Stark H.U. 2012 - Decision Making in Models of Social Interaction

Diss. ETH No.
19639
Decision making in models

of social interaction
A dissertation submitted to the

ETH ZURICH
for the degree of

DOCTOR OF SCIENCES
presented by
HANS-ULRICH STARK
Diplom-Verkehrswirtschaftler, Technische Universitat Dresden
born September 23, 1977
citizen of Germany
accepted on the recommendation of
Prof. Dr. Dr. Frank Schweitzer, examiner
Prof. Dr. Dirk Helbing, co-examiner
2012
Contents
I Introduction 2
1 Modeling and analyzing social interaction 3

1.1 Micromotives and macrobehavior . . . . . . . . . . . . . . . . . . . . . . 4
1.2 The Weidlich-Haag approach . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Dynamical xed points and their stability . . . . . . . . . . . . . . . . . . 7
1.3.1 A one-dimensional ow . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.2 Multi-dimensional ows . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Decisions through social influence 12

2.1 Decision theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Models of social inuence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 The voter model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 Decision making in (evolutionary) game theory 23

3.1 Classical game theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.1 Denitions and terminology . . . . . . . . . . . . . . . . . . . . . . 23
3.1.2 Coordination- vs. cooperation problems . . . . . . . . . . . . . . . 29
3.1.3 Symmetrical 22 games the framework . . . . . . . . . . . . . . . 32
3.2 Experimental game theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3 Evolutionary game theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.3.1 Evolutionary dynamics . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.2 The conundrum of cooperation . . . . . . . . . . . . . . . . . . . 45
ii
4 Social influence in evolutionary game theory 47
II Models and analyses 52
5 Slower is faster: Fostering consensus formation by heterogeneous inertia 53

5.1 (Social) Inertia in the voter model . . . . . . . . . . . . . . . . . . . . . . . 54
5.2 Preliminary considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.3 Results of binary inertia . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.4 Results of multiple inertia states . . . . . . . . . . . . . . . . . . . . . . . . 68
5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.6 Outlook: inertia-eects in another nonlinear voter model . . . . . . . . . . 79
6 Social influence in social dilemmas 84

6.1 Homogeneous, dynamic random networks . . . . . . . . . . . . . . . . . . . 85
6.2 Evolutionary dynamics with imitators and contrarians . . . . . . . . . . . . 86
6.3 Co-evolution in the spatial Prisoners Dilemma . . . . . . . . . . . . . . . . 89
6.4 The role of imitation in social dilemmas . . . . . . . . . . . . . . . . . . . 95
6.4.1 Motivation and literature . . . . . . . . . . . . . . . . . . . . . . . . 95
6.4.2 The evolution of cooperation through imitation . . . . . . . . . . . 98
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
7 Dilemmas of partial cooperation 116

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.2 Symmetrical 22 games . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.3 Derivation of PCD in repeated games . . . . . . . . . . . . . . . . . . . . . 123
7.3.1 Partial cooperation in repeated games . . . . . . . . . . . . . . . . 123
7.3.2 From social dilemmas to PCD . . . . . . . . . . . . . . . . . . . . . 124
7.4 Derivation of PCD in evolutionary games . . . . . . . . . . . . . . . . . . . 126
7.5 Evolution of partial cooperation . . . . . . . . . . . . . . . . . . . . . . . . 129
7.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
iii
III Conclusion 135
Appendix: Transformations of the standard cooperation model 140
iv
Preface It has always been a hard task to explain in two or three sentences what the
research of my PhD thesis is about. Fortunately, there is room for some more sentences
here and I want to take the opportunity of this Preface to do my best in explaining this
as untechnically as possible.
As is obvious from the title, this thesis deals with social interaction, which is maybe the
most intuitive part: at least two individuals encounter each other and interact in any arbi-
trary way. Such individuals can be humans or animals and the kind of interaction ranges
from talking, helping, and ghting to rather indirect interactions like consumers reacting
to special supermarket oers, animals using the same food resources, or a species pro-
ducing a biological side product which supports another species. These examples already
indicate a number of relevant research elds, e.g. social sciences, psychology, economics,
management, biology, and ecology, thereby pointing out the broad interdisciplinarity of
the research topic.
In the behavioral sciences, one would then formulate questions like:
(i) How do human beings decide/behave in certain situations and what are their reasons,
motivation, and decision mechanisms?
But this is not the kind of question which is addressed in this thesis. In fact, the knowledge
gained by behavioral sciences is used in order to investigate questions of the type:
(ii) Given the behavior of individuals, which consequences arise from the combination of
individual behaviors for a community of many interacting individuals?
In many cases, the consequence of social interaction is not just the sum of individual
actions, but might be interdependent on each other. Let me give a simple example: if a
re emerges during a theater performance, people should get out of the theater as fast as
possible. However, if all the people do their best to do so, the result can be a fatal panic
situation, i.e. the worst outcome where the evacuation can get stuck. This example is also
suited to illustrate that careful investigations of this type may lead to counter-intuitive
results: it can theoretically and practically be shown that deliberately placing obstacles
in front of an emergency exit can help avoiding fatal clogging situations. In other words,
putting obstacles in front of an emergency exit can improve the evacuation during an
emergency event (see Helbing et al. (2005a) for details and other results).
Modeling social interaction is very important for the research in this example, but also
in many other cases. By modeling I mean the abstraction of real social behaviors and
interactions into mathematical equations that capture the most important characteristics.
v
Solving these equations analytically or by means of numerical computer simulations, one
gains a quantitative understanding of causes and eects under the conditions covered
by the equations. Another advantage is that one can eciently change these conditions
and investigate dierences between dierent scenarios. Such an approach also allows to
investigate questions of the type:
(iii) How can we understand and explain facts in our environment, that seem to be con-
tradictory to current knowledge?
Such a question is closely connected to the previous ones. If one understands the main
driving forces leading to the status quo, a lot is learned about the behavior of subjects
under consideration (type (i) questions). This in turn improves the models of social
interaction, which then are used to answer questions of type (ii).
In this thesis, questions of the type (ii) and (iii) will be discussed. In the models, social
individuals are characterized by one or more certain attributes and a behavior for the
interaction with other individuals. To illustrate this, let me mention a simple example: the
famous segregation model of Thomas C. Schelling (Schelling, 1978). There are two types of
individuals that dier in one ethnical trait or, for example, in their religious beliefs. They
all live in one area - maybe a small town - and the arrangement of houses is simplied to a
regular grid, where houses are located at the intersections of lines. Hence, every house is
surrounded by four other houses, but not all of them are inhabited. As a simple behavioral
rule, it is assumed that an individual moves to a random, empty house if more than half
of its neighbors do not share its belief. With this simplistic model, Schelling described
how uncoordinated actions (individuals move to a random node) can lead to a state with a
high level of global ordering (segregation). Whereas Schelling initially analyzed this model
on a sheet of paper, other, still simplistic models require more quantitative involvement.
In combination with the fact that there are parallels between social individuals and the
physical interaction of atoms or other particles, the interdisciplinarity of this thesis subject
is even more broad. Especially in the last decade, a vital exchange of ideas and models
between the sciences mentioned above and mathematics, physics, chemistry, and computer
science has been and still is developing. One of the results is the emergence of a new
scientic branch called socio- or econophysics, where methods of physics are applied to
social or economic systems. The content of the present work does also exhibit this kind of
interdisciplinarity.
Finally, let me introduce the specic kinds of decision making within the social interaction
that are investigated in this thesis social inuence and strategic decisions, where
the latter refers to following an arbitrary strategy in the game theoretic sense, and the
vi
former is restricted to individuals reactions to observed behavior. The dierence lies in
the characteristics of the individuals that are considered in the models. To some extent,
the considered individuals undergo a cultural evolution in the course of this work: they
start as individuals with zero condence. They do not consciously decide for a specic
alternative, but only imitate other individuals. Imitation is a very simple form of decision
making and when I speak of imitation throughout this work, I mean the most primitive
form of imitation: an individual does not imitate because it tries to copy a successful
behavior, but it just imitates regardless of success considerations. One example is herding
behavior, which might have its origin in the absence of relevant information or cognitive
abilities. Moreover, these primal individuals are also zero condent because they are
not tied to a once taken decision, but readily imitate another one. Let us say that these
individuals are entirely susceptible to social influence. There are a number of models (some
of them will be introduced in the main body of this thesis) dealing with such behavior
and the main conclusion is quite intuitive: individual tendencies to imitate each other
lead to highly homogeneous decisions among individuals. However, there are aspects that
are not straight-forward and one can even nd counter-intuitive results that need to be
understood. At least one of such surprising results will be presented and carefully analyzed
within this thesis.
The considered individuals then develop a certain kind of condence. Unlike before, they
take a decision for some reason and their readiness to switch decisions is reduced compared
to the zero condent individuals. Moreover, their condence might evolve in time an
individual that took one and the same decision many times already is assumed to be highly
condent about it and has an inertia to change to another decision. It is one of the main
aims of this thesis to thoroughly investigate how such behavior inuences the theoretical
results of the respective models.
In a next evolution step, the individuals develop more sophisticated decision behavior.
They well distinguish between the dierent decision alternatives and their consequences
(success) and directedly decide for one of them. They might even be capable of memorizing
past interactions and developing adaptive decision behavior. I call this strategic decision
making, because considerations about the possibilities of interaction partners, or their
behavior in the past, might enter the decision making process of an individual. Such a
success-oriented behavior is also essential for the individuals, because there is competition
in their environment. The success of taken decisions is evaluated and compared with
other individuals. Only successful individuals are able to survive the competition and to
reproduce. In consequence, successful behavior can spread in the environment, whereas
unsuccessful behavior might face extinction. In my own view, it has been exciting to
realize that there are respective models capable of capturing the essential principles of
vii
natural selection in this way.
Finally, I will compare the nature of the primal individuals with the one of self-condently
deciding individuals. Although it seems unfair, I will even put them together in the same
competitive environment and let them struggle for life, as Darwin has put it (Darwin,
1859). Unfair or not, it seems reasonable to assume that there is both zero-condent imi-
tation behavior and more strategic decision making present in Nature. The results of this
experiment disclose surprising consequences of this coexistence and yield new insight into
one of the most discussed, interdisciplinary questions in current science: if cooperative
behavior is individually suboptimal, how can such behavior be richly present in todays
world? In fact, this is one of the questions Darwin left open, because, following his theory,
natural selection favors the most successful behaviors in the long run. In this work, some
pieces of evidence will be added in order to overcome the seeming discrepance between
cooperation and evolution.
The present thesis concludes with a more conceptual work on game theoretic decision
situations themselves. The modeling of the individuals environment is of course very
important, because the validity of theoretical results is restricted to the specications of
the model. It is always a crucial, but not simple task to nd a reasonable compromise
between keeping models simple (in order to derive precise conclusions) and making them as
realistic as possible. In this nal part, I argue that with a dierent look at an established
methodology, one can cover a wider range of strategic decision situations and possibly
obtain more explanations of real phenomena. Besides some concrete results, this part
consists of a methodological argumentation and serves as an outlook on potential further
research in this eld.
viii
Abstract Investigating social interaction systems is highly complex because no social
individual behaves exactly like another and, in many cases, the action of an individual i
causes eects on other individuals, that in return may act to cause eects on the behavior
of individual i. Therefore, we nd the main sources of complexity; heterogeneity and the
existence of feedback loops. Both are sources of non-linear dynamics, which often lead to
unexpected, sometimes even paradoxical, results for dynamical systems. However, under-
standing the intrinsic, dynamical properties of social interaction systems is very important
for regulating such a system towards a desired state, or at least to avoid measures that
are disastrous for the social system. Moreover, there is a large overlap between social
interaction and the interaction between biological species or physical particles. Therefore,
especially in the last decade, there is growing interest in the interdisciplinary elds of
socio-physics (applying physical and mathematical laws to social interaction) and evo-
lutionary game theory (applying models of strategic decision theory to the evolution of
social behaviors, biological species, and ecological populations). Both approaches aim at
a quantication and analysis of complex systems, which is only possible if the complex-
ity is largely reduced. It results in a critical trade-o between complexity reduction and
the explanatory power of the analyzed model, which is an important part of the research
challenges in these elds.
In this thesis, we investigate social interaction systems from two perspectives. The rst is
called opinion dynamics, which emphasizes the role of social inuence in decision mak-
ing. The second is evolutionary game theory, where the decisions are not only depending
on social inuence, but might follow any decision rule. While the work within this thesis
contributes novel insight into both elds, the central aim is to reveal the large common
ground and to argue that knowledge gained in one eld is very likely to be relevant for the
other eld. Beginning with the eld of opinion dynamics, we investigate a co-evolutionary
dynamics of imitation and conviction, where some individuals quickly imitate their neigh-
bors, while others rather keep their own opinion. Furthermore, the behavior of individuals
steadily changes depending on their interaction history. We report on surprising results,
where a system with a higher level of conviction promotes the emergence of consensus
in comparison to a system where all individuals readily imitate. We then apply a very
similar co-evolutionary dynamics to evolutionary game theory, specically to the evolu-
tion of cooperating and non-cooperating individuals. Since cooperating individuals always
provide help to others at their own cost and following the selective rules of evolution, non-
cooperating individuals are generally expected to displace cooperating individuals in the
course of time. It is one of the most relevant, scientic problems to nd suitable expla-
nations of the high level of cooperation in Nature and our cultural environment. With
the applied co-evolutionary rule, where some individuals are more inuential than oth-
ix
ers and this characteristic again changes depending on the interaction history, we nd
that this promotes the persistent survival of cooperating individuals, although initially
cooperating and non-cooperating individuals have the same level of inuence. Therefore,
this co-evolutionary setup is a good example to illustrate the strong connection between
opinion dynamics and evolutionary game theory. In a further step, we combine the purely
imitational behavior of opinion dynamics and the more general decision behavior of game
theory together in a novel model of evolutionary game theory. Also in this approach we
nd remarkable results, indicating that the pure existence of imitational behavior might
explain the evolutionary success of cooperative strategies. We therefore can conclude with
the proposition we started from: opinion dynamics and (evolutionary) game theory are
two perspectives on the same matter social interaction systems.
x
Kurzfassung Die Untersuchung sozialer Interaktionssysteme ist komplex: Einerseits
verhalten sich keine zwei Individuen exakt gleich und andererseits hat eine Aktion eines
Individuums i Einuss auf die Aktion eines anderen Indivduums j, was wiederum zu
einem Einuss auf i fuhren kann. Damit sind die Hauptursachen von Komplexitat vorhan-
den: Heterogenitat und die Existenz von Ruckkoppelungseekten. Beides fuhrt zu nicht-
linearer Dynamik, die unerwartete, teilweise sogar paradoxe, Resultate hervorbringen
kann. Ein Verstandnis der inharenten dynamischen Eigenschaften sozialer Interaktionssys-
teme ist aber wichtig, um sie ezient regulieren zu konnen oder sie zumindest vor drastis-
chen Schaden bewahren zu konnen. Zudem gibt es Parallelen zwischen sozialer Interak-
tion und der Interaktion zwischen biologischen Arten oder physikalischen Teilchen, was
besonders im letzten Jahrzehnt zu einem wachsenden Interesse an interdisziplinaren Wis-
senschaften wie der Sozio-Physik (Anwendung physikalischer und mathematischer Metho-
den zur Beschreibung sozialer Interaktion) und der evolutionaren Spieltheorie (Anwendung
von Modellen strategischer Entscheidungstheorie auf die Evolution sozialen Verhaltens, bi-
ologischer Arten und okologischer Populationen) gefuhrt hat. Beide Disziplinen streben
eine Quantikation und Analyse komplexer Systeme an, wozu eine Verringerung der realen
Komplexitat vonnoten ist. Daraus ergibt sich ein Widerstreit zwischen der notwendigen
Reduktion von Modell-Komplexitat und der Aufrechterhaltung der Aussagekraft des un-
tersuchten Modells, was einen entscheidenden Teil der wissenschaftlichen Aufgaben in
diesem Bereich darstellt.
In der vorliegenden Arbeit werden soziale Interaktionssysteme aus zwei Perspektiven be-
trachtet. Die erste ist die Untersuchung von Meinungsbildungsprozessen, bei welcher
die Rolle von sozialem Einuss auf die Entscheidungen Einzelner hervorgehoben wird.
Die Zweite ist die der evolutionaren Spieltheorie, wobei Entscheidungen nicht nur durch
sozialen Einuss bestimmt werden, sondern jeglicher Art von Entscheidungsregel entsprin-
gen konnen. Wahrend die Arbeit in dieser Dissertation neue Erkenntnisse fur beide
Disziplinen hervorbringt, ist es das zentrale Anliegen, die gemeinsame Grundlage bei-
der Gebiete darzustellen und herauszuarbeiten, dass Resultate in einem Gebiet mit groer
Wahrscheinlichkeit auch Relevanz fur das jeweils andere Gebiet haben. Wir beginnen mit
Meinungsbildungsprozessen und untersuchen eine koevolutionare Dynamik von Imitation
und Uberzeugung, wobei einige Individuen bereitwillig ihre Nachbarn imitieren, wahrend
andere eher ihre Meinung behalten. Auerdem andert sich dieses Verhalten der Indi-
viduen in Abhangigkeit ihrer Interaktions-Historie. Wir beschreiben das uberraschende
Ergebnis, dass ein System mit einem hoheren Grad an Uberzeugung die Enstehung von
Konsens unterstutzt, wenn man es mit einem System von reiner, bereitwilliger Imitation
vergleicht. Anschlieend wenden wir eine vergleichbare koevolutionare Dynamik auf die
evolutionare Spieltheorie an, im Speziellen auf die Evolution von kooperierenden und nicht-
xi
1
kooperierenden Individuen. Da Kooperierende jederzeit anderen Individuen auf ihre eige-

nen Kosten helfen und die Dynamik den selektiven Regeln der Evolution folgt, sollten die
Kooperierenden im Laufe der Zeit von nicht-kooperierenden Individuen verdrangt werden.
Es ist eines der uberaus relevanten wissenschftlichen Problemstellungen herauszunden,
wie das hohe Ma an Kooperation in unserer naturlichen und kulturellen Umwelt zu
erklaren ist. Die koevolutionare Dynamik, bei welcher einige Individuen einussreicher
sind als andere und erneut die Auspragung dessen von der Interaktions-Historie abhangt,
fuhrt zu einer deutlich groeren Uberlebenswahrscheinlichkeit fur Kooperierende, obwohl
Kooperierende und Nicht-Kooperierende ursprunglich gleich viel Einuss haben. Daher
ist dieser koevolutionare Modell-Ansatz ein gutes Beispiel um die Verbindung zwischen
Meinungsbildungsprozessen und evolutionarer Spieltheorie zu veranschaulichen. In einer
weiteren Untersuchung kombinieren wir das rein imitative Verhalten von Meinungsbil-
dungsprozessen mit dem allgemeineren Entscheidungsverhalten der Spieltheorie in einem
neuartigen Modell-Ansatz der evolutionaren Spieltheorie. Auch in diesem Ansatz erhalten
wir beachtenswerte Resultate: Die bloe Existenz von Imitation kann unter Umstanden
den evolutionaren Erfolg von Kooperation erklaren. Zusammenfassend kann man die
aufgestellte Pramisse dieser Arbeit wiederholen: Meinungsbildungsprozesse und (evolu-
tionare) Spieltheorie sind zwei verschiedene Perspektiven auf ein und denselben Zusam-
menhang soziale Interaktionssysteme.
Part I
Introduction
The present thesis investigates models of social interaction from dierent perspectives.
This introductory Part consists of four Chapters. Chapter 1 aims at explaining and mo-
tivating the quantitative research approach taken in this work. In Chapters 2 and 3, we
mainly introduce the elds of opinion dynamics and (evolutionary) game theory, which
constitute the main subjects of this work. The rst one is what we refer to as models of
social inuence and the latter represent investigations on strategic decision situations.
Finally, Chapter 4 relates both elds to each other and reveals one of the goals of this
thesis: showing that both approaches mainly dier in their perspective on social interac-
tion and using this proposition to investigate the impact of social inuence on models of
evolutionary game theory.
Chapter 1
Modeling and analyzing social

interaction
Within this rst Chapter, we aim at preparing the general ground for the discussions,
models, and analyses of the present thesis. We want to motivate the central approaches
applied in the thesis and address their diculties and critical points. In particular, we
introduce the so-called micro-macro link, that forms the central fundament of all modeling
techniques within the presented material of Part II. Furthermore, based on an exemplary
model, we motivate and introduce the quantitative modeling approach taken within this
thesis and discuss and the ubiquitous challenge attached to it: nding a reasonable balance
between needed simplications and remaining explanatory power of the models.
Let us rst specify what exactly is meant by social interaction. A sociological lexicon
denes it as:
Interaktion, soziale, [1] die durch Kommunikation (Sprache, Symbole,

Gesten usw.) vermittelten wechselseitigen Beziehungen zwischen Personen und
Gruppen und die daraus resultiende wechselseitige Beeinussung ihrer Einstel-
lungen, Erwartungen und Handlungen. [...]
Lexikon zur Soziologie (Fuchs-Heinritz et al., 1994)
Freely translated by the author, this means that social interaction concerns the interactive
relations between individuals or groups, mediated via communication (language, symbols,
gestures, etc.) and the resulting, interactive inuence on the individual attitudes, expec-
tations, and actions.
3
4 Chapter 1. Modeling and analyzing social interaction
1.1 Micromotives and macrobehavior

Focusing on social interaction, it is less important how an individual reacts to a certain
situation and why it does so, but one is mainly interested in the resulting eects out
of the interaction between individuals. The particular interest of the present work is on
interaction outcomes that are not primarily intended and assessed by the involved social
entities (that generally will be called individuals in this work, but in special cases also
voters or players), but result from inherent (mathematical) properties of the social
system. When we speak of a (social) system here, in simple words, we mean the collectivity
of social interactions (that might be an emergent phenomenon itself). The heading of this
Section, which we borrowed from an excellent book by Thomas C. Schelling (Schelling,
1978), sketches the central theme of the present work: investigating macroscopic eects of
microscopic actions.
For this purpose, it is often necessary to reduce the complexity of real interaction scenar-
ios to much more simple interaction models. Only then it is possible to quantitatively
investigate the inuence of concrete factors on the dynamics of the system. In order to do
so, in this thesis methods of agent-based modeling, multi-agent simulation, statistical and
analytical methods, and evolutionary game theory are employed.
Schelling, in conclusion of his segregation model, wrote:
The model that is described is limited in the phenomena it can handle be-
cause it makes no allowance for speculative behavior, for time lags in behavior,
for organized action, or for misperception. It also involves a single area rather
than many areas simultaneously affected. But it can be built on to accommodate
some of those enrichments.
Thomas C. Schelling (Schelling, 1978)
The models and analyses of this thesis follow this modeling philosophy. We aim at a better
understanding of how the existence, structure, and dynamical properties of interactions
inuence the decisions of individuals in the system. Every day in life basically consists of
a steady sequence of decisions to be taken by individuals. Such decisions can be conscious
and well elaborated, or they can be taken unconscious, following primitive instincts. A
decision is the result of a selection among alternatives. If many individuals have to take a
common decision, e.g. a board of directors appointing a chief executive ocer, one speaks
of collective decision processes and social choice theory (Arrow, 1970; Suzumura, 1983).
Here, but also for most other decisions, it holds that an individual decision does not only
1.2. The Weidlich-Haag approach 5
aect the individual itself, but has eects on other individuals either. Therefore, in social
systems, economics, politics, etc. it is important to understand individual decision making
(i.e. on the microscopic level) and its eects on a system comprising many individuals
(the macroscopic level).
In sociology, economics, psychology, and political sciences, the individuals under consider-
ation are human beings. However, social interaction is not restricted to human beings, but
relevant for biological species in general. Especially the so called social insects, like ants
and bees, but also swarming birds and shes are good examples to illustrate the impor-
tance and consequences of interaction between individuals (Bonabeau et al., 1999; Choe
and Crespi, 1997). Therefore, models of social interaction are not necessarily tied to hu-
man investigations, but are frequently applied in biological and ecological sciences (Chave,
2001; Reichenbach et al., 2007).
In their book on Growing articial societies, Epstein and Axtell develop an extensive
computer model with the aim to simulate the evolution of a whole society from bottom
up (Epstein and Axtell, 1996). Especially the way they relate their work to the social
sciences has many parallels to the approach taken in this thesis.
We apply agent-based computer modeling techniques to the study of hu-

man social phenomena, including trade, migration, group formation, combat,
interaction with an environment, transmission of culture, propagation of dis-
ease, and population dynamics. Our broad aim is to begin the development
of a computational approach that permits the study of these diverse spheres
of human activity from an evolutionary perspective as a single social science,
a transdiscipline subsuming such fields as economics and demography. [...]
We believe that the methodology developed here can help to overcome [some
problems of social sciences]. This approach departs dramatically from the tra-
ditional disciplines, first in the way specific spheres of social behaviorsuch as
combat, trade, and cultural transmissionare treated, and second in the way
those spheres are combined.
Joshua M. Epstein and Robert Axtell (Epstein and Axtell, 1996)
1.2 The Weidlich-Haag approach

In 1983, a fundamental basis of sociodynamical models has been formalized by Weidlich
and Haag (1983). The state of a system is described by a small set of global variables, the
socio-conguration, and its change modeled by equations of motion. An important

example, also with respect to this thesis, is the master equation formulation. Master
equations are used to mathematically describe stochastic processes where the probability
of nding a system in a particular state cannot explicitly be written, but an equation for
its time evolution is given. It can be derived from the discrete Markov assumption,
according to which the state of a system at time t does only depend on the state at time
t 1. Let P (s, t) denote the probability of nding a particular system in state s at time
t. If the Markov assumption holds,
X
P (s, t) = P (s, t|s , t t) P (s , t t),
s
where the sum runs over all possible states of the system s and t is the size of the
discrete timestep. P (a|b) denotes the conditional probability of a when b is given. The
time-continuous Master equation describes the time evolution of P (s, t) for innitesimal
small timesteps (the limit of t 0) and has the form
" #
d P (s, t) X
P (s, t) = w(s|s , t) P (s , t) w(s |s, t) P (s, t) , s 6= s ,
dt s
where w(s|s , t) = limt0 P (s, t + t|s , t)/t is the conditional transition rate from
s s. This general master equation, which is a specic differential equation, can be
formulated in words: the probability of nding a system in state s at time t increases
with any inow into state s from any other state s and decreases with any outow
from state s to any other state s at time t. The inow is determined by all conditional
transition rates w(s|s , t), multiplied with the respective occurrence probabilities P (s , t),
and the outow analogously.
This formalism (where dening the transition rates species the respective sociodynamical
model), allows for many analytical insights into these dynamical systems. However, its
applicability is mostly restricted to systems with very few dynamical variables (like e.g.
in the mean-eld limit, where the state of a system is described by an aggregated
variable) and a limited number of possible state realizations. When this is not given,
other modeling techniques can be employed. Agent-based modeling (ABM) and multi-
agent simulations (MAS) are such techniques where microscopic agents (individuals) and
their interactions among each other are dened and the macroscopic evolution of a system
is observed through stochastic simulations (e.g. Monte Carlo methods). Such models
often allow for a better understanding of the microscopic processes leading to macroscopic
consequences, but have the disadvantage that analytical results are, if at all possible, hard
1.3. Dynamical fixed points and their stability 7
to obtain. It results a critical tradeo between the modeling principle keeping it simple
and stupid (KISS) and realism in the models.
It can scarcely be denied that the supreme goal of all theory is to make the
irreducible basic elements as simple and as few as possible without having to
surrender the adequate representation of a single datum of experience.
Albert Einstein (Einstein, 1934)
... or, as Einstein is often cited with respect to the above statement:
Things should be as simple as possible, but not simpler.
1.3 Dynamical fixed points and their stability

In the following Chapters, we will use dierential equations in order to describe the re-
spective model under consideration. These dierential equations constitute a dynamical
system and it is in most cases helpful to nd equilibrium states of the system so-called
xed points of the dynamics. As this is of interest in several parts of this thesis, we
shall explain the methodology in general:
Consider a dynamical system f (x, y, t) = x(t) + y(t), where all equations of motion are
given. They represent time-derivatives of the respective, time dependent state functions
x and y. This means, the time derivative of the system is given as f(x, y, t) = x(t) +
y(t), where the dot above the variables indicates time derivatives. Consider further that,
e.g. due to non-linear dependencies, the dynamical equations are not integrable, i.e. the
primitive functions x(t) and y(t) cannot be obtained. In this cases, one can at least learn
about the fixed points of the dynamics, i.e. states of the system where f(x, y, t) = 0 (also
called stationary solutions, there is no dynamics in such a point). A state of the system
f at time t is characterized by the values x and y at time t and, obviously, in a xed point
these values have to satisfy either x(t) = y(t) = 0 or x(t) = y(t).
However, having found a xed point in x0 and y0 , a small perturbation in x or y might drive
the system out of the xed point where the dynamical equations are non-zero. Therefore,
in addition to nding the xed points, one is interested in their stability. Let us use
the notion of a dynamical ow of the system f (x, y, t). In our case, this ow is two-
dimensional because both x and y can change. In a xed point f0 (x0 , y0 , t), there is no
dynamical ow. In any other state, there is a dynamical ow along either one or both
dimensions. To evaluate stability of a xed point, we consider a small perturbation of a

xed point which means we nd the system in a vicinal state f (x0 + 1 , y0 + 2 ), where
the absolutes of 1,2 are very small. If the dynamical ows in this vicinity drive the system
into the considered xed point, we call it stable and attractive (or asymptotically stable).
If the dynamical ows keep the system constantly within this vicinity, e.g. by oscillations,
the xed point would be marginally stable, but not attractive. Whenever a dynamical
system exhibits exactly one stable xed point, we can expect to nd it in this state (or in
its vicinity) after a sucient time of evolution. If a system exhibits more than one stable
xed point, its dynamical outcome depends on the initial state. Note that a system with
a nite state space must possess at least one stable xed point, as we will see later on.
If the dynamical ows in the complete vicinity of a xed point drive the system further
away from it, it is called unstable or repellent. Finally, a xed point can also be attractive
in a part of its vicinity, and repellent in another part. This is called a saddle point and
means that, dependent on the initial condition, a system might evolve into this state
or depart from it. Let us be more precise by considering a specic example of one-
dimensional ows and, afterwards, its general translation to multi-dimensional ows,
i.e. dynamical systems with one or more dynamical variables, respectively.
1.3.1 A one-dimensional flow
Assume that the state of a system is fully described by one variable x. The dynamical
equation shall be
dx (t)
x = x3 x. (1.1)
dt
The xed points are found by setting x = 0. There are three values of x fullling this
condition, namely x1 = 1, x2 = 0, and x3 = 1. These are the xed points of the
dynamical system constituted by Eq. (1.1). To nd out the stability, we linearize the
dynamics around the xed points, i.e. we derive x after the only state variable x and
obtain x = 3 x2 1. We evaluate this derivative at each xed point and the sign of this
result indicates stability or not. The argument can easily be understood by visualizing it,
as done in Fig. 1.3.1: Consider a positive slope (derivative) of x at a xed point, e.g. the
one of x1 = 1. This simply means x < 0 left of the xed point and x > 0 right of the
xed point. Hence, the dynamics at a value of x < x1 further decreases x, while a slightly
higher value lets x increase further, i.e. xed point x1 is unstable. The same holds for x3
where x = 2. The reversed scenario holds when the derivative is negative, like the one at
xed point x2 . It is therefore a stable and attractive xed point and its attractor region
is bounded by the two unstable xed points. A derivative equal to zero (not present at
a xed point in our example) would indicate a saddle point. The dynamics of Eq. (1.1)
can be analyzed as follows: The nal state of the system depends on the initial value of x.
If initially 1 < x < 1, the stable xed point will be reached after some time of system
evolution and the system will remain in this state. For another initial condition, x diverges
(x ).
2
dx(t) / dt
6
2 1 0 1 2
x
Figure 1.1: The dynamical function of Eq. (1.1) dependent on the state x. Arrows indicate
the direction of the dynamics in the vicinity of the three xed points.
Let us briey mention two extensions of the just described dynamics. The rst one regards
a statement from above regarding the necessity of a stable xed point in a limited state
space. If we assume a restricted state space for our example, e.g 2 < x < 2, the system
cannot diverge, but only run into this boundary states. Obviously, these are then also
stable xed points a small (valid) perturbation causes a dynamics driving the system
back to the boundary. Note that the applicability of the linearization is restricted in such
a case, because of the discontinuity of the dynamical function.
As a second extension, let us assume another parameter in the dynamics, e.g. assume
x = x3 x + a, which vertically shifts the function in Fig. 1.1. Of course, the values of
xed points and their stability now depend on a. Furthermore, the general characteristics
of the dynamical system might change at some values of the parameter: whereas we found
a stable xed point if a = 0, for a = 2 the system invariably diverges. Such qualitative,
parameter dependent changes in the dynamics are called bifurcations (see e.g. Redner
(2001); Strogatz (1994) for a classication of bifurcations).
1.3.2 Multi-dimensional flows
Here, we assume n dynamical variables x1 , ..., xn and their evolution functions
x1 = f1 (x1 , ..., xn ) (1.2)

x2 = f2 (x1 , ..., xn ) (1.3)
..
. (1.4)
xn = fn (x1 , ..., xn ). (1.5)
At rst, we have to compute the xed points, i.e. all sets of (x1 , ..., xn ), such that xi = 0
holds for all i {1, ..., n}. The stability analysis follows the same reasoning as above, but
is a little more involved. Like before, we want to nd out what happens if a system in
a xed point is perturbed. In the example above, we considered only one dimension of
dynamical ows (the x-axis) and the perturbation could only happen along this dimension
(left or right of the xed point). Here, we have n dimensions for the dynamics to move
along and for each dimension two possible directions of movement (in- or decreasing the
respective variable). Hence, to nd out about stability, we have to take into account all
possible partial derivatives. We do this by building the Jacobian matrix

x1 /x1 x1 /x2 ... x1 /xn
x2 /x1 x2 /x2 ... x2 /xn

J=
.. . (1.6)
.

