Flache A (2002) Learning dynamics in social dilemmas. Proc Natl Acad Sci
99:7229–7236
17 Learning 453
Macy M, Skvoretz J (1998) The evolution of trust and cooperation between strangers: a computa-
tional model. Am Sociol Rev 63:638–660
Macy M, Kitts J, Flache A, Benard S (2003) Polarization in dynamic networks: a Hopfield model
of emergent structure. In: Breiger R, Carley K (eds) Dynamic social network modeling and
analysis: workshop summary and papers. National Academy Press, Washington, DC,
pp 162–173
March JG, Simon HA (1958) Organizations. Wiley, New York
McPherson JM, Smith-Lovin L (1987) Homophily in voluntary organizations: status distance and
the composition of face to face groups. Am Sociol Rev 52:370–379
Nowak A, Vallacher RR (1998) Toward computational social psychology: cellular automata and
neural network models of interpersonal dynamics. In: Read SJ, Miller LC (eds) Connectionist
models of social reasoning and social behavior. Lawrence Erlbaum, Mahwah, pp 277–311
Offerman T (1997) Beliefs and decision rules in public good games. Kluwer, Dordrecht
Quinlan PT (1991) Connectionism and psychology: a psychological perspective on New connec-
tionist research. University of Chicago Press, Chicago
Roth AE, Erev I (1995) Learning in extensive-form games: experimental data and simple dynamic
models in intermediate term. Game Econ Behav 8:164–212
Rumelhart DE, McClelland JL (1988) Parallel distributed processing: explorations in the micro-
structure of cognition. MIT Press, Cambridge, MA
Schwartz SH, Ames RE (1977) Positive and negative referent others as sources of influence: a case
of helping. Sociometry 40:12–21
Sun R (ed) (2008) The Cambridge handbook of computational psychology. Cambridge University
Press, Cambridge
Taylor PD, Jonker L (1978) Evolutionary stable strategies and game dynamics. Math Biosci
40:145–156
Thorndike EL (1898) Animal intelligence: an experimental study of the associative processes in
animals, vol 8, Psychological review, monograph supplements. Macmillan, New York
Vanberg V (1994) Rules and choice in economics. Routledge, London
Williams GC (1966) Adaptation and natural selection. Princeton University Press, Princeton
Wilson EO (1988) On human nature. Harvard University Press, Cambridge, MA
Wynne-Edwards VC (1962) Animal dispersion in relation to social behaviour. Oliver & Boyd,
Edinburgh
Wynne-Edwards VC (1986) Evolution through group selection. Blackwell, Oxford
Young HP (1998) Individual strategy and social structure. An evolutionary theory of institutions.
Zeggelink E (1994) Dynamics of structure: an individual oriented approach. Soc Network
16:295–333
Chapter 18
Evolutionary Mechanisms
Edmund Chattoe-Brown and Bruce Edmonds
Why Read This Chapter? To learn about techniques that may be useful in
designing simulations of adaptive systems including Genetic Algorithms (GA),
Classifier Systems (CS) and Genetic Programming (GP). The chapter will also
tell you about simulations that have a fundamentally evolutionary structure – those
with variation, selection and replications of entities – showing how this might be
made relevant to social science problems.
Abstract After an introduction, the abstract idea of evolution is analysed into four
processes which are illustrated with respect to a simple evolutionary game. A brief
history of evolutionary ideas in the social sciences is given, illustrating the different
ways in which the idea of evolution has been used. The technique of GA is then
described and discussed including: the representation of the problem and the com-
position of the initial population, the fitness function, the reproduction process, the
genetic operators, issues of convergence, and some generalisations of the approach
including endogenising the evolution. GP and CS are also briefly introduced as
potential developments of GA. Four detailed examples of social science applications
of evolutionary techniques are then presented: the use of GA in the Arifovic
“cobweb” model, using CS in a model of price setting developed by Moss, the
role of GP in understanding decision making processes in a stock market model and
relating evolutionary ideas to social science in a model of survival for “strict”
churches. The chapter concludes with a discussion of the prospects and difficulties
of using the idea of biological evolution in the social sciences.
E. Chattoe-Brown (*)
Department of Sociology, University of Leicester, Leicester, UK
B. Edmonds
Centre for Policy Modelling, Manchester Metropolitan University, Manchester, UK

456 E. Chattoe-Brown and B. Edmonds
18.1 Introduction
There are now many simulations of complex social phenomena that have structures
or component processes analogous to biological evolution (see Arifovic 1994;
Chattoe 2006a; Dosi et al. 1999; Lomborg 1996; Nelson and Winter 1982; Oliphant
1996; Windrum and Birchenhall 1998 to get a flavour of the diversity of approach
and applications). Clearly the process of biological evolution is complex and has
resulted in the development of complex (and in several cases social) systems.
However, biological evolution follows very specific mechanisms and is clearly
not strictly isomorphic with social processes. For a start, biological evolution
occurs over larger time spans than most social processes. Further, it is unlikely,
as sociobiology (Wilson 1975) and evolutionary psychology (Buss 1998) are
sometimes supposed to imply, that the domain of social behaviour will actually
prove reducible to genetics. Thus it is not immediately apparent why evolutionary
ideas have had such an influence upon the modelling of social processes. Neverthe-
less, simulations of social phenomena have been strongly influenced by our under-
standing of biological evolution and this has occurred via two main routes: through
analogies with biological evolution and through computer science approaches.
In the first case, conceptions of evolution have been used as a way of under-
standing social processes and then simulations have been made using these
conceptions. For example Nelson and Winter (1982) modelled growth and change
in firms using the idea of random variation (new products or production processes)
and selective retention (whether these novelties in fact sustain profitability – the
survival requirement for firms – in an environment defined by what other firms are
currently doing).
In the second case, computer science has taken up the ideas of evolution and
applied it to engineering problems. Most importantly in Machine Learning, ideas
from biological evolution have inspired whole families of techniques in what has
become known as “Evolutionary Computation”. The most famous of these
techniques are Genetic Algorithms (Holland 1975; Mitchell 1996) and Genetic
Programming (Koza 1992a, 1994) discussed below. These algorithms have then
been applied in social simulations with different degrees of modification; from
using them unchanged as “off the shelf” plug-ins (for example to model learning
processes) to specifying simulation processes that use the core evolutionary idea
but are completely re-engineered for a particular modelling purpose or domain.
There is no a priori reason to suppose that a particular technique from computer
science will be the most appropriate algorithm in a social simulation (including
those with a biological inspiration) as we shall see below, but it certainly presents a
wealth of evolutionary ideas and results that are potentially applicable in some
form. Like any theory, the trick is to use good judgement and a clear specification in
applying an algorithm to a particular social domain (Chattoe 1998, 2006b).
What is certainly the case is that biological evolution offers an example of how
complex and self-organised phenomena can emerge using randomness, so it is
natural to look to this as a possible conceptual framework with which to understand
18 Evolutionary Mechanisms 457
social phenomena with similar properties. (In particular, while it may be reasonable
to assume deliberation and rationality in some social contexts, it is extremely
unlikely to apply to all social structures and phenomena. As such, some kind of
blind variation and retention – resulting in evolution – is probably the only well
defined theoretical alternative.) The extent to which evolution-like processes are
generally applicable to social phenomena is unknown (largely because this founda-
tional issue has not received much attention to date), but these processes certainly
are a rich source of ideas and it may be that there are some aspects of social
complexity that will prove to be explicable by models thus inspired. It is already the
case that many social simulation models have taken this path and thus have the
potential to play a part in helping us to understand social complexity (even if they
only serve as horrible examples).
This chapter looks at some of the most widely used approaches to this kind of
modelling, discusses others, gives examples and critically discusses the field along
with areas of potential development.
18.2 An Abstract Description of Biological Evolution
We will not provide full details of biological evolution as currently understood in

the Neo-Darwinian synthesis.1 Rather we will take from this a generalised model of
evolution that will potentially cover a variety of social processes. This description
will then be used to discuss an example from Evolutionary Game Theory (Vega-
Redondo 1996). This will unpack the implications of the abstract description and
demonstrate its generality. This generalisation is a preliminary to discussing evolu-
tionary simulations of social phenomena based on the abstract description as a
framework.
18.2.1 The Four Process Description
The basic components in the biological theory are the genotype (the set of
instructions or genome) and the phenotype (the “body” which the genotype specifies)
in which these instructions are embedded. The phenotype is constructed using
“instructions” encoded in the genotype. The phenotype has various capabilities
including reproduction. Maintenance of the phenotype (and the embedded genotype)
requires a number of potentially scarce inputs (food, water). The phenotypic
capabilities include management of inputs and outputs to the organism. Poor adapta-
tion of these capabilities with respect to either external or internal environment will
1
For details about this see any good textbook on biology (e.g. Dobzhansky et al. 1977).
result in malfunction and consequent death. The death of a particular phenotype also
ends its reproductive activity and removes the corresponding genotype from the
population. Variation occurs by mutation during reproduction, giving rise to novel
genotypes (and hence subsequent phenotypes) in the resulting offspring. Genotypic
variations are not selected directly by the environment but according to the overall
capabilities of their phenotypes. In biology, phenotype alterations cannot be trans-
mitted to the genotype for physiological reasons but in social systems this “Lamarck-
ian” adjunct to evolution (which is not, however, adequate to explain change in its
own right) is both possible and plausible. In particular it allows for combinations of
evolutionary learning at the social level and deliberate action at the individual level
(Chattoe 2006a).
A full specification of an evolutionary model requires descriptions of the fol-
lowing processes:
1. Generation of phenotypes: A specification of the genotypes and the phenotypes
these correspond to. This may not specify a 1–1 mapping between genotypes and
phenotypes but describe the process by which phenotypes are actually constructed
from genotypes. This is necessary when genotypes cannot be enumerated.
2. Capabilities of the phenotypes: A specification of ways in which phenotypes
may use their capabilities to affect the internal and external environment,
including the behaviour and numbers of other phenotypes. Lamarckian systems
include the capability to modify the genotype using environmental feedback
during the lifetime of the phenotype.
3. Mechanisms of reproduction and variation: A specification of the process by
which phenotypes reproduce including possible differences between ancestor
and successor genotypes resulting from reproduction. Reproduction may involve
a single ancestor genotype (parthenogenesis) or a pair (sexual reproduction). In
principle, multiple parents could be modelled if appropriate for particular social
domains (like policies decided by committees) though this approach has not
been used so far.
4. Mechanism of selection: A specification of all the processes impinging on the
phenotype and their effects. This is the converse of the second mechanism, the
capabilities of one phenotype form part of the selection process for the others.
Some processes, such as fighting to the death, can be seen as directly selective.
However, even indirect processes like global warming may interact with pheno-
typic capabilities in ways that affect fitness.
In these process specifications it may be convenient to distinguish (and model
separately) the “environment” as the subset of objects impinging on phenotypes
which display no processes of the first three types. Whether a separate representa-
tion of the environment is useful depends on the process being modelled. At one
extreme, a person in a desert is almost exclusively dealing with the environment. At
the other, rats in an otherwise empty cage interact almost entirely with each other.
Obviously, some of these specifications could be extremely complex depending
on the system being modelled. The division into system components is necessarily
imprecise but not arbitrary. It is based on the considerable observed integrity of

organisms relative to their environment. (This integrity is also observed in social
“organisms” like firms which have clearly – and often legally – defined boundaries.)
The first and third specifications involve processes internal to the organism while the
second and fourth represent the organism’s effect on the external world and the
converse.
Of course, social processes, even “evolutionary social processes” are not limited
by the above specification. For example what most closely corresponds to the
genotype might not be separable from what corresponds to the phenotype. Never-
theless, however, for a very broad class of evolutionary simulations it will be
necessary to implement something very similar to the above four categories.
18.2.2 Illustrative Example: A Simple Evolutionary Game
Despite the potential complexity of specifying complete models for biological

systems, this description can also be used to clarify and analyse relatively simple
evolutionary systems. In this section we shall provide a description for an evolu-
tionary game. The purpose is not to comment on Evolutionary Game Theory per se
but to show how the description raises issues relevant to our understanding of
evolutionary models.
For each agent, the genotype is one of a set of finite state automata producing a
single action in each period, for example the complete set of one and two-state
automata leading to the actions “Co-operate” (C) and “Defect” (D) in a Prisoner’s
Dilemma (see, for example, Lomborg 1996). The action is compared with the
action of a co-player (another agent) and the result is an adjustment to the “energy
level” for each agent depending on the game payoffs and chosen strategies. If
agents reach a certain energy level, they produce an exact copy. (This model
dispenses with variation and involves asexual reproduction.) If the energy level
of any agent reaches zero, it dies and is removed from the environment. Reproduc-
tion reduces the energy level considerably. Merely existing also does so but at a
much lower rate.
With some qualifications, this is an example of a complete description discussed
in the last section. It reveals some interesting things about the process of
constructing such descriptions.
Firstly, this model involves a very attenuated environment compared to real
social systems. Agents have a single external capability involving one of two
actions and thus affecting the energy levels of their co-players. The effect of
these actions is also the only environmental process that impinges on agents. The
model of the environment just consists of mechanisms for deciding when and which
agents will play, administering actions and energy changes, producing copies and
removing dead agents. In real social systems, exogenous events (both social and
environmental) are likely to be very important.
Secondly, the discussion of energy levels still sounds biological but this is
simply to make the interpretation of the example more straightforward in the
light of the recent discussion. As we shall show subsequently, the survival criterion
can just as easily be profitability or organisation membership levels.
Thirdly (and perhaps most importantly) there has been sleight of hand in the
description of the model. We have already described Lamarckism (genotype modi-
fication by the phenotype during the organism’s lifetime) and the construction of the
phenotype by the genotype during gestation is a fundamental part of the evolutionary
process. But in this example the genotype is effectively “reconstructing” the pheno-
type every time the finite state automaton generates an action. There is nothing
observable about a particular agent, given the description above, except the sequence
of actions they choose. There is no way for an agent to establish that another is
actually the same when it plays D on one occasion and C on another, or that two
plays of D in successive periods actually come from two different agents. In fact, this
is the point at which models of social evolution develop intuitively from the simple
description of biological evolution used so far. The capabilities of social agents
(such as consumers, families, churches and firms) include the “senses” that give
them the ability to record actions and reactions in memory. Furthermore, they have
mental capabilities that permit the processing of sense data in various ways, some
subset of which we might call rationality. The simple automata described above are
reactive, in that their actions depend in systematic ways on external stimuli, but they
can hardly be said to be rational or reflective in that their “decision process” involves
no choice points, internal representations of the world or “deliberation”. Such
distinctions shelve into the deep waters of defining intelligence, but the important
point is that we can make useful distinctions between different kinds of adaptability
based on the specifications of process we use in our models without compromising
the evolutionary framework we have set up. It is in this way that the complex
relationship between selection and reasoned action may begin to be addressed.
Thus, even in this simple example, one can see not only the general evolutionary
structure of the simulation but also that the conception differs in significant ways
from the corresponding biological process.
18.3 Evolutionary Ideas in the Social Sciences
From early on, since the publication of “The Origin of Species” (Darwin 1859),
Darwin’s ideas of evolution have influenced those who have studied social phenom-
ena. For example, Tarde (1884) published a paper discussing “natural” and “social”
Darwinism. This marked a shift from looking at the social organisation of
individuals to the patterns of social products (fashions, ideas, tunes, laws and so
on). Tarde (1903 p. 74) put it like this:
but self-propagation and not self-organisation is the prime demand of the social as well as
of the vital thing. Organisation is but the means of which propagation, of which generative
or imitative imitation, is the end.
However, it was from the latter half of the twentieth Century that the full force
of the analogy with biological evolution (as understood in the Neo-Darwinian
Synthesis) was felt in the social sciences. There were those who sought to under-
stand the continuous change in cultural behaviours over long time-scales in this
way, e.g. (Boyd and Richerson 1985; Campbell 1965; Cavalli-Sforza and Feldman
1973; Cloak 1975; Csányi 1989). Richard Dawkins coined the term ‘meme’ as a
discrete and identifiable unit of cultural inheritance corresponding to the biological
gene (Dawkins 1976, 1982), an idea which has influenced a multitude of thinkers
including (Costall 1991; Lynch 1996; Dennett 1990; Heyes and Plotkin 1989; Hull
1982, 1988; Westoby 1994). Another stream of influence has been the Philosophy
of Science via the idea that truth might result from the evolution of competing
hypotheses (Popper 1979), a position known as Evolutionary Epistemology since
(Campbell 1974). The ultimate reflection of the shift described by Tarde above is
that the human mind is “merely” the niche where memes survive (Blackmore 1999)
or which they exploit (as “viruses of the mind”, Dawkins 1993) – the human brain
is programmed by the memes, rather than using them (Dennett 1990). This fits in
with the idea of the Social Intelligence Hypothesis (Kummer et al. 1997) that the
biological reason the brain evolved is because it allows specific cultures to develop
in groups giving specific survival value with respect to the ecological niches they
inhabit (Reader 1970). All of these ideas hinge on the importance of imitation
(Dautenhahn and Nehaniv 2002), since without this process individual memes,
ideas or cultural patterns would be quickly lost.
Evolutionary theories are applied in a wide variety of disciplines. As mentioned
above, evolutionary theories are applied to culture and anthropology, as in the work
of Boyd and Richerson, Cavalli-Sforza and Feldman, and Csányi. The evolution of
language can be seen as an analogy to biological evolution, as described by
Hoenigswald and Wiener (1987). In computer science, Genetic Programming and
Genetic Algorithms (as well as the more rarely used Classifier Systems) are
descendants of the evolutionary view as well, for example in the work of several
individuals at the Santa Fe Institute (Holland 1975; Kauffman 1993). Learning
theories of humans, applied to individuals, groups and society can be tied to
evolutionary theory, as shown in the work of Campbell (1965, 1974). The work of
several philosophers of science also shows an evolutionary perspective on knowl-
edge, as in the work of Popper (1979) and Kuhn (1970). Such theories have been used
to account for brain development by Gerald Edelman (1992), and extended to the ms-
to-minutes time scale of thought and action by William Calvin (1996a, b). Evolu-
tionary theory (and in some cases, explicit modelling) is present in economics, often
tied to the development of technology, as in the work of Nelson and Winter (1982) or
to the evolution of institutions and practices as in the work of Dosi et al. (1999),
Hodgson (1993) and North (1990). Sociology too has used evolutionary ideas and
simulations to understand the evolution of social order (Lomborg 1996; Macy 1996),
changing populations of organisations (Hannan and Freeman 1993) and the survival
of so-called “strict” churches (Chattoe 2006a).
Interestingly, however, against these creative approaches must be set forces, in
particular social sciences that have slowed or marginalised their adoption. In
sociology, the conversion of functionalism (potentially a form of social evolution)

into a virtual religion was followed by a huge backlash against untestable grand
theory which made these ideas virtually beyond the pale for 20 years or so (Chattoe
2002; Runciman 1998). It is quite likely that confused associations with Social
Darwinism, eugenics and even Nazism have not helped the use of biological
analogies in social science from the 1940s until quite recently. In economics,
the focus on deliberate rationality and well defined equilibria has meant that
evolutionary approaches are judged ad hoc unless they can be reinterpreted to
support the core assumptions of economics. (This can be observed, for example,
in evolutionary approaches to equilibrium selection where the object is not to
understand the dynamics of the system but to support the claim that particular
equilibria are robust.) In psychology, while there appears to be no overt objection to
evolutionary approaches, it seems to be the case (perhaps for historical reasons) that
the main interest in these ideas is to explain behaviour using genetic accounts of
cognitive structure rather than using evolutionary analogies.
In subsequent sections, having shown that interest in evolutionary ideas is
widespread, we turn to technical details of various kinds of evolutionary algorithm,
their strengths, weaknesses and social applicability so the reader is able to evaluate
their use and consider applications in their own areas of research interest. We start
with the Genetic Algorithm, which is easiest to describe, then move to Genetic
Programming and the (more rarely used but in some sense more satisfactory as an
analogy) Classifier Systems. The final example doesn’t rely directly on the use of an
evolutionary algorithm but clearly attempts to model a social process using a
biological analogy.
18.4 The Basic Genetic Algorithm
This section describes the basic operation and limitations of the Genetic Algorithm.
This leads to a description of ways in which the Genetic Algorithm can be
generalised and a detailed discussion of one specific way of generalising it (Genetic
Programming) in the subsequent section.
18.4.1 What Is a Genetic Algorithm?
The Genetic Algorithm is actually a family of programmes developed by John

Holland (1975) and his co-workers at the University of Michigan. The following
algorithm describes the structure of a typical Genetic Algorithm. It is the different
ways in which various parts of the algorithm can be implemented which produces
the wide variety of Genetic Algorithms available. Each part of the algorithm will be
discussed in more detail in a subsequent section. For the purposes of illustration,
consider an attempt to solve the notorious Travelling Salesman Problem that
involves producing the shortest tour of a set of cities at known distances visiting
each once only (Grefenstette et al. 1985).
1. Represent potential solutions to the problem as data structures.
2. Generate a number of these solutions/structures and store them in a composite
data structure called the Solution Pool.
3. Evaluate the “fitness” of each solution in the Solution Pool using a Fitness
Function.
4. Make copies of each solution in the Solution Pool, the number of copies
depending positively on its fitness according to a Reproduction Function.
These copies are stored in a second (temporary) composite data structure called
the Breeding Pool.
5. Apply Genetic Operators to copies in the Breeding Pool chosen as “parents” and
return one or more of the resulting “offspring” to the Solution Pool, randomly
overwriting solutions which are already there. Repeat this step until some
proportion of the Solution Pool has been replaced.
6. Repeat steps 3, 4 and 5 until the population of the Solution Pool satisfies a
Stopping Condition. One such condition is that the Solution Pool should be
within a certain distance of homogeneity.
There is an obvious parallel between this algorithm and the process of biological
evolution that inspired it. The string representing a solution to a problem corresponds
to the genotype and each element to a gene. The Fitness Function represents the
environment that selects whole genotypes on the basis of their relative performance.
The Genetic Operators correspond to the processes causing genetic variation in
biology that allow better genes to propagate while poorer ones are selected out.
This class of Genetic Algorithms has a number of interesting properties (for further
discussion see Goldberg 1989).
1. It is evolutionary. Genetic Operators combine and modify solutions directly to
generate new ones. Non-evolutionary search algorithms typically generate
solutions “from scratch” even if the location of these solutions is determined by
the current location of the search process. The common Genetic Operators are
based on biological processes of variation. Genetic Operators permit short
subsections of parent solutions to be propagated unchanged in their offspring.
These subsections (called schemata) are selected through their effect on the
overall fitness of solutions. Schemata that produce high fitness for the solutions
in which they occur continue to be propagated while those producing lower fitness
tend to die out. (Note that while it is not possible to assign a meaningful fitness to
single genes, it is possible to talk about the relative fitness of whole genotypes
differing by one or more genes. By extension, this permits talk about successful
“combinations” of genes.) The Genetic Operators also mix “genetic material”
(different solutions in the Breeding Pool) and thus help to ensure that all the
promising areas of the Problem Space are explored continuously. These ideas
clearly resonate with the social production of knowledge, in science for example.
2. It is non-local. Each solution is potentially exploring a different area of the

Problem Space although solutions can “cluster” in promising areas to explore
them more thoroughly. This allows for societies to be “smarter” than their
members.
3. It is probabilistic. The Fitness Function ensures that fitter solutions participate in
Genetic Operators more often because they have more copies in the Breeding
Pool and are thus more likely to propagate their useful schemata. However,
it sometimes happens that a solution of low overall fitness contains useful
schemata. The probabilistic replacement of only a proportion of the Solution
Pool with new solutions means that a small number of poor solutions will
survive for sufficient generations that these schemata have a chance of being
incorporated into fitter solutions. This probabilistic approach to survival (when
coupled with non-locality and the use of Genetic Operators) means that the
Genetic Algorithm have an ability to avoid getting stuck on non-optimal peaks in
the Problem Space. Consider a problem space with two peaks, one higher than
the other. A simple hill climbing algorithm, if it happens to start “near” the lower
peak, will climb up it and then be stuck at a non optimal position. By contrast,
there is nothing to prevent the Genetic Operators from producing a new solution
somewhere on the higher peak. Once this happens, there is a possibility of
solutions fitter than those at the top of the lower peak and these will come to
dominate the population. The search process can thus “jump” from one peak to
another which most variants of hill climbing don’t do.
4. It is implicitly parallel. In contrast with the behaviour of serial search algorithms
that operate on a single best solution and improve it further, the Genetic
Algorithm uses a population of solutions and simultaneously explores the area
each occupies in the Problem Space. The results of these explorations are
repeatedly used to modify the direction taken by each solution. The parallelism
arises because the “side effects” of exploring the area surrounding each solution
affect all the other solutions through the functioning of Genetic Operators.
The whole is thus greater than the sum of its parts.
5. It is highly general. The Genetic Algorithm makes relatively few assumptions
about the Problem Space in advance. Instead, it tries to extract the maximum
amount of information from the process of traversing it. For example, non-
evolutionary heuristic search algorithms use features like the gradient (first
differential) which may not be calculable for highly irregular Problem Spaces.
By contrast, in the Genetic Algorithm all operations take place directly on a
representation of the potential solution. The Fitness Function also evaluates
fitness directly from solutions rather than using derived measures. Although no
search technique escapes the fact that all such techniques exploit some properties
of the problem space they are applied upon, in practice Genetic Algorithms are
good at finding acceptable solutions to hard problems (which, in some cases,
defeat other methods), albeit not always the best solution. Ironically, social
evolutionary learning may be better at finding the solutions to difficult problems
than rationality which struggles without high levels of knowledge about envi-
ronmental structure.
18.4.2 The Problem Representation and Initial Population
The most important step in developing a Genetic Algorithm also requires the most
human ingenuity. A good representation for solutions to the problem is vital to
efficient convergence. Some solutions have more obvious representations than
others do. In the Travelling Salesman Problem, for example, the obvious represen-
tation is an ordered list of numbers representing cities. For example, the solution
(1 4 3 2) involves starting at city 1, then going to city 4 and so on. Once a
representation has been developed, a number of solutions are generated and form
the initial population in the Solution Pool. These solutions can be generated
randomly or they may make use of some other (perhaps “quick and dirty”)
algorithm producing better than random fitness. The optimum size of the initial
population depends on the size of the Problem Space. A population of almost any
size will ultimately converge. But the efficiency of the Genetic Algorithm relies on
the availability of useful genetic material that can be propagated and developed by
the Genetic Operators. The larger the initial population, the greater the likelihood
that it will already contain schemata of an arbitrary quality. This must be set against
the increased computational cost of manipulating the larger Solution Pool. The
initial population must also be sufficiently large that it covers the Problem Space
adequately. One natural criterion is that any given point in the Problem Space
should not be more than a certain “distance” from some initial solution. A final
requirement for a good solution representation is that all “genes” should be
similarly important to overall fitness, rather than some “genes” being much more
important than others. Equivalent variations at different positions should have a
broadly similar effect on overall fitness. In the Travelling Salesman Problem all the
positions in the list are equivalent. They all represent cities. The efficiency of the
Genetic Algorithms relies on the exponential propagation of successful schemata
and this efficiency is impaired if schemata differ too much in importance as the
system then becomes “bottlenecked” on certain genes.
18.4.3 The Fitness Function
The Fitness Function is almost as important as the solution representation for the
efficiency of the Genetic Algorithm. It assigns fitness to each solution by reference
to the problem that solution is designed to solve. The main requirement for the
Fitness Function is that it must generate a fitness for any syntactically correct
solution. (These are commonly referred to as “legal” solutions.) In the Travelling
Salesman Problem, an obvious Fitness Function satisfying this requirement would
be the reciprocal of the tour length. The reciprocal is used because the definition of
the problem involves finding the shortest tour. Given this goal we should regard
shorter tours as fitter. More complicated problems like constrained optimisation can
also be handled using the Fitness Function. One approach is simply to reject all
solutions that do not satisfy the constraints. This involves assigning them a fitness
of 0. However, where solutions satisfying the constraints are sparse, a more efficient
method is to add terms to the Fitness Function reflecting the extent of constraint
satisfaction. These “penalty terms” lower the fitness of solutions that fail to satisfy
the constraints but do not necessarily reduce it to zero.
18.4.4 The Process of Reproduction
Reproduction is sometimes classified as a Genetic Operator in that it takes a number

of solutions (the Solution Pool) and produces a new set (the Breeding Pool).
However, it is a Genetic Operator of a special type in that it uses additional
information (the fitness of solutions and the Reproduction Function) in generating
that population. The Reproduction Function links the fitness of individual solutions
and the number of copies they produce. This process mimics the reproductive
success of fitter organisms in biological systems. The number of copies depends
on the type of Genetic Algorithm. Typically the fittest solutions in the Solution Pool
may produce two or three copies while the worst may produce none. In order that
potentially useful “genetic material” be retained, it is important that fitter solutions
do not proliferate too rapidly, nor less fit solutions die out too fast. Despite their low
fitness, poor solutions may contain useful schemata that need to be incorporated
into better solutions. Ensuring “adequate” survival for instrumental efficiency is a
matter of trial-and-error and depends on the problem and the type of Genetic
Algorithm being used. There are two main types of reproduction strategies.
In the first, the Holland-Type algorithm (Holland 1975), the copies of each
solution make up the Breeding Pool as described above. The Breeding Pool thus
contains more copies of fitter solutions. There are two main kinds of Reproduction
Function. The first is proportional fitness: here the number of copies produced for
each solution is equal to the size of the Solution Pool normalised by some function
according to the “share of fitness” accruing to each particular solution. Fitter
solutions, responsible for a larger share of total fitness, produce more copies. This
system is similar to that used in Replicator Dynamics (Vega-Redondo 1996): it is
performance relative to the average that determines the number of offspring. The
second possibility is rank-based fitness. In this case, the number of copies depends on
fitness rank. For example, the fittest five solutions may receive two copies each, the
least fit receive no copies and all others receive one. Both types of function have
probabilistic equivalents. Instead of determining the actual number of copies, the
function can determine the probability of drawing each type. The reproduction
operator is then applied repeatedly, drawing from the probability distribution until
the Breeding Pool is full. Clearly, this will still result in a greater proportion of fitter
solutions in the Breeding Pool. The Reproduction Function can be linear or arbi-
trarily complex. In practice, the “shape” of the Reproduction Function is chosen on
the basis of experience to optimise the performance of the Genetic Algorithm.
The second reproduction strategy, the GENITOR algorithm (Whitley 1989) does
not involve a Breeding Pool. Instead of copying solutions into the Breeding Pool
and then copying the results of Genetic Operators back again, the GENITOR takes
parent solutions sequentially from the Solution Pool, applies Genetic Operators and
returns the offspring immediately to the Solution Pool. The Solution Pool is kept
sorted by rank and new solutions are appropriately placed according to fitness.
A new solution either overwrites the solution with fitness nearest to its own or it is
inserted into the Solution Pool so that all solutions with lower fitness move down
one place and the solution with the lowest fitness is removed altogether. The
GENITOR algorithm ensures that fitter solutions are more likely to become parents
by using a skewed distribution to select them.
The differences between these strategies are instructive. The GENITOR algo-
rithm is more similar to the interaction of biological organisms. The parents produce
offspring that are introduced into a population that probably still contains at least one
parent. Fitness affects which parents will mate, rather than generating offspring from
all individuals in the Solution Pool. Even the “pecking order” interpretation of the
introduction of offspring seems relatively intelligible. By contrast, the Breeding Pool
in the Holland-Type algorithm seems to be an abstraction with little descriptive
plausibility. The Holland-Type algorithm effectively splits the process of reproduc-
tion into two parts: the proliferation of fitter individuals and the subsequent genera-
tion of variation in their offspring. In biological systems, both processes result from
the “same” act of reproduction. Furthermore, the differential production of offspring
emerges from the relative fitness of parents. It is not explicitly designed into the
system. In functional terms, both types of algorithm promote the survival of the
fittest through variation and selective retention. In instrumental terms, one is some-
times more suitable than the other for a particular Problem Space. In descriptive
terms, the GENITOR algorithm seems more appropriate to biological systems.
(It can also be given a more plausible behavioural interpretation in social contexts.)
18.4.5 The Genetic Operators
There are two main types of Genetic Operator that correspond to the biological
phenomena of recombination and mutation. These are the original Genetic
Operators developed by Holland (1975). Recombination Genetic Operators involve
more than one solution and the exchange of genetic material to produce offspring.
The commonest example is the Crossover Operator. Two solutions are broken at the
same randomly selected point (n) and the “head” of each solution is joined to the
“tail” of the other to produce two new solutions as shown in Fig. 18.1. Here ai
identifies an ordered set of k genes from one parent and bi identifies those from
another.
One of the two new solutions is then chosen with equal probability as the
offspring to be placed in the Solution Pool.
Fig. 18.1 The Crossover

Operator a1a2…an an+1…ak a1a2…anbn+1…bk
b1b2…bn bn+1…bk b1b2…bnan+1…ak
SPLIT
Fig. 18.2 The (Point) a1a2…an-1anan+1…ak

Mutation Operator
a1a2…an-1hnan+1…ak
Fig. 18.3 The Inversion

Operator a1a2an an+1...am am+1ak
a1a2anam…an+1am+1ak
REVERSE
Mutation Genetic Operators involve a single solution and introduce new genetic
possibilities. The two main kinds of mutation Genetic Operators used in Genetic
Algorithms correspond to so-called “large scale” and “point” mutation. In the
Mutation Operator (realising point mutation) one gene is altered to another value
from the legal range (selected with equal probability) and shown in Fig. 18.2 as hn.
One commonly used Genetic Operator (corresponding to large scale chromo-
somal mutation) is the Inversion Operator (see Fig. 18.3) which involves reversing
the order of a set of genes between two randomly selected points (n and m) in the
genotype.
The Inversion Operator provides an opportunity to discuss positional effects in
solution representations although these can arise in all Genetic Operators except
point mutation. It is not problematic to invert (reverse) the order in which a section
of a city tour takes place in the Travelling Salesman Problem. However, there may
be problem representations for which we have no reason to expect that the Inversion
Operator will generate solutions that are even syntactically correct let alone fit.
There are two solutions to the production of illegal solutions by Genetic Operators.
One is to use the penalty method. The other is simply to avoid unsuitable Genetic
Operators by design. Positional effects pose particular problems if genes have
different meanings, some representing one sort of object and some another. In
this case, inverted solutions are almost certain not to be legal. This difficulty will be
addressed further in the section on developing the Genetic Algorithm.
In biological systems, recombination ensures that the genes of sexually
reproduced offspring are different from those of both parents. Various forms of
mutation guarantee that entirely new genetic possibilities are also being introduced
continuously into the gene pool. In the Genetic Algorithm, the Genetic Operators
perform the same function, but the probability with which each is applied has to be
tuned to ensure that useful genetic material can be properly incorporated before it is
lost. Typically the Crossover Operator is applied with a high probability to each
solution and the Mutation Operator with a low probability to each gene leading to a
moderate probability of some mutation occurring in each solution. Other Genetic
Operators are applied with intermediate probability. These probabilities are
intended to reflect very approximately the relative importance of each process in
biological systems.
For instrumental uses of the Genetic Algorithm, the setting of probabilities is a
matter of experience. If the probabilities of application are too low, especially for
the Mutation Operator, there is a danger of premature convergence on a local
optimum followed by inefficient “mutation only” search. (In such cases, the
advantages of parallel search are lost and the Genetic Algorithm effectively reverts
to undirected serial search.) By contrast, if the probabilities of application are too
high, excessive mixing destroys useful schemata before they can be combined into
fit solutions.
There is a wide variety of other Genetic Operators discussed in the literature
(Goldberg 1989; Mitchell 1996), some developed descriptively from biological
systems and others designed instrumentally to work on particular problems. The
descriptive use of Genetic Operators in the example provided here means that
although it is important to bear the instrumental examples in mind, they should
not be regarded as definitive. The processes of variation that affect the social
analogues of genotypes should be established empirically just as they were for
biological genes.
18.4.6 Convergence
Because the Genetic Algorithm is a powerful technique, many of the problems it is

used to solve are very hard to tackle by other means. Although it is possible to test
the Genetic Algorithm by comparison with other techniques for simple problems,
there is a danger that conclusions about performance will not scale to more
complex cases. One consequence of this is the difficulty of defining satisfactory
conditions for convergence. Provided the Problem Space is suitable, a non-
evolutionary algorithm will find the best solution within a certain time. In the
same time, the Genetic Algorithm is only statistically likely to converge (though,
in practice, it will actually do so for a far larger class of problems). As a result,
unlike some iterative procedures, the Genetic Algorithm cannot simply be stopped
after a fixed number of generations. Instead, the properties of the Solution Pool
must be analysed to determine when the programme should stop. The simplest
method involves stopping when the fittest solution is “good enough”. Clearly, this
involves a value judgement external to the definition of the problem. Another
possibility is to stop the programme when the rate of change in best solution fitness
drops below a specified level. Unfortunately the behaviour of the Genetic Algo-
rithm means that improvements in fitness are often “stepped” as the Genetic
Operators give rise to whole ranges of new possibilities to be explored. For this
reason more sophisticated approaches analyse the Solution Pool continuously and
measure fitness in the whole population. Another advantage of this technique is
that it allows for the fact that convergence is never total because of the Mutation
Operator. There is always a certain amount of “mutation noise” in the Solution
Pool even when it has converged.
18.4.7 Developing the Genetic Algorithm
The previous section was intended to provide a summary of the main aspects of
design and a feel for the operation of a typical instrumental Genetic Algorithm (one
that is supposed to solve a pre-defined problem as efficiently as possible). In the
next two subsections, we describe a variety of generalisations that move the Genetic
Algorithm away from the instrumental interpretation and towards the possibility of
realistic description of certain social processes. This involves enriching the syntax
for solution representations, developing formal techniques for analysing the
behaviour of evolutionary models and making various aspects of the evolutionary
process endogenous. The fact that these generalisations develop naturally from
previous discussions suggests that a suitably sophisticated Genetic Algorithm might
serve as a framework for evolutionary models of (carefully chosen) social phenom-
ena. We shall try to show that Genetic Programming (as an extension of Genetic
Algorithms) is particularly suitable for this purpose.
18.4.7.1 Generalising the Solution Representation
In the simplest Genetic Algorithm, the solution representation is just a list of

numbers with a fixed length. Each gene (number) in the genotype (list) represents
an object like a city in the Travelling Salesman Problem. But there is no reason why
the Genetic Algorithm should be limited to solving problems using such a restricted
representation. The enrichment of the syntax for solution representations has
proceeded in three overlapping domains: the computational improvement of
programmes implementing Genetic Algorithms, the incorporation of useful insights
from biology and the study of theoretical requirements for the use of different
solution representations.
Developments of the first sort are those which broaden the capabilities of the
Genetic Algorithm itself. Instead of solutions of fixed length “hard coded” by the
programmer, Goldberg et al. (1990) have developed a “messy” Genetic Algorithm.
This evolves an encoding of optimal length by varying the lengths of potential
solutions as well as their encoding interpretations. Schraudolph and Belew (1992)
have also addressed this problem, developing a technique called Dynamic Parameter
Encoding that changes the solution encoding in response to an analysis of the current
Solution Pool. (This technique avoids the loss of efficiency that results from
premature convergence and the consequent failure of parallel search.) Finally,

Harvey (1993) has stressed the importance of variable length genotypes in systems
that are to display genuine increases in behavioural complexity.
Developments of the second sort have arisen from the study of biological systems.
Smith et al. (1992) have developed a Genetic Algorithm that produces a diverse
coexistent population of solutions in “equilibrium” rather than one dominated by a
single “optimal” solution. In this way, the coexistent population is capable of
generalisation. This approach also forms the basis of the Classifier Systems discussed
in Forrest (1991). Here groups of “IF [condition] THEN [action]” rules form
coexistent data structures that can jointly perform computational tasks. Belew
(1989, 1990) has developed this notion further by considering models in which the
solutions themselves take in information from the environment and carry out a
simple form of learning. Koza (1992b, c) considers the possibility of co-evolution.
This is a process in which the fitness of a solution population is not defined relative to
a fixed environment or Fitness Function but rather in terms of another population. He
applies this technique to game strategy learning by Genetic Programmes. Clearly this
development is important to models of social systems where we can seldom define,
let alone agree, a clear objective ranking of alternative social arrangements. In a
sense, it is the existence of a Fitness Function that identifies instrumental (rather than
descriptive) applications of Evolutionary Algorithms. The exception might be a
model in which different solutions to a problem were created “subconsciously” in
the style of a Genetic Algorithm but were then evaluated “rationally” by an agent.
For an example see Chattoe and Gilbert (1997).
Developments of the third sort involve the adaptation of formal systems such as
grammars to serve as solution representations. Antoinisse has developed a repre-
sentation and set of Genetic Operators that can be used for any problem in which
legal solutions can be expressed as statements in a formal grammar (Antoinisse
1991). Koza (1992a, 1994) has developed a similar though far more general
representation involving the syntax of computer languages. This approach (called
Genetic Programming) will receive detailed discussion in its own section shortly.
18.4.7.2 Making the Process of Evolution Endogenous
So far, most of the Genetic Algorithm generalisations discussed have been instru-
mental in their motivation and use. The abstractions and limitations in the simple
Genetic Algorithm have not been viewed as unrealistic but merely unhelpful (since
they are engineering solutions rather than attempts to describe and understand
complex social behaviour). The interesting question from the perspective of this
chapter is how it is possible to develop simulations based on Evolutionary
Algorithms which are not just instrumentally effective (allowing firms to survive
by learning about their market situation for example) but actually provide a con-
vincing (“descriptive”) insight into their decision processes and the complexity of
the resulting system. At the same time, the powerful self-organising capabilities
of Evolutionary Algorithms may serve to provide an alternative explanation of
observed stability (and instability) in social systems which do not (or cannot) involve
a high level of individual rationality. Despite the instrumental nature of most current
developments in Genetic Algorithms, the trend of these developments suggests an
important issue for the design of descriptive models.
Most of the developments discussed above can be characterised as making
various aspects of the process of evolution endogenous. Instead of exogenous
system level parameters that are externally “tuned” by the programmer for instru-
mental purposes, various parts of the evolutionary process become internalised
attributes of the individual solutions. They need not be represented in the solution
explicitly as numerical parameters. They are parameters in the more general sense
that they alter the process of evolution and may be adjusted by the programmer. For
example, the level of mutation may emerge from some other process (such as
endogenous copying of information through imitation) rather than being “applied”
to the solutions. Co-evolution provides a good example of this approach. In the
instrumental Genetic Algorithm, the Fitness Function is specified by the program-
mer and applied equally to all solutions, producing an answer to some question of
interest. To follow an old Darwinian example, this is equivalent to the deliberate
breeding of particular dog breeds. In co-evolving Genetic Algorithms, as in
biological evolution, there is no fixed Fitness Function. Fitness can only be
measured relative to the behaviour of other agents that constitute an important
part of the environment. This is equivalent to the production of the dog species by
biological evolution. Another example is provided by the Classifier Systems briefly
discussed above. The simple Genetic Algorithm assumes that the fitness of an
individual solution is independent of the fitness of other solutions. In practice, the
fitness of one solution may depend on the existence and behaviour of other
solutions. In biology, this is acknowledged in the treatment of altruism (Becker
1976; Boorman and Levitt 1980) and of group selection (Hughes 1988).
The use of solutions that are syntactically identical also abstracts from another
important feature of evolution. Because the solutions only differ semantically there
is no sense in measuring the relative “cost” of each. By contrast, when solutions
differ syntactically, selection pressure may operate to produce shorter solutions as
well as better ones. In descriptive models, “fitness” no longer measures an abstract
quantity but describes the efficient scheduling of all scarce resources used including
time. The less time is spent making decisions (provided they are sensible) the more
time can be spent on other things. To put this point in its most general terms,
organisms (and firms) are dynamic solutions to a dynamic environment while the
simple Genetic Algorithm is a static solution to a static environment. Since social
environments are dynamic, one way in which social agents can evolve or adapt is by
evolving or adapting their models of that environment. Thus, an important way in
which descriptive models can make the evolutionary process endogenous is by
simulating agents that develop and test their own interpretations of the world in an
evolutionary manner rather than being “gifted” with a fixed set of interpretations or
decision processes by the modeller (Dosi et al. 1999).
The value of making parts of the process specification endogenous can only be
assessed in specific cases using descriptive plausibility as the main criterion. For
example, if the rate of mutation can realistically be treated as fixed over the lifetime
of a given evolutionary process it makes little practical difference whether it is

represented as an extra global parameter or as part of the representation for each
solution. In such cases, instrumental considerations such as computational effi-
ciency may as well decide the matter. By contrast, making fitness endogenous
will probably have a major effect on the behaviour of the system. In particular, there
will be a tension between the descriptive plausibility of this change and the
“instrumental” desirability of convergence to a unique optimum facilitated by an
external Fitness Function.
This aspect of Genetic Algorithm design provides a new insight into the distinc-
tion between instrumental and descriptive models. Instrumental models are those
that allow the programmer to achieve her goals whatever they are. By contrast, the
only goal that is permitted to shape a descriptive model is that of effective
description as determined by empirical evidence. What determines the extent to
which mutation should be modelled as a process inhering in agents is the extent to
which the mutation process inheres in agents. Only once it has been shown that the
mutation rate does not vary significantly across agents should it be represented as an
environmental variable.
To sum up then, Genetic Algorithms constitute a broad class of powerful evolu-
tionary search mechanisms with an active research agenda. Some (but not all) of
the subsequent developments to the basic Genetic Algorithm are valuable to the
descriptive modelling of social systems. (In addition some developments may have
value in the characterisation of models. In the long term it may be possible to prove
formal convergence results for descriptively realistic systems.) We now turn to a
discussion of Genetic Programming, a significant variant of the Genetic Algorithm
based on the idea of “evolving” computer programmes which can both solve
instrumental problems and represent sets of practices agents use to address the
problems their environment creates for them.
18.5 Genetic Programming
The fundamental insight of Genetic Programming (Koza 1992a, 1994) is that

Evolutionary Algorithms do not need to be limited to static representations or
adaptation in a static environment. The approach originated in an instrumental
concern, the possibility of evolving efficient computer programmes rather than
having to design them explicitly (Koza 1991). However, it rapidly became clear
that the power of the technique could be extended to any process which could be
represented as an algorithm provided the fitness of different solutions could be
measured (Koza 1992d). The possibility of developing descriptive models of agents
was also considered early on (Koza 1992c). In most models of this kind, however,
the fitness of the programme representing an agent is assessed by its ability to fulfil
exogenous goals. Agents typically “compete” against the environment on an equal
footing rather than constituting that environment.
Fig. 18.4 An S-expression OR
AND AND
NOT NOT D0 D1
D0 D1
The potential of such an approach is tremendous. It involves the possibility of an

evolutionary process that operates on the richest representation language we can
envisage: the set of computable functions. These functions can model the capability
to collect, abstract, store and process data from the environment, transfer it between
agents and use it to determine action. Furthermore, we know that (in principle at
least) languages within the class of computable functions can also represent impor-
tant features of human consciousness like self-awareness and self-modification of
complex mental representations (Kampis 1991; Metcalfe 1994; Fagin et al. 1995).
A simple example illustrates the most common solution representation used in
Genetic Programming. This can be visualised as a tree structure and translates
exactly into the set of “S-expressions” available in the LISP programming language
(Friedman and Felleisen 1987). This is convenient for programming purposes
because LISP comes already equipped to perform operations on S-expressions and
can therefore easily and efficiently implement suitable Genetic Operators. The tree
structure in Fig. 18.4 is equivalent to the S-expression (OR (AND (NOT D0) (NOT
D1)) (AND D0 D1)). This is the definition of the XOR (exclusive or) function. For
obvious reasons, D0 and D1 are referred to as terminals and the set {AND, OR,
NOT} are referred to as functions. The choice of a suitable set of functions and
terminals (the equivalent of the solution representation in Genetic Algorithms) is a
key part of Genetic Programming. Functions are by no means limited to the logical
operators. They can also include mathematical operators and programming language
instructions. Similarly, terminals can represent numerical (or physical) constants, a
variety of “sensor” inputs from the environment (including the observable actions of
other agents) and “symbolic” variables like “true” and “false”.
The instrumental measurement of fitness involves providing the S-expressions
with different “inputs” (in this case truth values for D0 and D1) and assessing the
extent to which the desired “output” results. For example, in Koza (1991) a
programme to generate random numbers was tested by measuring the statistical
properties of the number sequences it generated and rewarding such features as
uncorrelated residuals. If S-expressions represent agents that are capable of action
in an environment, success can be measured by the ability to modify the relation-
ship between the agent and the environment in a certain way, for example by
following a trail successfully. (The further along the trail an agent gets the fitter
its programme.) It should be noted that the instrumental measurement of fitness

requires a fairly precisely defined problem and solution grammar. On the other
hand, the descriptive modelling of interaction need not. In order to do “well
enough” in the market, a firm only needs to make some profit in every period
sufficient to cover its costs. It may or may not have an internal goal to do better than
this or even to make as much profit as it possibly can but this goal is not required for
its survival (and may, in some cases, actually be counter-productive).
This discussion raises several potential difficulties with the descriptive use of
Genetic Programming. However, these appear to recede on further consideration of
the corresponding solutions to these problems in instrumental applications. The first
difficulty is designing Genetic Operators that are guaranteed to produce meaningful
offspring. In the S-expression representation, it is clear that a cut can be made at any
point on the tree and the crossing of two such fragmented parents will always result
in two legal offspring. However, the price to be paid for this advantage is that
solutions must have a hierarchical form. More complicated function sets, mixing
numerical and logical functions for example, must restrict crossover to prevent such
outcomes as (þ 4 TRUE) or (NOT 6).2
However, given the descriptive interpretation of Genetic Operators, it is plausi-
ble that agents should know the syntactic rules of combination for the set of
terminals and operators they possess. As such the relevant “descriptive” Genetic
Operators may execute rather more slowly than the simple instrumental ones but it
is not unreasonable to suppose that only syntactically correct trees will result.
However, this raises another interesting possibility for using Genetic Operators.
A good illustration is provided by a second difficulty with Genetic Programming,
that of “bloating”. This occurs because Genetic Programmes sometimes grow very
large and contain substantial amounts of syntactically redundant material. (If a tree
is trying to converge on a specific numerical value, for example, any sub trees
evaluating to 0 are syntactically redundant.) Bloating produces a number of
difficulties. Firstly, it slows down the evaluation of trees. Secondly, it becomes
harder to interpret the trees and assess their behavioural plausibility. Finally, it is
descriptively unsatisfactory. We do not expect real human decision processes to
contain pointless operations (although bureaucratically specified production pro-
cesses might, for example). Unfortunately, the obvious solution (the exogenous
penalisation of long solutions) lacks precision. It is not possible to establish how
long solutions to a particular problem “ought” to be without making arbitrary
assumptions. The result is an ungrounded trade-off between length and quality.
An interesting alternative is to introduce “purely syntactic” Genetic Operators.
These take no account of tree fitness but simply look for redundant material within
trees. For example, a Genetic Operator which replaced instances of the pattern
(* constant 0) with 0 would be very simple to implement.
2
There are other ways to solve this, e.g., implementing a strong typing (Montana 1995) or ensuring
that even mixed expressions have a coherent interpretation.
This approach allows firms (for example) to apply plausible syntactic knowledge
to the structure of their decision processes (“rationalisation” in the non pejorative
sense) without compromising the assumption (opposed to extreme economic ratio-
nality) that they cannot evaluate the fitness of a strategy without trying it in the
market.
It also suggests a possible solution to another persistent problem with Genetic
Programmes, that of interpretation. Even quite small trees are often hard to interpret
and thus to evaluate behaviourally. Application of syntactic Genetic Operators may
reduce the tree to a form in which it can be more easily interpreted. Another
approach might be to use a Genetic Programming instrumentally to interpret
trees, giving the greatest fitness to the shortest tree which can predict the output
of a decision process tree to within a certain degree of accuracy. Thus, in principle
at least, the Genetic Programming approach can be extended to include processes
that are behaviourally similar to abstraction and refinement of the decision process
itself.
As in the discussion of Genetic Algorithms above, we have kept this discussion
of Genetic Programming relatively technical with some digressions about its
general relevance to modelling social behaviour. In the final section of this chapter,
we will present some evolutionary models in social science specifically based on
Evolutionary Algorithms. This discussion allows us to move from general to
specific issues about the applicability of biological analogies to social systems. In
particular, we will try to show why models based on Genetic Programming and
some Classifier Systems are more behaviourally plausible than those based on
Genetic Algorithms.
18.6 Example Applications of Evolutionary Algorithms
18.6.1 Example Using Genetic Algorithms: The Arifovic

“Cobweb” Model
Arifovic (1994) is probably responsible for the best-known simulation of this type
representing the quantity setting decisions of firms to show convergence in a cobweb
model.3 She argues that the Genetic Algorithm both produces convergence over a
wider range of model parameters than various forms of rational and adaptive
learning, but also that it mimics the convergence behaviour of humans in experimen-
tal cobweb markets. Arifovic draws attention to two different interpretations of the
Genetic Algorithm and explores the behaviour of both. In the “single population
interpretation”, each firm constitutes a single genotype and the Genetic Algorithm
3
A cobweb model is one in which the amount produced in a market must be chosen before market
prices are observed. It is intended to explain why prices might be subject to periodic fluctuations in
certain types of markets.
operates over the whole market. In the “multiple population interpretation”, each firm
has a number of genotypes representing alternate solutions to the quantity setting
decision and operates its own “internal” Genetic Algorithm to choose between them.
She shows that using a basic Holland-type Genetic Algorithm, neither interpre-
tation leads to convergence on the Rational Expectations equilibrium for the
cobweb market. When she adds her “Election” Genetic Operator however, both
interpretations do so. The Election Operator involves using Crossover but then
evaluating the offspring for profitability on the basis of the price prevailing in the
previous period. The effect of this is to add some “direction” to the application of
Genetic Operators, in fact a hill climbing component. An offspring is only added to
the population if it would have performed better than its parents did in the previous
period. This approach does not require any implausible knowledge as it is based on
past events. However, it appears that the motivation for introducing the Election
Operator is instrumental, namely to ensure perfect convergence to the Rational
Expectations equilibrium (a goal of economic theory rather than a property of real
markets necessarily). Interestingly, the graphs shown in the paper suggest that the
Genetic Algorithm has done very well in converging to a stable (if mutation noise
augmented) price fairly close to the Rational Expectations equilibrium. In fact,
Arifovic shows how the Election Operator endogenously reduces the effective
mutation rate to zero as the system approaches the theoretical equilibrium. She
also points out that the Election Operator does not harm the ability of the Genetic
Algorithm to learn a new equilibrium if the parameters of the cobweb model
change. What she doesn’t explain is why the goal of the model should be to produce
the theoretical equilibrium.
In fact, there are problems with both of her models that serve as instructive
examples in the application of evolutionary ideas. The single population interpre-
tation seems to involve a standard Holland-type Genetic Algorithm even down to a
Breeding Pool that has no behavioural interpretation in real systems. There is also a
problem with the use of Genetic Operators that is general in Genetic Algorithms.
The way in which the bit strings are interpreted is very precise. If one firm uses
Crossover involving the price strategy of another, it is necessary to “gift” a common
representation to all firms and assume that firms know precisely where bit strings
should “fit” in their own strategies. Given the encoding Arifovic uses, inserting a bit
string one position to the left by mistake doubles the price it produces. In descrip-
tive terms, this seems to be the worst of both worlds. It is easy to see how one firm
could charge the same price as another, or (with more difficulty) acquire a “narra-
tive” strategy fragment like “keep investment in a fixed ratio to profit” but not how
firms could come to share a very precise arbitrary representation and copy instances
around exactly. More generally, encoding price in this way is just behaviourally
odd. It is hard to imagine what a firm would think it was doing if it took a “bit” of
one of its prices and “inserted it” into another. Of course, the effect would be to
raise or lower the price, but the way of going about it is very bizarre.
We think the reason for this is that an encoding is not a procedure that is
endogenously evolved. A Genetic Programme that calculates price by taking the
previous price of another firm, adding unit cost and then adding 2 is telling a firm
behaviourally how to determine price. These are “real” procedures given by the
ontology of what a firm knows: the set of operators and terminals. By contrast there
has to be reason why a firm would bother to encode its price as a bit string rather
than just operating on it directly. Unless this encoding is “gifted”, it is not clear how
(or why) the firm would develop it.
The multiple population interpretation is much more plausible in behavioural
terms since the problem representation only needs to be the same within a firm,
although the strangeness of “combining” prices remains. A firm applying Genetic
Operators to its own strategies can reasonably be assumed to know how they are
encoded however.
However, both interpretations come up against a serious empirical problem noted
by Olivetti (1994). Because the Election Operator is effectively a hill-climbing
algorithm, it fails to converge under quite small changes to the assumptions of the
model. In particular, Olivetti shows that the system doesn’t converge when a white
noise stochastic disturbance is added to the demand function. This suggests that
Arifovic has not understood the main advantage of the Genetic Algorithm and her
pursuit of instrumental convergence at the expense of behavioural plausibility is
actually counter-productive. In a sense, this is just a reprise of the previous instru-
mental insight. Genetic Algorithms perform better on difficult problems precisely
because they do not “hill climb” (as the Election Operator does) and can thus “jump”
from one optimum to another through parallel search.
18.6.2 Example Using Classifier Systems: The Moss Price

Setting Model
As discussed briefly above, Classifier Systems consist of sets of “IF [condition]

THEN [action]” rules that can collectively solve problems. They are evolutionary
because new rules are typically generated using a Genetic Algorithm to select,
recombine and mutate the most effective rules in the population. However, there is
one significant (and potentially problematic) difference between Classifier Systems
and Genetic Algorithms or Genetic Programming. This is the allocation of fitness to
the individual rules, frequently using the so-called Bucket Brigade algorithm. This
allows individual rules to “bid” fitness in order to take part in the set that is used to
solve the problem in a particular instance. Rules taking part in a successful outcome
then receive “recompense” also in terms of fitness. Unfortunately, the behavioural
interpretation for this algorithm is not clear. In addition, the system is required to
make decisions about how to “allocate” fitness between participating rules. This is
the “Credit Assignment Problem” recognised in Artificial Intelligence and it is hard
to produce effective general solutions. In particular, rules that are only used
occasionally may nonetheless be essential under specific circumstances. (It is
possible that an instrumental approach and lack of biological awareness have
created this problem but that it is not actually intrinsic to this kind of modelling.
In biological evolution there is no credit assignment. Phenotypic traits stand and fall
together.)
That said, the Classifier System has one definite advantage over both Genetic
Programming and Genetic Algorithms assuming these difficulties can be overcome.
This is that the individual rules may be much simpler (and hence more easily
interpreted behaviourally) than Genetic Programmes. This ease of interpretation
also makes it more plausible that individual rules (rather than sub trees from
Genetic Programmes or very precisely encoded bit strings from Genetic
Algorithms) might be transferred meaningfully between firms either by interpreta-
tion of observable actions or “gossip”. Interestingly, despite their advantages,
Classifier Systems are easily the least applied Evolutionary Algorithms for under-
standing social behaviour and this lacuna offers real opportunities for new research.
In what appears to be one of the earliest applications to firm decision making,
Moss (1992) compares a Classifier System and a (non-evolutionary) algorithm of his
own design on the task of price setting in a monopoly. His algorithm hypothesises
specific relationships between variables in the market and then tests these.
For example, if an inverse relationship between price and profit is postulated, the
firm experiments by raising price and seeing whether profit actually falls. If not, the
hypothesis is rejected and another generated. If it works, but only over a range, then
the hypothesis is progressively refined. The conclusion that Moss draws from this
approach illustrates an important advantage of Genetic Programming over Genetic
Algorithms and (some) Classifier Systems – that its solutions are explicitly based on
process and therefore explanatory. Moss points out that the simple Classifier System
simply evolves a price while his algorithm shows how the firm evolves a represen-
tation of the world that allows it to set a price. Although not doing it quite as
explicitly as his algorithm, a Genetic Programme may incorporate a stylised repre-
sentation of market relationships into the encoding of the decision process.
(Of course, in certain circumstances the firm may lack the operators and terminals
to deduce these relationships adequately or they may not form a reliable basis for
action. In this case, simpler strategies like “price following” – simply setting the
same price as another firm – are likely to result.)
To return to the point made by Moss, all the Classifier System models so far
developed to study firm behaviour seem to be “flat” and “hard coded”. By “flat” we
mean that only a single rule is needed to bridge the gap between information
received and action taken. In practice, the Classifier System paradigm is capable
of representing sets of rules that may trigger each other in complex patterns to
generate the final output. This set of rules may also encapsulate evolved knowledge
of the environment although “hard coding” prevents this. For example, we might
model the production process as a Classifier System in which the rules describe the
microstructure of the factory floor: where each worker went to get raw materials,
what sequence of actions they performed to transform them and where they put the
results. In such a model events (the arrival of a partially assembled computer at your
position on the production line) trigger actions (the insertion of a particular compo-
nent). However, running out of “your” component would trigger a whole other set
of actions like stopping the production line and calling the warehouse. The con-
struction of such “thick” Classifier Systems is a task for future research.
“Hard coding” implies that each rule bridges the gap between input and output in
the same way, suggesting the common representation of Genetic Algorithms with
its attendant behavioural implausibility. In the models described above decision-
makers do not have the option to add to the set of conditions or to change the
mappings between conditions and actions: changing price on the basis of customer
loyalty rather than costs for example. There is nothing in the Classifier System
architecture to prevent this, but all the current models seem to implement the
architecture in a simplified and behaviourally implausible way that makes it more
like a Genetic Algorithm than Genetic Programming in terms of “hard coding” of
representations and decision processes.
18.6.3 Example Using Genetic Programming: An Artificial

Stock Market
In this example (Edmonds 2002) there is a simulated market for a limited number of
stocks, with a fixed number of simulated traders and a single “market maker”. Each
trader starts off with an amount of cash and can, in each trading period, seek to buy
or sell each of the kinds of stock. Thus at any time a trader might have a mixture of
cash and amounts of each stock. A single market maker sets the prices of each stock
at the beginning of each trading period depending on the last price and the previous
amount of buying and selling of it. The ‘fundamental’ is the dividend paid on each
stock, which for each stock is modelled as a slowly moving random walk. There is a
transaction cost for each buy or sell action by the traders. Thus there is some
incentive to buy and hold stocks and not trade too much, but in general more money
can be made (or lost) in short-term speculation. The market is endogenous except
for the slowly changing dividend rate so that the prices depend on the buy and sell
actions and a trader’s success depends on “out-smarting” the other traders.
In the original Artificial Stock Market model (Arthur et al. 1997) each artificial
trader had a fixed set of price prediction strategies. At each time interval they would
see which of these strategies was most successful at predicting the price in the
recent past (fixed number of time cycles) and rely on the best of these to predict the
immediate future price movements. Depending on its prediction using this best
strategy it would either buy or sell.
In the model presented here each trader has a small population of action
strategies for each stock, encoded as a GP tree. In each time period each artificial
trader evaluates each of these strategies for each stock. The strategies are evaluated
against the recent past (a fixed number of time cycles) to calculate how much value
(current value based on cash plus stock holdings at current market prices) the trader
would have had if they had used this strategy (taking into account transactions costs
and dividends gained), assuming that the prices were as in the recent past. The
trader then picks the best strategy for each stock and (given constraints of cash and
holdings) tries to apply this strategy in their next buy and sell (or hold) actions.
At the end of each trading period the set of action strategy trees are slightly
evolved using the GP algorithm. That is to say that they are probabilistically
“remembered” in the next trading round depending on their evaluated success,
with a few of them crossed in a GP manner to produce new variations on the old
strategies and very few utterly new random strategies introduced. As a result of this
a lot of evolution of small populations is occurring, namely a population for each
trader and each stock. Here each GP tree represents a possible strategy that the
trader could think of for that stock. The Genetic Programming algorithm represents
the trader’s learning process for each stock, thinking up new variations of
remembered strategies, discarding strategies that are currently unsuccessful and
occasionally thinking up completely novel strategies. This is a direct implementa-
tion of Campbell’s model of creative thinking known as “Blind Variation and
Selective Attention” (Campbell 1965). A new trader will have relatively poor
strategies generally and will not necessarily have the feedback to choose the most
appropriate strategy for a new stock. By contrast, an expert will have both a good
set of strategies to choose from and better judgement of which to choose. These
aspects of social (evolutionary) learning are clearly important in domains where
there is genuine novelty which many traditional approaches do not handle well (or
in some cases at all.)
The nodes of the strategy trees can be any mixture of appropriate nodes and
types. This model uses a relatively rich set of nodes, allowing arithmetic, logic,
conditionals, branching, averaging, statistical market indices, random numbers,
comparisons, time lags and the past observed actions of other traders. With a
certain amount of extra programming, the trees can be strongly typed (Haynes
et al. 1996), i.e. certain nodes can take inputs that are only a specific type (say
numeric) and output a different type (say Boolean) – for example the comparison
“greaterThan”. This complicates the programming of the Genetic Operators but can
result in richer and more specific trees.
Below are a couple of examples of strategies in this version of the stock market
model. The output of the expression is ultimately a numeric value which indicates
buy or sell (for positive or negative numbers), but only if that buy or sell is of a
greater magnitude than a minimal threshold (which is a parameter, allowing for the
“do nothing” – hold – option).
• [minus [priceNow ‘stock-1’] [maxHistoricalPrice ‘stock-1’]] – Sell if price is
greater than the maximum historical price otherwise buy;
• [lagNumeric [2] [divide [doneByLast ‘trader-2’ ‘stock-3’] [indexNow]]] – Buy or
sell as done by trader-2 for stock-3 divided by the price index 3 time periods ago.
The field of Evolutionary Computation is primarily concerned with the effi-
ciency and effectiveness of its algorithms in solving explicitly posed problems.
However, efficiency is not the primary consideration here but rather how to make
such algorithms correspond to the behaviour of observed social actors. In this
model, a large population of strategies within each individual would correspond
to a very powerful ability in a human to find near-optimal strategies, which is
clearly unrealistic. Thus a relatively small population of strategies is “better” since
it does mean that particular traders get ‘locked-in’ to a narrow range of strategies
for a period of time (maybe they all do so badly that a random, novel strategy does
better eventually). This allows the emergence of “group think” and trading “styles”
that can reasonably be anticipated in real markets.
Other relevant issues might be that traders are unlikely to ever completely
discard a strategy that has worked well in the past. (Many evolutionary models
fail to take account of the fact that humans are much better at recall from structured
memory than they are at reasoning. Such a model might thus “file” all past
strategies but only have a very small subset of the currently most effective ones
in live memory. However, if things started going very badly, it would be easy to
choose not from randomly generated strategies but from “past successes”. It is an
interesting question whether this would be a more effective strategy.) Clearly
however the only ultimate tests are whether the resulting learning behaviour
sufficiently matches that of observed markets and whether the set of operators
and terminals can be grounded in (or at least abstracted from) the strategies used by
real traders. (Either test taken alone is insufficient. Simply matching behaviour may
be a coincidence while “realistic” trading strategies that don’t match behaviour
have either been abstracted inappropriately or don’t really capture what traders do.
It is quite possible that what they are able to report doing is only part of what they
actually do.)
Given such a market and trader structure what transpires is a sort of learning
“arms-race” where each trader is trying to “out-learn” the others, detecting the
patterns in their actions and exploiting them. The fact that all agents are following
some strategy at all times ensures that (potentially) there are patterns in existence to
be out-learned. Under a wide range of conditions and parameter settings one readily
observes many of the qualitative patterns observed in real stock markets – specula-
tive bubbles and crashes, clustered volatility, long-term inflation of prices and so
on. Based on the simulation methodology proposed by Gilbert and Troitzsch (2005)
and the idea of generative social science put forward by Epstein (2007), this
outcome shows how a set of assumptions about individual actions (how traders
implement and evolve their strategies) can potentially be falsified against aggregate
properties of the system such as price trends across the range of stocks. Such models
are an active area of research; a recent Ph.D., which surveys these, is (Martinez-
Jaramillo 2007).
18.6.4 Example: The Functional Survival of “Strict” Churches
There are clear advantages to using existing evolutionary algorithms to understand

complex social processes as we hope we have shown through the examples above.
Apart from an opportunity to discuss the “technicalities” of evolutionary algorithms
through looking at simple cases, it is valuable to have programs that can be used
“off the shelf” (rather than needing to be developed from scratch) and for which
there is an active research agenda of technical developments and formal analysis
which can be drawn on. However, the major downside of the approach has also
been hinted at (and will be discussed in more detail in the conclusion). Great care
must be exercised in choosing a domain of application for evolutionary algorithms
in understanding complex social systems. The more an evolutionary algorithm is
used “as is”, the smaller its potential domain of social application is likely to be.
Furthermore, while it is possible, by careful choice of the exact algorithm, to relax
some of the more socially unhelpful assumptions of evolutionary algorithms (the
example of an external Fitness Function and a separate Breeding Pool have already
been discussed), the danger is that some domains will simply require too much
modification of the basic evolutionary algorithm to the point where the result
becomes awkward or the algorithm incoherent. (A major problem with existing
models has been the inability of their interpretations to stand up to scrutiny. In some
cases, such as the Election Operator proposed by Arifovic, it appears that even the
designers of these models are not fully aware of the implications of biological
evolution.)
As suggested at the beginning of the chapter, the other approach, formerly rare
but now increasingly popular is to start not with an evolutionary algorithm but with
a social system and build a simulation that is nonetheless evolutionary based on the
structure of that. The challenge of choosing domains with a clear analogy to
biological evolution remains but is not further complicated by the need to unpick
and redesign the assumptions of an evolutionary algorithm. Such an example of a
“bespoke” evolutionary simulation is provided in this section.
Iannaccone (1994) puts forward an interesting argument to explain the poten-
tially counter-intuitive finding that “strict churches are strong”. It might seem that a
church that asked a lot of you, in terms of money, time and appropriate behaviour,
would be less robust (in this consumerist era) than one that simply allowed you to
attend on “high days and holidays” (choosing your own level of participation).
However, the evidence suggests that it is the liberal churches that are losing
members fastest. Iannaccone proposes that this can be explained by reflecting on
the nature of religious experience. The satisfaction that people get out of an act of
worship depends not just on their own level of involvement but also that of all other
participants. This creates a free rider problem for “rational” worshippers. Each
would like to derive the social benefit while minimising their individual contribu-
tion. Churches are thus constantly at the mercy of those who want to turn up at
Christmas to a full and lively church but don’t want to take part in the everyday
work (like learning to sing the hymns together) that makes this possible.
Iannaccone then argues that an interesting social process can potentially deal
with this problem. If we suppose that churches do things like demanding time,
money and appropriate behaviour from their worshippers, this affects the satisfac-
tion that worshippers can derive from certain patterns of activity. If the church can
somehow make non-religious activities less possible and less comfortable, it shifts
the time allocations of a “rational” worshipper towards the religious activities and
can simultaneously reward him or her with the greater social benefit that comes
from the church “guiding” its members in this way. To take a mildly contrived
example, Muslims don’t drink alcohol. They also dress distinctively. A Muslim
who wanted to drink couldn’t safely ask his friends to join him, could easily be seen
entering or leaving a pub by other Muslims and would probably feel out of place
and uncomfortable once inside (quite apart from any guilt the church had managed
to instill). The net effect is that Muslims do not spend much time in pubs (while
many others in the UK do) and have more time for religious activity. Of course, it is
easy to pick holes in the specifics of Iannaccone’s argument. Why would the
Muslim not dress up in other clothes? (That itself might need explanation though.)
Why not engage in another non religious activity that was not forbidden? Why
assume that only religious activities are club goods? (Isn’t a good night at the pub
just as much a result of collective effort?)
However, regardless of the details, the basic evolutionary point is this. Religious
groups set up relatively fixed “creeds” that tell members when and how to worship,
what to wear and eat, how much money must be given to the church and so on.
Given these creeds worshippers join and leave churches. To survive, churches need
worshippers and a certain amount of “labour” and income to maintain buildings,
pay religious leaders and so on. Is it in fact the case, as Iannaccone argues that the
dynamics of this system will result in the differential survival of strict churches at
the expense of liberal ones? This is in, fact, a very general framework for looking at
social change. Organisations like firms depend on the ability to sell their product
and recruit workers in a way that generates profit. Organisations like hospitals are
simultaneously required to meet external goals set by their funders and honour their
commitments to their “customers”: On one hand, the budget for surgery may be
exhausted. On the other, you can’t turn away someone who is nearly dead from a
car crash knowing they will never survive to the next nearest accident and emer-
gency department. This evolutionary interplay between organisations facing exter-
nal constraints and their members is ubiquitous in social systems.
Before reporting the results and discussing their implications, two issues must be
dealt with. Because this is a “two sided” process (involving worshippers and
churches) we must attend to the assumptions made about the behaviour of these
groups. In the model discussed here, it was assumed that churches were simply
defined by a fixed set of practices and did not adapt themselves. This is clearly a
simplification but not a foolish one. Although creeds do adapt, they often do so over
very long periods and this is a risky process. If worshippers feel that a creed is just
being changed for expedience (rather than in a way consistent with doctrine) they
may lose faith just as fast as in a church whose creed is clearly irrelevant to changed
circumstances. Speculatively, the great religions are those that have homed in on
the unchanging challenges and solutions that people face in all times and all
places while the ephemeral ones are those that are particular to a place or set of
circumstances. Conversely, the model assumes that worshippers are strictly rational
in choosing the allocations of time to different activities that maximise their satis-
faction. Again, this assumption isn’t as artificial as it may seem. Although we do not
choose religions like we choose baked beans, there is still a sense in which a religion
must strike a chord in us (or come to do so). It is hard to imagine that a religion that
someone hated and disbelieved in could be followed for long merely out of a sense
of duty. Thus, here, satisfaction is being used in a strictly subjective sense without
inquiring into any potential objective correlates. This life, for me, is better than that
life. In terms of predicting individual behaviour, this renders satisfaction a truism but
in the context of the model (and explaining the survival of different kinds of
churches) what matters is not what people happen to like but the fact that they
pursue it. To sum up, we could have represented the churches as more adaptive and
the worshippers as less adaptive but since we are interested in the interplay of their
behaviours (and, incidentally, this novel approach reveals a shortage of social
science data about how creeds change and worshippers participate in detail), there
is no definite advantage to doing so.
In a nutshell, the model works as follows (more details can be found in Chattoe
2006a). Each agent allocates their time to activities generating satisfaction
(and different agents like different things to different extents). They can generate
new time allocations in two main ways. One is by introspection, simply reflecting that
a bit more of this and a bit less of that might be nicer. The other is by meeting other
agents and seeing if their time allocations would work better. This means, for
example, that an agnostic who meets a worshipper from church A may suddenly
realise that leading their life in faith A would actually be much more satisfying than
anything they have come up with themselves. Conversely, someone “brought up in”
church B (and thus mainly getting ideas from other B worshippers about “the good
life”) may suddenly realise that a life involving no churchgoing at all is much better
for him or her (after meeting an agnostic). Of course, who you meet will depend on
which church you are in and how big the churches (and agnostic populations) are. It
may be hard to meet agnostics if you are in a big strict church and similarly, there are
those whom a more unusual religion might suit very well who will simply not
encounter its creed. Churches are created at a low rate and each one comes with a
creed that specifies how much time and money members must contribute and how
many non-religious activities are “forbidden”. Members can only have time
allocations that are compatible with the creed of the church. These allocations
determine the social benefits of membership discussed above. If a church cannot
meet minimum membership and money constraints, it disappears. Thus, over time,
churches come and go, differing in their “strictness”, and their survival is decided by
their ability to attract worshippers and contributions. Worshippers make decisions
that are reasonable (but not strictly rational in that they are not able instantaneously to
choose the best time allocation and church for them – which may include no church –
for any state of the environment). This system reproduces some stylised facts about
religion. New churches start small and are often (but not always) slow to grow.
Churches can appear to fade and then experience resurgences. There are a lot of small
churches and very few large ones.
What happens? In fact, there is almost no difference between the lifetimes of
liberal churches and mildly strict ones. What is clear however is that very strict
churches (and especially cults – which proscribe all non-religious activities) do not
last very long at all. It is important to be clear about this as people often confuse
membership with longevity. It is true that strict churches can grow very fast and (for
a while) very large but the issue at stake here is whether they will survive in the long
term. To the extent that the assumptions of the simulation are realistic, the answer
would appear to be no. Thus we have seen how it is possible to implement a
reasonably coherent biological analogy in a social context without using a pre-

existing evolutionary algorithm.
18.7 Conclusion: Using Biological Analogies to Understand

