Springer Complexity: Editorial and Programme Advisory Board

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
computational 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
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: J.A. Scott 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
Dirk Helbing
Editor
Social Self-Organization
Agent-Based Simulations and Experiments
to Study Emergent Social Behavior
123
Editor
Dirk Helbing
ETH Zurich, CLU E1
Chair of Sociology, in particular of Modeling and Simulation
Clausiusstrasse 50
8092 Zurich
Switzerland
ISSN 1860-0832 ISSN 1860-0840 (electronic)

ISBN 978-3-642-24003-4 ISBN 978-3-642-24004-1 (eBook)
DOI 10.1007/978-3-642-24004-1
Springer Heidelberg New York Dordrecht London
Library of Congress Control Number: 2012936485
c Springer-Verlag Berlin Heidelberg 2012

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)

Preface
Social systems are the most complex systems we know. They are even more complex
than physical or biological systems. Their complexity results not only from multiple
interactions of individuals, but also from the complexity of cognitive systems.
I, therefore, agree with August Comte that sociology is the queen of science, the
ultimate scientific challenge. I furthermore believe that this field will soon be one of
the most dynamic scientific areas, not only for its interesting fundamental questions,
but also because of the many practical problems humanity is facing in the twenty-
first century.
The crucial question is, how one can make substantial progress in a field as
complicated and multi-faceted as the social sciences. There are certainly different
possibilities, as Chap. 1 discusses. Nevertheless, it seems that many characteristic
features of complex social systems can be understood from simple models of social
interactions, and I am convinced that a number of challenging scientific puzzles
can be solved, using concepts from complexity theory, including self-organization,
coevolution, and emergence. Agent-based computational models and behavioral
experiments can reveal the mechanisms underlying such phenomena, and the role
that different factors play for them.
Complex systems often display a counter-intuitive behavior. For example, as
will be shown in this book, the same kinds of social interactions can lead to
opposite conclusions, when interactions occur with neighbors, friends, colleagues,
or business partners rather than with average interaction partners (see Chaps. 7
and 8). Therefore, a simple nonlinear model may explain phenomena, which even
complicated linear models may fail to reproduce. Nonlinear models are expected to
shed new light on such social phenomena. They may even lead to a paradigm shift
in the way we interpret society.
While part of this book is a compilation of recently published papers, some
chapters present variants of previous work or new materials, for example, on the
technique and future of agent-based computer simulation and on coordination games
in networks (a subject that is interesting to understand the competitive spreading
of innovations). The chapters do not need to be read in sequential order, but the
organization of this book has a clear logic:
v
vi Preface
• Chapter 1 discusses the issue of how to describe social systems. It highlights

different traditions in the social sciences and the respective advantages and
disadvantages of these approaches, stressing their complementary character. The
focus of this book, however, will be on simple models of social interactions,
explaining various aspects of social self-organization.
• Chapter 2 discusses the method of agent-based computational modeling, how to
do it right, and what are future perspectives of this approach. A particular focus is
put on the question, how to avoid mistakes and how to derive meaningful results.
• Chapter 3 demonstrates the concept of self-organization for the example of
pedestrian crowds. We analyze the spontaneous outbreak of social order under
everyday conditions and its breakdown at extreme densities. The chapter illus-
trates how fundamental research can lead to useful models that enable an
understanding of macroscopic outcomes of social interactions. It also summa-
rizes related empirical and experimental work as well as practical applications.
• Chapter 4 provides a second example of agent-based modeling, namely a
continuous opinion formation model. It shows why previous models did not solve
the puzzle why one finds global diversity in spite of local convergence. Strikingly,
it is the tendency of individualization which promotes pluralism through the self-
organization of groups.
• Chapter 5 turns the attention from mobility in opinion space to mobility in
geographical space. Assuming social interactions with neighboring locations,
where the outcome of these interactions is quantified by “payoffs” as this is
common in game theory, we find the self-organization of spatiotemporal patterns,
when success-driven mobility to neighboring locations occurs. Even when
starting with a uniform distribution in space, we observe interesting segregation
phenomena and different kinds of agglomeration phenomena, depending on the
respective payoff structure. These come about when success-driven mobility
increases local differences and thereby destabilizes a homogeneous distribution
in space.
• Chapter 6 focuses on the problem of cooperation in social dilemma situations,
where it appears more advantageous to selfish individuals to exploit others than
to cooperate with them. It is discussed in what ways social mechanisms can
effectively change the payoff structure and, thereby, the rules and character of
the “game” individuals are playing. It is shown that different mechanisms such as
repeated interactions, reputation effects, or social networking can imply different
routes to cooperation. The chapter also provides a classification of different kinds
of transitions to cooperation and shows that adaptive group pressure can promote
cooperation in the prisoner’s dilemma game even without changing the properties
of its equilibrium solutions.
• Chapter 7 combines the elements of Chaps. 5 and 6, i.e., it studies individuals
facing social dilemma situations in space, considering success-driven mobility.
While one would think that social cooperation and mobility are unrelated, it
surprisingly turns out that mobility is an important factor supporting human
sociality, and that it promotes a co-evolution of social behavior and social
environment. This model may shed new light on a number of fundamental
Preface vii
questions such the following: Why do similarly behaving people tend to agglom-
erate, e.g., form groups or cities (an observation which is often referred to as
“homophily”)? How are “social milieus” come about and why do they influence
the behavior of people? Why do they persist so long? And why are selfish people
more cooperative in reality than expected? What is the role of fluctuations and
exploratory behavior for the emergence of cooperation?1
• Chapter 8 looks at a further mechanism that has been suggested to promote
cooperation, namely “costly punishment.” It is shown that the consideration of
neighborhood interactions can resolve the so-called second-order free-rider puz-
zle that wonders why people would invest into the punishment of uncooperative
behavior, if they can profit from other people’s sanctioning efforts. The chapter
also suggests that the spreading of morals and double moral behavior can
be understood with concept of evolutionary game theory. The related system
dynamics shows quite a number of surprising features.2
• Chapter 9 studies effects of network interactions and transaction costs in
coordination games. Developing a percolation theoretical description for the
related system dynamics allows one to analytically understand the competitive
spreading of innovations, opinions, or products from a new scientific angle.
Furthermore, we point out that system-wide coordination is a double edge sword.
• Chapter 10 focuses on the implications of heterogeneity in the inclinations
of individuals. For this, we study the interaction of several populations with
incompatible preferences. This implies a large variety of different system
behaviors, such as the outbreak or breakdown of cooperation, the formation
of commonly shared norms, the evolution of subcultures, or the occurrence
of conflicts which may cause “revolutions”. Despite its richness, the model is
simple enough to facilitate an analytical understanding of the possible system
behaviors.3 It would be highly desirable to test the predictions of the model by
behavioral experiments.4
1
The following related paper may interest the reader as well: D. Helbing, W. Yu, and H. Rauhut
(2011) Self-organization and emergence in social systems: Modeling the coevolution of social
environments and cooperative behavior. Journal of Mathematical Sociology 35,177–208.
2
The reader may be interested in the related follow up work as well: D. Helbing, A. Szolnoki,
M. Perc, and G. Szabó (2010) Punish, but not too hard: how costly punishment spreads in the
spatial public goods game. New Journal of Physics 12, 083005; D. Helbing, A. Szolnoki, M. Perc,
and G. Szabó (2010) Defector-accelerated cooperativeness and punishment in public goods games
with mutations. Physical Review E 81(5), 057104.
3
D. Helbing and A. Johansson (2010) Evolutionary dynamics of populations with conflicting
interactions: Classification and analytical treatment considering asymmetry and power. Physical
Review E 81, 016112.
4
Recently, the emergence of social norms has been further investigated, and the related study may
be interesting for the reader as well: D. Helbing, W. Yu, K.-D. Opp, and H. Rauhut (2011) The
emergence of homogeneous norms in heterogeneous populations. Santa Fe Working Paper 11-01-
001, see http://www.santafe.edu/media/workingpapers/11-01-001.pdf.
viii Preface
• Chapter 11, therefore, discusses ways in which social experiments should and
could be done in the future, using the help of computers.
• Chapter 12 focuses on the example of a route choice experiment, in which
a sudden transition to turn-taking behavior is found after many interactions,
and it is shown that this transition can be understood with a reinforcement
learning model.
• Chapter 13 analyzes the same route choice game for the case of several
participating players. It is found that, rather than applying probabilistic strategies,
experimental participants develop specialized and almost deterministic strategies
over time. There is strong evidence of social differentiation. Furthermore, we
show how an individualized information system can be developed that supports
social adaptation and avoids a self-defeating prophecy effect.
• Chapter 14 then looks at social systems from a complex systems perspective.
In this way, it analyzes systemic socioeconomic risks and the underlying
mechanisms. In particular, it discusses factors that have probably contributed to
the current financial crisis.
• Chapter 15 addresses the question, how to manage the complexity of social
systems, considering that classical control concepts are known to fail.
• Chapter 16 finally tries to identify fundamental and real-world challenges in
economics, thereby suggesting questions and approaches for future research.5
Although these contributions have originally not been written as chapters of a book,
they are largely complementary to each other and follow a common approach that
tries to understand macroscopic behavioral patterns from interactions between many
individuals. All models pursue an agent-based approach, and for many chapters,
supplementary video animations are available at http://www.soms.ethz.ch . Most
of the models furthermore follow an evolutionary game theory perspective and
are mutually consistent. In fact, the wide spectrum of phenomena that can be
described by evolutionary game theoretical models, ranging from coordination and
cooperation over social norms and conflict up to revolutions, suggests that this
theoretical framework may be flexible enough to form the basis of a future integrated
theory of socioeconomic interactions. With this vision in mind, I hope the reader will
find this book inspiring.
Zurich Dirk Helbing
5
Beyond proposing research questions, the Visioneer White papers “From Social Data Mining to
Forecasting Socio-Economic Crises,” “From Social Simulation to Integrative System Design,” and
“How to Create an Innovation Accelerator” make suggestions how to foster scientific progress in
the socioeconomic sciences, see http://www.visioneer.ethz.ch. These are published in EPJ Special
Topics 195, 1–186 (2011).
Acknowledgments
The research reported in this book and its production were partially supported
by the Future and Emerging Technologies programme FP7-COSI-ICT of the
European Commission through the project QLectives (grant no. 231200), through
the Coordination and Support Action GSDP - Global System Dynamics and Policy
of the European Commission (grant no. 266723), and the FET Flagship Pilot Project
FuturICT (grant no. 284709).
ix
•
Contents
1 Modeling of Socio-Economic Systems . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1

2 Agent-Based Modeling .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 25
3 Self-organization in Pedestrian Crowds.. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 71
4 Opinion Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 101
5 Spatial Self-organization Through Success-Driven Mobility . . . . . . . . . . 115
6 Cooperation in Social Dilemmas . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 131
7 Co-evolution of Social Behavior and Spatial Organization . . . . . . . . . . . . 139
8 Evolution of Moral Behavior . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 153
9 Coordination and Competitive Innovation Spreading
in Social Networks .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 169
10 Heterogeneous Populations: Coexistence, Integration, or Conflict . . . 185
11 Social Experiments and Computing . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 201
12 Learning of Coordinated Behavior . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 211
13 Response to Information .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 239
14 Systemic Risks in Society and Economics. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 261
15 Managing Complexity.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 285
16 Challenges in Economics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 301
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 331
xi
Chapter 1
Modeling of Socio-Economic Systems
1.1 Introduction
When the “father of sociology”, August Comte, came up with the idea of a “social
physics”, he hoped that the puzzles of social systems could be revealed with a
natural science approach [1]. However, progress along these lines was very difficult
and slow. Today, most sociologists do not believe in his positivistic approach
anymore. The question is whether this proves the failure of the positivistic approach
or whether it just shows that social scientists did not use the right methods so far.
After all, social scientists rarely have a background in the natural sciences, while
the positivistic approach has been most successful in fields like physics, chemistry,
or biology.
In fact, recently new scientific communities are developing, and they are
growing quickly. They call themselves socio-physicists, mathematical sociologists,
computational social scientists, agent-based modelers, complexity or network sci-
entists. Researchers from the social sciences, physics, computer science, biology,
mathematics, and artificial intelligence research are addressing the challenges of
social and economic systems with mathematical or computational models and lab
or web experiments. Will they end up with resignation in view of the complexity of
social and economic systems, or will they manage to push our knowledge of social
systems considerably beyond what was imaginable even a decade ago? Will August
Comte’s vision of sociology as “the queen of the sciences” [2] finally become true?
My own judgement is that it is less hopeless to develop mathematical models for
social systems than most social scientists usually think, but more difficult than most
natural scientists imagine. The crucial question is, how substantial progress in a field
as complicated and multi-faceted as the social sciences can be made, and how the
current obstacles can be overcome? Moreover, what are these obstacles, after all?
The current contribution tries to make the controversial issues better understandable

This chapter reprints a previous publication to be cited as: D. Helbing, Pluralistic Modeling of
Complex System. Science and Culture 76(9/10), 399–417 (2010).
D. Helbing (ed.), Social Self-Organization, Understanding Complex Systems, 1

DOI 10.1007/978-3-642-24004-1 1, © Springer-Verlag Berlin Heidelberg 2012
2 1 Modeling of Socio-Economic Systems
to scientific communities with different approaches and backgrounds. While each

of the points may be well-known to some scientists, they are probably not so
obvious for others. Putting it differently, this contribution tries to build bridges
between different disciplines interested in similar subjects, and make thoughts
understandable to scientific communities with different points of views.
A dialogue between social, natural and economic sciences seems to be desirable
not only for the sake of an intellectual exchange on fundamental scientific problems.
It also appears that science is lacking behind the pace of upcoming socio-economic
problems, and that we need to become more efficient in addressing practical
problems [3]. President Lee C. Bollinger of New York’s prestigious Columbia
University formulated the challenge as follows: “The forces affecting societies
around the world ... are powerful and novel. The spread of global market systems
... are ... reshaping our world ..., raising profound questions. These questions
call for the kinds of analyses and understandings that academic institutions are
uniquely capable of providing. Too many policy failures are fundamentally failures
of knowledge” [4].
The fundamental and practical scientific challenges require from us that we do
everything we can to find solutions, and that we do not give up before the limits
or failure of a scientific approach have become obvious. As will be argued in the
Discussion and Outlook, different methods should be seen complementary to each
other and, even when inconsistent, may allow one to get a better picture than any
single method can do, no matter how powerful it may seem.
1.2 Particular Difficulties of Modeling Socio-Economic

Systems
When speaking about socio-economic systems in the following, it could be anything

from families over social groups or companies up to countries, markets, or the world
economy including the financial system and the labor market. The constituting sys-
tem elements or system components would be individuals, groups, or companies, for
example, depending on the system under consideration and the level of description
one is interested in.
On the macroscopic (systemic) level, social and economic systems have some
features that seem to be similar to properties of certain physical or biological
systems. One example is the hierarchical organization. In social systems, individuals
form groups, which establish organizations, companies, parties, etc., which make up
states, and these build communities of states (like the United States or the European
Union, for example). In physics, elementary particles form atoms, which create
molecules, which may form solid bodies, fluids or gases, which together make up
our planet, which belongs to a solar system, and a galaxy. In biology, cells are
composed of organelles, they form tissues and organs, which are the constituting
parts of living creatures, and these make up ecosystems.
1.2 Particular Difficulties of Modeling Socio-Economic Systems 3
Such analogies are certainly interesting and have been discussed, for example,
by Herbert Spencer [5] and later on in systems theory [6]. It is not so obvious,
however, how much one can learn from them. While physical systems are often well
understood by mathematical models, biological and socio-economic systems are
usually not. This often inspires physicists to transfer their models to biological and
socio-economic problems (see the discussion in Sect. 1.4.4: “The Model Captures
Some Features...”), while biologists, social scientists, and economists often find
such attempts “physicalistic” and inadequate. In fact, social and economic systems
possess a number of properties, which distinguish them from most physical ones:
1. The number of variables involved is typically (much) larger (considering that
each human brain contains about one thousand billion neurons).
2. The relevant variables and parameters are often unknown and hard to measure
(the existence of “unknown unknowns” is typical).
3. The time scales on which the variables evolve are often not well separated from
each other.
4. The statistical variation of measurements is considerable and masks laws of
social behavior, where they exist (if they exist at all).
5. Frequently there is no ensemble of equivalent systems, but just one realization
(one human history).
6. Empirical studies are limited by technical, financial, and ethical issues.
7. It is difficult or impossible to subdivide the system into simple, non-interacting
subsystems that can be separately studied.
8. The observer participates in the system and modifies social reality.
9. The non-linear and/or network dependence of many variables leads to complex
dynamics and structures, and sometimes paradoxical effects.
10. Interaction effects are often strong, and emergent phenomena are ubiquitous
(hence, not understandable by the measurement and quantification of the
individual system elements).
11. Factors such as a large degree of randomness and heterogeneity, memory, antic-
ipation, decision-making, communication, consciousness, and the relevance of
intentions and individual interpretations complicate the analysis and modeling
a lot.
12. The same applies to human features such as emotions, creativity, and
innovation.
13. The impact of information is often more decisive for the behavior of a socio-
economic system than physical aspects (energy, matter) or our biological
heritage.
14. The “rules of the game” and the interactions in a social or economic system may
change over time, in contrast to what we believe to be true for the fundamental
laws and forces of physics.
15. In particular, social systems are influenced by normative and moral issues,
which are variable.
For such reasons, social systems are the most complex systems we know. They are
certainly more complex than physical systems are. As a consequence, a considerable
fraction of sociologists thinks that mathematical models for social systems are
destined to fail, while most economists and many quantitatively oriented social
scientists seem to believe in models with many variables. Both is in sharp contrast
to the often simple models containing only a few variables that physicists tend to
propose. So, who is right? The following discussion suggests that this is the wrong
question. We will therefore analyze why different scientists, who apparently deal
with the same research subject, come to so dramatically different conclusions.
It is clear that this situation has some undesirable side effects: Scientists
belonging to different schools of thought often do not talk to each other, do not
learn from each other, and probably reject each others’ papers and project proposals
more frequently. It is, therefore, important to make the approach of each school
understandable to the others.
1.3 Modeling Approaches
1.3.1 Qualitative Descriptions
Many social scientists think that the 15 challenges listed above are so serious that
it is hopeless to come up with mathematical models for social systems. A common
view is that all models are wrong. Thus, a widespread approach is to work
out narratives, i.e. to give a qualitative (non-mathematical and non-algorithmic)
description of reality that is as detailed as possible. This may be compared with a
naturalist painting.
Narratives are important, as they collect empirical evidence and create knowl-
edge that is essential for modelers sooner or later. Good models require several steps
of intellectual digestion, and the first and very essential one is to create a picture of
the system one is interested in and to make sense of what is going on in it. This
step is clearly indispensable. Nevertheless, the approach is sometimes criticized for
reasons such as the following:
• Observation, description, and interpretation are difficult to separate from each
other, since they are typically performed by the same brain (of a single scientist).
Since these processes strongly involve the observer, it is hard or even impossible
to provide an objective description of a system at this level of detail. Therefore,
different scientists may analyze and interpret the system in different, subjective
ways. What is an important aspect for one observer may be an irrelevant detail
for another one, or may even be overlooked. In German, there is a saying
that “one does not see the forest amongst all the trees”, i.e. details may hide
the bigger picture or the underlying mechanisms. In the natural sciences, this
problem has been partially overcome by splitting up observation, description,
and interpretation into separate processes: measurements, statistical analysis, and
modeling attempts. Many of these steps are supported by technical instruments,
computers, and software tools to reduce the individual element and subjective
1.3 Modeling Approaches 5
influence. Obviously, this method can not be easily transferred to the study of
social systems, as individuals and subjective interpretations can have important
impacts on the overall system.
• Despite its level of detail, a narrative is often not suited to be translated into a
computer program that would reproduce the phenomena depicted by it. When
scientists try to do so, in many cases it turns out that the descriptions are
ambiguous, i.e. still not detailed enough to come up with a unique computer
model. In other words, different programmers would end up with different
computer models, producing different results. Therefore, Joshua Epstein claims:
“If you didn’t grow it, you didn’t explain it” [7] (where “grow” stands here
for “simulate in the computer”). For example, if system elements interact in
a non-linear way, i.e. effects are not proportional to causes, there are many
different possibilities to specify the non-linearity: is it a parabola, an exponential
dependence, a square root, a logarithm, a power law, ...? Or when a system shows
partially random behavior, is it best described by additive or multiplicative noise,
internal or external noise? Is it chaotic or turbulent behavior, or are the system
elements just heterogeneous? It could even be a combination of several options.
What differences would these various possibilities make?
1.3.2 Detailed Models
In certain fields of computational social science or economics, it is common to

develop computer models that grasp as many details as possible. They would
try to implement all the aspects of the system under consideration, which are
known to exist. In the ideal case, these facts would be properties, which have been
repeatedly observed in several independent studies of the kind of system under
consideration, preferably in different areas of the world. In some sense, they would
correspond to the overlapping part of many narratives. Thus, one could assume
that these properties would be characteristic features of the kind of system under
consideration, not just properties of a single and potentially quite particular system.
Despite it sounds logical to proceed in this way, there are several criticisms of
this approach:
• In case of many variables, it is difficult to specify their interdependencies in the
right way. (Just remember the many different possibilities to specify non-linear
interactions and randomness in the system.)
• Some models containing many variables may have a large variety of different
solutions, which may be highly dependent on the initial or boundary conditions,
or the history of the system. This particularly applies to models containing non-
linear interactions, which may have multiple stable solutions or non-stationary
ones (such as periodic or non-periodic oscillations), or they may even show
chaotic behavior. Therefore, depending on the parameter choice and the initial
condition, such a model could show virtually any kind of behavior. While one
may think that such a model would be a flexible world model, it would in fact be
just a fit model. Moreover, it would probably not be very helpful to understand
the mechanisms underlying the behavior of the system. As John von Neumann
pointed out: “With four parameters I can fit an elephant and with five I can make
him wiggle his trunk.” This wants to say that a model with many parameters can
fit anything and explains nothing. This is certainly an extreme standpoint, but
there is some truth in it.
• When many variables are considered, it is hard to judge which ones are indepen-
dent of each other and which ones not. If variables are mutually dependent, one
effect may easily be considered twice in the model, which would lead to biased
results. Dependencies among variables may also imply serious problems in the
process of parameter calibration. The problem is known, for example, from sets
of linear equations containing collinear variables.
• Models with many variables, particularly non-linear ones, may be sensitive
to the exact specification of parameters, initial, or boundary conditions, or to
small random effects. Phenomena like hysteresis (history-dependence) [8], phase
transitions [9] or “catastrophes” [10], chaos [11], or noise-induced transitions
[12] illustrate this clearly.
• Parameters, initial and boundary conditions of models with many variables
are hard to calibrate. If small (or no) data sets are available, the model is
under-specified, and the remaining data must be estimated based on “expert
knowledge”, intuition or rules of thumb, but due to the sensitivity problem, the
results may be quite misleading. The simulation of many scenarios with varying
parameters can overcome the problem in part, as it gives an idea of the possible
variability of systemic behaviors. However, the resulting variability can be quite
large. Moreover, a full exploration of the parameter space is usually not possible
when a model contains many parameters, not even with supercomputers.
• In models with many variables, it is often difficult to identify the mechanism
underlying a certain phenomenon or system behavior. The majority of variables
may be irrelevant for it. However, in order to understand a phenomenon, it is
essential to identify the variables and interactions (i.e. the interdependencies
among them) that matter.
1.3.3 Simple Models
Simple models try to avoid (some of) the problems of detailed models by restricting
themselves to a minimum number of variables needed to reproduce a certain effect,
phenomenon or system behavior. They are aiming at a better understanding of so-
called “stylized facts”, i.e. simplified, abstracted, or “ideal-typical” observations
(“the essence”). For example, while detailed descriptions pay a lot of attention to
the particular content of social norms or opinions and how they change over time in
relation to the respective cultural setting, simple models abstract from the content of
social norms and opinions. They try to formulate general rules of how social norms
come about or how opinions change, independently of their content, with the aim of
understanding why these processes are history-dependent (“hysteretic”) and in what
way they dependent on microscopic and macroscopic influences.
It is clear that simple models do not describe (and do not even want to describe)
all details of a system under consideration, and for this reason they are also called
minimal or toy models sometimes. The approach may be represented by a few
quotes. The “KISS principle” for building a model demands to “keep it simple and
straightforward” [13]. This is also known as Occam’s (or Ockham’s) razor, or as
principle of parsimony. Albert Einstein as well demanded [14]: “Make everything
as simple as possible, but not simpler”.
A clear advantage of simple models is that they may facilitate an analytical
treatment and, thereby, a better understanding. Moreover, it is easy to extend simple
models in a way that allows one to consider a heterogeneity among the system
components. This supports the consideration of effects of individuality and the
creation of simple “ecological models” for socio-economic systems. Nevertheless,
as George Box puts it: “Essentially, all models are wrong, but some are useful” [15].
The last quote touches an important point. The choice of the model and its degree
of detail should depend on the purpose of a model, i.e. its range of application. For
example, there is a large variety of models used for the modeling and simulation of
freeway traffic. The most prominent model classes are “microscopic” car-following
models, focussing on the interaction of single vehicles, “mesoscopic” gas-kinetic
models, describing the change of the velocity distribution of vehicles in space and
time, “macroscopic” fluid-dynamic models, restricting themselves to changes of
the average speed and density of vehicles, and cellular automata, which simplify
microscopic ones in favor of simulation speed. Each type of model has certain
ranges of application. Macroscopic and cellular automata models, for example, are
used for large-scale traffic simulations to determine the traffic situation on freeways
and perform short-term forecasts, while microscopic ones are used to study the
interaction of vehicles and to develop driver assistance systems. For some of these
models, it is also known how they are mathematically connected with each other,
i.e. macroscopic ones can be derived from microscopic ones by certain kinds of
simplifications (approximations) [16, 17].
The main purpose of models is to guide people’s thoughts. Therefore, models
may be compared with city maps. It is clear that maps simplify facts, otherwise they
would be quite confusing. We do not want to see any single detail (e.g. each tree) in
them. Rather we expect a map to show the facts we are interested in, and depending
on the respective purpose, there are quite different maps (showing streets, points
of interest, topography, supply networks, industrial production, mining of natural
resources, etc.).
One common purpose of models is prediction, which is mostly (mis)understood
as “forecast”, while it often means “the identification of implications regarding how
a system is expected to behave under certain conditions”. It is clear that, in contrast
to the motion of a planet around the sun, the behavior of an individual can hardly
be forecasted. Nevertheless, there are certain tendencies or probabilities of doing
certain things, and we usually have our hypotheses of what our friends, colleagues,
or family members would do in certain situations.
Moreover, it turns out that, when many people interact, the aggregate behavior
can sometimes be quite predictable. For example, the “wisdom of crowds” is based
on the statistical law of large numbers [18], according to which individual variations
(here: the independent estimation of facts) are averaged out.
Furthermore, interactions between many individuals tend to restrict the degree of
freedom regarding what each individual can or will do. This is, why the concept of
“social norms” is so important. Another example is the behavior of a driver, which is
constrained by other surrounding vehicles. Therefore, the dynamics of traffic flows
can be mathematically well understood [17, 19]. Nevertheless, one cannot exactly
forecast the moment in which free traffic flow breaks down and congestion sets
in, and therefore, one cannot forecast travel times well. The reason for this is the
history-dependent dynamics, which makes it dependent on random effects, namely
on the size of perturbations in the traffic flow. However, what can be predicted is
what are the possible traffic states and what are conditions under which they can
occur. One can also identify the probability of traffic flows to break down under
certain flow conditions, and it is possible to estimate travel times under free and
congested flow conditions, given a measurement of the inflows. The detail that
cannot be forecasted is the exact moment in which the regime shift from free to
congested traffic flow occurs, but this detail has a dramatic influence on the system.
It can determine whether the travel time is 5 or 40 min.
However, it is important to underline that, in contrast to what is frequently
stated, the purpose of developing models is not only prediction. Joshua Epstein,
for example, discusses 16 other reasons to build models, including explanation,
guiding data collection, revealing dynamical analogies, discovering new questions,
illuminating core uncertainties, demonstrating tradeoffs, training practitioners, and
decision support, particularly in crises [20].
Of course, not everybody favors simple models, and typical criticisms of them
are:
• It is usually easy to find empirical evidence, which is not compatible with simple
models (even though, to be fair, one would have to consider the purpose they
have been created for, when judging them). Therefore, one can say that simple
models tend to over-simplify things and leave out more or less important facts.
For this reason, they may be considered inadequate to describe a system under
consideration.
• Due to their simplicity, it may be dangerous to take decisions based on their
implications.
• It may be difficult to decide, what are the few relevant variables and parameters
that a simple model should consider. Scientists may even disagree about the
stylized facts to model.
• Simple models tend to reproduce a few stylized facts only and are often not able
to consistently reproduce a large number of observations. The bigger picture and
the systemic view may get lost.
• Making simple models compatible with a long list of stylized facts often requires
to improve or extend the models by additional terms or parameter dependencies.
Eventually, this improvement process ends up with detailed models, leaving one
with the problems specified in the related section.
• Certain properties and behaviors of socio-economic systems may not be under-
standable with methods that have been successful in physics: Subdividing the
system into subsystems, analyzing and modeling these subsystems, and putting
the models together may not lead to a good description of the overall system.
For example, several effects may act in parallel and have non-separable orders
of magnitude. This makes it difficult or impossible to start with a zeroth or first
order approximation and to improve it by adding correction terms (as it is done,
for example, when the falling of a body is described by the effect of gravitational
acceleration plus the effect of air resistance). Summing up the mathematical
terms that describe the different effects may not converge. It is also not clear
whether complex systems can be always understood via simple principles, as
the success of complexity science might suggest. Some complex systems may
require complex models to explain them, and there may even be phenomena, the
complexity of which is irreducible. Turbulence [21] could be such an example.
While it is a long-standing problem that has been addressed by many bright
people, it has still not been explained completely.
It should be added, however, that we do not know today, whether the last point
is relevant, how relevant it is, and where. So far, it is a potential problem one
should be aware of. It basically limits the realm, in which classical modeling will be
successful, but we have certainly not reached these limits, yet.
1.3.4 Modeling Complex Systems
Modeling socio-economic systems is less hopeless than many social scientists may
think [22]. In recent years, considerable progress has been made in a variety of
relevant fields, including:
• Experimental research [23–25]
• Data mining [26]
• Network science [27]
• Agent-based modeling [7, 28]
• The theory of complex systems (including emergent and self-organized phenom-
ena, or chaos) [29]
• The theory of phase transitions [9] (“catastrophes” [10]), critical phenomena
[30], and extreme events [31]
• The engineering of intelligent systems [32, 33]
These fields have considerably advanced our understanding of complex systems. In
this connection, one should be aware that the term “complexity” is used in many
different ways. In the following, we will distinguish three kinds of complexity:
1. structural,
2. dynamical, and
3. functional complexity
One could also add algorithmic complexity, which is given by the amount of
computational time needed to solve certain problems. Some optimization problems,
such as the optimization of logistic or traffic signal operations, are algorithmically
complex [34].
Linear models are not considered to be complex, no matter how many terms
they contain. An example for structural complexity is a car or airplane. They
are constructed in a way that is dynamically more or less deterministic and well
controllable, i.e. dynamically simple, and they also serve relatively simple functions
(the motion from a location A to another location B). While the acceleration of a
car or a periodic oscillation would be an example for a simple dynamics, examples
for complex dynamical behavior are non-periodic changes, deterministic chaos, or
history-dependent behaviors. Complex dynamics can already be produced by simple
sets of non-linearly coupled equations. While a planet orbiting around the sun
follows a simple dynamics, the interaction of three celestial bodies can already show
a chaotic dynamics. Ecosystems, the human body or the brain would be functionally
complex systems. The same would hold for the world wide web, financial markets,
or running a country or multi-national company.
While the interrelation between function, form and dynamics still poses great
scientific challenges, the understanding of structurally or dynamically complex
systems has significantly progressed. Simple agent-based models of systems with
a large number of interacting system elements (be it particles, cars, pedestrians,
individuals, or companies) show properties, which remind of socio-economic
systems. Assuming that these elements mutually adapt to each other through
non-linear or network interactions (i.e. that the elements are influenced by their
environment while modifying it themselves), one can find a rich, history-dependent
system behavior, which is often counter-intuitive, hardly predictable, and seemingly
uncontrollable. These models challenge our common way of thinking and help to
grasp behaviors of complex systems, which are currently a nightmare for decision-
makers.
For example, complex systems are often unresponsive to control attempts,
while close to “critical points” (also known as “tipping points”), they may cause
sudden (and often unexpected) phase transition (so-called “regime shifts”). These
correspond to discontinuous changes in the system behavior. The breakdown of
free traffic flow would be a harmless example, while a systemic crisis (such as
a financial collapse or revolution) would be a more dramatic one. Such systemic
crises are often based on cascade spreading through network interactions [35].
Hence, complex adaptive systems allow one to understand extreme events as a
result of strong interactions in a system (rather than as externally caused shocks).
Furthermore, the interaction of many system elements may give rise to interesting
self-organization phenomena and emergent properties, which cannot be understood
from the behaviors of the single elements or by adding them up. Typical examples
1.4 Challenges of Socio-Economic Modeling 11
are collective patterns of motion in pedestrian crowds or what is sometimes called

“swarm intelligence” [36].
Considering this, it is conceivable that many of today’s puzzles in the social sci-
ences may one day be explained by simple models, namely as emergent phenomena
resulting from interactions of many individuals and/or other system elements. It is
important to note that emergent phenomena cannot be explained by linear models
(which are most common in many areas of quantitative empirical research in the
social sciences and economics).
Unfortunately, there is no standard way to set up models of emergent phenomena.
On the one hand, there are many possible kinds of non-linear functional dependen-
cies (“interactions”) (see the end of the section on “Qualitative Descriptions”). On
the other hand, model assumptions that appear plausible do often not produce the
desired or expected effects.
In spite of these difficulties, taking into account time-dependent change, a non-
linear coupling of variables, spatial or network interactions, randomness, and/or
correlations (i.e. features that many social and economic models currently do not
consider to the necessary extent), can sometimes deliver unexpected solutions of
long-standing puzzles. For example, it turns out that representative agent models
(which are common in economics) can be quite misleading, as the same kinds of
interactions among the system components can imply completely different or even
opposite conclusions, when interactions take place in a socio-economic network
rather than with average (or randomly chosen) interaction partners [37]. Therefore,
models often produce counter-intuitive results, when spatio-temporal or network
interactions are relevant. A simple non-linear model may explain phenomena, which
complicated linear models may fail to reproduce. In fact, this generally applies
to systems that can show several possible states (i.e. systems which do not have
just one stable equilibrium). Out-of-equilibrium models are also required for the
description of systemic crises such as the current financial crisis [35].
1.4 Challenges of Socio-Economic Modeling
Many people before and after Popper have been thinking about the logic of scientific
discovery [38]. A wide-spread opinion is that a good model should be applicable to
measurements of many systems of a certain kind, in particular to measurements in
different areas of the world. The more observations a model can explain and the less
parameters it has, the more powerful it is usually considered to be.
Models with a few parameters can often be easier to calibrate, and cause-and-
effect relationships may be better identified, but one can usually not expect that these
models would provide an exact description of reality. Nevertheless, a good model
should make predictions regarding some possible, but previously unobserved system
behaviors. In this connection, prediction does not necessarily mean the forecast of
a certain event at a specific future point in time. It means a specific system behavior
that is expected to occur (or to be possible) under certain conditions (e.g. for certain
parameter combinations or certain initial conditions). When such conditions apply

and the system shows the expected behavior, this would be considered to verify the
model, while the model would be falsified or seriously questioned, if the predicted
system behavior would not occur. By experimentally challenging models based on
their predictions (implications), it has been possible in the natural sciences to rate
alternative models based on their quality in reproducing and predicting measured
data. Unfortunately, it turns out that this approach is less suited to identify “the right
model” of a social or economic system under consideration. As we will discuss
in the following, this is not only due to the smaller amount of data available on
most aspects of social and economic systems and due to experimental limitations
for financial, technical and ethical reasons...
1.4.1 Promises and Difficulties of the Experimental Approach
So far, it is very expensive to carry out social and economic experiments, for
example in the laboratory. While the study of human behavior under controlled
conditions has become a common research method not only in psychology, but also
in experimental economics and in sociology, the number of individuals that can
be studied in such experiments is limited. This implies a large degree of statistical
variation, which makes it difficult to determine behavioral laws or to distinguish
between different models. The statistical noise creates something like a foggy
situation, which makes it difficult to see what is going on. In physics, this problem
can be usually solved by better measurement methods (apart from uncertainty that
results from the laws of quantum mechanics). In social systems, however, there is an
irreducible degree of randomness. The behavior varies not only between individuals
due to their heterogeneity (different “personality”). It also varies from one instance
to another, i.e. the decision-making of an individual is usually not deterministic. This
could be due to various reasons: unknown external influences (details attracting the
attention of the individual) or internal factors (exploration behavior, memory effects,
decisions taken by mistake, etc.).
The large level of behavioral variability within and between individuals is
probably not only due to the different histories individuals have, but also due to
the fact that exploration behavior and the heterogeneity of behaviors are beneficial
for the learning of individuals and for the adaptability of human groups to various
environmental conditions. Applying a theory of social evolution would, therefore,
suggest that randomness is significant in social and economic systems, because
it increases system performance. Besides, heterogeneity can also have individual
benefits, as differentiation facilitates specialization. The benefit of a variation
between individuals is also well-known from ecological systems [39].
Besides impeding the discovery of behavioral laws, the limited number of
participants in laboratory experiments also restricts the number of repetitions and
the number of experimental settings or parameter combinations that can be studied.
Scanning parameter spaces is impossible so far, while it would be useful to detect
different system behaviors and to determine under which conditions they occur. It
can be quite tricky to select suitable system parameters (e.g. the payoff matrix in a
game-theoretical experiment).
Computer simulations suggest that one would find interesting results mainly,
if the parameters selected in different experimental setups imply different system
behaviors, i.e. if they belong to different “phases” in the parameter space (see
Fig. 1.1). To identify such parameter combinations, it is advised to perform com-
puter simulations before, in order to determine the phase diagram for the system
under consideration [25]. The problem, however, is that the underlying model is
unlikely to be perfect, i.e. even a good social or economic model is expected to
make predictions which are only approximately valid. As a consequence, the effect
one likes to show may appear for (somewhat) different parameter values, or it may
not occur at all (considering the level of randomness) [40].
1.4.2 Several Models Are Right
The above mentioned properties of socio-economic systems imply that it is difficult

to select the “right” model among several alternative ones. For an illustration, let us
take car-following models, which are used for the simulation of urban or freeway
traffic. Thanks to radar sensors, it has become possible to measure the acceleration
of vehicles as a function of the typical variables of car-following models, which are
the distance to the car ahead, the own speed, and the speed difference. When fitting
the parameters of various car-following models to data of such measurements, it
turns out that the remaining error between computer simulations and measurements
is about the same for most of the models. The calibration error varies between 12
and 17%, and according to the authors, “no model can be denoted to be the best”
[41]. When the error of different models (i.e. the deviation between model and data)
is determined for a new data set (using the model parameters determined with the
previous data set), the resulting validation error usually varies between 17 and 22%
(larger validation errors mainly result, when the calibration data set is overfitted)
[41]. Again, the performance of the different models is so similar that it would not be
well justified to select one of them as the “correct” model and exclude all the others.
A closer analysis shows that the parameters of the car-following dynamics vary
among different drivers, but the behavior of specific drivers also vary over time [42].
We have to assume that the same applies to basically all kinds of behavior, not only
for car driving. Moreover, it is likely that many behaviors (such as decision-making
behaviors) vary even more than car-following behavior does. As a consequence,
it would be even more difficult to distinguish between different models by means
of empirical or experimental data, which would mean that we may have to accept
several models to be (possibly) “right”, even when they are not consistent with each
other. In other words, the question “What is the best model?” or “How to choose
the model?” may not be decidable in a reasonable way, as is also suggested by the
a b
r = 2.0 r = 3.5
0.7 0.7
0.6 0.6
0.5 0.5
cost
cost
0.4 0.4
D D
0.3 0.3
0.2 PC 0.2 PC
D+PC
0.1 0.1
0.0 0.0
0 0.3 0.6 0.9 1.2 1.5 1.8 0 0.1 0.2 0.3 0.4 0.5 0.6
fine fine
c d
r = 4.4 r = 3.5 (enlarged)
0.7 0.03
0.6
0.5
D+C 0.02 D D+PC
cost
0.4
cost
0.3 PC PC
0.01
0.2
0.1 D+PC PD+PC
0.0 0.00
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.15 0.30 0.45
fine fine
Fig. 1.1 So-called “phase diagram”, showing the finally remaining strategies in the spatial public
goods game with cooperators (C), defectors (D), cooperators who punish defectors (PC) and
hypocritical punishers (PD), who punish other defectors while defecting themselves (after [37]).
Initially, each of the four strategies occupies 25% of the sites of the two-dimensional lattice,
in which individuals interact, and their distribution is uniform in space. However, due to their
evolutionary competition, two or three strategies die out after some time. The finally resulting state
depends on the punishment cost, the punishment fine, and the synergy r of cooperation (the factor
by which cooperation increases the sum of investments). The displayed phase diagrams are for
(a) r D 2:0, (b) r D 3:5, and (c) r D 4:4. (d) Enlargement of the small-cost area for r D 3:5.
Solid separating lines indicate that the resulting fractions of all strategies change continuously
with a modification of the punishment cost and punishment fine, while broken lines correspond to
discontinuous changes. All diagrams show that cooperators and defectors cannot stop the spreading
of costly punishment, if only the fine-to-cost ratio is large enough (see the PC area). Note that, in
the absence of defectors, the spreading of punishing cooperators is extremely slow and follows a
voter model kind of dynamics. A small level of strategy mutations (which continuously creates a
small number of strategies of all kinds, in particular defectors) can largely accelerate the spreading
of them. Furthermore, there are parameter regions where punishing cooperators can crowd out
“second-order free-riders” (non-punishing cooperators) in the presence of defectors (D+PC).
Finally, for low punishment costs, but moderate punishment fines, it may happen that “moralists”,
who cooperate and punish non-cooperative behavior, can only survive through an “unholy alliance”
with “immoral”, hypocritical punishers (PD+PC). For related videos, see http://www.soms.ethz.ch/
research/secondorder-freeriders or http://www.matjazperc.com/games/moral.html
next section. This situation reminds a bit of Gödel’s Undecidability Theorem [43],
which relates to the (in)completeness of certain axiom systems.
It may be tempting to determine the best model as the one which is most
successful, for example in terms of the number of citations it gets. However, success
is not necessarily an indicator of a good model. Let us take models used for stock
trading as an example. Clearly, even if the stock prices vary in a perfectly random
manner and if the average success of each model is the same over an infinite time
period; when different traders apply different trading models, they will be differently
successful at any chosen point in time. Therefore, one would consider some models
more successful than others, while this would be only a matter of luck. At other
points in time, different models would be the most successful ones.
Of course, if behaviors are not just random, some models should be better than
others, and it should eventually be possible to separate “good” from “bad” models
through the “wisdom of crowds” effect. However, the “wisdom of crowds” assumes
independent judgements, while scientists have repeated interactions. It has be shown
experimentally that this tends to create consensus, but that this consensus will
often deviate from the truth [44]. The problem results from social influence, which
creates a herding effect that can undermine the “wisdom of crowds”. Of course,
this mainly applies, when the facts are not sufficiently obvious, which however is
the case in the social sciences due to the high variability of observations, while the
problem is less pressing in the natural sciences thanks to the higher measurement
precision. Nevertheless, the physicist Max Planck is known for the quote: “Science
progresses funeral by funeral” [45]. Thomas Kuhn’s study of scientific revolutions
[46] suggests as well that scientific progress is not continuous, but there are sudden
paradigm shifts. This reveals the problem of herding effects. Even a collective
agreement is no guarantee for the correctness of a model, as the replacement of
classical mechanics by relativistic quantum theory shows. In other words, success is
no necessarily an indicator for good models. It may just be an indicator for what
model is most fashionable at a given time. The problem becomes worse by the
academic selection process that decides, what scientists make a career and which
ones not. It creates a considerable inertia in the adjustment to new knowledge, i.e.
scientific trends are likely to persist longer than what is justified by facts.
1.4.3 No Known Model Is Right
A typical approach in the natural sciences is to verify or falsify previously untested

predictions (implications) of alternative models by sometimes quite sophisticated
experiments. Only in the minority of cases, two alternative theories turn out to be the
same, like the wave and the particle picture of quantum mechanics. In most cases,
however, two theories A and B are non-identical and inconsistent, which means that
they should make different predictions in particular kinds of situations. Experiments
are performed to find out whether theory A or theory B is right, or whether both
of them deviate from the measurements. If the experimental data confirm theory
A and are not compatible with theory B (i.e. deviate significantly from it), one would
discard theory B forever. In this way, experiments are thought to narrow down the
number of alternative theories, until a single theory remains, which is considered to
be “true”.
When social or economic systems are modeled, the following situation is not
unlikely to happen: Scientists identify mutually incompatible predictions of theories
A and B, and it turns out that an experiment supports theory A, but not theory
B. One day, another scientist identifies a different set of incompatible predictions,
and another experiment supports theory B, but not theory A. Due to the inherent
simplifications of socio-economic models, for any model it should be easy to find
empirical evidence that contradicts it. What should one do in such cases? Giving up
on modeling would probably not be the best idea. Generalizing a model is always
possible, but it will usually end up with detailed models, which implies a number
of problems that have been outlined in the related section. One could also stay
with many particular models and determine their respective ranges of validity. This,
however, will never result in a holistic or systemic model. A possible way out would
be the approach of pluralistic modeling outlined in the Summary and Outlook.
Modeling in modern physics seems to face similar problems. While one would
expect that each experiment narrows down the number of remaining, non-falsified
models, one actually observes that, after each experiment, scientists come up with
a number of new models. As people say: “Each answered question raises ten new
ones.” In fact, there is an abundance of elementary particle models, and the same
applies to cosmological models. Many models require to assume the existence of
factors that have never been measured and perhaps will never be measured, such as
Higgs bosons, dark matter, or dark energy. We will probably have to live with the
fact that models are just models that never grasp all details of reality.
Moreover, as has been pointed out, understanding elementary particles and
fundamental forces in physics does not explain at all what is happening in the world
around us [47, 48]. Many emergent phenomena that we observe in the biological,
economic and social world will never be derived from elementary particle physics,
because emergent properties of a system cannot be understood from the properties
of its system components alone. They usually come about by the interaction of a
large number of system components. Let us be honest: Our textbooks do not even
explain the particular properties of water, as simple as H2 O molecules may be. (Of
course, this does not mean that this situation will remain forever – see e.g. H. Eugene
Stanley’s related work.)
Generally, there is still a serious lack in understanding the connection between
function, dynamics, and form. Emergence often seems to have an element of sur-
prise. The medical effect of a new chemical drug cannot be understood by computer
simulation alone. So far, we also do not understand emotions and consciousness,
and we cannot calculate the biological fitness of a species in the computer. The
most exciting open puzzles in science concern such emergent phenomena. It would
be interesting to study, whether social and economic phenomena such as trust,
solidarity, and economic value can be understood as emergent phenomena as
well [3].
1.4.4 The Model Captures Some Features, But May Be

Inadequate
Scientists are often prompted to transfer their methods to another areas of applica-
tion, based on analogies that they see between the behavior of different systems.
Systems science is based on such analogies, and physicists generalize their methods
as well. The question is how useful a “physicalist approach” can be, which transfers
properties of many-particle systems to social or economic systems, although indi-
viduals are certainly more intelligent than particles and have many more behavioral
degrees of freedom.
Of course, physicists would never claim that particle models could provide an
exact description of social or economic systems. Why, then, do they think the
models could make a contribution to the understanding of these systems? This
is, because they have experience with what can happen in systems characterized
by the non-linear interaction of many system components in space and time,
and when randomness plays a role. They know how self-organized collective
phenomena on the “macroscopic” (aggregate) level can results from interactions
on the “microscopic” (individual) level. And they have learned, how this can
lead to phase transitions (also called “regime shifts” or “catastrophes”), when a
system parameter (“control parameter”) crosses a critical point (“tipping point”).
Furthermore, they have discovered that, at a critical point, the system typically
shows a scale-free behavior (i.e. power laws or other fat-tail distributions rather
than Gaussian distributions).
It is important to note that the characteristic features of the system at the critical
point tend to be “universal”, i.e. they largely do not depend on the details of
the interactions. This is, why physicists think they can abstract from the details.
Of course, details are expected to be relevant when the system is not close to a
critical point. It should also be added, that there are a number of different kinds
of universal behavior, so-called universality classes. Nevertheless, many-particle
models may allow one to get a better understanding of regime shifts, which are not
so well understood by most established models in economics or the social sciences.
However, if the tipping point is far away, the usefulness of many-particle models
is limited, and more detailed descriptions, as they are favored by economists and
social scientists, appear to be more adequate.
Sometimes, it is not so clear how far analogies can carry, or whether they are
useful at all. Let us take neural network models. In a certain sense, they can be
used to model learning, generalization, and abstraction. However, the hope that they
would explain the functioning of the brain has been largely disappointed. Today, we
know that the brain works quite differently, but neural network theory has given birth
to many interesting engineering applications that are even commercially applied.
Let us consider models of cooperation based on coupled oscillators as a second
example. Without any doubt, the synchronization of cyclical behavior is among the
most interesting collective phenomena we know of, and models allow one to study if
and how groups of oscillators will coordinate each other or fall apart into subgroups
(which are not synchronized among each other, while the oscillators in each of them
are) [49]. Despite this analogy to group formation and group dynamics, it is not
clear, what we can learn from such models for social systems.
A similar point is sometimes raised for spin models, which have been proposed
to describe opinion formation processes or the emergence of cooperation in social
dilemma situations. In this connection, it has been underlined that social interactions
cannot always be broken down into binary interactions. Some interactions involve
three or more individuals at the same time, which may change the character of
the interaction. Nevertheless, similar phenomena have been studied by overlaying
binary interactions, and it is not fully clear how important the difference is.
Let us finally ask whether unrealistic assumptions are generally a sign of bad
models? The discussion in the section on “Simple Models” suggests that this is
not necessarily so. It seems more a matter of the purpose of a model, which
determines the level of simplification, and a matter of the availability of better
models, i.e. a matter of competition. Note, however, that a more realistic model
is not necessarily more useful. For example, many car-following models are more
realistic than fluid-dynamic traffic models, but they are not suited to simulate large-
scale traffic networks in real-time. For social systems, there are a number of different
modeling approaches as well, including the following:
• Physical(istic) modeling approach: Socio- and econo-physicists often abstract
social interactions so much that their models come down to multi-particle models
(or even spin models with two behavioral options). Such models focus on the
effect of non-linear interactions and are a special case of bounded rationality
models, sometimes called zero-intelligence models [50]. Nevertheless, they may
display features of collective or swarm intelligence [36]. Furthermore, they
may be suited to describe regime shifts or situations of routine choice [51],
i.e. situations where individuals react to their environment in a more or less
subconscious and automatic way. Paul Omerod, an economist by background,
argues as follows [52]: “In many social and economic contexts, self-awareness of
agents is of little consequence... No matter how advanced the cognitive abilities
of agents in abstract intellectual terms, it is as if they operate with relatively
low cognitive ability within the system... The more useful ‘null model’ in social
science agent modelling is one close to zero intelligence. It is only when this fails
that more advanced cognition of agents should be considered.”
• Economic modeling approach: Most economists seem to have quite the opposite
approach. Their concept of “homo economicus” (the “perfect egoist”) assumes
that individuals take strategic decisions, choosing the optimal of their behavioral
options. This requires individuals with infinite memory and processing capaci-
ties. Insofar, one could speak of an infinite-intelligence approach. It is also known
as rational choice approach and has the advantage that the expected behaviors of
individuals can be axiomatically derived. In this way, it was possible to build
the voluminous and impressive theory of mainstream economics. Again, the
reliability of this theory depends, of course, on the realism of its underlying
assumptions.
• Sociological modeling approach: Certain schools of sociologists use rational

choice models as well. In contrast to economists, however, they do not generally
assume that individuals would radically optimize their own profit. Their models
rather consider that, in social systems, exchange is more differentiated and multi-
faceted. For example, when choosing their behavior, individuals may not only
consider their own preferences, but the preferences of their interaction partner(s)
as well. In recent years, “fairness theory” has received a particular attention [53]
and often been contrasted with rational choice theory. These social aspects of
decision-making are now eventually entering economic thinking as well [54].
• Psychological modeling approach: Psychologists are perhaps least axiomatic and
usually oriented at empirical observations. They have identified behavioral para-
doxies, which are inconsistent with rational choice theory, at least its classical
variant. For example, it turns out that most people behave in a risk-averse way.
To account for their observations, new concepts have been developed, including
prospect theory [55], satisficing theory [56], and the concept of behavioral
heuristics [57]. In particular, it turns out that individual decisions depend on the
respective framing. In his Nobel economics lecture, Daniel Kahneman put it this
way: “Rational models are psychologically unrealistic... the central characteristic
of agents is not that they reason poorly, but that they often act intuitively. And
the behavior of these agents is not guided by what they are able to compute,
but by what they happen to see at a given moment.” Therefore, modern research
directions relate to the cognitive and neurosciences. These results are now finding
their way into economics via the fields of experimental, behavioral, and neuro-
economics.
In summary, there is currently no unified approach that scientists generally agree
on. Some of the approaches are more stylized or axiomatic. Others are in better
quantitative agreement with empirical or experimental evidence, but mathematically
less elaborated. Therefore, they are theoretically less suited to derive implications
for the behavior in situations, which have not been studied so far. Consequently,
all models have their strengths and weaknesses, no matter how realistic they may
be. Moreover, none of the mathematical models available so far seems to be
sophisticated enough to reflect the full complexity of social interactions between
many people.
1.4.5 Different Interpretations of the Same Model
A further difficulty of modeling socio-economic systems is that scientists may not

agree on the interpretation of a model. Let us discuss, for example, the multi-nomial
logit model, which has been used to model decision-making in a large variety of
contexts and awarded with the Nobel prize [58]. This model can be derived in a
utility-maximizing framework, assuming perfectly rational agents deciding under
conditions of uncertainty. The very same model, however, can also be derived in
other ways. For example, it can be linked to psychological laws or to distributions

of statistical physics [59]. In the first case, the interpretation is compatible with the
infinite-intelligence approach, while in the last case, it is compatible with the zero-
intelligence approach, which is quite puzzling. A comparison of these approaches
is provided by [59].
1.5 Discussion and Outlook
1.5.1 Pluralistic or Possibilistic Modeling and Multiple World

Views: The Way Out?
Summarizing the previous discussion, it is quite unlikely that we will ever have a
single, consistent, complete, and correct model of socio-economic systems. Maybe
we will not even find such a grand unified theory in physics. Recently, doubts along
these lines have even been raised by some particle physicists [60, 61]. It may be the
time to say good bye to a modeling approach that believes in the feasibility of a
unique, general, integrated and consistent model. At least there is no theoretical or
empirical evidence for the possibility of it.
This calls for a paradigm shift in the modeling approach. It is important to be
honest that each model is limited, but most models are useful for something. In
other words, we should be tolerant with regard to each others’ models and see where
they can complement each other. This does not mean that there would be separate
models for non-overlapping parts of the system, one for each subsystem. As has
been pointed out, it is hard to decide whether a particular model is valid, no matter
how small the subsystem is chosen. It makes more sense to assume that each model
has a certain validity or usefulness, which may be measured on a scale between 0
and 1, and that the validity furthermore depends on the part or aspect of the system
addressed. This validity may be quantified, for example, by the goodness of fit of a
given system or the accuracy of description of another system of the same kind.
As there are often several models for each part or aspect of a system, one could
consider all of them and give each one a weight according to their respective validity,
as determined statistically by comparison with empirical or experimental data.
Analogously to the “wisdom of crowds” [18], which is based on the law of large
numbers, this should lead to a better quantitative fit or prediction than most (or even
each) model in separation, despite the likely inconsistency among the models. Such
an approach could be called a pluralistic modeling approach [62], as it tolerates
and integrates multiple world views. It may also be called a possibilistic approach
[63], because it takes into account that each model has only a certain likelihood
to be valid, i.e. each model describes a possible truth. However, this should not be
misunderstood as an appeal for a subjectivistic approach. The pluralistic modeling
approach still assumes that there is some underlying reality that some, many, or all
of us share (depending on the aspect one is talking about).
1.5 Discussion and Outlook 21
As shocking as it may be for many scientists and decision-makers to abandon

their belief in the existence of a unique, true model, solution, or policy, the
pluralistic modeling approach is already being used. Hurricane prediction and
climate modeling are such examples [64]. Even modern airplanes are controlled
by multiple computer programs that are run in parallel. If they do not agree with
each other, a majority decision is taken and implemented. Although this may
seem pretty scary, this approach has apparently worked surprisingly well so far.
Moreover, when crash tests of newly developed cars are simulated in the computer,
the simulations are again performed with several models, each of which is based on
different approximation methods. Therefore, it is plausible to assume that pluralistic
modeling will be much more widely used in future, whenever a complex system
shall be modeled.
1.5.2 Where Social Scientists and Natural Scientists

or Engineers Can Learn from Each Other
It has been argued that all modeling approaches have their strengths and weaknesses,
and that they should be considered complementary to each other rather than being
each others’ enemies. This also implies that scientists of different disciplines may
profit and learn from each other. Areas of fruitful multi-disciplinary collaboration
could be:
• The modeling of socio-economic systems themselves.
• Understanding the impacts that engineered systems have on the socio-economic
world.
• The modeling of the social mechanisms that drive the evolution and spreading of
innovations, norms, technologies, products etc.
• Scientific challenges related to the questions how to manage complexity or to
design better systems.
• The application of social coordination and cooperation mechanisms to the cre-
ation of self-organizing technical systems (such as decentralized traffic controls
or peer-to-peer systems).
• The development of techno-social systems [65], in which the use of technology
is combined with social competence and human knowledge (such as Wikipedia,
prediction markets, recommender systems, or the semantic web).
Given the large potentials of such collaborations, it is time to overcome disci-
plinary boundaries. They seem to make less and less sense. It rather appears that
multi-disciplinary, large-scale efforts are needed to describe and understand socio-
economic systems well enough and to address practical challenges of humanity
(such as the financial and economic crisis) more successfully [66].
Acknowledgements The author is grateful for support by the ETH Competence Center “Coping
with Crises in Complex Socio-Economic Systems” (CCSS) through ETH Research Grant CH1-01
08-2 and by the Future and Emerging Technologies programme FP7-COSI-ICT of the European
Commission through the project Visioneer (grant no.: 248438).
References
1. A. Comte, Social Physics: From the Positive Philosophy (Calvin Blanchard, New York, 1856)
2. A. Comte, Course on Positive Philosophy (1830–1842)
3. D. Helbing, Grand socio-economic challenges, Working Paper, ETH Zurich (2010)
4. L.C. Bollinger, Announcing the Columbia committee on global thought, see http://www.
columbia.edu/content/announcing-columbia-committee-global-thought.html, last accessed on
March 6, 2012
5. H. Spencer, The Principles of Sociology (Appleton, New York, 1898; the three volumes were
originally published in serial form between 1874 and 1896)
6. L. von Bertalanffy, General System Theory: Foundations, Development, Applications (George
Braziller, New York, 1968)
7. J.M. Epstein, Generative Social Science. Studies in Agent-Based Computational Modeling
(Princeton University, 2006), p. 51
8. I.D. Mayergoyz, Mathematical Models of Hysteresis and their Applications (Academic Press,
2003)
9. H.E. Stanley Introduction to Phase Transitions and Critical Phenomena (Oxford University,
1987)
10. E.C. Zeeman (ed.), Catastrophe Theory (Addison-Wesley, London, 1977)
11. H.G. Schuster, W. Just, Deterministic Chaos (Wiley-VCH, Weinheim, 2005)
12. W. Horsthemke, R. Lefever, Noise-Induced Transitions: Theory and Applications in Physics,
Chemistry, and Biology (Springer, Berlin, 1983)
13. “KISS principle” at Wikipedia.org, see http://en.wikipedia.org/wiki/KISS principle, last
accessed on March 6, 2012
14. A. Einstein, “On the Method of Theoretical Physics”. The Herbert Spencer Lecture, delivered
at Oxford (10 June 1933); also published in Philosophy of Science 1(2), p. 165 (April 1934).
15. G.E.P. Box, N.R. Draper, Empirical Model-Building and Response Surfaces (Wiley, NJ, 1987),
pp. 74+424
16. D. Helbing, Derivation of non-local macroscopic traffic equations and consistent traffic
pressures from microscopic car-following models. Eur. Phys. J. B 69(4), 539–548 (2009), see
also http://www.soms.ethz.ch/research/traffictheory
17. D. Helbing, Traffic and related self-driven many-particle systems. Rev.Mod. Phys. 73, 1067–
1141 (2001)
18. F. Galton, Vox populi. Nature 75, 450–451 (1907)
19. D. Helbing, et al., see collection of publications on analytical traffic flow theory at http://www.
soms.ethz.ch/research/traffictheory, last accessed on March 6, 2012
20. J.M. Epstein, Why model? J Artif. Soc. Soc. Simulat. 11(4), 12 (2008). see http://jasss.soc.
surrey.ac.uk/11/4/12.html
21. P.A. Davidson, Turbulence (Cambridge University, Cambridge, 2004)
22. W. Weidlich, Sociodynamics: A Systemic Approach to Mathematical Modelling in the Social
Sciences (Dover, 2006)
23. J.H. Kagel, A.E. Roth, The Handbook of Experimental Economics (Princeton University,
Princeton, NJ, 1995)
24. F. Guala, The Methodology of Experimental Economics (Cambridge University Press,
New York, 2005)
25. D. Helbing, W. Yu (2010) The future of social experimenting. Proceedings of the National
Academy of Sciences USA (PNAS) 107(12), 5265–5266; see also http://www.soms.ethz.ch/
research/socialexperimenting
26. O. Maimon, L. Rokach, The Data Mining and Knowledge Discovery Handbook (Springer,
Berlin, 2005)
27. M.O. Jackson, Social and Economic Networks (Princeton University, Princeton, 2008)
28. N. Gilbert (ed.), Computational Social Science (Sage, CA, 2010)
References 23
29. J.H. Miller, S.E. Page, Complex Adaptive Systems: An Introduction to Computational Models
of Social Life (Princeton University, Princeton, NJ, 2007)
30. D. Sornette, Critical Phenomena in Natural Sciences. Chaos, Fractals, Selforganization and
Disorder: Concepts and Tools (Springer, Berlin, 2006)
31. S. Albeverio, V. Jentsch, H. Kantz (eds.), Extreme Events in Nature and Society (Springer,
Berlin, 2005)
32. D. Floreano, C. Mattiussi, Bio-Inspired Artificial Intelligence: Theories, Methods, and Tech-
nologies (MIT, Cambridge, MA, 2008)
33. S. Nolfi, D. Floreano, Evolutionary Robotics : The Biology, Intelligence, and Technology of
Self-Organizing Machines (MIT, Cambridge, MA, 2000)
34. D. Helbing, A. Deutsch, S. Diez, K. Peters, Y. Kalaidzidis, K. Padberg, S. Lämmer,
A. Johansson, G. Breier, F. Schulze, M. Zerial, Biologistics and the struggle for efficiency:
Concepts and perspectives. Adv. Complex Syst. 12(6), 533–548 (2009)
35. D. Helbing, System risks in society and economics. Sante Fe Institute Working Paper #09-12-
044 (2009). See http://www.santafe.edu/media/workingpapers/09-12-044.pdf, last accessed on
March 6, 2012
36. M. Moussaid, S. Garnier, G. Theraulaz, D. Helbing, Collective information processing and
pattern formation in swarms, flocks, and crowds. Top. Cognit. Sci. 1(3), 469–497 (2009)
37. D. Helbing, A. Szolnoki, M. Perc, G. Szabó, Evolutionary establishment of moral and double
moral standards through spatial interactions. PLoS Comput. Biol. 6(4), e1000758 (2010)
38. K.R. Popper, The Logic of Scientific Discovery (Hutchinson, 1959); original German version:
Logik der Forschung (Springer, Vienna, 1935)
39. D. Tilman, D. Wedin, J. Knops, Productivity and sustainability influenced by biodiversity in
grassland ecosystems. Nature 379, 718–720 (1996)
40. A. Traulsen, D. Semmann, R.D. Sommerfeld, H.-J. Krambeck, M. Milinski, Human strategy
updating in evolutionary games. Proc. Natl. Acad. Sci. USA (PNAS) 107(7), 2962–2966 (2010)
41. E. Brockfeld, R.D. Kühne, P. Wagner, Calibration and validation of microscopic traffic flow
models. Transport. Res. Board 1876, 62–70 (2004)
42. A. Kesting, M. Treiber, Calibrating car-following models by using trajectory data: Method-
ological study. Transport. Res. Record 2088, 148–156 (2008)
43. K. Gödel, On Formally Undecidable Propositions of Principia Mathematica and Related
Systems (Basic, New York, 1962)
44. J. Lorenz, H. Rauhut, F. Schweitzer, D. Helbing, How social influence undermines the wisdom
of crowds. Proc. Natl. Acad. Sci. USA (PNAS) 108(22), 9020–9025 (2011)
45. M. Planck: “An important scientific innovation rarely makes its way by gradually winning
over and converting its opponents, but rather because its opponents eventually die, and a new
generation grows up that is familiar with it.”
46. T.S. Kuhn, The Structure of Scientific Revolutions (University of Chicago, Chicago, 1962)
47. T. Vicsek, The bigger picture. Nature 418, 131 (2002)
48. L. Pietronero, Complexity ideas from condensed matter and statistical physics. europhysic-
snews 39(6), 26–29
49. A.S. Mikhailov, V. Calenbuhr, From Cells to Societies. Models of Complex Coherent Action
(Springer, Berlin, 2002)
50. R.A. Bentley, P. Ormerod, Agents, intelligence, and social atoms, in Creating Consilience:
Integrating the Sciences and the Humanities, ed. by M. Collard, E. Slingerland (Oxford
University Press, 2011)
51. H. Gintis, The Bounds of Reason: Game Theory and the Unification of the Behavioral Sciences
(Princeton University, Princeton, 2009)
52. P. Omerod, What can agents know? The feasibility of advanced cognition in social and eco-
nomic systems. In Proceedings of the AISB 2008 Convention on Communication, Interaction
and Social Intelligence 6 (Aberdeen, Scotland, 2008), pp. 17–20
53. E. Fehr, K.M. Schmidt, A theory of fairness, competition, and cooperation. Q. J. Econ. 114(3),
817–868 (1999)
54. B. Frey, Economics as a Science of Human Behaviour: Towards a New Social Science
Paradigm (Kluwer Academics, Dordrecht, 1999)
55. D. Kahneman, A. Tversky, Prospect theory: An analysis of decision under risk. Econometrica
47(2), 263–291 (1979)
56. H.A. Simon, A behavioral model of rational choice. Q. J. Econ. 69(1), 99–118 (1955)
57. G. Gigerenzer, P.M. Todd, and the ABC Research Group, Simple Heuristics That Make Us
Smart (Oxford University, New York, 2000)
58. D. McFadden, Conditional logit analysis of qualitative choice behaviour, in Frontiers of
Econometrics, ed. by P. Zarembka (Academic Press, New York, 1974), pp. 105–142
59. D. Helbing, Quantitative Sociodynamics. Stochastic Methods and Models of Social Interaction
Processes (Kluwer Academic, Dordrecht, 1995)
60. P. Woit, Not Even Wrong: The Failure of String Theory and the Search for Unity in Physical
Law for Unity in Physical Law (Basic, New York, 2006)
61. L. Smolin, The Trouble With Physics: The Rise of String Theory, The Fall of a Science, and
What Comes Next (Mariner, Boston, 2007)
62. J. Rotmans, M.B.A. van Asselt, Uncertainty management in integrated assessment modeling:
Towards a pluralistic approach. Environ. Monit. Assess. 69(2), 101–130 (2001)
63. D. Dubois, H. Prade, Possibilistic logic: a retrospective and prospective view. Fuzzy Set. Syst.
144(1), 3–23 (2004)
64. V. Lucarini, Towards a definition of climate science. Int. J. Environ. Pollut. 18(5), 413–422
(2002)
65. A. Vespignani, Predicting the behavior of techno-social systems. Science 325, 425–428 (2009)
66. D. Helbing, The FuturIcT knowledge accelerator: Unleashing the power of information for a
sustainable future, Project Proposal (2010), see http://arxiv.org/abs/1004.4969 and http://www.
futurict.eu, last accessed on March 6, 2012
Chapter 2
Agent-Based Modeling
Since the advent of computers, the natural and engineering sciences have enor-
mously progressed. Computer simulations allow one to understand interactions of
physical particles and make sense of astronomical observations, to describe many
chemical properties ab initio, and to design energy-efficient aircrafts and safer
cars. Today, the use of computational devices is pervasive. Offices, administrations,
financial trading, economic exchange, the control of infrastructure networks, and
a large share of our communication would not be conceivable without the use of
computers anymore. Hence, it would be very surprising, if computers could not
make a contribution to a better understanding of social and economic systems.
While relevant also for the statistical analysis of data and data-driven efforts to
reveal patterns of human interaction [1], we will focus here on the prospects of
computer simulation of social and economic systems. More specifically, we will
discuss the techniques of agent-based modeling (ABM) and multi-agent simulation
(MAS), including the challenges, perspectives and limitations of the approach. In
doing so, we will discuss a number of issues, which have not been covered by the
excellent books and review papers available so far [2–10]. In particular, we will
describe the different steps belonging to a thorough agent-based simulation study,
and try to explain, how to do them right from a scientific perspective. To some
extent, computer simulation can be seen as experimental technique for hypothesis
testing and scenario analysis, which can be used complementary and in combination
with experiments in real-life, the lab or the Web.

This chapter has been prepared by D. Helbing and S. Balietti under the project title “How to
Do Agent-Based Simulations in the Future: From Modeling Social Mechanisms to Emergent
Phenomena and Interactive Systems Design”.

26 2 Agent-Based Modeling
2.1 Why Develop and Use Agent-Based Models?
2.1.1 Potential of Computer Simulation in the Socio-Economic

Sciences
It is well-known that the ways in which social scientists analyze human behavior,
social interactions, and society vary largely. The methods range from qualitative
to quantitative ones, and among the quantitative ones, some communities prefer
detailed models with many variables and parameters, while others prefer simple
or simplified models with a few variables and parameters only. Reference [11]
discusses these different types of system description and their respective advantages
and disadvantages. Overall, each method has its justification, and the choice of
the proper method very much depends on the respective purpose. For example,
the elaboration of applications such as new systems designs often requires a quite
realistic and, hence, detailed description of all relevant aspects. In contrast, simple
models may be used to get a better understanding of how social mechanisms work.
They serve to reduce the complexity of a given system to an extent that allows to
guide our thinking and provide an intuition how certain changes in the system would
affect its dynamics and outcome.
The application of computational models is currently not common in the social
and economic sciences. This is perhaps because many people consider them
as intransparent and unreliable (as compared to analytical methods) and/or as
unsuitable for prediction. These points will be addressed later on. In fact, if
properly done, computer simulations can deliver reliable results beyond the range
of analytical tractability (see Sect. 2.3.6). Moreover, we will show in Sect. 2.4.1 that
prediction is not generally impossible for socio-economic systems and that it is even
not necessary to improve a system (e.g. to reduce instabilities or vulnerabilities),
see Sect. 2.4.2. Besides, the benefit of computational models is not restricted to
prediction. Joshua Epstein, for example, discusses 16 other reasons to build models,
including explanation, guiding data collection, revealing dynamical analogies,
discovering new questions, illuminating core uncertainties, demonstrating tradeoffs,
training practitioners, and last but not least decision support, particularly in crisis
situations [12].
In fact, computer models can naturally complement classical research methods
in the socio-economic sciences. For example, they allow one to test whether
mechanisms and theories used to explain certain observed phenomena are sufficient
to understand the respective empirical evidence, or whether there are gaps or
inconsistencies in the explanation. Moreover, they allow one to study situations,
for which analytical solutions cannot be found anymore, and to go beyond the
idealizations and approximations of simple models. Without the exploration of
model behaviors that can only be numerically determined, scientific analysis is often
restricted to unrealistic models and to situations, which may be of little relevance for
reality. For example, the financial crisis may have been the result of approximations
2.1 Why Develop and Use Agent-Based Models? 27
and simplifications of economic models, which were not sufficiently justified (for a
more detailed discussion of this point see [13]).
2.1.2 Equation-Based Versus Agent-Based Approach
Today, computer-simulation in the natural sciences and engineering mostly rely on

equation-based modeling (e.g. of the dynamics of gases, fluids, or solid bodies).
Such an approach would certainly be hard to transfer to the social sciences, as
most system behaviors have not been formalized mathematically. A method that
seems to be more suited for the computer simulation of socio-economic systems is
agent-based modeling (ABM) [2–6]. The corresponding computational technique
is called multi-agent simulation (MAS) or agent-based computational modeling
(“ABC modeling”). Depending on the problem of interest, agents may for example
represent individuals, groups, companies, or countries and their interactions. The
behaviors and interactions of the agents may be formalized by equations, but more
generally they may be specified through (decision) rules, such as if-then kind of
rules or logical operations. This makes the modeling approach much more flexible.
Besides, it is easily possible to consider individual variations in the behavioral rules
(“heterogeneity”) and random influences or variations (“stochasticity”).
To give a clearer picture, let us provide below a list of properties, which may be
given to an agent representing an individual:
• Birth, death, and reproduction
• Individual needs of resources (e.g. to eat and drink)
• Competition and fighting ability
• Toolmaking ability (e.g. the possibility to grow food, hunt etc.)
• Perception
• Curiosity, exploration behavior, ability for innovation
• Emotions
• Memory and future expectations
• Mobility and carrying capacity
• Communication
• Learning and teaching ability
• The possibility of trading and exchange
• The tendency to have relationships with other agents (e.g. family or friendship
ties etc.)
Agent-based computational models also appear ideal to study interdependencies
between different human activities (both, symbiotic and competitive relationships)
[14]. Therefore, they can shed new light on social and economic systems from
an “ecological” perspective [15]. In fact, evolutionary ecological models can also
reflect the feature of steady innovation, which is typical for socio-economic systems.
Moreover, such models appear particularly suited to study the sustainability and
resilience of systems. Finally, they can be well combined with other simulation
methods used in the natural and engineering sciences, including statistical physics,
biology, and cybernetics.
2.1.3 Scientific Agent-Based Models Versus Computer Games
Agent-based computational modeling is also well suited for visualization. Here, it

makes sense to distinguish agent-based models in science from computer games.
While the latter may actually be based on agent-based models, they are often
oriented at believability, i.e. at appearing as realistic as possible. However, computer
games usually do not care too much about making realistic assumptions. That is,
while the outcome may look realistic, the underlying mechanisms may not be well
justified. As a consequence, they may not be suited to understand the outcome of
the simulation, to draw conclusions, or to make predictions. Implications outside the
exact settings that the game was prepared for are likely to be unreliable. Therefore,
in many cases, such computer simulations will not produce useful knowledge
beyond what has been put into the model.
Scientific agent-based models, in contrast, do often not invest into believability.
That is, they intentionally make simplifications and may, for example, represent
people by circles, or purposefully restrict themselves to very few properties from
the list given in Sect. 2.1.2. Instead of focusing on a plausible appearance, they
try to represent a few characteristic features (such as certain kinds of interaction
mechanisms) more realistically. Such computer simulations should enable one to
draw conclusions about previously unexperienced scenarios, i.e. they should be in
reasonable agreement with later empirical observations or experimental results. In
other words, scientific simulations are more focused on getting the processes rather
than the visual representation right. They are interested in explanatory power.
Finally, agent-based simulations for engineering applications are often located
somewhere between the two archetypical cases discussed above. However, rather
than on the basis of the level of detail and believability, it also makes sense to classify
models as follows:
• Physical models assume that individuals are mutually reactive to current (and/or
past) interactions.
• Economic models assume that individuals respond to their future expectations
and take decisions in a selfish way.
• Sociological models assume that individuals respond to their own and other
people’s future expectations (and their past and current experiences as well).
2.1.4 Advantages of Agent-Based Simulations
Agent-based simulations are suited not only to reflect interactions between different
individuals (and other entities). They allow one to start off with the descriptive
power of verbal argumentation and to determine the implications of different
hypotheses. From this perspective, computer simulation can provide “an orderly
formal framework and explanatory apparatus” [16]. Other favorable features of
agent-based simulations are [17]: modularity, great flexibility, large expressiveness,
and the possibility to execute them in a parallelized way.
Agent-based models can be combined well with other kinds of models. For
example, when simulating the interaction with the environment, the environment
may be represented by a discrete or continuous field. Such an approach is pursued
within the framework of active walker models [18, 19]. One can easily couple
agent-based models with continuum models, such as gas-kinetic or fluid-dynamic
models. Such an approach is, for example, used to simulate the evacuation of people
in scenarios where poisonous gas spreads in the environment [20, 21]. A similar
approach would be applied, when weather, environmental, or climate simulations
would be combined with models of human response to the respective external
conditions.
In certain contexts, for reasons of computational efficiency it may also be
reasonable to replace an agent-based by an aggregate (“macroscopic”) simulation
approach. For example, traffic flows can not only be well represented by a car-
following (agent-based) model, but also by a fluid-dynamic one [22]. It is even
possible to relate the car-following models with fluid-dynamic ones in an analytical
way [23]. In other words, it is possible to construct a mathematical bridge between
the micro- and macro-level of description [24] – something which would be very
nice to have for economics and other fields as well.
In the economic sciences, multi-agent computer simulations make it possible to
overcome limitations of the theoretical concept of homo economicus (the “perfect
egoist”) [25], by relaxing idealized assumptions that are empirically not well enough
supported. They also offer a possibility to go beyond the representative agent models
of macroeconomics [26], and to establish a natural link between the micro- and
macro-level description, considering heterogeneity, spatio-temporal variability, and
fluctuations, which are known to change the dynamics and outcome of the system
sometimes dramatically [13].
Finally, agent-based simulations are suited for detailed hypothesis-testing, i.e.
for the study of the consequences of ex-ante hypotheses regarding the interactions
of agents. Insofar, one could say that they can serve as a sort of magnifying
glass or telescope (“socioscope”), which may be used to understand our reality
better. It is usually no problem to apply methods from statistics and econometrics
to simulation outcomes and to compare simulation results with actual data (after
processing them in a way reflecting the measurement process). Moreover, by
modeling the relationships on the level of individuals in a rule-based way, agent-
based simulations allow one to produce characteristic features of the system as
emergent phenomena without having to make a priori assumptions regarding the
aggregate (“macroscopic”) system properties.
2.1.5 Understanding Self-organization and Emergence
Agent-based simulations are a suitable tool to study complex systems. Complex sys-
tems are systems with many interacting entities and non-linear interactions among
them. Such systems may behave in a number of interesting and often unexpected
(sometimes even paradoxical) ways, which justifies to call them complex:
• They may have several stationary states (a phenomenon known as “multi-

stability”), and the resulting outcome may depend on the previous history (such
as the size of occurring perturbations, the “initial state”, etc., and such history-
dependencies are often called “hysteresis effect”) (see Figs. 2.1 and 2.3).
• They may be “out of equilibrium” and behave in non-stationary ways.
• They may “self-organize”, showing periodic or non-periodic oscillations,
“chaotic” or “turbulent” behavior, or spatio-temporal pattern formation (such
as stop-and-go waves in traffic flows).
• They are often robust to small perturbations, i.e. “relax” to their previous
behavior (the “stable attractor”).
• Consequently, they often resist external manipulation or control attempts.
• However, at so-called “tipping points”, small influences may cause a sudden and
often unexpected “systemic shift” (“phase transition”), after which the system
behaves substantially different (see Fig. 2.2).
• More generally, they may show new, “emergent” properties, which cannot be
understood from the properties of their system elements (“the system is more
than the sum of its parts”) (see Fig. 2.1).
• Correlations may determine the system dynamics, and neglecting them may lead
to completely wrong conclusions.
• During systemic shifts (so-called “phase transitions”) or due to a phenomenon
called “self-organized criticality” (SOC), cascading effects on all scales (i.e. of
any size) may occur, so that local factors may have a systemic (“global”) impact
(“critical phenomena”).
• Therefore, “extreme events” may happen with probabilities much higher than
expected according to a normal distribution, and are distributed according to
“(truncated) power laws” or other “fat tail distributions”.
• The system may have features such as reproduction, innovation, reinforcement
learning, an expectation-related dynamics, etc.
• There may be singularities after a finite time.
Many of the above features are results of strong interactions in real or abstract space,
or of network interactions within the system. Such interactions can often lead to
counter-intuitive behaviors [29, 30]. Here are a number of examples (most of them
related to traffic systems, as they are well-known to everybody):
• Even when all drivers try to drive fluently, vehicles are sometimes stopped by
“phantom traffic jams” (i.e. jams without an obvious reason such as an accident
or bottleneck) [31, 32].
• Stop-and-go traffic may occur despite the best attempts of drivers to move ahead
smoothly [33, 34].
• Even when the maximum road capacity is not reached, a temporary reduction in
the traffic flow can cause a lasting traffic jam [35].
• Traffic jams do not occur in the bottleneck area, but upstream of it.
• Under certain conditions, speed limits can speed up traffic [31]. Similar “slower-
is-faster effects” occur in urban traffic control, logistic systems, administrative
processes, etc., i.e. delays at appropriate times and of suitable durations may
reduce overall waiting times [36, 37].
a IDM
MLC
b IDM
c IDM
OCT
SGW V (km / h)
V (km / h) V (km / h) V (km / h) 160
0 140
0 0 120
50 50 50 100
100 100 100 80
60
40
–5 20
–5 -5 0
Lo
Lo
Lo
ca
ca
ca
0 1 0 1
tio
tio
)
tio
e (h )
n
e (h )
e (h
n
Tim
(km
Tim Tim
(km
(km
0 0 0
)
)
d IDM
WSP
e IDM
PLC
f IDM
HCT V (km / h)
V (km / h) V (km / h) V (km / h) 160
0 0 140
0 120
50 50 50 100
100 100 100 80
60
–5 –5 40
20
–5 0
Lo
Lo
ca
Lo
0 1 0 1 1
ca
tio
ca
(h) 0
tio
n
)
Time e (h )
tio
e (h
(km
Tim Tim
n
0 0
(km
(km
)
Fig. 2.1 Top: Freeway traffic constitutes a dynamically complex system, as it involves the non-
linear interaction of many independent driver-vehicle units with a largely autonomous behaviour.
Their interactions can lead to the emergence of different kinds of traffic jams, depending on
the traffic flow on the freeway, the bottleneck strength, and the initial condition (after [27]): a
moving cluster (MC), a pinned localized cluster (PLC), stop-and-go waves (SGW), oscillating
congested traffic (OCT), or homogeneous congested traffic (HCT). The different traffic patterns
were produced by computer simulation of a freeway with an on-ramp at location x D 0 km using
the intelligent driver model (IDM), which is a particular car-following model. The velocity as a
function of the freeway location and time was determined from the vehicle trajectories (i.e. their
spatio-temporal movement). During the first minutes of the simulation, the flows on the freeway
and the on-ramp were increased from low values to their final values. The actual breakdown of
free traffic flow was triggered by additional perturbations of the ramp flow. Bottom: In pedestrian
counterflows one typically observes a separation of the opposite walking directions. This “lane
formation” phenomenon has been reproduced here with the social force model [28]
• While pedestrians moving in opposite directions normally organize into lanes,

under certain conditions a “freezing-by-heating” effects or intermittent flows may
occur [38].
• The maximization of system efficiency may lead to a breakdown of capacity and
throughput [30].
2
|C| /|B| = (1−f)/f→ ←|C| / |B| = f / (1–f)
1.5
Regime
|C| / |B|
shifts
1
Sy
st
0.5 em
ic
ch
an
ge amics
m dyn
0 Syste
0 0.2 0.4 0.6 0.8 1
f
Fig. 2.2 Left: Illustration of the parameter-dependent types of outcomes in the social norms game
of two interacting populations with incompatible preferences (after [151]). f is the relative size of
population 1 and B D b c < 0 and C D b C c > 0 are model parameters, which depend on the
benefit b of showing the individually preferred behavior, while c is the reward of conforming
with the behavior of the respective interaction partner. Small arrows illustrate the vector field
.dp=dt; dq=dt / as a function of the fraction p of individuals in population 1 showing their
preferred behavior (on the horizontal axis) and the corresponding fraction q in population 2 (on
the vertical axis). Empty circles stand for unstable fix points (repelling neighboring trajectories),
black circles represent stable fix points (attracting neighboring trajectories), and crosses represent
saddle points (i.e. they are attractive in one direction and repulsive in the other). The basins of
attraction of different stable fix points are represented in different shades of grey (colors) [green D
population 1 sets the norm, blue D population 2 sets the norm, yellow D each population does what
it prefers, red D nobody shows the preferred behavior]. The solid red lines indicates the threshold
at which a continuous phase transition takes place, dashed lines indicate discontinuous phase
transitions. Right: When a complex system is manipulated (e.g. by external control attempts), its
system parameters, stability, and dynamics may be affected. This figure illustrates the occurrence
of a so-called “cusp catastrophe”. It implies a discontinuous transition (“regime shift”) in system
dynamics
• Sometimes, the combination of two dissatisfactory solutions can be the best

solution [39, 40].
• There is clearly a global diversity of opinions and behaviors, despite a rather
strong tendency of local convergence [41].
• In repeated interactions, the “wisdom of crowds” mechanism may lead to
collective error [42].
• One can often find cooperation among selfish individuals in social dilemma
situations (where it seems more profitable to exploit the cooperativeness of
others) [39].
• People often do not show the behavior they like and often show behaviors they
do not like [43].
Generally, there is still a serious lack of understanding regarding the connections
between the structure, dynamics, and function of complex networks such as techno-
socio-economic-environmental systems. Therefore, emergence often has an element
Small Perturbation Large Perturbation
OCT, OCT,
ΔQ = Qout–Qc3
HCT
Qup(vehicles / h / lane)
Qup(vehicles / h / lane)
SGW SGW
ΔQ = Qout–Qc4
HCT
MLC
Q
to
t = Qup = Qc1 Q
Q to
m t =
ax Q
ou
Q PLC t
FT to
t =
Q
FT c1
ΔQ (vehicles / h / lane) ΔQ (vehicles / h / lane)
Fig. 2.3 Schematic (idealized) phase diagrams of the traffic patterns expected for a particular
traffic model as a function of the freeway traffic flow Qup upstream of a ramp bottleneck and
the on-ramp flow Q (after [27]). The left figure is for negligible, the right figure for large
perturbations in the traffic flow. The situation for medium-sized perturbations can lie anywhere
between these two extremes. Different colors correspond to different possible traffic states (FT
D free traffic flow, PLC D pinned localized cluster, MLC D moving localized cluster, OCT D
oscillating congested traffic, SGW D stop-and-go waves, HCT D homogenous congested traffic,
see Fig. 2.1). The equations next to the separating lines can be analytically calculated, but are not
important here. For details see [27])
of surprise. So far, for example, we do not understand emotions and consciousness,

and we cannot even calculate the “fitness” of a behavioral strategy in the computer.
The most exciting open puzzles in science concern such emergent phenomena. It
would be interesting to study, whether social and economic phenomena such as
trust, solidarity, and economic value can be understood as emergent phenomena
as well [13]. Agent-based simulations appear to be a promising approach to make
scientific progress in these areas.
2.1.6 Examples of Agent-Based Models
The method of agent-based modeling is very versatile. It may be applied, for

example, to the following problems:
• Social influence and opinion formation [41, 44]
• Coalition formation [45, 46]
• Collective intelligence [47]
• Social networks [48–50]
• Group dynamics [51]
• Social cooperation [52, 53]
• Social norms [14, 54, 55]
• Social conflict [56, 57]
• Financial markets [58–60]

• Competition and cooperation between firms [61, 62]
• Micro-economic models [63, 64]
• Macro-economic models [65, 66]
• Organization and managerial decisions [67]
• Migration [68]
• Agglomeration and segregation [69, 70]
• Urban and regional development [71–73]
• Traffic dynamics [74, 75]
• Crowd dynamics [76, 77]
• Systemic risks in socio-economic systems [78, 79]
and more [80–84].
2.1.7 Social Super-Computing
Multi-agent simulations are well suited for parallelization. Therefore, using super-
computing power, it is possible in principle to run agent-based simulations with
millions (or, in future, even billions) of agents. The following examples give an idea
of the state-of-the art:
• A first successful application area was large-scale traffic simulation. The TRAN-
SIMS project [85–87] of the Los Alamos National Institute (LANL), for
example, has created agent-based simulations of whole cities such as Dallas [88]
or Portland [89]. The approach has been recently extended to the simulation
of the travel behavior of the 7.5 million inhabitants of Switzerland [90, 91].
These simulations are obviously based on parallel computing. They generate
realistic individual activity patterns according to detailed statistical panel data
(“travel diaries”) [92, 93], which are nowadays complemented by GPS data
and public mobility data (e.g. from Greater London Area’s OYSTER Card).
Other extensions look at interconnections between the traffic system and regional
development [72, 73].
• Recent applications are studying contingency plans for large-scale evacuations
of cities [94, 95]. The key aspect here is understanding the interdependency of
infrastructure systems [96, 97] and their vulnerabilities through natural disasters,
terrorist attacks, accidents, and other incidents. For example, the Los Alamos
National Laboratories have already established a Critical Infrastructure Protec-
tion Decision Support System [98]. Its advanced simulation capabilities have
already been extensively used during past emergencies.
• Large-scale simulations have also been applied to study and predict the spreading
of diseases. While previous epidemic spreading models such as the SIR model
[99–101] have neglected spatial interaction effects, recent models [102] take into
account effects of spatial proximity [20], air traffic [103] and land transport [104],
using TRANSIMS and other traffic simulation models. The current scientific
2.2 Principles of Agent-Based Modeling 35
development also tries to take into account behavioral changes, which may
reduce the spreading rate of diseases.
• Furthermore, there are attempts to model financial markets with agent-based
simulations. Two examples for this are the Santa-Fe Stock Market Simulator
[105] and U-mart [106]. Recent attempts are even heading towards the simulation
of the whole economic system (see for example the EU project EURACE
[107]). Other simulation studies are trying to understand the evolution of
cooperation [39], social norms [14, 54], language [108–110], and culture [3].
Such simulations explore the conditions under which trust, cooperation and other
forms of “social capital” can thrive in societies (see also [13, 111]). They also
show that the crust of civilization is disturbingly vulnerable. Such simulations
can reveal possible causes of breakdowns of social order. Examples for events
where this actually happened are as diverse as the war in former Yugoslavia,
lootings after earthquakes or other natural disasters, or the violent demonstrations
we have recently seen in some European countries.
It appears logical that supercomputing will be ultimately moving on from applica-
tions in the natural and engineering sciences to simulations of social and economic
systems, as more and more complex systems become understandable and the
required data become available. And it is obvious that virtual three-dimensional
worlds (such as Google Earth) are waiting to be filled with realistic life.
2.2 Principles of Agent-Based Modeling
After describing the potential of multi-agent simulations, let us now discuss the
principles of how to craft agent-based models. A thorough scientific study involves
a number of steps:
• First, one should clearly describe the evidence to be explained by the respective
study. What are the empirical or experimental data or observations to be repro-
duced, or what are the “stylized facts”, i.e. the simplified, idealized properties of
the system under consideration?
• Second, one should explain what is the purpose of the simulation? To understand
a phenomenon? To get a more accurate description? To make predictions? To
develop an application (e.g. a new traffic control)? In the social sciences, it is
common to formulate a scientific puzzle, i.e. to describe a problem that is hard to
understand and why. This could be an unexpected or even paradoxical individual
or system behavior. Emergent system behaviors are particularly interesting
candidates for the formulation of such a puzzle (“scientific mystery”).
• Next, one needs to decide how to choose the agents in the model. For example,
when the competition of companies shall be studied, it may not be necessary
to simulate all employees of all companies. It may be sufficient to choose the
companies as the agents of the model. In fact, it can be shown mathematically
(e.g. by eigenvalue analysis) that mutually coupled agents may jointly behave
like one entity, i.e. one agent. An example for this is the quasi-species concept in
the theory of evolution [112].
• After specifying the agents, one should formulate hypotheses regarding the
underlying socio-economic processes or fundamental mechanisms leading to the
particular system behavior that needs to be explained. Ideally, these mechanisms
should be sociologically or economically justified, i.e. there should be some
empirical evidence for the mechanisms on which the agent-based model is
based. The transfer of models from other sciences (such as spin, epidemic, or
synchronization models) requires particular justification beyond saying that the
resulting system behavior is reminiscent of features that have been observed
elsewhere.
• When specifying the mechanisms underlying the multi-agent simulation, one
should not put into the model assumptions what one wants to explain. The
mechanisms on which the multi-agent simulations are based should be (at least)
one level more elementary than the evidence to be understood. For example,
the rich-gets-richer effect [113] may be used as an ingredient, if class formation
shall be described. Moreover, “homophily” [114] may be assumed in models of
coalition formation or solidarity. Moreover, social network characteristics may
be used to explain the spreading of behaviors [115–117].
However, if the income distribution is to be explained, it is favorable not to
start with the rich-gets-richer effect, but instead with a mechanism that is purely
random and not biased in favor of anybody in the beginning. Moreover, even if
this is not realistic, it would be interesting to start the computer simulation with
identical wealth of everybody [118]. Furthermore, if social segregation is to be
explained, one should not assume “homophily” already, but to let it evolve in a
system that starts off with identical and non-segregated individuals [39, 119].
Finally, if group formation is to be explained, social network characteristics
should not be assumed as an input [41]. They should, for example, result from
certain rules regarding the formation and deletion of social ties [120, 121].
• Last but not least, one should compare the computer simulation results with
the empirical evidence. Here, one should avoid to be selective, i.e. one should
state what are the features that are correctly reproduced, and which ones are not.
Pointing out the limitations of a model is equally important as underlining its
explanatory power.
Note that, even though linear models may be sensitive to parameter variations
or perturbations, they cannot explain self-organization or emergent phenomena
in socio-economic systems. This underlines that the consideration of non-linear
interaction mechanisms is crucial to understand many observations in social and
economic systems [13].
When specifying the properties of agents and their interactions, it makes sense
to select some from the list given in Sect. 2.1.2, but (by far) not all. As was
pointed out before, the goal of scientific computer-simulation is usually not a
realistically looking computer game, but the explanation of a set of observations
from a minimum set of assumptions.
2.2 Principles of Agent-Based Modeling 37
It is definitely not obvious how each of the points on the list in Sect. 2.1.2
is modeled best. Typically, there are a number of plausible alternatives. To gain
an understanding and intuition of the studied system, simple assumptions (even
if idealized) are often preferable over detailed or complicated ones. For example,
rather than assuming that individuals would strictly optimize over all options when
decision are taken (as “homo economicus” would do), it seems justified to use
simple decision heuristics, as evidence from social psychology suggests [122].
However, it would obviously be interesting to compare the implications of both
modeling approaches (the classical economics and the heuristics-based one).
2.2.1 Number of Parameters and Choice of Model
When formulating the agent-based model, a number of aspects should be consid-

ered, as discussed in the following. The adequate number of model parameters and
variables depends on the purpose of the model and its required degree of realism
and accuracy. A model for the elaboration of practical applications tends to have
more parameters than a model aimed at the fundamental understanding of social
or economic mechanisms. However, one should keep in mind that parameter-rich
models are hard to calibrate and may suffer from over-fitting. Hence, their predictive
power may not be higher than that of simple models (see Sect. 2.3.2).
Generally, one can say that
• A model with meaningful parameters (which have a clear interpretation and are
measurable) should be favored over a model with meaningless fit parameters.
• The same applies to a model with operational variables as compared to a model
containing variables that cannot be measured.
• Given the same number of parameters, an explanatory model should be prefered
over a purely descriptive (“fit”) model.
• In case of a comparable goodness of fit in the model calibration step, one should
choose the model with the better predictive power (i.e. which better matches data
sets that have not been used for calibration).
• Given a comparable predictive power of two models, one should select the
simpler one (e.g. the one with analytical tractability or with fewer parameters)
according to Einstein’s principle that a model should be as simple as possible,
but not simpler.
The goodness of fit should be judged with established statistical methods, for
example with the adjusted R-value or similar concepts considering the number of
model parameters [123, 124]. This adjusted formula tries to compensate for the fact
that it is easier to fit a data set with more parameters. For a further discussion of
issues related to the calibration and validation of agent-based models see Sect. 2.3.2.
2.3 Implementation and Computer Simulation of Agent-Based

Models
2.3.1 Coding Multi-agent Simulations and Available Software

Packages
Agent-based simulations may be programmed from scratch in most computer

languages (e.g. C, C++, Java, etc.). This may be favorable from the perspective
of computational efficiency. However, there are also more user-friendly ways of
specifying agent-based models. This includes Swarm [125] and Repast [126], which
provide low-level libraries, as well as graphical modelling environments such as
Netlogo [127] or Sesam [128], which are more suited for beginners. Another
recent software tool for the simulation of large crowds is MASSIVE [129]. For a
comprehensive comparison of the features of available packages for agent-based
simulations see [130, 131].
2.3.1.1 Documentation
No matter how multi-agent simulations are implemented, it is crucial to document

the code and the simulation scenarios well. Not only would the code be useless for
any future collaboration partner, if it is not well documented. It is also quite easy to
lose the overview over all the changes one may have experimented with. There are
several publications, which report results that do not fit the model discussed in the
paper. Such accidents happen often for one of the following reasons:
• The computer code has not been properly debugged – checked for mistakes,
see Sect. 2.3.1.3 (e.g. variables are not properly initiated, or local variables are
overwritten by global variables, or variables are mixed up, or there is a mismatch
of different types of variables, such as integer values and real numbers).
• Changes are made in the code, but not well documented and forgotten.
• Readimade routines from software libraries are applied, but the preconditions for
their use are not satisfied.
2.3.1.2 Plausibility Tests
The most famous mistake in a computer code was the explosion of a space
shuttle caused by confusing a semicolon and a comma. However, not always do
programming mistakes become evident in such a dramatic way. In fact, the danger
is that they remain unnoticed. Computer codes do not automatically generate errors,
when something is incorrectly implemented. On the contrary, most computer codes
produce results independently of how reasonable the code or input is. Therefore, one
should better have a healthy distrust towards any computational output. Computer
codes always require verification and proper testing, and it is crucial to check
2.3 Implementation and Computer Simulation of Agent-Based Models 39
the plausibility of the output. Section 2.3.4.2 discusses some ways how this can
be done.
2.3.1.3 Common Mistakes and Error Sources
To avoid errors in the computer code, the following precautions should be taken:
• Structuring the computer code into subroutines allows one to keep a better
overview (and do a thorough, component-wise testing) as compared to one
monolithic code.
• Different variants of the computer code should be distinguished with different
version numbers,1 and the main features, particularly the changes with regard
to the last version, should be clearly documented in the beginning of the code
and, ideally, shown on the computer display and in the output file, when the
code is run.
• One should not include the parameters of the model in the computer code itself
(“hardcoding”), but instead they should be read from a separate input file, which
should also have an own version number. This parameter file should contain
information, with what version of the computer code the parameter file needs
to be run. It makes sense to write out the parameters and their values in the
output file, to reduce unnoticed errors (such as adding a new parameter, which is
erroneously initiated with the value of another parameter).
• Programming languages initiate variables and parameters in different ways
(and some of them with random contents), so that it may easily remain
unnoticed when a variable or parameter was not set. Therefore, all parameters
and variables should be initialized immediately, when they are introduced in
the code.
• Moreover, one should not make several significant changes at a time. For
example, it is common to change only one parameter value and compare the
new result with the previous one, before another parameter is changed. This
helps to check the plausibility of the computational results.
2.3.2 Specification of Initial and Boundary Conditions,