xn /x1 xn /x2 ... xn /xn
The n eigenvectors of this matrix stand for one dimension in the state space each and the
associated eigenvalues indicate the direction of the dynamical ow along this dimension.
Note that every possible line of movement can be derived as combination of all eigenvectors,
a so called superposition of the eigenvectors. Therefore, the n eigenvalues at a xed point
inform us about the dynamical ows in the n-dimensional vicinity of that xed point just
as the only derivative in our one-dimensional example from above: if all eigenvalues of
the Jacobian at a xed point are negative, the xed point is stable and attractive (from
every direction, the dynamical ows drive the system into this xed point). If there are
positive and negative eigenvalues, we have a saddle point. Exclusively positive eigenvalues
indicate an unstable xed point (a repellor). It can also happen that more than one xed
point form a line (or a plane) of xed points, which would get apparent by at least one
eigenvalue that equals zero.
The same classication holds for complex eigenvalues. Non-zero imaginary parts indicate
oscillations in the dynamics, but the sign of the real parts indicate whether the xed
point attracts or repels (i.e. whether the oscillations shrink or grow) in the same way as
described above. 1 If the eigenvalues are complex and the real parts equal zero, this gives
rise to oscillations that are constant in time, i.e. the xed point neither attracts nor repels
the dynamics.
1
Note that complex eigenvalues have equal real parts and for the imaginary parts only the signs are
opposite.
Chapter 2
Decisions through social influence
2.1 Decision theory

This thesis investigates the inuence of social interaction on aggregated outcomes of the
dynamics within a group of individuals. Although we basically focus on quite primitive
behavioral rules and our investigations are not necessarily focused on the interaction be-
tween human beings, let us rst mention some major concepts of (human) decision theory.
First, there is a general distinction between normative and descriptive decision theory; the
former seeks to understand how individuals should decide if they were rational (a term
that we shall explain more in detail below), and the latter observes and aims at explaining
how individuals actually do decide in certain circumstances.
Rationality involves the existence of utility- or preference functions, in order to normalize
the consequences (or success) of decisions. For example, the personal value of 100 Swiss
Francs surely diers for people of signicantly dierent incomes. With this normalization
of absolute values, one generally denes more is always better and individuals act with
the aim of maximizing their utility. These preferences and the associated decisions have
to fulll some propositions in order to be considered rational:
Transitivity
The preferences of individuals are consistent, i.e. if an individual prefers alternative
A over B and alternative B over C, it also prefers alternative A over C. This
proposition follows directly from the denition of utility.
Invariance
Decisions of individuals are independent of the representation of decision situations.
If, for example, alternative A is presented prior to alternative B or the other way
12
2.1. Decision theory 13
round must not inuence the decision.
Independence of irrelevant alternatives

If from the alternatives (A, B) an individual decides for A, the same individual must
not choose B from the alternatives (A, B, C).
The term rationality goes even beyond mostly it also involves the assumption of complete
and perfect knowledge, i.e. all information relevant for the decision are available to the
individual. In other cases, it is sucient to assume exact knowledge about the probabilities
of several alternative circumstances in order to allow for a rational decision. Especially in
classical economic theory, the models assume a perfectly rational, economic actor, which is
often called homo oeconomicus (see also Anderson (1991); Eisenfuhr and Weber (2002);
Simon (1955); Smith (1991); Tversky and Kahneman (1986)).
It is easy to criticize the rationality assumptions: in most situations, people do not have
all relevant information available, their capabilities to analyze all available information are
restricted, future consequences of decisions are subject to uncertainty, and, most impor-
tantly, we humans do not even decide consistently or rationally given the information we
have. One example: we arrive at the theater and realize that we forgot our ticket, that
we paid 100 Swiss Francs for, at home. Asked if we would buy a new one, most of us
would answer no. If we did not already have a ticket, but while queueing to buy one we
realize that we lost a 100 Swiss Francs note on the way to the theater, most of us would
still buy the ticket (Eisenfuhr and Weber, 2002). The economic decision is the same in
both cases, but somehow it is not the same to us. Besides this rather illustrative example,
much more empirical evidence has been collected to conclude that rationality is not well
suited to explain human decision behavior (Fehr and Gachter, 2002; Henrich et al., 2001;
Kagel and Roth, 1995).
The concept of bounded rationality, introduced by Herbert Simon (Gigerenzer and Selten,
2001; Simon, 1956, 1972), takes into account the limitations to information availability
and information processing. It has to be distinguished from irrationality in the sense of
making decisions that are harmful to the individual. It rather means that individuals
do not optimize before making the decision. For example, if many decision alternatives
are available, optimizing requires comparative eort and takes time. For this reason,
individuals stop the optimizing process at some point and make a (hopefully) near-
optimal decision. Simon named this satiscing, which refers to satisfaction with a
sucient, but not optimal decision.
Another approach within the framework of bounded rationality is to consider decision
making processes as following heuristics (Gigerenzer and Selten, 2001; Gigerenzer and
14 Chapter 2. Decisions through social influence
Todd, 1999; Tversky and Kahneman, 1974). Instead of trying to nd an optimal solution
for a certain decision situation, the whole situation is simplied to some important features.
This simplied situation can be analyzed much faster and, depending on the heuristics, a
more or less successful solution be found.
Decisions under social inuence clearly belong into the bounded rational category of mod-
els. For the game theoretic part of this thesis, we will generally assume that individuals
strive for maximizing their utility or that an evolutionary force favors individuals with a
higher utility values. Furthermore, the decision situations under consideration are mostly
very simple, there is only a small number of decision alternatives with clearly specied
consequences. In some parts of the argumentation, perfect information is also assumed.
There, rationality will play a role, but only for a rst assessment of the respective situation
(serving as a kind of benchmark). The most important ndings and conclusion do not rely
on the rationality assumption.
This short overview over some concepts of decision theory is aimed at introducing part
of the terminology used throughout this work. However, as mentioned before, there are
fundamental dierences in the research approach taken in the works cited above and the
context of this thesis. The main points of dierentiation are:
While decision theory seeks to understand and explain human behavior, we apply
dierent behavioral rules in our interaction models in order to investigate the dy-
namical properties and the eects of social interaction.
In our decision situations, we exclusively consider interdependent decisions, i.e. the

decision of one individual has consequences for and is inuenced by other individuals.
Depending on the specic decision rule under consideration, the individuals in the
models of this thesis are not necessarily humans.
2.2 Models of social influence

As discussed in the previous Section, there are many reasons to assume bounded rational
rather than rational decisions in the real world. If comparing all decision alternatives is
costly, non-rational decision making might be more ecient. One natural form of non-
rational decision making is imitation, i.e. trusting in: what is good for you cannot be
bad for me.
The following general statement, offered as a summary of many prior stud-

ies as they apply here, explains why imitation occurs: Nonrational decision
2.2. Models of social influence 15
making occurs because it economizes on decision-making resources.
Mark Pingle (Pingle, 1995)
Putting this argument further: in the absence of information, imitation might be the
best decision to be taken. Consider an individual in a burning building with dense wads
of smoke. The individual has no idea which direction to choose in order to leave the
building. Following some other individuals seems then the best option because there is a
chance that these individuals do have some information on the next exit. However, with
slightly changed assumptions, the opposite can be true. If there are too many individuals
heading in the same direction and there is evidence of more than one exit, it might be
the best to choose another direction in order to avoid clogging at the commonly chosen
exit. What these options have in common is that there is contentwise symmetry between
the alternatives the individual does not know where an exit is and, with respect to this,
every direction is as good as another. Therefore, the decisions are entirely dependent on
the decisions of others and independent of the alternatives as such.
We refer to this kind of decision making as decisions by social influence.
Social influence is defined as change in an individuals thoughts, feeling,

attitudes, or behaviors that results from interaction with another individual or
a group.
The Blackwell Encyclopedia of Social Psychology (Rashotte, 2007)
This is related to observational- or social learning in social psychology (Bandura,

1977; Miller and Dollard, 1998). There, the consequence of observation is not necessarily
imitation, but can also be the opposite if the observed consequences of a behavior are
negative. In general, such learning theories require the anticipation of consequences as
motivation to learn, which is not necessary for the social inuence mechanisms assumed
in this work.
However, we can classify the eects of social inuence into (i) imitation/herding/persuasion,
(ii) repulsion/avoidance/contrarian behavior, and (iii) compromising/averaging, as done
by Helbing (1995). However, the main focus here and in the literature is on positive social
influence, i.e. imitational processes. Positive social inuence and imitation is closely con-
nected to majority rules or majority voting. If interaction occurs between more than two
individuals, imitation is not straight-forward because an individual might observe dierent
behaviors. In that case, following the behavior of the majority represents positive social
inuence.
Imitational behavior can originate from dierent sources: dierent from compliance or
obedience, which assumes a form of authority or power to coerce individuals to behave
like others, conformity refers to the mere desire to not behave dierently from others.
Experiments with human subjects illustrated how strong, and sometimes dangerous, such
tendencies can grow: Asch (1951) showed that people sometimes deny the truth if all
other people in a group state something that objectively is wrong (conformity). Regard-
ing compliance and obedience, the experiments of Milgram (1974) and Zimbardo (2008)
are remarkable. In Milgrams setup, people were told to punish other people with electric
shocks whenever they did mistakes in remembering words. As the frequency of mistakes
increased, many test persons were ready to shock other people with life-threatening volt-
age, just because they were told to and although the other individual was just another
test person.1 In the experiment of Zimbardo, 24 test persons were divided into prisoners
and guards of a simulated jail. They were told to play these roles for two weeks with all
associated consequences regarding power of the guards and restrictions for the prisoners.
This order was taken so serious by the test persons, that the experiment had to be stopped
after a few days. Initially homogeneous test persons internalized their roles up to a level
that the experimenter could not take the responsibility any longer.2
Across the social sciences, social inuence has been subject to investigations from many
dierent elds and with dierent specications. In a relatively recent review, Mason et al.
(2007) conclude four dierent, fundamental dimensions on which social inuence models
dier:
(a) the pattern of connectivity and inuence assumed among individuals,
(b) the treatment of attitudes or behavioral responses as continuous versus discrete,
(c) the presence or absence of individual dierences in private information, and
(d) the assumptions made about the assimilative versus contrastive directional eorts of
social inuence.
The focus in (a) lies on network eects within social systems and accounts for the fact,
that not only direct inuence between two individuals matters, but also inuence via indi-
1
Of course, the individual to be threatened was not a test person, but an actor who did not receive a
real electric shock. However, this was not realized by the real test persons.
2
This experiment was also taken as basis of the novel Black box by Mario Giordano and the German
movie Das Experiment by Oliver Hirschbiegel.
2.2. Models of social influence 17
rect connections between individuals occurs (see also Wasserman (1994)). This has latest
become obvious since Milgram (1967) formalized the phrase of a small-world problem,
following which every person on Earth is connected with any other person via six degrees
of separation on average (Guare, 1990). Dimension (b) dierentiates between continu-
ous and discrete decisions or behaviors upon which social inuence works. An example
of continuous behaviors stems from the theory of reasoned action (Fishbein and Ajzen,
1975), where the relative strength of an individual intention to perform a certain behav-
ior (behavioral intention) is taken into account. On the other hand, discrete decisions
play an important role e.g. for investigations on innovation diusion, like in the model
of Granovetter (1978). In dimension (c), Mason et al. (2007) distinguish between models
where social inuence is the only source of inuence, like in (Granovetter, 1978), or where
there are also other sources of input for the individuals, as for example in the theory of
reasoned action (Fishbein and Ajzen, 1975) or in models of swarm intelligence (Kennedy
et al., 2001). Finally, dimension (d) reects the already mentioned topic of imitational
processes versus contrarian behavior. While assimilative or imitational processes are ex-
tensively discussed within this thesis, the seceder model of Dittrich et al. (2000) serves as
example to show how group formation can take place as consequence of local tendencies to
be dierent. For a much more comprehensive overview of theories and models regarding
social inuence, we point the interested reader to the review of Mason et al. (2007).
In order to understand the intrinsic properties of systems with strong social inuence,
a number of models have been developed that take the spread of opinions as sample
application. Early approaches in the social sciences showed that the existence of positive
social inuence (i.e. imitation behavior) tends to establish homogeneity (i.e. consensus)
among individuals (Abelson, 1964; French, 1956). The voter model, rigorously dened
by Liggett (1995), conrms these results. As we will see later on, it models dynamics
between two discrete opinions (e.g. pro or contra), where individuals tend to imitate
the opinion of an interaction partner. The same holds for models with a continuous
opinion space (Deuant et al., 2000; Hegselmann and Krause, 2002), where an individuals
opinion is one position on a continuous spectrum mostly the unit interval between zero
and one and an individual moves its opinion in the direction of an interaction partner
(compromising). There, however, it was also shown that a selection of interaction partners
(bounded condence) can lead to stable diversity of opinions, even when considering
positive social inuence. Bounded condence means that an individual does not interact
with any other individual, but only with those whose opinion is not too distinct from
their own one, i.e. opinions within a certain condence interval. Such results are in line
with dynamical models of social impact theory, as introduced by Latane (1981), Latane
and Wolf (1981) and Latane (1996). There, the fundamental argument is a spatial one:
interaction between near-by individuals is more likely than long-range interactions. This
was shown to create clearly separated clusters of homogeneous behavior and to diminish
but stabilize the share of individuals applying the minority behavior.
While modeling details for the voter model are introduced below, let us only mention some
of the most important models (a more elaborated review of the eld and its models can
be found in Castellano et al. (2007)): the Ising model of statistical mechanics originally
constitutes a spin system explaining order-disorder phase transitions in ferromagnets (Bin-
ney et al., 1993). In the model of Weidlich (1971), it found rst application in a social
context, but many other models apply the analogy between energy minimization of spins
and human tendencies to avoid social friction (see also the Galam model (Galam et al.,
1982)). One of them is the voter model (Liggett, 1995), which will extensively be dis-
cussed in Chapter 5. Opposing to e.g. the voter model, the Sznaijd model (Behera and
Schweitzer, 2003; Stauer et al., 2000) assumes an inside-out inuence where (two) like-
minded individuals are able to spread their opinion to other individuals. If the state space
of opinions cannot suciently be described by (two) discrete opinions, there foremost are
two models with a continuous opinion space: the Deffuant model (Deuant et al., 2000)
and the Hegselmann-Krause model (Hegselmann and Krause, 2002). They dier in that
the rst one considers bilateral interactions whereas in the latter one an individual in-
teracts with all other individuals within its condence interval. For a recent survey of
continuous opinion models with bounded condence, see Lorenz (2007).
2.3 The voter model

Consider a number of individuals that can have one out two possible opinions about a
certain topic, e.g. whether to vote yes or no in the next referendum. The topic might
be complex and individuals are not very decided on it. Hence, they repeatedly discuss
about it with some other individuals that they are connected to. Of course, individuals are
rather connected to others that they have some similarities with and so we assume that
the arguments of others sound convincing enough to adopt the same opinion. Based on
this, admittedly rather simplistic scenario, one is interested in the dynamical properties
of such a system.
In the original voter model (Dornic et al., 2001; Holley and Liggett, 1975; Liggett, 1995),
N voters are positioned at the sites of a regular, ddimensional lattice. This means
that every voter exclusively interacts with a xed set of nearest neighbors, e.g. the voters
at the four directly neighboring sites in a regular, 2-dimensional lattice. Each voter i has
one of two possible opinions at a time, i (t) = 1, which might change every time the
2.3. The voter model 19
voter interacts with one of its neighbors. A timestep in the evolution of the system of N
voters consists of N update events. In an update event, one voter is picked at random and
adopts the opinion of one randomly selected neighbor. Thus, the probability that voter i
adopts opinion , that will be denoted by WiV (), is equal to the frequency of opinion
in its neighborhood. Hence,

1 X
WiV (, t) WiV (|i , t) = 1 + j (t) , (2.1)
2 k
j{i}
where k is the number of neighbors each voter has, and {i} is the set of its neighbors.3 The
second term within the brackets yields the average opinion in the neighborhood, which
ranges from 1 to +1. Addition by 1 and division by 2 transfers this average opinion
into a probability between 0 and 1. As stated above, if 3 out of 4 neighbors have opinion
+1, the probability to randomly choose a +1 neighbor is 0.75, which is identical to the
probability to adopt this opinion. For this reason, the model is also called linear voter
model, because the adoption probabilities depend linearly on their frequencies (compare
with (Cox and Durrett, 1991; Schweitzer and Behera, 2009)).
Here, it gets obvious why the voters in this model are said to have zero condence: their
own opinion i does not enter Eq. (2.1), i.e. it does not have any inuence on a voters
decision. Note that this equation can also be applied to networks of dierent topology, as
we will do later on.
For the voter model, the 2-dimensional regular lattice was found to be a critical dimension,
i.e. depending on the dimension of the system, its dynamical properties change. Only in
systems with dimension 2 or below, coarsening takes place. This means that starting
from a completely disordered state, clusters of likeminded individuals form and grow over
time. Due to the probabilistic nature of the transition probabilities (2.1), these clusters
are neither spatially static nor separated from each other by sharp boarders. Instead, they
exhibit strong uctuations and move around in the system (compare with Fig. 2.1). This
critical coarsening without surface tension is an important characteristics of the voter
model it denes its own dynamical universality class (Dornic et al., 2001). However,
modifying the dynamics, e.g. implementing a third, intermediate opinion (Castello et al.,
2006), or including memory eects (DallAsta and Castellano, 2007), might change this
fluctuation driven dynamics into curvature driven dynamics (with surface tension).
3
Eq. (2.1) represents the node update rule of the voter model. There is another specification the link
update in which instead of one voter, two connected voters are chosen randomly for opinion update (a
link with its two end nodes). If this link is active, i.e. it connects two voters with different opinions, one
of the voters (chosen randomly) switches its opinion, thereby making the link inactive.
Finite voter model systems reach consensus (one of the two absorbing states where all
voters have either opinion +1 or -1) in a nite time (compare with Fig. 2.1). The time to
reach consensus, T , depends on the size of the system and the topology of the neighbor-
hood network. For regular lattices with dimension d = 1, T N 2 , for d = 2, T ln N ,
and for d > 2, T N . The coarsening of systems with dimension d 2 drives the system
into consensus even in the thermodynamic limit, i.e. when the system size goes to innity.
In systems of any dimension larger than 2, no coarsening takes place and consensus might
not be reached in the thermodynamic limit (Slanina and Lavicka, 2003).
Another important feature of the voter model is the conservation of magnetization (Castel-
lano et al., 2003; Frachebourg and Krapivsky, 1996; Suchecki et al., 2005a), as compared
to other prototypical models, such as the Ising Model with Kawasaki dynamics (Gunton
et al., 1983). Let A(t) (B(t)) be the global frequency of voters with opinion +1 (1)
at time t. The average opinion of the system (also called magnetization analogous to
studies of spin systems in physics) can be computed as
M (t) = A(t) B(t). (2.2)
In the mean-eld limit, that we will study in more detail later on, we assume that the
change of the opinion of an individual voter only depends on the average frequencies of
the dierent opinions in the whole system. Therefore, we replace the local frequencies in
Eq. (2.1) by global ones, which leads to the adoption probabilities W V (+1| 1, t) = A(t)
and W V (1| + 1, t) = B(t). For the macroscopic dynamics, we can compute the change
in the global frequency of one opinion as
A(t + 1) A(t) = W V (+1| 1, t) B(t) W V (1| + 1, t) A(t)

= A(t) B(t) B(t) A(t)
0, (2.3)
i.e. the frequency of each opinion is conserved for every state of the system. Departing
from mean-eld calculations, the magnetization conservation generally holds. In a single
simulation run, a voter model system changes its magnetization due to uctuations and
eventually reaches consensus. However, ruling out these uctuations, we nd that the
exit probability of an opinion (or fixation probability, i.e. the probability that an initial
frequency of an opinion leads to consensus in this opinion) is exactly as high as the
initial frequency of that opinion. The considerations of this conservation law will play an
important role in the investigations of the following Sections.
2.3. The voter model 21
(a) t = 0 (b) t = 1200
(c) t = 3600 (d) t = 4600
(e) t = 6900 (f) t = 7100
Figure 2.1: Exemplary time evolution of the voter model on a 2-dimensional, regular
lattice, where every voter interacts with its 4 nearest neighbors. The system consists of
N = 10, 000 voters and the evolution time t is measured in generations, i.e. 10,000 single
updates correspond to 1 generation.
The order parameter, most often used in the voter model, is that of the average interface
density . It gives the relative number of links in the system that connect two voters with
dierent opinions and can be written as
1XX
(t) = 1 i (t) j (t) . (2.4)
4 i
j{i}
Because of its simple structure, the voter model allows for many analytical calculations
(Liggett, 1995; Redner, 2001) and, therefore, serves a comprehensive understanding of the
dynamics involved. Application areas of the voter model include coarsening phenomena
(Dornic et al., 2001), spin-glasses (Fontes et al., 2001; Liggett, 1995), species compe-
tition (Chave, 2001; Ravasz et al., 2004), and opinion dynamics (Holyst et al., 2001).
Based on the voter model, investigations were conducted to study interesting emergent
phenomena and relevant applications. Such works comprise the possibility of minority
opinion spreading (Galam, 2002; Tessone et al., 2004), dominance in predator-prey sys-
tems (Ravasz et al., 2004), forest growth with tree species competition (Chave, 2001),
and the role of bilingualism in the context of language competition (Castello et al., 2006).
The question of consensus times and their scaling for dierent system characteristics was
particularly addressed in several studies (Castellano et al., 2003; Liggett, 1995; Redner,
2001; Sood and Redner, 2005; Suchecki et al., 2005b).
Chapter 3
Decision making in (evolutionary)

game theory
At dierence with models of opinion dynamics, in the following we will consider a wider
range of decision behavior of individuals rather than imitation, avoidance, and compromise.
In this Chapter, we will rst introduce some fundamental concepts of classical game theory;
its origin, dierent specications, aims, and terminology. While formal, mathematical
denitions of games and concepts can be found in a number of introductory books (Aubin,
1979; Fudenberg and Tirole, 1991; Guth, 1999; Kreps, 1990; Luce and Raia, 1957; Osborne
and Rubinstein, 1994), we will focus on contentwise explanations that are aimed at allowing
a better understanding of the discussion below. After mentioning important questions
addressed by non-cooperative game theory and providing examples of social and economic
systems, two elds basing on classical game theory are introduced: experimental game
theory and evolutionary game theory. Within this thesis, we mainly deal with evolutionary
game theory. Concepts and analyses of classical game theory are introduced and refered
to at several places in the present work, but always with the aim of analyzing the outcome
of evolutionary dynamics applying particular game theoretic models.
3.1 Classical game theory
3.1.1 Definitions and terminology
Decision making in game theory generally follows a strategy, where the space of strategies
covers all possible decision behaviors from simple (e.g. random or always deciding the
same) to complex (e.g. history dependent and/or anticipating), from stupid to smart (e.g.
23
24 Chapter 3. Decision making in (evolutionary) game theory
involving learning mechanisms). By common denition:
A strategy denes a set of moves or actions a player will follow in a given

game. A strategy must be complete, dening an action in every contingency,
including those that may not be attainable in equilibrium. For example, a
strategy for the game of checkers would dene a players move at every possible
position attainable during a game. Such moves may be random, in the case of
mixed strategies.
Dictionary of game theory (Shor, 2005)
First we have to note that therefore also decisions through social inuence are strategic,
because the (re-)action of an individual is at every time dened by the observed action
of interaction partners. However, the available strategies in models of social inuence are
restricted to those being a function of the interaction partners decisions. Strategies in
the general sense of game theory can be any kind of decision plan, including the repeated
application of one and the same action or random decisions. The conceptual dierence
of decisions by social inuence compared to general strategic decision making is that
individuals cannot assign a specic decision to any decision point, but the decision will
always be dependent on the action applied by the interaction partner. In this Chapter we
consider strategic decision making in the general sense, i.e. individuals can choose from
the whole strategy space. To summarize:
Decisions by pure social inuence are taken frequency dependent and utility inde-
pendent and the decision alternatives themselves are symmetric, i.e. apart from
the comparison with interaction partners, there is no dierence for the individuals
between one or the other decision alternative.
Strategic decisions in general can be taken frequency independent, i.e. the decisions
are not necessarily symmetric for the individuals. This allows for utility maximiza-
tion in the strategy selection process of the individuals, dependent on the available
information and analytic capacity.
Decision making by social inuence is a subset of general strategies, where infor-

mation on utility might not be available, individuals are not able to process these
information, or utility considerations do not play a role because of the absence of
competition.
3.1. Classical game theory 25
In 1944, John von Neumann and Oskar Morgenstern rst applied their own mathematical
framework to interdependent decision making in economic environments (Von Neumann
and Morgenstern, 1944). This work is nowadays widely reckoned as the hour of birth of
game theory.
A game, in the sense of game theory, is dened by the number of players (N 2),
the complete set of decision alternatives to be chosen among by each player (at least 2
alternatives per player), and the utility (referred to as payoff ) for every player in every
possible outcome of the game. In order to analyze the games, we have to assume that
every player is rational and that rationality is common knowledge, i.e. any player i knows
that all players are rational; all players know that i knows that all players are rational;
i knows that all players know that i knows that all players are rational; ... and so on.
Furthermore, all individuals have perfect information about the rules and payo structure
of the game. A complete decision plan, that assigns a decision to every possible state of
the play, is called the strategy of a player.
Let us give some simple examples in order to clearly distinguish between games in colloquial
language and the situations covered by game theory:
Roulette in the casino does not fall into the denition of games. There are usually
many players on the table, but their decisions do not inuence each other everybody
plays for himself against the casino and the odds of a player are independent of the
decisions of other players. Given the rules of roulette, the casino is the interaction
partner, but it has no strategic choice. It is chance that decides how much is won
by whom.
The case is not so simple when considering an ordinary lottery. The situation is
almost the same as for roulette, but now players can try to maximize their payo
in case they win. Their chances to win is still independent of other decisions, but
the amount that can be won depends on how many other players won in the same
category (e.g. the category of 5 correct numbers all players with 5 correct num-
bers share the jackpot of this category). Therefore, it makes sense to consider the
decisions of others when deciding for a set of numbers.
A straight-forward example of games fullling the denitions of game theory is the

famous board game chess, where 2 players have a strategic interaction and the success
of a player depends on its own decisions and the decisions of the opponent.
Whereas chess is a game in which exclusively the playing skills of players decide
the winner, this is not prerequisite to this extent. The game rock-paper-scissors
is another example of a strategic game. Two players have to decide for one of the
symbols rock, paper, or scissors. Rock wins over scissors, scissors wins over paper,
and paper wins over rock. Despite of mental abilities, there are no skills that help
winning, but still it is one of the strategic interactions in the sense of game theory.
Figure 3.1: Presentation of the rock-paper-scissors game in extensive form (top) and in
normal form (bottom). In the upper picture, big, black circles represent decision nodes
where the respective player has to decide for one of its alternatives. The dashed ellipse
indicates an information set: a player does not know which of the decision nodes within
one information set is its current one. Therefore, the example game is a simultaneous one.
The small, black circles at the ends of branches indicate the possible outcomes of the game
and the payos are given in brackets: the rst one for player 1 and the second one for
player two. The lower picture contains exactly the same information about the game in
form of a payo bimatrix.
There is a substantial dierence between cooperative- and non-cooperative game theory. In

cooperative game theory (see e.g. Luce and Raia (1957); Osborne and Rubinstein (1994)),
players are allowed to have preplay discussions in order to coordinate their strategies. To
be precise, players have the possibility to conclude binding contracts. This means that
some players arrange their strategies with each other and the agreements are enforceable
by the contract. In cooperative game theory, one is mainly interested in which coalitions
will form, can be maintained, and are joined by which players of the game. The present
work exclusively deals with non-cooperative game theory, where every player decides on
its own without discussing strategies with others. In fact, it is sucient to assume that
there is no possibility for the players to conclude binding contracts.
In order to explain the terminology used in the respective parts of the thesis, let us briey
mention some other dierentiations:
Simultaneous vs. sequential games