Social Systems
Having presented a number of case studies of evolutionary algorithms in different

application areas, we are now in a position to draw some general conclusions about
the design and use of evolutionary simulations. Despite the fact that some have
claimed that a generalised version of evolution (Blind Variation and Selective
Retention) is the basic template for human creativity (Campbell 1965) and that it
is plausible that some processes similar to biological evolution do occur in human
societies, it is unlikely that these processes will be direct translations of biological
evolution in all its details. For this reason, we would propose that research into
evolutionary models proceeds as follows (although it is inevitable that there will be
some backward and forward interplay between the stages for reasons discussed
below):
1. Start with your substantive research domain of interest (linguistics, stock
markets, the rise and fall of religious groups) and consider the general arguments
for representing these (or parts of them) in evolutionary terms. While it is seldom
spelt out explicitly, there are actually rather few candidate “general social
theories” to explain the dynamic interaction of choice and change. Unless one
believes that individuals have the power and knowledge required for rational
action to benefit them (and note that this condition isn’t met in situations as
simple as the two person one shot Prisoner’s Dilemma), evolution is really the
only coherent and completely specified theory available.4 Thus (and obviously
the authors are biased in this regard) if you believe that agents act on imperfect
knowledge in an independently operating5 environment (such that there often is
a gap between what you expect to happen and what happens, however effec-
tively you collect and process data about your environment), it is worth consid-
ering an evolutionary approach. We would argue that these conditions are met in
most social settings but economists would disagree.
2. Consider the explicit specification of an evolutionary process for your particular
domain of research (perhaps using the four process specification above as a
guide). The key choice made in this context is a “coherent” object of selection
(OOS) whose presence or absence is empirically accessible. This makes
4
In fact, it might be argued that it is the only one. Rational choice cannot contend with novelty or
the origin of social order. By focusing on relative performance, no matter how absolutely poor,
evolution can produce order from randomness.
5
This independence comes both from other social actors and physical processes like climate and
erosion.
organisations and firms with any kind of formal status particularly suitable. For
informal groups like families, for example, it is much less clear what constitutes a
“unit”. (Is it, in a traditional society setting, that they physically survive, or, in a
modern setting, that they still cohabit or are still on speaking terms? The problems
here are evident.) Interestingly, individuals (while obviously “physically” coher-
ent) are still problematic as objects of selection. Unless the model involves “bare”
survival, it is less obvious what happens when an agent is “selected”. However,
examples still exist, such as who is trading in particular markets. Most of the rest
of the evolutionary process specification follows naturally from the choice of an
OOS. It then becomes fairly clear what the resource driving selection is (food for
tribal groups, profit for firms, membership for voluntary organisations, attention
for memes), what causes the birth and death of OOS (sexual reproduction, merger,
religious inspiration, bankruptcy, lack of interest or memorability and so on) and
what variation occurs between OOS.
This last is an interesting area and one where it is very important to have a
clearly specified domain of application. For example, consider industrial
organisation. Textbook economic theory creates in the mind an image of the
archetypal kettle factory (of variable size), selling kettles “at the factory gates”
directly to customers and ploughing profits straight back into growth and better
technology. In such a world, a firm that is successful early on can make lots of
poor judgements later because it has efficient technology, market dominance,
retained profit and so on. As such, evolutionary pressure rapidly ceases to operate.
Further, this kind of firm does not “reproduce” (it merely gets larger) and even
imitation of its strategy by other firms (that are smaller and poorer) may not cause
the effective “spread” of social practices required by an evolutionary approach.
(What works for the dominant firm may actually be harmful to smaller
“followers”.)
By contrast, we can see the more modern forms of competition by franchises
and chains (Chattoe 1999) or the more realistic detail of “supply chain produc-
tion” as much more naturally modelled in evolutionary terms. In the first case,
firms do directly “reproduce” a set of practices (and style of product, décor,
amount of choice and so on) from branch to branch. More successful chains have
more branches. Furthermore, the “scale” of competition is determined by the
number of branches and it is thus reasonable to say that successful business
practices proliferate. Wimpy may drive out “Joe Smith’s Diner” from a particu-
lar town but Joe Smith is never a real competitor with the Wimpy organisation
even if he deters them from setting up a branch in that town. This means that
selection pressure continues to operate with chains at any scale competing with
other chains at similar scales. Short of outright monopoly, there is never a
dominant market position that is stable.6
6
This is probably because the market is spatially distributed and the only way of making additional
profits is by opening more branches (with associated costs). There are no major economies of scale
to be exploited as when the kettle factory simply gets bigger and bigger with all customers
continuing to bear the transport costs.
In the second case, we can see how open ended evolution may create new
opportunities for business and that supply chains as a whole constitute
“ecologies” (Chattoe-Brown 2009). Initially, each firm may transport its own
goods to market but once markets are sufficiently distant and numerous, there
may be economies of scale in offering specialist transport and logistics services
(for example, all goods going from Bristol to Cardiff in 1 week may be carried
by a single carter or a firm may create a distribution infrastructure so not all
goods are transported directly from origin to destination but via cost saving
looped routes.) Again, it is clear how the organisations here must operate
successful practices that satisfy both suppliers (those who want to deliver
goods) and customers (those who want to receive them) and, further, how the
nature of the business environment may change continuously as a consequence
of innovation (whether technical or social). The creation of the refrigerated ship
or the internal combustion engine may foreclose some business opportunities
(like raising animals in the city or harness making) and give rise to others which
may or may not be taken up (spot markets, garages).
These examples show several things. Firstly, it is necessary to be very clear
what you are trying to understand as only then can the fitness of the evolutionary
analogy be assessed. Secondly, it is useful to have a systematic way (Chattoe
1998, 2006b) of specifying evolutionary models since these stand or fall on their
most implausible assumptions (particularly in social sciences which aren’t very
keen on this approach).7 Thirdly, there are a lot more opportunities for evolution-
ary modelling than are visible to the “naked eye”, particularly to those who take
the trouble to develop both domain knowledge and a broad evolutionary perspec-
tive. Considering the ubiquity of branch competition and intermediate production
in the real world, the economic literature is amazingly distorted towards the
“autonomous kettle factory view” and simulation models of realistic market
structures are scarcer still (though this is just starting to change). The price of
adopting a novel method is scepticism by one’s peers (and associated difficulties
in “routine” academic advancement) but the rewards are large domains of unex-
plored research opportunities and the consequent possibility for real innovation.
Finally, whilst it is always possible to use an evolutionary algorithm as a “black
box learning system” within the “mind” of an agent or organisation, there is a
design issue about interpreting this kind of model discussed previously. The
choice of learning algorithm can make a crucial difference in simulations
(Edmonds and Moss 2001) and one cannot simply assume that any learning
algorithm will do.
3. Explore whether data for your chosen domain is available (or can readily be got
using standard social science methods).8 If it is available, does it exist at both the
7
More informally, “the assumptions you don’t realise you are making are the ones that will do you
in”.
8
In a way, it is a black mark against simulation that this needs to be said. Nobody would dream of
designing a piece of statistical or ethnographic work without reference to the availability or
accessibility of data!
individual level and in aggregate? For example, is there observational data about
firm price setting practices (in board meetings for example) and long term
historical data about the birth, death and merger of firms in a particular industry
and their prices over time? Because simulation is a relatively new method, it is
still possible to build and publish exploratory (or less flatteringly “toy”) models
of evolutionary processes but it is likely to get harder and may become impossi-
ble unless the evolutionary model or the application domain is novel. It is almost
certainly good scientific practice to make the accessibility of data part of the
research design but it does not follow from this that only models based on
available data are scientific. The requirement of falsifiability is served by the
data being collectable “in principle”, not already collected. The best argument to
support claims of scientific status for a simulation is to consider (as a design
principle) how each parameter could be calibrated using existing data or existing
research methods. (The case is obviously weaker if someone has to come up with
a new data collection method first although it helps if its approach or
requirements can be sketched out a priori.)
This aspect of research design also feeds into the decision about whether to
use an existing evolutionary algorithm and if so, which one. The emerging
methodology of social simulation (Gilbert and Troitzsch 2005, pp. 15–18,
Epstein 2007) is to make a set of empirically grounded hypotheses at the
micro level (firms set prices thus) and then to falsify this ensemble at the
macro level. (The real distribution of survival times for firms is thus: It does or
does not match the simulated distribution of survival times produced by the
model.) A problem will arise if it is hard to interpret the simulated price setting
practices. Suppose, for example, we use GP to model the evolution of trading
strategies in stock markets. We may use interviews or observation of real traders
to decide what terminals and operators are appropriate but, having let the
simulation run and observed plausible aggregate properties, we may still not
know (and find it extremely hard to work out because of the interpretation issue)
whether the evolved strategies used are actually anything like those which
traders would (or could) use.
Equating the empirical validation of the GP grammar with the validation of
the strategies evolved from it is bit like assuming that, because we have derived a
Swahili grammar from listening to native speakers that we are then qualified to
decide when Swahili speakers are telling the truth (rather than simply talking
intelligibly). It muddles syntax and semantics. The design principle here is then
to consider how the chosen evolutionary algorithm will be interpreted to estab-
lish the validity of evolved practices. (Creative approaches may be possible here
like getting real traders to design or choose GP trees to trade for them or getting
them to “critique” what are effectively verbal “translations” of strategies derived
from apparently successful GP trees as if they were from real traders.) In this
regard, it is their potential ease of interpretation that makes the relative neglect
of CS models seem more surprising in evolutionary modelling.
4. Having first got a clear sense of what needs to be modelled, it is then possible to
choose a modelling technique in a principled way. As the analysis of case studies
suggest, the danger with a “methods-led” approach is that the social domain will
be stylised (or simply falsified) to fit the method. A subsidiary difficulty with the
methods-led approach is that even if the researcher is wise enough to use a
modified evolutionary algorithm to mirror a social process accurately (rather
than distorting or abstracting the domain to fit the method), inadequate technical
understanding may render the modified algorithm incoherent or ineffective. It is
thus very important to understand fully any methods you plan to apply particu-
larly with regard to any instrumental assumptions they contain. (In making
convergence her goal for the GA cobweb model, Arifovic introduced an election
operator which actually rendered the GA less effective in solving hard problems.
This issue would probably have been foreseen in advance by a competent
instrumental user of the GA technique. The muddle arose from the interface
between social description and the GA as a highly effective instrumental
optimisation device.)
Having chosen a modelling technique, all its supporting assumptions must
also be examined in the light of the application domain. For example, it is very
important not to confuse single and multiple population interpretations of a GA:
Do firms each have multiple candidate pricing strategies and choose them by an
evolutionary process or is there one overall evolutionary process, in which single
pricing strategies succeed and fail with the associated firms “carrying” them?
Each model (or some combination) might be justified on empirical grounds but
only if the difference in interpretation is kept clearly in mind. Although we are
sceptical that systems of realistic social complexity would allow this, the
principled choice of methods means that it is even possible that some domains
would not require simulation at all but could be handled by mathematical models
of evolution like replicator dynamics (Weibull 1995) or stochastic models
(Moran 1962).
By contrast, however, if the properties of the chosen social domain are too far
from a standard evolutionary algorithm (such that it can neither be used whole-
sale or deconstructed without collapsing into incoherence), the best solution is to
build a bespoke evolutionary model as was done for the “strict churches” case
study. (At the end of the day, evolutionary algorithms were themselves “evolved”
in a completely different engineering environment and we would not therefore
expect them to apply widely in social systems. Thus, great care needs to be taken
to use them only where they clearly do apply and thus have real value.) With free
and widely used agent based modelling packages like NetLogo9 and associated
teaching materials (Gilbert and Troitzsch 2005; Gilbert 2007), this is now much
easier than it was. Ten years ago, one reason to use an existing algorithm was
simply the significant cost of building your own from scratch. To sum up this
strategy of research, the decision to use, modify or build an evolutionary algo-
rithm from scratch should be a conscious and principled one based on a clear
understanding of the domain and existing social science data about it.
9
http://ccl.northwestern.edu/netlogo/
The final piece of advice is not technical or methodological but presentational. In

applying a novel method, be prepared to suffer equally at the hands of those who
don’t understand it and those who do! One of the hardest things to do in academia is
to strike a balance between rejecting ill founded criticisms or those that translate to
“I just don’t like this” without also rejecting real objections that may devalue
months (or even years) of your effort (and still, frustratingly for you, be part of
“good science”). To judge criticisms in a novel area, you must be especially well
informed and thus confident of your ground. For example, there is no clear cut
evidence for Lamarckism (modification of the genotype by the phenotype during
the life of the organism in a way that can then be transmitted by reproduction) in
biology but in social systems such processes are ubiquitous. (Someone discovers a
good way to discipline children. Those who were thus disciplined do the same thing
to their children. This is an acid test because, with hindsight, the “victims” have to
see it as beneficial, and thus not have been warped by it, even if it was hateful at the
time. Punishments so nasty that the victims won’t inflict them, or so ineffective that
the parents stop bothering, will die out.) Failure to understand this issue may either
set you on the path of non-Lamarckian (and thus quite possibly implausible)
evolutionary models of social systems or of apologising mistakenly for building
Lamarckian models which don’t “truly” reflect biological evolution (when that was
never the design criterion for using biological analogies in social science anyway).
The best way to address these issues is hopefully to follow the systematic
procedure outlined above. This minimises the chances that you will miss things
which critics can use to reject your models (and if they are hostile enough, your
whole approach) and ensures that by justifying the models to yourself, you can
actually justify them to others. In popular scientific folklore Darwin (still the
greatest evolutionist) spent a considerable period trying to anticipate all possible
objections to his theory and see how valid they were (and what counters he could
provide) before he presented his work. Given how fraught the acceptance of his
theory has been anyway, imagine if he had not troubled to take that step!
We hope we have shown, by the use of diverse case studies and different
evolutionary modelling techniques, both the considerable advantages and (poten-
tially avoidable) limitations of this approach and encourage interested readers to
take these ideas forward both in developing new kinds of models and applying
evolutionary models to novel domains. The field is still wide open and we are always
pleased to hear from potential students, co-workers, collaborators, supporters or
funders!
Acknowledgements Edmund Chattoe-Brown acknowledges the financial support of the Eco-

nomic and Social Research Council as part of the SIMIAN (http://www.simian.ac.uk) node of the
National Centre for Research Methods (http://www.ncrm.ac.uk).
Further Reading
(Gilbert and Troitzsch 2005) is a good general introduction to social science

simulation and deals with evolutionary techniques explicitly, while (Gilbert 2007)
is recommended as an introduction of this kind of simulation for studying evolution
in social systems. For deeper introductions to the basic techniques see (Goldberg
1989), which is still an excellent introduction to GA despite its age (for a more up-
to-date introduction see: Mitchell 1996), and (Koza 1992a, 1994) for a very
accessible explanation of GP with lots of examples. (Forrest 1991) is a good
introduction to techniques in Classifier Systems.
More details about the four example models are given in the following: (Chattoe
2006a) shows how a simulation using an evolutionary approach can be related to
mainstream social science issues, (Edmonds 2002) gives an example of the appli-
cation of a GP-based simulation to an economic case, and (Moss 1992) is a
relatively rare example of a classifier based model.
References
Antoinisse H (1991) A grammar based genetic algorithm. In: Rawlins G (ed) Foundations of
genetic algorithms: proceedings of the first workshop on the foundations of genetic algorithms
and classifier systems, Indiana University. Morgan Kaufmann, San Mateo, 15–18 July 1990,
pp 193–204
Arifovic J (1994) Genetic algorithm learning and the cobweb model. J Econ Dyn Control 18:3–28
Arthur WB, Holland JH, LeBaron B, Palmer R, Tayler P (1997) Asset pricing under endogenous
expectations in an artificial stock market. In: Arthur WB, Durlauf SN, Lane DA (eds) The
economy as a complex evolving system II (Santa Fe Institute studies in the science of
complexity, proceedings volume XXVII) Addison-Wesley, Reading, pp 15–44
Becker G (1976) Altruism, egoism and genetic fitness: economics and sociobiology. J Econ Lit
14:817–826
Belew R (1989) When both individuals and populations search: adding simple learning to the
genetic algorithm. In: Schaffer J (ed) Proceedings of the third international conference on
genetic algorithms, George Mason University. Morgan Kaufmann, San Francisco, 4–7 June
1989, pp 34–41
Belew R (1990) Evolution, learning and culture: computational metaphors for adaptive search.
Complex Syst 4:11–49
Blackmore S (1999) The meme machine. Oxford University Press, Oxford
Boorman S, Levitt P (1980) The genetics of altruism. Academic, St Louis
Boyd R, Richerson PJ (1985) Culture and the evolutionary process. University of Chicago Press,
Chicago
Buss DM (1998) Evolutionary psychology: the new science of the mind. Allyn and Bacon, Boston
Calvin W (1996a) How brains think: evolving intelligence, then and now. Basic Books, New York
Calvin W (1996b) The cerebral code: thinking a thought in the mosaics of the mind. MIT Press,
Cambridge, MA
Campbell DT (1965) Variation and selective retention in socio-cultural evolution. In: Barringer
HR, Blanksten GI, Mack RW (eds) Social change in developing areas: a reinterpretation of
evolutionary theory. Schenkman, Cambridge, MA, pp 19–49
Campbell DT (1974) Evolutionary epistemology. In: Schlipp PA (ed) The philosophy of Karl R.
Popper, vol XIV, The library of living philosophers. Open Court, LaSalle, pp 412–463
Cavalli-Sforza L, Feldman M (1973) Cultural versus biological inheritance: phenotypic transmis-
sion from parents to children. Hum Genet 25:618–637
Chattoe E (1998) Just how (un)realistic are evolutionary algorithms as representations of social
processes? J Artif Soc Soc Simul 1(3). http://www.soc.surrey.ac.uk/JASSS/1/3/2.html
Chattoe E (1999) A co-evolutionary simulation of multi-branch enterprises. Paper presented at the
European meeting on applied evolutionary economics, Grenoble, 7–9 June. http://webu2.
upmf-grenoble.fr/iepe/textes/chatoe2.PDF
Chattoe E (2002) Developing the selectionist paradigm in sociology. Sociology 36:817–833
Chattoe E (2006a) Using simulation to develop and test functionalist explanations: a case study of
dynamic church membership. Br J Sociol 57:379–397
Chattoe E (2006b) Using evolutionary analogies in social science: two case studies. In: Wimmer
A, Kössler R (eds) Understanding change: models, methodologies and metaphors. Palgrave
Macmillan, Basingstoke, pp 89–98
Chattoe E, Gilbert N (1997) A simulation of adaptation mechanisms in budgetary decision making.
In: Conte R, Hegselmann R, Terna P (eds) Simulating social phenomena (Lecture notes in
economics and mathematical systems), vol 456. Springer, Berlin, pp 401–418
Chattoe-Brown E (2009) The implications of different analogies between biology and society for
effective functionalist analysis (Draft paper). Department of Sociology, University of Leices-
ter, Leicester
Cloak FT (1975) Is a cultural ethology possible? Hum Ecol 3:161–182
Costall A (1991) The meme meme. Cult Dyn 4:321–335
Csányi V (1989) Evolutionary systems and society: a general theory of life, mind and culture.
Duke University Press, Durham
Darwin CR (1859) On the origin of species by means of natural selection. John Murray, London
Dautenhahn K, Nehaniv CL (eds) (2002) Imitation in animals and artifacts. MIT Press,
Cambridge, MA
Dawkins R (1982) Organisms, groups and memes: replicators or vehicles? In: Dawkins R (ed) The
extended phenotype. Oxford University Press, Oxford
Dawkins R (1993) Viruses of the mind. In: Dahlbohm B (ed) Dennett and his critics. Blackwell,
Malden, pp 13–27
Dennett D (1990) Memes and the exploitation of imagination. J Aesthet Art Critic 48:127–135
Dobzhansky T, Ayala FJ, Stebbins GL, Valentine JW (1977) Evolution. W.H. Freeman,
San Francisco
Dosi G, Marengo L, Bassanini A, Valente M (1999) Norms as emergent properties of adaptive
learning: the case of economic routines. J Evol Econ 9:5–26
Edelman G (1992) Bright air, brilliant fire: on the matter of the mind. Basic Books, New York
Edmonds B (2002) Exploring the value of prediction in an artificial stock market. In: Butz VM,
Sigaud O, Gérard P (eds) Anticipatory behaviour in adaptive learning systems (Lecture notes in
Edmonds B, Moss S (2001) The importance of representing cognitive processes in multi-agent
models. In: Dorffner G, Bischof H, Hornik K (eds) Artificial neural networks: ICANN 2001,
international conference Vienna, 21–25 Aug 2001, proceedings (Lecture notes in computer
science), vol 2130. Springer, Berlin, pp 759–766
Epstein JM (2007) Generative social science: studies in agent-based computational modelling.
Fagin R, Halpern J, Moses Y, Vardi M (1995) Reasoning about knowledge. MIT Press,
Cambridge, MA
Forrest S (1991) Parallelism and programming in classifier systems. Pitman, London
Friedman DP, Felleisen M (1987) The little LISPER. MIT Press, Cambridge, MA
Sage, London
Gilbert N, Troitzsch KG (2005) Simulation for the social scientist, 2nd edn. Open University Press,
Buckingham
Goldberg DE (1989) Genetic algorithms in search, optimization and machine learning. Addison-
Wesley, Boston
Goldberg DE, Deb K, Korb B (1990) Messy genetic algorithms revisited: studies in mixed size and
scale. Complex Syst 4:415–444
Grefenstette J, Gopal R, Rosmaita B, Van Gucht D (1985) Genetic algorithms for the travelling
salesman problem. In: Grefenstette J (ed) Proceedings of the first international conference on
genetic algorithms and their applications, Carnegie Mellon University, Pittsburgh. Lawrence
Erlbaum, Hillsdale, 24–26 July 1985, pp 160–168
Hannan MT, Freeman J (1993) Organizational ecology. Harvard University Press, Cambridge, MA
Harvey I (1993) Evolutionary robotics and SAGA: the case for hill crawling and tournament
selection. In: Langton CG (ed) Artificial life III: proceedings of the workshop on artificial life,
Santa Fe, June 1992. Addison-Wesley, Boston, pp 299–326
Haynes T, Schoenefeld D, Wainwright R (1996) Type inheritance in strongly typed genetic
programming. In: Angeline PJ, Kinnear JE (eds) Advances in genetic programming 2. MIT
Press, Boston, pp 359–376
Heyes CM, Plotkin HC (1989) Replicators and interactors in cultural evolution. In: Ruse M (ed)
What the philosophy of biology is: essays dedicated to David Hull. Kluwer, Amsterdam
Hodgson G (1993) Economics and evolution: bringing life back into economics. Polity Press,
Cambridge, MA
Hoenigswald HM, Wiener LS (1987) Biological metaphor and cladistics classification. Francis
Pinter, London
Holland JH (1975) Adaptation in natural and artificial systems. University of Michigan Press,
Ann Arbor
Hughes A (1988) Evolution and human kinship. Oxford University Press, Oxford
Hull DL (1982) The naked meme. In: Plotkin HC (ed) Learning development and culture: essays
in evolutionary epistemology. Wiley, New York
Hull DL (1988) Interactors versus vehicles. In: Plotkin HC (ed) The role of behaviour in evolution.
MIT Press, Cambridge, MA
Iannaccone L (1994) Why strict churches are strong. Am J Sociol 99:1180–1211
Kampis G (1991) Self-modifying systems in biology: a new framework for dynamics, information
and complexity. Pergamon Press, Oxford
Kauffman SA (1993) The origins of order, self-organization and selection in evolution. Oxford
University Press, Oxford
Koza JR (1991) Evolving a computer program to generate random numbers using the genetic
programming paradigm. In: Belew R, Booker L (eds) Proceedings of the fourth international
conference on genetic algorithms, UCSD, San Diego. Morgan Kaufmann, San Francisco,
13–16 July 1991, pp 37–44
Koza JR (1992a) Genetic programming: on the programming of computers by means of natural
selection. A Bradford Book/MIT Press, Cambridge, MA
Koza JR (1992b) Genetic evolution and co-evolution of computer programmes. In: Langton C,
Taylor C, Farmer J, Rassmussen S (eds) Artificial life II: proceedings of the workshop on
artificial life, Santa Fe, February 1990. Addison-Wesley, Redwood City, pp 603–629
Koza JR (1992c) Evolution and co-evolution of computer programs to control independently
acting agents. In: Meyer J-A, Wilson S (eds) From animals to animats: proceedings of the first
international conference on simulation of adaptive behaviour (SAB 90), Paris. A Bradford
Book/MIT Press, Cambridge, MA, 24–28 Sept 1990, pp 366–375
Koza JR (1992d) A genetic approach to econometric modelling. In: Bourgine P, Walliser B (eds)
Economics and cognitive science: selected papers from the second international conference on
economics and artificial intelligence, Paris. Pergamon Press, Oxford, 4–6 July 1990, pp 57–75
Koza JR (1994) Genetic programming II: automatic discovery of reusable programs. A Bradford
Book/MIT Press, Cambridge, MA
Kuhn TS (1970) The structure of scientific revolutions. University of Chicago Press, Chicago
Kummer H, Daston L, Gigerenzer G, Silk J (1997) The social intelligence hypothesis. In: Weingart
P, Richerson P, Mitchell SD, Maasen S (eds) Human by nature: between biology and the social
sciences. Lawrence Erlbaum, Hillsdale, pp 157–179
Lomborg B (1996) Nucleus and shield: the evolution of social structure in the iterated prisoner’s
dilemma. Am Sociol Rev 61:278–307
Lynch A (1996) Thought contagion, how belief spreads through society: the new science of
memes. Basic Books, New York
Macy M (1996) Natural selection and social learning in the prisoner’s dilemma: co-adaptation
with genetic algorithms and artificial neural networks. Sociol Meth Res 25:103–137
Martinez-Jaramillo S (2007) Artificial financial markets: an agent based approach to reproduce
stylized facts and to study the red queen effect (PhD thesis). Centre for Computational Finance
and Economic Agents (CCFEA), University of Essex
Metcalfe J (ed) (1994) Metacognition: knowing about knowing. A Bradford Book/MIT Press,
Cambridge, MA
Mitchell M (1996) An introduction to genetic algorithms. A Bradford Book/MIT Press,
Cambridge, MA
Montana DJ (1995) Strongly typed genetic programming. Evol Comput 3(2):199–230
Moran PAP (1962) The statistical processes of evolutionary theory. Clarendon, Oxford
Moss S (1992) Artificial Intelligence models of complex economic systems. In: Moss S, Rae J
(eds) Artificial intelligence and economic analysis: prospects and problems. Edward Elgar,
Cheltenham, pp 25–42
Nelson RR, Winter SG Jr (1982) An evolutionary theory of economic change. Belknap Press of
Harvard University Press, Cambridge, MA
North DC (1990) Institutions, institutional change and economic performance. Cambridge Uni-
versity Press, Cambridge
Oliphant M (1996) The development of Saussurean communication. Biosystems 37:31–38
Olivetti C (1994) Do genetic algorithms converge to economic equilibria? (Discussion paper, 24).
Department of Economics, University of Rome “La Sapienza”, Rome
Popper KR (1979) Objective knowledge: an evolutionary approach. Clarendon, Oxford
Reader J (1970) Man on earth. Collins, London
Runciman WG (1998) The selectionist paradigm and its implications for sociology. Sociology
32:163–188
Schraudolph N, Belew R (1992) Dynamic parameter encoding for genetic algorithms. Mach Learn
9:9–21
Smith R, Forrest S, Perelson A (1992) Searching for diverse co-operative populations with genetic
algorithms (TCGA report, 92002). The clearinghouse for genetic algorithms, Department of
Engineering Mechanics, University of Alabama, Tuscaloosa
Tarde G (1884) Darwinisme naturel et Darwinisme social. Revue Philosophique XVII:607–637
Tarde G (1903) The laws of imitation. Henry Holt, New York
Vega-Redondo F (1996) Evolution, games and economic behaviour. Oxford University Press,
Oxford
Weibull J (1995) Evolutionary game theory. MIT Press, Cambridge, MA
Westoby A (1994) The ecology of intentions: how to make memes and influence people:
culturology. http://ase.tufts.edu/cogstud/papers/ecointen.htm
Whitley D (1989) The GENITOR algorithm and selection pressure: why rank-based allocation of
reproductive trials is best. In: Schaffer J (ed) Proceedings of the third international conference
on genetic algorithms, George Mason University. Morgan Kaufmann, San Francisco, 4–7 June
1989, pp 116–121
Wilson EO (1975) Sociobiology: the new synthesis. Harvard University Press, Cambridge, MA
Windrum P, Birchenhall C (1998) Developing simulation models with policy relevance: getting to
grips with UK science policy. In: Ahrweiler P, Gilbert N (eds) Computer simulations in science
and technology studies. Springer, Berlin, pp 183–206
Part IV
Applications
Chapter 19
Agent-Based Modelling and Simulation
Applied to Environmental Management
Christophe Le Page, Didier Bazile, Nicolas Becu, Pierre Bommel,

François Bousquet, Michel Etienne, Raphael Mathevet, Véronique Souchère,
Guy Trébuil, and Jacques Weber
Why Read This Chapter? To understand the recent shift of paradigms

prevailing in both environmental modelling and renewable resources management
that led to the emerging rise in the application of ABMS. Also, to learn about a
practical way to characterize applications of ABMS to environmental management
and to see this framework applied to review a selection of recent applications of
C. Le Page (*) • D. Bazile • F. Bousquet • G. Trébuil

CIRAD, UPR47 Green, F-34398 Montpellier, France
e-mail: [email protected]; [email protected]; [email protected];
[email protected]
N. Becu
CNRS, UMR8586 Prodig, 75005 Paris, France
P. Bommel
PUC University, Rio de Janeiro, Brazil
CIRAD, UPR47 Green, F-34398 Montpellier, France
M. Etienne
INRA, Unité d’Ecodéveloppement, Site Agroparc, 84914 Avignon cedex 9, Avignon, France
R. Mathevet
Centre d’Ecologie Fonctionnelle et Evolutive, UMR 5175 – CNRS 1919 Route de Mende, F34293
Montpellier, France
V. Souchère
UMR SADAPT INRA AgroParisTech, B.P. 1, 78850 Thiverval Grignon, France
J. Weber
CIRAD, Direction Régionale Ile de France, 42 rue Scheffer, F-75116 Paris, France

500 C. Le Page et al.
ABMS from various fields related to environmental management including the

dynamics of land use changes, water, forest and wildlife management, agriculture,
livestock productions and epidemiology.
Abstract The purpose of this chapter is to summarize how agent-based modelling
and simulation (ABMS) is being used in the area of environmental management.
With the science of complex systems now being widely recognized as an appropri-
ate one to tackle the main issues of ecological management, ABMS is emerging as
one of the most promising approaches. To avoid any confusion and disbelief about
the actual usefulness of ABMS, the objectives of the modelling process have to be
unambiguously made explicit. It is still quite common to consider ABMS as mostly
useful to deliver recommendations to a lone decision-maker, yet a variety of
different purposes have progressively emerged, from gaining understanding
through raising awareness, facilitating communication, promoting coordination or
mitigating conflicts. Whatever the goal, the description of an agent-based model
remains challenging. Some standard protocols have been recently proposed, but
still a comprehensive description requires a lot of space, often too much for the
maximum length of a paper authorized by a scientific journal. To account for the
diversity and the swelling of ABMS in the field of ecological management, a review
of recent publications based on a lightened descriptive framework is proposed. The
objective of these descriptions is not to allow the replication of the models but
rather to characterize the types of spatial representation, the properties of the
agents, the features of the scenarios that have been explored, and also to mention
which simulation platforms were used to implement them (if any). This chapter
concludes with a discussion of recurrent questions and stimulating challenges
currently faced by ABMS for environmental management.
19.1 Introduction
In this chapter, we state that there is a combined shift in the way of thinking in both
ecosystem management and ecological modelling fields. Over the last 20 years, the
status of computer simulation in the field of renewable resources management has
changed. This chapter investigates how agent-based modelling and simulation
(ABMS) may have contributed to this evolution and what are the challenges it
has to face for such a combination to remain fruitful.
Biosphere 2, an artificial closed ecological system built in Arizona (USA) in the
late 1980s, was supposed to test if and how people could live and work in a closed
biosphere. It proved to be sustainable for eight humans for 2 years, when low
oxygen level and wild fluctuations in carbon dioxide led to the end of the experi-
ence. Biosphere 2 represents the quest for “engineering Nature” that has fascinated
a lot of people (including a non-scientific audience) during the second part of the
last century. The human aspect of this “adventure” mainly dealt with the psycho-
logical impact on a few people living in enclosed environments. In the real world,
19 Agent-Based Modelling and Simulation Applied to Environmental Management 501
the relationships between human beings and the biosphere are based on tight
linkages between cultural and biological diversity. Launched around 20 years
before the Biosphere 2 project, the Man and Biosphere Program (MAB) of
UNESCO is seeking to improve the global relationship between people and their
environment. This is now the kind of approach – in line with the Millennium
Development Goal #7 from the United Nations – that is attracting more and more
interest.
In ecological management, the place of people directly involved in the man-
agement scheme is now widely recognized as central, and the impact of their
activities has both to be considered as promoting and endangering different types
of biodiversity. At the same time, ABMS has progressively demonstrated its
ability to explicitly represent the way people are using resources, the impact of
this management on plant and animal dynamics and the way ecosystems adapt to
it. The next section discusses how both trends have been reinforcing each other in
more detail.
The third section of this chapter gives a review of recent applications of ABMS
in the field of environmental management. To avoid confusion due to the
co-existence of multiple terms not clearly distinguishable, we use ABMS here
as an umbrella term to refer indifferently to what authors may have denominated
“agent-based modelling”, “multi-agent simulation”, or even “multi-agent based
simulation” (also the name of an international workshop where applications
dealing with environmental management are regularly presented). Our review is
covering the dynamics of land use changes, water, forest and wildlife manage-
ment, but also agriculture, livestock productions and epidemiology. We are
focusing here on models with explicit consideration of the stakeholders (in this
chapter this is how we will denominate people directly concerned by the local
environmental management system). Bousquet and Le Page (2004) proposed a
more extensive review of ABMS in ecological modelling. For a specific review of
ABMS dealing with animal social behaviour, see Chap. 22 in this handbook
(Hemelrijk 2013).
19.2 A Shift in Intertwined Paradigms
During the last two decades, evidences accumulate that the interlinked fields of
ecosystem management and environmental modelling are changing from one way
of thinking to another. This is a kind of paired dynamics where agents of change
from one field are fostering the evolution of conceptual views in the other one.
A survey of the articles published in “Journal of Environmental Management” and
“Ecological Modelling” – just to refer to a couple of authoritative journals in those
fields – clearly reveals this combined shift of paradigms. Another indication from
the scientific literature was given when the former “Conservation Ecology” journal
was renamed “Ecology and Society” in June 1997.
Among ecologists, it has become well accepted that classical equilibrium

theories are inadequate and that ecosystems are facing cycles of adaptive change
made of persistence and novelty (Holling 1986). Concepts from the sciences of
complexity are now widely adopted in ecology (Levin 1998), and the perception of
ecosystems as complex adaptive systems, in which patterns at higher levels emerge
from localized interactions and selection processes acting at lower levels, has begun
to affect the management of renewable resources (Levin 1999).
Beyond the standard concept of “integrated renewable resource management”,
the challenge is now to develop a new “integrative science for resilience and
sustainability” focusing on the interactions between ecological and social
components and taking into account the heterogeneity and interdependent dynamics
of these components (Berkes and Folke 1998). The relationships between
stakeholders dealing with the access and use of renewable resources are the core
of these intertwined ecological and social dynamics that are driving the changes
observed in many ecosystems.
Panarchy is a useful concept to understand how renewable resources manage-
ment is affected by this new paradigm in ecology. It has been formalized as the
process by which ecological and social systems grow, adapt, transform, and
abruptly collapse (Gunderson and Holling 2002). The back loop of such changes
is a critical time when uncertainties arise and when resilience is tested and
established (Holling 2004). This new theoretical background is making sense to
social scientists working on renewable resources management (Abel 1998) and to
interdisciplinary groups expanding ecological regime shifts theory to dynamics in
social and economic systems (Kinzig et al. 2006).
For a long period, the mainstream postulate in ecological modelling has been
that science should first help to understand the “natural functioning” of a given
ecosystem, so that the impacts of external shocks due to human activities
(“anthropic pressures”) could be monitored. Models were mainly predictive, ori-
ented towards decision makers who were supposed to be supported by powerful
tools (expert systems, decision-support systems) in selecting the “best”, “optimal”
management option. Nowadays, command-and-control approaches are seen as
“being worse than inadequate” (Levin 1999).
Evidently, there is a growing need for more flexible (usable and understandable
by diverse participants) and adaptive (easily modified to accommodate unforeseen
situations and new ideas) models that should allow any involved stakeholders
(ecosystem and resource managers among others) to gain insights through explora-
tion of simulation scenarios that mimic the challenges they face. Similar to the role
of metaphor in narratives, such simulation models do not strive for prediction
anymore, but rather aim at sparking creativity, facilitating discussion, clarifying
communication, and contributing to collective understanding of problems and
potential solutions (Carpenter et al. 1999). To underline the change of status of
simulation models used in such a way, the term “companion modelling” has been
proposed (ComMod 2003; Etienne 2011).
In recent years, ABMS has attracted more and more attention in the field of
environmental management (Bousquet and Le Page 2004; Hare and Deadman
2004). Recent compilations of experiences have been edited (Gimblett 2002;

Janssen et al. 2002; Bousquet et al. 2005; Perez and Batten 2006). We propose to
review recent ABMS applications in the field of environmental management based
on a simplified framework presented in the next section.
19.3 A Framework for Characterizing Applications of ABMS

to Environmental Management
To standardize the description of ecological models based on the interactions

between elementary entities (individual-based models and agent-based models),
Grimm and others (Grimm et al. 2006) have recently proposed a protocol based on a
three blocks sequence: overview, design concepts, details (ODD). It is a kind of
guideline for authors wishing to publish their model whose fulfilment corresponds
to an entire article devoted to communicating the details of the model. Hare and
Deadman (2004) also proposed a first classification scheme from the analysis of
11 case studies. Revisiting some elements from these two contributions, we propose
here to successively give some insights about: (1) the purpose of the model; (2) the
way the environment is represented; (3) the architecture of the different agents;
(4) the implementation (translation of the conceptual model into a computer
programme); (5) the simulation scenarios.
19.3.1 What Is the Model’s Purpose?
As recommended by Grimm and the 27 other participants to the collective design of

the ODD protocol (2006), a concise formulation of the model’s purpose has to be
stated first: it is crucial to understand why some aspects of reality are included while
others are ignored. The reasons leading to start a modelling process are not always
clearly given. Is it mainly to gain understanding and increase scientific knowledge?
Is it more about raising awareness of stakeholders who do not have a clear picture of
a complex system? Does it aim at facilitating communication or supporting deci-
sion? The more the information about the model’s purpose will be precise, the less
confusion and disbelieving about its real usefulness will remain.
19.3.2 How Is the Environment Represented?
Applications of ABMS to investigate environmental management issues are relying

on a fundamental principle: they represent interacting social and ecological dynamics.
On one hand, agents represent some sort of stakeholders, at the level of individual
people or/and at more aggregated levels some groups of individuals defining (lato
sensu) institutions (social groups such as families; economic groups such as farmers’
organizations; political groups such as non-governmental organizations). On the other
hand, the environment, holding some sort of renewable resources, stands for the
landscape. The renewable resources are contributing to define the landscape, and in
turn the way the landscape is structured and developed influences the renewable
resources. Typically, the resources are modified by direct actions of agents on their
environment, whereas the resources also exhibit some intrinsic natural dynamics
(growth, dispersal, etc.. . .). At the same time, agents’ decisions are somehow
modified by the environment as the state of resources evolves. In such situations,
the implementation of the social dynamics is performed through defining the
behaviours of agents, whereas the implementation of the natural dynamics is
commonly ascribed to the spatial entities defining the environment. Furthermore,
spatial entities viewed as “management entities” can support the reification of the
specific relationships between a stakeholder using the renewable resource and the
renewable resource itself.
Yet some applications of ABMS to environmental management do not repre-
sent any spatially-explicit natural dynamics. The data related to the environmental
conditions (i.e. the overall quantity of available resource), used by the agents to
make their decisions, are managed just like any other kind of information. But
even when the environmental conditions are not spatially-explicit, the explicit
representation of space can help to structure the interactions among the agents.
For instance, in the simulation of land use changes, the cognitive reasoning of
agents can be embedded in cellular automata (CA) where each cell (space portion)
represents a decisional entity that considers its neighbourhood to evaluate the
transition function determining the next land use. Typically, acquaintances
are then straightforwardly set from direct geographical proximity. It is still
possible to stick on CA with more flexible “social-oriented” ways to define the
acquaintances, but as soon as decisional entities control more than a single space
portion, they have to be disembodied from the spatial entities. The FEARLUS
model proposed by Polhill et al. (2001) and described in Sect. 19.4 is a good
illustration of such a situation.
Applications of ABMS explicitly representing some renewable resources have to
deal with the fact that renewable resources are usually heterogeneously scattered over
the landscape being shaped by their patterns. Any irregularities or specificities in the
topological properties of the environment legitimate to incorporate a spatially-
explicit representation of the environment. The combined use of geographic infor-
mation systems (GIS) and ABMS is a promising approach to implement such
integration (Gimblett 2002), particularly when there is a need to refer explicitly to
an actual landscape in a realistic way. More generally, the relationship between an
artificial landscape and a real one can be pinpointed by referring to 3 levels of
proximity: (1) none, in case of theoretical, abstract landscape; (2) intermediate,
when the reference to a given landscape is implicit; (3) high, when the reference to
a given landscape is explicit. Theoretical ABMS applications frequently use purely
abstract landscapes, such as in the well-known sugar and spice virtual world of
Sugarscape (Epstein and Axtell 1996). In the intermediate case, the implicit reference
to a given landscape may exist through matching proportions in the composition of

the landscape and similar patterns in its spatial configuration. When the reference to
an actual landscape is explicit, the use of GIS is required to design the environment.
An example of such realistic representation of a given landscape is given by Etienne
et al. (2003). Characterizing the relationship between the simulated environment and
the reality is a good way to estimate to what extent the model may provide a wide
scope: the rule “the more realistic, the less generic” is hardly refutable.
From a technical point of view, the representation of space in ABMS
applications with spatially-explicit representation of the environment could be
either continuous or discrete. Most of the time, the representation of space is
based on a raster grid (the space is regularly dissected into a matrix of similar
elementary components), less frequently it is made of a collection of vector
polygons. The continuous mode is quite uncommon in ABMS. This is related to
the standard scheduling of ABMS that relies on either a discrete-time approach or
on a discrete-event approach. Therefore, dealing with time (regular or irregular)
intervals, the spatial resolution of the virtual landscape can be chosen so that the
elementary spatial entity (defined as the smallest homogeneous portion of space)
can be used as the unit to characterize distances or neighbourhood. Using a discrete
mode to represent the space allows to easily define aggregated spatial entities that
directly refer to different ecological scales relevant to specific natural or social
dynamics, as well as to the management units specifically handled by the
stakeholders. The corresponding spatial entities are interrelated according to a
hierarchical organization, through aggregations.
19.3.3 How Are the Agents Modelled?
As we restrict our study to the sole applications with explicit consideration of the
stakeholders, by “agent” we mean a computer entity representing a kind of stake-
holder (individual) or a group of stakeholders. We stick with that operational
definition even when the authors opt for another terminology and propose to
characterize each kind of agent by considering two aspects: internal reasoning
leading to decision-making and interactions with the other agents (coordination).
19.3.3.1 Internal Reasoning
Decision-making is the internal process that specifies how an agent behaves. It

encompasses two dimensions: sophistication (from reactive to cognitive) and
adaptiveness (through evaluation). What is called the “behaviour” of an agent
refers to a wide range of notions. In some situations, the behaviour of an agent is
simply and straightforwardly characterized by the value of a key parameter, as in
the theoretical exploration of the tragedy of the commons by Pepper and Smuts
(2000) where agents are either restrained (intake rate of resource is set to 50 %) or
Collector Attitude
-threshold = 15
-collect Wood()
1 adopts 1
-go Back To Pile() +free Destinations()
-empty Basket() +evaluate()
-evaluate Attitude()
+step()
Disrespectful Respectful
-perception Range = 3
+evaluate()
+is Watched()
+evaluate()
+free Destinations()
Fig. 19.1 Design pattern of the driftwood collector agents (Thébaud and Locatelli 2001)
unrestrained (intake rate of resource is set to 99 %) in their foraging activity, when

all the other biological functions (perception, movement, reproduction, mortality)
are the same. In some other cases, the behaviour of a given agent does not only
depend on internal characteristics, like the driftwood collector agents proposed by
Thébaud and Locatelli (2001) who are stealing wood collected by other agents only
when their attitude is still disrespectful (internal property) and when they cannot be
observed (no peer pressure). Whatever the factors determining the behaviour of an
agent are, this behaviour may or may not change over time. When the behaviour is
simply characterized by a value of a key parameter, the adaptiveness can be taken
into account without any particular architectural design. For more sophisticated
behaviours, it becomes necessary to use a design pattern linking the agent to its
behavioural attitude. With such a design pattern, the different behavioural attitudes
are made explicit through corresponding subclasses, as shown in Fig. 19.1 with an
example taken from the Dricol model (Thébaud and Locatelli 2001). The adoption
of a particular attitude is updated according to some evaluation function.
Regarding the degree of sophistication of the decision-making process, the
so-called reactive agents implement a direct coupling between perception (often
related to little instantaneous information) and action. The forager agents of Pepper
and Smuts (2000) mentioned above are typical reactive agents. On the opposite
side, cognitive agents implement more complex decision-making processes by
explicitly deliberating about different possibilities of action and by referring to
specific representations of their environment, which is of particular importance for
applications of ABMS to environmental management (Bousquet and Le Page
2004). An example of such agents is given by Becu et al. (2003): farmer agents
evaluate direct observations and messages received from others (social network),
update their knowledge base, and evaluate options according to their objectives (see
the corresponding architecture in Fig. 19.2).
Objectives
Representation
- Improve soil fertility
(Believes on the state of
- Earn money
the system)
- Soil fertility (good/bad)
- Crop result (good/bad)
- Manure (expensive) Cognition
(Believes on the functioning of the Decision making

system)
- Crop choice
- Soil fertility depends on its - Soil input choice

structure and pH
- Crop yield is an indicator of soil

fertility
Perception
- Manure improves soil fertility
Observations - Crop yields depend on soil fertility
- Soil structure - Garlic cannot grow on low pH soil
- Crop yield
Execution
- Crop quality
Information - Plant
- Use soil input

- Soil pH
- Request soil pH test
- Manure price
- Exchange information
with others
Fig. 19.2 Architecture of farmer agents from the Catchscape model (Becu et al. 2003)
19.3.3.2 Interactions with Other Agents (Coordination)
Bousquet (2001) synthesized his general approach of multi-agent systems to study

environmental management issues with a diagram (see Fig. 19.3). We will refer
here to the three kinds of interactions depicted in Fig. 19.3 to describe the types of
agents implemented in applications of ABMS to environmental management.
The deliberative process of one agent is quite often influenced by some other
closely related agents. The proximity may be either spatial (local neighbourhood) or
social (acquaintances). In situations like the Dricol model (Thébaud and Locatelli
2001), what matters is just the presence of other agents in the surroundings. When
more information about the related agents is needed, then the rules to access this
information have to be specified. It is often assumed that the information is directly
accessible through browsing the agents belonging to a given network. This
corresponds to “Ie” in Fig. 19.3: other agents perceived through the environment
are considered as part of the environment of one agent.
Fig. 19.3 Interactions

between agents via the collective
collective level
level
environment (Ie), through
peer to peer communication
(Ii) and via the collective Ic
Ic
level (Ic) (Bousquet 2001)
individual level
Ii
Ie
Ie
environment
environment
Agents may strictly control the access to their internal information unless they
intentionally decide to communicate it (“Ii” in Fig. 19.3). Then the sharing of
information has to go through direct exchanges of peer-to-peer messages, with a
specified protocol of communication.
Relating agents directly to the collective level (“Ic” in Fig. 19.3) is most often
achieved via the notion of groups to which they can belong and for which they can
be representative to outsiders. Inspired by the Aalaadin meta-model proposed by
Ferber and Gutknecht (1998), recent applications of ABMS to agricultural water
(Abrami 2004) and waste water (Courdier et al. 2002, described in the next section)
management as well as to epidemiology (Muller et al. 2004, described in the next
section) illustrate how both notions of group and role are useful to handle levels of
organisation, and behaviours within levels of organisation. Even when «Ic» are not
implemented through specific features of the agents’ architecture, the mutual
influence of both collective and individual levels is fundamental in renewable
resources management. On one hand, individuals’ behaviours are driven by collec-
tive norms and rules; on the other hand, these norms and rules are evolving through
agents’ interactions. This individuals-society dynamics linkage, introducing the
notion of institution, relies on the representation of common referents (Bousquet
2001). Such “mediatory objects” are for instance the water temples involved in the
coordination of a complex rice-terraces system in Bali (Lansing and Kremer 1993;
Janssen 2007).
19.3.4 Implementation
We believe it is useful to indicate whether a simulation platform was used or not to

implement the model. Nowadays, some established generic tools such as Ascape,
Cormas, Mason, (Net)(Star)Logo, Repast, or Swarm are being used by large
communities of users. Intending to release researchers from low-level technical-

operational issues, their development is boosted by their comparisons performed
through the implementation of simple benchmark models (Railsback et al. 2006)
and the analysis of their abilities to fulfil identified requirements (Marietto et al.
2003). The maintainers and developers of such generic platforms have also taken
into consideration some sensitive technical aspects (floating point arithmetic,
random numbers generators, etc.) recently pointed out (Polhill et al. 2006), and
provide some elements to help users to escape these numerical traps. Additionally
to the benefit of not having to re-implement basic functionalities from scratch,
it may also happen that a previous model, made available online in the library of
existing models, presents some similarities with the new model to be developed.
Nevertheless, using a generic platform is not a panacea. It may incidentally lead
to poorer presentation of the developed models if the authors (wrongly) assume that
any reader is aware of the platform’s general principles. A new research stream
(model-to-model comparison) recently emerged from the fact that it is very difficult
to replicate simulation models from what is reported in publications (Hales et al.
2003). Reproducing results, however, is a sine qua non condition for making
ABMS a more rigorous tool for science. It may be achieved through a better
description of individual models, but also through the maintenance and develop-
ment of strong communities of users sharing the same tools for implementation.
This kind of stimulating diversity of the simulation platforms may contribute to
identify some generic ‘shorthand’ conventions that could minimize the effort to
describe the model rigorously and completely (Grimm et al. 2006).
19.3.5 Simulation Scenarios
Some ABMS platforms like NetLogo, for simplicity purposes, merge in the same
(unique) implementation file the definition of the domain entities with the specifi-
cation of the simulation scenario. In their ODD protocol, Grimm et al. (2006)
suggest to state the elements pertaining to the scheduling of a “standard” simulation
scenario in the first part (overview) of the sequential protocol, then come back to
some specific design concepts characterizing the domain entities (discussed here in
Sect. 19.3.3) in the second part (design) of the sequential protocol, and finally to
describe the initialization of a standard scenario in the third and last part (details).
Yet, a clear separation between model and simulation should be promoted when
seeking genericity. At the level of the agents, focusing on the description of their
internal structure and potential behaviour may help to identify some modules of
their architecture that could be re-used in other contexts. At the level of the
initialization and scheduling of the simulation, the same benefit can be expected:
for instance, generating parameter values for a population of agents from a statisti-
cal distribution, or creating an initial landscape fitting some schematic patterns
(Berger and Schreinemachers 2006).
The notion of “standard” scenario is not always very easily recognizable. Some
authors prefer to start by presenting what they call “reference” scenarios that
correspond to “extreme” situations. For instance, whenever the structure of a
given model makes sense to mention it, a “no agents” simulation scenario should
be available without any modifications, i.e. just initializing the number of agents to
zero. These scenarios can be used either as a verification step in the modelling
process (to test that the implementation is a faithful translation of the conceptual
model), or as a reference to compare the outputs obtained from more plausible
simulation scenarios. More generally, simulation scenarios have to address valida-
tion by questioning the results through looking back at the system under study.
In ABMS, validation is a multi-dimensional notion. Depending on the purpose
assigned to the model, the focus will be mainly set on: (1) checking if the simulated
data are fitting available real datasets; (2) looking for comparable processes
observed in other case studies; (3) evaluating to what extent the stakeholders accept
the model and its outputs as a fair representation of their system. For a more
detailed discussion of validation see Chap. 8 in this volume (David 2013).
Another essential dimension of simulation scenarios relates to the model output
used to observe them. Confronting the interpretations of the same simulated results
built from specific stakeholders’ viewpoints may be an effective way to share the
different opinions and then highlight the need to improve the agents coordination
mechanisms, or even to achieve a compromise (Etienne et al. 2003).
19.4 A Review of Recent Applications of ABMS

to Environmental Management
To classify the recent applications of ABMS in environmental management is not

an easy task, as the range of covered topics is wide: dynamics of land use changes,
water, forest and wildlife management, but also agriculture, livestock productions
and epidemiology. Some topics like epidemiology can easily be treated separately.
Some others are likely to be appearing simultaneously in some case studies,
especially for those dealing with multiple uses of the same renewable resource
and/or representing landscape with several land-use types. This latter situation
frames an entire research field in human geography that is focusing on the dynamics
of land-use/cover changes (LUCC). In the classification proposed below, some
applications clearly related to LUCC are listed in other subsections (mainly in
“agriculture” and in “forest”). Conversely, the “LUCC” subsection contains
applications that are related to some other topics specifically addressed later on.
Finally, some other topics like biodiversity are multi-dimensional and thus the
related case studies can be split into several other topics (for instance, biodiversity
related to endangered species are reversed into “wildlife”). Whenever it undoubt-
edly exists for an application, we are mentioning the relevance to other topics.
A specific category has been added to group the examples of ABMS addressing
theoretical issues in ecological management.
As the number of publications related to the use of ABMS in ecological

management is booming, it is almost impossible to analyze all of them. So for
each of the categories presented above, we had to select only a few representative
case studies to be briefly described by referring as much as possible to the elements
discussed in previous sections (the case studies that were not selected to be
analyzed are just mentioned in the introduction paragraph of each category).
Following Hare and Deadman (2004) who proposed a taxonomy of ABMS in
environmental management as a first step to provoke discussion and feedback,
our purpose here is to contribute to the framing of a practical bibliographic
survey by proposing some key characteristics useful for comparing applications
of ABMS in ecological management. See Appendix for a table recording the key
characteristics of the selected case studies.
19.4.1 Theoretical Issues in Environmental Management
Thébaud and Locatelli (2001) have designed a simple model of driftwood collec-
tion to study the emergence of resource sharing conventions; Pepper and Smuts
(2000) have investigated the evolution of cooperation in an ecological context with
simple reactive agents foraging either almost everything or just half of a renewable
resource. Schreinemachers and Berger (2006) have compared respective
advantages of heuristic and optimizing agent decision architectures; Rouchier
et al. (2001) have compared economic and social rationales of nomad herdsmen
securing their access to rangelands; Evans and Kelley (2004) have compared
experimental economics and ABMS results to explore land-use decision-making
dynamics; Soulié and Thébaud (2006) represent a virtual fishery targeting different
species in different areas to analyze the effects of spatial fishing bans as manage-
ment tools.
19.4.2 Dynamics of Land-Use/Cover Changes
Parker et al. (2003) have recently reviewed the application of multi-agent systems
to better understand the forces driving land-use/cover change (MAS/LUCC). Their
detailed state of the art presents a wide range of explanatory and descriptive
applications. Since this authoritative paper has been published, new applications
related to LUCC have continued to flourish. For instance, Caplat et al. (2006) have
simulated pine encroachment in a Mediterranean upland, Matthews (2006) has
proposed a generic tool called PALM (People and Landscape Model) for simulating
resource flows in a rural livestock-based subsistence community; LUCITA (Lim
et al. 2002), an ABM representing colonist household decision-making and land-
use change in the Amazon Rainforest, has been developed further (Deadman et al.
2004); Bonaudo et al. (2005) have designed an ABM to simulate the pioneers fronts
in the same Transamazon highway region, but at a lower scale; Manson (2006;
2005) has continued to explore scenarios of population and institutional changes in
the Southern Yucatan Peninsular Region of Mexico; Huigen (2004, 2006) has
developed MameLuke to simulate settling decisions and behaviours in the San
Mariano watershed, the Philippines. Below we describe in more detail a selection of
applications that are also characterized in the overview table presented in the
Appendix.
19.4.2.1 FEARLUS, Land-Use and Land Ownership Dynamics
FEARLUS, an abstract model of land use and land ownership implemented with
Swarm, has been developed to improve the understanding of LUCC in rural
Scotland by simulating the relative success of imitative versus non-imitative pro-
cess of land use selection in different kinds of environment (Polhill et al. 2001).
An abstract regional environment is defined as a toroidal raster grid made out of
8-connex land parcels, each being characterized by fixed biophysical conditions.
The same external conditions that vary over time apply in the same way to all land
parcels. These two factors are determining the economic return of a given land use
at a particular time and place. The land manager agents decide about the land uses
of the land parcels they own (initially a single one) according to a specific selection
algorithm. During the simulation, they can buy and sell land parcels (landless
managers leave the simulation; new ones may enter it by buying a land parcel).
Simulation scenarios were defined on several grids by pairing selection algorithms
from the predefined sets of 5 imitative and 5 non-imitative selection algorithms.
19.4.2.2 Greenbelt to Control Residential Development
This ABM, the simplest version of which being strictly equivalent to a mathemati-
cal model, has been developed to investigate the effectiveness of a greenbelt located
beside a developed area for delaying residential development outside the greenbelt
(Brown et al. 2004). The environment is represented as an abstract cellular lattice
where each cell is characterized by two values: a constant one to account for
aesthetic quality and a variable one to denote the proximity to service centres.
Service centres are called agents but actually they are more passive entities as they
do not exhibit any decision making. Residential agents, all equipped with the same
aesthetic and service centre preferences, decide their location among a set of
randomly selected cells according to a given utility function. The Swarm platform
was used for implementation. Scenarios, scheduled with periodic introductions of
new agents, are based on the values of residential agents’ preferences and on the
spatial distribution of aesthetic quality.
19.4.2.3 LUCC in the Northern Mountains of Vietnam
Castella et al. (2005a) developed the SAMBA model under the Cormas simulation
platform to simulate the land use changes during the transition period of decollec-
tivisation in a commune of the northern Vietnam uplands (Castella et al. 2005a, b;
Castella and Verburg 2007). This simple and adaptable model with heuristic value
represented the diversity of land use systems during the 1980s as a function of
household demographic composition and paddy field endowment in the lowland
areas. The environment in which agents make decisions was made of a 2,500 cell
grid, and 6 different land use types could be attributed to each cell, representing a
plot of 1,000 m2, also characterized by its distance to the village. While there were
no coordination among farmer agents with reactive behaviour in the early version of
the model, interactions among them was added later and the model coupled to a GIS
to extrapolate the dynamics to the regional landscape level (Castella et al. 2005b).
The simulated scenarios tested the effects of the size of the environment, the overall
population and household composition, and the rules for the allocation of the paddy
fields on the agricultural dynamics and differentiation among farming households.
More recently, this process-oriented model was compared to a spatially explicit
statistical regression-based pattern-oriented model (CLUE-s) implemented at the
same site. While SAMBA better represented the land use structure related to
villages, CLUE-s captured the overall pattern better. Such complementarity
supports a pattern-to- process modelling approach to add knowledge of the area
to empirically calibrated models (Castella and Verburg 2007).
19.4.2.4 Competing Rangeland and Rice Cropping Land-Uses in Senegal
To test the direct design and use of Role-Playing Games (RPG) and ABMS with
farmers and herders competing for land-use in the Senegal River Valley, participa-
tory simulation workshops were organized in several villages (D’Aquino et al.
2003). The ABM used during the last day of the workshops was straightforwardly
implemented with the Cormas platform from the characteristics and rules collec-
tively agreed the day before when crafting and testing a RPG representing
stakeholders’ activities related to agriculture and cattle raising. The environment
is set as a raster grid incorporating soil, vegetation and water properties of the
village landscape as stated by the stakeholders (a GIS was used only to clear
ambiguities). The same crude rules defined and applied during the RPG were
used to implement the autonomous reactive farmer agents. After displaying the
scenario identified during the RPG, new questions emerged and were investigated
by running the corresponding simulation scenarios. The hot debates that emerged
demonstrate the potential of these tools for the improvement of collective processes
about renewable resources management.
19.4.2.5 Landscape Dynamics in the Méjan Plateau, Massif Central, France
Etienne et al. (2003) developed a multi-agent system in order to support a companion

modelling approach on landscape dynamics in the Méjan plateau of the Massif
Central, the mountain range of central France (Etienne et al. 2003). The purpose of
the model is to support the coordination process among stakeholders concerned
with pine encroachment. The environment is a cellular automaton coming from the
rasterisation of a vector map. Several procedures account for vegetation changes
due to pine encroachment according to natural succession trends and range, timber
or conservation management decisions. The three agents types (sheep farmers,
foresters and the National Park) are concerned by this global biological process
but it affects their management goals in a very different way (sheep production,
timber production, nature conservation). The model is used to simulate and com-
pare collectively contrasting management scenarios arising from different
agreements. Simulation results were used to support the emergence of collective
projects leading to a jointly agreed management plan.
19.4.2.6 GEMACE: Multiple Uses of the Rhone River Delta,

Southern France
This ABM developed with the Cormas platform simulates the socio-economic
dynamic between hunting managers and farmers in the Camargue (Rhone river
delta, southern France), through the market of the wildfowling leasing system, in
interaction with ecological and spatial dynamics (Mathevet et al. 2003a). A CA
represents an archetypal region based on a spatial representation of the main types
of estates, distributed around a nature reserve. Each cell is characterized by water
and salt levels through land relief, land-use history, infrastructure, spatial
neighbourhood, and current land use. A wintering duck population, heteroge-
neously distributed in its habitats, is affected by various factors such as land-
use changes, wetland management, hunting harvest, and disturbance. Land-use
decisions are made at farmland level by farmers and hunting managers that are
communicating agents. Their strategy, farming or hunting oriented, is based on crop
rotation, allocation of land use and water management, and may change according
to some specific representations and values related to farming and hunting. Scenario
runs allowed discussing the structuring of the waterfowl hunting area resulting from
the individual functioning of farms in conjunction with a nature reserve and other
hunting units and the conservation policy.
19.4.3 Water Management
In the field of sustainable development, water management resources are an issue of

major importance. ABMS dealing with water management are used to simulating
the management of irrigated ecosystems, to represent the interactions among
stakeholders by capturing their views and formalizing the decision-making

mechanisms (especially negotiation processes), to capture the socioeconomic
aspects of potable water management and evaluate scenarios based on alternative
control measures, etc. For instance, Haffner and Gramel (2001) have investigated
strategies for water supply companies to deal with nitrate pollution; Janssen (2001)
has simulated the effects of tax rates related to the intensive use of phosphorus on
lake eutrophication; Becu et al. (2003) have developed CATCHSCAPE to simulate
the impact of upstream irrigation management on downstream agricultural viability
in a small catchment of Northern Thailand; Krywkow et al. (2002) have simulated
the effects of river engineering alternatives on the water balance of the Meuse river
in the Netherlands, and have related this hydrological module to stakeholders’
negotiations and decisions. Below we describe in more detail a selection of
applications that are also characterized in the table presented in the Appendix.
19.4.3.1 SHADOC: Viability of Irrigated Systems in the Senegal River Valley
To examine how existing social networks affect the viability of irrigated systems in
the Senegal River Valley, the SHADOC ABM focuses on rules used for credit
assignment, water allocation and cropping season assessment, as well as on organi-
zation and coordination of farmers in an irrigation scheme represented as a place of
acquisition and distribution of two resources: water and credit (Barreteau and
Bousquet 2000; Barreteau et al. 2004). The model used a spatially non-explicit
representation: all plots are subject to the same hydrological cycle regardless of
their exact geographical position. The societal model is structured with three types
of group agents in charge of credit management, watercourse and pumping station
management. As far as individual agents (farmers) are concerned, the model
employs a four-level social categorization with different types of farmers according
to their own cultivation objective. Each agent acts according to a set of rules local to
him. Each agent also has its own point of view about the state of the system and
especially its potential relations with other agents. SHADOC was first designed as a
tool for simulating scenarios of collective rules and individual behaviours.
19.4.3.2 MANGA: Collective Rules of Water Allocation in a Watershed
MANGA has been developed to test the economics, environmental, and ethical
consequences of particular water rules in order to improve the collective manage-
ment of water resources according to agricultural constraints, different actors’
behaviours, and confrontation of decision rules of each actor (Le Bars et al.
2005). Their modelling approach takes into account cognitive agents (farmers or
water supplier) trying to obtain the water they need via negotiation with the others
as a result of its individual preferences, rationality, and objectives. The MANGA
model used a spatially non-explicit representation for coupling social and environ-
mental models. To implement the decision-making process of the cognitive agents,
the authors used the BDI formalism and more particularly the PRS architecture.
During simulations, MANGA allows to test several water allocation rules based on
water request, irrigated corn area or behaviour evolution.
19.4.3.3 SINUSE: Water Demand Management in Tunisia
Sinuse is a simulator conceived to simulate the interactions between a water table

and the decisions of farmers in Tunisia (Feuillette et al. 2003). The farmers’
decisions are driven by economic objectives, but the dynamics of the system is
mainly dependent on the interactions among agents. The agents interact through
message sending to exchange land, and to team up to build wells. They also interact
through imitation and influence on the land price. They interact through the
environment as they share a common resource, the water table which has its own
dynamics and depends on the number of active wells. The model was developed
with Cormas platform. Simulations study the influence of various policies such as
subsidies for improved irrigation equipment.
19.4.3.4 Water Management and Water Temple Networks in Bali
Do irrigation systems necessarily need a centralized authority to solve complex

coordination problems? An ancestral Balinese system of coordination based on
villages of organized rice farmers (subaks) linked via irrigation canals has served as
a case study to investigate this question (Janssen 2007; Lansing and Kremer 1993).
Actions to be done on each specific date for each subak are traditionally related to
offerings to temples. The original model was recently re-implemented to deeper
investigate why the temple level would be the best level for coordination. The
environment is set as a network of 172 subaks, together with a network of 12 dams
allocating the water to the subaks. Each subak has up to 4 neighbouring subaks. It
selects one cropping plan out of 49 predefined ones. The corresponding water
demand is affecting the runoff between dams. Harvests are affected by water
stresses and pest outbreaks. The densities of pest in subaks are changing due to
local growth (related to the presence of rice) and migration (based on a diffusion
process). Six simulation scenarios based on the level of social coordination were
explored by Lansing and Kremer. Additionally, to the two extreme scenarios
defined with a single group of all 172 subaks (full synchronization) and 172
separate groups (no synchronization), 4 intermediate scenarios were tested, based
on groups defined from the existing system of temples.
19.4.3.5 Sharing Irrigation Water in the Lingmuteychu Watershed, Bhutan
Raj Gurung and colleagues used ABMS, following the companion modelling
approach, to facilitate water management negotiations in Bhutan (Gurung et al.
2006). A conceptual model was first implemented as a role-playing game to
validate the proposed environment, the behavioural rules, and the emergent
properties of the game. It was then translated into a computerized multi-agent
system under the Cormas platform, which allowed different scenarios to be
explored. Communicating farmer-agents exchanged water and labour, either within
a kinship network or among an acquaintance network. Different modes of commu-
nication (intra-village and inter-village) were simulated and a communication
observer displayed the exchange of water among farmers.
19.4.4 Forestry
Applications of ABMS in forestry are either focusing on LUCC issues or on

management issues. For instance, Moreno et al. (2007) have simulated social and
environmental aspects of deforestation in the Caparo Forest Reserve of Venezuela;
Nute et al. (2004) have developed NED-2, an agent-based decision support system
that integrates vegetation growth, wildlife and silvi-culture modules to simulate
forest ecosystem management plans and perform goal analysis on different views of
the management unit. Below we describe in more detail a selection of applications
that are also characterized in the table presented in the Appendix.
19.4.4.1 Deforestation and Afforestation in South-Central Indiana
Hoffmann et al. (2002) propose an original way of using ABMS to improve

scientific knowledge on the interactions between human activities and forest
patterns in Indiana, during the last 200 years (Hoffmann et al. 2002). The environ-
ment is a raster artificial landscape randomly generated from the 1820’s land-cover
ratio between crops, fallows and forests, and randomly calculated slopes. Farmer is
the only type of agent identified but they can behave differently according to two
potential goals (utility maximizing or learning reinforcement) and two actions:
deforestation or afforestation. Simulations are used to check through statistical
analysis of a high number of runs, the impact of ecological (slope), social
(stakeholders goals) or economic (agricultural prices, returns) factors in changing
land-use patterns.
19.4.4.2 Forest Plantation Co-management
This ABMS modeling approach links social, economic and biophysical dynamics to
explore scenarios of co-management of forest resources in Indonesia (Purnomo and
Guizol 2006; Purnomo et al. 2005). The purpose is to create a common dynamic
representation to facilitate negotiations between stakeholders for growing trees.
The environment is a simplified forest landscape (forest plots, road, agricultural
land) represented on a cellular automaton. Each stakeholder has explicit
communication capacities, behaviours and rationales from which emerge specific

actions that impact landscape dynamics. The model is used to simulate different
types of collaboration between stakeholders and both biophysical and economic
indicators are provided to measure the impact of each scenario on forest landscape
and smallholders incomes. Simulations results are supposed to support the selection
of the system of governance providing the best pathway to accelerate plantation
development, local community poverty alleviation and forest landscape
improvement.
19.4.5 Wildlife
Understanding how human activities impact on the population of animals in the

wild is a concern shared by conservationists, by external harvesters (hunters,
fishermen) and by local people. Viewed as a source of food or as an emblem of
biodiversity, management schemes first have to ensure the viability of the popula-
tion. Viewed as competitors for the living space of local people, management
schemes have to control the population. For instance, Zunga et al. (1998) have
simulated conflicts between elephants and people in the Mid-Zambezi Valley;
Galvin et al. (2006) have used ABMS to analyse how the situation in the
Ngorongoro Conservation Area (NCA) in northern Tanzania could be modified to
improve human welfare without compromising wildlife conservation value; Jepsen
et al. (2005) have investigated the ecological impacts of pesticide use in Denmark.
Below we describe in more detail a selection of applications that are also
characterized in the table presented in the Appendix.
19.4.5.1 Water Management in Mediterranean Reedbeds
Using the Cormas platform, the authors have developed an ABM to be used in
environmental planning and to support collective decision-making by allowing
evaluation of the long-term impact of several water management scenarios on the
habitat and its fauna of large Mediterranean reedbeds (Mathevet et al. 2003b).
A hydro-ecological module (water level, reedbed, fish, common and rare bird
populations) is linked to a socio-economic module (reed market and management).
Each cell was assigned a type of land use, land tenure and topography from a GIS to
create a virtual reedbed similar to the studied wetland. Five types of interacting agents
represent the users of the wetland. They are characterized by specific attributes
(satisfaction, cash amount, estates etc.) and exploit several hydro-functional units.
The behaviour of the agents depends on their utility function based on their evalua-
tion of the access cost to the reedbed and on their beliefs. The ecological and socio-
economic consequences of individual management decisions go beyond the estates,
and relate to the whole system at a different time scale.
19.4.5.2 Giant Pandas in China
Using data from Wolong Nature Reserve for giant pandas (China), this ABM
simulates the impact of the growing rural population on the forests and panda
habitat (An et al. 2005). The model was implemented using Java-Swarm 2.1.1 and
IMSHED that provides a graphical interface to set parameters and run the program.
It has three major components: household development, fuelwood demand, and
fuelwood growth and harvesting. The simulated landscape was built from GIS data.
Two resolutions were identified for sub-models requiring extensive human demo-
graphic factors and for landscape sub-models. Both person and household are
cognitive agents that were defined from socioeconomic survey. They allowed
simulating the demographic and household dynamics. Agents interact with each
other and their environment through their activities according to a set of rules.
The main interaction between humans and the environment is realized through
fuelwood collection according to demand. This model was used to test several
scenarios and particular features of complexity, to understand the roles of socio-
economic and demographic factors, identifying particular areas of special concern,
and conservation policy.
19.4.5.3 Traditional Hunting of Small Antelopes in Cameroon
To investigate the viability of populations of blue duikers, a small antelope tradi-

tionally hunted by villagers in the forests of Eastern Cameroon, an ABM has been
developed with the Cormas platform (Bousquet et al. 2001). The raster spatial
grid was defined by reading data from a GIS map corresponding to a village that
was surveyed during several months. Each cell represents 4 ha (the size of the blue
duiker habitat) and is characterized by a cover (river, road or village) and a
reference to a hunting locality. The population dynamics of the blue duiker is
simulated through the implementation of biological functions (growth, age-
dependent natural mortality, migration, reproduction) applied to all the individual
antelope agents. Hunter agents decide on the location of their traps by selecting one
hunting locality out of the four they use. A first set of simulation scenarios was
based on unilateral decisions of hunter agents, all of them following the same
general rule. Coordination among kinship groups of hunters was introduced in a
second set of experiences.
19.4.5.4 Whale-Watching in Canada
To investigate the interactions between whale-watching boats and marine mammals

in the Saguenay St. Lawrence Marine Park and the adjacent Marine Protected Area
in the St. Lawrence estuary, in Quebec, this ABM was implemented with the RePast
platform (Anwar et al. 2007). A raster grid defined from a GIS database represents
the landscape of the studied area. The boats are cognitive agents and whales are
simple reactive agents. Several simulations were run to explore various decision
strategies of the boat agents and how these strategies can impact on whales. For
each simulation, the happiness factor was used as an indicator of how successful the
boat agents were in achieving their goals. Results showed that cooperative
behaviour that involves a combination of innovator and imitator strategies based
on information sharing yields a higher average happiness factor over non-
cooperative and purely innovators behaviours. However, this cooperative
behaviour creates increased risk for the whale population in the estuary.
19.4.6 Agriculture
ABMS applied to agriculture is mainly focussing on decision-making processes at

the farm level (typically, agents represent households). Economic aspects usually
play a pivotal role and standard procedures like linear programming are often used
to represent individual choices among available production, investment, marketing
alternatives, etc. This economic module is then embedded into a more integrated
framework to explicitly represent spatial and social aspects. For instance, to
investigate technology diffusion, resource use changes and policy analysis in a
Chilean region, Berger (2001) has connected an economic sub-model based on
recursive linear programming to an hydrological sub-model; Ziervogel et al. (2005)
have used ABMS to assess the impact of using seasonal forecasts among
smallholder farmers in Lesotho; Sulistyawati et al. (2005) have analyzed the
consequence at the landscape level of swidden cultivation of rice and the planting
and tapping of rubber by Indonesian households whose demography and economic
welfare are simulated. Below we describe in more detail a selection of applications
that are also characterized in the table presented in the Appendix.
19.4.6.1 Agricultural Pest and Disease Incursions in Australia
To analyse the effectiveness and regional economic implications of alternative

management strategies for a range of different scenarios of a disease incursion in
the agricultural sector, an ABM has been developed with the Cormas platform and
applied to the case of the wheat disease Karnal bunt in a region of south eastern
Queensland, Australia (Elliston and Beare 2006). A cellular spatial grid allows
representing the spread of the pest across neighbouring paddocks and a range of
potential transmission pathways including the wind, farm inputs and agents (farmers,
contractors and quarantine officers) through their movement over the spatial grid.
Farmers make cropping decisions about planting, spraying for weeds, harvesting and
the use of contract labour. They can directly identify and report signs of a Karnal bunt
incursion on their property. The incursion can also be detected from quality inspec-
tion when farm production reaches the collective storage unit. Then a quarantine
response, based on a recursive checking in the neighbourhood of infected farms, is

implemented by officer agents. Simulation scenarios are based on one hand on levels
of farmer detection and reporting and on the other hand on the way the disease was
first introduced into the system (limited and slowly expanding incursion versus
potentially rapid expansion from a wide use of contaminated fertiliser).
19.4.6.2 AgriPolis: Policy Impact on Farms Structural Changes

in Western Europe
Agripolis is the evolution of a model developed by Balmann (1997). It describes the

dynamics of an agricultural region composed of farms managed by farmers
(an agent represents both of these concepts) (Happe et al. 2006). The Landscape
and the Market are the other agents. The farm agent has a cognitive decision-
making process: this process corresponds to the traditional modelling in agricultural
economics, where agents try to maximise their income. The land market is the
central interaction institution between agents in Agripolis. Farm agents extend their
land by renting land from farm landowners. The allocation of land is done through
auctions. The Agripolis model was used to study a region in southwest Germany.
A sensitivity analysis is done to analyze the relationship between policy change and
determinants of structural changes such as the interest rate, managerial abilities and
technical change.
19.4.6.3 Adaptive Watershed Management in the Mountains

of Northern Thailand
This companion modelling experiment aims at facilitating a learning process

among Akha highlanders about the socio-economic aspects (i.e. allocation of
formal & informal credit, on & off-farm employment) of the expansion of planta-
tion crops to mitigate soil erosion risk on steep land (Barnaud et al. 2007). Farmers’
individual decision-making regarding investment in perennial crops, assignment of
family labour to off-farm wage earning activities and search for credit were
modelled. The simulated scenarios looked at the effects of the duration of the
grace period of the loans, the distribution of formal credit amongst the 3 types of
farms, and different structures of the networks of acquaintances for informal credit.
Two main indicators were used to analyze the results of the simulations for each
type of farm: (1) the total area under plantation crops (ecological indicator) and (2)
the proportion of bankrupt farms leaving the agricultural sector (socio-economic
indicator).
19.4.7 Livestock Management
Janssen et al. (2000) have used adaptive agents to study the co-evolution of
management and policies in a complex rangeland system; another work lead by
Janssen (2002) has investigated the implications of spatial heterogeneity of
grazing pressure on the resilience of rangelands; Bah et al. (2006) have
simulated the multiple uses of land and resources around drillings in Sahel
under variable rainfall patterns; Milner-Gulland et al. (2006) have built an
ABM of livestock owners’ decision-making, based on data collected over
2 years in five villages in southeast Kazakhstan. Below we describe in more
detail a selection of applications that are also characterized in the table presented
in the Appendix.
19.4.7.1 Rangeland Patterns in Australia
To evaluate general behaviours of rangeland systems in Australia, this ABM

represents a landscape made of enterprises, which are cognitive agents that repre-
sent a commercial grazing property (Gross et al. 2006). Each property is defined by
an area, the quality of land in each patch, and its livestock. Behaviours of the
enterprise agents are defined by a strategy set comprised of a set of rules, which
evolves over time to represent learning. A government agent has an institutional
strategy set that also varies through time. Its main roles are to collect taxes and
deliver drought relief in the form of interest payment subsidies. The biophysical
sub-models allow simulating plant and livestock dynamics. Pastoral decisions are
made by the enterprise agents according to a set of rules. The variation in the level
of financial weakness leads to the adoption of a new strategy by an enterprise. Each
one is randomly associated with a rate of learning. Implemented in the Cþþ
programming language, the model is fed by inputs of historical data. The
simulations emphasize consequences of interactions between environmental
heterogeneity and learning rate.
19.4.7.2 Collective Management of Animal Wastes in La Reunion
To investigate the collective management of livestock farming wastes on La

Reunion Island, an ABM called Biomas has been developed with the Geamas
platform (Courdier et al. 2002). Biomas simulates the organization of transfers of
organic materials between two kinds of agents: surplus-producing farms (i.e. farms
where livestock raising activity dominates) and deficit farms (i.e. predominantly
crop production). The environment is represented as a network of “situated
objects”. Their association with the Geamas agents enable the agents to act on the
environment (e.g. “crop” agents are linked to “plot” situated objects). Some
situated objects like “road sections” are only related to other situated objects.
Graphs of connected situated objects allow representing itineraries. The agents in

Biomas are interacting through direct exchanges of messages; they are also linked
to a “Group” agent through a membership process. The “Group” agent is responsi-
ble for imposing the management constraints on all its members or individual
agents by means of contracts, and implements a penalty system in the case of
disregard of the regulations. The simulation scenarios are based on the constraints
and regulations defined at this “Group” level.
19.4.8 Epidemiology
Models developed for the spread of infectious diseases in human populations are
typically implemented assuming homogeneous population mixing, without a spa-
tial dimension, social (and network) dimension, or symptom-based behaviour.
ABMS offers great potential to challenge these assumptions. Recently Ling
Bian (2004) proposed a conceptual framework for individual-based spatially
explicit epidemiological modelling, discussing four aspects: (1) population
segments or unique individuals as the modelling unit, (2) continuous process or
discrete events for disease development through time, (3) travelling wave or
network dispersion for transmission of diseases in space, and (4) interactions within
and between night time (at home) and day time (at work) groups. As an illustration,
she compares a simple population-based model of influenza to an equivalent
schematic individual-based one. This abstract model has been utilised by Dunham
(2005) to develop a generic ABMS tool. Recently, the fear of bioterrorism has
also stimulated intensive studies in the USA, see for instance BioWar, developed
by Carley et al. (2006) to simulate Anthrax and smallpox attacks on the scale of
a city.
19.4.8.1 Bovine Leukemia
Two methodologies (system dynamics and agent-based) used to simulate the spread
of a viral disease (Bovine Leukemia) are compared (Bagni et al. 2002). The purpose
is, through “what-if” analysis, to assess the system’s behaviour under various
conditions and to evaluate alternative sanitary policies. Based from the same set
of Unified Modelling Language (UML) diagrams, Vensim and Swarm are the two
platforms that have been used to implement the conceptual model. The environ-
ment represents in an abstract way a dairy farm segmented into sectors. “Cow” and
“Farm sector” are the two types of autonomous agents in this model. The integra-
tion at the farm level is directly achieved through the “model swarm”. Scenarios
focus particularly on the number of cows detected as positive at sanitary controls
(as opposed to the total number of infected cows).
Fig. 19.4 The representation

of space as clustering of three F8 V1
kinds of “location” agents: C6
F3
village (pentagons), cocoa C7
plantations (hexagons) and F1

C4
forest (rectangles) (Muller F2
et al. 2004)
C5
C3 V2
C1 F4
V3 F7
F5 F6
C2
F8
F9
19.4.8.2 Malaria in Haiti
To assess the impact of education on malaria healthcare in Haiti, an ABM with a

realistic representation of Haiti has been designed (Rateb et al. 2005). The environ-
ment is set as a raster-grid with cells characterized by land-covers (sea, road, land,
mountain, city, school, and hospital) associated with specific contamination
probabilities (this is how mosquitoes are represented in the model). Apart from
an epidemiological status, autonomous agents (representing individual people) are
characterized by a mobility capability and an education score which value
corresponds to the time agents take to attribute existing symptoms to malaria and
therefore to go to hospital. Implemented in StarLogo, three scenarios based on the
number of schools and hospitals have been discussed.
19.4.8.3 Sleeping Sickness in Cameroon
To understand the spread of Human African Trypanosomiasis, and ultimately to

elaborate a tool to evaluate risk and test control strategies, an ABM has been
developed with the MadKit platform and tested with data from one village in
Southern Cameroon (Muller et al. 2004). The space is not explicitly represented
in this model. This is due to the meta-model associated with the MadKit platform:
the system under study has to be described through “agent-group-role” interactions
(Ferber and Gutknecht 1998). Hence, surprisingly, locations can only be depicted as
agents here (see Fig. 19.4). They are characterized by a proportional surface area
and a number of animals.
Location agents, as “groups”, are responsible for “enrolling” tsetse and human
agents that will, as members of the same group, be able to interact through the
sending of “bite” messages. The probability for a human agent to be bitten is
inversely proportional to the number of animals. Simulation scenarios are based on

the organisation of space, to investigate the effect of the size and number of
transmission areas.
19.4.9 General Considerations
To fill in the table presented in the Appendix from the description of the models
found in publications was not always easy. This is partly due to the fact that the
elements to be detailed in the columns of the table require further refinements and
more precise definitions. But this can also be attributed to the heterogeneity in the
way model contents are detailed by the authors. The lack of a general framework to
document such kind of models is patent, and all designers of ABM should become
aware and refer to the framework proposed by Grimm et al. (2006). Even when the
code of the implemented model is published (in appendices of articles or on a
website), it is quite challenging and time-consuming to dive into it to retrieve
specific information. This difficulty has triggered a bias: we have tended to select
the applications we know better. As the co-authors of this chapter all belong to the
same scientific network, the representativeness of the selected case studies may be
questioned. This kind of task – a systematic survey based on a set of unambiguously
defined characteristics – should be undertaken at the whole scientific community
level, in a continuous way. Ideally, it should use an effective tool for mass
collaborative authoring like wiki.
The environment, abstract or realistic, is most often represented as a raster grid.
The spatial resolution, when it makes sense to define it precisely (for realistic
simulated landscapes), is always clearly related to a key-characteristic of one of
the model’s components.
The number of applications with interactions involving the collective level is
rather low. This does not necessarily imply cognitive agents. In the model of
Bousquet et al. (2001), for instance, the collective level is related to the kinship
structure of the small antelopes’ population; when a young individual becomes
mature, it leaves the parental habitat territory and starts to move around to look for a
potential partner to establish a new family group in an unoccupied habitat. The
group left by the young adult is affected in such a way that the reproduction can be
activated again.
In our review, theoretical case studies are less numerous than empirical case
studies. The prevalence of theoretical case studies is only significant for the LUCC
category. It suggests that the proportion of empirical applications of ABMS is
gaining ground compared to theoretical and abstract contributions. As analyzed
by Janssen and Ostrom (2006), this could be explained by the fact that theoretical
models, more frequent at the beginning, have demonstrated that ABMS can provide
novel insights to scientific inquiry. The increased availability of more and more
relevant ecological and socio-economics data then paved the way to the rise of
empirically-based ABMS.
19.5 Why ABMS Is More and More Applied to Environmental

Management
If ABMS is becoming more and more popular in environmental modelling, it is

mainly because it demonstrates a potential to overcome the limitations of other kinds
of models to take into account elements and processes that can hardly be ignored to
consider the underlying research questions. Another aspect has to be stressed:
ABMS is structurally an integrative modelling approach. It can easily be expressed
with other modelling formalisms and tools. Additionally, to the evidential use of CA
as a way to represent the space in ABMS applications dealing with environmental
management, several other fruitful associations with complementary tools (GIS to
handle spatial requests, Linear Programming modules directly used by agents to
perform maximization of utility functions, etc.) have already been explored. Beyond
technical aspects, ABMS can also be seen as a methodological step of a wider
approach, like in companion modelling when it is jointly used with role-playing
games to allow stakeholders’ participation in the design of the tools used during
participatory simulation workshops (Bousquet et al. 2002).
19.5.1 Getting Rid of Empirically Implausible Assumptions
In ecology, traditional general population models are assuming that: (1) individuals
are identical; (2) the interaction between individuals is global; (3) the spatial
distribution of individuals is uniform. Required to ensure analytical tractability,
these overly simplified assumptions significantly limit the usefulness of such
population-based approaches. The assumption of “perfect mixing” on which
population-based modelling approaches rely (two individuals randomly picked
can be inter-changed) is only valid when the environment is homogeneous or
when all individuals facing the same environmental conditions react in exactly
the same way. One way to account for heterogeneity is to define subpopulations as
classes of similar individuals (for instance based on their age). Then a distribution
function of the individual states is sufficient. But when interactions between
individuals are depending on the local configuration of the environment (including
the other individuals), the spatial heterogeneity and the inter-individual variability
(two key drivers of evolution) can not be left out anymore. Spatially-explicit
individual-based models (IBM) allow representing any kind of details critical to
the system under study, thus relaxing assumptions distorting the reality in an
unacceptable manner. This is the main reason why for more than two decades
now (DeAngelis and Gross 1992; Huston et al. 1988; Judson 1994; Kawata and
Toquenaga 1994), IBM has been more and more widely used in ecological
modelling (for a recent guideline to make IBM more coherent and effective, see
Grimm and Railsback 2005).
When it comes to include human decision-making processes into models, the

standard way consists in assuming that all individuals equally informed (perfect
information sharing) exhibit a standard behaviour based on rationality to achieve
optimization. It is well known that renewable resources management addresses self-
referential situations. The success of an individual strategy highly depends on the
ability to “best-guess” what the other individuals may do over time (Batten 2007).
This is closely related to the notion of “representation” defined by Rouchier et al.
(2000) as the understanding an agent has of what it perceives, and that enables it to
evaluate and then to choose the actions it can undertake on its environment. Do all
agents agree, or do they have very different approaches to the same object or agent?
One of the main interests of ABMS is to offer the possibility to explore the second
option.
More generally, ABMS is a valid technical methodology to take into account
heterogeneity in parameter values and in behaviours. In abstract and theoretical
ABMS, standard statistical distribution functions can be used to assign particular
parameter values to the different instances of agents created from the same class.
Railsback et al. (2006) have compared how the main generic ABMS simulation
platforms handle the initialization of one attribute of a population of agents from a
random normal distribution (see version #14 of their benchmark “StupidModel”).
For more realistic applications of ABMS, Berger and Schreinemachers (2006)
recently introduced a straightforward approach to empirical parameterization
using a common sampling frame to randomly select observation units for both
biophysical measurements and socioeconomic surveys. The heterogeneity in
behaviours is usually considered with each agent having to select one behavioural
module from a set of existing ones. From a conceptual design point of view,
heterogeneity of behaviours is easier to represent with a hierarchy of classes.
Subclasses of a generic agent class are a proper design when a given agent does
not update its behaviour over time. To account for such an adaptive ability, the
agent class has to be linked to a hierarchy of behaviours, as shown in Fig. 19.1.
Beyond selecting an alternative out of a predefined set of options, it is even possible
to define innovative agents equipped with some evolutionary programming to drive
the creation of new behavioural patterns by recombining elementary behavioural
components.
19.5.2 Dealing with Multiple Nested Levels
The seminal paper of Simon (1973) envisions hierarchical organizations as adaptive

structures and not only as top-down sequences of authoritative control. This view
was instilled in ecology by Allen and Starr (1982), who promoted the idea that
biotic and abiotic processes at work in ecosystems are developing mutually
reinforcing relationships over distinct ranges of scales. Each level, made from
components interacting at the same time scale, communicates some information
to the next higher and slower level. Reciprocally, any level can contribute to
maintain the stability of faster and smaller levels. In the field of environmental
management, both social and biophysical systems are characterized by hierarchical,

nested structures. For example, family members interact to form a household,
which may interact with other households in a village through political and
economic institutions. Populations formed of individual species members aggregate
to form communities, which, in turn, collectively define ecosystems. Holling (2001)
nicely illustrates this with two mirroring examples: on one side the components of
the boreal forest represented over time and space scales (from needle to landscape);
on the other side the institutional hierarchy of rule sets (from the decisions of small
groups of individuals to constitution and culture) represented along dimensions of
the number of people involved and the turnover times.
These ideas have been conceptualized to frame the emerging paradigm of
“panarchy” (Gunderson and Holling 2002): the hierarchical structure of socio-
ecological systems is exhibiting never-ending cycles of growth, accumulation,
restructuring and renewal. The key-concepts of this heuristic model are undoubtedly
expanding the theoretical understanding of environmental management. What is their
concrete contribution to the evolution of ecological modelling? To what extent can
ABMS claim to represent them in a better way than other kinds of models?
The aggregation between hierarchical levels is very difficult to model in a purely
analytical or statistical framework. In ecology, aggregation methods are applicable
for models involving two levels of organisation (individual and population) and
their corresponding time scales (fast and slow) to reduce the dimension of the initial
dynamical system to an aggregated one governing few global variables evolving at
the slow time scale (Auger et al. 2000). The reverse way is much more difficult to
integrate to models. How to account for the influence of changes at the global level
on transitions at the microscopic level? Moreover, how to simulate both ways
simultaneously at work? The main challenge deals with the coordination and
scheduling of the different processes running at different levels: at the collective
level, explicit decisions about temporarily giving back the control to lower-level
component entities, and conversely decisions from lower-level entities to create a
group and to give the control to it. In the scientific community of ABMS, these
ideas have directly inspired the production of conceptual organizational meta-
models like Aalaadin (Ferber and Gutknecht 1998), specific features in generic
simulation platforms like the threaded scheduling of agents in Swarm, as well as
applications like simulating hydrological processes (runoff, erosion and infiltration
on heterogeneous soil surfaces) with “waterball”, pond and river agents (Servat
et al. 1998).
19.5.3 Beyond Decision Support Systems: Exploring New

Dimensions in the Way to Use Models
As they represent complex adaptive systems which are unpredictable as a whole,