Interaction Network and Parameters
The simulation of agent-based models requires a number of specifications before

the simulations can be run:
1
There exist a number of software packages aimed at supporting developers in versioning the code.
They automatise several operations such as assigning sequential versions numbers, comparing
different versions of files, undo or merging changes on the same files, etc. For examples, see
[132–135].
• It is necessary to define the interaction network. For example, agents may

interact in space, and their interaction frequency or strength may depend on
their distance. However, in many cases, spatial proximity is just one factor or
even irrelevant. In many cases, one needs to consider friendship networks or
other interaction networks. It is well-known that the statistical properties of the
interaction network may matter for the dynamics of the system dynamics. As a
consequence, it may also be needed to run the simulation for different interaction
networks. For example, interactions in square grids and in hexagonal grids may
sometimes lead to qualitatively different outcomes. Moreover, the number of
interaction partners may be relevant for the system dynamics, as may be a
heterogeneity in the number of interaction partners or the existence of loops in
the interaction network.
• One needs to specify the initial values of all variables (and if memory effects are
considered as well, also their previous history). In many cases, it is common to
make simplifying assumptions (e.g. all individuals are characterized by identical
initial values, or more realistically, the initial variables are assumed to vary
according to a particular distribution (e.g. a normal distribution). It is advised
to test the sensitivity of the model with respect to the initial condition.
• Furthermore, boundary conditions may have to be specified as well. For example,
if a simulation of agents is performed in space, one needs to decide how the
rules should look like at the boundary. One may decide to use a finite world
such as a chess board. However, this may cause artificial behavior close to
the boundaries of the simulation area. Therefore, in many cases one assumes
“periodic” boundary conditions, which corresponds to a simulation on the
surface of a torus. This is often simpler than the simulation on the surface of
a sphere. Note, however, that an improper choice of the boundary conditions can
sometimes produce artifacts, and that the simulated system size may affect the
result. Therefore, one needs to have an eye on this and may have to test different
specifications of the boundary conditions.
• Finally, it is necessary to specify the open model parameters, i.e. to calibrate the
model. If the initial and boundary conditions, or the structure of the interaction
network cannot be independently measured, they have to be treated as model
parameters as well. In case there are enough empirical data, the parameters can
be determined by minimizing an “error function”, which measures the difference
between simulation results and empirical data. Note that the choice of a suitable
error function may be relevant, but non-trivial. An improperly chosen function
may not be able to differentiate well between good and bad models. For example,
minimizing the error between actual and simulated trajectories of pedestrians
does not manage to distinguish well between a simple extrapolation model and
a circular repulsion model for pedestrians, although the latter avoids collisions
and is much more realistic [136]. Or maximizing throughput may produce some
unrealistic effects, such as the occurrence of pedestrians or drivers, who get
trapped in a certain place [137], while a minimization of delay times ensures
that everybody gets forward. Therefore, not only the model, but also the error
function must be well chosen.
If there are not enough (or even no) empirical or experimental measurements
to calibrate the model with, on can still try to find parameters, which “do the job”,
i.e. which deliver plausibly looking simulation results. Such a qualitative simu-
lation approach has,for example, been pursued in the early days of pedestrian
modeling. Despite the simplicity of this method, it was surprisingly successful
and managed to reproduce a number of striking self-organization phenomena
observed in pedestrian crowds [136].
Finally note that, even when empirical or experimental data are available, the
size of empirical data sets typically does not allow one to determine the model
parameters accurately. Particularly, if the model contains many parameters,
the reliability of the parameter estimates tends to be poor. In the worst case,
this can imply dangerously misleading model predictions. To get an idea of
the confidentiality intervals of parameters, one should therefore determine all
parameter combinations which are compatible with the error bars of the empirical
data to be reproduced.
2.3.2.1 Model Validation
A high goodness of fit during model calibration does not necessarily imply a high
predictive power [138,139], i.e. a good fit of new data sets. In many cases, one faces
the problem of over-fitting (i.e. the risk of fitting noise or irrelevant details in the
data). Therefore, it is necessary to determine the “predictive power” of a model by
a suitable validation procedure.
In the ideal case, the model parameters can be measured independently (or can, at
least be estimated by experts). If the parameters have a concrete meaning, it is often
possible to restrict the parameters to a reasonable range. In case of non-meaningful
fit parameters, however, this is not so simple (or even unfeasible).
One way of validating a model is to subdivide the empirical or experimental
data into two non-overlapping parts: a calibration and a validation dataset. The
calibration dataset is used to determine the model parameters, and the validation
dataset is used to measure the goodness of fit reached with the model parameters
determined in the calibration stage. In order to make this calibration and validation
procedure independent of the way in which the original dataset is subdivided,
the whole procedure should be performed either for all possible subdivisions into
calibration and validation datasets or for a representative statistical sample of all
possibilities. As each of these subdivisions delivers a separate set of parameters in
the calibration step, this results in a distribution of model parameters, which are
consistent with the data. From these distributions, one can finally determine average
or most likely model parameters as well as confidence intervals. Furthermore, the
distribution of goodness-of-fit values reached in the related validation steps reflects
the predictive power of the model.
Another way of judging the power of a model is to determine the number of
stylized facts that a model can reproduce. It sometimes makes sense to prefer a
model that reproduces many different observations qualitatively well (for example,
different observed kinds of traffic patterns) over a model whose goodness of fit is
quantitatively better (for example, in terms of reproducing measured travel times).
This applies in particular, if the model, which appears to be quantitatively superior,
is not well consistent with the stylized facts [27]. To distinguish models of different
quality, it can also be useful to measure the goodness of fit with several different
error functions.
2.3.2.2 Sensitivity Analysis
As empirically or experimentally determined parameter values have a limited

accuracy, one should also carry out a sensitivity analysis. For this, the simulation is
performed with modified parameters to determine, how robust the simulation results
are with respect to the choice of parameters. The sensitivity can be measured with
Theil’s inequality coefficient. Note that the parameters should be varied at least
within the range of the confidence interval that determines the range, within which
the actual parameters may vary according to the calibration dataset.
Besides determining the robustness of the model to the parameter specification,
it is also recommended to test the robustness to the model assumptions themselves.
By simulating variants of the model, one can figure out, which conclusions
stay qualitatively the same, and which ones are changing (e.g., if the network
characteristics or the system size or the learning rule are modified).
2.3.3 Performing Multi-agent Simulations
2.3.3.1 Choice of the Time Discretization
The first problem relates to the choice of the time step t. In models that are discrete
in time (such as deterministic cellular automata), the time step is assumed to be
fixed, and the question may matter or not. When the model contains differential
equations, however, the choice of t is always relevant. The rule is that large
t speed up the simulation, but may lead to wrong results. For example, the
discrete logistic equation xnC1 D rxn .1 xn / may behave very differently from
the continuous one dx=dt D ax.1 x/ (the former one may perform a chaotic
motion, while the latter one evolves smoothly in time).
To determine a suitable time discretization, t is reduced from moderate values
(like t D 1 or 0.1) to smaller values, until the results do not change significantly
anymore (e.g. less than 0.1% when t is chosen five times smaller).
2.3.3.2 Relevance of Considering Fluctuations
A widely recognized fact of socio-economic systems is that they do not behave

deterministically. There are always sources of fluctuations (“noise”), such as errors,
exploratory behavior, or the influence of unknown factors. For this reason, the
probabilistic behavior should be reflected in the model. Neglecting “noise” may lead
to misleading conclusions. For example, the zero-noise case may behave completely
different from a system with just a little bit of noise, no matter how small it may be
[119,140].2 In such cases, the result of the deterministic model without noise should
be considered as artificial and of no practical relevance.
Note that the significance of fluctuations in techno-socio-economic-
environmental systems is often wrongly judged. While large noise usually has a
destructive influence on a system, as expected, it is quite surprising that small noise
intensities can actually trigger structure formation or increase system performance
[39, 141–143].
The implementation of noise in computer simulations is not fully trivial, i.e.
mistakes can be easily made. First of all, no random number generator produces
random numbers in a strict sense, but rather quasi-random numbers. In other words,
there is a certain degree of statistical interdependency between computer-generated
random numbers, and this may create artifacts. The quality of random number
generators can be very different. Therefore, it makes sense to test the number
generator in advance.
Moreover, when adding a Gaussian noise to differential equations (e.g. dx=dt D
f .x; t/), it must be considered that the variance of the related diffusion process
increases linearly with time. This implies that the prefactor of the noise term in
p differential equation is not proportional to the time step t, but proportional to
the
t, which can be easily overlooked. This has to do with the particular properties
of stochastic equations. (As a consequence, the discretized version of the p above
differential equation with noise would be x.t C t/ x.t/ D t f .x; t/ C D t,
where D allows to vary the noise strength.)
2.3.3.3 Simulating Statistical Ensembles
In case of a model containing noise, one needs to determine statistical distribution(s)

(or, more simply, the averages and standard deviations), running the simulation
many times with different sets of random numbers (i.e. with a differently initialized
random number generator3). In other words, running a model with noise a single
time is of relatively little meaning. Sometimes, it may serve to illustrate a “typical”
system behavior (but then, the simulation run should not be selected by the authors;
it should be randomly chosen to avoid a bias through selective presentation).
In any case, a scientific paper should present the statistical distributions over
sufficiently many (typically at least 100 simulation runs) or the mean value and
variability (either one, two or three standard deviations, or quantiles, as it is done by
box plots).
2
The reason for this is that deterministic systems may easily get trapped in local optima, which can
be overcome by noise [119].
3
Some random number generators do this automatically by coupling to the clock.
2.3.3.4 Choice of the Discretization Method
Finally, note that appropriate discretization schemes are also not trivial. Depending
on how they are done, they may be quite inaccurate or inefficient. For example,
the simplest possible discretization of the differential equation dx=dt D f .x; t/,
namely x.t C t/ x.t/ D t f .x; t/, may converge slowly. In case of so-called
“stiff” systems of differential equations, the convergence may be so inefficient (due
to different time scales on which different variables in the system change) that it may
be completely useless. Certain discretizations may even not converge at all towards
the correct solution, if t is not chosen extremely small. (For example, the solution
procedure may face instabilities, which may be recognized by oscillatory values
with growing amplitudes.) In such cases, it might be necessary to choose particular
solution approaches. The situation for partial differential equations (which contain
derivatives in several variables, such as space and time), is even more sophisticated.
Normally, the trivial discretization does not work at all, and particular procedures
such as the upwind scheme may be needed [144].
2.3.3.5 Performance and Scalability
Multi-agent simulations may require a considerable computational power for the

following reasons:
• The dependence of the system dynamics on random fluctuations requires many
simulation runs.
• Multi-agent simulations may involve a large number of agents.
• The simulation of rational behavior (i.e. systematic optimization over future
developments resulting from behavioral alternatives) or of human cognitive
and psychological processes (such as personality, memory, strategic decision-
making, reflexivity, emotions, creativity etc.) is potentially quite resource
demanding.
• The calibration of model parameters to empirical data also requires many
simulation runs for various parameter combinations to determine the parameter
set(s) with the minimum error.
• The parameter space needs to be scanned to determine possible system behaviors
(see Sect. 2.3.4.5).
• Performing scenario analyses (with changed model assumptions) requires many
additional simulation runs. A typical example is the method of “sensitivity
analysis” to determine the robustness of a model (see Sect. 2.3.4.6).
• The visualization of simulation scenarios may require further substantial com-
puter power (see Sect. 2.3.4.4).
For the above reasons, the performance of a multi-agent simulations can matter
a lot. Unfortunately, this often makes it advantageous to write a specialized own
computer program rather than using a general-purpose agent-based simulation
platform. However, the performance can sometimes be increased by a factor of 100,

1,000, or more by a number of measures such as
• Using suitable compiler options to optimize the executable file of the computer
program.
• Avoiding output on a computer display (for example, by writing the numerical
results into a file every 100 or 1,000 time steps and visualizing the results
afterwards).
• Avoiding multiple entangled loops and performing loops in the right order (as is
favorable for read/write operations).
• Avoiding exponential functions, logarithms, and exponents, where possible.
• Applying an efficient numerical integration method together with a proper time
discretization (see Sect. 2.3.3.4).
• Using appropriate parameter values (for example, dividing by small numbers
often causes problems and, considering limitations of numerical accuracy, may
create almost any output).
• Starting with well-chosen initial conditions (e.g. an approximate analytical
solution).
• Considering that there are simple ways of determining certain quantities (e.g. the
standard deviation can be easily determined from the sum and sum of squares
of data values; a moving average can be determined by adding a new value
and subtracting the last value; or an exponential average can be determined by
multiplying the previous value of the average with a factor q < 1 and adding
.1 q/ times the new value).
The performance is often measured by the analysis of how the required compu-
tational time increases with the number N of agents in the system. In the ideal
case, such a scalability analysis gives a linear (or constant) dependency on the
“system size” N . In many cases, however, the computational time scales like a
polynomial, or, what is even worse, like an exponential function. Hence, evaluating
how a computer code scales with system size allows one to distinguish efficient
implementations from inefficient ones. In computer science, the performance of
algorithms is expressed by the “complexity” of an algorithm. For example, the term
NP-hard refers to an algorithm which does not scale polynomially, which means
that the computer time required for simulations explodes with the system size. In
such cases, only moderate system sizes are numerically tractable on a PC. Larger
systems may still be treated by parallelization of the computer code and parallel
processing on dozens or hundreds of processors, but NP-hard problems can be too
demanding even for the biggest supercomputers. However, it is sometimes possible
to reduce the complexity considerably by applying reasonable approximations. For
example, the simulation of pedestrian crowds can be significantly accelerated by
assuming that pedestrians do not interact if their distance is greater than a certain
value. Moreover, many optimization problems can be approximately solved by using
suitable “heuristics” [122, 145–147].
2.3.4 Presentation of Results
2.3.4.1 Reproducibility
As it is customary in other scientific areas, the result of multi-agent simulations

must be presented in a way that allows other scientists to reproduce the results
without having to ask the authors for details. In the ideal case, the source code
underlying the computer simulations is published as supplementary information in
a well documented form.
In order to be reproducable, a publication must contain all the information
discussed in Sect. 2.3.2 (including the initial and boundary condition, kind of
interaction network, and model parameters). Furthermore, it must be specified how
the noise was implemented. The update rule (such as parallel or random sequential
update) and the order of update steps must be provided as well as the full set of rules
underlying the agent-based model. Any relevant approximations must be pointed
out, and it may make sense to specify the numerical solution method and the way,
in which random numbers and statistical distributions were produced.
For the sake of comfort, one should consider to provide parameter values in
tables or figure captions. Moreover, besides specifying them, it is desirable to
use meaningful names for each parameter and to explain the reasons, why the
rules underlying the multi-agent simulation, certain initial and boundary conditions,
particular network interactions, etc. were used.
2.3.4.2 Plausibility Considerations
As underlined before, there are quite a number of mistakes that can be made in
multi-agent simulations (see Sects. 2.3.1.3 and 2.3.2). Therefore, the computer code
and its single subroutines should be carefully checked. Moreover, it should be
described what plausibility checks have been performed. For example, as has been
pointed out in Sect. 2.3.1.2, the model may have exact or approximate solutions in
certain limiting cases (e.g. when setting certain parameters to 0, 1 or very large
values).
Furthermore, the computational results should have the right order of magnitude,
and change in a plausible way over time or when the model parameters are modified.
In addition, one should take into account that some variables are restricted to a
certain range of values. For example, probabilities must always be between 0 and 1.
Besides, there may be so-called “constants of motion”. For example, probabilities
must add up to 1 at any point in time, or the number of vehicles in a closed road
network must stay the same. There may also be certain quantities, which should
develop monotonously in time (e.g. certain systems have an entropy or Lyapunov
function). All these features can be used to check the plausilibilty of simulation
results. It may be enough to determine the values of such quantities every 1,000
time steps.
Finally, any unexpected results must be tracked back to their origin, to make
sure they are well understood. Seemingly paradoxical results must be carefully
studied, and their origins and underlying mechanisms must be clearly identified and
explained.
2.3.4.3 Error Bars, Statistical Analysis, and Significance of Results
Like empirical data, the data resulting from multi-agent simulations should be
subject to statistical analysis. In particular, it is not sufficient to present single
simulation runs or mean values of many simulation runs. No matter how far apart
two mean values may be, this does not necessarily mean that the difference is
statistically significant. The judgment of this requires a proper statistical analysis
(such as a variance analysis) [148].
A minimum requirement to judge the significance of results is the presentation of
error bars (and it should be stated whether they display one, two, or three standard
deviations). Box plots (i.e. presenting the median, minimum and maximum value,
and quantiles) are likely to give a better picture. Based on error bars or box plots,
the significance of simulation results (and differences between different simulation
settings) can be often visually assessed, but a thorough statistical analysis is clearly
favorable.
When performing statistical analyses, it must be taken into account that the
frequency distributions resulting in multi-agent simulations may not be of Gaus-
sian type. One may find multi-modal distributions or strongly skewed, “fat-tail”
distributions such as (truncated) power laws. We also point out that fitting power
laws is tricky and that straight-forward fitting approaches may easily lead to wrong
exponents and confidence intervals. Besides, one speaks of a power law only if there
is a linear relationship in a log-log plot over at least two orders of magnitude (i.e.
over a range of the horizontal axis that spans a factor 100). Deviations from power
laws may be as meaningful as the power laws themselves and should be pointed out
in the related discussion of results.
2.3.4.4 Visualization
Besides presenting tables with statistical analyses, visualization is a useful way of

presenting scientific data, which is widely applied in the natural and engineering
sciences. Some classical ways of reprensation are
• Time-dependent plots
• Two- or three-dimensional plots indicating the interdependencies (or correla-
tions) between two or three variables (which may require to project the full
parameter space into two or three dimensions)
• Pie charts or frequency distributions
• Representations of relative shares (percentages) and their changes over time
• Snapshots of spatial distributions or videos of their development in time

• Illustrations of network dependencies
In some sense, visualization is the art of transferring relevant information about
complex interdependencies to the reader quickly and in an intuitive way. Today,
large-scale data mining and massive computer simulations steadily create a need
for new visualization techniques and approaches (see [111] for a more detailed
discussion).
2.3.4.5 Scanning of Parameter Spaces and Phase Diagrams
An important way of studying the behavior of socio-economic systems with agent-

based models is the scanning of the parameter space. As was pointed out in
Sect. 2.1.5, systems composed of many non-linearly interacting agents are likely
to produce a number of self-organized or emergent phenomena. It is, therefore,
interesting to determine the conditions under which they occur. In particular, it is
relevant to identify the separating lines (or hypersurfaces) in the parameter space
that separate different system behaviors (“system states”) from each other, and to
determine the character of the related “phase transitions”: One typically checks
whether it is a continuous (“second-order”) transition, at which systemic properties
start to change smoothly, or a discontinuous (“first-order”) transition, where a
sudden regime shift occurs (see Fig. 2.2). The latter transitions occur, for example,
in cases of hysteresis (“history dependence”), and their possible types have been
studied by “catastrophe theory” [149] (see Fig. 2.2). Scientific paradigm shifts, for
example, are typically first-order transitions [150], as are revolutions in societies
[151, 152].
The parameter dependencies of the different kinds of system states (“phases”)
and their separating lines or hypersurfaces are usually represented by “phase
diagrams”. A particularly interesting case occurs, if the system displays multi-
stability, i.e. were different stable states are possible, depending on the respective
initial condition or history. For example, in traffic theory, various kinds of congested
traffic states may result, depending on the size of perturbations in the traffic flow
(see Fig. 2.3). That circumstance allows one to understand systems, which show
apparently inconsistent behaviors. In fact, it is quite common for social systems that
a certain characteristic system behavior is reported in one part of the world, while
another one is observed elsewhere [153]. There may nevertheless be a common
theory explaining both system behaviors, and history-dependence may be the reason
for the different observations.
Scanning parameter spaces typically requires large computational resources.
Even when using a computer pool, it may require weeks to determine a two-
dimensional phase diagram at reasonable accuracy. Varying more parameters will
consume even more time. However, one way of determining interesting areas of
the parameter space is to use the “overlay method”, which simulates interactions in
two-dimensional space, but additionally varies the parameter values in horizontal
Fig. 2.4 Overlay phase diagram for an evolutionary public goods game with success-driven
migration as well as prosocial and antisocial punishment, representing different behaviors by
different colors (recent work with Wenjian Yu; see Chap. 8 for an explanation of the different
behaviors such as ‘moralists’ and ‘immoralists’). The prevalent behavior depends on the size of the
model parameters (here: the punishment cost and fine), which is varied along the axes. One can see
that people showing the same behavior tend to form clusters (a phenomenon called “homophily”).
Moreover, cooperators (mainly green moralists and a few blue non-punishing cooperators) spread
above a hyperbolic kind of line. Below it, defectors (red or black) florish. The spreading of
moralists above a certain punishment level gets rid of the conventional free-rider and the second-
order free-rider problem. Mobility speeds up the convergence to the finally resulting strategy
distribution. It also increases the green area of moralists, i.e. it pushes the hyperbolic separation
line to lower punishment values. Defectors who punish non-punishers (grey) occur around the
separating line. Defectors who punish defectors (yellow immoralists) occur in separation from
each other (as loners). They require enough space, which they mainly find at low densities, or
when mobility creates areas of low density. In the mixed phase of black and red, and in the mixed
phase of blue and green, there is only a slow logarithmic coarsening, because the payoffs are the
same. (This looks like a coexistence of two strategies, if the simulations are not run long enough.)
The occurence of black defectors who punish cooperators can explain the existence of antisocial
punishers. Black antisocial punishers can exist basically at all punishment levels, if they can cluster
together and are not in direct neighborhood to moralists
and vertical direction. In this way, one may get a quick impression of the spatio-
temporal dynamics in different areas of the considered two-dimensional parameter
space (see Fig. 2.4). After identifying the approximate location of separating lines
between different phases (i.e. qualitatively different system behaviors), a fine-
grained analysis (e.g. scaling analysis) can be made to reveal the detailed behavior
next to the phase separation lines (or hysersurfaces).
2.3.4.6 Sensitivity Analysis
Preparing phase diagrams already provide good hints regarding the sensitivity of an
agent-based model to parameter changes. Generally, within a given phase, there is
not much variation of the system behavior, given the chosen parameter combinations
are not too close to the lines (or hypersurfaces) separating it from other phases.
However, when a transition line (or hypersurface) is crossed, significant changes in
the system behavior are expected, particularly if the transition is of first order (i.e.
discontinuous).
Beyond determining the phase diagram, the sensitivity can also be measured in
terms of Theil’s inequality coefficient [154]. It measures how different two time-
dependent solutions are, when the model parameters (or initial conditions) are
slightly changed. In a similar way, one may study how sensitive the model is towards
the consideration of fluctuations (“noise”). A characterization can also be made by
determining the Lyapunov exponents [155].
However, a multi-agent simulation may not only be sensitive to parameter
changes. It may also be sensitive to minor modifications of the agent-based
model itself. For example, slight changes of the interaction network (by adding or
subtracting nodes or links) may impact the system behavior. Analyses of failure and
attack tolerance demonstrate this very well [156]. To investigate so-called k-failures,
one randomly removes k agents or links from the system and studies changes in the
system performance. Similarly, one may investigate the impact of adding k links or
agents. The method is capable to reveal certain kinds of “structural instabilities”. A
further kind of structural instabilities may be caused by modifications in the rules
determining the interactions of agents. Such modifications may reflect innovations,
but also inaccuracies of the model as compared to reality. For example, “unknown
unknowns” are factors overlooked in the model, but they may be discovered to a
certain extent by varying the model assumptions. They may also be identified by
comparing models of different researchers, focusing on their incompatible features.
2.3.5 Identification of the Minimum Model Ingredients
One important part of scientific analysis is the identification of the minimum set of
rules required to explain certain empirical observations or “stylized facts” derived
from them (i.e. simplified, idealized, characteristic features). Particularly in models
with many parameters, it is sometimes difficult to understand the exact mechanism
underlying a certain phenomenon. Therefore, one should attempt to successively
reduce the model to a simpler one with less terms, parameters and/or variables, in
order to find out under what conditions the phenomenon of interest disappears. It
is clear that simplifications of the model will often reduce the level of quantitative
agreement with empirical data. However, in many cases one is mainly interested in
questions such as:
• Does the system have multiple stable states?
• Does the model behave in a history-dependent way?
• Does it produce an oscillatory behavior or a stable equilibrium?

• Is the statistical distribution Gaussian, bimodal, multi-modal or heavily skewed,
e.g. a (truncated) power law?
• What kinds of observed patterns can the model reproduce?
• Is a linear or an equilibrium model sufficient to reproduce the observations, or
does it miss out important facts?
• Are spatial (neighborhood) interactions important or not?
• Is a heterogeneity of agent properties relevant for the explanation or not?
• Does small or moderate noise have a significant influence on the system behavior
or not?
• Are correlations important, or is a mean field approximation (“representative
agent approach”), assuming well-mixed interactions good enough?
• How important are the degree distribution or other specifics of the interaction
network?
From statistical physics, it is known that all these factors may play a significant role
for the system behavior [13], but this is not always the case. Therefore, the required
ingredients of a model and appropriate level of sophistication very much depend
on the phenomena to be described, on the purpose of the model, and the desired
accuracy.
2.3.6 Gaining an Analytical Understanding
Besides providing a clearer understanding and intuition, simplifications also have

another advantage: they can make models mathematically better tractable. In
order to derive the essential phenomena in an analytical way (e.g. by means of
stability or perturbation analyses), a radical simplification may be needed, and as
a consequence, the resulting model will usually not reproduce empirical details,
but just the stylized facts (such as the occurrence of certain kinds of instabilities,
patterns, phase transitions/regime shifts, or phase diagrams with certain topological
features) (see [27] for an example).
One advantage of analytical tractability is the circumstance that one can often
derive parameter dependencies or scaling relationships. Frequently, such parameter
dependencies are not obvious, and even a numerical analysis may not give a good
picture, if several parameters are involved.
Analytical treatments often allow one to determine the location of stationary
points and their stability properties. From this information, one can derive the
fundamental features of the phase diagram, which gives a pretty good picture of
the possible system behaviors. Therefore, the fundamental properties of a system
may indeed be analytically understood. This is nicely illustrated by the example
of multi-population evolutionary games [151]. Also the properties of freeway traffic
flow and the possible congestion patterns have been analytically understood, despite
their complexity [27].
2.3.7 Some Problems and Limitations of Computational

Modeling
Despite all the virtues of mathematical modeling, one should not forget some
possible problems. So far, it is not known what phenomena can be understood by
agent-based models, and what are the fundamental limits of this approach. It is
conceivable that there exist phenomena, which are irreducibly complex [11]. For
example, the method of physics to reduce most observations to the behavior of
individual particles and pair interactions may not be fully appropriate in socio-
economic systems. Some phenomena require a more integrated treatment of the
interactions between many agents. Public good games are just one example [157].
Recent models of pedestrian interactions are also turning away from pair interaction
approaches in favor of heuristics that respond to an integrated visual pattern [158].
The corresponding behaviors can still be treated by agent-based models, but one
must be aware that they may have fundamental limitations as well.
Overestimating the power of models can be quite harmful for society, as the
financial crisis has shown. For this reason, it is important to state known limitations
of a model or, in other words, its range of validity. It should be made clear what the
purpose of a particular model is, e.g. whether it serves to understand stylized facts
or scientific puzzles better, or whether the model aims at predictions or real-world
applications. When it comes to “predictions”, it should be said whether they are
meant to be “forecasts” (in time) or “model implications” (in the sense of system
states that are expected to occur when model parameters or initial conditions etc.
are changed in a certain way).
In order to assess the reliability of a model, it is favorable to derive a number of
predictions (or implications) of a model, which may later be verified or falsified.
It is also good to point out advantages and disadvantages with regard to other
existing models. This can give a better picture of the strengths and weaknesses of
a certain approach, and guide further research aimed at developing models that can
consistently explain a large set of empirical observations. It also helps to be aware
of crucial modeling assumptions, on which the validity of the model implications
depends. This is particularly important for models, which are practically applied.4
It must be underlined that specifying a model correctly is not simple. In many
cases, different plausible mechanisms may exist that promise to explain the same
observations [159]. For example, there are many possible explanation of power laws
[160]. Moreover, if empirical data vary considerably (as it is common for socio-
economic data), it may be difficult to decide empirically, which of the proposed
models is the best [138, 139]. It may very well be that all known models are
wrong, or that their parameter specification is wrong (as it happened in financial
risk models that were calibrated with historical data). Actually, due to the implicit
4
One should be aware that this may sooner or later happen to any model, if it promises to be useful
to address real-world phenomena.
2.4 Practical Application of Agent-Based Models: Potentials and Limitations 53
simplifications and approximations of most models, this is true most of the time
[161], and this is why it is so important to state what a model is aimed to be
useful for, and what are its limits. In other words, usually, models contain a grain of
truth, but are not valid in every single respect. Consequently, a pluralistic modeling
approach makes sense [11], and overlaying the implications of several models may
give better results than the best available model itself (due to the “wisdom of
crowds” effect or the law of large numbers).
2.4 Practical Application of Agent-Based Models:

Potentials and Limitations
Having discussed potential problems and limitations of computational modeling in

several passages of Sect. 2.3.3, and in particular Sect. 2.3.7, we should not forget to
point out, where agent-based models of socio-economic systems may actually be
more powerful than one would think.
2.4.1 Stylized Facts and Prediction in Socio-Economic Systems
Some people think that the lack of collections of stylized facts in the socio-economic
sciences may be a result of the non-existence of such facts due to the great degree
of flexibility of social and economic systems. However, there are actually a number
of stylized facts with a surprisingly large range of validity:
1. The Fisher equation of financial mathematics (which determines the relationship
between nominal and real interest rates under inflation) [162]
2. The fat tail character of many financial and economic distributions [163, 164]
3. The Matthew effect (i.e. the rich-gets-richer effect) [113, 160]
4. Dunbar’s number (limiting the number of people one can have stable social
relations with) [109]
5. Pareto’s principle (according to which roughly 80% of an effect comes from
about 20% of the causes) [165]
6. Zipf’s law (determining the distribution of city rank sizes and many other things)
[166]
7. The gravity law (describing the distribution of trade flows and migration)
[167–169]
8. Goodhart’s law (according to which any observed statistical regularity, e.g. a
risk model, breaks down once pressure is placed upon it for control purposes)
[170, 171]
Complementary, a list of stylized facts regarding social norms can be found
in [43].
Even when accepting that the above “laws” tend to apply,5 many people doubt
the possibility to predict the behavior of socio-economic systems, while they believe
that this is a precondition for crisis relief. However, both is not exactly true:
1. As pointed out before, there are many other purposes of modeling and simulation
besides prediction [12]. For example, models may be used to get a picture of the
robustness of a system to perturbations (i.e. applied to perform simulated “stress
tests”, considering the effect of interactions among the entities constituting the
system).
2. “Model predictions” should be better understood as “model implications” rather
than “model forecasts”, e.g. a statements that say what systemic outcomes (e.g.
cooperation, free-riding, or conflict) are expected to occur for certain (regulatory)
boundary conditions [14, 54] (see Sect. 2.4.2.2).
3. It is important to recognize that forecasts are possible sometimes. A famous
example is Moore’s law regarding the performance of computer chips (which
is ultimately a matter of the innovation rate [172]). Moreover, while the detailed
ups and downs of stock markets are hard to predict, the manipulation of interest
rates by central banks leads, to a certain extent, to foreseeable effects. Also the
sequence in which the real-estate market in the US affected the banking system,
the US economy, and the world economy was quite logical. That is, even though
the exact timing can often not be predicted, causality networks allow one to
determine likely courses of events (see Fig. 2.5). In principle, this enables us
to take counter-actions in order to avoid or mitigate the further spreading of the
crisis [173].
4. As weather forecasts show, even unreliable short-term forecasts can be useful
and of great economic value (e.g. for agriculture). Another example illustrating
this is a novel self-control principle for urban traffic flows, which was recently
invented [40, 174]. Although its anticipation of arriving platoons is of very
short-term nature, it manages to reduce travel times, fuel consumption, and
vehicle emissions.
In conclusion, prediction is limited in socio-economic systems, but more powerful
than many people believe. Moreover, in certain contexts, it is not necessary to
forecast the course of events. For example, in order to reduce problems resulting
from bubbles in financial markets, it is not necessary to predict the time of their
bursting or even to know about their concrete existence. The reason for this will
be explained in Sect. 2.4.2.3. Furthermore, Sec.2.4.2.2 shows how the often raised
problem of self-destroying prophecies can be avoided.
5
Probably, nobody would claim that they are always true.
Fig. 2.5 Illustration of cascading effects in techno-socio-economic systems triggered by forest

fires (after [173]). Note that the largest damage of most disasters is caused by cascading effects,
i.e. the systemic impact of an over-critical local perturbation
2.4.2 Possibilities and Limitations in the Management

of Socio-Economic Systems
2.4.2.1 Paradigm Shift from Controlling Systems to Managing Complexity
When trying to improve socio-economic systems, first of all, it must be stressed

that the idea to control socio-economic systems is not only inadequate – it is
also not working well. As socio-economic systems are complex systems, cause
and effect are usually not proportional to each other. In many cases, complex
systems tend to resist manipulation attempts (cf. “Goodhart’s law”), while close to
so-called “tipping points” (or “critical points”), unexpected “regime shifts” (“phase
transitions”, “catastrophes”) may happen. Consequently, complex systems cannot
be controlled like a technical system (such as a car) [29].
The above property of systemic resistance is actually a result of the fact that
complex systems often self-organize, and that their behavior is robust to not-
too-large perturbations. While forcing complex systems tends to be expensive (if
systemic resistance is strong) or dangerous (if an unexpected systemic shift is
caused), the alternative to support the self-organization of the system appears to
be promising. Such an approach “goes with the flow” (using the natural tendencies
in the system) and is resource-efficient. Therefore, a reasonable way to manage
complexity is to guide self-organization or facilitating coordination [29, 175].
In a certain sense, this self-organization or self-control approach moves away
from classical regulation to mechanism design [176]. Regulation often corresponds
to changing the boundary conditions, while mechanism design changes the interac-
tions in the system in a way that reduces instabilities (e.g. due to delays) and avoids
that the system is trapped in local optima (and continues its evolution to a system-
optimal state). For example, slightly modifying the interaction of cars by special
driver assistant systems can stabilize traffic flows and avoid bottleneck effects to a
certain degree [177].
2.4.2.2 Self-destroying Prophecies and Proper Design of Information

Systems
It is often pointed out that socio-economic systems would not be predictable,

because the reaction of people to information about the system would destroy
the validity of the forecast. If done in the wrong way, this is actually true. Let
us illustrate this by an example: Assume that all car drivers are given the same
information about existing traffic jams. Then, drivers may most likely over-react,
i.e. more drivers may use an alternative road than is required to reach a system-
optimal distribution of traffic flows [178].
However, as has been shown by laboratory route choice experiments, an almost
optimal route choice behavior may be reached by an information system that gives
user-specific recommendations. In other words, some drivers would be asked to stay
on the congested road, and others to leave it. When the recommendation system
compensates for the fact that not everyone follows the recommendations, one can
avoid over- or under-reactions of drivers in congestion scenarios [178].
One crucial issue of such individualized recommender systems, however, is
their reliability. An unreliable system or one that is systematically biased, will
be only poorly followed, and people will eventually compensate for biases [178].
Therefore, it is also essential to design the information system in a way that is fair
to everyone. That is, nobody should have a systematic advantage. Nevertheless, the
system should be flexible enough to allow a trading of temporary (dis)advantages.
For example, somebody who was asked to take the slower road on a given day (and
who would have a right to use the faster road on another day), may still use the faster
road. However, he or she would have to pay a fee for this, which would be earned
by somebody else, who would exchange his or her “ticket” for the faster road for a
“ticket” for the slower road. In such a way, the system optimum state could still be
maintained [178].
In summary, the above described information system would cheat nobody, and
it would be flexible and fair. Only people who use the faster road more often than
average would have to pay a road usage fee. A normal driver would either pay
nothing on average: While he or she will pay on some days (when being under a
pressure of time, while the recommendation asks to take the slower road), the same
amount of money can be earned on other days by taking the slower road. In other
words, fair usage would be free of charge on average, and drivers would still have a
freedom of route choice. The primary goal of the system would not be to suppress
traffic flows through road pricing, but the pricing scheme would serve to reach a
system optimal traffic state.
2.4.2.3 New Approaches and Designs to Manage Complexity
It is quite obvious that there is no single scheme, which allows one to manage all
kinds of complex systems optimally, independently of their nature. The success of
a management concept very much depends on the characteristics of the system, e.g.
its degree of predictability. The systems design must account for this.
If long-term forecasts are possible, there must obviously be an almost determinis-
tic relationship between input- and output-variables, which allows one to change the
temporal development and final outcome of the system. If long-term predictability
is not given, management attempts must be oriented at a sufficiently frequent re-
adjustment, which requires a suitable monitoring of the system.
As the example of weather forecasts shows, even unreliable short-term forecasts
can be very useful and economically relevant (e.g. for agriculture).6 The success
principle in case of short-term forecasts is the flexible adjustment to the local
conditions, while well predictable systems often perform well with a fixed (rather
than variable) organization principle.
In systems where no forecasts over time are possible at all, it may still be feasible
to improve the system behavior by modifying the statistical distribution. Taking the
time-dependencies at stock markets for illustration, introducing a Tobin tax would
reduce excessive levels of speculation (e.g. high-frequency trading). Moreover,
introducing “noise” (further sources of unpredictability), could destroy undesirable
correlations and impede insider trading [179].
6
Another example is the “self-control” of urban traffic flows, which is based on a special, traffic-
reponsive kind of decentralized traffic light control [40], see Sect. 2.4.2.1.
These are just a few examples illustrating that there actually are possibilities to
influence systems involving human behavior in a favorable way. A more detailed
discussion of the issue of managing complexity is given in [29, 180–183].
2.4.2.4 Changing the Rules of the Game and Integrative Systems Design
Generally, there are two ways of influencing the dynamics and outcome of a system
by changing the “rules of the game”. If the interactions in the system are weak, the
system dynamics can be well influenced by modifying the boundary conditions of
the system (i.e. by regulatory measures). However, if the interactions are strong,
as in many social and economic processes, the self-organization of the system
dominates the external influence. In this case, a modification of interactions in the
system (“mechanism design”) seems to be more promising. (A good example for
this is a traffic assistance system that reduces the likelihood of congestion by special
driver assistance systems [177].) Of course, regulatory measures and mechanism
design may also be combined with each other.
While mechanism design is relatively common in computer science and some
other areas (e.g. in evolutionary game theory, mathematics, and partly in physics), it
seems that these methods have not been extensively applied in the social sciences so
far. For example, there are many different mechanisms to match supply and demand,
and it would be interesting to know what systemic properties they imply (such as
the level of stability, the efficiency of markets, the resulting wealth distribution, the
creation of investment opportunities, etc.). It is also not clear how to reach the best
combination of top-down and bottom-up elements in decision-making processes,
and how to find the best balance between centralized and decentralized coordination
approaches. All this poses interesting and practically relevant challenges that
determine the prosperity and well-being of societies (see also [111]).
Moreover, in the past, mechanism design has been applied basically to subsys-
tems, i.e. parts of the complex overall system we are living in. However, due to the
interconnection of all sorts of (sub-)systems (e.g. of the traffic, supply, industrial,
environmental, health and social systems), measures in one (sub)system may have
undesired side effects on other (sub)systems. In fact, for fundamental reasons, it
is quite frequent that taking the best action in one (sub)system affects another
(sub)system in a negative way. Such partial improvements will usually not promote
an optimal state of society (whatever the optimal state of society may be). A good
example for this is the poor performance of traffic light controls that optimize locally
(without a coordination with neighboring intersections) [177]. Unfortunately, such
kinds of problems are not at all restricted to traffic systems. Undesirable feedback
effects (like spill-over effects and mutual obstructions) are quite common for many
networked systems, such as logistic or production systems, or even administrative
processes [174].
2.4.2.5 Choice of the Goal Function
Improving a socio-economic system is far from trivial. Even if one would have a
perfectly realistic and predictive model, the result may largely depend on the chosen
goal function. An improper choice of the goal function can cause more harm than
benefit. For example, maximizing the efficiency of a system may make it vulnerable
to breakdowns [30]. Besides, there is a danger of misusing models to promote
individual interests that are not compatible with human well-being. However, the
following goals appear to be widely acceptable:
• Increase the self-awareness of society
• Reduce vulnerability and risk
• Increase resilience (the ability to absorb societal, economic, or environmental
shocks)
• Avoid loss of control (sudden, large and unexpected systemic shifts)
• Develop contingency plans
• Explore options for future challenges and opportunities
• Increase sustainability
• Facilitate flexible adaptation
• Promote fairness
• Increase social capital and the happiness of people
• Support social, economic and political inclusion and participation
• Balance between central and decentral (global and local) control
• Protect privacy and other human rights, pluralism and socio-bio-diversity
• Support collaborative forms of competition and vice versa (“coopetition”)
• Promote human well-being
If several goals shall be promoted at the same time, the question arises how
to perform such multi-goal optimization. Most optimization methods used today
eliminate heterogeneity in the system, i.e. there is one optimal solution which is
applied to everyone. For socio-economic systems, this appears to be particularly
problematic, as it tends to reduce socio-diversity and innovation. Besides, it is
promoting average performance rather than individual strengths. A way to overcome
this problem is suggested in the following.
The crucial question in this connection is, how to translate the performance
values Xij of the alternative systems i (measured on multiple scales j by goal
functions gj ) into one scale. Traditionally, this is done by weighting each criterion
j or goal function with a certain factor wj . This results in the overall individual
performance X
xi D wj xij ; (2.1)
j
where xij D Xij =hXij ii is the value Xij , scaled by the average performance hXij ii
of all alternatives i . The overall individual performance values xi can ordered on a
one-dimensional scale, i.e. ranked. Such an approach, however, promotes average
performance rather than excellence, since excellence is typically characterized by
extreme values on one or a few rating scales, but not on all of them.
In order to reward individual strengths of alternative system designs, one may

proceed as follows: Political decision-makers could choose the weight they would
like to attribute to each criterion or goal function, say w1 D 0:35, w2 D 0:25,
w3 D 0:25, and w4 D 0:15 (assuming only four relevant goals in this example). An
index that is favorable with respect to individual strengths, would for example be
X
yi D wj xij C 0:1.yi1 C yi 2 yi 3 yi 4 /; (2.2)
j
where the values yij correspond to the values xij , sorted according to their size
in descending order. This formula overweights the particular strengths of each
individual system i , and it is possible that different alternative systems perform
equally well. Putting this into the context of the European Union for illustration,
each country could choose a systemic design which fits the respective national
strengths best. Hence, the “pluralistic” goal function (2.2) overcomes a number of
problems of the optimization methods that are predominantly used today (namely,
one-dimensional ranking scales, which measure the performance in an individually
non-differentiated way, which typically creates one winner and many losers).
2.5 Summary, Discussion, and Outlook
In this contribution, we have presented an overview, how to do agent-based model-

ing (ABM) and multi-agent simulations (MAS) properly and how to avoid a number
of traps associated with this research approach. In particular, we have discussed
the potentials, limitations, and possible problems. Multi-agent simulations can be
used for hypothesis testing and to get a better understanding of complex systems.
They are flexible and allow one to reflect many characteristic features of techno-
socio-economic-environmental systems in a natural way (including heterogeneity
and network interactions). As a result, one can expect insights into the different
possible “states” or behaviors of a system and the preconditions for their occurrence.
In particular, phase diagrams facilitate the representation of different characteristic
phases and of transitions between them. Considering the possibility of multi-
stability and history-dependence, phase diagram are also a promising approach to
make sense of seemingly inconsistent empirical evidence.
We have underlined how essential it is to proceed carefully when modeling
and simulating socio-economic systems. In particular, a publication should clearly
describe:
• The research question (challenge, “puzzle”, “mystery”) addressed, including the
purpose of the model
• The research methodology/approach used
• The assumptions underlying the agent-based model
• The current empirical or experimental evidence
2.5 Summary, Discussion, and Outlook 61
• Implications or predictions that allow others to assess the explanatory power (e.g.
through lab or Web experiments)
• The expected range of validity and limitations of the approach
Agent-based models can be published in a number of journals (searching for “agent-
based model” or “multi-agent simulation” at http://scholar.google.com will give a
good overview). However, studies presenting multi-agent simulations are currently
still hard to publish in mainstream economic or social science journals, possibly
because most authors do not back up their computational results with analytical
ones. Proceeding more carefully when performing multi-agent simulations, as
suggested in this contribution, will most likely increase the interest in this approach
over time, particularly as it can produce significant results beyond the range of
phenomena that are understandable through analytical methods. Finally, multi-agent
simulations can also be useful to identify interesting experimental setups [142],
considering in particular that experiments are costly and restricted to a small number
of conditions and repetitions.
2.5.1 Future Prospects and Paradigm Shifts
Great prospects for agent-based modeling do not only result from the experience
gained with multi-agent simulations, the availability of user-friendly simulation
platforms, greater computer power, and improved visualization techniques. We also
expect a number of paradigm shifts:
• The social sciences are currently experiencing a transition from a data-poor
to a data-rich situation. This allows one to verify or falsify models, calibrate
their parameters, and to move to data-driven modeling approaches [1, 158, 184].
Moreover, it will be possible to improve the level of detail, accuracy, and scale
of agent-based models by orders of magnitude. At the same time, thanks to the
availability of user-friendly simulation tools, the development times for multi-
agent simulations will shrink dramatically.
• The application of methods from statistical physics and the theory of complex
systems to socio-economic data generates a chance of moving beyond descriptive
(fit) models towards explanatory models. The improved data situation supports
the comprehension of the inner relationships and significant patterns of complex
systems.
• New possibilities to mine real-time data (e.g. text mining of news, blogs, twitter
feeds, etc.) create the opportunity to move from measurements with a delay (such
as the classical ways of determining the gross national product or the number of
people, who have got the flu) towards reliable real-time estimates (“nowcasting”)
[185, 186]. Furthermore, using particular properties of spreading processes in
networks, it seems even possible to achieve two-week forecasts (based on
the method of “health sensors”) [187]. More generally, “reality mining” will
facilitate multi-agent simulations of realistic scenarios, the determination of
model parameters (and other relevant model inputs) on the fly, and the timely
determination of advance warning signs. It will also help to avoid destabilizing
delays and to increase the efficiency of crisis response measures. (Delays are a
major problem in the efficiency of disaster response management and mitigation
[188].)
• Multi-agent simulations will integrate measurement-based data-mining and
model-based simulation approaches. This approach goes beyond feeding in
real-time inputs (such as initial and boundary conditions, parameters, and
network characteristics) into multi-agent simulations: it performs a data-driven
pattern recognition and modeling in parallel to related computer simulations
and, thereby, combines the strengths of both methods to reach the optimum
accuracy and predictability. For example, integrating two incomplete sets of
traffic information (cross-sectional measurements and floating car data) and a
fluid-dynamic real-time traffic simulation facilitates to reduce the delay times
between the formation and detection of traffic jams by 50%, and to double the
reliability of such traffic information.
• It will be possible to move from a batch processing of alternative simulation
scenarios to interactive real-time specifications and scenario analyses. This will
facilitate to explore policy options and “parallel worlds” (i.e. possible futures),
as the situation evolves and pressing decisions must be taken. For example,
evacuation scenarios in response to certain disasters have to be developed and
evaluated quickly. More generally, interactive supercomputing would facilitate
more flexible contingency plans that are tailored to the actual situation of crisis.
• Multi-agent simulations could be directly coupled with lab and web experiments.
In fact, the decisions of agents in computer simulations could be taken by real
people. Serious multi-player online-games provide the opportunity of involving
a large number of people into the analysis of complex data and the exploration
of realistic decision-making scenarios in virtual worlds, which realistically map
possible future worlds. In this way, agent-based simulation approaches may be
applied for crowd sourcing and eGovernance applications, to make use of the
“wisdom of crowds”. For example, one could populate the three-dimensional
virtual model of a new shopping center, railway station, or airport in order to find
out, how well the architectures would fulfil its function, and to determine which
design is favored by the future users.
• In the medium-term future, one can expect a confluence of real and virtual
worlds. For example, Google Earth and similar virtual representations of the real
world could be populated with simulated people or real ones. In fact, people
agreeing to share their GPS coordinates could be represented in these worlds
directly, to the level of detail they like. An augmented reality approach would
allow people to share information about their interests, backgrounds, values, etc.
The amount of information shared may be decided interactively or by the kinds
of interaction partners (e.g. people are expected to share private information
more openly with people they consider to be their friends). Such augmented
reality tools will be able to serve as “translator” or “adaptor” for people with
different languages or cultural backgrounds, helping them to make themselves
References 63
understandable to each other. The resulting techno-social systems would also

offer many new opportunities for social and economic participation, both in the
virtual and in the real world.
Given this development, we envision a new way of performing socio-economic
research, which may be called “Social Supercomputing”. This approach would
facilitate the integration of different kinds of data (e.g. demographic, socio-
economic, and geographic data) and different kinds of simulation approaches (e.g.
agent-based and equation-based ones), at an unprecedented scale and level of detail.
It would enable simulations and interactivity on all scales. In fact, the computer
simulation of techno-socio-economic systems on a global scale is a vision that
appears to become feasible within the next 10 to 15 years, with unprecedented
opportunities for societies and economies, if done in the right way. Some of the
challenges for computer scientists and other researchers on the way are described in
the Visioneer White papers (particularly Sects. 3.3 and 3.4.2 of [189]).
References
1. D. Helbing, S. Balietti, From social data mining to forecasting socio-economic crisis.

visioneer white paper (2010) http://www.visioneer.ethz.ch.
2. N. Gilbert, K.G. Troitzsch, Simulation for the Social Scientist. (Open University Press,
England, 2005)
3. J.M. Epstein, R. Axtell, Growing artificial societies. (Cambridge, MA, 1996)
4. J.M. Epstein, Generative Social Science: Studies in Agent-Based Computational Modeling.
Princeton Studies in Complexity. (Princeton University Press, Princeton, 2007)
5. N.R. Jennings, On agent-based software engineering. Artif. Intell. 117(2), 277–296 (2000)
6. A.M. Uhrmacher, D. Weyns, Multi-Agent Systems: Simulation and Applications. (CRC Press,
Inc., Boca Raton, FL, USA, 2009)
7. N. Gilbert, S. Bankes, Platforms and methods for agent-based modeling. Proc. Natl. Acad.
Sci. USA 99(Suppl 3), 7197 (2002)
8. M.W. Macy, R. Willer, From factors to actors: computational sociology and agent-based
modeling. Annu. Rev. Sociol. 28(1), 143–166 (2002)
9. J.P. Davis, K.M. Eisenhardt, C.B. Bingham, Developing theory through simulation methods.
Acad. Manag. Rev. 32(2), 480 (2007)
10. R.A. Bentley, P. Ormerod, Agents, intelligence, and social atoms, in Creating Consilience:
Integrating the Sciences and the Humanities, ed. by M. Collard, E. Slingerland (Oxford
University Press, 2011)
11. D. Helbing, Pluralistic modeling of complex systems (2010) CCSS-10-009
12. J.M. Epstein, Why model? J. Artif. Soc. Soc. Simulat. 11(4), 12 (2008). http://jasss.soc.surrey.
ac.uk/11/4/12.html
13. D. Helbing, S. Balietti, Fundamental and real-world challenges in economics. Sci. Cult. 76,
399–417 (2010). Special Issue: 15 year of econophysics research
14. D. Helbing, A. Szolnoki, M. Perc, G. Szabó, Evolutionary establishment of moral and double
moral standards through spatial interactions. PLoS Comput. Biol. 6(4), e1000758, 04 (2010)
15. R. May, A. McLean, Theoretical Ecology: Principles and Applications, 3rd edn. (Oxford
University Press, USA, 2007)
16. I. Foster, A two-way street to science’s future. Nature 440, 419 (2006)
17. C.S. Taber, R.J. Timpone, Computational Modelign. (Sage, London, 1996)
18. F. Schweitzer, Brownian Agents and Active Particles. On the Emergence of Complex Behavior
in the Natural and Social Sciences. (Springer, Berlin, 2003)
19. D. Helbing, J. Keltsch, P. Molnar, Modelling the evolution of human trail systems. Nature
388(6637), 47–50 (1997)
20. J.M. Epstein, Modelling to contain pandemics. Nature 460(7256), 687 (2009)
21. J. Parker, J. Epstein, A global-scale distributed agent-based model of disease transmission,
ACM Transactions on Modelling and Computer Simulation (TOMACS) 22(1), (2011)
22. M. Treiber, A. Hennecke, D. Helbing, Congested traffic states in empirical observations and
microscopic simulations. Phys. Rev. E 62, 1805 (2000)
23. D. Helbing, Derivation of non-local macroscopic traffic equations and consistent traffic
pressures from microscopic car-following models. Eur. Phys. J. B 69(4), 539–548 (2009)
24. R.K. Sawyer, Artificial Societies. Socio. Meth. Res. 31(3), 325 (2003)
25. H.J. Aaron, Distinguished lecture on economics in government: Public policy, values,
and consciousness. J. Econ. Perspect. 8(2), 3–21 (1994). http://ideas.repec.org/a/aea/jecper/
v8y1994i2p3-21.html
26. A.P. Kirman, Whom or what does the representative individual represent? J. Econ. Perspect.
6(2), 117–36 (1992). http://ideas.repec.org/a/aea/jecper/v6y1992i2p117-36.html
27. D. Helbing, M. Treiber, A. Kesting, M. Schönhof, Theoretical vs. empirical classification and
prediction of congested traffic states. Eur. Phys. J. B 69(4), 583–598 (2009)
28. D. Helbing, Verkehrsdynamik. (Springer, Berlin, 1997)
29. D. Helbing, Managing Complexity: Insights, Concepts, Applications, 1st edn. (Springer,
Berlin, 2007)
30. D. Helbing, Systemic risks in society and economics (2009). Sante Fe Institute, working paper
09-12-044.
31. R. Kuehne, Freeway control using a dynamic traffic flow model and vehicle reidentification
techniques. Transport. Res. Record 1320, 251–259 (1991)
32. B.S. Kerner, P. Konhäuser, Cluster effect in initially homogeneous traffic flow. Phys. Rev. E
48(4), 2335+ (1993)
33. R.E. Chandler, R. Herman, E.W. Montroll, Traffic dynamics: studies in car following. Oper.
Res., 165–184 (1958)
34. R. Herman, E.W. Montroll, R.B. Potts, R.W. Rothery, Traffic dynamics: analysis of stability
in car following. Oper. Res. 7(1), 86–106 (1959)
35. D. Helbing, M. Treiber, Critical discussion of “synchronized flow”. Cooper. Transport. Dyn.
1, 2.1–2.24 (2002)
36. D. Helbing, T. Seidel, S. Laemmer, K. Peters, Econophysics and Sociophysics - Trends and
Perspectives, chapter Self-organization principles in supply networks and production systems
(Wiley, Weinheim, 2006), pp. 535–558
37. D. Helbing, A. Mazloumian, Operation regimes and slower-is-faster effect in the control of
traffic intersections. Eur. Phys. J. B 70(2), 257–274 (2009)
38. Dirk Helbing, Illés J. Farkas, and Tamás Vicsek, Freezing by Heating in a Driven Mesoscopic
System. Phys. Rev. Lett. 84(6), 1240–1243 (2000)
39. D. Helbing, W. Yu, The outbreak of cooperation among success-driven individuals under
noisy conditions. Proc. Natl. Acad. Sci. 106(10), 3680–3685 (2009)
40. S. Laemmer, D. Helbing, Self-control of traffic lights and vehicle flows in urban road
networks. JSTAT (2008). P04019
41. Michael Mäs, Andreas Flache, and Dirk Helbing, Individualization as Driving Force of
Clustering Phenomena in Humans. PLoS Comput. Biol. 6(10), e1000959+ (2010)
42. J. Lorenz, H. Rauhut, F. Schweitzer, D. Helbing, How social influence undermines the wisdom
of crowds. Proc. Natl. Acad. Sci. USA (PNAS) 108(22), 9020–9025 (2011)
43. D. Helbing, W. Yu, K.-D. Opp, H. Rauhut, The emergence of homogeneous norms in
heterogeneous populations. Santa Fe Working Paper 11-01-001 (2011), see http://www.
santafe.edu/media/workingpapers/11-01-001.pdf, last accessed on March 6, 2012
44. C. Nardini, B. Kozma, A. Barrat, Who’s talking first? consensus or lack thereof in coevolving
opinion formation models. Phys. Rev. Lett. 100(15), 158701 (2008)
References 65
45. J.S. Sichman, Depint: Dependence-based coalition formation in an open multi-agent scenario.
JASSS J. Artif. Soc. Soc. Simulat. 1(2) (1998)
46. M.E. Gaston, M. des Jardins, Agent-organized networks for dynamic team formation.
In Proceedings of the fourth international joint conference on Autonomous agents and
multiagent systems, AAMAS ’05, pages 230–237, New York, NY, USA, 2005. ACM
47. E. Bonabeau, M. Dorigo, G. Theraulaz, Inspiration for optimization from social insect
behaviour. Nature 406(6791), 39–42 (2000)
48. G. Szabo, G. Fath, Evolutionary games on graphs. Phys. Rep. 446(4-6), 97–216 (2007)
49. P.J. Carrington, J. Scott, S. Wasserman, Models and Methods in Social Network Analysis.
(Cambridge University Press, Cambridge, 2005)
50. P. Holme, G. Ghoshal, Dynamics of networking agents competing for high centrality and low
degree. Phys. Rev. Lett. 96(9), 098701 (2006)
51. M. Sierhuis, J.M. Bradshaw, A. Acquisti, R. van Hoof, R. Jeffers, A. Uszok, Human-Agent
Teamwork and Adjustable Autonomy in Practice. In Proceedings of the seventh international
symposium on artificial intelligence, robotics and automation in space (I-SAIRAS), 2003
52. Robert Axelrod, The Complexity of Cooperation: Agent-Based Models of Competition and
Collaboration, 1st printing edn. (Princeton University Press, Princeton, 1997)
53. S. Bowles, H. Gintis, The evolution of strong reciprocity: cooperation in heterogeneous
populations. Theor. Popul. Biol. 65(1), 17–28 (2004)
54. D. Helbing, A. Johansson, Cooperation, norms, and conflict: A unified approach (2009) SFI
Working Paper
55. Heiko Rauhut and Marcel Junker, Punishment deters crime because humans are bounded in
their strategic decision-making. J. Artif. Soc. Soc. Simulat. 12(3), 1 (2009)
56. L-E. Cederman, Modeling the size of wars: From billiard balls to sandpiles. Am. Polit. Sci.
Rev. 97, 135–150 (2003)
57. L-E. Cederman, Emergent Actors in World Politics: How States and Nations Develop and
Dissolve. (Princeton University Press, Princeton, NJ, 1997)
58. C. Hommes, Chapter 23 Heterogeneous Agent Models in Economics and Finance, in
Handbook of Computational Economics, vol. 2 (Elsevier, 2006), pp. 1109–1186
59. B. LeBaron, Building the santa fe artificial stock market. Physica A, 2002. Working Paper,
Brandeis University.
60. M. Raberto, S. Cincotti, S.M. Focardi, M. Marchesi, Agent-based simulation of a financial
market. Phys. Stat. Mech. Appl. 299(1-2), 319–327 (2001)
61. J. Zhang, Growing Silicon Valley on a landscape: an agent-based approach to high-tech
industrial clusters. Entrepreneurships, the New Economy and Public Policy, pp. 71–90 (2005)
62. Robert Axtell, The emergence of firms in a population of agents. Working Papers 99-03-019,
Santa Fe Institute, March 1999
63. T. Kaihara, Multi-agent based supply chain modelling with dynamic environment. Int. J. Prod.
Econ. 85(2), 263–269 (2003)
64. C. Preist, Commodity trading using an agent-based iterated double auction, in Proceedings of
the third annual conference on Autonomous Agents (ACM 1999), pp. 131–138.
65. Lebaron, Blake, Tesfatsion, Leigh, Modeling macroeconomies as open-ended dynamic
systems of interacting agents. Am. Econ. Rev. 98(2), 246–250 (2008)
66. Leigh Tesfatsion, Agent-based computational economics: growing economies from the
bottom up. Artif. Life 8(1), 55–82 (2002)
67. J.R. Harrison, Z. Lin, G.R. Carroll, K.M. Carley, Simulation modeling in organizational and
management research. Acad. Manag. Rev. 32(4), 1229 (2007)
68. B.S. Onggo, Parallel discrete-event simulation of population dynamics. In Proceedings of
Winter Simulation Conference 2008 (Miami, FL, USA), pp. 1047–1054
69. Y. Mansury, M. Kimura, J. Lobo, T.S. Deisboeck, Emerging patterns in tumor systems:
simulating the dynamics of multicellular clusters with an agent-based spatial agglomeration
model. J. Theor. Biol. 219(3), 343–370 (2002)
70. D. Osullivan, J.M Macgill, C. Yu, Agent-based residential segregation: a hierarchically
structured spatial model. Agent 2003 Challenges in Social Simulation (2003)
71. M. Batty, Cities and Complexity: Understanding Cities with Cellular Automata, Agent-Based
Models, and Fractals. (The MIT Press, Cambridge, 2007)
72. D. Helbing, K. Nagel, The physics of traffic and regional development. Contemp. Phys. 45(5),
405–426 (2004)
73. V. Killer, K.W. Axhausen, D. Guth, C. Holz-Rau, Understanding regional effects of travel
times in switzerland and germany 1970-2005. Jonkoping, Sweden, August 2010. 50th
European Regional Science Association (ERSA)
74. C.R. Binder, C. Hofer, A. Wiek, R.W. Scholz, Transition towards improved regional
wood flows by integrating material flux analysis and agent analysis: the case of appenzell
ausserrhoden, switzerland. Ecol. Econ. 49(1), 1–17 (2004)
75. H. Dia, An agent-based approach to modelling driver route choice behaviour under the
influence of real-time information. Transport. Res. C Emerg. Tech. 10(5-6), 331–349 (2002)
76. C.M. Henein, T. White, Agent-based modelling of forces in crowds, in Multi-agent and Multi-
agent-Based Simulation (Springer, Berlin, 2005), pp. 173–184
77. M. Batty, Agent-based pedestrian modelling, in Advanced spatial analysis: the CASA book of
GIS, page 81 (2003)
78. D. Delli Gatti, Emergent macroeconomics : an agent-based approach to business fluctuations.
79. M. Aoki, H. Yoshikawa, Reconstructing Macroeconomics: A Perspective from Statistical
Physics and Combinatorial Stochastic Processes (Japan-US Center UFJ Bank Monographs
on International Financial Markets), 1st edn. (Cambridge University Press, Cambridge, 2006)
80. R. Conte, R. Hegselmann, P. Terna, Simulating Social Phenomena. (Springer, Berlin, 1997)
81. R. Sun (ed.), Cognition and Multi-Agent Interaction: From Cognitive Modeling to Social
Simulation, 1st edn. (Cambridge University Press, Cambridge, 2008)
82. N. Gilbert, Agent-Based Models. (Sage Publications, Inc, 2007)
83. V. Grimm, E. Revilla, U. Berger, F. Jeltsch, W.M. Mooij, S.F. Railsback, H.H. Thulke,
J. Weiner, T. Wiegand, D.L. DeAngelis, Pattern-oriented modeling of agent-based complex
systems: lessons from ecology. Science 310(5750), 987 (2005)
84. R. Axelrod, Advancing the art of simulation in the social sciences, in Handbook of Research
on Nature Inspired Computing for Economy and Management, ed. by Jean-Philippe Rennard
(Hersey, PA, 2005)
85. K. Nagel, Parallel implementation of the transims micro-simulation. Parallel Comput. 27(12),
1611–1639 (2001)
86. M. Rickert, K. Nagel, Dynamic traffic assignment on parallel computers in transims. Future
Generat. Comput. Syst. 17(5), 637–648 (2001)
87. K. Nagel, R.J. Beckman, C.L. Barrett, Transims for urban planning. Los Alamos Unclassified
Report (LA-UR) 98-4389, Los Alamos National Laboratory (1999)
88. K. Nagel, C.L. Barrett, Using microsimulation feedback for trip adaptation
for realistic traffic in Dallas. Santa Fe Working Paper 97-03-028 (1997), see
http://www.santafe.edu/media/workingpapers/97-03-028.pdf, last accessed on March 6,
2012
89. P.M. Simon, J. Esser, and Nagel K. Simple queueing model applied to the city of portland.
Int. J. Mod. Phys. C 10(5), 941–960 (1999)
90. B. Raney, N. Cetin, A. Vollmy, M. Vrtic, Axhausen. K., K. Nagel, An agent-based microsim-
ulation model of swiss travel: First results. Network Spatial. Econ. 3, 23–41 (2003)
91. M. Balmer, K.W. Axhausen, K. Nagel, An agent based demand modeling framework for
large scale micro-simulations (2005) Working paper, 329, Institute for Transport Planning
and Systems (IVT), ETH Zurich, Switzerland.
92. D. Charypar, K. Nagel, Generating complete all-day activity plans with genetic algorithms.
Transportation 32(4), 369–397 (2005)
93. K.W. Axhausen, T. Garling, Activity-based approaches to travel analysis: conceptual frame-
works, models, and research problems. Transport Rev. 12(4), 323–341 (1992)
94. G. Laemmel, D. Grether, K. Nagel, The representation and implementation of time-dependent
inundation in large-scale microscopic evacuation simulations. Transport. Res. C Emerg. Tech.
18(1), 84–98 (2010)
References 67
95. H. Taubenboeck, N. Goseberg, N. Setiadi, G. Laemmel, F. Moder, M. Oczipka, H. Klupfel,

Roland Wahl, T. Schlurmann, G. Strunz, J. Birkmann, K. Nagel, F. Siegert, F. Lehmann,
S. Dech, A. Gress, R. Klein, “last-mile” preparation for a potential disaster - interdisciplinary
approach towards tsunami early warning and an evacuation information system for the coastal
city of padang, indonesia. Nat. Hazards Earth Syst. Sci. 9(4), 1509–1528 (2009). http://www.
nat-hazards-earth-syst-sci.net/9/1509/2009/
96. J. Min, J.E. Beyler, Brown T.H., Y. Jun Son, A.T. Jones, Toward modeling and simulation of
critical national infrastructure interdependencies. IIE Trans., 57–71 (2007) special issue on
Industrial Engineering and Operations Research in Homeland Security.
97. C. Barrett, S. Eubank, M. Marathe, Modeling and simulation of large biological, information
and socio-technical systems: An interaction based approach, in Interactive Computation, ed.
by D. Goldin, S.A. Smolka, P. Wegner (Springer, Berlin, 2006), pp. 353–392. 10.1007/3-540-
34874-3 14
98. National Infrastructure Simulation and Analysis Center (NISAC), see http://www.lanl.gov/
programs/nisac/, last accessed on March 6, 2012
99. W. Kermack, A. Mckendrick, Contributions to the mathematical theory of epidemics–i. Bull.
Math. Biol. 53(1), 33–55 (1991)
100. W. Kermack, A. Mckendrick, Contributions to the mathematical theory of epidemics–ii. the
problem of endemicity. Bull. Math. Biol. 53(1), 57–87 (1991)
101. W. Kermack, A. McKendrick, Contributions to the mathematical theory of epidemics–iii.
further studies of the problem of endemicity. Bull. Math. Biol. 53(1), 89–118 (1991)
102. V. Colizza, A. Barrat, M. Bartholemy, A. Vespignani, The role of the airline transportation
network in the prediction and predictability of global epidemics. Proc. Natl. Acad. Sci. USA
103(7), 2015–2020 (2006)
103. L. Hufnagel, D. Brockmann, T. Geisel, Forecast and control of epidemics in a globalized
world. Proc. Natl. Acad. Sci. USA 101(42), 15124–15129 (2004)
104. S. Eubank, H. Guclu, Anil K.V.S., M.V. Marathe, A. Srinivasan, Z. Toroczkai, N. Wang,
Modelling disease outbreaks in realistic urban social networks. Nature 429(6988), 180–184
(2004)
105. Artificial stock market, see http://sourceforge.net/projects/artstkmkt/, last accessed on March
6, 2012
106. U-mart. http://www.u-mart.org/, last accessed on March 6, 2012
107. Eurace. http://www.eurace.org/, last accessed on March 6, 2012
108. A. Cangelosi, D. Parisi, Simulating the evolution of language. (Springer, Berline, 2002)
109. R. Dunbar, Grooming, gossip, and the evolution of language. (Harvard University Press,
Cambridge, 1998)
110. M.A. Nowak, D.C. Krakauer, The evolution of language. Proc. Natl. Acad. Sci. USA 96(14),
8028 (1999)
111. D. Helbing, S. Balietti, How to create an innovation accelerator. Eur. Phys. J. Special Topics
195, 101–136 (2011)
112. M. Eigen, P. Schuster, The Hypercycle. Naturwissenschaften 65(1), 7–41 (1978)
113. R.K. Merton, The Matthew effect in science. The reward and communication systems of
science are considered. Science (New York, N.Y.) 159(810), 56–63 (1968)
114. M. Mcpherson, L. Smith-Lovin, J.M. Cook, Birds of a Feather: Homophily in Social
Networks. Annu. Rev. Sociol. 27, 415–444 (2001)
115. N.A. Christakis, J.H. Fowler, The spread of obesity in a large social network over 32 years.
New Engl. J. Med. 357(4), 370–379 (2007)
116. N.A. Christakis, J.H. Fowler, The Collective Dynamics of Smoking in a Large Social
Network. New Engl. J. Med. 358(21), 2249–2258 (2008)
117. K.P. Smith, N.A. Christakis, Social Networks and Health. Annu. Rev. Sociol. 34(1), 405–429
(2008)
118. T. Chadefaux, D. Helbing, How wealth accumulation can promote cooperation. PLoS One
5(10), e13471, 10 (2010)
119. D. Helbing, W. Yu, H. Rauhut, Self-organization and emergence in social systems: Modeling
the coevolution of social environments and cooperative behavior. J. Math. Sociol. 35, 177–208
(2011)
120. A.P. Fiske, Structures of Social Life: The Four Elementary Forms of Human Relations:
Communal Sharing, Authority Ranking, Equality Matching, Market Pricing. (Free Press,
1991)
121. Mark Granovetter, The strength of weak ties: a network theory revisited. Sociol. Theor. 1,
201–233 (1982)
122. G. Gigerenzer, P.M. Todd, ABC Research Group, Simple Heuristics That Make Us Smart
(Oxford University Press, New York, 2000)
123. W.H. Greene, Econometric Analysis. (Prentice Hall, Upper Saddle River, NJ., 2008). Particu-
larly Chap. 7.4: Model selection criteria
124. F. Diebold, Elements of forecasting, 3rd edn. (Thomson/South-Western, Mason, Ohio, 2003)
125. Swarm, http://www.swarm.org/, last accessed on March 6, 2012
126. Repast, recursive porous agent simulation toolkit. http://repast.sourceforge.net/, last accessed
on March 6, 2012
127. Netlogo, http://ccl.northwestern.edu/netlogo/, last accessed on March 6, 2012
128. Sesam, shell for simulated agent systems. http://www.simsesam.de/, last accessed on March
6, 2012
129. Massive software. http://www.massivesoftware.com/, last accessed on March 6, 2012
130. Wikipedia: Comparison of agent-based modeling software. http://en.wikipedia.org/wiki/
Comparison of agent-based modeling software, last accessed on March 6, 2012
131. S.F. Railsback, S.L. Lytinen, S.K. Jackson, Agent-based simulation platforms: review and
development recommendations. Simulation 82(9), 609–623 (2006)
132. CVS Concurrent Version System. http://www.nongnu.org/cvs/, last accessed on March 6,
2012
133. Subversion, http://subversion.apache.org/, last accessed on March 6, 2012
134. Bazaar, http://bazaar.canonical.com/en/, last accessed on March 6, 2012
135. Git - fast version control system. http://git-scm.com/
136. D. Helbing, A. Johansson, Pedestrian, crowd and evacuation dynamics, in Encyclopedia of
Complexity and Systems Science (Springer, Berlin, 2010), pp. 6476–6495
137. D. Helbing, A. Johansson, L. Buzna, Modelisation Du Traffic - Actes du groupe de travail,
chapter New design solutions for pedestrian facilities based on recent empirical results and
computer simulations, pp. 67–88. (2003) Actes No. 104
138. E. Brockfeld, R.D. Kuehne, P. Wagner, Toward benchmarking of microscopic traffic flow
models. Transport. Res. Record 1852, 124–129 (2004)
139. E. Brockfeld, R.D. Kuehne, P. Wagner, Calibration and validation of microscopic traffic flow
models. Transport. Res. Record 1876, 62–70 (2004)
140. D. Helbing, M. Schreckenberg, Cellular automata simulating experimental properties of
traffic flows. Phys. Rev. E 59, R2505–R2508 (2010)
141. A.G. Hoekstra, J. Kroc, P.M.A. Sloot (eds.), chapter Game theoretical interactions of moving
agents, in Simulating Complex Systems by Cellular Automata (Springer, Berlin, 2010),
pp.219–239
142. D. Helbing, W. Yu, The future of social experimenting. PNAS 107(12), 5265–5266 (2010)
143. D. Helbing, T. Platkowski, Self-organization in space and induced by fluctuations. Int. J.
Chaos Theor. Appl. 5(4), (2000)
144. R.J. LeVeque, Numerical Methods for Conservation Laws. (Birkhauser, Basel, 1992)
145. K. De Jong, Evolutionary computation: a unified approach, in GECCO ’08: Proceedings of
the 2008 GECCO conference companion on Genetic and evolutionary computation (ACM,
New York, NY, USA, 2008), pp. 2245–2258
146. P. Ciaccia, M. Patella, P. Zezula, M-tree: An Efficient Access Method for Similarity Search
in Metric Spaces. VLDB J. 426–435 (1997)
147. L.A. Wolsey, An analysis of the greedy algorithm for the submodular set covering problem.
Combinatorica 2(4), 385–393 (1982)
References 69
148. D. Freedman, R. Pisani, R. Purves, Statistics. (W. W. Norton & Company, New York, 1997)
149. E.C. Zeeman, Catastrophe Theory: Selected Papers, 1972–1977 (Addison-Wesley, 1980)
150. T.S. Kuhn, The Structure of Scientific Revolutions. (University Of Chicago Press, Chicago,
1962)
151. D. Helbing, A. Johansson, Evolutionary dynamics of populations with conflicting interac-
tions: Classification and analytical treatment considering asymmetry and power. Phys. Rev. E
81, 016112+ (2010)
152. D. Helbing, A. Johansson, Cooperation, Norms, and Revolutions: A Unified Game-
Theoretical Approach. PLoS ONE 5, e12530 (2010)
153. Benedikt Herrmann, Christian Thöni, and Simon Gächter. Antisocial Punishment Across
Societies. Science 319(5868), 1362–1367 (2008)
154. D. Helbing, Quantitative Sociodynamics: Stochastic Methods and Models of Social Interac-
tion Processes. (Springer, Berlin, 2010)
155. V. Nemytsky, V. Stepanov, Qualitative theory of differential equations. (Princeton University
Press, Princeton NJ, 1960)
156. R. Albert, H. Jeong, A.-L. Barabasi, Error and attack tolerance of complex networks. Nature
406, 378–382 (2000)
157. M. Olson, The Logic of Collective Action : Public Goods and the Theory of Groups. Harvard
economic studies, v. 124. (Harvard University Press, Cambridge, MA, 1971)
158. M. Moussaı̈d, D. Helbing, S. Garnier, A. Johansson, M. Combe, G. Theraulaz, Experimental
study of the behavioural mechanisms underlying self-organization in human crowds. Proc.
Biol. Sci. B: Biol. Sci. 276(1668), 2755–2762 (2009)
159. M. Treiber, A. Kesting, D. Helbing, Three-phase traffic theory and two-phase models
with a fundamental diagram in the light of empirical stylized facts. Transport. Res. B:
Methodological, 44(8–9), 983–1000 (2010)
160. A.-L. Barabási, R. Albert, Emergence of Scaling in Random Networks. Science 286(5439),
509–512 (1999)
161. John P. Ioannidis, Why Most Published Research Findings Are False. PLoS Med. 2(8), e124+
(2005)
162. I. Fisher, The Theory of Interest. (Augustus M. Kelley Publishers, NJ, 1930)
163. B.B. Mandelbrot, The variation of certain speculative prices. J. Bus. 36, 394–419 (1963)
164. R.N. Mantegna, E.H. Stanley, An Introduction to Econophysics: Correlations and Complexity
in Finance. (Cambridge University Press, New York, 1999)
165. V. Pareto, Translation of Manuale di economia politica (“Manual of political economy”).
(A.M. Kelley, New york, 1971)
166. George Zipf, The economy of geography (Addison-Wesley Publishing Co. Inc, Cambridge,
1949), pp.347–415
167. Ravenstein E. The birthplaces of the people and the laws of migration. The Geographical
Magazine III, pp. 173–177, 201–206, 229–233 (1876)
168. G.K. Zipf, The p1 p2/d hypothesis: on the intercity movement of persons. Am. Soc. Rev. 11,
677–686 (1946)
169. J. Tinbergen, Shaping the world economy: suggestions for an international economic policy.
(Twentieth Century Fund, New York, 1962)
170. C.A.E. Goodhart, Monetary relationships: a view from Threadneedle Street. Papers monetary
Econ. 1 (1975)
171. J. Danielsson, The emperor has no clothes: Limits to risk modelling. J. Bank. Finance 26(7),
1273–1296 (2002)
172. D. Helbing, M. Treiber, N.J. Saam, Analytical investigation of innovation dynamics consid-
ering stochasticity in the evaluation of fitness. Phys. Rev. E 71, 067101 (2005)
173. D. Helbing, H. Ammoser, C. Kuehnert, chapter Disasters as extreme events and the
importance of network interactions for disaster response management, in Extreme Events in
Nature and Society (Springer, Berlin, 2005), pp. 319–348
174. Verfahren zur Koordination von vernetzten Abfertigungsprozessen oder zur Steuerung des
Transports von mobilen Einheiten innerhalb eines Netzwerkes, (2010), see http://www.patent-
de.com/20100805/DE102005023742B4.html, last accessed on March 6, 2012
175. D. Helbing, A. Deutsch, S Diez, K Peters, Y Kalaidzidis, K. Padberg-Gehle, S. Laemmer,

A. Johansson, G. Breier, F. Schulze, M. Zerial, Biologistics and the struggle for efficiency:
concepts and perpesctives. Adv. Complex Syst. 12(06), 533+ (2009)
176. L. Hurwicz, Optimality and informational efficiency in resource allocation processes, in
Mathematical Methods in the Social Sciences, ed. by K.J. Arrow, S. Karlin, P. Suppes
(Stanford University Press, Stanford, CA, 1960), pp. 27–46
177. A. Kesting, M. Treiber, M. Schonhof, D. Helbing, Adaptive cruise control design for active
congestion avoidance. Transport. Res. C Emerg. Tech. 16(6), 668–683 (2008)
178. D. Helbing, chapter Dynamic decision behavior and optimal guidance through information
services: Models and experiments, in Human Behaviour and Traffic Networks (Springer,
Berlin, 2004), pp. 47–95
179. D. Helbing, M. Christen, Physik der Finanzmärkte (Wirtschaftswoche, December 22, 2010)
180. N. Wiener, Cybernetics, Second Edition: or the Control and Communication in the Animal
and the Machine. (The MIT Press, Cambridge, MA, 1965)
181. B. Fabien, Analytical System Dynamics: Modeling and Simulation. (Springer, Berlin, 2008)
182. S. Skogestad, I. Postlethwaite, Multivariable Feedback Control: Analysis and Design,
2nd edn. (Wiley-Interscience, New York, 2005)
183. A.L. Fradkov, I.V. Miroshnik, V.O. Nikiforov, Nonlinear and Adaptive Control of Complex
Systems. (Springer, Berlin, 1999)
184. A. Johansson, D. Helbing, P. Shukla, Specification of the social force pedestrian model by
evolutionary adjustment to video tracking data. Adv. Complex Syst. 10, 271–288 (2007)
185. Google flu trends. http://www.google.org/flutrends/, last accessed on March 6, 2012
186. J. Vernon Henderson, A. Storeygard, D.N. Weil, Measuring economic growth from outer
space. Technical Report 15199, National Bureau of Economic Research, July (2009) http://
www.nber.org/papers/w15199.
187. N.A. Christakis, J.H. Fowler, Social network sensors for early detection of contagious
outbreaks. PloS one 5(9), e12948+, September (2010)
188. L. Buzna, K. Peters, H. Ammoser, C. Kuehnert, D. Helbing, Efficient response to cascading
disaster spreading. Phys. Rev. E 75, 056107 (2006)
189. D. Helbing, S. Balietti, From social simulation to intgrative system design. visioneer white
paper, 2010. http://www.visioneer.ethz.ch.
Chapter 3
Self-organization in Pedestrian Crowds
3.1 Introduction
The emergence of new, functional or complex collective behaviors in social systems

has fascinated many scientists. One of the primary questions in this field is how
cooperation or coordination patterns originate based on elementary individual
interactions. While one could think that these are a result of intelligent human
actions, it turns out that much simpler models assuming automatic responses can
reproduce the observations very well. This suggests that humans are using their
intelligence primarily for more complicated tasks, but also that simple interactions
can lead to intelligent patterns of motion. Of course, it is reasonable to assume that
these interactions are the result of a previous learning process that has optimized
the automatic response in terms of minimizing collisions and delays. This, however,
seems to be sufficient to explain most observations.
Note, however, that research into pedestrian and crowd behavior is highly multi-
disciplinary. It involves activities of traffic scientists, psychologists, sociologists,
biologists, physicists, computer scientists, and others. Therefore, it is not surprising
that there are sometimes different or even controversial views on the subject, e.g.
with regard to the concept of “panic”, the explanation of collective, spatio-temporal
patterns of motion in pedestrian crowds, the best modeling concept, or the optimal
number of parameters of a model.
In this contribution, we will start with a short history of pedestrian modeling
and, then, introduce the wide-spread “social force model” of pedestrian interactions
to illustrate further issues such as, for example, model calibration by video tracking
data. Next, we will turn to the subject of crowd dynamics, since one typically
finds the formation of large-scale spatio-temporal patterns of motion, when many

This chapter reprints parts of a previous publication with kind permission of the copyright owner,
Springer Publishers. It is requested to cite this work as follows: D. Helbing and A. Johansson
(2010) Pedestrian, crowd and evacuation dynamics, in: Encyclopedia of Complexity and Systems
Science (Springer, New York), Vol. 16, pp. 6476–6495.

72 3 Self-organization in Pedestrian Crowds
pedestrians interact with each other. These patterns will be discussed in some detail
before we will turn to evacuation situations and cases of extreme densities, where
one can sometimes observe the breakdown of coordination. Finally, we will address
possibilities to design improved pedestrian facilities, using special evolutionary
algorithms.
3.2 Pedestrian Dynamics
3.2.1 Short History of Pedestrian Modeling
Pedestrians have been empirically studied for more than four decades [1–3]. The
evaluation methods initially applied were based on direct observation, photographs,
and time-lapse films. For a long time, the main goal of these studies was to
develop a level-of-service concept [4], design elements of pedestrian facilities [5–8],
or planning guidelines [9, 10]. The latter have usually the form of regression
relations, which are, however, not very well suited for the prediction of pedestrian
flows in pedestrian zones and buildings with an exceptional architecture, or in
challenging evacuation situations. Therefore, a number of simulation models have
been proposed, e.g. queueing models [11], transition matrix models [12], and
stochastic models [13], which are partly related to each other. In addition, there
are models for the route choice behavior of pedestrians [14, 15].
None of these concepts adequately takes into account the self-organization
effects occurring in pedestrian crowds. These are the subject of recent experimental
studies [8,16–20]. Most pedestrian models, however, were formulated before. A first
modeling approach that appears to be suited to reproduce spatio-temporal patterns
of motion was proposed by Henderson [21], who conjectured that pedestrian crowds
behave similar to gases or fluids (see also [22]). This could be partially confirmed,
but a realistic gas-kinetic or fluid-dynamic theory for pedestrians must contain
corrections due to their particular interactions (i.e. avoidance and deceleration
maneuvers) which, of course, do not obey momentum and energy conservation.
Although such a theory can be actually formulated [23,24], for practical applications
a direct simulation of individual pedestrian motion is favourable, since this is more
flexible. As a consequence, pedestrian research mainly focusses on agent-based
models of pedestrian crowds, which also allow one to consider local coordination
problems. The “social force model” [25, 26] is maybe the most well-known of
these models, but we also like to mention cellular automata of pedestrian dynamics
[27–33] and AI-based models [34, 35].
3.2.2 The Social Force Concept
In the following, we shall shortly introduce the social force concept, which repro-
duces most empirical observations in a simple and natural way. Human behavior
3.2 Pedestrian Dynamics 73
often seems to be “chaotic”, irregular, and unpredictable. So, why and under what
conditions can we model it by means of forces? First of all, we need to be confronted
with a phenomenon of motion in some (quasi-)continuous space, which may be also
an abstract behavioral space such as an opinion scale [36]. Moreover, it is favourable
to have a system where the fluctuations due to unknown influences are not large
compared to the systematic, deterministic part of motion. This is usually the case
in pedestrian traffic, where people are confronted with standard situations and react
“automatically” rather than taking complicated decisions, e.g. if they have to evade
others.
This “automatic” behavior can be interpreted as the result of a learning process
based on trial and error [37], which can be simulated with evolutionary algorithms
[38]. For example, pedestrians have a preferred side of walking, since an asymmet-
rical avoidance behavior turns out to be profitable [25, 37]. The related formation
of a behavioral convention can be described by means of evolutionary game theory
[25, 39].
Another requirement is the vectorial additivity of the separate force terms
reflecting different environmental influences. This is probably an approximation, but
there is some experimental evidence for it. Based on quantitative measurements for
animals and test persons subject to separately or simultaneously applied stimuli of
different nature and strength, one could show that the behavior in conflict situations
can be described by a superposition of forces [40, 41]. This fits well into a concept
by Lewin [42], according to which behavioral changes are guided by so-called
social fields or social forces, which has later on been put into mathematical terms
[25, 43]. In some cases, social forces, which determine the amount and direction
of systematic behavioral changes, can be expressed as gradients of dynamically
varying potentials, which reflect the social or behavioral fields resulting from the
interactions of individuals. Such a social force concept was applied to opinion
formation and migration [43], and it was particularly successful in the description
of collective pedestrian behavior [8, 25, 26, 37].
For reliable simulations of pedestrian crowds, we do not need to know whether
a certain pedestrian, say, turns to the right at the next intersection. It is sufficient
to have a good estimate what percentage of pedestrians turns to the right. This can
be either empirically measured or estimated by means of route choice models [14].
In some sense, the uncertainty about the individual behaviors is averaged out at
the macroscopic level of description. Nevertheless, we will use the more flexible
microscopic simulation approach based on the social force concept. According to
this, the temporal change of the location r˛ .t/ of pedestrian ˛ obeys the equation
dr˛ .t/
D v˛ .t/: (3.1)
dt
Moreover, if f˛ .t/ denotes the sum of social forces influencing pedestrian ˛ and if
˛ .t/ are individual fluctuations reflecting unsystematic behavioral variations, the
velocity changes are given by the acceleration equation
d v˛
D f˛ .t/ C ˛ .t/: (3.2)
dt
A particular advantage of this approach is that we can take into account the flexible
usage of space by pedestrians, requiring a continuous treatment of motion. It turns
out that this point is essential to reproduce the empirical observations in a natural
and robust way, i.e. without having to adjust the model to each single situation and
measurement site. Furthermore, it is interesting to note that, if the fluctuation term
is neglected, the social force model can be interpreted as a particular differential
game, i.e. its dynamics can be derived from the minimization of a special utility
function [44].
3.2.3 Specification of the Social Force Model
The social force model for pedestrians assumes that each individual ˛ is trying to
move in a desired direction e˛0 with a desired speed v0˛ , and that it adapts the actual
velocity v˛ to the desired one, v0˛ D v0˛ e˛0 , within a certain relaxation time ˛ . The
systematic part f˛ .t/ of the acceleration force of pedestrian ˛ is then given by
1 0 0 X X
f˛ .t/ D .v˛ e˛ v˛ / C f˛ˇ .t/ C f˛i .t/; (3.3)
˛ i
ˇ.¤˛/
where the terms f˛ˇ .t/ and f˛i .t/ denote the repulsive forces describing attempts to
keep a certain safety distance to other pedestrians ˇ and obstacles i . In very crowded
situations, additional physical contact forces come into play (see Sect. 3.4.3).
Further forces may be added to reflect attraction effects between members of a group
or other influences. For details see [37].
First, we will assume a simplified interaction force of the form

f˛ˇ .t/ D f d˛ˇ .t/ ; (3.4)
where d˛ˇ D r˛ rˇ is the distance vector pointing from pedestrian ˇ to ˛.

Angular-dependent shielding effects may be furthermore taken into account by a
prefactor describing the anisotropic reaction to situations in front of as compared to
behind a pedestrian [26, 45], see Sect. 3.2.4. However, we will start with a circular
specification of the distance-dependent interaction force,
d˛ˇ
f .d˛ˇ / D A˛ ed˛ˇ =B˛ ; (3.5)
kd˛ˇ k
where d˛ˇ D kd˛ˇ k is the distance. The parameter A˛ reflects the interaction
strength, and B˛ corresponds to the interaction range. While the dependence on ˛
explicitly allows for a dependence of these parameters on the single individual, we
will assume a homogeneous population, i.e. A˛ D A and B˛ D B in the following.
Otherwise, it would be hard to collect enough data for parameter calibration.
Elliptical specification: Note that it is possible to express (3.5) as gradient of an

exponentially decaying potential V˛ˇ . This circumstance can be used to formulate a
generalized, elliptical interaction force via the potential
V˛ˇ .b˛ˇ / D AB eb˛ˇ =B ; (3.6)
where the variable b˛ˇ denotes the semi-minor axis b˛ˇ of the elliptical equipotential
lines. This has been specified according to
q
2b˛ˇ D .kd˛ˇ k C kd˛ˇ .vˇ v˛ /tk/2 k.vˇ v˛ /tk2 ; (3.7)
so that both pedestrians ˛ and ˇ are treated symmetrically. The repulsive force is
related to the above potential via
dV˛ˇ .b˛ˇ /
f˛ˇ .d˛ˇ / D rd˛ˇ V˛ˇ .b˛ˇ / D rd˛ˇ b˛ˇ .d˛ˇ /; (3.8)
db˛ˇ
where rd˛ˇ represents the gradient with respect to d˛ˇ . Considering the chain rule,
p p
kzk D z2 , and rz kzk D z= z2 D z=kzk, this leads to the explicit formula

kd˛ˇ k C kd˛ˇ y˛ˇ k 1 d˛ˇ d˛ˇ y˛ˇ
f˛ˇ .d˛ˇ / D Aeb˛ˇ =B C
2b˛ˇ 2 kd˛ˇ k kd˛ˇ y˛ˇ k
(3.9)
with y˛ˇ D .vˇ v˛ /t. We used t D 0:5 s. For t D 0, we regain the expression
of (3.5).
The elliptical specification has two major advantages compared to the circular
one: First, the interactions depend not only on the distance, but also on the relative
velocity. Second, the repulsive force is not strictly directed from pedestrian ˇ to
pedestrian ˛, but has a lateral component. As a consequence, this leads to less
confrontative, smoother (“sliding”) evading maneuvers. Note that further velocity-
dependent specifications of pedestrian interaction forces have been proposed [7,26],
but we will restrict to the above specifications, as these are sufficient to demonstrate
the method of evolutionary model calibration. For suggested improvements regard-
ing the specification of social forces see, for example, [46, 47].
3.2.4 Angular Dependence
In reality, of course, pedestrian interactions are not isotropic, but dependent on the
angle '˛ˇ of the encounter, which is given by the formula
v˛ d˛ˇ
cos.'˛ˇ / D : (3.10)
kv˛ k kd˛ˇ k
Generally, pedestrians show little response to pedestrians behind them. This can be
reflected by an angular-dependent prefactor w.'˛ˇ / of the interaction force [45].
Empirical results are represented in Fig. 3.2 (right). Reasonable results are obtained
for the following specification of the prefactor:

1 C cos.'˛ˇ /
w '˛ˇ .t/ D ˛ C .1 ˛ / ; (3.11)
2
where ˛ with 0 ˛ 1 is a parameter which grows with the strength of inter-

actions from behind. An evolutionary parameter optimization gives values 0:1
(see Sect. 3.2.5, i.e. a strong anisotropy. Other angular-dependent specifications split
up the interaction force between pedestrians into a component against the direction
of motion and another one perpendicular to it. Such a description allows for even
smoother avoidance maneuvers.
3.2.5 Evolutionary Calibration with Video Tracking Data
For parameter calibration, several video recordings of pedestrian crowds in different

natural environments have been used. The dimensions of the recorded areas were
known, and the floor tiling or environment provided something like a “coordinate
system”. The heads were automatically determined by searching for round moving
structures, and the accuracy of tracking was improved by comparing actual with
linearly extrapolated positions (so it would not happen so easily that the algorithm
interchanged or “lost” closeby pedestrians). The trajectories of the heads were
then projected on two-dimensional space in a way correcting for distortion by
the camera perspective. A representative plot of the resulting trajectories is shown
in Fig. 3.1. Note that trajectory data have been obtained with infra-red sensors
[48] or video cameras [49, 50] for several years now, but algorithms that can
simultaneously handle more than one thousand pedestrians have become available
Fig. 3.1 Video tracking used to extract the trajectories of pedestrians from video recordings close
to two escalators (after [45]). Left: Illustration of the tracking of pedestrian heads. Right: Resulting
trajectories after being transformed onto the two-dimensional plane
only recently [51]. For model calibration, it is recommended to use a hybrid method
fusing empirical trajectory data and microscopic simulation data of pedestrian
movement in space. In corresponding algorithms, a virtual pedestrian is assigned
to each tracked pedestrian in the simulation domain. One then starts a simulation
for a time period T (e.g. 1.5 s), in which one pedestrian ˛ is moved according to a
simulation of the social force model, while the others are moved exactly according
to the trajectories extracted from the videos. This procedure is performed for all
pedestrians ˛ and for several different starting times t, using a fixed parameter set
for the social force model.
Each simulation run is performed according to the following scheme:
1. Define a starting point and calculate the state (position r˛ , velocity v˛ , and
acceleration a˛ D d v˛ =dt) for each pedestrian ˛.
2. Assign a desired speed v0˛ to each pedestrian, e.g. the maximum speed during the
pedestrian tracking time. This is sufficiently accurate, if the overall pedestrian
density is not too high and the desired speed is constant in time.
3. Assign a desired goal point for each pedestrian, e.g. the end point of the
trajectory.
4. Given the tracked motion of the surrounding pedestrians ˇ, simulate the trajec-
tory of pedestrian ˛ over a time period T based on the social force model, starting
at the actual location r˛ .t/.
After each simulation run, one determines the relative distance error
kr˛simulated .t C T / r˛tracked .t C T /k
: (3.12)
kr˛tracked .t C T / r˛tracked .t/k
After averaging the relative distance errors over the pedestrians ˛ and starting
times t, 1 minus the result can be taken as measure of the goodness of fit (the
“fitness”) of the parameter set used in the pedestrian simulation. Hence, the best
possible value of the “fitness” is 1, but any deviation from the real pedestrian
trajectories implies lower values.
One result of such a parameter optimization is that, for each video, there is a
broad range of parameter combinations of A and B which perform almost equally
well [45]. This allows one to apply additional goal functions in the parameter
optimization, e.g. to determine among the best performing parameter values such
parameter combinations, which perform well for several video recordings, using
a fitness function which equally weights the fitness reached in each single video.
This is how the parameter values listed in Table 3.1 were determined. It turns out
that, in order to reach a good model performance, the pedestrian interaction force
must be specified velocity dependent, as in the elliptical model.
Note that our evolutionary fitting method can be also used to determine interac-
tion laws without prespecified interaction functions. For example, one can obtain
the distance dependence of pedestrian interactions without a pre-specified function.
For this, one adjusts the values of the force at given distances dk D kd1 (with k 2
f1; 2; 3; : : :g) in an evolutionary way. To get some smoothness, linear interpolation
Table 3.1 Interaction strength A and interaction range B resulting from our evolutionary
parameter calibration for the circular and elliptical specification of the interaction forces between
pedestrians (see main text), with an assumed angular dependence according to (3.11). A com-
parison with the extrapolation scenario, which assumes constant speeds, allows one to judge the
improvement in the goodness of fit (“fitness”) by the specified interaction force. The calibration
was based on three different video recordings, one for low crowd density, one for medium, and
one for high density (see [45] for details). The parameter values are specified as mean value ˙
standard deviation. The best fitness value obtained with the elliptical specification for the video
with the lowest crowd density was as high as 0.9
Model A [m/s2 ] B [m] Fitness
Extrapolation 0 – – 0.34
Circular 0.42 ˙ 0.26 1.65 ˙ 1.01 0.12 ˙ 0.07 0.40
Elliptical 0.04 ˙ 0.01 3.22 ˙ 0.67 0.06 ˙ 0.04 0.61
Fig. 3.2 Results of an evolutionary fitting of pedestrian interactions. Left: Empirically determined
distance dependence of the interaction force between pedestrians. An exponential decay fits the
empirical data quite well. The dashed fit curve corresponds to (3.5) with the parameters A D 0:53
and B D 1:0. Right: Angular dependence of the influence of other pedestrians. The direction along
the positive x axis corresponds to the walking direction of pedestrians, y to the perpendicular
direction (After [45])
is applied. The resulting fit curve is presented in Fig. 3.2 (left). It turns out that the
empirical dependence of the force with distance can be well fitted by an exponential
decay.
3.3 Crowd Dynamics
3.3.1 Analogies with Gases, Fluids, and Granular Media
When the density is low, pedestrians can move freely, and the observed crowd
dynamics can be partially compared with the behavior of gases. At medium and high
densities, however, the motion of pedestrian crowds shows some striking analogies
with the motion of fluids:
3.3 Crowd Dynamics 79
1. Footprints of pedestrians in snow look similar to streamlines of fluids [15].