Straight-forwardly, if all players decide at the same time (with identical information),
we call it a simultaneous game, otherwise a sequential one.
Games in normal form vs. extensive form

Every game can be presented in a decision tree, i.e. in extensive form. A game in
normal form is presented in a bimatrix of payos and can therefore only consider 2
players and simultaneous games (see Fig. 3.1 for a comparison of one and the same
game in both forms).
Symmetrical vs. asymmetrical games

If the initial situation (chances and risks, i.e. the payo structure) is identical for
all players, it is a symmetrical game, otherwise an asymmetrical one.
Constant- (e.g. zero-) sum games vs. variable-sum games

In a zero-sum game (e.g. the game in Fig. 3.1), any positive payo gain of a player is
paid by other players with negative payo. There is no gain or loss in the system
of all players. Accordingly, any constant system gain or loss can be distributed by a
constant-sum game. In a variable-sum game, the gain or loss of the whole of players
can vary between the dierent outcomes of the game.
The rock-paper-scissors game illustrated in Fig. 3.1 is a non-cooperative, simultaneous,

symmetrical, zero-sum game with two players. Therefore, it can be presented both in
extensive and in normal form. Since the game is symmetrical, without loss of information
it would be sucient to specify the payos of player 1 only (as we will do in such cases from
now on). The games mainly discussed in the present work are similar to that example,
but mostly variable-sum games.
Let us turn our attention to some analysis concepts that are particularly important for
the investigations reported below:
Nash equilibrium
Certainly, one of the most important solution concepts of game theory is the equi-
librium state dened by John F. Nash Jr., which was awarded the Nobel prize in
economics in 1994 (Nash, 1950, 1951). It states that a solution1 is an equilibrium
outcome of a game if none of the players has an incentive to unilaterally deviate from
this solution. A strategy of a player is called best answer if this strategy maximizes
the players payo for a particular strategy combination of all other players. Hence,
a Nash equilibrium consists of best answers only. If all best answers of an Nash equi-
librium are unique, i.e. any player would denitely get less payo when unilaterally
deviating, it is called a strict Nash equilibrium. The Nash theorem states that every
game possesses at least one such equilibrium in mixed strategies2 . In pure strategies,
a game might possess no Nash equilibrium (like in the rock-paper-scissors example),
a unique one, or multiple Nash equilibria (see below for examples).
Dominance
A strategy that is best answer to all strategy combinations by the others is called
a payoff dominant strategy. If this strategy is always the unique best answer, it
is called strictly payoff dominant. A strategy is called risk dominant (or maximin
strategy) if its minimal possible payo (worst case for the player) is higher than the
minimal possible payo of any other strategy.
Pareto efficiency
This is the most common eciency criterion for solutions of a game. A solution A
is called Pareto ecient if there is no other solution B that (i) yields a higher payo
for at least 1 player and (ii) yields at least equal payos for all players. If such a
solution B exists, one would speak of a Pareto dominance of B over A and A is Pareto
decient. In the example of Fig. 3.1, every solution is Pareto ecient. Note that
there is no dependency between this eciency criterion and a Nash equilibrium: a
Nash equilibrium can be Pareto ecient or not and a Pareto ecient solution might
be an equilibrium or not.
1
A final outcome of a game, i.e. a set of strategies of all players, is also called a solution.
2
The strategies considered so far are also called pure strategies. A mixed strategy consists of probabil-
ities for choosing any possible pure strategy. It is a randomized strategy where an agent does not decide
for one of its pure strategies, but assigns a probability to all of them and lets chance decide.
System optimality
We introduce the notion of system-optimal solutions in the sense of Schellings col-
lective total (Schelling, 1978). In a system optimum, the sum of all players payos
are the highest possible. In economic terms, one can also regard this the welfare
maximum. We will use this notion as an eciency criterion that is hardly used in
the literature so far, but will be crucial to the investigations presented in Chapter 7.
Note that a system optimal solution invariantly is Pareto ecient. It is a corollary
of the Pareto eciency denition that a Pareto dominating solution B has a higher
system payo than A. Therefore, system optimality of A precludes the existence of
a Pareto dominating solution B, which makes A Pareto ecient.
3.1.2 Coordination- vs. cooperation problems
Particularly in extensive form, game theory can abstract complex strategic situations
presenting enormous challenges not only for the considered players, but also for researchers
trying to analyze the game. A number of such examples can be found in Guth (1999), for
instance games dealing with the problem of insurances and labor markets.
However, it is maybe the simplest class of games that often is the base of investigations,
namely games with only 2 players and 2 choice alternatives each, where both players
have identical roles (symmetrical 22 games). In Section 3.1.3, we will provide a detailed
classication of such games when assuming that payos in dierent situations are never
exactly equal. For the moment, let us have a look at some simple examples to illustrate
strategic conicts of coordination and cooperation.
Since the games are symmetrical, we will specify them by providing a 2 by 2 payo matrix
which contains the payo for the player deciding for a row-strategy, given the column-
strategy of the opponent player. In a game with the payos
A B
! %
A 1 0
, (3.1)
B 0 1
players face a coordination problem. The game possesses 2 equivalent Nash-equilibria

and players have to synchronize their strategies in order to achieve one of the 2 desired
outcomes (A, A) or (B, B). Another coordination problem arises in the game
A B
! %
A 0 1
, (3.2)
B 1 0
where players now have to asynchronously coordinate their strategies to reach a mutually
protable outcome, e.g. (A, B). Therefore, we refer to it as asynchronous coordination
problem 3 . In both games (3.1) and (3.2), players have no conict of interest, but only a
problem due to the lack of communication possibilities. This is not anymore entirely true
for the game
A B
! %
A 2 0
. (3.3)
B 1 1
This game also has 2 Nash equilibria, but only (A, A) is both Pareto- and system-optimal.
This means that there is no coordination problem since it is clear to both players that
strategy A leads to an outcome that yields the highest possible payo for both players.
However, by slightly relaxing the strict rationality assumption, it is conceivable that one
player might want to win over the other, which only is possible by choosing B. Moreover, it
is sucient to assume one player that expects the other to think in this way. Then, strategy
B would stand for risk-averseness because it ensures at least a payo of 1 (maximin
strategy). We therefore refer to this scenario as a basic cooperation problem, where
cooperative behavior would mean to abandon such competition considerations.
As a fourth of such basic problems, let us have a look at the game
A B
! %
A 0 0
. (3.4)
B 1 0
To some extent, the problems of coordination and cooperation are both present in this
scenario. The only way for the players to obtain any payo is to coordinate either in
solution (A, B) or (B, A). But there is only 1 payo point at stake, i.e. in addition to the
coordination problem, one player would have to be cooperatively enough to abandon his
chance of receiving the payo. Here, a clear conict of interest arises. We refer to this as
a partial cooperation problem, which will play a crucial role in Chapter 7.
3
In the literature, it is mostly called anti-coordination problem. This might be misleading because
players effectively have to coordinate their strategies.
The so called Prisoners Dilemma game is surely the most prominent example of game
theory. It is a paradigmatic cooperation problem and its history is closely tied to the
founding of game theory and Nashs work on equilibria in games. It started with experi-
ments carried out by Merril Flood and Melvin Dresher in 1950 at the RAND corporation.
Although they used dierent payo values and their game was not even symmetric, the
essence of their experiment is captured by the matrix
C D
! %
C R S
, (3.5)
D T P
with the following specications (although this matrix is usually used for the Prisoners
Dilemma, we will later on use the same variables for any symmetrical 22 game, but
partially refrain from the following semantics): if both players choose to cooperate (C),
each gets the reward, e.g. R = 3 Dollar. If both do not cooperate, called defection D,
they get a punishment of P = 1 Dollar each. However, if only one player cooperates, he
leaves with nothing (sucker, S = 0) while the defector receives the temptation payo
T = 5 cent. Hence, the matrix reads
C D
! %
C 3 0
.
D 5 1
This game possesses the dominant strategy defection, leading to the strict Nash equilibrium
(D, D) a real dilemma because if the players would cooperate, both would receive 3 cents
instead of 1 cent. In general, the Prisoners Dilemma game is dened by T > R > P > S
and some references additionally require 2R > S + T .
When preparing a lecture on the very new idea of game theory, the supervisor of Nash,
Albert W. Tucker, illustrated the game by the following story and named it Prisoners
Dilemma: two suspects are caught by the police and interrogated separately. Both have
the option to confess the crime or not to do so . The police has no clear evidence against
them and oers that if only one suspect confesses, he will get free while the other is
sentenced the maximum imprisonment. However, if both confess, both are imprisoned
slightly below maximum. In contrast, their crime could not be proven if both would not
confess. Then, the only consequence would be some more time of remand4 .
4
The information regarding the history of the Prisoners Dilemma game are collected from various
sources, e.g. Holt and Roth (2004); Rasmusen (2001).
3.1.3 Symmetrical 22 games the framework
In the previous Section, we have seen that symmetrical 22 games can serve as simple
models to investigate strategic problems of coordination and cooperation. They will be the
model of choice in the respective part of the present thesis. Therefore, a full classication
of such games is presented in the following.
Let us base on the parameters in Eq. (3.5), i.e. we have two strategies (C, D) and the dif-
ferent combinations lead to one of four payo values (R, S, T, P ). Since we will elaborate
on dierent symmetrical 22 games, it is important to dene which strategy is regarded
cooperative and which defective. For the sake of convenient readability, we will use
the following simplication throughout this Section:
Let us only consider an encounter of different strategies leading to the payoffs S and T
(partial cooperation). In such a situation, the strategy which yields the lower payoff is
regarded the cooperative strategy (C), and the other one the defective strategy (D).
This denition has no contentwise consequences, it is one possible way of commonly naming
the dierent outcomes in all the games. Where it makes sense to speak of cooperative
behavior, this simplication yields the correct naming of strategies. For the other games,
it is maybe the most useful way to name the strategies likewise. In all the cases, cooperation
means to risk losing against the other player and defecting means holding the chance to
end up with a higher payo than the other. Therefore, this approach puts more weight on
a players relative payo with respect to the co-players payo. With these specications
in mind, we use the same variables commonly used for the Prisoners Dilemma game for
any symmetrical 22 game.
Let us briey comment on our denition of T > S: it diers from the one taken in
other works, where mostly R > P is xed. However, there is no restriction regarding the
considered games. As we will see below, we address the same (ordinally distinct) games as
the ones discussed elsewhere (Hauert, 2002; Helbing et al., 2005b; Rapoport, 1967; Stark
et al., 2008a; Tanimoto and Sagara, 2007). The reason for our diering approach will
become clear in Section 7: it leads to an integrative representation of all symmetrical
22 games where the connection between neighboring games can be explained (Fig. 7.2).
Having made the points of Section 7, we will come back to this issue in the respective
Discussion (Section 7.6).
A game is dened by the number of players, their set of strategies, the sequence of de-
cisions to be taken by the players, and the payos for all players and for every possible
strategy-combination. The class of games described here is one of the simplest: two play-
ers decide between two alternative strategies. The strategic situation is identical for both
players, (e.g. the one for the row-player in the matrix of Eq. (3.5)). After they decided
simultaneously, they receive a payo depending on their own strategy and the strategy of
the other player.
As stated above, we nd it convenient for the reader if we use the same variables commonly
used for the Prisoners Dilemma game also for an arbitrary symmetrical 22 game. Hence,
we impose the label cooperative to that strategy that yields the lower payo in an
encounter of dierent strategies, i.e. T > S always holds.
We assume that the absolute payo values are not decisive for the strategic situation, but
only the ranking of them (we will qualify this point later on). Since we dened T > S,
which eliminates equivalent rankings, one can discern 12 ordinally distinct games. Fig. 3.2
conveniently visualizes the phase space of symmetrical 22 games in a coordinate system
and includes exemplary payo matrices.
Each of the 12 rectangular or triangular parcels of the coordinate system (separated by
full lines) host one ordinal payo ranking. Within this classication scheme, we nd the
prominent Prisoners Dilemma game, which we already have introduced in Section 3.1.2.
This game is characterized by a strict Nash equilibrium that is not Pareto ecient and,
depending on the payo values, system-optimal or not. Let us briey introduce the single
games (compare with the payo matrices of Fig. 3.2):
The game of Chicken is also often called Hawk-Dove game or Snowdrift game. One
of the stories is about 2 teenagers and their test of courage. They arrange that they
frontally approach each other by cars and that they turn to the right if they want to avoid
a collision. Of course, the one that chooses this avoidance option is perceived as chicken
and the other one will get the respect of the teenagers. However, a frontal collision of the
cars could be fatal and is, therefore, the worst outcome. If both turn, they survive and
none of them loses his face against the other. Similar in spirit, the Hawk-Dove story is
about moderate and escalating ght-strategies in animals and the Snowdrift game about
shoveling snow from a car or waiting for the other person to do this. The game possesses
2 Nash equilibria in pure strategies, namely the partial cooperation solutions. Depending
on the payo values, these equilibria are system-optimal or not.
Battle-of-the sexes (not in the asymmetric version of Dawkins (1989)) sketches a situation
where a couple arranged to meet in the evening, but they forgot on what they agreed
on. One of them preferred to visit a sports event and the other had an opera visit as
preference. However, both would rather abandon their own preference than spending the
evening alone. If both abandon their preference, they yield the worst case of visiting the
unpreferred event alone. The best is a coordination where one of them chooses her/his
preference and the other other one abandons her/his preference. This game also possesses
Figure 3.2: Classication of symmetrical 22 games according to payo rankings. For

each area a respective payo matrix (in the form of matrix (3.5), with T > S) is given.
Nash equilibria are marked by bold payo numbers. Parcels separated by solid lines denote
dierent rankings of the payo values. Two-dimensionality is achieved by xing T > S
and classifying ordinal dierences only.
2 pure, system-optimal Nash equilibria in the partial cooperation solutions.

Leader denotes a similar situation, but now the worst case is if both chose their preferred
event maybe because then they start to doubt their relationship. The Nash equilibria
are the same as in the previous game.
The 4 games exemplied so far represent the four archetypes of Rapoport (1967) (Mar-
tyr, Exploiter, Hero, and Leader).
In the game Stag Hunt (see e.g. Skyrms (2004)), 2 hunters have to decide whether to hunt
a stag or a rabbit. The stag requires both hunters to be slain if only one goes for the
stag he will not succeed whereas the other gets the rabbit. If they hunt together, they
3.2. Experimental game theory 35
can either share the stag or have to compete for the rabbit. This game is also interesting
because it possesses 2 pure Nash equilibria in the synchronous solutions (C, C) and (D, D).
Only (C, C) is system-optimal, but (D, D) has the advantage of being risk-dominant.
The game of Route Choice reects important characteristics of (vehicular- or data-) traf-
c systems and was named and experimentally investigated in Helbing et al. (2005b)
and Stark et al. (2008a). Independent of these works, the same situation was also in-
cluded in the experimental setup of Kaplan and Rue (2007). A reasonable story could
be that 2 controllers have to send a large truck convoy along the same highway and know
about this fact. They have to decide whether to send the convoy on this highway anyway
or to use another route over dierent highways. If they both send it on the same path,
the highway will be overloaded and trac jams lead to costly time delays. However, one
route is much shorter and, therefore, making the detour would most probably take more
time than queuing on the short route. Like the Prisoners Dilemma, this game has 1 strict
Nash equilibrium that might be system-optimal or not.
Deadlock is very similar, but now there is an additional advantage of being the rst to
arrive. Therefore, R > S because being the last is worse than needing more time. The
Nash equilibrium is the same like in the Route Choice game.
The remaining games Harmony I and II and the game Own Goal are similarly trivial
in the sense that they all have a strict Nash equilibrium that is Pareto ecient, system
optimal, and risk-dominant. Only Harmony I gained some attention in the literature due
to the fact that there the equilibrium lies in cooperation. It is also referred to as By-
product mutualism because the cooperative act does not only benet the recipient, but
also the donor (see e.g. Bergstrom and Lachmann (2003); Clutton-Brock (2002); Hauert
(2002)).
3.2 Experimental game theory

In the previous Section about classical game theory, we mentioned the necessary assump-
tions underlying the solution concepts; rationality, common knowledge, and perfect infor-
mation. Moreover, these assumptions also imply that each individual is capable of storing
and optimally processing all the information available. It is obvious that these assumptions
are quite strong and that, even when restricting the analysis to human beings, individu-
als are not fully rational, not always all the necessary information is available, and if so,
individuals dier in their abilities to memorize information and in their analytical skills.
Hence, game theory serves us as important theory about strategic decision making, but
its ability to explain individual decision behavior has to be questioned.
In order to validate theory, experimental game theory (or experimental economics) analyze
real human decision behavior under laboratory conditions. In most of the cases, test
persons are invited into a computer lab, get instructions on how to play a game against
one or more other test persons, and take decisions which lead, depending on the game
and the decisions of the other(s), to payo scores. In order to simulate relevant decision
situations, where the individuals try to optimize the outcome of their decisions, it is
important to appropriately generate an incentive structure for the test persons, such that
they are motivated to try their best to be successful. Such incentives can potentially have
dierent forms, in most of the cases test persons obtain a success-dependent, monetary
reward, where payo scores are converted into a disbursement.
Although such techniques allow to test theoretical predictions under relaxation of theo-
retical assumptions, each experimental design still contains assumptions and, most impor-
tantly, the validity of results of such experiments signicantly depends on the quality of
the experimental design. This quality is inuenced by many factors: the selection of test
persons, the selection of information to the test persons, the clarity of these information
(did every test person understand the instructions correctly?), the design of the rules has
to be appropriate for the questions the experiment is aimed to answer, and many other
factors. The importance of these design factors cannot be overemphasized and we would
like to point the interested reader to the Handbook of experimental economics by Kagel
and Roth (1995), which constitutes a comprehensive survey of results, but also of the
methods to design and conduct laboratory experiments.
While there is a vast array of relevant and interesting insight into human decision behavior
resulting from experimental research, we here want to address only one general nding,
which also serves as motivation for some investigations contained in this thesis: human be-
ings consistently decide cooperatively in experiments. One of the most suited experiments
to test the hypothesis of selsh, payo maximizing, and rational humans is the ultimatum
game (UG), which has been conducted innumerous times with dierent specications. The
basic ultimatum game is rather simple: one of two individuals receives a certain amount
of money, say 10 Swiss Francs, from the experimenter. This individual, the proposer,
has to make an oer to the other individual how to share this 10 Francs among the two
individuals. The other individual, the responder, then decides about the oer and has
two options: (i) the responder acccepts the oer, then the experimenter splits the 10
Francs and pays the individuals according to the oer; (ii) the responder rejects the oer,
then the experimenter keeps the 10 Francs and both individuals get nothing. The game
theoretical prediction is straight: the proposer will oer the smallest possible, non-zero
amount to the responder and the responder will accept this oer, because its alternative is
to receive nothing. However, the empirical evidence is signicantly dierent. Even if test
3.3. Evolutionary game theory 37
persons do not know each other and will very likely never meet again, the average oer of
the proposer lies somewhere between 40 and 50% of the whole amount, while the rejection
rates even for more than minimum oers is larger than zero. Variants of the experimental
investigation of the UG, all of which conrm the mentioned result at least qualitatively,
comprise tests in dierent cultural environments by conducting the experiments in 15 dif-
ferent, small-scale societies (Henrich et al., 2001), raising the amount of money at stake
to a higher than usual level (Cameron, 1999), testing asymmetries in the experimental
setup (Kagel et al., 1996), or observing neural responses of the test persons (Sanfey et al.,
2003).
While this thesis does not contain experimental work, part of it is motivated by and
follows up on empirical work on the decision behavior of test persons in route choice sit-
uations (Helbing et al., 2005b; Stark et al., 2008a). Particularly with view on the design
of advanced traveller information systems and the optimal routing in data networks, ex-
periments have been designed and conducted where individuals repeatedly had to decide
for one of two alternative routes. One route could be seen as a highway and the other as
rural road. While the highway in principle is faster, it might be congested (because too
many of the test persons decided for this route), in which case the rural road can even
be faster than the congested highway. Omitting the details, the decision situation to the
test persons was of the category partial cooperation problems, which was introduced
in Eq. (3.4). Besides the general result, that individuals were able to develop alternating
cooperation strategies in order to exploit the system optimal distribution within the traf-
c system in a fair manner, several treatments have been conducted to investigate the
optimal conditions and the inuence of additional information on the decision behavior
of the test persons. Despite the interesting empirical nding that individuals are even
under dicult circumstances willing and able to cooperate, it turned out that this kind
of strategic decision situations contains relevant theoretical implications, which have not
been thoroughly discussed from a game theoretic point of view. Therefore, particularly
Chapter 7 deals with such situations and introduces partial cooperation dilemmas.
3.3 Evolutionary game theory
Natural evolution works on populations of reproducing individuals (species). Dierences

in the speed of reproduction are explained by dierences in reproductive fitness and lead
to augmentation and extinction of species, i.e. selection. Mutations are small variations
between ancestor and ospring in the course of reproduction mostly rare events of erro-
neous reproduction. These mutations allow for variation, thereby increasing competition
in the struggle for life, as Charles Darwin has put it (Darwin, 1859). With these 2 in-
gredients (have a glance at the incipient quotation of Section 3.3.2 for a recent proposal of
a third fundamental process), evolution incorporates a fascinating optimization procedure
that allowed to create ecient, complex life forms, including human civilization.
How complex Nature and its laws might be, evolution can be described and analyzed by
rather simple mathematical tools, and game theory is one of them. Instead of individuals
which are free to choose between dierent strategies, evolutionary game theory assumes
the individuals to have one strategy hardcoded in their genetic material the single
individual is restricted to the decisions following this strategy. Reproductive tness is
the equivalent of payo in classical game theory: individuals interact with each other
according to their hardcoded strategies, and the gain or loss in payo contributes to their
own reproductive tness, but thereby of course also to the reproductive tness of their
genotype. Compared to classical game theory, two extensions are most important: (i)
starting from any combination of applied strategies, we analyze the evolutionary dynamics
towards an equilibrium state or between equilibrium states, and (ii) the main solution
concept (equivalent to that of the Nash equilibrium) is the concept of evolutionarily stable
strategies (ESS), which will be explained below (Maynard Smith, 1982; Maynard Smith
and Price, 1973).
One of the the longstanding questions in evolutionary biology, which over the time also
entered other scientic domains like economics, social sciences, physics, etc. and will be
of particular importance also in this thesis, is that of the emergence and maintenance of
cooperative behavior in competitive environments. The methods of evolutionary game
theory (see e.g. Friedman (1991); Hofbauer and Sigmund (1998); Nowak and May (1992);
Szabo and Fath (2007); Taylor and Jonker (1978)) have since Maynard Smith and Price
(1973) been applied to innumerous investigations regarding the evolution of cooperation in
biology (Maynard Smith and Szathmary, 1995; Nowak, 2006b), social sciences (Fehr and
Fischbacher, 2003; Henrich et al., 2003), and economics (Friedman, 1991; Gintis, 2005;
Kreps, 1990). The advantage of evolutionary considerations is that the strong assumptions
of rationality and common knowledge, key in classical game theory, are not necessary any
more.
Evolutionary game theory diverges from classical game theory only when
it comes to the analysis of beliefs. In the classical theory, players have be-
liefs about one another which are grounded in, or at least consistent with, ideal
rationality and common knowledge; their strategy choices are rational in the
sense that they maximize subjectively expected utility, when subjective beliefs
are themselves rational. Evolutionary game theory does not require the ratio-
nality of beliefs.
Robert Sugden (Sugden, 2001)
So far, we have introduced that evolution results from genetic inheritance in combination
with selection and mutation processes. The transmission process of successful behaviors
here is exclusively vertical, i.e. an evolutionarily positive eect for a successful behavior
can only occur in future generations, by a more successful replication of genes applying
the respective bahavior. Particularly with view on human beings, evolution might also
work through horizontal transmission, i.e. by individuals changing their (not genetically
encoded) behaviors. This might be induced by learning mechanisms, trial and error
behavior, strategic reasoning, or similar mechanisms. Another mechanism of cultural
evolution is the propagation of experience over generations, which is similar to genetic
evolution in that the transmission of behaviors is vertical, though not hereditary. At
dierence with genetic evolution, the time scale of signicant changes in cultural evolution
can be much faster than in genetic evolution. On one hand, horizontal transmission
does not require time costly reproduction, but can occur instantanously. On the other
hand, mutations in biology are rather seldom events of small changes, whereas in cultural
evolution radical shifts in the behavior of individuals are not unlikely.
In Chapter 6.3, we report on results from a modeling framework that arguments in terms
of cultural evolution. There, dierences in the transmission of behaviors result from two
sources: (i) from dierent, co-evolving abilities of individuals to inuence other individuals
and (ii) from payo dierences between the decision alternatives. In the other parts of the
thesis that deal with evolutionary game theory, the focus is rather on genetic eveloution and
the terminolgy accordingingly chosen. However, due to the generality of the investigated
models and mechanisms, most of the results and conclusions hold for any kind of tness
or payo dependent evolution.
3.3.1 Evolutionary dynamics
In this Section, we introduce one of the main dynamical systems to study evolutionary
game theory. On one hand, this serves as a good example of the general way of modeling,
on the other hand we will apply this kind of dynamics at several places in the thesis.
We consider the general symmetrical 22 game of Eq. (3.5). The whole population of indi-
viduals consists of species C and D, where individuals of each species do unconditionally
cooperate or defect, respectively. Assuming an innitely large population and random
interactions between individuals (a well-mixed population), one can write the expected
payo, that is assumed to be equivalent to reproductive tness, of a species as
C = x R + (1 x) S
= x (R S) + S
D = x (T P ) + P, (3.6)
where x denotes the global frequency of C individuals and 1 x the global frequency of
D individuals. A mathematical expression for the tness-dependent evolution of species
in a well-mixed, innite population is the replicator equation (Hofbauer and Sigmund,
1998; Nowak, 2006a; Nowak and Sigmund, 2004), also called game dynamical equation.
Neglecting mutations, the dynamics is straight-forward: the population share of a species
depends on its relative tness in the population. If species a has a higher tness than
average, its share in the population fa will grow proportionally to that dierence. On the
other hand, it will shrink in case it is less t than average. This means, in general,
d fa (t)
fa (t) = fa (t) (a (t) (t)), (3.7)
dt
where denotes the average tness in the population. Eq. (3.7) is applied to all species in
the population and, therefore, replicator dynamics can involve arbitrarily many species. In
our two-species system, the state of the population is suciently described by x, because
the share of the D species is simply 1 x. Hence, the whole replicator dynamics in this
case reads
x = x (1 x) (C D )
= x (1 x) (x (R + P S T ) + S P ). (3.8)
In order to nd the xed points of this dynamics, we set x = 0 and nd the three solutions
(i) x = 0
(ii) x = 1
P S
(iii) x = R+P ST
.
To nd out which of them will be the evolutionary outcome, we need to know about the
stability of these xed points. In a state close to a xed point, will the dynamics drive
the system into this xed point or away from it? Only in the rst case, we can expect the
system to end up in this state. Applying the considerations of Section 1.3.1, we derive
d x
x = 3 x2 (R + P S T ) + 2 x (R + 2P 2S T ) + S P (3.9)
dx
and conclude stability if x < 0. Depending on the ranking of the payo values (R, S, T, P ),
our dynamics can lead to the following scenarios (assuming 4 distinct values of R, S, T, P ):5
Dominance of C: 1 unique stable xed point in x = 1.

This occurs when R > T and S > P . In this case, species C dominates species D.
Dominance of D: 1 unique stable xed point in x = 0.