ABMS applied to environmental management should caution about the large
uncertainties related to their predictive abilities (Bradbury 2002). Still, empirically-

based ABMS can be used as a decision-support system, for instance to assist
policymakers in prioritizing and targeting alternative policy interventions, as
Berger et al. (2006) did in Uganda and Chile. Nevertheless, when multiple
perceptions of the reality coexist, the statement “everything is defined by the reality
of the observed phenomena” can be questioned. Therefore, ABMS in the field of
ecological management should take some distance with the positivist posture that
designates the scientific knowledge as the only authentic one. Relating empirical
observations of phenomena to each other, in a way which is consistent with
fundamental theory, phenomenological modelling is a means to represent the
phenomena in a formalized and synthetic way. Descriptive rather than explanatory,
this approach does not truly gain understanding of the phenomena, but can claim, in
simple cases, to predict them. Parker et al. (2003) refer to these two distinct
explanatory and descriptive approaches to clarify the potential roles of ABMS in
LUCC. Anyway, the general rule “the more realistic the application, the more
descriptive the approach” may not necessarily always apply. Explanatory goals
can be assigned to models closely related to a real situation as well.
In contrast to the positivist approach, the constructivist approach refers to
“constructed” knowledge, contingent on human perception and social experience
and not necessarily reflecting any external “transcendent” realities. Starting “from
scratch” to collectively design a model is a straightforward implementation of the
constructivist approach. Among scientists, it will integrate within and between
disciplines. By involving stakeholders, instead of showing them a simplification
of their knowledge, the collective design of the model is seeking a mutual recogni-
tion of everyone’s representation. In such a context, ABMS is more a communica-
tion platform to facilitate collective learning than a turnkey itinerary for piloting
renewable resources management (Bousquet et al. 1999; ComMod 2003; Etienne
et al. 2003; Gurung et al. 2006).
19.6 Drawbacks, Pitfalls, and Remaining Challenges
19.6.1 Verification and Validation of ABMS
This is a problem challenging ABMS in general that is addressed in Chaps. 6 (Galán

et al. 2013) and 8 (David 2013) of this book. In the field of ecological management,
as in other fields of applications, some authors claim to intentionally bridle the
development of their agent-based model to design a strict equivalent to a mathe-
matical equations-based model, as a means to verify it (Brown et al. 2004). The
same process has been tested with mathematical representations of discrete

distributed systems like Petri nets (Bakam et al. 2001).
19.6.2 Capturing the Meta-rules Governing the Adoption

of Alternative Strategies
Nowadays, a set of tested and reliable tools and methods is available to better
understand decision rules of actors and integrate them in computer agents (Janssen
and Ostrom 2006). What remains much more challenging is to capture the rules
governing the changes in agents’ behaviours. An example of such a kind of “meta-
rule” is the “evaluateAttitude” of the Collector agent (see Fig. 19.1) defined by
(Thébaud and Locatelli 2001). In such a stylized model, the meta-rule is simply
based on a threshold value of a specific parameter (the size of the pile of collected
driftwood). When it comes to making the rules explicit to governing the changes of
behavioural rules of human beings in real situations, methods are still weak. The
meta-rules, if they exist, that control changes of strategies are difficult to grasp and
elicit, and by consequence to implement in an empirical-based ABM. One reason is
the time scale which is greater for these meta-rules than for decision rules, and
which makes direct observation and verification harder to carry.
19.6.3 Improving the Representation of Space
Representing space with a CA, by far the most frequent way in current applications
of ABMS in environmental management, is easy. But, as recently pointed out by
Bithell and Macmillan (2007), imposition of a fixed grid upon the dynamics may
cause important phenomena to be misrepresented when interactions between
individuals are mediated by their size, and may become too consumed by computer
resources when the system scale exceeds the size of individuals by a large factor.
How to handle discrete spatial data that is potentially completely unstructured, and
how to discover patterns of neighbourhood relationships between the discrete
individuals within it? New directions like particle-in-cell are suggested.
In the next few years, we can also expect more applications based on autono-
mous agents moving over a GIS-based model of the landscape, with rendering
algorithms determining what an individual agent is able to “see”. Already used to
simulate recreational activities (see for instance Bishop and Gimblett 2000),
behavioural responses to 3D virtual landscape may become more common.
Further Reading
1. The special issue of JASSS in 20011 on “ABM, Game Theory and Natural
Resource Management issues” presents a set of papers selected from a workshop
held in Montpellier in March 2000, most of them dealing with collective
decision-making processes in the field of natural resource management and
environment.
2. Gimblett (2002) is a book on integrating GIS and ABM, derived from a workshop
held in March 1998 at the Santa Fe Institute. It provides contributions from
computer scientists, geographers, landscape architects, biologists, anthropologists,
social scientists and ecologists focusing on spatially explicit simulation modelling
with agents.
3. Janssen (2002) provides a state-of-the-art review of the theory and application of
multi-agent systems for ecosystem management and addresses a number of
important topics including the participatory use of models. For a detailed review
of this book see Terna (2005).
4. Paredes and Iglesias (2008) advocate why agent based simulations provide a
new and exciting avenue for natural resource planning and management:
researches and advisers can compare and explore alternative scenarios and
institutional arrangements to evaluate the consequences of policy actions in
terms of economic, social and ecological impacts. But as a new field it demands
from the modellers a great deal of creativeness, expertise and “wise choice”, as
the papers collected in this book show.
Appendix
Topic and Issue
When multiple topics are covered by a case study, the first in the list indicates the
one we used to classify it. Within each topic we have tried to order the case studies
from the more abstract and theoretical ones to the more realistic ones. This
information can be retrieved from the issue: only case studies representing a real
system mention a geographical location.
1
http://jasss.soc.surrey.ac.uk/4/2/contents.html
Publications Model name Topic Issue Environment Agents Software
532
Polhill et al. 2001 FEARLUS LUCC Land-use and land ownership Raster 7,7 region Land manager (49) HeB Swarm
dynamics (10) Ie
Brown et al. 2004 LUCC Greenbelt to control residential Raster X,80 suburb Resident (10) Ho Ie R Swarm
development
Castella et al. 2005a SAMBA LUCC LU changes under decollecti- Raster 50,50 (25 Ha) Household (50) HeP Ie R Cormas
(Generic) vization in north Vietnam commune level
(Abstract)
Castella et al. 2005b SAMBA-GIS LUCC LU changes under Raster regional level Household (x) HeP Ie R Cormas
(More decollectivization (More realistic) (x ¼ nb household in ArcView
realistic) in north Vietnam each village)
D’Aquino et al. 2003 SelfCormas LUCC; Competing rangeland Raster 20,20 (5Ha) Farmer (20) HeB(2) Ie R Cormas
agriculture; and rice cropping village
livestock land-uses in Senegal
Etienne et al. 2003 Mejan LUCC; livestock; Pine encroachment Raster 23793 (1 ha) Farmer (40) HeB Ie,Ii,Ic C Cormas
forestry; on original landscapes plateau forester (2) HeB Ie,Ii,Ic MapInfo
wildlife in France R national park (1) Ho
Ie,Ii,Ic C
Mathevet et al. 2003a GEMACE LUCC; Competing hunting and hunting Raster region Farmer Ie, Ii hunting Cormas
agriculture; activities in Camargue manager Ie, Ii
wildlife (South of France)
Barreteau and Bousquet SHADOC Water; agriculture Viability of irrigated systems Farmer
2000; Barreteau et al. in the Senegal river valley
2004
Le Bars et al. 2005 Manga Water Farmer (n) Ie,Ii C water
supplier (1) Ic C
Feuillette et al. 2003 Sinuse Water; agriculture Water table level Raster 2400 (1Ha) Farmer HeP Ie,Ii,Ic Cormas
watershed
Lansing and Kremer Water; agriculture Water management and water temple Network 172 watershed Village (172) HeB(49)
1993; Janssen 2007 networks in Bali Ie,Ic R
Raj Gurung et al. 2006 Limbukha Water Negotiation of irrigation water sharing Grid 8,13 (10.4 Ha) Farmer (12) HeB Ii Cormas
management between two Bhutanese (Abstract) Village
communities
Hoffmann et al. 2002 LUCIM Forestry; LUCC Deforestation and afforestation in Raster 100 state Farmer (10) HeB Ie R
Indiana USA
C. Le Page et al.
Purnomo and Guizol Forestry Forest plantation comanagement in Raster 50,50 () Forest Developer Ie,Ii,Ic
2006 Indonesia massif Smallholder Ie,Ii,Ic C
Broker Ie,Ii,Ic
government Ie,Ii,Ic
Mathevet et al. 2003b ReedSim Wildlife Water management in Mediterranean Cormas
19
reedbeds
An et al. 2005 Wildlife Impact of the growing rural population on
the forests and panda habitat in China
Bousquet et al. 2001 Djemiong Wildlife Traditional hunting of small antelopes in Raster 2042 (4Ha) Hunter (90) HeP Ie,Ic R Cormas
Cameroon village antelop (7350) HeB(3)
Ie,Ic R
Anwar et al. 2007 Wildlife Interactions between whale-watching Raster Boat C whale R Repast
boats and whales in the St. Lawrence
estuary
Elliston and Beare 2006 Agriculture; Agricultural pest and disease incursions Cormas
epidemiology in Australia
Happe et al. 2006 AgriPolis Agriculture; Policy impact on structural changes of Raster 73439 (1Ha) Farms (2869) HeP Ic
lifestock W. Europe farms region
Barnaud et al. MaeSalaep 2.2 Agriculture; LU strategies in transitional swidden Raster (300 Ha) Farmer (12) HeP Ii R credit Cormas
(forthcoming) LUCC agricultural systems, Thailand (Realistic) village sources (3)
catchment
Gross et al. 2006 Lifestock Evaluation of behaviours of rangeland Entreprise C government
systems in Australia
Courdier et al. 2002 Biomas Livestock; Collective management of animal wastes Network region Livestock farm (48) HeP Ie, Geamas
agriculture in La Reunion Ii,Ic crop farm (59) HeP
Ie,Ii,Ic shipping agent
(34) HeP Ii
Bagni et al. 2002 Epidemiology; Evaluation of sanitary policies to control Abstract farm Farm sector Cow Swarm
livestock the spread of a viral pathology of
bovines
Rateb et al. 2005 Epidemiology Impact of education on malaria Raster country Inhabitant HeP R StarLogo
healthcare in Haiti
Muller et al. 2004 Epidemiology; Risk and control strategies of African Network village Villager Ic R MadKit
agriculture; Trypanosomiasis in Southern
livestock Cameroon
Agent-Based Modelling and Simulation Applied to Environmental Management
533
Environment
• First line: mode of representation, with the general following pattern:
½none; network; raster; vector NðxÞ
N indicates the number of elementary spatial entities (nodes of network, cells

or polygons), when raster mode, N is given as number of lines x number of
columns, unless some cells have been discarded from the rectangular grid
because they were out of bound (then only the total number is given), and (x)
indicates the spatial resolution.
• Second line: level of organization at which the issue is considered (for instance
village; biophysical entity (watershed, forest massif, plateau, etc.); city; conur-
bation; province, country, etc.)
Agents
One line per type of agent (the practical definition given in this paper applies,
regardless of the terminology used by the authors). The general pattern of informa-
tion looks like:
nameðxÞ½Ho; HeP; HeBðyÞ½Ie; Ii; Ic ½R; C
• (x) indicates the number of instances defined when initializing a standard

scenario, italic mentions that this initial number change during simulation.
• When x > 1, to account for the heterogeneity of the population of agents, we
propose the following coding: “Ho” stands for a homogeneous population
(identical agents), “He” stands for a heterogeneous population. “HeP” indicates
that the heterogeneity lies only in parameter values, while “HeB” indicates that
the heterogeneity lies in behaviours. In such a case, each agent is equipped with
one behavioural module selected from a set of (y) existing ones. Italic points out
adaptive agents updating either parameter value (HeP) or behaviour (HeB)
during simulation.
• [Ie, Ii, Ic] indicates the nature of relationships as defined in the text and shown in
Fig. 19.3.
• [R; C] indicates if agents are clearly either reactive or cognitive
References
Abel T (1998) Complex adaptive systems, evolutionism, and ecology within anthropology:
interdisciplinary research for understanding cultural and ecological dynamics. Georgia J
Ecol Anthropol 2:6–29
Abrami G (2004) Niveaux d’organisation dans la modélisation multi-agent pour la gestion de
ressources renouvelables. Application à la mise en oeuvre de règles collectives de gestion de
l’eau agricole dans la basse-vallée de la Drôme. Ecole Nationale du Génie Rural, des Eaux et
Forêts, Montpellier, France
Allen TFH, Starr TB (1982) Hierarchy: perspectives for ecological complexity. University of
Chicago Press, Chicago
An L, Linderman M, Qi J, Shortridge A, Liu J (2005) Exploring complexity in a
human–environment system: an agent-based spatial model for multidisciplinary and multiscale
integration. Ann Assoc Am Geogr 95(1):54–79
Anwar SM, Jeanneret C, Parrott L, Marceau D (2007) Conceptualization and implementation of a
multi-agent model to simulate whale-watching tours in the St. Lawrence estuary in Quebec,
Canada. Environ Model Software 22(12):1775–1787
Auger P, Charles S, Viala M, Poggiale J-C (2000) Aggregation and emergence in ecological
modelling: integration of ecological levels. Ecol Model 127:11–20
Bagni R, Berchi R, Cariello P (2002) A comparison of simulation models applied to epidemics.
J Artif Soc Soc Simul 5(3), 5
Bah A, Touré I, Le Page C, Ickowicz A, Diop AT (2006) An agent-based model to understand the
multiple uses of land and resources around drilling in Sahel. Math Comp Model 44(5–6):513–534
Bakam I, Kordon F, Le Page C, Bousquet F (2001) Formalization of a multi-agent model using colored
petri nets for the study of an hunting management system. In: Rash JL, Rouff C, Truszkowski W,
Gordon DF, Hinchey MG (eds) Formal approaches to agent-based systems, first international
workshop, FAABS 2000 (Lecture notes in computer science, 1871). Springer, Greenbelt, pp 123–132
Balmann A (1997) Farm based modelling of regional structural change: a cellular automata
approach. Eur Rev Agric Econ 24(1):85–108
Barnaud C, Promburom T, Trébuil G, Bousquet F (2007) An evolving simulation and gaming
process for adaptive resource management in the highlands of northern Thailand. Simulat
Gaming 38(3):398–420
Barreteau O, Bousquet F (2000) SHADOC: a multi-agent model to tackle viability of irrigated
systems. Ann Oper Res 94:139–162
Barreteau O, Bousquet F, Millier C, Weber J (2004) Suitability of multi-agent simulations to study
irrigated system viability: application to case studies in the Senegal River valley. Agr Syst
80:255–275
Batten D (2007) Are some human ecosystems self-defeating? Environ Model Software
22:649–655
Becu N, Perez P, Walker A, Barreteau O, Le Page C (2003) Agent based simulation of a small
catchment water management in northern Thailand: description of the CATCHSCAPE model.
Ecol Model 170(2–3):319–331
Berger T (2001) Agent-based spatial models applied to agriculture: a simulation tool for technol-
ogy diffusion, resource use changes and policy analysis. Agr Econ 25:245–260
Berger T, Schreinemachers P (2006) Creating agents and landscapes for multiagent systems from
random samples. Ecol Soc 11(2):19, http://www.ecologyandsociety.org/vol11/iss2/art19/
Berger T, Schreinemachers P, Woelcke J (2006) Multi-agent simulation for the targeting of
development policies in less-favored areas. Agr Syst 88:28–43
Berkes F, Folke C (eds) (1998) Linking ecological and social systems: management practices and
social mechanisms for building resilience. Cambridge University Press, Cambridge, UK
Bian L (2004) A conceptual framework for an individual-based spatially explicit epidemiological
model. Environ Plann B Plann Des 31(3):381–395
Bishop I, Gimblett R (2000) Management of recreational areas: GIS, autonomous agents, and
virtual reality. Environ Plann B Plann Des 27(3):423–435
Bithell M, Macmillan WD (2007) Escape from the cell: spatially explicit modelling with and
without grids. Ecol Model 200:59–78
Bonaudo T, Bommel P, Tourrand JF (2005) Modelling the pioneers fronts of the transamazon
highway region. In: Proceedings of CABM-HEMA-SMAGET 2005, Joint conference on
multi-agent modeling for environmental management, Bourg St Maurice, Les Arcs, 21–25
Mar 2005. http://smaget.lyon.cemagref.fr/contenu/SMAGET%20proc/PAPERS/
BonaudoBommelTourrand.pdf
Bousquet F (2001) Modélisation d’accompagnement. Simulations multi-agents et gestion des
ressources naturelles et renouvelables (Unpublished Habilitation à Diriger des Recherches).
Lyon I University, Lyon
Bousquet F, Le Page C (2004) Multi-agent simulations and ecosystem management: a review.
Ecol Model 176:313–332
Bousquet F, Barreteau O, Le Page C, Mullon C, Weber J (1999) An environmental modelling
approach: the use of multi-agents simulations. In: Blasco F, Weill A (eds) Advances in
environmental and ecological modelling. Elsevier, Paris, pp 113–122
Bousquet F, Le Page C, Bakam I, Takforyan A (2001) Multiagent simulations of hunting wild
meat in a village in eastern Cameroon. Ecol Model 138:331–346
Bousquet F et al (2002) Multi-agent systems and role games: collective learning processes for
ecosystem management. In: Janssen MA (ed) Complexity and ecosystem management: the
theory and practice of multi-agent systems. Edward Elgar, Cheltenham, pp 248–285
Bousquet F, Trébuil G, Hardy B (eds) (2005) Companion modeling and multi-agent systems for
integrated natural resources management in Asia. International Rice Research Institute, Los
Baños
Bradbury R (2002) Futures, prediction and other foolishness. In: Janssen MA (ed) Complexity and
ecosystem management: the theory and practice of multi-agent systems. Edward Elgar,
Brown DG, Page SE, Riolo R, Rand W (2004) Agent-based and analytical modeling to evaluate
the effectiveness of greenbelts. Environ Model Software 19:1097–1109
Caplat P, Lepart J, Marty P (2006) Landscape patterns and agriculture: modelling the long-term
effects of human practices on Pinus sylvestris spatial dynamics (Causse Mejean, France).
Landscape Ecol 21(5):657–670
Carley KM et al (2006) BioWar: scalable agent-based model of bioattacks. IEEE T Syst Man Cy
A: Syst Hum 36(2):252–265
Carpenter S, Brock W, Hanson P (1999) Ecological and social dynamics in simple models of
ecosystem management. Conserv Ecol 3(2), 4
Castella J-C, Verburg PH (2007) Combination of process-oriented and pattern-oriented models of
land use change in a mountain area of Vietnam. Ecol Model 202(3–4):410–420
Castella J-C, Boissau S, Trung TN, Quang DD (2005a) Agrarian transition and lowland–upland
interactions in mountain areas in northern Vietnam: application of a multi-agent simulation
model. Agr Syst 86:312–332
Castella J-C, Trung TN, Boissau S (2005b) Participatory simulation of land-use changes in the
northern mountains of Vietnam: the combined use of an agent-based model, a role-playing
game, and a geographic information system. Ecol Soc 10(1), 27
ComMod (2003) Our companion modelling approach. J Artif Soc Soc Simul 6(2). http://jasss.soc.
Courdier R, Guerrin F, Andriamasinoro FH, Paillat J-M (2002) Agent-based simulation of
complex systems: application to collective management of animal wastes. J Artif Soc Soc
D’Aquino P, Le Page C, Bousquet F, Bah A (2003) Using self-designed role-playing games and a
multi-agent system to empower a local decision-making process for land use management: the
SelfCormas experiment in Senegal. J Artif Soc Soc Simul 6(3). http://jasss.soc.surrey.ac.uk/6/

3/5.html
Deadman P, Robinson D, Moran E, Brondı́zio E (2004) Colonist household decision making and
land-use change in the Amazon rainforest: an agent-based simulation. Environ Plann B Plann
Des 31(5):693–709
DeAngelis DL, Gross LJ (eds) (1992) Individual-based models and approaches in ecology.
Chapman and Hall, New York
Dunham JB (2005) An agent-based spatially explicit epidemiological model in MASON. J Artif
Soc Soc Simulat 9(1). http://jasss.soc.surrey.ac.uk/9/1/3.html
Elliston L, Beare S (2006) Managing agricultural pest and disease incursions: an application of
agent-based modelling. In: Perez P, Batten D (eds) Complex science for a complex world:
exploring human ecosystems with agents. ANU E Press, Canberra, pp 177–189
Brookings Institution Press, Washington, DC
Etienne M (ed) (2011) Companion modelling: a participatory approach to support sustainable
development. Quae, Versailles
Etienne M, Le Page C, Cohen M (2003) A step-by-step approach to building land management
scenarios based on multiple viewpoints on multi-agent system simulations. J Artif Soc Soc
Evans TP, Kelley H (2004) Multi-scale analysis of a household level agent-based model of
landcover change. J Environ Manage 72:57–72
Ferber J, Gutknecht O (1998) A meta-model for the analysis and design of organizations in multi-
agent systems. In: Demazeau Y (ed) Proceedings of the 3rd international conference on multi-
agent systems (ICMAS’98), Cité des sciences – La Villette, Paris, 4–7 July 1998. IEEE, Los
Alamos, pp 128–135
Feuillette S, Bousquet F, Le Goulven P (2003) SINUSE: a multi-agent model to negotiate water
demand management on a free access water table. Environ Model Software 18:413–427
Galán J et al. (2013) Checking simulations: detecting and avoiding errors and artefacts. Chapter 6
in this volume
Galvin KA, Thornton PK, Roque de Pinho J, Sunderland J, Boone RB (2006) Integrated modeling
and its potential for resolving conflicts between conservation and people in the rangelands of
East Africa. Hum Ecol 34(2):155–183
Gimblett R (ed) (2002) Integrating geographic information systems and agent-based modeling
techniques for understanding social and ecological processes (Santa Fe Institute studies in the
sciences of complexity). Oxford University Press, Oxford
Grimm V, Railsback SF (2005) Individual-based modeling and ecology. Princeton University
Press, Princeton
Gross JE, McAllister RRJ, Abel N, Stafford Smith DM, Maru Y (2006) Australian rangelands as
complex adaptive systems: a conceptual model and preliminary results. Environ Model
Software 21(9):1264–1272
Gunderson LH, Holling CS (2002) Panarchy: understanding transformations in human and natural
systems. Island Press, Washington, DC
Gurung TR, Bousquet F, Trébuil G (2006) Companion modeling, conflict resolution, and
institution building: sharing irrigation water in the Lingmuteychu watershed, Bhutan. Ecol
Soc 11(2), 36
Haffner Y, Gramel S (2001) Modelling strategies for water supply companies to deal with nitrate
pollution. J Artif Soc Soc Simul 4(3). http://jasss.soc.surrey.ac.uk/4/3/11.html
Hales D, Rouchier J, Edmonds B (2003) Model-to-model analysis. J Artif Soc Soc Simul 6(4).
Happe K, Kellermann K, Balmann A (2006) Agent-based analysis of agricultural policies: an

illustration of the agricultural policy simulator AgriPoliS, its adaptation and behavior. Ecol Soc
11(1), 49
Hare M, Deadman P (2004) Further towards a taxonomy of agent-based simulation models in
environmental management. Math Comput Simul 64:25–40
Hemelrijk C (2013) Simulating complexity of animal social behaviour. Chapter 22 in this volume
Hoffmann M, Kelley H, Evans T (2002) Simulating land-cover change in south-central Indiana: an
agent-based model of deforestation and afforestation. In: Janssen MA (ed) Complexity and
Holling CS (1986) The resilience of terrestrial ecosystems; local surprise and global change. In:
Clark WC, Munn RE (eds) Sustainable development of the biosphere. Cambridge University
Press, Cambridge, pp 292–317
Holling CS (2001) Understanding the complexity of economic, ecological, and social systems.
Ecosystems 4:390–405
Holling CS (2004) From complex regions to complex worlds. Ecol Soc 9(1), 11
Huigen MGA (2004) First principles of the MameLuke multi-actor modelling framework for land
use change, illustrated with a Philippine case study. J Environ Manage 72:5–21
Huigen MGA, Overmars KP, de Groot WT (2006) Multiactor modeling of settling decisions and
behavior in the San Mariano watershed, the Philippines: a first application with the MameLuke
framework. Ecol Soc 11(2), 33
Huston M, DeAngelis DL, Post W (1988) New computer models unify ecological theory.
Bioscience 38(10):682–691
Janssen MA (2001) An exploratory integrated model to assess management of lake eutrophication.
Ecol Model 140:111–124
Janssen MA (ed) (2002) Complexity and ecosystem management: the theory and practice of multi-
agent systems. Edward Elgar, Cheltenham
Janssen MA (2007) Coordination in irrigation systems: an analysis of the Lansing–Kremer model
of Bali. Agr Syst 93:170–190
Janssen MA, Ostrom E (2006) Empirically based, agent-based models. Ecol Soc 11(2)
Janssen MA, Walker BH, Langridge J, Abel N (2000) An adaptive agent model for analysing
co-evolution of management and policies in a complex rangeland system. Ecol Model 131
(2–3):249–268
Janssen MA, Anderies JM, Stafford Smith M, Walker BH (2002) Implications of spatial heteroge-
neity of grazing pressure on the resilience of rangelands. In: Janssen MA (ed) Complexity and
Jepsen JU, Topping CJ, Odderskær P, Andersen PN (2005) Evaluating consequences of land-use
strategies on wildlife populations using multiple-species predictive scenarios. Agr Ecosyst
Environ 105:581–594
Judson OP (1994) The rise of the individual-based model in ecology. Trends Ecol Evol 9(1):9–14
Kawata M, Toquenaga Y (1994) From artificial individuals to global patterns. Trends Ecol Evol 9
(11):417–421
Kinzig AP et al (2006) Resilience and regime shifts: assessing cascading effects. Ecol Soc 11(1), 20
Krywkow J, Valkering P, Rotmans J, van der Veen A (2002) Agent-based and integrated
assessment modelling for incorporating social dynamics in the management of the Meuse in
the Dutch province of Limburg. In: Rizzoli AE , Jakeman AJ (eds) Integrated assessment and
decision support: proceedings of the first biennial meeting of the international environmental
modelling and software society iEMSs, 24–27 June 2002, University of Lugano, Switzerland,
Vol 2. iEMSs, Manno, pp 263–268
Lansing JS, Kremer JN (1993) Emergent properties of Balinese water temple networks: coadapta-
tion on a rugged fitness landscape. Am Anthropol 95(1):97–114
Le Bars M, Attonaty J-M, Ferrand N, Pinson S (2005) An agent-based simulation testing the
impact of water allocation on farmers’ collective behaviors. Simulation 81(3):223–235
Levin SA (1998) Ecosystems and the biosphere as complex adaptive systems. Ecosystems
1:431–436
Levin SA (1999) Towards a science of ecological management. Conserv Ecol 3(2), 6
Lim K, Deadman PJ, Moran E, Brondizio E, McCracken S (2002) Agent-based simulations of
household decision making and land use change near Altamira, Brazil. In: Gimblett R (ed)
Integrating geographic information systems and agent-based modeling techniques for under-
standing social and ecological processes (Santa Fe Institute studies in the sciences of complex-
ity). Oxford University Press, Oxford, pp 277–310
López Paredes A, Hernández Iglesias C (eds) (2008) Agent-based modelling in natural resource
management. INSISOC, Valladolid
Manson S (2005) Agent-based modeling and genetic programming for modeling land change in
the southern Yucatan Peninsular region of Mexico. Agr Ecosyst Environ 111:47–62
Manson S (2006) Land use in the southern Yucatan peninsular region of Mexico: scenarios of
population and institutional change. Comput Environ Urban Syst 30:230–253
Marietto MB, David N, Sichman JS, Coelho H (2003) Requirements analysis of agent-based
simulation platforms: state of the art and new prospects. In: Sichman JS, Bousquet F,
Davidsson P (eds) MABS’02, Proceedings of the 3rd international conference on multi-
agent-based simulation (Lecture notes in artificial intelligence, 2581). Springer, Berlin,
pp 125–141
Mathevet R, Bousquet F, Le Page C, Antona M (2003a) Agent-based simulations of interactions
between duck population, farming decisions and leasing of hunting rights in the Camargue
(Southern France). Ecol Model 165(2/3):107–126
Mathevet R, Mauchamp A, Lifran R, Poulin B, Lefebvre G (2003b) ReedSim: simulating
ecological and economical dynamics of Mediterranean reedbeds. In: Proceedings of
MODSIM03, International congress on modelling and simulation, integrative modelling of
biophysical, social, and economic systems for resource management solutions, Townsville,
14–17 July 2003. Modelling & Simulation Society of Australia & New Zealand. http://mssanz.
org.au/MODSIM03/Volume_03/B02/01_Mathevet.pdf
Matthews R (2006) The people and landscape model (PALM): towards full integration of human
decision-making and biophysical simulation models. Ecol Model 194:329–343
Milner-Gulland EJ, Kerven C, Behnke R, Wright IA, Smailov A (2006) A multi-agent system
model of pastoralist behaviour in Kazakhstan. Ecol Complex 3(1):23–36
Moreno N et al (2007) Biocomplexity of deforestation in the Caparo tropical forest reserve in
Venezuela: an integrated multi-agent and cellular automata model. Environ Model Software
22:664–673
Muller G, Grébaut P, Gouteux J-P (2004) An agent-based model of sleeping sickness: simulation
trials of a forest focus in southern Cameroon. C R Biol 327:1–11
Nute D et al (2004) NED-2: an agent-based decision support system for forest ecosystem
management. Environ Model Software 19:831–843
Parker DC, Manson SM, Janssen MA, Hoffmann MJ, Deadman P (2003) Multi-agent systems for
the simulation of land-use and land-cover change: a review. Ann Assoc Am Geogr 93
(2):314–337
Pepper JW, Smuts BB (2000) The evolution of cooperation in an ecological context: an agent-
based model. In: Kohler TA, Gumerman GJ (eds) Dynamics in human and primate societies
(Santa Fe Institute studies in the sciences of complexity). Oxford University Press, New York,
pp 45–76
Perez P, Batten D (eds) (2006) Complex science for a complex world: exploring human
ecosystems with agents. ANU E Press, Canberra
Polhill JG, Gotts NM, Law ANR (2001) Imitative versus non-imitative strategies in a land use
simulation. Cybern Syst 32(1–2):285–307
Polhill JG, Izquierdo LR, Gotts NM (2006) What every agent-based modeller should know about
floating point arithmetic. Environ Model Software 21:283–309
Purnomo H, Guizol P (2006) Simulating forest plantation co-management with a multi-agent
system. Math Comput Model 44:535–552
Purnomo H, Mendoza GA, Prabhu R, Yasmi Y (2005) Developing multi-stakeholder forest
management scenarios: a multi-agent system simulation approach applied in Indonesia. Forest
Policy Econ 7(4):475–491
Rateb F, Pavard B, Bellamine-BenSaoud N, Merelo JJ, Arenas MG (2005) Modeling malaria with
multi-agent systems. Int J Intell Info Technol 1(2):17–27
Rouchier J, Bousquet F, Barreteau O, Le Page C, Bonnefoy J-L (2000) Multi-agent modelling and
renewable resources issues: the relevance of shared representation for interacting agents. In:
Moss S, Davidsson P (eds) Multi-agent-based simulation, second international workshop,
MABS 2000, Boston, July; revised and additional papers (Lecture notes in artificial intelli-
gence, 1979). Springer, Berlin, pp 341–348
Rouchier J, Bousquet F, Requier-Desjardins M, Antona M (2001) A multi-agent model for
describing transhumance in North Cameroon: comparison of different rationality to develop
a routine. J Econ Dyn Control 25:527–559
Schreinemachers P, Berger T (2006) Land-use decisions in developing countries and their
representation in multi-agent systems. J Land Use Sci 1(1):29–44
Servat D, Perrier E, Treuil J-P , Drogoul A (1998) When agents emerge from agents: introducing
multi-scale viewpoints in multi-agent simulations. In: Sichman JS, Conte R, Gilbert N (eds)
Proceedings of the first international workshop on multi-agent systems and agent-based
simulation (Lecture notes in artificial intelligence, 1534). Springer, London, pp 183–198
Simon HA (1973) The organization of complex systems. In: Pattee HH (ed) Hierarchy theory: the
challenge of complex systems. Braziller, New York, pp 1–27
Soulié J-C, Thébaud O (2006) Modeling fleet response in regulated fisheries: an agent-based
approach. Math Comput Model 44:553–554
Sulistyawati E, Noble IR, Roderick ML (2005) A simulation model to study land use strategies in
swidden agriculture systems. Agric Syst 85:271–288
Terna P (2005) Review of “complexity and ecosystem management: the theory and practice of
multi-agent systems”. J Artif Soc Soc Simul 8(2). http://jasss.soc.surrey.ac.uk/8/2/reviews/
terna.html
Thébaud O, Locatelli B (2001) Modelling the emergence of resource-sharing conventions: an
agent-based approach. J Artif Soc Soc Simul 4(2). http://jasss.soc.surrey.ac.uk/4/2/3.html
Ziervogel G, Bithell M, Washington R, Downing T (2005) Agent-based social simulation: a
method for assessing the impact of seasonal climate forecast applications among smallholder
farmers. Agric Syst 83:1–26
Zunga Q et al (1998) Coupler Systèmes d’Information Géographiques et Systèmes Multi-Agents
pour modéliser les dynamiques de transfotmation des paysages. Le cas des dynamiques
foncières de la Moyenne Vallée du Zambèze (Zimbabwe). In: Ferrand N (ed) Modèles et
systèmes multi-agents pour la gestion de l’environnement et des territoires. Cemagref Editions,
Clermont-Ferrand, pp 193–206
Chapter 20
Assessing Organisational Design
Virginia Dignum
Why Read This Chapter? To understand the requirements of an adequate formal

model of organisations; in particular, to understand the main components of an
organisational model and how they might be formalised. Also to be aware of how
organisational change might be represented and assessed using formal models.
Abstract This chapter highlights the use of formal models for the analysis and
change of organisational designs. In order to survive, organisations must be able to
maintain a good fit with the environment, and therefore be able to determine when
and how changes should take place. Current work in Organisational Theory is taken
as a basis for the development of a theoretical model in which organisational
performance and change measures can be defined and analyzed. In addition, this
chapter presents a prototype framework for the simulation of reorganization.
20.1 Introduction
Organisations are complex dynamic goal-oriented processes, with outcomes depen-

dent both on the contextual situation as on the coordinated and purposeful action of
human beings in the environment. Providing organisation designers with knowl-
edge and tools that enable them to effectively identify, assess and respond to
environment changes is therefore of utmost importance for the survival of the
organisation. One of the main reasons for creating organisations is to provide the
means for coordination that enable the achievement of global goals. Organisations
can be seen as open systems in that they have to interact with their environment in
order to realize those goals (Aldrich and Pfeffer 1976). The design of organisations
must consider both structural and human components. Organisational structure has
V. Dignum (*)
Faculty Technology Policy and Management, section ICT, Delft University of Technology,
Delft, The Netherlands

542 V. Dignum
essentially two objectives (Duncan 1979): Firstly, it facilitates the flow of informa-
tion within the organisation in order to reduce the uncertainty of decision making.
Secondly, the structure of the organisation should integrate organisational behavior
across the parts of the organisation so that it is coordinated. This raises two
challenges: division of labor and coordination (Mintzberg 1993). The design of
organisational structure determines the allocation of resources and people to
specified tasks or purposes, and the coordination of these resources to achieve
organisational goals (Galbraith 1977). Ideally, the organisation is designed to fit
its environment and to provide the information and coordination needed for its
strategic objectives.
Organisational design mostly adopts a multi-contingency view, which says that
an organisation’s design should be chosen based on the multi-dimensional
characteristics of a particular environment or context. These characteristics include
structural (e.g. goals, strategy and structure) and human (e.g. work processes,
people, coordination and control) components. Contingency theory states that
even if there is not one best way to organize, some designs perform better than
others. In fact, Roberts states that the central problem of organisation design is
finding a strategy and an organisation structure that are consistent with each other.
This requires finding a dynamic balance in the face of strategic and organisational
changes (Roberts 2004). Given that there are no exact recipes to construct the
optimal organisation, organisational assessment is about evaluating a certain design
and determine its appropriateness given the organisation’s aims and constraints.
In order to access organisational design and performance, the fact that
organisations and their members are independent entities must be considered.
Taking this point of view, agents participating in organisations are not merely
fulfilling some roles of an organisation to realize global objectives set out by the
organisational structures. They have an independent existence (like in human
organisations people have an existence independent from the organisation). Follow-
ing this approach, it becomes very important to describe the exact relation between
the organisational entities on the one hand and the agents on the other hand.
The above considerations provide two different motivations for this work. On
the one hand, the need for a formal representation of organisations, with their
environment, objectives and agents in a way that enables to analyze their partial
contributions to the performance of the organisation in a changing environment.
These formalisms will enable the analysis of abstract organisation models and the
verification of formal properties such as flexibility or robustness (Grossi et al. 2007)
in dynamic environments (Dignum and Dignum 2007). On the other hand, the need
for such a model to be realistic enough to incorporate the more ‘pragmatic’
considerations faced by real organisations. Most existing formal models lack this
realism, e.g. either by ignoring temporal issues, or by taking a very restrictive view
on the controllability of agents, or by assuming complete control and knowledge
within the system (cf. van der Hoek and Wooldridge 2005; Santos et al. 1997). Such
instruments must moreover be both formal and realistic.
Organisational assessment requires that there is a set of parameters against
which organisational performance is to be evaluated. Organisation design processes
20 Assessing Organisational Design 543
usually start with an assessment of what is the current state and where there are gaps
between desired and actual outcomes. Design strategies will determine what is the
focus of the design process. Some organisations place a higher priority on effi-
ciency, preferring designs that minimize the costs of producing goods or services.
Other organisations emphasize effectiveness, focusing on generating revenues or
seizing leading-edge innovation in the marketplace (Burton et al. 2006). In practice,
the process of organisational assessment relies on individual experience and skill of
the assessor. Even if best practices exist, these are often not formal or sufficiently
detailed to enable an objective comparison and uniform application. In this chapter,
we will introduce an abstract model to describe organisations and their
environments that allows for formal analysis of the fit between organisational
design and environment characteristics. This enables the a priori comparison of
designs and their consequences and therefore supports the decision making process
on the choice of design. The chapter is organised as follows: In Sect. 20.2, we will
describe in detail the characteristics and requirements of the organisational design
process. In Sect. 20.3 we present the formal model for organisation specification.
The model is further extended in Sect. 20.4 to support the analysis of organisational
change. Finally, in Sect. 20.5 we describe a simulation framework that implements
this formalism. Conclusion and directions for future work are given in Sect. 20.6.
20.2 Organisational Design
Organisations are inherently computational and (distributed) computational

systems are inherently organisational (Carley 2002). As such, the use of agent
based models and simulation to formally capture and understand organisations has
already yielded positive results and promises further tremendous potential
(Davidsson 2002). Also, in the Multi-Agent Systems (MAS) research area there is
an increased realization that concepts from Organisation Theory are well applicable
to understand and structure MAS. Agent organisations can be understood as
complex entities where a multitude of agents interact, within a structured environ-
ment aiming at some global purpose. Models for the assessment of organisational
design must be able to separate organisational from individual concerns.
Agent-based models provide an abstract representation of organisations, their
environment, objectives and participating agents that enables the analysis of their
partial contributions to the performance of the organisation in a changing environ-
ment. The agent-based approach views an organisation as a set of mechanisms of
social order that regulate autonomous actors (agents) to achieve common goals
(Dignum 2004). The performance of an organisation in such framework is deter-
mined both by the structures of interactions between agents, and by the individual
characteristics of the agents. Organisational structures define the formal lines of
communication, allocation of information processing tasks, distribution of
decision-making authorities, and the provision of incentives. From an
organisational perspective, the main function of an individual agent is the
544 V. Dignum
enactment of a role that contributes to the global aims of the organisation. That is,
organisational goals determine agent roles and interaction norms. Agents are then
seen actors that perform role(s) described by the organisation design. However, the
very notion of agent autonomy refers to the ability of individual agents to determine
their own actions, plans and beliefs. From the agent’s perspective, its own
capabilities and aims determine the reasons and the specific way an agent will
enact its role(s), and the behavior of individual agents is motivated from their own
goals and capabilities (Dastani et al. 2003; Weigand and Dignum 2004). That is,
agents bring their own ways into the society as well.
Role-oriented approaches to agent systems assume this inherent dichotomy
between agent interests and global aims (Castelfranchi 1995; Dignum 2004;
Zambonelli et al. 2001; Artikis and Pitt 2001). This indicates two levels of
autonomy:
• Internal autonomy: interaction and structure of the society must be represented
independently from the internal design of the agents
• Collaboration autonomy: activity and interaction in the society must be
specified without completely fixing in advance the interaction structures.
The first requirement relates to the fact that in open environments, organisations
allow the participation of multiple heterogeneous entities, the number,
characteristics, aims and architecture of which are unknown to the designer, the
design of the organisation cannot be dependent on their individual designs. With
respect to the second requirement, fundamentally, a tension exists between the
goals of the organisation decision makers and the autonomy of the participating
entities. On the one hand, the more detail the organisation can use to specify its aims
and processes, the more requirements are possible to check and guarantee at design
time. This allows, for example, to ensure the legitimacy of the interactions, or that
certain rules are always followed. On the other hand, there are good reasons to
allow the agents some degree of freedom, basically to enable their freedom to
choose their own way of achieving collaboration, and as such increase flexibility
and adaptability.
Existing computational simulations of organisations, based on formal models
provide a precise theoretical formulation of relationships between variables can
inform strategic decisions concerning reorganization (Dignum et al. 2006) but are
often limited to a specific domain and difficult to validate (Harrison et al. 2007).
Another aspect to consider is the fact that organisations and their environments
are not static. Agents can migrate, organisational objectives can change, or opera-
tional behavior can evolve. That is, as circumstances change, organisational
structures must also be able to change, disappear or grow. In fact, organisations
are active entities, capable not only of adapting to the environment but also of
changing that environment, which leads to the question of how and why change
decisions are made in organisations (Gazendam and Simons 1998). Models for
organisations must therefore be able to describe dynamically adapting
organisations to changes in the environment. Such models will enable to understand
how different organisations can be designed from different populations of agents,
performing different tasks in changing environments, to meet various performance

goals (So and Durfee 1998).
To sum up, formal models for organisations that are able to deal with realistic
situations, must meet the following requirements (Dignum and Dignum 2007):
1. Represent notions of ability and activity of agents, without requiring knowl-
edge about the specific actions available to a specific agent (internal
autonomy).
2. Accept limitedness of agent capability.
3. Represent the abilities and activities of a group of agents.
4. Deal with temporal issues, and especially the fact that activity takes time.
5. Represent the concept of ‘being responsible’ for the achievement of a given
state of affairs.
6. Represent organisational (global) goals and their dependency on agents’ activ-
ity, by relating activity and organisational structure.
7. Represent organisations in terms of organisational roles that enable individuals
to act in accordance with their own capabilities and requirements (collaboration
autonomy).
8. Relate roles and agents (role enacting agents).
9. Deal with resource limitedness and the dependency of activity on resources.
10. Represent organisational dynamics (evolution of organisation over time,
changes in agent population).
11. Deal with normative issues (representation of boundaries for action and the
violation thereof).
The first six requirements above are related to the structural properties of an
organisation. The last requirements are related to the operational aspects of an
organisation, such as the notion that agent activity has a cost (that is, choosing one
or the other course of action is not only dependent on agent capabilities but also the
costs of an action must compare positively to its benefits), which is related to actual
performance of an agent within the organisational structure.
20.3 A Model for Organisation Design
As discussed in the previous sections, one of the main reasons for creating
organisations is efficiency, that is, to provide the means for coordination that enable
the achievement of global goals in an efficient manner. Such global objectives are
not necessarily shared by any of the individual participants, that can only be
achieved through combined action. In order to achieve its goals, it is thus necessary
that an organisation employs the relevant agents, and assures that their interactions
and responsibilities enable an efficient realization of the objectives. In its most
simple expression, an organisation consists of a set of agents and a set of objectives
in a given environment. Contingency theory states that organisational performance
can be seen as a dependent variable on these factors (So and Durfee 1998):
546 V. Dignum
1. Environment: this is the space in which organisations exist. This space is not
completely controllable by the organisation and therefore results of activity are
can not always be guaranteed. Two dimensions of environment are unpredict-
ability and (task) complexity. This class also includes the description of tasks
and resources (such as size and frequency).
2. Agents: are entities with the capability to act, that is to control the state of some
element in the environment. Agents can be defined by their capabilities.
3. Structure: describe the features of the organisation, such as roles, relationships
and strategy. Differentiating dimensions in this category are size, skills, degree
of centralization and formalization.
These are the main building blocks of organisational design and should be
included in any model that aims at understanding or specifying organisational
performance or behavior. We have developed a formal language LAO (Language
for Agent Organisations) based on these concepts to describe and analyze
organisational design and performance. The definition of the semantics of LAO is
presented in (Dignum and Dignum 2007). In the following, we give an overview of
the language.
20.3.1 Environment
In Organisation Theory, environment is commonly defined as the forces outside the

organisation that can impact it (Daft et al. 2010). The environment changes over
time are not fully controlled by the organisation nor the individuals that populate it.
For a formal representation of the environment, we use Kripke semantics to
describe the environment and its changes. In LAO, the environment is formally
represented as a partially ordered set of worlds W. A world describes an actual state
of the environment, and transitions indicate the effects of possible changes (the
opportunities and threats in terms of organisational theory concepts). The set of
atomic facts, Φ, describes the vocabulary (or ontology) of the organisation domain.
At each moment, the state of the organisation is given by a certain state of affairs
in the current world. Transitions between worlds represent changes in the situation.
Such changes can occur both by decision of the organisation or by external factors
affecting the environmental conditions. We use a branching time structure to
represent choice, that is, that there is more than one possible future for each world.
20.3.2 Agents and Groups
Agents are active entities that can make change happen, that is, agents can modify
the truth value of a fact. Transitions from one to a next world are partially controlled
by the agents. Intuitively, the idea is that, in organisations, changes are for some
part result of the intervention of (specific) agents. Formally, transitions between

worlds are labeled with the set of agents that influence the changes on that
transition. That is, for a transition r ¼ (w,w0 ),w,w0 2W, t(r) indicates the set of
agents that indeed contribute to the changes indicated by that transition. For each
world w2W and each agent a2A we can indicate the set of transitions from w for
which a has influence. Moreover, at any moment, unexpected change can also
occur, which is not result of the action of any of the agents in that world (i.e.
unlabelled or partially labelled transitions are possible).
The notions of agent capability and action have been widely discussed in MAS.
The intuition is that an agent possesses capabilities that make action possible. That
is, in order to talk about agent activity, or, that agent a possesses the ability to make
proposition φ hold in some next state in a path from the current world, we need to
establish the control of the agent over the (truth) value of φ. For each agent a we
partition the set of atomic facts Φ in any world w of M in two classes: the set of
atomic facts that agent a can control, Ca , and the set of atomic facts that a cannot
control, Ca. Control over a composite fact (i.e. not atomic fact), φ, is possible iff Ca
yields φ. For example, if we consider a world where the basic facts are sand-door
and paint-door, an agent a that only has the capability to perform sand-door, that is,
Ca (sand-door), will not be able to control the composite fact (sand-door ∧ paint-
door). Agents can control the state of variables in a world, represented by Ca φ iff the
agent controls the set of atomic facts that yield φ. Furthermore, LAO is defined such
that no agent can control the obvious (tautologies), and if an agent controls a fact φ
it also controls its negation ¬φ. Intuitively, the ability of an agent to realize a state of
affairs φ in a world w, depends not only on the capabilities of the agent but also on
the status of that world. Therefore, we define the ability of a, Ga φ to represent the
case in with the agent not only controls φ but is currently in a state in which it has
influence over some of the possible transitions. Only when the agent influences all
outgoing transitions from the current world and φ will hold in all worlds reachable
from the current world, we can talk about agent activity; in the other cases, there is a
(possibly unsuccessful) attempt. The attempt by agent a to realize φ is represented
by Ha φ. Finally, the stit operator, Ea φ (‘agent a sees to it that φ’) represents the
result of successful action (that is, φ holds in all worlds following the current one).
This notion of agent activity is based on that introduced by Pörn (1974) to represent
the externally ‘observable’ consequences of an action instead of the action itself,
and as such abstracts from internal motivations of the agents. Stit can be seen as an
abstract representation of the family of all possible actions that result in φ.
Agents are, by definition, limited on their capabilities, that is, the set of facts in
the world that they can control. This implies that in MAS certain states of affairs
can only be reached if two or more agents cooperate to bring that state to be. We
define control and action of a group of agents based on the combined atomic
capabilities of the agent in the group. For example, if a group contains two agents
a and b, such that Ca (sand-door) and Cb (paint-door), only working as a group is
possible to realize both. This is represented by Cfa;bg (sand-door ∧ paint-door).
Table 20.1 gives an overview of agent and group operators. We refer the reader
to (Dignum and Dignum 2007) for the formal specification of LAO.
548 V. Dignum
Table 20.1 Overview of agent and group operators

Agent Group Description
Capability Ca φ CS φ Control over the truth value of φ
Ability Ga φ GS φ Capability and influence over at least one outgoing transition
Attempt Ha φ HS φ Ability and φ holds in all worlds reached by an influenced transition
Action (stit) Ea φ ES φ Ability and φ holds in all next worlds
20.3.3 Structure
The organisational structure involves two basic concepts: roles and their relations
in terms of which the the overall behavior of the organisation is determined. The
specification of the overall behavior of the organisation concerns the optimization
of agents’ activities, the management and security of the information flow among
agents, and the enforcement of certain outcomes. According to Selznick (1948)
organisations “represent rationally ordered instruments for the achievement of
stated goals”, that is, organisations arise in order to achieve specific objectives,
and these objectives are pursued defining a number of sub-objectives contributing
to the overall purpose of the organisation. These sub-objectives identify the roles
that are played in the organisation. The relation between roles and overall
objectives of the organisation, i.e., the primitive decomposition of tasks within
the organisation, defines the essential form of organisational structure (Grossi et al.
2005). Roles are the basic units over which this structure ranges determining the
source of the “rational order” holding in the organisation. A formal definition of
organisation therefore needs to include the constructs to relate roles in the
organisation in a way that enables objectives to be passed to those agents that can
effectively realize them. Note that in our view roles and agents are independent
concepts. Role describe activities and services necessary to achieve organisations
goals and enable to abstract from the individuals that eventually will enact the role.
In this sense, in our model, roles can be seen as place-holders for agents and
represent the behavior expected from agents by design. On the other hand, agents
are active entities that are able to effectively enact roles in order to achieve
organisational objectives. That is, objectives of an organisation, DO are assigned
to roles but can only be achieved through agent action. In order to make this
possible, an organisation must in any case employ the relevant agents, that is agents
with the right capabilities. In the following, we will use the notion of role enaction
agent (Dignum 2004) to refer to the combination of agent and role.
In most cases, the objectives of the organisation are only known to a few of the
agents in the organisation, who may have no control over those objectives. It is
therefore necessary to organize agents in the organisation in a way that enables
objectives to be passed to those agents that can effectively realize them. Depen-
dency between agents a and b is represented by a b meaning that agent a is able
to delegate the realization of some state of affairs to agent b. Agents can be made
responsible for some of the objectives of the organisation, represented by Ra φ.1

Responsibility means that either the agent attempts to achieve that objective, or it
delegates the objective to another agent with whom it has a dependency relation.
That is, given a proposition φ, such that Ra φ, the order relation a O b indicates that
a is capable to pass its responsibility for φ to b, that is, Ca Rb φ.
20.3.4 Organisation Instance
Formally, given a world w 2 W, an organisation O is defined by a set of agents, a set

of objectives (missions or desires), and a set of capabilities or assets. The capabilities
of an organisation are defined by the capabilities of the agents that are part of the
organisation at a given moment. Moreover, the capabilities of agents and groups can
vary. Each world describes the capabilities of agents, so that by defining with agents
belong to an organisation, one is able to determine the potential capabilities of that
organisation. That is, given a organisation O such that AO is the set of agents in O,
organisational capability CO is defined as: CO ¼ CAO .
We are now able to give a formal definition of organisation as introduced in
LAO. Given a set of worlds W representing states of affairs in the environment, an
organisation O in a world w2W is defined as:

Ow ¼ AwO ; wO ; DwO ; SwO

where AwO ¼ fa1 ; . . . ; an g is the set of agents, AwO ; wO is a partial order relation on A
reflecting the structure of the organisation, DwO Φ is a set of objectives (states of
affairs to achieve), and SwO Φ is the set of current assets and characteristics of O. This
abstract definition of organisations enables to analyze some theoretical properties and
incorporates some realistic notions such as bounded capabilities of agents, coordination
and control. As an example, consider the organisation O ¼ ðA; O ; D; S0 Þ, where:
A is the set of agents {a,b,c,d}

O are dependencies between the agents
ðaO b; aO cO dÞÞ
D ¼ (sand-door ∧ paint-door) ∨ tile-floor ∨ paper-wall, is the objective of O
S0 ¼ fRa D; Cb ðsand-doorÞ; Cd ðpaint-doorÞ; Cfb;cg ðtile-floorÞg, is the initial state
Note that the initial organisational state S0 indicates the capabilities of the agents
in A and that agent a is responsible for the achievement of the organisational goal.
This example also shows that organisations are dependent on the capabilities of
their agents to achieve their objectives. In this case, without agent b, the
1
In fact, agents are responsible for the objectives associated with the roles it enacts
550 V. Dignum
Table 20.2 Example of s0 Ra ((sand-door ∧ paint-door) ∨ tile-floor ∨ paper-wall)

organisational activity
s1 Ea Rb (sand-door) ^ Ea Rc (paint-door)
s2 Rb (sand-door) ^ Rc (paint-door)
... ...
si Rb (sand-door) ^ Ec Rd (paint-door)
siþ1 Rb (sand-door) ^ Rd (paint-door)
... ...
sn1 Eb (sand-door) ∧ Ed (paint-door)
sn Sand-door ∧ paint-door
organisation can never achieve its goals. A possible strategy for this organisation to
realize its objectives is depicted in Table 20.2.
20.4 Performance and Change
The model presented in the previous section describes snapshots of organisations,

that is, the state of the organisation at a given moment (world). In order to keep
effective, organisations must strive to maintain a good fit in a changing environ-
ment. Changes in the environment lead to alterations on the effectiveness of the
organisation and therefore to the need to reorganize, or in the least, to the need to
consider the consequences of that change to the organisation’s effectiveness and
efficiency. On the other hand, organisations are active entities, capable not only of
adapting to the environment but also of changing that environment. This means that
organisations are in state of, to a certain degree, altering environment conditions to
meet their aims and requirements, which leads to the question of how and why
reorganization decisions should be reached. The flexibility of an organisation is
defined as the combination of the changeability of an organisational characteristic
(structure, technology, culture) and the capabilities of management to change that
characteristic (Gazendam and Simons 1998).
In Organisational Theory, the concept of adaptation can mean different things,
ranging from strategic choice to environmental determinism. Strategic choice refers
to the planned pursuit of ends based on a rational assessment of available means and
conditions, resulting on a explicit decision to change the organisation. Determin-
istic views on adaptation, on the other hand, explain organisational change as
(involuntary) response to environmental requirements. In this chapter, we treat
adaptation as a design issue that requires an (explicit) action resulting in the
modification of some organisational characteristics. Such decisions can be of two
kinds: proactive, preparing the organisation in advance for an expected future, and
reactive, making adjustments after the environment has changed (Dignum et al.
2004).
In terms of the formal model of organisations introduced in the previous
sections, changes are represented as (temporal) transitions between two different
worlds. Given a world w 2 W, many different events may happen that change some
proposition in that world resulting in a different world (the relation T between two
worlds represents this). Because not all parameters in w are controllable by the
organisation, the current state of the organisation is not necessarily, and in fact in
most cases not, completely controlled by the organisation itself. That is, changes are
not always an effect of (planned) organisational activity. We distinguish between
exogenous and endogenous change. In the exogenous situation, changes occur
outside the control, and independently of the actions, of the agents in the
organisation, whereas that in the endogenous case, changes that are result of activity
explicitely taken by agents in the organisation.
In summary, reorganization consists basically of two activities. Firstly, the
formal representation and evaluation of current organisational state and its ‘dis-
tance’ to desired state, and, secondly, the formalization of reorganization strategies,
that is, the purposeful change of organisational constituents (structure, agent popu-
lation, objectives) in order to make a path to desired state possible and efficient. In
the following section, we describe how LAO can be used to support reorganization
decisions. In Sect. 20.4.2, we will extend LAO to accommodate and reason about
changing organisations.
20.4.1 Analysis of Reorganisation in LAO
In general, at any given moment, an organisation cannot fully control its state nor
the way it will evolve. Formally, given organisation O ¼ ðA; O ; D; S0 Þ , this is
represented by: SO 6¼ CO , or 9φ 2 SO : :CO φ. That is, there are elements of the
desired state that are not achievable by the capabilities currently existing in the
organisation. Some parameters may be controlled by external parties, but others
may be completely uncontrollable (disasters, natural forces, etc.). External agents
which can control parameters that are part of the organisation state are said to have
an influence on the organisation. Organisations are constantly trying to identify the
optimal design for their environment, and will choose a change strategy that they
believe will improve their current situation. Figure 20.1 shows the relation between
organisation desires and states, and controllable parameters both a generic situation,
and the ideal situation.
Ideally, in a successful organisation the set of desires DO will be a subset of the
organisational state SO , which is under the control of the agent in the organisation.
Reorganization activities aim therefore at aligning these sets, either by attempting
to change the current state, or by altering the set of desires.
Intuitively, performance is a value function on the environment (current world),
agents and organisational capability, and on the task (desired state of affairs). The
higher the performance the better organisational objectives are achieved in a given
state. That is, reorganization decisions start with the evaluation of current
organisational performance. In order to specify organisational performance in
LAO, we define a function θ with signature W PðA Þ Φ7 !R, such that
552 V. Dignum
a b
SO DO
CA
CAA
DO
COO
COO SSOO
Current situation Ideal situation
Legend: SO : current state of organization O

DO : desired state of organization O
CA : scope of control of all agents
CO : scope of control of agents in O
Fig. 20.1 Organisational state and organisational control
θðw; G ; φÞ returns the value of the performance in world w of structured group

G for φ, indicating how well G can realize φ. The function θ can be seen as the
cost associated with a transition in the world W. We assume that for each agent and
each world, the performance for each atomic proposition p is fixed. That is, 8w,a,p,
∃c 2 R : θ (w,a,p) ¼ c. The function θ has the following properties:
1. θðw; G ; pÞ + θðw; G ; qÞ θðw; G ; p ^ qÞ, for atomic propositions p, q
2. :CG φ ! ðθðw; G ; φÞ ¼ 1Þ
3. θðw; G ; φÞ θðw; G ; φ ^ ψÞ
4. ðφ ! ψ Þ ! ðθðw; G ; ψÞ θðw; G ; φÞÞ
Informally, the first property above, represents the fact that an agent can get
‘tired’. That is, the cost of performing a combination of activities can be higher than
the sum of the costs associated with each activity, for that agent (Cholvy et al.
2005). The second property indicates that performance is impossible if the capabil-
ity is non-existent. Properties 3 and 4 are associated with the logical relations
between propositions. Some authors have used finite state machines to describe
organisation transitions (Matson and DeLoach 2005). An important difference
between our work and such approaches is the temporal dimension of our model,
that is, an organisation can never move to a previous state even if conceptually
equivalent, all states are different in time.
The function θ provides a way to establish the cost of achievement of a certain
state of affairs, given the current state and group of agents. Based on this function,
different strategies can be adopted that determine the course of action to follow in
order to choose a organisation with better performance with respect to a given goal.
In the next subsection, we describe the actions that can be taken in order to
effectively change the organisation. In the case of endogenous reorganization
processes, it must be possible for agents within the organisation to monitor the
organisation performance, as part of the process of deciding on reorganization.
We therefore specify the capability and responsibility of agents for checking
performance as follows: Ca check-perform ðθðw; G ; φÞ; υÞ, and Ra check-perform
ðθðw; G ; φÞ; υÞ . We abbreviate these expressions to Ca check-perform and Ra
check-perform to indicate the capability, resp. responsibility to check the overall
performance. Similar expressions are also defined for groups of agents.
20.4.2 Acting for Change
In this section, we are concerned with planned reorganization. That is, we want to
specify the activities that are consciously taken by an organisation in order to
attempt to modify its characteristics. Note that, under planned reorganization, we
consider both endogenous reorganization and exogenous reorganization, in which
reorganization is achieved by activity outside of the system, for example by the
designer (and thus off-line).
In human organisations, reorganization strategies take different forms, such as
hiring new personnel, downsizing, training or reassigning tasks or personnel
(Carley and Svoboda 1996). In principle, organisations can also decide to modify
their mission or objectives. However, studies suggest that such second-order
change, that is, a shift from one strategic orientation to another, is atypical even
in times of highly dynamic environmental behavior (Fox-Wolfgramm et al. 1998).
Because organisations aim at making certain states of affairs to be the case, and
only agents can bring affairs to be, it is important for the organisation to make sure
it ‘employs’ and organizes an adequate set of agents ðAO ; O Þ such that the
combined action of those agents has the potentiality to bring about the desired
state of affairs DO. The dependency relation O between agents must allow for the
desired states to be achieved, that is, dependencies must be sufficient for responsi-
bilities to be passed to the appropriate agents, that is, the agents that have the
necessary capabilities. If that is not the case, the organisation should take the steps
needed to decide and implement reorganization, such that the resulting organisation
O0 is indeed able to realize its objectives DO . In practice, reorganization activities
can be classified in three groups:
• Staffing: Changes on the set of agents: adding new agents, or deleting agents
from the set. This corresponds to HRM activities in human organisations (hiring,
firing and training).
• Structuring: Changes on the ordering structure of the organisation.
Corresponding to infra-structural changes in human organisations: e.g. changes
in composition of departments or positions.
554 V. Dignum
• Strategy: Changes on the objectives of the organisation: adding or deleting

desired states. Corresponding to strategic (or second-order) changes in human
organisations: modifications on the mission, vision, or charter of the
organisation.
Given an organisation O ¼ ðAO ; O ; DO ; SO Þ and an agent a, such that Ca φ,
a possible initial definition for the reorganization operations over O is:
• Staff þ ðO; aÞ : Adding agent a to organisation O results in organisation O0 ¼
ðAO [ fag; O ; DO ; SO [ Ca φÞ.
• Staff ðO; aÞ: Ifa 2 AO, removing agent a from organisation O results in organisation
O0 ¼ ðAO =fag; O =depa ; DO ; SO =Ca φÞ where depa ¼ fðx¼O yÞ: x ¼ a _ y ¼ ag
• Struct þ ðO; ða ¼ bÞÞ: If a; b 2 AO, adding dependency ðaO bÞ to organisation O
results in organisation O0 ¼ ðAO ; O [ fða bÞg; DO ; SO Þ
• Struct ðO; ða bÞÞ: If ðaO bÞ, removing dependency ðaO bÞ from organisa-
tion O results in organisation O0 ¼ ðAO ; O =fða bÞg; DO ; SO Þ
• Strategþ ðO; dÞ: If ðd ^ DO Þ is consistent, adding objective d to organisation O
results in organisation O0 ¼ ðAO ; O ; DO [ fdg; SO Þ
• Strateg ðO; dÞ: If d 2 DO, removing objective d from organisation O, results in
organisation O0 ¼ ðAO ; O ; DO =d; SO Þ
The above definition gives a fairly simple (naive) description of model updates.
This is especially the case for the strategic reorganization operations. The solution
chosen above is to specify that strategic decisions cannot be realized if they yield in
a contradiction. In reality, a more elaborate formal treatment of reorganization,
must consider belief revision problems that result from the addition and/or deletion
of model components. This is an issue for future research and will not be further
discussed in this chapter.
The operations described above enable the reorganization of an agent
organisation. The issue of deciding about reorganization still remains. That is,
how do organisations reach the decision to reorganize? What should then be
reorganized? When should one reorganize? An informal foundation for reasoning
about dynamic reorganization is given in (Dignum et al. 2004). In Sect. 20.4.3 we
will further discuss this issue. However, nothing is yet said about how to reach the
decision to reorganize, and how to monitor the state of an organisation wrt its
desires and environment conditions.
Reorganization operations are just operations, that is, effect the value of some
variables. That is, reorganization can also be either endogenous or exogenous. In
the exogenous case, reorganization lays outside the control of any agent in the
world, and is often realized by action of the system designer. In the endogenous
case, agents are able to achieve the states resulting from reorganization operations
(that is, can control reorganization results). In this case, we can specify an agent
a such that Ca σ where σ is one of the reorganization results specified above.
For example, the fact that agent a is able to hire other agents is represented by
Ca staff þ ðO; bÞ. We can thus cover both the concept of engineered reorganization
(exogenous) and dynamic reorganization (endogenous). Furthermore, as discussed
in (Dignum et al. 2004; Barber and Martin 2001), in the endogenous case, reorgani-
zation can either be collaborative, that is, there is a group of agents G A, such
that CG ρ, or role-based, that is, a single agent controls the reorganization result, Ca ρ,
where ρ is one of the reorganization operations defined above.
20.4.3 Deciding About Change
An explicit, planned reorganization strategy must take into account the current
performance and determine which characteristics of the organisation should be
modified in order to achieve a better performance. The general idea behind reorga-
nization strategies, is that one should be able to evaluate the utility of the current
state of affairs (that is, what happens if nothing changes), and the utility of future
states of affairs that can be obtained by performing reorganization actions. The
choice is then to choose the future with the highest utility.
By applying the θ function defined in Sect. 20.4.1, one is able to determine the
current and future performance of the organisation, and also calculate the cost of
reorganization. This can be used to decide about when to reorganize. In order to
enable endogenous reorganization, agents in the organisation must be able to
evaluate the performance of the organisation, that is, have the capability Ca check-
perform ðθðw; G ; φÞ; υÞ , and responsibility Ra check-perform ðθðw; G ; φÞ; υÞ .
Moreover, expressions describing different reorganization decisions can be
expressed in the same way. For instance, Ra minimize-perform indicates that agent
a is responsible to minimize organisation performance costs.
The general idea is that one evaluates the utility of the current state, and the
utility of all following states obtained by performing controlling or structuring
actions. The choice is then to move into the state with the highest utility.
A possible strategy to decide whether to realize reorganization operation ρ is if
θðw; O; ρÞ þ θðw0 ; O0 ; DO Þ θðw; O; DO Þ , where O0 is the organisation in
w0 resulting from the realization of ρ on organisation O in w. Informally, this
strategy says that if the cost of reorganization plus the cost of achieving DO by
the reorganized organisation is less than the cost of achieving DO without
reorganizing, then reorganization should be chosen. Another issue to take into
account on the reorganization decision is the frequency of reorganization. In
particular when considering human organisations, it cannot be expected that the
organisation can be reorganized every time a change occurs. Such information can
be included in the calculation of the θ action by adding a ‘penalty’ for frequent
reorganization. In (Dignum et al. 2006) we describe some experiments concerning
frequency of change.
556 V. Dignum
20.5 Evaluation Through Simulation
We have applied the model presented in this chapter to formally describe several
existing agent systems that simulate organisations and organisational adaptation.
Those systems were taken from the literature and include such different domains as
architectures for Command and Control (Kleinman et al. 2005), RoboSoccer
(Hübner et al. 2004), market regulation (Nissen 2001) and sensor networks (Matson
and DeLoach 2004). The original work on those cases use different frameworks and
were developed independently by other research groups. For simplicity reasons, we
will use here the example of the builders organisation described in Sect. 20.3.
We have developed a simulation framework, Organisation Ontology Simula-
tor (OOS) that provides an ontology for LAO concepts (Dignum and Tick 2007;
Dignum 2010). Organisation instances are then specified as instances of the
organisation ontology. OOS converts the ontology instances into a custom Java-
based format, which can be used by the Repast2 simulation environment to
visualize and run the simulations. In addition, OOS provides a library of agent
behaviors based on the 3APL agent platform3, which allows to use software agents
with cognitive BDI capabilities within the Repast environment. A simulation is
created as follows. First, a specific organisation is created as instance of the
organisation ontology. Agent instances are associated with standard agent
behaviors existing in the 3APL library of OOS. A more detailed description of
OOS can be found in (Tick 2007).
The case study described in this section aims at illustrating the framework in
order to assess organisational design in a fictive situation by designing two suffi-
ciently different environments, and two organisations that exploit the differences
between environments. The idea is that in the case objectives are incongruent with
organisation, i.e. in the current organisational situation objectives cannot be
achieved, organisations will need to reorganize in order to fulfil their missions.
For instance, organisation O is incongruent with DO if CO 6¼ DO. Using the builders
organisation introduced in Sect. 20.3 we define the following types of environments
and organisations:
• Environments: Provide the tasks to be taken up by the organisation. We define
three types of tasks: paint-door, paper-wall, tile-floor. Frequency of tasks, set in
the simulation at 0.25, indicates how often the organisation will be requested to
perform a new task. Two different types of environments are identified as
follows:
Heterogenous (Het): all task types occur with the same probability.
Homogenous (Hom): mostly one type of task: tasks of type paint-door occur in
80 % of the cases; other tasks occur with 10 % probability each.
2
http://repast.sourceforge.net/
3
http://www.cs.uu.nl/3apl/
Fig. 20.2 Organisation ontology for the builder scenario
• Organisations: In this simulation we assume a hierarchical structure and two

different roles: manager and worker. We populate the organisation with 1
manager agent and 12 worker agents. We further specify two types of
organisations as follows:
Specialists (Spec): each worker can only perform tasks of one type (four workers
for each type). Time to complete a task is calculated randomly as Ts 2 ½50; 100.
Generalists (Gen): all workers can perform all tasks, however take longer to
complete a task: Tg ¼ 2 Ts 25.
The idea is that a manager receives tasks to be executed and distributes those to a
worker with capability to perform that task and is currently free. If no capable agent
is free, the manager keeps tasks in a waiting list. There is a penalty for keeping tasks
in the waiting list, resulting in less profit. Each completed task results in a profit of x
for the organisation. For each waiting task the organisation looses x/10 per simula-
tion tick (representing working days). In total, four different scenarios are obtained,
as follows: Het-Gen, Het-Spec, Hom-Gen, Hom-Spec.
OOS was used to generate simulations of the four scenarios above, each
corresponding to a different instantiation of the organisation ontology depicted in
Fig. 20.2. Each scenario, with a length of 400 time units, was run 10 times, resulting
in 40 different simulations. For each scenario, we calculated the number of
completed tasks, the number of waiting tasks, and the average completion length.
Figure 20.3 shows the results obtained.
As can be expected, the performance of Generalist organisations is not
influenced by the type of environment. Specialist organisations do not perform
558 V. Dignum
Average Completed Tasks Average Waiting Tasks

45 80
40 70
35
60
30
50
25
40
20
30
15
10 20
5 10
0 0
1 11 21 31 41 1 11 21 31 41
Simulation Length (x10) Simulation Length (x10)
Het-Gen Het-Spec Hom-Gen Hom-Spec Het-Gen Het-Spec Hom-Gen Hom-Spec
Fig. 20.3 Organisational performance
Average Waiting Tasks

40
20
0
1 51 101 151 201 251 301 351 401
-20
-40
-60
-80
-100
Simulation length
Profit-Reorg (x10) Profit-HomSpec (x10) Waiting-Reorg Waiting-HomSpec
Fig. 20.4 Reorganization effect
well in Homogeneous environments (workers capable of paint-door are too busy,

other workers mostly idle). In such case, reorganization should be considered. We
provided manager agents with the capability to check organisational performance
and adapt accordingly. In the experiment, manager agents of Specialist
organisations in a Homogeneous environment monitor the number of waiting
tasks and perform a reorganization action when that number reaches a given
threshold. Specifically, when the manager notices that there are ten or more waiting
tasks, it will train one, if any, idle agent to perform the task paint-door. Formally,
this is represented as: Em staff ðO; aÞ for some agent a, followed by Em staff þ ðO; pÞ
where agent p has capabilities Cppaint-door. Cost reorganization, that is, the cost of
training one agent is fixed at x*3, where x is the task profit. Figure 20.4 shows
the difference of performance between organisation Hom-Spec and organisation
Hom-Spec-Reorg endowed with such reorganization capabilities, for task comple-

tion times Ts 2 ½25; 50 and task frequency 25 %.
As it can be seen, reorganization results in fewer waiting tasks and therefore
higher profit due to less penalties. Depending on the values for profit, penalties and
reorganization costs, simulations can be used to determine an effective threshold.
20.6 Conclusions
In this chapter, we presented a theoretic model to describe organisations and their

environments that enables the formal analysis of the fit between organisational
design and environment characteristics. This enables the a priori comparison of
designs and their consequences and therefore supports the decision making process
on the choice of design. Reorganization is needed in order to enable systems to
enforce or adapt to changes in the environment. This issue has been discussed by
many researchers in both Organisational Theory as in Distributed Systems,
resulting mostly in domain-oriented empiric solutions. The lack, in most cases, of
a formal basis makes difficult the development of theories about reorganization,
prevents the comparison or approaches and results, and makes difficult the adapta-
tion of models to other domains or situations. The theoretical model presented in
this chapter provides a formal, provable representation of organisations, with their
environment, objectives and agents in a way that enables to analyze their partial
contributions to the performance of the organisation in a changing environment.
Moreover, the model includes enough realistic concepts that enable it to represent
the more ‘pragmatic’ considerations faced by real organisations. The semantics of
LAO, the model for organisational concepts and the reorganization process
presented in this chapter are based on modal temporal logic. We further introduced
a framework for the rapid development of organisational simulations. This frame-
work provides a structured efficient way to deploy many different organisational
and environment designs, by using ontologies describing organisation structures,
environment characteristics and agent capabilities, and provides semi-automatic
means to generate simulations from the ontology instances. Moreover, the simula-
tion framework incorporates the cognitive characteristics of agents. We are cur-
rently finalizing the development of the framework and in the near future will be
using it for the modelling and analysis of realistic organisations.
The simple scenario presented in Sect. 20.5 demonstrates the use of simulations
to support the process of assessment of organisational performance. By varying the
types and capabilities of the agents and the characteristics of the environment, is
possible to analyze the fit of organisational structures and the potential of different
reorganization decisions to the performance of the organisation. Currently, we are
extending the evaluation of the model and the OOS tool to the analysis of more
realistic organisations. In parallel, we are also engaged on the extension of LAO
and the specification of a full axiomatization for LAO.
560 V. Dignum
Acknowledgements The research presented in this chapter is funded by the Netherlands

Organisation for Scientific Research (NWO), through Veni-grant 639.021.509. The author is
grateful to Christiaan Tick and Daghan Acay for their contribution to the development of the
OOS simulation framework.
Further Reading
For readers interested in organisational design in general, (Burton et al. 2006) give a
step-by-step approach whereas (Ostrom 2005) provides a framework for analysing
diverse economic, political and social institutions. The use of agent-based models
to investigate how internal policies and procedures affect the performance of
organisations is demonstrated by (Carley 2002). Applying organisational design
principle to multi-agent systems, Dignum and Paget (2013) give an overview of the
state of the art, while Horling and Lesser (2004) provide a survey of current
paradigms.
References
Aldrich H, Pfeffer J (1976) Environments of organizations. Annu Rev Sociol 2:79–105

Artikis A, Pitt J (2001) A formal model of open agent societies. In: Proceedings of the fifth
international conference on autonomous agents, Montreal, 28 May–01 June 2001. ACM Press,
Barber KS, Martin CE (2001) Dynamic reorganization of decision-making groups. In: Proceedings
of the fifth international conference on autonomous agents, Montreal, 28 May–01 June 2001.
ACM Press, New York, pp 513–520
Burton R, DeSanctis G, Obel B (2006) Organizational design: a step by step approach. Cambridge
Carley K (2002) Computational organisation science: a new frontier. PNAS 99(3):7257–7262
Carley K, Svoboda D (1996) Modeling organisational adaptation as a simulated annealing process.
Sociol Method Res 25(1):138–168
Castelfranchi C (1995) Guarantees for autonomy in cognitive agent architecture. In: Wooldridge
MJ, Jennings NR (eds) Intelligent agents, ECAI-94 workshop on agent theories, architectures
and languages, Amsterdam, 8–9 Aug 1994. Springer, Berlin, pp 56–70
Cholvy L, Garion C, Saurel C (2005) Ability in a multi-agent context: a model in the situation
calculus. In: Toni F, Torroni P (eds) CLIMA’05 proceedings of the 6th international confer-
ence on computational logic in multi-agent systems. Springer, Berlin, pp 23–36
Daft R, Murphy J, Willmott H (2010) Organization theory and design. South-Western Cengage
Learning, Andover
Dastani M, Dignum V, Dignum F (2003) Role assignment in open agent societies. In: Proceedings
of the second international joint conference on autonomous agents and multiagent systems,
AAMAS 2003, 14–18 July 2003, Melbourne. ACM Press, New York, pp 489–496
Davidsson P (2002) Agent based social simulation: a computer science view. J Artif Soc Soc Simul
Dignum V (2004) A model for organizational interaction: based on agents, founded in logic. PhD
thesis, SIKS dissertation series 2004–1, Utrecht University
Dignum V (2010) Ontology support for agent-based simulation of organisations. Multiagent Grid
Syst 6:191–208
Dignum V, Dignum F (2007) A logic for agent organization. In: Proceedings of the international
workshop on formal aspects of MAS, Durham, 3–7 Sept 2007
Dignum V, Paget J (2013) Multi-agent organizations. In: Weiss G (ed) Multiagent systems,
2nd edn. MIT Press, Cambridge
Dignum V, Tick C (2007). Agent-based analysis of organisations: formalization and simulation.
In: Proceedings of IEEE/WIC/ACM international conference on intelligent agent technology
(IAT’07), Silicon Valley, 2–5 Nov 2007. IEEE, pp 244–247
Dignum V, Dignum F, Sonenberg L (2004) Towards dynamic organisation of agent societies. In:
Vouros G (ed) Workshop on coordination in emergent agent societies, ECAI 2004. pp 70–78
Dignum V, Dignum F, Sonenberg L (2006) Exploring congruence between organisational struc-
ture and task performance: a simulation approach. In: Boissier O et al (eds) Coordination,
organization, institutions and norms in multi-agent systems, ANIREM’05/OOOP’05, Utrecht,
25–26 July 2005, revised selected papers (Lecture notes in computer science, 3913). Springer,
Berlin, pp 213–230
Duncan R (1979) What is the right organisational structure: decision tree analyis provides the
answer. Organ Dyn 7(3):59–80
Fox-Wolfgramm S, Boal K, Hunt J (1998) Organizational adaptation to institutional change: a
comparative study of first-order change in prospector and defender banks. Adm Sci Q 43
(1):87–126
Galbraith J (1977) Organization design. Addison-Wesley, Reading
Gazendam H, Simons JL (1998) An analysis of the concept of equilibrium in organization theory.
In: Proceedings of the computational and mathematical organization theory workshop,
Montreal, 25–26 Apr 1998
Grossi D, Dignum F, Dastani M, Royakkers L (2005) Foundations of organizational structures in
multiagent systems. In: Dignum F et al (eds) Proceedings of the 4th international joint
conference on autonomous agents and multiagent systems (AAMAS 2005), 25–29 July
2005, Utrecht. ACM Press, pp 690–697
Grossi D, Dignum F, Dignum V, Dastani M, Royakkers L (2007) Structural aspects of the
evaluation of agent organisations. In: Noriega P et al (eds) Coordination, organizations,
institutions, and norms in agent systems II, AAMAS 2006 and ECAI 2006 international
workshops, COIN 2006, Revised selected papers (Lecture notes in computer science, 4386).
Harrison R, Lin Z, Carroll G, Carley K (2007) Simulation modeling in organizational and
management research. Acad Manage Rev 32(4):1229–1245
Horling B, Lesser V (2004) A survey of multi-agent organizational paradigms. Knowl Eng Rev 19
(4):281–316
Hübner JF, Sichman JS, Boissier O (2004) Using the moise + for a cooperative framework of
MAS reorganisation. In: Bazzan A, Labidi S (eds) Advances in artificial intelligence – SBIA
2004, Proceedings of the 17th Brazilian symposium on artificial intelligence (Lecture notes in
computer science, 3171). Springer, Berlin, pp 481–517
Kleinman D, Levchuk G, Hutchins S, Kemple W (2005) Scenario design for the empirical testing
of organisational congruence (Technical report ADA440390). Department of informational
science, Naval Postgraduate School, Monterey
Matson E, DeLoach SA (2004) Enabling intra-robotic capabilities adaptation using an
organisation-based multiagent system. In: Proceedings of the 2004 I.E. international confer-
ence on robotics and automation (ICRA 2004)
Matson E, DeLoach SA (2005) Formal transition in agent organizations. In: Thompson C,
Hexmoor H (eds) 2005 International conference on integration of knowledge intensive
multiagent systems (KIMAS’05): modeling, exploration, and engineering, 18–21 Apr 2005,
Waltham. IEEE, pp 235–240
562 V. Dignum
Mintzberg H (1993) Structures in fives: designing effective organizations. Prentice Hall,

Englewood Cliffs
Nissen M (2001) Agent-based supply chain integration. Inf Technol Manage 2(3):289–312
Ostrom E (2005) Understanding institutional diversity. Princeton Universiy Press, Princeton
Pörn I (1974) Some basic concepts of action. In: Stenlund S (ed) Logical theory and semantical
analysis. Reidel, Dordrecht, pp 93–101
Roberts J (2004) The modern firm: organizational design for performance and growth. Oxford
University Press, Oxford
Santos F, Jones A, Carmo J (1997) Action concepts for describing organised interaction. In:
Proceedings of the 30th annual Hawaii international conference on system sciences (HICCS-
30), 7–10 Jan 1997, Maui, vol 5. IEEE Computer Society Press, pp 373–382
Selznick P (1948) Foundations of the theory of organization. Am Sociol Rev 13(1):25–35
So Y, Durfee E (1998) Designing organisations for computational agents. In: Carley K, Prietula M,
Gasser L (eds) Simulating organizations. Cambridge, MA: The MIT Press, pp 47–64
Tick C (2007) Organization structure analysis by simulation. Master’s thesis, Utrecht University
van der Hoek W, Wooldridge M (2005) On the logic of cooperation and propositional control.
Artif Intell 164(1–2):81–119
Weigand H, Dignum V (2004) I am autonomous, you are autonomous. In: Nickles M, Rovatsos M,
Weiss G (eds) Agents and computational autonomy, potential, risks, and solutions, vol 2969,
Lecture notes in computer science. Springer, Berlin, pp 227–236
Zambonelli F, Jennings N, Wooldridge M (2001) Organisational abstractions for the analysis and
design of multi agent systems. In: Ciancarini P, Wooldridge M (eds) Agent-oriented software
engineering, vol 1957, Lecture notes in computer science. Springer, Berlin, pp 235–251
Chapter 21
Distributed Computer Systems
David Hales
Why Read This Chapter? To understand how simulating social complexity

might be used in the process of designing distributed computer systems.
Abstract Ideas derived from social simulation models can directly inform the
design of distributed computer systems. This is particularly the case when systems
are “open”, in the sense of having no centralised control, where traditional design
approaches struggle. In this chapter we indicate the key features of social simula-
tion work that are valuable for distributed systems design. We also discuss the
differences between social and biological models in this respect. We give examples
of socially inspired systems from the currently active area of peer-to-peer systems
and finally we discuss open areas for future research in the field.
21.1 Introduction
Massive and open distributed computer systems provide a major application area
for ideas and techniques developed within social simulation and complex systems
modelling. In the early years of the twenty-first century there has been an explosion
in global networking infrastructure in the form of wired and wireless broadband
connections to the internet encompassing both traditional general purpose computer
systems, mobile devices and specialist appliances and services. The challenge is to
utilise such diverse infrastructure to provide novel services that satisfy user needs
reliably. Traditional methods of software design and testing are not always appli-
cable to this challenge. Why is this? And what can social simulation and complexity
perspectives bring to addressing the challenge? This chapter answers these
questions by providing a general overview of some of the major benefits of
approaching design from socially inspired perspectives in addition to examples of
D. Hales (*)
University of Bologna, Bologna, Italy

564 D. Hales
applications in the area of peer-to-peer (P2P) systems and protocols. Finally we will
speculate on possible future directions in the area.
This chapter is not an exhaustive survey of the area, for example, we have not
discussed the application of social network analysis techniques to web graphs and
social networks that are constructed within, or facilitated by, distributed software
systems (Staab et al. 2005). Both these are active areas. Further, we have not
discussed the active research area based on randomised “Gossiping” approaches,
where information is diffused over networks through randomised copying of
information between adjacent nodes (Wang et al. 2007).
21.2 What’s Wrong with Traditional Design Approaches?
Traditional design approaches to systems and software often assume that systems
are essentially “closed” – meaning they are under the control of some administra-
tive authority that can control access, authenticate users and manage system
resources such as issuing software components and updates. Consider the simplest
situation in which we have a single computer system that is not connected to a
network that is required to solve some task. Here, design follows a traditional
process of analysis of requirements, specification of requirements then design and
iterative refinement, until the system meets the specified requirements. User
requirements are assumed to be discoverable and translatable into specifications
at a level of detail that can inform a design. The designer has, generally, freedom to
dictate how the system should achieve the required tasks via the coordination of
various software components. This coordination generally follows an elaborate
sequencing of events where the output from one component becomes the input to
others. The design task is to get that sequence right.
In “open systems” it is assumed that there are multiple authorities. This means that
components that comprise the system cannot be assumed to be under central control.
An extreme example of this might be an open peer-to-peer (P2P) system in which
each node in the system executes on a user machine under the control of a different
user. Popular file-sharing systems operate in this way, allowing each user to run any
variant of client software they choose. These kinds of systems function because the
client software implements publicly available peer communication protocols,
allowing the nodes to interconnect and provide functionality. However, due to the
open nature of such systems, it is not possible for the designer, a priori, to control the
sequence of processing in each node. Hence the designer needs to consider what
kinds of protocols will produce acceptable system level behaviours under plausible
assumptions of node behaviour. This requires a radically different design perspective
in which systems need to be designed as self-repairing and self-organising systems in
which behaviour emerges bottom-up rather than being centrally controlled.
One term for this approach, based on self-organisation and emergence, is so-
called Self-Star (or Self-*) systems (Babaoglu et al. 2005). The term is a broad
expression that aims to capture all kinds self-organising computer systems that
continue to function acceptably under realistic conditions.
21 Distributed Computer Systems 565
But what kinds of design approach can be employed? Currently there is no

accepted general theory of self-organisation or emergence – rather there are some
interesting models at different levels of abstraction that capture certain phenomena.
Many such models have been produced within the biological sciences to explain
complex self-organising biological phenomena. Biological systems, particularly
co-evolving systems, appear to evidence many of the desirable properties required
by self-* computer systems. Hence, several proposed self-* techniques have drawn
on biological inspiration (Babaoglu et al. 2006).
21.3 Socially Versus Biological Inspiration
It is useful to ask in what way social organisation differs from the biological level.
In this section we briefly consider this question with regard to desirable properties
for information systems. An important aspect of human social systems (HSS) is
their ability (like biological systems) to both preserve structures – with
organisations and institutions persisting over time – and adapt to changing
environments and needs. The evolution of HSS is not based on DNA, but rather
on a complex interplay between behaviour, learning, and individual goals. Here we
present some distinguishing aspects of HSS.
21.3.1 Rapid Change
A feature of HSS is the speed at which re-organisations can occur. Revolutions in

social organisation can take place within the lifetime of a single individual. Hence
although HSS often show stable patterns over long periods, rapid change is possi-
ble. The ability to respond rapidly would appear to be a desirable property in rapidly
changing information system environments; however, for engineering purposes,
one must ensure that such fast changes (unlike revolutions!) can be both predicted
and controlled.
21.3.2 Non-Darwinian Evolution
HSS do not evolve in a Darwinian fashion. Cultural and social evolution is not
mediated by random mutations and selection of some base replicator over vast time
periods, but rather follows a kind of collective learning process. That is, the
information storage media supporting the change by learning – and hence (as
noted above), both the mechanisms for change and their time scale – are very
different from those of Darwinian evolution. Individuals within HSS can learn both
directly from their own parents (vertical transmission), from other members of the
566 D. Hales
older generation (diagonal transmission), or from their peers (horizontal transmis-

sion). Hence, new cultural traits (behaviours, beliefs, skills) can be propagated
quickly through a HSS. This can be contrasted with simple Darwinian transmission
in which, typically, only vertical transmission of genetic information is possible.
Although it is possible to characterise certain processes of cultural evolution based
on the fitness of cultural replicators (Boyd and Richerson 1985) or memes
(Dawkins 1976) it is important to realise such replicators are not physical – like
DNA – but part of a socio-cognitive process – passing through human minds – and
may follow many kinds of selective process (Lumsden and Wilson 1981). The
problems of using the idea of biological evolution in the social sciences are
discussed in more detail in Chap. 18 (Chattoe-Brown and Edmonds 2013) of this
volume.
21.3.3 Stable Under Internal Conflict
HSS exist because individuals need others to achieve their aims and goals. Produc-
tion in all HSS is collective, involving some specialisation of roles. In large modern
and post-modern HSS roles are highly specialised, requiring large and complex co-
ordination and redistribution methods. However, although HSS may sometime
appear well integrated, they also embody inherent conflicts and tensions between
individual and group goals. What may be in the interests of one individual or group
may be in direct opposition to another. Hence, HSS embody and mediate conflict on
many levels.
This aspect is highly relevant to distributed and open information systems. A
major shift from the closed monolithic design approach is the need to deal with and
tolerate inevitable conflicts between sub-components of a system. For example,
different users may have different goals that directly conflict. Some components
may want to destroy the entire system. In open systems this behaviour cannot be
eliminated and hence needs in some way to be tolerated.
21.3.4 Only Partial Views and Controversy
Although HSS are composed of goal directed intelligent agents, there is little
evidence that individuals or groups within them have a full view or understanding
of the HSS. Each individual tends to have a partial view often resulting from
specialisation within, and complexity of, the system. Such partial views, often
dictated by immediate material goals, may have a normative (how things “should”
be) character rather than a more scientific descriptive one (how things “are”).
Consequently, the ideas that circulate within HSS concerning the HSS itself tend
to take on an “ideological” form. Given this, social theories are rarely as consensual
as those from biological sciences. Thus, social theories include a high degree of
controversy, and they lack the generally accepted foundational structure found in
our understanding of biology. However, from an information systems perspective,
such controversy is not problematic: we do not care if a given social theory is true
for HSS or not; we only care if the ideas and mechanisms in the theory can be
usefully applied in information systems. This last point, of course, also holds for
controversial theories from biology as well (e.g. Lamarckian evolution).
21.3.5 Trust and Socially Beneficial Norms
In trying to understand the stability of socially functional behaviour, much work

within the social sciences has focused on the formation and fixation of “norms” of
behaviour. Many researchers working with Multi-Agent Systems (MAS) have
attempted to create artificial versions of norms to regulate MAS behaviours –
although much of these have not been based on theories from HSS (although see
Conte and Paolucci 2002). Certainly the establishment and stability of beneficial
norms (such as not cheating one’s neighbour) is a desirable property visible in all
stable HSS (Hales 2002). This point (the existence and power of norms) is of course
closely related to the previous point, which notes that norms can influence under-
standing and perception.
It is widely agreed that, in HSS, many observed behaviours do not follow the
same pattern as would be expected from simple Darwinian evolution or individual
“rational” behaviour – in the sense of maximising the chance of achieving individ-
ual goals. Behaviour is often more socially beneficial and co-operative or altruistic,
generally directed towards the good of the group or organisation within which the
individual is embedded. (We note the widespread appearance of altruistic behavior
among many species of social mammals – such that, once again, we speak here of a
difference in degree between HSS and other social animals.) Many theories and
mechanisms have been proposed by social scientists for this kind of behaviour
(Axelrod 1984), with many of these formalised as computer algorithms; further-
more, several of these have already been translated for use in information systems
(Cohen 2003; Hales and Edmonds 2005).
21.3.6 Generalised Exchange and Economics
Almost all HSS evidence some kind of generalised exchange mechanisms (GEM) –
i.e. some kind of money. The emergence of GEM allows for co-ordination through
trade and markets. That, is, collective co-ordination can occur where individual
entities (individuals or firms) behave to achieve their own goals. It is an open (and
perhaps overly simplified) question whether certain norms are required to support
GEM or, rather, most norms are created via economic behaviour within GEM
(Edmonds and Hales 2005). Certainly, the formation and maintenance of GEM
568 D. Hales
would be an essential feature of any self-organised economic behaviour within

information systems – currently many information systems work by assuming an
existing GEM a priori, i.e. they are parasitic on HSS supplying the trust and norms
required. Such systems require trusted and centralised nodes before they can
operate because they do not emerge such nodes in on-going interaction. However,
given that GEM exist, a huge amount of economic theory, including new evolu-
tionary economics and game theory, can be applied to information systems.
21.4 What Can Social Simulation Offer the Designer?
Social simulation work has the potential to offer a number of insights that can be
applied to aid design of distributed computer systems. Social simulators have had
no choice but to start from the assumption of open systems composed of autono-
mous agents – since most social systems embody these aspects. In addition, much
social simulation work is concerned with key aspects of self-* systems such as:
• Emergence and self-organisation: understanding the micro-to-macro and the
macro-to-micro link. Phenomena of interest often result from bottom-up pro-
cesses that create emergent structures that then constrain or guide (top-down) the
dynamics of the system.
• Cooperation and trust: getting disparate components to “hang together” even
with bad guys around. In order for socially integrated cooperation to emerge it is
generally necessary to employ distributed mechanisms to control selfish and
free-riding behaviour. One mechanism for this is to use markets (see Chap. 23 in
this volume, Rouchier 2013) but there are other methods.
• Evolving robust network structure: Constructing and maintaining functional
topologies robustly. Distributed systems often form dynamic networks in
which the maintenance of certain topological structures improves system level
performance.
• Constructing adaptive/evolutionary heuristics rather than rational action
models. Models of both software and user behaviour in distributed systems are
based on implicit or explicit models. Traditional approaches in game theory and
economics have assumed rational action but these are rarely applicable in
distributed systems.
These key aspects have import into two broad areas of system design. Firstly,
simulation models that produce desirable properties can be adapted into distributed
system protocols that attempt to reproduce those properties. Secondly, models of
agent behaviour, other than rational action approaches, can be borrowed as models
of user behaviour in order to test existing and new protocols.
Currently, however, it is an open question as to how results obtained from social
simulation models can be productively applied to the design of distributed infor-
mation systems. There is currently no general method whereby desirable results
from a social simulation model can be imported into a distributed system. It is
certainly not currently the case that techniques and models can be simply “slotted
into” distributed systems. Extensive experimentation and modification is required.
Hence, in this chapter we give specific examples from P2P systems where such an
approach has been applied.
21.5 What About Agent-Orientated Design Approaches?
Multi-Agent System (MAS) design approaches have previously been proposed