2. At borderlines between opposite directions of walking one can observe “viscous
fingering” [52, 53].
3. The emergence of pedestrian streams through standing crowds [7,37,54] appears
analogous to the formation of river beds [55, 56].
At high densities, however, the observations have rather analogies with driven
granular flows. This will be elaborated in more detail in Sects. 3.4.3 and 3.4.4. In
summary, one could say that fluid-dynamic analogies work reasonably well in nor-
mal situations, while granular aspects dominate at extreme densities. Nevertheless,
the analogy is limited, since the self-driven motion and the violation of momentum
conservation imply special properties of pedestrian flows. For example, one usually
does not observe eddies.
3.3.2 Self-organization of Pedestrian Crowds
Despite its simplifications, the social force model of pedestrian dynamics describes
a lot of observed phenomena quite realistically. Especially, it allows one to explain
various self-organized spatio-temporal patterns that are not externally planned,
prescribed, or organized, e.g. by traffic signs, laws, or behavioral conventions
[7, 8, 37]. Instead, the spatio-temporal patterns discussed below emerge due to
the non-linear interactions of pedestrians even without assuming strategical con-
siderations, communication, or imitative behavior of pedestrians. Despite this, we
may still interpret the forming cooperation patterns as phenomena that establish
social order on short time scales. It is actually surprising that strangers coordinate
each other within seconds, if they have grown up in a similar environment. People
from different countries, however, are sometimes irritated about local walking
habits, which indicates that learning effects and cultural backgrounds still play a
role in social interactions as simple as random pedestrian encounters. Rather than
on particular features, however, in the following we will focus on the common,
internationally reproducible observations.
3.3.2.1 Lane Formation
In pedestrian flows one can often observe that oppositely moving pedestrians are
forming lanes of uniform walking direction (see Fig. 3.3) [8, 20, 25, 26]. This
phenomenon even occurs when there is not a large distance to separate each other,
e.g. on zebra crossings. However, the width of lanes increases (and their number
decreases), if the interaction continues over longer distances (and if perturbations,
e.g. by flows entering or leaving on the sides, are low; otherwise the phenomenon
of lane formation may break down [57]).
Fig. 3.3 Self-organization of pedestrian crowds. Left: Photograph of lanes formed in a shopping
center. Computer simulations reproduce the self-organization of such lanes very well. Top right:
Evaluation of the cumulative number of pedestrians passing a bottleneck from different sides.
One can clearly see that the narrowing is often passed by groups of people in an oscillatory way
rather than one by one. Bottom right: Multi-agent simulation of two crossing pedestrian streams,
showing the phenomenon of stripe formation. This self-organized pattern allows pedestrians to
pass the other stream without having to stop, namely by moving sidewards in a forwardly moving
stripe (After [8])
Lane formation may be viewed as segregation phenomenon [58, 59]. Although

there is a weak preference for one side (with the corresponding behavioral conven-
tion depending on the country), the observations can only be well reproduced when
repulsive pedestrian interactions are taken into account. The most relevant factor
for the lane formation phenomenon is the higher relative velocity of pedestrians
walking in opposite directions. Compared to people following each other, oppositely
moving pedestrians have more frequent interactions until they have segregated into
separate lanes by stepping aside whenever another pedestrian is encountered. The
most long-lived patterns of motion are the ones which change the least. It is obvious
that such patterns correspond to lanes, as they minimize the frequency and strength
of avoidance maneuvers. Interestingly enough, as computer simulations show, lane
formation occurs also when there is no preference for any side.
Lanes minimize frictional effects, accelerations, energy consumption, and delays
in oppositely moving crowds. Therefore, one could say that they are a pattern
reflecting “collective intelligence”. In fact, it is not possible for a single pedestrian
to reach such a collective pattern of motion. Lane formation is a self-organized
collaborative pattern of motion originating from simple pedestrian interactions.
Particularly in cases of no side preference, the system behavior cannot be understood
3.3 Crowd Dynamics 81
by adding up the behavior of the single individuals. This is a typical feature of

complex, self-organizing systems and, in fact, a wide-spread characteristics of social
systems. It is worth noting, however, that it does not require a conscious behavior
to reach forms of social organization like the segregation of oppositely moving
pedestrians into lanes. This organization occurs automatically, although most people
are not even aware of the existence of this phenomenon.
3.3.2.2 Oscillatory Flows at Bottlenecks
At bottlenecks, bidirectional flows of moderate density are often characterized by

oscillatory changes in the flow direction (see Fig. 3.3) [8, 26]. For example, one can
sometimes observe this at entrances of museums during crowded art exhibitions or
at entrances of staff canteens during lunch time. While these oscillatory flows may
be interpreted as an effect of friendly behavior (“you go first, please”), computer
simulations of the social force model indicate that the collective behavior may again
be understood by simple pedestrian interactions. That is, oscillatory flows can even
occur in the absence of communication, although it may be involved in reality.
The interaction-based mechanism of oscillatory flows suggests to interpret them as
another self-organization phenomenon, which again reduces frictional effects and
delays. That is, oscillatory flows have features of “collective intelligence”.
While this may be interpreted as result of a learning effect in a large number of
similar situations (a “repeated game”), our simulations suggest an even simpler,
“many-particle” interpretation: Once a pedestrian is able to pass the narrowing,
pedestrians with the same walking direction can easily follow. Hence, the number
and “pressure” of waiting, “pushy” pedestrians on one side of the bottleneck
becomes less than on the other side. This eventually increases their chance to occupy
the passage. Finally, the “pressure difference” is large enough to stop the flow
and turn the passing direction at the bottleneck. This reverses the situation, and
eventually the flow direction changes again, giving rise to oscillatory flows.
At bottlenecks, further interesting observations can be made: Hoogendoorn and
Daamen [60] report the formation of layers in unidirectional bottleneck flows. Due
to the partial overlap of neighboring layers, there is a zipper effect. Moreover, Kretz
et al. [61] have observed that the specific flow through a narrow bottleneck decreases
with a growing width of the bottleneck, as long as it can be passed by one person at
a time only. This is due to mutual obstructions, if two people are trying to enter the
bottleneck simultaneously. If the opening is large enough to be entered by several
people in parallel, the specific flow stays constant with increasing width. Space is
then used in a flexible way.
3.3.2.3 Stripe Formation in Intersecting Flows
In intersection areas, the flow of people often appears to be irregular or “chaotic”.

In fact, it can be shown that there are several possible collective patterns of motion,
among them rotary and oscillating flows. However, these patterns continuously com-
pete with each other, and a temporarily dominating pattern is destroyed by another
one after a short time. Obviously, there has not evolved any social convention that
would establish and stabilize an ordered and efficient flow at intersections.
Self-organized patterns of motion, however, are found in situations where
pedestrian flows cross each other only in two directions. In such situations, the
phenomenon of stripe formation is observed [62]. Stripe formation allows two flows
to penetrate each other without requiring the pedestrians to stop. For an illustration
see Fig. 3.3. Like lanes, stripes are a segregation phenomenon, but not a stationary
one. Instead, the stripes are density waves moving into the direction of the sum of the
directional vectors of both intersecting flows. Naturally, the stripes extend sidewards
into the direction which is perpendicular to their direction of motion. Therefore, the
pedestrians move forward with the stripes and sidewards within the stripes. Lane
formation corresponds to the particular case of stripe formation where both direc-
tions are exactly opposite. In this case, no intersection takes place, and the stripes do
not move systematically. As in lane formation, stripe formation allows to minimize
obstructing interactions and to maximize the average pedestrian speeds, i.e. simple,
repulsive pedestrian interactions again lead to an “intelligent” collective behavior.
3.4 Evacuation Dynamics
While the previous section has focussed on the dynamics of pedestrian crowds in
normal situations, we will now turn to the description of situations in which extreme
crowd densities occur. Such situations may arise at mass events, particularly in
cases of urgent egress. While most evacuations run relatively smoothly and orderly,
the situation may also get out of control and end up in terrible crowd disasters
(see Table 3.2). In such situations, one often speaks of “panic”, although, from a
scientific standpoint, the use of this term is rather controversial. Here, however,
we will not be interested in the question whether “panic” actually occurs or not.
We will rather focus on the issue of crowd dynamics at high densities and under
psychological stress.
3.4.1 Evacuation and Panic Research
Computer models have been also developed for emergency and evacuation situa-
tions [32, 63–71]. Most research into panic, however, has been of empirical nature
(see, e.g. [72–74]), carried out by social psychologists and others.
With some exceptions, panic is thought to occur in cases of scarce or dwindling
resources [75, 76], which are either required for survival or anxiously desired. They
are usually distinguished into escape panic (“stampedes”, bank or stock market
panic) and acquisitive panic (“crazes”, speculative manias) [77, 78], but in some
cases this classification is questionable [79].
3.4 Evacuation Dynamics 83
Table 3.2 Incomplete list of major crowd disasters since 1970 after J. F. Dickie in
[89], http://www.crowddynamics.com/Main/Crowddisasters.html, http://SportsIllustrated.CNN.
com/soccer/world/news/2000/07/09/stadium disasters ap/, and other internet sources, excluding
fires, bomb attacks, and train or plane accidents. The number of injured people was usually a
multiple of the fatalities
Date Place Venue Deaths Reason
1971 Ibrox, UK Stadium 66 Collapse of barriers
1974 Cairo, Egypt Stadium 48 Crowds break barriers
1982 Moscow, USSR Stadium 340 Re-entering fans after last
minute goal
1988 Katmandu, Nepal Stadium 93 Stampede due to hailstorm
1989 Hillsborough, Stadium 96 Fans trying to force their
Sheffield, UK way into the stadium
1990 New York City Bronx 87 Illegal happy land social
club
1990 Mena, Saudi Arabia Pedestrian tunnel 1;426 Overcrowding
1994 Mena, Saudi Arabia Jamarat bridge 266 Overcrowding
1996 Guatemala City, Stadium 83 Fans trying to force their
Guatemala way into the stadium
1998 Mena, Saudi Arabia 118 Overcrowding
1999 Kerala, India Hindu shrine 51 Collapse of parts of the
shrine
1999 Minsk, Belarus Subway station 53 Heavy rain at rock concert
2001 Ghana, West Africa Stadium >100 Panic triggered by tear gas
2005 Wai, India Religious procession 150 Overcrowding (and fire)
2005 Bagdad, Iraque Religious procession >640 Rumors regarding suicide
bomber
2005 Chennai, India Disaster area 42 Rush for flood relief
supplies
2006 Pilippines Stadium 79 Rush for game show tickets
2006 Ibb, Yemen Stadium 51 Rally for Yemeni president
It is often stated that panicking people are obsessed by short-term personal

interests uncontrolled by social and cultural constraints [76, 77]. This is possibly
a result of the reduced attention in situations of fear [76], which also causes that
options like side exits are mostly ignored [72]. It is, however, mostly attributed to
social contagion [73, 75–84], i.e., a transition from individual to mass psychology,
in which individuals transfer control over their actions to others [78], leading to
conformity [85]. This “herding behavior” is in some sense irrational, as it often leads
to bad overall results like dangerous overcrowding and slower escape [72,78,79]. In
this way, herding behavior can increase the fatalities or, more generally, the damage
in the crisis faced.
The various socio-psychological theories for this contagion assume hypnotic
effects, rapport, mutual excitation of a primordial instinct, circular reactions, social
facilitation (see the summary by Brown [83]), or the emergence of normative
support for selfish behavior [84]. Brown [83] and Coleman [78] add another
explanation related to the prisoner’s dilemma [86, 87] or common goods dilemma
[88], showing that it is reasonable to make one’s subsequent actions contingent
upon those of others. However, the socially favourable behavior of walking orderly
is unstable, which normally gives rise to rushing by everyone. These thoughtful
considerations are well compatible with many aspects discussed above and with
the classical experiments by Mintz [75], which showed that jamming in escape
situations depends on the reward structure (“payoff matrix”).
Nevertheless and despite of the frequent reports in the media and many published
investigations of crowd disasters (see Table 3.2), a quantitative understanding of the
observed phenomena in panic stampedes was lacking for a long time. The following
sections will close this gap.
3.4.2 Situations of “Panic”
Panic stampede is one of the most tragic collective behaviors [73–75, 77, 78, 80–84],
as it often leads to the death of people who are either crushed or trampled
down by others. While this behavior may be comprehensible in life-threatening
situations like fires in crowded buildings [72, 76], it is hard to understand in cases
of a rush for good seats at a pop concert [79] or without any obvious reasons.
Unfortunately, the frequency of such disasters is increasing (see Table 3.2), as
growing population densities combined with easier transportation lead to greater
mass events like pop concerts, sport events, and demonstrations. Nevertheless,
systematic empirical studies of panic [75, 90] are rare [76, 77, 79], and there
is a scarcity of quantitative theories capable of predicting crowd dynamics at
extreme densities [32, 63, 64, 67, 68, 71]. The following features appear to be typical
[57, 91]:
1. In situations of escape panic, individuals are getting nervous, i.e. they tend to
develop blind actionism.
2. People try to move considerably faster than normal [9].
3. Individuals start pushing, and interactions among people become physical in
nature.
4. Moving and, in particular, passing of a bottleneck frequently becomes incoor-
dinated [75].
5. At exits, jams are building up [75]. Sometimes, intermittent flows or arching
and clogging are observed [9].
6. The physical interactions in jammed crowds add up and can cause dangerous
pressures up to 4,500 Newtons per meter [72,89], which can bend steel barriers
or tear down brick walls.
7. The strength and direction of the forces acting in large crowds can suddenly
change [51], pushing people around in an uncontrollable way. This may cause
people to fall.
8. Escape is slowed down by fallen or injured people turning into “obstacles”.
9. People tend to show herding behavior, i.e., to do what other people do [76, 81].
10. Alternative exits are often overlooked or not efficiently used in escape situations
[72, 76].
3.4.3 Force Model for Panicking Pedestrians
ph
Additional, physical interaction forces f˛ˇ come into play when pedestrians get so
close to each other that they have physical contact (i.e. d˛ˇ < r˛ˇ D r˛ C rˇ , where
r˛ means the “radius” of pedestrian ˛) [91]. In this case, which is mainly relevant to
panic situations, we assume also a “body force” k.r˛ˇ d˛ˇ / n˛ˇ counteracting body
compression and a “sliding friction force” .r˛ˇ d˛ˇ / vtˇ˛ t˛ˇ impeding relative
tangential motion. Inspired by the formulas for granular interactions [92, 93], we
assume
ph
f˛ˇ .t/ D k.r˛ˇ d˛ˇ /n˛ˇ C .r˛ˇ d˛ˇ /vtˇ˛ t˛ˇ ; (3.13)
where the function .z/ is equal to its argument z, if z 0, otherwise 0. Moreover,
t˛ˇ D .n2˛ˇ ; n1˛ˇ / means the tangential direction and vtˇ˛ D .vˇ v˛ / t˛ˇ the
tangential velocity difference, while k and represent large constants. (Strictly
speaking, friction effects already set in before pedestrians touch each other, because
of the psychological tendency not to pass other individuals with a high relative
velocity, when the distance is small.)
The interactions with the boundaries of walls and other obstacles are treated anal-
ogously to pedestrian interactions, i.e., if d˛i .t/ means the distance to obstacle or
boundary i , n˛i .t/ denotes the direction perpendicular to it, and t˛i .t/ the direction
tangential to it, the corresponding interaction force with the boundary reads
f˛i D fA˛ expŒ.r˛ d˛i /=B˛ C k.r˛ d˛i /g n˛i .r˛ d˛i /.v˛ t˛i / t˛i :
(3.14)
Finally, fire fronts are reflected by repulsive social forces similar those describing
walls, but they are much stronger. The physical interactions, however, are
qualitatively different, as people reached by the fire front become injured and
immobile (v˛ D 0).
3.4.4 Collective Phenomena in Situations of “Panic”
Inspired by the observations discussed in Sect. 3.4.2, we have simulated situations

of “panic” escape in the computer, assuming the following features:
1. People are getting nervous, resulting in a higher level of fluctuations.
2. They are trying to escape from the source of panic, which can be reflected by a
significantly higher desired velocity v0˛ .
3. Individuals in complex situations, who do not know what is the right thing to do,
orient at the actions of their neighbours, i.e. they tend to do what other people
do. We will describe this by an additional herding interaction.
We will now discuss the fundamental collective effects which fluctuations, increased
desired velocities, and herding behavior can have according to simulations. Note
that, in contrast to other approaches, we do not assume or imply that individuals in
panic or emergency situations would behave relentless and asocial, although they
sometimes do.
3.4.4.1 Herding and Ignorance of Available Exits
If people are not sure what is the best thing to do, there is a tendency to show a
“herding behavior”, i.e. to imitate the behavior of others. Fashion, hypes and trends
are examples for this. The phenomenon is also known from stock markets, and
particularly pronounced when people are anxious. Such a situation is, for example,
given if people need to escape from a smoky room. There, the evacuation dynamics
is very different from normal leaving (see Fig. 3.4).
Fig. 3.4 Left: Normal leaving of a room, when the exit is well visible. Snapshots of a video-
recorded experiment with ten people after (a) t D 0 s (initial condition), (b) t D 1 s, (c) t D 3 s,
and (d) t D 5 s. The face directions are indicated by arrows. Right: Escape from a room with
no visibility, e.g. due to dense smoke or a power blackout. Snapshots of an experiment with test
persons, whose eyes were covered by masks, after t D 0 s (initial condition), t D 5 s, (c) t D 10 s,
and (d) t D 15 s (After [18])
Under normal visibility, everybody easily finds an exit and uses more or less
the shortest path. However, when the exit cannot be seen, evacuation is much less
efficient and may take a long time. Most people tend to walk relatively straight
into the direction in which they suspect an exit, but in most cases, they end up at a
wall. Then, they usually move along it in one of the two possible directions, until
they finally find an exit [18]. If they encounter others, there is a tendency to take a
decision for one direction and move collectively. Also in case of acoustic signals,
people may be attracted into the same direction. This can lead to over-crowded
exits, while other exits are ignored. The same can happen even for normal visibility,
when people are not well familiar with their environment and are not aware of the
directions of the emergency exits.
Computer simulations suggest that neither individualistic nor herding behavior
performs well [91]. Pure individualistic behavior means that each pedestrian finds
an exit only accidentally, while pure herding behavior implies that the complete
crowd is eventually moving into the same and probably congested direction, so
that available emergency exits are not efficiently used. Optimal chances of survival
are expected for a certain mixture of individualistic and herding behavior, where
individualism allows some people to detect the exits and herding guarantees that
successful solutions are imitated by small groups of others [91].
3.4.4.2 “Freezing by Heating”
Another effect of getting nervous has been investigated in [57]. Let us assume the
individual fluctuation strength, i.e. the standard deviation of the noise term ˛ , is
given by
˛ D .1 n˛ / 0 C n˛ max ; (3.15)
where n˛ with 0 n˛ 1 measures the nervousness of pedestrian ˛. The parameter
0 means the normal and max the maximum fluctuation strength. It turns out that,
at sufficiently high pedestrian densities, lanes are destroyed by increasing the
fluctuation strength (which is analogous to the temperature). However, instead of
the expected transition from the “fluid” lane state to a disordered, “gaseous” state,
a “solid” state is formed. It is characterized by a blocked, “frozen” situation so that
one calls this paradoxial transition “freezing by heating” (see Fig. 3.5). Notably
Fig. 3.5 Result of the noise-induced formation of a “frozen” state in a (periodic) corridor used by
oppositely moving pedestrians (after [57])
enough, the blocked state has a higher degree of order, although the internal energy
is increased [57].
The preconditions for this unusual freezing-by-heating transition are the driving
term v0˛ e˛0 =˛ and the dissipative friction v˛ =˛ , while the sliding friction force is
not required. Inhomogeneities in the channel diameter or other impurities which
temporarily slow down pedestrians can further this transition at the respective
places. Finally note that a transition from fluid to blocked pedestrian counter flows
is also observed, when a critical density is exceeded, as impatient pedestrians enter
temporary gaps in the opposite lane to overtake others [31,57]. However, in contrast
to computer simulations, resulting deadlocks are usually not permanent in real
crowds, as turning the bodies (shoulders) often allows pedestrians to get out of the
blocked area.
3.4.4.3 Intermittent Flows, Faster-Is-Slower Effect, and “Phantom Panic”
If the overall flow towards a bottleneck is higher than the overall outflow from it, a
pedestrian queue emerges [94]. In other words, a waiting crowd is formed upstream
of the bottleneck. High densities can result, if people keep heading forward, as this
eventually leads to higher and higher compressions. Particularly critical situations
may occur if the arrival flow is much higher than the departure flow, especially if
people are trying to get towards a strongly desired goal (“aquisitive panic”) or away
from a perceived source of danger (“escape panic”) with an increased driving force
v0˛ e˛0 =. In such situations, the high density causes coordination problems, as several
people compete for the same few gaps. This typically causes body interactions
and frictional effects, which can slow down crowd motion or evacuation (“faster
is slower effect”).
A possible consequence of these coordination problems are intermittent flows.
In such cases, the outflow from the bottleneck is not constant, but it is typically
interrupted. While one possible origin of the intermittent flows are clogging and
arching effects as known from granular flows through funnels or hoppers [92, 93],
stop-and-go waves have also been observed in more than 10 m wide streets and in the
44 m wide entrance area to the Jamarat Bridge during the pilgrimage in January 12,
2006 [51], see Fig. 3.6. Therefore, it seems to be important that people do not move
continuously, but have minimum strides [25]. That is, once a person is stopped,
he or she will not move until some space opens up in front. However, increasing
impatience will eventually reduce the minimum stride, so that people eventually
start moving again, even if the outflow through the bottleneck is stopped. This will
lead to a further compression of the crowd.
In the worst case, such behavior can trigger a “phantom panic”, i.e. a crowd
disaster without any serious reasons (e.g., in Moscow, 1982). For example, due to
the “faster-is-slower effect” panic can be triggered by small pedestrian counterflows
[72], which cause delays to the crowd intending to leave. Consequently, stopped
pedestrians in the back, who do not see the reason for the temporary slowdown,
are getting impatient and pushy. In accordance with observations [7, 25], one may
model this by increasing the desired velocity, for example, by the formula
Fig. 3.6 Top: Long-term photograph showing stop-and-go waves in a densely packed street. While
stopped people appear relatively sharp, people moving from right to left have a fuzzy appearance.
Note that gaps propagate from left to right. Middle: Empirically observed stop-and-go waves in
front of the entrance to the Jamarat Bridge on January 12, 2006 (after [51]), where pilgrims moved
from left to right. Dark areas correspond to phases of motion, light grey to stop phases. The
“location” coordinate represents the distance to the beginning of the narrowing, i.e. to the cross
section of reduced width. Bottom left: Illustration of the “shell model” (see [94]), in particular
of situations where several pedestrians compete for the same gap, which causes coordination
problems. Bottom right: Simulation results of the shell model. The observed stop-and-go waves
result from the alternation of forward pedestrian motion and backward gap propagation
v0˛ .t/ D Œ1 n˛ .t/v0˛ .0/ C n˛ .t/vmax

˛ : (3.16)
Herein, vmax
˛ is the maximum desired velocity and v0˛ .0/ the initial one, correspond-
ing to the expected velocity of leaving. The time-dependent parameter
v˛ .t/
n˛ .t/ D 1 (3.17)
v0˛ .0/
reflects the nervousness, where v˛ .t/ denotes the average speed into the desired
direction of motion. Altogether, long waiting times increase the desired speed v0˛ or
driving force v0˛ .t/e˛0 =, which can produce high densities and inefficient motion.
This further increases the waiting times, and so on, so that this tragic feedback
can eventually trigger so high pressures that people are crushed or falling and
trampled. It is, therefore, imperative, to have sufficiently wide exits and to prevent
counterflows, when big crowds want to leave [91].
3.4.4.4 Transition to Stop-and-Go Waves
Recent empirical studies of pilgrim flows in the area of Makkah, Saudi Arabia,
have shown that intermittent flows occur not only when bottlenecks are obvious.
On January 12, 2006, pronounced stop-and-go waves have been even observed
upstream of the 44 m wide entrance to the Jamarat Bridge [51]. While the pilgrim
flows were smooth and continuous (“laminar”) over many hours, at 11:53 am stop-
and-go waves suddenly appeared and propagated over distances of more than 30 m
(see Fig. 3.6). The sudden transition was related to a significant drop of the flow, i.e.
with the onset of congestion [51]. Once the stop-and-go waves set in, they persisted
over more than 20 min.
This phenomenon can be reproduced by a recent model based on two continuity
equations, one for forward pedestrian motion and another one for backward gap
propagation [94]. The model was derived from a “shell model” (see Fig. 3.6) and
describes very well the observed alternation between backward gap propagation
and forward pedestrian motion.
3.4.4.5 Transition to “Crowd Turbulence”
On the same day, around 12:19, the density reached even higher values and the
video recordings showed a sudden transition from stop-and-go waves to irregular
flows (see Fig. 3.7). These irregular flows were characterized by random, unintended
displacements into all possible directions, which pushed people around. With a
certain likelihood, this caused them to stumble. As the people behind were moved
by the crowd as well and could not stop, fallen individuals were trampled, if they did
not get back on their feet quickly enough. Tragically, the area of trampled people
grew more and more in the course of time, as the fallen pilgrims became obstacles
Fig. 3.7 Pedestrian dynamics at different densities. Left: Representative trajectories (space-time
plots) of pedestrians during the laminar, stop-and-go, and turbulent flow regime. Each trajectory
extends over a range of 8 m, while the time required for this stretch is normalized to 1. To
indicate the different speeds, symbols are included in the curves every 5 s. While the laminar
flow (top line) is fast and smooth, motion is temporarily interrupted in stop-and-go flow (medium
line), and backward motion can occur in “turbulent” flows (bottom line). Right: Example of the
temporal evolution of the velocity components vx .t / into the average direction of motion and vy .t /
perpendicular to it in “turbulent flow”, which occurs when the crowd density is extreme. One can
clearly see the irregular motion into all possible directions characterizing “crowd turbulence”. For
details see [51]
for others [51]. The result was one of the biggest crowd disasters in the history of
pilgrimage.
How can we understand this transition to irregular crowd motion? A closer look
at video recordings of the crowd reveals that, at this time, people were so densely
packed that they were moved involuntarily by the crowd. This is reflected by random
displacements into all possible directions. To distinguish these irregular flows from
laminar and stop-and-go flows and due to their visual appearance, we will refer to
them as “crowd turbulence”.
As in certain kinds of fluid flows, “turbulence” in crowds results from a
sequence of instabilities in the flow pattern. Additionally, one finds a sharply peaked
probability density function of velocity increments
Vx D Vx .r; t C / Vx .r; t/; (3.18)
which is typical for turbulence [95], if the time shift is small enough [51]. One also
observes a power-law scaling of the displacements indicating self-similar behaviour
[51]. As large eddies are not detected, however, the similarity with fluid turbulence is
limited, but there is still an analogy to turbulence at currency exchange markets [95].
Instead of vortex cascades like in turbulent fluids, one rather finds a hierarchical
fragmentation dynamics: At extreme densities, individual motion is replaced by
mass motion, but there is a stick-slip instability which leads to “rupture” when
the stress in the crowd becomes too large. That is, the mass splits up into clusters
of different sizes with strong velocity correlations inside and distance-dependent
correlations between the clusters.
“Crowd turbulence” has further specific features [51]. Due to the physical
contacts among people in extremely dense crowds, we expect commonalities with
granular media. In fact, dense driven granular media may form density waves,
while moving forward [96], and can display turbulent-like states [97,98]. Moreover,
under quasi-static conditions [97], force chains [99] are building up, causing strong
variations in the strengths and directions of local forces. As in earthquakes [100,101]
this can lead to events of sudden, uncontrollable stress release with power-law
distributed displacements. Such a power-law has also been discovered by video-
based crowd analysis [51].
3.4.5 Some Warning Signs of Critical Crowd Conditions
Turbulent waves are experienced in dozens of crowd-intensive events each year all
over the world [102]. Therefore, it is necessary to understand why, where and when
potentially critical situations occur. Viewing real-time video recordings is not very
suited to identify critical crowd conditions: While the average density rarely exceeds
values of six persons per square meter, the local densities can reach almost twice as
large values [51]. It has been found, however, that even evaluating the local densities
is not enough to identify the critical times and locations precisely, which also applies
to an analysis of the velocity field [51]. The decisive quantity is rather the “crowd
pressure”, i.e. the density, multiplied with the variance of speeds. It allows one to
identify critical locations and times (see Fig. 3.8).
There are even advance warning signs of critical crowd conditions: The crowd
accident on January 12, 2006 started about 10 min after “turbulent” crowd motion
set in, i.e. after the “pressure” exceeded a value of 0.02/s2 (see Fig. 3.8). Moreover,
it occurred more than 30 min after stop-and-go waves set in, which can be easily
detected in accelerated surveillance videos. Such advance warning signs of critical
Fig. 3.8 Left: Snapshot of the on-line visualization of “crowd pressure”. Red colors (see the
lower ellipses) indicate areas of critical crowd conditions. In fact, the sad crowd disaster during
the Muslim pilgrimage on January 12, 2006, started in this area. Right: The “crowd pressure” is
a quantitative measure of the onset of “crowd turbulence”. The crowd disaster started when the
“crowd pressure” reached particularly high values. For details see [51]
crowd conditions can be evaluated on-line by an automated video analysis system.

In many cases, this can help one to gain time for corrective measures like flow
control, pressure-relief strategies, or the separation of crowds into blocks to stop the
propagation of shockwaves [51]. Such anticipative crowd control could increase the
level of safety during future mass events.
3.4.6 Evolutionary Optimization of Pedestrian Facilities
Having understood some of the main factors causing crowd disasters, it is interesting
to ask how pedestrian facilities can be designed in a way that maximizes the
efficiency of pedestrian flows and the level of safety. One of the major goals during
mass events must be to avoid extreme densities. These often result from the onset
of congestion at bottlenecks, which is a consequence of the breakdown of free flow
and causes an increasing degree of compression. When a certain critical density is
increased (which depends on the size distribution of people), this potentially implies
high pressures in the crowd, particularly if people are impatient due to long delays
or panic.
The danger of an onset of congestion can be minimized by avoiding bottlenecks.
Notice, however, that jamming can also occur at widenings of escape routes [91].
This surprising fact results from disturbances due to pedestrians, who try to overtake
each other and expand in the wider area because of their repulsive interactions.
These squeeze into the main stream again at the end of the widening, which acts
like a bottleneck and leads to jamming. The corresponding drop of efficiency E is
more pronounced:
1. If the corridor is narrow.
2. If the pedestrians have different or high desired velocities.
3. If the pedestrian density in the corridor is high.
Obviously, the emerging pedestrian flows decisively depend on the geometry of
the boundaries. They can be simulated on a computer already in the planning phase
of pedestrian facilities. Their configuration and shape can be systematically varied,
e.g. by means of evolutionary algorithms [28, 103] and evaluated on the basis of
particular mathematical performance measures [7]. Apart from the efficiency
1 X v˛ e˛0
ED (3.19)
N ˛ v0˛
we can, for example, define the measure of comfort C D .1 D/ via the discomfort
!
1 X .v˛ v˛ /2 1 X v˛ 2
DD D 1 : (3.20)
N ˛ .v˛ /2 N ˛ .v˛ /2
The latter is again between 0 and 1 and reflects the frequency and degree of
sudden velocity changes, i.e. the level of discontinuity of walking due to necessary
avoidance maneuvers. Hence, the optimal configuration regarding the pedestrian
requirements is the one with the highest values of efficiency and comfort.
During the optimization procedure, some or all of the following can be varied:
1. The location and form of planned buildings.
2. The arrangement of walkways, entrances, exits, staircases, elevators, escalators,
and corridors.
3. The shape of rooms, corridors, entrances, and exits.
4. The function and time schedule (Recreation rooms or restaurants are often
continuously frequented, rooms for conferences or special events are mainly
visited and left at peak periods, exhibition rooms or rooms for festivities require
additional space for people standing around, and some areas are claimed by
queues or through traffic.)
In contrast to early evolutionary optimization methods, recent approaches allow to
change not only the dimensions of the different elements of pedestrian facilities,
but also to vary their topology. The procedure of such algorithms is illustrated in
Fig. 3.9. Highly performing designs are illustrated in Fig. 3.10. It turns out that,
for an emergency evacuation route, it is favorable if the crowd does not move
completely straight towards a bottleneck. For example, a zigzag design of the
evacuation route can reduce the pressure on the crowd upstream of a bottleneck
(see Fig. 3.11). The proposed evolutionary optimization procedure can, of course,
not only be applied to the design of new pedestrian facilities, but also to a reduction
of existing bottlenecks, when suitable modifications are implemented.
Fig. 3.9 The evolutionary optimization based on Boolean grids uses a two-stage algorithm
(see [104] for details). Left: In the “randomization stage”, obstacles are distributed over the
grid with some randomness, thereby allowing for the generation and testing of new topologies
(architectures). Right: In the “agglomeration stage”, small nearby obstacles are clustered to form
larger objects with smooth boundaries. After several iterations, the best performing designs are
reasonably shaped. See Fig. 3.10 for examples of possible bottleneck designs
3.5 Future Directions 95
Fig. 3.10 Two examples of improved designs for cases with a bottleneck along the escape route
of a large crowd, obtained with an evolutionary algorithm based on Boolean grids (after [104]).
People were assumed to move from left to right only. Left: Funnel-shaped escape route. Right:
Zig-zag design
Fig. 3.11 Left: Conventional design of a stadium exit in an emergency scenario, where we assume
that some pedestrians have fallen at the end of the downwards staircase to the left. The dark areas
indicate high pressures, since pedestrians are impatient and pushing from behind. Right: In the
improved design, the increasing diameter of corridors can reduce waiting times and impatience
(even with the same number of seats), thereby accelerating evacuation. Moreover, the zigzag design
of the downwards staircases changes the pushing direction in the crowd. Computer simulations
indicate that the zig-zag design can reduce the average pressure in the crowd at the location of the
incident by a factor of two (After [8])
3.5 Future Directions
In this contribution, we have presented a multi-agent approach to pedestrian and

crowd dynamics. Despite the great effort required, pedestrian interactions can be
well quantified by video tracking. Compared to other social interactions they turn
out to be quite simple. Nevertheless, they cause a surprisingly large variety of
self-organized patterns and short-lived social phenomena, where coordination or
cooperation emerges spontaneously. For this reason, they are interesting to study,
particularly as one can expect new insights into coordination mechanisms of social
beings beyond the scope of classical game theory. Examples for observed self-
organization phenomena in normal situations are lane formation, stripe formation,
oscillations and intermittent clogging effects at bottlenecks, and the evolution of

behavioral conventions (such as the preference of the right-hand side in continental
Europe). Under extreme conditions (high densities or panic), however, coordination
may break down, giving rise to “freezing-by-heating” or “faster-is-slower effects”,
stop-and-go waves or “crowd turbulence”.
Similar observations as in pedestrian crowds are made in other social systems
and settings. Therefore, we expect that realistic models of pedestrian dynamics will
also promote the understanding of opinion formation and other kinds of collective
behaviors. The hope is that, based on the discovered elementary mechanisms of
emergence and self-organization, one can eventually also obtain a better under-
standing of the constituting principles of more complex social systems. At least the
same underlying factors are found in many social systems: non-linear interactions
of individuals, time-dependence, heterogeneity, stochasticity, competition for scarce
resources (here: space and time), decision-making, and learning. Future work will
certainly also address issues of perception, anticipation, and communication.
Acknowledgements The authors are grateful for partial financial support by the German Research
Foundation (research projects He 2789/7-1, 8-1) and by the “Cooperative Center for Communi-
cation Networks Data Analysis”, a NAP project sponsored by the Hungarian National Office of
Research and Technology under grant No. KCKHA005.
References
Primary Literature
1. B.D. Hankin, R.A. Wright, Oper. Res. Q. 9, 81–88 (1958)
2. S.J. Older, Traffic. Eng. Contr. 10, 160–163 (1968)
3. U. Weidmann, Transporttechnik der Fußgänger, (Institut für Verkehrsplanung,
Transporttechnik, Straßen- und Eisenbahnbau, ETH Zürich, 1993)
4. J.J. Fruin, Designing for pedestrians: A level-of-service concept, in Highway Research
Record, Number 355: Pedestrians (Highway Research Board, Washington, D.C., 1971),
pp. 1–15
5. J. Pauls, Fire Technol. 20, 27–47 (1984)
6. W.H. Whyte, City. Rediscovering the Center (Doubleday, New York, 1988)
7. D. Helbing, Verkehrsdynamik (Springer, Berlin, 1997)
8. D. Helbing, L. Buzna, A. Johansson, T. Werner, Transport. Sci. 39(1), 1–24 (2005)
9. W.M. Predtetschenski, A.I. Milinski, Personenströme in Gebäuden – Berechnungsmethoden
für die Projektierung – (Rudolf Müller, Köln-Braunsfeld, 1971)
10. Transportation Research Board, Highway Capacity Manual, Special Report 209 (Transporta-
tion Research Board, Washington, D.C., 1985)
11. S.J. Yuhaski Jr., J.M. Macgregor Smith, Queueing Syst. 4, 319–338 (1989)
12. D. Garbrecht, Traffic Q. 27, 89–109 (1973)
13. N. Ashford, M. O’Leary, P.D. McGinity, Traffic. Eng. Contr. 17, 207–210 (1976)
14. A. Borgers, H. Timmermans, Socio-Econ. Plann. Sci. 20, 25–31 (1986)
15. D. Helbing, Stochastische Methoden, nichtlineare Dynamik und quantitative Modelle sozialer
Prozesse, Ph.D. thesis (University of Stuttgart, 1992, published by Shaker, Aachen, 1993)
16. D. Helbing, M. Isobe, T. Nagatani, K. Takimoto, Phys. Rev. E 67, 067101 (2003)
References 97
17. W. Daamen, S.P. Hoogendoorn, in Proceedings of the 82nd Annual Meeting at the Trans-
portation Research Board (CDROM, Washington D.C., 2003)
18. M. Isobe, D. Helbing, T. Nagatani, Phys. Rev. E 69, 066132 (2004)
19. A. Seyfried, B. Steffen, W. Klingsch, M. Boltes, J. Stat. Mech. P10002 (2005)
20. T. Kretz, A. Grünebohm, M. Kaufman, F. Mazur, M. Schreckenberg, J. Stat. Mech. P10001
(2006)
21. L.F. Henderson, Transport. Res. 8, 509–515 (1974)
22. R.L. Hughes, Transport. Res. B 36, 507–535 (2002)
23. D. Helbing, Complex Syst. 6, 391–415 (1992)
24. S.P. Hoogendoorn, P.H.L. Bovy, Transport. Res. Record. 1710, 28–36 (2000)
25. D. Helbing, Behav. Sci. 36, 298–310 (1991)
26. D. Helbing, P. Molnár, Phys. Rev. E 51, 4282–4286 (1995)
27. P.G. Gipps, B. Marksjö, Math. Comp. Simul. 27, 95–105 (1985)
28. K. Bolay, Nichtlineare Phänomene in einem fluid-dynamischen Verkehrsmodell (Master’s
thesis, University of Stuttgart, 1998)
29. V.J. Blue, J.L. Adler, Transport. Res. Record. 1644, 29–36 (1998)
30. M. Fukui, Y. Ishibashi, J. Phys. Soc. Jpn. 68, 2861–2863 (1999)
31. M. Muramatsu, T. Irie, T. Nagatani, Physica A 267, 487–498 (1999)
32. H. Klüpfel, M. Meyer-König, J. Wahle, M. Schreckenberg, in Theory and Practical Issues on
Cellular Automata, ed. by S. Bandini, T. Worsch (Springer, London, 2000)
33. C. Burstedde, K. Klauck, A. Schadschneider, J. Zittartz, Physica A 295, 507–525 (2001)
34. S. Gopal, T.R. Smith, in Spatial Choices and Processes, ed. by M.M. Fischer, P. Nijkamp,
Y.Y. Papageorgiou (North-Holland, Amsterdam, 1990), pp. 169–200
35. C.W. Reynolds, in From Animals to Animats 3: Proceedings of the Third International
Conference on Simulation of Adaptive Behavior, ed. by D. Cliff, P. Husbands, J.-A. Meyer,
S. Wilson (MIT Press, Cambridge, Massachusetts, 1994), pp. 402–410
36. D. Helbing, Behav. Sci. 37, 190–214 (1992)
37. D. Helbing, P. Molnár, I. Farkas, K. Bolay, Environ. Plann. B 28, 361–383 (2001)
38. J. Klockgether, H.-P. Schwefel, in Proceedings of the Eleventh Symposium on Engineering
Aspects of Magnetohydrodynamics, ed. by D.G. Elliott (California Institute of Technology,
Pasadena, CA, 1970), pp. 141–148
39. D. Helbing, in Economic Evolution and Demographic Change. Formal Models in Social
Sciences, ed. by G. Haag, U. Mueller, K.G. Troitzsch (Springer, Berlin, 1992), pp. 330–348
40. N.E. Miller, in Personality and the behavior disorders, ed. by J.McV. Hunt, Vol. 1 (Ronald,
New York, 1944)
41. N.E. Miller, in Psychology: A Study of Science, ed. by S. Koch, Vol. 2 (McGraw Hill,
New York, 1959)
42. K. Lewin, Field Theory in Social Science (Harper & Brothers, New York, 1951)
43. D. Helbing, J. Math. Sociol. 19(3), 189–219 (1994)
44. S. Hoogendoorn, P.H.L. Bovy, Optim. Contr. Appl. Meth. 24(3), 153–172 (2003)
45. A. Johansson, D. Helbing, P.K. Shukla, Specification of the social force pedestrian model by
evolutionary adjustment to videotracking data. Advances in Complex Systems (ACS), 10(2)
271–288 (2007)
46. T.I. Lakoba, D.J. Kaup, N.M. Finkelstein, Simulation 81(5), 339–352 (2005)
47. A. Seyfried, B. Steffen, T. Lippert, Physica A 368, 232–238 (2006)
48. J. Kerridge, T. Chamberlain, in Pedestrian and Evacuation Dynamics ’05, ed. by N. Waldau,
P. Gattermann, H. Knoflacher, M. Schreckenberg (Springer, Berlin, 2005)
49. S.P. Hoogendoorn, W. Daamen, P.H.L. Bovy, in Proceedings of the 82nd Annual Meeting
at the Transportation Research Board (CDROM, Mira Digital Publishing, Washington D.C.,
2003)
50. K. Teknomo, Microscopic pedestrian flow characteristics: Development of an image process-
ing data collection and simulation model (PhD thesis, Tohoku University Japan, Sendai, 2002)
51. D. Helbing, A. Johansson, H.Z. Al-Abideen, Phys. Rev. E 75, 046109 (2007)
52. L.P. Kadanoff, J. Stat. Phys. 39, 267–283 (1985)
53. H.E. Stanley, N. Ostrowsky (eds.), On Growth and Form (Martinus Nijhoff, Boston, 1986)
54. T. Arns, Video films of pedestrian crowds (Stuttgart, 1993)
55. H.-H. Stølum, Nature 271, 1710–1713 (1996)
56. I. Rodrı́guez-Iturbe, A. Rinaldo, Fractal River Basins: Chance and Self-Organization
(Cambridge University, Cambridge, England, 1997)
57. D. Helbing, I. Farkas, T. Vicsek, Phys. Rev. Lett. 84, 1240–1243 (2000)
58. T. Schelling, J. Math. Sociol. 1, 143–186 (1971)
59. D. Helbing, T. Platkowski, Int. J. Chaos Theor. Appl. 5(4), 47–62 (2000)
60. S.P. Hoogendoorn, W. Daamen, Transpn. Sci. 39(2), 147–159 (2005)
61. T. Kretz, A. Grünebohm, M. Schreckenberg, J. Stat. Mech. P10014 (2006)
62. K. Ando, H. Oto, T. Aoki, Railway Res. Rev. 45(8), 8–13 (1988)
63. K.H. Drager, G. Løvås, J. Wiklund, H. Soma, D. Duong, A. Violas, V. Lanèrès, in the
Proceedings of the 1992 Emergency Management and Engineering Conference (Society for
Computer Simulation, Orlando, Florida, 1992), pp. 101–108
64. M. Ebihara, A. Ohtsuki, H. Iwaki, Microcomput. Civ. Eng. 7, 63–71 (1992)
65. N. Ketchell, S. Cole, D.M. Webber, C.A. Marriott, P.J. Stephens, I.R. Brearley, J. Fraser,
J. Doheny, J. Smart, in Engineering for Crowd Safety, ed. by R.A. Smith, J.F. Dickie (Elsevier,
Amsterdam, 1993), pp. 361–370
66. S. Okazaki, S. Matsushita, in Engineering for Crowd Safety, ed. by R.A. Smith, J.F. Dickie
(Elsevier, Amsterdam, 1993), pp. 271–280
67. G.K. Still, New computer system can predict human behaviour response to building fires. Fire
84, 40–41 (1993)
68. G.K. Still, Crowd Dynamics (Ph.D. thesis, University of Warwick, 2000)
69. P.A. Thompson, E.W. Marchant, Modelling techniques for evacuation, in Engineering for
Crowd Safety, ed. by R.A. Smith, J.F. Dickie (Elsevier, Amsterdam, 1993), pp. 259–269
70. G.G. Løvås, On the importance of building evacuation system components, IEEE Trans. Eng.
Manag. 45, 181–191 (1998)
71. H.W. Hamacher, S.A. Tjandra, in Pedestrian and Evacuation Dynamics, ed. by M. Schreck-
enberg, S.D. Sharma (Springer, Berlin, 2001), pp. 227–266
72. D. Elliott, D. Smith, Football stadia disasters in the United Kingdom: Learning from
tragedy?, Industrial & Environmental Crisis Quarterly 7(3), 205–229 (1993)
73. B.D. Jacobs, P. ’t Hart, in Hazard Management and Emergency Planning, Chap. 10, ed. by
D.J. Parker, J.W. Handmer (James & James Science, London, 1992)
74. D. Canter (ed.), Fires and Human Behaviour (David Fulton, London, 1990)
75. A. Mintz, J. Abnorm. Norm. Soc. Psychol. 46, 150–159 (1951)
76. J.P. Keating, Fire J., 57–61+147 (May/1982)
77. D.L. Miller, Introduction to Collective Behavior, Fig. 3.3 and Chap. 9 (Wadsworth, Belmont,
CA, 1985)
78. J.S. Coleman, Foundations of Social Theory, Chaps. 9 and 33 (Belkamp, Cambridge, MA,
1990)
79. N.R. Johnson, Panic at “The Who Concert Stampede”: An empirical assessment, Soc. Prob.
34(4), 362–373 (1987)
80. G. LeBon, The Crowd (Viking, New York, 1960 [1895])
81. E. Quarantelli, Sociol. Soc. Res. 41, 187–194 (1957)
82. N.J. Smelser, Theory of Collective Behavior, (The Free Press, New York, 1963)
83. R. Brown, Social Psychology (The Free Press, New York, 1965)
84. R.H. Turner, L.M. Killian, Collective Behavior, 3rd edn. (Prentice Hall, Englewood Cliffs,
NJ, 1987)
85. J.L. Bryan, Fire J., 27–30+86–90 (Nov./1985)
86. R. Axelrod, W.D. Hamilton, Science 211, 1390–1396 (1981)
87. R. Axelrod, D. Dion, Science 242, 1385–1390 (1988)
88. N.S. Glance, B.A. Huberman, Sci. Am. 270, 76–81 (1994)
89. R.A. Smith, J.F. Dickie (eds.), Engineering for Crowd Safety (Elsevier, Amsterdam, 1993)
90. H.H. Kelley, J.C. Condry Jr., A.E. Dahlke, A.H. Hill, J. Exp. Soc. Psychol. 1, 20–54 (1965)
References 99
91. D. Helbing, I. Farkas, T. Vicsek, Nature 407, 487–490 (2000)

92. G.H. Ristow, H.J. Herrmann, Phys. Rev. E 50, R5–R8 (1994)
93. D.E. Wolf, P. Grassberger (eds.), Friction, Arching, Contact Dynamics (World Scientific,
Singapore, 1997)
94. D. Helbing, A. Johansson, J. Mathiesen, M.H. Jensen, A. Hansen Phys. Rev. Lett. 97, 168001
(2006)
95. S. Ghashghaie, W. Breymann, J. Peinke, P. Talkner, Y. Dodge, Nature 381, 767–770 (1996)
96. G. Peng, H.J. Herrmann, Phys. Rev. E 49, R1796–R1799 (1994)
97. F. Radjai, S. Roux, Phys. Rev. Lett. 89, 064302 (2002)
98. K.R. Sreenivasan, Nature 344, 192–193 (1990)
99. M.E. Cates, J.P. Wittmer, J.-P. Bouchaud, P. Claudin, Phys. Rev. Lett. 81, 1841–1844 (1998)
100. P. Bak, K. Christensen, L. Danon, T. Scanlon, Phys. Rev. Lett. 88, 178501 (2002)
101. P.A. Johnson, X. Jia, Nature 437, 871–874 (2005)
102. J.J. Fruin, in Engineering for Crowd Safety, ed. by R.A. Smith, J.F. Dickie (Elsevier,
Amsterdam, 1993), pp. 99–108
103. T. Baeck, Evolutionary Algorithms in Theory and Practice (Oxford University Press,
New York, 1996)
104. A. Johansson, D. Helbing, in Pedestrian and Evacuation Dynamics 2005, ed. by N. Waldau,
P. Gattermann, H. Knoflacher, M. Schreckenberg (Springer-Verlag, Berlin, 2007), pp. 267–
272
Chapter 4
Opinion Formation
4.1 Introduction
Many biological systems exhibit collective patterns, which emerge through simple
interactions of large numbers of individuals. A typical example is agglomeration
phenomena. Such clustering dynamics have been found in systems as different as
bacterial colonies [1], gregarious animals like cockroaches [2], fish schools [3],
flocks of birds [4], and animal groups [5]. Similar phenomena are observed in
ecosystems [6] and human populations, as examples ranging from the formation of
pedestrian groups [7] to the formation of urban agglomerations demonstrate [8, 9].
Recently, numerous studies on the structure of human interaction networks
[10–12] demonstrated that clustering is not restricted to physical or geographical
space. For instance, clustering has been extensively studied in networks of email
communication [13], phone calls [12], scientific collaboration [14] and sexual
contacts [15]. It is much less understood, however, how and what conditions
clustering patterns emerge in behavioral or opinion space. Empirical studies suggest
that opinions differ globally [16, 17], while they cluster locally within geographical
regions [18], socio-demographic groups [19], or Internet communities [20]. In
addition, research on dynamics in work teams demonstrates that even groups of
very small size often show high opinion diversity and can even suffer from opinion
polarization [21, 22].
Opinion clustering is defined as the co-existence of distinct subgroups (clusters)
of individuals with similar opinions, while opinions in different subgroups are
relatively large. The gaps in our theoretical understanding of opinion clustering
are pressing since both local consensus and global diversity are precarious. On the
one hand, cultural diversity may get lost in a world where people are increasingly
exposed to influences from mass media, Internet communication, interregional

This chapter reprints a previously published paper and should be cited as follows: M. Mäs,
A. Flache, and D. Helbing (2010) Individualization as driving force of clustering phenomena in
humans. PLoS Comput. Biol. 6(10), e1000959.

102 4 Opinion Formation
migration, and mass tourism, which may promote a universal monoculture [23, 24],
as the extinction of languages suggests [25]. On the other hand, increasing indi-
vidualization threatens to disintegrate the social structures in which individuals are
embedded, with the possible consequence of the loss of societal consensus [26, 27].
This is illustrated by the recent debate on the decline of social capital binding
individuals into local communities [28].
Early formal models of social influence imply that monoculture is unavoidable,
unless a subset of the population is perfectly cut off from outside influences
[29]. Social isolation, however, appears questionable as explanation of pluralism.
In modern societies, distances in social networks are quite short on the whole,
and only relatively few random links are required to dramatically reduce network
distance [10].
Aiming to explain pluralism, researchers have incorporated the empirically well-
supported observation of “homophily”, i.e. the tendency of “birds of a feather to
flock together” [30, 31], into formal models of social influence [32]. These models
typically assume “bounded confidence” (BC) in the sense that only those individuals
interact, whose opinions do not differ more than a given threshold level [33, 34]. As
Fig. 4.1a illustrates, BC generates opinion clustering, a result that generalizes to
model variants with categorical rather than continuous opinions [32, 35]. However,
clustering in the BC-model is sensitive to “interaction noise”: A small random
chance that agents may interact even when their opinions are not similar, causes
monoculture again (see Fig. 4.1b).
To avoid this convergence of opinions, it was suggested that individuals would
separate themselves from negatively evaluated others [19, 36, 37]. However, recent
empirical results do not support such “negative influence” [38]. Scientists also tried
to avoid convergence by “opinion noise”, i.e. random influences, which lead to
arbitrary opinion changes with a small probability. Assuming uniformly distributed
opinion noise [39] leads to sudden, large, and unmotivated opinion changes of
individuals, while theories of social integration [26,27,40,41] and empirical studies
of individualization [42, 43] show a tendency of incremental opinion changes
rather than arbitrary opinion jumps. Incremental opinion changes, however, tend
to promote monoculture, even in models with categorical rather than continuous
opinions [44]. Figure 4.1 demonstrates that adding a “white noise” term (N.0; /)
to an agent’s current opinion in the BC model fails to explain opinion clustering.
Weak opinion noise ( D 5) triggers convergence cascades that inevitably end in
monoculture. Stronger noise restores opinion diversity, but not clustering. Instead,
diversity is based on frequent individual deviations from a predominant opinion
cluster (for D 18). However, additional clusters can not form and persist, because
opinion noise needs to be strong to separate enough agents from the majority
cluster – so strong that randomly emerging smaller clusters cannot stabilize.
In conclusion, the formation of persistent opinion clusters is such a difficult
puzzle that all attempts to explain them had to make assumptions that are difficult to
justify by empirical evidence. The solution proposed in the following, in contrast,
aims to reconcile model assumptions with sociological and psychological research.
The key innovation is to integrate another decisive feature into the model, namely
4.1 Introduction 103
Fig. 4.1 Opinion dynamics produced by the bounded confidence (BC) model [33] with and
without noise. Populations consist of 100 agents. Opinions vary between 250 and 250. Initial
opinions are uniformly distributed. For visualization, the opinion scale is divided into 50 bins of
equal size. Color coding indicates the relative frequency of agents in each bin. (a) Dynamics of the
BC-model without noise [33] over ten iterations. At each simulation event, one agent’s opinion
is replaced by the average opinion of those other agents who hold opinions oj .t / within the
focal agent’s confidence interval (oi .t / oj .t / oi .t / C ). For D 0:05, one finds
several homogeneous clusters, which stabilize when the distance between all clusters exceeds
the confidence threshold . (b) Computer simulation of the same BC-model, but considering
interaction noise. Agents that would otherwise not have been influential, now influence the focal
agent’s opinion with a probability of p D 0:01. This small noise is sufficient to eventually generate
monoculture. (c) Simulation of the BC-model with opinion noise. After each opinion update, a
random value drawn from a normal distribution with an average of zero and a standard deviation
of (abbreviated by N.0; /) is added to the opinion. For weak opinion noise ( D 5), one
cluster is formed, which carries out a random walk on the opinion scale. When the opinion noise
is significantly increased ( D 18), there is still one big cluster, but many separated agents exist
as well (cf. Fig. 4.4). With even stronger opinion noise ( D 20), the opinion distribution becomes
completely random
the “striving for uniqueness” [42, 43]. While individuals are influenced by their
social environment, they also show a desire to increase their uniqueness when
too many other members of society hold similar opinions. We incorporate this
assumption as a white noise term in the model. However, in contrast to existing
models we assume that noise strength is not constant but adaptive. To be precise,
we assume that the impact of noise on the opinion of an individual is the stronger
the less unique the individual’s opinion is compared to the other members of the
population. Consumer behavior regarding fashions illustrates the adaptiveness of
opinion noise: When new clothing styles are adopted by some people, they often
tend to be imitated by others with similar spirit and taste (the “peer group”).
However, when imitation turns the new style into a norm, people will seek to
increase their uniqueness. This will sooner or later lead some individuals to invent
new ways to dress differently from the new norm.
Adaptive noise creates a dynamic interplay of the integrating and disintegrat-

ing forces highlighted by Durkheim’s classic theory of social integration [26].
Durkheim argued that integrating forces bind individuals to society, motivating
them to conform and adopt values and norms that are similar to those of others.
But he also saw societal integration as being threatened by disintegrating forces
that foster individualization and drive actors to differentiate from one another
[27, 40, 41]. The “Durkheimian opinion dynamics model” proposed in the fol-
lowing can explain pluralistic clustering for the case of continuously varying
opinions, although it incorporates all the features that have previously been found
to undermine clustering: (1) a fully connected influence network, (2) absence of
bounded confidence, (3) no negative influence, and (4) white opinion noise. From
a methodological viewpoint, our model builds on concepts from statistical physics,
namely the phenomenon of “nucleation” [45], illustrated by the formation of water
droplets in supersaturated vapor. However, by assuming adaptive noise, we move
beyond conventional nucleation models. The model also resembles elements of
Interacting Particle Systems [46] like the voter model and the anti-voter model
[47–50] which have been used to study dynamics of discrete opinions (“pro” and
“contra”). However, we focus here on continuous opinions like the degree to which
individuals are in favor of or against a political party.
Computational simulation experiments reveal that, despite the continuity of
opinions in our model, it generates pluralism as an intermediate phase between
monoculture and individualism. When the integrating forces are too strong, the
model dynamics inevitably implies monoculture, even when the individual opinions
are initially distributed at random. When the disintegrating forces prevail, the result
is what Durkheim called “anomie”, a state of extreme individualism without a social
structure, even if there is perfect consensus in the beginning. Interestingly, there is
no sharp transition between these two phases, when the relative strength of both
forces is changed. Instead, we observe an additional, intermediate regime, where
opinion clustering occurs, which is independent of the initial condition. In this
regime, adaptive noise entails robust pluralism that is stabilized by the adaptiveness
of cluster size. When clusters are small, individualization tendencies are too weak to
prohibit a fusion of clusters. However, when clusters grow large, individualization
increases in strength, which triggers a splitting into smaller clusters (“fission”). In
this way, our model solves the cluster formation problem of earlier models. While
in BC models, white noise causes either monoculture or fragmentation (Fig. 4.1c),
in the Durkheimian opinion dynamics model proposed here, it enables clustering.
Therefore, rather than endangering cluster formation, noise supports it. In the
following, we describe the model and identify conditions under which pluralism
can flourish.
4.2 Model
The model has been elaborated as an agent-based model [51] addressing the
opinion dynamics of interacting individuals. The simulated population consists of N
agents i , representing individuals, each characterized by an opinion oi .t/ at time t.
4.2 Model 105
The numerical value for the opinion varies between a given minimum and maximum
value on a metric scale. We use the term “opinion” here, for consistency with the
literature on social influence models. However, oi .t/ may also reflect behaviors,
beliefs, norms, customs or any other cardinal cultural attribute that individuals
consider relevant and that is changed by social influence. The dynamics is modeled
as a sequence of events. Every time t 0 D k=N (with k 2 f1; : : : ; N g), the computer
randomly picks an agent i and changes the opinion oi .t/ by the amount
X
N

oj .t/ oi .t/ wij .t/
j D1
j ¤i
oi D C i .t/: (4.1)
X
N
wij .t/
j D1
j ¤i
The first term on the rhs of (4.1) models the integrating forces of Durkheim’s theory.
Technically, agents tend to adopt the weighted average of the opinions oj .t/ of all
other members j of the population. Implementing homophily, the social influence
wij .t/ that agent j has on agent i is the stronger, the smaller their opinion distance
dij .t/ D joj .t/ oi .t/j is. Formally, we assume
wij .t/ D edij .t /=A D ejoj .t /oi .t /j=A : (4.2)
The parameter A represents the range of social influence of agents. For small
positive values of A, agents are very confident in their current opinion and are
mainly influenced by individuals who hold very similar opinions, while markedly
distinct opinions have little impact. The higher A is, however, the more are agents
influenced by individuals with considerably different opinions and the stronger are
the integrating forces in our Durkheimian theory.
The disintegrating forces on the opinion of agent i are modeled by a noise term
i .t/. Specifically, the computer adds a normally distributed random value i .t/
(“white noise”) to the first term on the rhs of (4.1). While we assume that the
mean value of the random variable i .t/ is zero, the standard deviation has been
specified as
X
N
i .t/ D s edij .t / : (4.3)
j D1
The larger the standard deviation, the stronger are the individualization ten-
dencies of an agent. Following Durkheim’s theory, (4.3) implements noise in an
adaptive way: Accordingly, an agent’s striving for individualization is weak, if there
are only a few others with similar opinions. Under such conditions, there is no need
to increase distinctiveness. However, if many others hold a similar opinion, then
individuals are more motivated to differ from others.
By including the focal agent i in the sum of (4.3), we assume that there is
always some degree of opinion noise, even when agent i holds a perfectly unique
opinion. These fluctuations may have a variety of reasons, such as misjudgments,
trial-and-error behavior, or the influence of exogenous factors on the individual
opinion. Furthermore, this assumption reflects Durkheim’s notion that the seeking
for uniqueness is a fundamental feature of human personality, which can not be
suppressed completely [26, 52].
We use the parameter s of (4.3) to vary the strength of the disintegrating forces
in society. The higher the value of s, the higher is the standard deviation of the
distribution, from which i .t/ is drawn, and the stronger are the disintegrating
forces. Finally, to keep the opinions of the agents within the bounds of the opinion
scale, we set the value of i .t/ to zero, if the bounds of the opinion space would be
left otherwise.
4.3 Results
We have studied the Durkheimian opinion dynamics model with extensive computer
simulations, focusing on relatively small populations (N D 100), because in this
case it is reasonable to assume that all members may interact with each other.
For bigger populations one would have to take into account the topology of the
social interaction network as well. Such networks would most likely consist of
segregated components (“communities”), which are not or only loosely connected
with each other [12–15]. Existing social influence models can explain how under
such conditions each community develops its own shared opinion (see Fig. 4.1a).
However, according to these models opinion clustering is only stable when there
is no interaction between communities [29, 33], an assumption that appears not
to be empirically correct in an increasingly connected world. Therefore, we focus
on a setting for which the lack of connectedness is guaranteed to be excluded as
explanation of clustering and study model dynamics in relatively small and complete
interaction networks.
To illustrate the model dynamics, Fig. 4.2 shows three typical simulation runs
for different strengths s of disintegrating forces, while the strength A D 2 of the
integrating force is kept constant. In each run, all agents start with an opinion in the
middle of the opinion scale (oi .0/ D 0), i.e. conformity. This is an initial condition
for which the classical BC-model does not produce diversity. Figure 4.2a shows
typical opinion trajectories for a population in which the integrating forces are
much stronger than the disintegrating forces. Consequently, the population develops
collective consensus, i.e. the variation of opinions remains small, even though not
all agents hold exactly the same opinion. Triggered by the random influences i .t/,
the average opinion performs a characteristic random walk.
When the disintegrating force prevails, the pattern is strikingly different.
Figure 4.2b shows that for large noise strengths s, the initial consensus breaks
up quickly, and the agents’ opinions are soon scattered across the entire opinion
4.3 Results 107
Fig. 4.2 Opinion trajectories of three representative simulation runs with 100 agents generated
by the Durkheimian model. In all three runs, the opinions are restricted to values between 250
and 250, and all agents hold the same opinion initially (oi .0/ D 0 for all i ). In all runs, we assume
the same social influence range A D 2, but vary the strength s of the disintegrating force. (a)
Monoculture, resulting in the case of a weak disintegrating force (s D 0:4). Agents do not hold
perfectly identical opinions, but the variance is low. We studied dynamics over 10.000 iterations.
(b) Anomie (i.e. extreme individualism), generated by a very strong disintegrating force (s D 6).
Agents spread over the complete opinion scale. The black line represents the time-dependent
opinion of a single, randomly picked agent, showing significant opinion changes over time, which
is in contrast to the collective opinion formation dynamics found in the monocultural and pluralistic
cases (a) and (b). (c) For a moderate disintegrating force (s D 1:2), the population quickly
disintegrates into clusters. As long as these clusters are small, they are metastable. However,
clusters perform random walks and can merge (e.g. around iteration 5,500). As the disintegrating
force grows with the size of a cluster, big clusters eventually split up into subclusters (e.g. around
iteration 7,000). The additional graph, in which each agent’s opinion trajectory is represented by a
solid black line, is an alternative visualization of the simulation run with s D 1:2. It shows that the
composition of clusters persists over long time periods
space. Simulation scenarios A and B are characteristic for what Durkheim referred
to as states of social cohesion and of anomie. Interestingly, however, pluralism
arises as a third state in which several opinion clusters form and coexist. Figure 4.2c
shows a typical simulation run, where the adaptive noise maintains pluralism despite
the antagonistic impacts of integrating and disintegrating forces – in fact because
of this. In the related region of the parameter space, disintegrating forces prevent
global consensus, but the integrating forces are strong enough to also prevent the
population from extreme individualization. This is in pronounced contrast to what
we found for the BC-model with strong noise (Fig. 4.1c). Instead, we obtain a
number of coexisting, metastable clusters of a characteristic, parameter-dependent
size. Each cluster consists of a relatively small number of agents, which keeps the
disintegrating forces in the cluster weak and allows clusters to persist. (Remember
that the tendency of individualization according to (4.3) increases, when many
individuals hold similar opinions.) However, due to opinion drift, distinct clusters
may eventually merge. When this happens, the emergent cluster becomes unstable
and will eventually split up into smaller clusters, because disintegrating forces
increase in strength as a cluster grows.
Fig. 4.3 Conditions of clustering, monoculture and anomie. The figure shows the dependence of
the average number of clusters in the Durkheimian model on the strength s of the disintegrating
force and the range A of social influence. To generate it, we conducted computer simulations with
N D 100 agents, starting with initial consensus (oi .0/ D 0 for all i ). We restricted opinions
to values between 250 and 250. We varied the strength s of the disintegrating force between
s D 0:4 and s D 8 in steps of 0.4. A varied between A D 0:2 and A D 4 in steps of 0.2.
For each parameter combination, we conducted 100 independent replications and assessed the
average number of clusters formed after 250,000 iterations (see z-axis and the color scale). The
two transparent (gray) surfaces depict the inter-quartile range, which indicates a small variance
in the number of clusters (and also typical cluster sizes) for each parameter combination. The
horizontal grids indicate the borders of the three phases, as defined by us. An average cluster
size below 1.5 indicates monoculture. Values between 1.5 and 31 reflect clustering. Finally, values
above 31 correspond to opinion distributions that cannot be distinguished from random ones and
represent a state of anomie
Strikingly, the state of diversity, in which several opinion clusters can coexist,
is not restricted to a narrow set of conditions under which integrating and disinte-
grating forces are balanced exactly. Figure 4.3 demonstrates that opinion clusters
exist in a significant area of the parameter space, i.e. the clustering state establishes
another phase, which is to be distinguished from monoculture and from anomie.
To generate Fig. 4.3, we conducted a simulation experiment in which we varied
the influence range A and the strength s of the disintegrating force. For each
parameter combination, we ran 100 replications and measured the average number
of clusters that were present after 250,000 iterations. To count the number of clusters
in a population, we ordered the N agents according to their opinion. A cluster was
defined as a set of agents in adjacent positions such that each set member was
separated from the adjacent set members by a maximum of 5 scale points (D opinion
4.4 Discussion 109
range/N ). Figure 4.3 shows that, for large social influence ranges A and small noise
strengths s, the average number of clusters is below 1.5, reflecting monoculture
in the population. In the other extreme, i.e. for a small influence range A and
large noise strengths s, the resulting distribution contains more than 31 clusters,
a number of clusters that cannot be distinguished from purely random distributions.
Following Durkheim, we have classified such cases as anomie, i.e. as the state of
extreme individualism. Between these two phases, there are numerous parameter
combinations, for which the number of clusters is higher than 1.5 and clearly smaller
than in the anomie phase. This constitutes the clustering phase. Figure 4.3 also
shows that, for each parameter combination, there is a small variance in the number
of clusters, which is due to a statistical equilibrium of occasional fusion and fission
processes of opinion clusters (see Fig. 4.2c).
The same results were found, when starting the computer simulations with a
uniform opinion distribution. This demonstrates that the simulations were run long
enough (250,000 iterations) to obtain reliable results. It also suggests that clustering
is an attractor in the sense that the model generates clustering independent of
the initial distribution of opinions. In addition, we performed additional statistical
tests with the simulation outcomes to make sure that the existence of clusters in
our model indeed indicates pluralism and not fragmentation, a state in which a
population consists of one big cluster and a number of isolated agents (see Fig. 4.4).
To illustrate, Fig. 4.4a plots the size of the biggest cluster in the population versus
the number of clusters (see the blue areas). For comparison, the yellow area depicts
the corresponding distribution for randomly fragmented opinion distributions. The
figure shows that the distributions hardly overlap and that the Durkheimian model
generates clustering rather than fragmentation. In clear contrast, Fig. 4.4b reveals
that the opinion distributions generated by the noisy BC-model are fragmented and
not clustered.
4.4 Discussion
The phenomenon of self-organized clustering phenomena in biological and social

systems is widespread and important. With the advent of mathematical and com-
puter models for such phenomena, there has been an increasing interest to study
them also in human populations. The work presented here focuses on resolving the
long-standing puzzle of opinion clustering.
The emergence and persistence of pluralism is a striking phenomenon in a world
in which social networks are highly connected and social influence is an ever present
force that reduces differences between those who interact. We have developed a
formal theory of social influence that, besides anomie and monoculture, shows
a third, pluralistic phase characterized by opinion clustering. It occurs, when all
individuals interact with each other and noise prevents the convergence to a single
opinion, despite homophily.
a b
Fig. 4.4 Comparison of the (a) Durkheimian model and (b) the noisy BC-model. Figures plot the
size of the biggest cluster versus the number of clusters and compare it to the case of random
fragmentation in all simulation runs that resulted in more than one and less than 32 clusters.
Figure 4.4a is based on the simulation experiment with the Durkheimian model underlying Fig. 4.3.
Figure 4.4b is based on an experiment with the BC-model [33] where we varied the bounded-
confidence level between 0.01 and 0.15 in steps of 0.02 and the noise level between 5 and 50 in
steps of 5. We conducted 100 replications per parameter combination and measured the number of
clusters and the size of the biggest cluster after 250,000 iterations. White solid lines represent the
average size of the biggest cluster. The dark blue area shows the respective interquartile range and
the light blue area the complete value range. For comparison, we generated randomly fragmented
opinion distributions of N D 100 agents where n agents hold random opinions (N.0; 50/) and the
remaining N n agents hold opinion oi D 0 and form one big cluster. We varied the value of n
between 0 and 100 in steps of 1 and generated 1,000 distributions per condition. The average size of
the biggest cluster of the resulting distributions is shown by the thin yellow-black line. (The curve
stops at 22, since this is the highest number of clusters generated.) The bold yellow-black lines
represent the related interquartile range. We find that the value range of the Durkheimian model
(blue area) hardly overlaps with the interquartile range of the fragmented distributions (yellow
area). This demonstrates that the Durkheimian model shows clustering rather than fragmentation.
In contrast, Fig. 4.4b illustrates that the distributions of the noisy BC-model and the results for
random fragmentation overlap
Our model does not assume negative influence, and it behaves markedly dif-
ferent from bounded confidence models, in which white opinion noise produces
fragmentation rather than clustering. Furthermore, our model does not rely on the
problematic assumption of classical influence models that agents are forevermore
cut-off from influence by members of distinct clusters. In order to demonstrate
this, we studied model predictions in a setting where all members of the population
interact with each other. However, empirical research shows that opinion clustering
tends to coincide with clustered network structures [20] and spatial separation [18].
It would therefore be natural to generalize the model in a way that it also considers
the structure of real social networks. Such a model is obtained by replacing the
values wij .t/ by wij .t/aij , where aij are the entries of the adjacency matrix (i.e.
aij D 1, if individuals i and j interact, otherwise aij D 0). Then, the resulting
opinion clusters are expected to have a broad range of different sizes, similar to
what is observed for the sizes of social groups.
Our model highlights the functional role that “noise” (randomness, fluctuations,
or other sources of variability) plays for the organization of social systems. It
4.4 Discussion 111
furthermore shows that the combination of two mechanisms (deterministic integrat-

ing forces and stochastic disintegrating forces) can give rise to new phenomena.
We also believe that our results are meaningful for the analysis of the social
integration of our societies. According to Durkheim’s theory of the development
of societies [26], traditional human societies are characterized by “mechanical
solidarity”. In these societies, individuals are strongly integrated in very homo-
geneous communities which exert strong influence on the behavior and opinions
of individuals. According to Durkheim, however, these regulating social structures
dissolve as societies turn modern. In addition, Durkheim [26] and contemporary
social thinkers [27] argue that in modern and globalized societies individuals
are increasingly exposed to disintegrating forces, which foster individualization
[26]. As a consequence, the social forces which let individuals follow societal
norms may lose their power to limit individual variation. Durkheim feared that
the high diversity could disintegrate societies as they modernize [26]. That is,
extreme individualization in modern societies may obstruct the social structures that
traditionally provided social support and guidance to individuals.
Today, modern societies are highly diverse, but at the same time they are far from
a state of disintegration as foreseen by Durkheim. He argued that this is possible
if societies develop what he called “organic solidarity”. In this state societies are
highly diverse but at the same time the division of labor creates a dense web
of dependencies which integrate individuals into society and generate sufficient
moral and social binding [26]. Strikingly, our formal model of Durkheim’s theory
revealed another possibility which does not require additional integrating structures
like the division of labor: Besides monoculture and anomie, there is a third,
pluralistic clustering phase, in which individualization prevents overall consensus,
but at the same time, social influence can still prevent extreme individualism.
The interplay between integrating and disintegrating forces leads to a plurality of
opinions, while metastable subgroups occur, within which individuals find a local
consensus. Individuals may identify with such subgroups and develop long-lasting
social relationships with similar others. Therefore, they are not isolated and not
without support or guidance, in contrast to the state of disintegration that Durkheim
was worried about.
We have seen, however, that pluralism and cultural diversity require an approx-
imate balance between integrating and disintegrating forces. If this balance is
disturbed, societies may drift towards anomie or monoculture. It is, therefore,
interesting to ask how the current tendency of globalization will influence society
and cultural dynamics. The Internet, interregional migration, and global tourism,
for example, make it easy to get in contact with members of distant and different
cultures. Previous models [24, 35] suggest that this could affect cultural diversity
in favor of a monoculture. However, if the individual striving for uniqueness is
sufficiently strong, formation of diverse groups (a large variety of international
social communities) should be able to persist even in a globalizing world. In
view of the alternative futures, characterized by monoculture or pluralism, further
theoretical, empirical, and experimental research should be performed to expand our
knowledge of the mechanisms that will determine the future of pluralistic societies.
Acknowledgements We thank Tobias Stark, Heiko Rauhut, Jacob G. Foster and Michael Macy
as well as the members of the Norms and Networks cluster at the Department of Sociology at
the University of Groningen and the members of the Cooperative Relations and Social Networks
Seminar at the Department of Sociology at Utrecht University for their constructive comments.
Author Contributions Conceived and designed the experiments: MM AF DH. Performed the
experiments: MM. Analyzed the data: MM. Wrote the paper: MM AF DH.
References
1. E. Ben-Jacob, O. Schochet, A. Tenenbaum, I. Cohen, A. Czirok, T. Vicsek, Generic Modeling

of Cooperative Growth-Patterns in Bacterial Colonies. Nature 368, 46–49 (1994)
2. R. Jeanson, C. Rivault, J.L. Deneubourg, S. Blanco, R. Fournier, C. Jost, G. Theraulaz, Self-
organized aggregation in cockroaches. Anim. Behav. 69, 169–180 (2005)
3. J. Gautrais, C. Jost, G. Theraulaz, Key behavioural factors in a self-organised fish school model.
Ann. Zool. Fenn. 45, 415–428 (2008)
4. M. Ballerini, N. Calbibbo, R. Candeleir, A. Cavagna, E. Cisbani, et al., Interaction ruling
animal collective behavior depends on topological rather than metric distance: Evidence from
a field study. Proc. Natl. Acad. Sci. USA 105, 1232–1237 (2008)
5. I.D. Couzin, J. Krause, N.R. Franks, S.A. Levin, Effective leadership and decision-making in
animal groups on the move. Nature 433, 513–516 (2005)
6. Y. Iwasa, V. Andreasen, S.A. Levin, Aggregation in model ecosystems. 1. Perfect Aggregation.
Ecol. Model 37, 287–302 (1987)
7. M. Moussaid, N. Perozo, S. Garnier, D. Helbing, G. Theraulaz, The walking behaviour of
pedestrian social groups and its impact on crowd dynamics. PLoS One 5(4), e10047 (2010)
8. H.A. Makse, S. Havlin, H.E. Stanley, Modeling Urban-Growth Patterns. Nature 377, 608–612
(1995)
9. M. Batty, The size, scale, and shape of cities. Science 319, 769–771 (2008)
10. D.J. Watts, S.H. Strogatz, Collective dynamics of ‘small-world’ networks. Nature 393, 440–
442 (1998)
11. A.L. Barabasi, R. Albert, Emergence of scaling in random networks. Science 286, 509–512
(1999)
12. G. Palla, A.L. Barabasi, T. Vicsek, Quantifying social group evolution. Nature 446, 664–667
(2007)
13. D. Liben-Nowell, J. Kleinberg, Tracing information flow on a global scale using Internet chain-
letter data. Proc. Natl. Acad. Sci. USA 105, 4633–4638 (2008)
14. M.E.J. Newman, Coauthorship networks and patterns of scientific collaboration. Proc. Natl.
Acad. Sci. USA 101, 5200–5205 (2004)
15. F. Liljeros, C.R. Edling, L.A.N. Amaral, H.E. Stanley, Y. Aberg, The web of human sexual
contacts. Nature 411, 907–908 (2001)
16. M.P. Fiorina, S.J. Abrams, Political polarization in the American public. Annu. Rev. Polit. Sci.
11, 563–588 (2008)
17. P. DiMaggio, J. Evans, B. Bryson, Have Americans’ social attitudes become more polarized?
Am. J. Sociol. 102, 690–755 (1996)
18. E.L. Glaeser, B.A. Ward, Myths and realities of American political geography. J. Econ.
Perspect. 20, 119–144 (2006)
19. N.P. Mark, Culture and competition: Homophily and distancing explanations for cultural
niches. Am. Sociol. Rev. 68, 319–345 (2003)
20. D. Lazer, et al., Computational Social Science. Science 323, 721–723 (2009)
References 113
21. F.J. Milliken, L.L. Martins, Searching for Common Threads: Understanding the Multiple
Effects of Diversity in Organizational Groups. Acad. Manag. J. 21, 402–433 (1996)
22. P.C. Early, E. Mosakowski, Creating hybrid team cultures: an empirical test of transnational
team functioning. Acad. Manag. J. 43, 26–49 (2000)
23. T.L. Friedman, The World is Flat. A brief history of the twenty-first century (Farrar, Straus and
Giroux, New York, 2005)
24. M.J. Greig, The end of geography? Globalization, communication and culture in the interna-
tional system. J. Conflict Res. 46, 225–243 (2002)
25. W.J. Sutherland, Parallel extinction risk and global distribution of languages and species.
Nature 423, 276–279 (2003)
26. E. Durkheim, The Division of Labor in Society (The Free Press, New York, 1997 [1893])
27. U. Beck, in Reflexive Modernization. Politics, Tradition and Aestherics in the Modern Social
Order, ed. by U. Beck, A. Giddens, S. Lash (Polity Press, Cambridge, 1994)
28. M. McPherson, L. Smith-Lovin, M.E. Brashears, Social isolation in America: Changes in
core discussion networks over two decades. Am. Sociol. Rev. 71, 353–375; also consult the
discussion on this article in Am. Sociol. Rev. 74, 4 (2006)
29. R.P. Abelson, in Contributions to Mathematical Psychology, ed. by N. Frederiksen,
H. Gulliksen (Rinehart Winston, New York, 1964), pp. 142–160
30. M. McPherson, L. Smith-Lovin, J.M. Cook, Birds of a feather: Homophily in social networks.
Annu. Rev. Sociol. 27, 415–444 (2001)
31. S. Aral, L. Muchnik, A. Sundararajan, Distinguishing influenced-based contagion from
homophily-driven diffusion in dynamic networks. Proc. Natl. Acad. Sci. USA 106, 21544–
21549 (2009)
32. A. Nowak, J. Szamrej, B. Latané, From private attitude to public opinion: a dynamic theory of
social impact. Psychol. Rev. 97, 362–376 (1990)
33. R. Hegselmann, U. Krause, Opinion dynamics and bounded confidence models. Analysis, and
Simulation J. Artif. Soc. S 5 (2002)
34. G. Deffuant, S. Huet, F. Amblard, An individual-based model of innnovation diffusion mixing
social value and individual benefit. Am. J. Soc. 110, 1041–1069 (2005)
35. R. Axelrod, The dissemination of culture - A model with local convergence and global
polarization. J. Conflict Res. 41, 203–226 (1997)
36. M.W. Macy, J. Kitts, A. Flache, S. Benard, in Dynamic Social Network Modelling and
Analysis, ed. by R. Breiger, K. Carley, P. Pattison (The National Academies Press, Washington,
DC, 2003), pp. 162–173
37. A. Flache, M. Mäs, How to get the timing right? A computational model of how demographic
faultlines undermine team performance and how the right timing of contacts can solve the
problem. Comput. Math. Organ. Theory 14, 23–51 (2008)
38. Z. Krizan, R.S. Baron, Group polarization and choice-dilemmas: How important is self-
categorization? Eur. J. Soc. Psychol. 37, 191–201 (2007)
39. M. Pineda, R. Toral, E. Hernandez-Garcia, Noisy continuous-opinion dynamics. J. Stat. Mech.
P08001 (2009)
40. M.J. Hornsey, J. Jetten, The individual within the group: Balancing the need to belong with the
need to be different. Pers. Soc. Psychol. Rev. 8, 248–264 (2004)
41. V.L. Vignoles, X. Chryssochoou, G.M. Breakwell, The distinctiveness principle: Identity,
meaning, and the bounds of cultural relativity. Pers. Soc. Psychol. Rev. 4, 337–354 (2000)
42. R. Imhoff, H.P. Erb, What motivates nonconformity? Uniqueness seeking blocks majority
influence. Pers. Soc. Psychol. B 35, 309–320 (2009)
43. C.R. Snyder, H.L. Fromkin, Uniqueness. The Human Pursuit of Difference (Plenum Press,
New York and London, 1980)
44. K. Klemm, V.M. Eguiluz, R. Toral, M.S. Miguel, Global culture: A noise-induced transition in
finite systems. Phys. Rev. E 67, 045101(R) (2003)
45. H.E. Stanley, Introduction to Phase Transitions and Critical Phenomena (Oxford University
Press, Oxford and New York, 1971)
46. W. Weidlich, The statistical description of polarization phenomena in society. Br. J. Math. Stat.
Psychol. 24, 251–266 (1971)
47. T.M. Liggett, Interacting Particle Systems (Springer, New York, 1985)
48. K. Sznajd-Weron, J. Sznajd, Opinion Evolution in Closed Community. Int. J. Mod. Phys. C 11,
1157–1165 (2000)
49. S. Galam, Heterogeneous beliefs, segregation, and extremism in the making of public opinions.
Phys. Rev. E 71, 46123 (2005)
50. L. Behera, F. Schweitzer, On spatial consensus formation: Is the Sznajd model different from
a voter model? International Journal of Modern Physics C 14(10), 1331–1354 (2003)
51. E. Bonabeau, Agent-based modeling: Methods and techniques for simulating human systems.
Proc. Natl. Acad. Sci. USA 99, 7280–7287 (2002)
52. E. Durkheim, Suicide. A study in Sociology (The Free Press, New York, 1997 [1897])
Chapter 5
Spatial Self-organization
Through Success-Driven Mobility
5.1 Introduction
Although the biological, social, and economic world are full of self-organization
phenomena, many people believe that the dynamics behind them is too complex
to be modelled in a mathematical way. Reasons for this are the huge number
of interacting variables, most of which cannot be quantified, plus the assumed
freedom of decision-making or large fluctuations within biological and socio-
economic systems. However, in many situations, the living entities making up these
systems decide for some (more or less) optimal behavior, which can make the latter
describable or predictable to a certain extend [1–10]. This is even more the case
for the behavior shown under certain constraints like, for example, in pedestrian
or vehicle dynamics [11–13]. While pedestrians or vehicles can move freely at
small traffic densities, at large densities the interactions with the others and with
the boundaries of the street confines them to a small spectrum of moving behaviors.
Consequently, empirical traffic dynamics can be reproduced by simulation models
surprisingly well [11–17].
In this connection, it is also interesting to mention some insights gained in
statistical physics and complex systems theory: Non-linearly interacting variables
do not change independently of each other, and in many cases there is a separation
of the time scales on which they evolve. This often allows to “forget” about the
vast number of rapidly changing variables, which are usually determined by a small
number of “order parameters” and treatable as fluctuations [18, 19]. In the above
mentioned examples of traffic dynamics, the order parameters are the traffic density
and the average velocity of pedestrians or vehicles.

This chapter reprints a previous publication with kind permission of the copyright owner, the
Technique & Technologies Press Ltd. It is requested to cite this work as follows: D. Helbing and
T. Platkowski, Self-organization in space and induced by fluctuations. International Journal of
Chaos Theory and Applications 5(4), 47–62 (2000).