This occurs when T > R and P > S (like in the Prisoners Dilemma). In this case,
species D dominates species C.
Bistability: 2 stable xed points in x = 0 and x = 1 and 1 unstable xed point in

x = (P S)/(R + P S T ).
This is the case if R > T and P > S and corresponds to the coordination problems.
Coexistence: 1 stable xed point at x and 2 unstable xed points at x = 0 and

x = 1.
The reversed bistability scenario, where T > R and S > P . This corresponds to our
asynchronous coordination problems.
We can also relate this discussion to the concept of evolutionarily stable strategies (ESS,
(Maynard Smith, 1982; Maynard Smith and Price, 1973)). A strategy (a species, a geno-
type) is called evolutionarily stable if a population, consisting almost only of individuals
applying this strategy, cannot be invaded by a few individuals with another strategy. In
our case of only two possible strategies in the system (C and D), this implies the following:
For C to be an ESS against D, the rst of Maynard Smiths conditions is that the payo
of the solution (C, C) must be higher or equal to the solution (D, C), i.e. a D strategist
must not gain a higher payo than an C strategist when playing against C. If equality
holds, an additional condition must be fullled: the payo of (C, D) must be higher than
the payo of (D, D), i.e. a C strategist must be able to exploit D strategists in this case.
If the rst condition is a strict inequality or both conditions are fullled, strategy C is ESS
5
see also Hauert (2002); Nowak (2006b)
against strategy D. Let us connect this with the above considerations regarding stable
xed points and strict inequality of payo values: strategy C is ESS if R > T . Strategy D
is ESS if P > S. If only one strategy is ESS, we nd dominance of this strategy. If both
are ESS, we nd bistability. Finally, the absence of an ESS leads to coexistence with a
proportion x of individuals of species C and a proportion 1 x of individuals of species
D.
Fig. 3.3 uses color encoding to illustrate the evolutionary outcomes of Eq. (3.8) for all
possible payo rankings: red areas indicate the stability of cooperation, blue areas the
stability of defection, light blue, green, yellow, and orange colors indicate an interior
stable xed point. Note that in this Figure, we only consider stable xed points under
the assumption of equal initial conditions, i.e. x(t0 ) = 0.5. Therefore, in the bistability
region (R > T P > S), we show the stable xed point whose attractor region includes
the initial condition. Whereas here we only show the existence of this dierent dynamical
regimes, in Chapter 7 we will go into details of the symmetrical 22 games included in
Fig 3.3.
Let us again briey comment on the illustration of games in Fig. 3.3: we use the clas-
sication scheme presented in Fig. 3.2, which diers from the commonly used one, e.g.
in Hauert (2002) and Helbing et al. (2005b). There, the payo values R and P are kept
constant (with R > P ) and S, T are varied on the axes of the coordinate system. Of
course, the contained payo rankings (games) are the same, but the spatial arrangement
is dierent. However, the region where all four dynamical regimes are neighboring each
other (around the coordinate R = T = 1, P = S = 0, investigations often are restricted
to this interesting region, e.g. in Santos et al. (2006) and Roca et al. (2008)) is similarly
existent in both representations. The reason why we use a dierent representation in this
thesis is that it better supports and illustrates the argumentation of Chapter 7 and it is
more convenient for the reader if one and the same representation is used throughout the
whole work.
For the analysis of innite, well mixed populations, the results depicted in Fig. 3.3 will
be the reference scenario for the considerations in Chapter 6. In Fig. 3.4, we depict some
basic results of structured populations, but beyond that point the interested reader to two
comprehensive surveys, that contain these and many more results (Hauert, 2002; Roca
et al., 2008). Since, in such systems, analytical solutions are hard to obtain (if obtainable
at all), Fig. 3.4 shows results of numerical simulations. For comparison, panel (a) shows
the case of a fully connected network, i.e. the nite population analogon to Fig. 3.3. The
other panels show the cases of a random network in which every agent interacts with 8
neighbors (b) and a regular, 2-dimensional lattice with 4 (c) and 8 (d) nearest neighbors,
S = 0, T = 1
2 1
0.9
1.5
0.8
0.7
1
0.6
P
0.5 0.5
0.4
0
0.3
0.2
0.5
0.1
1 0
1 0.5 0 0.5 1 1.5 2
R
Figure 3.3: Analytically computed equilibrium fraction of cooperators according to repli-

cator dynamics between cooperators and defectors in an innite, well-mixed population.
In case of bistability (R > T P > S), the stable xed point with the bigger attractor
region is displayed. Every R, P -coordinate constitutes one specic payo matrix (a game)
with xed values of S = 0, T = 1.
respectively. Compared to Fig. 3.3, the region around the point P = S, R = T is amplied,
because there the important eects happen.
What can been seen is that structure matters in the evolutionary dynamics and that
spatial structure, i.e. when the nearest neighbors of my nearest neighbors are second
nearest neighbors to me, increases the inuence of cooperators in parts of the Prisoners
Dilemma, but acts detrimental to cooperation in parts of the Chicken game (Hauert, 2002;
Roca et al., 2008).
Let us nally remark on the algorithmic details applied in Fig. 3.4. The dynamics in such
simulations has two steps: (i) individuals in the system collect payos from interactions
with connected individuals (neighbors). This can be dierently realized, e.g. such that
each individual interacts with one randomly chosen neighbor or such that each individ-
ual interacts with all of its neighbors. Then, (ii) individuals revise their applied decision
0 0
0.2 0.2
0.4 0.4
0.6 0.6
0.8 0.8
1 1
1.2 1.2
1.4 1.4
1.6 1.6
1.8 1.8
2 2
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 1 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3
(a) all-to-all (b) random, k = 8
0 0
0.2 0.2
0.4 0.4
0.6 0.6
0.8 0.8
1 1
1.2 1.2
1.4 1.4
1.6 1.6
1.8 1.8
2 2
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
(c) 2-d, k = 4 (d) 2-d, k = 8
Figure 3.4: Simulation results: equilibrium fraction of cooperators after the evolution of
cooperators and defectors in a system of 400 agents. Considered are dierent population
structures: (a) fully connected networks; (b) random networks with 8 interactions per
agent; (c) 2-dimensional, regular lattices with 4 neighbors and (d) 8 neighbors, respectively.
Compared to Fig. 3.3, the most interesting region around P = S = 0, R = T = 1 is
magnied.
according to the observed success of other individuals.6 Also for this process, several re-
alizations are applied in the literature, e.g. the Fermi rule, which will be introduced and
6
This explanation uses cultural evolution as terminology. However, the same algorithms hold for
genetic evolution, where the terminology would include dying individuals and the appearance of offspring
at empty sites of the network.
applied in Chapter 6.3, the Moran rule (Moran, 1962), where also underperforming deci-
sions can be selected with low probability, or the so-called unconditional imitation process,
which deterministically choses the best performing decision from the neighborhood.7 For
the results of Fig. 3.4, we implemented the discrete equivalent of the replicator equation,
called replicator rule or proportional imitation rule (Hauert, 2002; Helbing, 1992; Roca
et al., 2008; Schlag, 1998). There, an individual imitates another individual only if it has
a higher tness, and if so, it imitates with a probability proportional to the (positive)
tness dierence between the other individual and itself.
3.3.2 The conundrum of cooperation

Evolutionary biologists are fascinated by cooperation. We think that this
fascination is entirely justified, because cooperation is essential for construc-
tion. Whenever evolution constructs a new level of organization, cooperation
is involved. The very origin of life, the emergence of the first cell, the rise of
multicellular organisms, and the advent of human language are all based on
cooperation. A higher level of organization emerges, whenever the competing
units on the lower level begin to cooperate. Therefore, we propose that coopera-
tion is a third fundamental principle of evolutionary dynamics besides mutation
and selection.
C. Taylor and M.A. Nowak (Taylor and Nowak, 2007)
Cooperative behavior is dened as an action that induces a tness-benet b to another

individual and a certain tness-cost c to the cooperator himself. The payo matrix
C D
! %
C b c c
, (3.10)
D b 0
with b > c > 0, denotes a specic Prisoners Dilemma and is often used as the standard
cooperation model in evolutionary biology. It illustrates the puzzle of cooperation: coop-
eration is dominated by defection and, therefore, selection should lead to the extinction
7
It is important to note that also this unconditional imitation rule is different from what we discuss
in this thesis as plain imitation within the context of social influence. Whereas the latter is a behavior of
imitation without fitness or payoff dependence, the imitation rule in evolutionary game theory considers
fitness dependent imitation.
of cooperative behavior. The question is then: how could cooperation lead to the ma-
jor transitions in evolution (Maynard Smith and Szathmary, 1995) and why are there
so many examples of cooperation in biology (e.g. single meerkats give alarm calls when
enemies approach, thereby exposing immediate danger to themselves, many animals help
other parents to breed ospring, for details and more examples see e.g. Clutton-Brock
(2002); Doebeli and Hauert (2005) and references therein)? Moreover, consider the evo-
lutionary success of mankind and the fact that this success is based on the high level of
cooperation within human civilization.
Hence, even if not obvious in the rst place, there must be mechanisms accounting for
the evolutionary success of cooperation, i.e. cooperation seems to induce an evolutionary
advantage to the individuals applying cooperative behavior. Many scientists from various
elds have addressed these questions in the last decades. Without any claim of complete-
ness, some of the most important concepts shall be mentioned: Hamiltons inclusive fitness
theory has led to the concept of kin selection, which can explain cooperation among closely
related individuals (Hamilton, 1963); Trivers reciprocal altruism (Trivers, 1971) and Ax-
elrods pioneering work on the iterated Prisoners Dilemma (Axelrod, 1984) emphasize
the importance of repeated interactions; Nowak and Sigmunds concept of indirect reci-
procity (Nowak and Sigmund, 1998) bases on reputation mechanisms that help directing
cooperative behavior at other cooperators, similarly to tag systems (Riolo et al., 2001)
and green beard eects (Dawkins, 1989; Hamilton, 1964; Traulsen and Nowak, 2007).
Other concepts comprise punishment (Fehr and Gachter, 2002), volunteering (Hauert et al.,
2002), network eects (Nowak and May, 1992; Santos et al., 2006, 2008), etc..
Based on Nowak (2006b), Taylor and Nowak (2007) implement ve evolutionary concepts
into the general Prisoners Dilemma payo matrix (3.5), i.e. they quantify the eects of
the dierent mechanisms by respectively modifying the basic payo matrix. Using these
matrices, they are able to calculate the conditions for cooperation to prevail under nat-
ural selection within the particular evolutionary scenarios. We will go into the details of
these investigations later on in Chapter 7, where we discuss the characteristics of par-
tial cooperation dilemmas and provide the evolutionary analysis with respect to these
mechanisms.
Particularly in Chapter 6.4 of this thesis, we will contribute to the unraveling of the
conundrum of cooperation by suggesting and anlyzing another mechanism to allow for
the evolution of cooperation. However, both Chapter 6 and Chapter 7 are focused on
investigations regarding the evolution of cooperation and disclose both new results on
the possibilty of cooperation to evolve (Chapter 6), and a novel modeling framework to
analyze cooperation dilemmas (Chapter 7).
Chapter 4
Social influence in evolutionary game

theory
In the previous Chapters, we discussed models of social interaction from two perspectives:
one in which the contentwise dierences between the alternative decisions are neglected
and agents do only adapt to the decisions of their interaction partners mostly they
tended to avoid being dierent from the others (imitational processes in models of opinion
dynamics). The second perspective was one in which agents experience explicit utility
consequences of their decisions (also dependent on the decisions of the others) and suc-
cessful decision making has an advantage for the individuals in the evolution of a system
(evolutionary game theory). As consequent third perspective, we will investigate the ef-
fect of implementing decisions based on social inuence in the competitive environments
of evolutionary game theory.
In fact, opinion dynamics and game theory are similar frameworks. Both consist of in-
teracting individuals that hold one opinion/strategy out of a set of alternatives. In both
cases, social interactions inuence the individual decisions and one is mainly interested in
the macroscopic outcome of such systems. Although the nature of interactions is quite
dierent (simple adaptation processes vs. the possibility of payo maximization), both
are aimed at modeling aspects of social interaction. It is therefore not surprising that
many methods can be used in both frameworks (Hauert and Szabo, 2005; Szabo and Fath,
2007). Moreover, microscopic imitation behavior in connection with spontaneous opinion
changes has been shown to lead to the replicator dynamics already more than a decade
ago (Helbing, 1992; Schlag, 1998). However, as regards content, both perspectives on so-
cial interaction are brought together rather seldomly, although we would argue that both
forms are present in real social systems.
To illustrate the commonalities and dierences further, let us have a look at a specic
47
48 Chapter 4. Social influence in evolutionary game theory
example: the voter model actually is also a model of evolutionary game theory, although
a very boring one. Consider the competition (through the replicator equation) between
two species in a game dened by R = S = T = P = 0 (in Eq. 3.5). In this system,
there is no selection pressure as all species have an equal, constant tness. Hence, every
state is a xed point, because both species have the same rate of reproduction, which
preserves global frequencies. In nite systems (with the replicator rule), neutral drift
might lead to dominance of one species which constitutes the voter model dynamics. In
order to build the bridge to more interesting evolutionary games, let us explain the same
matter dierently: consider now the coordination problem described by Eq. (3.1), i.e.
R = P = 1, S = T = 0. There, individuals of the same species can mutually increase their
tness, while an interaction between individuals of dierent species has no tness eect. In
order to describe the voter model dynamics, instead of the replicator dynamics we have to
assume that the evolutionary dynamics has no tness dependence, but is only frequency
dependent. One could assume a small poulation in a large system, where the required
resources are not scarce and both species can reproduce without selection pressure. After
some time of evolution, the whole population will reach a size at which resources start
to get scarce, i.e. competition occurs and the dynamics might assume the replicator
dynamics. This competitive system is then characterized by bistability 1 , where only the
two states of dominance of either species are stable xed points. For nite systems, this
means that the absorbing dominance state is not reached by neutral drift, but much faster
by the directed evolutionary dynamics (see Fig 4.1).
These considerations show on one hand the large common ground by menas of the general
modeling framework, but on the other hand also the relevant dierence: in general there is
no competition in opinion dynamics. Or, dierently formulated, individuals are not capa-
ble of assessing payo dierences. If voters in the voter model try to establish consensus,
they should only imitate successful individuals, because they seem to be in a (local or
global) majority. This would lead to a much faster consensus in the system than their
unconditional imitation of any individual. The distinction between these two behaviors
is essential for this thesis: the imitation of successful decisions is what is referred to as
imitation rule in the game theoretic literature. However, it is important to note that
this is not equivalent to the puristic concept of imitation by social inuence, that we
refer to when we speak of imitation in this work particularly in Capter 6.4.
In the literature, one approach to combine both research elds was undertaken by Galeotti
and Goyal (2007). There, the population consists of two levels: one of economic actors and
one of social entities. The latter ones are interconnected through a social network where
1
Please refer to the distinction of dynamical scenarios in Section 3.3.1.
49
1
Wealth CG
0.9 Wealth VM
Magnetization (o) / Wealth (x)
0.8 Magnetization VM
Magnetization CG
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 1 2 3 4
10 10 10 10 10
Time (generations)
Figure 4.1: Simulation results of two systems with N = 100 individuals, where every
individual is connected to all other individuals (fully connected network). The magenta
colored symbols display the results of the voter model dynamics (as explained in the
text, without selection pressure). The blue symbols represent results from the replicator
dynamics in the coordination problem, i.e. in a game with R = P = 1, S = T = 0. Circles
show the development of the absolute magnetization in the system, a measure for the
homogeneity of opinions in the system, where the value zero indicates equal frequencies and
the value one consensus (M = |12x|, if x is the relative frequency of one opinion). Crosses
denote the wealth in each system in terms of the expected payo of one individual
when playing the coordination problem game with a random other individual. Values are
averaged over 10,000 simulation runs for each system.
information is exchanged. The economic actors try to eciently spread information in the
social network. For this purpose, they can apply dierent strategies leading to a particular
payo. The focus of this work lies on the inuence of dierent network characteristics on
ecient information strategies (e.g. advertisement). In a nutshell, however, strategic
interaction only takes place between the economic actors and the individuals subject to
social inuence are merely part of the environment.
In the context of Chapter 6.4, we are specically interested in the role of social inuence
(as dened in Section 2) on evolutionary dynamics. Regarding our main model, we will
explain the straight-forward implementation of pure imitation behavior in the competition
50 Chapter 4. Social influence in evolutionary game theory
between cooperators and defectors and clearly relate it to similar, yet distinct approaches
in the literature. Most interestingly, the results presented in this Chapter show that
imitation is not able to dominate in a population, but due to its remarkable survival
abilities, it aects the evolutionary dynamics drastically. The novel approach here is to
combine the models of evolutionary game theory and of pure social inuence by separately
considering both forms of decision making (decisions through pure social inuence and
decisions completely independent of social inuence) exposed to the same evolutionary
rules. Within this thesis, there are two fundamental models forming the starting points of
the investigations: the voter model for the social inuence part and the replicator dynamics
applied to the evolution of cooperators and defectors in symmetrical 22 games for the
evolutionary game theory part. However, these two models are not only fundamental for
this thesis, but are also paradigmatic, interdisciplinary models for the scientic literature.
The approach in Chapter 6.4 is to straight-forwardly combine both models (put voters
together with cooperators and defectors) in order to address the question of how (positive)
social inuence aects the evolution of cooperation.
While a more detailed motivation of assuming social inuence decisions in evolutionary
dynamics is presented in Chapter 6.4, let us address some reasons for this assumption:
rst, decisions based on social inuence may reect asymmetries in either the information
or the capabilities of individuals. If an individual is not able to reasonably or consciously
decide on its own, imitating its environment might be the only reasonable strategy. Second,
imitation can be seen as an unconditional form of reciprocal behavior, in close connection
to the literature on the tit-for-tat strategy or other forms of direct reciprocity (Axelrod,
1984; Fehr and Gachter, 1998; Nowak and Sigmund, 1993). Third, imitation might be
motivated by egalitarian considerations (Dawes et al., 2007; Fehr et al., 2008), as at least
for our symmetrical games, the same decisions lead to equal payos. Fourth, imitation
might be a consequence of the persuasive power of other individuals (e.g. teaching
capabilities (Szolnoki and Szabo, 2007)). Fifth, imitation is relevant because it is an
empirical observation in social, economical, and biological systems (e.g. herding or
swarming behavior (Bonabeau et al., 1999; Choe and Crespi, 1997; Helbing et al., 2000;
Kirman, 1993; Trotter, 2005)).
Based on this argumentation on the presence and relevance of decisions purely through so-
cial inuence, i.e. independent of tness or payo considerations, we present in Chapter 6
dierent approaches to combine models of social inuence with models of evolutionary
game theory. In a rst approach to do so, we only assume imitators and contrarians,
i.e. individuals entirely susceptible to either positive or negative social inuence, in all
evolutionary, symmetrical 22 games. While the results of these investigations are less
attractive in that they mirror the evolutionary outcome of the reference dynamics be-
51
tween cooperators and defectors (see Fig. 3.3), they might contribute to the explanation
why negative social inuence plays are far smaller role in the social interactions of human
beings than herding and imitation does. In the main part of Chapter 6, we then inves-
tigate the already described mixture of imitational behavior with that of unconditional
cooperators and defectors in evolutionary games. Apart from this, we further connect the
opinion dynamics part of this thesis with the game theoretic part by testing a similar co-
evolutionary mechanism as proposed and discussed for the voter model in Chapter 5 in a
model of an evolutionary, spatial Prisoners Dilemma game with heterogeneous teaching
abilities by Szolnoki and Szabo (2007).
Part II
Models and analyses

This Part of the thesis is devoted to the specic models under consideration and the
results obtained by novel approaches. In Chapter 5, we introduce and investigate models
of social inuence. In particular, extensions of the voter model are investigated with the
aim to account for individual memory eects in the dynamics of opinion spreading (see
also Stark et al. (2008b) and Stark et al. (2008c)). In Chapter 6 we turn to evolutionary
game theory, but in fact continue the investigations of Chapter 5 with a dierent focus.
On one hand, we apply a similar, memory dependent interaction mechanism to a model
for studying the evolution of cooperation and analyze the results of this modication (see
also Szolnoki et al. (2009)). On the other hand, we investigate the combination of purely
imitational and more self-condent decision behavior in a novel model of evolutionary
game theory and nd that the inclusion of puristic social inuence remarkably aects the
evolutionary dynamics in social dilemmas (Stark and Tessone, 2010). Finally, Chapter 7
focuses on the particular models of evolutionary game theory themselves. The framework
of symmetrical 22 games is particularly often applied because of its ability to abstract
dierent strategic decision situations by relatively simple models. Whereas the Prisoners
Dilemma and other social dilemmas are well studied, we will introduce another class of
dilemmas that represent a social dilemma only in repeated and evolutionary games (see
also Stark (2010)). Besides presenting explicit results, this Chapter 7 also serves as an
outlook on promising future research.
Chapter 5
Slower is faster: Fostering consensus

formation by heterogeneous inertia
In this Chapter, we will investigate the eects of implementing social inertia into one of
the paradigmatic models of opinion dynamics: the voter model, which has been intro-
duced in Section 2.3. In our extension of the voter model, which we will introduce in
Section 5.1, voters are not zero-condent as in the standard voter model, but have a
kind of conviction regarding their own opinion which makes them inertial to change their
opinions. This inertia evolves with persistence time, i.e. the time elapsed since their last
opinion change. It seems natural that such an inertia would slow down the process of
consensus formation in the voter model. In remarkable disagreement with this hypothe-
sis, we will present simulation results of and analytical insight into our main nding, the
slower-is-faster eect on reaching consensus through inertial voters. We investigate in
depth under which circumstances the eect can be observed and introduce a theoretical
framework, that allows for the understanding of the phenomenon. For the purpose of a
complete picture, we both present the results of a simpler model, where there are only two
dierent levels of individual inertia present in the system (a zero and a non-zero inertia
below one, Section 5.3), and a model with an almost continuous spectrum of inertia val-
ues (ranging between zero and a maximum value below one, Section 5.4). Although the
observed eect is qualitatively very similar in both model setups, its realization diers in
some aspects and it is worth presenting the whole analysis within this thesis. After con-
cluding our investigations on the voter model, we propose further research on the inertia
mechanism in another dynamical model. Results of preliminary investigations allow the
conjecture of a related, but qualitatively dierent phenomenon caused by the dynamical
inertia mechanism.
53
54 Chapter 5. Slower is faster
5.1 (Social) Inertia in the voter model

Dierent from the standard voter model described in Section 2.3, we here consider that
voters additionally are characterized by a parameter i [0, 1], an inertia to change their
opinion. The basic probabilities to choose an opinion are those from the voter model, i.e.
equal to the respective frequencies in a voters neighborhood. However, the probability
to chose another than the current opinion is multiplied by i . Compared to the standard
voter model, this reduces opinion changes and increases the probability that a voter keeps
its opinion. Therefore, we have to distinguish between the probability that voter i changes
its opinion
Wi (i |i , i ) = (1 i ) WiV (i |i ) (5.1)
and the complementary probability of sticking to its previous opinion
Wi (i |i , i ) = 1 Wi (i |i , i ). (5.2)
In this setting, i represents the strength of condence that voter i has regarding its
opinion. WiV denotes the basic probabilities from the standard voter model.
In particular, we consider that the longer a voter has been keeping its current opinion, the
less likely it will change to the other one. Throughout this work, the linear relationship
i = min( i , s ) (5.3)
is used. Here, i denotes the persistence time of agent i, i.e. the number of update steps
since its last opinion change. The individual inertia increases by the growth rate until
it saturates at s < 1. This saturation below 1 avoids trivial frozen states where agents
will never change their opinion again. Note that the results presented in the following do
qualitatively not depend on the exact form of function (5.3), as long as i monotonically
increases with persistence time and saturates below 1.
For = 0, the standard voter model is regained. For values of > 0, our extension
diers from the standard voter model in that it considers the current opinion of voters
as an important decisive factor. The voters do not only act based on the frequencies
in their neighborhood, but take their own current opinion into particular account. In a
manner of speaking, we put some condence on the originally zero condent voters.
This general idea can also be compared to the models of continuous opinion dynamics
(see Deuant et al. (2000); Hegselmann and Krause (2002); Lorenz (2007)): already in the
5.1. (Social) Inertia in the voter model 55
basic continuous models, the current opinion of a deciding individual is of high importance.
More precisely, it is as decisive as the average opinion in the considered neighborhood
because the updated opinion is the average of both. The concept of bounded condence
emphasizes this importance because individuals do only approach opinions that are not
too far away from their own current one. Therefore, bounded condence can also be
interpreted as a kind of inertia that tend to let individuals keep their own current opinion.
However, the parameter regulating the condence interval is generally kept constant in
time whereas, in our model, individuals change their decision behavior dependent on their
history.
While the standard voter model is also called linear voter modelbecause the probabil-
ity to adopt an opinion is linearly dependent on the frequency of that opinion in the
neighborhoodthe extended version of the voter model considered here is, apart from the
= 0 case, a nonlinear voter model (Cox and Durrett, 1991; Schweitzer and Behera,
2009). However, as the inertia values are heterogeneous and co-evolving together with the
individuals opinions in the system, it is non of the nonlinear voter models as described
by Cox and Durrett (1991) or Schweitzer and Behera (2009), but constitutes a dynamical,
nonlinear voter model.
As already mentioned, an important implication of this inertia mechanism is that, starting
from a homogeneous population of voters, a heterogeneity of decision behaviors emerges
and evolves in time. Such a heterogeneity, although without evolution and only with
one voter being dierent from all the others, was also investigated by Kacperski and
Holyst (1997). However, these authors do not assume a linear dependency of adoption
probabilities on their frequencies (like in the standard voter model), but one following an
S-shaped Fermi function (corresponding to Glauber dynamics, that we will compare our
model with in Section 5.6).
The inertia parameter , that reduces the probability of state changes, can have dierent
interpretations in the various elds of application of the voter model: it may characterize
molecules that are less reactive, the permanent alignment of spins in a magnet, etc. In
economics, changes may be discarded due to transition- or sunk costs (investments in
the current solution, that will be lost in case of a switch to another solution). In social
applications, there are at least two interpretations for the parameter : (i) within the
concept of social inertia, which deals with a habituation of individuals and groups to
continue their behavior regardless of possible advantages of a change, (ii) to reect a
(subjective or objective) conviction regarding a view or an opinion. Originally, the latter
point served as a motivation for us to study the implications of built-in conviction in a
simple imitation model like the voter model. Will the systems, dependent on the level of
conviction, still reach a consensus state, or can we observe a (meta-) stable segregation of
opinions? How do the ordering dynamics and the emergent opinion patterns look like?
Our investigations focus on the average time to reach consensus, i.e. the number of
timesteps the system evolves until it reaches an equilibrium state in which all voters
have the same opinion. Taking into account the inertia introduced to the VM, we would
assume that the time to reach consensus shall be increased because of the slowed-down
voter dynamics. Counter-intuitively, we nd that increasing inertia in the system can
decrease the time to reach consensus. This result resembles the faster-is-slower eect
reported in a dierent context by Helbing et al. (2000). In their work on panic situations,
they explain why rooms can be evacuated faster if people move slower than a critical value
through the narrow exit door. When individuals try to get out as fast as they can, this
results in clogging eects in the vicinity of the door, which decreases the overall evacuation
speed. Note that although the phrase slower-is-faster is appropriate for both ndings,
our eect has to be clearly distinguished from the one described by Helbing et al.. In
their generalized force model, an individual increase in the desired velocity would have a
contrary eect on the microscopic level, i.e. all individuals would get slower and thereby
the macroscopic dynamics would be decelerated. In our case, microscopic changes produce
the counter-intuitive eect only on the macroscopic level.
5.2 Preliminary considerations
Before we go into the thorough investigations of the model, let us have a look at three
simplied models in order to obtain a rst hypothesis on the eects of the time to reach
consensus in the voter model extended by social inertia. Generally, the expected time to
consensus, T , can be found by an expected value formula: Let Pc (t) be the conditional
probability to reach consensus in the tth update event if consensus has not been reached,
yet. The expected value is found by summing for every timestep the evolution time t
multiplied by the probability to reach consensus exactly at this time (which is Pc (t),
multiplied by the probability that no consensus was reached before t). Assuming that
there is no consensus in the initial condition (t = 0, otherwise simply T = 0), we can
write

X t1
Y
T = Pc (1) + t Pc (t) (1 Pc ( )). (5.4)
t=2 =1
The rst term is simply Pc (1) since it is multiplied by t = 1 and our presumption that
5.2. Preliminary considerations 57
there is no consensus yet. All the other terms contain the factors t and the product of
complementary probabilities 1 Pc (t) for all times before t (the probability that there is
no consensus at time t 1).
Let us assume a system comprising N = 3 voters. Initially, the system is not in consensus,
which invariably means that two voters share the opinion and one has the opposite opinion.
According to the voter model dynamics, Pc = 2/9 at every update event t before consensus
is reached (probability of choosing the voter in minority as focal voter (1/3) times the
probability that it changes its opinion (2/3)). In this case, Eq. (5.4) reduces to

X
T,0 = Pc + t Pc (1 Pc )t1 . (5.5)
t=2
Therefore, T,0 follows a geometric distribution with the expected value
1
T,0 = , (5.6)
Pc
and yields the result 9/2 = 4.5 for the voter model dynamics. Let us now consider three
versions of our inertial voter model:
1. Homogeneous, static inertia:

If all voters have an equal, xed inertia , the time independent consensus probabil-
ity reduces from Pc = 9/2 to Pc, = (1 ) 2/9. Since Pc is time independent before
consensus is reached, we can use Eq. (5.6) to calculate T, and, for 0 < 1, we
obtain T, > T,0 . For example, if = 0.1, the expected consensus time would be
5 update events.
2. Dynamic, persistence time dependent inertia:

This is the scenario of our persistence-time dependent inertia mechanism. Here,

we need the inertia i (t) of the voter in minority for the calculation of Pc (t) (the
conditional probability of consensus in this case), which now is time dependent.
!
Therefore, we have to use Eq. (5.4). To prove that also in this case T T,0 , we
have to show that

X t1
Y
X

t Pc (t) [1 Pc ( )] t Pc (1 Pc )t1 ,
t=2 =1 t=2

using Eqs. (5.4) and (5.5) and the fact that Pc (1) = Pc .