(Wooldridge and Jennings 1995), which attempt to address some of the design
issues raised by open systems. Those approaches start with a “blank sheet” design
approach rather than looking for biological or social inspiration. The focus there-
fore has tended to be on logical foundations, proof, agent languages and communi-
cation protocols. For example, the BDI agent framework starts from the assumption
that agents within a system follow a particular logical architecture based on “folk
psychological” cognitive objects – such as beliefs or intentions (Rao and Georgeff
1991). However, such approaches have difficulty scaling to large societies with
complex interactions particularly where the effects of emergence and self-
organisation are important. A more recent approach within MAS work has been
to look towards self-organising approaches using simulation to capture processes of
emergence (Brueckner et al. 2006). In this work heavy use has been made of
biological and socially inspired approaches.
21.6 Examples of Socially Inspired P2P Systems
Here we give very brief outlines of some P2P protocols that have been directly
inspired by social simulation models. While P2P systems have not been the only
distributed systems that benefited from social inspiration we have focused on this
particular technology because it is currently, at the time of writing, a very active
research area and increasingly widely deployed on the internet.
21.6.1 Reciprocity Based BitTorrent P2P System
BitTorrent (Cohen 2003) is an open P2P file sharing system that draws directly from
the social simulation work of Robert Axelod (1984) on the evolution of coopera-
tion. The protocol is based on a form of the famous Tit-For-Tat (TFT) strategy
popularised by Axelrods computer simulation tournaments. Strategies were com-
pared by having agents play the canonical Prisoner’s Dilemma (PD) game.
The PD game captures a social dilemma in the form of a minimal game in which
two players each select a move from two alternatives (either to cooperate or defect)
570 D. Hales
then each player receives a score (or pay-off). If both players cooperate then both
get a reward payoff (R). If both defect they are punished, both obtaining payoff P. If
one player selects defect and the other selects cooperate then the defector gets T (the
“temptation”), the other receives S (the “sucker”). When these pay-offs, which are
numbers representing some kind of desirable utility (for example, money), obey the
following constraints: T > R > P > S and 2R > T + S then we say the game
represents a Prisoner’s Dilemma. When both players cooperate this maximizes
the collective good but when one player defects and another cooperates this
represents a form of free riding with the defector gaining a higher score (T) at the
expense of the co-operator (S).
Axelrod asked researchers to submit computer programs to a “tournament”
where they repeatedly played the PD against each other accumulating payoffs.
The result of the tournaments was that a simple strategy, TFT, did remarkably well
against the majority of other submitted programs – although other strategies can
also survive within the complex ecology that occurs when there is a population of
competing strategies.
TFT operates in environments where the PD is played repeatedly with the same
partners for a number of rounds. The basic strategy is simple: a player starts by
cooperating then in subsequent rounds copies the move made in the previous round
by its opponent. This means defectors are punished in the future: the strategy relies
on future reciprocity. To put it another way, the “shadow” of future interactions
motivates cooperative behaviour in the present. In many situations this simple
strategy can outperform pure defection.
In the context of BitTorrent the basic mechanism is simple: files are split into
small chunks (about 1 MB each) and downloaded by peers, initially, from a single
hosting source. Peers then effectively “trade” chunks with each other using a TFT-
like strategy – i.e. if two peers offer each other a required chunk then this equates to
mutual cooperation. However if either does not reciprocate then this is analogous to
a defect and the suckered peer will retaliate in future interactions.
The process is actually a little more subtle because each peer is constantly looking
at the upload rate/download rate from each connected peer in time – so it does not
work just by file chunk but by time unit within each file chunk. While a file is being
downloaded between peers, each peer maintains a rolling average of the download
rate from each of the peers it is connected to. It then tries to match its uploading rate
accordingly. If a peer determines that another is not downloading fast enough then it
may “choke” (stop uploading) to that other. Figure 21.1 shows a schematic diagram
of the way the Bittorrent protocol structures population interactions.
Additionally, peers periodically try connecting to new peers randomly by
uploading to them – testing for better rates. This means that if a peer does not
upload data to other peers (a kind of defecting strategy) then it is punished by other
peers in the future (by not sharing file chunks) – hence a TFT-like strategy based on
punishment in future interactions is used.
Axelrod used the TFT result to justify sociological hypotheses such as under-
standing how fraternisation broke out between enemies across the trenches of
WW1. Cohen has applied a modified form of TFT to produce a file sharing system
Fig. 21.1 A schematic of a portion of a BitTorrent system. The trackers support swarms of peers
each downloading the same file from each other. Thick lines indicate file data. They are constantly
in flux due to the application of the TFT-like “choking” protocol. Trackers store references to each
peer in each supported swarms. It is not unknown for trackers to support thousands of swarms and
for swarms to contain hundreds of peers. Arrows show how peers might, through user intervention,
move between swarms. Generally at least one peer in each swarm would be a “seeder” that holds a
full copy of the file being shared
resistant to free riding. However, TFT has certain limitations, it requires future
interactions with the same individuals and each has to keep records of the last move
made by each opponent. Without fixed identities it is possible for hacked clients to
cheat BitTorrent. Although it appears that widespread cheating has not actually
spread in the population of clients. It is an open question as to why this might be
(but see (Hales and Patarin 2006) for a hypothesis).
21.6.1.1 Group Selection Based P2P Systems
Recent work, drawing on agent-based simulations of cooperative group formation

based on “tags” (social labels or cues) and dynamic social networks suggests a
572 D. Hales
mechanism that does not require reciprocal arrangements but can produce coopera-
tion and specialisation between nodes in a P2P (Riolo et al. 2001; Hales and
Edmonds 2005). It is based on the idea of cultural group selection and the well-
known social psychological phenomena that people tend to favour those believed to
be similar to themselves – even when this is based on seemingly arbitrary criteria
(e.g. supporting the same football team). Despite the rather complex lineage, like
TFT the mechanism is refreshingly simple. Individuals interact in cliques (subsets
of the population). Periodically, if they find another individual who is getting higher
utility than themselves they copy them – changing to their clique and adopting their
strategy. Also, periodically, individuals form new cliques by joining with a ran-
domly selected other.
Defectors can do well initially, suckering the co-operators in their clique – but
ultimately all the co-operators leave the clique for pastures new – leaving the
defectors all alone with nobody to free-ride on. Those copying a defector (who
does well initially) will also copy their strategy, further reducing the free-riding
potential in the clique. So a clique containing any free riders quickly dissolves but
those containing only co-operators grow.
Given an open system of autonomous agents all cliques will eventually be
invaded by a free rider who will exploit and dissolve the clique. However, so
long as other new cooperative cliques are being created cooperation will persist
in the overall population. In the context of social labels or “tags” cliques are defined
as those individuals sharing particular labels (e.g. supporting the same football
team). In the context of P2P systems the clique is defined as all the other peers each
peer is connected to (its neighbourhood) and movement between cliques follows a
process of network “re-wiring”.
Through agent-based simulation, the formation and maintenance of high levels
of cooperation in the single round PD and in a P2P file-sharing scenario has been
demonstrated (Hales and Edmonds 2005). The mechanism appears to be highly
scalable with zero scaling cost – i.e. it does not take longer to establish cooperation
in bigger populations. Figure 21.2 shows the evolution of cooperative clusters
within a simulated network of peer nodes. A similar approach was presented by
Hales and Arteconi (2006) that produced small-world connected cooperative
networks rather than disconnected components.
In addition to maintaining cooperation between nodes in P2P, the same group
selection approach has been applied to other areas such as the coordination of
robots in a simulated warehouse scenario and to support specialisation between
nodes in a P2P job sharing system (Hales and Edmonds 2003; Hales 2006).
21.6.1.2 Segregation Based P2P Systems
The model of segregation proposed by Thomas Schelling is well known within

social simulation (Schelling 1969, 1971). The model demonstrates how a macro-
structure of segregated clusters or regions robustly emerges from simple local
behaviour rules. Schelling’s original model consists of agents on a two dimensional
Fig. 21.2 Evolution of the connection network between nodes playing the Prisoner’s Dilemma.
From an initially random topology composed of all nodes playing the defect strategy (dark shaded
nodes), components quickly evolve, still containing all defect nodes (a). Then a large cooperative
component emerges in which all nodes cooperate (b). Subsequently the large component begins to
break apart as defect nodes invade the large cooperative component and make it less desirable for
cooperative nodes (c). Finally an ecology of cooperative components dynamically persists as new
components form and old components die (d). Note: the cooperative status of a node is indicated
by a light shade
grid. Each grid location can hold a single agent or may be empty. Each agent
maintains a periodically updated satisfaction function. Agents take one of two
colours that are fixed. An agent is said to be satisfied if at least some proportion
of adjacent agents on the grid have the same colour, otherwise the agent is said to be
not satisfied. Unsatisfied agents move randomly in the grid to a free location. The
main finding of the model is that even if the satisfaction proportion is very low, this
574 D. Hales
Fig. 21.3 Example run of the Schelling segregation model. From an initially random distribution
of agents (a) clusters emerge over successive iterations (b) and (c). In this example the satisfaction
proportion is 0.5, meaning that agents are unsatisfied if less than 50 % of neighbours in the grid are
a different colour (Taken from Edmonds and Hales 2005)
still leads to high levels of segregation by colour – i.e. large clusters of agents
emerge with the same colour. Figure 21.3 shows an example run of the model in
which clusters of similar colours emerge over time.
The results of the segregation model are robust even when nodes randomly leave
and enter the system – the clusters are maintained. Also agents in the segregation
model only require local information in order to decide on their actions. These
properties are highly desirable for producing distributed information systems and
therefore it is not surprising that designs based on the model have been proposed.
Sing and Haahr (2006) propose a general framework for applying a modified
form of Schelling’s model to topology adoption in a P2P network. They show how a
simple protocol can be derived that maintains a “hub-based” backbone topology
within unstructured networks. Hubs are nodes in a network that maintain many
links to other nodes. By maintaining certain proportions of these within networks it
is possible to improve system performance for certain tasks. For many tasks linking
the hubs to form a backbone within the network can further increase performance.
For example, the Gnutella1 file-sharing network maintains high-bandwidth hubs
called “super-peers” that speed file queries and data transfer between nodes.
Figure 21.4 shows an example of a small network maintaining a hub backbone.
In the P2P model nodes represent agents and neighbours are represented by
explicit lists of neighbour links (a so-called view in P2P terminology). Nodes adapt
their links based on a satisfaction function. Hub nodes (in the minority) are only
satisfied if they have at least some number of other hubs in their view. Normal
nodes are only satisfied if they have at least one hub node in their view. Hence
different node types use different satisfaction functions and exist in a network rather
than lattice. It is demonstrated via simulation that the network maintains a stable
and connected topology supporting a hub backbone under a number of conditions,
including dynamic networks in which nodes enter and leave the network over time.
1
http://www.gnutella.com
Fig. 21.4 An example of a hub based peer-to-peer topology. Hubs are linked and perform
specialist functions that improve system performance (Taken from Singh and Haahr 2006)
The approach presented by Sing and Haahr is given as a general approach

(a template design pattern) that may be specialised to other P2P application areas
rather than just self-organising hub topologies. For example, they apply the same
pattern to decrease bandwidth bottlenecks and increase system performance of a P2P
by clustering similar nodes based on bandwidth capacity (Singh and Haahr 2004).
21.7 Possible Future Research
In the following sections we give a brief outline of some promising possible areas
related to socially inspired distributed systems research.
21.7.1 Design Patterns
Social simulators and distributed systems researchers currently constitute very

different communities with different backgrounds and goals. A major problem for
moving knowledge between these disciplines is the different language, assumptions
and outlets used by them. One promising approach for communicating techniques
from social simulation to distributed systems designers is to develop so-called
“design patterns” which provide general templates of application for given
techniques. This approach has been influential within object-oriented programming
and recently biologically inspired approaches have been cast as design patterns
(Gamma et al. 1995; Babaoglu et al. 2006). Design patterns are not formal and need
576 D. Hales
not be tied to a specific computer language but rather provide a consistent frame-
work and nomenclature in which to describe techniques that solve recurrent
problems the designer may encounter. At the time of writing few, if any, detailed
attempts have been made to present techniques from social simulation within a
consistent framework of design patterns.
21.7.1.1 The Human in the Loop: Techno-social Systems
Most distributed and open systems function via human user behaviour being
embedded within them. In order to understand and design such systems some
model of user behaviour is required. This is particularly important when certain
kinds of user intervention are required for the system to operate effectively. For
example, for current file sharing systems (e.g. BitTorrent) to operate users are
required to perform certain kinds of altruistic actions such as initially uploading
new files and maintaining sharing of files after they have been downloaded (so
called “seeding”). Web2.0 systems often require users to create, upload and main-
tain content (e.g. Wikipedia). It seems that classical notions of rational action are
not appropriate models of user behaviour in these contexts. Hence, explicitly or
implicitly such distributed systems require models of user behaviour which capture,
at some level, realistic behaviour. Such systems can be viewed as techno-social
systems – social systems that are highly technologically mediated.
One promising method for understanding and modelling such systems is to make
use of the participatory modelling approach discussed in Chap. 10 (Barreteau et al.
2013). In such a system user behaviour is monitored within simulations of the
technical infrastructure that mediates their interactions. Such an approach can
generate empirically informed and experimentally derived behaviour models
derived from situated social interactions. This is currently, at the time of writing,
an underdeveloped research area.
Interestingly, from the perspective of distributed systems, if it is possible to
model user behaviour at a sufficient level of detail based on experimental result then
certain aspects of that behaviour could be incorporated into the technological
infrastructure itself as protocols.
21.7.1.2 Power, Leadership and Hierarchy
A major area of interest to social scientists is the concept of power – what kinds of
process can lead to some individuals and groups becoming more powerful than
others? Most explanations are tightly related to theories of inequality and economic
relationships; hence this is a vast and complex area.
Here we give just a brief very speculative sketch of recent computational work,
motivated by sociological questions, that could have significant import into under-
standing and engineering certain kinds of properties (e.g. in peer-to-peer systems),
in which differential power relationships emerge and may, perhaps, be utilised in a
functional way. See Chap. 25 in this volume (Geller and Moss 2013) for a detailed
overview of modelling power and authority in social simulation.
Fig. 21.5 Forms of

‘hierarchy’, ‘leadership’ and
unequal wealth distribution
have been observed to emerge
in simulated interaction
networks (From Zimmerman
et al. 2001). Nodes play PD-
like games with neighbours
and break connections based
on a simple satisfaction rule.
Hierarchies are produced in
which some nodes are more
connected and hence can
affect the network
dramatically by their
individual actions – a form of
‘topological power’
Interactions in human society are increasingly seen as being situated within

formal and informal networks (Kirman and Vriend 2001). These interactions are
often modelled using the abstraction of a game capturing interaction possibilities
between linked agents (Zimmerman et al. 2001). When agents have the ability to
change their networks based on past experience and some goals or predisposition,
then, over time, networks evolve and change.
Interestingly, even if agents start with more-or-less equal endowments and
freedom to act, and follow the same rules, vastly unequal outcomes can be pro-
duced. This can lead to a situation in which some nodes become objectively more
powerful than other nodes through topological location (within the evolved net-
work) and exploitative game interactions over time.
Zimmermann et al. (2001) found this in their simulations of agents playing a
version of the Prisoner’s Dilemma on an evolving network. Their motivation and
interpretation is socio-economic: agents accumulate ‘wealth’ from the payoffs of
playing games with neighbours and make or break connections to neighbours based
on a simple satisfaction heuristic similar to a rule discussed in Kirman (1993).
Figure 21.5 shows an example of an emergent stable hierarchical network
structure. Interestingly, it was found that, over time, some nodes accumulate
large amounts of ‘wealth’ (through exploitative game behaviour) and other nodes
become ‘leaders’ by being at the top of a hierarchy. These unequal topological and
wealth distributions emerge from simple self-interested behaviour within the net-
work. Essentially, leaders, through their own actions, can re-arrange the topology of
the network significantly whereas those on the bottom of the hierarchy have little
‘topological power’.
578 D. Hales
The idea of explicitly recognising the possibility of differential power between

sub-units in self-* systems and harnessing this is an idea rarely discussed in
engineering contexts but could offer new ways to solve difficult co-ordination
problems.
Considering P2P applications, one can envisage certain kinds of task in which
differential power would be required for efficient operation. Consider e.g. two
nodes negotiating an exchange on behalf of their group or follower nodes. This
might be more efficient than individual nodes having to negotiate with each other
every time they wished to interact. Or consider a node reducing intra-group conflict
by imposing a central plan of action.
We mention the notion of engineering emergent power structures, briefly and
speculatively here, because we consider power to be an under-explored phenome-
non within evolving information systems. Agents, units or nodes are often assumed
to have equal power. It is rare for human societies to possess such egalitarian
properties and perhaps many self-* like properties are facilitated by the application
of unequal power relationships. We consider this a fascinating area for future work.
21.8 Summary
We have introduced the idea of social inspiration for distributed systems design and
given some specific examples from P2P systems. We have argued that social
simulation work can directly inform protocol designs. We have identified some of
the current open issues and problem areas within this research space and pointed out
promising areas for future research.
Increasingly, distributed systems designers are looking to self-organisation as a
way to address their difficult design problems. Also, there has been an explosive
growth in the use of such systems over the Internet, particularly with the high take-
up of peer-to-peer systems. Also the idea of “social software” and the so-called
“Web2.0” approach indicate that social processes are becoming increasingly central
to system design. We believe that the wealth of models and methods produced with
social simulation should have a major impact in this area over the coming decade.
Further Reading
The interested reader could look at the recent series of the IEEE Self-Adaptive and
Self-Organising systems (SASO) conference proceedings which started in 2007 and
have been organised annually (http://www.saso-conference.org/). To get an idea of
current work in social simulation a good place to start is the open access online
Journal Artificial Societies and Social Simulation (JASSS), see http://jasss.soc.
surrey.ac.uk/JASSS.html.
References

Babaoglu O et al (eds) (2005) Self-star properties in complex information systems: conceptual and
practical foundations, vol 3460, Lecture notes in computer science. Springer, Berlin
Babaoglu O et al (2006) Design patterns from biology for distributed computing. ACM T Auton
Adapt Syst 1(1):26–66
Boyd R, Richerson P (1985) Culture and the evolutionary process. University of Chicago Press,
Chicago
Brueckner S, Di Marzo Serugendo G, Hales D, Zambonelli F (eds) (2006) Engineering self-
organising systems. In: Proceedings of the 3rd workshop on engineering self-organising
applications (EOSA’05) (Lecture notes in artificial intelligence, 3910). Springer, Berlin
Cohen B (2003) Incentives build robustness in BitTorrent. In: Chuang J, Krishnan R (eds)
Proceedings of the first workshop on the economics of peer-2-peer systems, 5–6 June 2003,
Berkley. http://www2.sims.berkeley.edu/research/conferences/p2pecon/papers/s4-cohen.pdf
Conte R, Paolucci M (2002) Reputation in artificial societies: social beliefs for social order.
Kluwer, Amsterdam
Edmonds B, Hales D (2005) Computational simulation as theoretical experiment. J Math Sociol 29
(3):209–232
Gamma E, Helm R, Johnson R, Vlissides J (1995) Design patterns: elements of reusable object-
oriented software. Addison-Wesley, Reading
Geller A, Moss S (2013) Modeling power and authority: an emergentist view from Afghanistan.
Hales D (2002) Group reputation supports beneficent norms. J Artif Soc Soc Simul 5(4). http://
Hales D (2006) Emergent group-level selection in a peer-to-peer network. Complexus 3:108–118
Hales D, Arteconi S (2006) SLACER: a self-organizing protocol for coordination in P2P networks.
IEEE Intell Syst 21(2):29–35
Hales D, Edmonds B (2003) Evolving social rationality for MAS using “Tags”. In: Rosenchein JS
et al (eds) Proceedings of the 2nd international conference on autonomous agents and multi-
agent systems, Melbourne, July 2003 (AAMAS 2003). ACM Press, New York, pp 497–503
Hales D, Edmonds B (2005) Applying a socially-inspired technique (tags) to improve cooperation
in P2P networks. IEEE T Syst Man Cy A Syst Hum 35(3):385–395
Hales D, Patarin S (2006) How to cheat BitTorrent and why nobody does. In: Jost J, Reed-Tsochas
F, Schuster P (eds) ECCS2006, Proceedings of the European conference on complex systems,
towards a science of complex systems. http://www.cabdyn.ox.ac.uk/complexity_PDFs/
ECCS06/Conference_Proceedings/PDF/p19.pdf
Kirman A (1993) Ants, rationality and recruitment. Q J Econ 108:137–156
Kirman AP, Vriend NJ (2001) Evolving market structure: an ACE model of price dispersion and
loyalty. J Econ Dyn Control 25(3/4):459–502
Lumsden C, Wilson E (1981) Genes, mind and culture. Harvard University Press, London
Rao AS, Georgeff MP (1991) Modeling rational agents within a BDI-architecture. In: Fikes R,
Sandewall E (eds) Proceedings of knowledge representation and reasoning (KR&R-91).
Morgan Kaufmann, San Mateo, pp 473–484
Riolo R, Cohen M, Axelrod R (2001) Evolution of cooperation without reciprocity. Nature
414:441–443
Rouchier J (2013) Markets. Chapter 23 in this volume
Schelling T (1969) Models of segregation. Am Econ Rev 59:488–493
Schelling T (1971) Dynamic models of segregation. J Math Sociol 1(1):143–186
580 D. Hales
Singh A, Haahr M (2004) Topology adaptation in P2P networks using Schelling’s model. In: Oh
JC, Mosse D (eds) Proceedings of the workshop on games and emergent behaviours in
distributed computing environments, co-located with PPSN VIII, Birmingham, 18–22 Sept
2004
Singh A, Haahr M (2006) Creating an adaptive network of hubs using Schelling’s model. Commun
ACM 49(3):69–73
Staab S et al (2005) Social networks applied. IEEE Intell Syst 20(1):80–93
Wang F-Y, Carley KM, Zeng D, Mao W (2007) Social computing: from social informatics to
social intelligence. IEEE Intell Syst 22(2):79–83
Wooldridge M, Jennings NR (1995) Intelligent agents: theory and practice. Knowl Eng Rev
10(2):115–152
Zimmermann MG, Eguiluz VM, San Miguel M (2001) Cooperation, adaptation and the emergence
of leadership. In: Kirman A, Zimmermann JB (eds) Economics with heterogeneous interacting
agents. Springer, Berlin, pp 73–86
Chapter 22
Simulating Complexity of Animal Social
Behaviour
Charlotte Hemelrijk
Why Read This Chapter? To get an overview of simulation models aimed at

understanding animal social behaviour, such as travelling, foraging, dominance, or
task division. The chapter also provides an analysis of the kinds of insight each
simulation provides, how specific these insights are and whether they are testable.
Abstract Complex social phenomena occur not only among humans, but also
throughout the animal kingdom, from bacteria and amoebae to non-human
primates. At a lower complexity they concern phenomena such as the formation
of groups and their coordination (during travelling, foraging, and nest choice) and at
a higher complexity they deal with individuals that develop individual differences
that affect the social structure of a group (such as its dominance hierarchy, domi-
nance style, social relationships and task division). In this chapter, we survey
models that give insight into the way in which such complex social phenomena
may originate by self-organisation in groups of beetle larvae, in colonies of ants and
bumblebees, in groups of fish, and groups of primates. We confine ourselves to
simulations and models within the framework of complexity science. These models
show that the interactions of an individual with others and with its environment lead
to patterns at a group level that are emergent and are not coded in the individual
(genetically or physiologically), such as the oblong shape of a fish school, specific
swarming pattern in ants, the centrality of dominants in primates and the task
division among bumble bees. The hypotheses provided by these models appear to
be more parsimonious than usual in the number of adaptive traits and the degree of
cognitive sophistication involved. With regard to the usefulness of these
simulations, we discuss for each model what kind of insight it provides, whether
it is biologically relevant, and if so, whether it is specific to the species and
environment and to what extent it delivers testable hypotheses.
C. Hemelrijk (*)
Professor of Self-organisation in social systems, Behavioural ecology and self-organisation,
BESO, Centre for Ecological and Evolutionary Studies, University of Groningen, Centre for Life
Sciences, room 72.05.54, Nijenborgh 7, 9747 AG GRONINGEN

582 C. Hemelrijk
22.1 Introduction
Many complex social phenomena of human behaviour are also observed in

animals. For instance, humans coordinate their movement while searching for
the right restaurant. They also make and follow paths on lawns in between
different university buildings in order to cross between them. This resembles
the path marking and following behaviour of ants as they forage and select
different food sources. Furthermore, a division of tasks is found in both human
social organisations and in large colonies of social insects, such as honey bees.
Human social relationships are diverse, and so are the social relationships of
primates. In groups of humans and animals, competition results in stable
relationships in which one individual consistently beats the other. Within a
group, individuals can be ordered in a dominance hierarchy. Furthermore, domi-
nant relationships are also found between groups and between classes of
individuals, such as between the sexes. Further, societies may differ in their
dominance style, such as whether they are egalitarian or despotic. In summary,
despite the hugely inferior cognitive capacities of animals they show a number of
complex social phenomena that resemble those of humans. From this the question
arises whether or not these complex phenomena originate by self-organisation in
the same way as both humans and animals. Therefore, it is important to survey
complex social behaviour in animals in addition to that of humans.
We divide the chapter in subsections dealing with different (groups of) complex
phenomena ordered in increasing complexity: From group formation (which is
simple in animals) and coordination under various circumstances (such as during
travelling, foraging, selection of shelter and nest site) we continue to the social
organisation of the group (its dominance hierarchy, dominance style, dominance
classes, social relationships, personality style and task division).
We mainly confine ourselves to individual-based models that are spatially
explicit. Individuals in the model are steered by behavioural rules that are fixed
or that are based on parameters that change (compare self-organisation with(out)
structural changes, Pfeifer and Scheier 1999); they react to their local environment
only. The environment may contain only other individuals, or may also contain
food. Food may be continuously abundant, or it may be depleted and may re-grow
or not after being eaten.
In general, we discuss for each model whether it has led to a new perspective in
the study of the phenomenon; whether this perspective is more ‘parsimonious’ than
previous explanations in terms of cognitive sophistication and the number of
specific behavioural adaptations; whether the new explanation concerns a general
principle or is specific to a certain species or environment; and whether it produces
hypotheses that can easily be tested in real animals. In the evaluation we also
provide literature for further reading and indicate important areas for future
modelling.
22 Simulating Complexity of Animal Social Behaviour 583
22.2 Group Formation and Coordination
Everywhere in nature groups are formed. They are formed among individuals of all
kinds of species in all stages of their lives and in a few cases groups contain several
species. Groups may differ in size, longevity, the degree of heterogeneity and also
in the way they are formed. We distinguish two processes of group formation,
namely social and environmental processes. Social processes concern forms of
social attraction, and environmental processes relate to attraction to food resources,
shelter and the like. Below we discuss models for each process separately and for
their combination.
22.2.1 Social Attraction
Attraction is often mediated through visual and chemical cues. Visual attraction is
important in many species of fish and birds. Chemical attraction (through
pheromones) occurs among single celled individuals (such as bacteria and
amoebae), ants and beetle larvae.
One of the simplest aggregation processes has been studied in certain
experiments and models of cluster formation of beetle larvae (Camazine et al.
2001). In their natural habitat, these beetle larvae live in oak trees and grouping
helps them to overcome the production of toxic resin by the host tree. They emit a
pheromone and move towards higher concentrations of this chemical. In the
experimental setup, larvae are kept in a Petri dish. In the model, the behavioural
rules of following a gradient are represented in individuals that roam in an environ-
ment that resembles a Petri dish. It appears that both in the model and in the
experiment the speed of cluster growth and the final number and distribution of
clusters depend on the initial distribution and density of the larvae. By means of
both model and experimental data it has been shown that cluster growth can be
explained by a positive feedback. A larger group emits more pheromone and
therefore attracts more members. Consequently, its size increases in turn emitting
more pheromone, etcetera. This process is faster at a higher density, because
individuals meet each other more frequently. Thus clusters appear sooner. Further-
more, growth is faster at higher densities, because more individuals are available for
growing.
The location of the clusters has been studied. After starting from a random
distribution of individuals, a single cluster remains in the centre of the Petri dish.
This location is due to the higher frequency with which individuals visit the centre
than the periphery. Starting from a peripheral cluster, there is an attraction to both
the peripheral cluster and the centre. Thus, there is a kind of competition between
clusters to attract additional individuals. The final distribution of clusters may
consist of one cluster (at the centre, the periphery or at an intermediate location),
or of two clusters (peripheral and in the centre). The final distribution depends on
584 C. Hemelrijk
self-organisation and on three factors, namely, the initial density of individuals,

initial distribution of clusters, and whether there is social attraction. Although the
model is devised for the specific experimental setup and species, the mechanisms of
group formation and growth shown in the model are found in many animal species.
For instance, similar patterns are observed in models and experiments of
cockroaches.
Cockroaches aggregate by sensing the relative humidity. They tend to move
towards lower humidity. Here larger groups grow faster, because individuals have a
higher tendency to stop and form a group if they collide with a larger number of
others. Besides, they rest longer if more individuals are resting close by. They use
the relative humidity as a cue to estimate the number of individuals. A lower
relative humidity correlates with a higher number of individuals close by (Jeanson
et al. 2005). Despite the different underlying process of social attraction (phero-
monal, visual or based on relative humidity) pattern formation is similar to that of
beetle larvae.
22.2.2 Foraging
A model that deals with group formation solely through environmental causes
concerns the splitting and merging of groups as it is found in the so-called
fission-fusion system of primates, in particular that of spider monkeys (Ramos-
Fernández et al. 2006). It relates the pattern of group formation to various
distributions of food, because particularly in these fission-fusion societies, the
food distribution may have an effect on grouping. In the model, the food distribu-
tion is based on a distribution of resources in forests that follows an inverse power
law as described by Enquist and co-authors (Enquist and Niklas 2001; Enquist et al.
1999). The foragers maximise their food intake by moving to larger food patches
that are closer by (they minimise the factor distance divided by patch size). Further,
they do not visit patches that they have visited before. Individuals have no social
behaviour and they meet in groups only because they accidentally visit the same
food patch. For a distribution with both large and small trees (patches) in roughly
equal numbers the model leads to a frequency distribution of subgroup sizes that
resembles that of spider monkeys. Results hold, however, only if foragers have
complete knowledge of the environment in terms of patch size and distance. The
resemblance is not found if individuals only know (a random) part of the environ-
ment. Furthermore, if they have complete knowledge, individuals meet with certain
partners more often than expected if the choice of the food patch is a random
choice. In this way, certain networks of encounters develop. Another serious
shortcoming of the model is that food can only be depleted; there is no food renewal
and individuals do not return to former food patches. The main conclusion is that
the ecological environment influences grouping: Ecologies that differ in the varia-
tion of their patch sizes (high, medium or low variance) also differ in the distribu-
tion of subgroup sizes. This model delivers mainly a proof of principle.
22.2.3 Combination of Social and Environmental Processes
The relation between ecology and sub-group formation is also shown in a simula-
tion study of orang-utans in the Ketambe forest in Indonesia (te Boekhorst and
Hogeweg 1994a). The predictions for and resemblance to empirical data that are
delivered in this model (te Boekhorst and Hogeweg 1994a) are more specific than
those in the model of spider monkeys discussed above (Ramos-Fernández et al.
2006). Here, realistic patterns of grouping arise from simple foraging rules in
interaction with the structure of the forest and in the presence of two social rules.
The environment and population are built to resemble the Ketambe area in the
composition of the community and the size of the habitat. Two main categories of
food trees are distinguished, figs (very large and non-seasonal) and fruit trees (small
and seasonal). There are 480,000 trees of which 1,200 bear fruit at the same time,
once a year for ten consecutive days unless depleted earlier by the individuals. The
crop size and spatial distribution of trees are specified: fig trees are clustered and
fruit trees distributed randomly. Furthermore there are randomly distributed sources
of protein. These are renewed immediately after being eaten. In the fruiting season
extra individuals migrate into the area. With regard to the behavioural rules,
individuals first search for a fig tree. If this is not found, while moving approxi-
mately in the direction of others, because other individuals may indicate the
presence of food on a distant fig tree, individuals look for fruits close by. Upon
entering a tree, the individual feeds until satiated or the tree is emptied. Next the
individual rests close by and starts all over again later on. A further social rule
causes adult males to avoid each other.
The resulting grouping patterns in the model resemble those of real orang-utans
in many respects: Temporary aggregations are found in the enormous fig trees. This
happens because in these big trees individuals may feed until satiated and then leave
separately. However, when feeding in the much smaller fruit trees, food is insuffi-
cient and therefore individuals move to the next tree. This causes them to travel
together. Thus, travelling bands develop mainly in the fruit season. In this season
parties are larger than when fruit is scarce. Groupings emerge as a consequence of
simple foraging rules in interaction with the forest structure. Thus, the model makes
clear that differences between forest structures will lead to different grouping
patterns in apes. This is a parsimonious explanation, because it is not necessary to
think in terms of costs and benefits of sociality to explain these grouping patterns,
rather they arise as a side effect of feeding behaviour. The empirical data analysis
and ideas for the model of orang-utans were inspired by findings in another model
on grouping in chimpanzees.
This model of chimpanzees concerns both foraging and social attraction and is
meant to explain group formation, in particular the fission-fusion society (te
Boekhorst and Hogeweg 1994b). It offers new explanations for the relatively
solitary life of females and the numerous and large subgroups of males. Subgroups
of males had so far been supposed to form in order to join in defending the
community. To explain such cooperation, males were believed to be genetically
586 C. Hemelrijk
LOOK FOR
CHIMP
males
females
CHIMP NO CHIMP
both
CHIMP
LOOK FOR
=
FRUIT
ELSE
UNTIL
FRUIT NO FRUIT
GOTO CHIMP LOOK FOR FORWARD

PROT
GOTO FRUIT PROT NO PROT

CHIMP
=
ESTRUS
EAT GOTO PROT
FOLLOW REST
Fig. 22.1 Behavioural rules of artificial chimpanzees (te Boekhorst and Hogeweg 1994b)
related to each other. Further, the solitary life of females was attributed to competi-
tion for food. However, the model shows that their patterns of grouping may arise in
an entirely different way: they may result from a difference between the sexes in
diet and in the priority of foraging versus reproduction. The model resembles a
certain community of chimpanzees in Gombe in its community composition (num-
ber of males and females, 11 and 14 respectively, and the number of females that
are synchronously in oestrus) and its habitat size (4 km2), its number of trees, the
number of trees that bear fruit at the same time, the length of their fruit bearing,
their crop size, and the speed of their depletion. The number of protein resources is
modelled more arbitrarily, approximating such forests. With regard to their
behavioural rules, individuals of both sexes look for fruit and when satiated have
a rest close to the tree (Fig. 22.1). If not satiated they move towards the next fruit
tree. They continue to do this until they are satiated. If there is no fruit on a specific
tree, females (but not males) look for proteins before searching another fruit tree.
Males have one priority over finding food: finding females. Whenever they see a
chimpanzee, they approach it to investigate whether it is a female in oestrus. If this
is the case, males in the model follow her until she is no longer in oestrus.
In the model, patterns of grouping resemble those of real chimpanzees. Male
groups emerge because all males have the same diet which differs from that of
females and because they will follow the same female in oestrus together. Further-
more, in the model, as in real chimpanzees, males appear to travel over longer
distances per day than females. They do so because they are travelling in larger
groups and this leads to faster depletion of food sources and therefore they have to
visit a larger number of trees in order to become satiated. This interconnection
between the number of trees visited, group size and distance covered led to the
hypothesis for orang-utans that different kinds of groupings (aggregation and travel
bands) originate in different trees (fig and fruit). Note that these results were robust
for a large range of different ecological variables and different compositions of the
community. Here, a difference between the sexes in their diet and their priorities for
sex and food appeared essential. With regard to the chimpanzee community, the
authors conclude that, to explain its fission-fusion structure, the genetically based
theory that kin-related males are jointly defending the community is not needed. In
fact, in subsequent DNA studies no special genetic relatedness was found among
cooperating male chimpanzees in Kibale (Goldberg and Wrangham 1997). Instead,
chimpanzee-like party structures may emerge by self-organisation if chimpanzees
search for food and for mates in a forest. Besides, the model can be used to explain
the frequent bi-sexual groups observed in bonobos as being caused by their
prolonged period of oestrus. Whereas these models have been controversial
among primatologists for a long time, their usefulness is slowly becoming accepted
(Aureli et al. 2008).
22.2.4 Group Coordination and Foraging
Groups of social insects, for instance ants, are remarkably efficient in foraging.
They collectively choose a food source that is closer rather than one (of the same
quality) that is further away (Deneubourg and Goss 1989). Their choice is made
without comparing the distance to different food sources. Experiments and models
show that ants use trail pheromones as a mechanism of collective ‘decision’
making: Individuals mark their path with pheromone and at crossings follow the
path that is more strongly marked. As they return to the nest sooner when the food
source is close by, they obviously imprint the shorter path more often with
pheromones. This results in a positive feedback: as the shorter path receives
stronger markings, it also receives more ants, etcetera. Thus, the interaction
between ants and their environment results in the adaptive and efficient exploitation
of food sources. The ‘preference’ for a food source nearby rather than one further
away is a side-effect of pheromonal marking. This marking also helps a single ant to
find its way and may initially have been an adaptation to cause the ant to return to its
nest. It is actually more than just that, because its intensity is adapted to the quality
of the food source. The process of marking and following may also lead to mistakes.
For instance, if a path is initially developed to a food source of low quality, and later
to a food source of high quality is introduced elsewhere, due to ‘historical con-
straint’ the ants may remain fixated on the source of low quality even if it is further
from the nest.
Army ants live in colonies of 200,000 individuals; they are virtually blind; they
travel in a swarm with about 100,000 workers back and forth to collect food.
588 C. Hemelrijk
a b c d
Step 1100 Step 1300
Fig. 22.2 Foraging patterns of two species of army ants, Eciton burchelli and Eciton rapax,
empirical data and models. Empirical data: Eciton burchelli (a) and Eciton rapax (b); models: (c) a
few food clumps, (d) frequent occurrence of single food items, (e) network of nodes (After
Deneubourg and Goss 1989)
Different species of army ants display highly structured patterns of swarming that
may be species specific (Fig. 22.2a, b). For example, Eciton burchelli has a more
dispersed swarm than E. rapax. Such species-specific differences in swarming are
usually regarded as a separate adaptation, which is assumed to be based on
corresponding differences in the underlying behavioural tendencies of coordina-
tion. Deneubourg and co-authors (Deneubourg and Goss 1989; Deneubourg et al.
1998), however, give a simpler explanation. In a model, they have shown that such
markedly different swarming patterns may arise from a single rule-system of laying
and following pheromone trails when ants are raiding food sources with different
spatial distributions (Fig. 22.2c, d). The authors have shown this in a model in
which ‘artificial ants’ move in a network of points (Fig. 22.2e), and mark their path
with pheromones. When choosing between left and right, they prefer the more
strongly marked direction. By introducing different distributions of food in the
model (either uniformly distributed single insects or sparsely distributed colonies of
insects), different swarming patterns arise from the interaction between the flow of
ants heading away from the nest to collect food and the spatial distribution of the
foragers returning with the food. These different swarm types are remarkably
similar to those of the two species of army ants mentioned above (for empirical
confirmation see Franks et al. 1991). Therefore, these different swarm forms reflect
the variation in diet of the different species. Thus, the explanation of the model is
more parsimonious than if we assume the different swarm forms to arise from a
specific adaptation in rules of swarming. In summary, this model teaches us the
effects of the environment on swarm coordination.
Fig. 22.3 Behavioural areas

of avoidance (separation), esion are
alignment and attraction Coh a
(cohesion) with a dead
(‘blind’) angle at the back ment ar
ign e
(From Hemelrijk and Al a
Hildenbrandt 2008). The
ration a
separation angle is indicated pa
re
Se
as σ
a
σ
With regard to its evolution, natural selection shapes the necessary traits for the
successful marking and following of trails depending on the size of the food source
and other environmental characteristics (for an evolutionary model of this, see Solé
et al. 2001).
22.2.5 Group Coordination in a Homogeneous Environment
Even in environments (such as the open sea, savannah and sky) that are virtually
uniform without environmental structure, remarkable coordination is observed in
the swarms of many animal species, e.g. of insects, fish, birds and ungulates.
Coordination appears flexible even in swarms of a very large size (for instance,
of up to ten million individuals in certain species of fish). Swarming behaviour has
been modelled in several ways. The most simplistic representations of emergent
phenomena have used partial differential equations. Slightly more complex
behaviour has been obtained using particle-based models derived from statistical
mechanics in physics (Vicsek et al. 1995). These models have been used to explain
the phase transition of unordered to ordered swarms in locusts (Buhl et al. 2006).
Yet the biologically most relevant results come from models wherein individuals
coordinate with their local neighbours by following only three rules based on zones
of perception (Fig. 22.3): They avoid neighbours that are close by (separation),
align to others up to an intermediate distance (alignment) and approach those
further away (cohesion). These models have been applied to describe herds of
ungulates (Gueron et al. 1996), schools of fish (Couzin et al. 2002; Hemelrijk and
Kunz 2005; Huth and Wissel 1992, 1994; Kunz and Hemelrijk 2003; for a review
see Parrish and Viscido 2005), and swarms of birds (Hildenbrandt et al. 2010;
Reynolds 1987; Hemelrijk & Hildenbrandt 2011; Hemelrijk & Hildenbrandt 2012).
590 C. Hemelrijk
Through these models we obtain insight into a number of important biological

aspects of swarming, which have mainly been related to schools of fish.
Firstly, we get insight into the coordination of schools. Schools coordinate with
remarkable flexibility even into the absence of a leader and without a directional
preference. The direction of their movement is merely the consequence of the
location and heading of others in the school. With regard to the question whether
there is a leader in a swarm, such a leader fish is supposed to be located at the front
(Bumann and Krause 1993). However, neither in such models of fish schools nor in
real swarms individuals appear to have a fixed location. Instead frontal locations are
continuously switched in the model on average every 2 s, in real fish every 1.4 s
(Gnatapogon elongates, Huth and Wissel 1994). Thus, there cannot be consistent
leaders.
Furthermore, if a number of individuals have a directional preference (for
instance, for certain food sources or breeding locations), but most of them do not,
those with such a preference will automatically lead the school. Huse and co-
authors (Huse et al. 2002) showed that even if the percentage of individuals that
has a certain preferred direction is very small (though above 7 %), this may
influence the direction of the entire school.
If individuals prefer a different direction, for instance, because they aim to go to
different food locations, the school may react differently depending on the degree to
which two preferred directions differ (Couzin et al. 2005): If the directions differ
little, the group will follow the average direction between the two (‘a compro-
mise’). If the angle between both directions is large, the group will either follow the
direction that is preferred by the majority or in the absence of a ‘convincing
majority’ it will randomly choose one of the two. In these examples of swarming
the models help us to understand the processes that determine the direction in which
a school is heading.
Secondly, we obtain insight into the segregation of individuals that differ in a
certain trait. In a school, individuals may be segregated, for example, according to
size. Usually this is attributed to an intentional or genetic preference for being near
individuals of the same size or body form. Models show, however, that this
segregation may also arise directly as a side-effect of differences in body size
without any preference (e.g. see Couzin et al. 2002; Hemelrijk 2005; Kunz and
Hemelrijk 2003). This may, for instance, arise because larger individuals due to
their larger body have a larger range at which they avoid others who are too close.
Thus, by avoiding smaller individuals more often than the reverse, large individuals
may end up at the periphery leaving the small ones in the centre (as has been found
in water insects (Romey 1995)).
Thirdly, natural schools of fish show a number of traits that are believed to be
helpful in the protection against predators: Their shape is oblong and their density
is highest at the front. Bumann et al. (1997) argue that an oblong form and high
frontal density protect against predation: the oblong shape reduces the size of the
frontal area, where predators are supposed to attack and high frontal density
protects individuals against approaching predators. Hemelrijk and co-authors
(Hemelrijk and Hildenbrandt 2008; Hemelrijk and Kunz 2005) have noted that it
is unlikely that individual fish actively organise themselves so as to create these two
patterns. Therefore, the authors studied in a model whether these patterns might
arise by self-organisation as a side-effect of their coordinated movements. This
indeed appeared to be the case and their emergence appeared to be robust (inde-
pendent of school size and speed). These patterns come about because, during
travelling, individuals avoid collisions mostly by slowing down (as they do not
perceive others to be directly behind them, in the so-called blind angle, Fig. 22.3).
Consequently, they leave a gap between their former forward neighbours and
subsequently these former neighbours move inwards to be closer together. Thus,
the school becomes oblong.
Furthermore, when individuals fall back, a loose tail builds up and this automat-
ically leaves the highest density at the front. In the model it appears that larger
schools are relatively more oblong because they are denser and so more individuals
fall back to avoid collision. Faster schools appear to be less oblong. This arises
because fast individuals have greater difficulty to turn, thus, the path of the school is
straighter, the school is more aligned (polarised) and, therefore, fewer individuals
fall back. Consequently, the core of a faster school is denser and the tail is looser
than they are in slower schools. Recent tests in real fish (mullets) confirm the
specific relationships between group shape, and its size, density, and polarisation as
found in the model (Hemelrijk et al. 2010). Although this indicates that the shape in
real schools develops in a similar way, it is still necessary to investigate shape and
frontal density in more species and to study the effects of different speeds on these
traits.
Fourth, in real animals predation and attacks on swarms result in a spectacular
range of behavioural patterns of evasion by schooling prey. These patterns are
supposed to confuse the predator. They have been labelled ‘tight ball’, ‘bend,
‘hourglass’, ‘fountain effect’, ‘vacuole’, ‘split’, ‘join’, and ‘herd’. They have
been described for schools of several prey species and predators (Axelsen et al.
2001; Lee 2006; Nottestad and Axelsen 1999; Parrish 1993; Pitcher and Wyche
1983). Most of these patterns may be obtained in a model by simple behavioural
rules of prey and predator (e.g., see Inada and Kawachi 2002). Many of the different
patterns of evasion result in self-organisation in models of schooling which are built
in such a way that upon detecting the predator, individuals compromise between
their tendency to avoid the predator, and to coordinate with their group members.
Though these models do not exactly fit real data, they give us insight into how
specific collective evasion patterns may arise.
22.3 Social Organisation
Although groups may be beneficial for their members in so far as they provide
protection against predation, they also result in competition for food, mates and
space. If individuals meet for the first time, such competitive interactions may
initially have a random outcome. Over time, however, a dominance hierarchy
592 C. Hemelrijk
develops, whereby certain individuals are consistently victorious over others and
are said to have a higher dominance value than others (Drews 1993). Individuals
may perceive the dominance value of the other from the body posture of the other
(in primates) or from their pheromone composition (in insects).
With regard to the question which individual becomes dominant, there are two
extremely opposing views: dominance as a fixed trait by inherited predisposition
and dominance by chance and self-organisation. While some argue for the impor-
tance of predisposition of dominance by its (genetic) inheritance (Ellis 1991),
others reject this for the following reasons: Experimental results show that the
dominance of an individual depends on the order of its introduction in a group
(Bernstein and Gordon 1980). Dominance changes with experience, because the
effects of victory and defeat in conflicts are self-reinforcing, the so-called winner-
loser effect. This implies that winning a fight increases the probability of victory in
the next fight and losing a fight increases the probability of defeat the next time.
This effect has been established empirically in many animal species and is
accompanied by psychological and physiological changes, such as hormonal
fluctuations (Bonabeau et al. 1996a; see Chase et al. 1994; Hemelrijk 2000; for a
recent review, see Hsu et al. 2006; Mazur 1985).
The self-reinforcing effects of fighting are the core of a model called DomWorld,
which concerns individuals that group and compete (Hemelrijk 1996, 1999b;
Hemelrijk and Wantia 2005). It leads to a number of emergent phenomena that
have relevance for many species, in particular for primates and more specifically
macaques.
22.3.1 The Basic DomWorld Model
The model DomWorld consists of a homogeneous world in which individuals are

grouping and competing. It does not specify what they compete about. Grouping is
implemented by making individuals react to others in different manners depending
on the distance to the other (Fig. 22.4). An individual is attracted to others far away
(within MaxView), it continues to follow its direction if it perceives others at an
intermediate distance (within NearView), and it decides whether or not to perform a
competitive dominance interaction if it encounters others close by (within
PerSpace) (Hemelrijk 2000). After winning a dominance interaction, it chases
away the other and after losing a fight, it flees from it.
Dominance interactions are implemented after the DoDom interactions by
Hogeweg (Hogeweg and Hesper 1985), which were extended to reflect differences
in the intensity of aggression between species and between the sexes and with the
decision whether or not to attack depending on risks involved (Hemelrijk 1998,
1999b). Each individual has a dominance value which indicates the individual’s
capacity to win. At the beginning of a competitive interaction both individuals
display their dominance value and observe that of the other. The outcome of the
fight depends on the relative dominance of both partners and on chance.
Select nearest partner Others ≤ NearView?
YES NO YES NO
Dominance interaction
Others ≤ PerSpace? Move on

Ego wins Ego loses
Others ≤ MaxView?
Flee from Move to

opponent other YES NO
Go to Turn (SearchAngle
opponent at random to right or left)
Fig. 22.4 Flowchart of the behavioural rules of individuals in DomWorld. PerSpace, NearView
and MaxView indicate close, intermediate and long distances, respectively
The probability of winning is higher the higher the dominance value of an individ-
ual in relation to that of the other. Initially, the dominance values are the same for
all individuals. Thus, during the first encounter, chance decides who wins. After
winning, the dominance of the winner increases and that of the loser decreases.
Consequently, the winner has a greater chance to win again (and vice versa) which
reflects the self-reinforcing effects of the victories (and defeats) in conflicts of real
animals.
We allow for rank reversals; when, unexpectedly, a lower-ranking individual
defeats a higher-ranking opponent, this outcome has a greater impact on the
dominance values of both opponents, which change with a greater amount than
when, as we would expect, the same individual conquers a lower-ranking opponent
(conform to detailed behavioral studies on bumble bees by van Honk and Hogeweg
1981). Furthermore, in their decision whether or not to attack, we made individuals
sensitive to the risks, i.e. the ‘will’ to undertake an aggressive interaction (instead of
remaining non-aggressively close by) increases with the chance to defeat the
opponent, which depends on the relative dominance ranks of both opponents
(Hemelrijk 1998).
We also represented the intensity of aggression in which primate societies differ
(in some species individuals bite, in other species they merely approach, threaten
and slap) as a scaling factor (called StepDom) that weighs the changes in domi-
nance value after a fight more heavily if the fight was intense (such as biting) than if
the fight was mild (involving threats and slaps, or merely approaches and retreats)
(Hemelrijk 1999b). In several models, we distinguished two types of individuals
representing males and females (Hemelrijk et al. 2003). We gave males a higher
initial dominance value and a higher intensity of aggression (reflecting their larger
594 C. Hemelrijk
High Intensity
35
2
25 4 6 3 1.5
1.5 5 1.5
4 5 5 71.54 2
15 7 2 1.5
4
4 5
1 374 6 4 1.54
41.5 2
1 1.53
3 1.5 6 2 7 7 4 5 1.5
6 4 4
6 5 6 3 6
6
5
3
1.5 3 4 7
75
5
6 6 3 5 1
4 77366
6 6
6 46 633 5 4 71.5 4
5 1.5
31.5
6
7 3 7 1.57 6 5
47 7
7 4
8
4 355 666 3 1.5 5 7
4 7 3
1.5 71 3 7 41.55 5
76
4 2
7 4
Y 67 6
2 36
6 5 1.5
21.5 3 5 4 5 3 1.5
7 1.5 5 1 31.5
1.5
54 14 3
5 1.5 3
-5 2 1 3
5
4 1.5
2 1.5
2
7
3 77 2
24
2 5
3
3
2 17 2 5
5 1
5 1
15
1.5 7 1 6
4 2
-15 2 1.5
6 6 7
3
1 2 5
1.5
-25
-35
-35 -25 -15 -5 5 15 25 35
X
Fig. 22.5 Spatial-social structure. Darker shading indicates areas with individuals of decreasing
dominance rank
body size, stronger muscular structure and larger canines than those of females). In
fights between both types of individuals, the intensity of the fight was determined
by the initiator (the attacker).
22.3.2 Spatial Structure
The major advantage of group life is supposedly to be protection against predators.

Central positions are supposed to be safest, because here individuals are shielded
off by other group-members from predators approaching from the outside. There-
fore, according to the well-known ‘selfish herd’ theory of Hamilton (1971),
individuals have evolved a preference for a position in the centre, the so-called
‘centripetal instinct’. If competition for this location is won by dominants,
dominants will end up in the centre. This is thought to be the main reason why in
many animal species dominants are seen to occupy the centre. However, in
DomWorld this spatial structure emerges, even though such a preference is lacking
and there is no ‘centripetal instinct’ nor threat of predation (Hemelrijk 2000).
The spatial configuration, with dominant individuals in the centre and
subordinates at the periphery of the group (Fig. 22.5), emerges in the model due
to a feedback between the dominance hierarchy and the spatial location of the
individuals of different rank. During the development of the hierarchy, some
individuals become permanent losers. Such low-ranking individuals must end up
at the periphery by being constantly chased away. This automatically leaves
dominants in the centre. Also, in real animals, such a spatial structure may occur
although its members have no centripetal instinct nor experience a threat of
predation. For instance, in the elegant experiments with fish by Krause (1993),
central dominants were observed, although no centre-oriented locomotion appeared

(Krause and Tegeder 1994). Furthermore, this spatial structure has been described
in hammerhead sharks in spite of the absence of any predatory threat (Klimley
1985). Thus, the model provides a new way of understanding spatial structure.
22.3.3 Dominance Style: Egalitarian and Despotic Societies
High dominance ranking is supposed to be associated with benefits such as priority

of access to mates, food and safe locations. If benefits are strongly biased towards
higher-ranking individuals, the society is called ‘despotic’, whereas if access to
resources is more equally distributed, it is called ‘egalitarian’. These terms have
been used to classify social systems of many animal species (such as insects, birds
and primates). Egalitarian and despotic species of primates, such as macaques,
appear to differ in many other traits too, such as in group density, their intensity and
frequency of aggression and in their frequency and patterns of affiliation
(grooming). Usually these differences are explained from the perspective of
optimisation of single traits by natural selection. However, Thierry (1990b)
suggests that in macaques the many behavioural differences can be traced back to
two inherited differences, namely degree of nepotism (i.e. cooperation among kin)
and intensity of aggression. Note that despotic macaques display aggression of a
higher intensity, i.e. they bite more often, whereas egalitarian macaques display
aggression that is milder, they only threaten and slap.
The model DomWorld presents an even simpler hypothesis (Hemelrijk 1999b),
namely that a mere difference in intensity of aggression produces both types of
societies. By increasing the value of one parameter, namely that of intensity of
aggression, the artificial society switches from a typically egalitarian society to a
despotic one. For instance, compared to egalitarian artificial societies, despotic ones
are more loosely grouped, showing a higher frequency of attack, their behaviour is
more rank-related, aggression is more asymmetric, spatial centrality of dominants is
clearer and female dominance over males is greater. All these differences between
despotic and egalitarian societies arise via feedback between the development of
the hierarchy and spatial structure and this happens only at a high intensity of
aggression. The steep hierarchy develops as a consequence of the high aggression
intensity, because each outcome has a stronger impact than at a low intensity and it
is strengthened further via a mutual feedback between the hierarchy and the spatial
structure with dominants in the center and subordinates at the periphery (Hemelrijk
1999b, 2000).
Pronounced rank-development causes low-ranking individuals to be continu-
ously chased away by others and thus the group spreads out (1 in Fig. 22.6).
Consequently, the frequency of attack diminishes among the individuals (2 in
Fig. 22.6) and therefore the hierarchy stabilizes (3 in Fig. 22.6). While low-ranking
individuals flee from everyone, this automatically leaves dominants in the center,
and thus a spatial-social structure develops (Fig. 22.5). Since individuals of similar
596 C. Hemelrijk
Spatial
Distance
1 + 2 –
5 Dominance Aggression
Differentation Frequency
+ 6+ 3+
Uni-
directionality
+ + +
Spatial 4 + Dominance
Centrality Stability
Fig. 22.6 Interconnection between variables causing spatial-social structuring at a high aggres-
sion intensity
dominance are treated by others in more or less the same way, similar individuals
remain close together; therefore, they interact mainly with others of similar rank;
thus, if a rank reversal between two opponents occurs, it is only a minor one
because opponents are often similar in dominance. In this way the spatial structure
stabilizes the hierarchy and it maintains the hierarchical differentiation (4 and 5 in
Fig. 22.6). Also, the hierarchical differentiation and the hierarchical stability
mutually strengthen each other (6 in Fig. 22.6).
In short, the model (Hemelrijk 1999b) makes it clear that changing a single
parameter representing the intensity of aggression, may cause a switch from an
egalitarian to a despotic society. Since all the differences resemble those found
between egalitarian and despotic societies of macaques, this implies that in real
macaques these differences may also be entirely due to a single trait, intensity of
aggression. Apart from intense aggression, such as biting, however, a high fre-
quency of aggression can also cause this switch (Hemelrijk 1999a). A higher
frequency of aggression also leads to a steeper hierarchy. This in turn, results in a
clearer spatial structure which again strengthens the development of the hierarchy
and this has the cascade of consequences as described for aggression intensity.
One of the shortcomings of the model may be considered as the lack of individ-
ual recognition among group members. In a subsequent model, this was corrected
by having each individual keep a record of the dominance value of each group
member. This value was updated depending on its experiences gained with other
group members (Hemelrijk 1996, 2000). With regard to the development of spatial
structure, hierarchy and dominance style remained similar, but patterns were
weaker than in the case of a direct perception of dominance without individual
recognition. Weaker patterns arise due to the contradictory experiences that differ-
ent individuals have with specific others. This will impede a hierarchical develop-
ment and thus weaken the accompanying consequences. Even though this may be
more realistic for certain modelling purposes, it is more useful to have clearer
patterns in a simpler model. Such a caricature is more helpful for building upon
understanding and developing new ideas.
Dominance style is usually considered to be species specific, but Preuschoft and
colleagues (Preuschoft et al. 1998) raised the question whether competitive regimes
(egalitarian versus despotic) should not rather be considered as sex specific. In their
study of the competitive regime of both sexes of Barbary macaques, they found an
unexpected sex difference: males behave in an egalitarian way whereas females are
despotic. It is unexpected that the sex with the larger body size and fiercer aggression
evolved a more egalitarian dominance style. Therefore, it seems to be a separate
adaptation. However, the same difference in dominance style between the sexes was
also found in DomWorld: males appear to be more egalitarian than females. The
unexpectedly stronger egalitarianism of males in the model is due to yet another
difference between the sexes, the higher initial dominance of males compared to
females (which relates to differences in body size). Consequently, single events of
victory and defeat have less impact on their overall power or dominance. Therefore,
they lead to less hierarchical differentiation than among females, who are much
smaller and weaker and on whom each victory and defeat therefore has more impact.
The greater the sexual difference in initial dominance between the sexes, the more
egalitarian the males behave among themselves compared to the behaviour of the
females among themselves. The conclusion of this study is that the degree of sexual
dimorphism may influence the competitive regime of each sex, in the model and in
real primates. Further empirical studies are needed to investigate whether the degree
of sexual dimorphism is directly proportional to the steepness of the gradient of the
hierarchy of females compared to males.
With regard to its evolution, dominance style is supposed to be a consequence of
different degrees of competition within and between groups (van Schaik 1989).
According to this theory, when competition within groups is high and that between
groups is low, despotic societies evolve and if it is reversed, and competition between
groups is high, egalitarian groups emerge. In line with this, DomWorld already shows
that high competition within a group leads to a despotic society (Hemelrijk 1999b).
However, in a subsequent study, competition between groups appears to favour
despotic rather than egalitarian groups (Wantia 2007). To study effects of competi-
tion between groups, the DomWorld model was extended to several groups (i.e.
GroupWorld, Wantia 2007). Here, as in real primates, usually individuals of high
rank participate in encounters between groups (e.g. see, Cooper 2004).
The model generates a number of unexpected results. Firstly, among groups of
the same dominance style, competition between groups does not affect dominance
style, since it happens at a very low frequency compared to competition within
groups. However, in competition between groups of different dominance style,
remarkable results came to light. Unexpectedly, under most conditions groups with
a more despotic dominance style were victorious over others with a more egalitar-
ian style. This arose due to the greater power of individuals of the highest rank of
the despotic group compared to that of the egalitarian group. In the model, this is a
consequence of the steeper hierarchy in despotic groups compared to that in
598 C. Hemelrijk
egalitarian groups. In reality this effect may be even stronger, because higher-
ranking individuals may also obtain relatively more resources in a despotic group
than in an egalitarian one. The outcome of fights between groups depends, however,
on the details of the fights between groups and the composition of the group. When
participants of inter-group fights fought in dyads or in coalitions of equal size, the
despotic group out-competed the egalitarian one. If, however, individuals of egali-
tarian groups, for one reason or another, fought in coalitions of a larger size or if
their coalitions included more males than those of the despotic groups, the egalitar-
ian group had a chance to win. Thus, the main conclusion of the study is that it
depends on a number of factors simultaneously, which dominance style will be
favoured. Therefore, this model suggests that group composition and details of what
happens in fights between groups should be studied in order to increase our
understanding of the origination of dominance style.
22.3.4 Distribution of Affiliation, Grooming
Grooming (to clean the pelage of dirt and parasites) is supposed to be altruistic and
therefore it should be rewarded in return (Trivers 1971). Using the theory of social
exchange, Seyfarth (1977) argues that female primates try to exchange grooming
for receipt of support in fights. For this, they direct the supposedly altruistic
grooming behaviour more towards higher- than towards lower-ranking partners.
As a consequence of this supposed competition for grooming partners of high rank,
and by being defeated by dominant competitors, females will end up grooming
close-ranking partners most frequently and are groomed themselves most often by
females ranking just below them. According to him, this explains the following
grooming patterns that are apparent among female primates: (a) high ranking
individuals receive more grooming than others and (b) most grooming takes place
between individuals that are adjacent in rank.
DomWorld presents a simpler alternative for the explanation by Seyfarth
(Hemelrijk 2002b) in which a mental mechanism of social exchange for future
social benefits is absent, for newer updates see Puga-Gonzalez et al. (2009),
Hemelrijk and Puga-Gonzalez (2012). This is important, because it is doubtful
whether grooming involves any real costs at all (see Wilkinson 1988).
The core of the argument is that individuals more often groom those whom they
encounter more frequently. In the case of spatial structure with dominants in the
centre and individuals in closer proximity to those that are closer in dominance (as
described for primates (e.g. see Itani 1954)), these patterns will follow automati-
cally. Individuals more often groom those that are nearby in rank (as found by
Seyfarth 1977). Further, because dominants are more often in the centre, they are
frequently surrounded on all sides by others. Subordinates, however, are at the edge
of the group, and therefore have nobody on one side. Consequently, dominants
simply meet others more frequently than subordinates do (as is shown in
DomWorld Hemelrijk 2000). Therefore, dominants are more often involved in

grooming than subordinates (as found by Seyfarth).
In support of the model, grooming patterns in despotic macaques appear to be
more dominance oriented, and in egalitarian species grooming is more often
directed at anyone (Thierry et al. 1990). To establish the relevance of this model-
based hypothesis for real primates, it should be tested further in more species and
groups, whether or not the patterns of grooming among individuals of similar rank
and of the receipt of grooming by individuals of higher rank occur especially in
groups with centrally located dominants, and less so in those with a weak spatial
structure.
22.3.5 Dominance Relationships Between the Sexes
Most primate species live in bi-sexual groups. Apart from the order of
Lemuriformes, males are usually dominant. However, variation occurs and even
if males are larger than females, females are sometimes dominant over a number of
them (Smuts 1987). To study whether this may arise through chance and the self-
reinforcing effects of dominance interactions, the sexes are modelled in
DomWorld. When, for the sake of simplicity, the sexes in DomWorld are distin-
guished only in terms of an inferior fighting capacity of females as compared to that
of males, then, surprisingly, males appear to become less dominant over females at
a high intensity of aggression than at a low intensity (Hemelrijk 1999b; Hemelrijk
et al. 2003). This is due to the stronger hierarchical differentiation, which causes the
hierarchy to develop in a more pronounced way. This implies that the hierarchy is
also more strongly developed for each sex separately. Since this differentiation
increases the overlap in dominance between the sexes, it causes more females to be
dominant over males than in the case of a hierarchy that has been less developed
(Fig. 22.7a, b).
Similarly to females, adolescent males of despotic macaques have more ease
than those of egalitarian species in outranking adult females (Thierry 1990a).
Thierry explains this as a consequence of the stronger cooperation to suppress
males among related females of despotic macaques than of egalitarian ones, and
van Schaik (1989) emphasizes the greater benefits associated with coalitions for
females of despotic species than egalitarian ones. However, DomWorld explains
greater female dominance, as we have seen, simply as a side-effect of more
pronounced hierarchical differentiation.
Differences in overlap of ranks between the sexes may affect sexual behavior.
Males of certain species, such as Bonnet macaques, have difficulty in mating with
females to which they are subordinate (Rosenblum and Nadler 1971). Therefore,
following the model, we expect that despotic females (because they outrank more
males) have fewer males to mate with than egalitarian ones. In line with this, in
macaques despotic females are observed to mate with fewer partners and almost
exclusively with males of the highest ranks (Caldecott 1986); this observation is
600 C. Hemelrijk
40
a Low intensity female
male
Dominance values
30
20
0.8 c
Female dominance over males

10
0.6
0
40
b High intensity
0.4
Dominance values
30
0.2
20
0.0
M. tonkeana
M. tonkeana
M. mulatta
M. mulatta
M. mulatta
M. mulatta
M. mulatta
M. mulatta
M. mulatta
M. arctoides
M. arctoides
M. arctoides
M. arctoides
M. fuscata
M. assamensis
M. thibetana
M. thibetana
M. thibetana
M. nemestrina
M. fascicularis
M. fascicularis
M. fascicularis
10
0
0 60 120 180 240
Time units
Fig. 22.7 Female dominance over males or degree of rank overlap: (a, b) female dominance over
males over time and aggression intensity (Hemelrijk 1999b), (c) degree of rank overlap of females
with males (i.e. female dominance over males) in groups of egalitarian macaques (in black) and in
groups of despotic macaques (in grey) (Hemelrijk et al. 2008)
attributed to the evolution of a more pronounced female preference in despotic than

in egalitarian species of macaques. The explanation derived from the model,
however, is simpler. In a subsequent study, in support of this hypothesis, the relative
dominance position of both sexes in egalitarian and despotic macaque species
indeed appeared to differ in macaques as expected: females of despotic species
were dominant over a significantly higher percentage of the males in their group
than females of egalitarian species (Fig. 22.7c).
In a similar way as high aggression intensity, a high frequency of aggression in
the model also results in more female dominance. A higher frequency of aggression
in the model can be obtained by increasing the SearchAngle over which individuals
search for others (Fig. 22.4). Due to a greater SearchAngle they return to the others
sooner, the group becomes denser and thus the frequency of aggression is higher. A
difference in group density may be of relevance to the difference in female
dominance observed between common chimpanzees and bonobos (also known as
pygmy chimpanzees (Stanford 1998)). Despite their similar sexual dimorphism,
female dominance in pygmy chimpanzees is higher than in common chimpanzees.
This is usually attributed to more intensive coalition formation among pygmy
females against males. However, in line with DomWorld, we may also explain it
as a side-effect of the difference in density between both species (Hemelrijk 2002a,
2003). Density is high in groups of pygmy chimpanzees. Due to the higher density
there is a higher frequency of aggression and according to DomWorld this may
result in more female dominance over males. This hypothesis should be tested by
a Hierarchy #Males 7 b asyn 13
below Female synchronous
Female dominance to males

6
asyn 52
1+2 = 3
5
asyn 5
4 no attraction
2
3
2
0 +
1
U=5
0
Time units
50 c 50 d
asynchronous asynchronous
#MM interactions
#MF interactions
40 40
30 30
20 20
10 10
0 0
no syn asyn5 asyn13 asyn52 no syn asyn5 asyn13 asyn52
Attraction Attraction
Fig. 22.8 Female dominance over males and sexual attraction. (a) Measurement of female
dominance; (b) female dominance over males over time without attraction (control) and when
attracted to females that cycle synchronously and asynchronously; (c) number of interactions
between sexes with(out) attraction during (a) synchronously cycling; (d) number of interactions
among males with(out) attraction during (a) synchronously cycling. Asyn asynchronous cycles,
syn synchronous cycling; 5, 13, 52 are duration of estrus period
comparing different groups of bonobos and by studying the relationship between

female dominance and frequency of aggression.
Sexual attraction in real animals is usually thought to be accompanied by
strategies of exchange. For instance, chimpanzee males are described as exchang-
ing sex for food with females (Goodall 1986; Stanford 1996; Tutin 1980). Yet, in
spite of detailed statistical studies, we have found no evidence that males obtain
more copulations with, or more offspring from those females with whom they share
their food more often (Hemelrijk et al. 1992, 1999, 2001; Meier et al. 2000). Male
tolerance of females seems to increase during the females’ period of estrus even
without noticeable benefits. Thus, we need another explanation for male tolerance
of females. DomWorld provides us with such an alternative hypothesis. ‘Sexual
attraction’ of males to females is implemented in such a way that males have a
greater inclination to approach females than individuals of their own sex. In the
model (and in the preceding models and empirical studies of Fig. 22.7), we measure
the relative dominance position of females compared to males by counting the
number of males ranking below each female and calculating this figure
(Mann–Whitney U-value, Fig. 22.8a). It appears that this value of relative female
602 C. Hemelrijk
dominance to males increases with sexual attraction as an automatic consequence

of the more frequent encounters between the sexes (Fig. 22.8b, synchronous). This
result is in line with the observation that female dominance in chimpanzees
increases when males are sexually attracted to the females (Yerkes 1940). The
question of whether female dominance over males also increases during sexual
attraction in other species should be studied in the future.
Whereas the examples mentioned above concern species in which females are
synchronously sexually attractive (tumescent), in other species they cycle asyn-
chronously. In the model, however, female dominance over males is relatively
similar regardless of whether they are attractive synchronously or asynchronously
(Fig. 22.8b, synchronous, asyn). The process leading to increased female domi-
nance differs, however, for the two conditions. If single females are attractive in
turn, many males cluster close to a single female. Consequently, in contrast to
synchronous tumescence, the frequency of interaction between the sexes remains
similar to that when females are not attractive to males, but the frequency of male-
male interactions is increased markedly (Fig. 22.8c, d). Due to the higher frequency
of interactions among males, the differentiation of the male hierarchy is stronger
than without attraction and this causes certain males to become subordinate to some
females (Fig. 22.8b, asyn).
Furthermore, the adult sex ratio (or percentage of males) in the group influences
the relative dominance of females compared to that of males (Hemelrijk et al.
2008). Female dominance appears to be higher when there are more males in the
group (Fig. 22.9a). This arises from a shift in the relative number of intra- and inter-
sexual interactions. A higher proportion of males causes both sexes to interact more
often with males. Due to the males’ higher intensity of aggression, this causes a
greater differentiation of the dominance values of both females and males. Conse-
quently, at a high intensity of aggression, the hierarchy of females overlaps more
with that of males and thus, the dominance position of females is higher in relation
to males than if there are fewer males in the group. Subsequent analysis of these
patterns in real primates has confirmed that female dominance increases with a
higher percentage of males in the group (Fig. 22.9b). It appeared that in line with
the preceding modelling results, in groups of a despotic species (rhesus macaques) a
higher percentage of males appeared to be correlated with greater female domi-
nance over them, whereas such a correlation was absent among groups of an
egalitarian species (stump-tailed macaques).
22.3.6 Strategies of Attack
When real animals are brought together for the first time, they perform dominance
interactions only during a limited period. This has been empirically established in
several animal species, e.g. chickens (Guhl 1968) and primates (Kummer 1974).
The interpretation is that individuals fight to reduce the ambiguity of their
relationships (Pagel and Dawkins 1997); once these are clear, energy should be
saved. On the other hand, it has also been suggested that individuals should
0.25 a 1.0 b HapGri (N) EulMac (F)
Female dominance over males

0.20 0.8 EulFul (N)
Female dominance
CepApe (N) VarVar (N)

0.15 0.6
PanPan (C)
0.10 0.4
SemEnt (N) MacThi (N)
MacNem (N) MacAss (N)

0.05 0.2 SaiSci (C)
CalJac (N)
MacTon CerAet (N)
MacMul (C) PanTro (C)
MacFas (N) AloPal (F) MacArc (F)
MacFus (C) ManSph (C)
0.00 0.0
0 20 40 60 80 100 0.0 0.1GorGor 0.2 0.3 0.4 0.5 0.6 0.7
(C)
% Males (in groups of 10) Percentage of males
Fig. 22.9 Percentage of males in the group and female dominance over males. (a) Model (average
and SE), (b) real data of primates. 6-letter codes indicate species. Environmental conditions:
N natural, F free-ranging, C captive condition
continuously strive for a higher ranking and therefore always attack, unless an
opponent is clearly believed to be superior (e.g. see Datta and Beauchamp 1991).
In the DomWorld model, we compare these popular ethological views with each
other and a control strategy, in which individuals invariably attack others upon
meeting them (this is called the ‘Obligate’ attack strategy). Here, the ‘Ambiguity-
Reducing’ strategy is a symmetrical rule in which individuals are more likely to
attack opponents that are closer in rank to themselves. In the so-called ‘Risk
Sensitive’ strategy, the probability of an attack is higher when the opponent is of
a lower rank (Hemelrijk 1998). Remarkably, it appears that, with time, the fre-
quency of aggression decreases in all three attack strategies, at least when groups
are cohesive and the intensity of aggression is sufficiently high (Fig. 22.10).
This decrease of aggression is a direct consequence of the ‘Ambiguity-Reducing’
strategy, but unexpectedly it also develops in the other two. Because of the high
intensity of aggression each interaction has a strong impact on the rank of
both partners, and thus, a steep hierarchy develops. This automatically implies that
some individuals become permanent losers and that, by fleeing repeatedly, they move
further and further away from others (bottom row in Fig. 22.10). The increased
distance among individuals in turn results in a decrease of the frequency of encounters
and hence aggression. This provides a test for the real world: it has to be examined
whether the development of the dominance hierarchy is accompanied not only by
a reduction of aggression but also by an increase in inter-individual distances.
22.3.7 Personality Types
Two of these attack strategies, i.e. ‘risk-sensitive’ and ‘obligate’ attack (Hemelrijk
2000) resemble the attack strategies of individuals of different personality types,
604 C. Hemelrijk
Ambiguity reducing Obligatory attacking Risk sensitive

75 85 46
65 42
75
Frequency of attack
55
38
45 65
34
35
55
30
25
45
15 26
5 35 22
15 18
17
14 17
16
16
13
Mean distance
15 15
12
14 14
11
13 13
10
12
12
9 11
11
8 10
-20 40 100 160 220 280 -20 40 100 160 220 280 -20 40 100 160 220 280
Time units Time units Time units
Fig. 22.10 Development of frequency of aggressive interactions (top half) and mean distance
(lower half) among individuals for different attack strategies and intensities of aggression (loga-
rithmic line fitting). Open circles represent StepDom of 0.1,closed blocks of 1.0
namely those of the cautious and the bold personality, respectively (Koolhaas
2001). When groups in the model consist of both types (mixed groups), the
differentiation of dominance values appears to be greater among individuals that
are attacking obligatorily than risk-sensitively due to their higher frequency of
attack (Hemelrijk 2000). Consequently, obligate attackers rise very high and
descend very low in the hierarchy (resulting in a bimodal distribution of dominance
values, fat lines in Fig. 22.11a), whereas risk-sensitive attack leads to less variation, a
unimodal distribution of values (fat lines in Fig. 22.11a), and therefore to more
intermediate dominance positions. Further, among risk-sensitive individuals, the
average dominance is slightly higher than among those that always attack (obligato-
rily). This is due to higher ‘intelligence’ of the risk-sensitive attack-strategy, because
these individuals will attack, especially when the risk of losing the fight is minimal.
This resembles the distribution of dominance in mixed groups of a bird species,
the great tits (Verbeek et al. 1999). Here, bold individuals become very high up in the
dominance hierarchy or descend very low, whereas cautious individuals have inter-
mediate ranks that on average are above those of bold individuals. Differences in high
and low rank were explained by different stages in which the individuals were. With
regard to the moulting of their feathers and to a difference in tendency to attack from
a familiar territory or an unfamiliar one and to a difference in speed of recovery from
defeats. DomWorld shows that there is no need to add causes based on different
stages of moulting, or on familiarity to a territory or to different speed of recovery.
Thus, the model produces a far simpler explanation for the distribution of dominance
a b
CV
60 RS
0.8
OBL
50
0.7
Dominance
40
30 0.6
20 0.5
10
0.4
0
0 50 100 150 200 250 0 3.33 16.67 50 83.33 96.67 100
Time 6.67 33.33 66.67 93.33
% OBL in groups of 30 (rest: RS)
Fig. 22.11 Dominance distribution and personality types: (a) example of hierarchical develop-
ment of mixed group (fat lines: obligate attackers, dotted lines: risk-sensitive individuals).
N ¼ 10, 5 of each type. (b) hierarchical differentiation in mixed groups with different ratios of
obligately attacking (OBL) and risk-sensitive (RS) individuals. CV Mean coefficient of variation
of DomValues. Box ¼ S.E., whiskers ¼ S.D
values in these groups of great tits. To verify these results empirically, differences in
risk sensitivity of both types of personality need to be confirmed empirically.
Secondly, the model provides us with an alternative explanation for the
associations between dominance behavior and personality style in great tits found
by Dingemanse and de Goede (2004). This association appears to differ among
individuals that own a territory and those that do not; whereas among territory
owners bolds were dominant over cautious ones, the reverse held for those without
a territory. To explain this the authors use a context-specific argument in which they
need an additional trait, namely speed of recovery from defeat (Carere et al. 2001).
They argue that particularly among those individuals without a territory, bolds have
more difficulty in recovering from defeat than cautious ones and that therefore, they
become low in rank, whereas territory-owners do not suffer this setback and,
therefore, they become higher in rank.
Alternatively, a simpler explanation, in line with our model, may apply. Here,
we start from existing dominance relationships and suppose that these exert a
decisive influence on the question who will obtain a territory (instead of the other
way around). We assume that, because territories are limited in numbers, the
higher-ranking individuals (say the top half of them) will acquire them, whereas
individuals in the lower part of the hierarchy are unable to get one. Due to the bi-
modal distribution of dominance values among the bold birds, and the uni-modal
distribution of the cautious ones, the most extreme dominance positions in the
colony will be occupied by bold-ones, and the cautious are located in the middle of
the hierarchy. Thus, among individuals in the top half of the hierarchy (the territory-
owners) the bolds will rank above the cautious, whereas in the bottom half of the
hierarchy, namely among the individuals without a territory, the reverse is true
(Fig. 22.11a). For this explanation to be proven correct, we must verify whether
territory owners belong to the upper half of the dominance hierarchy or not.
606 C. Hemelrijk
Thus, DomWorld produces new explanations for dominance relationships of these

‘personality styles’ in great tits.
An important question regarding personality is how different types, bold and
cautious may co-exist, and why one type did not take over. Although there are a
number of explanations for various species, none of them applies to primates. Since
in primates group survival and individual survival depend on competition within
and between groups, Wantia and Hemelrijk (Wantia 2007) have studied the two
personality types in these contexts. They have found that risk-sensitive individuals
out-competed obligate-attackers in fights within groups, but that in fights between
groups the obligate attackers did better: the higher the percentage of individuals that
attacked obligatorily in fights between groups, the greater the chance of the group
winning (Fig. 22.11b). The better performance within groups of risk-sensitive
individuals was due to their more cautious and deliberate strategy: to attack when
the chance of winning was high. Greater success by obligate attackers in fights
between groups was a consequence of the higher dominance value of the highest
ranking individuals in groups with more obligate attackers. This is due to the
steeper hierarchy as a consequence of the higher frequency of aggression in groups
with more individuals that carry out obligatory attacks. Thus, whereas risk-sensitive
individuals out-compete obligate-attackers in conflicts within groups, the reverse
happens in conflicts between groups. Since competition within and between groups
is essential for primate societies (van Schaik and van Hooff 1983), and the success
of both attack strategies depends on these contexts, we may imagine that a similar
differential performance may contribute to the co-existence of bold and cautious
primates.
22.3.8 Distribution of Tasks
It seems a miracle that a colony of social insects consisting of tens of thousands of

individuals is able to cope with the huge socio-economic demands of foraging,
building, cleaning nests and nursing the brood. It appears that group-members are
somehow able to divide the work efficiently among them. Such a division of labour
is flexible, i.e. the ratio of workers performing different tasks varies according to
changes in the needs and circumstances of the colony. There are several ways in
which this task-division may arise. Different mechanisms may operate in different
species (for a review, see Beshers and Fewell 2001). Task-division may be based on
a genetic difference in predisposition (e.g. Moritz et al. 1996; Robinson 1998;
Robinson and Page 1988), or a response-threshold to perform certain tasks
(Bonabeau et al. 1996b). Such a threshold may be combined with a self-reinforcing
learning process (Gautrais et al. 2002; Theraulaz et al. 1998). Thus, after
performing, the threshold is lessened (by learning) and after a long period of no
involvement in the task, it increases by forgetting.
The execution of tasks may also be a consequence of dominance values, as is
shown in a model of bumblebees (Bombus terrestris) developed by Hogeweg and
Hesper (1983, 1985). This was based on an earlier experimental study (van Honk
and Hogeweg 1981) that showed that during the growth of the colony, workers
develop into two types, the low-ranking, so-called ‘common’, and the high-ranking,
so-called ‘elite’, workers. The activities carried out by the two types differ notice-
ably: Whereas the ‘common’ workers mainly forage and take rest, the ‘elite’
workers are more active, feed the brood, interact with each other and with the
queen, and sometimes lay eggs. In their study of the minimal conditions needed for
the formation of the two types of workers (on the assumption that all workers are
identical when hatching), Hogeweg and Hesper (1983, 1985) used an individual-
based model based on biological data concerning the time of the development of
eggs, larvae, pupae, etc. Space in the model is divided into two parts, peripheral
(where inactive common workers doze for part of the time) and central (where the
brood is, and all interactions take place). The rules of the artificial, adult
bumblebees operate ‘locally’ in so far as their behaviour is triggered by what
they encounter. What they encounter in the model is chosen randomly from what
is available in the space in which the bumblebee finds itself. For instance, if an adult
bumblebee meets a larva, it feeds it, if it meets a pupa of the proper age, it starts
building a cell in which a new egg can be laid, etc. All workers start with the same
dominance value after hatching, with only the queen gifted with a much higher
dominance rank. When an adult meets another, a dominance interaction takes place,
the outcome of which (victory or defeat) is self-reinforcing. Dominance values of
the artificial bumblebees influence almost all their behavioural activities (for
instance, individuals of low rank are more likely to forage).
This model automatically, and unexpectedly, generates two stable classes, those
of ‘commons’ (low-ranking) and ‘elites’ (high-ranking) with their typical conduct.
This differentiation only occurs if the nest is divided into a centre and a periphery
(as in real nests).
The flexibility of the distribution of tasks among individuals manifests when we
take half the work force out of the model. In line with observations in similar
experiments with real bumblebees, this reduction in work force causes the
remaining ones to take over the work. In the model this arises because the decreased
number of workers reduces the frequency of encounters among them and increases
encounters between them and the brood (which has not been reduced in number).
An increased rate of encounters with brood induces workers to collect food more
frequently. Therefore, workers are absent from the nest more often and, conse-
quently, they meet other workers less frequently.
In real bumblebees the queen switches from producing sterile female offspring
to fertile offspring (males and females) at the end of the season. Note that whereas
females are produced by fertilised eggs, males are produced from unfertilised eggs.
Usually females will develop into sterile workers, but if they are fed extra
‘queenfood’ during larval development, they will develop into queens. The switch
seems to take place at an optimal moment, because it occurs at the time when the
colony is at its largest and can take care of the largest number of offspring. Oster
and Wilson (1978) point out that it is difficult to think of a possible external signal
that could trigger such a switch, because it takes 3 weeks to raise fertile offspring
and during these 3 weeks there must still be enough food.
608 C. Hemelrijk
Hogeweg and Hesper (1983) discovered that no such external signal is needed in
their bumble bee model, but that the switch originates automatically as if scheduled
by a socially regulated ‘clock’; it arises from the interaction between colony growth
and stress development of the queen as follows. During the development of the
colony, the queen produces a certain pheromone that inhibits the extra feeding
behaviour of larvae by ‘elite’ workers (that lead to queens) and worker-ovipositions
(i.e. unfertilised male eggs). Just before she is killed, she can no longer suppress the
‘elite’ workers from feeding larvae to become queens and from laying drone-eggs,
because the colony has grown too large. Consequently, individual workers meet the
queen less often and are less subjected to the dominance of the queen, so they start
to lay unfertilised (drone) eggs. Furthermore, the stress on the queen increases
whenever she has to perform dominance interactions with workers during her egg
laying. When the stress on the queen has reached a certain threshold value she
switches to producing male eggs (drones).
Because generative offspring are also sometimes found in small nests, van der
Blom (1986) challenges the notion that dominance interactions induce stress in the
queen and thus lead to this switch. However, in the model, it is not the number of
workers that causes the switch: Hogeweg and Hesper (1985) have shown that in
small nests the switch also appears to occur at the same time in the season as in large
nests. For this they studied the bumblebee model for a reduced speed of growth.
They found that the switch occurs at a similar moment due to the following
complicated feedback process. If the colony grows faster, the heavy duties of caring
for the brood leave the workers little time to interact with each other and the queen.
Consequently, the dominance hierarchy among workers only develops weakly.
Therefore, if workers interact with the queen they do not pose much of a threat to
her and as a result, the queen is not severely stressed and the colony can grow very
large. In contrast, in a slowly growing colony, the small number of brood gives the
workers little work and leaves them time to interact with each other. Consequently,
their dominance relationships are clearly differentiated. Furthermore, by often
interacting with the queen, they become on average higher-ranking themselves.
In the end, the queen receives as much stress from frequent interactions with a few
high-ranking workers in a slowly growing nest as from few interactions with very
many low-ranking workers in a fast growing nest. Consequently, the switch to
reproductive offspring takes place at about the same moment in both nests in the
model.
22.4 Evaluation
The power of these kinds of models is the generation of phenomena that are
emergent. These emergent phenomena lead to explanations that are more parsimo-
nious than usual, because the patterns emerge from the interaction among
individuals and their environment rather than from the cognition of an individual.
These explanations can be tested empirically.
Considering the kind of questions posed in the models discussed above, it

becomes clear that most of them can only be studied using certain kinds of agent-
based models. Other kinds of models, such as partial differential equations based on
density functions or even individual-based models of fluids and gases cannot
incorporate the complexity of the rules and/or emergent effects.
The behavioural rules of the agents in the agent-based models that were treated
here were in all cases biologically inspired. In some cases, behavioural rules were
based on precise experimental results specific to a certain species, such as in the
case of group formation (Camazine et al. 2001; Jeanson et al. 2005). Usually,
however, parameters were tuned only loosely to the real system (e.g. the angle of
vision in fish is set at about 270 and that of mammals at 120 ). Sometimes,
mathematical equations were used that were developed for a different species,
such as in the case of dominance interactions. Here, the equations were initially
derived for the self-reinforcing effects in bumblebees (van Honk and Hogeweg
1981) and subsequently extended and applied to explain social systems in primates
(Hemelrijk 1999b). In all cases the behavioural rules are supposed to capture the
essentials of those of real animals.
If, however, macro patterns are obtained that resemble patterns observed in the
real system, this is still no proof of the correctness of the rules in reflecting
behavioural rules of real animals, because different sets of rules may lead to the
same macro-pattern. The relevance of the rules and other aspects of the model could
be studied further by investigating additional hypotheses and comparing further
results of the model to natural systems.
Agent-based models appear to be particularly useful in showing the
consequences of interactions among individuals for social structure and the reverse,
i.e. from social structure to social interactions. In this way, they teach us about the
integration of traits. This is particularly apparent when the models include a
representation of space (time is always represented). The representation of time
and space is of general importance, because all real individuals live in time and
space and thus it is unrealistic to ignore these aspects. It appears that by
representing the interaction of individuals with their social and physical environ-
ment in time and space, patterns result in such a way that explanations become
unusually parsimonious in terms of their specific adaptive traits and the cognition
needed (in line with findings of situated cognition Pfeifer and Scheier 1999).
These explanations give a new insight into the phenomenon of social complexity
that may either be very general (e.g. conceptual model) or specific to the type of
animal and the kind of behaviour under investigation. Usually, it leads to
hypotheses that are easily tested in the real system. These hypotheses make
unexpected connections between variables, such as between the production of
sterile and reproductively-active offspring with colony dynamics in bumblebees.
Testing can be done by natural observation, for example, by studying the effects of
different group sizes on group shape in case of fish. Furthermore, testing can be
done by experimental interventions, such as by putting together individuals of both
sexes in different group compositions to study effects on relative dominance of
females to males.
610 C. Hemelrijk
While ignoring models of certain topic areas where no emergent effects

appeared (so-called output-oriented models), we have here surveyed a large number
of the most important models of social complexity. Of course, it is impossible to
discuss them all. Note that we have not treated models concerning biological
evolution since only a few are related to social systems (such as Kunz et al. 2006;
Oboshi et al. 2003; Ulbrich et al. 1996). This is due to the fact that evolution usually
happens on a far larger time scale than phenomena of social complexity. However,
see Chap. 18 in this handbook (Chattoe-Brown and Edmonds 2013) on how
concepts from biological evolution have influenced recent approaches in program-
ming simulation models.
Furthermore, we have not treated a number of other complex phenomena, such
as synchronisation of lighting behaviour of insects, temperature regulation of nests
and the building of nests, because in these cases no agent-based models have been
used, or the behaviour is insufficiently social. These topics are, however, covered
by Camazine and co-authors (Camazine et al. 2001).
22.5 Future Work
With regard to areas that may be important for future work, I suggest (1) models of
self-organisation that help us understand specific social systems better by being
closely tuned to empirical systems, such as fish schools, insect colonies or primate
groups, (2) models that present social behaviour and its ontogenetical development
in greater detail and (3) evolutionary models that include also spatial effects.
Acknowledgments I would like to thank Andreas Flache and Rineke Verbrugge for comments on
an earlier draft. I thank Daan Reid for correcting the English.
Further Reading
For further reading I recommend the book on self-organisation in biological

systems by Camazine and co-authors (Camazine et al. 2001). This forms an
extensive introduction to almost all topics treated above and more (with the
exception of task division and social systems of primates). Furthermore, new
extensions of a number of the models are treated by Sumpter (2006). For teaching
I suggest the well-written book by Resnick (1994). This book has been used in
secondary schools and teaches to think in terms of complexity and self-organisation
in general. For more advanced readers, I recommend “Self-organisation and Evo-
lution of Social Systems” (Hemelrijk 2005). This is an edited book and contains
recent articles on modelling of social systems, including those of humans.
References
Aureli F et al (2008) Fission-fusion dynamics: new research frameworks. Curr Anthropol