116 5 Spatial Self-organization Through Success-Driven Mobility
Another discovery is that, by proper transformations or scaling, many different

models can by mapped onto each other, i.e. they behave basically the same
[13, 18–20]. That is, a certain class of models displays the same kinds of states,
shows the same kinds of transitions among them, and can be described by the
same “phase diagram”, displaying the respective states as a function of some
“control parameters” [18, 19]. We call such a class of models a “universality
class”, since any of these models shows the same kind of “universal” behavior, i.e.,
the same phenomena. Consequently, one usually tries to find the simplest model
having the properties of the universality class. While physicists like to call it a
“minimal model”, “prototype model”, or “toy model”, mathematicians named the
corresponding mathematical equations “normal forms” [18, 19, 21, 22].
Universal behavior is the reason of the great success of systems theory [23–25]
in comparing phenomena in seemingly completely different systems, like physical,
biological, or social ones. However, since these systems are composed of different
entities and their corresponding interactions can be considerably different, it is not
always easy to identify the variables and parameters behind their dynamics. It can
be helpful to take up game-theoretical ideas, here, quantifying interactions in terms
of payoffs [2–7, 26, 27]. This can be applied to positive (profitable, constructive,
cooperative, symbiotic) or negative (competitive, destructive) interactions in socio-
economic or biological systems, but to attractive and repulsive interactions in
physical systems as well [28].
In the following, we will investigate a simple model for interactive motion in
space allowing to describe (1) various self-organized agglomeration phenomena,
like settlement formation, and segregation phenomena, like ghetto formation,
emerging from different kinds of interactions and (2) fluctuation-induced ordering
or self-organization phenomena.
Noise-related phenomena can be quite surprising and have, therefore, recently
attracted the interest of many researchers. For example, we mention stochastic
resonance [29], noise-driven motion [30, 31], and “freezing by heating” [32].
The issue of order through fluctuations has already a considerable history.
Prigogine has discussed it in the context of structural instability with respect to
the appearance of a new species [33, 34], but this is not related to the approach
considered in the following.
Moreover, since both, the initial conditions and the interaction strengths in our
model are assumed independent of the position in space, the fluctuation-induced
self-organization discussed later on must be distinguished from so-called “noise-
induced transitions” as well, where we have a space-dependent diffusion coefficient
which can induce a transition [35].
Although our model is related to diffusive processes, it is also different from
reaction-diffusion systems that can show fluctuation-induced self-organization phe-
nomena known as Turing patterns [36–41], which are usually periodic in space. The
noise-induced self-organization that we find seems to have (1) no typical length
scale and (2) no attractor, since our model is translation-invariant. This, however, is
not yet a final conclusion and still subject to investigations.
5.2 Discrete Model of Interactive Motion in Space 117
We also point out that, in the case of spatial invariance, self-organization

directly implies spontaneous symmetry-breaking, and we expect a pronounced
history-dependence of the resulting state. Nevertheless, when averaging over a
large ensemble of simulation runs with different random seeds, we again expect
a homogeneous distribution, since this is the only result compatible with translation
invariance.
Finally, we mention that our results do not fit into the concept of noise-induced
transitions from a metastable disordered state (local optimum) to a stable ordered
state (global optimum), which are, for example, found for undercooled fluids,
metallic glasses, or some granular systems [42–44].
5.2 Discrete Model of Interactive Motion in Space
Describing motion in space has the advantage that the essential variables like
positions, densities, and velocities are well measurable, which allows to calibrate,
test, and verify or falsify the model. Although we will focus on motion in “real”
space like the motion of pedestrians or bacteria, our model may also be applied to
changes of positions in abstract spaces, e.g. to opinion changes on an opinion scale
[7, 46]. There exist, of course, already plenty of models for motion in space, and we
can mention only a few [1, 7, 11–21, 26, 28, 30–32, 35–41, 43, 45–53]. Most of them
are, however, rather specific for certain systems, e.g., for fluids or for migration
behavior.
For simplicity, we will restrict the following considerations to a one-dimensional
space, but a generalization to higher dimensions is straightforward. The space is
divided into I equal cells i which can be occupied by the entities. We will apply
periodic boundary conditions, i.e. the space can be imagined as a circle. In our
model, we group the N entities ˛ in the system into homogeneously behaving
subpopulations a. If nai .t/ denotes the number of entities of subpopulation a in
cell i at time t, we have the relations
X X
nai .t/ D Na ; Na D N: (5.1)
i a
We will assume that the numbers Na of entities belonging to the subpopulations a do

not change. It is, however, easy to take additional birth and death processes and/or
transitions of individuals from one subpopulation to another into account [1].
In order not to introduce any bias, we start our simulations with a completely
uniform distribution of the entities in each subpopulation over the I cells of the
system, i.e., nai .0/ D nahom D Na =I , for which we choose a natural number. At times
t 2 f1; 2; 3; :::g, we apply the following update steps, using a random sequential
update (although a parallel update is possible as well, which is more efficient [54,
55], but normally less realistic [56] due to the assumed synchronous updating):
1st step: For updating of the state of entity ˛, given it is a member of

subpopulation a and located in cell i , determine the so-called (expected) “success”
according to the formula
X
Sa .i; t/ D Pab nbi .t/ C ˛ .t/ : (5.2)
b
Here, Pab is the “payoff” in interactions of an entity of subpopulation a with an

entity of subpopulation b. The payoff Pab is positive for attractive, profitable, con-
structive, or symbiotic interactions, while it is negative for repulsive, competitive,
or destructive interactions. Notice that Pab is assumed to bePindependent of the
position (i.e., translation invariant), while the total payoff b Pab nbi .t/ due to
interactions depends on the distribution of entities over the system. The latter is
an essential point for the possibility of fluctuation-induced self-organization. We
also point out that, in formula (5.2), pair interactions are restricted to the cell in
which the individual is located. Therefore, we do not assume spin-like or Ising-like
interactions, in contrast to other quantitative models proposed for the description of
social behavior [9, 10].
The quantities ˛ .t/ are random variables allowing to consider individual
variations of the success, which may be “real” or due to uncertainty in the evaluation
or estimation of success. In our simulation program, they are uniformly distributed
in the interval Œ0; Da , where Da is the fluctuation strength (not to be mixed up with
a diffusion constant). However, other specifications of the noise term are possible
as well.
2nd step: Determine the (expected) successes Sa .i ˙ 1; t/ for the nearest
neighbors .i ˙ 1/ as well.
3rd step: Keep entity ˛ in its previous cell i , if Sa .i; t/ maxfS.i 1; t/; S.i C
1; t/g. Otherwise, move to cell .i 1/, if S.i 1; t/ > S.i C 1; t/, and move to cell
.i C 1/, if S.i 1; t/ < S.i C 1; t/. In the remaining case S.i 1; t/ D S.i C 1; t/,
jump randomly to cell .i 1/ or .i C 1/ with probability 1/2.
If there is a maximum density max D Nmax =I of entities, overcrowding can be
avoided by introducing a saturation factor
Nj .t/ X
c.j; t/ D 1 ; Nj .t/ D naj .t/ ; (5.3)
Nmax a
and performing the update steps with the generalized success
Sa0 .j; t/ D c.j; t/Sa .j; t/ (5.4)
instead of Sa .j; t/, where j 2 fi 1; i; i C1g. The model can be also easily extended
to include long distance interactions, jumps to more remote cells, etc. (cf. Sect. 5).
5.3 Simulation Results 119
5.3 Simulation Results
We consider two subpopulations a 2 f1; 2g and N1 D N2 D 100 entities in

each subpopulation, which are distributed over I D 20 cells. The payoff matrix
.Pab / will be represented by the vector P D .P11 ; P12 ; P21 ; P22 /, where we will
restrict ourselves to jPab j 2 f1; 2g for didactical reasons. For symmetric interactions
between subpopulations, we have Pab D Pba , while for asymmetric interactions,
there is Pab ¤ Pba , if a ¤ b. For brevity, the interactions within the same
population will be called self-interactions, those between different populations
cross-interactions.
To characterize the level of self-organization in each subpopulation a, we can,
for example, use the overall successes
1 XX a
Sa .t/ D ni .t/ Pab nbi .t/ ; (5.5)
I2 i
b
the variances
1 X a
Va .t/ D Œni .t/ nahom 2 ; (5.6)
I2 i
or the alternation strengths
1 X a
Aa .t/ D Œni .t/ nai1 .t/2 : (5.7)
I2 i
5.3.1 Symmetric Interactions
By analogy with a more P complicated model [28] it is expected that the global
overall success S.t/ D a Sa .t/ is an increasing function in time, if the fluctuation
strengths Da are zero. However, what happens at finite noise amplitudes Da is
not exactly known. One would usually expect that finite noise tends to obstruct or
suppress self-organization, which will be investigated in the following.
We start with the payoff matrix P D .2; 1; 1; 2/ corresponding to positive
(or attractive) self-interactions and negative (or repulsive) cross-interactions. That
is, entities of the same subpopulation like each other, while entities of different
subpopulations dislike each other. The result will naturally be segregation (“ghetto
formation”) [1, 57], if the noise amplitude is small. However, segregation is
suppressed by large fluctuations, as expected (see Fig. 5.1).
However, for medium noise amplitudes Da , we find a much more pronounced
self-organization (segregation) than for small ones (compare Fig. 5.2 with Fig. 5.1).
The effect is systematic insofar as the degree of segregation (and, hence, the overall
success) increases with increasing noise amplitude, until segregation breaks down
above a certain critical noise level.
Fig. 5.1 Resulting distribution of entities at t D 4; 000 for the payoff matrix P D .2; 1; 1; 2/
at small fluctuation strength Da D 0:1 (left) and large fluctuations strength Da D 5 (right)
Fig. 5.2 As Fig. 5.1, but with medium fluctuation strength Da D 2 (left) and Da D 3 (right)
Let us investigate some other cases: For the structurally similar payoff matrix
.1; 2; 2; 1/, we find segregation as well, which is not surprising. In contrast,
we find agglomeration for the payoff matrices .1; 2; 2; 1/ and .2; 1; 1; 2/. This
agrees with intuition, since all entities like each other in these cases, which
makes them move to the same places, like in the formation of settlements [1], the
development of trail systems [17, 52, 53], or the build up of slime molds [34, 50].
More interesting is the case corresponding to the payoff matrix .1; 2; 2; 1/,
where the cross-interactions are positive (attractive), while the self-interactions
are negative (repulsive). One may think that this causes the entities of the same
subpopulation to spread homogeneously over the system, and in all cells would
result an equal number of entities of both subpopulations, which is compatible with
mutual attraction. However, this homogeneous distribution turns out to be unstable
with respect to fluctuations. Instead, we find agglomeration! This result is more
intuitive if we imagine one subpopulation to represent women and the other one men
(without taking this example too serious). While the interaction between women and
men is normally strongly attractive, the interactions among men or among women
may be considered to be weakly competitive. As we all know, the result is a tendency
of young men and women to move into cities. Corresponding simulation results
Fig. 5.3 As Fig. 5.1, but for the payoff matrix P D .1; 2; 2; 1/ and Da D 0:05 (left), Da D
1:5 (middle), and Da D 5 (right)
for different noise strengths are depicted in Fig. 5.3. Again, we find that the self-
organized pattern is destroyed by strong fluctuations in favour of a more or less
homogeneous distribution, while medium noise strengths further self-organization.
For the payoff matrices .2; 1; 1; 2/ and .2; 1; 1; 2/, i.e. cases of strong
negative self-interactions, we find a more or less homogeneous distribution of
entities in both subpopulations, irrespective of the noise amplitude. In contrast, the
payoff matrix .1; 2; 2; 1/ corresponding to negative self-interactions but even
stronger negative cross-interactions, leads to another self-organized pattern. We may
describe it as the formation of lanes, as it is observed in pedestrian counterflows
[12, 28] or in sheared granular media with different kinds of grains [47]. While
both subpopulations tend to separate from each other, at the same time they tend to
spread over all the available space (see Fig. 5.4), in contrast to the situation depicted
in Figs. 5.1 and 5.2. Astonishingly enough, a medium level of noise again supports
self-organized ordering, since it helps the subpopulations to separate from each
other.
We finally mention that a finite saturation level suppresses self-organization in a
surprisingly strong way, as is shown in Fig. 5.5. Instead of pronounced segregation,
we will find a result similar to lane formation, and even strong agglomeration will
be replaced by an almost homogeneous distribution.
Noise-Induced Ordering
A possible interpretation for noise-induced ordering would be that fluctuations

allow the system to leave local minima (corresponding to partial agglomeration
or segregation only). This could trigger a transition to a more stable state with
Fig. 5.4 As Fig. 5.1, but for

the payoff matrix
P D .1; 2; 2; 1/ and
Da D 0:05 (top), Da D 0:5
(middle), and Da D 5
(bottom)
Fig. 5.5 Resulting

distribution of entities at
t D 4; 000 with saturation
level Nmax D 50. Top:
P D .2; 1; 1; 2/ and
Da D 3. Bottom:
P D .1; 2; 2; 1/ and
Da D 1:5
Fig. 5.6 Temporal evolution

of the distribution of entitities
within subpopulation a D 2
for P D .2; 1; 1; 2/ and
Da D 3 (top),
P D .1; 2; 2; 1/ and
Da D 1:5 (middle), and
P D .1; 2; 2; 1/ and
Da D 0:5 (bottom)
more pronounced ordering. However, although this interpretation is consistent with

a related example discussed in [28], the idea of a step-wise coarsening process is not
supported by the temporal evolution of the distribution of entities (see Fig. 5.6) and
the time-dependence of the overall success within the subpopulations (see Fig. 5.7).
This idea is anyway not applicable to segregation, since, in the one-dimensional
case, the repulsive clusters of different subpopulations cannot simply pass each other
in order to join others of the same subpopulation.
According to Figs. 5.6 and 5.7, segregation and agglomeration rather take place
in three phases: First, there is a certain time interval, during which the distribution of
entities remains more or less homogeneous. Second, there is a short period of rapid
self-organization. Third, there is a continuing period, during which the distribution
and overall success do not change anymore. The latter is a consequence of the short-
range interactions within our model, which are limited to the nearest neighbors.
Therefore, the segregation or aggregation process practically stops, after separate
peaks have evolved. This is not the case for lane formation, where the entities
redistribute, but all cells remain occupied, so that we have ongoing interactions.
This is reflected in the non-stationarity of the lanes and by the oscillations of the
overall success.
We suggest the following interpretation for the three phases mentioned above:
During the first time interval, which is characterized by a quasi-continuous dis-
tribution of entities over space, a long-range pre-ordering process takes place.
After this “phase of preparation”, order develops in the second phase similar to
crystallization, and it persists in the third phase. The role of fluctuations seems

of the overall success within
both subpopulations for
P D .2; 1; 1; 2/ and
Da D 3 (top),
P D .1; 2; 2; 1/ and
Da D 1:5 (middle), and
P D .1; 2; 2; 1/ and
Da D 0:5 (bottom)
to be the following: An increased noise level avoids a rash local self-organization

by keeping up a quasi-continuous distribution of entities, which is required for a
redistribution of entities over larger distances. In this way, a higher noise level
increases the effective interaction range by extending the first phase, the “interaction
phase”. As a consequence, the resulting structures are more extended in space (but
probably without a characteristic length scale, see Introduction).
It would be interesting to investigate, whether this mechanism has something
to do with the recently discovered phenomenon of “freezing by heating”, where a
medium noise level causes a transition to a highly ordered (but energetically less
stable) state, while extreme noise levels produce a disordered, homogeneous state
again [32].
5.3.2 Asymmetric Interactions
Even more intriguing transitions than in the symmetric case can be found for
asymmetric interactions between the subpopulations. Here, we will focus on the
payoff matrix .1; 2; 2; 1/, only. This example corresponds to the curious case,
where individuals of subpopulation 1 weakly dislike each other, but strongly like
individuals of the other subpopulation. In contrast, individuals of subpopulation 2
weakly like each other, but they strongly dislike the other subpopulation. A good
example for this is hard to find. With some good will, one may imagine subpopula-
tion 1 to represent poor people, while subpopulation 2 corresponds to rich people.
What will be the outcome? In simple terms, the rich are expected to agglomerate in
a few areas, if the poor are moving too nervously (see Fig. 5.8). In detail, however,
the situation is quite complex, as discussed in the next paragraph.
Noise-Induced Self-organization
At small noise levels Da , we will just find more or less homogeneous distributions of
the entities. This is already different from the cases of agglomeration, segregation,
and lane formation we have discussed before. Self-organization is also not found
at higher noise amplitudes Da , as long as we assume that they are the same in both
subpopulations (i.e., D1 D D2 ). However, given that the fluctuation amplitude D2 in
subpopulation 2 is small, we find an agglomeration in subpopulation 2, if the noise
level D1 in subpopulation 1 is medium or high, so that subpopulation 1 remains
homogeneously distributed. The order in subpopulation 2 breaks down, as soon as
we have a relevant (but still small) noise level D2 in subpopulation 2 (see Fig. 5.8).
Hence, we have a situation where asymmetric noise with D1 ¤ D2 can facilitate
self-organization in a system with completely homogeneous initial conditions and
Fig. 5.8 Distributions for P D .1; 2; 2; 1/ and D1 D D2 D 0:5 (top left), D1 D 50, D2 D
0:5 (top right), D1 D 5; 000, D2 D 0:5 (bottom left), D1 D 5; 000, D2 D 5 (bottom right)

of the distribution of entitities
within subpopulation a D 2
(top) and of the overall
successes (bottom) for
P D .1; 2; 2; 1/ and
D1 D 50, D2 D 0:5
interaction laws, where we would not have ordering without any noise. We call
this phenomenon noise-induced self-organization. It is to be distinguished from
the noise-induced increase in the degree of ordering discussed above, where we
have self-organization even without noise, if only the initial conditions are not fully
homogeneous.
The role of the noise in subpopulation 1 seems to be the following: Despite of
the attractive interaction with subpopulation 2, it suppresses an agglomeration in
subpopulation 1, in particular at the places where subpopulation 2 agglomerates.
Therefore, the repulsive interaction of subpopulation 2 with subpopulation 1
is effectively reduced. As a consequence, the attractive self-interaction within
subpopulation 2 dominates, which gives rise to the observed agglomeration.
The temporal development of the distribution of entities and of the overall
success in the subpopulations gives additional information (see Fig. 5.9). As in
the case of lane formation, the overall success fluctuates strongly, because the
subpopulations do not separate from each other, causing ongoing interactions.
Hence, the resulting distribution is not stable, but changes continuously. It can,
therefore, happen, that clusters of subpopulation 2 merge, which is associated with
an increase of overall success in subpopulation 2 (see Fig. 5.9).
5.4 Conclusions
We have proposed a game theoretical model for self-organization in space, which

is applicable to many kinds of biological, economic, and social systems with
various types of profitable or competitive self- and cross-interactions between
5.4 Conclusions 127
subpopulations of the system. Depending on the structure of the payoff matrix,

we found several different self-organization phenomena like agglomeration, segre-
gation, or lane formation. It turned out that medium noise strengths can increase
the resulting level of order, while a high noise level leads to more or less
homogeneous distributions of entities over the available space. The mechanism of
noise-induced ordering in the above discussed systems with short-range interactions
seems to be the following: Noise extends a “pre-ordering” phase by keeping up a
quasi-continuous distribution of entities, which allows a long-range ordering. For
asymmetric payoff matrices, we can even have the phenomenon of noise-induced
self-organization, although we start with completely homogeneous distributions and
homogeneous (translation-invariant) payoffs. However, the phenomenon requires
different noise amplitudes in both subpopulations. The role of noise is to suppress
agglomeration in one of the subpopulations, in this way reducing repulsive effects
that would suppress agglomeration in the other subpopulation.
We point out that all the above results can be semi-quantitatively understood by
means of a linear stability analysis of a related continuous version of the model
[28]. This continuous version indicates that the linearly most unstable modes are
the ones with the shortest wave length, so that one does not expect a characteristic
length scale in the system. This is different from reaction-diffusion systems, where
the most unstable mode has a finite wave length, which gives rise to the formation
of periodic patterns. Nevertheless, the structures evolving in our model are spatially
extended, but non-periodic. The spatial extension is increasing with the fluctuation
strength, unless a critical noise amplitude is exceeded.
For a better agreement with real systems, the model can be generalized in
many ways. The entities may perform a biased or unbiased random walk in
space. One can allow random jumps to neigboring cells with some prescribed
probability. This probability may depend on the subpopulation, and thus we can
imitate different mobilities of the considered subpopulations. Evolution is slowed
down by introducing a threshold, fixed or random, so that the entities change to other
cells only if the differences in the relevant successes are bigger than the imposed
threshold. The model can be also generalized to higher dimensions, with expected
interesting patterns of self-organized structures.
In general, the random variables ˛ .t/ in the definition of the success functions
can be allowed to have different variances for the considered cell i and the
neighboring cells, with the interpretation that the uncertainty in the evaluation of the
success in the considered cell is different (e.g. smaller) than that in the neighboring
cells. Moreover, the uncertainties can be different for various subpopulations, which
could reflect to some extent their different knowledge and behavior.
One can as well study systems with more than two subpopulations, the influence
of long-range interactions, etc. The entities can also be allowed to jump to more
remote cells. As an example, the following update rule could be implemented: Move
entity ˛ from cell i to the cell .i C l/ for which
Sa00 .i C l; t/ D d jlj c.i C l; t/Sa .i C l; t/ (5.8)
is maximal (jlj D 0; 1; : : : ; lmax /. If there are m cells in the range f.x lmax /; : : : ,
.x C lmax /g with the same maximal value, choose one of them randomly with
probability 1=m. According to this, when spontaneously moving to another cell, the
entity prefers cells in the neighborhood with higher success. The indirect interaction
behind this transition, which is based on the observation or estimation of the success
in the neighborhood, is short-ranged if lmax I , otherwise long-ranged. Herein,
lmax denotes the maximum number of cells which an entity can move within one
time step. The factor containing d with 0 < d < 1 allows to consider that it is
less likely to move for large distances, if this is not motivated by a higher success.
A value d < 1 may also reflect the fact that the observation or estimation of the
success over large distances becomes more difficult and less reliable.
Acknowledgements D.H. thanks Eörs Szathmáry and Tamás Vicsek for inspiring discussions
and the German Research Foundation (DFG) for financial support by a Heisenberg scholarship.
T.P. is grateful to the Alexander-von-Humboldt Foundation for financial support during his stay in
Stuttgart.
References
1. W. Weidlich, Physics and social science—The approach of synergetics. Phys. Reports 204,
1–163 (1991)
2. J. von Neumann, O. Morgenstern, Theory of Games and Economic Behavior. (Princeton
University, Princeton, 1944)
3. R. Axelrod, W.D. Hamilton, The evolution of cooperation. Science 211, 1390–1396 (1981)
4. R. Axelrod, D. Dion, The further evolution of cooperation. Science 242, 1385–1390 (1988)
5. J. Hofbauer, K. Sigmund, The Theory of Evolution and Dynamical Systems. (Cambridge
University Press, Cambridge, 1988)
6. N.S. Glance, B.A. Huberman, The dynamics of social dilemmas. Sci. Am. 270, 76–81 (1994)
7. D. Helbing, Quantitative Sociodynamics. Stochastic Methods and Models of Social Interaction
Processes. (Kluwer Academics, Dordrecht, 1995)
8. F. Schweitzer (ed.), Self-Organization of Complex Structures, in From individual to Collective
Dynamics. (Gordon and Breach, Amsterdam, 1997)
9. M. Lewenstein, A. Nowak, B. Latané, Statistical mechanics of social impact. Phys. Rev. A 45,
763–776 (1992)
10. S. Galam, Rational group decision making. Physica A 238, 66–80 (1997)
11. D. Helbing, Verkehrsdynamik [Traffic Dynamics]. (Springer, Berlin, 1997)
12. D. Helbing, P. Molnár, Social force model for pedestrian dynamics. Phys. Rev. E 51,
4282–4286 (1995)
13. D. Helbing, A. Hennecke, M. Treiber, Phase diagram of traffic states in the presence of
inhomogeneities. Phys. Rev. Lett. 82, 4360–4363 (1999)
14. D. Helbing, B.A. Huberman, Coherent moving states in highway traffic. Nature 396, 738–740
(1998)
15. M. Treiber, D. Helbing, Macroscopic simulation of widely scattered synchronized traffic states.
J. Phys. A: Math. Gen. 32, L17-L23 (1999)
References 129
16. M. Treiber, A. Hennecke, D. Helbing, Congested traffic states in empirical observations and
microscopic simulations. Phys. Rev. E 62(2), 1805–1824 (2000)
17. D. Helbing, J. Keltsch, P. Molnár, Modelling the evolution of human trail systems. Nature 388,
47–50 (1997)
18. H. Haken Synergetics. (Springer, Berlin, 1977)
19. H. Haken Advanced Synergetics. (Springer, Berlin, 1983)
20. D. Helbing, D. Mukamel, G.M. Schütz, Global phase diagram of a one-dimensional driven
lattice gas. Phys. Rev. Lett. 82, 10–13 (1999)
21. P. Manneville, Dissipative Structures and Weak Turbulence. (Academic Press, New York,
1990)
22. E.C. Zeeman (ed.), Catastrophe Theory. (Addison-Wesley, London, 1977)
23. L. von Bertalanffy, General System Theory. (Braziller, New York, 1968)
24. W. Buckley, Sociology and Modern Systems Theory. (Prentice-Hall, Englewood Cliffs, NJ,
1967)
25. A. Rapoport, General System Theory. Essential Concepts and Applications. (Abacus Press,
Tunbridge Wells, Kent, 1986)
26. R. Feistel, W. Ebeling, Evolution of Complex Systems. (Kluwer Academic, Dordrecht, 1989)
27. D. Helbing, Stochastic and Boltzmann-like models for behavioral changes, and their relation
to game theory. Physica A 193, 241–258 (1993)
28. D. Helbing, T. Vicsek, Optimal self-organization. New J. Phys. 1, 13.1–13.17 (1999)
29. L. Gammaitoni, P. Hänggi, P. Jung, F. Marchesoni, Stochastic resonance. Rev. Modern Phys.
70, 223–288 (1998)
30. J. Łuczka, R. Bartussek, P. Hänggi, White noise induced transport in periodic structures.
Europhys. Lett. 31, 431–436 (1995)
31. P. Reimann, R. Bartussek, R. Häußler, P. Hänggi, Brownian motors driven by temperature
oscillations. Phys. Lett. A 215, 26–31 (1996)
32. D. Helbing, I.J. Farkas, T. Vicsek, Freezing by heating in a driven mesoscopic system. Phys.
Rev. Lett. 84, 1240–1243 (2000)
33. G. Nicolis, I. Prigogine, Self-Organization in Nonequilibrium Systems. From Dissipative
Structures to Order through Fluctuations. (Wiley, New York, 1977)
34. I. Prigogine, Order through fluctuation: Self-organization and social system, in Evolution
and Consciousness. Human Systems in Transition, ed. by E. Jantsch and C.H. Waddington
(Addison-Wesley, Reading, MA, 1976), pp. 93–130
35. W. Horsthemke, R. Lefever, Noise-Induced Transitions. (Springer, Berlin, 1984)
36. A.M. Turing, The chemical basis of morphogenesis. Phil. Trans. Roy. Soc. Lond. B237, 37–72
(1952)
37. J.D. Murray, Lectures on Nonlinear Differential Equation-Models in Biology. (Clanderon
Press, Oxford, 1977)
38. P.C. Fife, Mathematical aspects of reacting and diffusing systems. (Springer, New York, 1979)
39. E. Convay, D. Hoff, J.A. Smoller, Large time behavior of systems of nonlinear diffusion
equations. SIAM J. Appl. Math. 35, 1–16 (1978)
40. D.A. Kessler, H. Levine, Fluctuation-induced diffusive instabilities. Nature 394, 556–558
(1998)
41. H. Zhonghuai, Y. Lingfa, X. Zuo, X. Houwen, Noise induced pattern transition and spatiotem-
poral stochastic resonance. Phys. Rev. Lett. 81, 2854–2857 (1998)
42. A. Rosato, K.J. Strandburg, F. Prinz, R.H. Swendsen, Why the Brazil nuts are on top: Size
segregation of particulate matter by shaking. Physical Review Letters 58, 1038–1041 (1987)
43. J.A.C. Gallas, H.J. Herrmann, S. Sokołowski, Convection cells in vibrating granular media.
Phys. Rev. Lett. 69, 1371–1374 (1992)
44. P.B. Umbanhowar, F. Melo, H.L. Swinney, Localized excitations in a vertically vibrated
granular layer. Nature 382, 793–796 (1996)
45. J. Keizer, Statistical Thermodynamics of Nonequilibrium Processes. (Springer, New York,
1987)
46. D. Helbing, Boltzmann-like and Boltzmann-Fokker-Planck equations as a foundation of

behavioral models. Physics A 196, 546–573 (1993)
47. S.B. Santra, S. Schwarzer, H. Herrmann, Fluid-induced particle-size segregation in sheared
granular assemblies. Phys. Rev. E 54, 5066–5072 (1996)
48. Ben-Jacob, E., O. Schochet, A. Tenenbaum, I. Cohen, A. Czirók, T. Vicsek, Generic modelling
of cooperative growth patterns in bacterial colonies. Nature 368, 46–49 (1994)
49. Ben-Jacob, E., From snowflake formation to growth of bacterial colonies, Part II: Cooperative
formation of complex colonial patterns. Contemp. Phys. 38, 205–241 (1997)
50. D.A. Kessler, H. Levine, Pattern formation in Dictyostelium via the dynamics of cooperative
biological entities. Phys. Rev. E 48, 4801–4804 (1993)
51. T. Vicsek, A. Czirók, E. Ben-Jacob, I. Cohen, O. Shochet, Novel type of phase transition in a
system of self-driven particles. Phys. Rev. Lett. 75, 1226–1229 (1995)
52. F. Schweitzer, K. Lao, F. Family, Active random walkers simulate trunk trail formation by ants.
BioSystems, 41, 153–166 (1997)
53. E.M. Rauch, M.M. Millonas, D.R. Chialvo, Pattern formation and functionality in swarm
models. Phys. Lett. A, 207, 185–193 (1995)
54. S. Wolfram, Cellular automata as models of complexity. Nature 311, 419–424 (1984)
55. D. Stauffer, Computer simulations of cellular automata. J. Phys. A: Math. Gen. 24, 909–927
(1991)
56. B.A. Huberman, N.S. Glance, Evolutionary games and computer simulations. Proc. Nat. Acad.
Sci. USA 90, 7716–7718 (1993)
57. T.C. Schelling, Dynamic models of segregation. J. Math. Sociol. 1, 143–186 (1971)
Chapter 6
Cooperation in Social Dilemmas
6.1 Introduction
Game theory goes back to von Neumann [1], one of the superminds of quantum
mechanics. Originally intended to describe interactions in economics, sociology,
and biology [1–4], it has recently become a quickly growing research area in
physics, where methods from non-linear dynamics and pattern formation [5–11],
agent-based or particle-like models [11–13], network theory [14–18] and statistical
physics [19–21] are applied. There are even quantum theoretical contributions [22].
When two entities characterized by the states, “strategies”, or “behaviors” i and
j interact with each other, game theory formalizes the result by payoffs Pij , and the
structure of the payoff matrix .Pij / determines the kind of the game. The dynamics
of a system of such entities is often delineated by the so-called replicator equations
X X
dp.i; t/
D p.i; t/ Pij p.j; t/ p.l; t/Plj p.j; t/ (6.1)
dt j j;l
[3, 4]. p.i; t/ represents the relative frequency P

of behavior i in the system, which
increases
P when the expected “success” Fi D j Pij p.j; t/ exceeds the average
one, i Fi p.i; t/.
Many collective phenomena in physics such as agglomeration or segregation
phenomena can be studied in a game-theoretical way [11–13]. Applications also
include the theory of evolution [23, 24] and the study of ecosystems [25–27].
Another exciting research field is the study of mechanisms supporting the coopera-
tion between selfish individuals [1–4] in situations like the “prisoner’s dilemma”

This chapter reprints a previous publication with kind permission of the Americal Physical
Society. It is requested to cite this work as follows: D. Helbing and S. Lozano (2010) Phase
transitions to cooperation in the prisoner’s dilemma. Physical Review E 81(5), 057102. DOI:
10.1103/PhysRevE.81.057102.

132 6 Cooperation in Social Dilemmas
or public goods game, where they would usually defect (free-ride or cheat).
Contributing to public goods and sharing them constitute ubiquitous situations,
where cooperation is crucial, for example, in order to maintain a sustainable use
of natural resources or a well-functioning health or social security system.
In the following, we will give an overview of the stationary solutions of the repli-
cator equations (6.1) and their stability properties. Based on this, we will discuss
several “routes to cooperation”, which transform the prisoner’s dilemma into other
games via different sequences of continuous or discontinuous phase transitions.
These routes will then be connected to different biological or social mechanisms
accomplishing such phase transitions [28]. Finally, we will introduce the concept
of “equilibrium creation” and distinguish it from routes to cooperation based
on “equilibrium selection” or “equilibrium displacement”. A new cooperation-
promoting mechanism based on adaptive group pressure will exemplify it.
6.2 Stability Properties of Different Games
Studying games with two strategies i only, the replicator equations (6.1) simplify,
and we remain with
dp.t/ ˚
D p.t/Œ1 p.t/ 1 Œ1 p.t/ 2 p.t/ ; (6.2)
dt
where p.t/ D p.1; t/ represents the fraction of cooperators and 1 p.t/ D p.2; t/
the fraction of defectors. 1 D P12 P22 and 2 D P21 P11 are the eigenvalues
of the two stationary solutions pDp1 D 0 and p D p2 D1 . If 0 < 1 =.1 C2 /<1,
there is a third stationary solution p D p3 D 1 =.1 C 2 / with eigenvalue
3 D .1 p3 /1 . For the sake of our discussion, we imagine an additional
fluctuation term .t/ on the right-hand-side of (6.2), reflecting small perturbations
of the strategy distribution.
Four different cases can be classified [3, 4]: (1) If 1 < 0 and 2 > 0, the
stationary solution p1 corresponding to defection by everybody is stable, while the
stationary solution p2 corresponding to cooperation by everyone is unstable. That is,
any small perturbation will drive the system away from full cooperation towards full
defection. This situation applies to the prisoner’s dilemma (PD) defined by payoffs
with P21 > P11 > P22 > P12 . According to this, strategy i D 1 (“cooperation”)
is risky, as it can yield the lowest payoff P12 , while strategy i D 2 (“defection”) is
tempting, since it can give the highest payoff P21 . (2) If 1 > 0 and 2 < 0, the
stationary solution p1 is unstable, while p2 is stable. This means that the system
will end up with cooperation by everybody. Such a situation occurs for the so-called
harmony game (HG) with P11 > P21 > P12 > P22 , as mutual cooperation gives
the highest payoff P11 . (3) If 1 > 0 and 2 > 0, the stationary solutions p1 and
p2 are unstable, but there exists a third stationary solution p3 , which turns out to be
stable. As a consequence, the system is driven towards a situation, where a fraction
6.3 Phase Transitions and Routes to Cooperation 133
p3 of cooperators is expected to coexist with a fraction .1 p3 / of defectors. Such a

situation occurs for the snowdrift game (SD) (also known as hawk-dove or chicken
game). This game is characterized by P21 > P11 > P12 > P22 and assumes that
unilateral defection is tempting, as it yields the highest payoff P21 , but also risky,
as mutual defection gives the lowest payoff P22 . (4) If 1 < 0 and 2 < 0, the
stationary solutions p1 and p2 are both stable, while the stationary solution p3 is
unstable. As a consequence, full cooperation is possible, but not guaranteed. In fact,
the final state of the system depends on the initial condition p.0/ (the “history”):
If p.0/ < p3 , the system is expected to end up in the stationary solution p1 , i.e.
with full defection. If p.0/ > p3 , the system is expected to move towards p2 D 1,
corresponding to cooperation by everybody. The history-dependence implies that
the system is multistable (here: bistable), as it has several (locally) stable solutions.
This case is found for the stag hunt game (SH) (also called assurance). This game is
characterized by P11 > P21 > P22 > P12 , i.e. cooperation is rewarding, as it gives
the highest payoff P11 in case of mutual cooperation, but it is also risky, as it yields
the lowest payoff P12 , if the interaction partner is uncooperative.
6.3 Phase Transitions and Routes to Cooperation
When facing a prisoner’s dilemma, it is of vital interest to transform the payoffs

in such a way that cooperation between individuals is supported. Starting with
the payoffs Pij0 of a prisoner’s dilemma, one can reach different payoffs Pij , for
example, by introducing strategy-dependent taxes Tij D Pij0 Pij > 0. When
increasing the taxes Tij from 0 to Tij0 , the eigenvalues will change from 01 D
0
P12 P22
0
and 02 D P210
P11
0
to 1 D 01 C T22 T12 and 2 D 02 C T11 T21 . In
this way, one can create a variety of routes to cooperation, which are characterized
by different kinds of phase transitions. We define route 1 [PD!HG] by a direct
transition from a prisoner’s dilemma to a harmony game. It is characterized by a
discontinuous transition from a system, in which defection by everybody is stable,
to a system, in which cooperation by everybody is stable (see Fig. 6.1a). Route 2
[PD!SH] is defined by a direct transition from the prisoner’s dilemma to a stag hunt
game. After the moment t , where 2 changes from positive to negative values, the
system behavior becomes history-dependent: When the fluctuations .t/ for t > t
exceed the critical threshold p3 .t/ D 1 =Œ1 C 2 .t/, the system will experience a
sudden transition to cooperation by everybody. Otherwise one will find defection by
everyone, as in the prisoner’s dilemma (see Fig. 6.1b). In order to make sure that the
perturbations .t/ will eventually exceed p3 .t/ and trigger cooperation, the value
of 2 must be reduced to sufficiently large negative values. It is also possible to
have a continuous rather than sudden transition to cooperation: We define route 3
[PD!SD] by a transition from a prisoner’s dilemma to a snowdrift game. As 1 is
changed from negative to positive values, a fraction p3 .t/ D 1 .t/=Œ1 .t/ C 2 of
cooperators is expected to result (see Fig. 6.1c). When increasing 1 , this fraction
a p b c
p p
1 1 1
COEX
DEFECT COOP DEFECT BISTAB DEFECT
PD HG PD PD SD
SH
0 Route 1 0 Route 2 0 Route 3
d p
e p
f p
1 1 1
COEX BISTAB COEX BISTAB
DEF. COOP DEF. COOP DEF COOP
PD HG PD HG PD HG
SD SH SD SH
0 Route 4 0 Route 5 0 Route 6
Fig. 6.1 Schematic illustration of the phase transitions defining the different routes to cooperation.
The order parameter is the stationary frequency of cooperators, while the control parameters are
the parameters r, w, k, m, or q in Nowak’s cooperation-enhancing rules [28] (see main text) or,
more generally, (non-)linear combination of the model parameters b and c. Solid lines represent
stable stationary proportions of cooperators, dashed lines unstable fix points. Diagonal lines show
the additional stationary solution p3 , where 0 p3 1 (p D proportion of cooperators; DEFECT
D defection is stable, i.e. everybody defects; COOP D cooperation is stable, i.e. everybody
cooperates; COEX D mixture of defectors with a proportion p3 of cooperators; BISTAB D
cooperation is stable if p3 < p.0/, where p.0/ means the initial proportion of cooperators,
otherwise everybody defects)
rises continuously. One may also implement more complicated transitions. Route 4,
for example, establishes the transition sequence PD!SD!HG (see Fig. 6.1d),
while we define route 5 by the transition PD!SH!HG (see Fig. 6.1e). One
may also implement the transition PD!SD!HG!SH (route 6, see Fig. 6.1f),
establishing a path-dependence, which can guarantee cooperation by everybody
in the end. (When using route 2, the system remains in a defective state, if the
perturbations do not exceed the critical value p3 .)
6.4 Relationship with Cooperation-Supporting Mechanisms
We will now discuss the relationship of the above introduced routes to cooperation
with biological and social mechanisms (“rules”) promoting the evolution of cooper-
ation. Martin A. Nowak performs his analysis of five such rules with the reasonable
specifications T D b > 0, R D b c > 0, S D c < 0, and P D 0 in the limit
of weak selection [28]. Cooperation is assumed to require a contribution c > 0 and
to produce a benefit b > c for the interaction partner, while defection generates no
payoff (P D 0). As most mechanisms leave 1 or D .1 C 2 /=2 unchanged,
6.5 Further Kinds of Transitions to Cooperation 135
we will now focus on the payoff-dependent parameters 1 and (rather than 1 and
2 ). The basic prisoner’s dilemma is characterized by 01 D c and 0 D 0.
According to the Supporting Online Material of [28], kin selection (genetic
relatedness) tranforms the payoffs into P11 D P11 0
C r.b c/, P12 D P12 0
C br,
P21 D P21 cr, and P22 D P22 . Therefore, it leaves unchanged and increases
0 0
1 by T22 T12 D br, where r represents the degree of genetic relatedness.

Direct reciprocity (repeated interaction) does not change 1 , but it reduces by
12 .b c/Œ1=.1 w/ 1 < 0, where w is the probability of a future interaction.
Network reciprocity (clustering of individuals playing the same strategy) leaves
unchanged and increases 1 by H.k/, where H.k/ is a function of the number k
of neighbors. Finally, group selection (competition between different populations)
increases 1 by .b c/.m 1/, where m is the number of groups, while is not
modified. However, 1 and may also change simultaneously. For example, indirect
reciprocity (based on trust and reputation) increases 1 by cq and reduces by
12 .b c/q < 0, where q quantifies social acquaintanceship.
Summarizing this, kin selection, network reciprocity, and group selection pre-
serve D 0 and increase the value of 1 (see route 1 in Fig. 6.2). Direct
reciprocity, in contrast, preserves the value of 1 and reduces (see route 2a in
Fig. 6.2). Indirect reciprocity promotes the same transition (see route 2b in Fig. 6.2).
Supplementary, one can analyze costly punishment. Using the payoff specifications
made in the Supporting Information of [29], costly punishment changes by
.ˇ C /=2 < 0 and 1 by [29], i.e. when is increased, the values of
and 1 are simultaneously reduced (see route 2c in Fig. 6.2). Here, > 0 represents
the punishment cost invested by a cooperator to impose a punishment fine ˇ > 0 on
a defector, which decreases the payoffs of both interaction partners. Route 3 can be
generated by the formation of friendship networks [30]. Route 4 may occur by kin
selection, network reciprocity, or group selection, when starting with a prisoner’s
dilemma with 0 < 0 (rather than 0 D 0 as assumed before). Route 5 may be
generated by the same mechanisms, if 0 > 0. Finally, route 6 can be implemented
by time-dependent taxation (see Fig. 6.2).
6.5 Further Kinds of Transitions to Cooperation
The routes to cooperation discussed so far change the eigenvalues 1 and 2 ,

and leave the stationary solutions p1 and p2 unchanged. However, transitions
to cooperation can also be generated by shifting the stationary solutions or
creating new ones, as we will show now. For this, we generalize the replicator
equation (6.2) by replacing 1 with f .p/ and with g.p/, and by adding a
term h.p/, which can describe effects of spontaneous transitions like mutations.
To guarantee 0 p.t/ 1, we must have h.p/ D v.p/ pw.p/ with
functions
w.p/ v.p/ 0. The resulting equation is dp=dt D F .p.t// with
F p D .1 p/Œf .p/ 2g.p/pp C h.p/, and its stationary solutions pk are
given by F .pk / D .1 pk /Œf .pk / 2g.pk /pk pk C h.pk / D 0. The associated
λ1
λ1 = 2λ
4
1
COEX
COOP 5
SD
HG
3
6 0 λ
DEFECT
2b PD
2a p
1
2c
BISTAB COEX
DEFECT BISTAB
SH DEFECT
PD
0 K0 K
Fig. 6.2 Phase diagram of expected system behaviors, based on an analysis of the game-dynamical
replicator equation (6.2) as a function of the parameters and 1 . The different routes to
cooperation are illustrated by arrows. Terms in capital letters are defined in Fig. 6.1. Inset: Stable
stationary solutions (solid lines) and unstable ones (broken lines) as functions of the parameter K,
when the reward depends on the proportion of cooperators. The bifurcation at the “tipping point”
K D K0 “inverts” the system behavior (see main text)
eigenvalues k D dF .pk /=dp determining the stability of the stationary solutions

pk are
k D .1 2pk /.fk 2pk gk / C pk .1 pk /.fk0 2pk gk0 2gk / C h0k ;
where we have used the abbreviations fk D f .pk /, gk D g.pk /, hk D h.pk /.

fk0 D f 0 .pk /, gk0 D g 0 .pk / and hk D h0 .pk / are the derivatives of the functions
f .p/, g.p/ and h.p/ in the points p D pk .
Classification. We can now distinguish different kinds of transitions from defec-
tion to cooperation: If the stationary solutions p1 D 0 and p2 D 1 of the prisoner’s
dilemma are modified, we talk about transitions to cooperation by equilibrium
displacement. This case occurs, for example, when random mutations are not weak
(h ¤ 0). If the eigenvalues 1 or 2 of the stationary solutions p1 D 0 and p2 D 1
are changed, we speak of equilibrium selection. This case applies to all routes to
cooperation discussed before. If a new stationary solution appears, we speak of
equilibrium creation. The different cases often appear in combination with each
other (see the Summary below). In the following, we will discuss an interesting case,
where cooperation occurs solely through equilibrium creation, i.e. the stationary
solutions p1 and p2 of the replicator equation for the prisoner’s dilemma as well as
their eigenvalues 1 and 2 remain unchanged. We illustrate this by the example of
6.6 Summary 137
an adaptive kind of group pressure that rewards mutual cooperation (T11 < 0) or
sanctions unilateral defection (T21 > 0). Both, rewarding and sanctioning reduces
the value of 2 , while 1 remains unchanged. Assuming here that the group pressure
vanishes, when everybody cooperates (as it is not needed then), while it is maximum
when everybody defects (to encourage cooperation) [31], we may set f .p/ D 01
and g.p/ D 0 KŒ1 p.t/, corresponding to 2 .t/ D 02 2KŒ1 p.t/. It is
obvious that we still have the two stationary solutions p1 D 0 and p2 D 1 with the
eigenvalues 1 D 01 < 0 and 2 D 20 01 > 0 of the original prisoners dilemma
with parameters 01 and 02 orq0 . However, for large enough values of K [namely
for K > K0 D 0 C j01 j C j01 j.20 C j01 j/], we find two additional stationary
solutions s

1 0 1 0 2 j01 j
p˙ D ˙ : (6.3)
2 2K 2 2K 2K
p is an unstable stationary solution with p1 < p < pC and D dF .p /=dp > 0,
while pC is a stable stationary solution with p < pC < p2 and C D
dF .pC /=dp < 0 (see inset of Fig. 6.2). Hence, the assumed dependence of the
payoffs on the proportion p of cooperators generates a bistable situation (BISTAB),
with the possibility of a coexistence of a few defectors with a large proportion pC of
cooperators, given K > K0 . If p.0/ < p , where p.0/ denotes the initial condition,
defection by everybody results, while a stationary proportion pC of cooperators is
established for p < p.0/ < 1. Surprisingly, in the limit K ! 1, cooperation is
established for any initial condition p.0/ ¤ 0 (or through fluctuations).
6.6 Summary
We have discussed from a physical point of view what must happen that social or
biological, payoff-changing interaction mechanisms can create cooperation in the
prisoner’s dilemma. The possible ways are (1) moving the stable stationary solution
away from pure defection (routes 3, 4, and 6), (2) stabilizing the unstable solution
(routes 1, 2, 4, 5 and 6), or (3) creating new stationary solutions, which are stable
(routes 3, 4 and 6). Several of these points can be combined. If (1) is fulfilled, we
speak of “equilibrium displacement”, if their eigenvalues change, we called this
“equilibrium selection”, and if (3) is the case, we talk of “equilibrium creation”.
The first case can result from mutations, the second one applies to many social or
biological cooperation-enhancing mechanisms [28]. We have discussed an interest-
ing case of equilibrium creation, in which the outcome of the replicator equation is
changed, although the stationary solutions of the PD and their eigenvalues remained
unchanged. This can, for example, occur by adaptive group pressure [31], which
introduces an adaptive feedback mechanism and thereby increases the order of non-
linearity of the replicator equation. Surprisingly, already a linear dependence of the
payoff values Pij on the endogeneous dynamics p.t/ of the system is enough to
destabilize defection and stabilize cooperation, thereby inverting the outcome of the
prisoner’s dilemma.
Acknowledgements This work was partially supported by the Future and Emerging Technologies
programme FP7-COSI-ICT of the European Commission through the project QLectives (grant no.:
231200).
References
1. J. von Neumann, O. Morgenstern, Theory of Games and Economic Behavior (Princeton

University, Princeton, 1944)
2. R. Axelrod, The Evolution of Cooperation (Basic, New York, 1984)
3. J. Hofbauer, K. Sigmund, Evolutionary Games and Population Dynamics (Cambridge
4. J.W. Weibull, Evolutionary Game Theory (MIT Press, Cambridge, MA, 1996)
5. N.F. Johnson, P.M. Hui, R. Jonson, T.S. Lo, Phys. Rev. Lett. 82, 3360 (1999)
6. D. Challet, M. Marsili, R. Zecchina, Phys. Rev. Lett. 84, 1824 (2000)
7. G. Szabó, C. Hauert, Phys. Rev. Lett. 89, 118101 (2002)
8. C. Hauert, M. Doebeli, Nature 428, 643 (2004)
9. J.C. Claussen, A. Traulsen, Phys. Rev. Lett. 100, 058104 (2008)
10. C.P. Roca, J.A. Cuesta, A. Sánchez, Phys. Rev. Lett. 97, 158701 (2006)
11. D. Helbing, W. Yu, PNAS 106, 3680 (2009)
12. D. Helbing, T. Vicsek, New J. Phys. 1, 13 (1999)
13. D. Helbing, T. Platkowski, Europhys. Lett. 60, 227 (2002)
14. G. Szabó, G. Fath, Phys. Rep. 446, 97 (2007)
15. J.M. Pacheco, A. Traulsen, M.A. Nowak, Phys. Rev. Lett. 97, 258103 (2006)
16. F.C. Santos, J.M. Pacheco, Phys. Rev. Lett. 95, 098104 (2005)
17. J. Gómez-Gardeñes, M. Campillo, L.M. Florı́a, Y. Moreno, Phys. Rev. Lett. 98, 108103 (2007)
18. S. VanSegbroeck, F.C. Santos, T. Lenaerts, J.M. Pacheco, Phys. Rev. Lett. 102, 058105 (2009)
19. J. Berg, A. Engel, Phys. Rev. Lett. 81, 4999 (1998)
20. A. Traulsen, J.C. Claussen, C. Hauert, Phys. Rev. Lett. 95, 238701 (2005)
21. H. Ohtsuki, M.A. Nowak, J.M. Pacheco, Phys. Rev. Lett. 98, 108106 (2007)
22. J. Eisert, M. Wilkens, M. Lewenstein, Phys. Rev. Lett. 83, 3077 (1999)
23. M. Eigen, P. Schuster, The Hypercycle (Springer, Berlin, 1979)
24. R.A. Fisher, The Genetical Theory of Natural Selection (Oxford University Press, Oxford,
1930)
25. M. Opper, S. Diederich, Phys. Rev. Lett. 69, 1616 (1992)
26. V.M. de Oliveira, J.F. Fontanari, Phys. Rev. Lett. 89, 148101 (2002)
27. J.Y. Wakano, M.A. Nowak, C. Hauert, PNAS 106, 19 (2009)
28. M.A. Nowak, Science 314, 1560 (2006)
29. A. Traulsen, C. Hauert, H. De Silva, M.A. Nowak, K. Sigmund, PNAS 106(3), 709 (2009)
30. H. Ohtsuki, M.A. Nowak, J. Theor. Biol. 243, 86–97 (2006)
31. O. Gurerk, B. Irlenbusch, B. Rockenbach, Science 312, 108–111 (2006)
Chapter 7
Co-evolution of Social Behavior and Spatial
Organization
7.1 Introduction
While the availability of new data of human mobility has revealed relations with
social communication patterns [1] and epidemic spreading [2], its significance for
the cooperation among individuals is still widely unknown. This is surprising,
as migration is a driving force of population dynamics as well as urban and
interregional dynamics [3–5].
Below, we model cooperation in a game-theoretical way [6–8], and integrate a
model of stylized relocations. This is motivated by the observation that individuals
prefer better neighborhoods, e.g. a nicer urban quarter or a better work environment.
To improve their situation, individuals are often willing to migrate. In our model of
success-driven migration, individuals consider different alternative locations within
a certain migration range, reflecting the effort they are willing or able to spend on
identifying better neighborhoods. How favorable a new neighborhood is expected
to be is determined by test interactions with individuals in that area (“neighborhood
testing”). The related investments are often small compared to the potential gains
or losses after relocating, i.e. exploring new neighborhoods is treated as “fictitious
play”. Finally, individuals are assumed to move to the tested neighborhood that
promises to be the best.
So far, the role of migration has received relatively little attention in game theory
[9–16], probably because it has been found that mobility can undermine cooperation
by supporting defector invasion [11, 12]. However, this primarily applies to cases,
where individuals choose their new location in a random (e.g. diffusive) way. In
contrast, extending spatial games by the specific mechanism of success-driven
migration can support the survival and spreading of cooperation. As we will show,

This chapter reprints a previous publication with kind permission of the National Academy of
Sciences of the USA. It is requested to cite this work as follows: D. Helbing and W. Yu, The
outbreak of cooperation among success-driven individuals under noisy conditions. Proceedings of
the National Academy of Sciences USA 106(8), 3680–3685 (2009).

140 7 Co-evolution of Social Behavior and Spatial Organization
it even promotes the spontaneous outbreak of prevalent cooperation in a world

of selfish individuals with various sources of randomness (“noise”), starting with
defectors only.
7.2 Model
Our study is carried out for the prisoner’s dilemma game (PD). This has often
been used to model selfish behavior of individuals in situations, where it is risky
to cooperate and tempting to defect, but where the outcome of mutual defection is
inferior to cooperation on both sides [7, 17]. Formally, the so-called “reward” R
represents the payoff for mutual cooperation, while the payoff for defection on both
sides is the “punishment” P . T represents the “temptation” to unilaterally defect,
which results in the “sucker’s payoff” S for the cooperating individual. Given the
inequalities T > R > P > S and 2R > T CS , which define the classical prisoner’s
dilemma, it is more profitable to defect, no matter what strategy the other individual
selects. Therefore, rationally behaving individuals would be expected to defect when
they meet once. However, defection by everyone is implied as well by the game-
dynamical replicator equation [10], which takes into account imitation of superior
strategies, or payoff-driven birth-and-death processes. In contrast, a coexistence of
cooperators and defectors is predicted for the snowdrift game (SD). While it is also
used to study social cooperation, its payoffs are characterized by T > R > S > P
(i.e. S > P rather than P > S ).
As is well-known [17], cooperation can, for example, be supported by repeated
interactions [7], by intergroup competition with or without altruistic punishment
[18–20], and by network reciprocity based on the clustering of cooperators [21–23].
In the latter case, the level of cooperation in two-dimensional spatial games is further
enhanced by “disordered environments” (approximately 10% unaccessible empty
locations) [24], and by diffusive mobility, provided that the mobility parameter
is in a suitable range [16]. However, strategy mutations, random relocations, and
other sources of stochasticity (“noise”) can significantly challenge the formation
and survival of cooperative clusters. When no mobility or undirected, random
mobility are considered, the level of cooperation in the spatial games studied by
us is sensitive to noise (see Figs. 7.1d and 7.3c), as favorable correlations between
cooperative neighbors are destroyed. Success-driven migration, in contrast, is a
robust mechanism: By leaving unfavorable neighborhoods, seeking more favorable
ones, and remaining in cooperative neighborhoods, it supports cooperative clusters
very efficiently against the destructive effects of noise, thus preventing defector
invasion in a large area of payoff parameters. We assume N individuals on a
square lattice with periodic boundary conditions and L L sites, which are either
empty or occupied by one individual. Individuals are updated asynchronously, in a
random sequential order. The randomly selected individual performs simultaneous
interactions with the m D 4 direct neighbors and compares the overall payoff with
that of the m neighbors. Afterwards, the strategy of the best performing neighbor
7.2 Model 141
a b c
No noise
d e f
Noise 1
Imitation only Migration only Imitation and migration
Fig. 7.1 Representative simulation results for the spatial prisoner’s dilemma with payoffs
T D 1:3, R D 1, P D 0:1, and S D 0 after t D 200 iterations. The simulations are for
49 49-grids with 50% empty sites. At time t D 0 we assumed 50% of the individuals to be
cooperators and 50% defectors. Both strategies were homogeneously distributed over the whole
grid. For reasons of comparison, all simulations were performed with identical initial conditions
and random numbers (red D defector, blue D cooperator, white D empty site, green D defector
who became a cooperator in the last iteration, yellow D cooperator who turned into a defector).
Compared to simulations without noise (top), the strategy mutations of noise 1 with r D q D 0:05
do not only reduce the resulting level of cooperation, but also the outcome and pattern formation
dynamics, even if the payoff values, initial conditions, and update rules are the same (bottom):
In the imitation-only case with M D 0 that is displayed on the left, the initial fraction of 50%
cooperators is quickly reduced due to imitation of more successful defectors. The result is a
“frozen” configuration without any further strategy changes. (a) In the noiseless case, a certain
number of cooperators can survive in small cooperative clusters. (d) When noise 1 is present,
random strategy mutations destroy the level of cooperation almost completely, and the resulting
level of defection reaches values close to 100%. The illustrations in the center show the migration-
only case with mobility range M D 5: (b) When no noise is considered, small cooperative clusters
are formed, and defectors are primarily located at their boundaries. (e) In the presence of noise 1,
large clusters of defectors are formed instead, given P > 0. The illustrations on the right show
the case, where imitation is combined with success-driven migration (here, M D 5): (d) In the
noiseless case, cooperative clusters grow and eventually freeze (i.e. strategy changes or relocations
do not occur any longer). (f) Under noisy conditions, in contrast, the cooperative clusters continue
to adapt and reconfigure themselves, as the existence of yellow and green sites indicates
is copied with probability 1 r (“imitation”), if the own payoff was lower. With
probability r, however, the strategy is randomly “reset”: Noise 1 assumes that an
individual spontaneously chooses to cooperate with probability q or to defect with
probability 1 q until the next strategy change. The resulting strategy mutations
may be considered to reflect deficient imitation attempts or trial-and-error behavior.
As a side effect, such noise leads to an independence of the finally resulting level
of cooperation from the initial one at t D 0, and a qualitatively different pattern
formation dynamics for the same payoff values, update rules, and initial conditions
(see Fig. 7.1). Using the alternative Fermi update rule [22] would have been possible
as well. However, resetting strategies rather than inverting them, combined with
values q much smaller than 1/2, has here the advantage of creating particularly
adverse conditions for cooperation, independently of what strategy prevails. Below,
we want to learn, if predominant cooperation can survive or even emerge under such
adverse conditions.
“Success-driven migration” has been implemented as follows [9, 25]: Before the
imitation step, an individual explores the expected payoffs for the empty sites in the
migration neighborhood of size .2M C 1/ .2M C 1/ (the Moore neighborhood of
range M ). If the fictitious payoff is higher than in the current location, the individual
is assumed to move to the site with the highest payoff and, in case of several sites
with the same payoff, to the closest one (or one of them); otherwise it stays put.
7.3 Results
Computer simulations of the above model show that, in the imitation-only case of
classical spatial games with noise 1, but without a migration step, the resulting
fraction of cooperators in the PD tends to be very low. It basically reflects the
fraction rq of cooperators due to strategy mutations. For r D q D 0:05, we find
almost frozen configurations, in which only a small number of cooperators survive
(see Fig. 7.1d). In the migration-only case without an imitation step, the fraction
of cooperators changes only by strategy mutations. Even when the initial strategy
distribution is uniform, one observes the formation of spatio-temporal patterns, but
the patterns get almost frozen after some time (see Fig. 7.1e).
It is interesting that, although for the connectivity structure of our PD model
neither imitation only (Fig. 7.1d) nor migration only (Fig. 7.1e) can promote
cooperation under noisy conditions, their combination does: Computer simulations
show the formation of cooperative clusters with a few defectors at their boundaries
(see Fig. 7.1f). Once cooperators are organized in clusters, they tend to have more
neighbors and to reach higher payoffs on average, which allows them to survive
[9, 10, 25]. It will now have to be revealed, how success-driven migration causes the
formation of clusters at all, considering the opposing noise effects. In particular, we
will study, why defectors fail to invade cooperative clusters and to erode them from
within, although a cooperative environment is most attractive to them.
7.3 Results 143
To address these questions, Fig. 7.2 studies a “defector’s paradise” with a

single defector in the center of a cooperative cluster. In the noisy imitation-only
spatial prisoner’s dilemma, defection tends to spread up to the boundaries of the
cluster, as cooperators imitate more successful defectors (see Figs. 7.2a–d). But
if imitation is combined with success-driven migration, the results are in sharp
contrast: Although defectors still spread initially, cooperative neighbors who are
M steps away from the boundary of the cluster can now evade them. Due to this
defector-triggered migration, the neighborhood reconfigures itself adaptively. For
example, a large cooperative cluster may split up into several smaller ones (see
Figs. 7.2e–h). Eventually, the defectors end up at the boundaries of these cooperative
clusters, where they often turn into cooperators by imitation of more successful
cooperators in the cluster, who tend to have more neighbors. This promotes the
spreading of cooperation [9, 10, 25]. Since evasion takes time, cooperative clusters
could still be destroyed when continuously challenged by defectors, as it happens
under noisy conditions. Therefore, let us now study the effect of different kinds of
randomness [10,26]. Noise 1 (defined above) assumes strategy mutations, but leaves
a b c d
Imitation only
t=0 t=5 t = 20 t = 200
e f g h
With migration
t=0 t=5 t = 20 t = 200
Fig. 7.2 Representative simulation results after t D 200 iterations in the “defector’s paradise”
scenario, starting with a single defector in the center of a cooperative cluster at t D 0. The
simulations are performed on 49 49-grids with N D 481 individuals, corresponding to a circle
of diameter 25. They are based on the spatial prisoner’s dilemma with payoffs T D 1:3, R D 1,
P D 0:1, S D 0 and noise parameters r D q D 0:05 (red D defector, blue D cooperator,
white D empty site, green D defector who became a cooperator, yellow D cooperator who turned
into a defector in the last iteration). For reasons of comparison, all simulations were carried out
with identical initial conditions and random numbers. (a–d) In the noisy imitation-only case with
M D 0, defection (red) eventually spreads all over the cluster. The few remaining cooperators
(blue) are due to strategy mutations. (e–h) When we add success-driven motion, the result is very
different. Migration allows cooperators to evade defectors. That triggers a splitting of the cluster,
and defectors end up on the surface of the resulting smaller clusters, where most of them can be
turned into cooperators. This mechanism is crucial for the unexpected survival and spreading of
cooperators
the spatial distribution of individuals unchanged (see Fig. 7.3a). Noise 2, in contrast,
assumes that individuals, who are selected with probability r, move to a randomly
chosen free site without considering the expected success (random relocations).
Such random moves may potentially be of long distance and preserve the number
of cooperators, but have the potential of destroying spatial patterns (see Fig. 7.3b).
Noise 3 combines noise 1 and noise 2, assuming that individuals randomly relocate
with probability r and additionally reset their strategy as in noise 1 (see Fig. 7.3c).
While cooperation in the imitation-only case is quite sensitive to noise (see
Figs. 7.3a–c), the combination of imitation with success-driven motion is not (see
Fig. 7.3d–f): Whenever an empty site inside a cluster of cooperators occurs, it is
more likely that the free site is entered by a cooperator than by a defector, as long
as cooperators prevail within the migration range M . In fact, the formation of small
cooperative clusters was observed for all kinds of noise. That is, the combination of
imitation with success-driven migration is a robust mechanism to maintain and even
spread cooperation under various conditions, given there are enough cooperators in
the beginning.
It is interesting, whether this mechanism is also able to facilitate a spontaneous
outbreak of predominant cooperation in a noisy world dominated by selfishness,
without a “shadow of the future” [7,27]. Our simulation scenario assumes defectors
only in the beginning (see Fig. 7.4a), strategy mutations in favor of defection,
and short-term payoff-maximizing behavior in the vast majority of cases. In order
to study conditions under which a significant fraction of cooperators is unlikely,
our simulations are performed with noise 3 and r D q D 0:05, as it tends to
destroy spatial clusters and cooperation (see Fig. 7.3c): By relocating 5% randomly
chosen individuals in each time step, noise 3 dissolves clusters into more or
less separate individuals in the imitation-only case (see Figs. 7.3b+c). In the case
with success-driven migration, random relocations break up large clusters into
many smaller ones, which are distributed all over the space (see Figs. 7.3b+c
and 7.4b). Therefore, even the clustering tendency by success-driven migration
can only partially compensate for the dispersal tendency by random relocations.
Furthermore, the strategy mutations involved in noise 3 tend to destroy cooperation
(see Figs. 7.3a+c, where the strategies of 5% randomly chosen individuals were
replaced by defection in 95% of the cases and by cooperation otherwise, to create
conditions favoring defection, i.e. the dominant strategy in the prisoner’s dilemma).
Overall, as a result of strategy mutations (i.e. without the consideration of imitation
processes), only a fraction rq D 0:0025 of all defectors turn into cooperators in each
time step, while a fraction r.1 q/ 0:05 of all cooperators turn into defectors
(i.e. 5% in each time step). This setting is extremely unfavorable for the spreading
of cooperators. In fact, defection prevails for an extremely long time (see Figs. 7.4b
and 7.5a). But suddenly, when a small, supercritical cluster of cooperators has
occurred by coincidence (see Fig. 7.4c), the fraction of cooperators spreads quickly
(see Fig. 7.5a), and soon cooperators prevail (see Figs. 7.4d and 7.5b). Note that this
spontaneous birth of predominant cooperation in a world of defectors does not occur
in the noisy imitation-only case and demonstrates that success-driven migration can
overcome the dispersive tendency of noises 2 and 3, if r is moderate and q has a
7.3 Results 145
a b c
Imitation only
t = 200
Noise 1, M = 0 Noise 2, M = 0 Noise 3, M = 0

d e f
With migration
t = 200
Noise 1, M = 5 Noise 2, M = 5 Noise 3, M = 5
Fig. 7.3 Representative simulation results for the invasion scenario with a defector in the center of
a cooperative cluster (“defector’s paradise”). The chosen payoffs T D 1:3, R D 1, P D 0:1, and
S D 0 correspond to a prisoner’s dilemma. The simulations are for 49 49-grids with N D 481
individuals, corresponding to a circle of diameter 25 (red D defector, blue D cooperator, white D
empty site, green D defector who became a cooperator, yellow D cooperator who turned into a
defector in the last iteration). Top: Typical numerical results for the imitation-only case (M D 0)
after t D 200 iterations (a) for noise 1 (strategy mutations) with mutation rate r D 0:05 and
creation of cooperators with probability q D 0:05, (b) for noise 2 (random relocations) with
relocation rate r D 0:05, and (c) for noise 3 (a combination of random relocations and strategy
mutations) with r D q D 0:05. As cooperators imitate defectors with a higher overall payoff,
defection spreads easily. The different kinds of noise influence the dynamics and resulting patterns
considerably: While strategy mutations in (a) and (c) strongly reduce the level of cooperation,
random relocations in (b) and (c) break up spatial clusters, leading to a dispersion of individuals in
space. Their combination in case (c) essentially destroys both, clusters and cooperation. Bottom:
Same for the case of imitation and success-driven migration with mobility range M D 5 (d) for
noise 1 with r D q D 0:05, (e) for noise 2 with r D 0:05, and (f) for noise 3 with r D q D 0:05.
Note that noise 1 just mutates strategies and does not support a spatial spreading, while noise
2 causes random relocations, but does not mutate strategies. This explains why the clusters in
Fig. 7.3d do not spread out over the whole space and why no new defectors are created in Fig. 7.3e.
However, the creation of small cooperative clusters is found in all three scenarios. Therefore, it is
robust with respect to various kinds of noise, in contrast to the imitation-only case
a b c d
With migration
t=0 t = 5000 t = 19140 t = 40000
Fig. 7.4 Spontaneous outbreak of prevalent cooperation in the spatial prisoner’s dilemma with
payoffs T D 1:3, R D 1, P D 0:1, S D 0 in the presence of noise 3 (random relocations
and strategy mutations) with r D q D 0:05. The simulations are for 49 49-grids (red D
defector, blue D cooperator, white D empty site, green D defector who became a cooperator,
yellow D cooperator who turned into a defector in the last iteration). (a) Initial cluster of defectors,
which corresponds to the final stage of the imitation-only case with strategy mutations according to
noise 1 (see Fig. 7.2d). (b) Dispersal of defectors by noise 3, which involves random relocations.
A few cooperators are created randomly by strategy mutations with the very small probability
rq D 0:0025 (0.25%). (c) Occurrence of a supercritical cluster of cooperators after a very
long time. This cooperative “nucleus” originates by random coincidence of favorable strategy
mutations in neighboring sites. (d) Spreading of cooperative clusters in the whole system. This
spreading despite the destructive effects of noise requires an effective mechanism to form growing
cooperative clusters (such as success-driven migration) and cannot be explained by random
coincidence. See the supplementary video for an animation of the outbreak of cooperation for
a different initial condition
finite value. That is, success-driven migration generates spatial correlations between
cooperators more quickly than these noises can destroy them. This changes the
outcome of spatial games essentially, as a comparison of Figs. 7.2a–d with 7.4a–
d shows.
The conditions for the spreading of cooperators from a supercritical cluster
(“nucleus”) can be understood by configurational analysis [26, 28] (see Fig. 7.1),
but the underlying argument can be both, simplified and extended: According to
Fig. 7.6a, the level of cooperation changes when certain lines (or, more generally,
certain hyperplanes) in the payoff-parameter space are crossed. These hyperplanes
are all of the linear form
n1 R C n2 S D n3 T C n4 P; (7.1)
where nk 2 f0; 1; 2; 3; 4g. The left-hand side of )7.1) represents the payoff of the
most successful cooperative neighbor of a focal individual, assuming that this has
n1 cooperating and n2 defecting neighbors, which implies n1 C n2 m D 4.
The right-hand side reflects the payoff of the most successful defecting neighbor,
assuming that n3 is the number of his/her cooperating neighbors and n4 the number
of defecting neighbors, which implies n3 C n4 m D 4. Under these conditions,
the best-performing cooperative neighbor earns a payoff of n1 R C n2 S , and the
best-performing defecting neighbor earns a payoff of n3 T C n4 P . Therefore, the
focal individual will imitate the cooperator, if n1 R C n2 S > n3 T C n4 P , but copy
7.3 Results 147
a R = 1, T = 1.3, P = 0.1, S = 0, r = q = 0.05, M = 5

1
0.9
0.8
0.7
Fraction of Cooperators
0.6
350
Distances of Cooperators
0.5 300
Total Migratory
250
0.4 200
150
0.3
100
0.2 50
0
0 1 2 3 4 5
0.1 Iteration x 104
0
0 1 2 3 4 5
Iteration x 104
b 1
r = 0.05, q = 0.05
0.8
Fraction of Cooperators
0.6
r = 0.1, q = 0.05
0.4
0.2
0
0 0.5 1 1.5 2
Iteration x 105
Fig. 7.5 Representative example for the outbreak of predominant cooperation in the prisoner’s
dilemma with payoffs T D 1:3, R D 1, P D 0:1, S D 0, in the presence of noise 3 with
r D q D 0:05. The simulations are for 49 49-grids with a circular cluster of defectors and no
cooperators in the beginning (see Fig. 7.4a). (a) After defection prevails for a very long time (here
for almost 20,000 iterations), a sudden transition to a large majority of cooperators is observed.
Inset: The overall distance moved by all individuals during one iteration has a peak at the time
when the outbreak of cooperation is observed. Before, the rate of success-driven migration is very
low, while it stabilizes at an intermediate level afterwards. This reflects a continuous evasion of
cooperators from defectors and, at the same time, the continuous effort to form and maintain
cooperative clusters. The graph displays the amount of success-driven migration only, while the
effect of random relocations is not shown. (b) Evaluating 50 simulation runs, the error bars
(representing three standard deviations) show a large variation of the time points when prevalent
cooperation breaks out. Since this time point depends on the coincidence of random cooperation
in neighboring sites, the large error bars have their natural reason in the stochasticity of this
process. After a potentially very long time period, however, all systems end up with a high level
of cooperation. The level of cooperation decreases with the noise strength r, as expected, but
moderate values of r can even accelerate the transition to predominant cooperation. Using the
parameter values r D 0:1 and q D 0:2, the outbreak of prevalent cooperation takes often less than
200 iterations
b 1
R = 1, T = 1.3, r = q = 0.05, ρ = 0.5
Snowdrift Game
0.5
M=0
0
S
M=1
−0.5 M=2
Prisoner’s Dilemma
M=5
–1
–1 –0.5 0 0.5 1
P
Fig. 7.6 Dependence of the fraction of cooperators for given payoff parameters T D 1:3 and
R D 1 on the parameters P and S. The area above the solid diagonal line corresponds to the
snowdrift game, the area below to the prisoner’s dilemma. Our simulations were performed for
grids with L L D 99 99 sites and N D L2 =2 individuals, corresponding to a density
D N=L2 D 0:5. At time t D 0 we assumed 50% of the individuals to be cooperators and
50% defectors. Both strategies were homogeneously distributed over the whole grid. The finally
resulting fraction of cooperators was averaged at time t D 200 over 50 simulation runs with
different random realizations. The simulations were performed with noise 3 (random relocations
with strategy mutations) and r D p D 0:05. An enhancement in the level of cooperation (often
by more than 100%) is observed mainly in the area with P 0:4 < S < P C 0:4 and P < 0:7.
Results for the noiseless case with r D 0 are shown in Fig. 7.5. (a) The fraction of cooperators is
represented by color codes (see the bar to the right of the figure, where dark orange, for example,
corresponds to 80% cooperators). It can be seen that the fraction of cooperators is approximately
constant in areas limited by straight lines (mostly triangular and rectangular ones). These lines
correspond to (7.1) for different specifications of n1 , n2 , n3 , and n4 (see main text for details).
(b) The light blue area reflects the parameters for which cooperators reach a majority in the
imitation-only case with M D 0: For all payoffs P and S corresponding to a prisoner’s dilemma,
cooperators are clearly in the minority, as expected. However, taking into account success-driven
migration changes the situation in a pronounced way: For a mobility range M D 1, the additional
area with more than 50% cooperators is represented by dark blue, the further extended area of
prevailing cooperation for M D 2 by green color, and for M D 5 in yellow. If M D 5, defectors
are in the majority only for parameter combinations falling into the red area. This demonstrates
that success-driven migration can promote predominant cooperation in considerable areas, where
defection would prevail without migration. For larger interaction neighborhoods m, e.g. m D 8,
the area of prevalent cooperation is further increased overall (not shown). Note that the irregular
shape of the separating lines is no artefact of the computer simulation or initial conditions. It results
by superposition of the areas defined by (7.1), see Fig. 7.6a
7.4 Discussion 149
the strategy of the defector if n1 R C n2 S < n3 T C n4 P . Equation (7.1) is the

line separating the area where cooperators spread (above the line) from the area of
defector invasion (below it) for a certain spatial configuration of cooperators and
defectors (see Fig. 7.6a). Every spatial configuration is characterized by a set of nk -
parameters. As expected, the relative occurrence frequency of each configuration
depends on the migration range M (see Fig. 7.6b): Higher values of M naturally
create better conditions for the spreading of cooperation, as there is a larger choice
of potentially more favorable neighborhoods.
Figure 7.6b also shows that success-driven migration extends the parameter
range, in which cooperators prevail, from the parameter range of the snowdrift
game with S > P to a considerable parameter range of the prisoner’s dilemma.
For this to happen, it is important that the attraction of cooperators is mutual, while
the attraction of defectors to cooperators is not. More specifically, the attraction
of cooperators is proportional to 2R, while the attraction between defectors and
cooperators is proportional to T CS . The attraction between cooperators is stronger,
because the prisoner’s dilemma usually assumes the inequality 2R > T C S .
Besides the speed of finding neighbors to interact with, the time scales of
configurational changes and correlations matter as well: By entering a cooperative
cluster, a defector triggers an avalanche of strategy changes and relocations, which
quickly destroys the cooperative neighborhood. During this process, individuals
may alter their strategy many times, as they realize opportunities by cooperation
or defection immediately. In contrast, if a cooperator joins a cooperative cluster,
this will stabilize the cooperative neighborhood. Although cooperative clusters
continuously adjust their size and shape, the average time period of their existence
is longer than the average time period after which individuals change their strategy
or location. This coevolution of social interactions and strategic behavior reflects
features of many social environments: While the latter come about by individual
actions, a suitable social context can make the average behavior of individuals more
predictable, which establishes a reinforcement process. For example, due to the
clustering tendency of cooperators, the likelihood of finding another cooperator in
the neighborhood of a cooperator is greater than 1/2, and also the likelihood that a
cooperator will cooperate in the next iteration.
7.4 Discussion
It is noteworthy that all the above features – the survival of cooperation in a

large parameter area of the PD, spatio-temporal pattern formation, noise-resistance,
and the outbreak of predominant cooperation – can be captured by considering
a mechanism as simple as success-driven migration: Success-driven migration
destabilizes a homogeneous strategy distribution (compare Fig. 7.1c with 7.1a and
Fig. 7.1f with 7.1d). This triggers the spontaneous formation of agglomeration
and segregation patterns [29], where noise or diffusion would cause dispersal in
the imitation-only case. The self-organized patterns create self-reinforcing social
environments characterized by behavioral correlations, and imitation promotes the
further growth of supercritical cooperation clusters. While each mechanism by

itself tends to produce frozen spatial structures, the combination of imitation and
migration supports adaptive patterns (see Fig. 7.1f). This facilitates, for example,
the regrouping of a cluster of cooperators upon invasion by a defector, which is
crucial for the survival and success of cooperators (see Fig. 7.2e–h).
By further simulations we have checked that our conclusions are robust with
respect to using different update rules, adding birth and death processes, or intro-
ducing a small fraction of individuals defecting unconditionally. The same applies
to various kinds of “noise”. Noise can even trigger cooperation in a world full of
defectors, when the probability of defectors to turn spontaneously into cooperators is
20 times smaller than the probability of cooperators to turn into defectors. Compared
to the implications of the game-dynamical replicator equation, this is remarkable:
While the replicator equation predicts that the stationary solution with a majority
of cooperators is unstable with respect to perturbations and the stationary solution
with a majority of defectors is stable [10], success-driven migration inverts the
situation: The state of 100% defectors becomes unstable to noise, while a majority
of cooperators is stabilized in a considerable area of the payoff parameter space.
Our results help to explain why cooperation can be frequent even if individuals
would behave selfishly in the vast majority of interactions. Although one may think
that migration would weaken social ties and cooperation, there is another side of it
which helps to establish cooperation in the first place, without the need to modify the
payoff structure. We suggest that, besides the ability for strategic interactions and
learning, the ability to move has played a crucial role for the evolution of large-scale
cooperation and social behavior. Success-driven migration can reduce unbalanced
social interactions, where cooperation is unilateral, and support local agglomeration.
In fact, it has been pointed out that local agglomeration is an important precondition
for the evolution of more sophisticated kinds of cooperation [30]. For example,
the level of cooperation could be further improved by combining imitation and
success-driven migration with other mechanisms such as costly punishment [19,20],
volunteering [22], or reputation [31–33].
Acknowledgements The authors would like to thank Christoph Hauert, Heiko Rauhut,
Sergi Lozano, Michael Maes, Carlos P. Roca, and Didier Sornette for their comments.
References
1. M.C. Gonzáles, C.A. Hidalgo, A.L. Barabási, Understanding individual human mobility
patterns. Nature 453, 779–782 (2008)
2. L. Hufnagel, D. Brockmann, T. Geisel, The scaling laws of human travel. Nature 439, 462–465
(2006)
3. M. Batty, Cities and Complexity (MIT Press, Cambridge, MA, 2005)
4. W. Weidlich, Sociodynamics. A Systematic Approach to Mathematical Modelling in the Social
Sciences (Harwood Academic, Amsterdam, 2000)
References 151
5. D. Pumain (ed.), Spatial Analysis and Population Dynamics (John Libbey Eurotext, France,
1991)
6. J.V. Neumann, O. Morgenstern, Theory of Games and Economic Behavior (Princeton Univer-
sity, Princeton, NJ, 1944)
7. R. Axelrod, The Evolution of Cooperation (Basic Books, New York, 1984)
8. B. Skyrms, Evolution of The Social Contract (Cambridge University, New York, 1996)
9. A. Flache, R. Hegselmann, Do irregular grids make a difference? Relaxing the spatial regularity
assumption in cellular models of social dynamics. Artif. Soc. Soc. Simulat. 4(4) (2001)
10. J.M. Epstein, Zones of cooperation in demographic prisoner’s dilemma. Complexity 4(2),
36–48 (1998)
11. L.A. Dugatkin, D.S. Wilson, ROVER: A strategy for exploiting cooperators in a patchy
environment. Am. Naturalist 138(3), 687–701 (1991)
12. M. Enquist, O. Leimar, The evolution of cooperation in mobile organisms. Animal Behav. 45,
747–757 (1993)
13. J.-F. Le Galliard, R. Ferrière, U. Dieckmann, Adaptive evolution of social traits: Origin,
trajectories, and correlations of altruism and mobility. Am. Naturalist 165(2), 206–224 (2005)
14. T. Reichenbach, M. Mobilia, E. Frey, Mobility promotes and jeopardizes biodiversity in rock-
paper-scissors games. Nature 448, 1046–1049 (2007)
15. C.A. Aktipis, Know when to walk away: contingent movement and the evolution of coopera-
tion. J. Theor. Biol. 231, 249–260 (2004)
16. M.H. Vainstein, A.T.C. Silva, J.J. Arenzon, Does mobility decrease cooperation? J. Theor. Biol.
244, 722–728 (2007)
17. M.A. Nowak, Five rules for the evolution of cooperation. Science 314, 1560–1563 (2006)
18. A. Traulsen, M.A. Nowak, Evolution of cooperation by multilevel selection. Proc. Natl. Acad.
Sci. (USA) 103, 10952–10955 (2006)
19. E. Fehr, S. Gächter, Altruistic punishment in humans. Nature 415, 137–140 (2002)
20. R. Boyd, H. Gintis, S. Bowles, P.J. Richerson, The evolution of altruistic punishment. Proc.
Natl. Acad. Sci. (USA) 100, 3531–3535 (2003)
21. M.A. Nowak, R.M. May, Evolutionary games and spatial chaos. Nature 359, 826–829 (1992)
22. G. Szabó, C. Hauert, Phase transitions and volunteering in spatial public goods games. Phys.
Rev. Lett. 89, 118101 (2002)
23. C. Hauert, M. Doebeli, Spatial structure often inhibits the evolution of cooperation in the
snowdrift game. Nature 428, 643–646 (2004)
24. M.H. Vainstein, J.J. Arenzon, Disordered environments in spatial games. Phys. Rev. E 64,
051905 (2001)
25. D. Helbing, W. Yu, Migration as a mechanism to promote cooperation. Adv. Complex Syst.
11(4), 641–652 (2008)
26. H.P. Young, Individual Strategy and Social Structure: An Evolutionary Theory of Institutions
(Princeton University, Princeton, NJ, 1998)
27. N.S. Glance, B.A. Huberman, The outbreak of cooperation. J. Math. Soc. 17(4), 281–302
(1993)
28. C. Hauert, Fundamental clusters in spatial 2 2 games. Proc. R. Soc. Lond. B 268, 761–769
(2000)
29. T.C. Schelling, Dynamic models of segregation. J. Math. Sociol. 1, 143–186 (1971)
30. J.L. Deneubourg, A. Lioni, C. Detrain, Dynamics of aggregation and emergence of coopera-
tion. Biol. Bull. 202, 262–267 (2002)
31. M.A. Nowak, K. Sigmund, Evolution of indirect reciprocity by image scoring. Nature 393,
573–577 (1998)
32. M. Milinski, D. Semmann, H.J. Krambeck, Reputation helps solve the “tragedy of the
commons”. Nature 415, 424–426 (2002)
33. B. Rockenbach, M. Milinski, The efficient interaction of indirect reciprocity and costly
punishment. Nature 444, 718–723 (2006)
Chapter 8
Evolution of Moral Behavior
8.1 Introduction
Public goods such as environmental resources or social benefits are particularly

prone to exploitation by non-cooperative individuals (“defectors”), who try to
increase their benefit at the expense of fair contributors or users, the “cooperators”.
This implies a tragedy of commons [1]. It was proposed that costly punishment
of non-cooperative individuals can establish cooperation in public goods dilemmas
[2–8], and it is effective indeed [9–11]. Nonetheless, why would cooperators choose
to punish defectors at a personal cost [12–14]? One would expect that evolutionary
pressure should eventually eliminate such “moralists” due to their extra costs
compared to “second-order free-riders” (i.e. cooperators, who do not punish). These,
however should finally be defeated by “free-riders” (defectors). To overcome this
problem [15, 16], it was proposed that cooperators who punish defectors (called
“moralists” by us) would survive through indirect reciprocity [17], reputation effects
[18] or the possibility to abstain from the joint enterprize [19–21] by “volunteering”
[22, 23]. Without such mechanisms, cooperators who punish will usually vanish.
Surprisingly, however, the second-order free-rider problem is naturally resolved,
without assuming additional mechanisms, if spatial or network interactions are
considered. This will be shown in the following.
In order to study the conditions for the disappearance of non-punishing coop-
erators and defectors, we simulate the public goods game with costly punishment,
considering two cooperative strategies (C, M) and two defective ones (D, I). For
illustration, one may imagine that cooperators (C) correspond to countries trying
to meet the CO2 emission standards of the Kyoto protocol [24], and “moralists”
(M) to cooperative countries that additionally enforce the standards by international
pressure (e.g. embargoes). Defectors (D) would correspond to those countries

This chapter reprints a previous publication, which should be cited as follows: D. Helbing,
A. Szolnoki, M. Perc, and G. Szabó, Evolutionary establishment of moral and double moral
standards through spatial interactions. PLoS Computational Biology 6(4), e1000758 (2010).

154 8 Evolution of Moral Behavior
ignoring the Kyoto protocol, and immoralists (I) to countries failing to meet the
Kyoto standards, but nevertheless imposing pressure on other countries to fulfil
them. According to the classical game-theoretical prediction, all countries would
finally fail to meet the emission standards, but we will show that, in a spatial setting,
interactions between the four strategies C, D, M, and I can promote the spreading of
moralists. Other well-known public goods problems are over-fishing, the pollution
of our environment, the creation of social benefit systems, or the establishment and
maintenance of cultural institutions (such as a shared language, norms, values, etc.).
Our simplified game-theoretical description of such problems assumes that
cooperators (C) and moralists (M) make a contribution of 1 to the respective
public good under consideration, while nothing is contributed by defectors (D)
and “immoralists” (I), i.e. defectors who punish other defectors. The sum of all
contributions is multiplied by a factor r reflecting synergy effects of cooperation,
and the resulting amount is equally shared among the k C 1 interacting individuals.
Moreover, moralists and immoralists impose a fine ˇ=k on each defecting individual
(playing D or I), which produces an additional cost =k per punished defector to
them (see Methods for details). The division by k scales for the group size, but for
simplicity, the parameter ˇ is called the punishment fine and the punishment cost.
Given the same interaction partners, an immoralist never gets a higher payoff
than a defector, but does equally well in a cooperative environment. Moreover, a
cooperator tends to outperform a moralist, given the interaction partners are the
same. However, a cooperator can do better than a defector when the punishment
fine ˇ is large enough.
It is known that punishment in the public goods game and similar games can
promote cooperation above a certain critical threshold of the synergy factor r
[11, 25]. Besides cooperators who punish defectors, Heckathorn considered “full
cooperators” (moralists) and “hypocritical cooperators” (immoralists) [26]. For
well-mixed interactions (where individuals interact with a representative rather
than local strategy distribution), Eldakar and Wilson find that altruistic punishment
(moralists) can spread, if second-order free-riders (non-punishing altruists) are
excluded, and that selfish punishers (immoralists) can survive together with altruis-
tic non-punishers (cooperators), provided that selfish nonpunishers (defectors) are
sufficiently scarce [27].
Besides well-mixed interactions, some researchers have also investigated the
effect of spatial interactions [5,11,28,29], since it is known that they can support the
survival or spreading of cooperators [30] (but this is not always the case [31, 32]).
In this way, Brandt et al. discovered a coexistence of cooperators and defectors
for certain parameter combinations [11]. Compared to these studies, our model
assumes somewhat different replication and strategy updating rules. The main point,
however, is that we have chosen long simulation times and scanned the parameter
space more extensively, which revealed several new insights, for example, the
possible coexistence of immoralists and moralists, even when a substantial number
of defectors is present initially. When interpreting our results within the context of
moral dynamics [33], our main discoveries for a society facing public goods games
may be summarized as follows:
8.2 Results 155
1. Victory over second-order free-riders: Over a long enough time period, moral-
ists fully eliminate cooperators, thereby solving the “second-order free-rider
problem”. This becomes possible by spatial segregation of the two cooperative
strategies C and M, where the presence of defectors puts moralists in a advan-
tageous position, which eventually allows moralists to get rid of non-punishing
cooperators.
2. Who laughs last laughs best effect: Moralists defeat cooperators even when the
defective strategies I and D are eventually eliminated, but this process is very
slow. That is, the system behavior changes its character significantly even after
very long times. This is the essence of the “who laughs last laughs best effect”.
The finally winning strategy can be in a miserable situation in the beginning, and
its victory may take very long.
3. Lucifer’s positive side effect: By permanently generating a number of defectors,
small mutation rates can considerably accelerate the spreading of moralists.
4. Unholy collaboration of moralists with immoralists: Under certain conditions,
moralists can survive by profiting from immoralists. This actually provides
the first explanation for the existence of defectors, who hypocritically punish
other defectors, although they defect themselves. The occurrence of this strange
behavior is well-known in reality and even experimentally confirmed [34, 35].
These discoveries required a combination of theoretical considerations and exten-
sive computer simulations on multiple processors over long time horizons.
8.2 Results
For well-mixed interactions, defectors are the winners of the evolutionary compe-
tition among the four behavioral strategies C, D, M, and I [36], which implies a
tragedy of the commons despite punishment efforts. The reason is that cooperators
(second-order free-riders) spread at the cost of moralists, while requiring them for
their own survival.
Conclusions from computer simulations are strikingly different, if the assump-
tion of well-mixed interactions is replaced by the more realistic assumption of
spatial interactions. When cooperators and defectors interact in space [5, 11, 37–44],
it is known that some cooperators can survive through spatial clustering [45].
However, it is not clear how the spatiotemporal dynamics and the frequency of
cooperation would change in the presence of moralists and immoralists. Would
spatial interactions be able to promote the spreading of punishment and thereby
eliminate second-order free-riders?
In order to explore this, we have scanned a large parameter space. Figure 8.1
shows the resulting state of the system as a function of the punishment cost
and punishment fine ˇ after a sufficiently long transient time. If the fine-to-cost
ratio ˇ= and the synergy factor r are low, defectors eliminate all other strategies.
a r = 2.0 b r = 3.5
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
cost
cost
D D
0.3 0.3
0.2 M 0.2 M
D+M
0.1 0.1
0.0 0.0
0 0.3 0.6 0.9 1.2 1.5 1.8 0 0.1 0.2 0.3 0.4 0.5 0.6
fine fine
c r = 4.4 d r = 3.5 (enlarged)

0.7 0.03
0.6
0.5
D+C 0.02 D
0.4 D+M
cost
cost
0.3 M M
0.01
0.2
0.1 D+M M+I
0.0 0.00
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.15 0.30 0.45
fine fine
Fig. 8.1 Phase diagrams showing the remaining strategies in the spatial public goods game with
cooperators (C), defectors (D), moralists (M) and immoralists (I), after a sufficiently long transient
time. Initially, each of the four strategies occupies 25% of the sites of the square lattice, and
their distribution is uniform in space. However, due to their evolutionary competition, two or
three strategies die out after some time. The finally resulting state depends on the synergy r of
cooperation, the punishment cost , and the punishment fine ˇ. The displayed phase diagrams are
for (a) r D 2:0, (b) r D 3:5, and (c) r D 4:4. (d) Enlargement of the small-cost area for r D 3:5.
Solid separating lines indicate that the resulting fractions of all strategies change continuously with
a modification of the model parameters ˇ and , while broken lines correspond to discontinuous
changes. All diagrams show that cooperators cannot stop the spreading of moralists, if only the fine-
to-cost ratio is large enough. Furthermore, there are parameter regions where moralist can crowd
out cooperators in the presence of defectors. Note that the spreading of moralists is extremely slow
and follows a voter model kind of dynamics [47], if their competition with cooperators occurs
in the absence of defectors. Therefore, computer simulations had to be run over extremely long
times (up to 107 iterations for a systems size of 400 400). For similar reasons, a small level
of strategy mutations (which permanently creates a small number of strategies of all kinds, in
particular defectors) can largely accelerate the spreading of moralists in the M phase, while it
does not significantly change the resulting fractions of the four strategies [53]. The existence of
immoralists is usually not relevant for the outcome of the evolutionary dynamics. Apart from
a very small parameter area, where immoralists and moralists coexist, immoralists are quickly
extinct. Therefore, the 4-strategy model usually behaves like a model with the three strategies
C, D, and M only. As a consequence, the phase diagrams for the latter look almost the same like
the ones presented here [46]
8.2 Results 157
However, for large enough fines ˇ, cooperators and defectors are always eliminated,
and moralists prevail (Fig. 8.1).
At larger r values, when the punishment costs are moderate, we find a coex-
istence of moralists with defectors without any cooperators. To understand why
moralists can outperform cooperators despite additional punishment costs, it is
important to analyze the dynamics of spatial interactions. Starting with a homo-
geneous strategy distribution (Fig. 8.2a), the imitation of better-performing neigh-
bors generates small clusters of individuals with identical strategies (Fig. 8.2b).
“Immoralists” die out quickly, while cooperators and moralists form separate clus-
ters in a sea of defectors (Fig. 8.2c). The further development is determined by the
interactions at the interfaces between clusters of different strategies (Figs. 8.2d–f). In
the presence of defectors, the fate of moralists is not decided by a direct competition
with cooperators, but rather by the success of both cooperative strategies against
a b c
d e f
Fig. 8.2 Elimination of second-order free-riders (non-punishing cooperators) in the spatial public
goods game with costly punishment for r D 4:4, ˇ D 0:1, and D 0:1. (a) Initially, at
time t D 0, cooperators (blue), defectors (red), moralists (green) and immoralists (yellow) are
uniformly distributed over the spatial lattice. (b) After a short time period (here, at t D 10),
defectors prevail. (c) After 100 iterations, immoralists have almost disappeared, and cooperators
prevail, since cooperators earn high payoffs when organized in clusters. (d) At t D 500, there is
a segregation of moralists and cooperators, with defectors in between. (e) The evolutionary battle
continues between cooperators and defectors on the one hand, and defectors and moralists on the
other hand (here at t D 1; 000). (f) At t D 2; 000, cooperators have been eliminated by defectors,
and a small fraction of defectors survives among a large majority of moralists. Interestingly, each
strategy (apart from I) has a time period during which it prevails, but only moralists can maintain
their majority. While moralists perform poorly in the beginning, they are doing well in the end. We
refer to this as the “who laughs last laughs best” effect
invasion attempts by defectors. If the ˇ= -ratio is appropriate, moralists respond

better to defectors than cooperators. Indeed, moralists can spread so successfully
in the presence of defectors that areas lost by cooperators are quickly occupied by
moralists (supplementary Video S1). This indirect territorial battle ultimately leads
to the extinction of cooperators (Fig. 8.2f), thus resolving the second-order free-rider
problem.
In conclusion, the presence of some conventional free-riders (defectors) supports
the elimination of second-order free-riders. However, if the fine-to-cost ratio is
high, defectors are eliminated after some time. Then, the final struggle between
moralists and cooperators takes such a long time that cooperators and moralists
seem to coexist in a stable way. Nevertheless, a very slow coarsening of clusters is
revealed, when simulating over extremely many iterations. This process is finally
won by moralists, as they are in the majority by the time the defectors disappear,
while they happen to be in the minority during the first stage of the simulation (see
Fig. 8.2). We call this the “who laughs last laughs best effect”. Since the payoffs
of cooperators and moralists are identical in the absence of other strategies, the
underlying coarsening dynamics is expected to agree with the voter model [47].
Note that there is always a punishment fine ˇ, for which moralists can out-
compete all other strategies. The higher the synergy factor r, the lower the
ˇ= -ratio required to reach the prevalence of moralists. Yet, for larger values of r,
the system behavior also becomes richer, and there are areas for small fines or
high punishment costs, where clusters with different strategies can coexist (see
Figs. 8.1b–d). For example, we observe the coexistence of clusters of moralists and
defectors (see Fig. 8.2 and supplementary Video S1) or of cooperators and defectors
(see supplementary Video S2).
Finally, for low punishment costs but moderate punishment fines and synergy
factors r (see Fig. 8.1d), the survival of moralists may require the coexistence
with “immoralists” (see Fig. 8.3 and supplementary Video S3). Such immoralists
are often called “sanctimonious” or blamed for “double moral standards”, as they
defect themselves, while enforcing the cooperation of others (for the purpose of
exploitation). This is actually the main obstacle for the spreading of immoralists, as
they have to pay punishment costs, while suffering from punishment fines as well.
Therefore, immoralists need small punishment costs to survive. As cooperators
die out quickly for moderate values of r, the survival of immoralists depends on the
existence of moralists they can exploit, otherwise they cannot outperform defectors.
Conversely, moralists benefit from immoralists by supporting the punishment of
defectors. Note, however, that this mutually profitable interaction between moralists
and immoralists, which appears like an “unholy collaboration”, is fragile: If ˇ is
increased, immoralists suffer from fines, and if is increased, punishing becomes
too costly. In both cases, immoralists die out, and the coexistence of moralists
and immoralists breaks down. Despite this fragility, “hypocritical” defectors, who
punish other defectors, are known to occur in reality. Their existence has even
been found in experiments [34, 35]. Here, we have revealed conditions for their
occurrence.
8.3 Discussion 159
a b c
d e f
Fig. 8.3 Coexistence of moralists and immoralists for r D 3:5, ˇ D 0:12, and D 0:005,
supporting the occurrence of individuals with “double moral standards” (who punish defectors,
while defecting themselves). (a) Initially, at time t D 0, cooperators (blue), defectors (red),
moralists (green) and immoralists (yellow) are uniformly distributed over the spatial lattice.
(b) After 250 iterations, cooperators have been eliminated in the competition with defectors (as
the synergy effect r of cooperation is not large enough), and defectors are prevailing. (c–e) The
snapshots at t D 760, t D 2;250, and t D 6;000 show the interdependence of moralists and
immoralists, which appears like a tacit collaboration. It is visible that the two punishing strategies
win the struggle with defectors by staying together. On the one hand, due to the additional
punishment cost, immoralists can survive the competition with defectors only by exploiting
moralists. On the other hand, immoralists support moralists in fighting defectors. (f) After 12,000
iterations, defectors have disappeared completely, leading to a coexistence of clusters of moralists
with immoralists
8.3 Discussion
In summary, the second-order free-rider problem finds a natural and simple

explanation, without requiring additional assumptions, if the local nature of most
social interactions is taken into account and punishment efforts are large enough.
In fact, the presence of spatial interactions can change the system behavior so
dramatically that we do not find the dominance of free-riders (defectors) as in the
case of well-mixed interactions, but a prevalence of moralists via a “who laughs
last laughs best” effect (Fig. 8.2). Moralists can escape disadvantageous kinds of
competition with cooperators by spatial segregation. However, their triumph over
all the other strategies requires the temporary presence of defectors, who diminish
the cooperators (second-order free-riders). Finally, moralists can take over, as they
have reached a superiority over cooperators (which is further growing) and as they
can outcompete defectors (conventional free-riders).
Our findings stress how crucial spatial or network interactions in social systems
are. Their consideration gives rise to a rich variety of possible dynamics and a
number of continuous or discontinuous transitions between qualitatively different
system behaviors. Spatial interactions can even invert the finally expected system
behavior and, thereby, explain a number of challenging puzzles of social, economic,
and biological systems. This includes the higher-than-expected level of cooperation
in social dilemma situations, the elimination of second-order free-riders, and the
formation of what looks like a collaboration between otherwise inferior strategies.
By carefully scanning the parameter space, we found several possible kinds of
coexistence between two strategies each:
• Moralists (M) and defectors (D) can coexist, when the disadvantage of coopera-
tive behavior is not too large (i.e. the synergy factor is high enough), and if the
punishment fine is sufficiently large that moralists can survive among defectors,
but not large enough to get rid of them.
• Instead of M and D, moralists (M) and immoralists (I) coexist, when the
punishment cost is small enough. The small punishment cost is needed to ensure
that the disadvantage of punishing defectors (I) compared to non-punishing
defectors (D) is small enough that it can be compensated by the additional
punishment efforts contributed by moralists.
• To explain the well-known coexistence of D and C [11], it is useful to remember
that defectors can be crowded out by cooperators, when the synergy factor
exceeds a critical value (even when punishment is not considered). Slightly below
this threshold, neither cooperators nor defectors have a sufficient advantage to get
rid of the other strategy, which results in a coexistence of both strategies.
Generally, a coexistence of strategies occurs, when the payoffs at the interface
between clusters of different strategies are balanced. In order to understand why the
coexistence is possible in a certain parameter area rather than just for an infinitely
small parameter set, it is important to consider that typical cluster sizes vary with
the parameter values. This also changes the typical radius of the interface between
the coexisting strategies and, thereby, the typical number of neighbors applying the
same strategy or a different one. In other words, a change in the shape of a cluster
can partly counter-balance payoff differences between two strategies by varying the
number of “friends” and “enemies” involved in the battle at the interface between
spatial areas with different strategies (see Fig. 8.4).
Finally, we would like to discuss the robustness of our observations. It is well-
known that the level of cooperation in the public goods game is highest in small
groups [10]. However, we have found that moralists can crowd out non-punishing
cooperators also for group sizes of k C 1 D 9, 13, 21, or 25 interacting individuals,
for example. In the limiting case of large groups, where everybody interacts with
everybody else, we expect the outcome of the well-mixed case, which corresponds
to defection by everybody (if other mechanisms like reputation effects [11] or
abstaining are not considered [20]). That is, the same mechanisms that can create
cooperation among friends may fail to establish shared moral standards, when
8.3 Discussion 161
a c
b d
Fig. 8.4 Dependence of cluster shapes on the punishment fine ˇ in the stationary state, supporting
an adaptive balance between the payoffs of two different strategies at the interface between
competing clusters. Snapshots in the top row were obtained for low punishment fines, while
the bottom row depicts results obtained for higher values of ˇ. (a) Coexistence of moralists and
defectors for a synergy factor r D 3:5, punishment cost D 0:20, and punishment fine ˇ D 0:25.
(b) Same parameters, apart from ˇ D 0:4. (c) Coexistence of moralists and immoralists for
r D 3:5, D 0:05, and ˇ D 0:12. (d) Same parameters, apart from ˇ D 0:25. A similar change
in the cluster shapes is found for the coexistence of cooperators and defectors, if the synergy factor
r is varied
spatial interactions are negligible. It would therefore be interesting to study, whether

the fact that interactions in the financial system are global, has contributed to the
financial crisis. Typically, when social communities exceed a certain size, they need
sanctioning institutions to stabilize cooperation (such as laws, an executive system,
and police).
Note that our principal discoveries are not expected to change substantially for
spatial interactions within irregular grids (i.e. neighborhoods different from Moore
1.0
0.8 D
C
0.6 I
fractions
M
0.4
0.2
0.0
0.3 0.4 0.5
fine
Fig. 8.5 Resulting fractions of the four strategies C, D, I, and M, for random regular graphs as
a function of the punishment fine ˇ. The graphs were constructed by rewiring links of a square
lattice of size 400 400 with probability Q, thereby preserving the degree distribution (i.e. every
player has four nearest neighbors) [49]. For small values of Q, small-world properties result,
while for Q ! 1, we have a random regular graph. By keeping the degree distribution fixed,
we can study the impact of randomness in the network structure independently of other effects. An
inhomogeneous degree distribution can further promote cooperation [37]. The results displayed
here are averages over ten simulation runs for the model parameters r D 3:5, D 0:05, and
Q D 0:99. Similar results can be obtained also for other parameter combinations
neighborhoods) [48]. In case of network interactions, we have checked that small-

world or random networks lead to similar results, when the degree distribution
is the same (see Fig. 8.5). A heterogeneous degree distribution is even expected
to reduce free-riding [37] (given the average degree is the same). Finally, adding
other cooperation-promoting mechanisms to our model such as direct reciprocity
(a shadow of the future through repeated interactions [50]), indirect reciprocity [17]
(trust and reputation effects [11, 18]), abstaining from a joint enterprize [19–23],
or success-driven migration [51], will strengthen the victory of moralists over
conventional and second-order free-riders.
In order to test the robustness of our observations, we have also checked the effect
of randomness (“noise”) originating from the possibility of strategy mutations. It is
known that mutations may promote cooperation [52]. According to the numerical
analysis of the spatial public goods game with punishment, the introduction of
rare mutations does not significantly change the final outcome of the competition
between moralists and non-punishing cooperators. Second-order free-riders will
always be a negligible minority in the end, if the fine-to-cost ratio and mutation
rate allows moralists to spread. While a large mutation rate naturally causes a
uniform distribution of strategies, a low level of strategy mutations can be even
8.3 Discussion 163
beneficial for moralists. Namely, by permanently generating a number of defectors,

small mutation rates can considerably accelerate the spreading of moralists, i.e. the
slow logarithmic coarsening is replaced by another kind of dynamics [53]. Defectors
created by mutations play the same role as in the D C M phase (see Figs. 8.1–8.2).
They put moralists into an advantage over non-punishing cooperators, resulting in a
faster spreading of the moralists (which facilitates the elimination of second-order
free-riders over realistic time periods). In this way, the presence of a few “bad
guys” (defectors) can accelerate the spreading of moral standards. Metaphorically
speaking, we call this “lucifer’s positive side effect”.
The current study paves the road for several interesting extensions. It is possible,
for example, to study antisocial punishment [54], considering also strategies which
punish cooperators [55]. The conditions for the survival or spreading of antisocial
punishers can be identified by the same methodology, but the larger number of
strategies creates new phases in the parameter space. While the added complexity
transcends what can be discussed here, the current study demonstrates clearly how
differentiated the moral dynamics in a society facing public goods problems can be
and how it depends on a variety of factors (such as the punishment cost, punishment
fine, and synergy factor). Going one step further, evolutionary game theory may
even prove useful to understand how moral feelings have evolved.
Furthermore, it would be interesting to investigate the emergence of punishment
within the framework of a coevolutionary model [56–58], where both, individual
strategies and punishment levels are simultaneously spread. Such a model could, for
example, assume that individuals show some exploration behavior [52] and stick to
successful punishment levels for a long time, while they quickly abandon unsuccess-
ful ones. In the beginning of this coevolutionary process, costly punishment would
not pay off. However, after a sufficiently long time, mutually fitting punishment
strategies are expected to appear in the same neighborhood by coincidence [51].
Once an over-critical number of successful punishment strategies have appeared
in some area of the simulated space, they are eventually expected to spread. The
consideration of success-driven migration should strongly support this process [51].
Over many generations, genetic-cultural coevolution could finally inherit costly
punishment as a behavioral trait, as is suggested by the mechanisms of strong
reciprocity [59].
Appendix A: Methods
We study the public goods game with punishment. Cooperative individuals (C and
M) make a contribution of 1 to the public good, while defecting individuals (D and I)
contribute nothing. The sum of all contributions is multiplied by r and the resulting
amount equally split among the k C1 interacting individuals. A defecting individual
(D or I) suffers a fine ˇ=k by each punisher among the interaction partners,
and each punishment requires a punisher (M or I) to spend a cost =k on each
defecting individual among the interaction partners. In other words, only defectors
and punishing defectors (immoralists) are punished, and the overall punishment is
proportional to the sum of moralists and immoralists among the k neighbors. The
scaling by k serves to make our results comparable with models studying different
groups sizes.
Denoting the number of so defined cooperators, defectors, moralists, and
immoralists among the k interaction partners by NC , ND , NM and NI ,
respectively, an individual obtains the following payoff: If it is a cooperator, it gets
PC D r.NC C NM C 1/=.k C 1/ 1, if a defector, the payoff is PD D r.NC C
NM /=.k C 1/ ˇ.NM C NI /=k, a moralist receives PM D PC .ND C NI /=k,
and an immoralist obtains PI D PD .ND C NI /=k. Our model of the spatial
variant of this game studies interactions in a simple social network allowing for
clustering. It assumes that individuals are distributed on a square lattice with
periodic boundary conditions and play a public goods game with k D 4 neighbors.
We work with a fully occupied lattice of size L L with L D 200 : : : 1; 200
in Fig. 8.1 and L D 100 in Figs. 8.2–8.4 (the lattice size must be large enough
to avoid an accidental extinction of a strategy). The initial strategies of the L2
individuals are equally and uniformly distributed. Then, we perform a random
sequential update. The individual at the randomly chosen location x belongs to
five groups. (It is the focal individual of a Moore neighborhood and a member
of the Moore neighborhoods of four nearest neighbors). It plays the public goods
g
game with the k interaction partners of a group g, and obtains P ga payoff Px in
all five groups it belongs to. The overall payoff is Px D g Px . Next, one of
the four nearest neighbors is randomly chosen. Its location shall be denoted by y
and its overall payoff by Py . This neighbor imitates the strategy of the individual at
location x with probability q D 1=f1CexpŒ.Py Px /=Kg [45]. That is, individuals
tend to imitate better performing strategies in their neighborhood, but sometimes
deviate (due to trial-and-error behavior or mistakes) [31]. Realistic noise levels lie
between the two extremes K ! 0 (corresponding to unconditional imitation by the
neighbor, whenever the overall payoff Px is higher than Py ) and K ! 1 (where
the strategy is copied with probability 1/2, independently of the payoffs). For the
noise level K D 0:5 chosen in our study, the evolutionary selection pressure is high
enough to eventually eliminate poorly performing strategies in favor of strategies
with a higher overall payoff. This implies that the resulting frequency distribution
of strategies in a large enough lattice is independent of the specific initial condition
after a sufficiently long transient time. Close to the separating line between M and
D+M in Fig. 8.1, the equilibration may require up to 107 iterations (involving L2
updates each).
Acknowledgements D.H. would like to thank for useful comments by Carlos P. Roca,
Moez Draief, Stefano Balietti, Thomas Chadefaux, and Sergi Lozano.
Author Contributions Conceived and designed the experiments: DH AS MP GS. Performed the
experiments: DH AS MP GS. Wrote the paper: DH AS MP GS.
References 165
References
1. G. Hardin, The tragedy of the commons. Science 162, 1243–1248 (1968)

3. E. Fehr, U. Fischbacher, The nature of human altruism. Nature 425, 785–791 (2003)
4. P. Hammerstein (ed.), Genetic and Cultural Evolution of Cooperation (MIT Press, Cambridge,
MA, 2003)
5. M. Nakamaru, Y. Iwasa, The evolution of altruism and punishment: Role of selfish punisher.
J. Theor. Biol. 240, 475–488 (2006)
6. C.F. Camerer, E. Fehr, When does “economic man” dominate social behavior. Science 311,
47–52 (2006)
7. O. Gurerk, B. Irlenbusch, B. Rockenbach, The competitive advantage of sanctioning institu-
tions. Science 312, 108–111 (2006)
8. K. Sigmund, C. Hauert, M.A. Nowak, Reward and punishment. Proc. Natl. Acad. Sci. USA 98,
10757–10762 (2001)
9. J. Henrich, R. Boyd, Why people punish defectors. J. Theor. Biol. 208, 79–89 (2001)
Natl. Acad. Sci. USA 100, 3531–3535 (2003)
11. H. Brandt, C. Hauert, K. Sigmund, Punishing and reputation in spatial public goods games.
Proc. R. Soc. Lond. Ser. B 270, 1099–1104 (2003)
12. T. Yamagishi, The provision of a sanctioning system as a public good. J. Pers. Soc. Psychol.
51, 110–116 (1986)
13. E. Fehr, Don’t lose your reputation. Nature 432, 449–450 (2004)
14. A.M. Colman, The puzzle of cooperation. Nature 440, 744–745 (2006)
15. J.H. Fowler, Second-order free-riding problem solved? Nature 437, E8-E8 (2005)
16. K. Panchanathan, R. Boyd, Reply. Nature 437, E8-E9 (2005)
17. K. Panchanathan, R. Boyd, Indirect reciprocity can stabilize cooperation without the second-
order free rider problem. Nature 432, 499–502 (2004)
18. M. Milinski, D. Semmann, H.-J. Krambeck, Reputation helps to solve the “tragedy of the
commons”. Nature 415, 424–426 (2002)
19. J.H. Fowler, Altruistic punishment and the origin of cooperation. Proc. Natl. Acad. Sci. USA
102, 7047–7049 (2005)
20. H. Brandt, C. Hauert, K. Sigmund, Punishing and abstaining for public goods. Proc. Natl.
Acad. Sci. USA 103, 495–497 (2006)
21. C. Hauert, A. Traulsen, H. Brandt, M.A. Nowak, K. Sigmund, Via freedom to coercion: The
emergence of costly punishment. Science 316, 1905–1907 (2007)
22. C. Hauert, S. De Monte, J. Hofbauer, K. Sigmund, Volunteering as red queen mechanism for
cooperation in public goods game. Science 296, 1129–1132 (2002)
23. D. Semmann, H.-J. Krambeck, M. Milinski, Volunteering leads to rock-paper-scissors dynam-
ics in a public goods game. Nature 425, 390–393 (2003)
24. M. Milinski, R.D. Sommerfeld, H.J. Krambeck, F.A. Reed, J. Marotzke, The collective-risk
social dilemma and the prevention of simulated dangerous climate change. Proc. Natl. Acad.
Sci. USA 105, 2291–2294 (2008)
25. K. Sigmund, The Calculus of Selfishness (Princeton University Press, Princeton, 2010)
26. D.D. Heckathorn, The dynamics and dilemmas of collective action. Am. Soc. Rev. 61, 250–277
(1996)
27. O.T. Eldakar, D.S. Wilson, Selfishness as second-order altruism. Proc. Natl. Acad. Sci. USA
109, 6982–6986 (2008)
28. M. Nakamaru, Y. Iwasa, The evolution of altruism by costly punishment in the lattice structured
population: Score-dependent viability versus score-dependent fertility. Evol. Ecol. Res. 7,
853–870 (2005)
29. T. Sekiguchi, M. Nakamaru, Effect of the presence of empty sites on the evolution of
cooperation by costly punishment in spatial games. J. Theor. Biol. 256(2), 297–304 (2009)
31. A. Traulsen, et al., Human strategy updating in evolutionary games. Proc. Natl. Acad. Sci. USA
107, 2962–2966 (2010)
32. M.A. Nowak, C.E. Tarnita, T. Antal, Evolutionary dynamics in structured populations. Phil.
Trans. R. Soc. B 365, 19–30 (2010)
33. M.D. Hauser, Moral Minds: How Nature Designed Our Universal Sense of Right and Wrong
(Ecco, New York, 2006)
34. A. Falk, E. Fehr, U. Fischbacher, Driving forces behind informal sanctions. Econometrica 73,
2017–2030 (2005)
35. M. Shinada, T. Yamagishi, Y. Ohmura, False friends are worse than bitter enemies: “Altruistic”
punishment of in-group members. Evol. Hum. Behav. 25, 379–393 (2004)
Rev. Lett. 89, 118101 (2002)
37. F.C. Santos, M.D. Santos, J.M. Pacheco, Social diversity promotes the emergence of coopera-
tion in public goods games. Nature 454, 213–216 (2008)
38. J.Y. Wakano, M.A. Nowak, C. Hauert, Spatial dynamics of ecological public goods. Proc. Natl.
Acad. Sci. USA 106, 7910–7914 (2009)
40. M.A. Nowak, S. Bonhoeffer, R.M. May, More spatial games. Int. J. Bifurcat. Chaos 4, 33–56
(1994)
41. M.A. Nowak, S. Bonhoeffer, R.M. May, Spatial games and the maintenance of cooperation.
Proc. Natl. Acad. Sci. USA 91, 4877–4881 (1994)
42. M.A. Nowak, S. Bonhoeffer, R.M. May, Robustness of cooperation. Nature 379, 125–126
(1996)
43. C.G. Nathanson, C.E. Tarnita, M.A. Nowak, Calculating evolutionary dynamics in structured
populations. PLoS Comput. Biol. 5, e1000615 (2009)
44. J.M. Pacheco, F.L. Pinheiro, F.C. Santos, Population structure induces a symmetry breaking
favoring the emergence of cooperation. PLoS Comput. Biol. 5, e1000596 (2009)
45. G. Szabó, C. Tőke, Evolutionary prisoner’s dilemma game on a square lattice. Phys. Rev. E 58,
69–73 (1998)
46. D. Helbing, A. Szolnoki, M. Perc, G. Szabó, Punish, but not too hard: how costly punishment
spreads in the spatial public goods game. New J. Phys. 12(8), 083005 (2010)
47. I. Dornic, H. Chaté, J. Chave, H. Hinrichsen, Critical coarsening without surface tension: The
universality class of the voter model. Phys. Rev. Lett. 87, 045701 (2001)
48. A. Flache, R. Hegselmann, Do irregular grids make a difference? Relaxing the spatial regularity
assumption in cellular models of social dynamics. J. Artif. Soc. Soc. Simul. 4, 4 (2001). See
http://www.soc.surrey.ac.uk/JASSS/4/4/6.html
49. G. Szabó, A. Szolnoki, R. Izsák, Rock-scissors-paper game on regular small-world networks.
J. Phys. A: Math. Gen. 37, 2599–2609 (2004)
51. D. Helbing, W. Yu, The outbreak of cooperation among success-driven individuals under noisy
conditions. Proc. Natl. Acad. Sci. USA 106, 3680–3685 (2009)
52. A. Traulsen, C. Hauert, H.D. Silva, M.A. Nowak, K. Sigmund, Exploration dynamics in
evolutionary games. Proc. Natl. Acad. Sci. USA 106, 709–712 (2009)
53. D. Helbing, A. Szolnoki, M. Perc, G. Szabó, Defector-accelerated cooperativeness and
punishment in public goods games with mutations. Phys. Rev. E 81(5), 057104 (2010)
54. B. Herrmann, C. Thöni, S. Gächter, Antisocial punishment across societies. Science 319,
1362–1367 (2008)
55. D.G. Rand, H. Ohtsuki, M.A. Nowak, Direct reciprocity with costly punishment: Generous
tit-for-tag prevails. J. Theor. Biol. 256, 45–57 (2009)
56. G. Szabó, A. Szolnoki, V. Jeromos, Selection of dynamical rules in spatial Prisoner’s Dilemma
games. EPL 87, 18007 (2009)
References 167
57. F.C. Santos, J.M. Pacheco, T. Lenaerts, Cooperation prevails when individuals adjust their
social ties. PLoS Comput. Biol. 2, 1284–1291 (2006)
58. M. Perc, A. Szolnoki, Coevolutionary games - A mini review. BioSystems 99, 109–125 (2010)
59. S. Bowles, H. Gintis, The evolution of strong reciprocity: Cooperation in heterogeneous
populations. Theor. Popul. Biol. 65, 17–28 (2004)
Chapter 9
Coordination and Competitive Innovation
Spreading in Social Networks
9.1 Introduction
The analysis of percolation in random media has become a very popular framework
over the last decades to address a wide variety of phenomena in disordered systems,
such as oil mining in porous reservoirs, fire spreading in forests, fracture patterns
in rocks, electromagnetic properties of composite materials, etc [1]. More recently,
it has also been applied to shed light on social phenomena, namely the diffusion
of opinions [2] and innovations [3] in social networks. All of the aforementioned
systems can be modeled as percolation problems. More precisely, they can be
abstracted as a network of nodes representing the topology of the random medium,
wherein nodes can be either “empty” or “occupied”, depending on the state of their
neighbors. Starting from an initial condition where some nodes are occupied, an
occupied node becomes empty if the number of its occupied neighbors goes below
a threshold k, the index of the percolation process (k D 2 for standard percolation
[4] and k 3 for bootstrap or k-core percolation [5, 6]). The underlying switching
dynamics is therefore assumed to be unidirectional.
Here we introduce a percolation model that generalizes this powerful theoretical
approach. Our extension assumes that nodes are of two types A or B, and that
a node changes type when the number of neighbors of the same type is less
than k. Consequently both changes A-to-B and B-to-A are possible, i.e. we are
considering a bi-directional percolation dynamics instead. Figure 9.1 provides an
example which illustrates the fundamental difference between both percolation
processes. The problem we want to address is the competition between innovations
[7]. Competition between products, tools or technical standards is ubiquitous. Well-
known examples are railway gauges, keyboard layouts, computer operating systems,
high-definition video standards, e-book readers, etc. The reasons that determine

This chapter has been prepared by C. Roca, Moez Draief, and D. Helbing under the project title
“Percolate or die: Multi-percolation decides the struggle between competing innovations”.