First, we have to note that Pc (t) Pc for all timesteps t, i.e. the conditional
probability to reach consensus is maximal in a system without inertia and lower
in any system with non-zero inertia. Let us now describe the eect of the rst
appearance of inertia at timestep s on the expected consensus time compared to the
case with no inertia at s. We have Pc (s) < Pc , but s1
Q s1
=1 [1 Pc ( )] = (1 Pc ) ,
i.e. the exit probability to reach consensus at time s is reduced by inertia. The
dierence Pc Pc (s) is distributed over later timesteps via the term t1
Q
=1 [1 Pc ( )].
At time t = s + 1,
t
Y
[1 Pc ( )] = (1 Pc )t1 (1 Pc )
=1
> (1 Pc )t . (5.7)
Assuming that inertia is only present at timestep s and vanishes again afterwards,
Eq. (5.7) also holds for any other timestep larger than s (stochastic dominance
of the term t =1 [1 Pc ( )] over the term (1 Pc )t for any t > s). On the other
Q
hand, the conditional probability to reach consensus reassumes the value of a system

without inertia, i.e. Pc (t) = Pc for all t > s. Hence, for the event of non-zero inertia

only at one timestep, this proves a positive eect on the expected consensus time T
compared to the case that this event would not have happened (T,0 ).
This is a very special inertia mechanism, where inertia pops up one timestep and
vanishes afterwards. Now, we want to show that the same positive consensus time
eect appears for any inertia mechanism, including our persistence-time dependent
one. To show this, we consider the scenario from above as new reference scenario
and nd out the consequence of another event of non-zero inertia after timestep
s. Obviously, the same argumentation holds and, therefore, proves a positive eect
on the consensus time in this case compared to the new reference case, which al-
ready implied a positive consensus time eect compared to the non-inertial voter
!
model. This induction can be continued arbitrarily and proves that T > T,0 for any
persistence time dependent inertia mechanism at work (q.e.d.).
3. Heterogeneous, static inertia:

Here, we consider that every voter has a static, but individual inertia value , i . As in

the previous case, we need the inertia value of the minority voter to calculate Pc (t).
Although the individual inertia values are not time dependent in this scenario, the
expected inertia value of the minority voter changes by the evolution of the system.
In the rst update step, due to the random initial condition, it is simply the average
5.3. Results of binary inertia 59
of inertia values in the system. Afterwards, the voter in minority (and the respective
inertia value) is not anymore random, but depends on the evolution of the system.
To write down these expected values is not convenient, but also not necessary. We
can use the proof of the previously considered scenario and extend it to an arbitrary
inertia mechanism, thereby including the present scenario. The restrictive equality

Pc (1) = Pc , with which we expressed that, in the rst update step, no voter has
a non-zero inertia, can simply be neglected. The considered event of a non-zero
inertia value at only one timestep can also happen at t = 1 and will lead to the same
consequences as described above. Therefore, the proof above includes this scenario,
too.
Note that, in the course of calculating the eect on the expected consensus time in these
three scenarios, we achieved a proof for every possible inertial voter model. We have shown
that, in a system of N = 3 voters, where opinion updates follow the inertial voter model
dynamics (Eq. (5.1)), the expected consensus time invariably increases by: (i) the presence
of a non-zero inertia value at a single timestep, (ii) the number of such timesteps with a
non-zero inertia value, and (iii) the magnitude of the inertia values. In simple words, the
higher the level of inertia in such a system is, the longer the consensus process takes.
These considerations support the hypothesis that extending the voter model by any form of
inertial behavior will invariably lead to an increase of the expected time to reach consensus
if consensus will be reached at all by the dynamics. Note that N = 3 is the highest
number of voters in the system where such simple calculations are possible. Already for
4 voters, one needs to involve a more detailed expression of the frequencies of opinions in
order to compute the values Pc (t). So far, there was a minority opinion with a frequency
of 1/3, and a majority opinion with frequency 2/3. With N = 4, there would also be the
possibility of equal frequencies, leading to Pc (t) = 0. In general, for N > 3, Pc (t) depends
on dynamic frequencies, which makes the computations much more complicated.
5.3 Results of binary inertia

Having analyzed the implementation of an arbitrary inertia mechanism in a small voter
model system, we now want to investigate the results of our persistent time dependent
mechanism in a multi-agent system, i.e. for N > 3. For the sake of simplicity, let us rst
consider that the inertia growth rate is larger than the saturation value of inertia, i.e. in
s holds in Eq. (5.3). Since we then only have one non-zero inertia value, we simply
denote it by and the respective function reads
(
0 , if = 0
i ( ) = . (5.8)
, if > 0
At time t = 0, and in every timestep after voter i has changed its opinion, the persis-
tence time is reset to zero, i = 0, and the inertia has the minimum constant value 0 .1 .
Whenever a voter keeps its opinion, its inertia increases to . We will study two distinct
scenarios later on: (i) xed social inertia where 0 = is a constant value for all voters.
(ii) 0 < , a scenario in which inertia grows for larger persistence times.
It would be expected that including inertial behavior in the model would invariably lead
to a slowing-down of the ordering dynamics. We will show that, contrary to this intuition,
these settings can lead to a much faster consensus.
We performed extensive computer simulations in which we investigated the time to reach
consensus, T , for systems of N voters. We used random initial conditions with equally dis-
tributed opinions and an asynchronous update mode, i.e. on average, every voter updates
its opinion once per timestep. The numerical results correspond to regular ddimensional
lattices (von-Neumann neighborhood) with periodic boundary conditions, and small-world
networks with a homogeneous degree distribution.
In order to obtain a benchmark for the results of our model, we rst consider the case of a
fixed and homogeneous inertia value 0 = . In the limit 0, we recover the standard
voter model, while for = 1 the system gets frozen in its initial state. For 0 < 1,
the time to reach global consensus is aected considerably; the systems still always reach
global consensus, but this process is decelerated for increasing values of . This can be
conrmed by computer simulations which assume a constant inertia equal for all voters
(see left panel in Fig. 5.1).
In the right panel of Fig. 5.1, we depict the evolution of the interface density , as intro-
duced in Eq. (2.4), for both the standard voter model and the inertial voter model with
0 = = 0.5. Dierences between these cases can be seen in the very beginning and
at about 103 time steps, right before the steep decay of disorder in the system. There,
the ordering process is slower in the voter model with inertia than in the standard voter
model.
This behavior can be well understood by analyzing Eq. (5.1). There, we nd that the ratio
between the opinion changes in the standard voter model and the inertial voter model is
given by W (|, )/W V (|, 0) = (1 ). Consequently, it is reasonable to infer
1
Note that the results of this Chapter are qualitatively robust against changes in the concrete function
i ( ). For more details on this, refer to the next Section
0
10
5 simulations
10
theory
1
10 ==0
0
0 = = 0.5
<>

T
4
10 0
10
2
10 1
10
2
10 <>( = 0,t)
<>( = 0.5,t) / 0.5
3
10
3 0
10 10
1
10
2
10
3 4
10
10 3
2 1 0
10 0 1 2 3 4
10 10 10 10 10 10 10 10
0 = timesteps
Figure 5.1: Left: Average time to consensus T in the voter model with a xed and
homogeneous inertia value 0 = . The line corresponds to the theoretical prediction
T () = T ( = 0)/(1 ). Details are given in the text. Right: comparison of the
development of the average interface density in the voter model and the model with
xed inertia. Right, inset: collapse of the curves when the time scale is rescaled according
to t t/(1 ). In both panels, the simulations results stem from averaging over 500
sample runs, where the system size is N = 30 30 and the voters are placed on a two-
dimensional, regular lattice.
that the characteristic time scale for a voter model with xed inertia will be rescaled as
t tV /(1). As can be seen in Fig. 5.1, there is good agreement between this theoretical
prediction and computer simulations in both: the average time to consensus (see panel
(a)), and the time evolution of the interface density (inset of panel (b)). In conclusion:
analyzing a benchmark model with xed inertia, the considerations of Section 5.2 are
conrmed.
As second benchmark, let us have a look at fixed, but heterogeneous inertia values within
the system. This means that voters have dierent inertia values, but these do not change
over time. Since heterogeneity endogeneously appears in our proposed model of persistence
time dependent inertia, this benchmark is important to obtain another reference with
the mere existence of heterogeneity in inertia values. For this purpose, we initialized
simulations of the voter model in fully connected networks, i.e. where every voter is
connected to all other voters. Every voter is assigned with a xed, individual inertia
value. In panel (a), we see the average results of systems with N = 200 voters. We measure
the time-to-consensus in systems with dierent non-zero inertia values and with dierent
frequencies of individuals holding the non-zero inertia value. The results are similar to
those of the rst benchmark: consensus is always reached, but with increasing the level of
1.5
1.4

2.8
Average timetoconsensus T
10
1.3
T () / T (0)
2.6
10

1.2

2.4
10
1.1
2.2
10
N = 25
1 N = 49
0 N = 100
10 0
1 10 N = 400
10 10
1
0.9 3
2 2 2 1 0
Frequency of voters with 10 10 10 10 10 10
Nonzero inertia value
nonzero inertia Maximum individual inertia
(a) (b)
Figure 5.2: Average consensus times in simulations of the voter model embedded into
fully connected networks. (a) Surface of average consensus times when varying both the
non-zero inertia value and the frequency of voters holding the non-zero inertia value.
The system size is N = 200 and averages are obtained from 4000 simulation runs. (b)
Here, the inertia values are not binary, but continuous between 0 and the maximum inertia
value and all the voters have a non-zero inertia. Results are presented for dierent system
sizes N and the number of simulation runs varies between 10000 for the smallest system
size and 2000 for the largest system size.
inertia in the system (in either way) the time needed to reach the ordered state is generally
not decreasing, but increasing for larger levels of inertia in the system. In panel (b), we also
show results of continuous instead of binary inertia values, for dierent sizes of the voter
model system. Here, we assign each voter with a random inertia value (i ) between zero
and the maximal inertia value (). Completing the investigations of benchmark models,
also these simulation results indicate that exogeneously induced inertia in the voter model
leads to the expected increase in consensus times.
We now turn our attention to the case where the individual inertia values evolve with
respect to the persistence time according to Eq. (5.8). Without loss of generality, we
x 0 = 0. Other choices simply decelerate the overall dynamics as described in the
previous subsection. Note that increasing increases the level of social inertia in the
voter population. Fig. 5.3 shows the average time to reach consensus as a function of the
parameter , namely the maximum inertia value reached by the voters when the system
is embedded in regular lattices of dierent dimensions. In Fig. 5.3, it is apparent that, for
lattices of dimension d 2, the system exhibits a noticeable reduction in the time to reach
consensus for intermediate values of the control parameter . We observe that there is a
critical value of such that the average time to reach consensus has a minimum. Especially
compared to the results of the previous section, this result is against the intuition that a
slowing-down of the local dynamics would lead to slower global dynamics. Furthermore,
it is also apparent that the larger the dimension of the lattice, the more pronounced the
phenomenon is. Increasing the dimension of the system enlarges the region of , in which
consensus times decrease with increasing . This fact accounts for the main dierences in
the shapes of the functions in panels (b-c) of Fig. 5.3.
(a) (b)
7
10 6
10
6
10
5
5 10
10
T
4 4
10 10
3
10 3
10
2
10
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(c) (d)
5 4
10 10
4
10 3
10
T
3
10
2
2 10
10
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 5.3: Average time to reach consensus T as a function of the maximum inertia value
. Panels (a)-(d) show the results for dierent system sizes in one-, two-, three-, and four-
dimensional regular lattices, respectively. The results are averaged over 104 realizations.
The system sizes for the dierent panels are the following: (a) N = 50 (), N = 100 (),
N = 500 (); (b) N = 302 (), N = 502 (), N = 702 (); (c) N = 103 (), N = 153
(), N = 183 (); (d) N = 44 (), N = 54 (), N = 74 ().
Fig. 5.3(a) shows the results for a one-dimensional lattice, where the phenomenon is not
present at all. For this network topology, it is found that all the curves collapse according
to a scaling relation T (, N ) = T ()/N 2 , which is in accordance with the scaling of
5
10
4
10
T
3
10
2
10
0 0.2 0.4 0.6 0.8 1

Figure 5.4: Average time to reach consensus T as a function of the maximum inertia
value in small world networks (see text for details). The symbols represent dierent
rewiring probabilities , starting with a 2-dimensional, regular lattice ( = 0). The
curves corresponds to = 0 (), = 0.03 (), = 0.1 (), and = 0.9 (). System
size is N = 302 and results are averaged over 104 realizations.
consensus times dependent on system size in the standard voter model (T N 2 , as

introduced in Section 2.3).
In Fig. 5.4, we plot T as a function of the maximum inertia value for dierent small-
world networks (Watts and Strogatz, 1998). The small-world networks are constructed as
follows: starting with a two-dimensional regular lattice, two edges are randomly selected
from the system and, with probability , their end nodes are exchanged (Maslov et al.,
2003). In this procedure, the number of neighbors remains constant for every voter. It
can be seen that the phenomenon of lower consensus times for intermediate inertia values
is also present in small world networks. Furthermore, increasing the rewiring probability
leads to larger reductions of the consensus times at the optimal value c . This implies
that the formation of spatial congurations, such as clusters, is not the origin of this
slower-is-faster eect.
Finally, we show the results on a fully-connected network, i.e. where every voter has N 1
neighbors. The results are shown in Fig. 5.5. As can be seen, the time to reach consensus
is signicantly decreased for intermediate values of .
As already mentioned, the results of Figs. 5.4 and 5.5 indicate that the spatial clustering
4
10
N = 100
N = 200
N = 1000
3
T 10
2
10
0 0.2 0.4 0.6 0.8 1

in fully connected networks of dierent size. Results are averaged over 104 realizations.
plays no important role for the voters aging2 and, therefore, for the qualitative behavior
observed. Hence, we will investigate the dynamics in the so-called mean-eld limit, which
eectively means that all voters are connected with each other and the system is innitely
large (N ). For this purpose, we now use the global frequencies of opinions to
calculate the transition probability Wi (i |i , i ) in Eq. (5.1). As we will see, this allows
us to analytically approach the respective dynamics.
Let us rst introduce the quantities a and b : a0 (t) (b0 (t)) as the fraction of voters with
opinion +1 (1) and persistence time i = 0, i.e. inertia state i (0) = 0 = 0, and a1 (t)
(b1 (t)) represent the fraction of voters with opinion +1 (1) and persistence time i 1,
i.e. inertia state i (t) = . Thus, voters with opinion +1 that changed their opinion in
the last update step would contribute to the quantity a0 (t), without an opinion change
they would contribute to a1 (t). The global frequency of opinion +1, let us denote it by A,
at time t is given by
A(t) = a0 (t) + a1 (t). (5.9)
Fig. 5.6 illustrates the possible transitions of voters from one fraction to another, visual-
izing the above described.
In the mean eld limit, we can write down the equations for the change in the single
2
By aging we mean the possibility to build up higher persistence times that in turn lead to increasing
inertia values.
a0 a1
b0 b1
Figure 5.6: Illustration of the four fractions a and b and the possible transitions of a
voter.
quantities, i.e. the evolution equations, as
a0 (t + 1) a0 (t) = W (+1| 1, 0) b0 (t) + W (+1| 1, ) b1 (t)

[W (1| + 1, 0) + W (+1| + 1, 0)] a0 (t), (5.10)
a1 (t + 1) a1 (t) = W (+1| + 1, 0) a0 (t) W (1| + 1, ) a1 (t), (5.11)
where all quantities at time t + 1 only depend on quantities at time t, i.e. the system
only has a memory of 1 timestep. For this reason, our system represents a Markov-
chain model (Helbing, 1993; Van Kampen, 2007). For the expressions in Eqs. (5.10)
and (5.11), the global probabilities for changing opinion and for sticking to the current
opinion, respectively, are easily found by using Eqs. (2.1) and (5.1),
W (1| + 1, 0) = W V (1| + 1) = B(t),

W (+1| + 1, 0) = W V (+1| + 1) = A(t),
W (1| + 1, ) = (1 ) W V (1| + 1) = (1 ) B(t),
W (+1| + 1, ) = 1 (1 ) W V (+1| + 1) = B(t) + A(t). (5.12)
The missing expressions are analogous to the ones given above and can be derived by
consistently exchanging a b, A B, and +1 1 in Eqs. (5.10), (5.11) and (5.12).
After some steps of straight-forward, but not very illustrative, algebra, we get for the
example of the +1 opinion

a0 (t + 1) a0 (t) = A(t) b0 (t) + (1 ) b1 (t) a0 (t), (5.13)
a1 (t + 1) a1 (t) = A(t) a0 (t) + B(t) a1 (t) ( 1), (5.14)
and the analogous equations for opinion 1

b0 (t + 1) b0 (t) = B(t) a0 (t) + (1 ) a1 (t) b0 (t),
b1 (t + 1) b1 (t) = B(t) b0 (t) + A(t) b1 (t) ( 1).
The global frequency of the +1 opinion evolves as the sum of Eqs. (5.13) and (5.14)
which yields, again after some steps of straight-forward algebra, the change in the global
frequency
A(t + 1) A(t) = [b0 (t) a1 (t) a0 (t)b1 (t)] . (5.15)
For = 0, i.e. the standard voter model, we obtain the general conservation of magne-
tization that we already have seen in Eq. (2.3). For > 0, everything depends on the
quantities a (t) and b (t). If there is no heterogeneity of social inertia in the system, i.e. if,
at some time, either a0 (t) + b0 (t) = 1 or a1 (t) + b1 (t) = 1, then there also is no dynamics in
the magnetization. The same holds if both products in the squared brackets of Eq. (5.15)
are equally high. This is true if A = B and the ratio of inertial voters is the same within
the two global frequencies, i.e. if a0 (t) = b0 (t).
In the remaining congurations of these four quantities, there is a dynamics in the mag-
netization of the system. This implies that, even if the global frequencies of the opinions
are the same (A = B = 0.5), we can nd an evolution towards full consensus at one of
the opinions. Interestingly, the opinion whose frequency is increasing can be the minority
opinion in the system. In general, at every timestep opinion +1 has an increasing share
of voters in the system whenever its internal ratio of inertial voters satises the inequality
a1 b1
> . (5.16)
a0 b0
Analogously, B increases if inequality (5.16) is reversed. However, the complete process

is nonlinear and, therefore, it is not possible to derive the nal outcome of the dynamics
from Eq. (5.16).
Note that condition (5.16) is evidence of the important role of the heterogeneity of voters
on the dynamics in the system. More precisely, the main driving force of the observed
slower-is-faster eect is the voters heterogeneity with respect to their inertia.
In order to have an analytical estimation of the eect of social inertia on the times to
consensus, we initialize the system in a situation just after the symmetry is broken. In
particular, we articially set the initial frequencies to dier slightly, i.e. we set a0 (0) =
1/2 + N 1 and, hence, b0 (0) = 1/2 N 1 .3 Then we iterate according to Eqs. (5.13)
and (5.14). Furthermore, we assume that the consensus is reached whenever for one opinion
a0 (t) + a1 (t) N 1 holds.4 This is due to the fact that for a system of size N , if the
frequency of the minority state falls below N 1 , the absorbing state is reached. As we are
interested in the eect of dierent inertia-levels, we again use as control parameter and
compare the results with computer simulations of the identical setup of our inertial voter
model. In Fig. 5.7, the lines correspond to this theoretical analysis, where a qualitative
agreement can be seen with the simulation results (quantitative dierences result from
nite size uctuations in the simulations, i.e. from the fact that the numerical simulations
are run with nite system sizes, whereas the analytical solution assumed innite systems).
5.4 Results of multiple inertia states

To be more general, we also present the results and analysis of our inertia mechanism
with more than two inertia states. For the sake of simplicity, the results presented here
assume that the individual inertia i increases linearly with persistence time i , being
the strength of this response, until it reaches a saturation value s , i.e.
(i ) = min ( i , s ) .
Choosing s < 1 avoids trivial frozen states of the dynamics 5 . The rate of inertia growth
determines the number of timesteps until the maximal inertia value is reached, denoted
as
3
We also calculated the theoretical predictions for breaking the symmetry in the other way, namely
by setting a0 (0) = 1/2 N 1 and a1 (0) = N 1 . Here, again opinion +1 is favored, but this time just
by a higher fraction of inertial voters. The initial frequencies of opinions are equal. Qualitatively, this
procedure leads to the same theoretical predictions.
4
With the described initial condition, +1 can be the only consensus opinion.
5
The results presented here are qualitatively independent of the exact functional relation i (i ), as
long as a monotonously increasing function with a saturation below 1 is considered. This statement stems
from additional analyzes in the course of these investigations, e.g. using a multi nomial logit function.
5.4. Results of multiple inertia states 69
4
10
N = 100
N = 200
N = 1000
theory (N = 100)
theory (N = 1000)
3
10
T
2
10
0 0.2 0.4 0.6 0.8 1

in fully connected networks of dierent size. Symbols show the simulation results for
dierent system sizes (averaged over 104 realizations), lines the results of the theoretical
estimation (explained in the text).
s = (s /) .
Increasing increases the level of inertia within the voter population, thereby slowing-
down the microscopic dynamics. Like in the discussion above, one would intuitively assume
an increase of the average time to reach consensus. Interestingly, this is not always the case
as simulation results of T () show for dierent network topologies (see Fig. 5.8). Instead,
it is found that there is an intermediate value , which leads to a global minimum of T .
For < , consensus times decrease with increasing values. Only for > , higher
levels of inertia result in increasing consensus times.
In panel (a) and in parts of panel (b) of Fig. 5.8, one can also clearly see a global maximum
in the consensus times depending on the control parameter . The fact that such a
maximum does not occur in fully connected networks and, especially, that it vanishes
in small-world networks for high rewiring probabilities, indicates that this phenomenon
results from spatial congurations in the respective systems. In contrast to this, the
minima occur for every of the investigated systems and we will focus on this very general
phenomenon in the following. Although it might be interesting to investigate the respective
spatial congurations leading to these maxima in the results, this will not be part of the
more general investigations conducted here.
(a) (b) (c) 2

10
4
3 10
4 10 T
10
5 T

10 0
10
2
T 10
10
1
1 2
4 10 1 2 3 4 10
10 10 10 10 10 10
3
10
1 0
10
1
10
N 3
10 2
3 10
10
2 2
10 10
4 3 2 1 0 4 3 2 1 0 2 1 0
10 10 10 10 10 10 10 10 10 10 10 10 10
Figure 5.8: Average consensus times T for varying values of the inertia slope and xed
saturation value s = 0.9. Sample sizes vary between 103 104 simulation runs. Filled,
black symbols always indicate the values of T at = 0. (a) 2d regular lattices (ki = 4)
with system sizes N = 100 (), N = 400 (), N = 900 (). The inset shows how
consensus time scales with system size in regular lattices at = for d = 1 (), d = 2
(), d = 3 (), d = 4 (). (b) Small-world networks obtained by randomly rewiring a 2d
regular lattice with probability: () pr = 0, () pr = 0.001, () pr = 0.01, () pr = 0.1,
() pr = 1. The system size is N = 900. (c) Fully connected networks (mean eld case,
ki = N 1) with system sizes N = 100 (), N = 900 (), N = 2500 (), N = 104 ().
Lines represent the numerical solutions of Eqs. (5.18), (5.19), (5.20) with the specications
in the text. The inset shows the collapse of the simulation curves by scaling and T as
explained in the text.
Before we analyze the dynamics similarly to the previous Section, let us provide some tech-
nical details of the simulation results: for a two-dimensional lattice, shown in Fig. 5.8(a),
we nd that the value of , for which a system nds its minimum in the average consensus
times, scales with the system size as 1/ ln N . Simulations of regular lattices in other
dimensions show that the non-monotonous eect on the consensus times is amplied in
higher dimensional systems. Being barely noticeable for d = 1, the ratio between T ( )
and T ( = 0) (i.e. the standard voter model) decreases for d = 3 and d = 4. We further
compare the scaling of T with system size N for the standard and the modied voter
model. For regular lattices of dierent dimensions (d), we have for the standard voter
model
d = 1: T N 2
d = 2: T N log N
d > 2: T N
Note that for d > 2, the systems do not always reach an ordered state in the thermody-
namic limit. In nite systems, however, one nds the above written scaling relation.
In the modied voter model, we instead nd that T ( ) scales with system size as a
power-law, T ( ) N (see inset in Fig. 5.8(a)); where = 1.99 0.14 for d = 1 (i.e., in
agreement with the standard voter model); = 0.98 0.04 for d = 2; = 0.5 0.08 for
d = 3; and = 0.3 0.03 for d = 4. For xed values of > , the same scalings apply.
These results suggest (at least approximative) the following the scalings in the modied
voter model:
d = 1: T N 2
d = 2: T N
d = 3: T N 1/2
d = 4: T N 1/3 .
In order to cope with the network topology, in Fig. 5.8(b) we plot the dependence of the
consensus times T for small-world networks built with dierent rewiring probabilities.
Again, the degree of each node is kept constant by randomly selecting a pair of edges and
exchanging their ends with probability p (Maslov et al., 2003). It can be seen that the
eect of reduced consensus times for intermediate values of still exists and is amplied by
increasing the randomness of the network. This result implies that the spatial extension
of the system, e.g. in regular lattices, does not play a crucial role in the emergence of
this phenomenon. This can be conrmed by investigating the case shown in Fig. 5.8(c),
in which the neighborhood network is a fully-connected one (the solid lines correspond to
a theoretical approximation introduced below). The inset shows the results of a scaling
analysis, exhibiting the collapse of all the curves by applying the scaling relations =
| ln( N ) 1 |, and T = T / ln(N/) , with = 1.8(1), 1 = 1.5(1), = 7.5(1). This
shows that the location of the minimum, as well as T , scales logarithmically with N .
For the analytical approach to these results, we need more than the four quantities used
in the previous Section. Now, the global frequencies a (t) (b (t)) stand for voters with
opinion +1 (1) and various persistence times (compare Figs. 5.6 and 5.9).
In analogy to Eq. (5.9), the frequencies satisfy
X X
A(t) = a (t), B(t) = b (t).

a0 a1 a2 aT
b0 b1 b2 bT
Figure 5.9: General illustration of the fractions a and b and the possible transitions of
a voter. The fractions aT and bT contain all voters with a persistence time i s , i.e.
voters with maximal inertia.
Since the dynamics satisfy the Markov assumption, the rate equations for the evolution of
these subpopulations in the mean-eld limit are given by the master equation
Xh i
a (t) = (a |a )a + (a |b )b

Xh i
(a |a ) + (b |a ) a . (5.17)

Due to symmetry, the expressions for b (t) are obtained by consistently exchanging a
b , i.e.
Xh i
b (t) = (b |b )b + (b |a )a

Xh i
(b |b ) + (a |b ) b .

These equations are very general and we will identify their constituents in detail below.
Note that most of the terms in Eq. (5.17) vanish because only two transitions are possible
for a voter: (i) it changes its state, thereby resetting its to zero, or (ii) it keeps its current
state and increases its persistence time by one. Case (i) is associated with the transition
rate (b0 |a ), that in the mean-eld limit reads
(b0 |a ) = (1 ( )) B(t).
B(t) is the frequency of voters with the opposite state that trigger this transition, while
the prefactor (1 ( )) is due to the inertia of voters of class a to change their state. For
case (ii),
(a +1 |a ) = 1 (b0 |a ),
since no voter can remain in the same subpopulation. I.e., in the mean-eld limit, the
corresponding transition rates are
(a +1 |a ) = A(t) + ( )B(t).
Therefore, if > 0, Eq. (5.17) reduces to
a (t) = (a |a 1 ) a 1 (t) a (t)

h i
= A(t) + ( 1)B(t) a 1 (t) a (t). (5.18)
On the other hand, the fraction of voters with = 0 evolves as
X
a0 (t) = (a0 |b )b (t) a0 (t)

h i
= A(t) B(t) IB (t) a0 (t). (5.19)
Due to the linear dependence of the transition rates on inertia, the terms involving can
be comprised into IB (t) and IA (t). The latter expressions stand for the average inertia of
voters with state 1 and +1, respectively, i.e.
X
IA (t) = ( )a (t)

X
IB (t) = ( )b (t). (5.20)

Expressions (5.18), (5.19), (5.20), and the corresponding ones for subpopulations b can
be used to give an estimate of the time to reach consensus in the mean-eld limit. Let us
consider an initial state a0 (t) = A(0) = 1/2 + N 1 and b0 (t) = B(0) = 1/2 N 1 , i.e.
voters with state +1 are in slight majority. By neglecting uctuations in the frequencies
(which drive the dynamics in the standard voter model), these equations are iterated until
B(t ) < N 1 (i.e. for a system size N , if the frequency of the minority state falls below
N 1 , the absorbing state is reached). Then, we assume T = t . The full lines in Fig. 5.8(c)
show the results of this theoretical approach, exhibiting the minimum and displaying good
agreement with the simulation results for large values of . For low values of , uctuations
drive the system faster into consensus compared to the deterministic approach.
Inserting Eqs. (5.18) and (5.19) into the time-derivative of Eq. (5.17) yields, after some
straight-forward algebra and in analogy with Eq. 5.15, the time evolution of the global
frequencies
A(t) = IA (t) B(t) IB (t) A(t). (5.21)
Compared to the standard voter model, the magnetization conservation is now broken
because of the inuence of the evolving inertia in the two possible states. For ( ) =
(i.e. a time-independent inertia, that includes the standard voter model, = 0), we
regain the magnetization conservation. This can be shown by reconsidering Eqs. (5.9) and
(5.20):
X
IA (t) = a (t)

X
= a (t)

= A(t). (5.22)
Then, considering the analogous considerations for IB (t), Eq. (5.21) reads
A(t) = IA (t) B(t) IB (t) A(t)

= A(t) B(t) B(t) A(t)
= [A(t) B(t) B(t) A(t)]
= 0, (5.23)
i.e., independent of the state of the system, there is no change in the global frequency of
opinion A and, consequently, also not in the frequency of opinion B. The same amount of
voters changing from opinion +1 to 1, change from 1 to +1.
However, let us return to the case of a persistence-time dependent inertia. Interestingly

enough, Eq. (5.21) implies that the frequency A(t) grows i.
IA (t)/A(t) > IB (t)/B(t).
When the time dependence of the inertia on the persistence time is a linear one, as assumed
in this Section, inserting Eqs. (5.18, 5.19) into Eq. (5.20), we obtain an equation for the
time evolution of IA (t) up to rst order in :
IA (t) = A(t) IA (t) + A2 (t) IA (t) + O(2 , aT ). (5.24)