49:627–654
Axelsen BE, Anker-Nilssen T, Fossum P, Kvamme C, Nottestad L (2001) Pretty patterns but a
simple strategy: predator–prey interactions between juvenile herring and Atlantic puffins
observed with multibeam sonar. Can J Zool – Rev Canadienne De Zoologie 79:1586–1596
Bernstein IS, Gordon TP (1980) The social component of dominance relationships in rhesus
monkeys (Macaca mulatta). Anim Behav 28:1033–1039
Beshers SN, Fewell JF (2001) Models of division of labour in social insects. Annu Rev Entomol
46:413–440
Blom Jvd (1986) Reproductive dominance within the colony of Bombus terrestris. Behaviour
97:37–49
Bonabeau E, Theraulaz G, Deneubourg J-L (1996a) Mathematical models of self-organizing
hierarchies in animal societies. Bull Math Biol 58:661–717
Bonabeau E, Theraulaz G, Deneubourg J-L (1996b) Quantitative study of the fixed threshold
model for the regulation of division of labour in insect societies. Proc R Soc Lond B
263:1565–1569
Buhl J et al (2006) From disorder to order in marching locusts. Science 312:1402–1406
Bumann D, Krause J (1993) Front individuals lead in shoals of 3-spined sticklebacks
(Gasterosteus-Aculeatus) and Juvenile Roach (Rutilus-Rutilus). Behaviour 125:189–198
Bumann D, Krause J, Rubenstein D (1997) Mortality risk of spatial positions in animal groups: the
danger of being in the front. Behaviour 134:1063–1076
Caldecott JO (1986) Mating patterns, societies and ecogeography of macaques. Anim Behav
34:208–220
Camazine S et al (2001) Self-organization in biological systems. Princeton University Press,
Princeton/Oxford, Princeton studies in complexity
Carere C, Welink D, Drent PJ, Koolhaas JM, Groothuis TGG (2001) Effect of social defeat in a
territorial bird (Parus major) selected for different coping styles. Physiol Behav 73:427–433
Chase ID, Bartelomeo C, Dugatkin LA (1994) Aggressive interactions and inter-contest interval:
how long do winners keep winning? Anim Behav 48:393–400
Cooper MA (2004) Inter-group relationships. In: Thierry B, Singh M, Kaumanns W (eds)
Macaque societies: a model for the study of social organization. Cambridge University
Press, Cambridge
Couzin ID, Krause J, James R, Ruxton GD, Franks NR (2002) Collective memory and spatial
sorting in animal groups. J Theor Biol 218:1–11
Couzin ID, Krause J, Franks NR, Levin SA (2005) Effective leadership and decision-making in
animal groups on the move. Nature 433:513–516
Datta SB, Beauchamp G (1991) Effects of group demography on dominance relationships among
female primates. 1. Mother-daughter and sister-sister relations. Am Nat 138:201–226
Deneubourg JL, Goss S (1989) Collective patterns and decision-making. Ethol Ecol Evol
1:295–311
Deneubourg JL, Goss S, Franks N, Pasteels JM (1998) The blind leading the blind: modelling
chemically mediated army ant raid patterns. J Insect Behav 2:719–725
Dingemanse NJ, De Goede P (2004) The relation between dominance and exploratory behavior is
context-dependent in wild great tits. Behav Ecol 15:1023–1030
Drews C (1993) The concept and definition of dominance in animal behaviour. Behaviour
125:283–313
Ellis L (1991) A biosocial theory of social stratification derived from the concepts of pro/
antisociality and r/K selection. Polit Life Sci 10:5–44
Enquist BJ, Niklas KJ (2001) Invariant scaling relations across tree-dominated communities.
Nature 410:655–660
612 C. Hemelrijk
Enquist BJ, West GB, Charnov EL, Brown JH (1999) Allometric scaling of production and life-
history variation in vascular plants. Nature 401:907–911
Franks NR, Gomez N, Goss S, Deneubourg J-L (1991) The blind leading the blind in army ant
raiding patterns: testing a model of self-organization (Hymenopter: Formicidae). J Insect
Behav 4:583–607
Gautrais J, Theraulaz G, Deneubourg JL, Anderson C (2002) Emergent polyethism as a conse-
quence of increased colony size in insect societies. J Theor Biol 215:363–373
Goldberg TL, Wrangham RW (1997) Genetic correlates of social behaviour in wild chimpanzees:
evidence from mitochondrial DNA. Anim Behav 54:559–570
Goodall J (1986) The chimpanzees of Gombe: patterns of behaviour. Belknapp Press of Harvard
University Press, Cambridge MA/London
Gueron S, Levin SA, Rubenstein DI (1996) The dynamics of herds: from individuals to
aggregations. J Theor Biol 182:85–98
Guhl AM (1968) Social inertia and social stability in chickens. Anim Behav 16:219–232
Hamilton WD (1971) Geometry for the selfish herd. J Theor Biol 31:295–311
Hemelrijk CK (1996) Dominance interactions, spatial dynamics and emergent reciprocity in a
virtual world. In: Maes P, Mataric MJ, Meyer J-A, Pollack J, Wilson SW (eds) Proceedings of
the fourth international conference on simulation of adaptive behavior, vol 4. The MIT Press,
Cambridge, MA, pp 545–552
Hemelrijk CK (1998) Risk sensitive and ambiguity reducing dominance interactions in a virtual
laboratory. In: Pfeifer R, Blumberg B, Meyer J-A, Wilson SW (eds) From animals to animats 5,
Proceedings of the fifth international conference on simulation on adaptive behavior, Zurich,
17–21 Aug 1998. MIT Press, Cambridge, MA, pp 255–262
Hemelrijk CK (1999) An individual-oriented model on the emergence of despotic and egalitarian
societies. Proc Royal Soc Lond B Biol Sci 266:361–369
Hemelrijk CK (1999a) Effects of cohesiveness on intersexual dominance relationships and spatial
structure among group-living virtual entities. In: Floreano D, Nicoud J-D, Mondada F (eds)
Advances in artificial life. Fifth European conference on artificial life (Lecture notes in
artificial intelligence, 1674). Springer, Berlin, pp 524–534
Hemelrijk CK (1999b) An individual-oriented model on the emergence of despotic and egalitarian
societies. Proc R Soc Lond B Biol Sci 266:361–369
Hemelrijk CK (2000) Towards the integration of social dominance and spatial structure. Anim
Behav 59:1035–1048
Hemelrijk CK (2002a) Self-organising properties of primate social behaviour: a hypothesis on
intersexual rank-overlap in chimpanzees and bonobos. In: Soligo C, Anzenberger G, Martin
RD (eds) Primatology and anthropology: into the third millennium. Wiley, New York,
pp 91–94
Hemelrijk CK (2002b) Understanding social behaviour with the help of complexity science.
Ethology 108:655–671
Hemelrijk CK (2003) Female co-dominance in a virtual world: ecological, cognitive, social and
sexual causes. Behaviour 140:1247–1273
Hemelrijk CK (ed) (2005) Self-organisation and evolution of social systems. Cambridge Univer-
sity Press, Cambridge
Hemelrijk CK, Hildenbrandt H (2008) Self-organized shape and frontal density of fish schools.
Ethology 114:245–254
Hemelrijk CK, Kunz H (2005) Density distribution and size sorting in fish schools: an individual-
based model. Behav Ecol 16:178–187
Hemelrijk CK, Wantia J (2005) Individual variation by self-organisation: a model. Neurosci
Biobehav Rev 29:125–136
Hemelrijk CK, van Laere GJ, van Hooff J (1992) Sexual exchange relationships in captive
chimpanzees? Behav Ecol Sociobiol 30:269–275
Hemelrijk CK, Meier CM, Martin RD (1999) ‘Friendship’ for fitness in chimpanzees? Anim
Behav 58:1223–1229
Hemelrijk CK, Meier CM, Martin RD (2001) Social positive behaviour for reproductive benefits in
primates? A response to comments by Stopka et al. (2001). Anim Behav 61:F22–F24
Hemelrijk CK, Wantia J, Daetwyler M (2003) Female co-dominance in a virtual world: ecological,
cognitive, social and sexual causes. Behaviour 140:1247–1273
Hemelrijk CK, Wantia J, Isler K (2008) Female dominance over males in primates: self-
organisation and sexual dimorphism. PLoS One 3, e2678
Hemelrijk CK, Hildenbrandt H, Reinders J, Stamhuis EJ (2010) Emergence of oblong school
shape: models and empirical data of fish. Ethology 116:1099–1112
Hemelrijk CK, Hildenbrandt H (2011) Some causes of the variable shape of flocks of birds. PLoS
One 6:e22479. doi:10.1371/journal.pone.0022479
Hemelrijk CK, Hildenbrandt H (2012) Schools of fish and flocks of birds: their shape and internal
structure by self-organization. Interface Focus 2:726–737. doi:10.1098/rsfs.2012.0025
Hemelrijk CK, Puga-Gonzalez I (2012) An individual-oriented model on the emergence of support
in fights, its reciprocation and exchange. PLoS One 7:e37271. doi:10.1371/journal.
pone.0037271
Hildenbrandt H, Carere CC, Hemelrijk CK (2010) Self-organized aerial displays of thousands of
starlings: a model. Behav Ecol 21(6):1349–1359
Hogeweg P, Hesper B (1983) The ontogeny of interaction structure in bumble bee colonies: a
MIRROR model. Behav Ecol Sociobiol 12:271–283
Hogeweg P, Hesper B (1985) Socioinformatic processes: MIRROR modelling methodology.
J Theor Biol 113:311–330
Hsu Y, Earley RL, Wolf LL (2006) Modulation of aggressive behaviour by fighting experience:
mechanisms and contest outcomes. Biol Rev 81:33–74
Huse G, Railsback S, Ferno A (2002) Modelling changes in migration pattern in herring: collective
behaviour and numerical domination. J Fish Biol 60:571–582
Huth A, Wissel C (1992) The simulation of the movement of fish schools. J Theor Biol
156:365–385
Huth A, Wissel C (1994) The simulation of fish schools in comparison with experimental data.
Ecol Model 75(76):135–145
Inada Y, Kawachi K (2002) Order and flexibility in the motion of fish schools. J Theor Biol
214:371–387
Itani J (1954) The monkeys of Mt Takasaki. Kobunsha, Tokyo
Jeanson R et al (2005) Self-organized aggregation in cockroaches. Anim Behav 69:169–180
Klimley AP (1985) Schooling in Sphyrna lewini, a species with low risk of predation: a non-
egalitarian state. Zeitschrift fürTierpsychology 70:279–319
Koolhaas JM (2001) How and why coping systems vary among individuals. In: Broom DM (ed)
Coping with challenge: welfare including humans. Dahlem University Press, Berlin
Koolhaas JM, de Boer SF, Buwalda B, van der Vegt BJ, Carere C, Groothuis AGG (2001) How
and why coping systems vary among individuals. In: Broom DM (ed) Coping with challenge.
Welfare including humans. Dahlem University Press, Berlin
Krause J (1993) The effect of ‘Schreckstoff’ on the shoaling behaviour of the minnow: a test of
Hamilton’s selfish herd theory. Anim Behav 45:1019–1024
Krause J, Tegeder RW (1994) The mechanism of aggregation behaviour in fish shoals: individuals
minimize approach time to neighbours. Anim Behav 48:353–359
Kummer H (1974) Rules of dyad and group formation among captive baboons (theropithecus
gelada). S. Karger, Basel
Kunz H, Hemelrijk CK (2003) Artificial fish schools: collective effects of school size, body size,
and body form. Artif Life 9:237–253
Kunz H, Züblin T, Hemelrijk CK (2006) On prey grouping and predator confusion in artificial fish
schools. In: Rocha LM et al (eds) Proceedings of the tenth international conference of artificial
life. MIT Press, Cambridge, MA, pp 365–371
Lee SH (2006) Predator’s attack-induced phase-like transition in prey flock. Phys Lett A
357:270–274
614 C. Hemelrijk
Mazur A (1985) A biosocial model of status in face-to face primate groups. Soc Forces
64(2):377–402
Meier C, Hemelrijk CK, Martin RD (2000) Paternity determination, genetic characterization and
social correlates in a captive group of chimpanzees (Pan troglodytes). Primates 41:175–183
Moritz RFA, Kryger P, Allsopp MH (1996) Competition for royalty in bees. Nature 384(6604), 31
Nottestad L, Axelsen BE (1999) Herring schooling manoeuvres in response to killer whale attacks.
Can J Zool – Rev Canadienne de Zoologie 77:1540–1546
Oboshi T, Kato S, Mutoh A, Itoh H (2003) Collective or scattering: evolving schooling behaviors
to escape from predator. In: Standish RK, Bedau M, Abbass HA (eds) Artificial life VIII. MIT
Press, Cambridge, MA, pp 386–389
Oster GF, Wilson EO (1978) Caste and ecology in social insects. Princeton University Press,
Princeton
Pagel M, Dawkins MS (1997) Peck orders and group size in laying hens: ‘future contracts’ for non-
aggression. Behav Processes 40:13–25
Parrish JK (1993) Comparison of the hunting behavior of 4 piscine predators attacking schooling
prey. Ethology 95:233–246
Parrish JK, Viscido SV (2005) Traffic rules of fish schools: a review of agent-based approaches. In:
Hemelrijk CK (ed) Self-organisation and the evolution of social behaviour. Cambridge Uni-
versity Press, Cambridge, pp 50–80
Pfeifer R, Scheier C (1999) Understanding intelligence. MIT Press, Cambridge, MA
Pitcher TJ, Wyche CJ (1983) Predator avoidance behaviour of sand-eel schools: why schools
seldom split? In: Noakes DLG, Lindquist BG, Helfman GS, Ward JA (eds) Predators and prey
in fishes. Dr. W. Junk Publications, Dordrecht, pp 193–204
Preuschoft S, Paul A, Kuester J (1998) Dominance styles of female and male Barbary maceques
(Macaca sylvanus). Behaviour 135:731–755
Puga-Gonzalez I, Hildenbrandt H, Hemelrijk CK (2009) Emergent patterns of social affiliation in
primates, a model. PLoS Comput Biol 5:e1000630. doi:10.1371/journal.pcbi.1000630
Ramos-Fernández G, Boyer D, Gómez VP (2006) A complex social structure with fission-fusion
properties can emerge from a simple foraging model. Behav Ecol Sociobiol 60(4):536–549
Resnick M (1994) Turtles, termites and traffic jams: explorations in massively parallel
microworlds. MIT Press, Cambridge, MA
Reynolds CW (1987) Flocks, herds and schools: a distributed behavioral model. Comput Graph
21:25–36
Robinson GE (1998) From society to genes with the honey bee. Am Scientist 86:456–462
Robinson GE, Page RE (1988) Genetic determination of guarding and undertaking in honey-bee
colonies. Nature 333:356–358
Romey WL (1995) Position preferences within groups: do whirligigs select positions which
balance feeding opportunities with predator avoidance? Behav Ecol Sociobiol 37:195–200
Rosenblum LA, Nadler RD (1971) The ontogeny of sexual behavior in male bonnet macaques. In:
Ford DH (ed) Influence of hormones on the nervous system. Karger, Basel, pp 388–400
Seyfarth RM (1977) A model of social grooming among adult female monkeys. J Theor Biol
65:671–698
Smuts BB (1987) Gender, aggression and influence. In: Smuts BB, Cheney DL, Seyfarth RM,
Wrangham RW, Struhsaker TT (eds) Primate societies. Chicago University Press, Chicago,
pp 400–412
Solé RV, Bonabeau E, Delgado J, Fernádez P, Marı́n J (2001) Pattern formation and optimization
in army ant raids. Artif Life 6:219–226
Stanford CB (1996) The hunting ecology of wild chimpanzees: implications for the evolutionary
ecology of Pliocene hominids. Am Anthropol 98:96–113
Stanford CB (1998) The social behaviour of chimpanzees and bonobos. Curr Anthropol
39:399–420
Sumpter DJT (2006) The principles of collective animal behaviour. Philos Trans R Soc B Biol Sci
361:5–22
te Boekhorst IJA, Hogeweg P (1994a) Effects of tree size on travel band formation in orang-utans:
data-analysis suggested by a model. In: Brooks RA, Maes P (eds) Artificial life, vol IV. The
MIT Press, Cambridge, MA, pp 119–129
te Boekhorst IJA, Hogeweg P (1994b) Self structuring in artificial ‘CHIMPS’ offers new
hypotheses for male grouping in chimpanzees. Behaviour 130:229–252
Theraulaz G, Bonabeau E, Deneubourg J-L (1998) Threshold reinforcement and the regulation of
division of labour in social insects. Proc R Soc Lond B 265:327–333
Thierry B (1990a) Feedback loop between kinship and dominance: the macaque model. J Theor
Biol 145:511–521
Thierry B (1990b) The state of equilibrium among agonistic behavior patterns in a group of
Japanese macaques (Macaca fuscata). C R Acad Sci Paris 310:35–40
Thierry B, Gauthier C, Peignot P (1990) Social grooming in Tonkean macaques (Macaca
tonkeana). Int J Primatol 11:357–375
Trivers RL (1971) The evolution of reciprocal altruism. Q Rev Biol 46:35–57
Tutin CEG (1980) Reproductive behaviour of wild chimpanzees in the Gombe National Park.
J Reprod Fertil Suppl 28:43–57
Ulbrich K, Henschel J R, Jeltsch F, Wissel C (1996) Modelling individual variability in a social
spider colony (Stegodyphus dumicola: Eresidae) in relation to food abundance and its alloca-
tion. Revue suisse de Zoologie (Proceedings of the XIIIth international congress of arachnol-
ogy, Geneva, 3–8 Sept 1995), hors série, 1, pp 661–670
van Honk C, Hogeweg P (1981) The ontogeny of the social structure in a captive Bombus terrestris
colony. Behav Ecol Sociobiol 9:111–119
van Schaik CP (1989) The ecology of social relationships amongst female primates. In: Standen V,
Foley GRA (eds) Comparative socioecology, the behavioural ecology of humans and other
mammals. Blackwell, Oxford, pp 195–218
van Schaik CP, van Hooff JARAM (1983) On the ultimate causes of primate social systems.
Behaviour 85:91–117
Verbeek MEM, de Goede P, Drent PJ, Wiepkema PR (1999) Individual behavioural characteristics
and dominance in aviary groups of great tits. Behaviour 136:23–48
Vicsek T, Czirók A, Ben-Jacob E, Cohen I, Shochet O (1995) Novel type of phase transition in a
system of self-driven particles. Phys Rev Lett 75:1226–1229
Wantia J (2007) Self-organised dominance relationships: a model and data of primates. In:
Wiskunde en Natuurwetenschappen. Rijksuniversiteit Groningen, Groningen, p 115
Wilkinson GS (1988) Reciprocal altruism in bats and other mammals. Ethol Sociobiol 9:85–100
Yerkes RM (1940) Social behavior of chimpanzees: dominance between mates in relation to
sexual status. J Comp Psychol 30:147–186
Chapter 23
Agent-Based Simulation as a Useful Tool
for the Study of Markets
Juliette Rouchier
Why Read This Chapter? To understand the various elements that might be
needed in a simulation of a market, including: some options for implementing
learning by agents in a market simulation, the role and kinds of market indicators,
and kinds of buyer-seller interaction (bargaining, loyalty-based, and reputation-
based). The chapter will give you an overview of the complexity of markets,
including multi-good markets and complicated/decentralised supply chains. It will
help you understand financial markets (especially in contrast to markets for tangible
goods) and the double-auction mechanism. Finally, it will give some indications of
how such models either have informed or might inform the design of markets.
Abstract This chapter describes a number of agent-based market models. They can
be seen as belonging to different trends in that different types of markets are presented
(goods markets, with or without stocks, or financial markets with diverse price
mechanisms, or even markets with or without money), but they also represent different
aims that can be achieved with the simulation tool. For example, they show that it is
possible to develop precise interaction processes to include loyalty among actors; try to
mimic as well as possible the behaviour of real humans. After they have been recorded
in experiments; or try to integrate psychological data to show a diffusion process. All
these market models share a deep interest in what is fundamental in agent-based
simulation, such as the role of interaction, inter-individual influence, and learning,
which induces a change in the representation that agents have of their environment.
J. Rouchier (*)
GREQAM, Marseille, France

618 J. Rouchier
23.1 Introduction
In recent years, there has been a growing recognition that an analysis of economic
systems – which includes markets – as being complex could lead to a better
understanding of the participating individuals’ actions. In particular, one element
that could be incorporated in such analyses is the heterogeneity of agents and of
their rationality. For example, the existence of multiple prices on a market for the
same product, sold at the same moment and at the same place, cannot be captured in
an equilibrium model, whereas it appears in real life and can be reproduced easily in
an agent-based model (Axtell 2005).
This issue is at the centre of much debate among economists. In classical
economy, agents are considered as rational, having a perfect knowledge of their
environment, and hence are homogeneous. This view of agents stayed unchallenged
for a while. Friedman (1953) argued, for example, that non-rational agents would
be driven out of the market by rational agents, who would trade against them and
simply earn higher profits. However, in the 1960’s, the view on rationality evolved.
Even Becker (1962) suspected that agents could be irrational and yet produce the
same results as rational agents (i.e. the negative slope of market demand curves).
However, the author who definitely changed the view on economic agents is Simon,
who stated that any individual could be seen as a “bounded rational” agent, which
means that it has an imperfect knowledge and has limited computing abilities
(Simon 1955). In most markets, agents do not have perfect knowledge of the
behaviour and preferences of other agents, which makes them unable to compute
an optimal choice. If they do have perfect knowledge, they will require unlimited
computational capacity in order to calculate their optimal choices.1
Indeed, for some contemporary authors, the understanding that one can get of
real market dynamics is more accurate if the assumption of a representative agent or
of the homogeneity of agents is dropped at the same time as the perfect knowledge
assumption (Kirman 2001). The way bounded rationality is approached can be very
formal and tentatively predictive (Kahneman and Tversky 1979), but some authors
go further by stating that the notion of rationality has to be abandoned. The correct
view is that agents possess rules of behaviours that they select thanks to diverse
mechanisms, which are most of the time based on past evaluation of actions
(Kirman 2001). Some authors also compare the different results one can get from
a representative agent approach and a bounded rationality approach for agents,
and try to integrate both simplicity and complexity of these two points of view
(Hommes 2007).
1
For example, an issue that anyone representing learning (not only on markets) has to face is the
exploration-exploitation dilemma. When an action gives a reward that is considered as “good”, the
agent performing it has to decide either to continue with this action – and hence possibly miss
other, more rewarding actions – or to search for alternative actions, which implies indeterminate
results. Leloup (2002), using the multi-armed bandit (Rothschild 1974) to represent this dilemma,
showed that a non-optimal learning procedure could lead to a better outcome than an optimal – but
non computable – procedure.
23 Agent-Based Simulation as a Useful Tool for the Study of Markets 619
To complete the view that agents have bounded rationality in a market and that
they evolve through time, it is necessary to consider one more aspect of their
belonging to a complex system: their interactions. The seminal works on markets
with interacting heterogeneous agents date back to the beginning of the 1990’s
(e.g. Palmer et al. 1993; Arthur et al. 1997b); many of them were collected in the
post-proceedings of a conference at the Santa Fe Institute (Arthur et al. 1997a).
Since then, a large part of research in agent-based simulation concerns market
situations, as can for example be seen in the cycles of WEHIA/ESHIA conferences
and ESSA conferences. The former ones gather many physicists who apply
dynamic systems techniques to representing heterogeneous interacting agents to
deal with economics issues, and they tend to buy the “agent-based computational
economics” (ACE) approach of simulated markets promoted by Leigh Tesfatsion2
(Phan 2003). In the latter ones, not only economy, but also sociology, renewable
resources management, and computer science participate and often try to generate
subtle representations of cognition and institutions and a strong view of agents as
computerized independent entities to deal with broad social issues and are thus
being closer to multi-agent social simulation (MAS). We will refer to both terms
(ACE or MAS) separately.
Not only the techniques can be different when studying markets with distributed
agents, but also the aims can be. Some try to infer theoretical results about rationality
and collective actions (Vriend 2000) or about market processes (Weisbuch et al.
2000). Others want to create algorithms to represent human rationality on markets,
and try to assess the value of these algorithms by comparing the simulated behaviour
with actions of real humans in order to understand the latter (Hommes and Lux 2008;
Duffy 2001; Arthur 1994). Eventually some explorations about the impact of diverse
rationalities in a market context enable the identification of possible worlds using a
sort of artificial society approach (Rouchier et al. 2001).
Being part of a handbook, this chapter should provide tools to be able to build
and use agent-based simulation techniques to create artificial markets and analyze
results. However, a “know-how” description of the building of artificial market is so
dependent on the type of issue that is addressed, that it was decided here to rather
establish a classification of the type of markets and modelling techniques that can
be found in agent-based simulation research. We are interested in representations of
markets that focus on a micro-representation of decision processes, with limited
information or limited computational abilities for agents and which take place in a
precise communication framework. This form of representation forces authors to
focus on new notions such as interaction or information gathering, and this induces
new issues such as the bases for loyalty or the exploration-exploitation dilemma.
The use of stylized facts is fundamental in this branch, where modellers try to
mimic some elements of the world in their research, focusing either on
representations of reasoning that are inferred from observed micro behaviours, or
2
http://www.econ.iastate.edu/tesfatsi/ace.htm.
620 J. Rouchier
trying to mimic global behaviour through learning algorithms, thereby to stay closer
to common view orthodox economics.
Contrary to this view, which does not distinguish among individual rationalities
and assumes aggregate, centralized knowledge and decision-making, researchers
involved in the use of multi-agent simulation usually try to understand the local
point of view of agents, and its influence on global indicators. Since the agent is
then seen as unable to have complete knowledge, it has to accumulate data about its
environment and treat those data according to its aims. The study of markets is
interesting when it comes to this type of analysis because markets display a much
more simple set of possible actions and motivations than many other social settings.
Income and reproduction of activity are direct motivations, which imply choices in
the short and the long term, prices and quantities are what have to be chosen (as well
as sometimes acquaintances, in the case of bilateral bargaining) and information is
limited to offers and demands, as well as responses to these two types of proposals.
This chapter presents the main fields of application for multi-agent simulation
dealing with markets, to show how researchers have focused on different aspects of
this institution and to conclude on the great interest of agent-based simulation when
trying to understand the very dynamics of these social environments.
In the next section, we will describe the main notions that are covered by the
term “market” in agent-based literature, and also the main ways to represent
rationality and learning that can be found. Subsequently, the main topics of market
studies will be described in three parts. In Sect 23.3, agents are on a market and
actually meet others individually, having private interactions with each other.
Choices that have to be modelled are about matching business partners, as well as
buying or selling decisions. In all other sections, agents are facing an aggregate
market and they have to make decisions based on global data, sometimes associated
to networks. In part 4 agents are either consumers or producers in a large market.
In part 5 we will deal with auctions, financial markets and market design.
23.2 Market and Agents’ Reasoning
Although the study of market(s) has been predominantly carried on by economists,

ethnographers and sociologists have also been active in this domain (Geertz et al.
1979; White 1988), and the field is now being developed through field studies
and analysis. The main difference for those two approaches is that economists
generally build models that are rather abstract and formal, whereas ethnologists and
sociologists describe actual markets after observing them and generally produce
models that are based on classifications of large amount of data. The notion of
market itself has a double meaning, even more important with the increasing use of
internet: it is at the same time the institution that enables individuals to coordinate
their actions through the fixing of a price or, alternatively, a physical place where
buyers meet sellers. There is no easy decision to choose the relevant scale to study
when dealing with market, neither is the limit of observation that is needed in the
supply chain easy to set. In agent-based simulation, markets are represented as

closed societies with a specified set of agents (market being open or closed to new
entry) that are possibly interconnected.
Simulations can be based on very specific case studies in order to describe as
accurately as possible the behaviour of real actors, but they can also be mainly
theoretic in an attempt to generate expected theoretical results. In all cases,
a simulated market cannot be implemented as described in neoclassical economic
theory since agents need to be independently specified and interact directly with
one another during the simulation. For example, to develop a model where individ-
ual agents have to make a decision, a demand curve (that gives for any price of a
product, the number of agents that are ready to buy or sell at this price) cannot be
imposed on the model but has to be derived from the determinants of agent
behaviour. One approach is to distribute reservation prices (the maximum price
for buying or minimum for selling) among agents which can then be used to
aggregate demand and offer curves.
The main elements to define a market model are given in the next section. We
will then describe a few approaches to rationality and learning for agents in markets
that depend on the type of market and goods that are represented. A similar analysis,
more oriented towards consumers’ behaviour can be found in (Jager 2007).
23.2.1 Main Elements to Build an Artificial Market
Several dimensions are important on a market, and each description of an element is

easy to relate to a dimension in the building of an artificial system with agents
(simply put: the market institution and the agents’ rules of behaviour). Axtell (2005)
proposes a very abstract view of decentralized exchange on an agent-based market,
where he gives no explanation of the bargaining process that organizes the exchange,
but shows the existence of a computationally calculable equilibrium to increase all
agents’ utility. Here, the aim is to find out actual processes that can be used by
modellers to represent markets.
The first distinction that can be made when developing a model is to know if one
is building the representation of a speculative market or a goods market. What I call
a speculative market, typically a financial one, is such that agents who have a
commodity can keep it, sell it or buy it. They have to anticipate prices and wait to
perform their actions so as to make the highest profit. Seminal works on agent-
based simulation were related to speculative markets, which display interesting
regularities in their stylized facts. A large body of literature has developed on this
topic, which is also due to the fact that data to calibrate models are more easily
available than for non-speculative markets. On a goods market, agents have only
one role, to sell or buy a certain number of units of one or more products, and they
usually have a reservation price to limit the prices they can accept. The products can
622 J. Rouchier
be perishable (with an intrinsic value that decreases over time) or durable so that
stock management is an issue.
Then, both types of market can be organized either through auctions (with
diverse protocols: double-auction, ascending, descending, with posted prices or
continuous announcements) or via pair-wise interactions which imply face-to-face
negotiation (with many different protocols, such as “take-it-or-leave-it”, one shot
negotiation, a series of offers and counter-offers, the possibility for buyers to
explore several sellers or not). In the design of the protocol, it can also be important
to know if money exists in the system or if agents exchange one good for another
directly.
Agents’ cognition must entail some choice algorithm. Agent-based simulation is
almost always used to design agents with bounded rationality because agents have
limited computational abilities or limited access to information. They have tasks to
perform, within a limited framework, and have to make decisions based on the
context they can perceive. Most of the time, they are given a function, equivalent to
a utility function in economics, that associates a value to each action, enabling the
agents to classify the profit it gets and hence to compare actions. First, an agent
must have constraints in its actions, in order to be able to make arbitration between
all possible options:
– To each commodity is associated a reserve price: if a buyer (resp. seller) goes on
a market, there is a maximum (resp. minimum) price it is willing to pay for the
good.
– The importance of obtaining a commodity can be indicated by the price of entry
to the market. Agents have a greater incentive to buy and get a 0 profit, rather
than not buying. The constraint for selling can be the same.
– In some papers, the price is not represented in the system, and the acquisition of a
product is limited by a utility function, where the agent acquires the product only
if it makes enough profit.
– In the case of negotiation, time constraints are usually put on buyers who can
visit a limited number of sellers, having hence a limit on their search for a
commodity.
– There can be a discount factor: at each period, the risk of seeing the market close
is constant and hence agents never know if they will be able to trade at the next
period.
The type of decisions that agents have to perform on a market:
• For buyers: how many units to buy, who to visit, how to decide to stay in a queue
depending on its length, which price to propose/accept, which product to accept,
and more fundamentally to participate or not.
• For sellers: how to deal with the queue of buyers (first-come-first-served or with
a preference for some buyers), which offer to make or accept, and in the case of
repeated markets how many units to buy for the next time step, in the case of a
market where quality is involved: which type of product to propose or produce
for the next time step.
Due to the increasing complexity when adding another type of decision to a

model, it is rare that all of these decisions will be made in one single model. For
example, although interactions could potentially be represented in a continuous
way, I know of no model where it is the case: all choices and meetings are made and
messages sent at discrete time steps.
23.2.2 Agents’ Learning
As said before, in most markets that are studied with agent-based models, the
central element is that agents are heterogeneous in knowledge as well as in need.
This situation can be decided from initialization or can emerge during the course
of the simulation, while agents learn. Another element that is rarely given at
initialization but is acquired by agents while learning is information about other
agents’ characteristics. In most cases, this learning takes place as a result of action,
at the same time as the acquisition of an object or the acquisition of money.
The way learning is organized is often linked to a performance of actions, with a
selection of actions that “satisfice” or give the best performance. On a market, it is
often assumed that agents are interested in getting the highest payoff for their
individual actions: the performance is either the profit that agents get from their
sells or the utility they get from consuming the product. In most models, learning
agents have a set of pre-defined actions they can take and they have to select the one
they like the best, following a probabilistic choice.
One of the simplest learning models is reinforcement learning (Erev et al. 1998;
Bendor et al. 2001; Macy and Flache 2002; see also Chap. 17 (Macy et al. 2013) in
this volume), where agents attribute a probability of choice for each possible action
that follows a logit function. The algorithm includes a forgetting parameter and a
relative weight attributed to exploitation (going to actions known as having high
value) and exploration (the random choice part in the system). The exploration
parameter can be fixed during the simulation, where the level of randomness has to
be chosen, or can vary during the simulation (decrease) so that there is a lot of
exploration at the beginning of the simulation whereas as time passes agents focus
on the “best” actions. This issue of which level of exploration and exploitation to
put in a learning algorithm is of current concern in the literature about markets
(Moulet and Rouchier 2008).
Another model of rationality for agents is based on a representation of strategies
of other agents, fictitious play: a distribution of past actions is built by each agent and
they can then infer the most probable set of actions of others, and hence choose their
optimal behaviour (Boylan and El-Gamal 1993). The EWA model has been pro-
posed by Camerer (Camerer and Ho 1999) to gather both characteristics of these
models: the agent not only learns what profit it got for each action, but also computes
notional gains for each possible action, and attributes the resulting notional profit to
each of those possible actions.
624 J. Rouchier
A slightly more complex representation of knowledge commonly found in the

literature is the classifier-system where each decision is made by considering a
context, a choice and the past profit made in this precise context by making this
special choice (Moulet and Rouchier 2007). This type of algorithm is very similar to
what (Izquierdo et al. 2004) call case-based learning, but it does not seem to be
applied to market situations. In general the number of possible actions is fixed from
the beginning, but the classifier system can be associated to a genetic algorithm that
generates new rules over the time (Kopel and Dawid 1998). Genetic Algorithm
learning is also a quite usual way to represent learning, where the information that is
used by agents to estimate the profit of each rule can be based on actual past actions
(Vriend 2000) or also on the imaginary profit of all possible actions considering the
actions of others (Hommes and Lux 2008). The presence of other agents can also be
relevant information when agents use imitation or social learning.
Brenner (2006) undertook an extensive review of common learning processes.3
In this paper an interesting element arises: the way “satisficing” rationality can be
developed by having agents not look for the best action but for one which enables
them to get a “good enough” profit; the notion of “good enough”, called “aspiration
level”, can then evolve during the simulation (Cyert and March 1963).
An alternative to learning algorithms that are only based on profit is to consider
that agents have a social utility, which need not have the same dimensionality as
profit. Indeed, it has been demonstrated that translating the social utility in costs or
profits that can be added to monetary profit gives a dynamics of learning and
behaviour that is radically different from a situation where agents reason in two
dimensions, social and monetary, with a lexicographic ordering (Rouchier et al.
2001). A way to implement social utility without lexicographic ordering is to
include in the utility the appreciation of similarity of members of the network,
such as in consumer choice models (Delre et al. 2007).
In most models, action and information gathering are made in one action, and
circulation of information as such is not really modelled. The reason is certainly
because it would take modellers too far from the neoclassical economic approach to
market, where the only information is the observation of transactions, sometimes of
intermediate prices as in auctions or, sometimes, bargaining. One model of a market
by Moss (2002) represents communication among agents before exchange takes
place and Rouchier and Hales (2003) (whose model evolved into the one of
Rouchier (2013)) also allocate one period out of three every time step for agents
to look for information.
3
The main objection to Brenner’s exposition is the lack of homogeneity of notation, which makes
the algorithms difficult to compare and maybe to implement.
23.2.3 Indicators and Method
Several types of modelling can be found in papers about markets, just like in any
other application domain of simulation. Some prefer to work at a purely abstract
level, while others try to fit as well as possible data that they extract from observa-
tion and experience. Whatever the approach, indicators that are often observed in
markets are prices, efficiency (the total profit that is extracted from agents com-
pared to the maximum profit that could be extracted), and relative power of
different agents. The notion of convergence is central to the modelling of markets,
since most research refers to economics and has to compare results to economic
static equilibrium. In some cases, what is observed is interaction patterns, which
can be represented as the random part of the agents’ choice when interacting,
e.g. the number of different sellers that a buyer meets in a number of steps.
In bargaining models and in general exchange models, the (non-)existence of an
exchange is also something that is observed. Sometimes, the cognition of agents
themselves is observed: their belief about the others’ preferences, or even the
distribution of propensities to choose sellers.
Among all the indicators that can be observed in the model, these last ones
cannot be observed in real life, and hence cannot be compared to human systems.
In a lot of models, agents’ behaviour is compared to that of humans in order to
establish the validity of the cognitive model. The data are very rarely extracted from
real life situation (although it sometimes happens), but are mainly constructed via
experiments. Experimental economists control all information circulation and
record all actions of agents. It is thus possible to compare in a very precise and
quantitative way the global behaviour of the group and individual behaviour, on one
side with artificial agents and on the other side with human agents. Real life
situation can also be seen as the mix of human and of artificial agents, such as in
financial online markets.
Other researchers do not want to match data too precisely. As Vriend (2007)
says, agent-based models, like any other models, are abstract settings that have to be
interpreted as such. The comparison between real and abstract data should go
through a first step which is the building of stylized facts that are already a summary
of human behaviours, where only the most striking elements are integrated. Vriend
is much more interested in the reaction of his abstract model to changes in
parameters, and in its self-consistency. It could be said that by construction, a
model can only capture a small part of human cognition, which is full of long-term
experiments and memories, and should not be compared to quantitative data
without caution.
Eventually, some researchers want their model to influence real life and try to
use the results they find to give advices on the way to build markets. Different ways
to fit models with real life will be found in each example of a model – be it to fit
precisely, to fit stylized facts, or to be an abstract study of the effect of some rules in
a social setting.
626 J. Rouchier
23.3 Buyer-Seller Interactions
In the literature about agent-based markets, a great attention has been given to the
analysis of local interactions, hardly ever studied in classical economics, apart from
rare exceptions (Rubinstein and Wolinsky 1985). An aim is to reproduce as well as
possible the features of real markets. It is indeed to be noticed in real observation
that buyers display regularity in the choice of sellers with whom to interact and that
this regularity emerges in time with experience – this attempt to reproduce patterns
of interaction is the only way to understand, rather than just describe, how
individuals coordinate (Weisbuch et al. 2000). Authors study local bargaining
processes with great care, as well as repetition of interactions over several days,
where choices are not only based on prices but also on the fact that buyers need to
buy and sellers need to sell. The basic feature of these models is pair-wise
interactions on markets with several sellers and buyers where prices are not fixed
and result from negotiation only. The number of visits buyers can undertake, the
way sellers manage queues, or the number of steps it takes to negotiate a price are
different in all these systems that focus only on small parts of the whole set of
stylized facts that are observable on such markets. Some usual aspects are described
in this section and the subsequent choices for modelling the organization of pair-
wise interactions are given.
23.3.1 Bargaining Processes
Brenner (2002) studies agents’ learning in a bilateral bargaining market, focusing

on the convergence of prices and the dynamics of bargaining. There is one com-
modity in the market, and buyers and sellers meet at every time step to exchange it.
Each buyer can choose one seller for each step, with the selection based on the price
being acceptable. The sellers respond to buyers waiting in their queue in order of
arrival by proposing a price. Buyers have to decide who to visit; sellers have to
decide on the first price to propose and the number of subsequent proposals if the
buyer rejects the offer, bargaining being costly for both agents. All decisions are
made following reinforcement learning based on past experience. Hence, all
choices are based on the satisfaction that is associated with each past action and
on a rigidity variable. A buyer will continue to choose a seller as long as he is
satisfied. His probability to change depends on his expectations with another agent.
A seller also calculates the probability to change behaviour depending on his belief
about what he would gain by performing another choice.
The rigidity parameter, which is the opposite of noise in the system, has a great
impact on results. If rigidity is high, buyers keep visiting the same seller. The cost of
bargaining also is important: if it is relatively high, sellers learn to offer the price
they know to be acceptable to buyers, and they do not bargain after a few rounds. In
this system, since the relations are so individual, the convergence of the price
overall is not very fast and can be highly variable for a long time, although
converging in the end. The model is extremely sensitive to all parameters that
define aspiration levels for agents.
Brenner’s paper is of the class that compares simulation to theoretical results.
Here, sub-game equilibria are used to compare the possible outcomes and their
efficiency to the generated prices and behaviours. There is no reference to any real
world data. However, it is interesting that both micro behaviour (the number of
bargaining steps) and macro data (prices) are of importance, justifying an agent-
based analysis.
Influenced by this paper, but referring to real data, Moulet and Rouchier (2008)
reported a bargaining model based on two sets of data: qualitative, from a field
study in the wholesale market of Marseilles (Rouchier and Mazaud 2004), and
quantitative, giving all proposals, intermediate and final prices for a series of
transactions in the same market (Kirman et al. 2005). Like the previous model,
the market gathers buyers and sellers who meet at every time step. However, buyers
can visit several sellers in one market opening. The choice for a buyer has several
dimensions: to decide which seller to visit, to decide to accept an offer or to reject it,
to propose a counter-offer or leave, and which value to counter-offer. Sellers must
choose the first price to offer, to accept buyer’s counter-offers or not, and the value
of the second offer they can make. In this model, decisions evolve following
classifier system learning, where each agent evaluates a list of possible options
following his past experience. The results that are produced are compared with
indicators derived from real-world data: values of offers and counter-offers of the
agents that vary depending of the kind of product that is purchased and ex post
bargaining power of sellers (which is the difference between the first offer and the
price of transaction compared to the difference between the counter-offer and the
price of transaction).
In the simulations, the values that are obtained fit the data quite well in that the
observed bargaining sequences and agents’ behaviours are reproduced. The two
main parameters are the number of sellers that agents can meet (from one to four)
and the speed of learning of sellers. The relative importance of learning for the
agents can be seen as situating them in a negotiation for in-season goods and a
negotiation for out-of-season goods. The model produces results similar to those of
out-of-season goods when agents have to learn repeatedly, when there is no known
market price but a lot of heterogeneity in the buyers’ readiness to pay. In the case of
in-season goods, the market price is more settled, and agents do not explore the
possible values of goods as much, relying instead on their experience. Between the
different in-season goods, the main difference could be the number of visits buyers
make, but this number tends to reduce after a learning period, when buyers have
selected their preferred seller. This aspect of the model – the growing loyalty of
agents – is not the centre of the research and was represented mainly with the aim of
matching the actual behaviours of the market actors. Other papers, described in the
following section, are more focused on this issue.
Another direction for the study of bargaining processes is related to the creation
of robots or artificially adaptive agents (AAA) to participate in electronic
628 J. Rouchier
commerce (Oliver 1997). Such models focus on complicated negotiations in that

they integrate several dimensions of trade in the deal: price, quantity and delivery
time. The main argument for the value of the algorithm that is proposed in the paper
is that the agents learn to negotiate at least “as well as humans”, which means that as
many negotiations lead to an agreement as in human bargaining situations so that
profit is extracted from both sides of the bargaining. The bargaining consists of
several steps, where a customer reacts to the first offer by comparing its profit to a
threshold, and the offer is accepted if it is higher than the threshold and rejected
with a counter-offer otherwise. Clearly, such models capture satisficing and
bounded rationality rather than profit maximization. The bargaining can then
carry on with several successive offers being made by customer and seller.
Strategies for accepting and counter-offering evolve through a Genetic Algorithm.
Five different games are used to test the learning, in a population of 20 agents with 3
rounds of bargaining at most and each agent is given 20 chromosomes for decision-
making. It is then proven that AAA perform better than random, that agents are able
to learn general strategies that can be used against different bargaining partners, and
eventually that AAA perform as well as humans (depending on the game, some-
times better and sometimes worse, maybe depending on affective values for
humans) in terms of number of agreements that are reached. This is an interesting
result to consider when one wants to introduce artificial agents into electronic
markets, since one wants to be able to reach as many agreements as possible.
23.3.2 Loyalty
Loyalty is present in quite a few models of markets where agents interact repeat-
edly. It is popular to deal with this topic with agents, mainly because it is related to
two main advances of agent-based modelling: heterogeneity and interactions. There
exist two representations of this loyalty in the literature: either fixed loyalties,
assumed in order to understand its impact (Rouchier 2013), or emerging loyalties,
as the result of endogenous interactions. Vriend refers to “endogenous interactions”
when he uses individual learning to generate an evolution of interactions among
agents (Vriend 2007). The idea is that agents learn to select which actions to
perform as well as which agent to interact with; it is clear that this can lead to the
apparition of loyalty, and that it can take different regular patterns.
One main field where this loyalty issue has been important is the study of
perishable goods markets (fruits and vegetables and fish). The participants of these
markets are very dependent on the regularity – which implies predictability – of their
interactions. The main reason is that buyers need to purchase goods almost every
day: they have very little ability to stock and they must provide their customers with
all possible goods (a retailer can become unattractive to his customers just because
of the lack of one commodity). In case of shortage, they need to have good relations
with a seller to make sure the commodity will be available to them. Conversely,
Rouchier (2013) shows in a model that the presence of loyal agents in a perishable
goods market is necessary for the sellers to predict the right number of goods to
provide every day. In this artificial market, two types of buyers interact with sellers:
those that look for the cheapest prices (“opportunistic”) and those that are faithful
and try to get the product rather than to get it cheap (“loyal”). To be able to be
opportunistic, agents first gather information about prices, and then decide on the
seller they want to meet to make the best transaction. The more opportunistic agents
are present in the market, the more garbage is produced and shortage occurs.
Although there is some randomization of needs for the buyers, the presence of
loyal agents makes the sellers estimate their stocks in the best way. This result
holds for different learning algorithms (step by step learning, simple reinforcement
learning and classifier systems). Considering that the field study on the fruits and
vegetables market of Marseilles, France, showed that most of the agents are loyal
(according to the definition of the model: first loyal and then try to find all the goods
in a minimum of visits to sellers), this result can give a functional explanation of
their action choices.
In a slightly different context, Rouchier et al. (2001) have represented the shape
of emerging patterns of relations that could be created by two types of rationality
with agents. The situation is a market-like situation, where offers are dependent on
the situation in the preceding step, since the commodity is a renewable resource.
Agents are herdsmen and farmers, with the latter selling access rights to their land.
If none or if too many herdsmen are using the same land, it will get depleted, and
hence the offer will be reduced. Two types of micro behaviour are defined: either the
herdsmen choose the cheapest farmers, or they choose the ones that offered them
access most often. In the first case, the simulations resulted in depletion of the
resource, with congestion of demand for the cheapest farmers. The links that were
created were highly stable (once an agent found the cheapest it would not change),
but on the other hand agents could not readapt when there was a shock in the
resource quantity (a drought) because everyone would converge to the same
farms. With the second rationality, agents had a representation of a “good” farmer,
which was only based on individual experience, and hence they would be heteroge-
neous. They would also have several “good” farmers to visit in case one was not
available. This made them much more flexible in their choice, avoiding depletion of
the resource, so everyone was better off. The macro situation, although the process is
different, also shows that a loyal micro-behaviour is a help to repartition of goods
were there can be shortages. In this setting the loyal micro-behaviour also enables
a more equal repartition of gain among farmers as well as herdsmen.
Kirman and Vriend (2001) explored the emergence of loyalty in an artificial
market based on a field study in the fish market of Marseille. The aim is to see
loyalty emerge, and in addition to see which emergent behaviour sellers display.
They use classifier systems to represent learning, where their agents can have a lot of
different actions, some of which are, a priori, not good for their profit. Through the
exploration of good and bad possible actions, they select those that bring the highest
profit in the past. Some buyers learn to be loyal, and those that learn this get higher
profit than others in the long run (it is actually a co-evolution where sellers learn to
offer lower prices to those that are loyal). The buyers are then differentiated: their
630 J. Rouchier
reservation price is heterogeneous (for example to represent that they do not sell
their fish to the same population, some are in rich neighbourhood, some in poor
ones). Sellers on the market learn to discriminate, and they offer higher prices to
those that have higher reservation prices. Eventually, some of the sellers get
themselves specialized since only low prices buyers can visit them. Using a very
basic learning where agents are not rational but learn by doing, the results are very
satisfying because they reproduce stylized facts of the fish market.
A third model represents the same market but refers more to quantitative data of
this market (Weisbuch et al. 2000). The data represents sales that took place over
more than 3 years and concern 237,162 individuals. From these it is possible to
observe that most buyers who are faithful to a seller buy a lot of quantities every
month. The model was built in two parts: one which is simple enough to generate
analytical results and a second that displays more realistic hypotheses. In the
analytical model, agents use the logit function to select their action (basic rein-
forcement learning), which means that their choice depends on a β value, between
0 and 1, which decrease gives a higher propensity to randomly test all sellers and
which increase induces a higher propensity to look for the best past interaction.
Agents can either imitate others or only base their choice on their own learning. The
results can be found using the mean field approach, coming from physics. It is
shown that there are radically different behaviours – either totally loyal or totally
“shop around agents” depending non-linearly on β.
The model becomes more complex with sellers being able to sell at two different
prices, high and low. What can happen in this system is that a buyer becomes loyal
to a seller when the price is low, and remains loyal even after the price has switched
to high. The only important thing is that, as seen before, the product is actually
provided. One indicator that is used to synthesize diverse information of the model
is “order”, which is defined as the number of loyal agents. The more regular the
agents, the more ordered the society. Although the results, once interpreted, are
very coherent with real data in terms of possible states of the market, it is a bit
difficult to understand precisely some concepts of the paper because it refers mainly
to physics indicators that are translated into social indicators, but this translation is
not always straightforward.
23.3.3 Reputation of Sellers
Pinyol et al. (2007) have developed a market model as a benchmark for a

reputation-based learning algorithm for agents in a social system. The model
integrates quality and judgment of a relationship. Reputation is used in the group
to enable agents to gather enough information in a context when it is scarce. The
market that is used is a rather simple institution, where buyers have to select one
seller at each time step to buy one unit of a commodity. The quality of the
commodity is different for each seller. For a buyer, the acquisition of a commodity
of lower quality will give less utility than the acquisition of a commodity of high
quality. Sellers have a limited quantity of units, which they can sell at any period
(the commodity is non-perishable) and they disappear from the system when they
have sold everything. The most important information for buyers is the quality of
the commodity that each seller offers. However, when the number of sellers is large,
this information cannot be acquired efficiently if the buyer has to meet a seller to
learn of the quality of his commodity. This is why information circulates among
buyers, who communicate once every time step. A buyer who meets a seller forms
an image of this seller; a buyer who gets information about a seller has access to a
reputation of this seller. When giving information to another buyer, an agent can
decide to give the direct knowledge it has (the image it formed of a seller) or the
reputation it has already received (which is more neutral since it is not its own
evaluation). Reputation can also circulate about the buyers, and in that case
concerns the validity of the knowledge they give about sellers. When a buyer is
not satisfied with the information given by another buyer, it can also retaliate and
cheat when this very agent asks him a question.
Pinyol et al. (2007) describe in detail the choices that agents make when asking a
seller for a commodity, asking another buyer for information about a seller or a
buyer, answering a question and the lying process.
The simulated market contains a large number of sellers (100 for 25 buyers).
Simulation runs are defined by (a) the type of information that is used by agents
(only image or image and reputation) and (b) the number of bad-quality sellers in
the system (99, 95, 90, and 50). The addition of reputation to the system makes the
difference between a normal learning mechanism where buyers select their
favourite seller and a learning mechanism where agents aggregate information
of different quality to (maybe) increase their performance. The results show that
globally the agents indeed learn more efficiently when using reputation, in that the
average quality that is bought is higher. The quantity of information that circulates
is much higher and this enables buyers to increase their utility. This social control
mechanism is especially important when quality is really scarce (1 % of good
sellers). This result is all the more interesting since this is a very rare case of a
simulated market where communication among buyers is represented, although this
behaviour is commonly observed in real life situations. For a more in-depth
discussion of reputation see Chap. 15 in this handbook (Giardini et al. 2013).
23.4 Consumers, Producers and Chains
Another way to look at the notion of goods markets is to consider large markets,
where individual interactions are not important for the agents who do not record the
characteristics of the ones they meet, but only the fact that they can or cannot
perform an exchange. A large market can indeed include numerous goods, that are
distributed among different other agents and not necessarily easy to access. Another
632 J. Rouchier
interest in large market is to study endogenous preferences for goods, and imagine
their evolution depending on the type of good and some cognitive characteristics of
agents. Eventually some authors are interested in the coordination process within
the supply chain itself, where the issue is about the amount of information that each
agent has to use to anticipate the needs of distant, end consumers.
23.4.1 Multi-goods Economy
As examples of a market with several goods we will discuss two very abstract
models where agents have to produce one commodity and consume others, which
they can acquire only through exchanges with other agents. The first model was
built to produce speculative behaviours in agents, which means acquiring a product
that has no value for consumption but only a value for exchange (Duffy 2001); the
second model’s aim is to witness the emergence of commonly used equivalence
value for the goods, which is interpreted as relative prices (Gintis 2006). Both
models are interesting for their pure description of an abstract economy with
minimalist but sufficient assumptions to induce economic exchanges. In the
works cited here, the methodology used in realising the models is slightly different:
one is purely abstract whereas the other tries to refer to experimental results and
mimic human players behaviours.
In his paper, John Duffy (2001) designs a model that was originally proposed by
Kiyotaki and Wright (1989) to induce some agents to store a good that is not their
designated consumption good and is more costly to store than their own produced
commodity, because they think it easier to exchange with others. There are three
different goods in the economy, agents need to consume one unit of good to
increase their utility and produce one unit of good each time they have consumed
one. There are also three types of agents: agent type 1 needs good 1 and
produces good 2; agent type 2 consumes good 2 and produces good 3; agent type
3 consumes good 3 and produces good 1 (in short, agent type i consumes good i and
produces good i + 1 modulo 3). Hence agents have to exchange when they want to
consume and not all agents can be satisfied by just one exchange. Indeed, if two
agents exchange their own production goods, one can be satisfied but the other
would get a useless good, which is neither its own production good nor its
consumption good. In this economy, only bilateral trading exists and it takes
place after a random pairing of agents. This involves that some agents must keep
a good for at least one time step after production before getting their consumption
good.
In this economy, speculation is defined as the storage of the good i + 2, since the
agent does exchange to get this good which it has not produced, only because of the
chances to use it as an exchange good at the next time step. The economy is made
non-symmetric by having different costs for the storage of goods, here 0 < c1 <
c2 < c3. The original Kiyotaki and Wright model is all about calculating, given the
storage costs, the discount factor (the probability that the economy stops at the end
of a time step) and the utility of consumption, whether agents decide to get the most
expensive good to store or keep their production good. In an economy with perfect
information, the expected profit for each type of agent depends on the proportion of
agents of type i holding good i + 1 (their production good), which is 1 – proportion
of agents of type i holding good i + 2 (their “speculative” good).
With this model John Duffy tries to investigate how agents could learn which
choice to make when they are able to acquire a good they do not consume.
Especially agents of type 1 are those that should hesitate, since good 3 is the
most expensive. A lot of models have been built on this topic (Basci 1999) already,
but what Duffy wants to produce is a setting that is close to laboratory experiments
he has been leading (Duffy and Ochs 1999) in order to be able to judge if his agents
are behaving in a way which is coherent with that of human actors. So, from a
theoretical setting he builds experiments and simulations, and compares all the
results this technique produces. In this paper he therefore proposes an algorithm
that is close to his intuition of what individuals should do (he has also asked
questions to people involved in his experiments), and he then tries to mimic the
results of his experiments, at a global level and a local level. He also proposes some
original settings where he mixes human agents with artificial agents to test at the
same time his algorithm and how much he can make humans change behaviour
depending on the stimuli they get. He is satisfied with his results, where his model
of learning enables to reproduce human behaviour correctly. Unfortunately, the
reproduction of his model is not so straightforward (Rouchier 2003), but all in all,
this description of a very basic economy with few goods and where agents learn in
an intuitively plausible way is a very interesting example of market for agent-based
modellers.
The paper by Gintis (2006) presents similarities, although the aim and the central
question are different. The economy that is presented can be seen in a very general
way but is only implemented in one setting, which is described here. In the
economy there are three goods and 300 agents. Each agent can produce one good
and needs to consume both goods it cannot produce; hence it is forced to exchange
with other agents. At the beginning of each period, an agent only holds the good it
produces (in a quantity that it can choose and which is costless) and can meet two
agents, each producing one of the good he needs to acquire. Each agent has a
representation of “prices”, which is here defined as the equivalence quantity
between two goods. There is no common knowledge of prices and each agent has
its own representation. When an agent meets another agent who can provide him
with the needed good, he offers to trade, by sending as a message its representation
of relative “prices”. The exchange takes place at this exchange rate if it is accept-
able to both agents and the exchanged quantities are the highest quantity that both
can exchange. Agents cannot choose who they meet, they just pick randomly from
the other producers’ groups. After exchanging, they can consume, which gives
them utility, and defines a performance for each individual. Learning in this system
is an event that takes place every 20 periods, where the 5 % of least performing
634 J. Rouchier
agents (who get the lowest utility) copy the price representation of the highest
performing agents.
What is observed in the system is the efficiency of the represented market,
meaning the sum of all profits, compared to a setting where prices would be public.
When prices are public, all exchanges can take place since all agents agree right
away on the equivalence that is proposed and there is no refusal in exchange. In the
long-term, the system converges to the highest possible efficiency, so although
the agents have private prices, these prices get to be close enough to have regular
exchanges. This result in itself is not very surprising in terms of simulation
(considering the process at stake), but is interesting in economics since it gives,
at least, a process to attain to a common knowledge which is often presupposed
to exist.
23.4.2 Adoption by Consumers
The study of the behaviour of large numbers of consumers facing the introduction
of new products on a market is a topic that is very interesting to approach with
agent-based simulation, since it allows, once more, looking for the influence of
heterogeneity of individuals and of networks in the evolution of global results.
Wander Jager is a prominent figure in this area of research, positioned between
psychology and marketing. In a paper with Marco Janssen, he presents the basic
model (Janssen and Jager 2001). The idea behind the study of the acquisition of a
new product in a group is that agents have a preference that is based on two main
parameters: the individual preference for the consumption of the product and the
interest that the agent has to consume the same product as his acquaintances. Hence,
a utility function depends on these two parameters, and this will influence an
agent’s decision to buy a new product. Agents are heterogeneous in such a system,
and the representation of “early adopters” (in the real world: people who buy a
product when it is just released) is modelled by a low need to conform to others’
behaviour. On the opposite end of the spectrum, some agents buy a good only
because a lot of their acquaintances have already acquired it.
In (Janssen and Jager 2001) the influence of network size and topology is tested,
as well as the influence of the utility brought by the product consumption and the
way agents choose their action. One aspect that is studied is the type of cognitive
process that can be used by the agent (repetition of the same action; deliberation to
find a new action; imitation – where other agents’ consumption is imitated; social
comparison – where other agents are imitated based on their utility). This indicator
is quite rare and shows the psychological grounding of the paper. It is interesting to
observe that the cognitive process changes with the utility gained by the consump-
tion of the considered product. Agents with a lot of links are very important for the
spreading of product adoption: in a small world network, many more products get
adopted than in a scale free network. A discussion is open here about the type of
products, which certainly influence the way people copy others – it will be different
for milk and for computers or clothes.
This last question is actually developed in a different paper: in (Delre et al.
2007), the question that is at stake is to determine how to advertise efficiently
depending on the type of good. Is it better to advertise a lot at the beginning of a
campaign, or after a moment; is it better to advertise to a large number of people or
to disseminate information among only a few agents?
Two products are differentiated: a brown product, which is a high-tech and quite
fancy product that can be compared with other agents of the network (e.g. CD, DVD
player), and a white product which is related to basic needs and is not really
compared (fridge or washing machine). Agents are gathered in networks of differ-
ent topologies. In this model, the heterogeneity in the utility formula is similar to
the one in the preceding paper: each one is defined by a propensity to be influenced
by others and to be an adopter of a new technology.
The first finding of this model is that the timing of promotion is important to the
success of the campaign. For a first launch, the best strategy is to “throw gravel”, i.e.
to do a little advertising to many distant small and cohesive groups of consumers,
who will then convince their network neighbours. Another element is that not too
many people must be reached at first, since if they see that others have not adopted
the good yet, they might not want it and become impossible to convince afterwards.
This is mainly true for the white good, where it is better to advertise broadly when at
least 10 % of agents have already adopted the good, whereas with brown goods
adoption is much faster and a campaign helps the takeoff.
The issue of the adoption of a practice within a social context has also been
studied to understand the adoption of electronic commerce for consumers (Darmon
and Torre 2004). The issue at stake is that it should be logical that everyone turns to
electronic commerce, which radically reduces transaction and search costs, but we
observe that a very small proportion of items are as yet traded via the internet. This
is mainly because consumers have not developed special abilities that are
associated with this form of interaction, and do not know how to reduce the risk
of performing a bad transaction. To study the dynamics of adoption of electronic
commerce and learning of agents in a risky setting, a simulation model has been
built. The market is composed of agents who can produce a good and exchange it
for the good they want to consume (hence all agents are at the same time producers
and end consumers). Agents and goods are located on a circle; the location of an
agent defines the “quality” of the good it produces. For consumption, each agent is
defined by an interval of quality: when consuming a good whose quality is within
this interval (not including its own production good), it will get a strictly positive
profit. The cost of production is heterogeneous and can be constant during the
simulation or evolving.
When trading on the traditional market, an agent can identify the quality of a
product offered by another agent, but it has to talk to many others before finding out
who it can exchange with (depending on the number of agents and of its interval of
choice). The authors also added a notion of friction, which is a probability of failing
to trade when two agents meet. In the electronic market, an agent sees all other
636 J. Rouchier
agents at once (no search cost), but cannot identify precisely the quality that is
offered and evaluates it with an error interval. Hence it potentially accepts goods
with zero utility. Agents are heterogeneous in their ability to perceive the quality and
their learning algorithm. If an agent learns via individual learning, then it eliminates
an agent from its list of potential sellers whenever the previous trade brought no
utility. If agents learn through collective learning, then a part of the whole society
belongs to a community that shares information about the quality (location on the
circle) of the seller they met at this time step; the agents not belonging to the
community learn individually. In some simulations, for both types of learning,
agents forget a randomly chosen part of what they learnt at each time step.
In the case of individual learning, the dynamics produced depend on the produc-
tion cost, which can either change at all time steps (and hence all agents have the
same average production cost over the simulation) or which can be constant and
delineate populations with high or low production costs. When the production cost
changes at each time step, the main result is that eventually all agents switch to the
electronic market, with the move happening in stages. Those that have a good
appreciation of quality go to the electronic market very fast because it is more
profitable for them. Their departure from the traditional market reduces the proba-
bility of exchange for the remaining agents, who eventually move to the electronic
market as well. When production costs are heterogeneous, some agents cannot
switch from traditional to electronic because of their inadequate production cost.
Hence they never learn how to identify quality and stay in the traditional market.
When agents forget part of what they have learnt, then the size of the electronic
market does not get as large as with perfect learning, and a number of agents do
not switch.
When agents participate in a community and exchange their information, the
highest number of agents will switch to the electronic market and overall the lowest
number of agents is excluded from exchange. Three groups are created: agents
belonging to the community, who get the highest pay-off; agents with low produc-
tion cost or high expertise who can go on the electronic market and make a high
profit; and the remaining agents, which sometimes cannot exchange This result is
rather coherent with what could be expected, but it is interesting to have it created
with this location-based representation of quality that each individual wants to
attain. It is especially clear that there is little risk that traditional markets should
disappear if the main assumption of the model – that agents need an expertise that
takes long to acquire before switching to the electronic market – is true.
23.4.3 Decentralized Supply Chain
Supply chains are an important aspect of economics, and they are often difficult to
consider, mainly because their dynamics spread in two directions: (1) along the
length of the chain, suppliers have to adapt to demand and buyers have to adapt to
the speed of production so as to be able to provide the end market with the right
quantity of goods; (2) another dimension is the fact that suppliers as well as buyers
in the chain are substitutable and that each actor is itself in a market position and
can choose between several offers or demands. In existing agent-based models,
only the first issue is treated. The structure of these models is a series of agents
(firms) that are each linked to two agents: a supplier and a client (except for first
supplier and end consumer, of course, who are linked to only one firm). Each agent
has to decide on its production level at each time step, knowing that it needs to use
goods from the preceding firm in the production process. It must then anticipate the
demand to order enough, before being able to transform the needed quantity.
Of course, each firm takes some time (number of steps) to transform the product
and be able to sell and there is a cost in storing its own production when it is not
completely sold.
One very important issue of these chains is at the centre of most research: how to
avoid the so-called bullwhip effect. This effect is a mechanical dynamics that comes
from the slow spreading of information and delay in answer because of the length of
the production process in each firm. When there is variability in demand coming
from end consumers, this variability increases a great deal when it goes up the
chain, right up to the first producer who exhibits the highest variability. It can be
very annoying for organizations to be trapped in such negative dynamics. Several
authors propose algorithms for artificial agents that have to deal with the issue of
anticipating demand at each stage of the chain. For example Lin and Lin (2006)
describe a system where artificial agents can interact with real agents (and hence be
integrated in a real-life company to help decision makers) in order to choose the
right level of production and order to reduce costs. Several learning algorithms are
tested and their efficiency attested, even in situations where the environment is
dynamically evolving. The same issue is dealt with by others, for example Kawagoe
and Wada (2005), who propose another algorithm. They also propose a method to
statistically evaluate the bullwhip effect. Their method is different from the usual
frequency-based statistical measurement (like stochastic dominance) but is based
on descriptive statistics.
23.5 Financial Markets and Auctions
Financial markets have been one of the first examples that were developed to prove
the relevance of agent-based modelling. Arthur et al. (1997b) indeed reproduced
some important stylized facts of asset exchanges on a market and this paper is
always cited as the first important use of this modelling technique for studying
markets. Contrary to models that were presented before, there is no direct interac-
tion among agents in these models, only observation of price patterns. One rare
example presented here is an attempt to link a financial market to a consumer
market such as the ones seen in previous sections. Another type of market that does
not integrate any interaction in the economy is the representation of auctions.
638 J. Rouchier
23.5.1 Financial Markets
The literature on financial markets is very important in agent-based simulation, and

dates back to the 1990’s (Arthur 1991, 1994; Arifovic 1996; Arthur et al. 1997a).
This holds also true for the related branch of research, which is called econophysics:
the use of physics techniques to deal with economic issues in systems that are
composed of a huge number of simple interacting actors (Levy et al. 2000).
A comprehensive review of this topic is (Lux 2009). It describes at the same time
the main stylized facts that can be found in financial markets (and hence are meant to
be reproduced by simulation) and some models that are candidate for explaining
these facts. Another review (Samanidou et al. 2007) describes several agent-based
simulations models dealing with financial markets; unfortunately, these models do
not achieve to reproduce the very general statistical regularities of these markets.
As usual, in the following I will describe only a selection of models and ways to
represent agents’ learning in the context of financial markets. The basic structure of
the market, which defines the type of choice the agent has to make, can vary as well
as the aim and methodology of the researcher building these models and this is why
discussing a few representative examples in detail seems a better idea than presenting
very generic results.
One reason for using agent-based models is to be able to represent populations of
heterogeneous agents. What is very often found is the representation of two types
of agents with different reactions to information: chartists and fundamentalists.
Fundamentalists base their investment decisions upon market fundamentals such as
dividends, earnings, interest rates or growth indicators. In contrast, technical traders
pay no attention to economic fundamentals but look for regular patterns in past
prices and base their investment decision upon simple trend following trading rules.
Computer simulations such as those of the Santa Fe artificial stock market (LeBaron
et al. 1999; but see also e.g. Kirman 1991; Lux and Marchesi 1999, 2000) have
shown that rational, fundamental traders do not necessarily drive out technical
analysts, who may earn higher profits in certain periods. An evolutionary competi-
tion between these different trader types, where traders tend to follow strategies that
have performed well in the recent past, may lead to irregular switching between the
different strategies and result in complicated, irregular asset price fluctuations.
Brock and Hommes (1998) have shown in simple, tractable evolutionary systems
that rational agents and/or fundamental traders do not necessarily drive out all other
trader types, but that the market may be characterized by perpetual evolutionary
switching between competing trading strategies. Non-rational traders may survive
evolutionary competition in the market; see for example (Hommes 2001) for a
survey.
In (Hommes and Lux 2008), the chosen market model is the so-called cobweb
model, which is a prediction model on a market, not an actual model of selling and
buying for agents. The model offers, however, a rational expectation value, which
serves as a benchmark. The methodology is to try to fit agents’ behaviour in an
artificial world to real behaviours of individuals in experiments. The game is such
that participants of the experiments have no clear idea of the structure of the market
but still have to predict the price of the next period. They neither know how many
other agents are present, nor the equation that calculates the future price based on
the realised price and the expectations of all participants. The simulations are made
based on rather simple models of agents including a genetic algorithm, simple
learning that copies past prices, and reinforcement learning. What interests the
authors most is the GA learning, which is the only one to fit stylized facts in
different treatments. What the GA learns about is a 40 bit string of 0 and 1
representing two values, α (the first 20 bits) and β (the remaining 20 bits), that
predict the price at t + 1 depending on the price at t with p(t + 1) ¼ α + β
(p(t) α).
There are three runs both for experiments and for simulations, with one parame-
ter defining the stability of the price (high, medium or low). The genetic algorithm
being varied for different mutation rates is proven to be largely better than other
learning procedures that have been implemented. “Better” means here that it fits the
stylized facts that have been produced by humans in experiments: (1) the mean
price is close to rational expectation, and the more stable the market, the closer the
mean price is to this rational expectation value; (2) there is no significant linear
autocorrelation in realized market prices. The reason for the good fit of the GA
given by the authors is really interesting because it is not obvious to imagine how
GAs, which are random learning processes with selection, should be similar to
human learning. The authors assume that the good fit is based on two facts: the fact
that successes are selected positively, and that there is heterogeneity in the
strategies among the set that agents can use. Once the assessment of the model is
done, it is used to question the stability of the results of the learning process. One
question that arises is to wonder whether humans would adapt the same way when
interacting in a very large group as they do in a small group of six. This opens many
questions about the scalability of results concerning market dynamics.
In our second example the interaction of agents is direct and not necessarily via
the price system, as is usual in financial markets. Hoffmann et al. (2007) indeed
consider that many agent-based simulations still take little interest in representing
actual behaviours of decision makers in financial markets. They argue that
(Takahashi and Terano 2003) is the first paper to integrate theories that come from
behavioural finance and represent multiple types of agents, such as overconfident
traders. In their own paper, Hoffmann et al. (2007) present their platform SimStock-
Exchange™, with agents performing trades and making decision according to news
they perceive and prices they anticipate. They argue that their model is based on
several theories that are empirically sound and that they validated their model results
against data over several years from the Dutch market. As usual, the platform allows
many variations (increase the number of different shares, of agents, change the
characteristics of agents) but is tested only with some values of parameters.
Agents receive news that they forget after one time step, and then can perform
two types of action: either sell their stock (if they expect to lose at the next time
step) or buy more shares (in the opposite case). To make sure that they are not
making mistakes, agents can use risk-reducing strategies, which can be clarifying
640 J. Rouchier
strategies (such as collecting more data) or simplifying strategies (i.e. imitating

other agents), as well as purely individual (the first one) or social (the latter). In the
presented simulation, strategies are always social, and hence agents’ confidence, C,
determines their use of risk-reducing strategies; the confidence values were
deduced from empirical studies. Each agent is also defined by a tendency R to
perform a simplifying strategy or a clarifying one. R and C are evaluated on the
basis of surveys made with investors. Agents are imbedded in networks of two
different topologies (torus, scale-free); agents may acquire information from their
links or choose to imitate them. The market itself is designed as an order book,
where proposals for sells and buys are written down with quantity and price, and are
erased as soon as an agent answers positively to a particular proposal. The market
price is the average of all proposed bids and asks of the order book – hence it is not
a realized price (average transactions’ price) but an aggregation of desired prices
for agents.
In the results, some statistical properties of the stock exchange have been
reproduced. For example, with weekly data of Dutch stock exchange, linear auto-
correlation can be observed, and this is better reproduced when the torus-shaped
network is used rather than the scale-free one is used. With regard to volatility
clustering, the torus network differs from both the scale-free network and the real
data. This can be due to the high speed of information circulation reducing the
shocks that it can cause. The main aspect of the SimStockExchange that needs
improvement is the news arrival, which is a normal distribution around the present
price. This might have a large impact since the use of different networks integrates
the importance of information spreading.
23.5.2 Relation Between Two Markets
Sallans et al. (2003) report a model integrating two types of markets: a financial
market and a goods market in the same system. Consumers, financial traders and
production firms are interacting and the aim is to understand how these two markets
influence each other. The good is perishable, and hence needs to be purchased
regularly. Consumers make purchase decisions; firms get income from sales and
update products and pricing policies based on performances; traders have shares,
which they can hold, sell or buy. Firms decide upon the features of their products,
which are represented as two binary strings of 10 bits. In choosing actions, the firm
agent uses an action-value function that integrates expectations about future
rewards (firms are not myopic agents) by taking into account the evolution of the
price of its share in the financial market and the profit made by selling products on
the goods market. Consumers have preferences for particular features of a product
and its price, and compare any available product to these preferences: they can
choose not to buy if the product is too different from their preferences. In the
financial market, agents build expectations and built representations of future
values by projecting actual and past values into the future. They are divided into
two groups: fundamentalists (use past dividend for projection) and chartists (use the
history of stock prices); they are also heterogeneous regarding their time horizon.
The market clearing mechanism is a sealed bid-auction and the price is chosen to
maximize the number of exchanges (and randomly among different prices if they
produce the same trade volume).
Agents from the financial market and firm agents have different views on the
future of a firm, and evaluate future gains in a different way, which might impact
the firm’s performance negatively. The simulations’ aim is to prove that the model
can be used, in certain parameter settings, to reproduce stylized facts of markets.
Although the central issue is very interesting, the paper itself is not as helpful as
it could be to understanding the dynamics of two markets. In particular, the stylized
facts are not very explicit in the paper (appear only once at the end, when obtained
results are given). They are classical in financial market analysis, but not clearly
shown here: low autocorrelations in stock returns, high kurtosis in marginal return,
volatility clustering. Hypotheses on behaviour are never explained, hence there is
no understanding of why the stylized facts can be achieved, apart from doing some
random exploration of the parameter space. Thus, while the main issue of the paper
is fascinating, the results are a bit frustrating and the reciprocal influences of these
two markets, so important in our real world, stay hidden.
23.5.3 Double Auctions
In economics double auction is a very fascinating topic, since it is an extremely

stable market protocol in which predictions can be translated from theory to real
life, which is not really the case for most economic systems. When putting real
people in a double-auction setting, one can observe that the convergence to
equilibrium price occurs. This does not mean that this protocol is efficient, since
a lot of exchanges take place out of equilibrium price, but at least there is a tendency
for the group to converge to a price where the highest number of exchange can be
performed, and hence the highest global profit can be extracted. Many authors have
therefore wanted to reproduce a double-auction market in an artificial society in
order to understand the source of this high efficiency.
The CDA (continuous double auction) is a two-sided progressive auction. At any
moment in time, buyers can submit bids (offers to buy) and sellers can submit asks
(offers to sell). Both, buyers and sellers may also accept an offer made by others. If
a bid or ask is accepted, a transaction occurs at the offer price. An improvement rule
is imposed on new offers entering the market, requiring submitted bids (asks) at a
higher (lower) price than the standing bid (ask). Each time an offer is satisfying for
one of the participants, she announces the acceptance of the trade at the given price,
and the transaction is completed. Once a transaction is completed the market is
cleared (meaning there is no standing bid or ask any more) and the agents who have
traded leave the market. At that moment, similar to the opening of the market, the
first offer can take any value, and this proposed price imposes a constraint on any
642 J. Rouchier
following offer. When the market closes, after a time decided beforehand, agents
who have not yet traded are not allowed to continue. In this market protocol all
market events are observed by all (bid, ask, acceptance and remaining time before
market closing) and hence are said to be common knowledge.
Using this double-action setting, a seminal paper by Gode and Sunder (1993,
2004) shows the strength of institutional constraints on the actions of agents. In
their model agents are perfectly stupid from an economics point of view, since they
have no understanding of their own interest and only follow a simple rule without
any strategic planning. These so-called “zero-intelligence” agents are not allowed
to sell (buy) lower (higher) than their reservation price, and they have to bid within
the limits that have been put by others. With this rule, convergence of prices is
obtained very fast. The approach in this paper is quite original in the behavioural
economics literature in the sense that it is close to an “artificial life approach”. The
authors do not pretend to study human rationality, but instead focus on the abstract
reproduction of phenomena. It is interesting to note that is not so easy to design a
double-auction market, especially in its continuity. Indeed, in a real situation, if two
individuals have close reservation prices, they will often be able to buy or sell at the
same moment. Who will be first is not obvious, since people have different aspira-
tion for profit. Gode and Sunders randomly choose an agent between all buyers who
can buy or make a bid, and then randomly pick a seller among those who can sell or
make an offer. After trying several methods they decided on random selection,
explaining that this is a good approximation to continuous double auctions.
Their work is widely criticized because (a) they are not interested in rationality
but in a specific market protocol and (b) it cannot be generalized to other protocols
(Brenner 2002). However, their result is important and led a lot of researchers to
question it. For example (Brewer et al. 2002) shows that humans are able to have
markets converge when the context changes a lot, which Gode and Sunders’ agents
cannot do. They organize a double-auction market, in which agents participate in
the public market but also receive offers from the experimenter privately. Only one
offer is made at a time, and it is the same for all agents that are proposed the offer,
since the equilibrium has to stay the same. The global equilibrium (which value is
described in the paper) is thus constant, but individuals can have incentives not to
participate in the public market if the offer is interesting. This does change the
performance of zero-intelligence a lot, since the prices do not converge anymore in
simulations led with this new protocol. On the opposite, humans performing
experiments attain convergence, which could mean that only very specific
institutions constrain participants enough so that they have no choice but to
converge, even while not understanding more than the rules (zero-intelligence).
Inspired by Gode and Sunder, but also by the theoretical model of Easley and
Ledyard (1993), Rouchier and Robin (2006) tried to establish the main elements a
rational agent would need to be able to choose the right action in the context of
a double auction. To differentiate among different possible learning procedures,
a comparison with some experimental results was made. The learning procedure
chosen is a simple algorithm that consists of making the agent revise its reservation
price towards past average perceptible prices, depending on two variable elements.
First, the duration after which an agent would change its reservation price (i.e. a
buyer (seller) accepting higher (lower) prices), called the “stress time”, could
change – increasing after a successful transaction and decreasing after a day with
no transaction. Second, the agent could either only perceive its own transactions or
those of any successful transactions in the market. The paper demonstrated that
agents learn faster to converge to the equilibrium price (making the highest global
pay-off) if they did not revise their stress-time and had a global perception of prices.
This quick learning would at the same time correspond best to the speed of
convergence that could be found in experiments. What is a bit surprising in this
result is that more “clever” agents (reacting to risk and failure from one day to
another) would neither copy human behaviour well nor get to the equilibrium price
very fast.
23.6 Market Design/Agent Design
In a chapter of the handbook for computational economics (Tesfatsion and Judd

2007), Robert Marks (2007) reviews recent work in market design using agent-
based simulation. Market design is the branch of economic research aiming to
provide insights about which protocol, i.e. interaction structure and information
circulation rules, is the best to obtain certain characteristics of a market. As said
repeatedly in this chapter, this choice is crucial in having certain parts of a popula-
tion gain more power than others, or having efficiency attained in a short time.
Hence, many scientists have been thinking about this issue, using game theory
(Roth et al. 1991), as well experimental economics, and more recently computa-
tional exploration. As seen before, sophisticated agents are not the ones who do best
in market situations or copy human behaviour closest.
When designing a market protocol it is important to see two challenges. First the
“aim” of the protocol needs to be clear since not all positive aspects can be achieved
in a single protocol (see for example Myerson and Satterthwaite 1983). For
example, using Dutch auction has the advantage of being fast, whereas double
auction is good because it extracts the highest global profit for all. On the other
hand, one might wish to extract the highest profit for buyers only, for example.
LeBaron (2001) explains that the fitness of a model is as important as all other
elements (what is traded, the motivations of agents, how the interaction and
information circulation is organised, etc.). To achieve this trade-off between differ-
ent characteristics is already a huge choice before starting the design.
Then one has to think on how to achieve this aim. It is indeed not easy to know
how individuals will react to an interaction and information constraint. The basic
use of agent-based simulation can then be to either test a certain agent behaviour
and compare protocols to see what difference it makes in prices or other indicators
(Moulet and Kirman 2008) or to test different learning algorithms in the same
setting (Chan and Shelton 2001). Both approaches are uniquely developed using
644 J. Rouchier
agent-based simulation, and can indeed help understand the relation between
participant behaviour and market protocol.
Many models, be it for computer scientists or economists, were designed to fit
the context of the electricity market, which is crucial since problems can be very
severe for society (when there are huge unpredicted shortages) and the variations in
price can be very fast. The agents in those models are not designed to represent
human rationality but to try to be as optimal as possible in the adaptation to the
electricity market. Many market protocols can be used, although auctions (which
are theoretically the most efficient of all market protocols) are most common.
Bidding behaviours, but also the number of sellers and buyers, and the capacity
to produce and sell (Nicolaisen et al. 2001) have an impact on the efficiency and this
can be explored. As said before, what is explored is the impact of the protocol on
efficiency and market power. Two ways of learning are commonly used for the
agents, and authors sometimes disagree over which one to choose: either social with
a Genetic Algorithm, or individual with reinforcement learning. While it is already
well known that this has a huge impact on global results (Vriend 2000), in this
chapter we cannot decide on the best choice to make. However, to our view, most
results cannot really be extended to real life design since the representation of
learning for agents can be badly adapted to the application context (necessity to
have long learning in case of GA or even reinforcement learning).
One original approach that is cited by Marks (2007) is the “evolutionary
mechanism design” (Phelps et al. 2002), where the strategies of three types of
actors – sellers, buyers and auctioneers – are all submitted to evolution and
selection (the fitness of the auctioneer’s strategy being linked to the total profit
of the participants). This approach is logically different since the protocol itself (via
the auctioneer) is what evolves to get to a better result, with the characteristics
of the participants being fixed (relative number of each and relative production and
demand).
It is interesting to note that another branch of research deals with the representa-
tion of individual agents on large markets, and is also quite close to an idea of
design of markets, but from the opposite perspective: by introducing agents into
real markets. Computer scientists interested in the analysis of cognition have the
goal of making artificial agents as efficient as possible in a context of bidding in
auctions, both from the point of view of the seller and the buyer (Kephart and
Greenwald 2002). They are usually not interested in understanding human
behaviour and decisions, but rather in explaining the properties that can emerge
in markets in which many artificial learning agents interact (with each other or
humans), differentiating their strategies, getting heterogeneous payoffs and creating
interesting price dynamics. The focus lies mainly on information treatment. This
applied approach is interesting in that many of its algorithms can also be used for
economic analysis in the framework of models of the type that have been explored
here. However, the aim is slightly different, since the indicator in the latter case is
the individual success of a strategy, whereas the indicators for the previous works
on markets are based on global properties of the system.
23.7 Concluding Remarks
This chapter is not a general review of market simulation in recent research; instead
of giving many examples, we focused on a few to show the diversity of questions,
models, rationality and eventual results that can be found in the literature, coming
from different backgrounds (classical economy, experimental economy, computer
science). The representation of a market is always linked to the purpose of the
simulation study, and there is never just one way forward. The quantity and
substitutability of goods, the possibility to interact with one or several sellers,
with other buyers, the memory of the agents themselves – all depends on the type
of issue, and this is why we have built the chapter in this manner: to give some ideas
of the issues that have been addressed up until now with agent-based simulation.
What is noticeable is the real difference between this approach and the classical
approach in economics, where the dynamics are not regarded as a central question.
However, the achievements with this new approach are now numerous enough to
prove that agent-based simulation can participate in a better understanding of
market protocols and behaviours of individuals on the market, and enhance the
institutional choices of politics. What can be noted in conclusion is that several
issues are still at stake when it comes to the representation of markets.
First, like with most simulation models, the temporal issue is huge. Most models
use discrete time to advance the simulation. This can lead to problems, for example
in an auction, where different agents might act precisely at the same time and have a
different impact on prices than when they act sequentially. Some people are
specifically working on this issue and build platforms that support a simulated
continuous time4 (Daniel 2006).
Another technical issue is the one of learning sequences of actions. In a situation
where agents evaluate their actions with profit, if they have to perform several
actions in a row (i.e. choosing a seller and then accepting a price or not), it is
impossible to decide which of these actions is the reason for a success or a failure.
Facing this issue, economic papers describe agents that associate the profit to all
actions, as if they were separated. This is clearly not very satisfying in terms of
logic, but no alternative modelling has been proposed yet.
Finally, there is a conceptual gap in all the cited models. As yet, another element
has never been taken into account in the representation of agents’ reasoning on
markets, which would fit in models where agents try to maximize their profit by
choosing the best strategies. In this case, they can scan past actions and the
following profits, or their past possible profit with all actions they could have
undertaken and then select the best action in all contexts. While the latter strategy
is a bit more general than the first one, neither lets the agents imagine that a change
in their action will modify other agents’ behaviour as well. This is strange enough,
since a lot of people interested in game theory have been working on agents in
markets, but none of them have produced models of anticipation of others’ choices.
4
Natlab, which can be found at: http://www.complexity-research.org/natlab.
646 J. Rouchier
In markets where bargaining is central, it could however be a central feature in the

understanding of real human behaviour.
Acknowledgements I wish to thank Bruce Edmonds for his patience, and Scott Moss and Sonia
Moulet for their advices.
Further Reading
Arthur (1991) is one of the first models incorporating learning agents in a market.
Lux (1998) describes a model of speculation on an asset market with interacting
agents. Duffy (2001) was the first attempt to link experimental data to simulation
results in order to evaluate the kind of learning within a speculative environment.
Jefferies and Johnson (2002) give a general overview of market models including
their structures and learning by agents. Moulet and Rouchier (2007) use data on
negotiation behaviours from a real market in order to fit the parameters of a two-
sided learning model. Finally, Kirman (2010) summarises many interesting
dimensions that can be captured using agent-based models.
References
Arifovic J (1996) The behaviour of the exchange rate in the genetic algorithm and experimental
economies. J Polit Econ 104(3):510–541
Arthur WB (1991) On designing economic agents that behave like human agents: a behavioural
approach to bounded rationality. Am Econ Rev 81:353–359
Arthur WB (1994) Inductive reasoning and bounded rationality. Am Econ Rev 84:406
Arthur WB, Durlauf S, Lane D (eds) (1997a) The economy as an evolving complex system II
(Santa Fe institute studies in the sciences of complexity, XXVII). Addison-Wesley, Reading
Arthur WB, Holland J, LeBaron B, Palmer R, Taylor P (1997b) Asset pricing under endogenous
expectations in an artificial stock market. In: Arthur et al (1997a), pp 15–44.
Axtell R (2005) The complexity of exchange. Econ J 115(June):193–210
Basci E (1999) Learning by imitation. J Econ Dyn Control 23:1569–1585
Becker G (1962) Irrational behaviour and economic theory. J Polit Econ 70:1–13
Bendor J, Mookherjee D, Ray D (2001) Aspiration-based reinforcement learning in repeated
interaction games: an overview. Int Game Theory Rev 3(2–3):159–174
Boylan R, El-Gamal M (1993) Fictitious play: a statistical study of multiple economic
experiments. Game Econ Behav 5:205–222
Brenner T (2002) A behavioural learning approach to the dynamics of prices. Comput Econ
19:67–94
Brenner T (2006) Agent learning representation: advice on modelling economic learning. In:
Tesfatsion and Judd (2006), Chapter 18
Brewer PJ, Huang M, Nelson B, Plott C (2002) On the behavioural foundations of the law of
supply and demand: human convergence and robot randomness. Exp Econ 5:179–208
Brock WA, Hommes C (1998) Heterogeneous beliefs and routes to chaos in an asset pricing
model. J Econ Dyn Control 22:1235–1274
Camerer C, Ho TH (1999) Experience-weighted attraction learning in normal form games.