170 9 Coordination and Competitive Innovation Spreading in Social Networks
a b c
d e f
Fig. 9.1 Comparison of unidirectional (panels A–C) vs bi-directional percolation (panels D–F).
In unidirectional percolation, occupied nodes (in black) become empty (in white) when they have
less than k occupied neighbors (in this example k D 2). In the end there is no occupied node
that survives the percolation pruning. With bi-directional percolation, both white and black nodes
switch color when they have less than k D 2 neighbors of their same color. The end result in this
case is an all-black graph, with no white nodes surviving the competitive bi-directional percolation
process. All black nodes end up with two black neighbors and they are connected, hence they
form a percolating cluster. Notice that although both cases have the same initial condition and
percolation index k, the outcome is opposite
the outcome of these fierce competitions have puzzled researchers of different

disciplines for a long time [8, 9]. Previous work has highlighted the combined
influence of intrinsic benefits of each option together with costs incurred due to
switching [10]. In addition, it has been suggested that social structure, i.e. the
network of social relationships in a group or population, would play a crucial
role [11]. So far, however, there has been little analytical work that elucidates
the outcome of such competitions. In this work we show that the competition
between innovations can be understood as a bi-directional percolation process,
which ultimately determines the fate of the options in contest.
9.2 Model
To start with, let us consider a simple model with two competing options, A and B
(for example Blu-ray Disc vs HD DVD), whose benefits to individuals depend
on intrinsic factors as well as on the acceptance by others in a certain social
neighborhood. This can be modeled as a coordination problem [12, 13], in which
individuals choosing one of the two options A or B obtain a payoff A D q xQ A and
9.2 Model 171
B D .1 q/xQ B respectively. The relative advantage of one option over the other
is represented by the parameter q, where 0 q 1. Quantities xQ A and xQ B give,
respectively, the proportion of people adhering to option A or B among the social
acquaintances who have an influence on the individual’s decision, such us family
members, friends, co-workers, etc (xQ A C xQ B D 1 for every individual). In addition,
we consider that changing option entails some switching cost, which is called cA for
a follower of option A who changes to B, and cB in the opposite case. Thus, A- and
B-individuals have the following effective payoff matrices

q 0 q cB cB
AW BW ; (9.1)
cA 1 q cA 0 1q
where we follow the standard convention for symmetric games: rows represent own
strategy and columns that of the interaction partner.
For the moment, we assume that individuals are able to assess to a good extent
the benefits and drawbacks of options A and B, and also the degree of penetration of
each option in their social neighborhood, i.e. we assume a low level of uncertainty
in the decision-making process (more on this important point later). Therefore,
individuals choose a best response to the current state of their social context
according to the payoffs expected from (9.1). As a consequence, A-individuals
change to option B if the proportion of A-neighbors is below a certain threshold,
namely xQ A < 1q cA , while B-individuals switch if the proportion of B-neighbors
is less than certain value, xQ B < q cB .
This defines an evolutionary game [14,15], which consists in a dynamical system
with state variable x, the global density of the followers of one of the options.
We set x D xB without loss of generality. Disregarding the effect of social
structure for the moment, the evolution of xB can easily be calculated assuming
a well-mixed population [16], equivalent to the mean field hypothesis in physics
or the representative agent in economics. It posits that every individual has, in her
social neighborhood, a proportion of A- or B-individuals equal to the respective
global densities, i.e. for every individual and at any time xQ B D xB . Under this
assumption, the population rapidly reaches an equilibrium with stationary value
xB D limt !1 xB .t/
8
ˆ if xB0 < q cB ;
<0;
ˆ
xB D xB0 ; if q cB xB0 q C cA ; (9.2)
ˆ
:̂1; 0
if x > q C c ;
B A
where xB0 represents the initial density of individuals following option B. Equa-
tion (9.2) shows that under well-mixed conditions switching costs induce the
appearance of a heterogeneous state, in which both competing options keep a share
of the population. If costs are left out (cA D cB D 0), then we find the standard
solution of a coordination game, with an unstable equilibrium at xB D xB0 , which
separates the basins of attraction of the stable equilibria xB D 0 and xB D 1 [17].
Let us now consider this model embedded in a social network [18, 19], that
is, in a network of social relationships which determines who interacts with
whom [14, 20, 21]. Here we use regular random networks [22], which are networks
where nodes are linked at random and where each node has the same number of
neighbors, or degree, z. Such networks are known to have the most neutral effect
on evolutionary games [23], just fixing neighborhood size and preserving the social
context of individuals. They avoid particular effects that some topological features,
such as clustering or degree heterogeneity, may have [24], which could obscure the
processes we want to reveal here.
9.3 Results
Figure 9.2 displays simulation results for this model, showing the stationary density
of B-individuals xB as a function of their initial density xB0 (see the Materials and
Methods section for full details about the simulations). Notably, there are large
differences to mean field theory, which predicts a heterogeneous population for a
much wider range of initial conditions. In order to understand this deviation we
have to consider the time evolution of the model. To that purpose, it is better to start
with a simpler case, setting one of the switching costs with so large a value that
it prevents the switching of individuals following that strategy. For example, let us
set cA 1 q, so that only B-individuals can change to option A. This switching
takes place when the proportion of B-individuals in the neighborhood of the focal
individual satisfies xQ B < q cB . Hence the subsequent dynamics exactly coincides
with the pruning of nodes of a standard site percolation process with unidirectional
dynamics (see Fig. 9.1), A- and B-individuals corresponding to empty and occupied
nodes, respectively.
When B-nodes become A-nodes, they leave other B-nodes with fewer B-
neighbors. The process repeats until it stabilizes to a subset of B-nodes, all of which
have .q cB / or more B-neighbors. When the size of this subset is a non-negligible
fraction of the size of the full graph, or infinite in the case of infinite graphs, then
percolation is said to occur [4]. The appearance of a percolating cluster constitutes
a phase transition, and it takes place when the initial density of occupied nodes is
larger than a critical density. In our case, the index of the percolation process of
B-individuals switching to option A, called kB , is given by
kB D dz.q cB /e: (9.3)
Herein, dxe denotes the smallest integer equal or larger than x. Conversely,
considering only the transitions of A-individuals to option B, we have another
percolation process with index kA , whose value is given by
kA D dz.1 q cA /e: (9.4)

9.3 Results 173
simulation
0.8 calculation
mean field theory
0.6
only A only B
X*
percolates percolates
0.4
both A and B
0.2 percolate
0
0 0.2 0.4 0.6 0.8 1
X0
Fig. 9.2 The fate of two options in contest is determined by the underlying process of multi-
percolation, taking place on the social network. The graph shows the stationary density of
B-individuals x as a function of the initial density x 0 (simulation results, black squares).
Parameter q D 0:5, so both options A and B are intrinsically equally good. Switching costs are also
symmetric, with values cA D cB D 0:25. As a result, both percolation indices have the same value
kA D kB D 3. Interactions take place in a regular random network of degree z D 9. The difference
with the prediction of mean theory (dashed line) demonstrates the crucial role played by the social
network. Labels indicate the possible regions of behavior, depending on the percolation of one
option, the other or both. Notice that a heterogeneous population is sustainable only when both
options percolate, but this case occurs for a significant range of initial conditions. Small arrows
near abscissa axes mark the critical density to attain percolation of each strategy, as predicted by a
calculation based on standard unidirectional percolation. The discrepancy with simulation results
highlights the fact that the mutual interference between both percolation processes changes the
percolation thresholds. This is confirmed by an analytical calculation that takes into account this
interplay (solid line)
Note that k D 0 and k D 1 are degenerate cases, whereas k D 2 is the index of

standard percolation and k 3 corresponds to bootstrap or k-core percolation.
The actual dynamics of our model is given by the competition between these
two percolation processes. The dynamics is therefore bi-directional, with both
transitions A-to-B and B-to-A taking place simultaneously. A calculation based on
standard unidirectional percolation, applied to each process separately, estimates
the percolation thresholds only poorly, as the arrows in Fig. 9.2 show. It is also
possible, however, to take into account the interference between both processes, with
a recursive calculation on the switching times of individuals. Figure 9.3 shows the
excellent agreement of this calculation with the computer simulations, and it clearly
demonstrates that mutual interference between both percolation processes occurs.
Interestingly, we find that this interplay supports the success of the dominated
Fig. 9.3 Stationary density x of B-individuals for a coordination game on a regular random
network of size N D 104 with switching costs cA D cB D 0:25, parameter q D 0:5, and initial
condition x 0 D 0:3. The theoretically determined values according to formulas (9.12) and (9.20)
match the values determined in computer simulations of the multi-percolation process perfectly
well, despite the complicated, ragged nature of the curve
option, i.e. it allows some individuals following the minor option to percolate
with initial conditions for which unidirectional percolation does not occur (range
0:36 . xB0 . 0:38 for option B, and range 0:62 . xB0 . 0:64 for option A).
Individuals who have switched obviously promote percolation of the newly acquired
option, as the switching increases the density of that option in the neighborhoods of
adjacent nodes. The time scale of switching for the major option is much faster
than for the minor one [25]. This implies that the pruning of nodes for the major
option is virtually done by the time the pruning of the minor option proceeds,
with the consequence that only changes of major to minor option have time to
effectively foster percolation of the latter option. More importantly, this analytical
theory confirms that the competition between options A and B gives rise to a bi-
directional percolation process, which allows a simple rationale for the outcome: In
the context of competing innovations, percolation of an option means survival and,
as long as it is the only one that percolates, it also implies dominance of the market.
We refer the reader to Sec. 9.4 for details of the analytical calculations and further
discussion.
The joint influence of switching costs and social networks becomes most salient
when one of the options is intrinsically superior to the other, i.e. when q ¤ 0:5.
Figure 9.4 shows an example, displaying again the asymptotic density of
B-individuals xB as a function of their initial density xB0 (see solid line and black
squares). In this case, the asymmetry of the game results in different percolation
indices for each option, namely kA D 4 and kB D 1 (see (9.3) and (9.4)), which
9.3 Results 175
mean field
0.8 T=0
T = 0.05
T = 0.07
T = 0.10
0.6 T = 0.17
T = 0.33
x*
T=1
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
x0
Fig. 9.4 Noise only has an effect on the multi-percolation process when the amount is large. The
graph shows the stationary density of B-individuals x as a function of the initial density x 0 , for
different amounts of uncertainty or noise in the decision rule of individuals (simulation results,
see line styles and symbols in legend). Compared to Fig. 9.2, in this case q D 0:25, so option
B is superior to option A. Switching costs are equal for both options, with values cA D cB D
0:2. The social network is a regular random network of degree z D 6. For T D 0 (no noise),
the asymmetry in intrinsic value between the options translates into different percolation indices,
kA D 4 and kB D 1, which causes different kinds of transitions to homogeneous population
(A- or B-dominated). This fact favors option B compared to the mean field prediction. Additional
curves show results for non-zero amounts of noise. Moderate noise does not change the result
qualitatively and, strikingly, larger amounts reinforce the superior option B rather than yielding an
more balanced outcome
causes a continuous transition towards an A-dominated population (xB0 . 0:1), but a

discontinuous one in the B-dominated case (xB0 & 0:2). The difference between both
transitions [26] originates from the characteristic transition of standard percolation,
in the former case, versus that of bootstrap percolation, in the latter. Interestingly,
the net effect of this imbalance between the two competing percolation processes
is a fostering of the superior option B. Note that if the same game, with or without
switching costs, took place on a well-mixed population, a symmetric outcome
around xB0 D q D 0:25 would result instead (see (9.2) and dashed red line in
Fig. 9.4). Let us finally address the issue of uncertainty or noise in the decision rule
of individuals. To this end, we assume that individuals choose options stochastically,
with a probability that follows a multi-nomial logit model [27, 28], which is also
known in physics as Fermi rule [29, 30]. Specifically, if the expected variation in
payoff resulting from a change of option is , then the probability of switching
strategy is assumed to be 1=.1 C exp.=T //. The parameter T determines the
amount of noise in the decision process. In the limit T ! 0, noise disappears

and we have the deterministic dynamics used so far. Additional curves in Fig. 9.4
display the influence of noise, showing that the qualitative behavior of the model
remains the same for low to moderate amounts of noise. It is striking, however, that
the evolution of the population is biased towards the superior option B, i.e. noise
reinforces the imbalance between options rather than washing it out.
9.4 Discussion
In conclusion, we have shown that the competition between several options gives
rise to bi-directional (or, more generally, multi-directional) percolation processes
which can be analytically understood. Multi-percolation thus provides a powerful
theoretical tool to understand the problem of competition between innovations in
networks. It offers predictions about the survival and dominance of the options
in contest, as well as insights into their dynamical evolution in time. Our general
finding is that percolation of an option implies its survival and, if only one option
percolates, it will be eventually supported by everyone. The latter may be favorable,
for example when it promotes a shared technical standard. Nevertheless, it could
also create monopolies or endanger pluralism. Our conclusions are expected to be
also relevant to the results of marketing and political campaigns, and to the diffusion
of opinions [11] and behavior [31]. Model variants or extensions may also describe
the spread of health behavior, such as obesity, smoking, depression or happiness
[32]. We recognize that the practical applicability of this theory requires a good
knowledge of the social network and a sound modeling of the decision making
process of individuals [33], but given the increasing availability of massive social
data, such information may be available for example systems, soon.
Appendix A: Methods
All the simulation results reported in the main text have been obtained according
to the procedures described in the following. Population size is 104 individuals.
Regular random networks are generated by randomly assigning links between
nodes, ensuring that each node ends up with exactly the same number of links.
Results have been obtained with synchronous update. This means that, first, the
next strategy is calculated for all individuals and, then, it is updated for them all at
once. We have also checked the influence of asynchronous update, which assigns
next strategies to individuals proceeding one by one in random order, but we have
found no significant qualitative difference in the results. The times of convergence
allowed for the population to reach a stationary state are 102 steps for the main
model (best response rule) and 103 steps for the model with noise (multi-nomial
logit or Fermi decision rule). We have verified that these are sufficiently large values.
9.4 Discussion 177
In the first case, population reaches a frozen state and, in the second one, results do
not change if a time of convergence 104 steps is used instead. Results plotted in
the graphs correspond, for each data point, to an average of 100 realizations. Each
realization is carried out with a newly generated random network, where strategies
were randomly assigned to individuals in accordance with the initial densities of
both strategies.
Appendix B: Analytical Theory
We study a coordination game with switching costs. A-individuals have the

following payoff matrix
q 0
; (9.5)
cA 1 q cA
while B-individuals have
q cB cB
: (9.6)
0 1q
All three parameters satisfy q; cA ; cB 2 Œ0; 1.
We call x t the global density of B-individuals in the population at time step t,
x its initial value and x its asymptotic value (x D limt !1 x t ). At time t D 0
0
individuals are randomly assigned options A or B, with probabilities xA0 D .1 x 0 /

or xB0 D x 0 respectively. Given an individual i , we call xQ i the proportion of B-
individuals in her neighborhood. Then, an A-individual i changes to option B when
her neighborhood satisfies
xQ i > q C cA ; (9.7)
whereas a B-individual j switches strategy when
xQ j < q cB : (9.8)
The above strategy switchings correspond to the removal of nodes (pruning) of

two competing site percolation processes, whose indices are respectively
kA D dz.1 q cA /e (9.9)
and
kB D dz.q cB /e; (9.10)
wherein dxe denotes the smallest integer equal or larger than x.
In the pruning of nodes associated with a site percolation process, the nodes that
remain in the stationary state are those which have k or more neighbors of the same
strategy and hence they belong to a percolating cluster of that strategy. We call
pA and pB the probabilities that a node belongs to a percolating cluster assuming
independent percolation processes for transitions A ! B and B ! A, respectively.
Obviously, pA xA D 1 x 0 and pB xB D x 0 . Both probabilities can be

calculated for infinite Bethe lattices [5, 6]. They offer a rough approximation of the
behavior of the model, assuming no interference between both processes, which
yields the following prediction for the asymptotic density of B-individuals
x D 1 x 0 C pB pA : (9.11)
This approximation obviously fails to account for the change in the percolation
probabilities that the simulation results reflect. Therefore, we propose the following
analytical theory, which takes into account the interplay between both competing
percolation processes.
Let us first define a simplified version of the model, which we call the 1T-model.
This model is the same as the original one, with the only difference that each node
can only switch strategy once. That is, nodes change strategy according to (9.7) and
(9.8), but once they have changed, they stick to the new strategy, no matter what
(9.7) and (9.8) dictate. Note that this idea can be generalized to a nT-model, where
each node is allowed to switch at most n times. Our original model could also be
called 1T-model, corresponding to an unbounded value of n.
In fact, we can assume in the original model that most nodes will switch only
once or never. For a node to switch, it is required that the number of neighbors of
the same strategy is below a given threshold k. Once it has switched, it will have
a number of neighbors of the (new) strategy larger than z k (z is the degree of
the network), which usually means a large number of them. So it is reasonable to
expect that the node will be locked in the new strategy forever, as it would require
a large number of changes in its neighborhood to switch back. The great similarity
between the simulation results of the original 1T-model and the 1T-model supports
this intuition (see Figs. 9.5–9.8).
The 1T-model can be calculated for an infinite Bethe lattice, with a recursive
procedure that we present in the following. From now on, let us denote by X one
of the options, A or B, and by Y the other one, i.e. Y is the only element of the
set fA; Bg fX g. The index of the percolation process with transitions X ! Y is
denoted kX . The fundamental recursive property of the 1T-model is that the time
of switching of any X -node is 1C (the kX -th greater time of switching among its
X -neighbors). For example, with a percolation process for transitions B ! A with
index kB D 3, if a B-node has six neighbors that are all B-nodes and whose times
of switching are t D 2, 2, 3, 5, 7, 7, respectively, then the node will switch at
time t D 6. Notice that if an X -node belongs to a percolating cluster then its time
of switching is unbounded, because it has kX or more neighbors that also have
unbounded switching times.
First, we calculate the switching probabilities of a node, conditioned to the event
of being connected to a node of the same or the other type. Thus, we define rXt as
the probability of a node being of type X and switching at time t, conditioned to
the event of being connected to an X -node until time t 1, with t 1. Similarly,
we define sXt as the probability of a node being of type X and switching at time t,
9.4 Discussion 179
conditioned to the event of being connected to a Y -node, i.e. of the opposite type,
until time t 1, with t 1.
Second, we calculate the probabilities of a node being of type X and switching at
time t, which we represent by pXt , given the conditional probabilities of switching
of its child nodes until time t 1, namely rX1 ; : : : ; rXt 1 and sY1 ; : : : ; sYt 1 .
Third, in the stationary state X -nodes will be either nodes that have been X -
nodes since the beginning or initial Y -nodes that have switched option. Hence the
stationary density of B-nodes x can be expressed as
1
X 1
X
x D x0 pBt C pAt : (9.12)
t D1 t D1
To calculate the probabilities frAt ; sAt ; pAt ; rBt ; sBt ; pBt g we need to consider all the
possible configurations of neighborhood of a node. To this end, we classify the
neighbors at time t into one of these types:
[1] Nodes of the same type that switch at time t 1.
[2] Nodes of the same type that switch at time t or later.
[3] Nodes of the same type that switch at time t 2 or before.
[4] Nodes of the other type that switch at time t or later.
[5] Nodes of the other type that switch at time t 1 or before.
Note that neighbors of type Œ1 are the ones that trigger switching of the focal node
at time t. For a given configuration, the number of neighbors of each type appears as
an exponent in the corresponding combinatorial expression. In (9.13)–(9.21) below,
we use the exponents i; j; k; m; n for types Œ1 to Œ5, Prespectively. P
In addition, we define RX0
D SX0 D 0, and RX t
D t D1 rX and SXt D t D1 sX ,
for t 1. The degree of the Bethe lattice is z. The calculation proceeds according
to the following (9.13)–(9.21)
kX
!
X 2
1 z1
rX D .xX0 /j C1 .xY0 /z1j ; (9.13)
j D0
j
kX
!
X 1
z1
sX1 D .xX0 /j C1 .xY0 /z1j : (9.14)
j D0
j
For t > 1
!
X z1
rXt D xX0 .r t 1 /i .xX0 RX
t 1 j t 2 k
/ .RX / .xY0 SYt 1 /m .SYt 1 /n ;
i; j; k; m; n X
.i;j;k;m;n/2R
! (9.15)
X z1
sXt D xX0 .r / .xX RX / .RX / .xY SY / .SYt 1 /n ;
t 1 i 0 t 1 j t 2 k 0 t 1 m
i; j; k; m; n X
.i;j;k;m;n/2S
(9.16)
where the sets of exponents R and S are
R D f .i; j; k; m; n/ 2 f0; 1; 2; : : : ; z 1g5 W i C j C k C m C n D z 1;

i C j C k kX 1;
i C j C n kX 1;
j C n kX 2 g; (9.17)
S D f .i; j; k; m; n/ 2 f0; 1; 2; : : : ; z 1g5 W i C j C k C m C n D z 1;

i C j C k kX ;
i C j C n kX ;
j C n kX 1 g: (9.18)
kX
!
X 1
z
pX1 D .xX0 /j C1 .xY0 /zj : (9.19)
j D0
j
For t > 1
!
X z
pXt q D xX0 .r t 1 /i .xX0 RX
t 1 j t 2 k 0
/ .RX / .xY SYt 1 /m .SYt 1 /n ;
i; j; k; m; n X
.i;j;k;m;n/2P
(9.20)
where the set of exponents P is
P D f .i; j; k; m; n/ 2 f0; 1; 2; : : : ; zg5 W i C j C k C m C n D z;

i C j C k kX ;
i C j C n kX ;
j C n kX 1 g: (9.21)
As the probabilities of switching decay exponentially with time, it is enough to

calculate a finite number of terms in the sums of (9.12). For the results reported
in this work 100 terms were used, which yields an excellent agreement with the
computer simulations (see Figs. 9.5–9.8).
Notice that (9.15)–(9.18) assume that nodes of the other strategy that have
switched (exponent n) have done so before or at the same time as those of the
own strategy (exponent k). This simplification thus avoids considering the full set
of possible histories of switchings in times previous to t 1. This assumption is
based on the separation of time scales between the switchings of both percolation
processes. For unidirectional percolation, the probability of a node switching at
time t decreases exponentially with the difference between the initial density
and the critical percolation density. That is, the nearer the systems starts to the
percolation threshold, the slower is the pruning of nodes [25]. As a consequence,
9.4 Discussion 181
simulation
simulation 1T-model
0.8 calculation
calc. without interference
0.6
x*
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
x0
Fig. 9.5 Comparison of simulation with analytical results. Asymptotic density of B-nodes x as a
function of the initial density x 0 . Model parameters are the same as in Fig. 9.2: relative preference
q D 0:5, switching costs cA D cB D 0:25, and network degree z D 9. The corresponding indices
of the percolation processes are kA D kB D 3
with bidirectional percolation, when an option is near its percolation threshold the
other is well beyond it, so that the switching of the latter will be exponentially
faster, which supports the assumption above. It is interesting to point out that the
proposed recursive scheme based on the switching times can be modified to carry
out a calculation neglecting interference, then becoming equivalent to (9.11). The
modification consists in setting sXt D 0, for any t 1, instead of using (9.14)
and (9.16).
Figures 9.5 and 9.6 show the results presented in Figs. 9.2 and 9.4 of the main
text, but this time including also the simulation results for the 1T-model and the
analytical results obtained with the calculation without interference. These figures
show the close similarity between the original model and the 1T-model, which
confirms that most nodes that switch in the original model do so only once. They
also demonstrate the high accuracy of the proposed approximation, as compared to
the calculation neglecting interference, which fails to reflect the actual percolation
thresholds properly. Figures 9.7 and 9.8 display two additional examples.
Finally, we want to point out that, strictly speaking, the networks considered
in the analytical calculations (namely, infinite Bethe lattices) differ from the ones
used in the computer simulations (finite regular random networks). Apart from the
different size, in the former case there are no closed paths or loops, whereas in the
latter there exists a (low) number of them. Figures 9.5–9.8 show, however, that these
0.8 simulation
simulation 1T-model
calculation
0.6
x*
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
x0
Fig. 9.6 Comparison of simulation with analytical results. Asymptotic density of B-nodes x as a
function of the initial density x 0 . Model parameters are the same as in Fig. 9.4: relative preference
q D 0:25, switching costs cA D cB D 0:2, and network degree z D 6. The corresponding indices
of the percolation processes are kA D 4, kB D 1
simulation
simulation 1T-model
0.8 calculation
0.6
x*
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
x0
Fig. 9.7 Comparison of simulation with analytical results. Asymptotic density of B-nodes x as
a function of the initial density x 0 . Model parameters: relative preference q D 0:5, switching
costs cA D cB D 0:2, and network degree z D 11. The indices of the percolation processes are
kA D kB D 4
References 183
0.8
0.6
x*
0.4
simulation
simulation 1T-model
calculation
0.2
0
0 0.2 0.4 0.6 0.8 1
x 0
Fig. 9.8 Comparison of simulation with analytical results. Asymptotic density of B-nodes x as
a function of the initial density x 0 . Model parameters: relative preference q D 0:3, switching costs
cA D cB D 0:2, and network degree z D 6. The indices of the percolation processes are kA D 3,
kB D 1
differences in network topology do not produce a significative discrepancy between

computational and analytical results.
Acknowledgements C. P. R. and D. H. were partially supported by the Future and Emerging

Technologies programme FP7-COSI-ICT of the European Commission through project QLectives
(grant no. 231200).
References
1. M. Sahimi, Applications of Percolation Theory (Taylor & Francis, PA, 1994)

2. J. Shao, S. Havlin, H.E. Stanley, Dynamic opinion model and invasion percolation. Phys. Rev.
Lett. 103, 018701 (2009)
3. J. Goldenberg, B. Libai, S. Solomon, N. Jan, D. Stauffer, Maketing percolation. Physica A
284, 335–347 (2000)
4. D. Stauffer, A. Aharony, Introduction to Percolation Theory, second edn. (Taylor & Francis,
Philadelphia, 1991)
5. J. Chalupa, P.L. Leath, G.R. Reich, Bootstrap percolation on a bethe lattice. J. Phys. C: Solid
State Phys. 12, L31–L35 (1979)
6. S.N. Dorogovtsev, A.V. Goltsev, J.F.F. Mendes, k-core organization of complex networks.
Phys. Rev. Lett. 96, 040601 (2006)
7. W.B. Arthur, Competing technologies, increasing returns, and lock-in by historical events.
Econ. J. 99, 116–131 (1989)
8. E.M. Rogers, Diffusion of Innovations, 5th edn. (Simon and Schuster, NY, 2003)
9. H. Amini, M. Draief, M. Lelarge, Marketing in a random network. Network Contr. Optim.
LNCS 5425, 17–25 (2009)
10. P. Klemperer, Competition when consumers have switching costs: An overview with
applications to industrial organization, macroeconomics, and international trade. Rev. Econ.
Stud. 62, 515–539 (1995)
11. M.S. Granovetter, Threshold models of collective behavior. Am. J. Sociol. 83, 1420–1443
(1978)
12. R.B. Myerson, Game Theory: Analysis of Conflict (Harvard University Press, Cambridge,
1991)
13. B. Skyrms, The Stag Hunt and the Evolution of Social Structure (Cambridge University Press,
Cambridge, 2003)
15. H. Gintis, Game Theory Evolving: A Problem-Centered Introduction to Modeling Strategic
Interaction, 2nd edn. (Princeton University Press, Princeton, 2009)
16. J. Hofbauer, K. Sigmund, Evolutionary Games and Population Dynamics (Cambridge
17. D. Helbing, A mathematical model for behavioral changes through pair interactions, in
Economic Evolution and Demographic Change ed. by G. Hagg, U. Mueller, K.G. Troitzsch
(Springer, Berlin, 1992), pp. 330–348
18. S. Wasserman, K. Faust, Social Network Analysis: Methods and Applications (Cambridge
19. F. Vega-Redondo, Complex Social Networks (Cambridge University Press, Cambridge, 2007)
20. M. Nakamaru, S.A. Levin, Spread of two linked social norms on complex interaction networks.
J. Theor. Biol. 230, 57–64 (2004)
21. A. Galeotti, S. Goyal, M.O. Jackson, F. Vega-Redondo, L. Yariv, Network games. Rev. Econ.
Studies 77, 218–244 (2010)
22. M.E.J. Newman, D.J. Watts, S.H. Strogatz, Random graph models of social networks. Proc.
Natl. Acad. Sci. U.S.A. 99, 2566–2572 (2002)
23. C.P. Roca, J.A. Cuesta, A. Sánchez, Evolutionary game theory: Temporal and spatial effects
beyond replicator dynamics. Phys. Life Rev. 6, 208–249 (2009)
24. G. Szabó, G. Fáth, Evolutionary games on graphs. Phys. Rep. 446, 97–216 (2007)
25. G. Grimmett, Percolation, 2nd edn. (Springer, Berlin, 1999)
26. D. Achlioptas, R.M. D’Souza, J. Spencer, Explosive percolation in random networks. Science
323, 1453–1455 (2009)
27. D. McFadden, Conditional logit analysis of qualitative choice behavior, in Frontiers of
Econometrics, ed. by P. Zarembka (Academic Press, New York, 1974), pp. 105–142
28. J.K. Goeree, C.A. Holt, Stochastic game theory: For playing games, not just for doing theory.
Proc. Natl. Acad. Sci. U.S.A. 96, 10564–10567 (1999)
29. L.E. Blume, The statistical mechanics of strategic interaction. Games Econ. Behav. 5, 387–424
(1993)
30. A. Traulsen, M.A. Nowak, J.M. Pacheco, Stochastic dynamics of invasion and fixation. Phys.
Rev. E 74, 011909 (2006)
31. D. Centola, The spread of behavior in an online social network experiment. Science 329,
1194–1197 (2010)
32. K.P. Smith, N.A. Christakis, Social networks and health. Annu. Rev. Sociol. 34, 405–429
(2008)
33. A. Traulsen, D. Semmann, R.D. Sommerfeld, H.-J. Krambeck, M. Milinski, Human strategy
updating in evolutionary games. Proc. Natl. Acad. Sci. U.S.A. 107, 2962–2966 (2010)
Chapter 10
Heterogeneous Populations: Coexistence,
Integration, or Conflict
10.1 Introduction
In order to gain a better understanding of factors preventing or promoting cooper-

ation among individuals, biologists, economists, social scientists, mathematicians
and physicists have intensively studied game theoretical problems such as the
prisoner’s dilemma and the snowdrift game (also known as chicken or hawk-dove
game) [1–3]. These games have in common that a certain fraction of people or even
everyone is expected to behave uncooperatively (see Fig. 10.1). Therefore, a large
amount of research has focused on how cooperation can be supported by mecha-
nisms such as [3] repeated interactions [1], reputation [4], clusters of cooperative
individuals [5], costly punishment [6–12], or success-driven migration [13].
Unfortunately, comparatively little attention has been devoted to the problem of
cooperation between groups with different preferences (e.g. people of different gen-
der, status, age, or cultural background). Yet, what constitutes cooperative behavior
for one group might be considered non-cooperative by another. For example, men
and women appear to have different preferences many times, but they normally
interact among and between each other on a daily basis. It is also more and more
common that people with different religious beliefs live and work together, while
their religions request some mutually incompatible behaviors (in terms of working
days and free days, food one may eat or should avoid, headgear, or appropriate
clothing, etc.). A similar situation applies, when people with different mother
tongues meet, or businessmen from countries with different business practices make
a deal. Is it possible to identify factors determining whether two such populations
go their own way, find a common agreement, or end up in conflict? And what is

The content of this chapter has some overlap with the following two papers, which should be
cited instead: D. Helbing and A. Johansson, Evolutionary dynamics of populations with conflicting
interactions: Classification and analytical treatment considering asymmetry and power. Physical
Review E 81, 016112 (2010); D. Helbing and A. Johansson, Cooperation, norms, and revolutions:
A unified game-theoretical approach. PLoS ONE 5(10): e12530.

186 10 Heterogeneous Populations: Coexistence, Integration, or Conflict
Fig. 10.1 Illustration of the parameter-dependent types and outcomes of symmetrical 2 2 games
in a single population [14]. For prisoner’s dilemmas (PD) and two-person public goods games,
we have B < 0 and C < 0, and the expected outcome is defection by everybody. For snowdrift
games (SD), which are also called chicken or hawk-dove games, we have B > 0 and C < 0.
The stable stationary solution corresponds to a coexistence of a fraction p0 D jBj=.jBj C jC j/
of cooperators with a fraction 1 p0 of defectors (i.e. non-cooperative individuals). For harmony
games (HG), which are sometimes called coordination games as well, we have B > 0 and C > 0,
and everybody will eventually cooperate. Finally, for stag hunt games (SH), which are also called
assurance games, we have B < 0 and C > 0, and there is a bistable situation: If the initial fraction
p.0/ of cooperators is larger than p0 D jBj=.jBj C jC j/, everybody is expected to cooperate in
the end, otherwise everybody will eventually behave uncooperatively [15]. The arrows illustrate
different routes to cooperation [3, 15]. Route 1 belongs to the way in which kin selection, network
reciprocity, or group selection modify the payoff structure of the game. Route 2a corresponds
to the effect of direct reciprocity (due to the “shadow of the future” through the likelihood of
future interactions). Route 2b reflects the mechanism of indirect reciprocity (based on reputation
effects), and route 2c corresponds to costly punishment. Route 3 results for certain kinds of network
interactions [16]
the relevance of power in the rivalry of populations? Differences in power can, for
example, result from different sizes of the interacting populations, their material
resources (money, weapons, etc.), social capital (status, social influence, etc.), and
other factors (charisma, moral persuasion, etc.).
10.2 Model
As a mathematical approach to this problem, we propose game-dynamical repli-

cator equations for multiple populations [17, 18]. The crucial point is to adjust
them in a way that reflects interactions between individuals with incompatible
preferences (see Methods and [19]). These equations describe the time evolution
10.2 Model 187
of the proportions p.t/ and q.t/ of cooperative individuals in populations 1 and 2,

respectively, as individuals imitate more successful behaviors in their own popula-
tion. The success depends on the “payoffs” resulting from social interactions, i.e.,
on the own behavior and the behavior of the interaction partners.
In order to reflect incompatible interests of both populations, we assume that
population 1 prefers behavior 1 (e.g. everybody should be undressed at the beach),
while population 2 prefers behavior 2 (everybody should be properly dressed). If an
interaction partner shows the behavior preferred by oneself, this behavior is called
“cooperative”, otherwise uncooperative. In other words, behavior 1 is cooperative
from the viewpoint of population 1, but uncooperative from the viewpoint of
population 2 (and vice versa). Furthermore, if an individual of population 1 interacts
with an individual of population 2 and both display same behavior, we call this
behavior “coordinated”. Finally, if the majority of individuals in both populations
shows a coordinated behavior, we speak of “normative behavior” or a “behavioral
norm”.
To establish a behavioral norm, the individuals belonging to one of the pop-
ulations have to act against their own preferences, in particular as we assume
that preferred behaviors and population sizes do not change. (Otherwise, identical
behavior in both populations could simply result from the adaptation of preferences
or group membership, which can, of course, further promote consensus in reality.)
It is very interesting to study under what conditions interactions within and
between populations with incompatible preferences can lead to cooperation, con-
flict, or “normative behavior”. To address this question, we adjust the theoretical
framework of game-dynamical replicator equations to the case of two populations
playing 2 2 games, which are represented by four payoffs T , R, P , and S . In the
prisoner’s dilemma, for example, the meaning of these parameters is “Temptation”
to behave non-cooperatively, “Reward” for mutual cooperation, “Punishment” for
mutual non-cooperative behavior and “Sucker’s payoff” for a cooperative individual
meeting an uncooperative one. The related game-dynamical replicator equations
read
dp.t/
D p.t/Œ1 p.t/F p.t/; q.t/ (10.1)
dt
and
dq.t/
D q.t/Œ1 q.t/G p.t/; q.t/ ; (10.2)
dt
where the terms p.1 p/ and q.1 q/ can be interpreted as saturation factors. They
make sure that the proportions p.t/ and q.t/ of individuals pursuing their preferred
strategies stay within the range from 0 to 1. F .p; q/ and G.p; q/ are functions
reflecting the interactions between individuals. They include terms describing “in-
group” interactions (“self-interactions”, reflecting encounters with individuals of
the same population) and “out-group” interactions (among individuals belonging to
different populations) (see Methods).
For simplicity, we will focus on the case where both populations play the
same game. Then, the functions F and G only depend on the payoff-dependent
parameters B D S P , C D R T , and the relative power f of population 1 (see
Methods). Furthermore, we specify the power here by the relative population size
(see [19] for details). The parameter C may be interpreted as gain of coordinating
on one’s own preferred behavior (if greater than zero, otherwise as loss). B may be
viewed as gain when giving up coordinated, but non-preferred behavior. Despite the
simplifications made, this model for two populations with incompatible preferences
shows very interesting features, and it can be generalized in numerous ways (to
consider more populations, to treat heterogeneous, but compatible interactions, to
reflect that payoffs in out-group interactions may differ from in-group interactions,
to consider migration between populations, spatial interactions, learning, punish-
ment, etc.).
10.3 Results
We find that social interactions with incompatible interests do not necessarily pro-
duce conflict – they may even promote mutual coordination. Depending on the signs
of B and C , which determine the character of the game, we have four archetypical
situations: (1) The case B < 0 and C < 0 applies to the multi-population prisoner’s
dilemma (MPD), (2) in the multi-population harmony game (MHG), we have B > 0
and C > 0, (3) the multi-population snowdrift game (MSD) is characterized by
B > 0 and C < 0, and (4) in the multi-population stag hunt game (MSH), we
have B < 0 and C > 0. In the multi-population harmony game, everybody shows
cooperative behavior, while in the multi-population prisoner’s dilemma, everybody
is uncooperative in the end, as one may expect (see Fig. 10.2). However, as we
will show in the following, the dynamics and outcome of the multi-population
stag hunt and snowdrift games are in marked contrast to the one-population case.
This can be demonstrated by mathematical analysis of the stationary solutions of
(10.1) and (10.4) and their stability properties (see Methods and [19]). However,
our results can be more intuitively illustrated by figures and movies showing the
evolutionary equilibria of the games for various parameter values B and C , their
basins of attraction, and representative flow lines. Details are discussed below and
in the captions of Figs. 10.3 and 10.4.
10.3.1 Evolution of Normative Behavior in the Stag Hunt Game
The one-population stag hunt game is characterized by an equilibrium selection

problem [20]: Everyone is finally expected to cooperate, if the initial fraction of
cooperative individuals is above p0 D jBj=.jBj C jC j/, otherwise everybody
will behave uncooperative in the end [15]. The same applies to non-interacting
10.3 Results 189
Prisoner’s Dilemmas Harmony Games

a 1 b
q(t) 0.75
0.5
0.25
0
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
p(t) p(t)
Fig. 10.2 Vector fields (small arrows), sample trajectories (large arrows) and phase diagrams
(colored areas) for two interacting populations with incompatible preferences, when 80% of
individuals belong to population 1 (f D 0:8). (a) If B D C D 1, individuals are facing
prisoner’s dilemma interactions, and everybody ends up with non-cooperative behavior in each
population. (b) If B D C D 1, individuals are playing harmony games instead, and everybody
will eventually behave cooperatively. The results look similar when the same two-population games
are played with different values of f , jBj or jC j
populations. However, in the multi-population stag hunt game with incompatible

preferences and no self-interactions, it never happens that everybody or nobody
cooperates in both populations (otherwise there should be yellow or red areas in the
second part of Movie S2). Although both populations prefer different behaviors, all
individuals end up coordinating themselves on a commonly shared behavior. This
can be interpreted as self-organized evolution of a behavioral norm [21].
If self-interactions are taken into account as well, the case where everybody
or nobody cooperates in both populations is possible, but rather exceptional (see
Fig. 10.3). It requires that both populations have similar strengths (f 1=2)
and the initial levels of cooperation are comparable as well (see yellow area
in Fig. 10.3b). Under such conditions, both populations may develop separate
subcultures. Normally, however, both populations establish a commonly shared
norm and either end up with behavior 1 (green area in Fig. 10.3) or with behavior 2
(blue area).
Due to the payoff structure of the multi-population stag hunt game, it can be
profitable to coordinate oneself with the prevailing behavior in the other population.
Yet, the establishment of a norm requires the individuals of one population to
give up their own preferred behavior in favor of the one preferred by the other
population. Therefore, it is striking that the preferred behavior of the weaker
population can actually win through and finally establish the norm (see blue areas
in Figs. 10.3a,c,d). Who adapts to the preferred strategy of the other population
essentially depends on the initial fractions of behaviors. The majority behavior is
likely to determine the resulting behavioral norm, but a powerful population is in a
favorable position: The area of possible histories leading to an establishment of the
a b
1
q(t) 0.75
0.5
0.25
0
c1 d
0.75
q(t)
0.5
0.25
0
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
p(t) p(t)
(colored areas) for two interacting populations with incompatible preferences, playing stag hunt
games with B < 0 and C > 0 (right). p is the fraction of individuals in population 1 showing their
preferred, cooperative behavior 1, and q is the fraction of cooperative individuals in population
2 showing their preferred behavior 2. The vector fields show .dp=dt; dq=dt /, i.e. the direction
and size of the expected temporal change of the behavioral distribution, if the fractions of
cooperative individuals in populations 1 and 2 are p.t / and q.t /. Sample trajectories illustrate some
representative flow lines .p.t /; q.t // as time t passes. The flow lines move away from unstable
stationary points (empty circles or dashed lines) and are attracted towards stable stationary points
(black circles or solid diagonal lines). The colored areas represent the basins of attraction, i.e.
all initial conditions .p.0/; q.0// leading to the same fix point [yellow D (1,1), blue D (0,1),
green D (1,0)]. Saddle points (crosses) are attractive in one direction, but repulsive in another. The
model parameters are as follows: (a) jBj D jC j D 1 and f D 0:8, i.e. 80% of all individuals
belong to population 1, (b) jC j D 2jBj D 2 and f D 1=2, i.e. both populations are equally
strong, (c) jC j D 2jBj D 2 and f D 0:8, (d) 2jC j D jBj D 2 and f D 0:8. In the
multi-population stag hunt game (MSH), due to the asymptotically stable fix points at (1,0) and
(0,1), all individuals of both populations finally show the behavior preferred in population 1 (when
starting in the green area) or the behavior preferred in population 2 (when starting in the blue
area). This case can be considered to describe the evolution of a shared behavioral norm. Only for
similarly strong populations (f 1=2) and similar initial fractions p.0/ and q.0/ of cooperators
in both populations (yellow area), both populations will end up with population-specific norms
(“subcultures”), corresponding to the asymptotically stable point at (1,1). The route towards the
establishment of a shared norm may be quite unexpected, as the flow line starting with the white
circle shows: The fraction q.t / of individuals in population 2 who are uncooperative from the
viewpoint of population 1 grows in the beginning, but later on it goes down dramatically. Therefore,
a momentary trend does not allow to easily predict the final outcome of the struggle between two
interest groups
10.3 Results 191
norm preferred by population 1 tends to increase with power f (compare the size
of the green areas in Figs. 10.3bCc).
Note that the evolution of norms is one of the most fundamental challenges in
the social and economic sciences. Norms are crucial for society, as they reduce
uncertainty, bargaining efforts, and conflict in social interactions. They are like
social forces guiding our interactions in numerous situations and subtle ways,
creating an “invisible hand” kind of self-organization of society [21].
Researchers from various disciplines have worked on the evolution of norms,
often utilizing game-theoretical concepts [22–26], but it has been hard to reveal the
conditions under which behavioral consensus[27] is established. This is so, because
cooperation norms, in contrast to coordination norms, are not self-enforcing. In
other words, there are incentives for unilateral deviance. Considering the fact that
norms require people to constrain self-interested behavior [28] and to perform
socially prescribed roles, the ubiquity of norms is quite surprising. Yet, widespread
cooperation-enhancing mechanisms such as group pressure can transform prisoner’s
dilemmas into stag hunt interactions [3,15,29] (see Fig. 10.1). This creates a natural
tendency towards the formation of norms, whatever their content may be.
Our model sheds new light on the problem of whether, why and how a norm can
establish. In particular, it reveals that the dynamics and finally resulting state of the
system is not just determined by the payoff structure. It also crucially depends on the
power of populations and even on the initial proportions of cooperative individuals
(the “initial conditions”).
10.3.2 Occurrence of Conflict in the Snowdrift Game
In the one-population snowdrift game, there is one stable stationary point, corre-
sponding to a fraction p0 D jBj=.jBj C jC j/ of cooperative individuals [15]. If
this were transferable to the multi-population case, we should have p D q D p0 in
the limit of large times t ! 1. Instead, we find a variety of different outcomes,
depending on the values of the model parameters B, C , and f :
(a) The interactions between both populations shift the fraction of cooperative
individuals in each population to values different from p0 . If jBj D jC j, we
discover a line of infinitely many stationary points, and the actually resulting
stationary solution uniquely depends on the initial condition (see Fig. 10.4a).
This line satisfies the relation q D p only if f D 1=2, while for most parameter
combinations we have q ¤ p ¤ p0 . Nevertheless, the typical outcome in the
case jBj D jC j is characterized by a finite fraction of cooperative individuals
in each population.
(b) Conflicting interactions between two equally strong groups destabilize the
stationary solution q D p D p0 of the one-population case, and both
populations lose control over the final outcome. For jBj ¤ jC j, all stationary
points are discrete and located on the boundaries, and only one of these points is
a 1 b
0.75
q(t)
0.5
0.25
0
c 1 d
0.75
0.5
q(t)
0.25
0
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
p(t) p(t)
(colored areas) for two interacting populations with incompatible preferences, playing snowdrift
games with B > 0 and C < 0 (right). The representation is the same as in Fig. 10.3. In particular,
the colored areas represent again the basins of attraction, i.e. all initial conditions .p.0/; q.0//
leading to the same fix point [red D (0,0), salmon D .u; 0/, mustard D .v; 1/, rainbow colors D
.u; v/, with 0 < u; v < 1]. The model parameters are as follows: (a) jBj D jC j D 1 and f D 0:8,
i.e. 80% of all individuals belong to population 1, (b) jC j D 2jBj D 2 and f D 1=2, i.e.
both populations are equally strong, (c) jC j D 2jBj D 2 and f D 0:8, (d) 2jC j D jBj D 2
and f D 0:8. (a) In the multi-population snowdrift game (MSD), a mixture of cooperative and
uncooperative behaviors results in both populations, if jBj D jC j. (b) For jBj < jC j and equally
strong populations, everybody ends up with non-cooperative behavior in each population. (c) For
jBj < jC j and f 1=2 0, the weaker population 2 solidarizes with the minority of the
stronger population 1 and opposes its majority. (d) Same as (c), but now, all individuals in the
weaker population 2 show their own preferred behavior after the occurrence of a discontinuous
(“revolutionary”) transition of the evolutionary equilibrium from .u; 0/ to .v; 1/
an evolutionary equilibrium. If both populations have equal power (f D 1=2),

we always end up with non-cooperative behavior by everybody (if p0 < 1=2,
see Fig. 10.4b), or everybody is cooperative (if p0 > 1=2). Remarkably, there is
no mixed stable solution between these two extremes.
(c) The stronger population gains control over the weaker one, but a change of
the model parameters may induce a “revolutionary” transition. If jBj ¤ jC j
and population 1 is much stronger than population 2 (i.e., f 1=2 0),
we find a finite fraction of cooperative individuals in the stronger population,
while either 0% or 100% of the individuals are cooperative in the weaker
population. A closer analysis reveals that the resulting overall fraction of
10.4 Discussion 193
cooperative individuals fits exactly the expectation p0 of the stronger population

(see Methods), while from the perspective of the weaker population, the overall
fraction of cooperative individuals is largely different from p0 D jBj=.jBj C
jC j/. Note that the stronger population alone can not reach an overall level of
cooperation of p0 . The desired outcome can only be produced by effectively
controlling the behavior of the weaker population. This takes place in an
unexpected way, namely by polarization: In the weaker population 2, everyone
shows behavior 1 for p0 < 1=2 (see Fig. 10.4c), otherwise everyone shows
behavior 2 (see Fig. 10.4d). There is no solution in between these two extremes
(apart from the special case p0 D 1=2 for jBj D jC j).
It comes as a further surprise that the behavior in the weaker population is always
coordinated with the minority behavior in the stronger population. Due to the payoff
structure of the multi-population snowdrift game, it is profitable for the weaker
population to oppose the majority of the stronger population, which creates a tacit
alliance with its minority. Such antagonistic behavior is well-known from protest
movements [30] and finds here a natural explanation.
Moreover, when jC j changes from values greater than jBj to values smaller than
jBj, there is an unexpected, discontinuous transition in the weaker population 2
from a state in which everybody is cooperative from the point of view of population
1 to a state in which everybody shows the own preferred behavior 2 (see Movie
S1 and Methods). History and science [31] have seen many abrupt regime shifts of
this kind. Revolutions caused by class conflict provide ample empirical evidence
for their existence. Combining the theory of phase transitions with “catastrophe
theory” offers a quantitative scientific approach to interpret such revolutions as the
outcome of social interactions [32]. Here, their recurrence becomes understandable
in a unified and simple game-theoretical framework.
10.4 Discussion
Multi-population game-dynamical replicator equations provide an elegant and

powerful approach to study the dynamics and outcomes expected for groups with
incompatible interests. A detailed mathematical analysis reveals how interactions
within and between groups can substantially change the dynamics of various game
theoretical dilemmas. Generalizations to more than two behaviors or groups and to
populations with heterogeneous preferences are easily possible.
When populations with incompatible preferences interact among and between
each other, the signs of the payoff-dependent parameters B and C determine the
character of the game. The snowdrift and stag hunt games show a particularly rich
and interesting dynamics. For example, there is a discontinuous (“revolutionary”)
transition, when 1jBj=jC j changes its sign. On top of this, the power f has a major
influence on the outcome, and the initial distribution of behaviors can be crucial.
Note that such a rich system behavior is already found for the simplest setting
of our model and that the concept of multi-population game-dynamical equations
can be generalized in various ways to address a number of challenging questions
in the future: How can we gain a better understanding of a clash of cultures, the
outbreak of civil wars, or conflicts with ethnic or religious minorities? How can
we analytically study migration and group competition? When do social systems
become unstable and face a polarization of society? How can we understand the
emergence of fairness norms in bargaining situations?
Another interesting aspect of our model is that it makes a variety of quantitative
predictions. Therefore, it could be tested experimentally with iterated games in
the laboratory, involving several groups of people with random matching and
sufficiently many iterations. Suitable changes in the payoff matrices should be able
to confirm the mathematical conditions under which different archetypical types
of social phenomena or discontinuous transitions in the system behavior can occur
(see Methods and [19]): (1) the breakdown of cooperation, (2) in-group cooperation
(the formation of “sub-cultures”), (3) societal polarization and conflict with the
possibility of discontinuous regime shifts (“revolutions”), and (4) the evolution of
shared behavioral norms.
In the past, the problem of the evolution of norms has often been addressed by
studying one-population prisoner’s dilemma situations [21, 24], and by assuming
cooperation-enhancing mechanisms such as repeated interactions (a “shadow of the
future”, as considered in the mechanism of direct reciprocity) [1], or the sanctioning
of non-conforming behavior (“punishment”) [7–12, 26, 33, 34] These mechanisms
can, in fact, transform prisoner’s dilemmas into stag hunt games [15, 23, 29] (see
Fig. 10.1), which connects our approach with previous work addressing norms.
However, our model goes beyond studying the circumstances under which people
follow a preset norm, it considers situations where it is not clear from the beginning
what behavior would eventually establish as a norm (or whether any of the
behaviors would become a norm at all). When studying multi-population settings
with incompatible interests, we do not only have the problem of how cooperative
behavior can be promoted, as in the prisoner’s dilemma. We also have a normative
dilemma by the circumstance that the establishment of a behavioral norm requires
one population to adjust to the preferred behavior in the other population – against
its own preference.
Other cooperation-enhancing mechanisms such as kin selection (based on
genetic relationship) and group selection tend to transform a prisoner’s dilemma
into a harmony game [15] (see Fig. 10.1). Therefore, our findings suggest that
genetic relatedness and group selection are not ideal mechanisms to establish shared
behavioral norms. They rather support the formation of subcultures. Moreover, the
transformation of prisoner’s dilemma interactions into a snowdrift game is expected
to cause social conflict. Obviously, this has crucial implications for society, law and
economics [20, 35], where conflicts need to be avoided or solved, and norms and
standards are of central importance.
Take language as another example – probably the most distinguishing trait of
humans [36, 37]. Successful communication requires the establishment of a norm,
10.4 Discussion 195
how words are used (the “evolution of meaning”). It will, therefore, be intriguing to
study whether the explosive development of language and culture in humans is due
to their ability to transform interactions into norm-promoting stag hunt interactions.
From this point of view, repeated interactions due to human agglomeration in
settlements, the development of reputation mechanisms, and the invention of
sanctioning institutions should have largely accelerated cultural evolution [38].
Appendix A: Methods
Multi-population game-dynamical replicator equations describe the temporal evo-

lution of the proportions pia .t/ of individuals showing behavior i at time t in
population a. They assume that more successful behaviors spread, as these are
imitated by individuals of the same population at a rate proportional to the gain
in the expected success. The expected success is determined from the frequency
of interactions between two behaviors i and j , and by the associated payoffs Aab ij
(see [19]). Focusing on the above-mentioned social dilemmas, in the case of two
interacting populations a; b 2 f1; 2g and two behavioral strategies i; j 2 f1; 2g,
we assume the following for interactions within the same population a: If two
interacting individuals show the same behavior i , both will either receive the
payoff ra or pa . If we have ra ¤ pa , we call the behavior with the larger
payoff ra “preferred” or “cooperative”, the other behavior “non-cooperative” or
“uncooperative”. When one individual chooses the cooperative behavior and the
interaction partner is uncooperative, the first one receives the payoff sa and the
second one the payoff ta . To model conflicts of interests, we assume that population
a D 1 prefers behavior i D 1 and population 2 prefers behavior 2. Therefore, if an
individual of population 1 meets an individual belonging to population 2 and both
show the same behavior i D 1, the first one will earn R1 and the second one P2 ,
as behavior i D 1 is considered uncooperative in population 2. Analogously, for
i D 2 they earn P1 and R2 , respectively. If the interaction partners choose different
behaviors i and j , they earn Sa , when the behavior corresponds to their cooperative
behavior, otherwise they earn Ta .
Assuming constant preferences and fixed relative population strengths fa , the
resulting coupled game-dynamical replicator equations for the temporal evolution
of the proportion p.t/ D p11 .t/ of cooperative individuals in population 1 and the
fraction q.t/ D p22 .t/ of cooperative individuals in population 2 are given by (10.1)
and (10.2) with
F .p; q/ D b1 f C .c1 b1 /fp.t/ C C1 .1 f / C .B1 C1 /.1 f /q.t/ (10.3)
and
G.p; q/ D b2 .1 f / C .c2 b2 /.1 f /q.t/ C C2 f C .B2 C2 /fp.t/ (10.4)

(see [19]). Here, we have used the abbreviation f D f1 D 1 f2 . Moreover,

ba D sa pa , Ba D Sa Pa , ca D ra ta , and Ca D Ra Ta are payoff-dependent
model parameters, which can be positive, negative, or zero. Note that the above
equations describe the general case for two populations with interactions and/or
self-interactions playing any kind of 2 2 game. When setting pa D Pa D P ,
ra D Ra D R, sa D Sa D S , and ta D Ta D T (i.e. ba D Ba D B and
ca D Ca D C ), both populations play the same game. Moreover, the payoff depends
on the own behavior i and the behavior j of the interaction partner only, but not on
the population he/she belongs to. That is, in- and out-group interactions yield the
same payoff, and the preference of the interaction partner does not matter for it.
Sanctioning of non-conforming behavior. Individuals often apply group pressure
to support conformity and discourage uncoordinated behavior. This could be
modeled by subtracting a value ı from the off-diagonal payoffs S and T or by
adding ı to the diagonal elements R and P , resulting in the effective model
parameters ba D Ba D B ı and ca D Ca D C C ı. Therefore, if the group
pressure ı is large enough (namely, ı > jC j), a prisoner’s dilemma with B < 0 and
C < 0 is transformed into a stag hunt game with ba D Ba < 0 and ca D Ca > 0.
Summary of the main analytical results. (see [19] for more general formulas). All
of the multi-population games with interactions and self-interactions studied by us
have the four stationary solutions .p1 ; q1 / D .0; 0/, .p2 ; q2 / D .1; 1/, .p3 ; q3 / D
.1; 0/ and .p4 ; q4 / D .0; 1/, corresponding to the four corners of the p-q-space.
Their stability properties depend on the eigenvalues l D .1 2pl /F .pl ; ql / and
l D .1 2ql /G.pl ; ql /, where l 2 f1; 2; 3; 4g. Stable (attractive) fix points require
l < 0 and l < 0, unstable (repulsive) fix points l < 0 and l < 0. l l < 0
implies a saddle point. In the multi-population prisoner’s dilemma (MPD), the only
stable fix point is .0; 0/, while in the multi-population harmony game (MHG), it is
.1; 1/. In both games, .1; 0/ and .0; 1/ are always saddle points. For B; C < 0 (the
MPD) and B; C > 0 (the MHG), no further stationary points exist (see Fig. 10.2).
For the multi-population stag hunt game (MSH) with B < 0 and C > 0, we find:
– (0,1) and (1,0) are always stable fix points, see Fig. 10.3.
– (1,1) is a stable fix point for C =jBj > maxŒf =.1 f /; .1 f /=f (see
Fig. 10.3b).
– (0,0) is a stable fix point for C =jBj < minŒf =.1 f /; .1 f /=f .
For the multi-population snowdrift game (MSD) with B > 0 and C < 0 we have:
– (1,0) and (0,1) are always unstable fix points (see Fig. 10.4).
– (0,0) is a stable fix point for C =jBj > maxŒf =.1 f /; .1 f /=f (see
Fig. 10.4b).
– (1,1) is a stable fix point for C =jBj < minŒf =.1 f /; .1 f /=f .
Moreover, if B and C have different signs, further stationary points .pl ; ql / with
l 2 f5; 6; 7; 8g may occur on the boundaries, while inner points .p9 ; q9 / with
0 < p9 < 1 and 0 < q9 < 1 can only occur for B D C (see Figs. 10.3a
and 10.4a). As jC j is increased from 0 to high values, we find the following
additional stationary points for the MSH and MSD, where we use the abbreviations
10.4 Discussion 197
p5 D ŒBf C C.1 f /=Œ.B C /f , p6 D B=Œ.B C /f , q7 D B=Œ.B C /

.1 f /, and q8 D ŒB.1 f / C Cf =Œ.B C /.1 f /:
– .p5 ; 0/ and .0; q8 /, if f 1=2 and jC j=jBj .1 f /=f or if f 1=2 and
jC j=jBj f =.1 f /.
– .p5 ; 0/ and .p6 ; 1/, if f 1=2 and .1 f /=f < jC j=jBj < f =.1 f /, or
.1; q7 / and .0; q8 / if f 1=2 and f =.1 f / < jC j=jBj < .1 f /=f (see
Figs. 10.3cCd and 10.4cCd).
– .p6 ; 1/ and .1; q7 /, if f 1=2 and jC j=jBj f =.1 f / or if f 1=2 and
jC j=jBj .1 f /=f (see Figs. 10.3b and 10.4b).
For B < 0 < C (the MSH), these fix points are unstable or saddle points,
while they are stable or saddle points for C < 0 < B (the MSD). Obviously,
there are transitions to a qualitatively different system behavior at the points
jC j=jBj D .1f /=f and jC j=jBj D f =.1f /. Moreover, there are discontinuous,
“revolutionary” transitions, when jC j crosses the value of jBj, as the stability
properties of pairs of fix points are then interchanged. This can be followed from
the fact that the dynamic behavior and final outcome for the case jBj > jC j can
be derived from the results for jBj < jC j, namely by applying the transformations
B $ C , p $ .1 p/, and q $ .1 q/, which do not change the game-
dynamical equations (see (10.1)–(10.4)). Generally, discontinuous transitions in the
system behavior may occur, when the sign of 1 jBj=jC j changes, or if the sign
of B or C changes (which modifies the character of the game, for example from a
MPD to a MSH game).
Fraction of cooperators in the multi-population snowdrift game. When .p5 ; 0/ is
the stable stationary point, the average fraction of cooperative individuals in both
populations from the perspective of the stronger population 1 can be determined
as the fraction of cooperative individuals in population 1 times the relative size
f of population 1, plus the fraction 1 q5 D 1 of non-cooperative individuals in
population 2 (who are cooperative from the point of view of population 1), weighted
by its relative size .1 f /:
Bf C C.1 f / B
p5 f C .1 q5 / .1 f / D f C 1 .1 f / D :
.B C /f B C
Considering C < 0, this corresponds to the expected fraction p0 D jBj=.jBj C C /

of cooperative individuals in the one-population snowdrift game [15].
Acknowledgements The authors would like to thank for partial support by the EU Project
QLectives and the ETH Competence Center “Coping with Crises in Complex Socio-Economic
Systems” (CCSS) through ETH Research Grant CH1-01 08-2. They are grateful to Thomas
Chadefaux, Ryan Murphy, Carlos P. Roca, Stefan Bechtold, Sergi Lozano, Heiko Rauhut, Wenjian
Yu and further colleagues for valuable comments and to Sergi Lozano for drawing Fig. 10.1. D.H.
thanks Thomas Voss for his insightful seminar on social norms.
References

2. H. Gintis, Game Theory Evolving (Princeton University, Princeton, NJ, 2000)
4. M. Milinski, D. Semmann, H.J. Krambeck, Reputation helps solve the “tragedy of the
commons”. Nature 415, 424–426 (2002)
7. C. Hauert, A. Traulsen, H. Brandt, M.A. Nowak, K. Sigmund, Via freedom to coercion: The
emergence of costly punishment. Science 316, 1905–1907 (2007)
8. B. Rockenbach, M. Milinski, The efficient interaction of indirect reciprocity and costly
punishment. Nature 444, 718–723 (2006)
9. O. Gurerk, B. Irlenbusch, B. Rockenbach, The competitive advantage of sanctioning institu-
tions. Science 312, 108–111 (2006)
10. J. Henrich et al., Costly punishment across human societies. Science 312, 1767–1770 (2006)
11. J.H. Fowler, Altruistic punishment and the origin of cooperation. Proc. Natl. Acad. Sci. USA
102, 7047–7049 (2005)
Natl. Acad. Sci. USA 100, 3531–3535 (2003)
noisy conditions. Proc. Natl. Acad. Sci. USA 106(8), 3680–3685 (2009)
14. J.W. Weibull, Evolutionary Game Theory (MIT Press, Cambridge, MA, 1996)
15. D. Helbing, S. Lozano, Phase transitions to cooperation in the prisoner’s dilemma. Phys. Rev.
E 81(5), 057102 (2010)
16. H. Ohtsukia, M.A. Nowak, The replicator equation on graphs. J. Theor. Biol. 243, 86–97
(2006)
17. P. Schuster, K. Sigmund, J. Hofbauer, R. Gottlieb, P. Merz, Selfregulation of behaviour in
animal societies. III. Games between two populations with selfinteraction. Biol. Cyber. 40,
17–25 (1981)
18. D. Helbing, A mathematical model for behavioral changes by pair interactions, in Economic
Evolution and Demographic Change ed. by G. Haag, U. Mueller, K.G. Troitzsch (Springer,
Berlin, 1992), pp. 330–348
19. D. Helbing, A. Johansson, Evolutionary dynamics of populations with conflicting interac-
tions: Classification and analytical treatment considering asymmetry and power. Phys. Rev. E
81, 016112 (2010)
20. L. Samuelson, Chap. 5: The Ultimatum Game, in Evolutionary Games and Equilibrium
Selection (The MIT Press, Cambridge, 1998)
21. R. Axelrod, An evolutionary approach to norms. Am. Pol. Sci. Rev. 80(4), 1095–1111 (1986)
22. M. Hechter, K.D. Opp (eds.), particularly Chap. 4 Game-theoretical perspectives on the
emergence of social norms, in Social Norms, ed. by T. Voss (Russell Sage, New York, 2001),
pp. 105–136
23. C. Bicchieri, R. Jeffrey, B. Skyrms (eds.), The Dynamics of Norms (Cambridge University,
Cambridge, 2009)
24. J. Bendor, P. Swistak, The evolution of norms. Am. J. Sociol. 106(6), 1493–1545 (2001)
25. F.A.C.C. Chalub, F.C. Santos, J.M. Pacheco, The evolution of norms. J. Theor. Biol. 241,
233–240 (2006)
26. E. Ostrom, Collective action and the evolution of social norms. J. Econ. Perspect. 14(3), 137–
158 (2000)
27. P.R. Ehrlich, S.A. Levin, The evolution of norms. PLoS Biol. 3(6), 0943–0948 (2005)
28. K. Keizer, S. Lindenberg, L. Steg, The spreading of disorder. Science 322, 1681–1685 (2008)
29. B. Skyrms, The Stag Hunt and the Evolution of Social Structure (Cambridge University,
Cambridge, 2003)
References 199
30. K.D. Opp, Theories of Political Protest and Social Movements (Routledge, London, 2009)
31. T.S. Kuhn, The Structure of Scientific Revolutions (University of Chicago, Chicago, 1962)
32. W. Weidlich, H. Huebner, Dynamics of political opinion formation including catastrophe
theory. J. Econ. Behav. Organ. 67, 1–26 (2008)
33. E. Fehr, U. Fischbacher, S. Gächter, Strong reciprocity, human cooperation, and the enforce-
ment of social norms. Hum. Nat. 13, 1–25 (2002)
34. A. Whiten, V. Horner, F.B.M. de Waal, Conformity to cultural norms of tool use in
chimpanzees. Nature 437, 737–740 (2005)
35. K. Binmore, Natural Justice (Oxford University, New York, 2005)
36. M.A. Nowak, N.L. Komarova, P. Niyogi, Computational and evolutionary aspects of lan-
guage. Nature 417, 611–617 (2002)
37. V. Loreto, L. Steels, Emergence of language. Nat. Phys. 3, 758–760 (2007)
38. R. Boyd, P.J. Richerson, The Origin and Evolution of Cultures (Oxford University, Oxford,
2005)
Chapter 11
Social Experiments and Computing
11.1 Introduction
When Nowak and May published their computational study of spatial games in
1992, it soon became a scientific milestone [1]. They showed that altruistic (“coop-
erative”) behavior would be able to survive through spatial clustering. This finding,
also called “network reciprocity” [2], is enormously important, as cooperation is
the essence that keeps societies together. It is the basis of solidarity and social order.
When humans stop cooperating, this implies a war of everybody against everybody.
Understanding why and under what conditions humans cooperate is one of the
grand challenges of science [3], particularly in social dilemma situations (where col-
lective cooperation is beneficial, but individual free-riding is even more profitable).
How should humans otherwise be able to create public goods (such as a shared
culture or a public infrastructure), build up functioning social benefit systems, or
fight global warming collectively in the future? From a theoretical point of view,
Nowak and May’s work demonstrates that the representative agent paradigm of
economics (according to which interactions with others can be represented by
the interaction with average individuals) can be quite misleading. This paradigm
predicts that cooperation should completely disappear in social dilemma situations,
leading to a “tragedy of the commons”. If the world was really like this, social
systems would not work.

This chapter was written together with Wenjian Yu. It is an extended version of the following
Commentary, which the reader is requested to cite instead: D. Helbing and W. Yu, The future of
social experimenting. Proceedings of the National Academy of Sciences USA 107(12), 5265–5266
(2010).

202 11 Social Experiments and Computing
However, when the same interactions take place in a spatial setting, they can
cause correlations between the behaviors of neighboring individuals, which can
dramatically change the outcome of the system (as long as the interactions are local
rather than global). The effect is even more pronounced, when a success-driven
kind of mobility is considered in the model [4]. Spatio-temporal pattern formation
facilitates a co-evolution of the behaviors and the spatial organization of individuals,
creating a “social milieu” that can encourage cooperative behavior. In fact, some
long-standing puzzles in the social sciences find a natural solution, when spatial
interactions (and mobility) are taken into account. This includes the higher-than-
expected level of cooperation in social dilemma situations and the spreading of
costly punishment (the eventual disappearance of defectors and “second-order
free-riders”, i.e. cooperators who abstain from the punishment of non-cooperative
behaviors).
11.2 Experiment
Despite the importance of these topics, it took quite long until the effects of
game-theoretic interactions in two-dimensional space were tested in laboratory
experiments. A recent study of Traulsen et al. [5] now reports experiments of a
spatial prisoner’s dilemma game for the original setting of Nowak and May, while
the size of the spatial grid, the number of interaction partners and the payoff
parameters were modified for experimental reasons. According to their results,
spatial interactions have no significant effect on the level of cooperation. The reason
for this is that their experimental subjects did not show an unconditional imitation
of neighbors with a higher payoff, as it is assumed in many game-theoretic models.
In fact, it is known that certain game-theoretic results are sensitive to details
of the model such as the number of interaction partners, the inclusion of self-
interactions or not, or significant levels of randomness (see Figs. 11.1–11.4).
Moreover, researchers have proposed a considerable number of different strategy
update rules, which matter as well. Besides unconditional imitation, these include
the best response rule [6], multi-stage strategies such as tit for tat [7], win-stay-
lose-shift rules [8] and aspiration-dependent rules [9], furthermore probabilistic
rules such as the proportional imitation rule [10, 11], the Fermi rule [12], and
the unconditional imitation rule with a superimposed randomness (“noise”) [4].
In addition, there are voter [13] and opinion dynamics models [14] of various
kinds, which assume social influence. According to these, individuals would imitate
behavioral strategies, which are more frequent in their neighborhood. So, how do
individuals really update their behavioral strategies?
Traulsen et al. find that the probability to cooperate increases with the number
of cooperative neighbors as expected from the Asch experiment [15]. Moreover, the
probability of strategy changes increases with the payoff difference in a way that
can be approximated by the Fermi rule [12]. In the case of two behavioral strategies
11.2 Experiment 203
Fig. 11.1 Snapshot of a computer simulation of the spatial prisoner’s dilemma without self-
interactions, illustrating the representative dynamics of strategy updating on a 49 49 lattice.
Here, we assume an unconditional imitation of the best performing direct neighbor (given his/her
payoff was higher). Blue sites correspond to cooperative individuals, red sites to defecting ones.
The payoffs in the underlying prisoner’s dilemma were assumed as in the paper by Traulsen et al.
[5]. A video illustrating the dynamics of the game is available at http://www.soms.ethz.ch/research/
socialexperimenting. It reveals that the level of cooperation decays quickly, and defectors prevail
after a short time. Since the simulation assumes no randomness in the strategy updates, the spatial
configuration “freezes” quickly, i.e. it does not change anymore after a few iterations
only, it corresponds to the well-known multi-nomial logit model of decision theory

[16]. However, there is a discontinuity in the data as the payoff difference turns from
positive to negative values, which may be an effect of risk aversion [17]. To describe
the time-dependent level of cooperation, it is sufficient to assume unconditional
imitation with a certain probability, otherwise strategy mutations. The mutation rate
is surprisingly big in the beginning and exponentially decaying over time.
The most surprising fact is maybe not the high level of randomness, which is
quite typical for social systems. While one may expect that a large noise level
quickly reduces a high level of cooperation, it actually leads to more cooperation
than the unconditional imitation rule predicts (see Fig. 2 of [5]). This goes along
with a significantly higher average payoff than for the unconditional imitation rule.
In other words, the random component of the strategy update is profitable for
the experimental subjects. This suggests that noise in social systems may play a
functional role.
Fig. 11.2 Snapshot of a computer simulation of the spatial prisoner’s dilemma without self-
interactions, illustrating the representative dynamics of strategy updating according to (3) of
Traulsen et al. [5]. The lattice size, payoff parameters, and color coding are the same as before, but
individuals are performing random strategy updates with an exponentially decaying probability,
while unconditional imitation occurs only otherwise. Due to the presence of strategy mutations,
the spatial configuration keeps changing. Compared to Fig. 11.1, the level of cooperation drops
further, since metastable configurations are broken up by strategy mutations (“noise”). A related
video is available at http://www.soms.ethz.ch/research/socialexperimenting
11.3 Discussion
Given that Traulsen et al. do not find effects of spatial interactions, do we have to say
good bye to network reciprocity in social systems and to all the nice explanations
that it offers? Comparing Fig. 11.6 with Fig. 11.5 suggests that this is not the
case, since the experimental setting did not promote cooperative clusters for the
occurring noise levels (see Fig. 11.2), while a scenario with self-interactions would
have done so (see Fig. 11.4). Also the empirically confirmed spreading of obesity,
smoking, happiness, and cooperation in social networks [18–22] suggests that
effects of imitating neighbors (also friends or colleagues) are relevant, but probably
over longer time periods than 25 interactions. In fact, in contrast to the scenario
without self-interactions (see Fig. 11.5), according to formula (3) of Traulsen et al.
one would expect a sudden spreading of cooperation when the mutation rate has
decreased to low values (after about 40 iterations), given that self-interactions are
taken into account (see Fig. 11.6). To make the effect observable experimentally, it
11.3 Discussion 205
Fig. 11.3 Snapshot of a computer simulation of the spatial prisoner’s dilemma assuming uncon-
ditional imitation. Compared to Fig. 11.1, we take self-interactions into account, which supports
the spreading of cooperators. Since individuals are assumed to imitate unconditionally, there are
no strategy mutations. As a consequence, the spatial configuration freezes after a few iterations. A
related video is available at http://www.soms.ethz.ch/research/socialexperimenting
would be favorable to reduce the necessary number of iterations for its occurrence
and to control the noise level.
The particular value of the work by Traulsen et al. [5] is that it facilitates more
realistic computer simulations. Thereby it becomes possible to determine payoff
values and other model parameters, which are expected to produce interesting
effects (such as spatial correlations) after an experimentally accessible number of
iterations. In fact, experimental games can have qualitatively different outcomes,
which are hard to predict without extensive computer simulations scanning the
parameter space (see [23] for an example and a variety of related “phase diagrams”).
Such parameter dependencies could explain some of the apparent inconsistencies
between empirical observations in different areas of the world [24] (at least when
framing effects such as the expected level of reciprocity and their impact on the
effective payoffs [2] are taken into account). The progress in the social sciences by
understanding such parameter dependencies would be enormous. However, as the
effort to experimentally determine phase diagrams is prohibitive, one can only check
computationally predicted, parameter-dependent outcomes by targeted samples.
The future of social experimenting lies in the combination of computational and
experimental approaches, where computer simulations optimize the experimental
Fig. 11.4 Snapshot of a computer simulation of the spatial prisoner’s dilemma with self-
interactions, assuming strategy updates according to (3) of Traulsen et al. [5]. Initially, there is
a high probability of strategy mutations, but it decreases exponentially. As a consequence, an
interesting effect occurs: While the level of cooperation decays in the beginning, it manages to
recover later and becomes almost as high as in the noiseless case represented by Fig. 11.3. A related
video is available at http://www.soms.ethz.ch/research/socialexperimenting
setting and experiments are used to verify, falsify or improve the underlying model
assumptions.
Besides selecting parameter values which maximize the signal-to-noise ratio and
minimize the number of iterations after which the expected effect becomes visible,
one could try to reduce the level of randomness by experimental noise control.
For this, it would be useful to understand the origin and relevance of the observed
randomness. Do the experimental subjects make mistakes and why? Do they try to
optimize their behavioral strategies or do they apply simple heuristics (and which
ones)? Do they use heterogeneous updating rules? Or do they just show exploratory
behavior? [25,26] Is it useful to work with subjects who have some experience with
behavioral experiments (without having a theoretical background in them)? How
relevant is the homogeneity of the subject pool? What are potentials and dangers of
framing effects? How can effects of the individual histories of experimental subjects
be eliminated? Does it make sense to perform the experiment with a mixture of
experimental subjects and computer agents (where the noise level can be reduced
by implementing deterministic strategy updates of these agents)?
In view of the great theoretical importance of experiments with many iterations
and spatial interactions, more large-scale experiments over long time horizons
11.3 Discussion 207
Fig. 11.5 Average payoff of all individuals in the spatial prisoner’s dilemma without self-
interactions, displayed over the number of iterations. It is clearly visible that the initial payoff drops
quickly. In the noiseless case, it becomes constant after a few iterations, as the spatial configuration
freezes (see broken line). In contrast, in the case of a decaying rate of strategy mutations according
to (3) of Traulsen et al. [5], the average payoff keeps changing (see solid line). It is interesting
that the average payoff is higher in the noisy case than in the noiseless one for approximately
30 iterations, particularly over the time period of the laboratory experiment by Traulsen et al.
(covering 25 iterations). The better performance in the presence of strategy mutations could be a
possible reason for the high level of strategy mutations observed by them
would be desirable. This calls for larger budgets (as they are common in the
natural and engineering sciences), but also for new concepts. Besides connecting
labs in different countries via internet, one may consider to perform experiments
in “living labs” on the web itself [27]. It also seems worth exploring, how much
we can learn from interactive games such as Second Life or Lords of Warcraft
[28–30], which could be adapted for experimental purposes in order to create
well-controlled environments. According to a recent replication of the Milgram
experiment [31] with Avatars [32, 33], experiments with virtual humans may
actually be surprisingly well transferable to real humans. One can furthermore
hope that lab or web experiments will eventually become standardized measurement
instruments to determine indices like the local “level of cooperation” as a function
of time, almost like the gross domestic product is measured today. Knowing the
“index of cooperativeness” would be good, as it reflects social capital. The same
applies to the measurement of social norms, which are equally important for social
order as cooperation, since they determine important factors such as coordination,
adaptation, assimilation, integration, or conflict.
Fig. 11.6 Average payoff of all individuals in the spatial prisoner’s dilemma with self-interactions,
as a function of the number of iterations. Like in Fig. 11.5, the payoff drops considerably in the
beginning. In the noiseless case, it stabilizes quickly (broken line). In comparison, when strategy
mutations decay according to formula (3) of Traulsen et al. [5], the average payoff keeps decreasing
for some time (see solid line) and falls significantly below the payoff of the noiseless case.
However, after about 40 iterations, the average payoff recovers, which correlates with an increase
in the level of cooperation. Due to the pronounced contrast to the case without self-interactions
(see Fig. 11.5), it would be interesting to perform experiments with self-interactions. These should
extend over significantly more than 25 iterations, or the payoff parameters would have to be
changed in such a way that the average payoff recovers earlier. It is conceivable, however, that
experimental subjects would show a lower level of strategy mutations under conditions where
noise does not pay off (in contrast to the studied experimental setting without self-interactions
represented in Fig. 11.5)
References
2. M.A. Nowak, Five rules for the evolution of cooperation. Science 314, 1560–1563, and
supplementary information (2006)
3. E. Pennisi, How did cooperative behavior evolve? Science 309, 93 (2005)
4. D. Helbing, W. Yu, The outbreak of cooperation among success-driven individuals under noisy
conditions. Proc. Natl. Acad. Sci. USA 106(8), 3680–3685 (2009)
5. A. Traulsen et al., Human strategy updating in evolutionary games. Proc. Natl. Acad. Sci. USA
107, 2962–2966 (2010)
6. A. Matsui, Best response dynamics and socially stable strategies. J. Econ. Theor. 57, 343–362
(1992)
8. M. Nowak, K. Sigmund, A strategy of win-stay, lose-shift that outperforms tit-for-tat in the
Prisoner’s Dilemma game. Nature 364, 56–58 (1993)
9. M.W. Macy, A. Flache, Learning dynamics in social dilemmas. Proc. Natl. Acad. Sci. USA
99(Suppl. 3), 7229–7236 (2002)
References 209
10. D. Helbing, in Economic Evolution and Demographic Change, ed. by G. Haag, U. Mueller,
K.G. Troitzsch (Springer, Berlin, 1992), pp. 330–348
11. K.H. Schlag, Why imitate, and if so, how? A boundedly rational approach to multi- armed
bandits. J. Econ. Theor. 78(1), 130–156 (1998)
12. G. Szabo, C. Toke, Evolutionary prisoner’s dilemma game on a square lattice. Phys. Rev. E 58,
69–73 (1998)
13. I. Dornic, H. Chate, J. Chave, H. Hinrichsen, Critical coarsening without surface tension: The
universality class of the voter model. Phys. Rev. Lett. 87, 045701 (2001)
14. K. Sznajd-Weron, J. Sznajd, Opinion evolution in closed community. Int. J. Mod. Phys. C
11(6), 1157–1165 (2000)
15. S.E. Asch, Studies of independence and conformity: A minority of one against a unanimous
majority. Psychol. Monogr. 70(9), 1–70 (1956)
16. D. McFadden, in Frontiers of Econometrics, ed. by P. Zarembka (Academic Press, New York,
1974), pp. 105–142
17. D. Kahnemann, A. Tversky, Prospect theory: An analysis of decision under risk. Econometrica
47(2), 263–291 (1979)
18. K.P. Smith, N.A. Christakis, Social networks and health. Ann. Rev. Sociol. 34, 405–429 (2008)
19. N.A. Christakis, J.H. Fowler, The spread of obesity in a large social network over 32 years.
New Engl. J. Med. 357(4), 370–379 (2007)
20. N.A. Christakis, J.H. Fowler, The collective dynamics of smoking in a large social network.
New Engl. J. Med. 358(21), 2249–2258 (2008)
21. J.H. Fowler, N.A. Christakis, Dynamic spread of happiness in a large social network. Br. Med.
J. 337, a2338 (2008)
22. J.H. Fowler, N.A. Christakis, Cooperative behavior cascades in human social networks. Proc.
Natl. Acad. Sci. USA 107(12), 5334–5338 (2010)
23. D. Helbing, A. Szolnoki, M. Perc, G. Szabo, Evolutionary establishment of moral and
double moral standards through spatial interactions. Evolutionary establishment of moral and
double moral standards through spatial interactions. PLoS Comput. Biol. 6(4), e1000758
(2010). Supplementary videos are available at http://www.soms.ethz.ch/research/secondorder-
freeriders and http://www.matjazperc.com/games/moral.html
24. B. Herrmann, C. Thoni, S. Gachter, Antisocial punishment across societies. Science 319, 1362–
1367 (2008)
25. D. Helbing, M. Schonhof, H.U. Stark, J.A. Holyst, How individuals learn to take turns. Adv.
Complex Syst. 8, 87–116 (2005)
26. A. Traulsen, C. Hauert, H.D. Silva, M.A. Nowak, K. Sigmund, Exploration dynamics in
evolutionary games. Proc. Natl. Acad. Sci. USA 106, 709–712 (2009)
27. M.J. Salganik, P.S. Dodds, D.J. Watts, Experimental study of inequality and unpredictability
in an artificial cultural market. Science 311, 854–856 (2006)
28. W.S. Bainbridge, The scientific research potential of virtual worlds. Science 317, 472–476
(2007)
29. N.F. Johnson, C. Xu, Z. Zhao, N. Ducheneaut, N. Yee, G. Tita, P.M. Hui, Human group
formation in online guilds and offline gangs driven by a common team dynamic. Phys. Rev. E
79, 066117 (2009)
30. M. Szell, S. Thurner, Measuring social dynamics in a massive multiplayer online game. Social
Networks 32(4), 313–329 (2010)
31. S. Milgram, Behavioral study of obedience. J. Abnorm. Soc. Psychol. 67(4), 371–378 (1963)
32. M. Slater, et al., A virtual reprise of the Stanley Milgram obedience experiments. PLoS One 1,
e39 (2006)
33. M. Cheetham, A.F. Pedroni, A. Antley, M. Slater, L. Jancke, Virtual Milgram: Empathic
concern or personal distress? Front. Hum. Neurosci. 3, 29, 1–13 (2009)
Chapter 12
Learning of Coordinated Behavior
12.1 Introduction
Congestion is a burden of today’s traffic systems, affecting the economic prosperity

of modern societies. Yet, the optimal distribution of vehicles over alternative routes
is still a challenging problem and uses scarce resources (street capacity) in an
inefficient way. Route choice is based on interactive, but decentralized individual
decisions, which cannot be well described by classical utility-based decision models
[27]. Similar to the minority game [16, 39, 43], it is reasonable for different people
to react to the same situation or information in different ways. As a consequence,
individuals tend to develop characteristic response patterns or roles [26]. Thanks to
this differentiation process, individuals learn to coordinate better in the course of
time. However, according to current knowledge, selfish routing does not establish
the system optimum of minimum overall travel times. It rather tends to establish
the Wardrop equilibrium, a special user or Nash equilibrium characterized by
equal travel times on all alternative routes chosen from a certain origin to a given
destination (while routes with longer travel times are not taken) [71].
Since Pigou [53], it has been suggested to resolve the problem of inefficient road
usage by congestion charges, but are they needed? Is the missing establishment
of a system optimum just a problem of varying traffic conditions and changing
origin-destination pairs, which make route-choice decisions comparable to one-shot
games? Or would individuals in an iterated setting of a day-to-day route choice
game with identical conditions spontaneously establish cooperation in order to
increase their returns, as the folk theorem suggests [6]?
How would such a cooperation look like? Taking turns could be a suitable
solution [62]. While simple symmetrical cooperation is typically found for the

This chapter reprints a previous publication with kind permission of the copyright owner, World
Scientific. It is requested to cite this work as follows: D. Helbing, M. Schönhof, H.-U. Stark,
and J. A. Holyst, How individuals learn to take turns: Emergence of alternating cooperation in
a congestion game and the prisoner’s dilemma. Advances in Complex Systems 8, 87–116 (2005).