P
Here, aT = s a contains all subpopulations with maximum inertia. The last term,
O(2 , aT ), stands for the omitted terms that contain 2 and aT . This simplication is
viable if these variables are so small that they can be neglected without changing the
results, which will be taken care of below. Eqs. (5.21) and (5.24) correspond to the
complete macroscopic level description of this model. Hence, we can analyze the system
of equations by identifying its xed points. We nd the saddle point,
A = B = 1/2 ; IA = IB = /2 + O(2 ),
and two stable xed points, one at
A = 1; IA = s
and another at
B = 1; IB = s .
Note that the saddle point is close to the initial condition of the simulations. Neglecting
uctuations, the time to reach consensus has two main contributions: (i) the time to
escape from the saddle point, Ts ; and (ii) the time to reach the stable xed point, Tf ;
namely
T Ts + Tf .
We then linearize the system around the xed points and calculate the largest eigenvalues
s and f (for the saddle and the stable xed points, respectively) as a function of . In
Section 1.3, we already explained that the sign of the eigenvalues disclose the direction of
the dynamical ow along each dimension of the system. Furthermore, the magnitude of
the eigenvalues indicate the speed of the respective dynamical ows. At the saddle point,
we nd
p
s () = 1 + 20 + 4 2 2 1 + O(2 ),
which equals 0 at = 0 and monotonously increases with . For larger values of ,

where the rst order term expansion is no longer valid, numerical computations show
that s continues to increase monotonously with . This means that, for larger inertia
growth rates , the system will escape faster from the saddle point, thereby reducing the
contribution Ts to the consensus time T . On the other hand, for 0, s vanishes and
the system leaves the saddle point only due to uctuations.
Near the stable xed points the contribution of aT to Eq. (5.24) cannot be neglected
anymore. Solving the complete expressions with the help of the technical computing
software Mathematica, we obtain f,1 = s for < 1 s , whilst f,2 = 1 for
1 s . Interestingly, both reect dierent processes: the eigenvalue f,1 is connected
to voters sharing the majority state which are, at the level of s , inertial to adopt the
minority one (signalled by f,1 being constant). For 1 s , the largest eigenvalue
f,2 is related to voters with the minority state that are, for increasing , more inertial to
adopt the majority state (apparent by the decrease in |f,2 |).
The contributions Ts and Tf are two competing factors in the dynamics towards consensus.
Qualitatively, they can be understood as follows: in the beginning of the dynamics, the
inertia mechanism amplies any small asymmetry in the initial conditions. While this
causes faster time to consensus for (small) increasing values of , for suciently large
values of inertia growth, another process outweighs the former: the rate of minority voters
converting to the nal consensus state is considerably reduced, too. It is worth mentioning
that the phenomenon described here is robust against changes in the initial condition:
starting from IA = IB < s , it holds for any initial frequencies of opinions. Conversely,
starting from A = B = 1/2, it holds for any IA 6= IB .
5.5 Conclusions
The time for reaching a fully ordered state in a two-state system such as the voter model is
a problem that attracted attention from dierent elds in the last years. In this Chapter,
we studied the eect of social inertia in the voter model based on the assumption that
5.5. Conclusions 77
social inertia grows with the time the voter has been keeping its current opinion. We
focus our study on how the times to consensus vary depending on the level of inertia in
the population. In the two scenarios considered, this level of inertia was steered by the
value of the only non-zero inertia parameter and the rate of inertia growth , respectively.
In contradiction with the expectation that increasing inertia may lead to increasing times
to reach consensus (which has been underlined by benchmark investigations), we nd
that, for intermediate values of (), this inertia mechanism causes the system to reach
consensus faster than in the standard voter model. Interestingly, this phenomenon implies
that individuals reluctant to change their opinion can have a counter-intuitive eect on
the consensus process, which was studied for some particular cases before (Galam, 2005).
Furthermore, an inertial minority can overcome a less inertial majority in a similar fashion
as previously discussed by Galam (2002) and Tessone et al. (2004).
We show that the phenomenon discussed here is robust against the exact topology of the
neighborhood network as we nd it in regular lattices and small-world networks. In the
former it holds that the higher the dimension, the more noticeable the eect. Furthermore,
we found that the phenomenon also appears in random and fully-connected networks.
In simple words, this intriguing slower-is-faster eect can be understood as follows: Due
to uctuations, one of the opinions is able to acquire a slight majority of voters. Therefore,
voters of this opinion change less likely and, hence, the average inertia of this opinion will
be higher than the other. Since inertia reduces emigration, but not immigration, the
majority will become even larger. This development is enforced by higher values of ()
and constitutes a clear direction of the ordering dynamics (as opposed to the standard
voter model, where the magnetization is conserved), which intuitively can lead to a faster
reaching of consensus. However, for high values of (), this development is outweighed
by the high level of average inertia in the complete system, i.e. also within the minority
population of voters, which slows down the overall time scale of the ordering dynamics
(shown in Fig. 5.1).
For both approaches (the reduced model based on only two levels of inertia (Stark et al.,
2008c) and the one with slowly increasing inertia (Stark et al., 2008b)) we observed the
described slower-is-faster phenomenon. While the results for both model specications
are qualitatively very similar, they still dier in some details. Comparing Figs. 5.3, 5.4,
and 5.5 with Fig. 5.8, we recognize for example dierences in the typical shapes of the
curves. Furthermore, the interior global maximum of consensus times for spatially ex-
tended systems, which get apparent in Fig. 5.8(a), we can only observe in the model with
multiple inertia states, but not in the binary setup.
We provided and analyzed the dynamical equations that unveiled the phenomenons origin,
namely the described aging-mechanism that breaks the magnetization conservation. This
is dierent from the standard voter model, where magnetization is always conserved. We
showed that the break of magnetization conservation only holds when the voters build
up a heterogeneity with respect to their inertia to change opinion. Therefore, once the
symmetry between (a) the global frequencies of the two possible states and/or (b) the
proportions of inertial voters is broken, the favored state (opinion) achieves both (i) a
reinforcement of its average inertia and (ii) a fast recruitment of the less inertial state.
Both eects contribute to a faster deviation from the symmetric state. For some parameter
ranges, these mechanisms outweigh the increase in the time to reach consensus generated
by the high inertia of the state that disappeared in equilibrium.
The heterogeneity of inertia values plays an important role for the appearance of the
described eect. However, it is important to note that it is not the heterogeneity in itself,
which triggers the eect. This is particularly underlined by the benchmark investigations
around Fig. 5.2, where we see that heterogeneously distributed inertia alone does not
produce surprising results. As already explained, only the co-evolutionary dependency
between heterogeneous inertia and the opinions, where higher inertia has a positive eect
on the frequency of an opinion and vice versa, leads to the interesting results reported
above.
Due to the general analytical approach taken in our analysis, we emphasize that this phe-
nomenon is not restricted to the voter model, but is expected to appear in any spin system,
whenever the inertia mechanism is present. This conjecture is supported by preliminary
results of another dynamical system, which are presented as an outlook on further research
in the next subsection.
When introducing the implementation of inertia into the voter model, we have put some
analytical evidence on the intuitively expected eect of such an extension increasing
consensus times. Since these results of a small system with N = 3 voters are contradictory
to our ndings for larger systems sizes, we conclude an emergent phenomenon out of the
interactions in a multi agent system.
A more in depth investigation of the model in a 2-dimensional regular lattice would be
valuable. There, one could nd out about the reason for the local maximum in the
consensus times shown in the left panel of Fig. 5.8. In the course of our investigations, we
also obtained some evidence that one would nd the existence of surface tension in the
spatial dynamics of this model (see Fig. 5.10 and compare to DallAsta and Castellano
(2007)). This would be an important change compared to the uctuation driven dynamics
of the standard voter model.
5.6. Outlook: inertia-effects in another nonlinear voter model 79
(a) t = 0 (b) t = 100 (c) t = 200
(d) t = 1000 (e) t = 10, 000 (f) t = 20, 000
(g) t = 30, 000 (h) t = 35, 800 (i) t = 35, 900
Figure 5.10: Exemplary time evolution of the inertial voter model with multiple inertia
states and = 0.1, s = 0.9. Voters are positioned on a 2-dimensional, regular lattice,
where every voter interacts with its 4 nearest neighbors. The system consists of N =
10, 000 voters and the evolution time t is measured in generations, i.e. 10,000 single
updates correspond to 1 generation.
5.6 Outlook: inertia-effects in another nonlinear voter

model
In the following, we provide an outlook on a valuable follow-up investigation of the inertia

mechanism in the voter model. At the same time, the considerations below provide some
more evidence for the generality of the previous results. Let us consider another dynamical
update rule. Here, instead of the linear function in the voter model, the probability
to adopt opinion depends on its frequency following a logistic function (also called
multinomial logit function or Fermi function, see Fig. 5.11).
voter model rule

probability to choose , W ()
0.8 logistic rule, = 4

L
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
frequency of opinion
Figure 5.11: Comparison of the logistic decision function with the one of the linear voter
model.
Hence, the transition probabilities are
1
WiL () WiL (|i , t) = , (5.25)
exp( [1 2 WiV (|i , t)])
where is a parameter of inverse (social) temperature. Higher temperature means more

randomness in the decision of the voters and, in particular, even if no neighbor (all neigh-
bors) choose a certain opinion, the focal voter might decide in favor (against) this opinion.
Again, the frequencies of opinions in the neighborhood of the updating voter are decisive 6 ,
but this time the relationship is an exponential one. It is well-known that the dynamics un-
dergo a phase transition at a critical temperature Tc = 1/c (compare Glauber-Metropolis
dynamics in the Ising model (Landau and Binder, 2005)). For -values resulting in a
lower social temperature > c , one of the opinions will reach a stable majority and
the magnetization of the system (Eq. (2.2)) approaches 1 with increasing -values. If
Because of equality, we used the transition probability of the voter model WiV () to account for the
6
frequency of opinion in the neighborhood of individual i.

the social temperature is high ( < c ), the system reaches or remains in a completely
disordered state of zero magnetization.
Assuming an innite, fully connected network, i.e. an updating agent does not decide
based on a local neighborhood, but takes the global frequencies of opinions into account
(the mean-eld limit), leads to the master equation for the probability of nding an agent
with opinion
X
P = [(| )P ( |)P ] , (5.26)

where, according to Eq. (5.25), the transition rates are
(| ) = exp WiV (| , t) .

Again representing the frequency of the +1 opinion by A and the frequency of the 1
opinion by B = 1 A, we have
(A|B) = exp( A),

(B|A) = exp [ (1 A)] , (5.27)
i.e.
A = exp( A) B exp [ (1 A)] A.
Setting A = 0, the stationary solution fullls
A

ln 1A
= . (5.28)
2A 1
For this mean-eld considerations, Fig. 5.12 visualizes the above mentioned order-disorder
phase transition. The so-called magnetization is simply M = 2 A 1, i.e. the average
opinion in the system. Numerical simulations in nite, fully connected networks agree
with this analytical result (not shown here, but indicated in Fig. 5.13).
Note that implementing a homogeneous, xed inertia into this mean-eld system does
not change the stationary solution. The transition rates here are the ones in Eq. (5.27),
multiplied by (1 ) each. Hence, the dynamics reads
A = (1 ) {exp( A) B exp[ (1 A) A]} ,

0.8
0.6
0.4
"Magnetization" 0.2 T
c
0.2
0.4
0.6
0.8
1
0 0.1 0.2 0.3 0.4 0.5 0.6
"Temperature", 1/
Figure 5.12: Order-disorder phase transition for transition probabilities following a Fermi
function.
1
= 105
= 0.1
0.8 = 0.9
stat. solution
abs(Magnetization)
0.6
0.4
0.2
0
0 0.5 1 1.5
Temperature
Figure 5.13: Numerical simulations of an N = 900, fully connected network.
where the prefactor (1 ) has no inuence on the stationary solution.

However, we performed numerical simulations in nite, fully connected networks, where
we implemented our persistence-time dependent inertia mechanism. Fig. 5.13 includes the
stationary solution shown in Fig. 5.12 and simulation results for 3 dierent -values. The
result for = 105 is basically identical to a system without inertia at all and shows perfect
agreement with the stationary solution (except for not reaching zero magnetization due to
nite size). For larger values of inertia growth rates , we see a pronounced dierence to the
stationary solution: remarkably, the inertia mechanism may lead to considerably higher
(absolute) magnetization in a system of a given temperature and an increased critical
point of the temperature. Interestingly, this represents another form of the positive eect
on ordering dynamics than the one discussed before. Another intriguing parallel is that
we again nd a non-monotonous behavior dependent on . Note that the curve of the
highest value = 0.9 (leading to binary inertia in the system since, like before, we applied
a saturation at s = 0.9) is below the one for an intermediate value = 0.1.
Chapter 6
Social influence in social dilemmas
It is the aim of this Chapter to implement decision behavior solely dependent on social
inuence into the tness dependent environment modeled by evolutionary game theory. In
the main part of this Chapter, we present a simple and straight-forward implementation
of imitation into the evolutionary dynamics of cooperators and defectors (Section 6.4).
Before we do so, we present the motivation and results of two other investigations, which
were supposed to provide more insight into evolutionary dynamics: (i) we were interested
in the eects of assuming a population that is well-mixed (interactions can occur between
any two individuals), but an individual does only interact with a certain sample of the
population, instead of all other individuals (Section 6.1). We investigate this interaction
scheme for the reference case described in Section 3.3, but also for the dynamics of: (ii)
instead of cooperators and defectors, we assume social inuence deciders only, but two
dierent kinds: some of the individuals are imitators (positive social inuence) and the
others are contrarians (negative social inuence). Contrarians are the counterpart of
imitators in that they decide for the opposite of what others do (as discussed in the
literature (De Martino et al., 1999; Galam, 2004; Selten et al., 2004)).
Furthermore, we apply a similar co-evolutionary dynamics as the inertia mechanism for
the voter model into a standard model in evolutionary game theory - the spatial Pris-
oners Dilemma. The approach and the results presented there illustrate the close relation
between the elds in Chapter 5 and Chapter 6.
84
6.1. Homogeneous, dynamic random networks 85
6.1 Evolutionary dynamics in homogeneous, dynamic

random networks
Complementing the results of Section 3.3, we want to consider an interaction scheme that
could be called homogeneous, dynamic random networks. In each update step, a focal
individual interacts with a xed number k of other individuals that are randomly chosen
from the whole network. This means we have a well mixed population, but the number of
interactions is limited (instead of interacting with all other individuals, which is equivalent
to interacting with one representative individual). Although our implementation is
dierent, the concept is similar to the one of accelerating the selection rate compared
to the rate of interaction, as described by Roca et al. (2006). These authors report on
dramatical changes in the dynamics when increasing the selection rate as compared to
the usual mean eld approach. However, here we show that these results are restricted to
their implementation where only a part of the individuals interact at all and others have
the lowest tness, because they did not interact. In our implementation, all individuals
interact, but only with a limited sample of the population. The aim of this approach is
to nd out whether one nds other or more xed points of the dynamics that can lead to
dierent equilibrium states than the ones known from the mean-eld approach.
As a simple example, consider k = 2. In the limit of an innite population, the probability
to interact with two cooperators equals x2 . The probability to interact with two defectors
equals (1x)2 and, nally, the probability to interact with one cooperator and one defector
equals x (1 x). Hence, the expected payo of a cooperator when interacting with two
randomly chosen individuals reads
C,2 = 2 R x2 + 2(R + S) x (1 x) + 2 S (1 x)2 . (6.1)
Simplifying this expression, we nd that
C,2 = 2(x(R S) + S), (6.2)
which is exactly twice the payo of a cooperator in the usual mean eld approach (Eq. (3.6)).
Similarly, D,2 = 2 D and, therefore, the xed points of the replicator equation are not
aected. In general, the expected payo of a cooperator when interacting with k randomly
chosen individuals reads
86 Chapter 6. Social influence in social dilemmas
k
X k ki
C,k = x (1 x)i (i R + (k i) S)
i=0
i
k k
X k ki i
X k ki
= (R S) i x (1 x) + k S x (1 x)i
i=0
i i=0
i
= (R S)k x + k S
= k (x(R S) + S)
= k C , (6.3)
leading to the same result as for the k = 2 example: the xed points of the evolutionary
dynamics remain unchanged compared to the usual mean eld approach. It would be
interesting to combine the implementations of an accelerated selection time scale of Roca
et al. (2006) and the one considered here. However, considering a homogeneous, dynamic,
and random network structure as introduced here bears no new insight into the respective,
underlying dynamics.
6.2 Evolutionary dynamics with imitators and con-

trarians
The idea here is to investigate the evolutionary dynamics between imitators and contrar-
ians. This means we assume that individuals are not capable of consciously deciding for
one or the other alternative (C or D), but they only can either imitate the decisions of
others, or do the opposite.
The questions are: in which scenarios will contrarians dominate the system, thereby lead-
ing to persistent, partial cooperation in the system? Provided imitators dominate the
system, will they nd consensus in cooperation or defection of course dependent on the
payo structure of the underlying game? In fact: what will be the equilibrium fraction of
cooperators in all the games compared to the result depicted in Fig. 3.3?
In order to answer these questions, we replace cooperators and defectors by imitators and
contrarians. For this purpose, we have to write the expected payos of imitators and
contrarians (im , con ), because we assume replicator dynamics with the usual form
i = i (1 i) (im con ), (6.4)
where i is the frequency of imitators (compare to Section 3.3).

6.2. Evolutionary dynamics with imitators and contrarians 87
Let us rst assume the voter model dynamics as decision rule for the individuals, i.e.
imitators cooperate with a probability equal to the level of cooperation in the system
(pC,im (t) = x(t)) and defect with the with a probability equal to the level of defection
(pD,im (t) = 1 x(t)). Contrarians decide in the opposite way, i.e. pC,con (t) = 1 x(t) and
pD,con (t) = x(t). Note that the change in x, i.e. in the level of cooperation in the system,
is only a consequence of the frequency dynamics between imitators and contrarians.
An individual, imitator or contrarian, receives C = x (R S) + S if it cooperates and
D = x (T P ) + P if it defects (in accordance with Eq. (3.6)). Hence, the payos of the
individuals in the system read
im = pC,im C + pD,im D
= x C + (1 x) D ,
con = (1 x) C + x D . (6.5)
Applying these payos to Eq. (6.4), we can compute the xed points of the dynamics
by setting i = 0 and transform the result in order to obtain the equilibrium level of
cooperation in the system x. We have solved this with the technical computing software
Maple, and the results are quite interesting: we get the trivial xed points at i = 0
(leading to x = 1/2) and i = 1 (with any initial x as equilibrium level of cooperation,
according to the voter model dynamics) and another xed point, namely
P S
x= .
R+P ST
The xed points i = 0 is unstable and the xed point i = 1, where stable, leads to
the same equilibrium level of cooperation as in the reference case. The third xed point
is interesting in that we already know it from the dynamics between cooperators and
defectors (see Eq. (3.8) and the text immediately below it). Since this xed point is
independent of i, also its stability analysis does not change and we can conclude that the
outcome of the evolutionary dynamics between imitators and contrarians is identical to
the the reference case depicted in Fig. 3.3. In the regions where defectors or cooperators
dominate the reference model, imitators dominate the system and nd consensus in the
respective decision (full cooperation or full defection). In the coexistence region of the
reference model, also imitators and contrarians coexist with frequencies that lead to an
identical frequency of cooperators compared to the reference case. Finally, the bi-stability
region has turned into a dominance region of imitators, but in terms of the nal frequency
of cooperators, it is still a bi-stable region and the interior, unstable xed points correspond
to those from the reference case.
Although this result is entirely determined by the model denition, we did not expect it
due to the completely dierent decision process assumed in this model. However, so far it
seems like there are no additional dynamical properties to observe in this model setup.
In order to account for a higher level of stochasticity, let us also assume that individuals
decide according to a Fermi function (as we discussed for opinion dynamics in Section 5.6),
i.e. in a well mixed population, the probability to cooperate equals
1
pC (t) = . (6.6)
1 + exp( [1 2x(t)])
To dene our sub-populations, we will assume {b, b}, where imitators apply a of
b > 0 and contrarians have = b. Note that we do not expect qualitative changes by
varying the value of b as long as it is well above the critical temperature known from the
Glauber dynamics (which lies around 2, see Section 5.6). Moreover, the results should be
similar for any decision rule where imitators tend to follow the majority and contrarians
tend to join the minority in the system. Using Eqs. (6.6) and (3.6), the expected payo
of an individual reads
(C D )
= + D , (6.7)
1 + exp{ [1 2x(t)]}
i.e.
(C D )
im = + D
1 + exp{b [1 2x(t)]}
(C D )
con = + D . (6.8)
1 + exp{b [1 2x(t)]}
What we still need is the global number of cooperators x(t) not only to determine the
equilibrium level of cooperation, but also for the expected payos in Eqs. (6.8). The
respective equation reads
i 1i
x(i, t) = + , (6.9)
1 + exp{b [1 2x(i, t)]} 1 + exp{b [1 2x(i, t)]}
which can be solved numerically. Doing so and computing the xed points of this dynamics,
we again nd that the equilibrium fraction of cooperators is identical to the dynamics
between cooperators and defectors.
6.3. Co-evolution in the spatial Prisoners Dilemma 89
Furthermore, the same result holds for the homogeneous, dynamic random networks, where
we assumed that a focal individual interacts with k other individuals that are randomly
chosen from the whole system. All we have to do is applying Eq. (6.3) to Eq. (6.7), which
yields
k (C D )
,k = + k D
1 + exp{ [1 2x(t)]}
= k . (6.10)
In conclusion, in this Section we investigated the evolutionary consequences of assuming

that individuals in a symmetrical 22 game do not have sucient information or capabil-
ities in order to decide for one of the two decision alternatives, but only decide dependent
on the decision-frequencies in their environment. We allowed for two opposing behaviors:
imitation and contrarian behavior. In general, we found that this approach does not lead to
new results regarding the evolution of cooperative behavior in competitive environments.
6.3 Co-evolution of strategies and teaching abilities

in the spatial Prisoners Dilemma
In this Section, we bring together the idea and motivation of the persistence time depen-
dent inertia mechanism in opinion dynamics (see Chapter 5) and the tness dependent
dynamics of evolutionary game theory. In particular, we base our investigations on the
idea of heterogeneous teaching abilities within the player population, as it has been
introduced by Szolnoki and Szabo (2007). In their model of an evolutionary, spatial Pris-
oners Dilemma (i.e. where players repeatedly play the Prisoners Dilemma game only with
their xed set of nearest neighbors and successful players are more likely to spread their
own strategy to a neighboring player, compare to (Frean and Abraham, 2001a; Nowak
and May, 1992; Perc and Szolnoki, 2008; Roca et al., 2008; Schweitzer et al., 2002)), they
assume that some individuals have (xed) teaching abilities that allow them to disperse
their own decision (cooperate or defect) with an increased probability to their interaction
partners. This means that some individuals are more inuential than others and their
behavior has better chances to spread in the population. By doing so, cooperators can
improve their payo by converting defecting interaction partners into cooperative ones
and, contrarily, defectors detoriate their success by teaching neighboring cooperators to
defect. Thereby, the level of cooperation in the system can be enhanced compared to the
evolutionary, spatial Prisoners Dilemma game without teaching abilities.
These teaching abilities can have the same source as the inertia that we investigated
in the opinion dynamics part of this thesis, namely conviction. The only dierence is
that inertia leads to an increased probability to keep the current decision, while teaching
increases the probability to transfer the own current decision to another individual. Both
mechanisms, however, provide an evolutionary advantage to the respective decision. Due
to this equivalence of approaches, we further assume evolving instead of xed teaching
abilities. As for the opinion dynamics investigations, we connect the teaching abilities of
an individual to the persistence time of their current decision (to cooperate or defect).
One could also interpret this as gaining experience in the course of time, so aging, as we
already used this term for the opinion dynamics investigations, is a good example for such
kind of processes 1 .
For the model setup, we consider an evolutionary Prisoners Dilemma game, that is char-
acterized by the following payo matrix
C D
! %
C 1 0
. (6.11)
D b 0
With the relation 1 < b 2, this payo matrix bears the essential properties of a Prisoners
Dilemma while reducing the model complexity to one parameter b (Nowak and May, 1992).
This parameter adjusts the strength of the dilemma: is the temptation to defect large
(b >> 1), it is more dicult to cooperate for the individuals. If b is close to 1, the
dilemma is weak. Each player i on the regular L L square lattice is initially designated
either as a cooperator (si = C) or defector (si = D) with equal probability, and the game
is iterated forward in accordance with the Monte Carlo simulation procedure comprising
the following elementary steps: rst, a randomly selected player i acquires its payo pi
by playing the game with its nearest neighbors. Next, one randomly chosen neighbor,
denoted by j, also acquires its payo pj by playing the game with its four neighbors.
Finally, player i enforces its strategy si on player j with the tness dependent probability
1
W (si sj ) = wi , (6.12)
1 + exp[(pj pi )/K]
1
Note that although the investigations of this approach have been designed in consequence of the
argumentation built within this thesis and with the aim to support the main argument of this thesis
(models of social influence and strategic decision making are only different perspectives to look at social
interactions and combining them leads to new and relevant insight to the intrinsic properties of social
interaction), the major part of the research was performed outside this thesis project. Moreover, the
results crucially rely on the spatial network structure, while the main focus of this thesis is more general.
Hence, we will here only summarize the main findings and discuss their relevance, while pointing the
interested reader to the respective publication for all details (Szolnoki et al., 2009).
where K denotes the amplitude of noise (Szabo and Toke, 1998) or the intensity of se-
lection (Altrock and Traulsen, 2009; Traulsen et al., 2007). Note that this again denotes
a Fermi function as we used it in the previous Section and for opinion dynamics (see
Section 5.6 on Glauber dynamics). wi characterizes the enforcing strength (or teaching
ability (Szolnoki and Szabo, 2007)) of player i. One full Monte Carlo step involves all
players having a chance to pass their strategies to their neighbors once on average. The
teaching ability wi is related to the integer age ei = 0, 1, . . . , emax in accordance with the
function
wi = (ei /emax ) , (6.13)
where emax = 99 denotes the maximal possible age of a player. Thereby, wi is kept within
the unit interval and determines the level of heterogeneity in the ei wi mapping.
Evidently, = 0 corresponds to the classical spatial model, where wi = 1 for all players.
= 1 leads to a model where wi and ei have the same distribution, whereas 2 impose
a power law relation between the teaching ability and age.
Initially, each players age ei is randomly selected from a uniform distribution within the
interval [0, emax ]. In order to separate the inuences of dierent parameters, three scenarios
have been investigated:
I A players age does not evolve in time and determines the static level of articially
introduced heterogeneity.
II The age of all players is increased by 1 every full Monte Carlo step. Furthermore, we
reset ei = 0 for all players i whose age exceeded emax (a newborn follows the dead
player).
III Additionally to the previous setup, we also reset ei = 0 for all players which changed
their decision (to cooperate or to defect). From a biological aspect, the more success-
ful player replaces its neighbor with its own descendant. From a social viewpoint, a
player who has not changed its strategy seems to be more reliable than the one who
has just changed its strategy. Thus, the latter is considered to have less teaching
ability.
Note that setup III represents the implementation of our inertia mechanism. The resulting
dynamics is co-evolutionary as both the level of cooperation and the heterogeneity of
teaching abilities endogenously evolve in time.
0.8
0.6
xs
0.4
0.2
0
1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16
b
Figure 6.1: Promotion of cooperation due to the increasing heterogeneity in the ei wi

mapping via . Stationary fraction of cooperators xs is plotted in dependence on b for
= 0 (solid red line), = 1 (dashed red line) and = 2 (dotted blue line). In all three
cases K = 1.
Results of Monte Carlo simulations presented below were obtained on populations com-
prising 100 100 to 800 800 individuals, where the stationary fraction of cooperators
xs was determined within 105 to 106 full Monte Carlo steps. Moreover, since the co-
evolutionary aging process may yield highly heterogeneous distribution of ei , which may
be additionally amplied during the ei wi mapping, nal results were averaged over up
to 300 independent runs for each set of parameter values in order to take into account the
uctuating output and assure accuracy.
In agreement with previous results (Perc and Szolnoki, 2008), model I yields an increased
promotion of cooperation with increasing heterogeneity in the system, as is illustrated in
Fig. 6.1. There, the results for dierent variants of the Prisoners Dilemma are shown
by varying the temptation parameter b the larger the temptation to defect is, the more
dicult it is to maintain cooperation within the evolutionary Prisoners Dilemma. The
threshold of b, up to which a certain level of cooperation xs can be found in the nal
population, increases with the parameter , i.e. with the level of heterogeneity in teaching
abilities.
This preliminary result is more comprehensively illustrated in Fig. 6.2, where the whole
phase diagram of the model parameters b and K is presented for the values = 0 and
1.4
(a) (b)
1.3
1.2
D
b
1.1
D
C
1.0
C
0.9
0 0.5 1 1.5 2 0 0.5 1 1.5 2
K K
Figure 6.2: Full b K phase diagrams for the Prisoners Dilemma game with quenched
uniform distribution of ei , obtained by setting = 0 [panel (a)] and = 2 [panel (b)] in
the ei wi mapping. Solid green and red lines mark the borders of pure C and D phases,
respectively, whereas the region in-between the lines characterizes a mixed distribution
of strategies on the spatial grid. The dashed, blue line at b = 1 denotes the boarder
between the essential Prisoners Dilemma payo parameterization (above the line) and a
dilemma-free situation (below the line).
= 2. Compared to the homogeneous scenario (panel (a)), the dominance region of

cooperators (below the solid, green line) is shifted towards higher values of b and the
coexistence region (between the green and the red line) enlarged in the heterogeneous
scenario.
Turning to models II and III, Fig. 6.3 shows the eects of the dynamical heterogeneity
(model II, panel (a)) and the co-evolution between heterogeneity and the level of coop-
eration (model III, panel (b)). Most interestingly, we nd opposite results compared to
the case of static heterogeneity (Fig. 6.2, panel (b)): if the heterogeneity is dynamically
changing due to the deterministic aging of individuals, the positive eect for coopera-
tion is there, but less pronounced as for the static case. In contrast, the co-evolutionary
model III clearly enlarges the promotion of cooperation. In the plain model III, smallest
values of the noise parameter K already lead to coexistence in the nal population and
for K > 0.25, cooperators are even able to dominate the population for weak Prisoners
Dilemmas (i.e. for low values of b). Fig. 6.3(b) also shows that the observed eect is even
1.4
(a) (b)
1.3
1.2
b
D
1.1
C
1.0
0.9
0 0.5 1 1.5 2 0 0.5 1 1.5 2
K K
Figure 6.3: Full b K phase diagrams for the Prisoners Dilemma game incorporating
aging as a dynamical process. In both panels, = 2 (to be compared with panel (b) of
Fig. 6.2). (a) results for model II, where aging is dynamical, but follows a deterministic
protocol; (b) results for co-evolutionary model III, where the individual teaching values
wi evolve corresponding to the persistence time of strategies, i.e. individuals who change
their strategy are considered newborn (ei = 0) in the next timestep. Additionally, the
dashed lines indicate the results for a slightly changed scenario, where only 10 % of the
individuals increase their age per time step. To avoid ambiguity, C and D symbols are not
given in panel (b), but the respective regions can be inferred by the line colors according
to panel (a).
more pronounced when the time scales of aging and strategy adaption are separated, i.e.
if the evolution of strategies (or frequency of interactions) is faster than the evolution of
age.
The comparison between models I, II, and III also shows that it is the persistent time
dependent aging mechanism that produces the most remarkable results. This has to be
seen in connection with the results of Chapter 5, where a similar co-evolutionary process
has led to counter-intuitive results within the framework of opinion dynamics. It does
not only show that the aging mechanism developed in Chapter 5 is of even more general
interest as described there, but it is also an indication for the close connection between
the two scientic approaches to social interaction. It denotes a good example for the
usefulness of exchanging and mixing methods of opinion dynamics and game theory in
6.4. The role of imitation in social dilemmas 95
order to obtain general results regarding the dynamics of social interaction.