Econometrica 67:827–874
Chan NT, Shelton CR (2001) An electronic market-maker. Technical report, AI Memo 2001–005,
MIT AI Lab, Cambridge
Cyert RM, March JG (1963) A behavioural theory of the firm. Prentice-Hall, Englewood Cliffs
Daniel G (2006) Asynchronous simulations of a limit order book. PhD thesis, University of
Manchester, http://gillesdaniel.com/PhD.html
Darmon E, Torre D (2004) Adoption and use of electronic markets: individual and collective
learning. J Artif Soc Soc Simul, 7(2), http://jasss.soc.surrey.ac.uk/7/2/2.html
Delre SA, Jager W, Bijmolt THA, Janssen MA (2007) Targeting and timing promotional
activities: an agent-based model for the takeoff of new products. J Bus Res 60:826–835
Duffy J (2001) Learning to speculate: experiments with artificial and real agents. J Econ Dyn
Control 25:295–319
Duffy J, Ochs J (1999) Emergence of money as a medium of exchange: an experimental study. Am
Econ Rev 89:847–877
Easley D, Ledyard J (1993) Theories of price formation and exchange in double oral auction. In:
Friedman D, Rust J (eds) The double-auction market: institutions, theories, and evidence.
Addison-Wesley, Reading, pp 63–97
Erev I, Bereby-Meyer Y, Roth A (1998) Prediction how people play games: reinforcement
learning in games with unique strategy equilibrium. Am Econ Rev 88:848–881. http://ie.
technion.ac.il/Home/Users/erev/erev_roth_aer98.pdf
Friedman M (1953) The methodology of positive economics. In: Friedman M (ed) Essays in
positive economics. The University of Chicago Press, Chicago, pp 3–43 [Republished in Breit
W, Hochman HM (eds) (1971) Readings in microeconomics, 2nd edn. Holt, Rinehart and
Winston, New York, pp 23–47]
Geertz C, Geertz H, Rosen L (1979) Meaning and order in Moroccan society: three essays in
cultural analysis. Cambridge University Press, Cambridge
Giardini F, Conte R, Paolucci M (2013): Reputation. Chapter 15 in this volume
Gintis H (2006) The emergence of a price system from decentralized bilateral exchange. B. E.
J Theor Econ 6(1):1–15, http://works.bepress.com/hgintis/1/
Gode DK, Sunder S (1993) Allocative efficiency of markets with zero intelligence traders: markets
as a partial substitute for individual rationality. J Polit Econ 101:119–137
Gode DK, Sunder S (2004) Double-auction dynamics: structural effects of non binding price
controls. J Econ Dyn Control 28:1707–1731
Hoffmann AOI, Jager W, von Eije JH (2007) Social simulation of stock markets: taking it to the
next level. J Artif Soc Soc Simul 10(2), http://jasss.soc.surrey.ac.uk/10/2/7.html
Hommes C (2001) Financial markets as complex adaptive evolutionary systems. Quant Financ
1:149–167
Hommes C (2007) Bounded rationality and learning in complex markets (Working paper, 07–01),
Center for Nonlinear Dynamics in Economics and Finance (CeNDEF), University of
Amsterdam, http://www1.fee.uva.nl/cendef/publications/papers/Handbook_Hommes.pdf
Hommes C, Lux T (2008) Individual expectations and aggregate behavior in learning to forecast
experiments (Working paper, 1466), Kiel Institute for the World Economy http://www.ifw-
members.ifw-kiel.de/publications/individual-expectations-and-aggregate-behavior-in-learning-
to-forecast-experiments/KWP_1466_Individual%20Expectations.pdf
Izquierdo LR, Gotts NM, Gary Polhill J (2004) Case-based reasoning, social dilemmas, and a new
equilibrium concept. J Artif Soc Social Simulat 7(3). http://jasss.soc.surrey.ac.uk/7/3/1.html
Jager W (2007) The four P’s in social simulation, a perspective on how marketing could benefit
from the use of social simulation. J Bus Res 60:868–875
Janssen M, Jager W (2001) Fashion, habits and changing preferences: simulations of psychologi-
cal factors affecting market dynamics. J Econ Psychol 22:745–772
Jefferies P, Johnson NF (2002) Designing agent-based market models. http://arxiv.org/abs/cond-
mat/0207523
648 J. Rouchier
Kahneman D, Tversky A (1979) Prospect theory: an analysis of decision under risk. Econometrica
2:263–292
Kawagoe T, Wada S (2005) A counterexample for the bullwhip effect in a supply chain. In:
Mathieu P, Beaufils B, Brandouy O (eds) Artificial economics, agent-based methods in finance,
game theory and their applications (Lecture notes in economics and mathematical systems,
564), Springer, Berlin, pp 103–111
Kephart J, Greenwald A (2002) Shopbot economics. Auton Agent Multi-Agent Syst 5(3):255–287
Kirman A (1991) Epidemics of opinion and speculative bubbles in financial markets. In: Taylor M
(ed) Money and financial markets. Macmillan, London, pp 354–356, Chapter 17
Kirman A (2001) Some problems and perspectives in economic theory. In: Debreu G, Neuefeind
W, Trockel W (eds) Economic essays: a Festschrift for Werner Hildenbrand. Springer, Berlin,
pp 231–252
Kirman A (2010) Complex economics: individual and collective rationality. Routledge, London
Kirman A, Schulz R, Hardle W, Werwatz A (2005) Transactions that did not happen and their
influence on prices. J Econ Behav Org 56:567–591
Kirman A, Vriend NJ (2001) Evolving market structure: an ACE model of price dispersion and
loyalty. J Econ Dyn Control 25(3–4):459–502
Kiyotaki N, Wright R (1989) On money as a medium of exchange. J Polit Econ 97:924–954
Kopel M, Dawid H (1998) On economic applications of the genetic algorithm: a model of the
cobweb type. J Evol Econ 8(3):297–315
LeBaron B (2001) Evolution and time horizons in an agent-based stock market. Macroecon Dyn 5
(2):225–254
LeBaron B, Arthur WB, Palmer R (1999) The time series properties of an artificial stock market.
J Econ Dyn Control 23:1487–1516
Leloup B (2002) Dynamic pricing with local interactions: logistic priors and agent technology. In:
Arabnia HR, Mun Y (eds) Proceedings of the international conference on artificial intelligence,
IC-AI ’02, vol 1. CSREA Press, Las Vegas, 24–27 June 2002, pp 17–23
Levy M, Levy H, Solomon S (2000) Microscopic simulation of financial markets: from investor
behaviour to market phenomena. Academic, San Diego
Lin F-R, Lin S-M (2006) Enhancing the supply chain performance by integrating simulated and
physical agents into organizational information systems. J Artif Soc Soc Simul 9(4), http://
Lux T (1998) The socio-economic dynamics of speculative markets: interacting agents, chaos and
the fat tails of return distribution. J Econ Behav Organ 33(2):143–165
Lux T (2009) Stochastic behavioral asset-pricing models and the stylized facts. In: Hens T,
Schenk-Hoppé KR (eds) Handbook of financial markets: dynamics and evolution. Elsevier,
Burlington, pp 161–215
Lux T, Marchesi M (1999) Scaling and criticality in a stochastic multi-agent model of a financial
market. Nature 397:498–500
Lux T, Marchesi M (2000) Volatiliy clustering in financial markets: a micro-simulation of
interacting agents. Int J Theor Appl Financ 3:675–702
Macy M, Flache A (2002) Learning dynamics in social dilemmas. Proc Natl Acad Sci
99:7229–7236
Macy MW, Benard S, Flache A (2013) Learning. Chapter 17 in this volume
Marks R (2007) Market design using agent-based models. In: Tesfatsion L, Judd KL (eds)
Handbook of computational economics volume 2: agent-based computational economics.
Elsevier/North Holland, Amsterdam, pp 1339–1380
Moss S (2002) Policy analysis from first principles. PNAS 99(3):7267–7274
Moulet S, Kirman A (2008) Impact de l’organisation du marché: Comparaison de la négociation de
gré à gré et les enchères descendantes (Document de travail, 2008–56). GREQAM, Marseille
Moulet S, Rouchier J (2008) The influence of sellers’ beliefs and time constraint on a sequential
bargaining in an artificial perishable goods market. J Econ Dyn Control 32(7):2322–2348
Myerson RB, Satterthwaite MA (1983) Efficient mechanisms for bilateral trading. J Econ Theory
29:265–281
Nicolaisen J, Petrov V, Tesfatsion L (2001) Market power and efficiency in a computational
electricity market with discriminatory double-auction pricing. IEEE Trans Evolut Comput
5:504–523
Oliver J (1997) Artificial agents learn policies for multi-issue negotiation. Int J Electron Commer 1
(4):49–88
Palmer R, Arthur WB, Holland JH, LeBaron B, Taylor P (1993) Artificial economic life: a simple
model for a stock market. Physica D 70:264–274
Phan D (2003) From agent-based computational economics towards cognitive economics. In:
Bourgine P, Nadal J-P (eds) Cognitive economics. Springer, Berlin, pp 371–398
Phelps S, Parsons S, McBurney P, Sklar E (2002) Co-evolution of auction mechanisms and trading
strategies: towards a novel approach to microeconomic design. In: Smith RE, Bonacina C,
Hoile C, Marrow P (eds) Proceedings of ECOMAS-2002 workshop on evolutionary computa-
tion in multi-agent systems, at genetic and evolutionary computation conference (GECCO-
2002), International society for genetic and evolutionary computation
Pinyol I, Paolucci M, Sabater-Mir J, Conte R (2007) Beyond accuracy: reputation for partner
selection with lies and retaliation. In: Antunes L, Paolucci M, Norling E (eds) Multi-agent-
based simulation VIII, International workshop, MABS 2007, Honolulu, 15 May 2007, Revised
and invited papers. (Lecture notes in computer science, 5003). Springer, Berlin, pp 128–140
Roth A, Prasnikar V, Okuno-Fujizara M, Zamie S (1991) Bargaining and market behavior
in Jerusalem, Ljubljana, Pittsburgh, and Tokyo: an experimental study. Am Econ Rev
81(5):1068–1095
Rothschild M (1974) A two-armed Bandit theory of market pricing. J Econ Theory 9:185–202
Rouchier J (2003) Reimplementation of a multi-agent model aimed at sustaining experimental
economic research: the case of simulations with emerging speculation. J Artif Soc Soc Simul
Rouchier J, Hales D (2003) How to be loyal, rich and have fun too: the fun is yet to come (Working
Paper, 03B13). Groupement de Recherche en Économie Quantitative d’Aix-Marseille
(GREQAM), France, http://www.greqam.fr/IMG/working_papers/2003/03b13.pdf
Rouchier J, Mazaud J-P (2004) Trade relation and moral link in transactions among retailers and
wholesale sellers on the Arnavaux market. In: 11th world congress for social economics,
Albertville, 8–11 June 2004
Rouchier J, Robin S (2006) Information perception and price dynamics in a continuous double
auction. Simul Gaming 37:195–208
Rouchier J (2013) The interest of having loyal buyers in a perishable market. Comput Econ
41(2):151–170
Rubinstein A, Wolinksy A (1985) Equilibrium in a market with sequential bargaining.
Econometrica 53(5):1133–1150
Sallans B, Pfister A, Karatzoglou A, Dorffner G (2003) Simulation and validation of an integrated
markets model. J Artif Soc Soc Simul 6(4), http://jasss.soc.surrey.ac.uk/6/4/2.html
Samanidou E, Zschischang E, Stauffer D, Lux T (2007) Agent-based models of financial markets.
Rep Prog Phys 70:409–450
Simon H (1955) A behavioural model of rational choice. Q J Econ 69:99–118
Takahashi H, Terano T (2003) Agent-based approach to investors’ behavior and asset price
fluctuation in financial markets. J Artif Soc Soc Simul 6(3), http://jasss.soc.surrey.ac.uk/6/3/
3.html
650 J. Rouchier
Tesfatsion L, Judd KL (2007) Handbook of computational economics volume 2: agent-based

computational economics. Elsevier/North Holland, Amsterdam
Vriend NJ (2000) An illustration of the essential difference between individual and social learning
and its consequences for computational analyses. J Econ Dyn Control 24:1–19
Vriend NJ (2007) ACE models of endogenous interactions. In: Tesfatsion L, Judd KL (eds)
Handbook of computational economics volume 2: agent-based computational economics.
Elsevier/North Holland, Amsterdam, pp 1047–1079
Weisbuch G, Kirman A, Herreiner D (2000) Market organization and trading relationships. Econ J
110(463):411–436
White H (1988) Varieties of markets. In: Wellman B, Berkowitz SD (eds) Social structures:
a network approach. Cambridge University Press, New York
Chapter 24
Movement of People and Goods
Linda Ramstedt, Johanna Törnquist Krasemann, and Paul Davidsson
Why Read This Chapter? To gain an overview of approaches to the simulation of

traffic and transportation by way of representative examples. Also, to reflect the
characteristics and benefits of using social simulation as opposed to other methods
within the domain. The chapter will inform both researchers and practitioners in the
traffic and transportation domain of some of the applications and benefits of social
simulation and relevant issues.
Abstract Due to the continuous growth of traffic and transportation and a thus
increased urgency to analyze resource usage and system behavior, the use of
computer simulation within this area has become more frequent and acceptable.
This chapter presents an overview of modeling and simulation of traffic and
transport systems, and focuses in particular on the imitation of social behavior
and individual decision making in these systems. We distinguish between transport
and traffic. Transport is an activity where goods or people are moved
between points A and B while traffic is referred to as the collection of several
transports in a common network such as a road network. We investigate to what
extent and how the social characteristics of the users of these different traffic and
transport systems are reflected in the simulation models and software. Moreover,
we highlight some trends and current issues within this field and provide further
reading advice.
L. Ramstedt (*)
Transport and Society, Vectura Consulting, Solna, Sweden
J.T. Krasemann
School of Computing, Blekinge Institute of Technology, Karlshamn, Sweden
P. Davidsson
Department of Computer Science, Malmö University, Malmö, Sweden

652 L. Ramstedt et al.
24.1 Introduction
The continuous growth of traffic and transportation is increasing the interest in limiting
their negative impact on society. Moreover, the different stakeholders involved in traffic
and transportation are interested in utilizing the available resources in the best way
possible. This has stimulated the development and use of both various policies and
advanced transport infrastructure systems, as well as the deployment of information and
communication technologies. In order to examine the effects of such developments prior
to, or during implementations, computer simulations have shown to be a useful approach.
This chapter addresses modeling and simulation of traffic and transport systems, and
focuses in particular on the imitation of social behavior and individual decision-making.
Traffic and transport systems typically involve numerous different stakeholders
and decision-makers that control or somehow affect the systems, and the prevalence of
social influence is thus significant. In order to study the behavior of such systems and
model them it then becomes necessary to capture and include the significant social
aspects to some extent, in addition to the flow of traffic or transport units that constitute
the backbone of such systems. In this context, we consider a traffic or transport system
as a society, consisting of physical components (e.g., cars, buses, airplanes or parcels)
and social components (e.g., drivers, passengers, traffic managers, transport chain
coordinators, or even public authorities) where the interactions between them may
play an important role. The social components determine the physical flow in their
common environment (e.g., road or rail networks, and terminals) in line with external
restrictions, internal intentions and so forth. Depending on the purpose of the simula-
tion study and the sophistication and detail that is desired in the model, there are
different approaches to incorporate social influence in such systems.
The purpose of this chapter is to present an overview of when and how social
influence and individual behavior within the domain of traffic and transportation
have been modeled in simulation studies. We also provide further reading advice to
related approaches and highlight current issues. We divide the domain into trans-
port and traffic. Transport is an activity where goods or people are moved between
points in one or several traffic modes (road, rail, waterborne, or air). The types of
vehicles we consider are train, truck, car, bus, ship, ferry, and airplane. While
transport refers to the movement of something from one point to another, traffic
refers to the flow of different transports within a network. One train set is thus a
transport (and possibly also part of a transport chain) that takes part in the train
traffic flow. Hence, a transport can be part of several traffic networks (air, water-
borne, road or rail) and a traffic network comprises several transports.
Typical points of interest related to transportation in this context are, for
instance, to predict consequences of transport policies aimed at mediating govern-
mental goals or objectives of companies, and to design and synchronize bus
timetables at certain stations to facilitate passenger flows. A typical issue that is
interesting to study in the traffic context is the impact of driving assistance systems
where the driving behavior in road networks is studied.
In the following sections, the types of simulation approaches that exist within the
traffic and transportation domain, and the motivation behind them, are presented.
24 Movement of People and Goods 653
The social aspects of the studied systems and models, and how they have been
accounted for are also described here. Finally, a concluding discussion is presented
along with some further reading advice.
24.2 Traffic
This section addresses the modeling and simulation of individual and social behav-
ior within traffic systems. There are several types of simulation models developed
and applied which have different granularity; i.e., macro-, meso- and micro-level
models. The first two typically represent the traffic behavior by the use of equations
which are based on aggregated data. Thus, the traffic is modeled as a collection of
rather homogenous entities in contrast to the microscopic models which more in-
depth consider the individual characteristics of the traffic entities and how these
influence each other and the traffic system. Since this handbook mainly addresses
the modeling of social aspects, we will focus on microscopic models.
Within the domain of traffic system simulation, the dominating focus is on simula-
tion of road traffic and car driver behavior to evaluate the quality-of-service in road
traffic networks (Tapani 2008). The development and implementation of ADAS
(Advanced Driver Assistance Systems), ATIS (Advanced Traveler Information
Systems), and road traffic control regimes, such as congestion taxation, stipulates an
increasing interest in sophisticated simulation models. There are also approaches that
study the behavior of traffic systems during extraordinary situations such as urban
evacuations. Since users of road traffic systems normally do not communicate and
interact with each other directly but rather indirectly due to the restrictions of the
traffic system, the design of agent communication and protocols is mostly not consid-
ered in these simulation models. Focus is instead on individual behavior models.
The attention given to the simulation of social aspects in other modes of traffic
such as air and railway traffic is, however, very limited and we have not found any
publications focusing on this specific subject. One reason may be that the traffic in
these modes to a large extent is managed by central entities, such as, traffic
controllers. In this perspective, the behavior and interaction of the individual
vehicles are less interesting to study. However, the interaction between traffic
controllers in these traffic modes seems to be a domain that needs more attention.
Below we will provide a more in-depth presentation of social simulation in road
traffic. Since the movement of people associated with a vehicle is the focus in this
chapter, simulation of pedestrians and crowds will only be presented briefly in the
last section on related research.
24.2.1 Road Traffic Simulation
In road traffic simulations, there are several distinctions made. First, there is a
distinction made with respect to the problem area in focus and whether it concerns
urban, rural or motorway traffic. Furthermore, depending on the infrastructure

modeled, there is a distinction between intersection, road and network model. As
an example, urban traffic is often modeled as a network (Pursula 1999).
One can see that there are three main categories of simulation models with
increasing level of detail in driving behavior. First, there are the empirical macro-
scopic traffic models (e.g., software like TRANSYT provided by TRL Software) that
focus on traffic flow analysis. Then we have the extended microscopic models with
capabilities of representing individual driver behavior by use of different sub-models
or rules for speed adaptation, car-following, lane change, as well as intersection and
roundabout movements if relevant (e.g., VISSIM for urban and freeway environments
and TRARR, TWOPAS, and VTISim for rural road environments (Tapani 2008)).
The third type of models has an even more complex representation of driver behavior
by use of, e.g., neural networks (NN) (Lee et al. 2005; Panwai and Dia 2007) or
discrete choice models (Dia 2002; Lee et al. 2005). In some cases these behavior
models can be dynamically configured to imitate driver behavior adaptation and the
effect of learning from experience, see, e.g., the work by Rossetti et al. (2000).
These three types of models do complement each other, but with the growing
need to evaluate the impact of investments in ITS (Intelligent Transport Systems)
such as ATIS and ADAS, and policies such as road pricing, the need to reflect the
complexity of driver behavior in more detail becomes apparent (Tapani 2008). The
current behavior sub-models for acceleration, lane changing, car-following and so
forth are mainly equation-based with threshold values. Toledo (2007) claims in a
review of state-of-the-art in road traffic driver behavior modeling that, in many
cases these sub-models are insufficient to adequately capture the sophistication of
drivers and the impact of long-term driving goals and considerations. Hencefor-
ward, we will focus on the approaches that emphasize such advanced driving
behavior modeling and refer to Pursula (1999), Mahmassani (2005), Toledo
(2007) and Tapani (2008) for more information about related research.
In the third category of approaches mentioned, the individual drivers are usually
modeled as individual autonomous vehicles represented by intelligent agents. The
main differences between the equation- or rule-based behavior sub-models and the
agent-based models are the increased reasoning capabilities and planning horizon
considered. The traffic network flow could in either case be based on techniques like
cellular automata (CA) (Nagel and Schreckenberg 1992; Esser and Schreckenberg
1997) and queuing theory while the decision-making of the agent is based on a
possibly more diverse set of objectives and influencing parameters. Sometimes the
agents use simple decision rules acting in a reactive manner rather than having
sophisticated reasoning capabilities. We refer to Ehlert and Rothkrantz (2001), El
hadouaj et al. (2000), Wahle and Schreckenberg (2001), and Kumar and Mitra (2006),
for examples of such reactive behavior. However, these approaches are not considered
part of the third category in this context and this illustrates that agent technology does
not necessarily imply, but only enables, a representation of more complex behavior.
The approaches and their models often focus on a certain traffic infrastructure
depending on which aspect to study. Below we address three main topics: Simula-
tion of road traffic in intersections, evaluation of ATIS and ADAS, and
benchmarking of behavior modeling techniques.
24.2.1.1 Simulation of Road Traffic in Intersections
Urban traffic networks typically involve intersections with a complex coordination

and interplay between vehicles. This is challenging to model and it is difficult to
prevent deadlocks from occurring during the simulation. One way to handle this is
to use game theory and agents. Mandiau et al. (2008) propose a distributed coordi-
nation and interaction mechanism for the simulation of vehicles in T- and X-shaped
intersections in urban road traffic networks. The vehicles are represented by agents
who have a choice of one of two actions: brake or accelerate. The coordination
between the vehicles crossing the intersection is based on game theory and a 2x2
decision matrix for each pair of vehicles is used to compute the decisions. The
approach is also extended to involve a larger traffic volume of n agents (vehicles),
where the memory requirements then increases rapidly due to the n(n 1)/2 deci-
sion matrices. The mechanism does not prevent occurrence of deadlocks, but is able
to resolve them.
Bazzan (2005) does also propose a distributed game theory-based approach for
coordination of road traffic signals at X-shaped intersections where vehicles are
represented by autonomous intelligent agents. The agents have both local and
global goals and are adaptive in the sense that they are equipped with a memory
to register the outcome (i.e. payoff) of executed actions and via learning rules they
are able to incorporate their experience in the decision-making. Game theory has
also been used in this context to model the route choice behavior (Schreckenberg
and Selten 2004).
24.2.1.2 Evaluation of ATIS and ADAS
The benefits from and application of ATIS and ADAS have become more known
and common in modern road traffic systems. An implementation is, however, often
associated with large investments and limited insight in how the system and its
users would respond to such an implementation in the short and long term. The use
of simulation may then offer some information on possible consequences to guide
the stakeholders. A typical question to address is how the individual drivers would
change their choice of traveling with an increased access to traffic information. Dia
(2002) investigates the implications of subjecting drivers to real-time information
in a road traffic network. Each vehicle is represented by a BDI agent (Beliefs-
Desires-Intentions), which has a static route-choice multi-nomial logit model (Ben-
Akiva and Lerman 1985). Static refers to that the multi-nomial logit model is not
updated during simulation and thus not making use of driver/vehicle experience.
Panwai and Dia (2005) extend the approach using NN and then further improve it
by giving each agent a memory and a dynamic behavior model based on NN (Dia
and Panwai 2007). The behavior model is updated accordingly during the agent’s
journey allowing the agent to act on the information provided by an ATIS by re-
evaluating its strategy en-route and possibly changing route if possible and
beneficial.
Rossetti et al. (2000) focus on the departure time and route choice model of
commuters in a road traffic network using the DRACULA (Dynamic Route
Assignment Combining User Learning and microsimulAtion) simulation model.
The drivers are represented by autonomous BDI agents, some of which have access
to ATIS and thus more information about the current traffic situation than other
drivers. The agent behavior model is dynamic in the sense that it incorporates the
experience of the commuting driver on a daily basis, but does not allow the driver to
change strategy en-route. Before the commuter starts its journey and decides when
to depart and which route to choose, it compares the predicated cost of choosing its
usual daily route with the cost of the alternative route. If the cost of any of the
alternatives is significantly lower than the cost of the daily route, the driver chooses
that. Otherwise, it stays with its usual route. Once the agent has decided on
departure time and route, it will follow that guided by car-following and lane-
changing rules and cannot change its mind.
Chen and Zhan (2008) use the agent-based simulation system Paramics to
simulate the urban evacuation process for a road traffic network and the drivers’
route choice and driving behavior. The vehicles are modeled as individual agents
that have local goals to minimize their travel time, i.e. to choose the fastest way.
Based on traffic information, a car-following model, network queuing constraints
and a behavior profile (e.g. aggressive or conservative driving style) each agent
dynamically, en-route, re-evaluates its driving strategy and route choice.
24.2.1.3 Benchmarking of Behavior Modeling Techniques
Since a number of alternative methods to model and simulate road traffic exist, a
few studies have focused on comparisons to evaluate the strengths and weaknesses
of some alternatives. Lee et al. (2005) compare and evaluate the use of three
different route choice models for road traffic drivers with access to trip information:
A traditional multinomial logit model, a traditional NN approach and NN combined
with a genetic algorithm (GA). The initial attitude of the authors indicates a
preference for the combined NN-GA solution proposing that it is better suited to
consider the influence of non-linearity and obscurity in drivers’ decision-making.
Based on their simulation experiments and the mean square error of the different
route choice models, the NN-GA approach was said to be most appropriate.
Hensher and Ton (2000) make a similar comparison of the predictive potential of
NN and nested logit models for commuter mode choice. They could not make any
judgment about which method that is most appropriate but concluded that both
requires a lot of historic data to train or construct the models. The characteristics of
NN are that they are better at handling noisy or lack of data than discrete choice
models were discussed.
For more in-depth information about human behavior in traffic we refer to
Schreckenberg and Selten (2004) while Boxill and and Yu (2000) present an
overview of road traffic simulation models and Koorey (2002) an overview of
software.
24.3 Transportation
In this section, approaches to simulating transportation systems are described, which

include both transportation of freight and people. Transportation is often described as
road, rail, waterborne or air transportation, but transportation can also be intermodal.
Intermodal transportation refers to a transport chain of two or more modes of
transport where some modal shift activity takes place, for instance at a terminal
where loading and unloading of goods is done, or at a train station where passengers
transfer from train to bus. Supply chains are related to freight transportation, even if
the focus mainly is on the product and its refinement processes, in contrast to freight
transportation where the focus is on the vehicle and its operations.
Issues in the field of passenger transportation typically concern evaluation of
policies for more efficient bus timetables and pricing policies. Another field within
the domain is emergency transportation, which often concerns the planning of
resources, such as ambulances or fire engines, in order to serve people in need
efficiently with respect to costs, coverage equity and labor equity. Other issues are
the evaluation of different policies, such as dispatching policies. Goldberg (2004)
has reviewed operations research approaches for emergency transportation, and
claims that mathematical programming currently is the dominating method used.
He also states that simulation is a promising approach for future work related to
emergency transportation, especially for vehicle relocation and dispatching, due to
the complexity of the problem domain. However, we have not found any papers
describing approaches of social simulation in the field of emergency transportation.
In papers describing simulation approaches to transportation with a focus on
social aspects and individual behavior, we have only found models for road, rail,
and waterborne transportation, i.e., we have not found any air transportation
models.
24.3.1 Freight Transportation
Different approaches are used when modeling and simulating transportation.

A common approach when analyzing transportation is the so-called four-step
approach (production/attraction, distribution, modal split, and assignment) which
primarily is developed for passenger transportation, but also used for freight
transportation (de Jong and Ben-Akiva 2007). This approach is on the macroscopic
level, i.e. averaged characteristics of the population are in focus and aggregated
data is used, in contrast to a microscopic level, which would include more details on
the individuals of the population. A trend in the field of freight transport simulation
for predicting probable consequences of transport policies is to include more
details, see e.g. (de Jong and Ben-Akiva 2007) and (Hunt and Gregor 2008).
However, most of these approaches are still macroscopic approaches, but with
some microscopic characteristics (Ramstedt 2008).
Traditional simulation approaches for simulating supply chains are discrete-

event simulation and dynamic simulation (Terzi and Cavalieri 2004). In such
models the behavior of the individuals is often only modeled as a set of actions
related to a probability function of being executed, not capturing causal behavior of
the simulated system.1 Moreover, interactions between the individuals, such as
negotiations, are not explicitly modeled in traditional supply chain models. Since
modeling and simulating social aspects is the main concern here, we focus on
simulation approaches that address interactions between individuals.
Most simulation approaches of freight transportation have a descriptive purpose,
such as predicting the effects of different kinds of policies. For instance,
Gambardella et al. (2002) who simulate intermodal transportation make use of
multi-agent-based simulation to examine policies aimed at improving the
operations at terminals, while simulation approaches of supply chains
(Swaminathan et al. 1998; van der Zee and van der Vorst 2005) often focus on
evaluating strategies such as VMI (Vendor Managed Inventories). Such studies are
mainly of interest for private companies, even if they can be of interest for public
authorities as well. Another example is provided by Holmgren et al. (2012) which
studies the possible effects of transport policies in transport chains, which is of
interest for public authorities. In this approach the decision making of actors in
transport chains, for instance regarding traffic mode choice, selection of supplier,
etc., are simulated. New prerequisites as a consequence of transport policies have
the potential to change these decisions so that a different system behavior occurs.
There are also examples of models with a prescriptive purpose, such as to support
the transport planning in order to improve the efficiency of the usage of transport
resources (Fischer et al. 1999).
The simulated system in transport approaches typically consists of a network of
links and nodes served by resources such as vehicles, which have a spatial explicit-
ness and are time-dependent. In supply chains the focus is more on the nodes and
their processes, while the links are not explicitly considered.
In the domain of freight transportation and supply chains, the decision making of
stakeholders is typically modeled and simulated. To model the decision making,
agents representing real-world roles in transportation, such as customers, transport
planners, transport buyers, and producers are typically implemented. Only in a few
cases are physical entities, such as vehicles, also modeled as agents (Gambardella
et al. 2002). Tasks which are commonly performed by the agents are selecting
(1) which resources (e.g. terminals, transport and production resources) to use, and
(2) how these resources should be used considering time, cost, availability, etc.
These tasks are often performed as a consequence of a customer request with an aim
of satisfying the demand based on cost, time, and availability. The behaviors of the
agents are often implemented in terms of various types of algorithms or decision
rules. Typically the agents try to minimize their costs, e.g. labor or fuel costs, which
occur as a consequence of performing a transport task between two nodes.
1
See Chap. 3 in this handbook (Davidsson and Verhagen 2013) for further general discussion.
Of course restrictions and requirements from other agents – concerning for instance
time of delivery – are also taken into account. The implemented algorithms in the
agents can be rather complex, with optimization techniques and heuristics used to
compute the decisions and actions (Holmgren et al. 2012; Fischer et al. 1999).
If some of the simulated agents represent physical entities, the locations of the
entities as well as their characteristics are typically modeled. If the agents represent
decision makers, the responsibilities of the agents are typically modeled, e.g. the
responsibility of certain types of vehicles on certain infrastructure segments.
The interactions between the agents often take place as negotiations concerning,
for instance, the cost and time of performing a task, such as transportation between
two nodes. The negotiations are then carried out according to interaction protocols
and the corresponding information exchange between the agents is also modeled.
As an example, a customer agent requests information concerning possible trans-
port solutions, or a transport planning agent requests information concerning
available vehicles for the transport task (Holmgren et al. 2012).
The agent-based simulation models are implemented in different ways, in some
cases multi-agent-based simulation platforms are used (e.g. Strader et al. 1998),
while multi-agent system platforms are used in other cases (e.g. Holmgren et al.
2012; Fischer et al. 1999). It is also possible to implement the agent model without
any platform, see e.g. (Swaminathan et al. 1998; Gambardella et al. 2002; van der
Zee and van der Vorst 2005). Using such platforms may often facilitate the
implementation of the model. However, if the model is very complex it can also
cause problems due to the structure of the platform; see for example (Davidsson
et al. 2008) for further discussion.
24.3.2 Passenger Transportation
While more work has been done regarding modeling and simulation of freight
transportation, there are some approaches concerning the transportation of
passengers using rail and road. These are described here.
One example of the simulation of bus transportation is Meignan et al. (2007)
where the main purpose of the simulations is to evaluate bus networks. The bus
networks are assessed based on the interests of travelers, bus operators and authorities
with respect to, for instance, accessibility, travel time, passenger waiting time, costs,
and profit. Therefore, a potential type of user of the simulation model is the manager
of a bus network. The model includes a road network, a bus network, and a traveler
network. Bus stops, bus stations, bus lines and itineraries are part of the modeled
system. There are two agent types: buses and passengers. A typical task of the bus
agent is to perform a bus trip. The networks include spatial explicitness, and time is
important for the agents since the bus routes are determined by timetables, and the
passengers have a need for travel at certain points in time. The model combines
micro- and macroscopic approaches since the global traffic situation is taken into
consideration in addition to the individual transports. Traffic assignment and modal
choice of the overall demand is made on the macroscopic level with a discrete choice
model. Interactions take place between the buses and passengers. For instance, the
bus agents have to consider the travel demand of passenger agents, and the passenger
agents have to consider the available bus agents. Moreover, the actual loading and
unloading of passengers is one kind of interaction. A similar approach is presented by
Gruer et al. (2001) where the mean passenger waiting time at bus stops is evaluated.
In this approach buses, stops, and road sections are modeled as agents, with the focus
on the activities at the bus stops.
Work is also done in the taxi domain where for instance Jin et al. (2008) present a
simulation model for planning the allocation of taxis to people requesting taxi
transportation. The model can be used for evaluating different planning policies
for how the allocation should be made taking issues like vehicle usage, passenger
waiting time, travel time, and travel distance into consideration. Four types of
agents are included in the model; the user agent, the node-station agent, the taxi
agent, and the transport administration agent. The taxi agent represents the physical
taxis, while the other agents have different planning functions. The different agent
types have different goals and agreements are reached by negotiations between the
agents through the user agent.
An example of a social simulation of rail passenger transportation is presented
by Li et al. (2006). In this, an activity-based model is outlined for the evaluation of
pricing policies and how they affect the traveler behavior. This approach is similar
to an agent-based approach where passengers are modeled as agents. The focus in
the model is on the traveler behavioral model, where the characteristics and
preferences of travelers and their activities, in terms of activity schedules, are
modeled. Typical tasks the traveler agent performs are scheduling and planning
the journeys as well as executing these activities. The decisions are typically made
based on generalized costs. In the presented model the traveler agents can interact
with for instance a tool that provides information regarding available travel options
by sending requests of possible travel options.
24.3.3 Related Research
Related to the simulation of passenger transportation is the simulation of

pedestrians and crowds. Traditionally, pedestrian simulation has used techniques
such as flow-speed-density equations thus aggregating pedestrian movement into
flows rather than a crowd of possibly heterogeneous individuals (Rindsfüser and
Klügl 2007). Rindsfüser and Klügl (2007) propose an agent-based pedestrian
simulation model for the simulation of the movement and behavior of train
passengers at the railway station in Bern, Switzerland. A similar approach is
Zhang et al. (2008), in which the alighting and boarding of passengers in the
Beijing metro is modeled and simulated. A cellular automata approach is used in
this model. Pelechano et al. (2005) also discuss social aspects in the simulation of
crowds and provide a review of different crowd simulation approaches.
One special case of crowd simulation is the simulation of emergency evacuation

situations where animal flocking behavior models are one type of models that have
been applied (see Chap. 22 in this handbook (Hemelrijk 2013) for an overview of
modeling animal social behavior). For further information on simulation of emer-
gency evacuation we refer to (Santos and Aguirre 2004) which provides a review of
current simulation models.
Another type of related work is agent-based simulation of seating preferences of
passengers in buses and airplanes (Alam and Werth 2008). The different agents are
then representing different categories of people characterized by ethnicity, age,
cultural background, and their respective seating preferences.
24.4 Discussion
In papers concerning simulation of traffic and transportation, different arguments

are given to support the use of social simulation. A common argument for making
use of multi-agent-based simulation in the transportation domain is that it enables
capturing the complex interactions between individuals, such as coordination and
cooperation (Fischer et al. 1999; Meignan et al. 2007; Li et al. 2006), and conse-
quently the emergent behavior of the system. Moreover, including autonomous and
heterogeneous individuals and their behavior is also supported with the agent
approach, as well as modeling and simulating distributed decision making which
are important in the transportation domain (Meignan et al. 2007). Since multi-agent
systems provide a modular structure of the system, the possibility to easily
exchange or reuse different parts of the simulated system for different cases is
facilitated (Swaminathan et al. 1998; van der Zee and van der Vorst 2005).
As pointed out earlier, there are new phenomena in the road traffic networks
imposed by the increasing level of technologies facilitating driving as well as new
motivations behind controlling and supervising networks and the effects of such
support systems when the infrastructure capacity becomes a scarce resource.
Toledo (2007) argues that to capture the level of sophistication, the modeling
capabilities need to improve where the use of agents can make a contribution.
In contrast to well-established software and methods for traffic and transport
simulations, the data requirements are different and data is not available to the same
extent for the newer approaches mentioned in this chapter. In addition, the more
novel approaches are all different and are often developed, used and evaluated only
by the researchers themselves. The level of maturity and acceptance reached by the
traditional approaches and software by being used and evaluated by a large number
of researchers during a long time is naturally difficult to compete with at this point.
However, using social simulation where individuals and their interactions are
explicitly modeled provides opportunities for validation due to the natural,
structure-preserving representation of the system. For instance, the behavioral
models of drivers or decision makers can be validated by the actual drivers or
real-world decision makers.
It is possible to identify some general differences between traffic and transpor-

tation applications. In traffic approaches, the agents typically represent physical
entities actually involved in the movement, i.e. vehicles or drivers are modeled as
agents. In the simulation of freight transportation, on the other hand, the agents
typically represent decision makers such as customers or transport planners. There-
fore, the physical representation of the agents is not of the same importance in these
approaches. In freight transportation approaches several agent types are typically
necessary, as opposed to traffic approaches where typically only one agent type is
modeled. Trends in the approaches to passenger transportation are not as obvious.
However, passengers are typically represented as agents (one exception is the taxi
transportation approach), and in the bus and taxi transportation approaches,
vehicles are also represented as agents. Gruer et al. (2001) represent also bus
stops and road sections as agents. Thus, physical entities may be represented as
agents, like in the traffic approaches. The number of different agent types is smaller
than in freight transportation, but there are more agent types than in traffic
approaches.
The agent behavior models are typically of a lower level (more detailed) in the
traffic approaches than in the transport approaches. One reason for this difference is
that in traffic approaches the personalities of the drivers often have a larger impact
on the system and are therefore modeled with corresponding driving behaviors such
as aggressive or calm driving style. The decision making of the agents in traffic
approaches is typically based on different rules where the choices or the planning in
transport approaches are made based on the best performance metrics such as cost
or time. The agent behaviors sometimes also include learning aspects, which
provide the agents with a dynamic behavior by the use of, e.g., NN.
In transport approaches, the focus is not on modeling the different personalities
of decision makers, but rather on modeling the different types of decision-making
roles and their associated rational behavior. Moreover, in (freight) transport
approaches to model the negotiations and the interactions between the decision
makers is crucial in order to reach a solution. In traffic approaches the interactions
between the individuals are secondary, while the models of the individuals, their
individual behavior, and consequently the system behavior are focused.
As far as we have seen, social simulation is not commonly applied for all modes
in traffic and transportation. In the traffic domain, mainly road traffic is studied.
Road traffic includes social aspects in terms of interactions between the vehicles. In
air and rail traffic, the control typically takes place in a more centralized way with
common objectives, which explain why these traffic modes include less social
aspects in this context and therefore do not benefit as much from social simulation
studies. For waterborne traffic, the infrastructure is not a scarce resource in the same
sense as in the other modes; instead the bottlenecks appear in the ports or other
terminals. In the freight transportation domain the social aspects mainly concern the
interactions and decision making of the actors in transport chains. The most
common modes that are included here are road and rail, but also waterborne
transportation is sometimes included. The types of decisions that are most often
studied in the freight transport domain are related to planning and mode choice,
which are a consequence of the interactions between actors in transportation. In

passenger transportation, planning decisions are also sometimes simulated, but
sometimes operational behavior, such as loading of passengers, is also simulated
which is a consequence of the interactions between passengers and vehicles.
Further Reading
For further information about traffic simulation we refer the interested reader to
(Chung and Dumont 2009), (Tapani 2008), (Toledo 2007) and (Koorey 2002). Terzi
and Cavalieri (2004) provide a review of supply chain simulation, while Williams
and Raha (2002) present a review of freight modeling and simulation. For general
information about transport modeling, we suggest to read (Ortúzar and Willumsen
2001). For further information on how agent technologies can be used in the traffic
and transport area, see (Davidsson et al. 2005).
References
Alam SJ, Werth B (2008) Studying emergence of clusters in a bus passengers’ seating preference
model. Transport Res C Emerg Technol 16(5):593–614
Bazzan AL (2005) A distributed approach for coordination of traffic signal agents. Auton Agent
Multi-Agent Syst 10:131–164
Ben-Akiva M, Lerman SR (1985) Discrete choice analysis: theory and application to travel
demand. MIT Press, Cambridge, MA
Boxill S, Yu L (2000) An evaluation of traffic simulation models for supporting ITS development
(Technical report, SWUTC/00/167602-1). Center for transportation training and research.
Texas Southern University, Houston
Chen X, Zhan FB (2008) Agent-based modeling and simulation of urban evacuation: relative
effectiveness of simultaneous and staged evacuation strategies. J Oper Res Soc 59(1):25–33
Chung E, Dumont A-G (2009) Transport simulation: beyond traditional approaches. EFPL Press,
Lausanne
Davidsson P, Verhagen H (2013) Types of simulation. Chapter 3 in this volume
Davidsson P, Henesey L, Ramstedt L, Törnquist J, Wernstedt F (2005) An analysis of agent-based
approaches to transport logistics. Transport Res C Emerg Technol 13(4):255–271
Davidsson P, Holmgren J, Persson JA, Ramstedt L (2008) Multi agent based simulation of
transport chains. In: Padgham L, Parkes D, Müller J, Parsons S (eds) Proceedings of the 7th
international conference on autonomous agents and multiagent systems (AAMAS 2008),
12–16 May2008, Estoril. International Foundation for Autonomous Agents and Multiagent
Systems, Richland, pp 1153–1160
de Jong G, Ben-Akiva M (2007) A micro-simulation model of shipment size and transport chain
choice. Transport Res B 41(9):950–965
Dia H (2002) An agent-based approach to modelling driver route choice behaviour under the
influence of real-time information. Transport Res C 10(5–6):331–349
Dia H, Panwai S (2007) Modelling drivers’ compliance and route choice behavior in response to
travel information. Nonlinear Dynam 49:493–509
Ehlert PAM, Rothkrantz LJM (2001) Microscopic traffic simulation with reactive driving
agents. In: Proceedings of intelligent transportation systems 2001. IEEE, Oakland (CA)
USA, pp 860–865
El hadouaj S, Drogoul A, Espié S (2000) How to combine reactivity and anticipation: the case of
conflict resolution in a simulated road traffic. In: Moss S, Davidsson P (eds) Multi-agent-based
simulation: second international workshop, MABS 2000, Boston; Revised and additional
papers (Lecture notes in computer science, 1979). Springer, Berlin, pp 82–96
Esser J, Schreckenberg M (1997) Microscopic simulation of urban traffic based on cellular
automata. Int J Mod Phys C 8(5):1025–1036
Fischer K, Chaib-draa B, Müller JP, Pischel M, Gerber C (1999) A simulation approach based on
negotiation and cooperation between agents: a case study. IEEE Trans Syst Man Cybern C
Appl Rev 29(4):531–545
Gambardella LM, Rizzoli A, Funk P (2002) Agent-based planning and simulation of combined
rail/road transport. Simulation 78(5):293–303
Goldberg JB (2004) Operations research models for the deployment of emergency services
vehicles. EMS Manage J 1(1):20–39
Gruer P, Hilaire V, Koukam A (2001) Multi-agent approach to modeling and simulation of urban
transportation systems. In: Proceedings of the 2001 I.E. international conference on systems,
man, and cybernetics (SMC 2001), vol 4. Tucson, 7–10 Oct 2001, pp 2499–2504
Hemelriajk C (2013) Animal social behaviour. Chapter 22 in this volume
Hensher DA, Ton TT (2000) A comparison of the predictive potential of artificial neural networks
and nested logit models for commuter mode choice. Transport Res E 36(3):155–172
Holmgren J, Davidsson P, Ramstedt L, Persson JA (2012) TAPAS: a multi-agent-based model for
simulation of transport chains. Simulation Model Pract Theory 23:1–18, Elsevier
Hunt JD, Gregor BJ (2008) Oregon generation 1 land use transport economic model treatment of
commercial movements: case example. In: Hancock KL (Rap.) Freight demand modeling:
tools for public-sector decision making; Summary of a conference, Washington, DC, 25–27
Sep 2006 (TRB conference proceedings, 40). Transportation research board, Washington, DC,
pp 56–60
Jin X, Abdulrab H, Itmi M (2008) A multi-agent based model for urban demand-responsive
passenger transport services. In: Proceedings of the international joint conference on neural
networks (IJCNN 2008), Part of the IEEE world congress on computational intelligence,
WCCI 2008, 1–6 June 2008, IEEE, Hong Kong, China, pp 3668–3675
Koorey G (2002) Assessment of rural road simulation modelling tools (Research report, 245).
Transfund New Zealand, Wellington. http://ir.canterbury.ac.nz/bitstream/10092/1561/1/
12591251_LTNZ-245-RuralRdSimulatnTools.pdf
Kumar S, Mitra S (2006) Self-organizing traffic at a malfunctioning intersection. J Artif Soc Soc
Simul 9(4):3. http://jasss.soc.surrey.ac.uk/9/4/3.html
Lee S, Kim Y, Namgung M, Kim J (2005) Development of route choice behavior model using
linkage of neural network and genetic algorithm with trip information. KSCE J Civil Eng vol 9,
(4):321–327
Li T, Hofker F, Jansma F (2006) Passenger travel behavior model in railway network simulation.
In: Perrone LF, Wieland FP, Liu J, Lawson BG, Nicol DM, Fujimoto RM (eds) Proceedings of
the 2006 winter simulation conference, Baltimore
Mahmassani HS (ed) (2005) Transportation and traffic theory: flow, dynamics and human interac-
tion. In: Proceedings of the 16th international symposium on transportation and traffic theory.
Elsevier, Oxford, Great Britain
Mandiau R, Champion A, Auberlet J-M, Espié S, Kolski C (2008) Behaviour based on decision
matrices for a coordination between agents in a urban traffic simulation. Appl Int 28
(2):121–138
Meignan D, Simonin O, Koukam A (2007) Simulation and evaluation of urban bus-networks using
a multiagent approach. Simulat Model Pract Theory 15:659–671
Nagel K, Schreckenberg M (1992) A cellular automaton model for freeway traffic. J de Physique I
2(12):2221–2229
Ortúzar JD, Willumsen LG (2001) Modelling transport, 3rd edn. Wiley, Chichester
Panwai S, Dia H (2005) A reactive agent-based neural network car following model. In:
Proceedings of the 8th international IEEE conference on intelligent transportation systems,
Vienna, Austria, 13–16 Sept 2005. IEEE, Vienna, pp 375–380
Pelechano N, O’Brien K, Silverman B, Badler N (2005) Crowd simulation incorporating agent
psychological models, roles and communication (Technical report). Center for Human
Modeling and Simulation, University of Pennsylvania
Pursula M (1999) Simulation of traffic systems – an overview. J Geogr Inf Decis Anal 3(1):1–8
Ramstedt L (2008) Transport policy analysis using multi-agent-based simulation. Doctoral disser-
tation No 2008:09, School of Engineering, Blekinge Institute of Technology
Rindsfüser G, Klügl F (2007) Large-scale agent-based pedestrian simulation. In: Petta P et al (eds)
MATES 2007, LNAI 4687, Leipzig, pp 145–156
Rossetti R, Bampi S, Liu R, Van Vleit D, Cybis H (2000) An agent-based framework for the
assessment of driver decision-making. In: Proceedings of the 2000 I.E. intelligent transporta-
tion systems, IEEE, Dearborn, (MI) USA, IEEE, 387–392
Santos G, Aguirre BE (2004) A critical review of emergency evacuation simulation models. In:
Peacock RD, Kuligowski ED (eds) Proceedings of the workshop on building occupant move-
ment during fire emergencies. National Institute of Standards and Technology, Gaithersburg,
10–11 June 2004, pp 27–52
Schreckenberg M, Selten S (2004) Human behaviour and traffic networks. Springer, Berlin
Strader TJ, Lin FR , Shaw M (1998) Simulation of order fulfillment in divergent assembly supply
chains. J Artif Soc Soc Simul 1(2). http://jasss.soc.surrey.ac.uk/1/2/5.html
Swaminathan J, Smith S, Sadeh N (1998) Modeling supply chain dynamics: a multiagent
approach. Decis Sci J 29(3):607–632
Tapani A (2008) Traffic simulation modeling of rural roads and driver assistance systems.
Doctoral thesis, Department of science and technology, Linköping University, Linköping
Terzi S, Cavalieri S (2004) Simulation in the supply chain context: a survey. Comput Ind
53(1):3–17
Toledo T (2007) Driving behaviour: models and challenges. Transport Rev 27(1):65–84
van der Zee DJ, van der Vorst JGAJ (2005) A modelling framework for supply chain simulation:
opportunities for improved decision making. Decis Sci 36(1):65–95
Wahle J, Schreckenberg M (2001) A multi-agent system for on-line simulations based on real-
world traffic data. In: Proceedings of the 34th annual Hawaii international conference on
system sciences (HICSS’01). Hawaii, USA, IEEE
Williams I, Raha N (2002) Review of freight modelling. Final report, DfT Integrated Transport
and Economics Appraisal, Cambridge
Zhang Q, Han B, Li D (2008) Modeling and simulation of passenger alighting and boarding
movement in Beijing metro stations. Transport Res C 16:635–649
Chapter 25
Modeling Power and Authority:
An Emergentist View from Afghanistan
Armando Geller and Scott Moss
Why Read This Chapter? To understand how an evidence-driven approach using

agent-based social simulation can incorporate qualitative data, and the effects of
social complexity, to capture some of the workings of power and authority, even in
the absence of sufficient statistical data. This is illustrated with a model of Afghan
power structures, which shows how a data collection process, intuitive behavioral
models and epistemological considerations can be usefully combined. It shows
how, even with a situation as complex as that of Afghanistan, the object under
investigation can shape the theoretical and methodological approach rather that
the other way around.
Abstract The aim of this chapter is to provide a critical overview of state of the art
models that deal with power and authority and to present an alternative research
design. The chapter is motivated by the fact that research on power and authority is
confined by a general lack of statistical data. However, the literal complexity of
structures and mechanisms of power and authority requires a formalized and
dynamic approach of analysis if more than a narrative understanding of the object
of investigation is sought. It is demonstrated that evidence-driven and agent-based
social simulation (EDABSS) can contend with the inclusion of qualitative data and
the effects of social complexity at the same time. A model on Afghan power
structures exemplifying this approach is introduced and discussed in detail from
the data collection process and the creation of a higher order intuitive model to the
derivation of the agent rules and the model’s computational implementation.
EDABSS not only deals in a very direct way with social reality but also produces
A. Geller (*)
Scensei, LLC, 107 S. West St., Alexandria, VA 22314, USA
The School for Conflict Analysis and Resolution, George Mason University, 3330 N. Washington
Blvd., Arlington, VA 22201, USA
S. Moss
Scott Moss Associates, Brookfold, The Wash, Chapel-en-le-Frith, SK23 0QW, UK

668 A. Geller and S. Moss
complex artificial representations of this reality. Explicit socio-cultural and episte-

mological couching of an EDABSS model is therefore essential and treated as well.
25.1 Introduction
Notions such as “power” and “authority” are redolent with meaning yet hard to
define. As a result, the number of definitions of power and authority is overwhelm-
ing (Neumann 1950). Weber (1980:53) defines power as the “probability that one
actor within a social relationship will be in a position to carry out his own will in
spite of resistance.” Giddens (1976:111) understands power in a relational sense as
a “property of interaction, and may be defined as the capability to secure outcomes
where the realization of these outcomes depends upon the agency of others.” He
implies a major distinction between two types of resources in connection with
power, (1) control over material resources and (2) authoritative resources. Parsons
(1952:121–132) underlines the pragmatic character of the notion of power even
more by stating that power is the capacity to achieve social and societal objectives,
and as such can be seen as analogous to money. Power, consequently, is the basis of
a generalized capacity to attain goals and to invoke consequences upon another
actor (Moss 1981:163).
Neither these nor any other definitions predominate and the decision to apply a
particular one is subjective and, if not based on normative grounds, most likely
context-dependent. Hence, in this research it is not aimed at applying one particular
theoretical approach to power and authority, but instead it is argued that such an
approach can also be founded on available and contextual evidence. Evidence is
understood as information that is derived from case-studies, empirically tested
theories, the high quality media and engagement with stakeholders, domain experts
and policy analysts and makers.
For heuristic reasons – and in awareness of the plethora of conceptual
approaches to power – it is for the time being assumed that the social phenomena
of power and authority occur in a two- (or more-) sided relationship. It is also
assumed that power serves a purpose. What should be of interest to students of
power has been identified by Lasswell (1936): “Who gets what, when, how”. Who
and what describe and explain structures; when and how describe and explain
mechanisms1 and processes. However, Lasswell ignores an important aspect:
“why”. Why does someone get something at a particular moment in time in a
particular way? And more generally: Why did a particular condition of power form?
1
Schelling (1998) understands a “social mechanism [. . .][as] a plausible hypotheses, or set of
plausible hypotheses, that could be the explanation of some social phenomenon, the explanation
being in terms of interactions between individuals and other individuals, or between individuals
and some social aggregate.” Alternatively, a social mechanism is an interpretation, in terms of
individual behavior, of a model that abstractly reproduces the phenomenon that needs explaining.
25 Modeling Power and Authority: An Emergentist View from Afghanistan 669
Castelfranchi (1990) already noted in the year 1990 that social power is a lacuna
in social simulation and (distributed) artificial intelligence. This chapter shows that
although power relations are ubiquitous in social systems, only a small number of
relevant models have been developed. This is despite the fact that social simulation
and in particular evidence-driven and agent-based social simulation (EDABBS) are
valuable complementary techniques to orthodox approaches to the study of power
and authority.
A prime virtue of EDABSS is that it imposes on the modeler a requirement to
devise an unambiguous formal meaning for such notions as power and authority
and their corresponding concepts. The modeling process and thus formalization
should begin by formulating questions that structure the rich body of narratives that
are available to describe the explanandum:
• Under what conditions would you label someone as powerful or as being in a
position of authority?
• Having labeled someone as being powerful or in a position of authority, how
would you expect that person to behave?
• Having labeled someone as powerful or in a position of authority, how would
you expect yourself/others to behave towards that person?
These questions are not abstract. They form part of a data collection strategy and
aim at enriching general accounts of power and authority, such as that power is a
form of relationship that is exercised between two or more individuals, by more
specific and context-dependent forms of descriptions of power and authority. Often
these descriptions concern actor behavior. Models based on such descriptions can
be closer to the evidence, because this evidence can be straightforwardly translated
into “if-then” rules that can be implemented in logic-like fashion in rule-based or
standard procedural languages using their native if-then structures. The if part is
determined by the first question and the then part by instances of the second or third
questions. Our own preference is for an evidence-driven and declarative represen-
tation of power and authority in order to preserve as much as possible of the rich
data drawn from case-studies and evidence in general while maintaining conceptual
clarity.
The computational modeling procedure described in this article is inspired by
the idea to represent reality by means of modeling; it is driven by shortcomings and
advantages of other methodological approaches; and it has matured out of research
on cognitive decision making as declarative and thus mnemonic implementations
(Moss 1981, 1998, 2000; Moss and Edmonds 1997; Moss and Kuznetsova 1996;
Moss et al. 1996) and on contemporary conflicts (Geller 2006b). Classical herme-
neutic approaches, although they may be strong in argument, are often methodo-
logically weak. However, they have an important “serendipity” function and
creative role, very much like that of intuitive models (Outhwaite 1987). Traditional
empirical approaches, such as statistical and econometric modeling, do not repre-
sent reality and have difficulties to produce insight into mechanisms, processes and
structures (cf. Shapiro 2005; Hedström and Swedberg 1998).2 Moreover, regular
incorporation of poor proxies and the use of inadequate data does not contribute to
the plausibility of research results. However, rigorous formalization furnishes
desirable clarity and comparability. Finally, qualitative and case-study-based anal-
ysis produces deep structural and processual insight as well as “thick description”,
although at a high – for some too high – price of idiography and thus lack of
generalizability.
None of these approaches is incompatible with EDABSS. But we argue here that
a natural modeling procedure starts with an informed but intuitive and theoretical
model that needs to be validated against reality. The intuitive model is to be
validated at micro-level against largely qualitative evidence, so that a rich model
develops. This qualitatively validated and enriched model is then formalized as a
computational model. The social simulation’s output should enhance our under-
standing of reality, which, in turn, necessitates adjustments in the model design.
This process is the hermeneutic cycle of EDABSS (cf. Geller 2006b; Geller and
Moss 2008b).
A selection of models dealing with power and authority is reviewed in
Sect. 25.2. A critical appraisal of these models reveals the strengths and weaknesses
of past modeling approaches and underlines the necessity of the research design
presented here. The selection criteria are evidently influenced by our own modeling
approach and are therefore subjective. However, models have also been chosen on a
functional basis: How successful are they in describing and explaining their target
system? We have also tried to choose exemplary models from a wide range of
modeling philosophies, to elucidate conceptual differences amongst the discussed
models. Section 25.3 comprises the dialectical results of Sect. 25.2 and discusses
the analytical concepts and methodological tools applied here to analyze power and
authority. The materialization of our approach is presented in Sect. 25.4, where the
model implementation and simulated output is discussed. The model’s context is
conflict-torn Afghanistan. Section 25.5 concludes by embedding our modeling
framework into a broader context of comparable social phenomena, such as conflict
and organized crime, and promotes agent- and evidence-based social simulation as
an efficient approach for the study of power and authority.
25.2 What Can We Learn from (a Selection) of Models

on Power and Authority?
The development of a social simulation model can be informed by intuitive ideas,

theory or observation. For many simulations, the respective borderlines cannot be
drawn unambiguously. Nevertheless, such a classification is superior to a more
traditional one which only distinguishes between micro- and macro-models.
2
See for a promising corrective Sambanis (2004).
Although agent-based models entail explicit micro level foundations, for example,
the micro foundations of econometric models are at best implicit. More importantly
in relation to complexity, agent-based simulations often generate emergent phe-
nomena at macro level. From a modeling point of view it is therefore more
interesting to understand the level and nature of data that has guided the researcher
to conceptualize a model in a particular way, to what model output this conceptual-
ization has led and how a design helped to better understand mechanisms, processes
and structures in a target system.
25.2.1 Modeling Ideas
A variety of models on power and authority are implemented not strictly based
upon theory but rather on a mixture of intuition and existing theoretical research in
a particular field. These models promise to lend insight into a usually only little
defined social phenomenon in an explorative, and likely to be abstract, way and
therefore operate as an entrée into the object of investigation’s broader field. Robert
Axelrod’s emerging political actor model has been chosen because it epitomizes the
prototype of an explorative model; Rouchier et al.’s model still exemplifies the
want to explore, however on a more evidence-oriented basis.
25.2.1.1 Emerging Political Actors
In a well known agent-based model Axelrod (1995) reasons about the emergence of
new political actors from an aggregation of smaller political actors. His motivation
is to explain the restructuring of the global political landscape after the end of the
cold war and the fact that, although political scientists have a number of concepts
and theories to analyze the emergence of new political actors, they lack models that
account for this emergence endogenously.
In short, Axelrod’s model of emerging actors is a well structured and intelligible
but empirically ungrounded, conceptual model. The core of his model is a simple
dynamic of “pay or else” resulting in a “tribute system in which an actor can extract
resources from others through tribute payments, and use these resources to extract
still more resources.” (Axelrod 1995:21) The model consists of ten agents
distributed on a circle. Wealth is the only resource and is distributed to each
agent randomly at the beginning of the simulation. In each simulation iteration
three agents are randomly chosen to become active. When active, agents can ask
other agents for tribute. When asked, an agent has the choice between paying the
demanded tribute or to fight, depending on his and the demander’s resources. In
case of paying the tribute, a specified amount of wealth is transferred to the
demander; in case of fighting, each agent loses wealth equal to 25 % of his
opponents’ wealth. After three iterations all agents exogenously receive again
wealth.
The core of any agent-based model is the rules according to which the agents
behave. In Axelrod’s tribute model the agents have to decide when to demand
tribute and how to respond to such a demand. First, an active agent needs to decide
whom it should address. “A suitable decision rule [. . .] is to choose among the
potential targets the one that maximizes the product of the target’s vulnerability
multiplied by its possible payment” (Axelrod 1995:24). If no agent is vulnerable
enough, then no demand is made. The chosen agent responds by fighting only if
t would cost less than paying the tribute.
Agents can form alliances. During the course of the simulation, agents develop
commitments towards each other. Commitment between two agents increases if
one agent pays tribute to the other agent and vice versa and if two agents fight
another agent together. Commitment decreases if one agent fights at the opposite
side of another agent. Alliance building indirectly increases an agent’s resources
and outreach, as it can only make a demand to an agent if it is either adjacent or
indirectly connected via allied agents.
Axelrod gets six characteristic results: (1) The model does not converge to an
equilibrium. (2) The model’s history is fairly variable in terms of combinations of
wealthy actors, fighting frequency and overall wealth accumulation. (3) Agents do
not only share the resources with their allies, they also share the costs and thus the
risks. Major conflict can occur as agents can be dragged into fighting. (4) Because
of asymmetric commitment, fighting can erupt among allies. (5) An alliance can
host more than one powerful agent. (6) The initial wealth distribution is not an
indicator for an agent’s success. As an overall result, Axelrod reports that he was
able to breed new political actors of a higher organizational level. These new
political actors are represented in the model as so called clusters of commitment.
Axelrod’s model demonstrates the realist notion that states do not have friends,
but only interests. (Agents initially make commitments out of rational calculations –
“pay or else” –, not out of ideological considerations.) These interests, the model
reveals, are attended most efficiently by joining coalitions. Thus, an effective
powerful agent seeks cooperation of some form or the other.
Axelrod’s model is convincing as long as he concentrates on his main task –
letting new political agents emerge in the context of an explorative setup. But
interpreting a positive feedback effect resulting from a total of ten agents as
“imperial overstretch”, or interpreting fighting between two agents of the same
cluster of commitment as civil war rings a little hollow. As much as we can learn
from Axelrod about how to setup and present a simple, but innovative and explor-
ative model, as little can we learn about how to discuss it, as Axelrod continually
blurs the distinction between the model and reality which leads to over- and
misinterpretation of his results. Hence, a number of open questions remain, the
most important of which is whether the model would withstand even circumstan-
tially founded cross-validation. This is naturally related to the question of how
much explanatory power Axelrod’s model holds. With agents so abstract from
reality, agent behavior in Axelrod’s model simply cannot provide any explanatory
insight into real world actor behavior. And while it makes sense to claim that
complex macro level behavior can emerge from simple micro level behavior – as is
the case in Axelrod’s model – this macro outcome, again, is so abstract from reality
that it does not allow for any insight into the real international system. Conse-
quently, the emergent interplay between the micro- and the macro-level, so typical
for complex systems, cannot have the effects it would have in reality. This accounts
for another loss of explanatory power. Axelrod claims that his model’s objective is to
think in new ways about old problems. However, its lack of evidence-based
foundations and its degree of abstraction from reality might well foster stereotyped
perceptions of the international system instead of triggering critical reflection about it.
An alternative is for agents and their environment to be derived either from
an empirically well-tested theory or from qualitative real-world observations.
Cederman (1997), discussed below, has advanced Axelrod’s work into this
direction.
25.2.1.2 An Artificial Gift-Giving Society
Rouchier et al. (2001) reported a model of an artificial gift-giving society that is

loosely founded on ethnographic research. Gifts structure society and reproduce
habits and values. The giving of gifts can also be a means of redistribution. The
donation, reception and reciprocation of a gift create relationships, which can
become competitive. Thus, gifts can be a means to create authority, establish
hierarchies and uphold power structures.
The model’s goal is to create an artificial society in which reputation emerges.
The gift-giving society’s agents are either occupied with working to accumulate
resources, which enables them to give away gifts or by giving away gifts them-
selves. The motivation to give away gifts is twofold: On one hand gifts are given
away because agents act according to their self-esteem, i.e. the desire to be able to
make gifts that are acceptable to the group. On the other hand, agents give away
gifts because they want to increase their reputation by swaggering. Agents are fully
informed and share the same decision process to determine what actions to take.
The artificial society’s population consists of 50 agents.
Each agent has to decide at each time step what gift it wants to make to whom.
This decision process stands at the core of the model. Agents can give away gifts,
either for the sake of reputation or sharing. So called “sharing gifts” are less costly
than prestige gifts. The decision of whom to give what gift is socially embedded and
the agent’s rationality depends on its social position. Making a sharing gift to any
agent adds to the social inclusion of an agent; making a prestige gift to an agent who
is considered as being prestigious fosters a hierarchy among the agents. Therefore,
receiving a sharing gift represents social acceptance; receiving a prestige gift
represents the acceptance of social stratification. The better an agent is socially
integrated, i.e. the higher its self-esteem, the higher is its motivation to give away
gifts. At the same time, high social integration increases the likelihood that an agent
is able to give away prestige gifts.
All agents exchange their gifts after each agent has decided to whom to give its
gift. Subsequently the agents evaluate their ranks within the group and their
reputation on the basis of the gifts they have received. Self-esteem and reputation
are adapted according to the outcome of the gift-giving round. Then the next round
starts.
The authors discuss their findings with regard to donation-reception dynamics
evolving during the simulation, of which most are positive feedback loops
(Rouchier et al. 2001:5.1). An agent’s esteem increases with the number of gifts
it has made and received respectively. The higher this agent’s esteem is, the higher
the likelihood is that it receives even more gifts and that he can give gifts away. The
same holds true with an agent’s reputation, which increases with the reception of
prestige gifts. The higher this agent’s reputation is, the higher is the likelihood that
it receives even more prestige gifts. Moreover, the authors find that esteem and
prestige “go together”. “Both help the circulation of gifts, and then the creation of
prestige reputation.” (Rouchier et al. 2001:5.3) The gift-giving model provides
insight into the emergence of an elite group based on socially contextualized
decision processes. Within the gift-giving model the emergence of social power
is explained by two factors: (1) the willingness and ability to become socially
accepted and (2) the ambition to accumulate reputation. Gift-giving has been
introduced and computationally implemented by Rouchier et al. (2001) as a process
of resource accumulation and redistribution, an important variant of strategic
behavior in many other contexts.
The authors, while creating a naturalistic model, do not attempt to derive their
agent rules directly from qualitative, in this case likely anthropological research.
This would have allowed them to gain narrative insight into the dynamics of gift
giving and receiving and would have enabled them to directly cross-validate their
findings with accounts made by anthropologists. Micro-level explanation was not
one of the modelers’ main interests. Since the emergence of complex macro level
outcomes results from micro-level behavior and social interaction, the model is of
limited usefulness in the analysis of real world social complexity. Formulation of
agent rules from anthropological evidence would also have enabled the authors to
avoid the assumption that agents are totally informed, both, spatially and with
regard to the internal state of other agents. The cognitive processing of information
about other agents is an especially difficult task. For example, the model does not
address the question of when, as well as to whom, agents give gifts. It was not a
purpose of the Rouchier et al. (2001) model to capture strategic decision-making
using incomplete information within a realistic context. To have done so would
usefully have entailed reliance on more evidence-based information available in the
anthropological literature and might then have provided an empirically better
grounded account of social status and power.
25.2.2 Testing Theory
25.2.2.1 Power Games Are Not Games
Game theoretical applications comprise perhaps the most formalized, coherent and
theoretically unified approach to power, especially in international relations.3
Regularly applied in military-political analysis and international political economy,
game theory has a straightforward conception of the nation state as interdependent,
goal seeking and rational actor embedded in a context free of centralized, authori-
tative institutions (Snidal 1985). The application of game theoretical models in
international relations raises questions such as “Who are the relevant actors?”,
“What are the rules of the game?”, “What are the choices available to each
actor?”, “What are the payoffs in the game?” and “Is the issue best characterized
as single-play or repeated-play?” (Snidal 1985:26). Abstract and simplified as they
are, game-theoretical models nevertheless intend replicating a particular social
situation and aim at – if the actors’ preferences, strategies and payoffs are accu-
rately modeled – generating testable predictions and understanding of fundamental
processes governing the system of international relations.
If states are conceived as rational power maximizers in an anarchic system, then
we talk about the realist paradigm of international politics. “Rationality in this
Realist world centers on the struggle for power in an anarchic environment” (Snidal
1985:39). However, compared with reality it would be misleading to conceive of
states as self-defending entities of purely opposing interests. Rather, game theory
teaches us that states exhibit a strategic rationality that incorporates the awareness
that no state can choose its best strategy or attain its best outcome independently of
choices made by others. This awareness is the birth of cooperation. Moreover, from
so called iterated and dynamic games we learn that while states have incentives to
defect in the short-run, in the longer-run they can achieve benefits from cooperation
through time (Snidal 1985).
Game theory has often been criticized on the grounds that it is unrealistic. Game
theory can, of course, analyze very particular examples of world politics, such as
the Cuban Missile Crisis, nuclear war in a bipolar world or the General Agreement
on Trade and Tariffs (GATT). What can be expected, in more general terms, are
explanations for meta-phenomena or for abstract representations of particular
circumstances. We criticize game theoretical approaches in the social sciences
from an evidence-based point of view. The difference between a game theoretic
model and the target system is enormous. Everything that is a natural representation
3
Models not discussed in this subsection but of further interest to the reader are Alam et al. (2005),
Caldas and Coelho (1999), Guyot et al. (2006), Lustick (2000), Mosler (2006), Rouchier and
Thoyer (2006), Saam and Harrer (1999) and Younger (2005). Particularly highlighted should be
the work of (Mailliard and Sibertin-Blanc 2010) who merge a multiagent and social network
approach to the complexity and transactional nature of power with approaches to power from the
French school of sociology, and develop against this background a formal logic system.
of the international system is a special and difficult case in game theory, for
example, non-symmetric n-person games involving intermediate numbers of states,
and everything that is straightforwardly implementable in game theory is unrealis-
tic. Yet analogies are quickly drawn between model results and the model’s target
system. Hence, findings rely on oversimplified model ontologies, which may lead to
over-interpretation.
Game theory’s simplistic ontologies also affect a model’s simulation output, as
Moss (2001) argues. An analysis of state of the art game theory as represented by 14
game theoretic papers published in the Journal of Economic Theory in the year
1999 indicates that the game theoretic models’ assumptions (perfect information)
and implementations (e.g., Markov transition matrices and processes) preclude the
emergence of self-organized criticality as reported by Bak (1997) and cannot
capture the necessary interaction as a dynamic process. Game theoretic models
on markets do not entail statistical signatures found in empirical data on markets,
such as power law distributed data, a characteristic of self-organized criticality.
This critique applies to game theoretic models in international relations as well, as
one of the observed regularities in the international system is the power law
distribution of wars (Richardson 1948). Whereas Cederman (2003) replicated
Richardson’s (1948) findings by means of agent-based modeling, to our knowledge
there exists no game theoretical reproduction of this statistical signature.
Although game theory can make statements of the “when” and “why”, it cannot
say anything about the “how” and cannot give insight into the mechanisms and
processes underlying the emergence of power structures [54]. Therefore, game
theoretical modeling is, like most statistical or econometric modeling, a type of
black box modeling. A rare exception to this is Hirshleifer’s work (Hirshleifer
1991, 1995).
25.2.2.2 Exploring the Limits of Equation-Based Modeling
Hirshleifer (1991) published a seminal paper on the paradox of power. The paradox
of power states that in case of conflict, a weaker contestant is able to ameliorate his
position relative to the stronger actor because his inferiority makes him fight harder.
In other words: “[N]on-conflictual or cooperative strategies tend to be relatively
more rewarding for the better-endowed side.” (Hirshleifer 1991:178) Hirshleifer’s
modeling solution for this problem is based on the assumption that if there exists an
equilibrium market economy, then there must also exist an equilibrium outcome if
contestants in two-party interactions compete by struggle and organized violence.
The model’s assumptions are full information, a steady state, indifference towards
geographical factors and the non-destructive nature of fighting (Hirshleifer
1991:198).
Hirshleifer’s econometric model leads to a well-specified outcome from which a
number of unequivocal results can be derived. When the paradox of power applies
the “rich end up transferring income to the poor” and this “tends to bring about a
more equal distribution of [. . .] income.” (Hirshleifer 1991:197) However, he also
states that “the comparative advantage of the poor in conflictual processes can be
overcome when the decisiveness of conflict is sufficiently great that is, when a
given ratio of fighting effort is very disproportionately effective in determining the
outcome of conflict.” (Hirshleifer 1991:197) Hirshleifer validates his analytical
results by providing circumstantial evidence for a number of examples from
class-struggle within nation-states to firms (labor-management conflicts) and
protracted low-level combat.
If one accepts economic theory’s underlying assumptions, Hirshleifer’s results
are indeed compelling and apply to a wide range of social issues related to power.
The paradox of power identifies, perhaps correctly, the expectations weak and
strong actors can have when mutually entering power politics: When the stakes
are high, the weak are likely to get crushed by the strong. This is why the paradox of
power is likely to apply to more limited contests than to full fledged conflict.
What can we learn from Hirshleifer with regard to modeling power and author-
ity? He elegantly exemplifies state of the art model conceptualization, presentation
and discussion of results, including the model’s limitations. His presentation of the
analytical tools and his disclosure of the model’s underlying assumptions are
exemplary and render his work amenable to critique. The same applies to the
rigid and straightforward formalizations of the mechanisms underlying the paradox
of power, which are, moreover, well-annotated and embedded in theory. Last, but
not least, the model’s scope and its delimitations are clearly marked by referring to
a number of examples ranging from interstate war to the sociology of the family.
Hirshleifer informs us precisely of the outcomes from power struggles and of the
factors that cause these outcomes; he fails to produce analytical insight into the
structural arrangements and processes that lead to these outcomes as a consequence
of the methodology itself as is demonstrated in another paper of his.
In “Anarchy and Its Breakdown” Hirshleifer (1995) models anarchy, a social
state lacking auctoritas, as a fragile spontaneous order that may either dissolve into
formless amorphy or a more organized system such as hierarchy. Interesting for our
task is the fact that Hirshleifer produces results that indeed allow insight in terms of
structure. He can for example demonstrate that the state of anarchy is stable if no
actor is militarily dominant and if income is high enough to assure one’s own and
the group’s survival. However, concrete processual insight can again not be deliv-
ered and it is referred to circumstantial evidence to concretize particular aspects of
the model. In fact circumstantial evidence is arguably the only validation that is
feasible with these kinds of very abstract models as the statistical signature of the
model’s output is so ideal-typical that it is not comparable with empirical data. In
short, orthodox theoretically informed models, such as econometric or statistical
models, often do address those issues in which social scientists are really interested,
but cannot provide an explanation in involving complexity arising from social
interaction.
One reason for this has already been identified above with relation to game
theory, i.e. the preclusion of emergence of self-organized criticality. Another reason
is highly unrealistic general assumptions, i.e. rational choice, and more specific
unrealistic assumptions, for example the non-destructive nature of fighting (as such
also identified by Hirshleifer (1991:196–199) himself) or monolithic actors lacking

an internal state. Finally, methodological individualism ignores the micro–macro
link as well as the heterogeneous nature of political actors. While the particular
assumptions as well as the (homogeneous) agents could be chosen more realisti-
cally, methodological individualism is together with most statistical and economet-
ric modeling relying on homoskedasticity of data points and variances. Such
models cannot be used to describe or explain the evolution of power structures.
(For statistical models this holds true only, of course, if there is sufficient statistical
data that describes power relations.) With this regard, Cederman’s (1997) model is
a paradigmatic shift.
25.2.2.3 When, Why and How Nations Emerge and Dissolve
Cederman’s (1997) model has been chosen, because it applies the analytical rigor of
formalized approaches to agent-based modeling without relinquishing the latter’s
virtues. He raises fundamental questions regarding realist and methodologically
orthodox approaches to international relations. He introduces an agent-based simu-
lation of the emergence and dissolution of nations. His model is based on a 20 20
grid that is initially inhabited by “predator” and “status quo” states. Each period of
time a state is randomly assigned to receive resources for attack and defense. Given
they have an advantage in overall and local resources respectively, predator states
can attack and conquer neighboring states. Although Cederman applies neorealist
principles to a dynamic and historical meta-context, his findings challenge the
orthodox neorealist belief that applying alternative methods does make an episte-
mological difference. He reports that defense alliances and the dominance of
defense in the international system are paradoxically destabilizing. While defensive
systems of cooperation deter predator states from growing to predator empires, at
the same time they make possible the build-up of a hegemonic predator actor,
because once a predator has reached a critical mass it has, due to the defensive
nature of the system, no rivals.
Cederman (1997:136) asserts that “[s]tates have been mostly modeled as inter-
nally consolidated monoliths, albeit with emergent outer boundaries.” He confronts
this simplifying assumption by supplying his agents with an internal decision
making mechanism representing a nationalistic two-level-politics mechanism.
State leaders cannot be concerned only with foreign affairs anymore, but must
also take into consideration domestic issues (cf. Putnam 1988). Lazer (2001) has
stated that the insight gained from Cederman’s nationalist implementation is not as
striking as the one dating from his emergent polarity implementation. Perhaps this
is true in terms of contents, but in an epistemological perspective Cederman makes
an important point: His nationalist model inspirits previously dead states. It is not
enough to know that states do something, but from a social scientific point of view it
is essential to know why they do it and how. This affords realistic, i.e. evidence-
based assumptions and implementations.
Although Cederman (2003) later on introduces technological change and power

projection and Weidmann and Cederman (2005) introduce strategy and self-
evaluation into the decision making process, Cederman’s models conform to
traditional empirical perceptions of international relations. Consequentially, a
number of issues relevant to the study of power and authority in contemporary
conflicts remain untouched. The state, territorial conquest and consequentially the
redrawing of borderlines have been important explanatory factors for conflicts since
the emergence of territorial states, but they misconceive the nature of a great
number of conflicts throughout history and consequentially can only partially
explain the emergence of power structures in contexts where the nation state or
any other type of centralized political power has only played a marginal role. And
even where such well defined territory existed, Cederman cannot explain the causes
for conflict if they have been for example of ethnic, religious or economic character.
Other issues that should be taken into consideration are neo-patrimonialism,
anomic states, genocide, transnational organized crime and external intervention.
Models analyzing such a reality must be multivariate and causal, allowing for an
explorative framework, and – contrary to Cederman (1997) – be able to include
atheoretical and evidence-based information, which is often of a qualitative type.
25.2.3 Toward Implementing Reality
Modeling reality is not just about modeling particular cases. Modeling reality is
about the development of models that have explanatory power on both, the micro-
and macro-level and therefore give also insight into mechanisms, processes and
structures. A model can hardly claim to exhibit explanatory power when lacking
pivotal aspects of a perceived reality and when abstracting too much from this
reality. Every reasonable model of reality – i.e. a model that describes not only the
who, what and when, but also the why and how – must entail construct validity.4
25.2.3.1 Explaining the Causes for Social Unrest and Organized Violence
Kuznar and Frederick (2007) propose a model in which they explore the impact of
nepotism on social status and thus power. They rely on an innovative model
architecture supported by relevant research results, which are not framed by
dogmatic theory. An agent-based model is employed “to model the origins of the
sort of wealth and status distributions that seem to engender political violence.
In particular, we explore the minimum conditions required to evolve a wealth
4
A model not discussed in this section but that is of excellent quality, both in terms of content and
innovation is (Guyot et al. 2006). The authors analyze and discuss the evolution of power relations
on the basis of participatory simulations of negotiation for common pool resources.
distribution from the mild inequalities seen in ancestral human societies of hunter-
gatherers to the more extreme wealth inequalities typical of complex, stratified
societies.” (Kuznar and Frederick 2007:31)
Wealth is implemented as a variable that takes an ideal cultural form over which
actors would compete in a particular society. For hunter-gatherer societies wealth is
distributed in a sigmoid fashion, where agents in the convex parts of the distribution
have more to gain than to lose when taking chances and therefore are risk prone.
The model consists of three building blocks: the distribution of wealth, agent
interaction and nepotism. Wealth distributions in complex societies are, by contrast,
exponential, with sigmoid oscillations around the exponential curve. Kuznar and
Frederick (2007:32) term this an expo-sigmoid curve. Agent behavior is modeled
along the lines of a coordination game with two equilibria (either both players
defect, or both players cooperate) and a Nash optimum which is to play a mixed
strategy of join and defect. Kinship is inherent to nepotism. Thus, agents with many
kin and offspring perform better, i.e. have higher payoffs, in the coordination game
and exhibit higher fertility due to the effect of cultural success.
The result Kuznar and Frederick are getting is that the effects of nepotism
transform a sigmoid wealth distribution into an approximate expo-sigmoid distri-
bution. The explanation for this is the emergence of a distinct three class society:
the poor and their offspring get poorer, a small middle class gets richer without
changing status and the elites get richer and richer. Thus, the authors conclude the
positive feedback loop working between nepotism and cultural success increases
the structural inequality between a powerful elite and the rest of the population and
thus escalates social unrest and potential organized violence.
Whereas Axelrod (1995) over-interprets his model, Kuznar and Frederick
explore the full potential of their research design: A simple, intuitive and at the
same time thoroughly grounded model is presented well, specified together with
moderate but auspicious conclusions, which advance research and shed new light
on a problem. The model, however, would be even more compelling if the authors
would have had presented cross-validational results.
Model output should be, if possible, cross-validated against empirical data
originating from the target system (Moss and Edmonds 2005). There are three
strategies: (1) If the simulation leads to statistical output, this output is statistically
analyzed and the resulting significant signatures are compared with the statistical
signatures gained from data originating from the target system. Such signatures can
for example be a leptokurtic data distribution, clustered volatility, power laws or
distinct (e.g., small world) patterns in a social network. If the model yields output of
statistical nature but statistical data is not available for the target system, then
validation must rely on qualitative data. In this case, validation must either (2) seek
systematic structural and processual similarities between the model and the target
system, for example cross-network-analysis, or (3) find circumstantial evidence in
the target system that can also be found in the simulation. In case of empirical data
scarcity (3) is often the last resort.
25.2.3.2 Power, Resources, and Violence
Geller (2006a,b) developed an agent-based model of contemporary conflict

informed by evidence. The lack of a unified theory of contemporary conflict
motivated an intuitive and explanatory model of contemporary conflict. This
model is based on three types of interacting actors: a politician, businessman and
warrior. They engage in six interactions: (1) the politicization of the economy and
(2) the military, (3) the economization of politics and (4) the military, (5) the
militarization of politics and (6) the economy. To ascertain if this intuitive and simple
ontology can capture the main structural and processual characteristics of contempo-
rary conflicts, ten cases, such as Afghanistan, Chechnya and Sierra Leone, have been
analyzed in a primary validation procedure against the backdrop of the intuitive
model. The analytical results in the form of mini case-studies based on secondary
literature have been sent out to subject matter experts for a critical review. None of the
experts requested an essential revision of the analytical tool, the intuitive model.
As a next step, Geller enriched the theoretical background of the primarily
validated intuitive model by consulting more relevant literature for further specifi-
cation of the model’s structure and agency aspects in order to be able to model the
computational model’s agent rules. The basic idea is that politicians affiliate with
businessmen and warriors to make good business and get protection while
businessmen affiliate with politicians to get access to the lucrative political arena
and warriors seek political representation by politicians. Businessmen affiliate with
warriors for the same reason politicians do, to get protection, while warriors get
money for their provided services. Warriors can kill warriors affiliated with other
politicians or businessmen, whereas civilians are considered as being non-
constitutive to the intuitive model, as they are introduced into the computational
model in a reactive way, meaning that although they can affiliate themselves to
politicians, they can be forcibly recruited and ultimately killed by warriors.
The model offers insight into the dynamics of power in contemporary conflicts.
Contrary to the prevailing “greed and grievance” approach in current conflict
studies, Geller’s model demonstrates that a powerful agent is dependent on
businessmen and warriors at the same time. Powerful is who is most socially
embedded. He can also show that the main organizers of violence are the politicians
and that the warriors need not exhibit enough organizational capacity for a fully
fledged campaign of organized violence. Hence, the more fragmented the political
landscape is, the greater is the magnitude of organized violence. Geller’s results
gain importance as they are cross-validated against statistical data describing the
number of conflict related victims on a daily basis in Northern Ireland, Iraq and
Afghanistan. Both the simulation output and the real world data suggest that
conflict related victims are log-normally distributed over time (right-skewed),
exhibiting outbreaks of violence unpredictable in magnitude and timing.
Modeling always involves a degree of arbitrariness. A modeler’s task, then,
should be to reduce arbitrariness by making the model’s design as inter-subjectively
comprehensible as possible. Axelrod’s emerging actors model is a good, but
nevertheless simple, example of this. The more evidence oriented a model
becomes, the more difficult it becomes to justify the various omissions, inclusions
and abstractions. Procedural programming is cumbersome in responding to idio-

graphic challenges. As described in the next section, a declarative, rule-based
approach is better suited to the translation of evidence-based information of actor
behavior into agent rules.
25.2.4 Discussion
The synopsis presented above has revealed the many approaches through which the
social phenomena of power and authority can be scrutinized: in models based on
ideas interest oriented states and gift giving individuals have been implemented; in
highly formalized theoretical models agents are conceived as rationalist power
maximizers or as neo-realist states internally and externally struggling for survival;
detailed evidence collected from secondary literature is used for modeling pro-
cesses and structures that entail a high construct validity. Power structures are
complex as well as dynamic and emerge as a result of a multitude of structure
generating interactions among self- and group-oriented actors encompassing mani-
fold interests.
The discussed models allow for insight into dynamic model behavior, as well as
drawing structural conclusions with regard to their object of investigation whether it
is theoretical or empirical by nature. Those models that feature an individual agent
architecture also lend insight into aspects of agency. Only these cope with our
stipulation that the explanandi of social simulations of power and authority must
deal with structure and agency. Nevertheless, in most cases the agent rules have been
implemented on a basis, which is under-representing evidence, bringing about the
problem that structural emergence cannot be related clearly (i.e. intra-contextually)
to agent behavior. Consequently, the analysis of agent behavior cannot be related to
actor behavior in the target system. This lack of realism in model design renders
validation attempts of simulation results against reality less plausible and informa-
tive. By contrast, homologue models of the type advocated for in this chapter enable
the researcher to gain insight into structure and agency of the sort that is more
directly linkable to actor behavior. As a result, validation results become more
plausible and research can enter the hermeneutic cycle of EDABSS.
25.3 Evidence-Driven and Agent-Based Social Simulation
EDABSS models seek homology.5 The modeled mechanisms, structures and pro-
cesses aim at resembling the mechanisms, processes and structures identified in the
target system. This has two reasons: (1) Construct validity renders validation more
5
For a meta-theoretical discussion of what follows see Bhaskar (1979), Boudon (1998),
Cruickshank (2003), Outhwaite (1987), Sawyer (2005), Sayer (1992, 2000).
expressive. (2) An agent-based implementation of the type presented in the follow-

ing sections is more than a mere input–output model. It is an “exhibitionist”
model that allows to analytically focus on internal mechanisms, processes and
structures. From a socio-scientific standpoint this can only be of interest, if the
modeled mechanisms, processes and structures exhibit construct validity in com-
parison with the target system – otherwise the model is just an arcade game
(cf. Boudon 1998).
The key to homology lies in the agent design. It is the agents and their inter-
actions respectively that trigger the evolution of emergent phenomena. Thus, at
the bottom of EDABSS agent design lays an evidence gathering process. Posing the
right questions leads to a collection of data (evidence), which directly informs the
modeling of agent behavior and cognition. We have presented such questions in
Sect. 25.1.
25.3.1 Evidence-Based Modeling
The source of information for homologue models must be evidence-based. This

refers to the fact that all information that is used during the process of model design,
whether derived from a single case or from a theory, must be empirically valid
(see also Boero and Squazzoni 2005). The bulk of this data is of qualitative nature,
stemming from one or a number of case-studies. Case-studies that give concrete
information of actor behavior, in particular social circumstances, are of best use to
EDABSS modelers. Such a presupposition excludes assumption-laden concepts
such as rational choice or Belief-Desire-Intention (BDI). EDABSS’s higher rational
is to found models of social simulation on what is social reality and not what is
methodologically convenient or theoretically desirable (cf. Shapiro 2005). It would
be wrong to stipulate that all the details entailed in the dataset must also be
recognized in an EDABSS model. Modeling is an intellectual condensation process
and it is the modeler who decides what particular aspects of a social phenomenon
are crucial and need to be represented in a stylized way in the model.6
The extensive use of qualitative data in EDABSS can be a virtue of its own,
when statistical data is scarce or not available at all. This applies to a variety of
important topics in the social sciences, such as elites, power structures, conflict or
organized crime. Logically systematic statistical data collection in these areas of
research is difficult. Although the same holds true for qualitative data collection as
well, it is, nevertheless, better feasible. For example, researchers, journalists or
humanitarian aid workers very often have the opportunity to conduct an interview
or to make an observation. Often this data becomes available to the public.
6
The term condensation is alternatively denoted by Stachowiak (1973) as reduction and by Casti
(1997) as allegorization. Other important modeling principles are simplicity and pragmatism
(Lave and March 1975).
EDABSS therefore fills an important lacuna that is set between abstract statistical
modeling and idiographic case-study research as it incorporates the advantages of
both, formalization and context-sensitivity.
The integration of stakeholders in the modeling process plays an important role
in EDABSS. Integrating stakeholders in the modeling process can be rewarding and
delusive at the same time. Stakeholders are keepers of information that others do
not have. For example, if a stakeholder is a powerful person, then he can be
motivated in a semi-structured interview to reflect on why he thinks he is powerful,
how he is acting as a powerful person and how he expects others to behave towards
her/him. On the other hand, stakeholder’s accounts can be deliberately misleading.
Consequentially, EDABSS modelers have to be familiar with qualitative data
collection and analysis techniques.7
To presume that evidence-based modeling ignores theory is not justified.
Evidence-based modeling is guided by theory in many respects. First, critical
realism clearly defines a research project’s explanandi: mechanisms, processes
and structures. Second, it highly depends on the researcher – and is not generic to
evidence-based modeling – how much the research process is guided by theory.
Third, evidence based modeling seeks generalization by intensively studying a
single unit for the purpose of understanding a larger class of (similar) units
(Gerring 2004).8
25.3.2 Endorsements: Reasoning About Power and Authority
Whereas evidence-based models of social simulation incorporate a variety of