212 12 Learning of Coordinated Behavior
repeated Prisoner’s Dilemma [2, 3, 44–46, 49, 52, 55, 59, 64, 67, 69], emergent
alternating reciprocity has been recently discovered for the games Leader and Battle
of the Sexes [11].1 Note that such coherent oscillations are a time-dependent, but
deterministic form of individual decision behavior, which can establish a persistent
phase-coordination, while mixed strategies, i.e. statistically varying decisions, can
establish cooperation only by chance or in the statistical average. This difference
is particularly important when the number of interacting persons is small, as in the
particular route choice game discussed below.
Note that oscillatory behavior has been found in iterated games before:
• In the rock-paper-scissors game [67], cycles are predicted by the game-dynamical
equations due to unstable stationary solutions [28].
• Oscillations can also result by coordination problems [1,29,31,33], at the cost of
reduced system performance.
• Moreover, blinker strategies may survive in repeated games played by a mixture
of finite automata [5] or result through evolutionary strategies [11, 15, 16, 38, 39,
42, 43, 74].
However, these oscillation-generating mechanisms are clearly to be distinguished
from the establishment of phase-coordinated alternating reciprocity we are inter-
ested in (coherent oscillatory cooperation to reach the system optimum).
Our paper is organized as follows: In Sect. 12.2, we will formally introduce the
route choice game for N players, including issues like the Wardrop equilibrium
[71] and the Braess paradox [10]. Section 12.3 will focus on the special case of the
2-person route choice game, compare it with the minority game [1, 15, 16, 38, 39,
42, 43, 74], and discuss its place in the classification scheme of symmetrical 2 2
games. This section will also reveal some apparent shortcomings of the previous
game-theoretical literature:
• While it is commonly stated that among the 12 ordinally distinct, symmetrical
2 2 games [11, 57] only four archetypical 2 2 games describe a strategical
conflict (the Prisoner’s Dilemma, the Battle of the Sexes, Chicken, and Leader)
[11,18,56], we will show that, for specific payoffs, the route choice game (besides
Deadlock) also represents an interesting strategical conflict, at least for iterated
games.
• The conclusion that conservative driver behavior is best, i.e. it does not pay off
to change routes [7,65,66], is restricted to the special case of route-choice games
with a system-optimal user equilibrium.
• It is only half the truth that cooperation in the iterated Prisoner’s Dilemma
is characterized by symmetrical behavior [11]. Phase-coordinated asymmetric
reciprocity is possible as well, as in some other symmetrical 2 2 games [11].
New perspectives arise by less restricted specifications of the payoff values.
1
See Fig. 12.2 for a specification of these games.
12.2 The Route Choice Game 213
In Sect. 12.4, we will discuss empirical results of laboratory experiments with

humans [12, 18, 32]. According to these, reaching a phase-coordinated alternating
state is only one problem. Exploratory behavior and suitable punishment strategies
are important to establish asymmetric oscillatory reciprocity as well [11, 20].
Moreover, we will discuss several coefficients characterizing individual behavior
and chances for the establishment of cooperation. In Sect. 12.5, we will present
multi-agent computer simulations of our observations, based on a novel win-
stay, lose-shift [50, 54] strategy, which is a special kind of reinforcement learning
strategy [40]. This approach is based on individual historical experience [13]
and, thereby, clearly differs from the selection of the best-performing strategy in
a set of hypothetical strategies as assumed in studies based on evolutionary or
genetical algorithms [5, 11, 15, 16, 39, 42, 43]. The final section will summarize
our results and discuss their relevance for game theory and possible applications
such as data routing algorithms [35, 72], advanced driver information systems
[8, 14, 30, 37, 41, 63, 70, 73], or road pricing [53].
12.2 The Route Choice Game
In the following, we will investigate a scenario with two alternative routes between a
certain origin and a given destination, say, between two places or towns A and B (see
Fig. 12.1). We are interested in the case where both routes have different capacities,
say a freeway and a subordinate or side road. While the freeway is faster when it is
empty, it may be reasonable to use the side road when the freeway is congested.
The “success” of taking route i could be measured in terms of its inverse travel
time 1=Ti .Ni / D Vi .Ni /=Li , where Li is the length of route i and Vi .Ni / the
average velocity when Ni of the N drivers have selected route i . One may roughly
approximate the average vehicle speed Vi on route i by the linear relationship [24]
Origin
Route 1
Route 2
Destination
Fig. 12.1 Illustration of the investigated day-to-day route choice scenario. We study the dynamic
decision behavior in a repeated route choice game, where a given destination can be reached from
a given origin via two different routes, a freeway (route 1) and a side road (route 2)

Ni .t/
Vi .Ni / D Vi 1 max ;
0
(12.1)
Ni
where Vi0 denotes the maximum velocity (speed limit) and Nimax the capacity, i.e. the
maximum possible number of vehicles on route i . With Ai D Vi0 =Li and Bi D
Vi0 =.Nimax Li /, the inverse travel time then obeys the relationship
1=T .Ni / D Ai Bi Ni ; (12.2)
which is linearly decreasing with the road occupancy Ni . Other monotonously

falling relationships Vi .Ni / would make the expression for the inverse travel times
non-linear, but they would probably not lead to qualitatively different conclusions.
The user equilibrium of equal travel times is found for a fraction
N1e B2 1 A1 A2
D C (12.3)
N B1 C B2 N B1 C B2
of persons choosing route 1. In contrast, the system optimum corresponds to the

maximum of the overall inverse travel times N1 =T1 .N1 / C N2 =T2 .N2 / and is found
for the fraction
N1o B2 1 A1 A2
D C (12.4)
N B1 C B2 2N B1 C B2
of 1-decisions. The difference between both fractions vanishes in the limit N ! 1.
Therefore, only experiments with a few players allow to find out, whether the
test persons adapt to the user equilibrium or to the system optimum. We will see
that both cases have completely different dynamical implications: While the most
successful strategy to establish the user equilibrium is to stick to the same decision
in subsequent iterations [27, 65, 66], the system optimum can only be reached by a
time-dependent strategy (at least, if no participant is ready to pay for the profits of
others).
Note that alternative routes can reach comparable travel times only when the total
number N of vehicles is large enough to fulfil the relationships P1 .N / < P2 .0/ D A2
and P2 .N / < P1 .0/ D A1 . Our route choice game will address this traffic regime
and additionally assume N Nimax . The case Ni D Nimax corresponds to a
complete gridlock on route i .
Finally, it may be interesting to connect the previous quantities with the vehicle
densities i and the traffic flows Qi : If route i consists of Ii lanes, the relation with
the average vehicle density is i .Ni / D Ni =.Ii Li /, and the relation with the traffic
flow is Qi .Ni / D i Vi .Ni / D Ni =ŒIi Ti .Ni /.
In the following, we will linearly transform the inverse travel time 1=Ti .Ni / in
order to define the so-called payoff
Pi .Ni / D Ci Di Ni (12.5)
12.2 The Route Choice Game 215
for choosing route i . The payoff parameters Ci and Di depend on the parameters
Ai , Bi , and N , but will be taken constant. We have scaled the parameters so that
we have the payoff Pi .Nie / D 0 (zero payoff points) in the user equilibrium and the
payoff N1 P1 .N1o / C N2 P2 .N N1o / D 100N (an average of 100 payoff points)
in the system optimum. This serves to reach generalizable results and to provide a
better orientation to the test persons.
Note that the investigation of social (multi-person) games with linearly falling
payoffs is not new [33]. For example, Schelling [62] has discussed situations
with “conditional externality”, where the outcome of a decision depends on the
independent decisions of potentially many others [62]. Pigou has addressed this
problem, which has been recently focused on by Schreckenberg and Selten’s project
SURVIVE [7, 65, 66] and others [8, 41, 58].
The route choice game is a special congestion game [22, 47, 60]. More precisely
speaking, it is a multi-stage symmetrical N -person single commodity congestion
game [68]. Congestion games belong to the class of “potential games” [48], for
which many theorems are available. For example, it is known that there always
exists a Wardrop equilibrium [71] with essentially unique Nash flows [4]. This
is characterized by the property that no individual driver can decrease his or her
travel time by a different route choice. If there are several alternative routes from
a given origin to a given destination, the travel times on all used alternative routes
in the Wardrop equilibrium is the same, while roads with longer travel times are
not used. However, the Wardrop equilibrium as expected outcome of selfish routing
does not generally reach the system optimum, i.e. minimize the total travel times.
Nash flows are often inefficient, and selfish behavior implies the possibility of
decreased network performance.2 This is particularly pronounced for the Braess
paradox [10, 61], according to which additional streets may sometimes increase the
overall travel time and reduce the throughput of a road network. The reason for this
is the possible existence of badly performing Nash equilibria, in which no single
person can improve his or her payoff by changing the decision behavior.
In fact, recent laboratory experiments indicate that, in a “day-to-day route
choice scenario” based on selfish routing, the distribution of individuals over the
alternative routes is fluctuating around the Wardrop equilibrium [27,63]. Additional
conclusions from the laboratory experiments by Schreckenberg, Selten et al. are as
follows [65, 66]:
• Most people, who change their decision frequently, respond to their experience
on the previous day (i.e. in the last iteration).
• There are only a few different behavioral patterns: direct responders (44%),
contrarian responders (14%), and conservative persons, who do not respond to
the previous outcome.
2
For more details see the work by T. Roughgarden.
• It does not pay off to react to travel time information in a sensitive way, as
conservative test persons reach the smallest travel times (the largest payoffs) on
average.
• People’s reactions to short term travel forecasts can invalidate these. Neverthe-
less, travel time information helps to match the Wardrop equilibrium, so that
excess travel times due to coordination problems are reduced.
A closer experimental analysis based on longer time series (i.e. more iterations) for
smaller groups of test persons reveals a more detailed picture [26]:
• Individuals do not only show an adaptive behavior to the travel times on the
previous day, but also change their response pattern in time [26, 34].
• In the course of time, one finds a differentiation process which leads to the
development of characteristic, individual response patterns, which tend to be
almost deterministic (in contrast to mixed strategies).
• While some test persons respond to small differences in travel times, others only
react to medium-sized deviations, further people respond to large deviations,
etc. In this way, overreactions of the group to deviations from the Wardrop
equilibrium are considerably reduced.
Note that the differentiation of individual behaviors is a way to resolve the
coordination problem to match the Wardrop equilibrium exactly, i.e. which partici-
pant should change his or her decision in the next iteration in order to compensate
for a deviation from it. This implies that the fractions of specific behavioral response
patterns should depend on the parameters of the payoff function. A certain fraction
of “stayers”, who do not respond to travel time information, can improve the
coordination in the group, i.e. the overall performance. However, stayers can also
prevent the establishment of a system optimum, if alternating reciprocity is needed,
see (12.14).
12.3 Classification of Symmetrical 2 2 Games
In contrast to previous laboratory experiments, we have studied the route choice

game not only with a very high number of repetitions, but also with a small
number N 2 f2; 4g of test persons, in order to see whether the system optimum
or the Wardrop equilibrium is established. Therefore, let us shortly discuss how the
2-person game relates to previous game-theoretical studies.
Iterated symmetrical two-person games have been intensively studied [12, 18],
including Stag Hunt, the Battle of the Sexes, or the Chicken Game (see Fig. 12.2).
They can all be represented by a payoff matrix of the form P D .Pij /, where Pij
is the success (“payoff”) of person 1 in a one-shot game when choosing strategy
i 2 f1; 2g and meeting strategy j 2 f1; 2g. The respective payoffs of the second
person are given by the symmetrical values Pj i . Figure 12.2 shows a systematics of
the previously mentioned and other kinds of symmetrical two-person games [21].
12.3 Classification of Symmetrical 2 2 Games 217
P21
Prisoner′s Leader
Notation: Chicken Battle
Dilemma
1 2 of the Sexes
–200 0 P12
Strategy 1 0 P12 Stag Hunt Harmony Route Choice
–200
Strategy 2 P21 –200
Pure
Deadlock
Coordination
Fig. 12.2 Classification of symmetrical 2 2 games according to their payoffs Pij . Two payoff
values have been kept constant as payoffs may be linearly transformed and the two strategies of
the one-shot game renumbered. Our choice of P11 D 0 and P22 D 200 was made to define a
payoff of 0 points in the user equilibrium and an average payoff of 100 in the system optimum of
our investigated route choice game with P12 D 300 and P21 D 100
a 1 2 b Coop.Def. c Coop.Def. d 1 2
Strategy 1 0 P12 Cooperation 0 –300 Cooperation 0 –300 Route 1 0 300
Strategy 2 P21 –200 Defection 100 –200 Defection 500 –200 Route 2 –100 –200
Fig. 12.3 Payoff specifications of the symmetrical 2 2 games investigated in this paper. (a)
General payoff matrix underlying the classification scheme of Fig. 12.2. (b), (c) Two variants of the
Prisoner’s Dilemma. (d) Route choice game with a strategical conflict between the user equilibrium
and the system optimum
The relations
P21 > P11 > P22 > P12 ; (12.6)
for example, define a Prisoner’s Dilemma. In this paper, however, we will mainly
focus on the 2-person route choice game defined by the conditions
P12 > P11 > P21 > P22 (12.7)
(see Fig. 12.3). Despite some common properties, this game differs from the
minority game [16, 39, 43] or El Farol bar problem [1] with P12 ; P21 > P11 ; P22 ,
as a minority decision for alternative 2 is less profitable than a majority decision for
alternative 1. Although oscillatory behavior has been found in the minority game
as well [9, 15, 16, 36, 43], an interesting feature of the route choice experiments
discussed in the following is the regularity and phase-coordination (coherence) of
the oscillations.
The 2-person route choice game fits well into the classification scheme of
symmetrical 2 2 games. In Rapoport and Guyer’s taxonomy of 2 2 games [57],
the 2-person route choice game appears on page 211 as game number 7 together
with four other games with strongly stable equilibria. Since then, the game has
almost been forgotten and did not have a commonly known interpretation or name.
Therefore, we suggest to name it the 2-person “route choice game”. Its place in
the extended Eriksson-Lindgren scheme of symmetrical 2 2 games is graphically
illustrated in Fig. 12.2.
According to the game-theoretical literature, there are 12 ordinally distinct,
symmetric 2 2 games [57], but after excluding strategically trivial games in the
sense of having equilibrium points that are uniquely Pareto-efficient, there remain
four archetypical 2 2 games: the Prisoner’s Dilemma, the Battle of the Sexes,
Chicken (Hawk-Dove), and Leader [56]. However, this conclusion is only correct,
if the four payoff values Pij are specified by the four values f1; 2; 3; 4g. Taking
different values would lead to a different conclusion: If we name subscripts so that
P11 > P22 , a strategical conflict between a user equilibrium and the system optimum
results when
P12 C P21 > 2P11 : (12.8)
Our conjecture is that players tend to develop alternating forms of reciprocity if this
condition is fulfilled, while symmetric reciprocity is found otherwise. This has the
following implications (see Fig. 12.2):
• If the 2 2 games Stag Hunt, Harmony, or Pure Coordination are repeated
frequently enough, we expect always a symmetrical form of cooperation.
• For Leader and the Battle of the Sexes, we expect the establishment of asym-
metric reciprocity, as has been found by Browning and Colman with a computer
simulation based on a genetic algorithm incorporating mutation and crossing-
over [11].
• For the games Route Choice, Deadlock, Chicken, and Prisoner’s Dilemma both,
symmetric (simultaneous) and asymmetric (alternating) forms of cooperation are
possible, depending on whether condition (12.8) is fulfilled or not. Note that
this condition cannot be met for some games, if one restricts to ordinal payoff
values Pij 2 f1; 2; 3; 4g only. Therefore, this interesting problem has been
largely neglected in the past (with a few exceptions, e.g. [51]). In particular,
convincing experimental evidence of alternating reciprocity is missing. The
following sections of this paper will, therefore, not only propose a simulation
model, but also focus on an experimental study of this problem, which promises
interesting new results.
12.4 Experimental Results
Altogether we have carried out more than 80 route choice experiments with
different experimental setups, all with different participants. In the 24 two-person
[12 four-person] experiments evaluated here (see Figs. 12.4–12.15), test persons
were instructed to choose between two possible routes between the same origin
and destination. They knew that route 1 corresponds to a “freeway” (which may be
fast or congested), while route 2 represents an alternative route (a “side road”).
Test persons were also informed that, if two [three] participants would choose
12.4 Experimental Results 219
20000
Participant
2 Participant 1
Decisions
Cummulative Payoff
1 15000 Participant 2
1
2
2 10000
1
5000
2
N1
1 0
0
–5000
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Iteration t Iteration t
Fig. 12.4 Representative example for the emergence of coherent oscillations in a 2-person
route choice experiment with the parameters specified in Fig. 12.3d. Top left: Decisions of both
participants over 300 iterations. Bottom left: Number N1 .t / of 1-decisions over time t . Note that
N1 D 1 corresponds to the system optimum, while N1 D 2 corresponds to the user equilibrium
of the one-shot game. Right: Cumulative payoff of both players in the course of time t (i.e. as
a function of the number of iterations). Once the coherent oscillatory cooperation is established
(t > 220), both individuals have high payoff gains on average
20000
2
Participant
Participant 1
Cummulative Payoff
Decisions
1
2
2 10000
1
5000
2
N1
1 0
0
–5000
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Fig. 12.5 Representative example for a 2-person route choice experiment, in which no alternating
cooperation was established. Due to the small changing frequency of participant 1, there were not
enough cooperative episodes that could have initiated coherent oscillations. Top left: Decisions of
both participants over 300 iterations. Bottom left: Number N1 .t / of 1-decisions over time t . Right:
The cumulative payoff of both players in the course of time t shows that the individual with the
smaller changing frequency has higher profits
route 1, everyone would receive 0 points, while if half of the participants would
choose route 1, they would receive the maximum average amount of 100 points,
but 1-choosers would profit at the cost of 2-choosers. Finally, participants were
told that everyone could reach an average of 100 points per round with variable,
situation-dependent decisions, and that the (additional) individual payment after the
experiment would depend on their cumulative payoff points reached in at least 300
rounds (100 points = 0.01 EUR).
Let us first focus on the two-person route-choice game with the payoffs P11 D
P1 .2/ D 0, P12 D P1 .1/ D 300, P21 D P2 .1/ D 100, and P22 D P2 .2/ D 200
(see Fig. 12.3d), corresponding to C1 D 600, D1 D 300, C2 D 0, and D2 D 100.
For this choice of parameters, the best individual payoff in each iteration is obtained
by choosing route 1 (the “freeway”) and have the co-player(s) choose route 2.
Choosing route 1 is the dominant strategy of the one-shot game, and players are
tempted to use it. This produces an initial tendency towards the “strongly stable”
user equilibrium [57] with 0 points for everyone. However, this decision behavior
is not Pareto efficient in the repeated game. Therefore, after many iterations, the
players often learn to establish the Pareto optimum of the multi-stage supergame by
12 30
Absolute Frequency
Absolute Frequency
10 25
of Participants
of Participants
8 20
6 15
4 10
2 5
0 0
-100 -50 0 50 100 150 200 -100 -50 0 50 100 150 200 250 300
Average Payoff per Iteration Average Payoff per Iteration
(Iterations 1-50) (Iterations 250-300)
Fig. 12.6 Frequency distributions of the average payoffs of the 48 players participating in our
24 two-person route choice experiments. Left: Distribution during the first 50 iterations. Right:
Distribution between iterations 250 and 300. The initial distribution with a maximum close to
0 points (left) indicates a tendency towards the user equilibrium corresponding to the dominant
strategy of the one-shot game. However, after many iterations, many individuals learn to establish
the system optimum with a payoff of 100 points (right)
selecting route 1 in turns (see Fig. 12.4). As a consequence, the experimental payoff
distribution shows a maximum close to 0 points in the beginning and a peak at 100
points after many iterations (see Fig. 12.6), which clearly confirms that the choice
behavior of test persons tends to change over time. Nevertheless, in 7 out of 24 two-
person experiments, persistent cooperation did not emerge during the experiment.
Later on, we will identify reasons for this.
12.4.1 Emergence of Cooperation3 and Punishment
In order to reach the system optimum of .100C300/=2 D 100 points per iteration,
one individual has to leave the freeway for one iteration, which yields a reduced
payoff of –100 in favour of a high payoff of C300 for the other individual. To be
profitable also for the first individual, the other one should reciprocate this “offer”
by switching to route 2, while the first individual returns to route 1. Establishing
this oscillatory cooperative behavior yields 100 extra points on average. If the
other individual is not cooperative, both will be back to the user equilibrium of
0 points only, and the uncooperative individual has temporarily profited from the
offer by the other individual. This makes “offers” for cooperation and, therefore,
the establishment of the system optimum unlikely.
Hence, the innovation of oscillatory behavior requires intentional or random
changes (“trial-and-error behavior”). Moreover, the consideration of multi-period
decisions is helpful. Instead of just 2 one-stage (i.e. one-period) alternative deci-
sions 1 and 2, there are 2n different n-stage (n-period) decisions. Such multi-
3
The term cooperation is used here, because a coordination in time is only part of the problem.
Individuals are also facing a dilemma situation, in which selfish behavior tends to prevent
cooperation and temporal coordination.
30000
Participant 1
Participant
Cummulative Payoff
Decisions
1 20000
2 2
15000
1
10000
2 5000
N1
1 0
0
-5000
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Fig. 12.7 Representative example for a 2-person route choice experiment, in which participant
1 leaves the pattern of oscillatory cooperation temporarily in order to make additional profits.
Note that participant 2 does not “punish” this selfish behavior, but continues to take routes in an
alternating way. Top left: Decisions of both participants over 300 iterations. Bottom left: Number
N1 .t / of 1-decisions over time t . Right: Cumulative payoff of both players as a function of the
number of iterations. The different slopes indicate an unfair outcome despite of high average
payoffs of both players
11 12 21 22
2
11 0 3 300 300 600
3 1
1 2
4
Route 1 0 300 12 –100 –200 200 100

2 4
Route 2 –100 –200 21 –100 200 –200 100

1
22 –200 –300 –300 –400
Fig. 12.8 Illustration of the concept of higher-order games defined by n-stage strategies. Left:
Payoff matrix P D .Pij / of the one-shot 2 2 route choice game. Right: Payoff matrix
.2/
P.i1 i2 /;.j1 j2 / D .Pi1 j1 C Pi2 j2 / of the 2nd-order route choice game defined by 2-stage decisions
(right). The analysis of the one-shot game (left) predicts that the user equilibrium (with both
persons choosing route 1) will establish and that no single player could increase the payoff by
another decision. For two-period decisions (right), the system optimum (strategy 12 meeting
strategy 21) corresponds to a fair solution, but one person can increase the payoff at the cost of
the other (see arrow 1), if the game is repeated. A change of the other person’s decision can reduce
losses and punish this egoistic behavior (arrow 2), which is likely to establish the user equilibrium
with payoff 0. In order to leave this state again in favour of the system optimum, one person will
have to make an “offer” at the cost of a reduced payoff (arrow 3). This offer may be due to a
random or intentional change of decision. If the other person reciprocates the offer (arrow 4), the
system optimum is established again. The time-averaged payoff of this cycle lies below the system
optimum
stage strategies can be used to define higher-order games and particular kinds
of supergame strategies. In the two-person 2nd-order route choice game, for
example, an encounter of the two-stage decision 12 with 21 establishes the system
optimum and yields equal payoffs for everyone (see Fig. 12.8). Such an optimal
and fair solution is not possible for one-stage decisions. Yet, the encounter of
12 with 21 (“cooperative episode”) is not a Nash equilibrium of the two-stage
Proportion of Groups
0.8
0.6
0.4
0.2 Empirical Data

Fit Function
0
0 5 10 15 20 25 30 35 40 45 50
Required Cooperative Episodes n
Fig. 12.9 Cumulative distribution of required cooperative episodes until persistent cooperation
was established, given that cooperation occurred during the duration of the game as in 17 out of 24
two-person experiments. The experimental data are well approximated by the logistic curve (12.9)
with the fit parameters c2 D 3:4 and d2 D 0:17
game, as an individual can increase his or her own payoff by selecting 11 (see
Fig. 12.8). Probably for this reason, the first cooperative episodes in a repeated route
choice game (i.e. encounters of 12-decisions with 21-decisions in two subsequent
iterations) do often not persist (see Fig. 12.9). Another possible reason is that
cooperative episodes may be overlooked. This problem, however, can be reduced
by a feedback signal that indicates when the system optimum has been reached.
For example, we have experimented with a green background color. In this setup,
a cooperative episode could be recognized by a green background that appeared in
two successive iterations together with two different payoff values.
The strategy of taking route 1 does not only dominate on the first day (in the first
iteration). Even if a cooperative oscillatory behavior has been established, there is
a temptation to leave this state, i.e. to choose route 1 several times, as this yields
more than 100 points on average for the uncooperative individual at the cost of
the participant continuing an alternating choice behavior (see Figs. 12.7 and 12.8).
That is, the conditional changing probability pl .2j1; N1 D 1I t/ of individuals
l from route 1 to route 2, when the system optimum in the previous iteration
was established (i.e. N1 D 1) tends to be small initially. However, oscillatory
cooperation of period 2 needs pl .2j1; N1 D 1I t/ D 1. The required transition in the
decision behavior can actually be observed in our experimental data (see Fig. 12.10,
left). With this transition, the average frequency of 1-decisions goes down to 1/2 (see
Fig. 12.10, right). Note, however, that alternating reciprocity does not necessarily
require oscillations of period 2. Longer periods are possible as well (see Fig. 12.11),
but have occurred only in a few cases (namely, 3 out of 24 cases).
How does the transition to oscillatory cooperation come about? The estab-
lishment of alternating reciprocity can be supported by a suitable punishment
strategy: If the other player should have selected route 2, but has chosen route
1 instead, he or she can be punished by changing to route 1 as well, since
this causes an average payoff of less than 100 points for the other person (see
Fig. 12.8). Repeated punishment of uncooperative behavior can, therefore, reinforce
Proportion of 1-Decisions
1 1
Transition Probability
p1(2|1, N1 = 1;t)
0.8 p2(2|1, N1 = 1;t) 0.8
0.6 0.6
0.4 0.4
0.2 0.2 P1(1, t)
P2(1, t)
0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Fig. 12.10 Left: Conditional changing probability pl .2j1; N1 D 1I t / of person l from route 1
(the “freeway”) to route 2, when the other person has chosen route 2, averaged over a time window
of 50 iterations. The transition from initially small values to 1 (for t > 240) is characteristic and
illustrates the learning of cooperative behavior. In this particular group (cf. Fig. 12.4) the values
started even at zero, after a transient time period of t < 60. Right: Proportion Pl .1; t / of 1-decisions
of both participants l in the two-person route choice experiment displayed in Fig. 12.4. While the
initial proportion is often close to 1 (the user equilibrium), it reaches the value 1/2 when persistent
oscillatory cooperation (the system optimum) is established
20000
Participant 1
Participant
2
Cummulative Payoff
Decisions
1
2
2 10000
1
5000
2
N1
1 0
0
-5000
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Fig. 12.11 Representative example for a 2-person route choice experiment with phase-
coordinated oscillations of long (and varying) time periods larger than 2. Top left: Decisions
of both participants over 300 iterations. Bottom left: Number N1 .t / of 1-decisions over time t .
Right: Cumulative payoff of both players as a function of the number of iterations. The sawtooth-
like increase in the cumulative payoff indicates gains by phase-coordinated alternations with long
oscillation periods
cooperative oscillatory behavior. However, the establishment of oscillations also

requires costly “offers” by switching to route 2, which only pay back in case of
alternating reciprocity. It does not matter whether these “offers” are intentional or
due to exploratory trial-and-error behavior.
Due to punishment strategies and similar reasons, persistent cooperation is often
established after a number n of cooperative episodes. In the 17 of our 24 two-
person experiments, in which persistent cooperation was established, the cumulative
distribution of required cooperative episodes could be mathematically described by
the logistic curve
FN .n/ D 1=Œ1 C cN exp.dN n/ (12.9)
(see Fig. 12.9). Note that, while we expect that this relationship is generally valid,
the fit parameters cN and dN may depend on factors like the distribution of
participant intelligence, as oscillatory behavior is apparently difficult to establish
(see below).
12.4.2 Preconditions for Cooperation
Let us focus on the time period before persistent oscillatory cooperation is estab-
lished and denote the occurrence probability that individual l chooses alternative
i 2 f1; 2g by Pl .i /. The quantity pl .j ji / shall represent the conditional probability
of choosing j in the next iteration, if i was chosen by person l in the present one.
Assuming stationarity for reasons of simplicity, we expect the relationship
pl .2j1/Pl .1/ D pl .1j2/Pl .2/; (12.10)
i.e. the (unconditional) occurrence probability Pl .1; 2/ D pl .2j1/Pl .1/ of having

alternative 1 in one iteration and 2 in the next agrees with the joint occurrence
probability Pl .2; 1/ D pl .1j2/Pl .2/ of finding the opposite sequence 21 of
decisions:
Pl .1; 2/ D Pl .2; 1/: (12.11)
Moreover, if rl denotes the average changing frequency of person l until persistent
cooperation is established, we have the relation
rl D Pl .1; 2/ C Pl .2; 1/: (12.12)
Therefore, the probability that allQN players simultaneously change their decision
from one iteration to the next is N N
lD1 rl . Note that there are 2 such realizations
of N decision changes 12 or 21, which have all the same occurrence probability
because of (12.11). Among these, only the ones where N=2 players change from 1 to
2 and the other N=2 participants change from 2 to 1 establish cooperative episodes,
given that the system optimum corresponds to an equal distribution over both
alternatives. Considering that the number of different possibilities of selecting N=2
out of N persons is given by the binomial coefficient, the occurrence probability of
cooperative events is
Y N
1 N
Pc D N rl (12.13)
2 N=2
lD1
(at least in the ensemble average). Since the expected time period T until the
cooperative state incidentally occurs equals the inverse of Pc , we finally find the
formula
2 YN
1 N .N=2/Š 1
T D D2 : (12.14)
Pc NŠ rl
lD1
This formula is well confirmed by our 2-person experiments (see Fig. 12.12). It
gives the lower bound for the expected value of the minimum number of required
iterations until persistent cooperation can spontaneously emerge (if already the first
cooperative episode is continued forever).
Obviously, the occurrence of oscillatory cooperation is expected to take much
longer for a large number N of participants. This tendency is confirmed by our
4-person experiments compared to our 2-person experiments. It is also in agreement
Fig. 12.12 Comparison of 100

the required number of 2-Person Games

cooperative episodes y with
the expected number x of
cooperative episodes
10
(approximated as occurrence
time of persistent
cooperation, divided by the
expected time interval T until
a cooperative episode occurs 1
by chance). Note that the data
points support the
relationship y D x and,
thereby, formula (12.14) 0.1
0.1 1 10 100
Expected Cooperative Episodes
with intuition, as coordination of more people is more difficult. (Note that mean first
passage or transition times in statistical physics tend to grow exponentially in the
number N of particles as well.)
Besides the number N of participants, another critical factor for the cooperation
probability are the changing frequencies rl : They are needed for the exploration
of innovative strategies, coordination and cooperation. Although the instruction of
test persons would have allowed them to conclude that taking turns would be a
good strategy, the changing frequencies rl of some individuals was so small that
cooperation within the duration of the respective experiment did not occur, in
accordance with formula (12.14). The unwillingness of some individuals to vary
their decisions is sometimes called “conservative” [7, 65, 66] or “inertial behavior”
[9]. Note that, if a player never reciprocates “offers” by other players, this may
discourage further “offers” and reduce the changing frequency of the other player(s)
as well (see the decisions 50 through 150 of player 2 in Fig. 12.4).
Our experimental time series show that most individuals initially did not know a
periodic decision behavior would allow them to establish the system optimum. This
indicates that the required depth of strategic reasoning [19] and the related com-
plexity of the game for an average person are already quite high, so that intelligence
may matter. Compared to control experiments, the hint that the maximum average
payoff of 100 points per round could be reached “by variable, situation-dependent
decisions”, increased the average changing frequency (by 75 percent) and with
this the occurrence frequency of cooperative events. Thereby, it also increased the
chance that persistent cooperation established during the duration of the experiment.
Note that successful cooperation requires not only coordination [9], but also
innovation: In their first route choice game, most test persons discover the oscillatory
cooperation strategy only by chance in accordance with formula (12.14). The chang-
ing frequency is, therefore, critical for the establishment of innovative strategies:
It determines the exploratory trial-and-error behavior. In contrast, cooperation is
easy when test persons know that the oscillatory strategy is successful: When two
Decisions
2
Particip.
1 1 1
2 1
2 2
Participant
1
Decisions
2
2 1
N1
1 2
0 3
1
0 50 100 150 200 250 300 2
4
2 1
Decisions
Particip.
1
1
2 4
2 3
1
N1
2
2 1
N1
1
0 0
0 50 100 150 200 250 300 0 50 100 150 200 250
Fig. 12.13 Experimentally observed decision behavior when two groups involved in two-person
route choice experiments afterwards played a four-person game with C1 D 900, D1 D 300, C2 D
100, D2 D 100. Left: While oscillations of period 2 emerged in the second group (bottom), another
alternating pattern corresponding to n-period decisions with n > 2 emerged in the first group
(top). Right: After all persons had learnt oscillatory cooperative behavior, the four-person game
just required coordination, but not the invention of a cooperative strategy. Therefore, persistent
cooperation was quickly established (in contrast to four-person experiments with new participants).
It is clearly visible that the test persons continued to apply similar decision strategies (right) as in
the previous two-person experiments (left)
teams, who had successfully cooperated in 2-person games, had afterwards to play
a 4-person game, cooperation was always and quickly established (see Fig. 12.13).
In contrast, unexperienced co-players suppressed the establishment of oscillatory
cooperation in 4-person route choice games.
12.4.3 Strategy Coefficients
In order to characterize the strategic behavior of individuals and predict their

chances of cooperation, we have introduced some strategy coefficients. For this,
let us introduce the following quantities, which are determined from the iterations
before persistent cooperation is established:
• clk D relative frequency of a changed subsequent decision of individual l if the
payoff was negative (k D ), zero (k D 0), or positive (k D C).
• slk D relative frequency of individual l to stay with the previous decision if the
payoff was negative (k D ), zero (k D 0), or positive (k D C).
The Yule-coefficient
cl slC clC sl
Ql D (12.15)
cl slC C clC sl
with 1 Ql 1 was used by Schreckenberg, Selten et al. [65] to identify direct

responders with 0:5 < Ql 1 (who change their decision after a negative payoff
and stay after a positive payoff), and contrarian responders with 0:5 > Ql 1
(who change their decision after a positive payoff and stay after a negative one).
A random decision behavior would correspond to a value Ql 0. However, a

problem arises if one of the variables cl , slC , clC , or sl assumes the value 0.
Then, we have Ql 2 f1; 1g, independently of the other three values. If two of
the variables become zero, Ql is sometimes even undefined. Moreover, if the values
are small, the resulting conclusion is not reliable. Therefore, we prefer to use the
percentage difference
c cC
Sl D l l C l C (12.16)
cl C sl cl C sl
for the assessment of strategies. Again, we have 1 Sl 1. Direct responders
correspond to Sl > 0:25 and contrarian responders to Sl < 0:25. For 0:25
Sl 0:25, the response to the previous payoff is rather random.
In addition, we have introduced the Z-coefficient
cl0
Zl D ; (12.17)
cl0 C sl0
for which we have 0 Zl 1. This coefficient describes the likely response of

individual l to the user equilibrium. Zl D 0 means that individual l does not change
routes, if the user equilibrium was reached. Zl D 1 implies that person l always
changes, while Zl 0:5 indicates a random response.
Figure 12.14 shows the result of the 2-person route choice experiments (cooper-
ation or not) as a function of S1 and S2 , and as a function of Z1 and Z2 . Moreover,
Fig. 12.15 displays the result as a function of the average strategy coefficients
1 X
N
ZD Zl (12.18)
N
lD1
1 1
S-Coefficient of Participant 2
Z-Coefficient of Participant 2
0.8
0.8 Participant in Cooperating Group
0.6 Part. in Non-Cooperating Group
0.4 0.6
0.2
0.4
0
-0.2 0.2
Participant in Cooperating Group
-0.4 Part. in Non-Cooperating Group
0
-0.4 -0.2 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
S-Coefficient of Participant 1 Z-Coefficient of Participant 1
Fig. 12.14 Coefficients Sl and Zl of both participants l in all 24 two-person route choice
games. The values of the S-coefficients (i.e. the individual tendencies towards direct or contrarian
responses) are not very significant for the establishment of persistent cooperation, while large
enough values of the Z-coefficient stand for the emergence of oscillatory cooperation
1
Cooperating Groups
Groups without Cooperation
Mean Z-Coefficient
0.8
of Participants
0.6
0.4
0.2
0
-0.4 -0.2 0 0.2 0.4 0.6 0.8
Mean S-Coefficient of Participants
Fig. 12.15 S- and Z-coefficients averaged over both participants in all 24 two-person route
choice games. The mainly small, but positive values of S indicate a slight tendency towards
direct responses. However, the S-coefficient is barely significant for the emergence of persistent
oscillations. A good indicator for their establishment is a sufficiently large Z-value
and
1 X
N
SD Sl : (12.19)
N
lD1
Our experimental data indicate that the Z-coefficient is a good indicator for the
establishment of cooperation, while the S -coefficient seems to be rather insignifi-
cant (which also applies to the Yule coefficient).
12.5 Multi-agent Simulation Model
In a first attempt, we have tried to reproduce the observed behavior in our 2-person
route choice experiments by game-dynamical equations [28]. We have applied these
to the 2 2 route choice game and its corresponding two-, three- and four-stage
higher-order games (see Sect. 12.4.1). Instead of describing patterns of alternating
cooperation, however, the game dynamical equations predicted a preference for the
dominant strategy of the one-shot game, i.e. a tendency towards choosing route 1.
The reason for this becomes understandable through Fig. 12.8. Selecting routes
2 and 1 in an alternating way is not a stable strategy, as the other player can get a
higher payoff by choosing two times route 1 rather than responding with 1 and 2.
Selecting route 1 all the time even guarantees that the own payoff is never below the
one by the other player. However, when both players select route 1 and establish the
related user equilibrium, no player can improve his or her payoff in the next iteration
by changing the decision. Nevertheless, it is possible to improve the long-term
outcome, if both players change their decisions, and if they do it in a coordinated
way. Note, however, that a strict alternating behavior of period 2 is an optimal
strategy only in infinitely repeated games, while it is unstable to perturbations in
finite games.
12.5 Multi-agent Simulation Model 229
20000
Cummulative Payoff
2 Agent 1
Decisions
1 Agent 2
Agent
1 15000
2
2 10000
1
5000
2
N1
1 0
0
-5000
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Fig. 12.16 Representative example for a 2-person route choice simulation based on our proposed
multi-agent reinforcement learning model with Pavmax D 100 and Pavmin D 200. The parameter
l1 has been set to 0.25. The other model parameters are specified in the text. Top left: Decisions
of both agents over 300 iterations. Bottom left: Number N1 .t / of 1-decisions over time t . Right:
Cumulative payoff of both agents as a function of the number of iterations. The emergence of
oscillatory cooperation is comparable with the experimental data displayed in Fig. 12.4
It is known that cooperative behavior may be explained by a “shadow of

the future” [2, 3], but it can also be established by a “shadow of the past”
[40], i.e. experience-based learning. This will be the approach of the multi-agent
simulation model proposed in this section. As indicated before, the emergence
of phase-coordinated strategic alternation (rather than a statistically independent
application of mixed strategies) requires an almost deterministic behavior (see
Fig. 12.16). Nevertheless, some weak stochasticity is needed for the establishment
of asymmetric cooperation, both for the exploration of innovative strategies and
for phase coordination. Therefore, we propose the following reinforcement learning
model, which could be called a generalized win-stay, lose-shift strategy [50, 54].
Let us presuppose that an individual approximately memorizes or has a good
feeling of how well he or she has performed on average in the last nl iterations
and since he or she has last responded with decision j to the situation .i; N1 /.
In our success- and history-dependent model of individual decision behavior,
pl .j ji; N1 I t/ denotes agent l’s conditional probability of taking decision j at time
t C 1, when i was selected at time t and N1 .t/ agents had chosen alternative 1.
Assuming that pl is either 0 or 1, pl .j ji; N1 I t/ has the meaning of a deterministic
response strategy: pl .j ji; N1 I t/ D 1 implies that individual l will respond at time
t C 1 with the decision j to the situation .i; N1 / at time t.
Our reinforcement learning strategy can be formulated as follows: The response
strategy pl .j ji; N1 ; t/ is switched with probability ql > 0, if the average individual
payoff since the last comparable situation with i.t 0 / D i.t/ and N1 .t 0 / D N1 .t/
at time t 0 < t is less than the average individual payoff P l .t/ during the last nl
iterations. In other words, if the time-dependent aspiration level P l .t/ [40, 54] is
not reached by the agent’s average payoff since his or her last comparable decision,
the individual is assumed to substitute the response strategy pl .j ji; N1 I t/ by
pl .j ji; N1 I t C 1/ D 1 pl .j ji; N1 I t/ (12.20)
with probability ql . The replacement of dissatisfactory strategies orients at histor-

ical long-term profits (namely, during the time period Œt 0 ; t). Thereby, it avoids
short-sighted changes after temporary losses. Moreover, it does not assume a
comparison of the performance of the actually applied strategy with hypothetical

ones as in most evolutionary models. A readiness for altruistic decisions is also
not required, while exploratory behavior (“trial and error”) is necessary. In order to
reflect this, the decision behavior is randomly switched from pl .j ji; N1 I t C 1/ to
1 pl .j ji; N1 I t C 1/ with probability
!
0 1 Pav P l .t/
max
l .t/ D max l ; l 1: (12.21)
Pavmax Pavmin
Herein, Pavmin and Pavmax denote the minimum and maximum average payoff of all
N agents (simulated players). The parameter l1 reflects the mutation frequency for
P l .t/ D Pavmin , while the mutation frequency is assumed to be l0 l1 when the
max
time-averaged payoff P l reaches the system optimum P av .
In our simulations, no emergent cooperation is found for l0 D l1 D 0. l0 > 0 or
odd values of nl may produce intermittent breakdowns of cooperation. A small,
but finite value of l1 is important to find a transition to persistent cooperation.
Therefore, we have used the parameter value l1 D 0:25, while the simplest possible
specification has been chosen for the other parameters, namely l0 D 0, ql D 1, and
nl D 2.
The initial conditions for the simulation of the route choice game were specified
in accordance with the dominant strategy of the one-shot game, i.e. Pl .1; 0/ D 1
(everyone tends to choose the freeway initially), pl .2j1; N1 I 0/ D 0 (it is not
attractive to change from the freeway to the side road) and pl .1j2; N1 I 0/ D 1 (it is
tempting to change from the side road to the freeway). Interestingly enough, agents
learnt to acquire the response strategy pl .2j1; N1 D 1I t/ D 1 in the course of time,
which established oscillatory cooperation with higher profits (see Figs. 12.16 and
12.17).
Proportion of 1-Decisions
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 p1(2|1, N1 = 1) 0.2 P1(1, t)
p2(2|1, N1 = 1) P2(1, t)
0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Fig. 12.17 Left: Conditional changing probability pl .2j1; N1 D 1I t / of agent l from route 1 (the
“freeway”) to route 2, when the other agent has chosen route 2, averaged over a time window
of 50 iterations. The transition from small values to 1 for the computer simulation displayed in
Fig. 12.16 is characteristic and illustrates the learning of cooperative behavior. Right: Proportion
Pl .1; t / of 1-decisions of both participants l in the two-person route choice experiment displayed
in Fig. 12.16. While the initial proportion is often close to 1 (the user equilibrium), it reaches
the value 1/2 when persistent oscillatory cooperation (the system optimum) is established. The
simulation results are compatible with the essential features of the experimental data (see, for
example, Fig. 12.10)
12.5 Multi-agent Simulation Model 231

100 1
Proportion of Groups
0.8
10
0.6
0.4
1
0.2 Simulation Data
2-Agent Simulations Fit Function
0.1 0
0.1 1 10 100 0 5 10 15 20 25 30
Expected Cooperative Episodes Required Cooperative Episodes n
Fig. 12.18 Left: Comparison of the required number of cooperative episodes with the expected
number of cooperative episodes in our multi-agent simulation of decisions in the route choice
game. Note that the data points support formula (12.14). Right: Cumulative distribution of required
cooperative episodes until persistent cooperation is established in our 2-person route choice
simulations, using the simplest specification of model parameters (not calibrated). The simulation
data are well approximated by the logistic curve (12.9) with the fit parameters c2 D 7:9 and
d2 D 0:41
Note that the above described reinforcement learning model [40] responds only
to the own previous experience [13]. Despite its simplicity (e.g. the neglection of
more powerful, but probably less realistic k-move memories [11]), our “multi-
agent” simulations reproduce the emergence of asymmetric reciprocity of two
or more players, if an oscillatory strategy of period 2 can establish the system
optimum. This raises the question why previous experiments of the N -person
route choice game [27, 63] have observed a clear tendency towards the Wardrop
equilibrium [71] with P1 .N1 / D P2 .N2 / rather than phase-coordinated oscillations?
It turns out that the payoff values must be suitably chosen (see (12.8)) and that
several hundred repetitions are needed. In fact, the expected time interval T until a
cooperative episode among N D N1 C N2 participants occurs in our simulations
by chance is well described by formula (12.14), see Fig. 12.18. The empirically
observed transition in the decision behavior displayed in Fig. 12.10 is qualitatively
reproduced by our computer simulations as well (see Fig. 12.17). The same applies
to the frequency distribution of the average payoff values (compare Fig. 12.19 with
Fig. 12.6) or to the number of expected and required cooperative episodes (compare
Fig. 12.18 with Figs. 12.9 and 12.12).
12.5.1 Simultaneous and Alternating Cooperation

in the Prisoner’s Dilemma
Let us finally simulate the dynamic behavior in the two different variants of
the Prisoner’s Dilemma indicated in Fig. 12.3b,c with the above experience-based
reinforcement learning model. Again, we will assume P11 D 0 and P22 D 200.
14 35
12 30
Absolute Frequency
Absolute Frequency
of Participants
of Participants
10 25
8 20
6 15
4 10
2 5
0 0
-100 -50 0 50 100 150 200 -100 -50 0 50 100 150 200 250 300
Average Payoff per Iteration (Iterations 1-50) Average Payoff per Iteration (Iterations 250-300)
Fig. 12.19 Frequency distributions of the average payoffs in our computer simulations of the
2-person route choice game. Left: Distribution during the first 50 iterations. Right: Distribution
between iterations 250 and 300. Our simulation results are compatible with the experimental data
displayed in Fig. 12.6
10000
Cummulative Payoff Agent 1
Decisions
2 5000
1 Agent 2
Agent
1
2 0
2 -5000
1
-10000
2 -15000
N1
1 -20000
0
-25000
0 50 100 150 200 250 300 0 50 100 150 200 250 300
20000
Agent 1
Cummulative Payoff
Decisions
2 15000
1 Agent 2
Agent
1
2 10000
2 5000
1
0
2 -5000
N1
1 -10000
0
-15000
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Fig. 12.20 Representative examples for computer simulations of the two different forms of the
Prisoner’s Dilemma specified in Fig. 12.3b,c. The parameter l1 has been set to 0.25, while the
other model parameters are specified in the text. Top: Emergence of simultaneous, symmetrical
cooperation, where decision 2 corresponds to defection and decision 1 to cooperation. The system
optimum corresponds to Pavmax D 0 payoff points, and the minimum payoff to Pavmin D 200.
Bottom: Emergence of alternating, asymmetric cooperation with Pavmax D 100 and Pavmin D 200.
Left: Time series of the agents’ decisions and the number N1 .t / of 1-decisions. Right: Cumulative
payoffs as a function of time t
According to (12.8), a simultaneous, symmetrical form of cooperation is expected

for P12 D 300 and P21 D 100, while an alternating, asymmetric cooperation
is expected for P12 D 300 and P21 D 500. Figure 12.20 shows simulation
results for the two different cases of the Prisoner’s Dilemma and confirms the two
predicted forms of cooperation. Again, we varied only the parameter l1 , while
we chose the simplest possible specification of the other parameters l0 D 0,
ql D 1, and nl D 2. The initial conditions were specified in accordance with
the expected non-cooperative outcome of the one-shot game, i.e. Pl .1; 0/ D 0
(everyone defects in the beginning), pl .2j2; N1 I 0/ D 0 (it is tempting to continue
defecting), pl .1j1; N1 D 1I 0/ D 0 (it is unfavourable to be the only cooperative
player), and pl .1j1; N1 D 2I 0/ D 1 (it is good to continue cooperating, if the

other player cooperates). In the course of time, agents learn to acquire the response
strategy pl .2j2; N1 D 0I t/ D 0 when simultaneous cooperation evolves, but
pl .2j2; N1 D 1I t/ D 0 when alternating cooperation is established.
In this paper, we have investigated the N -person day-to-day route-choice game.

This special congestion game has not been thoroughly studied before in the case
of small groups, where the system optimum can considerably differ from the user
equilibrium. The 2-person route choice game gives a meaning to a previously
uncommon repeated symmetrical 2 2 game and shows a transition from the
dominating strategy of the one-shot game to coherent oscillations, if P12 C P21 >
2P11 . However, a detailed analysis of laboratory experiments with humans reveals
that the establishment of this phase-coordinated alternating reciprocity, which is
expected to occur in other 2 2 games as well, is quite complex. It needs either
strategic experience or the invention of a suitable strategy. Such an innovation is
driven by the potential gains in the average payoffs of all participants and seems
to be based on exploratory trial-and-error behavior. If the changing frequency
of one or several players is too low, no cooperation is established in a long
time. Moreover, the emergence of cooperation requires certain kinds of strategies,
which can be characterized by the Z-coefficient (12.18). These strategies can be
acquired by means of reinforcement learning, i.e. by keeping response patterns
which have turned out to be better than average, while worse response patterns
are being replaced. The punishment of uncooperative behavior can help to enforce
cooperation. Note, however, that punishment in groups of N > 2 persons is
difficult, as it is hard to target the uncooperative person, and punishment hits
everyone. Nevertheless, computer simulations and additional experiments indicate
that oscillatory cooperation can still emerge in route choice games with more than
two players after a long time period (rarely within 300 iterations) (see Fig. 12.21).
Altogether, spontaneous cooperation takes a long time. It is, therefore, sensitive
to changing conditions reflected by time-dependent payoff parameters. As a conse-
quence, emergent cooperation is unlikely to appear in real traffic systems. This is the
reason why the Wardrop equilibrium tends to occur. However, cooperation could be
rapidly established by means of advanced traveler information systems (ATIS) [8,
14, 30, 37, 41, 63, 70, 73], which would avoid the slow learning process described by
(12.14). Moreover, while we do not recommend conventional congestion charges, a
charge for unfair usage patterns would support the compliance with individual route
choice recommendations. It would supplement the inefficient individual punishment
mechanism.
Different road pricing schemes have been proposed, each of which has its own
advantages and disadvantages or side effects. Congestion charges, for example,
could discourage to take congested routes, which is actually required to reach
2 2
1 1
1 1
Participant
2 2
Decisions
Decisions
Agent
2 2
1 1
2 2
3 3
1 1
2 2
4 4
1 1
4 4
3 3
N1
N1
2 2
1 1
0 0
0 50 100 150 200 250 0 50 100 150 200 250
Fig. 12.21 Emergence of phase-coordinated oscillatory behavior in the 4-person route choice
game with the parameters specified in Fig. 12.13. Left: Experimental data of the decisions of four
unexperienced participants over 300 iterations. Right: Computer simulation with the reinforcement
learning model
minimum average travel times. Conventional tolls and road pricing may reduce the
trip frequency due to budget constraints, which potentially interferes with economic
growth and fair chances for everyone’s mobility.
In order to activate capacity reserves, we therefore propose an automated route
guidance system based on the following principles: After specification of their
destination, drivers should get individual (and, on average, fair) route choice
recommendations in agreement with the traffic situation and the route choice
proportions required to reach the system optimum. If an individual selects a faster
route instead of the recommended route it should use, it will have to pay an amount
proportional to the decrease in the overall inverse travel time compared to the system
optimum. Moreover, drivers not in a hurry should be encouraged to take the slower
route i by receiving the amount of money corresponding to the related increase in
the overall inverse travel time. Altogether, such an ATIS could support the system
optimum while allowing for some flexibility in route choice. Moreover, the fair
usage pattern would be cost-neutral for everyone, i.e. traffic flows of potential
economic relevance would not be suppressed by extra costs.
In systems with many similar routing decisions, a Pareto optimum characterized
by asymmetric alternating cooperation may emerge even spontaneously. This could
help to enhance the routing in data networks [72] and generally to resolve Braess-
like paradoxes in networks [17].
Finally, it cannot be emphasized enough that taking turns is a promising
strategy to distribute scarce resources in a fair and optimal way. It could be
applied to a huge number of real-life situations due to the relevance for many
strategical conflicts, including Leader, the Battle of the Sexes, and variants of Route
Choice, Deadlock, Chicken, and the Prisoner’s Dilemma. The same applies to their
N -person generalizations, in particular social dilemmas [23, 25, 40]. It will also be
interesting to find out whether and where metabolic pathways, biological supply
networks, or information flows in neuronal and immune systems use alternating
strategies to avoid the wasting of costly resources.
References 235
Acknowledgements D.H. is grateful for the warm hospitality of the Santa Fe Institute, where the
Social Scaling Working Group Meeting in August 2003 inspired many ideas of this paper. The
results shall be presented during the workshop on “Collectives Formation and Specialization in
Biological and Social Systems” in Santa Fe (April 20–22, 2005).
References
1. W.B. Arthur, Inductive reasoning and bounded rationality. Am. Econ. Rev. 84, 406–411 (1994)
2. R. Axelrod, D. Dion, The further evolution of cooperation. Science 242, 1385–1390 (1988)
3. R. Axelrod, W.D. Hamilton, The evolution of cooperation. Science 211, 1390–1396 (1981)
4. M. Beckmann, C.B., McGuire, C.B. Winsten, Studies in the Economics of Transportation
(Yale University Press, New Haven, 1956)
5. K.G. Binmore, Evolutionary stability in repeated games played by finite automata. J. Econ.
Theor. 57, 278–305 (1992)
6. K. Binmore, Fun and Games: A Text on Game Theory (Heath, Lexington, MA, 1992),
pp. 373–377
7. U. Bohnsack, Uni DuE: Studie SURVIVE gibt Einblicke in das Wesen des Autofahrers, Press
release by Informationsdienst Wissenschaft (January 21, 2005)
8. P.W. Bonsall, T. Perry, Using an interactive route-choice simulator to investigate driver’s
compliance with route guidance information. Transpn. Res. Rec. 1306, 59–68 (1991)
9. G. Bottazzi, G. Devetag, Coordination and self-organization in minority games: Experimental
evidence, Working Paper 2002/09, Sant’Anna School of Advances Studies, May 2002
10. D. Braess, Über ein Paradoxon der Verkehrsplanung [A paradox of traffic assignment
problems]. Unternehmensforschung 12, 258–268 (1968). For about 100 related references see
http://homepage.ruhr-uni-bochum.de/Dietrich.Braess/#paradox
11. L. Browning, A.M. Colman, Evolution of coordinated alternating reciprocity in repeated
dyadic games. J. Theor. Biol. 229, 549–557 (2004)
12. C.F. Camerer, Behavioral Game Theory: Experiments on Strategic Interaction (Princeton
University Press, Princeton, 2003)
13. C.F. Camerer, T.-H. Ho, J.-K. Chong, Sophisticated experience-weighted attraction learning
and strategic teaching in repeated games. J. Econ. Theor. 104, 137–188 (2002)
14. N. Cetin, K. Nagel, B. Raney, A. Voellmy, Large scale multi-agent transportation simulations.
Comput. Phys. Comm. 147(1–2), 559–564 (2002)
15. D. Challet, M. Marsili, Relevance of memory in minority games. Phys. Rev. E 62, 1862–1868
(2000)
16. D. Challet, Y.-C. Zhang, Emergence of cooperation and organization in an evolutionary game.
Physica A 246, 407–418 (1997)
17. J.E. Cohen, P. Horowitz Paradoxial behaviour of mechanical and electrical networks. Nature
352, 699–701 (1991)
18. A.M. Colman, Game Theory and its Applications in the Social and Biological Sciences, 2nd
edn. (Butterworth-Heinemann, Oxford, 1995)
19. A.M. Colman, Depth of strategic reasoning in games. Trends Cognit. Sci. 7(1), 2–4 (2003)
20. P.H. Crowley, Dangerous games and the emergence of social structure: evolving memory-based
strategies for the generalized hawk-dove game. Behav. Ecol. 12, 753–760 (2001)
21. A. Eriksson, K. Lindgren Cooperation in an unpredictable environment, in Proceedings of
Artificial Life VIII, ed. by R.K. Standish, M.A. Bedau, H.A. Abbass (MIT Press, Sidney, 2002),
pp. 394–399 and poster available at http://frt.fy.chalmers.se/cs/people/eriksson.html
22. C.B. Garcia, W.I. Zangwill, Pathways to Solutions, Fixed Points, and Equilibria (Prentice Hall,
New York, 1981)
23. N.S. Glance, B.A. Huberman, The outbreak of cooperation. J. Math. Sociol. 17(4), 281–302
(1993)
24. B.D. Greenshield, A study of traffic capacity, in Proceedings of the Highway Research Board,
Vol. 14 (Highway Research Board, Washington, D.C., 1935), pp. 448–477
26. D. Helbing, Dynamic decision behavior and optimal guidance through information services:
Models and experiments, in Human Behaviour and Traffic Networks, ed. by M. Schreckenberg,
R. Selten (Springer, Berlin, 2004), pp. 47–95
27. D. Helbing, M. Schönhof, D. Kern, Volatile decision dynamics: Experiments, stochastic
description, intermittency control, and traffic optimization. New J. Phys. 4, 33.1–33.16 (2002)
28. J. Hofbauer, K. Sigmund, The Theory of Evolution and Dynamical Systems (Cambridge
29. T. Hogg, B.A. Huberman, Controlling chaos in distributed systems. IEEE Trans. Syst. Man
Cybern. 21(6), 1325–1333 (1991)
30. T.-Y. Hu, H.S. Mahmassani, Day-to-day evolution of network flows under real-time informa-
tion and reactive signal control. Transport. Res. C 5(1), 51–69 (1997)
31. S. Iwanaga, A. Namatame, The complexity of collective decision. Nonlinear Dynam. Psychol.
Life Sci. 6(2), 137–158 (2002)
32. J.H. Kagel, A.E. Roth (eds.), The Handbook of Experimental Economics (Princeton University,
Princeton, NJ, 1995)
33. J.O. Kephart, T. Hogg, B.A. Huberman, Dynamics of computational ecosystems. Phys. Rev. A
40(1), 404–421 (1989)
34. F. Klügl, A.L.C. Bazzan, Route decision behaviour in a commuting scenario: Simple heuristics
adaptation and effect of traffic forecast. JASSS 7(1) (2004)
35. Y.A. Korilis, A.A. Lazar, A. Orda Avoiding the Braess-paradox in non-cooperative networks.
J. Appl. Probab. 36, 211–222 (1999)
36. P. Laureti, P. Ruch, J. Wakeling, Y.-C. Zhang, The interactive minority game: a Web-based
investigation of human market interactions. Physica A 331, 651–659 (2004)
37. K. Lee, P.M. Hui, B.H. Wang, N.F. Johnson, Effects of announcing global information in a
two-route traffic flow model. J. Phys. Soc. Japan 70, 3507–3510 (2001)
38. T.S. Lo, H.Y. Chan, P.M. Hui, N.F. Johnson, Theory of networked minority games based on
strategy pattern dynamics. Phys. Rev. E 70, 056102 (2004)
39. T.S. Lo, P.M. Hui, N.F. Johnson, Theory of the evolutionary minority game. Phys. Rev. E 62,
4393–4396 (2000)
99(Suppl. 3), 7229–7236 (2002)
41. H.S. Mahmassani, R.C. Jou, Transferring insights into commuter behavior dynamics from
laboratory experiments to field surveys. Transport. Res. A 34, 243–260 (2000)
42. R. Mansilla, Algorithmic complexity in the minority game. Phys. Rev. E 62, 4553–4557 (2000)
43. M. Marsili, R. Mulet, F. Ricci-Tersenghi, R. Zecchina, Learning to coordinate in a complex
and nonstationary world. Phys. Rev. Lett. 87, 208701 (2001)
44. J.M. McNamara, Z. Barta, A.I. Houston, Variation in behaviour promotes cooperation in the
prisoner’s dilemma game. Nature 428, 745–748 (2004)
45. F. Michor, M.A. Nowak, The good, the bad and the lonely. Nature 419, 677–679 (2002)
46. M. Milinski, D. Semmann, H.-J. Krambeck, Reputation helps solve the ‘tragedy of the
commons’. Nature 415, 424–426 (2002)
47. D. Monderer, L.S. Shapley, Fictitious play property for games with identical interests.
J. Econ. Theor. 1, 258–265 (1996)
48. D. Monderer, L.S. Shapley, Potential games. Game. Econ. Behav. 14, 124–143 (1996)
49. M.A. Nowak, A. Sasaki, C. Taylor, D. Fudenberg, Emergence of cooperation and evolutionary
stability in finite populations. Nature 428, 646–650 (2004)
50. M. Novak, K. Sigmund, A strategy of win-stay, lose-shift that outperforms tit-for-tat in the
Prisoner’s Dilemma game. Nature 364, 56–58 (1993)
51. M.A. Nowak, K. Sigmund The alternating prisoner’s dilemma. J. theor. Biol. 168, 219–226
(1994)
References 237
52. M.A. Nowak, K. Sigmund, Evolution of indirect reciprocity by image scoring. Nature 393,
573–577 (1998)
53. A.C. Pigou, The Economics of Welfare (Macmillan, London, 1920)
54. M. Posch, Win-stay, lose-shift strategies for repeated games—Memory length, aspiration levels
and noise. J. Theor. Biol. 198, 183–195 (1999)
55. D.C. Queller, Kinship is relative. Nature 430, 975–976 (2004)
56. A. Rapoport, Exploiter, leader, hero, and martyr: the four archtypes of the 2 2 game. Behav.
Sci. 12, 81–84 (1967)
57. A. Rapoport, M. Guyer, A taxonomy of 2 2 games. Gen. Syst. 11, 203–214 (1966)
58. P.D.V.G. Reddy, et al., Design of an artificial simulator for analyzing route choice behavior in
the presence of information system. J. Math. Comp. Mod. 22, 119–147 (1995)
59. R.L. Riolo, M.D. Cohen, R. Axelrod, Evolution of cooperation without reciprocity. Nature 414,
441–443 (2001)
60. R.W. Rosenthal, A class of games possessing pure-strategy Nash equilibria. Int. J. Game Theor.
2, 65–67 (1973)
61. T. Roughgarden, E. Tardos, How bad is selfish routing? J. ACM 49(2), 236–259 (2002)
62. T.C. Schelling, Micromotives and Macrobehavior (WW Norton and Co, New York, 1978),
pp. 224–231C237
63. M. Schreckenberg, R. Selten (eds.), Human Behaviour and Traffic Networks (Springer, Berlin,
2004)
64. F. Schweitzer, L. Behera, H. Mühlenbein, Evolution of cooperation in a spatial prisoner’s
dilemma. Adv. Complex Syst. 5(2/3), 269–299 (2002)
65. R. Selten, et al., Experimental investigation of day-to-day route-choice behaviour and network
simulations of autobahn traffic in North Rhine-Westphalia, in Human Behaviour and Traffic
Networks, ed. by M. Schreckenberg, R. Selten (Springer, Berlin, 2004), pp. 1–21
66. R. Selten, M. Schreckenberg, T. Pitz, T. Chmura, S. Kube, Experiments and simulations
on day-to-day route choice-behaviour. See http://papers.ssrn.com/sol3/papers.cfm?abstract
id=393841
67. D. Semmann, H.-J. Krambeck, M. Milinski, Volunteering leads to rock-paper-scissors
dynamics in a public goods game. Nature 425, 390–393 (2003)
68. P. Spirakis, Algorithmic aspects of congestion games, Invited talk at the 11th Colloquium
on Structural Information and Communication Complexity (Smolenice Castle, Slovakia, June
21-23, 2004)
Rev. Lett. 89, 118101 (2002)
70. J. Wahle, A.L.C. Bazzan, F. Klügl, M. Schreckenberg, Decision dynamics in a traffic scenario.
Physica A 287, 669–681 (2000)
71. J.G. Wardrop, Some theoretical aspects of road traffic research, in Proceedings of the Institution
of Civil Engineers II, Vol. 1 (1952), pp. 325–378
72. D.H. Wolpert, K. Tumer, Collective intelligence, data routing and Braess’ paradox. J. Artif.
Intell. Res. 16, 359–387 (2002)
73. T. Yamashita, K. Izumi, K. Kurumatani, Effect of using route information sharing to reduce
traffic congestion. Lect. Notes Comput. Sci. 3012, 86–104 (2004)
74. B. Yuan, K. Chen, Evolutionary dynamics and the phase structure of the minority game. Phys.
Rev. E 69, 067106 (2004)
Chapter 13
Response to Information
Optimal route guidance strategies in overloaded traffic networks, for example,

require reliable traffic forecasts (see Fig. 13.1). These are extremely difficult for
two reasons: First of all, traffic dynamics is very complex, but after more than
50 years of research, it is relatively well understood [23]. The second and more
serious problem is the invalidation of forecasts by the driver reactions to route
choice recommendations. Nevertheless, some keen scientists hope to solve this long-
standing problem by means of an iteration scheme [1, 2, 4–6, 8, 36, 42, 45]: If the
driver reaction was known from experiments [10, 11, 15, 20, 28–30, 32, 35, 37, 43],
the resulting traffic situation could be calculated, yielding improved route choice
recommendations, etc. Given this iteration scheme converges, it would facilitate
optimal recommendations and reliable traffic forecasts anticipating the driver
reactions. Based on empirically determined transition and compliance probabilities,
the new procedure developed in the following would even allow us to reach the
optimal traffic distribution in one single step and in harmony with the forecast. Let
us now quantify the success or payoff Pi of road users in terms of their inverse
travel times. If one approximates the average vehicle speed Vi on route i by the
linear relationship
ni .t/
Vi .ni / D Vi0 1 max ; (13.1)
ni
the inverse travel times obey the payoff relations Pi .ni / D Pi0 Pi1 ni with
Vi0 Vi0
Pi0 D and Pi1 D : (13.2)
Li ni Li
max

This chapter reprints parts of a previous publication with kind permission of the copyright owner,
Springer Publishers. It is requested to cite this work as follows: D. Helbing, Dynamic decision
behavior and optimal guidance through information services: Models and experiments. Pages
47–95 in: M. Schreckenberg and R. Selten (eds.) Human Behaviour and Traffic Networks (Springer,
Berlin, 2004).

240 13 Response to Information
Fig. 13.1 Schematic illustration of the day-to-day route choice scenario (from [21]). Each day,
the drivers have to decide between two alternative routes, 1 and 2. Note that, due to the different
number of lanes, route 1 has a higher capacity than route 2. The latter is, therefore, used by less cars
Herein, Vi0 denotes the maximum velocity (speed limit), ni the number of drivers
on route i , Li its length, and nmax
i its capacity, i.e. the maximum possible number
of vehicles on route i . For an improved approach to determine the travel times in
road networks see [22]. Note that alternative routes can reach comparable payoffs
(inverse travel times) only when the total number N.t/ of vehicles is large enough
to fulfil the relations P1 .N.t// < P2 .0/ D P20 and P2 .N.t// < P1 .0/ D P10 . Our
route choice experiment will address this traffic regime. Furthermore, we have the
capacity restriction N.t/ < nmax
1 C nmax
2 . N.t/ D n1
max
C nmax
2 would correspond to
a complete gridlock.
13.1 Experimental Setup and Previous Results
To determine the route choice behavior, Schreckenberg, Selten et al. [43] have
recently carried out a decision experiment (see Fig. 13.2). N test persons had to
repeatedly decide between two alternatives 1 and 2 (the routes) and should try to
maximize their resulting payoffs (describing something like the speeds or inverse
travel times). To reflect the competition for a limited resource (the road capacity),
the received payoffs
P1 .n1 / D P10 P11 n1 and P2 .n2 / D P20 P21 n2 (13.3)
went down with the numbers of test persons n1 and n2 D N n1 deciding for
alternatives 1 and 2, respectively. The user equilibrium corresponding to equal
payoffs for both alternative decisions is found for a fraction
13.1 Experimental Setup and Previous Results 241
Fig. 13.2 Schematic illustration of the decision experiment (from [21]). Several test persons
take decisions based on the aggregate information their computer displays. The computers are
connected and can, therefore, exchange information. However, a direct communication among
players is suppressed
n1 P1 1 P10 P20
D 1 2 1C
eq
f1 D (13.4)
N P1 C P2 N P11 C P21
of persons choosing alternative 1. The system optimum corresponds to the maximum

of the total payoff n1 P1 .n1 / C n2 P2 .n2 /, which lies by an amount of
1 P10 P20
(13.5)
2N P11 C P21
below the user optimum. Therefore, only experiments with a few players allow to
find out, whether the test persons adapt to the user or the system optimum. Small
groups are also more suitable for the experimental investigation of the fluctuations
in the system and of the long-term adaptation behavior. Schreckenberg, Selten
et al. found that, on average, the test groups adapted relatively well to the user
equilibrium. However, although it appears reasonable to stick to the same decision
once the equilibrium is reached, the standard deviation stayed at a finite level.
This was not only observed in treatment 1, where all players knew only their own
(previously experienced) payoff, but also in treatment 2, where the payoffs P1 .n1 /
and P2 .n2 / for both, 1- and 2-decisions, were transmitted to all players (analogous
to radio news). Nevertheless, treatment 2 could decrease the changing rate and
increase the average payoffs (cf. Fig. 13.3). For details regarding the statistical
analysis see [43].
To explain the mysterious persistence in the changing behavior and explore
possibilities to suppress it, we have repeated these experiments with more iterations
and tested additional treatments. In the beginning, all treatments were consecutively
applied to the same players in order to determine the response to different kinds
of information (see Fig. 13.3). Afterwards, single treatments and variants of them
have been repeatedly tested with different players to check our conclusions. Apart
from this, we have generalized the experimental setup in the sense that it was not
anymore restricted to route choice decisions: The test persons did not have any idea
of the payoff functions in the beginning, but had to develop their own hypothesis
about them. In particular, the players did not know that the payoff decreased with
the number of persons deciding for the same alternative.
In treatment 3, every test person was informed about the own payoff P1 .n1 / [or
P2 .n2 /] and the potential payoff
P2 .N n1 C N / D P2 .n2 / NP21 (13.6)
[or P1 .N n2 C N / D P1 .n1 / NP11 ] he or she would have obtained, if a

fraction of persons had additionally chosen the other alternative (here: D 1=N ).
Treatments 4 and 5 were variants of treatment 3, but some payoff parameters were
changed in time to simulate varying environmental conditions. In treatment 5, each
player additionally received an individual recommendation which alternative to
choose.
The higher changing rate in treatment 1 compared to treatment 2 can be
understood as effect of an exploration rate 1 required to find out which alternative
performs better. It is also plausible that treatment 3 could further reduce the
changing rate: In the user equilibrium with P1 .n1 / D P2 .n2 /, every player knew that
he or she would not get the same, but a reduced payoff, if he or she would change
the decision. That explains why the new treatment 3 could reach a great adaptation
performance, reflected by a very low standard deviation and almost optimal average
payoffs. The behavioral changes induced by the treatments were not only observed
on average, but for all single individuals (see Fig. 13.4). Moreover, even the smallest
individual cumulative payoff exceeded the highest one in treatment 1. Therefore,
treatment 3’s way of information presentation is much superior to the ones used
today.
13.2 Is It Just an Unstable User Equilibrium?
In this section, we will investigate why players changed their decision in the user
equilibrium at all. With P1 .1; t/ D hn1 .t/i=N and hni .t/i D ni .t/ (as ni .t/ are the
measured numbers of i -decisions at time t), we find the following balance equation
for the decision experiment:
13.2 Is It Just an Unstable User Equilibrium? 243
a Treatment 1 Treatment 2 Treatment 3 Treatment 4 Treatment 5

7
No. of 1-Decisions
6
5
4
3
0 500 1000 1500 2000 2500
b
Standard Deviation
0
0 500 1000 1500 2000 2500
c
4
Changing Rate
3
2
1
0
0 500 1000 1500 2000 2500
d
10
Average Payoffs
8 All Players
6 Change of Decision
4
2
0
-2
-4
0 500 1000 1500 2000 2500
Iteration t
Fig. 13.3 Overview of treatments 1 to 5 [21] (with N D 9 and payoff parameters P20 D 28,
P11 D 4, P21 D 6, and P10 D 34 for 0 t 1; 500, but a zick-zack-like variation between
P10 D 44 and P10 D 6 with a period of 50 for 1; 501 t 2; 500): (a) Average number of
decisions for alternative 1 (solid line) compared to the user equilibrium (broken line), (b) standard
deviation of the number of 1-decisions from the user equilibrium, (c) number of decision changes
from one iteration to the next one, (d) average payoff per iteration for players who have changed
their decision and for all players. The latter increased with a reduction in the changing rate, but
normally stayed below the payoff in the user equilibrium (which is 1 on average in treatments 4
and 5, otherwise 10). The displayed moving time-averages [(a) over 40 iterations, (b)–(d) over
100 iterations] illustrate the systematic response to changes in the treatment every 500 iterations.
Dashed lines in (b)–(d) show estimates of the stationary values after the transient period (to guide
the eyes), while time periods around the dotted lines are not significant. Compared to treatment
1, treatment 3 managed to reduce the changing rate and to increase the average payoffs (three
times more than treatment 2 did). These changes were systematic for all players (see Fig. 13.4).
In treatment 4, the changing rate and the standard deviation went up, since the user equilibrium
changed in time. The user-specific recommendations in treatment 5 could almost fully compensate
for this. The above conclusions are also supported by additional experiments with single treatments
Fig. 13.4 Comparison of the a

individual decision behaviors 9
in (a) treatment 1, (b)
8
treatment 2, and (c) treatment
3 (from [21]). The upper 7
values correspond to a 6
Player
decision for alternative 2, the 5
lower ones for alternative 1.
4
Note that some test persons
showed similar behaviors 3
(either more or less the same 2
or almost opposite ones), 1
although they could not talk 0 100 200 300 400 500
to each other. This shows that
Iteration t
there are some typical
strategies how to react to a b
certain information 9
configuration, i.e. to a certain 8
decision distribution. The
7
group has, in fact, to develop
complementary strategies in 6
Player
order to reach a good 5

adaptation performance. 4
Identical strategies would
3
perform poorly (as in the
minority game [12–14]). 2
Despite the mentioned 1
complementary behavior, 500 600 700 800 900 1000
there is a characteristic Iterationt
reaction to changes in the c
treatment. For example, 9
compared to treatment 2 all
8
players reduce their changing
rate in treatment 3 7
6
Player
5
4
3
2
1
1000 1100 1200 1300 1400 1500
Iteration t
hn1 .t C 1/i n1 .t/ D p.1j2; n1 I t/n2 .t/ p.2j1; n1 I t/n1 .t/: (13.7)
Assuming stationary transition probabilities p.2j1; n1 / (after a transient phase), the

equilibrium distribution corresponds to
hn1 .t C 1/i D hn1 .t/i D n1 .t/: (13.8)

13.2 Is It Just an Unstable User Equilibrium? 245
Consequently, the equilibrium condition
p.2j1; n1 /n1 .t/ D p.1j2; n1 /n2 .t/ (13.9)

eq eq
should be fulfilled for the user equilibrium n1 .t/ D f1 N and n2 .t/ D .1 f1 /N .
This, however, is generally not compatible with the assumption
p.2j1; n1 / / expŒP2 .N n1 C 1/ P1 .n1 / (13.10)
or similar specifications of the transition probability that increase monotonically

with the payoff P2 or the payoff difference P2 P1 ! Since normally
eq eq
p.2j1; f1 N / 1 f1
eq ¤ eq ; (13.11)
p.1j2; f1 N / f1
the test persons would have serious problems reaching the user equilibrium. The
decision distribution would possibly tend to oscillate around it, corresponding to
an unstable user equilibrium. We have tested this potential interpretation of the on-
going tendency to change the decision. Figure 13.5 compares the changing rates
and the standard deviations for a case where the equilibrium condition (13.9) should
be valid and another case where it should be violated. However, the changing rate
and standard deviation were higher in the first case, so that the hypothesis of an
eq
unstable equilibrium must be wrong. In the user equilibrium with n1 .t/ D f1 N D
N n2 .t/, the inflow p.1j2; n1 /n2 .t/ is, in fact, well balanced by the outflow
a b
10 10
9 No. of 1-Decisions 9
8 Standard Deviation
Changing Rate 8
Number of Players
Number of Players
7 7
6 No. of 1-Decisions
6
Standard Deviation
5 5 Changing Rate
4 4
3 3
2 2
1 1
0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Iteration Iteration
Fig. 13.5 Comparison of the group performance for treatment 2, when the user equilibrium
corresponds to (a) f1 D 50% 1-decisions (for P10 D 32, P11 D 5, P20 D 32, P21 D 5)
eq
or (b) f1 D 80% 1-decisions (for P10 D 42, P11 D 4, P20 D 22, P21 D 6). If the user
eq
eq
equilibrium were unstable for f1 ¤ 1=2, the changing rate and standard deviation should be
lower in (a) than in (b). The observation contradicts this assumption. The persistent changing rate
is also not caused by a difference between the system and the user optimum, since this is zero in
eq
(a) but one in (b). Instead, the higher changing rate for the symmetrical case f1 D 1=2 is for
statistical reasons (Remember that the variance of a binomial distribution B.N; p/ is Np.1 p/
and becomes maximal for p D 1=2)
a b
4 1
Treatment 1 Treatment 1
3 Treatment 2 Treatment 2
Treatment 3 Treatment 3
0.8
2
Occurence Probability
Netto Flux
1
0.6
0
-1 0.4
-2
0.2
-3
-4 0
4 5 6 7 8 3 4 5 6 7 8 9
Number n1 of Previous 1-Decisions Number n1 of Previous 1-Decisions
Fig. 13.6 (a) The netto flux p.1j2; n1 /n2 .t / p.1j2; n1 /n1 .t /, which reflects the systematic part
of decision changes, does not significantly depend on the treatment. As expected, it is zero in
the user equilibrium, positive below it, and negative above it. (b) The treatment can influence the
occurrence probability of decision distributions. Compared to treatments 1 and 2, the probability
distribution is much more sharply peaked for treatment 3, implying a significantly smaller level of
randomness during decision changes (a smaller “diffusion coefficient”)
p.1j2; n1 /n1 .t/, as Fig. 13.6 shows. By the way, the results displayed in Fig. 13.5
also disprove the idea that a difference between the user and the system optimum
may be the reason for the continuing changing behavior.
13.3 Explaining the Volatile Decision Dynamics
The reason for the pertaining changing behavior can be revealed by a more detailed
analysis of the individual decisions in treatment 3. Figure 13.7 shows some kind of
intermittent behavior, i.e. quiescent periods without changes, followed by turbulent
periods with many changes. This is reminiscent of volatility clustering in stock
market indices [19, 33, 38], where individuals also react to aggregate information
reflecting all decisions (the trading transactions). Single players seem to change their
decision to reach above-average payoffs. In fact, although the cumulative individual
payoff is anticorrelated with the average changing rate, some players receive higher
payoffs with larger changing rates than others. They profit from the overreaction in
the system. Once the system is out of equilibrium, all players respond in one way
or another. Typically, there are too many decision changes (see Figs. 13.7 and 13.9).
The corresponding overcompensation, which had also been predicted by computer
simulations [2, 5, 6, 20, 36, 46], gives rise to “turbulent” periods.
Finally, we note that the calm periods without decision changes tend to become
longer in the course of time. That is, after a very long time period the individuals
seem to learn not to change their behavior when the user equilibrium is reached.
This is not only found in Fig. 13.7, but also visible in Fig. 13.3c after about 800
13.4 Simulation of Reinforcement Learning and Emergence 247
a 1
3
2
4
Ranking
5
6
8
9
7
0 100 200 300 400 500
Iteration t
b 10 No. of 1-Decisions
Standard Deviation
8 Changing Rate
Number of Players
0
0 100 200 300 400 500
Iteration t
Fig. 13.7 Illustration of typical results for treatment 3 [21] (which was here the only treatment
applied to the test persons, in contrast to Fig. 13.3). (a) Decisions of all nine players. Players
are displayed from the top to the bottom in the order of increasing changing rate. Although the
ranking of the cumulative payoff and the changing rate are anticorrelated, the relation is not
monotonic. Note that turbulent or volatile periods characterized by many decision changes are
usually triggered by individual changes after quiescent periods (dotted lines). (b) The changing
eq
rate is mostly larger than the (standard) deviation from the user equilibrium n1 D f1 N D 6,
indicating an overreaction in the system
iterations. In larger systems (with more test persons) this transient period would take
even longer, so that this stabilization effect could not be observed by Schreckenberg,
Selten et al. [43].
13.4 Simulation of Reinforcement Learning and Emergence

of Individual Response Patterns
A close look at Fig. 13.8a reveals additional details of decision behavior:

• Some players change their decision more frequently than others.
• Some test persons show similar behaviors (e.g., players 8 and 9 or 1 and 7 for
t 400), while some display almost opposite behaviors (e.g., players 7 and 8).
Fig. 13.8 (a) Typical a 9

individual decision changes
of nine test persons exposed 8
to treatment 2 with the
7
parameters specified in
Fig. 13.3. (b) Simulation of 6
Player
decision changes based on a
5
model of reinforment learning
(see main text) with 4
parameter values ı D 0:01,
3
q0 D 0:4, and r D 0:995
2
1
0 100 200 300 400 500
Iteration t
b 9
8
7
6
Player
5
4
3
2
1
0 100 200 300 400 500
Iteration t
The second point is very surprising, as the players could not communicate with
each other. However, both observations can be explained by the conjecture that
the individuals develop different characteristic strategies how to react to specific
information. “Movers” and “stayers” or direct and contrary strategies have, in
fact, been observed by Schreckenberg, Selten et al. [43], and it is an interesting
question, how they arise. The group has to develop complementary strategies in
order to reach a good adaptation performance. As a consequence, if some players
do not react to changing conditions, others will take the chance to earn additional
payoff. This experimentally supports the behavior assumed in the theory of efficient
markets. Note that identical strategies would perform poorly, as in the minority game
[3, 12–14].
In order to reproduce the above described evolution of complementary strategies
and other observed features, we have developed a simulation model based on
reinforcement learning [18, 34]. At first, it appears reasonable to apply a learning
strategy that reproduces the “law of relative effect”, according to which the
probability p˛0 .i; t C 1/ of an individual ˛ to choose alternative i at time t C 1
3
Overreaction
0
0 500 1000 1500 2000 2500
Iteration t
Fig. 13.9 Measured overreaction, i.e., difference between the actual number of decision changes
(the changing rate) and the required one (the standard deviation) [21]. The overreaction can be
significantly influenced by the treatment, i.e. the way of information presentation. The minimum
overreaction was reached by treatment 5, i.e. user-specific recommendations
would reflect the relative frequency with which this alternative was successful in
the past. This can, for example, be reached by means of the reinforcement rule

1 q0 Œ1 p˛0 .i; t/ in case of a successful decision
p˛0 .i; t C 1/ D (13.12)
q0 p˛0 .i; t/ otherwise
[34], where the way in which a successful decision is defined may vary from one
situation to another. However, a probabilistic decision strategy, when applied by
all individuals, produces a large amount of stochastic fluctuations, i.e. the user
equilibrium is hard to maintain. More importantly, although the above learning
strategy may explain a specialization in the individual behaviors (i.e. different
decision probabilities, depending on the respective success history), it does not
allow to understand the state-dependent probability of decision changes (see
Fig. 13.12). We will, therefore, develop a model for the conditional (transition)
probability p˛ .i ji 0 ; n1 I t/ of individual ˛ to select alternative i , given that the
previous decision was i 0 and n1 individuals had taken decision 1. Furthermore, let
us assume that each individual updates this transition probability according to the
following scheme:
8
< maxŒ1 ı; p˛ .i ji 0 ; n1 I t/ C q.t C 1/ for a successful
0
p˛ .i ji ; n1 I t C 1/ D decision,
: 0
minŒı; p˛ .i ji ; n1 I t/ q.t C 1/ otherwise.
(13.13)
Due to the normalization of transition probabilities, we have the additional relation
p˛ .3 i ji 0 ; n1 I t C 1/ D 1 p˛ .i ji 0 ; n1 I t C 1/; (13.14)
as 3 i is the alternative of decision i 2 f1; 2g. The parameter ı 0 reflects

a minimum changing probability, which ensures that there is always a certain
readiness to adapt to a potentially changing environment. It is responsible for the
stochastic termination of quiescent phases, in which nobody changes the decision.
Our simulations were run with ı D 0:01, i.e. the minimum changing probability
was assumed to be 1%.
The parameter q.t/ denotes the size of the adaptation step, by which the
transition probability is increased in case of success or otherwise decreased, while
the minimum and maximum functions guarantee that the transition probabilities
p˛ .i ji 0 ; n1 I t C 1/ stay between the minimum value ı and the maximum value 1 ı.
A time-dependent choice such as
q.t/ D q0 r t (13.15)
with an initial value q0 of q (0 < q0 < 1) and a value of r slightly smaller

than 1 allow one to describe that the learning rate is large in the beginning,
when the different possible strategies are explored, but it eventually goes down,
as the optimum strategy becomes more and more clear. In the course of time, this
leads to the stabilization of a particular, history-dependent response pattern that
characterizes the individual decision strategy. The resulting response pattern shows
either a high likelihood to stay with the previous decision (with p˛ ı) or a high
likelihood to change it (with p˛ 1 ı), depending on the respective system state
n1 and previous decision i 0 . That is, the resulting strategy tends to be approximately
deterministic, reflecting that the individual believes to know what is the “right”
decision. This is markedly different from other decision models with reinforcement
learning [18, 34]. Nevertheless, when averaging over all occurring system states,
the individuals appear to play mixed strategies, i.e. they seem to show probabilistic
(rather than almost deterministic) decision behavior (see Fig. 13.12). Therefore, our
approach is expected to be consistent with the law of relative effect, but only in the
statistical sense. Altogether, formula (13.15) reflects the observed trial-and-error
behavior in the beginning (the “experimentation phase”), but a tendency to follow
learned strategies later on without significant changes. The parameters ı, q0 , and
r may, of course, be individual, but for reasons of simplicity we have assumed
identical values in our simulations.
The way, in which a successful decision is defined, may depend on the respective
situation or experiment. In our simulations of treatment 2, we have assumed that the
decision is valued as successful, when
Pi .ni .t C 1// P3i .N ni .t C 1// D P3i .n3i .t C 1// (13.16)
and
Pi .ni .t C 1// Pi 0 .ni 0 .t//; (13.17)
i.e. when the payoff was at least as large as for the other alternative 3 i and
not smaller than in the previous time step. The first decision was made randomly
with probability 1/2. The following decisions were also randomly chosen, but
in accordance with the respective transition probabilities, which were updated
according to the above scheme.
The simulation results are in good qualitative agreement with the features
observed in our experiments. We find an adaptation of the group to the user
equilibrium with an average individual payoff of approximately 8.5, as in our
experiments. Moreover, the changing rate is high in the beginning and decreases
in the course of time (see Fig. 13.8b). As experimentally observed, some players
change their decision more frequently than others, and we find almost similar or
opposite behaviors after some time. That is, our simulations allow to reproduce that
players develop individual strategies (i.e. response patterns, “roles”, or “characters”)
in favour of a good group performance.
By means of our simulations, we can not only reproduce the main experimental
observations. One can also optimize the group sizes and number of iterations of
decision experiments. The above simulation concept is now used to design new
experiments, which try to improve the system performance or even to establish the
social optimum by particular information strategies. In the following section, we
will, for example, introduce a possible concept for decision guidance.
13.4.1 Potentials and Limitations of “Decision Control”
To avoid overreaction, in treatment 5 we have recommended a number

eq
f1 .t C 1/N n1 .t/ of players to change their decision and the other ones to keep
it. These user-specific recommendations helped the players to reach the smallest
overreaction of all treatments (see Fig. 13.9) and a very low standard deviation,
although the payoffs were changing in time (see Fig. 13.10). Treatment 4 shows how
the group performance was affected by the time-dependent user equilibrium: Even
without recommendations, the group managed to adapt to the changing conditions
surprisingly well, but the standard deviation and changing rate were approximately
as high as in treatment 2 (see Fig. 13.3). This adaptability (the collective “group
intelligence”) is based on complementary responses (direct and contrary ones [43],
“movers” and “stayers”, cf. Fig. 13.4). That is, if some players do not react to the
changing conditions, others will take the chance to earn additional payoff. This
experimentally supports the behavior assumed in the theory of efficient markets, but
here the efficiency is limited by overreaction.
In most experiments, we found a constant and high compliance CS .t/ 0:92
with recommendations to stay, but the compliance CM .t/ with recommendations to
change (to “move”) [15, 31, 32, 44] turned out to vary in time. It decreased with the
reliability of the recommendations (see Fig. 13.11a), which again dropped with the
compliance.
Based on this knowledge, we have developed a model, how the competition for
limited resources (such as road capacity) could be optimally guided by means of
information services. Let us assume we had n1 .t/ 1-decisions at time t, but the
eq
optimal number of 1-decision at time t C 1 is calculated to be f1 .t C 1/N n1 .t/.
eq
Our aim is to balance the deviation f1 .t C 1/N n1 .t/ 0 by the expected net
number
a 7
Treatment 4 User Equilibrium No. of 1-Decisions
6
Number of Players
5
4
3
2
1
Changing Rate Standard Deviation
0
0 100 200 300 400 500
Iteration t
b 7
Treatment 5 User Equilibrium No. of 1-Decisions
6
Number of Players
5
4
3
Changing Rate Standard Deviation
2
1
0
0 100 200 300 400 500
Iteration t
Fig. 13.10 Representative examples for (a) treatment 4 and (b) treatment 5 (from [21]). The
displayed curves are moving time-averages over 20 iterations. Compared to treatment 4, the
user-specific recommendations in treatment 5 (assuming CM D CS D 1, R1 D 0, R2 D
eq
max.Œf1 .t C 1/N n1 .t / C B.t C 1/=n2 .t /; 1/, I1 D I2 D 1) could increase the group
adaptability to the user equilibrium a lot, even if they had a systematic or random bias B (see
Fig. 13.11a). The standard deviation was reduced considerably and the changing rate even more
hn1 .t C 1/i D hn1 .t C 1/ n1 .t/i D hn1 .t C 1/i n1 .t/ (13.18)

eq
of transitions from decision 2 to decision 1, i.e. f1 .t C1/N n1 .t/ D hn1 .t C1/i.
eq
In the case f1 .t C 1/N n1 .t/ < 0, indices 1 and 2 have to be interchanged.
Let us assume we give recommendations to fractions I1 .t/ and I2 .t/ of players
who had chosen decision 1 and 2, respectively. The fraction of changing recom-
mendations to previous 1-choosers shall be denoted by R1 .t/, and for previous
2-choosers by R2 .t/. Correspondingly, fractions of Œ1 R1 .t/ and Œ1 R2 .t/
receive a recommendation to stick to the previous decision. Moreover, Œ1 CM .t/
is the refusal probability of recommendations to change, while Œ1 CS .t/ is the
refusal probability of recommendations to stay. Finally, we denote the spontaneous
transition probability from decision 1 to 2 by pa .2j1; n1 I t/ and the inverse transition
probability by pa .1j2; n1 I t/, in case a player does not receive any recommendation.
a 1 b 20
Change
16 No Change
0.8 12 Follow Recommendation
Refuse Recommendation
Compliance Rate
Average Payoff
Recommendations to Stay
0.6 Recommendations to Change 4
All Recommendations
0
0.4 -4
-8
0.2 -12
-16 Follow Rec. Change
A B C D E Refuse Rec. Change
0 -20
0 100 200 300 400 500 0 50 100 150 200
Fig. 13.11 (a) In treatment 5, the compliance to recommendations to change dropped con-
siderably below the compliance to recommendations to stay. The compliance to changing
recommendations was very sensitive to the degree of their reliability, i.e. participants followed
recommendations just as much as they helped them to reach the user equilibrium (so that the
bias B did not affect the small deviation from it, see Fig. 13.10b). While during time interval
A, the recommendations would have been perfect, if all players had followed them, in time
interval B the user equilibrium was overestimated by B D C1, in C it was underestimated by
B D 2, in D it was randomly over- or underestimated by B D ˙1, and in E by B D ˙2.
Obviously, a random error is more serious than a systematic one of the same amplitude. Dotted non-
vertical lines illustrate the estimated compliance levels during the transient periods and afterwards
(horizontal dotted lines). (b) The average payoffs varied largely with the decision behavior. Players
who changed their decision got significantly lower payoffs on average than those who kept their
previous decision. Even recommendations could not overcome this difference: It stayed profitable
not to change, although it was generally better to follow recommendations than to refuse them. For
illustrative reasons, the third and fourth line were shifted by 15, while the fifth and sixth line were
shifted by 30 iterations (From [21])
This happens with probabilities Œ1 I1 .t/ and Œ1 I2 .t/, respectively. Both

transition probabilities pa .2j1; n1 I t/ and pa .1j2; n1 I t/ are functions of the number
n1 .t/ D N n2 .t/ of previous 1-decisions. The index a allows us to reflect different
strategies or characters of players. The fraction of players pursuing strategy a is then
denoted by Fa .t/. Applying methods summarized in [24, 25], the expected change
hn1 .t C 1/i of n1 is given by the balance equation
X
hn1 .t C 1/i D pa .1j2; n1 I t/Fa .t/Œ1 I2 .t/n2 .t/
a
X
pa .2j1; n1 I t/Fa .t/Œ1 I1 .t/n1 .t/
a
X
a
C fCM .t/R2 .t/ C Œ1 CSa .t/Œ1 R2 .t/gFa .t/I2 .t/n2 .t/
a
X
a
fCM .t/R1 .t/ C Œ1 CSa .t/Œ1 R1 .t/gFa .t/I1 .t/n1 .t/;
a
(13.19)
eq
which should agree with f1 .t C 1/N n1 .t/. We have evaluated the overall
transition probabilities
X
p.1j2; n1 I t/ D pa .1j2; n1 I t/Fa .t/ and
a
X
p.2j1; n1 I t/ D pa .2j1; n1 I t/Fa .t/: (13.20)
a
According to classical decision theories [7, 9, 24, 25, 41], we would expect that
the transition probabilities pa .2j1; n1 I t/ and p.2j1; n1 I t/ should be monotonically
increasing functions of the payoff P2 .N n1 .t//, the payoff difference P2 .N
n1 .t// P1 .n1 .t//, the potential payoff P2 .N n1 .t/ C N /, or the potential
payoff gain P2 .N n1 .t/ C N / P1 .n1 .t//. All these quantities vary linearly
with n1 , so that p.2j1; n1 I t/ should be a monotonic function of n1 .t/. A similar
thing should apply to p.1j2; n1 I t/. Instead, the experimental data point to transition
probabilities with a minimum at the user equilibrium (see Fig. 13.12a). That is, the
players stick to a certain alternative for a longer time, when the system is close
to the user equilibrium. This is a result of learning [16, 17, 26, 27, 39, 40]. In fact,
we find a gradual change of the transition probabilities in time (see Fig. 13.12b).
The corresponding “learning curves” reflect the players’ adaptation to the user
equilibrium. After the experimental determination of the transition probabilities
p.2j1; n1 I t/, p.1j2; n1 I t/ and specification of the overall compliance probabilities
X X
a
CM .t/ D CM .t/Fa .t/; CS .t/ D CSa .t/Fa .t/; (13.21)
a a
we can guide the decision behavior in the system via the levels Ii .t/ of information
dissemination and the fractions Ri .t/ of recommendations to change (i 2 f1; 2g).
These four degrees of freedom allow us to apply a variety of guidance strategies
depending on the respective information medium. For example, a guidance by
radio news is limited by the fact that I1 .t/ D I2 .t/ is given by the average
percentage of radio users. Therefore, (13.19) cannot always be solved by variation
of the fractions of changing recommendations Ri .t/. User-specific services have
much higher guidance potentials and could, for example, be transmitted via SMS.
Among the different guidance strategies fulfilling (13.19), the one with the minimal
statistical variance will be the best. However, it would already improve the present
situation to inform everyone about the fractions Ri .t/ of participants who should
change their decision, as users can learn to respond with varying frequencies (see
Fig. 13.12). Some actually respond more sensitively than others (see Fig. 13.4), so
that a group of users can reach a good overall performance based on individual
strategies.
The outlined guidance strategy could, of course, also be applied to reach the
system optimum rather than the user optimum. The values of n1 .t C 1/ would just
be different. Note, however, that the users would soon recognize that this guidance is
a 1 b 0.4
Transition Probability p(1|2,n1) p(2|1,n1=5;t)
Transition Probability p(2|1,n1) p(2|1,n1=6;t)
0.8 Decision Probability P(1|n1) p(2|1,n1=7;t)
Decision Probability P(2|n1) 0.3
Probability
0.6
0.2
0.4
0.1
0.2
0 0
4 5 6 7 8 0 100 200 300 400 500
Number n1 of Previous 1-Decisions Iteration t
Fig. 13.12 Illustration of group-averaged decision distributions P .i jn1 / and transition probabil-
ities p.i 0 ji; n1 I t / measured in treatment 3 (from [21]). (a) The probability P .1jn1 / to choose
alternative 1 was approximately 2/3, independently of the number n1 of players who had previously
chosen alternative 1. The probability P .2jn1 / to choose alternative 2, given that n1 players had
chosen alternative 1, was always about 1/3. In contrast, the group-averaged transition probability
p.1j2; n1 / describing decision changes from alternative 2 to 1 did depend on the number n1
of players who had chosen decision 1. The same was true for the inverse transition probability
p.2j1; n1 / from decision 1 to decision 2. Remarkably enough, these transition probabilities are not
monotonically increasing with the payoff or the expected payoff gain, as they do not monotonically
increase with n1 . Instead, the probability to change the decision shows a minimum at the user
eq
equilibrium n1 D f1 N D 6. Figures 13.4 and 13.8 suggest that this transition probability does
not reflect the individual transition probabilities. There rather seem to be typical response patterns
(see Sect. 13.4), i.e. some individuals react only to large deviations from the user equilibrium,
while others already react to small ones, so that the overall response of the group reaches a good
adaptation performance. (b) The reason for the different transition probabilities is an adaptation
process in which the participants learn to take fewer changing decisions, when the user equilibrium
is reached or close by, but more, when the user equilibrium is far away (The curves were
exponentially smoothed with ˛ D 0:05)
not suitable to reach the user optimum. Consequently, the compliance probabilities
would gradually go down, which would affect the potentials and reliability of the
guidance system.
In practical applications, we would determine the compliance probabilities Cj .t/
with j 2 fM; S g (and the transition probabilities) on-line with an exponential
smoothing procedure according to
Cj .t C 1/ D ˛Cj0 .t/ C .1 ˛/Cj .t/ with ˛ 0:1; (13.22)
where Cj0 .t/ is the percentage of participants who have followed their recommenda-
tion at time t. As the average payoff for decision changes is normally lower than for
staying with the previous decision (see Figs. 13.11 and 13.3d), a high compliance
probability CM is hard to achieve. That is, individuals who follow recommendations
to change normally pay for reaching the user equilibrium (because of the overreac-
tion in the system). Hence, there are no good preconditions to charge the players
for recommendations, as we did in another treatment. Consequently, only a few
players requested recommendations, which reduced their reliability, so that the
overall performance of the system went down.
13.4.2 Master Equation Description of Iterated Decisions
The description of decisions that are taken at discrete time steps (e.g. on a day-to-
day basis) is different from decisions in continuous time. We can, however, apply
the time-discrete master equation [24] with t D 1, if there is no need to distinguish
several characters a. As the number of individuals changing to the other alternative
is given by a binomial distribution, we obtain the following expression for the
configurational transition probability:
ˇ
P .n1 ; n2 /; t C 1 ˇ .n1 n1 ; n2 C n1 /; t
min.n1 n1 ;n2
/
X n2 C n1
D p.1j2; n1 n1 I t/n1 Ck
n1 C k
kD0

n k n1 n1
1 p.1j2; n1 n1 I t/ 2 p.2j1; n1 n1 I t/k
k
n n1 k
1 p.2j1; n1 n1 I t/ 1 : (13.23)
This formula sums up the probabilities that n1 Ck of n2 Cn1 previous 2-choosers
change independently to alternative 1 with probability p.1j2; n1 n1 I t/, while
k of the n1 n1 previous 1-choosers change to alternative 2 with probability
p.2j1; n1 n1 I t/, so that the net number of changes is n1 . If n1 < 0, the
roles of alternatives 1 and 2 have to be interchanged. Only in the limits p.1j2; n1
n1 I t/ 0 and p.2j1; n1 n1 I t/ 0 corresponding to t 0 do we get the
approximation
ˇ
P .n1 ; n2 /; t C 1 ˇ .n1 n1 ; n2 C n1 /; t
8
< p.1j2; n1 1I t/.n2 C 1/ if n1 D C1
p.2j1; n1 C 1I t/.n1 C 1/ if n1 D 1 (13.24)
:
0 otherwise,
which is relevant for the time-continuous master equation.