However, we have to note that the observed results can be explained by understanding
the local dynamics in the spatial system (as done by Szolnoki et al. (2009)). If teaching
occurs between local neighbors, newborn cooperators are guaranteed to have protection
by the protable interaction with (an)other cooperator(s), namely at least the individual
that taught to cooperate. Due to this protection, the newborns themselves have the
chance to gain teaching abilities and to further spread cooperation. On the other hand,
the same argument decreases the potential success of a newborn defector, also supporting
the eect. Therefore, we have to restrict the conclusion of a promotion of cooperation to
systems incorporating spatial structure.
In the following Section, we will investigate another approach to combine ideas of social
inuence models with those of evolutionary game theory, that overcome the aforementioned
restrictions.
6.4 The role of imitation in social dilemmas
6.4.1 Motivation and literature
In our main approach to combine social inuence with (evolutionary) game theory, we
assume a population which consists of three subpopulations: (unconditional) cooperators,
(unconditional) defectors, and imitators. It is important to note (and we will see this
below) that the imitators assumed here are not following the so-called proportional im-
itation rule or any other tness-dependent imitation rule. Our imitators, representing
social inuence in the model, are imitators in the plainest form: they deterministically
and with probability one imitate decisions from their interaction partners, irrespective
of the tness of the imitated individual. This is in contrast to the formerly mentioned
rules, where individuals do only imitate the decisions of successful other individuals. We
denote the frequency of cooperators by x and the frequency of imitators by i. Hence,
the frequency of defectors equals 1 x i. Further, we assume a well-mixed population
where two random individuals meet per update step. Within one update step, the two
individuals play m iterations of the game dened by the payo matrix in Eq. (3.5).
Cooperators apply C and defectors apply D in every iteration of the game. Imitators
take a random choice in the rst iteration, and imitate their interaction partner in the
subsequent iterations. If two imitators meet, they will apply the same decision at latest
from iteration 2 on. Here, we assume that it is the nature of imitators to try to apply the
same decision and they will achieve this goal somehow. Such behavior is consistent with
inequality aversion (Fehr and Schmidt, 1999) and egalitarian motives in humans (Dawes
et al., 2007; Fehr et al., 2008). In consequence, only one of the imitators switches its
decision in case they initially applied dierent decisions. This can also be seen as a
(fast) consensus process taking place in case of dierent initial decisions. In fact, such a
consensus process happens in every update step, where consensus might be found through
imitational decision update, or might not be found due to the interaction between dierent,
fully inertial individuals (a cooperator and a defector). Here, the connection between
the dierent perspectives on decision making taken in Chapters 2 and 3) becomes evident.
In a work on a voter model of the spatial Prisoners Dilemma (Frean and Abraham,
2001a), a very similar setup of imitators is investigated. The similarity between the ap-
proaches can be best illustrated by mentioning the few, but important dierences. In their
model, they do not consider pure cooperators and defectors, but the respective individuals
cooperate (defect) with a probability of 0.8 and apply the opposite decision with probabil-
ity 0.2. Besides this, the authors focus on the eects of spatial structure and, therefore, do
not investigate the general dynamical characteristics and their interesting results. Finally,
they restrict their investigations to the Prisoners Dilemma, whereas we are interested in
all the dierent scenarios modeled by symmetrical 22 games.
Let us mention more connections to previous work on the evolution of cooperation. On the
one hand, our imitators are similar to tit-for-tat (TFT) players that play a nite number
of prisoners dilemma games per interaction with another individual. To illustrate the
mechanism of direct reciprocity, (Taylor and Nowak, 2007) assumed the evolution of a
TFT population together with unconditional defectors. This leads to the payo matrix
TFT D
! %
TFT mR S + (m 1) P
, (6.14)
D T + (m 1) P mP
where m is the number of games played per interaction and T > R > P > S of the
prisoners dilemma. TFT becomes evolutionarily stable against defectors if m R > T +
(m 1) P , i.e. if
T P
m> .
RP
These TFT players correspond to imitators with the additional trait that they initially
choose the cooperative move C. The imitators in our setup do not have this specic trait.
This is consequent since we assume that these individuals cannot consciously decide for
one of the alternatives. Note that our imitators are therefore less advantaged than TFT
players since an encounter of two such individuals may lead to mutual cooperation or
mutual defection equally likely. Furthermore, our mechanism diers from this one in that
it also considers unconditional cooperators. We are interested in the eect of the presence
of simple imitators (herding behavior) on the evolution between cooperators and defectors.
For the TFT strategy, this was done by Imhof et al. (2005). Their work can be summarized
as follows: there are the three subpopulations of cooperators, defectors, and TFT players.
One interaction consists of m rounds of the game and, if m is large enough, domination of
TFT is the only stable xed point. However, they assume a complexity cost for TFT, which
destabilizes the equilibrium and, thereby, leads to domination of defectors. Based on this,
they introduce mutation (with a small probability, reproduction of one individual leads to
ospring of another type) into the dynamics. Basically, this leads to oscillations between
domination of cooperators, defectors, and TFT players, where the dynamics spends most
of the time in the state where TFT dominates. This shows that even when considering
a complexity cost, the reciprocal behavior of TFT can have evolutionary advantages.
However, such results always come with some open problems: TFT is an articial strategy
that always cooperates in the beginning of an interaction and then deterministically repeats
the previous action of the interaction partner. It also needs to be deterministic as otherwise
it only receives the average of all four payo values when interacting with itself.
Taking up this argument, the similar strategy win-stay-lose-shift was suggested as more
robust strategy representing reciprocal behavior (Imhof et al., 2007; Nowak and Sigmund,
1993). This strategy repeats its last move if the last payo was above an aspiration level
(win-stay) and switches to the other move if the payo was lower (lose-shift). In the
Prisoners Dilemma, this usually means that an individual switches after payo S or P
and repeats its move after payo R or T . But this only resolves the problem of correcting
erroneous moves when playing against alike. At the same time, it brings along a new
problem: without further assumptions it is dominated by defectors because it cooperates
every second game iteration. Only if selection is weak and the temptation to defect in
the Prisoners Dilemma is low (b/c > 3 for the standard cooperation model in Eq. 3.10),
win-stay-lose-shift can be favored by selection.
Note that our imitators, on average, receive the same payo with or without erroneous
moves, namely (R + P )/2. Therefore, such considerations are pointless in our model.
This fact, together with the results presented in the remainder of this Section, lead us to
the supposition that our imitators might be a more suitable concept for investigations on
reciprocal behavior in evolutionary game theory. In a sense, we would answer the question:
tit-for-tat or win-stay-lose-shift? (Imhof et al., 2007) by: imitation.
On the other hand, our mechanism can also be compared to the idea of teaching abilities
of some individuals as considered by Szolnoki and Szabo (2007), which we explained and
extended in the previous Section. There, the teaching abilities are exogenously assumed
and one could argue that this is a quite strong assumption. Furthermore, these results
could only be obtained in spatially structured systems, implying a further assumption. The
analogy between our approach and the teaching model by Szolnoki and Szabo is that our
imitation could be framed as consequence of teaching. Then, all individuals in our setting
are teachers, but some are resistant against being taught (unconditional cooperators and
defectors). Imitators are not resistant and learn the decision of the interaction partner.
Therefore, when elaborating on the eect of imitation on evolutionary dynamics below,
we will obtain a natural motivation for a similar mechanism, and more general results at
the same time.
6.4.2 The evolution of cooperation through imitation
In the previous Section, we explained, motivated, and embedded our concept of including
imitators into the evolutionary dynamics between cooperators and defectors. Imitators are
zero condent individuals who copy the decisions of their interaction partners. This might
be due to missing information regarding the payo structure, due to egalitarian considera-
tions (in the symmetrical games considered here, identical decisions lead to equal payos),
or simply due to herding behavior. Therefore, we consider a population consisting of three
species: cooperators, defectors, and imitators. Cooperators and defectors unconditionally
repeat the same decision all the time and with whom ever they interact. Imitators apply
at every interaction the same decision as the interaction partner. If two imitators meet,
they mutually cooperate or defect with equal probability.2 All three species are subject
to tness dependent selection, which is incorporated by the replicator dynamics in an in-
nite, well-mixed population. Hence, there is no noise in the system and we can write the
evolution equations for the frequencies of cooperators (x) and imitators (i) by
x = x (c ), (6.15)
i = i (i ). (6.16)
The frequency of defectors equals 1 x i and evolves as (x + i). The average payo in
the system is obtained by
2
While we motivated especially the specification of imitators by an infinitely repeated game per inter-
action in the introduction, we now do not need to assume repeated games any more. Within the model,
we assume a one-shot game per interaction between cooperators, defectors, and imitators as specified in
the text.
= x c + i i + (1 x i) d .
As usual, the model is described by the specication of the expected payos for an indi-
vidual of each subpopulation. Assuming m iterations of the game per interaction event,
the payo matrix is
C D I

R+S
C mR mS 2
+ (m 1) R
T +P
D mT mP + (m 1) P . (6.17)

2
T +R S+P R+S+T +P
I 2
+ (m 1) R 2
+ (m 1) P 4
+ (m 1) R+P
2
To keep the analysis as simple as possible, we assume that the number of iterations is
large enough such that the payo of the rst iteration is negligible (m ). Hence, we
can leave m out of the equations by considering the average expected payo per played
iteration. The resulting payo matrix
C D I

C R S R
D T P P (6.18)

R+P
I R P 2
leads to the expected payos for individuals
d = x T + (1 x) P, (6.19)
c = (x + i) R + (1 x i) S, (6.20)

i i
i = x+ R+ 1x P, (6.21)
2 2
dependent on their species. The considered evolutionary dynamics leads to the ve distinct
xed points:
(i) x = 0, i = 0
(ii) x = 1, i = 0
(iii) x = 0, i = 1
P S
(iv) x = P +RST
, i=0
(P R) (P S) 2 (P S) (RT )
(v) x = P 2 +R2 +2 S T P (S+T )R (S+T )
, i= P 2 +R2 +2 S T P (S+T )R (S+T )
,
where in all cases the frequency of defectors is d = 1 x i.
S = 0, T = 1 S = 0, T = 1
1 1
0.8 0.8
0.6 0.6
x
x
0.4 0.4
0.2 0.2
0
0
2 2
1 1
1 1
0 0 0 0
P 1 1 R P R
(a) (b)
Figure 6.4: Equilibrium fraction of cooperators x in the two non-trivial xed points.
The x-axis and the y-axis are the same as in the two-dimensional Figs. 3.2 and 3.3, i.e.,
together with our general denition of S = 0 and T = 1, they dene the respective
game. (a) Fraction of cooperators in the non-trivial xed point of the evolution between
cooperators and defectors only [xed point (iv) in the text]. (b) Fraction of cooperators in
the additional, non-trivial xed point through the inclusion of imitators [xed point (v) in
the text]. This one only exists for games with 0 < P < R < 1, i.e. only in the Prisoners
Dilemma. In this Figure, stability of xed points is not illustrated.
The rst three xed points are the trivial ones that result from the non-innovative nature
of the replicator equations. The fourth xed point (iv) we already know from the evolution
between cooperators and defectors, discussed in Section 3.3. Here, if stable, cooperators
and defectors coexist and imitators go extinct. In Fig. 6.4a, we show this xed point
depending on the game in terms of the equilibrium fraction of cooperators x. This xed
point exists in two regions of the game space, namely the two regions of bi-stability and
coexistence explained for the reference case around Fig. 3.3. There, the plane of xed
points in the region P < 0 is stable and leads to coexistence, whereas the plane in the region
R > 1 is unstable and separates the attractor regions of the trivial xed points. However,
the inclusion of imitators into this evolutionary scenario leads to another, non-trivial
xed point (v), where cooperators, imitators, and defectors may coexist (see Fig. 6.4b).
Interestingly, note that this xed point does only exist in the Prisoners Dilemma region
(0 < P < R < 1), which might be dicult to exactly see in Fig. 6.4b.
Of course, the most interesting question is now: which of the xed points are stable and,
hence, what are the stable outcomes of this evolutionary dynamics? In general, we apply
the stability analysis introduced in Section 1.3.2, i.e. we build the Jacobian
! %
x/x x/i
J= (6.22)
i/x i/i
and take the eigenvalues of J at a xed point as indicator for stability of the particular
xed point. We computed the two eigenvalues for each xed point with the help of the
technical computing software Mathematica, and found them to be:
(i) 1 = P, 2 = 0
(ii) 1,2 = 12 (1 R |R 1|)
(iii) 1,2 = 12 |P R|
P (R1)
(iv) 1,2 = P +R1
q 2 R)2 (R1)2
(v) 1,2 = [PP (P(P1)+R (R1)]2
,
where our general assumption S = 0, T = 1 is implemented. As discussed in Section 1.3.1,

the applied linearization method can be problematic for xed points at the boundary of
the state-space. This becomes apparent in the example of xed point (i), where 2 = 0
would imply a set of neighboring xed points, which is not the case here. However, we
can apply a slightly dierent approach for these three points (i-iii): the concept behind
this eigenvalue-approach is to nd the direction of dynamical ows along each dimension
of the system. The dimensions are specied by the eigenvectors that, however, might
only touch the limited state space (see eigenvector 2 in Fig. 6.5 the illustration of the
state space and eigenvectors). From the known direction of dynamical ows along these
two eigenvectors, we can infer the resulting ow-direction at any other point, which is
suciently close to the xed point, by superpositions of the eigenvectors as described in
Section 1.3.2. Since xed points (i-iii) are at the corners of the triangular state space,
alternatively we can nd out the dynamical ows exactly on the respective edges and

Figure 6.5: Illustration of the state space (triangular simplex) for the dynamics on the
relative frequencies of three species: cooperators x, imitators i, and defectors 1 x i.
The dynamical system can only be in a state within the triangle or exactly at its border.
The black circle denotes xed point (i), i.e. the domination of defectors. Dashed lines are
illustrations of eigenvectors, aimed at supporting the verbal explanation in the text.
apply the same argumentation of superposition vectors for any point between the edges.
The edges replace eigenvectors and the direction of ows replace the signs of eigenvalues.
Finding the direction of ows on the edges is straight-forward: one species (or genotype) is
not there so that we can consider a 22 game between the other species. If one species is
ESS against the other (see Section 3.3), the direction of ow close to its domination-corner
(the xed point where the species under consideration is the only one present) is towards
that corner, i.e. along this dimension of the dynamical system the dynamics ows into
that xed point. Hence, the particular domination xed point is stable if the respective
species is bilaterally ESS against both other species. For a better orientation, let us write
the 3 bilateral games:
C D C I D I
! % ! % ! %
C R S C R R D P P
R+P R+P
D T P I R 2
I P 2
With these guidelines, we are ready to investigate the stability of our xed points. Since
for xed point (i) 2 = 0, we use our ESS argumentation. Defectors are ESS against
cooperators if P > S (applying the rst of Maynard Smiths conditions (Maynard Smith,
1982)). This nding corresponds to 1 since S = 0, suggesting that this eigenvector indeed
runs along this edge. Defectors are ESS against imitators if P > (R + P )/2, i.e. if P > R
(the second of Maynard Smiths conditions). Therefore, xed point (i) is a stable xed
point of our dynamics in and only in games with P > S, R. This rst result is already
quite remarkable: one of the fundamental characteristics of social dilemmas is that mutual
cooperation is more protable than mutual defection, i.e. R > P . The conclusion is that
defectors cannot dominate in any of the social dilemmas. This is a dramatic change as
compared to the reference dynamics between cooperators and defectors (Fig. 3.3). There,
domination of defectors is the unique stable xed point in the Prisoners Dilemma and
part of the bistability in Stag Hunt and Pure Coordination I.
1.5
1
1, 2
0.5
0.5
1
1 0.5 0 0.5 1 1.5 2
R
Figure 6.6: Eigenvalues for the Jacobian at xed point (ii), i.e. at the state where coop-
erators dominate. The value of 1 is illustrated by the black line and 2 by the red one.
They only depend on the payo value R, provided that S = 0, T = 1.
Fig. 6.6 illustrates the eigenvalues at xed point (ii), where only cooperators are present in
the population. The immediate conclusion is that dominance of cooperators is not stable
if R < T = 1. This fact is identical to the reference case and also corresponds to the fact
that cooperators are ESS against defectors if R > T . On the other hand, cooperators are
ESS against imitators if R > P . The latter fact is the only dierence to the reference
case and only eects Pure Coordination II. Like in the reference case, cooperators can
dominate in Stag Hunt and Pure Coordination I (both are social dilemmas). Moreover,
we will see that domination of cooperators is the unique stable xed point in these games,
thereby resolving the social dilemmas.
Imitators can never dominate the system, which gets apparent by the eigenvalues of xed
point (iii). As long as R 6= P , it invariably is a saddle point. This is an interesting feature
of our model: we assume the existence of a species that considerably eects the dynamics,
but itself is under no circumstances able to dominate the system.
Regarding xed point (iv), the analysis is simple: since this xed point exists on the
state-space boundary i = 0 and both eigenvalues are identical, we only have to nd out
whether the xed point is stable in the bilateral game between cooperators and defectors.
This means, the stability of this xed point is identical to the reference case. It is stable
in Chicken, Battle of the Sexes, and Leader, and not stable in Stag Hunt and the Pure
Coordination games. Evaluating the eigenvalues (iv) in the regions where this xed point
exists (R < 1, P < 0 and R > 1, P > 0, see Fig. 6.4a) yields, of course, the same result:
they are negative in the rst case and positive in the latter one.
It remains xed point (v) in the interior of the triangular simplex. This one does only
exist in the Prisoners Dilemma, as is shown in Fig. 6.4b. Let us rstly remark that none
of the other xed points exhibited any form of stability in the Prisoners Dilemma a fact
that already indicates stability of this last xed point. Since the eigenvalues are square
roots of a negative number, they are purely imaginary, i.e. they are complex numbers with
a zero real part. As explained in Section 1.3.2, this leads to oscillations in the dynamics
with constant amplitude. For xed point (v) we can conclude marginal stability, but not
attractiveness. In the triangular simplex of the state space, we can visualize the dynamics
as a trajectory of global states beginning from an initial condition. This is done in Fig. 6.7
for several initial conditions.
The trajectories are computed by numerically integrating the dynamical equations in
Eqs. (6.15) and (6.16).3 As expected, constant oscillations can be seen by the closed orbits
around the interior xed point. A trajectory from an initial condition circles around the
xed point and exactly passes by the initial condition. However, this is not true for some
initial conditions where imitators are relatively rare. In such cases (refer to the grey and
red trajectories of Fig. 6.7), the dynamics lead to a common, constant orbit that consid-
erably keeps the system o the boundary i = 0. This fact has important implications for
the dynamics in nite systems, as we will see below.
The dynamics resulting from the evolutionary competition between cooperators, defectors,
and imitators in the Prisoners Dilemma are equivalent to the one described for the rock-
paper-scissors game by Frean and Abraham (2001b) and complemented by Berr et al.
(2009). Therefore, this quite unexpected result can be understood by the survival of
the weakest, stemming from the fact that a species frequency does not depend on its
own invasion rate, but on the invasion rate of the species it invades. Complementing this
analysis, let us discuss the reasons for this phenomenon by calculating two quantities:
3
We used the Runge-Kutta method for numerical integration.
C I
(a)
D D
C I C I
(b) (c)
Figure 6.7: Oscillations of species frequencies in three Prisoners Dilemma games: (a)
R = 0.8, P = 0.4, (b) R = 0.95, P = 0.05, (c) R = 0.65, P = 0.6. The corners of a
simplex indicate dominance of cooperators (C), defectors (C), and imitators (I), respec-
tively. Circles denote xed points of the dynamics a black lling denotes stability of the
xed point. Shown are dynamical trajectories of dierent initial conditions. Generally,
the trajectories form constant orbits around the stable xed point that include the initial
condition. Exceptional cases (red and grey trajectories) are discussed in the text.
Approximately, we can quantify this evolutionary advantage of imitators by computing

the dynamical forces at the boundaries of the simplex. On one hand, we compute the
invasion speed along the boundary. For example: at the boundary x = 0, imitators
dominate defectors and hence invade any subpopulation of defectors. The speed of this
invasion is proportional to the payo dierence i , where = i i +(1i) d . Therefore,
it diers for dierent states of the system i. On the other hand, we compute the invasion
potential of the (almost) non-existing species in a dimorph system consisting of the other
two species. For example, the invasion potential of imitators is proportional to the payo
0.2 0.4
cooperators at x + i = 1 cooperators at x = 0
defectors at i = 0 defectors at x + i = 1
imitators at x = 0 imitators at i = 0
0.15 0.2
Invasion potential
Invasion speed
0.1 0
0.05 0.2
0 0.4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x, i, 1x x, i, 1x
(a) (b)
Figure 6.8: Illustrations of dynamical forces at the simplex boundary for a Prisoners
Dilemma with R = 0.8, P = 0.4. (a): positive dierence between the expected payo
of the bilaterally dominant species and the average payo in the system the quantity
that reects the dynamical speed of invasion. (b): payo dierence between the (potential)
expected payo of the non-existent species and the average payo in the system reecting
the invasion potential of the non-existent species. In both panels, the abscissa denotes the
frequency of the pairwise dominant (invading) species.
dierence i , where = x c + (1 x) d . These quantities are illustrated in Fig. 6.8.

First, notice that cooperators invade imitators at a low rate when the frequency of imitators
is low (solid line in the left panel for x > 0.8). In this region of x, the invasion potential of
defectors turns positive and increases further for increasing x (dash-dot line in the right
panel). Hence, the very small frequency of defectors increases relatively fast as x increases
very slowly. In result, defectors are able to start invading the population considerably
before x = 1. Second, the invasion potential of imitators at the boundary i = 0 is always
close to zero (in the payo example chosen for Fig. 6.8, it is even strictly non-negative),
i.e. the part of the trajectory that mainly reects the invasion of defectors does not
approach the i = 0 boundary further, but is rather repelled from it. To summarize: (i)
the invasion of cooperators (that diminishes imitators) is interrupted by the invasion of
defectors considerably before imitators are close to extinction, (ii) in any population where
imitators are rare, the remaining imitators receive enough payo to at least approximately
keep their share. These two processes prevent imitators from being close to extinction.
We can compare this with the same considerations for the other two invasion processes:
cooperators get closer to extinction because the invasion speed of defectors (dash-dot in
the left panel of Fig. 6.8) is always higher than that of cooperators and the invasion po-
tential of imitators (solid line in the right panel) is relatively low for high values of 1 x.
Furthermore, the invasion potential of cooperators is negative until i 0.6, leading to
a further decrease of a small rest frequency of cooperators. Defectors also get closer to
extinction because the invasion potential of cooperators is comparable to that of defectors
for high values of i and x, respectively, but the invasion speed of imitators is higher than
that of cooperators in that region. Like the cooperators, defectors also have negative in-
vasion potential for lower values of x, which brings them even closer to extinction.
Let us summarize the stability analysis of our dynamics for any symmetrical 22 game.
Since, compared to the reference case depicted in Fig. 3.3, dominance of defectors lost
stability in some games, we do not have a region with bistability anymore. This makes the
color encoded representation (Fig. 6.9) more convenient, because we only have one stable
xed point per payo ranking. However, we have to keep in mind that the xed point for
the Prisoners Dilemma (the triangle with 0 < P < R < 1) is not attractive, but we have
oscillating coexistence of all three species there.
Most remarkably, we can summarize the eect of including an imitating species into the
evolutionary dynamics by concluding an immense increase in the level of cooperation in
social dilemmas. Whereas the Chicken game is unaected by this change, the dilemma
is completely resolved in Stag Hunt and Pure Coordination I. These are the two social
dilemmas where there is no temptation to defect, but there are two Nash equilibria and
the inecient one (mutual defection) is risk dominant against the ecient equilibrium
(mutual cooperation). For this games, the dynamics possess one unique stable xed point
in the domination of cooperators, i.e. irrespective of the initial condition, evolutionary
dynamics will exclusively favor cooperators. Reversely, for Pure Coordination II (not a
social dilemma), defectors become the unique stable xed point, thereby also increasing
the average payo in the population compared to the reference case.
In the Prisoners Dilemma, where there is both a temptation to defect and the risk dom-
inance of the inecient equilibrium, we nd persistent, oscillatory coexistence between
all three species. Therefore, a considerable level of cooperation is preserved both by the
survival of cooperators, but also by the behavior of imitators in the system (compare the
respective part of Fig. 6.9(d)).
Departing from the analytical results for innite populations, let us note some remarkable
results for nite populations. For the numerical simulations, we initialize a network with N
nodes and assign an individual of a randomly chosen species to each node of the network.
For some simulations, the probabilities for each species to be chosen are equal (1/3), but we
will also investigate diering initial conditions. Depending on the respective population
structure, individuals are connected to other individuals with whom they are able to
interact let us call them neighbors. An individual receives payo (tness) from an
S = 0, T = 1 S = 0, T = 1
2 2
1.5 1.5
1 1
P
P
0.5 0.5
0 0
0.5 0.5
1 1
1 0.5 0 0.5 1 1.5 2 1 0.5 0 0.5 1 1.5 2
R R
(a) (unconditional) cooperators (b) imitators
S = 0, T = 1 S = 0, T = 1
2 2 1
0.9
1.5 1.5
0.8
0.7
1 1
0.6
P
0.5 0.5
P
0.5
0.4
0 0
0.3
0.2
0.5 0.5
0.1
1 1 0
1 0.5 0 0.5 1 1.5 2 1 0.5 0 0.5 1 1.5 2
R R
(c) (unconditional) defectors (d) cooperative decisions
Figure 6.9: Color encoded equilibrium frequencies in the stable xed points: (a) coopera-
tors, (b) imitators, (c) defectors, and (d) individuals that apply the cooperative decision
at a time, i.e. cooperators plus imitators that meet a cooperator plus half of the imitators
that meet another imitator (x + x i + i2 /2). Red color indicates a frequency of 1, blue a
frequency of 0. The colorbar of (d) translates colors into frequencies for all panels (a-d).
For the Prisoners Dilemma (0 < P < R < 1), the displayed xed point is not attractive,
but produces stable oscillations around it.
interaction with a randomly chosen neighbor according to the payo matrix in Eq. (6.18).4
4
If two imitators meet, both receive either R or P , each with probability 1/2. The results essentially
remain unchanged when assuming that interactions occur with every neighbor of the focal individual
An evolutionary update step consists of N selection steps that proceed as follows: a random
node of the network is chosen for selection. The resident individual a can remain on this
node or it is replaced by a randomly chosen neighbor b. The probability for replacement
is proportional to the relative tness dierence between a and b:
b a
P (b|a) = max(0, ), (6.23)
max min
where the tness dierence is normalized by the maximal possible dierence of tness.
This selection rule is known as replicator rule, because it is the nite size equivalent of
the replicator equation.
At rst, let us remain in a well mixed population, i.e. we only switch to a nite number of
individuals, but still every individual can interact with all other individuals (unstructured
or fully connected network). In general, the results depicted in Fig. 6.10 conrm the
analytical results discussed above. The only dierence regards the Prisoners Dilemma.
There, the niteness of the system penalizes defectors even more than in innite systems
defectors almost always die out. The reason lies in the fact that the amplitude of
oscillations grows over time and the evolution of a single run always ends in domination
of one species.5 This happens by one species eventually dying out, leading to a bilateral,
evolutionary game between the other two species (see matrices in Eq. (6.4.2)). Depending
on which species died out rst, one of the other two will dominate the system: imitators if
cooperators vanish rst, cooperators if defectors vanish rst, and defectors in case imitators
die out rst. In agreement with the above discussion of the shape of oscillations in Fig. 6.7,
the latter scenario is not possible in most cases imitators are of the species that is best
protected against extinction in the dynamics between all three species.
In Fig. 6.11, we show the trajectories of single simulation runs, corroborating this con-
siderations. Each simplex displays one simulation with equal initial conditions one in
which imitators dominate in the end and one in which only cooperators survived. For this
particular Prisoners Dilemma (R = 0.8, P = 0.4), 68.5% of equally initialized simulation
runs ended up in dominance of cooperators and 31.5% in dominance of imitators. Defec-
tors always died out.6 The left simplex additionally displays two extreme initial conditions
in which imitators are very rare. Even in these cases, defectors do not survive. Whereas
imitators approximately keep their share initially, cooperators constantly diminish and
eventually die out rst. Therefore, testing initial conditions with x = i = 0.02, i.e. where
(instead of one interaction with a random neighbor).
5
The shown simulation results consider that individuals receive payoff from a random interaction with
one other individual. However, the results basically remain unchanged when considering that an individual
interacts with all other individuals, which avoids stochasticity in the payoff calculation.
6
Percentages from a sample of 200 simulation runs.
S = 0, T = 1 S = 0, T = 1
2 2
1.5
1 1
P
0.5
P
0 0
0.5
1 1
1 0 1 2 1 0 1 2
R R
(a) cooperators (b) imitators
S = 0, T = 1
2
1
P
1
1 0 1 2
R
(c) defectors
Figure 6.10: Equilibrium fraction of cooperators x (up left), imitators i (up right), and
defectors, respectively. Values are averaged from 200 sample runs. The system size is
N = 900 individuals. In the Prisoners Dilemma (0 < P < R < 1), defectors almost never
survive.
96% of the initial population are defectors, all simulation runs (out of 100 sample runs)
ended in dominance of imitators.
As stated above, the dynamics corresponds to the one investigated by Frean and Abraham
(2001b). Therefore, these results are also in agreement with the ndings there, where it
was shown that the niteness of the system leads to overshooting and, eventually, to
the extinction of two species.
However, coexistence of species is preserved in structured populations. In contrast to fully
connected networks, spatially structured and even random networks lead to shrinking os-
6.5. Summary 111
D D
C I C I
(a) (b)
Figure 6.11: Growing oscillations in single simulation runs that lead to extinction of 2
species. The networks of individuals are fully connected (well mixed) with system sizes
of N = 104 . The game is a Prisoners Dilemma with R = 0.8, P = 0.4 (S = 0, T = 1).
Initial conditions are marked by crosses: besides the two equal initializations, the other
two in the left panel are x = i = 0.02 and x = 0.39, i = 0.01, respectively.
cillations and relaxed, yet uctuating, frequencies of species. Fig. 6.12 shows an exemplary
time evolution of frequencies in a 2-dimensional, regular lattice. The inset amplies the
rst 500 timesteps, showing the decrease of amplitude in the oscillations. After this short
period, the average frequencies relax to a x value where they only uctuate around.
In Fig. 6.13, we compare representative trajectories in random networks and 2-dimensional,
regular lattices.
6.5 Summary
In this Chapter, we have investigated the eects of assuming decision behavior that is
exclusively based on social inuence in the dynamics of evolutionary games. In the rst two
Sections, we presented two research approaches with which we started our investigations.
In Section 6.1, we modied the network of social interaction: assuming that interaction
can possibly occur between all individuals in the system, at each interaction one individual
is inuenced by a xed number of randomly chosen individuals. Due to the randomness
in choosing interaction partners, this procedure controls the level of stochasticity in the
decisions and, related to the ndings of Roca et al. (2006), one could expect that the
evolutionary dynamics might change. However, in the example of replicator dynamics
0.8
0.5
0.7
0.4
0.3
0.6
0.2
frequencies
0.5 0 100 200 300 400 500
0.4
0.3
0.2
0.1
0 1 2 3 4 5
time 4
x 10
Figure 6.12: Representative time evolution of initially equal frequencies of coopera-
tors (red), imitators (magenta), and defectors (blue) in a numerical simulation of a 2-
dimensional, regular lattice with interactions between 8 nearest neighbors. In contrast
to fully connected networks, oscillations do not grow. In fact, they shrink and lead to a
uctuating, but qualitatively stable coexistence between cooperators (mostly in majority),
imitators, and defectors (always in minority after a short, initial stage that is amplied in
the inset of the gure). The system size is N = 104 individuals.
between cooperators and defectors in all symmetrical 22 games, we showed that the
dynamics are independent of this form of stochasticity and, therefore, identical to assuming
that every individual is inuenced by all other individuals (the usual mean-eld approach).
In conclusion, we did not use this kind of interaction structure for any other investigation.
In a second approach (Section 6.2), we were interested in the dynamics between individuals
that react to social inuence only, but in opposing ways: imitators and contrarians. We
modeled both behaviors by means of the linear voter model and a Fermi function, which
makes the system comparable to the Glauber dynamics. An individual was characterized
by its behavior (being an imitator or a contrarian) and the resulting decision (to cooperate
or to defect). The latter was only the consequence of its behavior and the global frequencies
of decisions. Interestingly, we found that, in all the games, the frequency of decisions to
cooperate in equilibrium is identical to the reference case, i.e. identical to the fraction of
cooperators in the xed points of the replicator dynamics with cooperators and defectors.
Although remarkable and worth being recorded, such a result implies no new insight into
6.5. Summary 113
D D
C I C I
(a) random, k = 4 (b) random, k = 8
D D
C I C I
(c) 2-d, k = 4 (d) 2-d, k = 8
Figure 6.13: Evolutionary trajectories in homogeneous, random networks and 2-

dimensional, regular lattices. k denotes the number of links every individual has in the net-
work. Every simplex shows one representative simulation run with equal initial conditions.
Panel (a) additionally displays a trajectory for the initial condition x = 0.9, i = 0.02.
the scientic questions addressed in Section 3.3.2: which are the evolutionary advantages
of cooperation that cannot straight-forwardly be explained by evolutionary principles?
Therefore, the focus of this Chapter lies on the results presented in the previous two
Sections. In Section 6.3, we reported and discussed the results of a model, where the
persistence time dependent aging mechanism of Chapter 5 has been implemented into
the tness dependent teaching model of Szolnoki and Szabo (2007). As for the voter
model considerations, individuals are heterogeneous in their decision behavior and this
heterogeneity is endogenously co-evolving in the system. The underlying model here was
an evolutionary, spatial Prisoners Dilemma and the individual heterogeneity resulted from
an ability to transfer the own decision to an interaction partner (inside-out inuence),
instead of an inertia to copy the decision of the interaction partner (outside-in inuence).
However, the motivation of both specications is still the same: individuals develop a kind
of conviction over time and this conviction tends to increase the abundance of the current
decision in the system. We found that this mechanism leads to an additional positive eect
for the evolution of cooperation as compared to the original teaching model. It could be
shown that this eect prots from the heterogeneity in general, but the important factor
is the specic aging mechanism, which already led to remarkable results in the social
inuence models of Chapter 5.
Particular attention should be paid to Section 6.4, where we analyzed another approach
in which the assumptions are intriguingly simple: in the classical evolutionary dynamics
between cooperators and defectors (the reference case of our studies), what is the eect
of the mere presence of imitators, i.e. individuals that are zero-condent and completely
susceptible to positive social inuence? The results of this approach are striking:
In three of the four social dilemmas, cooperation is considerably enhanced. Whereas