structural and processual information of the target system, the actor’s actual
reasoning process cannot be derived from the data. Alternatively, the concept of
endorsements is applied to couch an agent’s reasoning process.
Power relations, as aforementioned, are interactions between at least two actors.
The computational implementation of these interactions must be based on certain
grounds. This can be knowledge an actor has about another actor; it can also be
experiences an actor has made in the past with his environment. Endorsements are a
“natural” way of implementing reasoning about this knowledge or experience.9
They were introduced by Cohen (1985) as a device for resolving conflicts in rule-
based expert systems. Endorsements can be used to describe cognitive trajectories
aimed at achieving information and preferential clarity over an agent or object from
the perspective of the endorsing agent himself. We use endorsements exactly in
7
The literature on qualitative data research has grown considerably in the last years and the
interested reader is referred to, amongst many others, Silverman (2004).
8
We are well aware of the ongoing discussion on induction with regard to case-study research and
the interested reader may refer, amongst others, to Gomm et al. (1992), Eckstein (1992) and Stakes
(1978).
9
See for a more complete treatment of endorsements Alam et al. (2010).
this sense, namely to capture a process of reasoning about preferences and the
establishment of a preferential ordering (Moss 1995, 1998, 2000; Moss and
Edmonds 2005).
Endorsements capture an agent’s (the endorser’s) reasoning process about other
agents (the endorsees). That process projects the endorser’s internal preferences
onto the endorsee. These preferences are represented by an endorsement
scheme which is a collection of categories of possible characteristics of other
agents. These categories of endorsements amount to a partial preference ordering
of characteristics perceived in other agents. The ranking of collections of
endorsements is an essentially arbitrary process. Cohen (1985) used a lexicographic
ordering so that the preferred object (in this case, agent) would be that with the
largest number of endorsements in the most valuable category. If several objects
were tied at the top level, the second level of endorsements would be used to break
the tie and then, if necessary, the third or fourth levels, etc. An alternative is to allow
for a large number of less important endorsements to dominate over a small number
of more important endorsements. One way of achieving this is to calculate endorse-
ment values E for each endorsee as in Eq. 25.1 where b is the number base (the
number of endorsements in one class that will be a matter of indifference with a
single endorsement of the next higher class) and ei is the endorsement class of the
ith endorsement token (Moss 1995). In choosing from among several other agents,
an agent would choose the endorsee with the highest endorsement value E.
X X
E¼ be i bjei j (25.1)
ei 0 ei <0
The process of endorsing an agent must be thought of as being embedded in an

agent’s environmental context, i.e. his neighboring agents (see Fig. 25.1). The
endorsement process allows an agent to find the agent most appropriate to him –
it does not (and cannot) seek the best of all agents.
The main advantage in applying the idea of endorsements lies in the fact that
they allow for combining the efficiency properties of numerical measures, with the
richness and subtleties of non-numerical measures of interest or belief (Moss 1995).
The choice of endorsements and the conditions in which each endorsement will
be attached is entirely context dependent. Agents concerned with critical incidence
management in water supply (Moss 1998) obviously have different criteria than
agents embedded in the context of contemporary conflicts (Geller and Moss 2008a).
While the former might be interested in actions and information models leading to
the successful resolution of a complicated allocation problem, agents in models of
contemporary conflict might be interested in with whom they should cooperate and
whom they should shun or even fight. For example, it is of importance to an agent to
know if its vis-à-vis is of the same ethnicity, religion and kin; if it has lived a similar
past; if it is reliable, corrupt or wealthy. Power and authority relations depend on
knowledge about these kinds of questions. Section 25.4 addresses this as well as the
question of how to translate the evidence into an adequate endorsement scheme in
more detail.
Fig. 25.1 Schematic representation of the embeddedness of the endorsement process
25.3.3 Declarative Implementations of Agent-Based Models
A program is declarative if there is a set of statements on a database, rules have a set

of conditions, which are statements with some values left open as variables, and
consequents exist, which are another set of statements. When all of the statements
in the conditions of a rule are matched by statements on the database, then the
variables are given their specific values from the database statements and the
consequent statements are added to the database. When a set of conditions are
satisfied and a rule fires (that is puts its consequents on the database), then the state
of the environment as represented by the database is changed and perhaps other
rules will now be able to fire and so on until all rules have fired and no further
matches of conditions can be found on the database. The sequence of rules that will
fire and the particular instantiations of their variable values are determined only as
the program is running. The sequence of actions represents the process of agent
behavior and leads in each case to a new state of the environment. If all agents are
implemented declaratively, then they will be changing the state of the environment
for one another and the pattern of rules, and therefore actions of all the agents taken
together will be influenced by one another.
In these circumstances, the outcomes of such a model are usually impossible to
predict with any exactitude.10 Frequently, such models exhibit the sort of episodic
volatility associated in the first section with complexity. The same effect can be
achieved by other means, but declarative representations of agents have a number
of virtues in terms of ease of development as new evidence becomes available and
in terms of yielding comprehensible outputs stored as statements on the databases.
10
Hence, a declarative model architecture does not allow easily for exact simulation replication.
Other disadvantages are that declarative models tend to be computationally expensive and
ontologically complex.
The assumption that the fulfillment of conditions triggers the execution of

consequents marks, in a homologue model, a natural representation of actor behav-
ior. Each actor behaves according to a defined set of rules. A rule fires only when
the set of conditions attributed to this rule is satisfied. Accordingly, an agent’s
behavior is governed on the basis of the fulfillment of conditions. It is fairly
straightforward to translate information on actor behavior obtained during the
data collection process into conditions.
Recall the following question from the opening paragraph: “Under what
conditions would you label someone as powerful?” A possible answer to this
could be “If a person belongs to a well known family.” The condition for an
agent providing the above-mentioned answer to label another agent as powerful is
fulfilled if this other agent belongs to a well-known family (whereas well known
could translate into socially well connected). Similarly, an agent might provide the
answer “If the agent has won five consecutive battles.” In this case an agent has to
follow another agent’s battle history to be able to tell if it can label him powerful or
not. The translation for a possible answer to the second question is analogous.
“Having labeled someone as being powerful, how would you expect that person to
behave?” could be answered by saying “That the person is generous.” However, the
translation now includes two conditions. Firstly, an agent must have been already
labeled powerful. Secondly, an agent must have experienced the powerful agent
being generous to it.
If the conditions of a rule are satisfied, then its consequents are put into effect.
Recall the third question stated in the introduction: “Having labeled someone as
powerful, how would you expect yourself/others to behave towards that person?”
A possible answer could be “Then I would subordinate myself to that person.” The
translation reads as follows: If an agent has labeled another agent as powerful, it
then (as a consequence) subordinates itself to this agent.
Analogously to the examples given, all the collected information that bears
relevance to the modeling process can be translated into declarative program
code. This translation process is essentially an operationalization and formalization
process of (sometimes vague) bits of qualitative information. Power and authority
are not implemented as predefined entities, but are “grown” artificially from a
number of evidence-based rules that are proxies for dimensions of power and
authority. Agents become powerful as a consequence of a variety of causally
interconnected conditions and consequents.
25.4 Modeling Power and Authority: A Case

from Afghanistan
This section presents an implementation of what has been discussed above theoret-
ically. Reflections of power and authority in contemporary conflict are presented,
from which an intuitive, but evidence-informed model of power and authority in
Afghanistan is derived. Against the background of this model and on the basis of
qualitative data, answers in the form of evidence to the questions posed in the
opening paragraph are presented. From these answers the agent rules are being
developed and translated into program code.
25.4.1 Power and Authority in Contemporary Conflicts
The anthropogenic nature of power structures (Popitz 1992) has been shown for a
variety of conflict regions, including Afghanistan (Bayart et al. 1999; Reno 1998;
Roy 1994, 1995).11 Sofsky (2002) argued that conflict societies are societies sui
generis. They function according to their own social laws and are structurally and
processually disjointed from societies lacking a comparable degree of organized
violence. In conflict-torn societies virtually anything goes. This can be illustrated
by the concept of anomie. Anomie is the situation in which the upper and lower
normative boundaries for the aspirations of members of a society are thrown awry
(Marks 1974). An anomic situation emerges when the means to attain a specific
goal, such as accumulation of wealth or power, run out of social control (Merton
1938). Accordingly, in a space emptied of restricting norms, i.e. an anomie,
virtually everything goes along with the creation of power structures to one’s
own ideas and interests.
Anomic spaces are political spaces lacking strong modern institutions, such as
the state’s monopoly on organized violence, stability of the law and protection of
property rights. In these circumstances only highly adaptive stakeholders prevail.
The socio-structural outcomes of this organizational process are manifold and so
are the adopted means that serve one’s interests.
In contemporary conflict societies this outcome is neo-patrimonialism (Geller
2006a; Medard 1990; Reno 1998). Weber (1980) understands patrimonial power as
being based on authority, suppressed subjects and paid military organizations, by virtue
of which the extent of a ruler’s arbitrary power, as well as grace and mercy increases.
Stakeholders interested in gaining power in contemporary conflict settings have to act
neo-patrimonially to accumulate and redistribute material as well as social resources.
The range of related activities is broad and includes corruption, clientelism, patronage,
nepotism, praebendism and so forth (cf. Medard 1990).
25.4.2 An Intuitive Model of Power and Authority

in Afghanistan
Anthropogeneity, anomie and neo-patrimonialism – or any other theoretical context

relevant to a particular research project – have eminent ramifications for the
11
Parts of this and the next paragraph have been taken from Geller and Moss (2007).
perception of power and authority and henceforth for the development of the model
at hand. The evidence presented below should therefore corroborate the implicit
claim that anthropogeneity, anomie and neo-patrimonialism amalgamate to a
framework describing Afghan power structures and functioning as an evidence-
informed theoretical framework that can be filled with the intricacies of the Afghan
case. In the beginning of a research project, such a framework model provides a
theoretically informed ontological entrée for the object of investigation.
The actual information the model at hand rests upon is derived either from data
collected by ourselves or from relevant secondary data sources. The collected
primary data stems from semi-structured interviews conducted with urban Afghan
elites between May 2006 and October 2007. The secondary data stems from case
studies, most of which are of anthropological type, reports published by Non
Governmental Organizations (NGOs), International Organizations, such as the
United Nations (UN) and the International Committee of the Red Cross (ICRC),
or the print media.
Although 27 years of conflict accentuated two important factors in Afghan
society, namely ethnicity and religion, the traditional organizational principle of
the qawm rested sound (Azoy 2003; Roy 1994, 1995; Shahrani 1998). Less men-
tioned, however, is a decline of norms and values in Afghan society leading to a
Hobbesian form of society (Tarzi 1993). Today’s Afghanistan can be characterized
as an anomie (Geller 2010).
The causes for this development are complex, but nevertheless directly linked to
the Jihad of 1979–1989. Although trends of neo-patrimonial politics are already
recognizable in the very beginning of the Jihad – and are indeed a characteristic of
Afghan politics throughout history – the war’s fundamental goals started to mutate
with its increasing duration. Some of the adopted means of warfare have been
traditional, such as organized violence, intrigue, alliance formation and dissolution;
others have been “imported”, such as religious extremism and radicalization of
ethnicity (Geller 2010; Roy 1998).
The concept of qawm is context dependent, defined by such social dimensions as
family, kinship, ethnicity and occupational groups, and also more abstract but
related concepts such as solidarity, rivalry, cooperation and conflict. The notion
of qawm also underlines the fluidity and contextual dependency of social relations
in Afghanistan. Hence, qawm ontologies are devised to codify individual actors,
their behavior and relations between actors, as well as social processes and
structures arising as a result of social interaction (Dorronsoro 2005:10–11).12
Each of those aspects is pertinent to the development of our model, which represent
and clarify social processes associated with these overlapping identity spheres and
the actors acting within them.
12
Monsutti (2004) “explores the basis of cooperation in a situation of war and migration” amongst
the Hazara in Afghanistan through the concepts of solidarity and reciprocity. Nancy Tapper (1991)
“reveals the structure of competition and conflict for the control of political and economic
resources” through the concept of marriage.
The notion of qawm varies not only in the literature, but also amongst Afghans
themselves. It can mean (extended) family, tribe, descent group, ethnicity, “people
like us” (Tapper 2008:88), an “occupational group” (Roy 1992:75), “persons who
mutually assist each other” (Canfield 1973:35) and it can connote a complex
interpersonal “network” (Roy 1995:22; Dorronsoro 2005:10) of political, social,
economic, military, and cultural relations (Mousavi 1997:46–48; Tapper 2008;
Glatzer 1998:174; Rasuly-Paleczek 1998:210–214; Roy 1995:21–25; Shahrani
1998:218–221).13 In fact, our interview data suggests that these meanings are not
mutually exclusive: qawm do not have clear boundaries nor do they divide Afghan
society into mutually exclusive groups. “[A]n individual always belongs to more
than one [qawm].” (Canfield 1988:194)
qawm face competition with other qawm and internal competition amongst
members of a qawm (Azoy 2003; Mousavi 1997:46–48; Roy 1994:74,
1995:21–22). qawm need to be sustained, and it is an Afghan leader’s ability to
redistribute resources that makes him powerful and eventually successful (Roy
1994:74). The ability to create a qawm for a particular aim is also perceived as a
demonstration of power (Azoy 2003:36). qawm still “have a powerful and perva-
sive effect on contemporary political discourse and the behavior of Afghans”
(Shahrani 1998:220) and have during the years of conflict often been misused by
new elites for the pursuit of conflict and criminal aims (Canfield 1988; Rasuly-
Paleczek 1998:210–214; Roy 1994; Rubin 1992; Shahrani 2002; Tapper 2008).
Manifestations of such abusive behavior are for example corruption, drug produc-
tion and smuggling, nepotism, massive organized violence, crime, ethnic, political
and religious radicalization (Giustozzi 2007; Glatzer 2003; Rubin 1992, 2007;
Schetter et al. 2007). qawm are not the cause for conflict in Afghanistan, as these
causes are manifold (Shahrani 2002:716; Dorronsoro 2005), but we will explore the
usefulness of the notion of qawm in analyzing and understanding conflict in
Afghanistan.
Figure 25.2 depicts an informed intuitive and ideal-typical representation of a
qawm. It consists of ten actor types: politicians, religious leaders, commanders
(meritocratic title for a militia leader), businessmen, warriors, civilians, farmers,
drug farmers, organized criminals and drug dealers. An important abstraction from
reality is that in our model each actor has its distinct role, whereas in reality actors
may incorporate a variety of roles. For example, a commander can be a (military)
commander, a politician and a drug lord at the same time. We proxy individual role
pluralism by mutual interdependence, i.e. each actor has virtues another actor may
be in need of and vice versa, leading to mutual cooperation and interdependence.
This, of course, is also a common pattern in reality, where there is no clear
distinction between role incorporation and cooperation.
13
Whether a qawm denotes a group or a network is not clear from the evidence. Following
Tapper’s (2008) argument, a qawm can take the form of a group or a network, depending on the
context.
Fig. 25.2 A case-study informed intuitive model of a qawm
The following examples explain the qawm-model in terms of agency. If a

politician is in need of military protection, he approaches a commander. In return,
a commander receives political appreciation by mere cooperation with a politician.
If a businessman wants to be awarded an official construction contract by the
government, he relies on a politician’s political connections. In return, the politician
receives a monetary provision, for example, bribes. If a politician wants beneficial
publicity, he asks a religious leader for support. The religious leader, in return,
becomes perceived as a religious authority. If a warrior seeks protection and
subsistence for his family, he lends his services to a commander, who, in return,
provides him with weapons, clothes, food and/or money. If an organized criminal
wants to carry drugs, he relies on the transport business of a businessman who, in
return, receives a share of the drugs sold. If a drug farmer needs protection for his
poppy fields, he a affiliates with a commander, who, in return, receives a tithe of the
drugs sold to a local drug dealer. According to Azoy (2003), such interactions are
also guided by the following four social categories: kinship, residence, class, and
religion. Our model represents this neo-patrimonial behavior. The links between the
agents in Fig. 25.2 can also represent such categories.
The continuing existence of the qawm in times of severe social change as a
means to organize and manage power cannot baffle the fact that the qawm itself has
undergone configurational alteration. Protracted conflict deteriorated not only
social structure but also obliterated moral boundaries. Corruption, fraud, mistrust,
crude materialism, and the like systematically found their way into Afghan society.
The power of the qawm and of its members to constantly adapt to new states of
anomie epitomizes the anthropogenic nature of power and authority (Geller 2010).
25.4.3 An EDABSS Model of Power and Authority

in Afghanistan
25.4.3.1 Evidence
The collected interview data explicitly highlights three sources of power in current
Afghanistan: Ownership, reputation, and qawm (cf. Azoy 2003). Their meaning
manifests when mirrored against the notions of hisiyat and e’tibar. hisiyat and
e’tibar, two Dari words that roughly translate into “character” and “credit”. hisiyat
denotes qualities such as piety and wisdom; e’tibar is about meritocracy.
A powerful actor must dispose of hisiyat and e’tibar.
Traditionally, ownership can be defined as land, access to water, livestock and
women. References to landownership were made often during interviews, whereas
water and women have never been mentioned. Ownership of livestock was mainly
mentioned to serve reputational means in order to increase one’s own or someone
else’s reputation. The interview data and observations made during the field trips
suggest that a modern comprehension of ownership has become more materialistic
and less subsistence oriented; mundane symbols of power such as money, houses
and cars have increased in importance. This raises the issue of the sources for these
goods. While some Afghans undoubtedly were able to build up prospering
businesses or brought assets with them from exile, other sources remain dubious
and are likely to include organized crime, corruption and clandestinely working for
foreign countries. Thus, a generalized answer to “When would you label someone
as being powerful?” includes that, either being traditional or modern, or, more
likely, a combination of the both, a powerful individual must have access to
ownership resources. “[. . .][P]ower without wealth is all but possible.” (Azoy
2003:30) But “[w]ealth is just a means to achieve prestige.” (Roy 1994:74)
In Afghanistan, authority is nourished by reputation. It is the “ultimate source of
political authority.” (Azoy 2003:31) Reputation exhibits a static and dynamic
component. The static component is closely related with ancestry. It is important
what ethnicity, family or tribe someone belongs to. Hazaras, one of the four major
ethnicities in Afghanistan, are often regarded as working class, while Pashtun, the
largest ethnicity, are often perceived as a warrior elite. A family may be regarded as
politically powerful and/or religiously influential. To have roots in a political,
religious or scholarly family provides authority. This became obvious during
interview sessions in a variety of ways. Interviewees have regularly been
introduced or have introduced themselves by referring to their family either in a
political or religious context. Two outspoken authoritarian interviewees, for exam-
ple, claimed to be Sayyed, i.e. descendants from the Prophet, and scholars at the
same time. The dynamic aspect of reputation relates to an individual’s historicity.
Individual politico-historical background is important. Individual histories link
actors to different social groups and can provide them, depending on the social
context, with esteem, such as in the case of having been a Mujahedin, a resistance
fighter during the time of Soviet occupation. Thus a generic answer based on the
interview data collected to “when would you label someone as being powerful?”
would include that a powerful individual belongs to an important family and/or has
played an important role in his past.
The qawm is the epitomization of network-based power structures in
Afghanistan. A qawm is more than mere reference to an actor’s ethnicity or tribe;
it is more than an extended family. The qawm is an actively used instrument in the
pursuit of power. An appropriate translation would therefore be “power basis”: a
number of people that can be mobilized to achieve a particular political aim. With
changing political aims, this group of people may change as well. Thus a
generalized answer to the question “when would you label someone as powerful?”
includes that a powerful individual not only disposes of a qawm, but also exerts
control over it.
25.4.3.2 Model and Computational Implementation
Based on the evidence, how are a powerful actor and his entourage to be structurally
modeled? First, it needs to be defined who is powerful and who is not by disposi-
tion. The evidence presented above indicates that the following agent types should
be considered as being powerful in our model: Politicians, businessmen, religious
leaders, commanders and organized criminals. These agents are computationally
created as being powerful by definition, because they are politicians, religious
leaders, commanders, etc.14 Hence, social resources do not have to be distributed
explicitly, agents are “born” possessing them – or not.
Secondly, agents must, at the initialization of the model, be equipped with
material resources. These assets are given, pars pro toto, in the form of money,
drugs and land. The absolute amount of money distributed in the model is arbitrary
as real numbers are not available. Money is distributed log-normally amongst
agents of one particular agent type. A lognormal distribution of wealth also appears
for the case of Afghanistan a plausible assumption (Limpert et al. 2001). Hence, a
small number of agents are very rich, while most of the people are poor. The data
record for land holdings, which is better than the one for wealth, suggests that
holdings of land should be distributed log-normally as well (Wily 2004). Again, this
means that a small number of agents possess a lot of land, while a large number of
agents will only own little land. Last but not least, drug farmers receive some drugs
in the beginning of the simulation and harvest drugs during the simulation in
specified harvesting periods (UNODC 2006).
Thirdly, agents must be given an internal state, representing what has been
identified above as hisiyat and e’tibar. This internal state is an agent’s endorsement
scheme, which is depicted in Table 25.1. Some of these labels (the endorsements)
are static, others are dynamic. Static endorsements are attributed to each agent at
14
Note that we are not simulating the genesis of a powerful agent, but the emergence of power as a
network-like structure in an evidence-based, artificial society. See for the qualitative description of
such a genesis Giustozzi (2006).
Table 25.1 Crosstabular presentation of an agent’s endorsement scheme

Static Dynamic
hisiyat Intellectual/non-scholarly Loyal/disloyal
Shared-ethnicity/different-ethnicity Trustworthy/untrustworthy
Shared-religion/different-religion Is-neighbor/non-neighbor
Is-kin/non-kin Pious/sinful
Politico-military-background
e’tibar Reliable/unreliable
Successful/unsuccessful
Capable/incapable
the beginning of a simulation and cannot be changed during the course of the
simulation: an agent is either an intellectual or he is not; he is either a Tajik, another
important Afghan ethnicity, or he is not; he is either my brother or he is not, etc. For
the time being it is not implemented that an agent’s changing relationships are taken
into account to form an individual political profile, hence the politico-military
background is static. Some hisiyat endorsements are dynamic, such as loyalty,
trustworthiness, neighborhood or religiousness and can change their respective
values during a simulation run. All e’tibar endorsements are, by contrast, dynamic,
as e’tibar, i.e. meritocracy, is inherently dynamic and must call for a dynamic
conceptualization. Formerly reliable agents can become unreliable and previously
successful agents can become unsuccessful, etc.
The endorsement scheme does not only depict an agent’s internal state but also
denotes the categories in which an agent reasons about other agents. For example,
assume agent A is Tajik, a Mujahedin, successful and neighbor of agent B. Further
assume that agent B is Tajik as well, was also a Mujahedin and evidently is also
neighbor of A. It is then likely that agent B will look favorably at agent A and vice
versa and that the two will establish an affiliation with each other. The endorsement
scheme breaks down why an agent is more powerful than another agent: Because it
disposes over an internal state that is seen favorable by other agents. Or to put it
differently: Because it internalized a number of qualities that are perceived as
symbols of power by other agents.
It is important to note that although an individual agent is defined at the moment
of its creation as being powerful or not per se, nothing is said about how powerful it
is going to be, or spoken differently: How good of a neo-patrimonial agent it will be.
The agent’s performance as well as the social product of its performance, the qawm,
are emerging out of the social simulation model and are not computational artifacts.
25.4.3.3 Behaving Powerfully
Having clarified what makes actors powerful in Afghanistan, it is now important to

know how these powerful actors behave and how other actors behave towards
powerful actors. The corresponding questions from the introduction are: “Having
accepted someone being powerful and/or being an authority, how would you expect
that person to behave?” And: “Having labeled someone as powerful/as being in a
position of authority, how would you expect yourself/others to behave towards that
person?” The freedom of choice an actor has to behave towards a powerful agent
depends on a number of different factors (hisiyat and e’tibar) but is foremost based
on the distinction whether an actor is powerful himself or not. The two
constellations powerful actor/weak actor and powerful actor/powerful actor lead
to different outcomes in the social organization. In our understanding these are
patron-client-relationships (powerful/weak) or affiliations (powerful/powerful).
However, in both cases, accumulating and redistributing resources is in the center
of actor behavior.
Weak actors have only little choice of how to behave towards a powerful actor
and are forced into a patron-client relationship because of grievance. They are
either locked in into economic dependency or are not even a member of the
powerful actor’s qawm and thus cosmos. In both cases weak actors can only submit
themselves and must fully depend on the powerful actor’s gratitude. Consequen-
tially, in patron-client-relationships, powerful behavior is more of a mundane sort,
for example supporting a client’s family with food, clothes and housing. Depending
on what kind of client it is, the patron might ask for dogs body services in the case of
a civilian, for protective services in the case of a warrior, or for zakhat (tithe) in the
case of a farmer. In general, the patron demands loyalty for his support. The weak
actor exerts power over the powerful actor indirectly insofar as it is harmful for a
powerful actor’s reputation to pay subsidies irregularly or not at all. The weak actor
exerts power directly in case of grass-root opposition as a result of insensitive
politicization by the powerful actor. This social relationship of dependence is
standard in Afghanistan and every powerful man surrounds himself with such a
“service force”, be it small or big. Nevertheless, supporting the weak should not be
considered as being unimportant, as they provide a basis of broad social support.
Moreover, supporting the weak increases a powerful man’s e’tibar as the following
ideal-typical characterization of a Pashtun men highlights: He is a men of honor,
with prowess and pride, whose dignity does not forbid him to be attentive to
authorities as well as the weak (Janata and Hassas 1975:84; translation ours).
Powerful behavior between two or more powerful agents is of a different nature.
Powerful actors have the freedom to choose their behavior towards their vis-à-vis
and can either act cooperatively, conflictously or submissively. Which type of
behavior is chosen, is based on a deliberate but nevertheless fragile assessment of
who must be considered a supporter and who must be considered a spoiler or even a
foe. Hence, whatever the project is that is of concern to a powerful actor, it is
intensively discussed with those people from his social circle who he thinks should
be included in the decision making process (Azoy 2003).
The evaluation of a project’s – and ultimately of a powerful actor’s behavior –
supporters and opponents constitutes the initialization of a project-based qawm. The
organization of a qawm, whether it is a generic qawm based on kinship or a
temporal qawm, is a delicate operation. Exacerbating is the fact that potential
supporters as well as potential enemies are competitors against whom the powerful
actor must stand up – at all time. Hence, the creation and maintenance of qawm
explains the volatile nature of cooperation, the often and sudden changes of
alliances and the ubiquity of conflict (cf. Azoy 2003).15
In the case of cooperation, powerful actors establish affiliations between each
other. Affiliations are relations between qualitatively equals. Consequentially,
powerful agents do not give each other material support, but they provide each
other with social resources. While powerful actors in patron-client relationships
accumulate social resources for redistributing material resources, powerful actors in
affiliations mostly accumulate and redistribute social resources. Politicians guaran-
tee that commanders are not denigrated as “warlords”, commanders protect
politicians, religious leaders openly designate politicians of being pious, politicians
declare a religious leader their spiritual leader, businessmen – financially support a
politician’s campaign, politicians provide businessmen with lucrative state
contracts, etc. In short, powerful actors support each other in in-creasing their
hisiyat and e’tibar record. The opposite, of course, exists as well. Powerful actors
can actively engage in diminishing another actor’s hisiyat and e’tibar record.
In summary, a powerful actor is able to control his qawm economically and
socially. He is able to redistribute enough material and social resources to keep his
qawm alive. If one would have to measure the power an actor in Afghanistan has,
then it would not suffice to only count his material assets. A comprehensive
measure of power would include this actor’s reputation, measured in terms of his
ability to “call on the services of supporters to help him in whatever enterprise”
(Azoy 2003:32).
25.4.3.4 Model and Computational Implementation
The computational implementation of the behavior of a powerful agent is straight-

forwardly informed by the evidence presented above. Powerful agents want to
accumulate and redistribute material and social resources. For this reason their
hisiyat and e’tibar needs to be relatively superior to their competitors’. Hence, a
powerful agent’s aim must be to establish as many favorable relationships as
possible. He does this in two steps: First he reasons about which agents to endorse
and second he takes action on the basis of his decision he has taken in step 1.
Step 1 consists of the endorsement process as explained in Sect. 25.3.3 and as
contextualized in Sect. 25.4.3. Each agent continually checks all the agents visible
to him – not all of them are –, i.e. he projects his endorsement scheme upon them,
rates the corresponding values, calculates E, compares all Es against each other and
chooses the one with the highest E to endorse. According to the evidence, the model
implies that during the endorsement process those agents are more likely to
15
We do not consider the emergence of conflict in this chapter. See for a preliminary discussion
Gerring (2004).
Table 25.2 Crosstabular presentation of a sample endorsement between one civilian and two
politicians
Endorsements Civilian Politician1 Politician2
Same/different ethnicity (1) Tajik Hazara/–2* Pashtun/–2*
Same/different politico-military Mujahedin Mujahedin/4 Mujahedin/4
background (2)
Reliable/unreliable (3) – unreliable/–8* reliable/8
E – –6 10
The values in parentheses indicate the importance of an endorsement. Values on the left side of
each slash represent “reality”; values on the right side of each dash represent the “weighed reality”.
Labels marked with an asterisk have been multiplied with 1, because in Eq. 25.1 they are part of
the second Σ sign and are therefore subtracted
establish a relationship with each other if they are similar with regard to hisiyat
endorsements and who exhibit higher values with regard to e’tibar endorsements.
Two short examples shall clarify this, one for patron-client-relationships and one
for affiliations.
Powerful agents do not seek ordinary agents; rather they are sought out by the
latter. But even though ordinary agents are in misery, they try to choose the powerful
agent who is most suitable for them, given the choice. Assume a civilian, who is
Tajik, a Mujahedin and in need of material support (see for what follows also
Table 25.2). (In the simulation this is the case when the civilian’s holdings of
money are 0.) Before asking every politician for material support the agent can
see, it checks which of the visible politicians are most suitable to him. Assume that
there are two visible politicians. The civilian agent then projects its endorsement
scheme upon each of the two politician agents. Recall that the endorsement scheme
not only tells the civilian how important particular characteristics of these two
politicians are to it, but also what the categories of its own perceptions are. In
Eq. 25.1 b denotes the number base for which every agent is randomly assigned a
value b > 0. (For b ¼ 0 the expression bx is equal to 0, independently of the
exponent x.) In this example, the civilian is assigned b ¼ 2. In Eq. 25.1 ei denotes
the value of the ith endorsement token. This value differs for each endorsement
token and for each agent. In this example the endorsement tokens for the civilian
are same-ethnicity/different-ethnicity, same-politico-military-background/different-
politico-military-background and reliable/un-reliable. Each endorsement token is
randomly assigned a value ei. ei differs for each agent. In the present example the
following values for ei are assigned for the civilian: 1 for same-ethnicity, 1 for
different-ethnicity, 2 for same-politico-military-background, 2 for different-polit-
ico-military-background, 3 for reliable and 3 for unreliable. These values are on an
ordinal scale and represent the civilian’s endorsement scheme. The interpretation of
this endorsement scheme is that the civilian perceives the labels same-ethnicity/
different-ethnicity less important than the labels same-politico-military-back-
ground/different-politico-military-background and the labels same-politico-military-
background/different-politico-military-background less important than the labels
reliable/unreliable. Politician1 is a Hazara, a Mujahedin and has never been endorsed
reliable by the civilian. Politician2 is an Uzbek, a Mujahedin and has been endorsed
reliable in the past by the civilian. With regard to politician1 the civilian reasons as
follows: Politician1 is not of the same ethnicity, has the same politico-military-
background and is unreliable. Depending on this information and the values
assigned above, (Eq. 25.1) allows the calculation of E, which is: 22 2|1| 2|3|
¼ 6. For politician2 the civilian reasons as follows: Politician2 is not of the same
ethnicity, has the same politico-military-background and is reliable. Thus, E can be
calculated as follows: 22 + 23 2|1| ¼ 10. Because politician2 has a higher E than
politician1 the civilian decides to choose politician2. This procedure can be extended
to as many agents and to as large an endorsement scheme as necessary.
Once the civilian has taken its decision, which politician to ask for support, the
agent sends a message to this particular politician, requesting support. As long as
the politician has enough money to support the civilian, it accepts the support
request. The acceptance of the support request is tantamount to the establishment
of a patron-client-relationship. Each simulation iteration, the politician pays the
civilian a defined amount of money, as long as it has enough money to do so. In
return, the civilian is required to endorse the politician as being trustworthy. In the
event that the politician cannot pay the civilian anymore, the latter endorses
the politician being unreliable and untrustworthy, leading to a break-up of the
patron-client relationship. The civilian then has to seek another politician and the
endorsement process starts again. All requests for support by ordinary agents follow
this scheme.
Interactions between powerful agents processually do not differ substantially
from the scheme described above. Consider the case where a commander wants a
trustworthy assertion from a politician in order to not be denigrated as “warlord”.
Assume that two politicians are visible to the commander. Again, before sending
off a message to the most suitable politician, the commander assesses which of the
two politicians are most suitable to it, i.e. which one has the higher E. Although
different and further endorsements might apply in this case, the endorsement
process, as described in the civilian-politician case, does not change in essence.
Once the commander has chosen a suitable politician – one, for example, who is of
the same ethnicity, the same politico-military back-ground and who has a record of
being trustworthy – it then sends a message to this politician requesting to be
endorsed trustworthy. Because the commander has chosen a politician with a
trustworthy record during its endorsement process, the politician is able to do so,
i.e. to endorse the commander trustworthy. However, the politician demands a
service in return. When dealing with a commander, this service is naturally protec-
tion. Hence, the politician agent sends a message to the commander agent, stating
that it will endorse it trustworthy, but only if the commander agent can protect the
politician agent. Naturally, the commander can provide protection only, if there is at
least one warrior with whom he has a patron-client relationship, established
according to the procedure described above. Given he can provide protection, he
accordingly answers the politician’s message. This mutual fulfillment of conditions
triggers a number of mutual endorsements: The politician endorses the commander
as not only being trustworthy, as requested, but also as being capable, as he is
capable of providing protection; the commander endorses the politician as being
trustworthy. Mutually endorsing each other is tantamount to establishing an affilia-

tion. This affiliation holds as long as the conditions leading to it hold true: “Is the
commander still able to provide protection?”, “Is he still capable?”, “Is it still
trustworthy?”, “Is the politician still trustworthy?”. Both agents do this every
simulation iteration. Moreover, all relationships, whether they are patron-client
relationships or affiliations, break up when there is an agent (endorsee) found
who fits an endorser better, because agents continually scan their neighborhood
for better opportunities.
The example given above is representative for all the interactions between
powerful agents: Politicians request armed protection from commanders for a
trustworthiness assertion; commanders approach businessmen to invest money;
businessmen pay politicians for courtesy services; politicians request a pious
assertion from a religious leader for a trustworthiness assertion; religious leaders
ask commanders for protection and, in return, assert a commander as being pious.
All these affiliation interactions describe the neo-patrimonial usage of social capital
as well as the accumulation and redistribution of material resources.
The quintessential mechanism is the endorsement scheme. It is the interface
through which two agents communicate with each other and it is the raster that
filters those who are “endorsables” from those who are “un-endorsables”. If two
agents match to establish a patron-client relationship or an affiliation, is ultimately
decided via the mechanism of the endorsement scheme. Hence, the evidence that is
enshrined in the endorsement scheme (see Sect. 25.4.3 and there in particular
Table 25.1) finally leads to the emergence of social reality.
25.4.3.5 Emerging Structures: Simulation Results
It would be beyond the scope of this chapter to discuss the simulation results and their
validation in detail (see Chaps. 8, 9 and 12) in this volume, also Moss 2007 for more
on how to do this).16 Figure 25.3 depicts the output of a representative simulation run
at time t ¼ 100. There are ten different agent types and the total number of agents is
190: 6 politicians, 6 religious leaders, 6 businessmen, 6 organized criminals, 6
commanders, 10 drug dealers, 35 drug farmers, 35 farmers, 70 civilians and 20
warriors. In summary, these effects represent the emergence of higher order organi-
zational structures or qawm out of the micro-processes and -structures introduced
above.
Agents affiliated with each other are linked via a line. Two distinct but never-
theless interconnected clusters of agents are apparent in the network. Each cluster
consists of a variety of agent types. This means that the depicted clusters are not
homogeneous organizations of power, but rather heterogeneous concentrations of
power generated by mutually dependent and interacting agents/actors. Agents
16
This paragraph is a condensed version of Geller and Moss (2008a). See also Geller and Moss
(2007; 2008b).
farmer-2
civilian-55 farmer-24
drug-farmer-17
civilian-69
civilian-33
civilian-28 civilian-21
civilian-65
civilian-60 warrior-18 civilian-52
drug-farmer-5 drug-farmer-27 civilian-64 warrior-25
civilian-45 politician-4 farmer-4
businessman-0
farmer-6 drug-farmer-1
drug-dealer-0
warrior-21civilian-31 civilian-20
civilian-44
warrior-10 commander-2 warrior-22
organised-criminal-5
businessman-5
civilian-2 drug-dealer-5
religious-leader-1 religious-leader-3
civilian-25
religious-leader-4 civilian-30 civilian-22
warrior-4 drug-farmer-7
civilian-35 politician-1 civilian-5 civilian-38 civilian-23
civilian-11
drug-dealer-7
civilian-66 farmer-14 civilian-24 drug-farmer-9 farmer-33 farmer-30
civilian-46 civilian-26 drug-farmer-32
civilian-3drug-farmer-18
civilian-59
civilian-8drug-farmer-0 drug-farmer-11
farmer-23 drug-farmer-26
civilian-9religious-leader-2
civilian-34 warrior-12 drug-farmer-14
civilian-43 politician-3
farmer-10
drug-farmer-20 farmer-27
warrior-24
warrior-23
civilian-18 farmer-13
warrior-3 civilian-7
drug-dealer-9
drug-farmer-8 drug-farmer-24 drug-farmer-28
civilian-61
farmer-9 civilian-48 farmer-7 drug-farmer-6
farmer-16
warrior-0 drug-farmer-3 civilian-54 civilian-62
drug-dealer-4
drug-farmer-25
civilian-36 farmer-34 civilian-39
civilian-29
civilian-27 drug-farmer-22 drug-dealer-6
warrior-5
commander-5
warrior-17 farmer-8 farmer-5warrior-27
drug-dealer-1 farmer-32
civilian-51
warrior-2 warrior-26 civilian-49
civilian-17 civilian-4warrior-19 farmer-18 farmer-26farmer-0 drug-farmer-12
farmer-17 warrior-9
commander-0
religious-leader-0
farmer-1
warrior-14 drug-dealer-2
civilian-15 drug-farmer-2
farmer-19
drug-farmer-30 civilian-42
organised-criminal-3 civilian-14
politician-2 civilian-58
warrior-7 businessman-1
farmer-22 drug-farmer-13 organised-criminal-4 warrior-11
farmer-20
farmer-3
civilian-67warrior-16 drug-farmer-16
drug-farmer-33 organised-criminal-0
drug-farmer-31 warrior-13 civilian-53
drug-farmer-29
civilian-16
businessman-3
religious-leader-5 civilian-19 drug-dealer-3 farmer-29
commander-1 businessman-2 drug-farmer-23
farmer-31 warrior-15 civilian-57
farmer-28
civilian-37 civilian-40 businessman-4
warrior-8 drug-dealer-8
politician-5 drug-farmer-19
warrior-20 civilian-12
civilian-63 drug-farmer-21 drug-farmer-4
farmer-15 drug-farmer-34 farmer-21 civilian-1 warrior-6 commander-3
organised-criminal-2
organised-criminal-1 farmer-25
commander-4
civilian-68 politician-0
farmer-12
drug-farmer-15 civilian-32
drug-farmer-10
civilian-6
warrior-1 farmer-11
civilian-41
Fig. 25.3 Artificial Afghan power structures depicted as a relational network of Afghan agents
assumed to be more powerful than others, i.e. politicians, commanders, religious

leaders and organized criminals, are prevalent in each of the two dense clusters. The
reasons for the evolution of this network of clustered affiliations are manifold:
Agents affiliate because they share the same ethnicity or religion; because they have
established a business relationship or because they seek protection with a com-
mander. But in general, the clusters can be perceived as emergent properties of
agent neo-patrimonial behavior as reified by our agent rules. The model generates
data of the sort we expected and its output can therefore be considered as artificial
representations of qawm.
A cogent argument for this claim is the fact that a cross-validational analysis
between the network constructed from simulation output and a real network,
constructed by Fuchs (2005), was successfully conducted. Cross-validation
provides a link between the model and its target system, i.e. reality (Moss and
Edmonds 2005). Fuchs (2005) has collected open source data on Afghan power
structures for the years 1992–2004/2005, which can be compared with the
AfghanModel network depicted in Fig. 25.3.17
17
Fuchs (2005) only collected data of elites. In order to compare her results with those generated
from the simulation presented here, all ordinary agents, i.e. non-elites, had to be removed from the
network to meaningfully calculate the desired network measures. Note that the simulation
parameters remained unchanged. Note also that the two networks vary in size: 62 agents partici-
pate in the Fuchs network and 30 in the AfghanModel network. This can lead to boundary
specification problems.
The densities for the Fuchs and the AfghanModel networks are 0.0593 and
0.1943* (only strongmen, marked with *) and 0.0697 (208 agents, marked
with {). The clustering coefficient for Fuchs is 0.428; while it is 0.542* and
0.432{ respectively for the AfghanModel. Both the Fuchs and the AfghanModel
networks tend to be small world. They exhibit sub-networks that are characterized
by the presence of connections between almost any two nodes within them. Most
pairs of nodes in both networks are connected by at least one short path. There is an
over-abundance of hubs in both networks. The average geodesic distances for the
Fuchs and the AfghanModel networks, 2.972, 3.331* and 2.628{, are relatively
small compared to the number of nodes. Erdős-Rényi random networks of equal
size and density as the Fuchs and the AfghanModel networks exhibit lower cluster-
ing coefficients than the Fuchs and the AfghanModel networks, namely 0.060,
0.212* and 0.072{ – an indication that neither the clustering of the Fuchs nor of
the AfghanModel network is random. The two Erdős-Rényi random net-works have
geodesic distances (2.094*, 2.276{) which are of a comparable order as the geodesic
distances that can be found in the AfghanModel network (3.331*, 2.628{). Thus, a
structural and functional equivalence based on qualitative evidence between the
model and the target system is observable.
25.5 Conclusions
The main task of this chapter was to provide a critical overview of state of the art
models that deal in alternative ways with power and authority, and to present an
alternative research design that overcomes the inefficiencies and shortcomings of
these approaches. The work presented is motivated by the fact that research on
power structures is confined on one hand by a general lack of statistical data. On the
other hand, the literal complexity of power structures requires a formalized
and dynamic approach of analysis if more than a narrative understanding of the
object under investigation is sought. The case of Afghanistan has only been
instrumentalized to exemplify such an approach, which would work without
doubt for any other comparable contexts.
With regard to the case in hand the explanandum is power and authority in
Afghanistan. The analysis focuses on power relations dominated by elites. The
qawm, a fluid and goal-oriented solidarity network, has been identified in the data
available to us as the pivotal structural and functional social institution to manage
power and authority in Afghanistan. The totality of qawm in Afghanistan do not
form a unified system of power but a cosmos of mutually interacting power
systems. This qualitative analytical result has been reproduced by our simulation
results and was subsequently cross-validated with independent out-of-sample
network-data. This cosmos is a root source for political volatility and unpredict-
ability and ultimately an important explanatory factor for conflict in Afghanistan.
For the time being, the latter only outlines an a priori statement that needs further
corroboration.
The proposed approach starts with an evidence-informed but intuitive model of

power and authority in Afghanistan. Such a model provides an intellectual entrée,
identifies and defines a target system, isolates relevant actors and generic actor
behavior, and helps to address appropriate research questions. Based on this intui-
tive model, agent structures and rules are developed according to the evidence that
is available. The aim is to develop a homologue, i.e. a construct valid model, which
allows translation of the information describing the model with regard to structure
and processes into program code. Weak agents have been implemented according
to the rational that they seek affiliation with a powerful agent because of grievance
and their will to survive; a powerful agent is implemented according to the rational
that he affiliates himself with other powerful agents in terms of functional comple-
mentarity to subsist his solidarity network, i.e. his qawm, with the aim of
consolidating or increasing his power. The general underlying notion of such
behavior is neo-patrimonialism. The simulation results are not self explanatory
but need to be validated against reality. This not only provides new insight into the
target system but also obliges to restart the research process as new evidence has
become available. This is the hermeneutic circle of EDABSS, in which the manner
of how to present simulation results meaningfully to stakeholders still needs to be
considered as being in its experimental stage (though companion modeling has
shown a very fruitful way forward). The approach introduced here does also
constitute a consequent implementation of generative social science, for which,
nevertheless, the research design still needs further formalization and clarification.
Special consideration was given to a cognitive process called “endorsements”.
The idea of endorsements serves two aims: firstly, to differentiate and define the
relevant dimensions agents reason about and, secondly, to implement agent cogni-
tion in a natural way (being able to use the mnemonic tokens found in the evidence
informing the model). Endorsements as they were introduced here lack two impor-
tant features: that agent types should have randomized, type-specific endorsement
schemes, and that Eq. 25.1 to calculate E should allow for continuous data
formalization. Alam et al. (2010) have proposed a solution for these problems.
What has been gained by this approach? First, tribute has been paid to reality by
taking it seriously into account. Although evidence-driven modeling is about
abstraction and formalization like any other modeling technique, it does it on the
basis of evidence. Secondly, while agent-based modeling accounts in general for an
epistemological shift from an intra-modeling view and disburdens modeling from
serving only a mere input–output function, evidence-driven modeling facilitates
cross-validation also between model and target system inter-agent mechanisms and
on the systems’ respective meso-levels. Thirdly, the inclusion and generation of
narratives in evidence-driven modeling opens up new ways of engaging
stakeholders and domain experts in the modeling process and in informing policy
analysts and makers in their decision-making.
However, the heuristic usefulness and value of the applied approach can only be
determined against the actual object under investigation, i.e. Afghanistan. What has
been found by modeling Afghan power structures evidence-driven and agent-based
that would have not been detected by applying only a hermeneutic analysis?
Modeling requires not only abstraction, but also formalization and thus disambigu-
ation of the evidence describing the case at hand. In particular we were forced to
dissect the notion of qawm, an inherently context-dependent and fuzzy concept,
with regard to actor behavior and cognition, assigning clear meaning to all
mechanisms deemed to be relevant to the model. This being only a beneficial side
effect of evidence-driven social simulation, emerging social processes and
structures are in the focus of description and analysis, social dynamics that could
hardly be made graspable by pure narrative analysis. Here this is the evidence-
driven generation of a model-based demonstration of an autopoietic system of
power structures taking the form of a small world network.
From the point of view of complexity science it is interesting to observe how the
introduction of localized mechanisms of power generate order on a higher level, or
to put it differently: How neo-patrimonial behavior and processes of accumulation
and redistribution in a state of anomie create a social structure which can be found
in Afghanistan, i.e. qawm. Further, EDABSS clarifies an important aspect of
emergentism: It is not sufficient to state that in agent-based modeling agents
constitute their own environment (cf. Cederman 2001). In lieu thereof it should
be replenished that agents and the emergent effects stemming from these agents’
interactions as a whole constitute an agent’s environment. Agent behavior on the
micro-level and social structurization on the macro-level cannot be thought of as
disjointed entities, but emblematize the wholeness of what is in essence a complex
system. EDABSS therefore is not only a methodological solution to the
micro–macro gap problem, but also an implementational one.
EDABSS implies more than inductive reasoning about social phenomena. It
constitutes a consequent implementation of the generative social science paradigm.
Evidence is the starting point of the research process and evidence denotes its end.
It is the nature of the object under investigation that shapes the theoretical and
methodological approach and not vice versa. This is perhaps EDABSS’ most
important virtue that it takes reality and its subjects seriously: during evidence
collection, model development and validation.
Acknowledgments We would like to thank Bruce Edmonds, Seyed M. Mussavi Rizi, Martin
Neumann, and Flaminio Squazzoni for thoughtful and helpful comments. We also thank Zemaray
Hakimi for translation and facilitator skills, Sayyed Askar Mousavi for advice in the data
collection process, Shah Jamal Alam, Ruth Meyer and Bogdan Werth for modeling support, and
the Bibliotheca Afghanica for access to its library.
Further Reading
Whereas the literature on power and authority is overwhelming, published work on

power and authority and modeling and simulation is, comparatively speaking,
meager. For further reading we suggest Alam et al. (2005) and Rouchier et al.
(2001) for models concerned with the emergence of structures and authority in gift
exchange; Geller and Moss (2008a) and Alam et al. (2008) for empirical models
relevant to power and authority; Axelrod (1995) and Cederman (1997) for
applications of modeling power to conflict in international relations; Mailliard
and Sibertin-Blanc (2010) for a good discussion of multi-agent simulation and
power from a sociological perspective; and finally Guyot et al. (2006) for a
participatory modeling approach with relevance to power and authority.
References
Alam SJ, Hillebrandt F, Schillo M (2005) Sociological implications of gift exchange in multiagent
systems. J Artif Soc Soc Simul 8(3), http://jasss.soc.surrey.ac.uk/8/3/5.html
Alam SJ, Meyer R, Ziervogel G, Moss S (2008) The impact of HIV/AIDS in the context of
socioeconomic stressors: an evidence-driven approach. J Artif Soc Soc Simul 10(4), http://
Alam SJ, Geller A, Meyer R, Werth B (2010) Modelling contextualized reasoning in complex
societies with ‘endorsements’. J Artif Soc Soc Simul 13(4), http://jasss.soc.surrey.ac.uk/13/4/6.
html
Axelrod R (1995) A model of the emergence of new political actors. In: Gilbert N, Conte R (eds)
Artificial societies: the computer simulation of social life. Routledge, London/New York,
pp 19–39
Azoy GW (2003) Game and power in Afghanistan, 2nd edn. Waveland Press, Long Grove
Bak P (1997) How nature works: the science of self organized criticality. Oxford University Press,
Oxford
Bayart J-F, Ellis S, Hibou B (1999) The criminalization of the state in Africa. Indiana University
Press, Bloomington/Indianapolis
Bhaskar R (1979) The possibility of naturalism: a philosophical critique of the contemporary
human sciences. The Harvester Press, Sussex
Boero R, Squazzoni F (2005) Does empirical embeddedness matter? Methodological issues on
agent-based models for analytical social science. J Artif Soc Soc Simul 8(4), http://jasss.soc.
Boudon R (1998) Social mechanisms without black boxes. In: Hedström P, Swedberg R (eds)
Social mechanism: an analytical approach to social theory. Cambridge University Press,
Caldas JC, Coelho H (1999) The origin of institutions: socio-economic processes, choice, norms
and conventions. J Artif Soc Soc Simul 2(2), http://jasss.soc.surrey.ac.uk/2/2/1.html
Canfield RL (1973) Faction and conversion in a plural society: religious alignments in the Hindu
Kush. University of Michigan, Ann Arbor
Canfield RL (1988) Afghanistan’s social identities in crisis. In: Digard J-P (ed) Le fait ethnique en
Iran et en Afghanistan. Editions du Centre National de la Recherche Scientifique, Paris,
pp 185–199
Castelfranchi C (1990) Social power: a point missed in multi-agent, DAI and HCI. In: Proceedings
of the first European workshop on modelling autonomous agents in a multi-agent world,
Elsevier, Cambridge, pp 49–62
Casti JL (1997) Would-be worlds: how simulation is changing the frontiers of science. Wiley,
New York
Cederman L-E (1997) Emergent actors in world politics: how states and nations develop and
dissolve. Princeton University Press, Princeton
Cederman L-E (2001) Agent-based modeling in political science. Pol Methodol 10(1):16–22
Cederman L-E (2003) Modeling the size of wars: from billiard balls to sandpiles. Am Pol Sci Rev
97(1):135–150
Cohen PR (1985) Heuristic reasoning about uncertainty: an artificial intelligence approach. Pitman
Advanced Publishing Program, Boston
Cruickshank J (2003) Introduction. In: Cruickshank J (ed) Critical realism: the difference that it
makes. Routledge, London/New York, pp 1–14
Dorronsoro G (2005) Revolution unending. Hurst, London
Eckstein H (1992) Case study and theory in political science. In: Gomm R, Hammersley M, Foster P
(eds) Case study method: key issues, key texts. Sage, London, pp 119–164
Fuchs C (2005) Machtverhältnisse in Afghanistan: Netzwerkanalyse des Beziehungssystems
regionaler Führer. M.A. Thesis, University of Zurich, Zurich
Geller A (2006a) The emergence of individual welfare in Afghanistan. Paper presented at the 20th
international political science association world congress, Fukuoka, 9–13 July 2006
Geller A (2006b) Macht, Ressourcen und Gewalt: Zur Komplexität zeitgenössischer Konflikte;
Eine agenten-basierte Modellierung. vdf, Zurich
Geller A (2010) The political economy of normlessness in Afghanistan. In: Schlenkhoff A,
Oeppen C (eds) Beyond the “wild tribes” – understanding modern Afghanistan and its
Diaspora. Columbia University Press, New York, pp 57–70
Geller A, Moss S (2007) The Afghan nexus: anomie, neo-Patrimonialism and the emergence of
small-world networks. In: Proceedings of UK social network conference 2007, University of
London, 13–14 July 2007, pp 86–88
Geller A, Moss S (2008a) Growing Qawm: an evidence-driven declarative model of Afghan power
structures. Adv Complex Syst 11(2):321–335
Geller A, Moss S (2008b) International contemporary conflict as small-world phenomenon and the
hermeneutic net. Paper presented at the annual international studies association conference,
San Francisco, 26–29 Mar 2008
Gerring J (2004) What is a case study and what is it good for? Am Pol Sci Rev 98(2):341–354
Giddens A (1976) New rules of sociological method: a positive critique of interpretative
sociologies. Hutchinson, London
Giustozzi A (2006) Genesis of a “Prince”: the rise of Ismael Khan in Western Afghanistan,
1979–1992, vol 4, Crisis states working papers series. Crisis States Research Centre, London
School of Economics, London
Giustozzi A (2007) War and peace economies of Afghanistan’s strongmen. Int Peacekeeping
14(1):75–89
Glatzer B (1998) Is Afghanistan on the brink of ethnic and tribal disintegration? In: Maley W (ed)
Fundamentalism reborn? Afghanistan and the Taliban. New York University Press, New York,
pp 167–181
Glatzer B (2003) Afghanistan (Studien zur Länderbezogenen Konfliktanalyse). Friedrich-Ebert
Stiftung/Gesellschaft für Technische Zusammenarbeit
Gomm R, Hammersley M, Foster P (1992) Case study and theory. In: Gomm R, Hammersley M,
Foster P (eds) Case study method: key issues, key texts. Sage, London, pp 234–258
Guyot P, Drogoul A, Honiden S (2006a) Power and negotiation: lessons from agent-based
participatory simulations. In: Stone P, Weiss G (eds) Proceedings of the fifth international
joint conference on autonomous agents and multiagent systems (AAMAS). ACM, New York,
pp 27–33
Guyot P, Drogoul A, Honiden S (2006b) Power and negotiation: lessons from agent-based
participatory simulations. In: Stone P, Weiss G (eds) Proceedings of the fifth international
joint conference on autonomous agents and multiagent systems (AAMAS). ACM Press,
Hirshleifer J (1991) The paradox of power. Econ Polit 3(3):177–200
Hirshleifer J (1995) Anarchy and its breakdown. J Polit Econ 103(1):26–52
Janata A, Hassas R (1975) Ghairatman – Der gute Pashtune: Exkurs über die Grundlagen des
Pashtunwali. Afghanistan J 2(2):83–97
Kuznar LA, Frederick W (2007) Simulating the effect of nepotism on political risk taking and
social unrest. Comp Math Org Theory 13(1):29–37
Lasswell HD (1936) Who gets what, when, how. Meridian Books, New York
Lave CA, March JG (1975) An introduction to models in the social sciences. Harper & Row,
New York
Lazer D (2001) Review of Cederman LE (1997) Emergent actors in world politics: how states and
nations develop. J Artif Soc Soc Simul 4(2), http://jasss.soc.surrey.ac.uk/4/2/reviews/lazer.
html
Limpert E, Stahel WA, Abbt M (2001) Log-normal distributions across the sciences: keys and
clues. Bioscience 51(5):341–352
Lustick IS (2000) Agent-based modeling of collective identity: testing constructivist theory. J Artif
Mailliard M, Sibertin-Blanc C (2010) What is power? Perspectives from sociology, multi-agent
systems and social network analysis. Paper presented at the second symposium on social
networks and multiagent systems (SNAMAS 2010), De Montfort University, Leicester,
29 Mar 2010 ftp://ftp.irit.fr/IRIT/SMAC/DOCUMENTS/PUBLIS/SocLab/SNAMAS-10.pdf
Marks SR (1974) Durkheim’s theory of anomie. Am J Sociol 80(2):329–363
Medard J-F (1990) L’etat patrimonialise. Politique Africaine 39:25–36
Merton RK (1938) Social structure and anomie. Am Sociol Rev 3(5):672–682
Monsutti A (2004) Cooperation, remittances, and kinship among the Hazaras. Iran Stud 37
(2):219–240
Mosler H-J (2006) Better be convincing or better be stylish? A theory based multi-agent simula-
tion to explain minority influence in groups via arguments or via peripheral cues. J Artif Soc
Soc Simul 9(3), http://jasss.soc.surrey.ac.uk/9/3/4.html
Moss S (1981) An economic theory of business strategy: an essay in dynamics without equilib-
rium. Martin Robertson, Oxford
Moss S (1995) Control metaphors in the modelling of decision-making behaviour. Comp Econ
8(3):283–301
Moss S (1998) Critical incident management: an empirically derived computational model. J Artif
Moss S (2000) Canonical tasks, environments and models for social simulation. Comp Math
Organ Theory 6(3):249–275
Moss S (2001) Game theory: limitations and an alternative. J Artif Soc Soc Simul 4(2), http://jasss.
soc.surrey.ac.uk/4/2/2.html
Moss S (2007) Alternative approaches to the empirical validation of agent-based models. Techni-
cal report CPM-07-178, Centre for Policy Modelling, Manchester Metropolitan University,
Manchester
Moss S, Edmonds B (1997) A knowledge-based model of context-dependent attribute preferences
for fast moving consumer goods. Omega Int J Manage Sci 25(2):155–169
Moss S, Edmonds B (2005) Sociology and simulation: statistical and qualitative cross-validation.
Am J Sociol 110(4):1095–1131
Moss S, Kuznetsova O (1996) Modelling the process of market emergence. In: Owsinski JW,
Nahorski Z (eds) Modelling and analysing economies in transition. MODEST, Warsaw
Moss S, Gaylard H, Wallis S, Edmonds B (1996) SDML: a multi-agent language for organiza-
tional modelling. Comp Math Org Theory 4(1):43–69
Mousavi SA (1997) The Hazaras of Afghanistan: an historical, cultural, economic and political
study. St. Martin’s Press, New York
Neumann FL (1950) Approaches to the study of power. Polit Sci Quart 65(2):161–180
Outhwaite W (1987) New philosophies of social science: realism, hermeneutics and critical
theory. Macmillan Education, London
Parsons T (1952) The social system. Tavistock, London
Popitz H (1992) Phänomene der Macht, 2nd edn. Mohr Siebeck, Tübingen
Putnam RD (1988) Diplomacy and domestic politics: the logic of two-level games. Int Organ
42(3):427–460
Rasuly-Paleczek G (1998) Ethnic identity versus nationalism: the Uzbeks of North-Eastern
Afghanistan and the Afghan state. In: Atabaki T, O’Kane J (eds) Post-Soviet Central Asia.
Tauris Academic Studies, London/New York, pp 204–230
Reno W (1998) Warlord politics and African states. Lynne Rienner, Boulder/London
Richardson LF (1948) Variation of the frequency of fatal quarrels with magnitude. J Am Stat
Assoc 43(244):523–546
Rouchier J, Thoyer S (2006) Votes and lobbying in the European decision-making process:
application to the European regulation on GMO release. J Artif Soc Soc Simul 9(3), http://
Rouchier J, O’Connor M, Bousquet F (2001) The creation of a reputation in an artificial society
organised by a gift system. J Artif Soc Soc Simul 4(2), http://jasss.soc.surrey.ac.uk/4/2/8.html
Roy O (1992) Ethnic identity and political expression in Northern Afghanistan. In: Gross J-A (ed)
Muslims in Central Asia: expressions of identity and change. Duke University Press, Durham/
London, pp 73–86
Roy O (1994) The new political elite of Afghanistan. In: Weiner M, Banuazizi A (eds) The politics
of social transformation in Afghanistan, Iran and Pakistan. Syracuse University Press,
Syracuse, pp 72–100
Roy O (1995) Afghanistan: from holy war to civil war. The Darwin Press, Princeton
Roy O (1998) Has Islamism a future in Afghanistan? In: Maley W (ed) Fundamentalism reborn?
Afghanistan and the Taliban. New York University Press, New York, pp 199–211
Rubin BR (1992) Political elites in Afghanistan: Rentier State Building, Rentier State Wrecking.
Int J Middle East Stud 24(1):77–99
Rubin BR (2007) Saving Afghanistan. Foreign Aff 86(1):57–78
Saam NJ, Harrer A (1999) Simulating norms, social inequality, and functional change in artificial
societies. J Artif Soc Soc Simul 2(1), http://jasss.soc.surrey.ac.uk/2/1/2.html
Sambanis N (2004) Using case studies to expand economic models of civil war. Perspect Polit
2(2):259–279
Sawyer RK (2005) Social emergence, societies as complex systems. Cambridge University Press,
Cambridge
Sayer A (1992) Method in social science: a realist approach, 2nd edn. Routledge, London/
New York
Sayer A (2000) Realism and social science. Sage, London/Thousand Oaks/New Delhi
Schelling TC (1998) Social mechanisms and social dynamics. In: Hedström P, Swedberg R (eds)
Social mechanism: an analytical approach to social theory. Cambridge University Press,
Schetter C, Glassner R, Karokhail M (2007) Beyond warlordism: the local security architecture in
Afghanistan. Internationale Politik und Gesellschaft 2:136–152
Shahrani MN (1998) The future of state and the structure of community governance in
Afghanistan. In: Maley W (ed) Fundamentalism reborn? Afghanistan and the Taliban.
New York University Press, New York, pp 212–242
Shahrani MN (2002) War, factionalism, and the state in Afghanistan. Am Anthropol 104
(3):715–722
Shapiro I (2005) The flight from reality in the human sciences. Princeton University Press,
Princeton
Silverman D (ed) (2004) Qualitative research: theory, method and practice, 2nd edn. Sage, London
Snidal D (1985) The game theory of international politics. World Pol 38(1):25–57
Sofsky W (2002) Zeiten des Schreckens: Amok, Terror, Krieg. S. Fischer, Frankfurt a.M
Stachowiak M (1973) Allgemeine Modelltheorie. Springer, Wien/New York
Stakes RE (1978) The case study method in social inquiry. Edu Res 7(2):5–8
Tapper N (1991) Bartered brides: politics, gender and marriage in an Afghan tribal society.
Cambridge University Press, Cambridge
Tapper R (2008) Who are the Kuchi? Nomad self-identities in Afghanistan. J R Anthropol Inst
14:97–116
Tarzi SM (1993) Afghanistan in 1992: a Hobbesian state of nature. Asian Surv 33(2):165–174
UNODC (2006) Afghanistan’s drug industry: structure, functioning, dynamics, and implications
for counter-narcotics policy. United Nations Office on Drugs and Crime/The World Bank
Weber M (1980) Wirtschaft und Gesellschaft: Grundriss der verstehenden Soziologie, 5th edn.
J.C.B. Mohr (Paul Siebeck), Tübingen
Weidmann N, Cederman L-E (2005) Geocontest: modeling strategic competition in geopolitical
systems. In: Troitzsch K (ed) Proceedings of the European social simulation association annual
conference (ESSA 2005). Koblenz, Germany, pp 179–185
Wily LA (2004) Looking for peace on the Pastures: rural land relations in Afghanistan.
Afghanistan Research and Evaluation Unit, Kabul
Younger S (2005) Violence and revenge in Egalitarian societies. J Artif Soc Soc Simul 8(4), http://
Chapter 26
Human Societies: Understanding Observed
Social Phenomena
Bruce Edmonds, Pablo Lucas, Juliette Rouchier, and Richard Taylor
Why Read This Chapter? To get an overview of the different ways in which
simulation can be used to gain understanding of human societies and to gain insight
into some of the principle issues that impinge upon such simulation, including the
difficulties these cause. The chapter will go through the various specific goals one
might have in doing such simulation, giving examples of each. It will provide a
critical view as to the success at reaching these various goals and hence inform
about the current state of such simulation projects.
Abstract The chapter begins by briefly describing two contrasting simulations: the
iconic system dynamics model publicised under the “Limits to Growth” book and a
detailed model of 1st millennium Native American societies in the south west of the
US. These are used to bring out the issues of: abstraction, replicability, model
comprehensibility, understanding vs. prediction, and the extent to which
simulations go beyond what is observed. These issues and difficulties result in
three “dimensions” in which simulation approaches differ. These issues are each
rooted in some fundamental difficulties in the project of simulating observed
societies that are then briefly discussed. The core of the chapter is a look at 15
different possible simulation goals, both abstract and concrete, giving some
examples of each and discussing them. The different inputs and results from such
B. Edmonds (*)
Centre for Policy Modelling, Manchester Metropolitan University
P. Lucas
Geary Institute, University College Dublin
J. Rouchier
School of Economics, Aix-Marseille University
R. Taylor
Oxford Branch, Stockholm Environment Institute

710 B. Edmonds et al.
simulations are briefly discussed as to their importance for simulating human

societies. The chapter ends with a brief critical assessment of the state of the art
in terms of using simulation techniques for helping to understand human societies.
26.1 Introduction
Understanding social phenomena is hard. There is all the complexity found in other
fields of enquiry but with additional difficulties due to our embedding as part of
what we are studying.1 Despite these, understanding our own nature is naturally
important to us, and our social aspects are a large part of that nature. Indeed some
would go as far as saying that our social abilities are the defining features of our
species (e.g. Dunbar 1998). The project of understanding human societies is so
intricate that we need to deploy any and all means at our disposal. Simulation is but
one tool in this vast project, but it has the potential to play an important part.
This chapter considers how and to what extent computer simulation helps us to
understand the social complexity we see all around us. It will start by discussing
two simulations in order to raise the key issues that this project involves, before
moving on to highlight the difficulties of understanding human society in more
detail. The core of the chapter is a review of some of the different ways that
simulation can be used for, with examples of each. It then briefly discusses a
conceptual framework, which will be used to organise the different ways that
simulations are used. It ends with a frank assessment of the actual progress in
understanding human societies using simulation techniques.
26.1.1 Example 1: The Club of Rome’s “Limits to Growth” (LTG)
In the early 1970s, on behalf of an international group under the name “The Club of
Rome” a simulation study was published (Meadows et al. 1972) with the attempt to
convince humankind that there were some serious issues facing it, in terms of a
coming population, resource and pollution catastrophe. To do this they developed a
system-dynamics model of the world. They chose a system-dynamics model
because they felt they needed to capture some of the feedback cycles between the
key factors – factors that would not come out in simple statistical projections of the
available data. They developed this model and ran it, publishing the findings – a
number of model-generated future scenarios – for a variety of settings and
variations. The book (“Limits to Growth”) considered the world as a single system,
and postulated some relationships between macro variables, such as population,
available resources, pollution etc. Based on these relations it simulated what might
happen if the feedback between the various variables were allowed to occur.
1
This embedding has advantages as well, such as prior knowledge.
26 Human Societies: Understanding Observed Social Phenomena 711
The results of the simulations were the curves that resulted from this model as the
simulation was continued for future dates. The results indicated that there was a
coming critical point in time and that a lot of suffering would result, even if
mankind managed to survive it.
The book had a considerable impact, firmly placing the possibility that human-
kind could not simply continue to grow indefinitely. It also attracted considerable
criticism (e.g. Cole et al. 1973) mainly based on the plausibility of the model’s
assumptions and the sensitivity of its results to those relationships. (For example it
assumed that growth will be exponential and that delay loops are extended). The
book presented the results of the simulations as predictions – a series of what-if
scenarios. Whilst they did add caveats and explore various possible versions of their
model, depending on what connections there turned out to be in the world-system,
the overall message of the book was unmistakeable: that if we did not change what
we were doing, by limiting our own economic consumption and population, disaster
would result. This was a work firmly in the tradition of Malthus (1798) who, 175
years earlier, had predicted a constant state of near-starvation for much of the world
based upon a consideration of the growth processes of population and agriculture.
The authors clearly hoped that by using a simulation (albeit a simplistic one by
present standards) they would be able to make the potential feedback loops real to
people. Thus this was a use of simulation to illustrate an understanding that the
authors of LTG had. However, the model was not presented as such, but as
something more scientific in some sense.2 A science-driven study that predicted
such suffering was a definite challenge to those who thought the problem was less
severe.3 By publishing their model and making it easy for others to replicate and
analyse it, they offered critics a good opportunity for counter-argumentation.
The model was criticised on many different grounds, but the most effective was
that the model was sensitive to the initial settings of some parameters (Vermeulen and
de Jongh 1976). This raised the question of whether the model had to be finely tuned
in order to get the behaviour claimed and thus, since the parameters were highly
abstract and did not directly correspond to anything measurable, the applicability of
the model to the world we live in was questioned. Its critics assumed that since this
model did not, hence, produce reliable predictions that it could be safely ignored. It
also engendered the general perception that predictive simulation models are not
credible tools for understanding human socio-economic changes – especially for long
term analyses – and discouraged their use in supporting policy-making for a while.
2
The intentions of the authors themselves in terms of what they thought of the simulation itself are
difficult to ascertain and varied between the individuals, however this was certainly how the work
was perceived.
3
Or those whose vested interests may have led them to maintain the status quo concerning the
desirability of continual economic growth.
26.1.2 Example 2: Modelling 1st Millennium Native American

Society
A contrasting example to the Club of Rome model is the use of simulation models to
assess and explore explanations of population shifts among the Native American
nations in the pre-Columbian era. This has been called “Generative Archaeology”
(GA) by Tim Kohler (2009). Here a spatial model of a population was developed where
movement was fitted to a wealth of archaeological and climatologically data, in order to
find and assess possible explanations of the size, distribution and change in populations
that existed in the first millennium AD in the Southwest US. This case offers a picture
of settlement patterns in the context of relatively high-resolution reconstructions of
changes in climate and resources relevant to human use of these landscapes.
The available data in this case is relatively rich, allowing many questions to be
answered directly. However, many interesting aspects are not directly answerable
from a static analysis of the data, for example those about the possible social
processes that existed. The problem is that different archaeologists can inspect
the same settlement pattern and generate different candidate processes
(explanations) for its generation. Here agent-based modelling helps infer the social
processes (which cannot be directly observed) from the detailed record over time.
This is not a direct or certain inference, since there are still many unknowns
involved in that process.
In Kohler et al. (2005, 2008)4 the use of agent-based modelling (ABM) has been
mainly to see what patterns one should expect if households were approximately
minimizing their caloric costs for access to adequate amounts of calories, protein, and
water. The differences through time in how well this expectation fits the observed
record and the changing directions of departure from those expectations provide a
completely novel source of inference on the archaeological record. Simulations using
the hypothesis of local food sharing during periods of mild food shortage may be
compared to the fit in a simulation where food sharing does not occur. In this way we
can get indirect evidence as to whether food sharing took place.
The ABM has hence allowed a comparison of a possible process with the recorded
evidence. This comparison is relative to the assumptions that are built into the model,
which tend to be plausible but questionable. However despite the uncertainties
involved, one is able to make a useful assessment of the possible explanations and
the assumptions are explicitly documented. This approach to processes that involve
complex interaction would be impossible to do without a computer simulation. At the
very least, such a process reveals new important questions to ask (and hence new
evidence to search for) and the times when the plausible explanations are demonstra-
bly inadequate. However for any real progress in explanation of such cases, a very
large amount of data seems to have been required.
4
For details of the wider project connected with these papers, see the Village Ecodynamics
Project, http://village.anth.wsu.edu.
26.1.3 Some Issues That the Aforementioned Examples Illustrate
The previous examples raise a few issues, common to much social simulation
modelling of human societies. These will now be briefly defined and discussed as
an introduction to the problem of understanding social phenomena using simulation.
1. Abstraction. Abstraction5 is a crucial step in modelling observed social phenom-
ena, as it involves choices about what aspects are or are not salient in relation to
the problem and what level of analysis is appropriate. The LTG example, being a
macro-model, assumes that distributive aspects such as geography and local
heterogeneity are less important with respect to feedbacks among global growth
variables. In this model all the detail of a whole world is reduced to the
interaction of a few numeric variables. The GA model was more specific and
detailed, including an explicit 2D map of the area and the position of settlements
at different times in the past. It is fair to say that the LTG model was driven by
the goals of its modellers, i.e. showing that the coming crisis could be sharp due
to slow feedback loops, the GA model is more driven by the data that was
available, with the model being used to try and answer a number of different
questions afterwards.
2. Replicability. Replicability is the extent to which a published model has been
described in a sufficiently comprehensive and transparent way so that the
simulation experiments can be independently reproduced by another modeller.
Replicability may be considerably easier if care is taken to verify the initial
model by means of debugging tests, and also if the original source code is
effectively available and is well commented. Here the LTG model was readily
replicable and hence open to inspection and criticism. The GA models are
available to download and inspect, but its very complexity makes it hard to
independently replicate (although it has been done).
3. Understanding the model. A modeller’s inferential ability is the extent to which
one can understand one’s own model. Evidence suggests that humans can fully
track systems only for about two or three variables and five or six states (Klein
1998); for higher levels of complexity, additional tools are required. In many
simulations, especially those towards the descriptive end of the spectrum, the
agents have many behavioural rules which may interact in complicated and
unpredictable ways. This makes simulations very difficult to fully understand
and check. Even in the case of a simple model, such as the LTG model, there can
be unexpected features (such as the fine tuning that was required). Although the
GA was rich in its depiction of the space and environment the behavioural rules
of sub-populations was fairly simple and easy to follow at the micro-level.
However this does not rule out subtle errors and complexities that might result
from the interaction of the micro-elements. Indeed this is the point of such a
5
Abstraction can be defined as “ignoring or hiding details to capture some kind of commonality
between different instances” (Free On-line Dictionary of Computing).
simulation that we can’t work out these complex outcomes ourselves, but require
a computer program to do it.
4. Prediction vs. Understanding. The main lesson to be drawn from the history of
formal modelling is that, for most complex systems it is impossible to model with
accuracy their evolution beyond an immediate timeframe. Whilst the broad trends
and properties may be to some degree forecast the particulars, e.g. the timing and
scale of changes in the aggregate variables, generally they cannot (Moss 1999).
The LTG model attempted to forecast the future, not in terms of the precise levels
but in the presence of a severe crisis – a peak in population followed by a crash.
The GA does not aim to strongly predict anything, but rather it seeks to establish
plausible explanations for the data that is known. Most simulations of human
society restrict themselves to establishing explanations, the simulations providing
a chain of causation that shows that the explanation is possible.6
5. Going beyond what is known. In social science there are a multitude of gaps in
our knowledge and social simulation methods may be well placed to address some
of these gaps. Given some data, and some plausible assumptions, the simulations
can be used to perform experiments that are consistent with the data and
assumptions, and then inspected to answer other questions. Clearly this depends
on the reliability of the assumptions chosen. In the GA case this is very clear, a
model with a food-sharing rule and one without can be compared to the data,
seeing which one fits it better. The LTG model attempts something harder, making
severe assumptions about how the aggregate variables relate, it “predicts” aspects
of the future. In general: the more reliable the assumptions and data (hence: the
less ambitious the attempt at projection), the more credible the result.
A social scientist who wants to capture key aspects of observed social phenom-
ena in a simulation model faces many difficulties. Indeed, the differences between
formal systems and complex, multi-facetted and meaning-laden social systems are
so fundamental that some criticise any attempt to bridge this gap (e.g. Clifford
1986). Simulators have to face these difficulties which have an impact as to how
social simulation is done and how useful (or otherwise) such models may be. We
briefly consider six of these difficulties here.
• Firstly, there is the sheer difference in nature between formal models (i.e. com-
puter programs) that modellers use as compared to the social world that we
observe. The former are explicit, precise, with a formal grammar, predictable at
the micro-level, reproducible and work in (mostly) the same way regardless of the
computational context. The later is vague, fluid, uncertain, subjective, implicit and
imprecise – which often seems to work completely differently in similar
situations, and whose operation seems to rely on the rich interaction of meaning
in a way that is sometimes explicable but usually unpredictable. In particular the
6
Although in many cases this is dressed up to look like prediction, such as the fitting to out-of-
sample data. Prediction has to be for data unknown to the modeller, otherwise the model will be
implicitly fitted to it.
gap between essentially formal symbols with precise but limited meaning and the
rich semantic associations of the observed social world (for example as expressed
in natural language) is particularly stark. This gap is so wide that some
philosophers have declared it unbridgeable (e.g. Lincoln and Guba 1985, Guba
and Lincoln 1994).
• Secondly there is the sheer variability, complication and complexity of the social
world. Social phenomena seem to be at least as complex as biological phenom-
ena but without the central organising principle of evolution as specified in the
neo-Darwinian Synthesis. If there are any general organising principles (and it is
not obvious that this is the case) then there are many of these, each with differing
(and sometimes overlapping) domains of application. In that sense, it is clear that
a model will always capture only a small part of the phenomenon among many
other related aspects, hence reducing drastically the possibility to predict with
any degree of certainty.
• Then there is the sheer lack of adequate multifaceted data about social phenom-
ena. Social simulators always seem to have to choose between longitudinal
studies, narrative data, cross-sectional surveys or time-series data. Having all
of these datasets about a single social process or event is to date very unlikely.
There does not seem to be the emphasis on data collection and measurement in
the social sciences that there is in some other sciences and certainly not the
corresponding prestige for those who collect it or invent ways of doing so.
• There is the more mundane difficulty of building, checking, maintaining, and
analysing simulations (Galán et al. 2009). Even the simplest simulations are
beyond our complete understanding, indeed that is often why we need them,
because there is no other practical way to find out the complex ramifications of a
set of interacting agents. This presence of emergent outcomes in the simulations
makes them very difficult to check. Ways to improve confidence that our
simulations in fact correspond to our intentions for them7 include: unit testing,
debugging, and the facility for querying the database of a simulation (see Chap.
8 (David 2013) in this handbook). Perhaps the strictest test is the independent
replication of simulations – working from the specifications and checking their
results at a high degree of accuracy (Axtell et al. 1996). However such replica-
tion is usually very difficult and time-consuming, even in relatively simple cases
(Edmonds and Hales 2003).
• Another difficulty is that of the inevitability of background assumptions in all we
do. There are always a wealth of facts, processes and affordances to give
meaning to, and provide the framework for, the foreground actions and causal
chains that we observe. Many of these are not immediately apparent to us since
they are part of the contexts we inhabit and so are not perceptually apparent. This
is the same as other fields, as it has been argued elsewhere; the concept of
causation only makes sense within a context (Edmonds 2007). However it does
7
In terms of design and implementation, if one has a good reference case in terms of observed data
then you can also check one’s simulation against this.
seem that context is more critical in the social world than others, since it can not
only change the outcomes of events but their very meaning (and hence the kind
of social outcome). Whilst in other fields it might be acceptable to represent
extra-contextual interferences as some kind of random distribution or process,
this is often manifestly inadequate with social phenomena (Edmonds and
Hales 2005).
• The uncertainty behind the foreground assumptions in social simulation is also
problematic. Even when we are aware of all of the assumptions they are often
either too numerous to include in a single model or else we simply lack any
evidence as to what they should be. For example, there are many social simula-
tion models which include some version of inference, learning or decision-
making within the agents of the model, even when there is no evidence as to
whether this actually corresponds to that used by the observed actors. It seems
that often it is simply hoped that these details will not happen to matter much in
the end – thus becoming a rarely checked, and sometimes wrong, aspect of
simulations (Edmonds 2001; Rouchier 2001).
• Finally there is a difficulty from the nature of simulation itself. Simulation will
demonstrate possible processes that might follow from a given situation (relative
to the assumptions on which the simulation is built). It does not show all the
possibilities, since it could happen that a future simulation will produce the same
outcomes from the same set-up in a different way (e.g. using a different
cognitive model). Thus simulation differs in terms of its inferential power
from analytic models (e.g. equation-based ones), where the simplicity of the
model can allow formal proofs of a general formulation of outcomes that may
establish the necessity of conditions as well as their adequacy. This difficulty is
the same that plagues many mathematical formulations since, in their raw form,
they are often unsolvable and hence either one has to use numerical simulation
of results (in which case one is back to a simulation) or one has to make
simplifying assumptions (in which case, depending on the strength of these
assumptions, one does not know if the results still apply to the original case).
These difficulties bring up the question of whether some aspects of societies can
be at all understood by means of modelling. This hypothesis asserting that simula-
tion is a credible method to better explore, understand or explain social processes, is
implicitly tested in the current volume and is discussed in some detail below. We
are not going to take any strong position but will restrict ourselves to considering
examples within the context of their use.8 Agent-based social simulation is not a
magic-bullet and is not yet a mature technique. It is common sense in the social
simulation community that best results will be achieved by combining social
simulation with other research methods.
8
Obviously we suspect it can be a useful tool otherwise we would not be bothering with it.
26.2 Styles of Modelling and Their Impact on Simulation Issues
26.2.1 Models of Evidence vs. Models of Ideas
One response to the above difficulties is not to model social phenomena directly,
but rather to restrict ourselves to modelling ideas about social phenomena. This is a
lot easier, since our ideas are necessarily a lot simpler and more abstract than the
phenomena itself (which can be formalized with the notion of pattern modelling
(Grimm et al. 2005) rather than strict adequacy to data). Some ideas need
modelling, in the sense that the ramifications of the ideas are themselves complex.
These kinds of models can be used to improve our understanding of the ideas, and
later this understanding can be applied in a rich, flexible and context-sensitive way.
This distinction is made clear in (Edmonds 2001).
Of course to some extent any model is a compact abstraction of the final target of
modelling – there will, presumably, be some reason why one conceptualises what
one is modelling in terms of evidence or experience by someone, and there will
always be some level of theory/assumption that motivates the decision as to what
can be safely left out of a model. Thus all models are somewhat about ideas and,
hopefully, all models have some relation to the evidence. However there is still a
clear difference between those models that take their primary structure from an
idea, and those whose primary considerations come from the available evidence.
For example, the former tend to be a lot simpler than the later. The later will tend to
have specific motivations for each feature whilst the former will tend to be
motivated in general terms. These two kinds of simulation have a close parallel
with the theoretical and phenomenological models identified by Cartwright (1993).
Unfortunately these kinds of model are often conflated in academic papers. This
seems frequently not deliberate, but rather due to the strong theoretical spectacles
(Kuhn 1962) that simulation models seem to provide. There is nothing like devel-
oping and playing with a simulation model for a while to make one see the world in
terms of that model – it is not only that the model is your creation and best effort in
formulating an aspect of the social world, but one has interacted with it and changed
it to include the features that you, the modeller, think it should have. Nevertheless,
whatever the source it can take some careful “reading between the lines” to
determine the exact nature of any model, and what it purports to represent.
26.2.2 Modelling as Representation of Social Phenomena vs.