The potential use of (13.23) is the calculation of the statistical variation of the
decision distribution or, equivalently, the number n1 of 1-choosers. It also allows
13.5 Summary and Outlook 257
one to determine the variance, which the optimal guidance strategy should minimize
in favour of reliable recommendations.
13.5 Summary and Outlook
With the described decision experiments, we have explored different and identified
superior ways of information presentation that facilitate to guide user decisions
in the spirit of higher payoffs. By far the least standard deviations from the user
equilibrium could be reached by presenting the own payoff and the potential payoff,
if the respective participant (or a certain fraction of players) had additionally
chosen the other alternative. Interestingly, the decision dynamics was found to be
intermittent similar to the volatility clustering in stock markets, where individuals
also react to aggregate information. This results from the desire to reach above-
average payoffs, combined with the immanent overreaction in the system. We have
also demonstrated that payoff losses due to a volatile decision dynamics (e.g., excess
travel times) can be reduced via user-specific recommendations by a factor of three
or more. Such kinds of results will be applied to the route guidance on German
highways (see, for example, the project SURVIVE conducted by Nobel prize winner
Reinhard Selten and Michael Schreckenberg). Optimal recommendations to reach
the user equilibrium follow directly from the derived balance equation (13.19) for
decision changes based on empirical transition and compliance probabilities. The
quantification of the transition probabilities requires a novel stochastic description
of the decision behavior, which is not just driven by the potential (gains in) payoffs,
in contrast to intuition and established models. To understand these findings, one has
to take into account reinforcement learning, which can also explain the emergence
of individual response patterns (see Sect. 13.4).
Obviously, it requires both, theoretical and experimental efforts to get ahead
in decision theory. In a decade from now, the microscopic theory of human
interactions will probably have been developed to a degree that allows one to
systematically derive social patterns and economic dynamics on this ground. This
will not only yield a deeper understanding of socio-economic systems, but also
help to more efficiently distribute scarce resources such as road capacities, time,
space, money, energy, goods, or our natural environment. One day, similar guidance
strategies as the ones suggested above may help politicians and managers to stabilize
economic markets, to increase average and individual profits, and to decrease the
unemployment rate.
Acknowledgements This study was partially supported by the ALTANA-Quandt foundation. The
author wants to thank Prof. Aruka, Prof. Selten, and Prof. Schreckenberg for their invitations and
fruitful discussions, Prof. Kondor and Dr. Schadschneider for inspiring comments, Tilo Grigat for
preparing some of the illustrations, Martin Schönhof and Daniel Kern for their help in setting up
and carrying out the decision experiments, and the test persons for their patience and ambitious
playing until the end of our experiments. Hints regarding manuscript-related references are very
much appreciated.
References
1. J. Adler, V. Blue, Towards the design of intelligent traveler information systems. Transport.
Res. C 6, 157–172 (1998)
2. R. Arnott, A. de Palma, R. Lindsey, Does providing information to drivers reduce traffic
congestion? Transport. Res. A 25, 309–318 (1991)
3. W.B. Arthur, Inductive reasoning and bounded rationality. Am. Econ. Rev. 84, 406–411
(1994)
4. Articles in Route Guidance and Driver Information, IEE Conference Publications, Vol. 472
(IEE, London, 2000)
5. W. Barfield, T. Dingus, Human Factors in Intelligent Transportation Systems (Erlbaum,
Mahwah, NJ, 1998)
6. M. Ben-Akiva, A. de Palma, I. Kaysi, Dynamic network models and driver information
systems. Transport. Res. A 25, 251–266 (1991)
7. M. Ben-Akiva, D.M. McFadden, et al., Extended framework for modeling choice behavior.
Market. Lett. 10, 187–203 (1999)
8. M. Ben-Akiva, J. Bottom, M.S. Ramming, Route guidance and information systems. Int. J.
Syst. Contr. Engin. 215, 317–324 (2001)
9. M. Ben-Akiva, S.R. Lerman, Discrete Choice Analysis: Theory and Application to Travel
Demand (MIT Press, Cambridge, MA, 1997)
10. P. Bonsall, P. Firmin, M. Anderson, I. Palmer, P. Balmforth, Validating the results of a route
choice simulator. Transport. Res. C 5, 371–387 (1997)
11. P. Bonsall, The influence of route guidance advice on route choice in urban networks.
Transportation 19, 1–23 (1992)
12. D. Challet, M. Marsili, Y.-C. Zhang, Modeling market mechanism with minority game.
Physica A 276, 284–315 (2000)
13. D. Challet, Y.-C. Zhang, Emergence of cooperation and organization in an evolutionary game.
Physica A 246, 407ff (1997)
14. D. Challet, Y.-C. Zhang, On the minority game: Analytical and numerical studies. Physica A
256, 514–532 (1998)
15. P.S.-T. Chen, K.K. Srinivasan, H.S. Mahmassani, Effect of information quality on compliance
behavior of commuters under real-time traffic information. Transport. Res. Record 1676,
53–60 (1999)
16. Y.-W. Cheung, D. Friedman, Individual learning in normal form games: Some laboratory
results. Games Econ. Behav. 19(1), 46–76 (1997)
17. I. Erev, A.E. Roth, Predicting how people play games: Reinforcement learning in experimen-
tal games with unique, mixed strategy equilibria. Am. Econ. Rev. 88(4), 848–881 (1998)
18. D. Fudenberg, D. Levine, The Theory of Learning in Games (MIT Press, Cambridge, MA,
1998)
19. S. Ghashghaie, W. Breymann, J. Peinke, P. Talkner, Y. Dodge, Turbulent cascades in foreign
exchange markets. Nature 381, 767–770 (1996)
20. R. Hall, Route choice and advanced traveler information systems on a capacitated and
dynamic network. Transport. Res. C 4, 289–306 (1996)
21. D. Helbing, M. Schönhof, D. Kern, Volatile decision dynamics: Experiments, stochastic
description, intermittency control, and traffic optimization. New J. Phys. 4, 33.1–33.16 (2002)
22. D. Helbing, A section-based queueing-theoretical traffic model for congestion and travel time
analysis, J. Phys. A: Math. Gen. 36(46), L593-L598 (2003)
23. D. Helbing, Traffic and related self-driven many-particle systems. Rev. Mod. Phys. 73, 1067–
1141 (2001)
24. D. Helbing, Quantitative Sociodynamics (and references therein) (Kluwer Academic,
Dordrecht, 1995)
25. D. Helbing, Stochastische Methoden, nichtlineare Dynamik und quantitative Modelle sozialer
Prozesse (Shaker, Aachen, 1996)
References 259
26. J.B. van Huyck, J.P. Cook, R.C. Battlio, Selection dynamics, asymptotic stability, and
adaptive behavior. J. Pol. Econ. 102(5), 975–1005 (1994)
27. J.B. van Huyck, R.C. Battlio, R.O. Beil, Tacit coordination games, strategic uncertainty, and
coordination failure. Am. Econ. Rev. 80(1), 234–252 (1990)
28. Y. Iida, T. Akiyama, T. Uchida, Experimental analysis of dynamic route choice behavior.
Transport. Res. B 26, 17–32 (1992)
29. A. Khattak, A. Polydoropoulou, M. Ben-Akiva, Modeling revealed and stated pretrip travel
response to advanced traveler information systems. Transport. Res. Record 1537, 46–54
(1996)
30. H.N. Koutsopoulos, A. Polydoropoulou, M. Ben-Akiva, Travel simulators for data collection
on driver behavior in the presence of information. Transport. Res. C 3, 143–159 (1995)
31. M. Kraan, H.S. Mahmassani, N. Huynh, Traveler Responses to Advanced Traveler Informa-
tion Systems for Shopping Trips: Interactive Survey Approach. Transport. Res. Record 1725,
116 (2000)
32. R.D. Kühne, K. Langbein-Euchner, M. Hilliges, N. Koch, Evaluation of compliance rates
and travel time calculation for automatic alternative route guidance systems on freeways.
Transport. Res. Record 1554, 153–161 (1996)
33. T. Lux, M. Marchesi, Scaling and criticality in a stochastic multi-agent model of a financial
market. Nature 397, 498–500 (1999)
99(Suppl. 3), 7229–7236 (2002)
35. H.S. Mahmassani, D.-G. Stephan, Experimental investigation of route and departure time
choice dynamics of urban commuters. Transport. Res. Records 1203, 69–84 (1988)
36. H.S. Mahmassani, R. Jayakrishnan, System performance and user response under real-time
information in a congested traffic corridor. Transport. Res. A 25, 293–307 (1991)
37. H.S. Mahmassani, R.C. Jou, Transferring insights into commuter behavior dynamics from
laboratory experiments to field surveys. Transport. Res. A 34, 243–260 (2000)
38. R.N. Mantegna, H.E. Stanley, Introduction to Econophysics: Correlations and Complexity in
Finance (Cambridge University, Cambridge, England, 1999)
39. J. Nachabar, Prediction, optimization, and learning in repeated games. Econometrica 65,
275–309 (1997)
40. S. Nakayama, R. Kitamura, Route Choice Model with Inductive Learning. Transport. Res.
Record 1725, 63–70 (2000)
41. J. de D. Ortúzar, L.G. Willumsen, Modelling Transport, Chap. 7: Discrete-Choice Models
(Wiley, Chichester, 1990)
42. M. Schreckenberg, R. Selten (eds.), Human Behaviour and Traffic Networks (Springer, Berlin,
2004)
43. M. Schreckenberg, R. Selten, T. Chmura, T. Pitz, J. Wahle, Experiments on day-to-day route
choice (and references therein), e-print http://vwitme011.vkw.tu-dresden.de/TrafficForum/
journalArticles/tf01080701.pdf, last accessed on March 8, 2012
44. K.K. Srinivasan, H.S. Mahmassani, Modeling Inertia and Compliance Mechanisms in Route
Choice Behavior Under Real-Time Information. Transport. Res. Record 1725, 45–53 (2000)
45. J. Wahle, A. Bazzan, F. Klügl, M. Schreckenberg, Decision dynamics in a traffic scenario.
Physica A 287, 669–681 (2000)
46. J. Wahle, A.L.C. Bazzan, F. Klügl, M. Schreckenberg, Anticipatory traffic forecast using
multi-agent techniques, in Traffic and Granular Flow ’99, ed. by D. Helbing, H.J. Herrmann,
M. Schreckenberg, D.E. Wolf (Springer, Berlin, 2000), pp. 87–92
Chapter 14
Systemic Risks in Society and Economics
14.1 Introduction
When studying systemic risks, i.e. risks that can trigger unexpected large-scale
changes of a system or imply uncontrollable large-scale threats to it, scientific
research has often focused on natural disasters such as earthquakes, tsunamis, hur-
ricanes, vulcano outbreaks, or on failures of engineered systems such as blackouts
of the electric power grids or nuclear accidents (as in Chernobyl). However, many
major disasters hitting human societies relate to social problems [1–4]: This includes
famines and other shortages of resources, wars, climate change, and epidemics,
some of which are related to population density and population growth. Financial
instabilities and economic crises are further examples of systemic risks.
Let us illustrate these risks by some numbers: World War I caused more than
15,000,000 victims, and World War II even 60,000,000 fatalities. The latter gener-
ated costs of 1,000 billion 1944 US$ and destroyed 1,710 cities, 70,000 villages,
31,850 industrial establishments, 40,000 miles of railroad, 40,000 hospitals, and
84,000 schools. Moreover, the world has seen many lossful wars ever since.
The current financial and economic crises triggered an estimated loss of 4-20
trillion US$.
Climate change is expected to cause natural disasters, conflicts for water, food,
land, migration, social and political instability. The related reduction of the world
gross domestic product is expected to amount to 0.6 trillion US$ per year or more.
Turning our attention to epidemics, the Spanish flu has caused 20-40 million deaths,
and SARS has triggered losses of 100 billion US$.
Considering these examples, one could in fact say “The major risks are social”,
but they are still poorly understood. In fact, we know much more about the origin

This chapter reprints a Case Study to be cited as D. Helbing (2010) Systemic Risks in Society
and Economics. International Risk Governance Council (irgc), see http://irgc.org/IMG/pdf/
Systemic Risks Helbing2.pdf.

262 14 Systemic Risks in Society and Economics
of the universe and about elementary particles than about the working of our socio-
economic system. This situation must be urgently changed (see Sect. 14.5).
It is obvious that mankind must be better prepared for the crises to come.
A variety of factors is currently driving the world out of equilibrium: Population
growth, climate change, globalization, changes in the composition of populations,
and the exploitation of natural resources are just some examples. As president
of New York’s Columbia University, Lee C. Bollinger formulated the problem
as follows: “The forces affecting societies around the world ... are powerful and
novel. The spread of global market systems ... are ... reshaping our world ...,
raising profound questions. These questions call for the kinds of analyses and
understandings that academic institutions are uniquely capable of providing. Too
many policy failures are fundamentally failures of knowledge” [5].
We certainly need to increase our capacity to gain a better understanding
of socio-economic systems, conditions triggering instabilities, alternative system
designs, ways to avoid or mitigate crises, and side effects of policy measures. This
contribution will shortly summarize the current knowledge of how systemic risks
emerge in society, and give a variety of relevant examples.
14.2 Socio-Economic Systems as Complex Systems
An important aspect of social and economic systems is that they are complex
systems (see Fig. 14.1) [6–38]. Other examples of complex systems are turbulent
fluids, traffic flows, large supply chains, or ecological systems. The commonality of
complex systems is that they are characterized by a large number of interacting
(mutually coupled) system elements (such as individuals, companies, countries,
cars, etc.) [7, 39–49]. These interactions are usually non-linear (see Sect. 14.2.1).
Typically, this implies a rich system behavior [7]. In particular, such systems tend to
behave dynamic rather than static, and probabilistic rather than deterministic. As a
consequence, complex systems can show surprising or even paradoxical behaviors.
The slower-is-faster effect [50, 51], according to which delays can sometimes speed
up the efficiency of transport systems, may serve as an example.
Moreover, complex systems are often hardly predictable and uncontrollable.
While we are part of many complex systems (such as traffic flows, groups or
crowds, financial markets, and other socio-economic systems), our perception of
them is mostly oversimplified [52, 53] or biased [54–56]. In fact, they challenge our
established ways of thinking and are currently a nightmare for decision-makers [52].
The following subsections will explain these points in more detail.
Note that there are at least three different ways in which the term “complexity”
is used:
1. Structural complexity applies, for example, to a car, which is a complicated
system made up of many parts. These parts, however, are constructed in a way
that makes them behave in a deterministic and predictable way. Therefore, a car
is relatively easy to control.
14.2 Socio-Economic Systems as Complex Systems 263
Fig. 14.1 Freeway traffic constitutes a dynamically complex system, as it involves the interaction
of many independent driver-vehicle units with a largely autonomous behavior. Their interactions
can lead to the self-organization of different kinds of traffic jams, the occurence of which is hard
to predict (after [57])
2. Dynamic complexity may be illustrated by freeway traffic. Here, the interaction

of many independent driver-vehicle units with a largely autonomous behavior
can cause the self-organization of different kinds of traffic jams, the occurence
of which is hard to predict (see Fig. 14.1).
3. Algorithmic complexity measures how the computer resources needed to simulate
or optimize a system scale with system size.
This chapter mainly focuses on dynamic complexity.
14.2.1 Non-Linear Interactions and Power Laws
Systems with a complex system dynamics are mostly characterized by non-linear

interactions among the elements or entities constituting the system (be it particles,
objects, or individuals). Non-linear interactions are typical for systems in which
elements mutually adapt to each other. That is, the elements are influenced by
their environment, but at the same time, they also have an impact on their
environment.
Non-linearity means that causes and effects are not proportional to each other.
A typical case is a system that is hardly responsive to control attempts, or which
shows sudden regime shifts when a “tipping point” is crossed [58–63] (see
Fig. 14.2). Examples for this are sudden changes in public opinion (e.g. from
smoking-tolerance to smoking bans, from pro- to anti-war mood, from a strict
banking secret to transparency, or from buying pickup trucks to buying environment-
friendly cars).
Fig. 14.2 Schematic illustration of one of the typical behaviors of complex systems: In regimes 1
and 2, a “cause” (such as a control attempt) has essentially no effect on the system, while at the
“tipping point”, an abrupt (and often unexpected) transition to a different system behavior occurs.
A recent example is the sudden large-scale erosion of the Swiss banking secret, after UBS had
handed over about 300 names of clients to an US authority
Fig. 14.3 When system components interact strongly, the normally distributed behavior of sepa-
rated system elements often becomes (approximately) power-law distributed. As a consequence,
fluctuations of any size can occur in the system, and extreme events are much more frequent than
expected. Note that power laws are typical for a system at a critical point, also known as “tipping
point”
14.2.2 Power Laws and Heavy-Tail Distributions
It is important to note that strong interactions among the system elements often
change the statistical distributions characterizing their behavior. Rather than normal
distributions, one typically finds (truncated) “power laws” or, more generally, so-
called heavy-tail distributions [48, 49, 58] (see Fig. 14.3 and Sect. 14.2.4). These
imply that extreme events occur much more frequently than expected. For example,
the crash of the stock market on Black Monday was a 35 sigma event (where
sigma stands for the standard deviation of the Dow Jones Index on a logarithmic
Fig. 14.4 Example of a blackout of the electrical power grid in Europe (from: EU project IRRIIS.
E. Liuf (2007) Critical Infrastructure protection, R&D view). To allow for the transfer of a
ship, one power line had to be temporarily disconnected in Northern Germany. This triggered an
overload-related cascading effect [80], during which many power lines went out of operation. As
a consequence, there were blackouts all over Europe (see black areas). The pattern illustrates how
counter-intuitive and hardly predictable the behavior of complex systems with network interactions
can be
scale). Other examples are the size distributions of floods, storms, earth quakes, or
wars [1–4]. Obviously, the occurence of the respective heavy-tail distributions is
highly important for the insurance business and for the risk assessment of financial
derivatives.
14.2.3 Network Interactions and Systemic Risks Through

Failure Cascades
A typical case of non-linear interactions are network interactions, which are ubiq-
uitous in socio-economic systems [64–79]. These imply feedback loops and vitious
circles or induce (often undesired) side effects [32]. (For example, the introduction
of cigarette taxes has promoted smuggling and other criminal activities.) Moreover,
network interactions are often the reason for a cascading of failure events. Examples
for this are epidemic spreading, the failure of the interbank market during a financial
crisis, the spreading of traffic congestion, or the blackout of an electrical power
system (see Fig. 14.4).
Failure cascades (which are also called chain reactions, avalanche or domino
effects) are the most common mechanism by which local risks can become systemic
[81–84] (see Fig. 14.5). Systemic failures are usually triggered by one of the
following reasons:
1. The parameters determining system stability are driven towards a so-called
“critical point” or “tipping point”, beyond which the system behavior becomes
16 0.0 40 0.0
over-critical
23 0.0 38 0.0 31 0.0
perturbation
10 0.0 24 0.0
21 0.0 5 0.0
9 0.0 42 0.0
17 0.0
9 0.0 50
0.0
14 0.0 15 0.0 48 0.0
44
30 0.0 22 0.0
1 0.0
26 0.0 42 0.0
8 0.0
6 0.0 3 0.0
12 0.0 28 0.0
11 0.0 19 0.0
47 0.0
4 0.0
2 0.0
35 0.0
33 0.0
27 0.0
37 20 0.0
46 0.0 0.0 49 0.0
45 0.0
feedback
loop
41 0.0
Fig. 14.5 Top: Schematic illustration of a networked system which is hit by an over-critical
perturbation (e.g. a natural disaster). The problem of feedback cycles is highlighted. They can
have “autocatalytic” (escalation) effects and act like vitious circles. Bottom: Illustration of
cascading effects in socio-economic systems, which may be triggered by the disruption (over-
critical perturbation) of an anthropogenic system. A more detailed picture can be given for specific
disasters. Note that the largest financial damage of most disasters is caused by such cascading
effects, i.e. the systemic impact of an over-critical perturbation (after [85])
unstable (see Sect. 14.2.1). For example, the destabilization of the former
German Democratic Republic (GDR) triggered off spontaneous demonstrations
in Leipzig, Germany, in 1989, which eventually caused the re-unification of
Germany. This “peaceful revolution” shows that systemic instability does not
necessarily imply systemic malfunctions. It can also induce a transition to a better
and more robust system state after a transient transformation period. Further
examples of spontaneous transitions by systemic destabilization are discussed
in Sects. 14.2.4, 14.3, and 14.4.1.
2. The system is metastable (i.e. robust to small perturbations, which quickly
disappear over time), but there occurs an over-critical perturbation (such as a
Fig. 14.6 The most efficient disaster response strategy depends on many factors such as the
network type (after [84]). Here, we have studied six different disaster response strategies for regular
grids, scale-free networks, and Erdös-Rényi random networks. The best strategy is a function of
the resources R available for disaster response management and the time delay tD before practical
measures are taken. Obviously, there is no single strategy, which always performs well. This makes
disaster response challenging, calling for scientific support
natural disaster), which harms the system functionality so much that this has
damaging effects on other parts of the system [84] (see Fig. 14.6).
3. The system is metastable, but there is a coincidence of several perturbations
in the network nodes or links such that their interaction happens to be over-
critical and triggers off additional failures in other parts of the system [83]. In
fact, disasters caused by human error [86, 87] are often based on a combination
of several errors. In networked systems, the occurence of this case is just a matter
of time.
14.2.4 Self-Organized or Self-Induced Criticality
A system may get into a critical state not only by external influences that are
affecting system stability. It is known that some endogeneous processes can
automatically drive the system towards a critical state, where avalanche or cascading
effects of arbitrary size appear (reflecting the characteristic heavy-tail statistics at
critical points, see Sect. 14.2.2). In such cases, the occurence of extreme events
is expected, and we speak of “self-induced” or “self-organized criticality” (SOC)
[88, 89].
It is likely that bankruptcy cascades can be understood in this way. The
underlying mechanism is that a company or bank tries to make a better offer to
customers or clients than the competing companies or banks do. This forces the
competitors to make better offers as well. Eventually, the profit margins in a free
market become so small that variations in the consumption rate can drive some
companies or banks out of business, which creates economic problems for other
companies or banks. Considering the interconnections between different companies
or banks, this mechanism can cause bankruptcy cascades. Eventually, the number
of competitors will be smaller, and as a consequence, they can take higher prices.
Therefore, their profits go up, which encourages new competitors to enter the
market. In this way, competition increases again and automatically drives the system
back to low profits and bankruptcies.
Another example concerns safety standards [86, 87]. These are usually specified
in such a way that normal perturbations would not cause serious harm or even
systemic failures. As a consequence, most man-made systems are constructed
in a way that makes them robust to small and moderate perturbations (in other
words: meta-stable). However, the requirement of cost efficiency excerts pressure
on decision-makers to restrict safety standards to what really appears to be needed,
and not more. Consequently, if a large-scale failure has not occurred in a long
time, decision-makers often conclude that the existing safety standards are higher
than necessary and that there is some potential to reduce costs by decreasing
them somewhat. Eventually, the standards are lowered so much that an over-
critical perturbation occurs sooner or later, which causes a systemic failure. As a
consequence, the safety-standards will be increased again, and the process will start
from the beginning.
As a third example, let us discuss man-made systems with capacity limits such
as traffic or logistic systems. These systems are often driven towards maximum
efficiency, i.e. full usage of their capacity. However, when reaching this point of
maximum efficiency, they also reach a tipping point, at which the system becomes
dynamically unstable [90]. This is known, for example, from freeway and railway
traffic. As a consequence, the system suffers an unexpected capacity drop due to
optimization efforts, shortly after the maximum performance was reached.
Similarly to freeway traffic, engineers also try to avoid the occurence of
congestion in urban traffic, which can be reached by re-routing strategies. A closer
analysis shows that this optimization leads again to a sudden breakdown of the
flow, once the maximum throughput is reached [91]. One may, therefore, conclude
that optimizing for the full usage of available system capacity implies the danger
of an abrupt breakdown of the system performance with potentially very harmful
consequences. To avoid this problem, one must know the capacity of the system and
avoid to reach it. This can be done by requiring to respect sufficient safety margins.
14.2.5 Limits of Predictability, Randomness, Turbulence

and Chaos
The large number of non-linearly coupled system components can lead to a complex
dynamics (see Fig. 14.7). Well-known examples for this are the phenomena of
turbulence [92] and chaos [42, 93], which make the dynamics of the system
unpredictable after a certain time period. A typical example are weather forecasts.
Fig. 14.7 Illustration of

various cases of non-linear
dynamics that can occur in
complex systems (from [98],
p. 504). Deterministic chaos
and turbulence constitute
further and even more
complicated cases of
non-linear system dynamics
The large sensitivity to small perturbations is sometimes called the “butterfly

effect”, suggesting that (in a chaotically behaving system) the flight of a butterfly
could significantly change the system behavior after a sufficiently long time.
A further obstacle for predicting the behavior of many complex systems is a
probabilistic or stochastic dynamics [94, 95], i.e. the importance of randomness.
In socio-economic systems, there is furthermore a tendency of self-fulfilling or
self-destroying prophecy effects [96] (and it is hard to say which effect will finally
dominate, see the current response of the population to the swine flu campaign).
Stock markets show both effects: On the one hand, the self-fulfilling prophecy
effect leads to herding behavior, which creates bubbles [97]. On the other hand,
the competition for the highest possible returns eventually destroys any predictable
gains (otherwise everybody could become rich without having to work, thereby
creating a “financial perpetuum mobile”). Altogether, this competition creates a
(more or less) “efficient” and unpredictable stock market. A generalization of this
principle is known as Goodhart’s law.
14.2.6 The Illusion of Control
Besides the difficulties to predict the future behavior of complex systems, there are
other effects which make them difficult to control:
Fig. 14.8 When a complex system is changed (e.g. by external control attempts), its system
parameters, stability, and dynamics may be affected. This figure illustrates the occurence of a
so-called “cusp catastrophe”. It implies discontinuous transitions (regime shifts) in the system
dynamics
1. On the one hand, big changes may have small or no effects (see Fig. 14.2) and,
when considering network interactions (see Sect. 14.2.3), even adverse effects.
This reflects the principle of Le Chatelier,according to which a system tends to
counteract external control attempts.
2. On the other hand, if the system is close to a “critical” or “tipping point”,
even small changes may cause a sudden “regime shift”, also known as “phase
transition” or “catastrophe” (see Figs. 14.2 and Sect. 14.8). In other words,
small changes can sometimes have a big impact, and often very unexpectedly
so. However, there are typically some early warning signals for such critical
transitions [99]. This includes the phenomenon of “slow relaxation”, which
means that it takes a long time to dampen out perturbations in the system, i.e.
to drive the system back to equilibrium.
Another warning signal of potential regime shifts are “critical fluctuations”,
which normally obey a heavy-tail distribution (see Sect. 14.2.2). In other words,
perturbations in the system tend to be larger than usual – a phenonenon which is
also known as “flickering”.
3. Control attempts may also be obstructed by “irreducible randomness”, i.e.
a degree of uncertainty or perturbation which cannot be eliminated (see
Sect. 14.2.5).
4. Delays are another typical problem that often cause a failure of control [100]. The
underlying reason is that delays may create an unstable system behavior (also
when people attempt to compensate delays by anticipation). Typical examples
are the breakdown of traffic flows and the occurence of stop-and-go traffic, which
result from delayed speed adjustments of drivers to variations in the vehicle
speeds ahead.
Since many control attempts these days are based on the use of statistics, but
compiling such statistics is time-consuming, delays may cause instabilities also
in other areas of society. Business cycles, for example, may result from such
delays as well (or may at least be intensified by them).
5. Finally, there is the problem of “unknown unknowns” [101], i.e. hidden factors
which influence the system behavior, but have not been noticed before. By
14.3 The Example of Financial Market Instability 271
definition, they appear unexpectedly. “Structural instabilities” [39] may create

such effects. The appearance of a new species in an ecosystem is a typical
example. In economics, this role is played by innovations or new products, which
happen to change the social or economic world. Well-known examples for this
are the invention of contraceptives, computers, or mobile phones.
14.2.7 The Logic of Failure
As a consequence of the above, complex systems cannot be controlled in the

conventional way (like pressing a button or steering a car). Such control attempts
will usually fail, as Doerner’s book “The Logic of Failure” has impressively
shown [52].
A typical failure scenario is as follows: A decision-maker tries to change the
social system. It turns out that the measure taken does not have any effect (see
Fig. 14.2). Therefore, he or she decides to intensify the measure. The effect may
still not be as expected. Hence, an even more forceful control attempt is made. As a
consequence, the system undergoes a sudden regime shift (see Figs. 14.2–14.8) and
the system organizes itself in a different way (but not necessarily in the desired way).
The decision-maker now tries to re-gain control and counteracts the unexpected
change. If the attempts to stabilize the system are delayed, this can even lead to an
oscillatory or chaotic system dynamics.
The right approach to influence complex systems is to support and strengthen
the self-organization and self-control of the system by mechanism design (see
Sect. 14.4.1). This basically means that coordination and cooperation in a complex
system will appear by itself, if the interactions among the system elements are well
chosen. That is, regulations should not specify what exactly the system elements
should do, but set bounds to actions (define “rules of the game”), which give the
system elements enough degrees of freedom to self-organize good solutions. If the
interaction rules are suitable, such an approach will usually lead to a much more
flexible and adaptive system behavior. Another advantage is “systemic robustness”,
i.e. the ability cope with challenges by external perturbations. Note however,
that everything depends on the interactions of the system elements. Unsuitable
interactions can, for example, cause that the system behaves dynamically unstable
or that it gets trapped in a suboptimal (“frustrated”) state. Hence, finding the right
interaction rules is a great challenge for decision-makers, and complex systems
scientists are needed to address them properly.
14.3 The Example of Financial Market Instability
One example of systemic risks that deserves more attention here is financial market
instability [102–108]. The recent financial crises shows very clearly how cascading
effects can lead to an uncontrollable dynamics and a relatively sudden systemic
crises. What started with local problems concerning subprime mortgages eventually
affected the mortgage companies, the home building industry, the financial markets,
the US economy, and the world economy. This crisis has been explained in many
ways. Widely discussed reasons include:
• The deregulation of financial markets.
• The explosive spread of derivatives (which reached a value of 15 times the gross
product of the world).
• The apparently “riskless” securization of risky deals by credit default swaps,
lowering lending standards.
• The opaqueness (intransparency) of financial derivatives.
• The failure of rating agencies due to the complexity of the financial products.
• Bad risk models (neglecting, for example, correlations and the heavy-tail charac-
ter of the fluctuations).
• Calibration of risk models with historical data not reflecting the actual situation.
• Insufficient net assets of banks.
• Low interest rates to fight previous crises.
• The growth of over-capacities and other developments with pro-cylical effects.
• Short-term incentive structures (“bonus schemes”) and “greed” of investment
bankers and managers.
Less debated, but not less relevant reasons are [109–111]:
• The complexity of the financial system is larger than what is knowable. For
example, many portfolios appear to contain too many different assets to support
a reliable optimization with the amount of data available [112].
• In the “arms race” between banks (and other agents) with the regulators,
regulators are sometimes in the weaker position. Therefore, financial market
instability may result from the fact that instability is beneficial for some interest
groups: It requires an unstable market to allow some people to become very
rich in a short time: Instability implies opportunities for good investments. When
GDP grows slowly, good returns mainly result from financial bubbles.
• The financial architecture has created a complex system, with a hard-to-predict
and hard-to-control dynamics. Financial products (“derivatives”) were con-
structed in a multi-level way, very much like a house of cards.
• The world-wide network interdependencies of all major banks have spread local
risks all over the system to an extent that produced a systemic risk. It created a
“global village” without any “firewalls” (security breaks).
• Delays in the adaptation of some markets build up disequilibria in the system
with the potential of earthquake-like stress reliefs. As examples for this, one may
take historical crashes in currency markets or recent drops in the values of certain
AAA-rated stocks.
• The financial and economic system are organized in a way that allows for the
occurrence of strong correlations. For example, when the strategies of companies
all over the world become more and more similar (due to “group think” [113] or
asking the same consultancy companies), a lack of variety (heterogeneity) results
14.3 The Example of Financial Market Instability 273
in the system. This can cause (more or less) that either no company fails or many
companies fail at the same time.
• An important factor producing herding effects [114, 115] and bubbles is the con-
tinuous information feedback regarding the investment decisions of others [116].
In this connection, it is important to underline that repeated interactions between
decision-makers supports consensus, but creates over-confidence (i.e. a false
feeling of safety, despite misjudgements of reality). Therefore, it undermines the
“wisdom of crowds” [117, 118]. This problem may be further intensified by the
public media which, in the worst case, may even create a mass hysteria.
• The price formation mechanism mixes material values and psychology in a single,
one-dimensional quantity, the price. Therefore, the price dynamics is sensitive to
factors such as trust, risk aversion, greed, and herding effects (the imitation of
the behavior of others) [54–56, 119].
• A stability of single banks does not imply that the banking system cannot enter
a state of systemic instability. (Monetary value is a matter of trust, and therefore
a single event such as the failure of Lehmann Brothers could induce that banks
were not anymore willing to lend money to each other. This triggered a liquidity
crises so big that it would have caused the failure of the world financial system,
if the central banks would not have quickly provided huge amounts of liquidity.)
• Lack of trust also reduces lending of cheap money to troubled companies, which
may drive them into bankruptcy, thereby increasing a bank’s problems.
• More generally, the economic system seems to have a tendency towards self-
organized critical behavior (see Sect. 14.2.4).
Many of the above factors have contributed to strong non-linear couplings in
the system. Furthermore, strong network interdependencies have been created
through the interbank markets and complex financial derivatives. These features
are already expected to imply cascade-like effects and a heavy-tail statistics (see
Sect. 14.2.2). This tendency is expected to be further amplified by anticipation
attempts in fluctuating markets. However, even more dangerous than the occurrence
of fluctuations in the markets is the occurence of strong correlations. These can
be promoted by economic cycles, herding effects, and the coupling of policies or
regulation attempts to global risk indicators.
The worldwide crisis in the automobile sector in 2009 and the quant meltdown in
August 2007 are good examples for the occurence of strong correlations. The latter
may be understood as follows [120]: Returns of hedge fonds largely depend on
their leverage. Therefore, there is an “evolutionary pressure” towards high leverage,
which can increase volatility. In case of huge price jumps, however, banks tend to
demand their loans back. This decreases the leverage of the affected hedge funds
and thereby their chances to perform well in the future. Hence, large system-wide
leverage levels are pre-requisites for collapses, and crises can emerge virtually “out
of nothing”, just through fluctuations. This example illustrates well how unsuitable
risk-averse policies can create pro-cyclical effects, through which banks may harm
their own interests.
14.4 Managing Complexity
Having discussed the particular challenges of complex systems, one may be left with
the impression that such systems are just too difficult for us to handle. However, in
the past decades, a variety of scientific techniques have been developed to address
these challenges. These include:
• Large-scale data mining.
• Network analysis.
• Systems dynamics.
• Scenario modeling.
• Sensitivity analysis.
• Non-equilibrium statistical physics.
• Non-linear dyamics and chaos theory.
• Systems theory and cybernetics.
• Catastrophe theory.
• The statistics of extreme events.
• The theory of critical phenomena and, maybe most prominently these days:
• Agent-based modeling [129–133].
The methods developed by these fields allow us to better assess the sensitivity
or robustness of systems and their dynamics, as will be shortly discussed in the
following. They have also revealed that complex systems are not our “enemies”.
In fact, they possess a number of favorable properties, which can be used to our
benefit.
14.4.1 How to Profit from Complex Systems
Understanding complex systems facilitates to utilize their interesting properties,

which however requires one to work with the system rather than against it
[121–128]. For example, complex systems tend to show emergent (collective)
properties, i.e. properties that the single system components do not have. This is, for
example, relevant for the possibility of collective intelligence [134–136]. One may
also benefit from the fact that complex systems tend to self-organize in a way, which
is adaptive to the enviroment and often robust and resource-efficient as well. This
approach has, for example, been successfully applied to develop improved design
principles for pedestrian facilities and other systems.
Technical control approaches based on self-organization principles become
more and more available now. While previous traffic control on highways and in
cities was based on a centralized optimization by supercomputers with expensive
measurement and control infrastructures, currently developed approaches are based
on decentralized coordination strategies (such as driver assistant systems or traffic
lights that are flexibly controlled by local traffic flows).
14.4 Managing Complexity 275
Fig. 14.9 One advantage of centralized control is quick large-scale coordination. However,
disadvantages result from the vulnerability of the network, a tendency of information overload, the
risk of selecting the wrong control parameters, and delays in adaptive feedback control. Because of
greater flexibility to local conditions and greater robustness to perturbations, decentralized control
approaches can perform better in complex systems with heterogeneous elements, large degree of
fluctuations, and short-term predictability (after [139])
Centralized structures can reach a quick information exchange among remote

parts of a system, but they become unstable beyond a certain critical size (as the
collapse of political states and many unsuccessful mergers of companies show).
In comparison, decentralized approaches are particularly suited to reach a flexible
adjustment to local conditions and local coordination [137]. Some decentralized
concepts for real-time control already exceed the performance of centralized ones,
particularly in complex, hardly controllable, fluctuating enviroments, which require
a quick and flexible response to the actual situation [138] (see Fig. 14.9). In fact,
in a strongly varying world, strict stability and control is not possible anymore or
excessively expensive (as the public spending deficits show). Therefore, a paradigm
shift towards more flexible, agile, adaptive systems is needed, possible, and overdue.
The best solutions are probably based on suitable combinations of centralized and
decentralized approaches.
In social systems, the principle of self-organization, which is also known as
principle of the “invisible hand”, is ubiquitous. However, self-organization does not
automatically lead to optimal results, and it may fail under extreme conditions (as is
known, for example, from financial and traffic systems as well as dense pedestrian
crowds).
A particularly important example of self-control is the establishment of social
norms, which are like social forces guiding the behavior of people. In this way,
social order can be created and maintained even without centralized regulations
such as enforced laws. Nevertheless, one must be aware that the principles on which
social cooperation and norms are based (for example, repeated interaction, trust
and reputation, or altruistic sanctioning of deviant behavior) are fragile. Simple
Fig. 14.10 Establishment of cooperation in a world with local interactions and local mobility (left)
in comparison with the breakdown of cooperation in a world with global interactions and global
mobility (right) (blue D cooperators, red D defectors/cheaters/free-riders) (after [140]). Note that
the loss of solidarity results from a lack of neighborhood interactions, not from larger mobility
computer simulations suggest, for example, that a change from repeated local
interactions (between family members, friends, colleagues, and neighbors) to non-
recurring interactions with changing partners from all over the world may cause
a breakdown of human cooperation [140]. Therefore, the on-going globalization
could potentially destabilize our social systems [141–143] (see Fig. 14.10), which
largely builds on norms and social cooperation. (Remember, for example, that the
breakdown of the interbank market, which almost caused a collapse of the world
financial system, was due to a breakdown of the network of trust.)
14.4.2 Reducing Network Vulnerability
In Sect. 14.2.3, we have seen that systemic risks are mostly based on cascade
spreading effects in networks. However, the vulnerability of networks to such
spreading events can be reduced. The following measures are often quite effective:
• The network structure can often been improved by redundancy, i.e. the provision
of alternatives, so that an over-critical perturbation would only occur, if several
nodes would fail or several links would break simultaneously.
• However, too much interconnectedness may be harmful, as it is provides
the “infrastructure” for the system-wide spreading of an unexpected problem.
Therefore, it makes sense to limit the degree of connectedness and the size of
networks (in order to avoid a “too big to fail” problem).
• Alternatively, one can introduce “firewalls”: Having several networks, each of
them characterized by strong links, while the connections between the networks
are weak, would allow to decouple the so defined supernetwork into several
subnetworks (see Fig. 14.11). This principle of decompartementalization allows
one to prevent the spreading of a problem over the whole system, if the
disconnection strategy is well chosen. The principle of firewalls to protect
Fig. 14.11 A networked system should be constructed in a way that allows its quick decom-
position or decompartementalization into weakly coupled (or, if necessary, even uncoupled)
subnetworks. In such a way, failure cascades all over the system (or large parts of it) can be
avoided, and most parts of it can be protected from damage
computer systems from malicious intrusion or the principle of electrical fuses

to protect an electrical network from overload could certainly be transferred to
other networked systems such as the financial system.
• For similar reasons, a heterogeneity (variety) among the nodes and/or links of
a network (in terms of design principles and operation strategies) will normally
increase its robustness.
• When fighting failure cascades in networks, a quick response to over-critical
perturbations is absolutely decisive. If the time delay of disaster response
management is small, its effectiveness depends in a complicated way on the
network structure, the amount of resources, and the strategy of distributing them
in the network (see Fig. 14.6). In case of significant delays, cascade spreading
can hardly be mitigated, even when large resources are invested.
• A moderate level of fluctuations may be useful to destroy potentially harmful
correlations (such as financial bubbles) in the system. Such fluctuations could be
created by central banks (for the purpose of “bubble control”) or by other regula-
tors, depending on the system. Note, however, that a large degree of fluctuations
can cause over-critical perturbations or coincidences of perturbations.
• An unhealthy degree of volatility can be lowered by introducing conservation
laws and/or frictional effects in the system. This is expected to dampen fluctu-
ations and, thereby, to reduce the likelihood of events that may trigger systemic
risks.
Rather than applying these concepts permanently, it can make sense to use them
adaptively, depending on the state of the system. When designing networked
systems according to the above principles, one can certainly profit from the
experience of physicists and engineers with other systems.
In this contribution, we have summarized properties of complex systems and

identified sources and drivers of systemic risks in socio-economic systems. Complex
systems cannot be easily controlled. They rather tend to follow a self-organized
eigendynamics, and conventional control attempts often have counter-intuitive and

unintended effects.
As the example of ecosystems shows, a networked system can have an aston-
ishing degree of robustness without any central control. Robustness just requires
the right interaction rules, which may be implemented, for example, by social
norms, laws, technological measures etc., depending on the system. Properly chosen
rules will lead to a self-regulation or self-control of the system, but improper
specifications can lead to low performance or systemic instability. For example,
if the failure rate of system elements is reduced, this may lead to larger systemic
failures later on. Moreover, it is probably good if the system is regularly exposed to
stress, as this is expected to strengthen its immunity to perturbations.
It was particularly underlined that, in any larger networked system, it is essential
to have “firewalls” (security breaks), which facilitate its quick decomposition or
decompartmentalization into disconnected or weakly connected subnetworks before
a failure cascade has percolated through the whole system or large parts of it.
Among the success stories of complex systems research, one may mention
the nobel prizes of Ilya Prigogine, Thomas Schelling, and Paul Krugmann. Some
examples for application areas of complexity science are [144–148]
• The organization of the internet.
• Modern epidemiology.
• The prevention of crowd disasters.
• Innovative solutions to improve traffic flow.
• The understanding of global climate change.
• The enhancement of the reliability of energy supply.
• Modern disaster response management.
• Prediction markets and other methods using the wisdom of crowds.
However, many socio-economic crises still occur, because the system dynamics is
not well enough understood, leading to serious management mistakes. In order to
support decision-makers, scientists need to be put in a better position to address
the increasing number of socio-economic problems. These mainly result from the
fact that social and economic systems are rapidly changing, i.e. in a transformation
process rather than in equilibrium.
We must close the gap between existing socio-economic problems and solutions,
and create conditions allowing us to come up with solutions before a problem
occurs. This requires to build up greater research capacities (a “socio-economic
knowledge accelerator”). It will also be necessary to establish a new study direction
(“integrative systems design”) to provide decision-makers with solid knowledge
regarding the behavior of complex systems, how to manage complexity in politics
and economy, and how to cope with crises.
Finally, scientists need to have access to better and more detailed data. Spe-
cial super-computing centers (as for climate research) would allow scientists to
simulate model societies and study the impact of policy measures before their
implementation. They would also support the development of contingency plans and
the investigation of alternative ways of organization (“plan B”). Such centers will
References 279
require a multi-disciplinary collaboration across the various relevant research areas,

ranging from the socio-economic over the natural to the engineering sciences. For
this, one needs to overcome the particular challenges of multidisciplinary research
regarding organization, funding, and publication.
Considering that we know more about the origin of the universe than about the
conditions for a stable society, a prospering economics, and enduring peace, we need
nothing less than an “Apollo project for the socio-economic sciences”. There is no
time to lose, since there are already signs of critical fluctuations indicating possible
regime shifts [149–154], which speak a clear language.
Acknowledgements This work was partially supported by the ETH Competence Center “Coping
with Crises in Complex Socio-Economic Systems” (CCSS) through ETH Research Grant CH1-01
08-2. The author would like to thank Peter Felten for creating many of the illustrations shown in
this contribution. Furthermore, the author is grateful for inspiring discussions with Kay Axhausen,
Stefano Battiston, Lubos Buzna, Lars-Erik Cederman, Hans Herrmann, Imre Kondor, Matteo
Marsili, Frank Schweitzer, Didier Sornette, Stefan Thurner, and many others.
References
1. D. Sornette, Critical Phenomena in Natural Sciences: Chaos, Fractals, Selforganization and

Disorder. (Springer, Berlin, 2006)
2. S. Albeverio, V. Jentsch, H. Kantz, eds., Extreme Events in Nature and Society. (Springer,
Berlin, 2006)
3. A. Bunde, J. Kropp, and H. J. Schellnhuber, eds., The Science of Disasters. Climate
Disruptions, Heart Attacks, and Market Crashes. (Springer, Berlin, 2002)
4. M. Buchanan, Ubiquity. Why Catastrophes Happen. (Three Rivers, New York, 2000)
5. Office of the President of Columbia University, Lee C. Bollinger (2005) Announcing the
Columbia Committee on Global Thought, see http://www.columbia.edu/cu/president/docs/
communications/2005-2006/051214-committee-global-thought.html
6. W. Weidlich, Sociodynamics: A Systematic Approach to Mathematical Modelling in the Social
Sciences. (Harwood Academic, Amsterdam, 2000)
7. H. Haken, Synergetics: Introduction and Advanced Topics. (Springer, Berlin, 2004)
8. D. Helbing, Quantitative Sociodynamics. (Kluwer Academic, Dordrecht, 1995)
9. R. Axelrod, The Complexity of Cooperation: Agent-Based Models of Competition and
Collaboration (Princeton University, Princeton, NJ, 1997)
10. F. Schweitzer, ed. Self-Organization of Complex Structures: From Individual to Collective
Dynamics (Gordon and Breach, Amsterdam, 1997)
11. D.S. Byrne, Complexity Theory and the Social Sciences: An Introduction (Routledge,
New York, 1998)
12. W. Buckley, Society–A Complex Adaptive System: Essays in Social Theory. (Routledge,
London, 1998)
13. S.Y. Auyang, Foundations of Complex-System Theories in Economics, Evolutionary Biology,
and Statistical Physics. (Cambridge University, Cambridge, 1999)
14. F. Schweitzer, ed. Modeling Complexity in Economic and Social Systems. (World Scientific,
Singapore, 2002)
15. R.K. Sawyer, Social Emergence: Societies As Complex Systems (Cambridge University,
Cambridge, 2005)
16. A.S. Mikhailov, V. Calenbuhr, From Cells to Societies. (Springer, Berlin, 2006)
17. B. Blasius, J. Kurths, L. Stone Complex Population Dynamics. (World Scientific, Singapore,
2007)
18. M. Batty, Cities and Complexity: Understanding Cities with Cellular Automata, Agent-Based
Models, and Fractals. (MIT Press, Cambridge, 2007)
19. S. Albeverio, D. Andrey, P. Giordano, A. Vancheri, The Dynamics of Complex Urban Systems.
(Physica, Heidelberg, 2008)
20. K. Mainzer, Thinking in Complexity: The Computational Dynamics of Matter, Mind, and
Mankind. (Springer, Berlin, 2007)
21. J.H. Miller, S.E. Page, Complex Adaptive Systems: An Introduction to Computational Models
of Social Life. (Princeton University, Princeton, NJ, 2007)
23. B. Castellani, F. Hafferty, Sociology and Complexity Science. (Springer, Berlin, 2009)
24. D. Lane, D. Pumain, S.E. van der Leeuw, G. West, eds. Complexity Perspectives in Innovation
and Social Change. (Springer, Berlin, 2009)
25. C. Castellano, S. Fortunato, V. Loreto, Statistical physics of social dynamics. Rev. Mod. Phys.
81, 591–646 (2009)
26. P.W. Anderson, K. Arrow, D. Pines, eds. The Economy as an Evolving Complex System.
(Westview, Boulder, Menzies, NK, 1988)
27. H.W. Lorenz, Nonlinear Dynamical Equations and Chaotic Economy. (Springer, Berlin,
1993)
28. P. Krugman, The Self-Organizing Economy. (Blackwell, Malden, MA, 1996)
29. W.B. Arthur, S.N. Durlauf, D.A. Lane, The Economy as an Evolving Complex System II.
(Westview, Boulder, Menzies, NK, 1997)
30. R.N. Mantegna, H.E. Stanley, An Introduction to Econophysics: Correlations and Complexity
in Finance. (Cambridge University, Cambridge, 1999)
31. D. Colander, ed. The Complexity Vision and the Teaching of Economics. (Elgar, Cheltenham,
UK, 2000)
32. J.D. Sterman, Business Dynamics: Systems Thinking and Modeling for a Complex World.
(Boston, McGraw Hill, 2000)
33. T. Puu, Attractors, Bifurcations, & Chaos. Nonlinear Phenomena in Economics. (Springer,
Berlin, 2003)
34. L.E. Blume, S.N. Durlauf, eds. The Economy as an Evolving Complex System III. (Oxford
University, Oxford, 2006)
35. B.K. Chakrabarti, A. Chakraborti, A. Chatterjee, Econophysics and Sociophysics. (Wiley,
New York, 2006)
36. A. Chatterjee, B.K. Chakrabarti, eds. Econophysics of Markets and Business Networks.
(Springer, Milan, 2007)
37. A.C.-L. Chian, Complex Systems Approach to Economic Dynamics. (Springer, Berlin, 2007)
38. V. Lacayo, What complexity science teaches us about social change, see http://www.
communicationforsocialchange.org/mazi-articles.php?id=333
39. G. Nicolis and I. Prigogine, Self-Organization in Nonequilibrium Systems. (Wiley, New York,
1977)
40. A.S. Mikhailov, Foundations of Synergetics I: Distributed Active Systems (Springer, Berlin,
1991)
41. A.S. Mikhailov, A.Y. Loskutov, Foundations of Synergetics II: Complex Patterns. (Springer,
Berlin, 1991)
42. S.H. Strogatz, Nonlinear Dynamics and Chaos: With Applications To Physics, Biology,
Chemistry, And Engineering. (Westview Press, 2001)
43. N. Boccara, Modeling Complex Systems. (Springer, Berlin, 2003)
44. J. Jost, Dynamical Systems: Examples of Complex Behaviour. (Springer, Berlin, 2005)
45. G. Nicolis, Foundations of Complex Systems. (World Scientific, New York, 2007)
46. P. Érdi, Complexity Explained. (Springer, Berlin, 2007)
47. C. Gros, Complex and Adaptive Dynamical Systems. (Springer, Berlin, 2008)
References 281
48. M. Schroeder, Fractals, Chaos, Power Laws. (Dover, 2009)

49. A. Saichev, Y. Malevergne, D. Sornette, Theory of Zipf’s Law and Beyond. (Springer, Berlin,
2010)
50. D. Helbing, I. Farkas, T. Vicsek, Simulating dynamical features of escape panic. Nature 407,
487–490 (2000)
51. D. Helbing, A. Mazloumian, Operation regimes and slower-is-faster effect in the control of
traffic intersections. Eur. Phys. J. B 70(2), 257–274 (2009)
52. D. Dorner, The Logic Of Failure: Recognizing and Avoiding Error in Complex Situations.
(Basic, New York, 1997)
53. D. Helbing, Managing complexity in socio-economic systems. Eur. Rev. 17(2), 423–438
(2009)
54. R.H. Thaler, ed. Advances in Behavioral Finance. (Russell Sage Foundation, New York, 1993)
55. C.F. Camerer, G. Loewenstein, M. Rabin, eds. Advances in Behavioral Economics. (Princeton
University, Princeton, NJ, 2004)
56. R.H. Thaler, ed. Advances in Behavioral Finance, Vol. II. (Princeton University, Princeton,
NJ, 2005)
57. M. Schönhof, D. Helbing, Empirical features of congested traffic states and their implications
for traffic modeling. Transport. Sci. 41(2), 135–166 (2007)
58. H.E. Stanley, Introduction to Phase Transitions and Critical Phenomena. (Oxford University,
1987)
59. N. Goldenfeld, Lectures On Phase Transitions and the Renormalization Group. (Westview,
Boulder, CO, 1992)
60. E.C. Zeeman, ed. Catastrophe Theory. (Addison-Wesley, London, 1977)
61. T. Poston, I. Stewart, Catastrophe Theory and Its Applications. (Dover, Mineola, 1996)
62. M. Gladwell, The Tipping Point: How Little Things Can Make a Big Difference. (Back Bay,
Boston, 2002)
63. V.I. Arnol’d, G.S. Wassermann, R.K. Thomas, Catastrophe Theory. (Springer, Berlin, 2004)
64. P. Doreian, F.N. Stokman, eds. Evolution of Social Networks. (Routledge, Amsterdam, 1997)
65. D.J. Watts, Small Worlds. (Princeton University, Princeton, NJ, 1999)
66. B.A. Huberman, The Laws of the Web: Patterns in the Ecology of Information. (MIT Press,
Cambridge, 2001)
67. R. Albert, L.-A. Barabasi, Statistical mechanics of complex networks. Rev. Mod. Phys. 74,
47–97 (2002)
68. S. Bornholdt, H.G. Schuster, Handbook of Graphs and Networks: From the Genome to the
Internet. (Wiley-VCH, Germany 2003)
69. S.N. Dorogovtsev, J.F.F. Mendes, Evolution of Networks. (Oxford University, Oxford 2003)
70. P.J. Carrington, J. Scott, S. Wassermann, Models and Methods in Social Network Analysis.
(Cambridge University, New York, 2005)
71. M. Newman, A.-L. Barabasi, D.J. Watts, eds. The Structure and Dynamics of Networks.
72. F. Vega-Redondo, Complex Social Networks. (Cambridge University, Cambridge, 2007)
73. M.O. Jackson, Social and Economic Networks. (Princeton University, Princeton, NJ, 2008)
74. J. Bruggeman, Social Networks. (Routledge, New York, 2008)
75. A. Barrat, M. Barthélemy, A. Vespignani, Dynamical Processes on Complex Networks.
(Cambridge University, Cambridge, 2008)
76. J. Reichardt, Structure in Complex Networks. (Springer, Berlin, 2009)
77. A.K. Naimzada, S. Stefani, A. Torriero, eds. Networks, Topology, and Dynamics. (Springer,
Berlin, 2009)
78. A. Pyka, A. Scharnhorst, eds. Innovation Networks. (Springer, Berlin, 2009)
79. T. Gross, H. Sayama, eds. Adaptive Networks. (Springer, Berlin, 2009)
80. I. Simonsen, L. Buzna, K. Peters, S. Bornholdt, D. Helbing, Transient dynamics increasing
network vulnerability to cascading failures. Phys. Rev. Lett. 100, 218701 (2008)
81. J. Lorenz, S. Battiston, F. Schweitzer, Systemic risk in a unifying framework for cascading
processes on networks. Eur. Phys. J. B 71(4), 441–460 (2009)
82. D. Helbing, C. Khnert, Assessing interaction networks with applications to catastrophe

dynamics and disaster management. Physica A 328, 584–606 (2003)
83. L. Buzna, K. Peters, D. Helbing, Modelling the dynamics of disaster spreading in networks.
Physica A 363, 132–140 (2006)
84. L. Buzna, K. Peters, H. Ammoser, C. Kühnert, D. Helbing, Efficient response to cascading
disaster spreading. Phys. Rev. E 75, 056107 (2007)
85. D. Helbing, H. Ammoser, C. Khnert, Disasters as extreme events and the importance of
network interactions for disaster response management. In [19], pp. 319–348 (2005)
86. J. Reason, Human Error. (Cambridge University, Cambridge, 1990)
87. J.R. Chiles, Inviting Disaster: Lessons From the Edge of Technology. (Harper, New York,
2002)
88. P. Bak, How nature works: the science of self-organized criticality. (Springer, Berlin, 1999)
89. H.J. Jensen, Self-Organized Criticality: Emergent Complex Behavior in Physical and Biolog-
ical Systems. (Cambridge University, Cambridge, 1998)
1141 (2001)
91. D. de Martino, L. Dall’Asta, G. Bianconi, M. Marsili, Congestion phenomena on complex
networks. Phys. Rev. E 79, 015101 (2009)
92. P.A. Davidson, Turbulence. (Cambridge University, Cambridge, 2004)
93. H.G. Schuster, W. Just, Deterministic Chaos. (Wiley-VCH, Weinheim, 2005)
94. N.G. van Kampen, Stochastic Processes in Physics and Chemistry. (North-Holland,
Amsterdam, 2007)
95. G. Deco, B. Schürmann, Information Dynamics. (Springer, Berlin, 2001)
96. R.A. Jones, Self-fulfilling Prophecies: Social, Psychological, and Physiological Effects of
Expectancies. (Lawrence Erlbaum, 1981)
97. R.E.A. Farmer, Macroeconomics of Self-fulfilling Prophecies. (MIT Press, Cambridge, MA,
1999)
98. J.D. Murray, Mathematical Biology, Vol. I. (Springer, Berlin, 2003)
99. M. Scheffer, J. Bascompte, W.A. Brock, V. Brovkin, S.R. Carpenter, V. Dakos, H. Held, E.H.
van Nes, M. Rietkerk, G. Sugihara, Early-warning signals for critical transitions. Nature 461,
53–59 (2009)
100. W. Michiels, S.-I. Niculescu, Stability and Stabilization of Time-Delay Systems. siam –
Society for Industrial and Applied Mathematics, (Philadelphia, USA, 2007)
101. N.N. Taleb, The Black Swan: The Impact of the Highly Improbable. (Random House,
New York, 2007)
102. E.E. Peters, Chaos and Order in the Capital Markets. (Wiley, New York, 1996)
103. S. Claessens, K.J. Forbes, eds. International Financial Contagion. (Kluwer Academic,
Dordrecht, 2001)
104. K. Ilinsiki, Physics of Finance. Gauge Modelling in Non-Equilibrium Pricing. (Wiley,
Chichester, 2001)
105. B.M. Roehner, Patterns of Speculation. (Cambridge University, Cambridge, 2002)
106. D. Sornette, Why Stock Markets Crash: Critical Events in Complex Financial Systems
107. B. Mandelbrot, R.L. Hudson, The Misbehavior of Markets: A Fractal View of Financial
Turbulence. (Basic, New York, 2006)
108. M. Faggini, T. Lux, eds. Coping with the Complexity of Economics. (Springer, Milan, 2009)
109. R.J. Breiding, M. Christen, D. Helbing Lost robustness. Naissance Newsletter, 8–14 (April
2009)
110. R.M. May, S.A. Levin, G. Sugihara, Ecology for bankers. Nature 451, 893–895 (2008)
111. S. Battiston, D. Delli Gatti, M. Gallegati, B. C.N. Greenwald, J.E. Stiglitz, Liaisons
dangereuses: Increasing connectivity, risk sharing and systemic risk, (2009) see http://www3.
unicatt.it/unicattolica/CentriRicerca/CSCC/allegati/delligatti.pdf
112. I. Kondor, I. Varga-Haszonits, Divergent estimation error in portfolio optimization and in
linear regression. Eur. Phys. J. B 64, 601–605 (2008)
References 283
113. G.A. Akerlof, R.J. Shiller, Animal Spirits. How Human Psychology Drives the Economy,
and Why It Matters for Global Capitalism. (Princeton University, Princeton, NJ, 2009)
114. G. Le Bon, The Crowd: A Study of the Popular Mind. (Dover, New York, 2002)
115. W. Trotter, Instincts of the Herd in Peace and War. (Cosimo Classics, New York, 2005)
116. C. Mackay, Extraordinary popular delusions and the madness of crowds (2003)
117. J. Surowiecki, The Wisdom of Crowds. (Anchor, New York, 2005)
118. P. Ball, Critical Mass. (Arrow, London, 2004)
119. R. Falcone, K.S. Barber, J. Sabater-Mir, M.P. Singh, eds. Trust in Agent Societies. (Springer,
Berlin, 2009)
120. S. Thurner, J.D. Farmer, J. Geanakoplos, Leverage causes fat tails and clustered volatility.
(2009) E-print http://arxiv.org/abs/0908.1555.
121. R. Axelrod, M.D. Cohen, Harnessing Complexity: Organizational Implications of a Scientific
Frontier. (Basic, New York, 2001)
122. H. Eisner, Managing Complex Systems: Thinking Outside the Box (Wiley, New York 2005)
123. L. Hurwicz, S. Reiter, Designing Economic Mechanisms. (Cambridge University, New York,
2006)
124. M. Salzano, D. Colander, eds. Complexity Hints for Economic Policy. (Springer, Berlin, 2007)
125. E. Schöll, H.G. Schuster, eds. Handbook of Chaos Control. (Wiley-VCH, Weinheim, 2008)
126. D. Grass, J.P. Caulkins, G. Feichtinger, G. Tragler, D.A. Behrens, Optimal Control of
Nonlinear Processes: With Applications in Drugs, Corruption, and Terror. (Springer, Berlin,
2008)
127. D. Helbing, eds. Managing Complexity: Insights, Concepts, Applications. (Springer, Berlin,
2008)
128. L.A. Cox Jr., Risk Analysis of Complex and Uncertain Systems. (Springer, New York, 2009)
129. M. Gallegati, A. Kirman, eds. Beyond the Representative Agent. (Edward Elgar, Cheltenham,
UK 1999)
130. A. Consiglio, eds. Artificial Markets Modeling. (Springer, Berlin, 2007)
131. M. Aoki, H. Yoshikawa, Reconstructing Macroeconomics. (Cambridge University Press,
Cambridge, 2007)
132. K. Schredelseker, F. Hauser, eds. Complexity and Artificial Markets. (Springer, Berlin, 2008)
133. D. Delli Gatti, E. Gaffeo, M. Gallegati, G. Giulioni, A. Palestrini, Emergent Macroeconomics.
An Agent-Based Approach to Business Fluctuations. (Springer, Milan, 2008)
134. J. Surowiecki, The Wisdom of Crowds. (Anchor, New York, 2005)
135. H. Rheingold, Smart Mobs. (Perseus, Cambridge, MA, 2003)
136. D. Floreano, C. Mattiussi, Bio-Inspired Artificial Intelligence: Theories, Methods, and
Technologies. (MIT Press, Cambridge, MA, 2008)
137. D. Helbing, A. Deutsch, S. Diez, K. Peters, Y. Kalaidzikis, K. Padberg, S. Lämmer,
A. Johansson, G. Breier, F. Schulze, M. Zerial, BioLogistics and the struggle for efficiency:
Concepts and perspectives. Advances in Complex Systems, (2009); in print, see e-print http://
www.santafe.edu/research/publications/wpabstract/200910041
138. S. Lämmer, D. Helbing, Self-control of traffic lights and vehicle flows in urban road networks.
J. Stat. Phys. (JSTAT), P04019 (2008)
139. K. Windt, T. Philipp, F. Böse, Complexity cube for the characterization of complex production
systems. Int. J. Comput. Integrated Manuf. 21(2), 195–200 (2007)
140. D. Helbing, W. Yu, H. Rauhut, Self-organization and emergence in social systems:
Modeling the coevolution of social environments and cooperative behavior, (2009); see
e-print http://www.santafe.edu/research/publications/wpabstract/200907026
141. E. Ostrom, Governing the Commons. The Evolution of Institutions for Collective Action.
142. G. Hardin, Living within Limits: Ecology, Economics, and Population Taboos. (Oxford
University, New York, 1995)
143. J.A. Baden, D.S. Noonan, eds. Managing the Commons. (Indiana University, Bloomington,
Indiana, 1998)
144. GIACS Complexity Roadmap, see http://users.isi.it/giacs/roadmap
145. The Complex Systems Society Roadmap, see http://www.soms.ethz.ch/research/

complexityscience/roadmap
146. ComplexityNET report on European Research Landscape, see http://www.soms.ethz.ch/
research/complexityscience/European Complexity Landscape D2.2 short report.pdf
147. EU report on Tackling Complexity in Science, see http://www.soms.ethz.ch/research/
complexityscience/EU complexity report.pdf
148. OECD Global Science Forum report on Applications of Complexity Science for Public Policy,
see http://www.oecd.org/dataoecd/44/41/43891980.pdf
149. J.A. Tainter, The Collapse of Complex Societies. (Cambridge University, Cambridge, 1988)
150. J.E. Stiglitz, Globalization and Its Discontents. (Norton, New York, 2003)
151. D. Meadows, J. Randers, D. Meadows, Limits to Growth. The 30-Year Update. (Chelsea Green
Publishing, Withe River Junction, Vermont, 2004)
152. J. Diamond, Collapse. How Societies Choose to Fall or Succeed. (Penguin, New York, 2005)
153. R. Costanza, L.J. Graumlich, W. Steffen, eds. Sustainability of Collapse? (MIT Press,
Cambridge, MA, 2007)
154. P. Krugman, The Return of Depression Economics and the Crisis of 2008. (Norton, New York,
2009)
Chapter 15
Managing Complexity
15.1 What Is Special About Complex Systems?
Many of us have been raised with the idea of cause and effect, i.e. some stimulus-
response theory of the world. Particularly, small causes would have small effects
and large causes would have large effects. This is, in fact, true for “linear systems”,
where cause and effect are proportional to each other. Such behavior is often found
close to the equilibrium state of a system. However, when complex systems are
driven far from equilibrium, non-linearities dominate, which can cause many kinds
of “strange” and counter-intuitive behaviors. In the following, we will mention a
few. We all have been surprised by these behaviors many times.
While linear system have no more than one stationary state (equilibrium) or one
optimal solution, the situation for non-linear systems is different. They can have
multiple stationary solutions or optima (see Fig. 15.1), which has several important
implications:
• The resulting state is history-dependent. Different initial conditions will not
automatically end up in the same state [1]. This is sometimes called “hysteresis”.
• It may be hard to find the best, i.e. the “global” optimum in the potentially
very large set of local optima. Many non-linear optimization problems are “NP
hard”, i.e. the computational time needed to determine the best state tends to
explode with the size of the system [2]. In fact, many optimization problems are
“combinatorially complex”.

This chapter reprints a previous publication with kind permission of the copyright owner,
Springer Publishers. It is requested to cite this work as follows: D. Helbing and S. Lämmer,
Managing complexity: An introduction. Pages 1–16 in D. Helbing (ed.) Managing Complexity
(Springer, Berlin, 2008).

286 15 Managing Complexity
Fig. 15.1 Illustration of linear and non-linear functions. While linear functions have one maxi-
mum in a limited area (left), non-linear functions may have many (local) maxima (right)
Fig. 15.2 Illustration of trajectories that converge towards (a) a stable stationary point, (b) a limit
cycle, and (c) a strange attractor
15.1.1 Chaotic Dynamics and Butterfly Effect
It may also happen that the stationary solutions are unstable, i.e. any small
perturbation will drive the system away from the stationary state until it is attracted
by another state (a so-called “attractor”). Such attractors may be other stationary
solutions, but in many cases, they can be of oscillatory nature (e.g. “limit cycles”).
Chaotically behaving systems [3] are characterized by “strange attractors”, which
are non-periodic (see Fig. 15.2). Furthermore, the slightest change in the trajectory
of a chaotic system (“the beat of a butterfly’s wing”) will eventually lead to a
completely different dynamics. This is often called the “butterfly effect” and makes
the behavior of chaotic systems unpredictable (beyond a certain time horizon), see
Fig. 15.3.
15.1.2 Self-organization, Competition, and Cooperation
Systems with non-linear interactions do not necessarily behave chaotically.

Often, they are characterized by “emergent”, i.e. spontaneous coordination or
15.1 What Is Special About Complex Systems? 287
Fig. 15.3 Illustration of the “butterfly effect”, i.e. the separation of neighboring trajectories in the
course of time
synchronization [4–6]. Even coordinated states, however, may sometimes be

undesired. A typical example for this are stop-and-go waves in freeway traffic [7],
which are a results of an instability of the traffic flow due to the delayed velocity
adjustments of vehicles.
Self-organization is typical in driven many-component systems [7] such as traf-
fic, crowds, organizations, companies, or production plants. Such systems have been
successfully modeled as many-particle or multi-agent systems. Depending on the
respective system, the components are vehicles, individuals, workers, or products
(or their parts). In these systems, the energy input is absorbed by frictional effects.
However, the frictional effect is not homogeneous, i.e. it is not the same everywhere.
It rather depends on the local interactions among the different components of the
system, which leads to spatio-temporal pattern formation.
The example of social insects like ants, bees, or termites shows that simple inter-
actions can lead to complex structures and impressive functions. This is often called
“swarm intelligence” [8]. Swarm intelligence is based on local (i.e. decentralized)
interactions and can be used for the self-organization and self-steering of complex
systems. Some recent examples are traffic assistance [9] systems or self-organized
traffic light control [9, 10]. However, if the interactions are not appropriate, the
system may be characterized by unstable dynamics, breakdowns and jamming, or it
may be trapped in a local optimum (a “frustrated state”).
Many systems are characterized by a competition for scarce resources. Then,
the question whether and how a system optimum is reached is often studied with
methods from “game theory” [11–13]. Instead of reaching the state that maximizes
the overall success, the system may instead converge to a user equilibrium, where
the success (“payoff”) of every system component is the same, but lower than it
could be. This happens, for example, in traffic systems with the consequence of
excess travel times [14]. In conclusion, if everybody tries to reach the best outcome
for him- or herself, this may lead to overall bad results and social dilemmas [15] (the
“tragedy of the commons” [16]). Sometimes, however, the system optimum can only
be reached by complicated coordination in space and/or time, e.g. by suitable turn-
taking behavior (see Fig. 15.4). We will return to this issue in Sect. 15.2.4, when we
discuss the “faster-is-slower” effect.
1
Participant
Decisions
2
1
2
2
0 50 100 150 200 250

Time period t
1
Participant
Decisions
2
1
2
2
0 50 100 150 200 250

Time period t
Fig. 15.4 Emergence of turn-taking behavior: After some time, individuals may learn to improve
their average success by choosing both possible options in an alternating and coordinated way
15.1.3 Phase Transitions and Catastrophe Theory
One typical feature of complex systems is their robustness with respect to pertur-
bations, because the system tends to get back to its “natural state”, the attractor.
However, as mentioned above, many complex system can assume different states.
For this reason, we may have transitions from one system state (“phase” or attractor)
to another one. These phase transitions occur at so-called “critical points” that are
reached by changes of the system parameters (which are often slowly changing
variables of the system). When system parameters come close to critical points,
small fluctuations may become a dominating influence and determine the future
fate of the system. Therefore, one speaks of “critical fluctuations” [1].
In other words, large fluctuations are a sign of a system entering an unstable
regime, indicating its potential transition to another system state, which may be hard
to anticipate. Another indicator of potential instability is “critical slowing down”.
However, once the critical point is passed, the system state may change quite rapidly.
The relatively abrupt change from one system state to an often completely different
one is studied by “catastrophe theory” [17]. One can distinguish a variety of different
types of catastrophes, but we cannot go into all these details, here.
15.1.4 Self-organized Criticality, Power Laws,

and Cascading Effects
At the critical point itself, fluctuations are not only dominating, they may even
become arbitrarily large. Therefore, one often speaks of “scale-free” behavior,
15.1 What Is Special About Complex Systems? 289
political and
economic fiscal problems
problems pogroms crime
black market
trade
evacuation
business public
security solidarity,
human external help
hoarding
resources
hospital maintaining
diseases, automatic teller public law
operation machines, panic
epidemics and order
cash desks coordination
buildings,
facilities
domestic orientation in
technique unknown areas
e.g.heating, disruption of disaster
hygiene anthropogenic systems management
cooker
(public)
sanitation
pressure of
transport system, electric power time, stress
outage (Blackout) detection,
circulation stop
transmission,
selection of
information
fuel supply
telecommunications
Fig. 15.5 Illustration of the interaction network in athropogenic systems. When the system is
seriously challenged, this is likely to cause cascading failures along the arrows of this network
(after [20])
which is typically characterized by power laws [18, 19]. Note that, for power laws,
the variance and the expected value (the average) of a variable may be undefined!
One possible implication of power laws are cascading effects. The classical
example is a sand pile, where more and more grains are added on top [21].
Eventually, when the critical “angle of repose” is reached, one observes avalanches
of sand grains of all possible sizes, and the avalanche size distribution is given by
a power law. The angle of repose, by the way, even determines the stability of the
famous pyramids in Egypt.
Cascading effects are the underlying reason for many disasters, where the failure
of one element of a system causes the failure of another one (see Fig. 15.5). Typical
examples for this dynamics are blackouts of electrical power grids and the spreading
of epidemics, rumors, bankruptcies or congestion patterns. This spreading is often
along the links of the underlying causality or interaction networks [20].
“Self-organized criticality” [21, 22] is a particularly interesting phenomenon,
where a system is driven towards a critical point. This is not uncommon for
economic systems or critical infrastructures: Due to the need to minimize costs,
safety margins will not be chosen higher than necessary. For example, they will
be adjusted to the largest system perturbation that has occurred in the last so-
and-so many years. As a consequence, there will be no failures in a long time.
But then, controllers start to argue that one could safe money by reducing the
standards. Eventually, the safety margins will be low enough to be exceeded by
some perturbation, which may finally trigger a disaster.
Waves of bankruptcies [23, 24] are not much different. The competition for
customers forces companies to make better and better offers, until the profits
have reached a critical value and some companies will die. This will reduce the
competitive pressure among the remaining companies and increase the profits again.
As a consequence, new competitors will enter the market, which eventually drives
the system back to the critical point.
15.2 Some Common Mistakes in the Management

of Complex Systems
The particular features of complex systems have important implications for orga-
nizations, companies, and societies, which are complex multi-component systems
themselves. Their counter-intuitive behaviors result from often very complicated
feedback loops in the system, which cause many management mistakes and unde-
sired side effects. Such effects are particularly well-known from failing political
attempts to improve the social or economic conditions.
15.2.1 The System Does Not Do What You Want It to Do
One of the consequences of the non-linear interactions between the components of a

complex system is that the internal interactions often dominate the external control
attempts (or boundary conditions). This is particularly obvious for group dynamics
[25, 26].
It is quite typical for complex systems that, many times, large efforts have no
significant effect, while sometimes, the slightest change (even a “wrong word”) has
a “revolutionary” impact. This all depends on whether a system is close to a critical
state (which will lead to the latter situation) or not (then, many efforts to change
the system will be in vain). In fact, complex systems often counteract the action.
In chemical systems, this is known as Le Chatelier’s principle.1 Therefore, if it is
necessary to change a system, the right strategy is to drive it to a critical point first.
Then, it will be easy to drive it into a new state, but the potential problem is that the
resulting state is often hard to predict.
Regarding such predictions, classical time series analysis will normally provide
bad forecasts. The problem of opinion polls to anticipate election results when the
mood in the population is changing, is well-known. In many cases, the expectations
of a large number of individuals, as expressed by the stock prices at real or virtual
stock markets, is more indicative than results of classical extrapolation. Therefore,
auction-based mechanisms have been proposed as a new prediction tool. Recently,
1
Specifically, Le Chatelier’s principle says: “If a chemical system at equilibrium experiences a
change in concentration, temperature, or total pressure, the equilibrium will shift in order to
minimize that change.”
15.2 Some Common Mistakes in the Management of Complex Systems 291
there are even techniques to forecast the future with small groups [27]. This,
however, requires to correct for individual biases by fitting certain personality
parameters. These reflect, for example, the degree of risk aversion.
15.2.2 Guided Self-organization Is Better Than Control
The previous section questions the classical control approach, which is, for example,
used to control machines. But it is also frequently applied to business and societies,
when decision-makers attempt to regulate all details by legislation, administrative
procedures, project definitions, etc. These procedures are very complicated and
time-consuming, sensitive to gaps, prone to failures, and they often go along
with unanticipated side effects and costs. However, a complex system cannot be
controlled like a bus, i.e. steering it somewhere may drive it to some unexpected
state.
Biological systems are very differently designed. They do not specify all proce-
dures in detail. Otherwise cells would be much too small to contain all construction
plans in their genetic code, and the brain would be too small to perform its incredible
tasks. Rather than trying to control all details of the system behavior, biology makes
use of the self-organization of complex systems rather than “fighting” it. It guides
self-organization, while forceful control would destroy it [28].
Detailed control would require a large amount of energy, and would need further
resources to put and keep the components of an artificial system together. That
means, overriding the self-organization in the system is costly and inefficient.
Instead, one could use self-organization principles as part of the management plan.
But this requires a better understanding of the natural behavior of complex systems
like companies and societies.
15.2.3 Self-organized Networks and Hierarchies
Hierarchies are a classical way to control systems. However, strict hierarchies are
only optimal under certain conditions. Particularly, they require a high reliability of
the nodes (the staff members) and the links (their exchange).
Experimental results on the problem solving performance of groups [29] show
that small groups can find solutions to difficult problems faster than any of their
constituting individuals, because groups profit from complementary knowledge and
ideas. The actual performance, however, sensitively depends on the organization
of information flows, i.e. on who can communicate with whom. If communication
is unidirectional, for example, this can reduce performance. However, it may
also be inefficient if everybody can talk to everyone else. This is, because the
number of potential (bidirectional) communicative links grows like N.N 1/=2,
where N denotes the number of group members. The number of communicative or
a b c
Fig. 15.6 Illustration of different kinds of hierarchical organization. As there are no alternative
communication links, strict hierarchies are vulnerable to the failure of nodes or links
group-dynamical constellations even grows as .3N 2N C1 C 1/=2. Consequently,

the number of possible information flows explodes with the group size, which may
easily overwhelm the communication and information processing capacity of indi-
viduals. This explains the slow speed of group decision making, i.e. the inefficiency
of large committees. It is also responsible for the fact that, after some transient
time, (communication) activities in large (discussion) groups often concentrate on
a few members only, which is due to a self-organized information bundling and
differentiation (role formation) process. A similar effect is even observed in insect
societies such as bee hives: When a critical colony size is exceeded, a few members
develop hyperactivity, while most colony members become lazy [30].
This illustrates the tendency of bundling and compressing information flows,
which is most pronounced in strict hierarchies. But the performance of strictly
hierarchical organizations (see Fig. 15.6) is vulnerable for the following reasons:
• Hierarchical organizations are not robust with respect to failure of nodes (due
to illness of staff members, holidays, quitting the job) or links (due to difficult
personal relationships).
• They often do not connect interrelated activities in different departments well.
• Important information may get lost due to the filtering of information implied by
the bundling process.
• Important information may arrive late, as it takes time to be communicated over
various hierarchical levels.
Therefore, hierarchical networks with short-cuts are expected to be superior to
strictly hierarchical networks [31–33]. They can profit from alternative information
paths and “small-world” effects [34].
Note that the spontaneous formation of hierarchical structures is not untypical
in social systems: Individuals form groups, which form companies, organizations,
and parties, which make up a society or nation. A similar situation can be found in
biology, where organelles form cells, cells form organs, and organs form bodies.
Another example is well-known from physics, where elementary particles form
nuclei, which combine to atoms with electrons. The atoms form chemical molecules,
which organize themselves as solids. These make up cellestial bodies, which form
solar systems, which again establish galaxies.
Obviously, the non-linear interactions between the different elements of the
system give rise to a formation of different levels, which are hierarchically ordered
one below another. While changes on the lowest hierarchical level are fastest,
changes on the highest level are slow.
On the lowest level, we find the strongest interactions among its elements. This
is obviously the reason for the fast changes on the lowest hierarchical level. If the
interactions are attractive, bonds will arise. These cause the elements to behave no
longer completely individually, but to form units representing the elements of the
next level. Since the attractive interactions are more or less “saturated” by the bonds,
the interactions within these units are stronger than the interactions between them.
The relatively weak residual interactions between the formed units induce their
relatively slow dynamics [35].
In summary, a general interdependence between the interaction strength, the
changing rate, and the formation of hierarchical levels can be found, and the
existence of different hierarchical levels implies a “separation of time scales”.
The management of organizations, production processes, companies, and politi-
cal changes seems to be quite different today: The highest hierarchy levels appear to
take a strong influence on the system on a relatively short time scale. This does not
only require a large amount of resources (administrative overhead). It also makes
it difficult for the lower, less central levels of organization to adjust themselves to
a changing environment. This complicates large-scale coordination in the system
and makes it more costly. Strong interference in the system may even destroy
self-organization in the system instead of using its potentials. Therefore, the re-
structuring of companies can easily fail, in particularly if it is applied too often.
A good example is given in [36].
Governments would be advised to focus their activities on coordination func-
tions, and on adaptations that are relevant for long time scales, i.e. applicable for
100 years or so. Otherwise the individuals will not be able to adjust to the boundary
conditions set by the government. If the government tries to adjust to the population
and the people try to adjust to the socio-economic conditions on the same time scale
of months or years, the control attempts are expected to cause a potentially chaotic
dynamics and a failure of control.
Anyway, detailed regulations hardly ever reach more fairness. They rather reduce
flexibility, and make the anyway required processes inefficient, slow, complicated,
and expensive. As a consequence, many people will not be able to utilize their rights
without external help, while a specialized minority will be able to profit from the
regulations or exploit them.
15.2.4 Faster Is Often Slower
Another common mistake is to push team members to their limits and have machines
run at maximum speed. In many cases, this will not maximize productivity and
throughput, but rather frustration. Most systems require some spare capacity to run
smoothly. This is well illustrated by queuing systems: If the arrival rate reaches the
service rate, the average waiting time will grow enormously. The same applies to
the variation of the waiting time. Jamming and full buffers will be an unfavorable,
but likely side effect. And there will be little reserves in case of additional demand.
The situation becomes even more difficult by dynamic interaction effects, when
a system is driven to its limits. In traffic systems, for example, this leads to a
“capacity drop”. Such a capacity drop occurs often unexpectedly and is a sign
of inefficiencies due to dynamical friction or obstruction effects. It results from
increasing coordination problems when sufficient space or time are lacking. The
consequence is often a “faster-is-slower effect” [38] (see Fig. 15.7). This effect has
been observed in many traffic, production, and logistic systems. Consequently, it is
often not good if everybody is doing his or her best. It is more important to adjust to
the other activities and processes in order to reach a harmonic and well coordinated
overall dynamics. Otherwise, more and more conflicts, inefficiencies and mistakes
will ruin the overall performance.
15.2.5 The Role of Fluctuations and Heterogeneity
Let us finally discuss the role of fluctuations and heterogeneity. Fluctuations are
often considered unfavorable, as they are thought to produce disorder. They can
also trigger instabilities and breakdowns, as is known from traffic flows. But in
some systems, fluctuations can also have positive effects.
While a large fluctuation strength, in fact, tends to destroy order, medium fluc-
tuation levels may even cause a noise-induced ordering (see Fig. 15.8). An eventual
increase in the degree of order in the system is particularly expected if the system
tends to be trapped in local minima (“frustrated states”). Only by means of fluctua-
tions, it is possible to escape these traps and to eventually find better solutions.
Fluctuations are also needed to develop different behavioral roles under initially
identical conditions. This eventually leads to a differentiation and specialization
(heterogeneity), which often helps to reach a better group performance [40] (see
Fig. 15.9).
Furthermore, the speed of evolution also profits from variety and fluctuations
(“mutations”). Uniformity, i.e. if everybody behaves and thinks the same, will lead
to a poor adaptation to changing environmental or market conditions. In contrast, a
large variety of different approaches (i.e. a heterogeneous population) will imply a
large innovation rate [41]. The innovation rate is actually expected to be proportional
to the variance of individual solutions. Therefore, strong norms, “monocultures”,
and the application of identical strategies all over the world due to the trend towards
globalization implies dangers.
This trend is re-inforced by “herding effects” [7]. Whenever the future is hard
to predict, people tend to orient at the behavior of others. This may easily lead to
wrong collective decisions, even of highly intelligent people. This danger can be
only reduced by supporting and maintaining a plurality of opinions and solutions.
Handler / Gripper
Chemical Water Chemical Water Gripper Chemical Water Dryer Input

Park Positions
Process 1 Process 2 Process 3 Process 4 Cleaner Process 5 Process 6 Process 7 Output
Process 4 Process 4
Process 3 Process 3
Process 2 Process 2
Process 1 Process 1
Handler 0 Handler 0
00:00:00
00:05:00
00:10:00
00:15:00
00:20:00
00:25:00
00:30:00
00:35:00
00:40:00
00:00:00
00:05:00
00:10:00
00:15:00
00:20:00
00:25:00
00:30:00
00:35:00
00:40:00
00:45:00
00:50:00
260
250
Optimized Parameters:
240
33.2 % more Throughput
230
220
Throughput [1 / h]
210
200
190
180
170
160
150
140
130
Change of Schedule
120
1.1.01
16.1.01
29.1.01
12.2.01
21.2.01
4.3.01
15.3.01
23.3.01
1.4.01
8.4.01
18.4.01
28.4.01
6.5.01
13.5.01
Time [Days]
Fig. 15.7 Top: Schematic representation of the successive processes of a wet bench, i.e a particular
supply chain in semiconductor production. Middle: The Gantt diagrams illustrate the treatment
times of the first four of several more processes, where we have used the same colors for processes
belonging to the same run, i.e. the same set of wafers. The left diagram shows the original schedule,
while the right one shows an optimized schedule based on the “slower-is-faster effect”. Bottom:
The increase in the throughput of a wet bench by switching from the original production schedule
to the optimized one was found to be 33%, in some cases even higher (after [37])
60
Subpopulation a = 1
Subpopulation a = 2
Number of Individuals
40
20
0
1 5 10 15 20
Spatial Location
60
Number of Individuals
40
20
0
1 5 10 15 20
Fig. 15.8 Illustration of frequency distributions of behaviors in space (after [39]). Top: Separation
of oppositely moving pedestrians perpendicularly to their walking direction for a low fluctuation
strength. Bottom: Noise-induced ordering for medium fluctuation levels leads to a clear separation
into two spatial areas. This reduces frictional effects and increases the efficiency of motion
15.3 Summary and Outlook
In this contribution, we have given a short overview of some properties and

particularities of complex systems. Many of their behaviors may occur unexpectedly
(due to “catastrophies” or phase transitions), and they are often counter-intuitive,
e.g. due to feedback loops and side effects. Therefore, the response of complex
systems to control attempts can be very different from the intended or predicted one.
References 297
6
Test Person
1
0 100 200 300 400 500
Iteration t
Fig. 15.9 Typical individual decision changes of nine test persons in a route choice experiment
with two alternative routes. Note that we find almost similar or opposite behaviors after some time.
The test persons develop a few kinds of complementary strategies (“roles”) in favour of a good
group performance (after [40])
Complex behavior in space and time is found for many multi-component systems
with non-linear interactions. Typical examples are companies, organizations, admin-
istrations, or societies. This has serious implications regarding suitable control
approaches. In fact, most control attempts are destined to fail. It would, however,
be the wrong conclusion that one would just have to apply more force to get control
over the system. This would destroy the self-organization in the system, on which
social systems are based.
We need to obtain a better understanding of how to make use of the natural
tendencies and behaviors at work. A management that supports and guides the
natural self-organization in the system would perform much more efficiently than
an artificially constructed system that requires continuous forcing. Companies and
countries that manage to successfully apply the principle of self-organization will
be the future winners of the on-going global competition.
In conclusion, we are currently facing a paradigm shift in the management of
complex systems, and investments into complexity research will be of competitive
advantage.
References
1. H. Haken, Synergetics (Springer, Berlin, 1977)

2. G. Ausiello, P. Crescenzi, G. Gambosi, et al., Complexity and Approximation – Combinatorial
optimization problems and their approximability properties (Springer, Berlin, 1999)
3. S.H. Strogatz, Nonlinear Dynamics and Chaos: With Applications to Physics, Biology,
Chemistry and Engineering (Perseus, New York, 2001)
4. Y. Kuramoto, Chemical Oscillations, Waves, and Turbulence (Springer, Berlin, 1984)
5. A. Pikovsky, M. Rosenblum, J. Kurths, Synchronization: A Universal Concept in Nonlinear
Sciences (Cambridge University Press, Cambridge, 2003)
6. S.C. Manrubia, A.S. Mikhailov, D.H. Zanette, Emergence of Dynamical Order. Synchroniza-
tion Phenomena in Complex systems (World Scientific, Singapore, 2004)
7. D. Helbing, Traffic and related self-driven many-particle systems. Rev. Mod. Phys. 73, 1067
(2001)
8. E. Bonabeau, M. Dorigo, G. Theraulaz, Swarm Intelligence: From Natural to Artificial
Systems. Santa Fe Institute Studies in the Sciences of Complexity Proceedings (1999)
9. A. Kesting, M. Schönhof, S. Lämmer, M. Treiber, D. Helbing, Decentralized approaches to
adaptive traffic control. In Managing Complexity: Insights, Concepts, Applications ed. by
D. Helbing (Springer, Berlin, 2008)
10. D. Helbing, S. Lämmer, Verfahren zur Koordination konkurrierender Prozesse oder zur
Steuerung des Transports von mobilen Einheiten innerhalb eines Netzwerkes [Method to
Coordinate Competing Processes or to Control the Transport of Mobile Units within a
Network]. Pending patent DE 10 2005 023 742.8 (2005)
12. J. von Neumann, O. Morgenstern, A. Rubinstein, H.W. Kuhn, Theory of Games and Economic
Behavior (Princeton University Press, Princeton, 2004)
13. T.C. Schelling, The Strategy of Conflict (Harvard University Press, Cambridge, 2006)
14. D. Helbing, M. Schönhof, H.-U. Stark, J.A. Holyst, How individuals learn to take turns:
Emergence of alternating cooperation in a congestion game and the prisoner’s dilemma. Adv.
Complex Syst. 8, 87 (2005)
15. N.S. Glance, B.A. Huberman, The dynamics of social dilemmas. Sci. Am. 270, 76 (1994)
16. G. Hardin, The Tragedy of the Commons. Science 162, 1243 (1968)
17. E.C. Zeeman, Catastrophe Theory (Addison-Wesley, London, 1977)
18. H.E. Stanley, Introduction to Phase Transitions and Critical Phenomena (Oxford University
19. M. Schroeder, Fractals, Chaos, Power Laws: Minutes from an Infinite Paradise (Freeman,
New York, 1992)
20. D. Helbing, H. Ammoser, C. Kühnert, Disasters as extreme events and the importance of net-
work interactions for disaster response management, in The Unimaginable and Unpredictable:
Extreme Events in Nature and Society, ed. by S. Albeverio, V. Jentsch, H. Kantz (Springer,
Berlin, 2005), pp. 319–348
21. P. Bak, C. Tang, K. Wiesenfeld, Self-organized criticality: An explanation of 1=f noise. Phys.
Rev. Lett. 59, 381 (1987)
22. P. Bak, How Nature Works: The Science of Self-Organized Criticality (Copernicus, New York,
1996)
23. A. Aleksiejuk, J.A. Hołyst, A simple model of bank bankruptcies. Physica A 299(1-2), 198
(2001)
24. A. Aleksiejuk, J.A. Hołyst, G. Kossinets, Self-organized criticality in a model of collective
bank bankruptcies. Int. J. Mod. Phys. C 13, 333 (2002)
25. S.L. Tubbs, A Systems Approach to Small Group Interaction (McGraw-Hill, Boston, 2003)
26. H. Arrow, J.E. McGrath, J.L. Berdahl, Small Groups as Complex Systems: Formation,
Coordination, Development, and Adaptation (Sage, CA, 2000)
27. K.-Y. Chen, L.R. Fine, B.A. Huberman, Predicting the Future. Inform. Syst. Front. 5, 47 (2003)
28. A.S. Mikhailov, Artificial life: an engineering perspective, in Evolution of Dynamical Struc-
tures in Complex Systems, ed. by R. Friedrich, A. Wunderlin (Springer, Berlin, 1992),
pp. 301–312
29. F-L. Ulschak, Small Group Problem Solving: An Aid to Organizational Effectiveness (Addison-
Wesley Reading Mass., MA, 1981)
References 299
30. J. Gautrais, G. Theraulaz, J.-L. Deneubourg, C. Anderson, Emergent polyethism as a conse-

quence of increased colony size in insect societies. J. Theor. Biol. 215, 363 (2002)
31. D. Helbing, H. Ammoser, C. Kühnert, Information flows in hierarchical networks and the
capability of organizations to successfully respond to failures, crises, and disasters. Physica
A 363, 141 (2006)
32. L.A. Adamic, E. Adar, Friends and neighbors on the web. Social Networks 25(3), 211–230
(2003)
33. D. Stauffer, P.M.C. de Oliveira, Optimization of hierarchical structures of information flow.
Int. J. Mod. Phys. C 17, 1367 (2006)
34. D.J. Watts, S.H. Strogatz, Collective dynamics of smallworld networks. Nature 393, 440 (1998)
35. D. Helbing, Quantitative Sociodynamics, in Stochastic Methods and Models of Social Interac-
tion Processes (Kluwer Academic, Dordrecht, 1995)
36. M. Christen, G. Bongard, A. Pausits, N. Stoop, R. Stoop, Managing autonomy and control
in economic systems. In Managing Complexity: Insights, Concepts, Applications ed. by
D. Helbing (Springer, Berlin, 2008)
37. D. Fasold, Optimierung logistischer Prozessketten am Beispiel einer Nassätzanlage in der
Halbleiterproduktion. MA thesis, TU Dresden (2001)
38. D. Helbing, T. Seidel, S. Lämmer, K. Peters, Self-organization principles in supply networks
and production systems, in Econophysics and Sociophysics - Trends and Perspectives, ed. by
B.K. Chakrabarti, A. Chakraborti, A. Chatterjee (Wiley, Weinheim, 2006), pp. 535–558
39. D. Helbing, T. Platkowski, Self-organization in space and induced by fluctuations. Int. J. Chaos
Theor. Appl. 5, 47–62 (2000)
Models and experiments, in Human Behaviour and Traffic Networks, ed. by M. Schreckenberg,
R. Selten (Springer, Berlin, 2004), pp. 47–95
41. D. Helbing, M. Treiber, N.J. Saam, Analytical investigation of innovation dynamics consider-
ing stochasticity in the evaluation of fitness. Physical Review E 71, 067101 (2005)
Chapter 16
Challenges in Economics
16.1 Introduction
“How did economists get it so wrong?”. Facing the financial crisis, this question
was brilliantly articulated by the Nobel prize winner of 2008, Paul Krugman,
in the New York Times [2]. A number of prominent economists even sees a
failure of academic economics [3]. Remarkably, the following declaration has been
signed by more than 2000 scientists [4]: “Few economists saw our current crisis
coming, but this predictive failure was the least of the field’s problems. More
important was the profession’s blindness to the very possibility of catastrophic
failures in a market economy . . . the economics profession went astray because
economists, as a group, mistook beauty, clad in impressive-looking mathematics,
for truth . . . economists fell back in love with the old, idealized vision of an
economy in which rational individuals interact in perfect markets, this time gussied
up with fancy equations . . . Unfortunately, this romanticized and sanitized vision
of the economy led most economists to ignore all the things that can go wrong.
They turned a blind eye to the limitations of human rationality that often lead to
bubbles and busts; to the problems of institutions that run amok; to the imperfections
of markets—especially financial markets—that can cause the economy’s operating
system to undergo sudden, unpredictable crashes; and to the dangers created when
regulators don’t believe in regulation. . . . When it comes to the all-too-human
problem of recessions and depressions, economists need to abandon the neat but
wrong solution of assuming that everyone is rational and markets work perfectly.”
Apparently, it has not always been like this. DeLisle Worrell writes: “Back in
the sixties . . . we were all too aware of the limitations of the discipline: it was

This chapter reprints part of a previous publication to be cited as: D. Helbing and S. Balietti,
Fundamental and Real-World Challenges in Economics. Science and Culture 76(9/10), 399–417
(2010).