as an Intervention in a Social Process
It must be said that some simulation models are not intended to represent anything
but rather created for another purpose, such as a tool for demonstrating an approach
or an intervention in a decision-making process. This may be deliberate and
explicit, or not, for various different reasons. Of course if a computer model does
not represent anything at all, it is not really a simulation but simply a computer
program, which may be presented in the style of a simulation. Also for a simulation
to be an effective tool for intervention it will have to have some credibility with the
participants.
However in some research the representation is either not the primary goal or
what they seek to represent is deliberately subjective in character. Thus in some
participatory approaches (see Chap. 11, Barreteau et al. 2013) it may be the primary
goal to raise awareness of an issue, to intervene in or facilitate a social process like a
negotiation or consensus process within a group of people. The modeller may not
focus as much on whether the model captures an objective reality but rather on how
stakeholders9 understood the issues and processes of concern and how this might
influence the outcomes. This does not mean that there will be no elements that are
objective and/or representative in character – for example such models might have
a well-validated hydrological component to them – but that the parts of the model
that are the focus are checked against the opinions of those being modelled or those
with an interest in the outcomes rather than any independent evidence.
Of course, this is a matter of degree – in a sense most social simulations are a
mixture of objective aspects linked to observations and other aspects derived from
theories, opinions, hypotheses and assumptions. In the participatory approaches the
modeller seeks not to put their own ideas forward but rather take the, possibly more
democratic, approach of being an expert facilitator expressing the stakeholders’
opinions and knowledge. Whilst some researchers might reject such ideas as too
“anecdotal” to be included in a formal model, it is not obvious that the stakeholders’
ideas about the nature of the processes involved (for example, how the key players
make decisions) are less reliable than the grander theories of academics. However
researchers do have a professional obligation to be transparent and honest about
their opinions, documenting assumptions to make them explicit and, at least, not
state things that they think are false. Thus, although participatory approaches are
not a world away from more traditional models of using simulation, they do have
some different biases and characteristics.
26.2.3 Context and Social Simulation
Human knowledge, but particularly human social knowledge is usually not context-
free. That is there is a set of background assumptions, facts, relationships and
meanings that may be necessary and generally known but not made explicit.
These background features can all be associated with the context of the knowledge
(Edmonds 1999). In a similar way, most social simulation happens within a
particular context as given, thus for example the environment in which racial
9
I.e. those who are part of, or can influence the social phenomenon in question.
segregation occurs might be obvious to all concerned. This context is sometimes

indicated in papers but is often left implicit. The context of a simulation is
associated with the uncountably many background assumptions that can be ignored,
either because they are irrelevant or fixed in that context. Social simulation would
probably be impossible if one was not able to assume a context whose associated
assumptions need not be questioned for a given model (Edmonds 2010). Without
such an effective restriction of scope every social simulation model would have to
include all potential aspects of human behaviour and social interaction. Whilst such
assumptions concerning the context are common to almost all fields of knowledge it
is particularly powerful in the social sciences due to the fact that we unavoidably
use our folk-knowledge10 of social situations to make sense of the studied social
phenomena. This process of (social) context identification is often automatic, so
that we correctly identify the appropriate context without expending much con-
scious thought. For this reason the context is often left implicit, despite the fact that
it can impact upon the use of simulation in understanding the phenomena. This
leaves the decisions as what to implement as foreground, deliberate decisions.
Choosing a social context which is relatively identifiable and self-contained is an
important factor if one is seeking to represent some evidence in a simulation. Being
able to include all the important factors of some social process and obtain some
evidence for their nature, allows the building of simulations that are not misleading
in the sense of not missing out factors that might critically change the outcomes.
Clearly the more restricted the context, the easier the representational task. How-
ever in this case one does not know whether what one learns from the simulation is
applicable in other contexts. Using a simulation developed for one context and
purpose for a different context and/or purpose might well lead to misleading
conclusions (Edmonds and Hales 2005; Lucas 2010).
Those simulations which are more focused on exploring an idea will often seek
to transcend context, in the hope that the models will have some degree of
generality – these often deliberately ignore any particular context. Although these
may seem general, their weakness can become apparent when its applicability is
tested. Here the ideas they represent might give some useful insights, but may be
misleading if taken as the defining feature of a specific case study. Clearly a
simulation that is claimed to have general applicability needs to have been validated
across the claimed scope before being relied upon by others. To date, no social
simulation has been found to be generally applicable beyond theoretical and
illustrational purposes (Lucas 2011).
10
Folk-knowledge is the set of widely-held beliefs about popular psychological and social
theories, this is sometimes used in a rather derogatory way even when the reliability of the
academic alternatives is unclear.
26.3 A Plethora of Modelling Purposes with Examples
Given the different purposes for which simulation models are used (Epstein 2008),
they will be considered in groups of those with similar goals. It is only relative to
their goals that simulation efforts can be judged. Nowadays it is widely acknowl-
edged that authors should state the purpose of their models clearly before how the
model is constituted (Grimm et al. 2006). Firstly however it is worth reviewing two
goals that are widely pursued in many other fields but have not been convincingly
attained with respect to the simulation of human society.
The first of these goals is that of predicting what will definitely happen in
unknown circumstances. In other words, social simulation cannot yet make accu-
rate and precise predictions. The nearest social simulations come (to our knowl-
edge) is predicting some outcomes in situations where the choices are very
constricted, and the data available is comprehensive. The clearest case of this is
the use of micro-simulation models to predict the final outcome of elections once
about 30 % of the results are already known (Curtis and Frith 2008). Thus this is
hardly new or unknown circumstances, and is not immune from surprises, since
sometimes their predictions are wrong. This model is a micro-simulation model that
relies on the balance between parties in each constituency and then translates the
general switches between parties (and non-voters) to the undeclared results. Thus
although it is a prediction, its very nature rules out counter-intuitive or surprising
predictions and comes more into the category of extending known data rather than
prediction. The gold standard for prediction is that of making predictions of
outcomes that are unexpected but true.11
The second goal simulations do not achieve is to decisively test sociological
hypotheses – in other words, they do not convincingly show that any particular idea
about what we observe occurring in human societies can be relied upon or compre-
hensively ruled out. Here the distinction between modelling what we observe and
modelling our ideas is important. A simulation that attempts to model what we observe
is a contingent hypothesis that may always be wrong. However social simulations of
evidence are always dependent on a raft of supportive assumptions – that the
simulation fails to reproduce the desired outcomes may be due to a failure of any of
its assumptions. Of course, if such a model is repeatedly tested against evidence and
fails to be proved wrong we may come to rely upon it more (Popper 1963) but this
success may be for other reasons (e.g. we simply have not tested it in sufficiently
diverse conditions). Hypothesis testing using simulations is always relative to the
totality of assumptions in the simulations and thus the gain in certainty is, at best,
incremental and relative.12 Thus the core assumptions of a field may be preserved by
adjusting “auxiliary” aspects (Lakatos and Musgrave 1970).
11
This is when prediction is actually useful, for if it only gives expected values one would not need
the simulation.
12
If a simulation is not directly related to evidence but is more a model of some ideas, then it might
be simple enough to be able to test hypotheses but these hypotheses will then be about the abstract
model and not about the target phenomena.
If a simulation is about ideas then a very restricted kind of test is possible: a

counter example. If it has been assumed that factor A will lead to result B, then one
might be able to show that this might not be the case in a plausible simulation.
Indeed the simulation may show that to obtain result B from factor A an additional
and implausible assumption C is necessary. This does prove that “it is not neces-
sarily the case that A leads to B”, but it may shift the burden of proof back onto those
who have assumed A will lead to B. This very restricted test is only useful if the
context of causation between A and B is appropriately identifiable. This case of
using a simulation to establish counter-examples is considered below.
A particular case of seeking for counter-examples is that of testing for the
“existence of a sufficient condition” for some particular results. For example, it
may be possible to show that there is no need to add some particular hypothesis to
see a phenomenon take place, as in economics where it can be shown that in many
cases the assumption of perfect rationality for agents does not need to be made.13
One might be disappointed that simulation provides neither predictions nor
proofs (in the stronger senses of those terms), but that does not stop them being
useful in other ways, which the sections below illustrate.
In the following, we look at how simulations might contribute to the understand-
ing of human societies in a number of different ways, with examples from the
literature. Unfortunately many articles describing social simulation research do not
make their goals explicit (as advocated by ODD, see Polhill et al. 2008 and Chap. 7,
Grimm et al. 2013), therefore the categorisation below is that of the chapter’s
authors and not necessarily the category that the authors of the papers discussed
would choose. Also it appears that some researchers have multiple purposes for
their simulations or simply have not thought about their goals clearly.
26.3.1 Abstract Goals
First we consider simulations that have more abstract goals i.e. these tend to be
more about ideas and theories than observed evidence (as discussed above in
Sect. 26.2.1).
26.3.1.1 Illustration of Ideas
Simulations can be good ways of making processes clear, because simulations are
what simulations do. Thus, if one can unpack a simulation to show how the
outcomes result from a setup or mechanism, then this can demonstrate an idea
clearly and dramatically. Of course, if how the outcomes emerge from the setup in a
13
This fact has lead some to argue that such assumptions of perfect rationality should be dropped
and that it might be better to adopt a more naturalistic representation of human’s cognition (Gode
and Sunder 1993; Kirman 2011).
simulation is opaque and/or difficult to understand then this is not an effective

technique. For this reason this tends to be done using relatively simple simulations
that are specifically designed to bring out the focus idea.
An example is (Rouchier and Thoyer 2006) which models voting and lobbying
in the EU decision making process. It does make fairly strong assumptions about
how the voting strategies might operate, but it does not pretend to be a descriptive
model. Instead it makes clear how the links between public opinion, lobbying
groups and elected representatives might operate at the national scale as well as
the EU one.
Another example is (Gode and Sunder 1993), a fairly simple demonstration that
in some cases market institutions are so constraining that agents do not even need to
be clever to achieve excellent results in this setting. They take the example of
Continuous Double Auction (CDA), a two-sided progressive auction, which is the
protocol that is most used in financial markets. At any moment, buyers can submit
bids (offers to buy). Similarly, sellers can submit asks (offers to sell). Both buyers
and sellers may propose an offer or accept the offer made by others. The main
constraint is an improvement rule, imposed on new offers entering the market,
which requires submitted bids/asks at a price higher/lower than the standing bids/
asks. Each time an offer is satisfied for one of the participants, he or she announces
the acceptance of the trade at the given price, and the transaction is completed.
Once a transaction is completed the agents who have traded leave the market and
the bid-offer process starts again following the same rule starting from any price.
The result of Gode and Sunder’s simulation is that even with completely stupid
agents, who know nothing of the market and only follow two constraints: the bid-
offer rule described above and not selling below or buying above their reservation
price, the market converges and enables agents to get excellent profits. This paper
shows how institutional constraints might act to ensure a reasonable allocation of
goods when agents are very clear about the value of things they want to sell or buy,
and that this does not require any other substantive rationality by the agents. Of
course this result cannot necessarily be extended to any observed markets, which
are most of the time complex, where the agents do have intelligence, where the
value of items might be unclear and where there might be many other social and
institutional mechanisms, but at least this result clarifies an idea about why
protocols of this kind might be important.
The OpenABM project14 has made significant progress in the development of a
community of people using illustrative models to facilitate the communication of
ideas. Working with others this group in particular promotes the educational value
of agent-based models.
A particular case of using a simulation to illustrate an idea is that of using a
simulation in teaching. Whilst demonstrating an idea to one’s peers might lead one
to choose a simulation that emphasises the idea’s generality and power, in teaching
one may well choose to simplify and highlight certain features of the idea that will
14
http://www.openabm.org.
be important later on. This is a matter of degree, but tends to result in simulations of
a slightly different kind.
For example, researchers at Oxford University Department of Computer Science
have developed a web application to assist students (particularly non-programmers)
in understanding the behaviour of systems of interactive agents (Kahn and Noble
2009). They model, for example, the dynamics of epidemics in schools and
workplaces, and effect of vaccination or school closing/quarantine periods upon
spread of disease in the population (Scherer and McLean 2002). The students can
quickly and easily test different policies and other parameter combinations, or for
more intensive sessions can work through a series of guided steps to build models
from pre-existing modular components or ‘micro-behaviours’ – a process called
‘composing’. The models can also be run, saved, and shared through a web browser
in order to facilitate discussion and collaboration as well as ownership of the ideas
and creative thinking.
26.3.1.2 Establishing the Possibility of a Process
A simulation can be used to show how a mechanism might result in certain

outcomes, and thus established that a proposed process is possible, demonstrated
by enfolding micro-processes in the simulation. This established plausibility of the
process is relative to the plausibility of the assumptions behind the simulation –
clearly if the simulation is one that could not convincingly be related to any
observed system then one would not have established that the process is possible
in any encountered system, but only be a theoretical possibility. This does not
require that the simulation is an accurate representation of any observed system
since all that is required is that one could imagine that a version of the target process
in the simulation could occur in a real system.
A classic example of this is Axelrod’s (1984, 1997) work on the evolution of
cooperation. Previous models in evolutionary biology had suggested that coopera-
tive behaviour would not be selected within an evolutionary setting, as any group of
co-operators would be vulnerable to a single non-cooperative invader or mutant.
Axelrod’s books described simulations in which a population of competing
individuals evolved, playing repeated games against others. Some cooperative
strategies, in particular ‘tit-for-tat’ (cooperate unless your partner did not last
time) where shown to survive and flourish in many game set-ups. Although the
simulations described were highly speculative and abstract, they did firmly estab-
lish that it was possible that cooperative strategies might evolve within an evolu-
tionary setting, where selfish strategies had a short-term advantage.
One use for establishing the possibility of a process is as a counter-example to an
existing assumption or accepted theory, if the process demonstrated contradicts the
assumption. Thus the simulations of Axelrod above can be seen as a counter-example
to the assumption that cooperative behaviour cannot survive in an evolutionary setting.
The particular case of the Schelling (1969, 1971) model can be classified in this
trend. Through very simple simulations, which Schelling ran by hand at the time, he
discovered that segregation could be attained at a group level although each
individual agent had no strong preference for segregation. This paper was impor-
tant, because it was one of the first examples of emergent phenomena applied to
social issues. But the most important element was the positive result obtained with
the model. Schelling used a very intuitive (though not necessarily realistic) way of
describing the change of location of agents in a city where they are surrounded by
neighbours, which can be of two different types: identical to themselves or differ-
ent. Each agent decides if it is satisfied with its location by judging if the proportion
of neighbours that are different is acceptable to it. If this is not the case it moves to a
new location. Even when each agent accepts up to 65 % of agents different to itself
in its neighbourhood, high levels of segregation in the global society of agents
result. This is a counter-example to the assumption that segregation results from a
high level of intolerance to those of different ethnic origins, since one can see from
the simulation that the apparition of high levels of segregation in cities could be due
to the movement of people at the edges of segregated areas who are in regions
dominated by those of different ethnicities. Of course, what this does not show is
that this is what actually causes segregation in cities, it merely undermines the
assumption that it must be due to high levels of intolerance.
26.3.1.3 Understanding the Properties of an Abstract Model
With some analytic mathematical models and very few, very simple simulation
models, one might seek to prove some properties of that model, for example the
parameter values under which a given end condition is reached. If this is not
possible (the usual case), then one has two basic options: to simplify the original
to obtain a model that is analytically tractable or to simulate it. If the simplifications
that are necessary to obtain a tractable model are well-understood and plausible,
then the simplified model might be trusted to approximate the original model
(although it is always wise to check). If it is the case that to obtain an analytically
tractable model one has to simplify so much that the relationship between the
simplified and the original model is suspect (for example, by adding implausibly
strong assumptions) then one cannot say that the simplified model is about the same
things as the original model. At best the simplified model might be used as an
analogy for what was being modelled – it cannot be relied upon to give correct
answers about the original target of modelling. In this case, if one wants to actually
model the original target of modelling, then simulation models are the only option.
In this case one might wish to understand the simulation itself by systematically
exploring its properties, such as doing parameter sweeps. In a sense this is a kind of
pseudo-maths, trying to get a grasp of the general model properties when analytic
proof is not feasible.
An example of such an exploration is (Biggs et al. 2009). This examined the
regime-shifts using the fisheries food web model, in particular looking at the
existence of turning points in a system with two attractors (piscivore and
planktivore dominated regimes). Anthropogenic drivers were modelled as gradual
changes in the amount of angling and shoreline development. Simulations were
carried out to investigate the onset of regime shifts in fish populations, the
possibilities to detect these changes and the effectiveness of management responses
to avert the shift. In relation to angling it was found that shifts could be averted by
reducing harvesting to zero at a relatively late stage (and well into the transition to
alternate regime) whereas with development it required action to be taken substan-
tially earlier, i.e. the lag time was substantially longer between taking action and the
resultant shift. The behaviour of different indicators to anticipate regime shifts was
examined. This is an example of a mathematical model with stochastic elements
that is solved numerically by means of a simulation.
Such stylised models, although based on well-understood processes, are
caricatures of real systems and have a number of simplifying assumptions. Never-
theless they may provide an insight that would be applicable to many types of real
world issues. In contrast to this some seek to understand the properties of some
fairly abstract models, aiming to uncover some structures and results that might be
quite generally applicable. This is directly analogous to mathematics that seeks to
establish some general structures, theorems and properties that might later be
usefully applied as part of the extensive menu of tools that mathematics applies.
In this case the usefulness of the exercise depends ultimately on the applicability of
the results in practice. The criteria by which pure mathematics are judged can be
seen as distinguishing those that are likely to be useful in the future: soundness,
generality, and importance.
An example of where the study of an abstract class of mechanisms has been
explored thoroughly to establish the general properties is the area of social influ-
ence, in particular the sub-case of opinion dynamics. It can be found in works that
use physics methodologies (Galam 1997) or more artificial life approaches
(Axelrod 1997). The topic in itself is extremely abstract and cannot be validated
against data in any direct manner. The notion of culture or opinion that is studied in
these models is considerably abstracted and so hard to accept for any sociologist
(von Randow 2003). In this area the most studied mechanism is the creation of
consensus or convergence of culture represented by a single real number or a binary
string (Galam 1997; Deffuant et al. 2000; Axelrod 1997). Many variations and
special cases of these classes of model have been undertaken, for a survey see
(Lorenz 2007). Some of these studies have indeed used a combination of parameter
sweeps of simulations and analytic approximations to give a comprehensive picture
of the model behaviour (Deffuant and Weisbuch 2007). Other merely seem to point
out possible variations of the model.
Sometimes the exploration of abstract properties of models can result in
surprises, showing behaviour that was contrary to expectations, so this category
can overlap with the one discussed in the next section (26.3.1.4).
26.3.1.4 Exploration of the Safety of Assumptions in Existing Models
This is similar to the previous goal, but instead of trying to establish the behaviour
of the model as it is, one might seek to explore what happens if any of the
assumptions in the model are changed or weakened. Thus here one is seeking to
explore a space of possibilities around the original model. The idea behind this is
often that one has a hypothesis about a particular assumption that the model is based
upon. For example one might suspect that one would get very different outcomes if
one varied some mechanism in the model in (what might seem) a trivial manner.
Another example is when one suspects that a certain assumption is unnecessary to
the outcomes and can be safely dropped. Thus for this goal one is essentially
comparing the behaviour of the original model with that of an altered, or extended
model.
For example, Izquierdo and Izquierdo (2006) carried out a systematic analysis of
the effect of making slight modifications to structural assumptions in the Prisoner’s
Dilemma game: in the population size, the mutation rate, the way that pairings were
made, etc. all of which produced large changes in the emergent outcome – the
frequency of strategies employed. The authors conclude that “the type of strategies
that are likely to emerge and be sustained in evolutionary contexts is strongly
dependent on assumptions that traditionally have been thought to be unimportant
or secondary” (Izquierdo and Izquierdo 2006:181).
How cooperation emerges in a social setting was first fashioned into a game-
theoretical problem by Axelrod (1984). The outcome was long thought to be
dependent upon the defining questions such as which strategies are available,
what are the pay-off values for each strategy, number of repetitions in a match,
etc. whereas other structural assumptions, thought to be unimportant, were ignored.
On further investigation, however, conclusions based on early work were shown to
be rather less general than would be desired, and sometimes actually contradicted
by later work.
A different case is explorations of the robustness of the simulation described in
(Riolo et al. 2001). This showed the emergence of a cooperative group in an
evolutionary setting similar to the Axelrod one mentioned above. Here each
individual had a characteristic (modelled as a number between 0 and 1) and a
tolerance in a similar range. Individuals are randomly paired and if the difference
between the partner’s and its own characteristic is less than or equal to their
tolerance then they cooperate, otherwise do not. As a result a group of individuals
with similar characteristics formed that effectively shared with each other. How-
ever later studies (Roberts and Sherratt 2002; Edmonds and Hales 2003) probed the
robustness of the model in a number of ways, but crucially by altering the rule for
cooperation from “cooperate if the difference between the partner’s and its own
characteristic is less than or equal to their tolerance” to “if the difference between
the partner’s and its own characteristic is strictly less than their tolerance”, i.e. from
“” to “<”. When this change is made the crucial result – the emergence of a
cooperative group – disappeared. It turned out that the (Riolo et al. 2001) effect
relied on the existence of a group of individuals with exactly the same characteristic
with whom they had to cooperate, since the smallest tolerance possible was 0.
When the existence of completely selfish individuals was made possible by this
change, the cooperation disappeared.
26.3.1.5 Exploring Counter-Factual Possibilities
We only observe a few of the possible configurations of the social phenomena

around us. Thus it is natural to wonder what might happen if events or processes
were other than as observed or known to be the case. This is the world of artificial
societies, where possible worlds loosely related to the one observed are explored.
Sometimes an analogy with artificial life is made, where alternative algorithmic
versions of life in the broadest sense are specified and experimented with – not life-
as-it-is but life-as-it-might-have-been.
An extreme example of this is Jim Doran’s model of a society with knowledge of
the future (Doran 1997) – this can be thought of as what a society might be like
whose members’ predictions of the future happen to be correct. Clearly this is a case
that does not hold in human society.
Such explorations might not contribute much to the understanding of our
society, but it may inform the design of distributed computational systems where
the components have a need to flexibly organise themselves in a way analogous, but
not identical to, how humans organise (see Chap. 21, Hales 2013).
26.3.2 Concrete Goals
Here we consider some of the goals that are more at the concrete and descriptive
end of the simulation spectrum. These tend to be more concerned to relate to
evidence and also to be much more specific. In the subsections below the “plausi-
bility” of assumptions, results and simulations is a frequent issue. The simulation of
human societies has not yet reached the situation where there is enough evidence to
obtain much more than simple plausibility. At this current stage of social simula-
tion, getting close enough to be deemed a “plausible” model is difficult enough, and
there is almost never data enough to justify a stronger claim. Thus claims of
anything stronger should be treated with appropriate scepticism.
26.3.2.1 Building Towards Realism
One common approach is to start with a fairly simple model that is easier to
understand and then to add aspects and mechanisms that are thought to be significant
aspects of an observed system. That is to build an additional level of realism to make
the model more plausible or useful in some way (e.g. as a thought experiment).
This is sometimes known as the TAPAS approach, i.e. to ‘take a previous model
and add something’. It is consistent with the engineering principle of “KISS” –
keep it simple, stupid. In this approach one starts simply and adds more features
or aspects one at a time if and only if the simple approach turns out to be inadequate
for some purpose.
Thus Izquierdo (2008) starts with some standard models of the iterated prisoner-
dilemma games and adds some more “realistic” features, such as case-based
learning and reasoning. A key idea in this is to maintain rigorous understanding
of the extended model, but take a step towards models that might eventually be
validated against observed data from human interactions.
Whether one would, in fact, reach useable and valid models by this means is
contested, with the alternative approach being to start with a complex model that
reflects the evidence as well as possible and then seek for understanding and
simplifications of this (Edmonds and Moss 2005).
To investigate the social aspects of socio-environmental systems, some highly
complicated models often have to be used that include the relevant biophysical
dynamics, coupled with social simulation. Rather than developing all such
components of the simulation system “from scratch” (and because the biophysical
parts are relatively universal), these systems have a modular architecture designed
to be reusable. It may therefore be more accurate to refer to the software as a
“toolkit” from which various sub-models can be configured depending on the
desired purpose of a particular study. In the area of land-use simulation, PALM
(Matthews 2006) is one such integrative model, and FEARLUS (Polhill et al. 2001,
2008) is part of another longstanding approach to socio-ecological modelling. With
each iteration the toolkit obtains further refinement and new features – whilst the
level of understanding of its user(s) increases. The social simulator is interested in
what additional complexity the human interaction part brings, and to what extent it
adds realism to the model’s behaviour when compared with observed evidence.
26.3.2.2 Extending Evidence to Extrapolate to Unobserved Cases
Data about social systems is often limited to measurements from a limited number
of observed cases. Thus there may well be many cases within a spectrum of
observed cases that one would like to estimate the outcome for. Of course, one
could use simple statistical techniques such as linear interpolation or similar to do
this, but such techniques depend upon assumptions concerning the regularity of the
results with respect to small changes in the set-up, which may be implausible for
some social systems. In this case one might simulate the system using plausible
assumptions, and validate it against the known observations, then find the outcomes
for set-ups that are different to those observed. For the results of this to be reliable
the simulation needs to be well validated; for it to correctly indicate the observed
cases; to not differ very much from the observed cases (in contrast to the case
described in Sect. 26.3.2.1); and for any unvalidated assumptions to be of a mild
and uncontroversial nature.
The plausibility of the results from such experiments depends upon the validity
of the original measurements as well as the generality of the assumptions (which
must be plausible for the unobserved as well as observed cases).
For example (Brown and Harding 2002) use a microsimulation model to extend
regional socio-demographic (census) data to cases that are not directly observed
(synthetic householder-level records for each spatial district). The extension is

attempted with assumptions that are thought as deliberately cautious.
The “Sienna” programme (Snijders et al. 2010) fits a particular class of dynamic
network model to “waves” of panel data. Simplifying a little, what happens is that
the modeller specifies some basic assumptions (e.g. symmetry of network links)
along with more than one set of panel data concerning the properties of the nodes at
certain points in time. The algorithm then finds the dynamic network model that is
consistent with the given specified constraints and that most closely fits the data.
This is directly analogous to the process of fitting a line to a set of values using
minimum total squared errors (or similar). What one gets out of this are some
“surprise free” projections to network and node properties for times other than those
given in the waves of panel data. This is not simulation in the same sense as other
simulations mentioned here, since what is simulated is not a kind of process (that is
given in the base specification of the family of models this technique uses) but
rather a set of structures and values that fit given data in a statistical sense. When
this technique is reliable and what its particular biases are, have not yet been
established.
26.3.2.3 Establishing the Consistency of a Process/Assumption

with Evidence
Oftentimes a social process is not included in a study because it is not considered to

be valid in the same way as a physical or biological principle might be. This is
particularly true in historical examples where social processes are less in evidence.
Going back to our second example of generative archaeology (Sect. 26.1.2), there
are few archaeological findings that suggest a particular social structure and set of
social processes, hence the need often for guesswork and the resulting coexistence
of many competing theories. However, as previously discussed, this is an area in
which social simulation can make an important contribution.
Perhaps the most well-known example is the Artificial Anasazi simulation model
(Axtell et al. 2002). The objective was to see if a model could be constructed broadly
consistent with available evidence – the number of households settled in part of the
U.S. South West region over the period 800–1350. The performance of the model
was impressive in its convergence upon the actual historical time-series after a
calibration of several parameters (a ‘fitting’ process), which suggested new social
explanations as to the apparent land abandonment after 1350 might be possible.
Interestingly a later paper (Janssen 2009) demonstrates that the model fit is mainly
explained by two parameters related only to the model’s carrying capacity. The
author argues that a more insightful basis might be to generalise the target domain,
working initially from the less concrete goals, rather than fitting a particular case and
focusing on one evident and quantifiable trend (such as population). If the evidence
base is broadened to include more ethnographic knowledge this approach would
resemble the pursuit of abstract goals as discussed in Sect. 26.4.1.4.
Data about real world social networks introduced at the design or validation
stages can be a valuable way of checking the consistency of a model. For example
Guimera et al. (2005) reconstruct the history of team collaborations in five different
scientific and artistic fields and the development of corresponding collaboration
networks. The authors develop and parameterise a probabilistic model of team
selection. Using real data on team sizes, along with estimation of probabilistic
parameters, to control the team assembly mechanism, the characteristics of the
resulting networks (the degree distribution and the largest component) are com-
pared with the real ones (independently for each of the five cases). The interest is in
the transition of the collaboration network from “isolated schools” to an “invisible
college” – the point at which the largest component of the network contains 50 % or
more of the nodes (which is the case for all representative fields). All simulated
network measurements are shown to be in close agreement with the real networks,
which establishes the plausibility of the proposed team selection mechanism.
However, being a probabilistic model it does not attribute any particular decision
process to this mechanism that might be able to reveal new questions.
Another example is in (White 1999), which attempts to evaluate some statistical
assumptions against data about marriage systems in different cultures using a
“controlled simulation”.
26.3.2.4 Analysis of Influence Factors
In any complex system it is very difficult to estimate the importance of different

factors on particular outcome measures or results. This is due to the “non linearity”
in many social systems where a normally insignificant factor can trigger a system-
wide change in behaviour. However, given a trusted simulation model of the
system, one can perform experiments to determine the importance of each factor
in the class of simulation set-ups that are run. Thus one does not have to determine
the relative importance of factors on an a priori basis; one can simply run the
experiments and measure the outcomes. Clearly this approach depends on having a
reliable simulation model.
In (Saqalli et al. 2010) a simulation model of the development over several
generations of a rural agrarian society is investigated to weigh the importance of
several different model parameters on simulation results. In the simulation
experiments reported, four parameters were assessed in relation to six state variables –
with measurements taken at the end of the run. The model was based on a case-study
of the Nigrien Sahel, typified as a low data situation where, in particular, little has
been published on the social factors governing access to economic activities (includ-
ing off-farm activities so often neglected as an important revenue generating source)
or on intra-household dynamics (which the authors recognise as having a complex
structure). The objective was to assess the robustness of results against variation in
socio-economic and biophysical parameters to show that it is “constrained by the
different parameters of its structure” (Saqalli et al. 2010: para. 3.6). This step provides
the researcher with an improved understanding of the range of outcomes possible
with the model and what might constitute a significant or meaningful difference when
comparing outcomes. It is worth noting, however that the single parameter approach
neglects any possible parametric interaction that could be identified from a pair-wise
analysis of influence factors.
A very different example of this is (Yang et al. 2009), which studies the factors
that influenced success in the system of Chinese civil service exams that existed in
the Imperial era in mainland China. The simulation model used historical data from
civil service records and some assumptions to assess the importance of factors such
as class, wealth and family connections in terms of success at passing this exam
(and hence obtaining a coveted civil service post). It is difficult to see how such
indications about events that are otherwise lost in the past could be obtained,
although this is open to the criticism of being unfalsifiable.
The disadvantages of this approach are that the assessment of influence is only as
good as the simulation model, and it only samples particular sets of initial
conditions – it does not rule out the case where very special values of parameters
cause totally different outcomes (unless one happens to be lucky and sample these).
26.3.2.5 Assessment of Policy Options
Recently more and more articles have appeared in the literature featuring ABMs
that address policy-making in contemporary issues such as developmental
sustainability and climate change adaptation. For example Berman et al. (2004)
consider eight employment scenarios defined by different policies for tourism and
government spending, as well as different climate futures, for an ABM case study of
sustainability in the Arctic community of Old Crow. Scenarios were developed with
the input of local residents: tourism being a policy option largely influenced by the
autonomous community of Old Crow (stemming from their land rights), and
attracting great local interest. In ABM, policy options are often addressed as a
certain type of scenario (scenarios are discussed in Sect. 26.3.2.9), embedding the
behaviour of actors within a few possible future contexts. The attraction of this
approach is that the model could potentially be used as a decision support tool, in a
form that is familiar to many analysts, to provide answers to very specific policy
questions. The merit is that it can improve the reckoning of human and social
factors and information into the issues at stake; the drawback is the multiplication
of uncertainties, not least of which is that we do not convincingly know how social
actors might adapt (even if the possible policy options are more concrete).
For example (Alam et al. 2007) investigates the outcomes indicated by a
complex, and detailed model of a particular village in South Africa. This model
in particular looks at many aspects of the situation, including: social network,
family structure, sexual network, HIV/AIDS spread, death, birth, savings clubs,
government grants and local employment prospects. It concludes with hypotheses
about this particular case. This does not mean that these outcomes will actually
occur, but this does provide a focus for future field research and may provide
thought for policy makers.15
26.3.2.6 Social Engineering: “Designing” Better Systems
Market design is the branch of economic research aiming to provide insights about
which market protocol, i.e. interaction structure and information circulation rules, is
the best to obtain certain characteristics of a market. Agent-based simulation seems to
be a good method to test several such protocols and see their influence on economic
performances, e.g. efficiency, fairness, power repartition (Marks 2007). Each proto-
col is already known for its advantages and disadvantages (e.g. Dutch auction is fast;
Double Auction extracts the highest global profit). Since not every good aspect can be
achieved with a single protocol, one has to choose the aim to attain (LeBaron 2006).
Then, assuming agents act rationally, it is possible to compare protocols to see what
difference it makes in prices or other indicators (e.g. Kirman and Moulet 2008). Many
studies were designed to fit the context of electricity markets (very crucial since
unpredicted shortages are a problem and prices can vary quickly) and are usually
treated by a comparison of protocol (Nicolaisen et al. 2001). One can also note the use
of “evolutionary mechanism design” (Phelps et al. 2002; March 2007) where
strategies of three types of actors – sellers, buyers and auctioneers – are all submitted
to evolution and selection and the actual organization of the market evolves while the
context of production and demand is fixed. In today’s economy more and more
artificial agents really interact – either on bidding on consumers’ sites or even in
financial markets (Kephart and Greenwald 2002) – so there is some convergence
between real markets with artificial markets and designed artificial systems which
utilise market mechanisms. For a more detailed discussion of modelling and design-
ing markets see Chap. 23 in this handbook (Rouchier 2013).
26.3.2.7 Data Integration
A mundane and sometimes overlooked aspect of the scientific process is simple

description. That is, recording what has been observed in a suitable form. Tradi-
tionally these forms have included the like of narratives, logs, videos,
measurements, and pictures. However simulations can also be used as a sort of
description, where the aim is not to express a theory about a mechanism, but rather
to integrate as much of the relevant evidence about what is observed as possible
about a particular target. Simulation has some advantages in such a process, since it
can allow the integration of several different kinds and levels of evidence within
one framework. To take some examples: aspects of narrative texts can be
15
Although in this particular case it did not, since the model indicated outcomes that the policy
makers preferred to ignore, being not compatible with the policy they had already fixed upon.
incorporated within the behavioural rules of an agent; the social network of sub
communities be compared to those that result from the simulation, the time-series
can be compared to the corresponding time-series derived from measurements on
the simulation outcomes, and survey data compared to the equivalent answers at
instances of the simulation runs. Such integration is far from easy, since some
aspects are programmed directly (e.g. agent behaviour) whilst others have to be
achieved in terms of the results (e.g. aggregate statistics about the outcomes).
Achieving any particular set of outcomes in a social simulation is difficult due to
the prevalence of unpredictable interactions and effects (i.e. emergence), so the
achievement of a data-integration model is not an easy one. Such models are not
entirely (or solely) a description since the structure of a simulation sometimes
brings into question the consistency of the various parts of the evidence. Thus if
it is difficult to square an account of how individuals behave with some of the
outcomes, one may be forced to make some choices, including possibly adding in
aspects that are not directly observed. This is alright as long as these are clearly
documented and can provide fertile issues for future data collection efforts. How-
ever, such data-integration models do not aim for generality beyond the case study
(or studies) focused on, and hence can avoid “high” theory to motivate simulation
features where this is not supported by the evidence with respect to the target case.
It is not that there is no theory in such simulations – any description or abstraction,
however mild, will rely on some theory but the point is that in a descriptive
simulation such theory is either well established, or relatively mundane.
Examples of simulations that intend to be descriptive in this sense include
(Christensen and Sasaki 2008) which aims at producing a simulation of the evacu-
ation from a particular building, with a view to a future evaluation of evacuation
plans and facilities, in particular with regard to disabled people. It uses many
particulars of the building structure, but makes assumptions (albeit of a plausible
variety) about how people behave when evacuating. Likewise (Terán et al. 2007)
aim to simulate land-use and users within a forest reserve with a view to producing
a computational representation of this. As in similar simulations there is a mixture
of assumptions that are backed by some evidence and some that are plausible
guesses. This simulation is loosely validated against some data and shows results
that confirm the results found in some other models. The ultimate use of this (and
similar models) is not described.
Such simulations can take a long time to construct, involving many iterations of
model development as well as being complicated and slow to run. The advantage
of such models is that they are a precise and coherent representation of a set of
evidence – in a sense an encapsulation of a particular case study.16 This can be the
basis for experiments and inspection that can lead to further abstraction steps,
resulting in theories of the processes observed within the data-integration model
being modelled in simpler models whose properties are easier to establish, but
whose outcomes can be checked against targeted experiments on the data-
integration model.
16
To be precise: a possible encapsulation of a particular set of evidence on the case study.
26.3.2.8 Finding New Questions and Areas of Ignorance, Hypothesis

Suggestion
Another use of a simulation is as an aid to good observation. That is, suggesting

issues and questions that should be sought in order to gain an adequate observa-
tional coverage. The simulation is developed as in the data-integration model
above, including different aspects of the observational evidence that are available.
It is often the case that it is only when one tries to simulate a process that some of
the gaps in knowledge become clear. Thus building a simulation as one is observing
can help direct the data-gathering research in order to complete an adequate
computational description. In this sense it forms a similar role to simulation in
some cognitive science (Newell 1990; Sun 2005).
For example, (Moss 1998) exhibits a simulation built on a mixture of bases: (a)
an assumed but plausible cognitive architecture which captures how one might
divide up a problem into sub-problems until they are doable, (b) some suggestions
elicited from an expert from the domain and (c) plausible guesses for the remainder.
This model attempted to examine behaviour in the face of crises (defined as when
one unwanted event causes another in an out-of-control chain), in particular how
the rotating of crisis-management teams and the information they pass on to the
next team might impact upon their effectiveness at fighting the crisis. The results
were not independently validated, but this is not the point of this simulation. As the
author says:
. . .results obtained with the North West Water model indicate a clear need for an investi-
gation of appropriate organizational structures and procedures to deal with full-blown
crises.
In contrast, (Younger 2005) is a very much more abstract model, which is only
loosely built upon evidence, but has the same broad aim of suggesting hypotheses – in
this case, hypotheses concerning the occurrence of violence and revenge within
egalitarian societies. Clearly the plausibility of the hypotheses or questions suggested
by a simulation will be greater when the simulation is more firmly rooted in evidence.
However, hypotheses and questions can be worthwhile investigating whatever their
source, and at least having a simulation grounds and defines the question in a precise
way, making clear what it might explain and the sort of other issues and questions that
might accompany it.
26.3.2.9 Creation/Critique of Scenarios
Berman et al. (2004) present an example of scenarios being used to constrain

models to produce simulations of the wider consequences of those scenarios (as
measured by relevant socio-economic or environmental indicators or by their
possible influence on human institutions) that can then be used to inform
discussions with stakeholders and may ultimately produce a better understanding
of such changes. Bharwani et al. (2005) use climate change scenarios to investigate
adaptive decision making among villagers in the Limpopo district of South Africa,
focusing on the use of seasonal forecast information in farming strategies. Data
from the Hadley Centre climate model – HadAM3 – showing a 100-year drying
trend with increasing potential evapotranspiration (PET), were used as model input
(providing PET and precipitation values). Results show a degree of resilience to
these changes is afforded when the forecast is correct 85 % of the time so that
farmers establish increased trust in, and use of, seasonal forecasts. They are able to
choose cropping strategies that are suited to climate change, though this
behavioural shift may only occur over a very long time-frame.
Bharwani et al. (2005) introduce the use of scenarios into the methodology in a
further and very interesting way: by postulating them as ‘drivers’ of actors’
decision-making processes. In this ethnographic approach, the authors combine
simplified scenarios across different domains (irrigation, forecast and market)
asking respondents what they would do under each scenario, in a given context.
This information was then used to produce the model rules for the agents’ decision-
making.
In either case, where conventional scenarios used in futures planning can seem
rather terse and lacking in specifics – which may be a limitation to their subsequent
use in policy discussion – simulation outputs that explore scenarios offer a great
deal of detailed information “that would be difficult to imagine otherwise” (Berman
et al. 2004: 410). Moreover, this can apply at different levels of analysis from trends
in macro variables down to the impacts on different sectors and regions, as well as
differentiated impacts for agents fitting any given ‘profile’ in which the analyst is
interested. Perhaps greater care has to be taken, however, in the use of model-
generated scenarios, to ensure that these are not taken as ‘more accurate
predictions’ by virtue of being ‘computed’ stories rather than conventional ‘imag-
ined’ stories.
Scenarios are often used in policy discussions, e.g. climate change. However
they are usually somewhat vague and/or only described in qualitative terms.
Simulations can be used to produce consistent scenarios or to produce models
that instantiate aspects of given scenarios (Taylor et al. 2009).
26.3.2.10 Intervention with Stakeholders
Instead of developing a simulation to represent some aspect of society, one can also
try to use a simulation to intervene in society. That is: use a simulation to change
some interaction between stakeholders, for example to facilitate collective decision
making or mutual understanding. One well-known approach is the Companion
Modelling approach which has been developed in the last decade (see Chap. 10;
Barreteau et al. 2013).
An example is demonstrated by Etienne (2003). Here, a model is used in
conjunction with a role-playing game to show chosen participants the issues that
can arise when several users compete on a pastoral resource. The building of the
model was an integration of multidisciplinary knowledge acquired on French
Mediterranean sylvopastoral systems into a model capable of representing the

interactions between ecological dynamics and social behaviours. In order to help
foresters and livestock farmers to better integrate these interactions into their
planning work, a multi-agent system was designed to simulate different manage-
ment strategies and to compare their impact on forest quality. This model was
coupled with a role-playing game (RPG) initially developed as a didactic support to
sylvopastoral training programmes and very soon, it proved useful in the
negotiations and interactions between livestock farmers and foresters involved in
the management of the same forest. The tool revealed itself flexible enough to make
it possible to play with actively involved stakeholders such as the current users of
the resource (local farmers and foresters), with potential regulators of the system
(managers or administrators), technical experts (extensionists, technicians) or
learners concerned with the topic (students, scientists).
This model is effectively an intervention between the livestock farmers and
foresters by being a subtle mediating tool, allowing the stakeholders to play at
decision making, to educate them in the possible effects of their decision-making
and to thus encourage debate and introspection. This model has also been used for
didactic purposes (Sect. 26.3.1.1) by getting agronomy students to play it.
26.4 Inputs and Results of Simulation Models
One method of assessing the use, and ultimately the success, of a simulation for
understanding aspects of society is to tease out what has gone into making a
simulation model, the input, and how the results from the simulation are interpreted
and used, the output. These, the input and the output together form the mapping
from the computer program and its calculation to and from the target of study. They
are crucial parts of what characterises a simulation, even if they tend to be described
in a less formal manner than the simulation code and behaviour.
26.4.1 Inputs
What is put “in” to the design of a model tends to be more explicitly distinguished
in papers than what comes “out”. This might be because the “job” of a simulator is
seen as a process of deciding what processes and structures will go in to a model and
because the inputs are under the control of the simulator in a way which the results
certainly are not, and hence can be displayed and talked about with greater
confidence. However, all social simulations are based on a raft of different
assumptions, settings and processes. These are somewhat separated out for analysis
here.
26.4.1.1 Evidence-Based Assumptions
If there is some evidence about the nature or extent of the processes that are being
observed then this can be used to inform the set-up or structure of a simulation. For
example, evidence from social psychology might be used to inform the specifica-
tion of the behavioural rules of a set of agents in a simulation, or the narrative
account of a participant used as the basis for programming a particular agent.17 Of
course, it is rare that such evidence constrains the possible settings and algorithms
completely but rather that it partially constrains these or constrains them in con-
junction with additional assumptions from another source. Clearly the more
assumptions can be constrained by evidence (either directly or as the result of
previous research) the better. The presence of other assumptions and inputs does not
make a simulation useless, especially if documented, it is just that simulation results
and usefulness are relative to the assumptions, so that if assumptions are included
that are completely misguided and critically affect the results, then this would
seriously limit their use with respect to the observed world.
26.4.1.2 Indirectly Inferred Settings
In situations where there are some parameters that are unknown, and where there is
a relative abundance of time-series data, one can attempt to infer the values of these
by seeing which parameter values result in the model giving the best fit to a segment
of the time-series data. This is a sort of evidence-based setting, but it often seems to
be used when the parameters concerned do not have any discernable meaning in
terms of the target of modelling. Thus when there is a tradition of using a certain
kind of decision or learning algorithm in an agent, then this might be “fitted” to an
initial segment of the data (so called “in sample” data) even when it is unlikely18
that the algorithm corresponds to how the target agents think. Thus the credibility
of this technique is dependent on the reliability of the other assumptions in the
model, and the meaning of the parameters being fitted. If the parameter was a
scaling parameter, then this might well be a sensible way to proceed.
17
This can either be done directly as a translation of an interview text into programmed rules or
used to check that such programming is correct by comparing the resulting behaviour of an agent
against what happens when the simulation is run. Thus there is not an absolutely clear distinction
between verification and validation from evidence. In a sense this second method is verification
since the programming is rejected until correct but, on the other hand, this is part of the production
of a simulation, which may only be completed later for its validation as a whole.
18
Unlikely in view of the psychological or sociological evidence about the target subjects.
26.4.1.3 Documented Theoretical Assumptions
Clearly, researchers do not invent all the details and algorithms of their model from
the ground up, but are doing their research with knowledge of certain approaches
and algorithms and within a community of other research, with established
techniques and traditions. Thus many parts of a simulation model will be based
on those of other models, or algorithms from other fields. Thus many models in
Economics will use a decision algorithm based on constrained comparisons of
predicted utility and other models might import techniques from the fields of
Artificial Intelligence or Evolutionary Computation. It seems impossible to
completely avoid all such theoretical assumptions; however there are distinctions
to be made in terms of the strength of the assumptions, the likely biases behind such
assumptions and the degree to which they are evidence-based.
“Strong” assumptions are those that are surprising or seem to specify conditions
that are rarely observed. Thus an assumption that an agent has in effect a perfect
model of the economy in its head is a very strong assumption, since even experts
find it difficult to understand the economy as a whole. Strong assumptions are often
introduced to allow for analytically solvable models to be specified and used, for
example the assumption of perfect information in game theory. Whilst analytically
tractable models were necessary when there was no other avenue for the precise
modelling of many kinds of phenomena, the advent of cheap computing power and
accessible simulation platforms means that often more appropriate methods are
now available, with analytic models possibly being used to check or understand the
reference simulation model, rather than being the focus. Clearly, other things being
equal, weaker assumptions are preferable to strong ones – the stronger an assump-
tion the more evidence is needed to justify its use. In any case all such assumptions
should be as fully documented as possible.
26.4.1.4 Explored Conditions
In much simulation work there will be a focus hypothesis or set of hypotheses that
are being investigated. In these cases it is usual to try the simulation using that
hypothesis and then compare the results to those coming from a version of the
simulation with a different hypothesis implemented. This provides evidence about
the possible effects of that hypothesis on the outcomes, allowing comparison with
evidence and possible subsequent inference as to which is more likely to be the
case. The clearest case of this is testing the significance of the inclusion of a
hypothesis against that of a “null” model19 to see if the properties of the results
that are deemed significant indeed result from the hypothesis or from other aspects
of the model. Thus a simulation of a stock-market might compare the results
19
A “null” model is a model version where the claimed causal mechanism is eliminated to see if
the resultant “effect” would have arisen as the result of background (e.g. random) mechanisms
anyway.
obtained with intelligent agents that notice patterns in pricing and try and exploit
these to agents that buy and sell at random. Unfortunately it is sometimes the case
that a simulation is presented purporting to show the significance of a hypothesis
without indicating what the comparison case is.
26.4.1.5 Randomness and Other Essentially Arbitrary Assumptions
A simulation modeller is often faced with deciding how to design a part of a

simulation model for which there is neither evidence nor any tradition of modelling
to guide them. In such a case one might simply make that aspect random. For
example, where it is unknown how a kind of choice is made in the modelled
situation it might be implemented as a random choice in a simulation model of
that situation.20 This is usually done in conjunction with a “Monte Carlo” approach
which runs the simulation a number of times and averages the resulting different
sets of outcomes. Presumably this is done under the assumption that the introduced
randomness will be averaged out leaving only the effects of the other design
settings – however this assumption is rarely proved but often simply remains a
hope. Of course, if it can be shown that the value of the particular input does not
influence those aspects of the results that are deemed significant by a series of
simulation experiments (or otherwise) then a random input or process might well be
acceptable. However, in this case, a constant value might be simpler and have the
same effect.21
We suspect that many uses of randomness in simulations are in the nature of a
programming “stub” – that is, a stand-in that the programmer intends (or intended)
to expand to a more plausible algorithm at a later date. Whilst this is perfectly
acceptable during model development and to some extent inevitable given that
researchers always have time constraints, such stubs are likely targets for criticism
by other researchers. At the very least some exploration of them to assess the extent
to which they affect those aspects of the results deemed significant is advisable.
Randomness can be considered as a special case of a broader class of
assumptions: those that are added into the model simply to get it to run, and for
no theoretical or evidence-related reason. Hopefully these are honestly declared
rather than “dressed up” under some other category, although often these are
excused under the broad umbrella of “simplicity”.22
20
Another option is to exhaustively try all the possibilities in a series of simulations or by using
techniques such as constraint logic programming but these are technically difficult and require a lot
of computational power.
21
There are possible reasons why a constant value might not work, for example when the input
provides some mechanism of symmetry-breaking.
22
There is nothing wrong with assumptions that had to be made due to constraints on resources,
such as time, expertise or computing power, but it is simply disingenuous to pretend that this is
sanctioned by a higher “virtue”.
26.4.1.6 Undocumented Assumptions
It is not feasible to document all of the assumptions in a model. Firstly, this might
take too much space in a single paper23 and secondly, many might be previously
established and well known to those in a particular field of work. However it is also
likely that researchers are simply not aware of all the assumptions inherent in their
simulation models, due to the limitation of human cognition.24 Clearly, it is part of
the job of other researchers to point out undocumented assumptions where these can
be shown to be significant.25
26.4.2 Outputs
A similar set of distinctions can be made about what comes out of a simulation, the
results. There is not an obligation to describe all the outputs from a simulation, but
rather one tends to get a sample of results, which typically is composed of: sample
results, sensitivity analyses, evidence of validation, and the outcomes from
experiments designed to test a hypothesis. However not all the details of the results
are considered as equally significant – we now consider each of these in order of
increasing significance.
Firstly, there are those aspects of the results that are considered as artefacts of the
model, for example the randomness that might have been input to the model.
Secondly, there are those features that might be considered to reflect some of the
model structure and the processes that result from them. These features may not
be one of those parts of the simulation that reflect what is being modelled, but
may be caused by theoretical or arbitrary assumptions that were put in. These
features of the results may well not be so much of a surprise to the modeller.
Thirdly, there are those features of the results that are interpreted as indicating
something about what is being modelled, for example they may suggest a
hypothesis about those phenomena. That is they indicate a possibility that may
be inherent in what is being modelled or that is possibly inherent in the target of
modelling. This may well go beyond what can be directly validated in the model
but, for example, track counter-factual possibilities concerning what might have
occurred.
23
However this is a poor excuse given that a technical paper which is relatively complete can
easily be archived and referenced along with a journal article or report.
24
Alternatively it may be because the simulation designers had not thought about what they were
doing.
25
It is trivial to point out that a simulation has missed out some assumption or other, but this is not
very useful. It is far more useful to point out how and why an assumption might be important and
for which purposes.
Lastly, there are those features that would be positively expected of the phenomena
being modelled. That is, if it were not present then it would be taken as evidence
that there was something amiss with the model. In other words, it is a necessity
of the phenomena. It is against this category of results that models are validated.
It is not easy to distinguish these different categories of significance in terms of
the results, since causation within a model can be very complicated, being a result
of many of the model aspects interacting together. It is also usually the case that the
modeller has hypotheses (or assumptions) about what aspects of the results are
significant in which ways. This is crucially useful information to impart to a reader
interested in the results. However, this is often left implicit.
One might justifiably criticise many social simulations in terms of the lack of
empirical grounding of both inputs and outputs. Many social simulations have only
the weakest connection with anything observed: the inputs are largely assumption-
based, and indeed often highly artificial; the outputs only relating in the broadest
way to any data and then only in terms of a few aspects of the possible outputs (i.e.
only a few selected aspects are deemed significant to what is observed and then in
the loosest, “hand waving”, manner). It may well be that simulating human society
is just very, very difficult, and one suspects that it is simply easier to stick to
considering abstract ideas.
26.5 An Assessment of Simulations for Understanding Social

Phenomena
As discussed in previous sections, complex social phenomena containing multiple

interacting actors can be computationally analysed using ABM. In this sense it is
worth noticing that evidence-driven modelling approaches tend to guide modellers
towards more up-to-date data and better understanding of social phenomena than
theory-driven approaches. To avoid using highly speculative (“strong”)
assumptions whilst modelling, it is essential to have very detailed knowledge of
the phenomena in question. That comprehensive grasp, backed with evidence, is
helpful to identify relevant model parameters, estimate configuration values and
evaluate simulation results.
Perhaps the biggest open challenge for modellers is to work out how results from
social simulation models can be useful beyond theory and hypothetical illustrations.
To harness the potential explanatory prowess of ABM in the social sciences, it
seems necessary to safeguard models with scrutinised findings from large amounts
of data about the social phenomenon. Nevertheless this is currently very difficult as
access to such datasets is often hampered either due to (Lucas 2011):
• Being non-existent, thus requiring funding to collect and process all relevant
information;
• Being unavailable because of privacy agreements, such as in non-disclosure
agreements;
• Or, if at hand, being incomplete or outdated.
Validating outputs that have no comparable evidence is an eminent issue, and
this seems only clarified by comparing simulation results with further data. In
(Lucas 2010, 2011) there is a discussion of a survey carried out in 200926 with 12
leading academic researchers – each of them having managed mid to long-term (3
to 5 years) social simulation projects in Europe and the United States – regarding
how their endeavours were modelled, applied and whether these have been useful
beyond theory. The questionnaire consisted of the following questions aimed at
eliciting views:
1. How were fieldwork findings used to guide the simulation development?
2. What were the contributions of simulation results to stakeholders?
3. Were simulation results regarded by them as useful as fieldwork findings?
4. What could have improved the chances of providing these via simulations?
Three businesses offering simulations, which take into account social behaviour,
were also approached, but all refused stating non-disclosure agreements. All
interviewees mentioned that their models targeted the scientific community and
could only, at best, provide plausible results regarding scenarios and that, albeit
coherently justified, nothing obtained in simulations could be regarded as directly
useful for policy-making purposes. All 12 cited that gathering detailed data about
the actual phenomena by interviewing stakeholders and reviewing existent litera-
ture helped to better understand their context and served as a good guide during the
modelling process. More than half (7) said stakeholders and policy-makers were not
interested in simulations per se, and that only real success cases (even those with
only anecdotal evidence) are what are taken into account in decisions. On the other
hand few (3) mentioned that simulations, despite their shortcomings, attracted
significant interest from stakeholders and policy-makers. In their view the
modelling process is a time-consuming task and occasionally even regarded it as
unproductive. This contrasts with positive experience with fieldwork, which –
despite also demanding a lot of time – a majority (10) confirmed was useful in
acquiring relevant new knowledge to practitioners. Engaging with stakeholders and
policy-makers was interpreted by all (12) as indispensable to improve their under-
standing of the actual social phenomena. Yet maintaining efficient interaction and
managing practitioners’ interest over many months of work engagement was
generally deemed as strenuous and difficult. This finding is partially supported by
other larger projects regarding the effective collaboration between policy-makers
26
Data was compiled until July 2009, on time for discussion at the 6th European Social Simulation
Association Conference in September, and then updated for the Artificial Intelligence & Society
Journal.
and researchers such as (Young and Mendizabal 2009). Some (4) cited that social
simulation models perhaps could be integrated in tools for aiding mediation of
group decisions. Lacking confidence about model results is another aspect raised by
all (12) interviewees, along with difficulties of coding or interpreting qualitative
data appropriately, plus communicating the model itself and its results intelligibly
to a non-technical audience. Some (6) interviewees said that they had no intention
to influence the social phenomenon in question, but only to model it plausibly.
Commissioned fieldwork, to date, has greater chances of being timely useful in
this sense, as resultant reports can provide very specific and up-to-date information
that is easily understood by stakeholders. The implication of such new information
is usually more quickly recognised than results obtained in simulation models, as
these usually deal with intricate processes over longer time scales. Claims that
social simulations could support, or guide, decision and policy-making seem only
possible with aid of in-house experienced modellers working in close liaison with
stakeholders. That is necessary as, meeting social simulation aims and objectives
(beyond theory) is still an experimental process of many trials and errors, which
requires good technical judgement to know what should be done when simulation
results have gone beyond the existing evidence (what is factually known).
A survey of agent-based modelling practices in the literature (Heath et al. 2009)
revealed that there were more applications in social sciences (24 %) than in public
policy (8 %) among the 279 articles published between 1998 and 2008. Of the 68
social science applications, a majority (66 %) used ABM as a hypothesis generator
for systems that were assumed to be less well-known, and the remaining articles
(34 %) used ABM as a mediator in order to represent a system that was moderately
understood and gain insights on the system’s characteristics and behaviour. In
contrast, only a small portion of the 23 public policy applications (4 %) used
ABM as a hypothesis generator, and most articles (96 %) used ABM as a mediator.
26.6 Conclusion
Simulation has undoubtedly helped to improve our understanding of human society,

although in a number of different and usually indirect ways. It is fair to say that, so
far at least, this has served to improve our understanding of some societal processes
and our ideas about society rather than directly in terms of being able to strongly
predict aspects of society or conclusively test hypotheses about society.
Simulation is not a replacement for other ways of understanding society27; it is
simply a flexible way of precisely modelling it in a way that can represent some of
the dynamic and complex aspects of it. It can be especially productive in conjunc-
tion with other approaches. For example, analytic models can be used to check the
outputs and properties of a simulation model and help us understand the model and,
27
At least, not in any of the cases we have yet come across.
conversely, a simulation used to probe and check some of the simplifications and
assumptions used in an analytic model. In participatory models, social science
techniques of engagement and elicitation can be used to inform the construction
of agent-based social simulations as well as the simulations suggesting what might
be usefully investigated in terms of the collection of new data.
Clearly social simulation has some way to go in terms of the maturity of its
method and the reporting and use of simulation models. There are still a number of
areas in which the methodology needs substantial improvement and standardising.
There are also significant unresolved issues, such as how to decide what level of
detail to include, and to what extent one should rely on prior theory.
We predict that simulation will be even more significant in helping us under-
stand human society in the future, in particular where it is used in close conjunction
with other relevant approaches.
Further Reading
The best general introduction to social simulation is (Gilbert and Troitzsch 2005)
which covers general issues and gives code examples. For a wider range of views on
social simulation the published papers from the US National Academy of Sciences
colloquium on “Adaptive Agents, Intelligence, and Emergent Human Organization:
Capturing Complexity through Agent-Based Modeling” (PNAS 2002) give a good
cross-section of the different approaches people take to this area. It is difficult to
point to further good sources as this topic is so diverse but the Journal of Artificial
Societies and Social Simulation has many accessible papers.
References
Alam SJ, Meyer R, Ziervogel G, Moss S (2007) The impact of HIV/AIDS in the context of
socioeconomic stressors: an evidence-driven approach. J Artif Soc Soc Simul 10(4), http://
Axelrod R (1997) The complexity of cooperation. Princeton University Press, Princeton
Axtell R, Axelrod R, Epstein JM, Cohen MD (1996) Aligning simulation models: a case study and
results. Comput Math Organ Theor 1(2):123–141
Axtell RL et al (2002) Population growth and collapse in a multi-agent model of the Kayenta
Anasazi in Long House Valley. Proc Natl Acad Sci 99(3):7275–7279
Berman M, Nicolson C, Kofinas G, Tetlichi J, Martin S (2004) Adaptation and sustainability in a
small arctic community: results of an agent-based simulation model. Arctic 57(4):401–414
Bharwani S et al (2005) Multi-agent modelling of climate outlooks and food security on a
community garden scheme in Limpopo, South Africa. Phil Trans R Soc B 360
(1463):2183–2194
Biggs R, Carpenter SR, Brock WA (2009) Turning back from the brink: detecting an impending
regime shift in time to avert it. Proc Natl Acad Sci (PNAS) 106:826–831
Brown L, Harding A (2002) Social modelling and public policy: application of microsimulation
modelling in Australia. J Artif Soc Soc Simul 5(4), http://jasss.soc.surrey.ac.uk/5/4/6.html
Cartwright N (1993) How the laws of physics lie. Oxford University Press, Oxford
Christensen K, Sasaki Y (2008) Agent-based emergency evacuation simulation with individuals
with disabilities in the population. J Artif Soc Soc Simul 11(3), http://jasss.soc.surrey.ac.uk/11/
3/9.html
Clifford J (1986) Writing culture: the poetics and politics of ethnography. University of California
Press, Berkeley
Cole HSD, Freeman C, Jahoda M, Pavitt KL (eds) (1973) Models of doom: a critique of the Limits
to Growth. Universe Books, New York
Curtis J, Frith D (2008) Exit polling in a cold climate: the BBC–ITV experience in Britain in 2005.
J R Stat Soc A Stat Soc 171(3):509–539
Deffuant G, Weisbuch G (2007) Probability distribution dynamics explaining agent model con-
vergence to extremism. In: Edmonds B, Hernandez C, Troitzsch KG (eds) Social simulation:
technologies, advances and new discoveries. IGI Publishing, Hershey, pp 43–60
Deffuant G, Neau D, Amblard F, Weisbuch G (2000) Mixing beliefs among interacting agents.
Adv Complex Syst 3(1):87–98
Doran JE (1997) Foreknowledge in artificial societies. In: Conte R, Hegselmann R, Tierna P (eds)
Simulating social phenomena. Lecture notes in economics and mathematical systems, 456.
Dunbar RIM (1998) The social brain hypothesis. Evol Anthropol 6(5):178–190
Edmonds B (1999) The pragmatic roots of context. In: Bouquet P, Serafini L, Brezillon P,
Benerecetti M, Castellani F (eds) Modelling and using context. Second international and
interdisciplinary conference, CONTEXT’99, Trento, Italy, 9–11 Sept 1999, Proceedings.
Lecture notes in artificial intelligence, 1688. Springer, Berlin, pp 119–132, http://cfpm.org/
cpmrep52.html
Edmonds B (2001) The use of models – making MABS actually work. In: Moss S, Davidsson P
(eds) Multi agent based simulation. Lecture notes in artificial intelligence, 1979. Springer,
Berlin pp 15–32
Edmonds B (2007) The practical modelling of context-dependent causal processes – a recasting of
Robert Rosen’s thought. Chem Biodivers 4(1):2386–2395
Edmonds B (2010) Context and social simulation. Paper presented at the IV edition of epistemo-
logical perspectives on simulation (EPOS2010), Hamburg, 23–25 June 2010, http://cfpm.org/
cpmrep210.html
Edmonds B, Hales D (2003) Replication, replication and replication – some hard lessons from
model alignment. J Artif Soc Soc Simul 6(4), http://jasss.soc.surrey.ac.uk/6/4/11.html
Edmonds B, Hales D (2005) Computational simulation as theoretical experiment. J Math Sociol
29(3):209–232
Edmonds B, Moss S (2005) From KISS to KIDS – an ‘anti-simplistic’ modelling approach. In:
Davidsson P et al (eds) Multi agent based simulation 2004. Lecture notes in artificial intelli-
gence, 3415. Springer, Berlin, pp 130–144
Epstein J (2008) Why model? J Artif Soc Soc Simul 11(4). http://jasss.soc.surrey.ac.uk/11/4/12.html
Etienne M (2003) SYLVOPAST: a multiple target role-playing game to assess negotiation
processes in sylvopastoral management planning. J Artif Soc Soc Simul 6(2), http://jasss.soc.
Galam S (1997) Rational group decision making: a random field Ising model at T ¼0. Physica A
238:66–80
Galán JM et al (2009) Errors and artefacts in agent-based modelling. J Artif Soc Soc Simul 12(1),
Gilbert N, Troitzsch K (2005) Simulation for the social scientist, 2nd edn. Open University Press,
New York
Gode DK, Sunder S (1993) Allocative efficiency of markets with zero intelligence traders: markets
as a partial substitute for individual rationality. J Polit Econ 110:119–137
Grimm V et al (2005) Pattern-oriented modeling of agent-based complex systems: lessons from
ecology. Science 310(5750):987–991
models. Ecol Model 198(1–2):115–126
Grimm V, Polhill JG, Touza J (2013) Documenting social simulation models: the ODD protocol as
a standard. Chapter 7 in this volume
Guba EG, Lincoln YS (1994) Competing paradigms in qualitative research. In: Denzin NK,
Lincoln YS (eds) Handbook of qualitative research. Sage, London, pp 105–117
Guimera R, Uzzi B, Spiro J, Amaral LA (2005) Team assembly mechanisms determine collabora-
tion network structure and team performance. Science 308(5722):697–702
Hales D (2013) Distributed computer systems. Chapter 21 in this volume
Heath B, Hill R, Ciarallo F (2009) A survey of agent-based modeling practices (January 1998 to
July 2008). J Artif Soc Soc Simul 12(9), http://jasss.soc.surrey.ac.uk/12/4/9.html
Izquierdo LR (2008) Advancing learning and evolutionary game theory with an application to
social dilemmas. PhD thesis, Manchester Metropolitan University, http://cfpm.org/theses/
luisizquierdo/
Izquierdo SS, Izquierdo LR (2006) On the structural robustness of evolutionary models of
cooperation. In: Corchado E, Yin H, Botti VJ, Fyfe C (eds) Intelligent data engineering and
automated learning – IDEAL 2006, 7th international conference, Burgos, Spain, 20–23 Sept
2006, Proceedings. Lecture notes in computer science 4224. Springer, Berlin, pp 172–182
Janssen MA (2009) Understanding artificial Anasazi. J Artif Soc Soc Simul 12(4), http://jasss.soc.
Kahn K, Noble H (2009) The modelling4All project – a web-based modelling tool embedded in
Web 2.0. In: Dalle O et al (eds) Proceedings of SIMUTools ’09, 2nd international conference
on simulation tools and techniques, Rome, 2–6 Mar 2009, ICST, Brussels, Article 50
Kephart JO, Greenwald AR (2002) Shopbot economics. Auton Agents Multi-Agent Syst 5
(3):255–287
Kirman A (2011) Complex economics: individual and collective rationality. Routledge, London
Kirman A, Moulet S (2008) Impact de l’organisation du marché: Comparaison de la négociation de
gré à gré et des enchères descendants. Working papers, halshs-00349034, HAL, Centre pour la
communication scientifique directe, http://halshs.archives-ouvertes.fr/docs/00/34/90/34/PDF/
DT2008-56.pdf
Klein G (1998) Sources of power: how people make decisions. MIT Press, Cambridge
Kohler T (2009) Generative archaeology: how even really simple models can help in understand-
ing the past. Invited talk at 6th conference of the European social simulation association,
University of Surrey, Guildford
Kohler TA, Gumerman GJ, Reynolds RG (2005) Simulating ancient societies. Sci Am 293
(1):76–82
Kohler TA, Varien MD, Wright A, Kuckelman KA (2008) Mesa Verde migrations: new archaeo-
logical research and computer simulation suggest why ancestral Puebloans deserted the
northern Southwest United States. Am Sci 96:146–153
Kuhn T (1962) The structure of scientific revolutions. University of Chicago Press, Chicago
Lakatos I, Musgrave A (eds) (1970) Criticism and the growth of knowledge. Cambridge Univer-
sity Press, Cambridge
LeBaron B (2006) Agent-based computational finance. In: Tesfatsion L, Judd KL (eds) Handbook
of computational economics, vol 2. North-Holland, pp 1187–1232
Lincoln YS, Guba EG (1985) Naturalistic inquiry. Sage, Beverly Hills
Lorenz J (2007) Continuous opinion dynamics under bounded confidence: a survey. Int J Mod
Phys C 18(12):1819–1838
Lucas P (2013) Conventional social behaviour amongst microfinance clients. PhD thesis, Centre
for Policy Modelling, Manchester Metropolitan University
Lucas P (2011) Usefulness of simulating social phenomena: evidence. AI Soc 26(4):355–362
Malthus T (1798) An essay on the principle of population. Johnson, London, Transcript available
online at http://socserv2.mcmaster.ca/~econ/ugcm/3ll3/malthus/popu.txt
Marks RE (2007) Validating simulation models: a general framework and four applied examples.
Comput Econ 30(3):265–290
Matthews RB (2006) The people and landscape model (PALM): towards full integration of human
decision-making and biophysical simulation models. Ecol Model 194:329–343
Meadows DH, Meadows D, Randers J, Behrens WWIII (1972) The Limits to Growth: a report for
the Club of Rome’s project on the predicament of mankind. Universe Books, New York
Moss S (1998) Critical incident management: an empirically derived computational model. J Artif
Moss S (1999) Relevance, realism and rigour: a third way for social and economic research. Report
no. CPM-99-56. Centre for Policy Modelling, Manchester Metropolitan University,
Manchester, http://cfpm.org/cpmrep56.html
Newell A (1990) Unified theories of cognition. Harvard University Press, Cambridge
Nicolaisen J, Petrov V, Tesfatsion L (2001) Market power and efficiency in a computational
electricity market with discriminatory double-auction pricing. IEEE Trans Evol Comput
5(5):504–523
Phelps S, McBurney P, Parsons S, Sklar E (2002) Co-evolutionary auction mechanism design: a
preliminary report. In: Padget J, Shehory O, Parkes D, Sadeh N, Walsh WE (eds) Agent-
mediated electronic commerce IV, designing mechanisms and systems. Lecture notes in
computer science, 2531. Springer, Berlin, pp 193–213
PNAS (2002) Colloquium papers. Proceedings of the National Academy of Sciences, 99(s3),
http://www.pnas.org/content/99/suppl.3.toc#ColloquiumPaper
Polhill JG, Gotts NM, Law ANR (2001) Imitative versus non-imitative strategies in a land use
simulation. Cybernet Syst 32(1–2):285–307
Polhill JG, Parker D, Brown D, Grimm V (2008) Using the ODD protocol for describing three
agent-based social simulation models of land-use change. J Artif Soc Soc Simul 11(2), http://
Popper K (1963) Conjectures and refutations: the growth of scientific knowledge. Routledge,
London
Riolo RL, Cohen MD, Axelrod R (2001) Evolution of cooperation without reciprocity. Nature
411:441–443
Roberts G, Sherratt TN (2002) Does similarity breed cooperation? Nature 418:499–500
Rouchier J (2001) Est-il possible d’utiliser une définition positive de la confiance dans les
interactions entre agents? Paper presented at Colloque Interactions, Toulouse, May 2001
Rouchier J (2013) Agent-based simulation as a useful tool for the study of markets. Chapter 23 in
this volume
Rouchier J, Thoyer S (2006) Votes and lobbying in the European decision-making process:
application to the European regulation on GMO release. J Artif Soc Soc Simul 9(3), http://
Rouchier J, Bousquet F, Requier-Desjardins M, Antona M, Econ J (2001) A multi-agent model for
describing transhumance in North Cameroon: comparison of different rationality to develop a
routine. J Econ Dyn Control 25(3–4):527–559
Saqalli M, Bielders CL, Gerard B, Defourny P (2010) Simulating rural environmentally and socio-
economically constrained multi-activity and multi-decision societies in a low-data context: a
challenge through empirical agent-based modeling. J Artif Soc Soc Simul 13(2), http://jasss.
Schelling T (1969) Models of segregation. Am Econ Rev 59(2):488–493
Schelling T (1971) Dynamic models of segregation. J Math Sociol 1:143–186
Scherer A, McLean A (2002) Mathematical models of vaccination. Br Med Bull 62(1):187–199
Snijders TAB, Steglich CEG, van de Bunt GG (2010) Introduction to actor-based models for
network dynamics. Soc Networks 32:44–60
Sun R (2005) Theoretical status of computational cognitive modelling. Technical report, Cogni-
tive Science Department, Rensselaer Polytechnic Institute, Troy, New York
Taylor R, Takama T, Taylor A, Ziervogel G (2009) Participatory validation of agent-based models
to support policy dialogue. In: Gilbert N, Edmonds B (eds) Proceedings of the 6th European
social simulation association conference, University of Surrey, Guildford
Terán O, Alvarez J, Ablan M, Jaimes M (2007) Characterising emergence of landowners in a forest
reserve. J Artif Soc Soc Simul 10(3), http://jasss.soc.surrey.ac.uk/10/3/6.html
Vermeulen PJ, de Jongh DCJ (1976) Parameter sensitivity of the ‘Limits to Growth’ world model.
Appl Math Model 1(1):29–32
von Randow G (2003) When the centre becomes radical. J Artif Soc Soc Simul 6(1), http://jasss.
White DR (1999) Controlled simulation of marriage systems. J Artif Soc Soc Simul 2(3), http://
Yang C, Kurahashi S, Kurahashi K, Ono I, Terano T (2009) Agent-based simulation on women’s
role in a family line on civil service examination in Chinese history. J Artif Soc Soc Simul 12
(2):5, http://jasss.soc.surrey.ac.uk/12/2/5.html
Young J, Mendizabal E (2009) Helping researchers become policy entrepreneurs – how to develop
engagement strategies for evidence-based policy-making. Briefing paper, Overseas Develop-
ment Institute, London
Younger S (2005) Violence and revenge in Egalitarian societies. J Artif Soc Soc Simul 8(4), http://
Index
A Asymptotic dynamics, 240, 252

Absorbing class, 254–259, 261, 262 Auction, 620, 621, 624, 637–645
Absorbing state, 240, 255, 257, 262 Authority, 667–704
Abstraction, 713, 717, 733 Autonomous (agent), 274, 280, 286
Accessibility, 254 Autonomy, 83, 90, 91
Accessory assumptions, 97, 107–110
Adaptation, 432–435, 446, 450
Afghanistan, 667–704 B
Agent, 566, 568, 569, 572–574, 577, 578 Back propagation, 436, 444–446
approach, 74, 81–92 Bargaining, 620, 621, 624–628, 646
organisations, 488, 543, 546 Bayesian updating, 451
strong notion, 84 BDI. See Belief-Desire-Intention (BDI)
weak notion, 82, 83 Behaviour, 501, 504–506, 508, 509, 511, 513,
Agent-based model/modelling, 96, 97, 515, 516, 518, 520, 522, 523, 527, 530,
101–106, 113, 366, 370, 373, 384 533, 534
Agent-based simulation, 443, 460 Behaviour modelling, 653–656, 661, 662
Agent-based simulation models (ABM) Belief-Desire-Intention (BDI), 569
tools, 55 Belief functions theory, 318
Agent-orientated design, 569 Biological models, 565, 566, 569
Aggregation processes, 583 BitTorrent, 569–576
Agriculture, 500, 501, 510, 513, 520–521, Boundary object, 204
532, 533 Bounded rationality, 299, 327, 618, 622, 628
Altruism, 598
Analogical model, 6
Ants, 582, 583, 587, 588 C
Aperiodic communicating class, 256 Case study, 49
Aperiodic state, 255–256 Causal explanation, 277, 278, 281
Applications, 8, 10 Causality, 175, 179, 187–192
Arm chair theorising, 4, 7 Causal pathways, 175, 190
Artefact, 95–113 Cellular automata, 504
Artificial life, 274 Church membership, 485
Artificial stock market, 480 Classifier system, 461, 462, 471, 472, 476,
Aspirations, 436, 439, 441 478–480, 492
Assessment, 710, 712, 731–732, 741–743 Closed communicating class, 254–257,
Assumptions, 41, 43–45, 47, 51, 52, 711, 712, 261, 262
714–731, 733, 736–741, 744 Cluster analysis, 180–181

Understanding Complex Systems, DOI 10.1007/978-3-540-93813-2,
750 Index
Clustering, 179, 180 DomWorld, 592–595, 597–601, 603,

Clustering coefficient, 410, 411, 415, 416, 604, 606
418, 419 Dynamic networks, 568, 574
Cognitive modelling, 6, 9, 366–367, 374, 375
Cognitive sophistication, 582
Collective choice, 587 E
Collective decision making, 293, 295 Economics, 6, 9, 10
Collectivism, 275 Eigenvector methods, 180
Communicating class, 254–256, 262 El Farol Bar problem, 302–303, 305, 330
Communication (of states), 254 Ellsberg’s paradox, 309–310, 330, 331
Companion modelling, 502, 516, 521, 526 Emergence, 137, 166, 169, 175, 275, 276,
Competition, 582, 583, 591, 594, 597, 598, 606 283, 285–287, 564, 565, 567–569
Compositional design, 58–67, 82 Entropy (statistics), 177, 178
Computer model/modelling, 100, 235–267 Environment, 499–534, 652, 654
Computer scientist, 102–106, 109–111 Environment model, 31, 32, 34
Computer simulation, 96, 97, 100, 101, 106, Epidemiology, 501, 508, 510, 523–525, 533
236–238, 242–250, 259–267 Epistemological perspectives, 138,
Conceptual model, 47 140, 169
Consolidation, 40–42, 45, 46, 50–51, 54, 55 Equifinality, 175, 190
Constraint satisfaction, 318, 321–323, 327, 333 Ergodicity, 257, 258, 261
Contingency, 229 Error, 95–113
Control, 652, 653, 662 Evaluation, 733
Coolworld, 245–249 Evidence, 712, 713, 716–721, 727–730,
Coordination, 300, 306–309, 329, 500, 505, 732–734, 737–743
507, 510, 513–516, 519, 528, 582–591 Evidence-driven agent-based social simulation,
Core assumption, 107, 108, 110 682–687
Correlation, 174, 176, 178, 180, 188–189 Evidence theory, 318, 322–327, 333
Counter-example, 721, 723, 724 Evolution, 432–435, 437–439, 450, 451,
Covering law approach, 277–279, 281, 285 565–569, 572, 573
Evolutionary models, 458, 459, 470, 476, 482,
486, 488–491
D Experimentation, 181, 189
Data, 710, 712–715, 717, 720, 725, 727–735, Explanation, 274–287, 712, 714
737, 741–744 Explicit, 402–404
Decision making, 505, 506, 511, 512, 515, 518, Exploration, 40–42, 45, 46, 50, 51, 54,
520–522, 527, 531 724–726, 739
Decomposition theorem, 255, 256 Exploration-exploitation, 618, 619
Deduction, 238, 242, 243, 266 External restrictions, 652
Degree distribution, 409, 411, 412, 418
Density, 419, 420, 423, 424, 426
Descriptive approach, 4 F
Descriptive models, 7, 10 Fieldwork, 742, 743
Design, 57–67, 74, 78–85, 92, 563–566, 568, Finance, 639
569, 576, 578 Fission-fusion system, 584
Design patterns, 575–578 Flow, 652, 654, 660
Diffusion, 177, 180 Forestry, 517–518, 532
Distributed artificial intelligence, 274 Formal approaches, 92
Distributed computing, 563–578 Formal language, 98, 99, 109
Documentation, 44, 45, 47, 48, 50–52 Formal models, 236–240, 242–244, 263, 266,
Documentation protocol, 118, 130 542, 544, 545, 550
Dominance Formal representation, 6
hierarchy, 582, 591, 594, 603–605, 608 Formal system, 238, 242, 244
relationships, 599–602, 605, 606, 608 Formation of groups, 583, 584, 609
style, 582, 595–598 Freight, 657–659, 662, 663
Index 751
Function, 237–242, 252, 253, 257–259, Individual-based model, 582, 609

262, 265 Individual decision making, 314, 652
Functionalism, 462 Individual model, 28, 32, 34
Induction, 243, 266
Inference engine, 100
G Influence games, 296, 302–303, 327, 330
GAM. See Geographical analysis machine Information dissemination, 215, 216
(GAM) Initial conditions, 245, 247–249, 252, 253,
Game of Chicken, 296, 300–301, 329 257–263, 266
Game theory, 293–295, 300, 327–329 Intelligent agents, 274
Generalisation, 243, 264, 266, 267 Intensity of aggression, 592, 593, 595, 596,
Genetic algorithms, 438, 442, 456, 461–474, 599, 602, 603
476–480 Intentions, 652, 655
Genetic programming, 461, 470, 471, 473–476, Interactions, 502–505, 507–508, 513, 514, 516,
478–482 517, 519, 521–526, 530, 533
Geographical analysis machine (GAM), 179 model, 31, 32, 34
Geographically weighted regression topology, 402
(GWR), 178 Inter-agent processes, 339, 341, 351, 352, 360
Geographic information systems (GIS), 504, Internal conflict, 566
505, 513, 518, 519, 526, 530–532 Internalisation, 338, 339, 341, 343, 347–360
Gradient of the hierarchy, 597 Interpretation, 273, 275, 276
Great tits, 604–606 Intersections, 654, 655
Grid, 403, 406–408 Intra-agent processes, 339, 351, 360
Groom, 595, 598–599 IPD. See Iterated prisoner’s dilemma (IPD)
Group decision support system, 198, 208 Irreducibility, 255, 261
Groups, 59–61, 67, 69–70, 72–80, 87–88, Iterated prisoner’s dilemma (IPD), 280
90, 92
Group selection, 571–572
GWR. See Geographically weighted regression K
(GWR) Knowledge specification, 65
Kuhn, T.S., 7
H
Hebbian learning, 447 L
Heterogeneity, 502, 522, 525–527, 534 Land use/land cover change, 501, 504,
Heuristics, 568, 577 510–514, 517, 518, 524, 532
Hierarchical differentiation, 596, 597, 599, 605 Lattice, 406–411
Hierarchy, 576–578 Leader, 590
History of social simulation, 18 Learning, 618–621, 623–624, 626, 629–631,
HSS. See Human social systems (HSS) 633, 635–639, 642–646
Human in the loop, 576 Limiting distribution, 257, 258, 261, 262
Human social systems (HSS), 565–568 Livestock, 501, 510, 511, 522–523, 532, 533
Hypotheses, 582, 609 Loyalty, 619, 627–630
I M
Illustration, 721–723 Market, 617–646
Image, 368–370, 374–382, 384, 386–394 Market design, 620, 643–644
Imitation, 652 Markov chain, 239, 247–263, 266, 267
Implementation, 29, 33, 34, 40, 46, 59, 60, 68, Markov property, 250
74, 77, 82, 91, 92 Mathematical analysis, 236, 237, 242–245,
Implicit, 402–404 249–267
Individual-and agent-based modelling, Mathematical tractability, 236, 260,
127–128 263–264, 266
752 Index
Mechanisms, 8–10, 278, 279, 281, 283–286 P

Mechanistic explanation, 278, 281, 284, Participatory, 718, 744
285, 287 Participatory modelling, 198, 199, 202, 206,
Methodological individualism, 282, 285, 210, 218, 228, 229
286, 342 Passengers, 652, 657, 659–663
Methodology, 8–9 Path marking, 582
Microanalytical simulation, 14, 15 Pattern, 712, 717, 739
Micro/macro, 710, 713, 714, 720, 723, 735 Patternation, 176
Minority game, 302, 303, 305–309, 330 Pattern recognition, 176, 177, 184
Mixed strategy, 295–297, 301 Peer-to-peer (P2P), 564, 569–575, 578
Model checking, 41, 46, 48, 50, 54 Periodic behaviour, 179
Modeller, 96, 98, 102–106, 109–111 Periodic communicating class, 255–256
Modelling, 96–106, 108, 109, 111–113, Periodic state, 255
712–714, 716–743 Personality style, 582, 605, 606
Modelling roles, 103, 109 Phase maps, 184–186
Model purpose, 41–44, 47, 48, 53 Pheromones, 583, 587, 588, 592, 608
Multi-agent modeling, 704 Pheromone trails, 587, 588
Multi-agent systems, 58–67, 85 Power, 567, 576–578, 667–704
Multidimensional scaling, 180 Predator, 590, 591, 594, 595
Multi-level systems, 79 Preference reversal, 312–315, 331
Multiple realization, 281, 284, 285 Preferences, 618, 622, 625, 631, 634, 640
Preferential attachment, 412, 413
Price setting, 479, 489
N Primates, 582, 584, 592, 593, 595, 597–599,
Narrative, 293–333 602, 603, 606, 609, 610
Narrative evidence, 5 Prisoner’s dilemma, 297–298, 302, 329, 569,
Nash equilibrium, 295, 297–299, 302 570, 573, 577
Natural resources management, 229 Probability distribution, 237, 240–242,
Neighbourhood effects, 178 245–249, 258, 260, 262, 263,
Netlogo, 239, 245 265, 266
Networks, 652–656, 658, 659, 661 Problem solving, 5, 6
Neural networks, 442–444, 446 Procedural rationality, 5
Neural networks, unsupervised, 318–321, Process, 711, 712, 715–719, 721–725, 727,
327, 333 729–730, 732–737, 739, 740, 742, 743
Neyman-Scott process, 421 Process specification, 59, 63–65, 74
Non-darwinian evolution, 565–566 Programmer, 96, 104–107, 109–112
Normal science, 7 Proof, 716, 721, 724
Normative agent-based models, 345, 361 Proximity, 598
Normative architectures, 353, 355–360 Pseudo-random number generator, 237, 238,
Norms, 335–361 240, 241
Pure strategy, 295–297, 300, 301
Purposes, 24–26, 34
O
Occupancy distribution, 258, 259, 261, 262
ODD. See Overview, design concepts, details R
(ODD) Random graph, 406, 409–411, 415, 416,
Ontology, 65, 85 419, 426
Open systems, 564, 566, 568, 569, 572, 576 Random seed, 240
Optimisation, 180 Random variable, 240, 241, 246
Optimisers, 5 Random walk, 251–256, 258
Organisational change, 544, 550 Rationality, 432, 436, 441
Organisations, 58, 60, 67–82 Realization, 281, 283–285
Overview, design concepts, details (ODD), Recurrence plots, 177, 183, 184, 186–187
118, 503, 509 Recurrent state, 255, 256
Index 753
Reduction, 275, 282, 287 models, 568

Regression modelling, 188, 189 network, 401–426
Re-implementation, 111, 112 norms, 567
Reinforcement, 432, 435–437, 440–442, process, 98–100
450, 451 relationships, 582
Renewable resources, 500, 502, 504, 510, 513, simulation, 370, 384, 387, 669, 670,
527, 529 682–687, 694, 703
Replication, 118, 121, 122, 128–130, 715 Socially-inspired computing, 569–575
Replicator dynamics, 437, 438 Sociophysics, 20
Representation, 717–719, 721, 723, 733 Spatial autocorrelation, 178, 189, 193
Representative agents, 84–85 Spatially explicit, 582
Reputation, 365–395, 630–631 Spatial structure, 402, 403, 420–426,
Requirement refinement, 78–81 594–596, 598
Resilience, 502, 522 Specification, 58, 59, 63–66, 68, 69, 71, 73, 74,
Revolutionary science, 7 80, 92
Robust networks, 568 Stakeholder involvement, 198, 213, 228, 230
Roles, 58, 60, 66–87 Stakeholders, 5, 7, 9, 652, 655, 658, 718,
Role theory, 340–342, 346, 347 734–735, 742, 743
Standard error, 246–248
State space, 250, 252, 255–257, 261
S Statistic, 240, 241, 243, 245, 249, 253,
Sammon mapping, 177, 180 258–263
Sampling, 257 Statistical confidence, 243, 266
Satisficers, 5 Steep hierarchy, 595, 603
Satisficing, 437 Stochastic, 437–441, 443, 444
Scale-free network, 405, 406, 411–414, Stochastic process, 237, 240–242, 254, 256, 267
418, 419 Strategy of attack, 603, 604
Segregation, 572–575, 590 Structured program theorem, 239
Self-*, 564, 565, 568, 578 Sub-goals, 5
Self-organisation, 564, 565, 568, 578, 582, 584, Substantive rationality, 4, 5, 7
587, 591, 592, 610 Supply chains, 620, 631, 636–637, 657,
Self-reinforcing, 592, 593, 606, 607, 609 658, 663
Semi-variograms, 178 Surface validity, 4
Sex ratio, 602 Swarm, 587–591
Shape of a school, 591 Symbolic system, 97–100
Simon, H., 4–7 System dynamics, 14, 19, 20
Simulation, 95–113
platforms, 508, 509, 516, 528
scenario, 502, 503, 509–510, 512, 513, 516, T
519, 521, 523, 525 Task division, 582, 606, 610
Simulation development process Thematician, 102–104, 106, 109, 110
computerized models, 136–138 Theory-like models, 7
conceptual models, 137, 140, 150 Theory testing, 502, 529
validation, 135–169 Time-homogeneous Markov chain, 250–251
verification, 136–141, 151, 154, 159, 163, Torus, 406–408, 420
166, 169 Tractability, 7
Slovic’s paradox, 312–314 Traffic, 652–656, 658, 659, 661–663
Small world network, 410–411, 415, 416 Transient dynamics, 252
Social Transient state, 255, 257
clock, 608 Transition matrix, 250, 251, 253, 256,
complexity, 609, 610, 674 257, 260
context, 719 Transportation, 652, 657–663
design, 730, 732 Trust, 368, 370–374, 381, 395, 567, 568
754 Index
U Verification, 97, 106–108, 113

Understanding, 273–287, 709–744 assertions, 142, 146, 147, 149
Users, 24–26, 33, 34 defensive programming, 146–149
Utility maximisation, 312 dynamic methods, 141–142
encapsulation, 142–143
good programming practices, 142–146, 154
V invariant, 147–149
Validation, 59, 82, 97, 106–109, 113, 510, model alignment, 139, 149–154,
529, 730, 737, 740 163, 166
context-specific models, 167, 168 model equivalence, 151–152
cross-element validity, 164 modularity, 142–143
cross-model validation, 162 replication, 149–154, 166
cross-sectional validity, 162–163 simulation failure, 141
event validity, 161 simulation fault, 141, 142
extreme condition tests, 161 software reuse, 144–146
face validity, 160 static methods, 141–142
historical validity, 160 Virtual experiment, 277, 279
KIDS, 167 Visual debugging, 182
KISS, 165, 167 Visualisation, 181–187
modelling strategies, 136, 155, 165–169 Visualisation, dynamic, 49, 53
prediction, 154, 156–159
retrodiction, 157–160, 164, 167
sensitivity analysis, 161–164, 169 W
structural similarity, 157, 159 Water, 501, 508, 510, 513–518, 532, 533
subjunctive models, 166 Wavelets, 177, 179, 180
turing tests, 160 Wild disjunction, 284
types of validity, 154, 155, 157–159 Wildlife, 501, 510, 517–520, 532, 533
Verbal argumentation, 98 Winner-loser effect, 594

Simulating Social Complexity - A Handbook-Springer-Verlag Berlin Heidelberg (2013)

Uploaded by

Copyright:

Available Formats

Simulating Social Complexity - A Handbook-Springer-Verlag Berlin Heidelberg (2013)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Simulating Social Complexity - A Handbook-Springer-Verlag Berlin Heidelberg (2013)

Uploaded by

Copyright:

Available Formats

Understanding Complex Systems

Editorial and Programme Advisory Board

Future scientific and technological developments in many fields will necessarily

For further volumes:

ISSN 1860-0832 ISSN 1860-0840 (electronic)

# Springer-Verlag Berlin Heidelberg 2013

Printed on acid-free paper

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

Part I Introductory Material

1 Introduction to the Handbook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

4 Informal Approaches to Developing Simulation Models . . . . . . . . . 39

10 Participatory Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

Part III Mechanisms

13 Utility, Games, and Narratives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293

19 Agent-Based Modelling and Simulation Applied to

22 Simulating Complexity of Animal Social Behaviour . . . . . . . . . . . . 581

Bruce Edmonds and Ruth Meyer

1.1 Simulating Social Complexity

B. Edmonds (*) • R. Meyer

B. Edmonds and R. Meyer (eds.), Simulating Social Complexity, 3

1.2 The Context, Going Back to Herbert Simon

This handbook is in memory of Herbert Simon, since he initiated several key

some adaptive or learning ability to individuals and allowing the individuals to

1.3 The Structure of the Handbook

1.3.1 Introductory Section

1.3.2 Methodology Section

visualising simulation results, which is vital if the programmer or a wider audience is

1.3.3 Mechanisms Section

1.3.4 Applications Section

Simon HA (1947) Administrative behavior: a study of decision-making processes in administrative

Why Read This Chapter? To understand the historical context of simulation in

2.2 The First Two Decades

K.G. Troitzsch (*)

B. Edmonds and R. Meyer (eds.), Simulating Social Complexity, 13

in formalising social science, was an early adopter of computer-assisted methods of

probabilities empirically estimated from official statistics. Transitions represent

of the models, perhaps because of dissatisfaction with the plausibility of their

2.3 Computer Simulation in Its Own Right

early forerunner of current simulation models of negotiation and decision making

automaton until they find themselves in an agreeable neighbourhood) the clustering

Table 2.1 Overview of important approaches to computational social science

Abelson RP, Bernstein A (1963) A computer simulation of community referendum controversies.

Paul Davidsson and Harko Verhagen

B. Edmonds and R. Meyer (eds.), Simulating Social Complexity, 23

3.2 Purposes of Simulation

We can identify a number of distinct purposes of simulation. In general terms,

– Education, where simulation is used to explain or illustrate a phenomenon and

3.3 Types of Systems Simulated

It is possible to categorize the systems being simulated:

– Organizations, which we here define as structures of persons related to each

3.4.2 Interaction Between Individuals

In dynamic micro-simulation simulated individuals are considered in isolation

the interaction topology is limited to the individuals immediately surrounding an

3.4.3 The Environment

The state of the environment is usually represented by a set of (global) parameters,

3.4.4 Factors to Consider When Choosing a Model

In contrast to some of the more traditional approaches, such as system dynamics,

Focus Aspect Options

Grimm V et al (2006) A standard protocol for describing individual-based and agent-based

E. Norling () • B. Edmonds () • R. Meyer