302 16 Challenges in Economics
static where the world was dynamic, it assumed competitive markets where few
existed, it assumed rationality when we knew full well that economic agents
were not rational . . . economics had no way of dealing with changing tastes
and technology . . . Econometrics was equally plagued with intractable problems:
economic observations are never randomly drawn and seldom independent, the
number of excluded variables is always unmanageably large, the degrees of free-
dom unacceptably small, the stability of significance tests seldom unequivocably
established, the errors in measurement too large to yield meaningful results . . . ” [5].
In the following, we will try to identify the scientific challenges that must
be addressed to come up with better theories in the near future. This comprises
practical challenges, i.e. the real-life problems that must be faced (see Sect. 16.2),
and fundamental challenges, i.e. the methodological advances that are required to
solve these problems (see Sect. 16.3). After this, we will discuss, which contribution
can be made by related scientific disciplines such as econophysics and the social
sciences.
The intention of this contribution is constructive. It tries to stimulate a fruitful
scientific exchange, in order to find the best way out of the crisis. According to our
perception, the economic challenges we are currently facing can only be mastered
by large-scale, multi-disciplinary efforts and by innovative approaches [6]. We fully
recognize the large variety of non-mainstream approaches that has been developed
by “heterodox economists”. However, the research traditions in economics seem
to be so powerful that these are not paid much attention to. Besides, there is no
agreement on which of the alternative modeling approaches would be the most
promising ones, i.e. the heterogeneity of alternatives is one of the problems, which
slows down their success. This situation clearly implies institutional challenges
as well, but these go beyond the scope of this contribution and will therefore be
addressed in the future.
16.2 Real-World Challenges
Since decades, if not since hundreds of years, the world is facing a number of
recurrent socio-economic problems, which are obviously hard to solve. Before
addressing related fundamental scientific challenges in economics, we will therefore
point out practical challenges one needs to pay attention to. This basically requires to
classify the multitude of problems into packages of interrelated problems. Probably,
such classification attempts are subjective to a certain extent. At least, the list
presented below differs from the one elaborated by Lomborg et al. [7], who
identified the following top ten problems: air pollution, security/conflict, disease
control, education, climate change, hunger/malnutrition, water sanitation, barriers to
migration and trade, transnational terrorism and, finally, women and development.
The following (non-ranked) list, in contrast, is more focused on socio-economic
factors rather than resource and engineering issues, and it is more oriented at the
roots of problems rather than their symptoms:
16.3 Fundamental Challenges 303
1. Demographic change of the population structure (change of birth rate, migra-

tion, integration. . . )
2. Financial and economic (in)stability (government debts, taxation, and inflation/
deflation; sustainability of social benefit systems; consumption and investment
behavior. . . )
3. Social, economic and political participation and inclusion (of people of
different gender, age, health, education, income, religion, culture, language,
preferences; reduction of unemployment. . . )
4. Balance of power in a multi-polar world (between different countries and
economic centers; also between individual and collective rights, political and
company power; avoidance of monopolies; formation of coalitions; protection
of pluralism, individual freedoms, minorities. . . )
5. Collective social behavior and opinion dynamics (abrupt changes in consumer
behavior; social contagion, extremism, hooliganism, changing values; break-
down of cooperation, trust, compliance, solidarity. . . )
6. Security and peace (organized crime, terrorism, social unrest, independence
movements, conflict, war. . . )
7. Institutional design (intellectual property rights; over-regulation; corruption;
balance between global and local, central and decentral control. . . )
8. Sustainable use of resources and environment (consumption habits, travel
behavior, sustainable and efficient use of energy and other resources, partici-
pation in recycling efforts, environmental protection. . . )
9. Information management (cyber risks, misuse of sensitive data, espionage,
violation of privacy; data deluge, spam; education and inheritance of culture. . . )
10. Public health (food safety; spreading of epidemics [flu, SARS, H1N1, HIV],
obesity, smoking, or unhealthy diets. . . )
Some of these challenges are interdependent.
16.3 Fundamental Challenges
In the following, we will try to identify the fundamental theoretical challenges that
need to be addressed in order to understand the above practical problems and to
draw conclusions regarding possible solutions.
The most difficult part of scientific research is often not to find the right answer.
The problem is to ask the right questions. In this context it can be a problem that
people are trained to think in certain ways. It is not easy to leave these ways and see
the problem from a new angle, thereby revealing a previously unnoticed solution.
Three factors contribute to this:
1. We may overlook the relevant facts because we have not learned to see them, i.e.
we do not pay attention to them. The issue is known from internalized norms,
which prevent people from considering possible alternatives.
2. We know the stylized facts, but may not have the right tools at hand to interpret
them. It is often difficult to make sense of patterns detected in data. Turning data
into knowledge is quite challenging.
3. We know the stylized facts and can interpret them, but may not take them
seriously enough, as we underestimate their implications. This may result from
misjudgements or from herding effects, i.e. from a tendency to follow traditions
and majority opinions. In fact, most of the issues discussed below have been
pointed out before, but it seems that this did not have an effect on mainstream
economics so far or on what decision-makers know about economics. This is
probably because mainstream theory has become a norm [8], and alternative
approaches are sanctioned as norm-deviant behavior [9, 10].
As we will try to explain, the following fundamental issues are not just a matter of
approximations (which often lead to the right understanding, but wrong numbers).
Rather they concern fundamental errors in the sense that certain conclusions
following from them are seriously misleading. As the recent financial crisis has
demonstrated, such errors can be very costly. However, it is not trivial to see what
dramatic consequences factors such as dynamics, spatial interactions, randomness,
non-linearity, network effects, differentiation and heterogeneity, irreversibility or
irrationality can have.
16.3.1 Homo Economicus
Despite of criticisms by several nobel prize winners such as Reinhard Selten

(1994), Joseph Stiglitz and George Akerlof (2001), or Daniel Kahneman (2002), the
paradigm of the homo economicus, i.e. of the “perfect egoist”, is still the dominating
approach in economics. It assumes that people would have quasi-infinite memory
and processing capacities and would determine the best one among all possible
alternative behaviors by strategic thinking (systematic utility optimization), and
would implement it into practice without mistakes. The Nobel prize winner of 1976,
Milton Friedman, supported the hypothesis of homo economicus by the following
argument: “irrational agents will lose money and will be driven out of the market
by rational agents” [11]. More recently, Robert E. Lucas Jr., the Nobel prize winner
of 1995, used the rationality hypothesis to narrow down the class of empirically
relevant equilibria [12].
The rational agent hypothesis is very charming, as its implications are clear and
it is possible to derive beautiful and powerful economic theorems and theories from
it. The best way to illustrate homo economicus is maybe a company that is run
by using optimization methods from operation research, applying supercomputers.
Another example are professional chess players, who are trying to anticipate the
possible future moves of their opponents. Obviously, in both examples, the future
course of actions can not be fully predicted, even if there are no random effects and
mistakes.
It is, therefore, no wonder that people have repeatedly expressed doubts regarding
the realism of the rational agent approach [13, 14]. Bertrand Russell, for example,
claimed: “Most people would rather die than think”. While this seems to be a rather
extreme opinion, the following scientific arguments must be taken seriously:
1. Human cognitive capacities are bounded [16, 17]. Already phone calls or
conversations can reduce people’s attention to events in the environment a lot.
Also, the abilities to memorize facts and to perform complicated logical analyses
are clearly limited.
2. In case of NP-hard optimization problems, even supercomputers are facing lim-
its, i.e. optimization jobs cannot be performed in real-time anymore. Therefore,
approximations or simplifications such as the application of heuristics may be
necessary. In fact, psychologists have identified a number of heuristics, which
people use when making decisions [18].
3. People perform strategic thinking mainly in important new situations. In normal,
everyday situation, however, they seem to pursue a satisficing rather than
optimizing strategy [17]. Meeting a certain aspiration level rather than finding
the optimal strategy can save time and energy spent on problem solving. In many
situation, people even seem to perform routine choices [14], for example, when
evading other pedestrians in counterflows.
4. There is a long list of cognitive biases which question rational behavior [19]. For
example, individuals are favorable of taking small risks (which are preceived as
“chances”, as the participation in lotteries shows), but they avoid large risks [20].
Furthermore, non-exponential temporal discounting may lead to paradoxical
behaviors [21] and requires one to rethink, how future expectations must be
modeled.
5. Most individuals have a tendency towards other-regarding behavior and fairness
[22, 23]. For example, the dictator game [24] and other experiments [25] show
that people tend to share, even if there is no reason for this. Leaving a tip for the
waiter in a restaurant that people visit only once is a typical example (particularly
in countries where tipping is not common) [26]. Such behavior has often been
interpreted as sign of social norms. While social norms can certainly change
the payoff structure, it has been found that the overall payoffs resulting from
them do not need to create a user or system optimum [27–29]. This suggests
that behavioral choices may be irrational in the sense of non-optimal. A typical
example is the existence of unfavorable norms, which are supported by people
although nobody likes them [30].
6. Certain optimization problems can have an infinite number of local optima or
Nash equilibria, which makes it impossible to decide what is the best strategy
[31].
7. Convergence towards the optimal solution may require such a huge amount
of time that the folk theorem becomes useless. This can make it practically
impossible to play the best response strategy [32].
8. The optimal strategy may be deterministically chaotic, i.e. sensitive to arbitrarily

small details of the initial condition, which makes the dynamic solution unpre-
dictable on the long run (“butterfly effect”) [33, 34]. This fundamental limit of
predictability also implies a limit of control—two circumstances that are even
more true for non-deterministic systems with a certain degree of randomness.
In conclusion, although the rational agent paradigm (the paradigm of homo eco-
nomicus) is theoretically powerful and appealing, there are a number of empirical
and theoretical facts, which suggest deficiencies. In fact, most methods used in
financial trading (such as technical analysis) are not well compatible with the
rational agent approach. Even if an optimal solution exists, it may be undecidable
for practical reasons or for theoretical ones [35, 36]. This is also relevant for the
following challenges, as boundedly rational agents may react inefficently and with
delays, which questions the efficient market hypothesis, the equilibrium paradigm,
and other fundamental concepts, calling for the consideration of spatial, network,
and time-dependencies, heterogeneity and correlations etc. It will be shown that
these points can have dramatic implications regarding the predictions of economic
models.
16.3.2 The Efficient Market Hypothesis
The efficient market hypothesis (EMH) was first developed by Eugene Fama [37]
in his Ph.D. thesis and rapidly spread among leading economists, who used it as
an argument to promote laissez-faire policies. The EMH states that current prices
reflect all publicly available information and (in its stronger formulation) that prices
instantly change to reflect new public information.
The idea of self-regulating markets goes back to Adam Smith [38], who believed
that “the free market, while appearing chaotic and unrestrained, is actually guided
to produce the right amount and variety of goods by a so-called “invisible hand”.”
Furthermore, “by pursuing his own interest, [the individual] frequently promotes
that of the society more effectually than when he intends to promote it” [39].
For this reason, Adam Smith is often considered to be the father of free market
economics. Curiously enough, however, he also wrote a book on “The Theory of
Moral Sentiments” [40]. “His goal in writing the work was to explain the source of
mankind’s ability to form moral judgements, in spite of man’s natural inclinations
towards self-interest. Smith proposes a theory of sympathy, in which the act of
observing others makes people aware of themselves and the morality of their own
behavior . . . [and] seek the approval of the “impartial spectator” as a result of
a natural desire to have outside observers sympathize with them” [38]. Such a
reputation-based concept would be considered today as indirect reciprocity [41].
Of course, there are criticisms of the efficient market hypothesis [42], and the
Nobel prize winner of 2001, Joseph Stiglitz, even believes that “There is not
invisible hand” [43]. The following list gives a number of empirical and theoretical
arguments questioning the efficient market hypothesis:
1. Examples of market failures are well-known and can result, for example, in cases
of monopolies or oligopolies, if there is not enough liquidity or if information
symmetry is not given.
2. While the concept of the “invisible hand” assumes something like an optimal
self-organization [44], it is well-known that this requires certain conditions,
such as symmetrical interactions. In general, however, self-organization does not
necessarily imply system-optimal solutions. Stop-and-go traffic [45] or crowd
disasters [46] are two obvious examples for systems, in which individuals
competitively try to reach individually optimal outcomes, but where the optimal
solution is dynamically unstable.
3. The limited processing capacity of boundedly rational individuals implies poten-
tial delays in their responses to sensorial inputs, which can cause such insta-
bilities [47]. For example, a delayed adaptation in production systems may
contribute to the occurrence of business cycles [48]. The same applies to the labor
market of specially skilled people, which cannot adjust on short time scales. Even
without delayed reactions, however, the competitive optimization of individuals
can lead to suboptimal individual results, as the “tragedy of the commons” in
public goods dilemmas demonstrates [49, 50].
4. Bubbles and crashes, or more generally, extreme events in financial markets
should not occur, if the efficient market hypothesis was correct (see next
subsection).
5. Collective social behavior such as “herding effects” as well as deviations of
human behavior from what is expected from rational agents can lead to such
bubbles and crashes [51], or can further increase their size through feedback
effects [52]. Cyclical feedbacks leading to oscillations are also known from the
beer game [53] or from business cycles [48].
16.3.3 Equilibrium Paradigm
The efficient market paradigm implies the equilibrium paradigm. This becomes
clear, if we split it up into its underlying hypotheses:
1. The market can be in equilibrium, i.e. there exists an equilibrium.
2. There is one and only one equilibrium.
3. The equilibrium is stable, i.e. any deviations from the equilibrium due to
“fluctuations” or “perturbations” tend to disappear eventually.
4. The relaxation to the equilibrium occurs at an infinite rate.
Note that, in order to act like an “invisible hand”, the stable equilibrium (Nash equi-
librium) furthermore needs to be a system optimum, i.e. to maximize the average
utility. This is true for coordination games, when interactions are well-mixed and
exploration behavior as well as transaction costs can be neglected [54]. However, it
is not fulfilled by so-called social dilemmas [49].
Let us discuss the evidence for the validity of the above hypotheses one by one:
1. A market is a system of extremely many dynamically coupled variables. Theoret-
ically, it is not obvious that such a system would have a stationary solution. For
example, the system could behave periodic, quasi-periodic, chaotic, or turbulent
[81–83, 85–87, 94]. In all these cases, there would be no convergence to a
stationary solution.
2. If a stationary solution exists, it is not clear that there are no further stationary
solutions. If many variables are non-linearly coupled, the phenomenon of multi-
stability can easily occur [55]. That is, the solution to which the system converges
may not only depend on the model parameters, but also on the initial condition,
history, or perturbation size. Such facts are known as path-dependencies or
hysteresis effects and are usually visualized by so-called phase diagrams [56].
3. In systems of non-linearly interacting variables, the existence of a stationary
solution does not necessarily imply that it is stable, i.e. that the system will
converge to this solution. For example, the stationary solution could be a focal
point with orbiting solutions (as for the classical Lotka-Volterra equations [57]),
or it could be unstable and give rise to a limit cycle [58] or a chaotic solution [33],
for example (see also item 1). In fact, experimental results suggest that volatility
clusters in financial markets may be a result of over-reactions to deviations from
the fundamental value [59].
4. An infinite relaxation rate is rather unusual, as most decisions and related
implemenations take time [15, 60].
The points listed in the beginning of this subsection are also questioned by empirical
evidence. In this connection, one may mention the existence of business cycles
[48] or unstable orders and deliveries observed in the experimental beer game [53].
Moreover, bubbles and crashes have been found in financial market games [61].
Today, there seems to be more evidence against than for the equilibrium paradigm.
In the past, however, most economists assumed that bubbles and crashes would
not exist (and many of them still do). The following quotes are quite typical for
this kind of thinking (from [62]): In 2004, the Federal Reserve chairman of the
U.S., Alan Greenspan, stated that the rise in house values was “not enough in our
judgment to raise major concerns”. In July 2005 when asked about the possibility
of a housing bubble and the potential for this to lead to a recession in the future,
the present U.S. Federal Reserve chairman Ben Bernanke (then Chairman of the
Council of Economic Advisors) said: “It’s a pretty unlikely possibility. We’ve never
had a decline in housing prices on a nationwide basis. So, what I think is more likely
is that house prices will slow, maybe stabilize, might slow consumption spending a
bit. I don’t think it’s going to drive the economy too far from it’s full path though.”
As late as May 2007 Bernanke stated that the Federal Reserve “do not expect
significant spillovers from the subprime market to the rest of the economy”.
According to the classical interpretation, sudden changes in stock prices
result from new information, e.g. from innovations (“technological shocks”). The
dynamics in such systems has, for example, been described by the method of
comparative statics (i.e. a series of snapshots). Here, the system is assumed to be in
equilibrium in each moment, but the equilibrium changes adiabatically (i.e. without
delay), as the system parameters change (e.g. through new facts). Such a treatment
of system dynamics, however, has certain deficiencies:
1. The approach cannot explain changes in or of the system, such as phase
transitions (“systemic shifts”), when the system is at a critical point (“tipping
point”).
2. It does not allow one to understand innovations and other changes as results of
an endogeneous system dynamics.
3. It cannot describe effects of delays or instabilities, such as overshooting,
self-organization, emergence, systemic breakdowns or extreme events (see
Sect. 16.3.4).
4. It does not allow one to study effects of different time scales. For example, when
there are fast autocatalytic (self-reinfocing) effects and slow inhibitory effects,
this may lead to pattern formation phenomena in space and time [63, 64]. The
formation of settlements, where people agglomerate in space, may serve as an
example [65, 66].
5. It ignores long-term correlations such as memory effects.
6. It neglects frictional effects, which are often proportional to change (“speed”)
and occur in most complex systems. Without friction, however, it is difficult to
understand entropy and other path-dependent effects, in particular irreversibility
(i.e. the fact that the system may not be able to get back to the previous state)
[67]. For example, the unemployment rate has the property that it does not go
back to the previous level in most countries after a business cycle [68].
16.3.4 Prevalence of Linear Models
Comparative statics is, of course, not the only method used in economics to describe
the dynamics of the system under consideration. As in physics and other fields, one
may use a linear approximation around a stationary solution to study the response
of the system to fluctuations or perturbations [69]. Such a linear stability analysis
allows one to study, whether the system will return to the stationary solution (which
is the case for a stable [Nash] equilibrium) or not (which implies that the system
will eventually be driven into a new state or regime).
In fact, the great majority of statistical analyses use linear models to fit empirical
data (also when they do not involve time-dependencies). It is know, however,
that linear models have special features, which are not representative for the rich
variety of possible functional dependencies, dynamics, and outcomes. Therefore,
the neglection of non-linearity has serious consequences:
1. As it was mentioned before, phenomena like multiple equilibria, chaos or
turbulence cannot be understood by linear models. The same is true for self-
organization phenomena or emergence. Additionally, in non-linearly coupled
systems, usually “more is different”, i.e. the system may change its behavior
fundamentally as it grows beyond a certain size. Furthermore, the system is often

hard to predict and difficult to control (see Sect. 16.3.8).
2. Linear modeling tends to overlook that a strong coupling of variables, which
would show a normally distributed behavior in separation, often leads to fat tail
distributions (such as “power laws”) [70, 71]. This implies that extreme events
are much more frequent than expected according to a Gaussian distribution. For
example, when additive noise is replaced by multiplicative noise, a number of
surprising phenomena may result, including noise-induced transitions [72] or
directed random walks (“ratchet effects”) [73].
3. Phenomena such as catastrophes [74] or phase transition (“system shifts”) [75]
cannot be well understood within a linear modeling framework. The same applies
to the phenomenon of “self-organized criticality” [79] (where the system drives
itself to a critical state, typically with power-law characteristics) or cascading
effects, which can result from network interactions (overcritically challenged
network nodes or links) [77,78]. It should be added that the relevance of network
effects resulting from the on-going globalization is often underestimated. For
example, “the stock market crash of 1987, began with a small drop in prices
which triggered an avalanche of sell orders in computerized trading programs,
causing a further price decline that triggered more automatic sales.” [80]
Therefore, while linear models have the advantage of being analytically solvable,
they are often unrealistic. Studying non-linear behavior, in contrast, often requires
numerical computational approaches. It is likely that most of today’s unsolved
economic puzzles cannot be well understood through linear models, no matter how
complicated they may be (in terms of the number of variables and parameters) [81–
94]. The following list mentions some areas, where the importance of non-linear
interdependencies is most likely underestimated:
• Collective opinions, such as trends, fashions, or herding effects.
• The success of new (and old) technologies, products, etc.
• Cultural or opinion shifts, e.g. regarding nuclear power, genetically manipulated
food, etc.
• The “fitness” or competitiveness of a product, value, quality perceptions, etc.
• The respect for copyrights.
• Social capital (trust, cooperation, compliance, solidarity, . . . ).
• Booms and recessions, bubbles and crashes.
• Bank panics.
• Community, cluster, or group formation.
• Relationships between different countries, including war (or trade war) and
peace.
16.3.5 Representative Agent Approach
Another common simplification in economic modeling is the representative agent

approach, which is known in physics as mean field approximation. Within this
framework, time-dependencies and non-linear dependencies are often considered,

but it is assumed that the interaction with other agents (e.g. of one company with all
the other companies) can be treated as if this agent would interact with an average
agent, the “representative agent”.
Let us illustrate this with the example of the public goods dilemma. Here,
everyone can decide whether to make an individual contribution to the public good
or not. The sum of all contributions is multiplied by a synergy factor, reflecting the
benefit of cooperation, and the resulting value is equally shared among all people.
The prediction of the representative agent approach is that, due to the selfishness of
agents, a “tragedy of the commons” would result [49]. According to this, everybody
should free-ride, i.e. nobody should make a contribution to the public good and
nobody would gain anything. However, if everybody would contribute, everybody
could multiply his or her contribution by the synergy factor. This example is
particularly relevant as society is facing a lot of public goods problems and would
not work without cooperation. Everything from the creation of public infrastructures
(streets, theaters, universities, libraries, schools, the World Wide Web, Wikipedia
etc.) over the use of environmental resources (water, forests, air, etc.) or of social
benefit systems (such as a public health insurance), maybe even the creation and
maintainance of a commonly shared language and culture are public goods problems
(although the last examples are often viewed as coordination problems). Even the
process of creating public goods is a public good [95].
While it is a well-known problem that people tend to make unfair contributions to
public goods or try to get a bigger share of them, individuals cooperate much more
than one would expect according to the representative agent approach. If they would
not, society could simply not exist. In economics, one tries to solve the problem
by introducing taxes (i.e. another incentive structure) or a “shadow of the future”
(i.e. a strategic optimization over infinite time horizons in accordance with the
rational agent approach) [96,97]. Both comes down to changing the payoff structure
in a way that transforms the public good problem into another one that does not
constitute a social dilemma [98]. However, there are other solutions of the problem.
When the realm of the mean-field approximation underlying the representative agent
approach is left and spatial or network interactions or the heterogeneity among
agents are considered, a miracle occurs: Cooperation can survive or even thrive
through correlations and co-evolutionary effects [99–101].
A similar result is found for the public goods game with costly punishment.
Here, the representative agent model predicts that individuals avoid to invest into
punishment, so that punishment efforts eventually disappear (and, as a consequence,
cooperation as well). However, this “second-order free-rider problem” is naturally
resolved and cooperation can spread, if one discards the mean-field approximation
and considers the fact that interactions take place in space or social networks [56].
Societies can overcome the tragedy of the commons even without transforming
the incentive structure through taxes. For example, social norms as well as group
dynamical and reputation effects can do so [102]. The representative agent approach
implies just the opposite conclusion and cannot well explain the mechanisms on
which society is built.
It is worth pointing out that the relevance of public goods dilemmas is probably
underestimated in economics. Partially related to Adam Smith’s belief in an
“invisible hand”, one often assumes underlying coordination games and that they
would automatically create a harmony between an individually and system optimal
state in the course of time [54]. However, running a stable financial system and
economy is most likely a public goods problem. Considering unemployment,
recessions always go along with a breakdown of solidarity and cooperation. Efficient
production clearly requires mutual cooperation (as the counter-example of countries
with many strikes illustrates). The failure of the interbank market when banks stop
lending to each other, is a good example for the breakdown of both, trust and
cooperation. We must be aware that there are many other systems that would not
work anymore, if people would lose their trust: electronic banking, e-mail and
internet use, Facebook, eBusiness and eGovernance, for example. Money itself
would not work without trust, as bank panics and hyperinflation scenarios show.
Similarly, cheating customers by selling low-quality products or selling products
at overrated prices, or by manipulating their choices by advertisements rather than
informing them objectively and when they want, may create profits on the short
run, but it affects the trust of customers (and their willingness to invest). The failure
of the immunization campaign during the swine flu pandemics may serve as an
example. Furthermore, people would probably spend more money, if the products
of competing companies were better compatible with each other. Therefore, on the
long run, more cooperation among companies and with the customers would pay off
and create additional value.
Besides providing a misleading picture of how cooperation comes about, there
are a number of other deficiencies of the representative agent approach, which are
listed below:
1. Correlations between variables are neglected, which is acceptable only for
“well-mixing” systems. According to what is known from critical phenomena
in physics, this approximation is valid only, when the interactions take place
in high-dimensional spaces or if the system elements are well connected.
(However, as the example of the public goods dilemma showed, this case does not
necessarily have beneficial consequences. Well-mixed interactions could rather
cause a breakdown of social or economic institutions, and it is conceivable that
this played a role in the recent financial crisis.)
2. Percolation phenomena, describing how far an idea, innovation, technology, or
(computer) virus spreads through a social or business network, are not well
reproduced, as they depend on details of the network structure, not just on the
average node degree [103].
3. The heterogeneity of agents is ignored. For this reason, factors underlying
economic exchange, perturbations, or systemic robustness [104] cannot be well
described. Moreover, as socio-economic differentiation and specialization imply
heterogeneity, they cannot be understood as emergent phenomena within a
representative agent approach. Finally, it is not possible to grasp innovation
without the consideration of variability. In fact, according to evolutionary theory,
the innovation rate would be zero, if the variability was zero [105]. Furthermore,
in order to explain innovation in modern societies, Schumpeter introduced the
concept of the “political entrepreneur” [106], an extra-ordinarily gifted person
capable of creating disruptive change and innovation. Such an extraordinary
individual can, by definition, not be modeled by a “representative agent”.
One of the most important drawbacks of the representative agent approach is that it
cannot explain the fundamental fact of economic exchange, since it requires one to
assume a heterogeneity in resources or production costs, or to consider a variation
in the value of goods among individuals. Ken Arrow, Nobel prize winner in 1972,
formulated this point as follows [107]: “One of the things that microeconomics
teaches you is that individuals are not alike. There is heterogeneity, and probably
the most important heterogeneity here is heterogeneity of expectations. If we didn’t
have heterogeneity, there would be no trade.”
We close this section by mentioning that economic approaches, which go beyond
the representative agent approach, can be found in Refs. [108, 109].
16.3.6 Lack of Micro-Macro Link and Ecological

Systems Thinking
Another deficiency of economic theory that needs to be mentioned is the lack of

a link between micro- and macroeconomics. Neoclassical economics implicitly
assumes that individuals make their decisions in isolation, using only the infor-
mation received from static market signals. Within this oversimplified framework,
macro-aggregates are just projections of some representative agent behavior, instead
of the outcome of complex interactions with asymmetric information among a
myriad of heterogeneous agents.
In principle, it should be understandable how the macroeconomic dynamics
results from the microscopic decisions and interactions on the level of producers
and consumers [81, 110] (as it was possible in the past to derive micro-macro
links for other systems with a complex dynamical behavior such as interactive
vehicle traffic [111]). It should also be comprehensible how the macroscopic
level (the aggregate econonomic situation) feeds back on the microscopic level
(the behavior of consumers and producers), and to understand the economy as a
complex, adaptive, self-organizing system [112, 113]. Concepts from evolutionary
theory [114] and ecology [115] appear to be particularly promising [116]. This,
however, requires a recognition of the importance of heterogeneity for the system
(see the the previous subsection).
The lack of ecological thinking implies not only that the sensitive network
interdependencies between the various agents in an economic system (as well
as minority solutions) are not properly valued. It also causes deficiencies in the
development and implementation of a sustainable economic approach based on
recycling and renewable resources. Today, forestry science is probably the best
developed scientific discipline concerning sustainability concepts [117]. Economic

growth to maintain social welfare is a serious misconception. From other scientific
disciplines, it is well known that stable pattern formation is also possible for a
constant (and potentially sustainable) inflow of energy [69, 118].
16.3.7 Optimization of System Performance
One of the great achievements of economics is that it has developed a multitude

of methods to use scarce resources efficiently. A conventional approach to this is
optimization. In principle, there is nothing wrong about this approach. Nevertheless,
there are a number of problems with the way it is usually applied:
1. One can only optimize for one goal at a time, while usually, one needs to meet
several objectives. This is mostly addressed by weighting the different goals
(objectives), by executing a hierarchy of optimization steps (through ranking
and prioritization), or by applying a satisficing strategy (requiring a minimum
performance for each goal) [119, 120]. However, when different optimization
goals are in conflict with each other (such as maximizing the throughput and
minimizing the queue length in a production system), a sophisticated time-
dependent strategy may be needed [121].
2. There is no unique rule what optimization goal should be chosen. Low costs?
High profit? Best customer satisfaction? Large throughput? Competitive advan-
tage? Resilience? [122] In fact, the choice of the optimization function is
arbitrary to a certain extent and, therefore, the result of optimization may vary
largely. Goal selection requires strategic decisions, which may involve normative
or moral factors (as in politics). In fact, one can often observe that, in the
course of time, different goal functions are chosen. Moreover, note that the
maximization of certain objectives such as resilience or “fitness” depends not
only on factors that are under the control of a company. Resilience and “fitness”
are functions of the whole system, in particularly, they also depend on the
competitors and the strategies chosen by them.
3. The best solution may be the combination of two bad solutions and may,
therefore, be overlooked. In other words, there are “evolutionary dead ends”, so
that gradual optimization may not work. (This problem can be partially overcome
by the application of evolutionary mechanisms [120]).
4. In certain systems (such as many transport, logistic, or production systems),
optimization tends to drive the system towards instability, since the point of
maximum efficiency is often in the neighborhood or even identical with the point
of breakdown of performance. Such breakdowns in capacity or performance can
result from inefficiencies due to dynamic interaction effects. For example, when
traffic flow reaches its maximum capacity, sooner or later it breaks down. As a
consequence, the road capacity tends to drop during the time period where it is
most urgently needed, namely during the rush hour [45, 123].
5. Optimization often eliminates reduncancies in the system and, thereby, increases

the vulnerability to perturbations, i.e. it decreases robustness and resilience.
6. Optimization tends to eliminate heterogeneity in the system [80], while hetero-
geneity frequently supports adaptability and resilience.
7. Optimization is often performed with centralized concepts (e.g. by using super-
computers that process information collected all over the system). Such cen-
tralized systems are vulnerable to disturbances or failures of the central control
unit. They are also sensitive to information overload, wrong selection of control
parameters, and delays in adaptive feedback control. In contrast, decentralized
control (with a certain degree of autonomy of local control units) may perform
better, when the system is complex and composed of many heterogeneous
elements, when the optimization problem is NP hard, the degree of fluctuations is
large, and predictability is restricted to short time periods [77, 124]. Under such
conditions, decentralized control strategies can perform well by adaptation to the
actual local conditions, while being robust to perturbations. Urban traffic light
control is a good example for this [121, 125].
8. Further on, today’s concept of quality control appears to be awkward. It leads to
a never-ending contest, requiring people and organizations to fulfil permanently
increasing standards. This leads to over-emphasizing measured performance
criteria, while non-measured success factors are neglected. The engagement into
non-rewarded activities is discouraged, and innovation may be suppressed (e.g.
when evaluating scientists by means of their h-index, which requires them to
focus on a big research field that generates many citations in a short time).
While so-called “beauty contests” are considered to produce the best results,
they will eventually absorb more and more resources for this contest, while
less and less time remains for the work that is actually to be performed, when
the contest is won. Besides, a large number of competitors have to waste
considerable resources for these contests which, of course, have to be paid by
someone. In this way, private and public sectors (from physicians over hospitals,
administrations, up to schools and universities) are aching under the evaluation-
related administrative load, while little time remains to perform the work that
the corresponding experts have been trained for. It seems naı̈ve to believe that
this would not waste resources. Rather than making use of individual strengths,
which are highly heterogeneous, today’s way of evaluating performance enforces
a large degree of conformity.
There are also some problems with parameter fitting, a method based on opti-
mization as well. In this case, the goal function is typically an error function or
a likelihood function. Not only are calibration methods often “blindly” applied
in practice (by people who are not experts in statistics), which can lead to
overfitting (the fitting of meaningless “noise”), to the neglection of collinearities
(implying largely variable parameter values), or to inaccurate and problematic
parameter determinations (when the data set is insufficient in size, for example,
when large portfolios are to be optimized [126]). As estimates for past data are
not necessarily indicative for the future, making predictions with interpolation
approaches can be quite problematic (see also Sect. 16.3.3 for the challenge of time
dependence). Moreover, classical calibration methods do not reveal inappropriate
model specifications (e.g. linear ones, when non-linear models would be needed,
or unsuitable choices of model variables). Finally, they do not identify unknown
unknowns (i.e. relevant explanatory variables, which have been overlooked in the
modeling process).
16.3.8 Control Approach
Managing economic systems is a particular challenge, not only for the reasons
discussed in the previous section. As large economic systems belong to the class of
complex systems, they are hard or even impossible to manage with classical control
approaches [76, 77].
Complex systems are characterized by a large number of system elements
(e.g. individuals, companies, countries, . . . ), which have non-linear or network
interactions causing mutual dependencies and responses. Such systems tend to
behave dynamic rather than static and probabilistic rather than deterministic. They
usually show a rich, hardly predictable, and sometimes paradoxical system behavior.
Therefore, they challenge our way of thinking [127], and their controllability is
often overestimated (which is sometimes paraphrased as “illusion of control”)
[80,128,129]. In particular, causes and effects are typically not proportional to each
other, which makes it difficult to predict the impact of a control attempt.
A complex system may be unresponsive to a control attempt, or the latter may
lead to unexpected, large changes in the system behavior (so-called “phase transi-
tions”, “regime shifts”, or “catastrophes”) [75]. The unresponsiveness is known as
principle of Le Chatelier or Goodhart’s law [130], according to which a complex
system tends to counteract external control attempts. However, regime shifts can
occur, when the system gets close to so-called “critical points” (also known as
“tipping points”). Examples are sudden changes in public opinion (e.g. from pro to
anti-war mood, from a smoking tolerance to a public smoking ban, or from buying
energy-hungry sport utilities vehicles (SUVs) to buying environmentally-friendly
cars).
Particularly in case of network interactions, big changes may have small, no,
or unexpected effects. Feedback loops, unwanted side effects, and circuli vitiosi are
quite typical. Delays may cause unstable system behavior (such as bull-whip effects)
[53], and over-critical perturbations can create cascading failures [78]. Systemic
breakdowns (such as large-scale blackouts, bankruptcy cascades, etc.) are often a
result of such domino or avalanche effects [77], and their probability of occurrence
as well as their resulting damage are usually underestimated. Further examples are
epidemic spreading phenomena or disasters with an impact on the socio-economic
system. A more detailed discussion is given in Refs. [76, 77].
Other factors contributing to the difficulty to manage economic systems are the
large heterogeneity of system elements and the considerable level of randomness
as well as the possibility of a chaotic or turbulent dynamics (see Sect. 16.3.4).

Furthermore, the agents in economic systems are responsive to information, which
can create self-fulfilling or self-destroying prophecy effects. Inflation may be viewed
as example of such an effect. Interestingly, in some cases one even does not know
in advance, which of these effects will occur.
It is also not obvious that the control mechanisms are well designed from a
cybernetic perspective, i.e. that we have sufficient information about the system
and suitable control variables to make control feasible. For example, central banks
do not have terribly many options to influence the economic system. Among them
are performing open-market operations (to control money supply), adjustments in
fractional-reserve banking (keeping only a limited deposit, while lending a large part
of the assets to others), or adaptations in the discount rate (the interest rate charged
to banks for borrowing short-term funds directly from a central bank). Nevertheless,
the central banks are asked to meet multiple goals such as:
• To guarantee well-functioning and robust financial markets.

• To support economic growth.
• To balance between inflation and unemployment.
• To keep exchange rates within reasonable limits.
Furthermore, the one-dimensional variable of “money” is also used to influence

individual behavior via taxes (by changing behavioral incentives). It is question-
able, whether money can optimally meet all these goals at the same time (see
Sect. 16.3.7). We believe that a computer, good food, friendship, social status, love,
fairness, and knowledge can only to a certain extent be replaced by and traded
against each other. Probably for this reason, social exchange comprises more than
just material exchange [131–133].
It is conceivable that financial markets as well are trying to meet too many goals
at the same time. This includes:
• To match supply and demand.

• To discover a fair price.
• To raise the foreign direct investment (FDI).
• To couple local economies with the international system.
• To facilitate large-scale investments.
• To boost development.
• To share risk.
• To support a robust economy, and
• To create opportunities (to gamble, to become rich, etc.).
Therefore, it would be worth stuyding the system from a cybernetic control

perspective. Maybe, it would work better to separate some of these functions from
each other rather than mixing them.
16.3.9 Human Factors
Another aspect that tends to be overlooked in mainstream economics is the relevance

of psychological and social factors such as emotions, creativity, social norms,
herding effects, etc. It would probably be wrong to interpret these effects just as a
result of perception biases (see Sect. 16.3.1). Most likely, these human factors serve
certain functions such as supporting the creation of public goods [102] or collective
intelligence [134, 135].
As Bruno Frey has pointed out, economics should be seen from a social science
perspective [136]. In particular, research on happiness has revealed that there are
more incentives than just financial ones that motivate people to work hard [133].
Interestingly, there are quite a number of factors which promote volunteering
[132].
It would also be misleading to judge emotions from the perspective of irrational
behavior. They are a quite universal and a relatively energy-consuming way of
signalling. Therefore, they are probably more reliable than non-emotional signals.
Moreover, they create empathy and, consequently, stimulate mutual support and
a readiness for compromises. It is quite likely that this creates a higher degree
of cooperativeness in social dilemma situations and, thereby, a higher payoff
on average as compared to emotionless decisions, which often have drawbacks
later on.
16.3.10 Information
Finally, there is no good theory that would allow one to assess the relevance
of information in economic systems. Most economic models do not consider
information as an explanatory variable, although information is actually a stronger
driving force of urban growth and social dynamics than energy [137]. While we
have an information theory to determine the number of bits required to encode a
message, we are lacking a theory, which would allow us to assess what kind of
information is relevant or important, or what kind of information will change the
social or economic world, or history. This may actually be largely dependent on
the perception of pieces of information, and on normative or moral issues filtering
or weighting information. Moreover, we lack theories describing what will happen
in cases of coincidence or contradiction of several pieces of information. When
pieces of information interact, this can change their interpretation and, thereby,
the decisions and behaviors resulting from them. That is one of the reasons why
socio-economic systems are so hard to predict: “Unknown unknowns”, structural
instabilities, and innovations cause emergent results and create a dynamics of
surprise [138].
16.4 Role of Other Scientific Fields 319
16.4 Role of Other Scientific Fields
16.4.1 Econophysics, Ecology, Computer Science
The problems discussed in the previous two sections pose interesting practical
and fundamental challenges for economists, but also other disciplines interested
in understanding economic systems. Econophysics, for example, pursues a phys-
ical approach to economic systems, applying methods from statistical physics
[81], network theory [139, 140], and the theory of complex systems [85, 87].
A contribution of physics appears quite natural, in fact, not only because of its
tradition in detecting and modeling regularities in large data sets [141]. Physics
also has a lot of experience how to theoretically deal with problems such as
time-dependence, fluctuations, friction, entropy, non-linearity, strong interactions,
correlations, heterogeneity, and many-particle simulations (which can be easily
extended towards multi-agent simulations). In fact, physics has influenced economic
modeling already in the past. Macroeconomic models, for example, were inspired
by thermodynamics. More recent examples of relevant contributions by physicists
concern models of self-organizing conventions [54], of geographic agglomeration
[65], of innovation spreading [142], or of financial markets [143], to mention just a
few examples.
One can probably say that physicists have been among the pioneers calling
for new approaches in economics [81, 87, 143–147]. A particularly visionary book
beside Wolfgang Weidlich’s work was the “Introduction to Quantitative Aspects of
Social Phenomena” by Elliott W. Montroll and Wade W. Badger, which addressed
by mathematical and empirical analysis subjects as diverse as population dynamics,
the arms race, speculation patterns in stock markets, congestion in vehicular traffic
as well as the problems of atmospheric pollution, city growth and developing
countries already in 1974 [148].
Unfortunately, it is impossible in our paper to reflect the numerous contributions
of the field of econophysics in any adequate way. The richness of scientific
contributions is probably reflected best by the Econophysics Forum run by Yi-Cheng
Zhang [149]. Many econophysics solutions are interesting, but so far they are not
broad and mighty enough to replace the rational agent paradigm with its large
body of implications and applications. Nevertheless, considering the relatively small
number of econophysicists, there have been many promising results. The probably
largest fraction of publications in econophysics in the last years had a data-driven
or computer modeling approach to financial markets [143]. But econophyics has
more to offer than the analysis of financial data (such as fluctuations in stock
and foreign currency exchange markets), the creation of interaction models for
stock markets, or the development of risk management strategies. Other scientists
have focused on statistical laws underlying income and wealth distributions, non-
linear market dynamics, macroeconomic production functions and conditions for
economic growth or agglomeration, sustainable economic systems, business cycles,
microeconomic interaction models, network models, the growth of companies,
supply and production systems, logistic and transport networks, or innovation

dynamics and diffusion. An overview of subjects is given, for example, by Ref. [152]
and the contributions to annual spring workshop of the Physics of Socio-Economic
Systems Division of the DPG [153].
To the dissatisfaction of many econophysicists, the transfer of knowledge often
did not work very well or, if so, has not been well recognized [150]. Besides
scepticism on the side of many economists with regard to novel approaches
introduced by “outsiders”, the limited resonance and level of interdisciplinary
exchange in the past was also caused in part by econophysicists. In many cases,
questions have been answered, which no economist asked, rather than addressing
puzzles economists are interested in. Apart from this, the econophysics work was
not always presented in a way that linked it to the traditions of economics and
pointed out deficiencies of existing models, highlighting the relevance of the new
approach well. Typical responses are: Why has this model been proposed and
not another one? Why has this simplification been used (e.g. an Ising model of
interacting spins rather than a rational agent model)? Why are existing models not
good enough to describe the same facts? What is the relevance of the work compared
to previous publications? What practical implications does the finding have? What
kind of paradigm shift does the approach imply? Can existing models be modified
or extended in a way that solves the problem without requiring a paradigm shift?
Correspondingly, there have been criticisms not only by mainstream economists,
but also by colleagues, who are open to new approaches [151].
Therefore, we would like to suggest to study the various economic subjects
from the perspective of the above-mentioned fundamental challenges, and to
contrast econophysics models with traditional economic models, showing that the
latter leave out important features. It is important to demonstrate what properties
of economic systems cannot be understood for fundamental reasons within the
mainstream framework (i.e. cannot be dealt with by additional terms within the
modeling class that is conventionally used). In other words, one needs to show why
a paradigm shift is unavoidable, and this requires careful argumentation. We are not
claiming that this has not been done in the past, but it certainly takes an additional
effort to explain the essence of the econophysics approach in the language of
economics, particularly as mainstream economics may not always provide suitable
terms and frameworks to do this. This is particularly important, as the number of
econophysicists is small compared to the number of economists, i.e. a minority
wants to convince an established majority. To be taken seriously, one must also
demonstrate a solid knowledge of related previous work of economists, to prevent
the stereotypical reaction that the subject of the paper has been studied already long
time ago (tacitly implying that it does not require another paper or model to address
what has already been looked at before).
A reasonable and promising strategy to address the above fundamental and
practical challenges is to set up multi-disciplinary collaborations in order to combine
the best of all relevant scientific methods and knowledge. It seems plausible that
this will generate better models and higher impact than working in separation,
16.4 Role of Other Scientific Fields 321
and it will stimulate scientific innovation. Physicists can contribute with their
experience in handling large data sets, in creating and simulating mathematical
models, in developing useful approximations, in setting up laboratory experiments
and measurement concepts. Current research activities in economics do not seem to
put enough focus on:
• Modeling approaches for complex systems [154].
• Computational modeling of what is not analytically tractable anymore, e.g. by
agent-based models [155–157].
• Testable predictions and their empirical or experimental validation [164].
• Managing complexity and systems engineering approaches to identify alternative
ways of organizing financial markets and economic systems [91, 93, 165], and
• An advance testing of the effectiveness, efficiency, safety, and systemic impact
(side effects) of innovations, before they are implemented in economic systems.
This is in sharp contrast to mechanical, electrical, nuclear, chemical and medical
drug engineering, for example.
Expanding the scope of economic thinking and paying more attention to these
natural, computer and engineering science aspects will certainly help to address the
theoretical and practical challenges posed by economic systems. Besides physics,
we anticipate that also evolutionary biology, ecology, psychology, neuroscience,
and artificial intelligence will be able to make significant contributions to the
understanding of the roots of economic problems and how to solve them. In
conclusion, there are interesting scientific times ahead.
16.4.2 Social Sciences
It is a good question, whether answering the above list of fundamental challenges

will sooner or later solve the practical problems as well. We think, this is a
precondition, but it takes more, namely the consideration of social factors. In
particular, the following questions need to be answered:
1. How to understand human decision-making? How to explain deviations from
rational choice theory and the decision-theoretical paradoxes? Why are people
risk averse?
2. How does consciousness and self-consciousness come about?
3. How to understand creativity and innovation?
4. How to explain homophily, i.e. the fact that individuals tend to agglomerate,
interact with and imitate similar others?
5. How to explain social influence, collective decision making, opinion dynamics
and voting behavior?
6. Why do individuals often cooperate in social dilemma situations?
7. How do indirect reciprocity, trust and reputation evolve?
8. How do costly punishment, antisocial punishment, and discrimination come

about?
9. How can the formation of social norms and conventions, social roles and
socialization, conformity and integration be understood?
10. How do language and culture evolve?
11. How to comprehend the formation of group identity and group dynamics? What
are the laws of coalition formation, crowd behavior, and social movements?
12. How to understand social networks, social structure, stratification, organiza-
tions and institutions?
13. How do social differentiation, specialization, inequality and segregation come
about?
14. How to model deviance and crime, conflicts, violence, and wars?
15. How to understand social exchange, trading, and market dynamics?
We think that, despite the large amount of research performed on these subjects, they
are still not fully understood. The ultimate goal would be to formulate mathematical
models, which would allow one to understand these issues as emergent phenomena
based on first principles, e.g. as a result of (co-)evolutionary processes. Such
first principles would be the basic facts of human capabilities and the kinds of
interactions resulting from them, namely:
1. Birth, death, and reproduction.
2. The need of and competition for resources (such as food and water).
3. The ability to observe their environment (with different senses).
4. The capability to memorize, learn, and imitate.
5. Empathy and emotions.
6. Signaling and communication abilities.
7. Constructive (e.g. tool-making) and destructive (e.g. fighting) abilities.
8. Mobility and (limited) carrying capacity.
9. The possibility of social and economic exchange.
Such features can, in principle, be implemented in agent-based models [158–163].
Computer simulations of many interacting agents would allow one to study the
phenomena emerging in the resulting (artificial or) model societies, and to compare
them with stylized facts [163, 168, 169]. The main challenge, however, is not to
program a seemingly realistic computer game. We are looking for scientific models,
i.e. the underlying assumptions need to be validated, and this requires to link
computer simulations with empirical and experimental research [170], and with
massive (but privacy-respecting) mining of social interaction data [141]. In the ideal
case, there would also be an analytical understanding in the end, as it has been
recently gained for interactive driver behavior [111].
Acknowledgements The authors are grateful for partial financial support by the ETH Competence
Center “Coping with Crises in Complex Socio-Economic Systems” (CCSS) through ETH Research
Grant CH1-01 08-2 and by the Future and Emerging Technologies programme FP7-COSI-ICT of
the European Commission through the project Visioneer (grant no.: 248438). They would like
to thank for feedbacks on the manuscript by Kenett Dror, Tobias Preis and Gabriele Tedeschi
References 323
as well as for inspiring discussions during a Visioneer workshop in Zurich from January 13
to 15, 2010, involving, besides the authors, Stefano Battiston Guido Caldarelli, Anna Carbone,
Giovanni Luca Ciampaglia, Andreas Flache, Imre Kondor, Sergi Lozano, Thomas Maillart, Amin
Mazloumian, Tamara Mihaljev, Alexander Mikhailov, Ryan Murphy, Carlos Perez Roca, Stefan
Reimann, Aki-Hiro Sato, Christian Schneider, Piotr Swistak, Gabriele Tedeschi, and Jiang Wu.
Last but not least, we are grateful to Didier Sornette, Frank Schweitzer and Lars-Erik Cederman
for providing some requested references.
References
1. P. Mancosu (ed.) From Brouwer to Hilbert: The Debate on the Foundations of Mathematics
in the 1920s. (Oxford University Press, New York, 1998)
2. Paul Krugman, “How did economists get it so wrong?”, The New York Times (September 2,
2009), see http://www.nytimes.com/2009/09/06/magazine/06Economic-t.html
3. D. Colander et al., The Financial Crisis and the Systemic Failure of Academic Economics
(Dahlem Report, Univ. of Copenhagen Dept. of Economics Discussion Paper No. 09–03,
2009), see http://papers.ssrn.com/sol3/papers.cfm?abstract id=1355882
4. See http://www.ipetitions.com/petition/revitalizing economics/?e and http://emmarogan.
wordpress.com/2009/10/23/the-financial-crisis-how-economists-went-astray/
5. DeLisle Worrell, What’s wrong with economics. Address of the Governor of the Central Bank
of Barbados on June 30, 2010, see http://www.bis.org/review/r100713c.pdf
6. D. Helbing, The FuturIcT knowledge accelerator: Unleashing the power of information for a
sustainable future, see http://arxiv.org/abs/1004.4969 and http://www.futurict.eu
7. Bjorn Lomborg, Global Crises, Global Solutions: Costs and Benefits. (Cambridge University
Press, Cambridge, 2nd ed., 2009)
8. R.H. Nelson, Economics as Religion: from Samuelson to Chicago and Beyond. (Pennsylvania
State University, 2001)
9. F.A. Von Hayek, Individualism and Economic Order. (University of Chicago Press, 1948)
10. F.A. Von Hayek, The Counter Revolution of Science. (The Free Press of Glencoe, London,
1955)
11. M. Friedman, The Case of flexible exchange rates, in: Essays in Positive Economics.
(University of Chicago Press, Chicago, 1953)
12. R. Lucas, Adaptive behavior and economic theory, J. Bus. 59, S401–S426 (1986)
13. G. Gigerenzer, R. Selten (eds.) Bounded Rationality. The Adaptive Toolbox. (MIT Press,
Cambridge, MA, 2001)
14. H. Gintis, The Bounds of Reason: Game Theory and the Unification of the Behavioral
Sciences. (Princeton University Press, Princeton, 2009)
15. F. Caccioli, M. Marsili, On information efficiency and financial stability. Preprint http://arxiv.
org/abs/1004.5014
16. H.A. Simon Models of Man. (Wiley, New York, 1957)
17. H.A. Simon, Models of Bounded Rationality. Vol. 1 (MIT Press, Boston, 1984)
18. G. Gigerenzer, P.M. Todd, and the ABC Research Group, Simple Heuristics That Make Us
Smart. (Oxford University Press, Oxford, 2000)
19. For a collection of cognitive biases see http://en.wikipedia.org/wiki/List of cognitive biases
20. D. Kahneman, P. Slovic, A. Tversky, Judgment under Uncertainty: Heuristics and Biases.
(Cambridge University Press, Cambridge, MA, 1982)
21. V.I. Yukalov, D. Sornette, Physics of risk and uncertainty in quantum decision making,
European Physical Journal B 71, 533–548 (2009)
22. E. Fehr, K.M. Schmidt, A theory of fairness, competition, and cooperation. Q. J. Econ. 114(3),
817–868 (1999)
23. J. Henrich, R. Boyd, S. Bowles, C. Camerer, Foundations of Human Sociality: Eco-

nomic Experiments and Ethnographic Evidence from Fifteen Small-Scale Societies. (Oxford
24. E. Hoffman, K. McCabe, V.L. Smith, Social distance and other-regarding behavior in dictator
games. Am. Econ. Rev. 86(3), 653–660 (1996)
25. R.O. Murphy, K. Ackermann, Measuring social value orientation. Judgement and decision
making 6(8), 771–781 (2011)
26. Ö.B. Bodvarsson, W.A. Gibson, Economics and restaurant gratuities: Determining tip rates.
Am. J. Econ. Soc. 56(2), 187–203 (1997)
27. J. Elster, The Cement of Society: A Study of Social Order. (Cambridge University Press,
Cambridge, 1989)
28. C. Horne, The Rewards of Punishment. A Relational Theory of Norm Enforcement. (Stanford
University Press, Stanford, 2009)
29. D. Helbing, W. Yu, K.-D. Opp, H. Rauhut, The emergence of homogeneous
norms in heterogeneous populations. Santa Fe Working Paper 11-01-001 (2011), see
http://www.santafe.edu/media/workingpapers/11-01-001.pdf, last accessed on March 6, 2012
30. D. Centola, R. Willer, M. Macy, The emperor’s dilemma: A computational model of self-
enforcing norms. Am. J. Soc. 110, 1009–1040 (2005)
31. J.M. Epstein, R.A. Hammond, Non-explanatory equilibria: An extremely simple game with
(mostly) unattainable fixed points. Complexity 7(4), 18–22 (2002)
32. C. Borgs, J. Chayesa, N. Immorlicab, A.T. Kalaia, V. Mirroknia, C Papadimitriou, The myth
of the folktheorem, Games and Economic Behavior 70(1), 34–43 (2010)
33. H.G. Schuster, Deterministic Chaos. (Wiley VCH, Weinheim, 2005)
34. For example, three-body planetary motion has deterministic chaotic solutions, although it
is a problem in classical mechanics, where the equations of motion optimize a Lagrangian
functional.
35. K. Gödel, On Formally Undecidable Propositions of Principia Mathematica and Related
Systems. (Basic, New York, 1962)
36. A.M. Turing, On computable numbers, with an application to the Entscheidungsproblem.
Proc. Lond. Math. Soc. 2.42, 230–265 (1936)
37. E. Fama, The Behavior of stock market prices. J. Bus. 38 34–105 (1965)
38. Wikipedia article on Adam Smith, see http://en.wikipedia.org/wiki/Adam Smith, downloaded
on July 14, 2010.
39. A. Smith (1776), An Inquiry into the Nature and Causes of the Wealth of Nations. (University
of Chicago Press, 1977)
40. A. Smith, The Theory of Moral Sentiments (1759)
41. M.A. Nowak, K. Sigmund, Evolution of indirect reciprocity. Nature 437, 1291–1298 (2005)
42. M.J. Mauboussin, Revisiting market efficiency: The stock market as a complex adaptive
system. J. Appl. Corp. Fin. 14, 47–55 (2005)
43. J. Stiglitz, There is no invisible hand. The Guardian (December 20, 2002), see http://www.
guardian.co.uk/education/2002/dec/20/highereducation.uk1
44. D. Helbing, T. Vicsek, Optimal self-organization. New J. Phys. 1, 13.1–13.17 (1999)
1141 (2001)
46. A. Johansson, D. Helbing, H.Z. A-Abideen, S. Al-Bosta, From crowd dynamics to crowd
safety: A video-based analysis. Advances in Complex Systems 11(4), 497–527 (2008)
47. W. Michiels, S.-I. Niculescu, Stability and Stabilization of Time-Delay Systems. (siam—
Society for Industrial and Applied Mathematics, Philadelphia, 2007)
48. D. Helbing, U. Witt, S. Lämmer, T. Brenner, Network-induced oscillatory behavior in material
flow networks and irregular business cycles. Phys. Rev. E 70, 056118 (2004)
51. T. Preis, H.E. Stanley, Switching phenomena in a system with no switches. J. Stat. Phys.
(JSTAT) 138, 431–446 (2010)
References 325
52. G.A. Akerlof, R.J. Shiller, Animal Spirits: How Human Psychology Drives the Economy, and
Why It Matters for Global Capitalism. (Princeton University Press, 2010)
53. J.D. Sterman, Testing behavioral simulation models by direct experiment. Manag. Sci. 33(12),
1572–1592 (1987)
54. D. Helbing, A mathematical model for behavioral changes by pair interactions. pp. 330–348
In: G. Haag, U. Mueller, and K.G. Troitzsch (eds.) Economic Evolution and Demographic
Change. Formal Models in Social Sciences. (Springer, Berlin, 1992)
55. H.N. Agiza, G.I. Bischib, M. Kopel, Multistability in a dynamic Cournot game with three
oligopolists. Math. Comput. Simulat. 51, 63–90 (1999)
56. D. Helbing, A. Szolnoki, M. Perc, G. Szabo, Evolutionary establishment of moral and double
moral standards through spatial interactions. PLoS Comput. Biol. 6(4), e1000758 (2010)
57. A.J. Lotka Elements of Mathematical Biology. (Dover, New York, 1956)
58. E. Hopf, Abzweigungen einer periodischen Lösung von einer stationären Lösung eines
Differentialgleichungssystems. Math. Naturwiss. Klasse 94, 1ff (1942)
Models and experiments. pp. 47–95 In: M. Schreckenberg, R. Selten (eds.) Human Behaviour
and Traffic Networks (Springer, Berlin, 2004)
60. H. Gintis, The dynamics of general equilibrium. Econ. J. 117, 1280–1309 (2007)
61. C.H. Hommes, Modeling the stylized facts in finance through simple nonlinear adaptive
systems. Proc. Natl. Acad. Sc. USA (PNAS) 99, Suppl. 3, 7221–7228 (2002)
62. The quotes were presented by Jorgen Vitting Andersen in his talk “Predicting moments of
crisis in physics and finance” during the workshop “Windows to Complexity” in Münster,
Germany, on June 11, 2010
63. Turing A.M., The chemical basis of morphogenesis. Phil. Trans. Roy. Soc. Lond. B 237, 37–
72 (1952)
64. J.D. Murray, Lectures on Nonlinear Differential Equation-Models in Biology. (Clanderon
65. W. Weidlich, M. Munz, Settlement formation I: A dynamic theory, Ann. Reg. Sci. 24, 83–106
(2000); Settlement formation II: Numerical simulation, Ann. Reg. Sci. 24, 177–196 (2000)
66. D. Helbing, T. Platkowski, Self-organization in space and induced by fluctuations. Int. J.
Chaos Theor. Appl. 5(4), 47–62 (2000)
67. J. Mimkes, Stokes integral of economic growth: Calculus and the Solow model. Physica A
389(8), 1665–1676 (2010)
68. O.J. Blanchard, L.H. Summers, Hysteresis and the European unemployment problem. NBER
Macroecon. Annu. 1, 15–78 (1986)
69. M. Cross, H. Greenside, Pattern Formation and Dynamics in Nonequilibrium Systems.
(Cambridge University, 2009)
70. D. Sornette, Critical Phenomena in Natural Sciences: Chaos, Fractals, Selforganization and
Disorder. (Springer, Berlin, 2006)
71. A. Bunde, J. Kropp, H.-J. Schellnhuber (eds.) The Science of Disasters. (Springer, Berlin,
2002)
72. W. Horsthemke, R. Lefever, Noise-Induced Transitions: Theory and Applications in Physics,
Chemistry, and Biology. (Springer, Berlin, 1983)
73. P. Reimann, Brownian motors: noisy transport far from equilibrium. Phys. Rep. 361, 57–265
(2002)
74. E.C. Zeeman (eds.) Catastrophe Theory (Addison-Wesley, London, 1977)
75. H.E. Stanley, Introduction to Phase Transitions and Critical Phenomena. (Oxford University,
1987)
76. D. Helbing, S. Lämmer, Managing complexity: An introduction. Pages 1–16 In: D. Helbing
(eds.) Managing Complexity: Insights, Concepts, Applications. (Springer, Berlin, 2008)
77. D. Helbing, Systemic risks in society and economics. SFI Working Paper 09-12-044, see
http://www.santafe.edu/media/workingpapers/09-12-044.pdf
78. J. Lorenz, S. Battiston, F. Schweitzer Systemic risk in a unifying framework for cascading
processes on networks. Eur. Phys. J. B 71(4), 441–460 (2009)
79. P. Bak, How Nature Works: The Science of Self-Organized Criticality. (Springer, Berlin, 1999)
80. Knowledge@Wharton, Why economists failed to predict the financial crisis (May 13,
2009), see http://knowledge.wharton.upenn.edu/article.cfm?articleid=2234 or http://www.
ftpress.com/articles/article.aspx?p=1350507
81. W. Weidlich, Physics and social science—the approach of synergetics. Phys. Rep. 204(1),
1–163 (1991)
82. T. Puu, Nonlinear Economic Dynamics (Lavoisier, 1991); Attractors, Bifurcations, & Chaos.
Nonlinear Phenomena in Economics (Springer, Berlin, 2003)
83. H.W. Lorenz, Nonlinear Dynamical Equations and Chaotic Economy. (Springer, Berlin,
1993)
84. P. Krugman, The Self-Organizing Economy. (Blackwell, 1996)
85. F. Schweitzer (ed.) Self-Organization of Complex Structures: From Individual to Collective
Dynamics (CRC Press, 1997); Modeling Complexity in Economic and Social Systems (World
Scientific, 2002)
86. R.H. Day, Complex Economic Dynamics. (MIT Press, Vol. 1: 1998; Vol. 2: 1999)
87. W. Weidlich, Sociodynamics: A Systematic Approach to Mathematical Modelling in the Social
Sciences. (CRC Press, Boca Raton, 2000)
88. J.D. Sterman, Business Dynamics: Systems Thinking and Modeling for a Complex World.
(McGraw Hill, 2000)
89. W.A. Brock, Growth Theory, Non-Linear Dynamics and Economic Modelling. (Edward Elgar,
2001)
90. S.Y. Auyang, Foundations of Complex-System Theories. (Cambridge University, 1998)
91. M. Salzano, D. Colander (eds.) Complexity Hints for Economic Policy. (Springer, Berlin,
2007)
92. D. Delli Gatti, E. Gaffeo, M. Gallegati, G. Giulioni, A. Palestrini, Emergent Economics.
93. M. Faggini, T. Lux (eds.) Coping with the Complexity of Economics. (Springer, Berlin, 2009)
94. J.B. Rosser, Jr., K.L. Cramer, J. Madison (eds.) Handbook of Research on Complexity.
(Edward Elgar Publishers, 2009)
95. M. Olson, The rise and decline of nations: economic growth, stagflation, and social rigidities.
(Yale University Press, New Haven, 1982)
96. R. Axelrod, The Evolution of Cooperation. (Basic Books, New York, 1984)
97. J.C. Harsanyi, R. Selten, A General Theory of Equilibrium Selection. (MIT Press, Cambridge,
MA, 1988)
98. D. Helbing, S. Lozano, Phase transitions to cooperation in the prisoner’s dilemma. Phys. Rev.
E 81(5), 057102 (2010)
100. F.C. Santos, M.D. Santos, J.M. Pacheco, Social diversity promotes the emergence of
cooperation in public goods games. Nature 454, 213–216 (2008)
noisy conditions. Proc. Natl. Acad Sci. USA (PNAS) 106(8), 3680–3685 (2009)
102. E. Ostrom, Governing the Commons. The Evolution of Institutions for Collective Action.
103. C.P. Roca, S. Lozano, A. Arenas, A. Sanchez, Topological traps control flow on real networks:
The case of coordination failures, PLoS One 5(12), e15210 (2010)
104. D. Helbing, Modelling supply networks and business cycles as unstable transport phenomena.
New J. Phys. 5, 90.1–90.28 (2003)
105. M. Eigen, The selforganization of matter and the evolution of biological macromolecules.
Naturwissenschaften 58, 465–523 (1971)
106. J. Schumpeter, The Theory of Economic Development (Harvard University Press, Cambridge,
1934)
107. Ken Arrow, In: D. Colander, R.P.F. Holt, J. Barkley Rosser (eds.) The Changing Face of
Economics. Conversations with Cutting Edge Economists (The University of Michigan Press,
Ann Arbor, 2004), p. 301.
References 327
108. M. Gallegati, A. Kirman, Beyond the Representative Agent. (Edward Elgar, Cheltenham,
1999)
109. A. Kirman, Economics with Heterogeneous Interacting Agents. (Springer, Berlin, 2001).
110. M. Aoki, Modeling Aggregate Behavior and Fluctuations in Economics. (Cambridge
University, Cambridge, 2002)
111. D. Helbing, Collection of papers on An Analytical Theory of Traffic Flows in Eur. Phys. J. B,
see http://www.soms.ethz.ch/research/traffictheory
112. W.B. Arthur, S.N. Durlauf, D. Lane (eds.) The Economy as An Evolving Complex System II.
(Santa Fe Institute Series, Westview Press, 1997)
113. L.E. Blume, S.N. Durlauf, (eds.) The Economy as an Evolving Complex System III. (Oxford
114. U. Witt, Evolutionary economics. pp. 67–73 In: S.N. Durlauf, L.E. Blume (eds.) The New
Palgrave Dictionary of Economics. (Palgrave Macmillan, 2008)
115. R.M. May, S.A. Levin, G. Sugihara, Ecology for bankers. Nature 451, 894–895 (2008)
116. E.D. Beinhocker, Origin of Wealth: Evolution, Complexity, and the Radical Remaking of
Economics. (Harvard Business Press, 2007)
117. R.A. Young, R.L. Giese (eds.), Introduction to Forest Ecosystem Science and Management.
(Wiley, New York, 2002)
118. H. Meinhardt, Models of Biological Pattern Formation. (Academic Press, London, 1982)
119. M. Caramia, P. Dell’Olmo, Multi-objective Management in Freight Logistics: Increasing
Capacity, Service Level and Safety with Optimization Algorithms. (Springer, New York, 2008)
120. Jean-Philippe Rennard (eds.) Handbook of Research on Nature-Inspired Computing for
Economics and Management. (IGI Global, 2006)
121. S. Lämmer, D. Helbing, Self-control of traffic lights and vehicle flows in urban road networks.
J. Stat. Mech. (JSTAT), P04019 (2008)
122. D. Helbing et al., Biologistics and the struggle for efficiency: Concepts and perspectives. Adv.
Comp. Syst. 12(6), 533–548 (2009)
123. D. Helbing, M. Moussaid, Analytical calculation of critical perturbation amplitudes and
critical densities by non-linear stability analysis of a simple traffic flow model. Eur. Phys.
J. B 69(4), 571–581 (2009)
124. K. Windt, T. Philipp, F. Böse, Complexity cube for the characterization of complex production
systems. Int. J. Comput. Integrated Manuf. 21(2), 195–200 (2007)
125. D. Helbing, S. Lämmer, Verfahren zur Koordination konkurrierender Prozesse oder zur
Steuerung des Transports von mobilen Einheiten innerhalb eines Netzwerkes (Method for
coordination of concurrent processes for control of the transport of mobile units within a
network), Patent WO/2006/122528 (2006)
126. I. Kondor, S. Pafka, G. Nagy, Noise sensitivity of portfolio selection under various risk
measures. J. Bank. Fin. 31(5), 1545–1573 (2007)
127. D. Dorner, The Logic of Failure: Recognizing and Avoiding Error in Complex Situations.
(Basic, New York, 1997)
128. K. Kempf, Complexity and the enterprise: The illusion of control. pp. 57–87 In: D. Helbing
(eds.) Managing Complexity: Insights, Concepts, Applications (Springer, Berlin, 2008)
129. D. Sornette, Nurturing breakthroughs: lessons from complexity theory. J. Econ. Interaction
and Coordination 3, 165–181 (2008)
130. C.A.E. Goodhart, Monetary Relationships: A View from Threadneedle Street. (Papers in
Monetary Economics, Reserve Bank of Australia, 1975); For applications of Le Chatelier’s
principle to economics see also P. A. Samuelson, Foundations of Economic Analysis. (Harvard
University, 1947)
131. A.P. Fiske, Structures of Social Life: The Four Elementary Forms of Human Relations. (The
Free Press, 1993)
132. E.G. Clary et al., Understanding and assessing the motivations of volunteers: a functional
approach. J. Pers. Soc. Psychol. 74(6), 1516–1530 (1998)
133. B.S. Frey, Happiness: A Revolution in Economics. (MIT Press, Cambridge, MA and London,
UK, 2008)
134. J. Surowiecki, The Wisdom of Crowds: Why the Many Are Smarter than the Few and How
Collective Wisdom Shapes Business, Economies, Societies, and Nations. (Doubleday, 2004)
135. C. Blum, D. Merkle (eds.), Swarm Intelligence. Introduction and Applications. (Springer,
Berlin, 2008)
136. B.S. Frey, Economics as a Science of Human Behaviour: Towards a New Social Science
Paradigm. (Springer, Berlin, 1999)
137. L.M.A. Bettencourt, J. Lobo, D. Helbing, C. Khnert, G.B. West, Growth, innovation, scaling
and the pace of life in cities. Proc. Natl. Acad. Sci. USA (PNAS) 104, 7301–7306 (2007)
138. R.R. McDaniel, Jr. D.J. Driebe (eds.), Uncertainty and Surprise in Complex Systems.
139. A.-L. Barabasi, Scale-free networks: A decade and beyond. Science 325, 412–413 (2009)
140. F. Schweitzer et al., Economic networks: The new challenges. Science 325, 422–425 (2009)
141. D. Lazer et al., Computational social science. Science 323, 721–723 (2009)
142. E. Bruckner, W. Ebeling, M.A.J. Montano, A. Scharnhorst, Hyperselection and innovation
described by a stochastic model of technological evolution; pP. 79–90 In: L.L. Leydesdorff,
P. van den Besselaar (eds.) Evolutionary Economics and Chaos Theory: New Directions in
Technology Studies. (Pinter, London, 1994)
143. R.N. Mantegna, H.E. Stanley, Introduction to Econophysics: Correlations and Complexity in
Finance. (Cambridge University Press, Cambridge, 2000).
144. D. Challet, M. Marsili, Y.-C. Zhang, Minority Games: Interacting Agents in Financial
Markets. (Oxford University, Oxford, New York, 2005)
145. J.-P. Bouchaud, Economics needs a scientific revolution. Nature 455, 1181 (2008)
146. J.D. Farmer, D. Foley, The economy needs agent-based modelling. Nature 460, 685–686
(2009)
147. M. Buchanan, Meltdown modelling. Nature 460, 680–682 (2009)
148. E.W. Montroll, W.W. Badger, Introduction to Quantitative Aspects of Social Phenomena.
(Gordon and Breach, 1974)
149. Y.-C. Zhang, Econophysics Forum, see http://www.unifr.ch/econophysics
150. B.M. Roehner, Fifteen years of econophysics: worries, hopes and prospects (2010), see http://
arxiv.org/abs/1004.3229
151. M. Gallegati, S. Keen, T. Lux, P. Omerod, Worrying trends in econophysics. Physica A 370,
1–6 (2006)
152. B.K. Chakrabarti, A. Chakraborti, A. Chatterjee (eds.) Econophysics and Sociophysics:
Trends and Perspectives. (Wiley VCH, Weinheim, 2006)
153. Aims and Scope of the Physics of Socio-Economic Systems Division of the German Physical
Society, see http://www.dpg-physik.de/dpg/gliederung/fv/soe/aims.html. For past events see
http://www.dpg-physik.de/dpg/gliederung/fv/soe/veranstaltungen/vergangene.html
154. D. Helbing, Pluralistic modeling of complex systems. Preprint http://arxiv.org/abs/1007.2818
(2010)
155. L. Tesfatsion, K.L. Judd (eds.), Handbook of Computational Economics, Vol. 2: Agent-Based
Computational Economics. (North-Holland, Amsterdam, 2006)
156. J.R. Harrison, Z. Lin, G.R. Carroll, K.M. Carley, Simulation modeling in organizational and
management research. Acad. Manag. Rev. 32(4), 1229–1245 (2007)
157. J.P. Davis, K.M. Eisenhardt, C.B. Bingham, Developing theory through simulation methods.
Acad. Manag. Rev. 32(2), 480–499 (2007)
158. T. Schelling, Micromotives and Macrobehavior. (Norton, New York, 1978)
159. R. Hegselmann, A. Flache, Understanding complex social dynamics: A plea for cellular
automata based modelling. J. Artif. Soc. Soc. Simul. 1, no. 3, see http://www.soc.surrey.ac.
uk/JASSS/1/3/1.html
160. N. Gilbert, S. Bankes, Platforms and methods for agent-based modeling. PNAS 99(3), 7197–
7198 (2002)
161. M.W. Macy, R. Willer, From factors to actors: Computational sociology and agent-based
modeling. Annu. Rev. Sociol. 28, 143–166 (2002)
References 329
162. R.K. Sawyer, Artificial societies. Multiagent systems and the micro-macro link in sociological
theory. Socio. Meth. Res. 31(3), 325–363 (2003)
(Princeton University, 2007)
164. J.H. Kagel, A.E. Roth, The Handbook of Experimental Economics. (Princeton University,
1997)
165. D. Helbing, Managing Complexity: Concepts, Insights, Applications. (Springer, Berlin, 2008)
166. L. Hurwicz, S. Reiter, Designing Economic Mechanisms. (Cambridge University, Cambridge,
2006)
167. W.G. Sullivan, E.M. Wicks, C.P. Koelling, Engineering Economy. (Prentice Hall, Upper
Saddle River, NJ, 2008)
168. D. Helbing, Quantitative Sociodynamics. (Kluwer Academic, Dordrecht, 1995)
169. C. Castellano, S. Fortunato, V. Loreto, Statistical physics of social dynamics. Rev. Mod. Phys.
81, 591–646 (2009)
170. D. Helbing, W. Yu (2010) The future of social experimenting. Proc. Natl. Acad.
Sci. USA (PNAS) 107(12), 5265–5266 (2010); see http://www.soms.ethz.ch/research/
socialexperimentingforalongerversion.
Index
ABC modeling, 27 Avoidance maneuvers, 76

Acceleration equation, 73 Awareness, 59
Adaptive feedback, 137
Adaptive systems, 275
Adaptor, 62 Badger, W.W., 319
Additive noise, 310 Balance of power, 303
Administrative processes, 58 Bankruptcy cascades, 267, 316
Advance warning signs, 62, 92 Battle of the sexes, 212
Agent-based modeling (ABM), 25–63, 72, Beauty contests, 315
274, 321 Bee hives, 292
Agglomeration, 34, 94, 126, 309, Beer game, 307, 308
319 Behavioral convention, 73
Algorithmic complexity, 263 Behavioral experiments, 206
Alternating cooperation, 231–233 Behavioral norms, 187, 190, 194
Alternations, 218, 223 Behaviour, 303
Analytical theory, 177–183 Bernanke, B., 308
Analytical understanding, 51 Best model, 13
Angular-dependent, 76 Best response, 202, 305
Animal groups, 101 Bi-directional percolation, 169
Anisotropy, 76 Bifurcation, 136
Anomie, 104, 107 Binomial distribution, 256
Antisocial punishers, 49 Birth, 322
Antisocial punishment, 163, 322 Bistability, 134, 137
Anti-voter model, 104 Blackouts, 261, 265, 316
Arms race, 272 Blinker strategies, 212
Arrow, 313 Bollinger, L.C., 2, 262
Artificial intelligence, 321 Boolean grids, 95
Asch experiment, 202 Booms, 310
Aspiration, 202 Bottleneck, 84
Asymmetric interactions, 124–126 Boundary conditions, 40
Asymmetric noise, 125 Bounded confidence (BC) model, 103
Attractor, 30, 286, 288 Boundedly rational agents, 306
Auction, 290 Box, G., 7
Augmented reality, 62 Box plots, 43, 47
Avalanches, 265, 289 Braess, D., 212, 215, 234
Avatars, 207 Braess paradox, 212, 215
Average performance, 59 Brandt, H., 154

DOI 10.1007/978-3-642-24004-1, © Springer-Verlag Berlin Heidelberg 2012
332 Index
Breakdowns, 268, 270, 287 Compromises, 318

of cooperation, 194 Computational modeling, 321
of performance, 314 Computer agents, 206
of solidarity, 312 Computer games, 28
Brown, R., 83 Computer simulations, 26–27, 205, 322
Bubble control, 277 Computing, 53
Bubbles, 269, 307, 310 Comte, A., 1
Business cycles, 270, 307, 308, 319 Conditional probability, 229
Butterfly effect, 269, 286, 287, 306 Conflicts, 185, 191, 194, 322
Conformity, 83, 322
Congestion, 211
Calibration, 40, 76–78 charges, 233
Calibration dataset, 41 games, 215
Capacity, 214 Connectedness, 276
drop, 268, 294 Consciousness, 321
limits, 268 Consensus, 102
Cascading Constants of motion, 46
effects, 55, 266, 288–290 Contingency plans, 59, 278
failures, 265–267, 277, 289 Control, 316
Catastrophe theory, 48, 193, 274, 288 attempts, 270
Catastrophes, 17, 56, 270, 316 parameters, 17, 116
Causality networks, 54 Controllability, 316
Central banks, 317 Conventions, 319, 322
Chain reactions, 265 Convergence, 109, 305
Chaos, 268–269 Cooperation, 131–138, 140, 144, 147, 286–287
Chaos theory, 274 Cooperative behavior, 187
Chaotic, 30, 308 clusters, 143
Chicken game, 185, 212 episode, 221
Classical control, 291 Cooperativeness, 207
Classification, 216–218 Coopetition, 59, 187
Clustering, 108, 109, 143 Coordinated behavior, 58, 169–183, 225
Coalition formation, 33, 322 games, 171, 307
Coarsening, 123 problem, 170
Co-evolution, 139–150, 311 Correlations, 30, 202, 272, 273, 309, 311, 312
Coexistence, 134, 159, 185–197 Costly punishment, 135, 202, 322
Cognitive biases, 305 Counter-intuitive behaviours, 11, 30, 278, 285,
Cognitive capacities, 305 290
Coherent oscillations, 233 Counterflows, 31
Coincidence, 267 Courses of events, 54
Coleman, J.S., 83 Crashes, 307, 310
Collective, 303 Crazes, 82
decision making, 321 Creativity, 318, 321
intelligence, 33, 80, 274 Crime, 322
opinions, 310 Crisis response, 62
Combinatorially complex, 285 Critical
Comfort C, 93 crowd conditions, 92–93
Communication, 322 fluctuations, 270, 288
Companies, 287, 319 infrastructures, 289
Competition, 169, 286–287, 322 perturbation, 266
Complementary strategies, 248 phenomena, 30, 274
Complex dynamics, 10 points, 10, 17, 56, 264, 265, 288, 316
Complex systems, 9–11, 285–290, 321 size, 275
Complexity, 262 slowing down, 288
Compliance, 251, 253 Criticality, 267
Index 333
Crowds, 287 Domino effects, 265

behavior, 322 Double moral standards, 158, 159
disasters, 91 Driver information systems, 213
dynamics, 34 Dunbar’s number, 53
pressure, 90–92 Durkheim, E., 104
Cultural Durkheimian model, 107, 110
backgrounds, 62 Dynamic complexity, 263
evolution, 195
Culture, 322
Cusp catastrophe, 270 Early warning signals, 270
Cybernetics, 27, 274 Earthquakes, 92, 272
catastrophe theory, 274 Ecological perspective, 27
control, 317 Ecological systems thinking, 313–314
Ecology, 313, 319–321
Economic
Daamen, W., 81 crises, 261
Data, 322 growth, 319
Data-mining, 62, 322 mechanisms, 37
Data-poor, 61 modeling approach, 18
Data-rich, 61 models, 28
Death, 322 instability, 303
Decentralized Econophysics, 319–321
control, 275 Ecosystems, 131
coordination, 274 E1 Farol bar problem, 217
Decision-making, 58, 321 Effects, 269, 278
Decision guidance, 251 Efficiency, 93
Decompartementalization, 276, 277 Efficient market hypothesis, 306–307
Decouple, 276 Efficient markets, 248
Defection, 132 eGovernance, 62
Delays, 270, 275 Eigendynamics, 277
Demographic change, 303 Eigenvalues, 132
Deregulation, 272 Einstein, A., 37
Derivatives, 272 Eldakar, O.T., 154
Design, 274 Emergence, 29–33, 220–223, 247–257
Design elements, 72 phenomena, 11, 36
Desired speed, 74 properties, 30, 274
Detailed models, 5–6 of punishment, 163
Differential equation, 44 Emotions, 318, 322
Differential game, 74 Empathy, 322
Differentiation, 294 Empirical evidence, 36
Diffusion, 320 Epidemic spreading, 265
Direct reciprocity, 135 Epstein, J.M., 8, 26
Directed random walks, 310 Equation-based modeling, 27
Disaster response, 267, 278 Equilibrium, 285
Disasters, 261 creation, 132, 136
Discontinuous transitions, 197 displacement, 132, 136
Discretization, 42, 44 model, 51
Discrimination, 322 paradigm, 307–309
Disequilibria, 272 selection, 132, 136
Disruption, 266 Errors, 42, 77
Distributions, 47 bars, 47
Diversity, 59 function, 40
Doerner, D., 271 Escape, 86
Dominance, 174 Escape route, 95
334 Index
Evacuation, 82–84 Free-riders, 153, 202

Evacuation dynamics, 82–94 Freeway traffic, 263
Evolution, 131, 153–164 Freezing by heating, 31, 87–88, 124
Evolutionary Frey, B.S., 318
algorithms, 73 Frictional effects, 277, 287
biology, 321 Frustrated states, 271, 287, 294
dead ends, 314 Functional, 46
game theory, 73 complexity, 10
optimization, 93–94 noise, 203
theory, 313
Excellence, 59
Exchange, 322 Gantt diagrams, 295
Experimental research, 322 Gases, 78–79
Experimental subjects, 206 Gaussian noise, 43
Experiments, 12, 202 Genetic relationship, 194
Explanatory models, 61 Ghetto formation, 119
Exploratory behavior, 42 Global market, 262
Exponential decay, 78 Globalization, 276, 294
Extreme events, 30, 274 Goal function, 59–60
Goal selection, 314
Godel’s undecidability theorem, 15
Failure, 265 Goodhart’s law, 53, 56, 269, 316
Failure cascades, 265–267, 277, 289 Goodness of fit, 37, 42
Fairness, 59, 305, 317 Granular media, 78–79
Fashions, 310 Gravity law, 53
Faster-is-slower effect, 88–90, 294 Greenspan, A., 308
Feedback effect, 58 Gridlock, 214
Feedback loops, 265, 290 Group
Fermi rule, 175, 202 dynamics, 33, 322
Financial instabilities, 261 formation, 310
Financial market instability, 271 identity, 322
Financial markets, 34, 317, 319 performance, 251
Fingering, 79 pressure, 132, 137
Firewalls, 276, 278 selection, 135
Firms, 34 Guided self-organization, 291
Fisher equation, 53 Guyer, M., 217
Fish schools, 101
Fitness, 77, 310
Fix point, 196 Harmony game, 132
Flexible, 275 Hawk-dove game, 185
adaptation, 59 Heavy-tail statistics, 273
adjustment, 275 Heckathorn, D.D., 154
Flocks of birds, 101 Herding, 86–87
Fluctuation-induced, 116 behaviour, 83, 85, 269
Fluctuations, 42–43, 115, 294 effects, 273, 294, 307, 310, 318
Fluctuations strength, 120 Heterodox, 302
Fluids, 78 Heterogeneity, 12, 27, 51, 162, 185–197, 206,
Folk theorem, 305 272, 277, 294, 311, 312, 315
Forecasts, 7, 11, 54, 239, 291 Heuristics, 45, 206
Forestry science, 313 Hierarchical organization, 2, 292
Fragmentation, 109 Hierarchies, 291–293
Framing effects, 206 History, 318
Free market, 306 History dependence, 48
Free-ride, 132 Homo economicus, 29, 304–306
Index 335
Homophily, 49, 102, 321 Irregular grids, 161

Hoogendoorn, S.P., 81 Irreversibility, 309
Human error, 267 Ising model, 320
Human factors, 318
Hypocritical co-operators, 154
Hypotheses, 36 Jamming, 287
Hysteresis, 48, 285
Kin selection, 135

Identical strategies, 248
Kretz, T., 81
Illusion of control, 269–271, 316
Krugman, P., 278, 301
Imitation, 142, 322
Kuhn, S., 15
Immoralists, 154
Incentive structures, 272
Indirect reciprocity, 135, 321
Individual Laboratory experiments, 202, 321
histories, 206 Lane formation, 31, 79–81, 121
performance, 59 Language, 194, 311, 322
strategies, 251 Law of relative effect, 248
Individualization, 107 Le Chatelier principle, 270, 278, 290, 316
Induced transitions, 116 Leader game, 212
Inequality, 322 Learning, 30, 73, 211–234, 254, 322
Information, 239, 318 curves, 254
management, 303 process, 73
overload, 315 Level-of-service concept, 72
systems, 56–57 Leverage, 273
In-group, 187 Lewin, K., 73
cooperation, 194 Limit cycle, 286
interactions, 187 Limitations, 36, 52–53
Inhibitory effects, 309 Linear models, 11, 36, 309–310
Initial condition, 40, 48 Linear systems, 285
Innovations, 30, 225, 271, 294, 321 Living labs, 207
dynamics, 320 Local optima, 56, 285, 305
spreading, 169–183, 319 Logic of failure, 271
Instabilities, 56, 261, 271, 318 Lomborg, B., 302
Institutional design, 303 Long-term forecasts, 57
Institutions, 322 Lords of Warcraft, 207
Integration, 185–197, 322 Lose-shift strategy, 229
Integrative systems design, 58, 278 Loss of control, 59
Intelligent driver model, 31 Lotka–Volterra equations, 308
Interacting particle systems, 104 Love, 317
Interaction, 74, 202 Low performance, 278
force, 74 Lucifer’s positive side effect, 155
networks, 40, 289 Lyapunov exponents, 50
range, 74 Lyapunov function, 46
strength, 74
Interbank market, 312
Interdependencies, 27 Macro-economic models, 34
Intermittent flows, 88–90 Mainstream economies, 320
Internet communication, 101 Managing complexity, 55–56, 274–277,
Interpretation, 318 285–297, 321
Intersecting flows, 81–82 Market dynamics, 322
Invisible hand, 191, 275, 307 Markets, 246, 248, 262, 269, 278, 306, 312,
Irreducible randomness, 270 317, 319, 322
336 Index
Mass hysteria, 273 Multi-stability, 30, 48

Mass media, 101 Multi-stage supergame, 219
Mass psychology, 83
Master equation, 256–257
Matthew effect, 53 Narratives, 4
May, R.M., 201 Nash equilibrium, 307, 309
Mean field approximation, 51, 311 Nash flows, 215
Measurement concepts, 321 Negative influence, 102
Measurements, 41, 321 Neoclassical economics, 313
Mechanism design, 56, 58, 271 Nervous, 84
Meltdown, 273 Netlogo, 38
Memorize, 322 Network, 11, 135, 265–267, 272, 274, 311, 319
Memory effects, 309 analysis, 274
Metastable, 266 interactions, 265–267, 311
Micro-economic models, 34 interdependencies, 272
Micro-macro link, 313–314 models, 319
Microscopic level, 313 reciprocity, 135
Migration, 34, 139 Neuroscience, 321
Milgram experiment, 207 Noise, 42, 102, 105, 140, 162, 175, 203
Minimal model, 116 level, 203
Mining, 322 term, 105
Minorities, 194 Noise-induced, 116, 121, 125, 294, 296, 310
Minority game, 244 ordering, 121, 294, 296
Mintz, A., 84 self-organization, 125
Mistakes, 206 transitions, 116, 310
Mobility, 139, 322 Non-linear, 263, 269, 274
Modeling, 1–21, 25–63, 274 dynamics, 269, 274
forecasts, 54 interactions, 263
predictions, 54 Normative behavior, 187–191
validation, 41–42 Nowak, M.A., 134, 201
Momentum conservation, 79 Nowcasting, 61
Monocultures, 102, 108, 294 NP-hard, 45, 285, 305, 315
Monopolies, 176 optimization problems, 305
Montroll, E.W., 319 problems, 45
Moore neighbourhood, 164 Nucleation, 104
Moral Nucleus, 146
behavior, 153–164
issues, 318
sentiments, 306 Occurrence probability, 224
Moralists, 153 Ockham’s razor, 7
More is different, 309 Offers, 223
Movers, 248 One-shot game, 233
Multi-agent simulation, 25, 228–233 Opaqueness, 272
Multi-disciplinary collaborations, 21, 320 Opinion
Multi-modal distributions, 47 clustering, 101
Multi-nomial logit model, 175 dynamics models, 103, 202, 321
Multiple world views, 20–21 formation, 33, 101–111
Multiplicative noise, 310 polarization, 101
Multi-population harmony game (MHG), shifts, 310
188 Optimization, 206, 247
Multi-population prisoner’s dilemma (MPD), Optimization goal, 314–316
188 Order parameters, 314
Multi-population snowdrift game (MSD), 188 Organization, 115
Multi-population stag hunt game (MSH), 188 Oscillations, 287
Index 337
Oscillatory, 30, 212, 223, 307 Planning guidelines, 15

behavior, 51, 223 Plausibility, 38–39, 72
flows, 81 Pluralism, 46–47, 294
Other-regarding behaviour, 51, 81, 213, 223 Pluralistic, 59, 102, 107, 176
Outbreak of, 305 goal function, 60
Out-group interactions, 140, 144, 147 modeling, 20, 53
Out-of-equilibrium models, 30, 187 Plurality, 20–21, 53, 60
Over-confidence, 11 Political entrepreneur, 194
Over-critical perturbation, 273 Possibilistic modeling, 313
Overcrowding, 266 Potential games, 20–21
Over-fitting, 83 Potential payoff, 215
Overlay phase diagram, 37 Power lawsr, 242
Overreaction, 49 Predictability, 17, 30, 47, 263–265, 288–290,
310
Prediction, 268–269
Panics, 249 Prediction markets, 201–207, 278
Paradigm shifts, 82, 84–85, 310 Predictive power, 7, 37
Paradoxes, 48 Pressures, 41
Paradoxical, 234 Prigogine, I., 84
behaviors, 262 Principle of parsimony, 270, 316
results, 47 Principles, 7
Parallel worlds, 47, 262 Prisoner’s dilemma, 270, 274, 290
Parameter fitting, 62 Probabilistic, 84, 131, 132, 140, 185, 212, 217,
Parameter space, 315 231–233
Parameter values, 48 Probabilistic decision, 262
Pareto optimum, 46 Pro-cyclical effects, 249
Pareto’s principle, 53 Production plants, 273
Partial differential equations, 234 Production systems, 287
Participation, 44 Properties, 58, 320
Path-dependent effect, 59, 303 Proportional imitation, 3, 27, 274
Pattern formation, 309 Protest movements, 202
Payoff, 287, 314 Psychological modeling approach, 193
Payoff matrices, 118, 131, 214, 287 Psychology, 19
Peace, 171 Public goods, 273, 321
Pedestrian, 303, 310 Public goods game, 132, 201
crowds, 80 Public health, 131
dynamics, 72–78 Public infrastructures, 303
Percolation, 31, 72–78, 80 Public media, 311
Percolation thresholds, 169, 172, 173, 312 Punishment, 273
Performance, 173 cost, 135, 154, 202, 322
Periodic, 44, 294 fine, 154
Perturbations, 308 Pushing, 140, 194, 220–223, 311
Phantom panic, 54, 266 Puzzles, 35, 84
Phantom traffic jams, 88–90
Phases, 14, 17, 30, 33, 48–49, 56, 133–134,
Qualitative descriptions, 33, 202
190, 205, 270, 288, 308, 316
Quality control, 4–5
diagrams, 14, 33, 48–49, 190, 205, 308
Quantiles, 315
separation, 49
Quasi-periodic, 43
transitions, 17, 30, 48, 56, 133–134, 270,
Queueing models, 308
288, 316
Quick response, 72
Physical models, 18
Physical(istic) modeling approach, 3, 17
Physicalist approach, 13 Randomization, 277
Planck, M., 28 Randomness, 94
338 Index
Random number generators, 162, 203, 268, Scenario analyses, 49

269 Scenario modeling, 62
Random relocations, 43 Schelling, T.C., 274
Rapoport, A., 144 Schreckenberg, M., 278
Ratchet effects, 217 Schumpeter, J., 215, 240
Rationality, 310 Scientific innovation, 313
Reality mining, 304 Scientific revolutions, 321
Real-time estimates, 61 Second life, 15
Recessions, 61 Second-order free-rider problem, 207
Reciprocity, 310 Second-order free-riders, 202, 311
Recommendations, 135, 163, 213, 218, 239, Security, 153
321 Security breaks, 303
Recycling, 57, 252 Segregation, 278
Redundancies, 313, 315 Segregation phenomenon, 34, 119, 322
Refusal probability, 276 Self-consciousness, 80
Regime shifts, 252, 270 Self-control, 321
Regional development, 10, 17, 56, 263, 270, Self-destroying prophecies, 54, 271, 275, 278
316 Self-fulfilling prophecy, 54, 56–57, 269, 317
Regression, 34 Self-organization, 30, 202, 273, 291
Reinforcement learning, 72 criticality, 30, 267, 273, 288–290, 310
Relaxation time, 229, 247–257 networks, 291
Relocations, 74 Self-regulation, 29–33, 36, 58, 71–96,
Renewable resources, 139, 144 115–128, 271, 275, 286–287
Repeated prisoner’s dilemma, 313 Selten, R., 278
Replicator equations, 212 Sensitivity, 215, 240
Representative agent approach, 51, 131, 187 Sensitivity analysis, 50, 269
Reproducibility, 310–313 Separation, time scales, 42
Reproduction, 46 Settlements, 115, 293
Reputation, 322 Shadow of the future, 309
Residual interactions, 321 Shell model, 194, 311
Resilience, 293 Short-term forecasts, 90
Resource-efficient, 59, 314 Side effects, 57
Response patterns, 274 Signaling, 265
Revolutions, 192, 247–257, 266 Similar behaviors, 318, 322
Reward, 15, 48, 193, 194 Simple models, 247
Risk aversion, 60, 140, 321 Singularities, 6–9
Risk models, 291 Skewed, 30
Road pricing, 272 Sliding friction force, 47
Robustness, 213, 233 Slow relaxation, 30, 295
Rock-paper-scissors game, 54, 271 Slower-is faster effects, 85
Roles, 212 “Small-world” effects, 292
Route choice game, 72, 211, 251, 322 Small-world effects, 270
Route guidance system, 211, 213–217, 239 Smith, A., 292
Routes to cooperation, 234 Snowdrift game, 306, 312
Rules of the game, 132–134 Social, 34–35, 133, 185, 191–193, 207, 321
benefit systems, 311
capital, 59, 207, 310
Saddle points, 58 cohesion, 107
Safety margins, 197 conflict, 33
Safety standards, 268, 289 contagion, 83
Sanctioning, 268 cooperation, 33
Scalability, 196 differentiation, 322
Scale-free behavior, 44–45 dilemma, 32, 131–138, 287, 307, 321
Scaling analysis, 17 experiments, 201–207
Index 339
factors, 318, 321 Structural instabilities, 10, 262

fields, 73 Structure formation, 50, 271, 318
force concept, 72–74 Stylized facts, 43
force model, 31, 74–75 Subcultures, 6, 35, 42, 50, 51, 53, 194, 322
forces, 73 Suboptimal, 190
influence, 33, 321 Subpopulations, 271
movements, 322 Success, 117
networks, 33, 169–183, 322 Success-driven migration, 118, 139, 213, 287
norms, 8, 33, 275, 318, 322 Success-driven mobility, 142
order, 79, 201, 207, 275 Sucker’s payoff, 115–128
physics, 1 Supercomputing, 140
roles, 322 Supercritical cluster, 35
structure, 322 Surprise, 146
supercomputing, 34–35, 63 Sustainability, 303, 318
welfare, 314 Swarm intelligence, 38, 287
Socialization, 1, 8, 31–33, 59, 63, 72–75, 79, Switching cost, 11, 287
83, 107, 131–138, 201–207, 275, Synchronization, 171
287, 307, 310, 311, 314, 318, 321, Synergy effects, 287
322 Systemic, 154
Sociological modeling approach, 322 Systems, 30, 261–279, 309
Sociological models, 19 design, 57, 273, 278
Solidarity, 28 dynamics, 34, 261–279
Spatial interactions, 201 optimum, 30, 214, 241, 309
Specialization (heterogeneity), 202 performance, 43
Spencer, H., 12, 294, 312, 322 shifts, 310
Spill-over effects, 3 size, 45
Spreading, 58 theory, 274
Spreading of cooperation, 169–183
Stability analysis, 204
Stable stationary solution, 309 Taking turns, 43, 45, 57, 214, 274, 310
Stadium exit, 137 Taxes, 211, 288
Stag hunt game (SH), 95 Taxonomy, 311
Stampedes, 133, 188–191 Techno-socio-economic-environmental
Stanley, H.E., 82 systems, 217
Stationary solutions, 16 Temptation, 60
Statistical Testing, 140
analysis, 132 Theil’s inequality coefficient, 38
ensembles, 47 Thermodynamics, 42, 50
physics, 43 Tipping points, 319
Stayers, 274 Tobin tax, 10, 17, 30, 56, 136, 263–265, 270,
Stiglitz, J.E., 216, 248 309
Stochastic models, 27 Too big to fail, 57
Stochasticity, 306 Toy model, 276
Stock markets, 72, 246 Trading, 116
Stop-and-go traffic, 269, 307 Traffic, 211, 287, 322
Stop-and-go waves, 270 assistance, 287
Strange attractor, 31, 33, 89–90 congestion, 211, 265
Strategy mutations, 286 dynamics, 34
Stratification, 143 forecasts, 239
Streamlines, 322 light controls, 58, 287
Stress reliefs, 79 Tragedy of the commons, 34, 265, 287
Stripe formation, 272 Trajectories, 153, 201, 287, 307, 311
Strong reciprocity, 80–82 Transaction costs, 76, 190
Structural complexity, 163 Transformation process, 307
340 Index
Transition, 278 Virtual humans, 322

Transition matrix models, 264 Virtual worlds, 62, 207
Translator, 72 Visualization, 47–48
Transport networks, 62 Vitious circles, 79
Traulsen, A., 320 Volatile decision dynamics, 265
Travel time, 233 Volatility clustering, 246
Traveler information systems, 202 Volunteering, 246
Treatment, 214 Von Neumann, J., 131
Trends, 241 Voter model, 104, 153, 202, 318
Trial-and-error behavior, 230, 310 Voting behavior, 104, 202
Trust, 164, 220, 225 Vulnerability, 315, 321
Turbulence, 276, 312, 321
Turbulent flow, 30, 268, 308
Turing patterns, 91 Wardrop equilibrium, 59, 275, 276
Turn-taking, 116 Wars, 215
Weather, 310
Web, 268
Unconditional imitation, 288 Weidlich, W., 62, 207
Uncontrollable, 203 Well-being, 319
Unemployment, 262, 271 Well-mixed interactions, 58, 59
Unholy collaboration, 309 Well-mixed population, 312
Universal, 155, 158 Who laughs last laughs best effect, 171
Universality class, 17 Wikipedia, 155
Unknown unknowns, 116 Wilson, D.S., 311
Unresponsiveness, 270, 318 Win-stay, 154
Unstable dynamics, 316 Win-stay-lose-shift, 229
Unstable stationary solution, 287 Wisdom of crowds, 202, 229
Upwind scheme, 137 World Wide Web, 8, 15, 32, 62
Urban agglomerations, 44 Worrell, D., 311
User equilibrium, 101
Yule-coefficient, 301
Validation dataset, 212, 214, 240
Vector fields, 41
Verification, 190 Z-coefficient, 226
Victory, 38 Zhang, Y.-C., 227
Video tracking, 155 Zig-zag design, 319
Violence, 76 Zipf’s law, 95

Springer Complexity: Editorial and Programme Advisory Board

Uploaded by

Copyright:

Available Formats

Springer Complexity: Editorial and Programme Advisory Board

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

Springer Complexity: Editorial and Programme Advisory Board

Uploaded by

Copyright:

Available Formats

Springer Complexity

Springer Complexity is an interdisciplinary program publishing the best research and

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)

c Springer-Verlag Berlin Heidelberg 2012

Printed on acid-free paper

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

• Chapter 1 discusses the issue of how to describe social systems. It highlights

Zurich Dirk Helbing

1 Modeling of Socio-Economic Systems . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1

D. Helbing (ed.), Social Self-Organization, Understanding Complex Systems, 1

to scientific communities with different approaches and backgrounds. While each

1.2 Particular Difficulties of Modeling Socio-Economic

When speaking about socio-economic systems in the following, it could be anything

1.3 Modeling Approaches

1.3.1 Qualitative Descriptions

1.3.2 Detailed Models

In certain fields of computational social science or economics, it is common to

1.3.3 Simple Models

1.3.4 Modeling Complex Systems

are collective patterns of motion in pedestrian crowds or what is sometimes called

1.4 Challenges of Socio-Economic Modeling

parameter combinations or certain initial conditions). When such conditions apply

1.4.1 Promises and Difficulties of the Experimental Approach

1.4.2 Several Models Are Right

The above mentioned properties of socio-economic systems imply that it is difficult

1.4.3 No Known Model Is Right

A typical approach in the natural sciences is to verify or falsify previously untested

1.4.4 The Model Captures Some Features, But May Be

• Sociological modeling approach: Certain schools of sociologists use rational

1.4.5 Different Interpretations of the Same Model

A further difficulty of modeling socio-economic systems is that scientists may not

other ways. For example, it can be linked to psychological laws or to distributions

1.5 Discussion and Outlook

1.5.1 Pluralistic or Possibilistic Modeling and Multiple World

As shocking as it may be for many scientists and decision-makers to abandon

1.5.2 Where Social Scientists and Natural Scientists

D. Helbing (ed.), Social Self-Organization, Understanding Complex Systems, 25

2.1 Why Develop and Use Agent-Based Models?

2.1.1 Potential of Computer Simulation in the Socio-Economic

2.1.2 Equation-Based Versus Agent-Based Approach

Today, computer-simulation in the natural sciences and engineering mostly rely on

2.1.3 Scientific Agent-Based Models Versus Computer Games

Agent-based computational modeling is also well suited for visualization. Here, it

2.1.4 Advantages of Agent-Based Simulations

2.1.5 Understanding Self-organization and Emergence

• They may have several stationary states (a phenomenon known as “multi-

• While pedestrians moving in opposite directions normally organize into lanes,

• Sometimes, the combination of two dissatisfactory solutions can be the best

Small Perturbation Large Perturbation

ΔQ (vehicles / h / lane) ΔQ (vehicles / h / lane)

of surprise. So far, for example, we do not understand emotions and consciousness,

2.1.6 Examples of Agent-Based Models

The method of agent-based modeling is very versatile. It may be applied, for

• Financial markets [58–60]

2.1.7 Social Super-Computing

2.2 Principles of Agent-Based Modeling

where ˛ with 0 ˛ 1 is a parameter which grows with the strength of inter-