the Chicken (or Snowdrift) game is unaected, this general statement holds for the
prominent Prisoners Dilemma game, Stag Hunt, and Pure Coordination I.
In the games Stag Hunt and Pure Coordination, cooperators invariantly are the
winners of the evolutionary race the only equilibrium state is characterized by the
dominance of cooperators. The temporary presence of imitators in the dynamics
paved the way for cooperation, thereby resolving the dilemma.
The Prisoners Dilemma:
In innite, well mixed populations, the dynamical characteristics assume the

one of the rock-paper-scissors game described by Frean and Abraham (2001b)
and Berr et al. (2009): cooperators, defectors, and imitators coexist persis-
tently. The frequencies in the population oscillate around the marginally stable
xed point in the interior of the state-space simplex. During these oscillations,
defectors and cooperators get considerably closer to the point of their extinction
than imitators.
Modifying the replicator dynamics in well mixed populations to nite system
sizes, these oscillations grow in the course of the dynamics, which lead to the
eventual extinction of two species, i.e. a nal state with only one species left
in the system. Most interestingly, cooperators have the highest probability to
dominate the population, whereas dominance of defectors is very unlikely.
In contrast, structured populations lead to damped oscillations, where the fre-
quencies relax at a point close to the interior xed point. Here, all three sub-
populations coexist persistently with a xed frequency (despite of some uctu-
ations). Only for extreme initial conditions, nite size uctuations can lead to
the dominance of one species.
6.5. Summary 115
Here, we assumed the existence of imitational behavior in evolutionary dynamics. In the

implementation of imitation, we avoid further assumptions: an imitator defects with a de-
fector, cooperates with a cooperator, and nds consensus in either cooperation or defection
with another imitator. It is important to note that we did not assume a given existence of
such imitators (in order to support cooperators against defectors), but imitators were sub-
ject to the same evolutionary rules as cooperators and defectors. Even though imitators
were only able to dominate under specic circumstances (well mixed, but nite population
in the Prisoners Dilemma), in many situations they survived long enough in order to aect
the dynamics considerably. More specically, they turned the tide in favor of cooperation
in three social dilemmas. With these ndings, we conclude a natural explanation for the
evolution of cooperation: the existence of pure (tness-independent) imitation behavior.
Chapter 7
Dilemmas of partial cooperation
The main focus of this Chapter lies on the problem of partial cooperation that deals
with situations where seemingly unselsh behavior is required by some, but not all in-
volved agents. Using a simple class of game theoretic models symmetrical 22 games
the subclass of partial cooperation dilemmas is dened and accurately related to the
broad literature on social dilemmas. Moreover, the 3 games belonging to this subclass are
shown to be instances of the standard cooperation model (e.g. in evolutionary biology)
when considering discounting eects of cooperation (according to Hauert et al. (2006),
see also Gardner et al. (2007); Queller (1985)). Consequently, according to Taylor and
Nowak (2007), we apply the evolutionary mechanisms of kin selection, group selection,
and network reciprocity in order to derive the respective conditions under which partial
cooperation can be a stable outcome of evolutionary dynamics in these scenarios.
7.1 Introduction
The emergence and maintenance of cooperative behavior in competitive environments is a

withstanding question in biology, economics, and social sciences, but it also attracts much
attention from mathematicians and physicists. Game theory, founded by Von Neumann
and Morgenstern (1944), has proven to be a powerful tool for describing and investigating
such real-life conicts. Certainly, one of the most important solution concepts of such con-
icts (represented in games) is that of the Nash equilibrium (Nash, 1950) where no player
has an incentive to unilaterally deviate from this state. If there is such an equilibrium solu-
tion that is not Pareto ecient, i.e. another solution is better for at least one player of the
game and not worse for any other player, one speaks of a social dilemma situation (Dawes,
1980; Hauert et al., 2006; Macy and Flache, 2002). If more than one equilibrium exists,
116
7.1. Introduction 117
the question is if the players are able to (anti-) coordinate 1 their actions in order to achieve
one of the equilibria and which equilibrium will be selected (Harsanyi and Selten, 1988;
Huyck et al., 1990; Samuelson, 1997). Maynard Smith (1982) extended the framework
by considering evolutionary scenarios and provided the concept of evolutionarily stable
strategies that is closely related to Nashs equilibrium. The eld of evolutionary game
theory (see e.g. Friedman (1991); Hofbauer and Sigmund (1998); Nowak and May (1992);
Szabo and Fath (2007); Taylor and Jonker (1978)) has since been applied to innumerous
investigations regarding the evolution of cooperation in biology (Maynard Smith, 1982;
Nowak, 2006a; Nowak and Sigmund, 2004), social sciences (Fehr and Fischbacher, 2003;
Henrich et al., 2003), and economics (Gintis, 2005; Kreps et al., 2001). The advantage
of evolutionary considerations is that one can relax the requirements regarding players
rationality that often is required in classical game theory. Note that in this Chapter
we successively apply both classical and evolutionary game theory in order to show the
existence and implications of partial cooperation dilemmas.
Although game theory provides an extensive framework for studying a variety of interde-
pendent decision situations, the simplest class of games is mostly applied to investigations,
namely games with only 2 players and 2 strategies each, where both players have identical
roles (symmetrical 2 2 games). In 1967, Rapoport concluded four archetypes of such
games (Rapoport, 1967) whereas the others are evaluated to be not of theoretical interest
a statement that, in this generality, will be subject to critical evaluation in this article.
Among these four games, the Prisoners Dilemma attracted most attention in various sci-
entic elds since it incorporates the conict between individually rational decisions and
collectively desired outcomes. Players have to decide whether to act cooperatively (C) or
to defect (D). If they both cooperate, each receives a payo R (reward) that is better
than what they would receive if both did not cooperate, P (punishment, R > P ). If
only one player defects, he receives T (temptation), whereas the other player receives S
(sucker). The dilemma here is that T > R and P > S, i.e. whatever decision the other
one took, a player is better o if he did not cooperate. Therefore, the only equilibrium
strategy is to defect, yielding the Pareto-decient outcome P for both. This can be best
1
Although both solutions require a form of coordinated acting of players, game theoretic literature
distinguishes between equilibria with equivalent actions of players (coordination games) and equilibria
where players take opposing actions (anti-coordination games). In some works (e.g. Browning and
Colman (2004); Neill (2003)), both situations are comprised in the term coordination and especially
coordinated turn-taking is commonly used to refer to alternating anti-coordination in repeated games.
118 Chapter 7. Dilemmas of partial cooperation
illustrated by the left of the matrices
C D C D
! % ! %
C R S C b c c (7.1)
D T P D b 0
where, due to symmetry, payos are given only for the row-player. A specic conguration
of the Prisoners Dilemma is often used as standard cooperation model in evolutionary bi-
ology (right matrix in Eq. (7.1)). Compared are 2 species or genetically encoded behaviors
that are subject to evolutionary selection. A behavior is called cooperative if individuals
endow a reproductive tness benet b (the evolutionary equivalent of payo) to other
individuals at a certain tness cost c to themselves (b > c > 0). The non-cooperative
counter-part is called defective as individuals will receive b from cooperators, but do not
act cooperatively. This parameterization (with or without the Prisoners Dilemma con-
dition b > c > 0) is special as it incorporates the equal-gains-from-switching property
(R T = S P , (Nowak and Sigmund, 1990)), i.e. the individual cost of cooperation
is independent of the other individuals behavior. Likewise, the individual benet of (re-
ceived) cooperation is constant, i.e. an individual, independent of its own behavior, gets
b more payo from a cooperator than from a defector. Therefore, a cooperative act can
suciently be dened as the action that provides a tness benet b > 0 to another indi-
vidual. Evolution of such behavior can easily be explained if cooperation induces a direct
tness benet also to the cooperator, i.e. if c < 0. If this is not the case, there must be
an indirect tness benet for the cooperator and it is a challenging research task to nd
the relevant mechanisms that provide indirect tness benets.
In the present Chapter, we consider a more general universe of situations by applying the
left parameterization of Eq. (7.1), i.e. games with equal-gains-from-switching are only a
subgroup of all considered situations. This implies that, in most cases, the tness eect
caused by a behavior is not static, but depends on the frequencies of behaviors in its
surrounding (see e.g. Nowak and Sigmund (2004)). In consequence, a general denition of
the cooperative act is more dicult, because one and the same act might have opposing
tness consequences. Mainly for illustrative reasons, the choice in this work (which will
be further discussed below) is to impose T > S. However, it is important to note that
this is a naming convention for the binary choices in the context of this Chapter and not
meant to semantically redene the notion of cooperation.
We use the dierent symmetrical 22 games as basis for repeated and evolutionary games.
In our argumentation on a specic subclass of games, we show the existence of a cooper-
ation dilemma both in classical game theory (with a xed set of players) as well as in
7.1. Introduction 119
evolutionary game theory (population dynamics). The specic kind of dilemma discussed
here has to be distinguished from the one of the Prisoners Dilemma (and other well known
social dilemmas). In particular, we focus on scenarios where partial cooperation plays an
important role. Partial cooperation means that, at a given time, only a part of the in-
dividuals apply the cooperative action, whereas the rest of the individuals do not, i.e.
the population of players partially consists of cooperators, and partially of defectors (1
cooperator and 1 defector in the two-person games).
Partial cooperation is, of course, relevant for games where the Nash equilibria are (C, D)
and (D, C). However, in this article, we will focus on whether max(2R, 2P ) > S +
T , or if this inequality is reversed. Verifying this inequality yields the system-optimal
outcome in the sense of Schellings collective total (Schelling, 1978), i.e. the solution of
the game that yields the highest possible overall payo. In a wide range of the literature,
where game theory is applied to research on cooperation, the importance of this system
optimum is neglected and mostly the cases where S + T is system optimal excluded from
the investigations. In our view, thereby, a specic subclass of symmetrical 22 games
received too little scientic attention, although these games are shown to exhibit interesting
strategic conicts in repeated and evolutionary games.
Within this Chapter, we will explicitly derive this subclass of games and introduce it as
partial cooperation dilemmas (PCD). The three members of this class are the games
Route Choice, Deadlock, and Prisoners Dilemma, but all three games exclusively
with the specication S + T > max(2R, 2P ). In repeated setups of these games, a suit-
able, Pareto-ecient solution is turn-taking (Browning and Colman, 2004; Duncan, 1972;
Helbing et al., 2005b; Neill, 2003; Stark et al., 2008a; Tanimoto and Sagara, 2007), i.e.
an anti-coordination of the players (where one player takes the opposite decision of the
other) and a permanent ipping between decision alternatives of both players. Despite its
eciency, this alternating behavior is not an equilibrium state in nitely repeated games,
i.e. players are permanently tempted to leave this solution. Hence, we nd a cooperation
dilemma in repeated games. In evolutionary terms, the most successful population of indi-
viduals would consist of cooperators (that provide help to others) and defectors (that only
receive help), but such a coexistence is not a xed point of the evolutionary dynamics.
This constitutes the evolutionary dilemma of cooperation.
In order to clarify the connection to the standard cooperation model (right matrix in
Eq. (7.1)), we show that all 3 partial cooperation dilemmas can be derived out of this
payo matrix by considering discounting eects of cooperation (Hauert et al., 2006). Fur-
thermore, we investigate known evolutionary concepts (Nowak, 2006b; Taylor and Nowak,
2007) with respect to their ability to sustain partial cooperation. Analytical conditions
under which partial cooperation can be a stable outcome of evolutionary dynamics are
presented.
Note that we are not the rst to deal with the particular games. However, the games
Deadlock and Route Choice are, with rare exceptions (Helbing et al., 2005b; Kaplan and
Rue, 2007; Stark et al., 2008a), almost completely neglected by literature so far. One
reason might be the choice of payo values and the respective conclusion of Rapoport that
was, for example, applied to investigations regarding the evolution of turn-taking behavior
only in the archetype games (Browning and Colman, 2004), but not to one of the partial
cooperation dilemmas. Most surprisingly, even the third member of this class, the promi-
nent Prisoners Dilemma with the specication S + T > 2R, is, despite some exceptions
(e.g. Kreps et al. (2001); Neill (2003); Schuler (1986)), often explicitly excluded from the
scientic investigations, although it is recognized as equivalent cooperation problem (see
also May (1987)). We conjecture that partial cooperation dilemmas are able to serve as
distinct and relevant models in the dierent scientic elds applying (evolutionary) game
theory.
The Chapter is organized as follows: in the following Section, we shortly repeat the intro-
duction to the framework of symmetrical 22 games and all its ordinally distinct instances
(as more extensively done in Section 3). These games serve us as basis for the argumen-
tation in this work which we start by considering repeated games in Section 7.3. First,
we illuminate a specic form of partial cooperation in repeated games: turn-taking.
Then, we discuss what constitutes a social dilemma and name four games as instances of
social dilemmas (largely in agreement with previous literature). Having dened all needed
terminology, we nally name three games that exhibit a partial cooperation dilemma
when they are repeated at least once. This (so far not yet concisely specied) kind of
dilemmas can be resolved by partial instead of full cooperation, which clearly separates
them from the well-known and often discussed class of social dilemmas. In this context,
partial cooperation can be realized by turn-taking or by applying an intermediate mixed
strategy.
Turning to evolutionary game theory, Section 7.4 starts by showing how partial cooperation
dilemmas are related to the standard cooperation model in evolutionary biology: they
have in common that the presence of cooperators can improve the tness of a species
compared to the case of prevailing defectors (constituting a social dilemma). The dierence
lies in the fact that, above a certain threshold, the benet induced by an additional
cooperator decreases with an increasing number of cooperators in the system (discounting
eects of cooperation (Hauert et al., 2006)). Considering an evolutionary scenario, a
dilemma even arises in one-shot games. A population that preserves an intermediate level
7.2. Symmetrical 22 games 121
of cooperation, e.g. through coexistence between cooperators and defectors, would be

optimal. However, how can cooperators survive at all when their reproductive tness is
below defectors? In order to shed some light on this question, in Section 7.5 we analytically
derive the conditions under which coexistence can be maintained by the three evolutionary
mechanisms of kin selection, group selection, and network reciprocity (Nowak, 2006b;
Taylor and Nowak, 2007).
Finally, Section 7.6 summarizes and discusses the role of partial cooperation dilemmas
and their relevance. We argue that such situations are likely to occur in reality, but many
solutions to cooperation problems are derived from classical social dilemmas and do not
work for partial cooperation dilemmas. Whereas similar issues have been addressed by a
number of publications for specic instances of such games, we here provide a holistic con-
ceptualization of partial cooperation dilemmas. Although we restrict our argumentation
to symmetrical 22 games, it can be applied to other classes of games as well.
7.2 Symmetrical 22 games

Let us briey repeat the main facts about the framework of symmetrical 22 games as
introduced in Section 3. We have two players with two strategies each and, hence, 4
distinct outcomes from one-shot game. They are represented by the payo matrix
C D
! %
C R S
, (7.2)
D T P
where the strategies are named cooperation (C) and defection (D), respectively. In order to
apply these names to all dierent game situations, we postulated the following denition:
Let us only consider an encounter of different strategies leading to the payoffs S and T
(partial cooperation). In such a situation, the strategy which yields the lower payoff is
regarded the cooperative strategy (C), and the other one the defective strategy (D).
We assume that the absolute payo values are not decisive for the strategic situation, but
only the ranking of them (we will qualify this point later on). Since we dened T > S,
which eliminates equivalent rankings, one can discern 12 ordinally distinct games. Fig. 7.1
conveniently visualizes the phase space of symmetrical 22 games in a coordinate system
and includes exemplary payo matrices. This gure extends the basic representation of
Fig. 3.2 by focusing on system-optimal solutions and the respective dierences between
variants of the same (ordinally distinct) game. The whole relevance of the included ex-
tensions will become clear below, where Fig. 7.1 will support the discussion on partial
cooperation dilemmas.
Figure 7.1: Classication of symmetrical 22 games according to payo ranking and

system-optimal solutions. Two-dimensionality is achieved by xing T > S and classifying
ordinal dierences only. Parcels separated by solid lines denote dierent rankings of the
payo values. The dashed lines divide the whole space in 2 regions according to whether
partial cooperation is system optimal (dark-grey background) or not. Social dilemma
games have a red background color. For each area a respective payo matrix (in form of
the left matrix in Eq. (7.1), with T > S) is given. Red payo matrices denote partial
cooperation dilemmas.
Each of the 12 rectangular or triangular parcels of the coordinate system (separated by

full lines) host one ordinal payo ranking. There, we nd the prominent games like
the Prisoners Dilemma and the Chicken game, which is also often called Hawk-Dove
game or Snowdrift game. Complemented by Leader and Battle-of-the sexes, these are
7.3. Derivation of PCD in repeated games 123
the four archetypes of Rapoport (1967) (Martyr, Exploiter, Hero, and Leader).
Harmony is also referred to as By-Product Mutualism. The game of Route Choice
reects important characteristics of (vehicular- or data-) trac systems and was named
and experimentally investigated by Helbing et al. (2005b) and Stark et al. (2008a). The
name Own Goal was chosen arbitrarily in order to not leave one parcel empty. The game
is trivial as any deviation from the dominant strategy hurts the deviant most. The names
of the other games are taken from the literature (see e.g. Skyrms (2004) and Szabo and
Fath (2007)).
7.3 Derivation of PCD in repeated games
7.3.1 Partial cooperation in repeated games
Taking turns, originally describing the sequential form of human conversation (Duncan,
1972), has a considerable impact on repeated games, too. Here, it means that players anti-
coordinate their actions over time such that both take dierent decisions, but switch their
decisions in an alternating manner. This is also called alternating cooperation (Helbing
et al., 2005b; Stark et al., 2008a), alternating reciprocity (Browning and Colman, 2004),
or ST-reciprocity (Tanimoto and Sagara, 2007). The games in the dark-grey area of
Fig. 7.1 are the ones with the system-optimal solutions in partial cooperation, i.e. one
of the players prots more than the other. We call this area turn-taking phase as, in
repeated games, taking turns would strengthen the relevance of this solutions because of
the fairness with respect to the equal average payos (see also Bornstein et al. (1997);
Browning and Colman (2004); Helbing et al. (2005b); Neill (2003); Stark et al. (2008a)).
In games outside the dark-grey region or exactly on the dashed lines, an equal distribution
of payos is provided by the system-optimal solution, i.e. a unique strategy leads to
equal and system-optimal payos both in one-shot and repeated games. Since there is a
signicant dierence between games with the same payo ranking depending on whether
they are within or outside the turn-taking phase, it is important to address them precisely.
For the Prisoners Dilemma game with S + T > 2R, the name Turn-Taking Dilemma was
already proposed (Neill, 2003). Accordingly, we will speak of the TT-Chicken, TT-Route
Choice, and the TT-Deadlock for the respective games in the turn-taking phase.
Partial cooperation in repeated games can also mean to apply an interior mixed strategy.
That means a player randomizes its decisions and applies a probability to cooperate.
This allows for any individual level of cooperation, but does not bear the possibility to
anti-coordinate with other players over time. We will refer back to this form of partial
cooperation in an example later on.
7.3.2 From social dilemmas to PCD
Strictly relying on Dawes (1980) ([...] (a) the social payo to each individual for defecting
behavior is higher than the payo for cooperating behavior, regardless of what the other
society members do, yet (b) all individuals in the society receive a lower payo if all
defect than if all cooperate.), only the Prisoners Dilemma constitutes a social dilemma.
However, according to a more recent denition by Macy and Flache (2002), we face a
social dilemma situation if players prefer
mutual cooperation over mutual defection (R > P ),
mutual cooperation over unilateral cooperation (R > S),
mutual cooperation over partial cooperation, even if players would receive the average
payo (2R > T + S), and
the defection outcome over the cooperation outcome for at least one given strategy
of the other player, i.e. T > R and/or P > S.
To put it in one sentence: if choosing defection can lead to a Pareto-decient Nash equilib-
rium, the game is called a social dilemma since players are either tempted to unilaterally
defect, are afraid of being unilaterally defected on, or even both. In our opinion, condition
2R > T + S is less important for this denition since the fact that players can get stuck
in a Pareto-decient solution might be sucient to constitute a dilemma. Let us remark
that in addition to the three games nominated by Macy and Flache (2002) Prisoners
Dilemma, Chicken, and Stag Hunt also Pure Coordination I belongs into this subclass.
Social dilemmas, i.e. games that fulll all the above mentioned conditions, are indicated
by a red background color in Fig. 7.1.
In this Chapter, we investigate another type of dilemma that is there only in repeated
games, but not in the one-shot game. In repeated versions of a symmetrical 22 game,
the sucient condition for a dilemma is that, in the underlying one-shot game, there is
a Nash equilibrium that is not system optimal. In addition to the social dilemmas, this
is true for the games TT-Dilemma, TT-Route Choice, and TT-Deadlock, i.e. each game
with S + T > max(2R, 2P ). Whereas this variant of the Prisoners Dilemma can also be
seen as a social dilemma, the other two one-shot games do not hold a dilemma since their
Nash equilibria are strict and Pareto ecient. Therefore, in the classication of Rapoport,
the payo rankings of these two games are assessed almost trivial.
7.3. Derivation of PCD in repeated games 125
However, in repeated setups of all the three games, players might take turns in order
to persistently exploit the system optimum while sharing the payos evenly among each
other. Of course, such a solution would imply a Pareto-improvement compared to the
persistently played one-shot Nash equilibrium (itself the only Nash equilibrium of the
denitely repeated game), hence the dilemma. This can be best illustrated by the payo
matrix for a twice played symmetrical 22 game:
CC CD DC DD

CC 2R R+S S+R 2S
CD R + T R + P
S+T S+P
. (7.3)

DC T + R T + S P +R P +S

DD 2T T +P P +T 2P
The three games (TT-Dilemma, TT-Route Choice, and TT-Deadlock) have in common
that T > R and P > S. It follows that the solution (DD, DD) is the unique and strict
Nash equilibrium. In contrast to the one-shot game, this strict equilibrium is Pareto-
dominated by the solutions (CD, DC) and (DC, CD), because there both players receive
S + T > 2 P .2
We call this dilemma situation partial cooperation dilemma (PCD), because an ecient
solution requires partial cooperation instead of full cooperation. Whereas here we only
use the pure possibility of taking-turns as argument to illustrate the existence of a social
dilemma in repeated games, we refer to other works that investigate how turn-taking can
emerge and be maintained (Browning and Colman, 2004; Helbing et al., 2005b; Kaplan and
Rue, 2007; Neill, 2003; Stark et al., 2008a; Tanimoto, 2008). Interestingly enough, turn-
taking in repeated games of PCD combines the game theoretical problems of cooperation
and anti-coordination (see also Neill (2003)). A similar argument holds for another form
of partial cooperation, namely interior mixed strategies. Although they are not as ecient
as coordinated turn-taking, a Pareto-improvement can still be achieved (the next Section
contains a corresponding quantication). Partial cooperation dilemmas are indicated by
red payo matrices in Fig. 7.1.
It is worth noticing that, in his analysis of binary choices, Schelling already pointed to
the relevance of the Route Choice game (which he, with respect to the Tragedy of the
Commons, named maybe somewhat misleading The Commons) and other underrepre-
sented conicts (Schelling, 1978). However, the relevance of the TT-Route Choice game
2
It is worth noting that the repeated Turn-Taking Dilemma possesses a particularly interesting feature:
its Nash equilibrium is twice Pareto-dominated. Hence, there are three solutions of interest: (i) the strict
equilibrium, (ii) mutual cooperation without anti-coordination efforts required, but still Pareto-dominated
by (iii) turn-taking, which is Pareto-efficient, but requires temporal anti-coordination.
for scientic investigations has only recently been pointed out by two independent, but
related papers (Helbing et al., 2005b; Kaplan and Rue, 2007). Furthermore, there are
only a few publications that explicitly have dealt with the Turn-Taking Dilemma; see
particularly Neill (2003); Schuler (1986), and, for example, Kreps et al. (2001).
7.4 Derivation of PCD in evolutionary games
So far, we have derived the notion of partial cooperation dilemma games with respect to
their relevance for repeated interactions. In the following, we will argue that this classica-
tion is also meaningful in evolutionary game theory. Particularly in evolutionary biology,
the evolution of cooperation under natural selection remains a not fully understood, sci-
entic topic. Here, cooperation means that an individual has a genetic trait that makes
it help another individual at a certain cost (in terms of reproductive tness) to itself.
In the standard model (see right matrix in Eq. (7.1)), every cooperator induces exactly
the same benet b, independent of the number of cooperators in the population. Queller
(1985) argued that this strict additivity of benets is not necessarily realistic and proposed
to implement a synergistic term in the inclusive tness model of Hamilton (1964), which
is closely related to the model depicted right in Eq. (7.1). This synergistic term accounts
for the possibility that one cooperator induces a benet of b, but two cooperators together
induce a benet larger than 2b. In response, Grafen (1984) showed that this synergy-
term has no eect if selection is weak. However, this result is restricted to the specic
model applied there. If synergistic eects lead to a change in the payo ranking of the
underlying game (investigated below), evolutionary dynamics will be aected also in the
limit of weak selection. More recently, the possibility of synergistic or discounting eects
in N -person social dilemmas has been discussed (Hauert et al., 2006), where discounting
is the consequent opposite to synergy. If one cooperator induces a benet of b, discounting
eects lead to a joint benet of less than 2b induced by two cooperators. Implementing
this concept into the framework of symmetrical 22 games, we obtain
C D
! %
C (1 + w)
. (7.4)
D 0
The parameter w determines whether cooperation has synergistic eects (w > 1), dis-
counting eects (w < 1), or none of both (w = 1). By specifying > = b > 0 and
= c, we nd the according implementation of the synergy/discounting-concept
7.4. Derivation of PCD in evolutionary games 127
into the standard cooperation model:
C D
! %
C w b c c
. (7.5)
D b 0
The question is now: what are the dierent possible scenarios when considering synergistic
and discounting eects of cooperation based on the standard model? For this purpose, let
us systematically vary the parameter w: For w [(b + c)/2b, (b + c)/b], which includes
w = 1, we regain the traditional Prisoners Dilemma game with 2R > S + T . Hence,
to a certain extent, this model covers both synergistic and discounting eects. However,
increasing w above (b + c)/b, the game eectively transforms into a Stag Hunt game.
That means if synergistic eects are strong enough, defection is not anymore a dominant
strategy. In evolutionary terms, we derive a bistable system where both strategies are
evolutionarily stable against each other. Most interestingly, the remaining three possible
games, generated by a discounting factor w < (b + c)/2b, are exclusively the partial
cooperation dilemmas. For w [c/b, (b + c)/2b], we nd a TT-Dilemma. For w [0, c/b],
the payo ranking is the one of TT-Deadlock. By the same reasoning, we obtain that
values w < 0 result in the TT-Route Choice game (see arrow 2 in Fig. 7.2).
Fig. 7.2 gives examples of payo transformations when considering dierent relations be-
tween and . In all scenarios, > 0 is respected such that synergistic and discounting
eects of cooperation are considered. One could likewise investigate synergistic and dis-
counting eects of defection, and the according transformations would lead to vertical
arrows in Fig. 7.2. The possibility to draw such straight arrows also show that the scheme
of Fig. 7.1 is integrative in the sense of visualizing relations between the games by spatial
arrangement. For example, the dierent mutualistic scenarios studied by Bergstrom and
Lachmann (2003) can be found in the parcels below the P = S line (arrow 1).
Let us return to arrow 2 in Fig. 7.2. Note that in all the dierent cases along this arrow,
the evolutionary problem of cooperation is addressed (helping behavior that induces a
tness-benet to the recipient and a tness-loss to the helping individual), but in dier-
ent environmental scenarios. Note further that the three partial cooperation dilemmas
possess a dominant strategy, just like the standard cooperation model. When, for ex-
ample, considering replicator dynamics (strategies that are more successful than average
increase their share in the population, see e.g. Hofbauer and Sigmund (1998) and Taylor
and Jonker (1978)) in innite, well mixed populations (interactions occur between random
individuals), a stable population would consist of defectors only. However, what makes
these games worth considering besides the standard cooperation model is that too many
Figure 7.2: Scheme of symmetrical 2 2 games according to Fig. 7.1. A possible position
of the standard cooperation model (right matrix in Eq. (7.1)) is indicated by the black
circle on arrow 2. The dotted arrows 1-4 indicate transformations of the payo matrices
when increasing the parameter w, i.e. the synergistic/discounting eects of cooperation
in the framework of matrix 7.4. In line with the argumentation of this article, T > S is
always respected.
cooperators may reduce overall tness. For example, a group (be it in the sense of group
selection, spatial clusters, or similar) consisting of cooperators only is not the most suc-
cessful group, but one in which cooperators and defectors coexist persistently would have
the highest group tness. Let us quantify the tness of a group as the expected payo
of a random interaction between two group members:
= 2 x2 R + 2 x (1 x) (S + T ) + 2 (1 x)2 P. (7.6)
In this equation, x is the frequency of cooperators in the group. If two cooperators

(defectors) meet (which randomly happens with probability x2 ((1 x)2 )), both receive
7.5. Evolution of partial cooperation 129
payo R (P ). If a cooperator and a defector meet (probability 2x (1 x)), one receives S

and the other T . has its maximum at
2P (S + T )
x = , (7.7)
2[(R + P ) (S + T )]
where 0 < x < 1, because S + T > max [2R, 2P, (R + P )] by the denition of partial
cooperation dilemmas. Let us remark that the value of x also corresponds to a mutually
optimal mixed strategy, i.e. if every player cooperates with probability to the amount of
x , the outcome is system-optimal and characterized by equal expected payos. This is in
contrast to the individually optimal strategy, which is x = 0. For the Prisoners Dilemma,
x = 1, i.e. only full cooperation would be mutually optimal. Due to the fact that x
is intermediate in partial cooperation dilemmas, the conceptual dierence to classical
social dilemmas becomes obvious: we do not ask the question how cooperation can achieve
evolutionary stability, but how an ecient coexistence of strategies can stabilize.3 In
fact, similar questions are addressed by many researchers seeking for explanations regard-
ing the huge biodiversity (Doebeli et al., 2004; Kerr et al., 2002; Reichenbach et al., 2007)
and variation in cooperation (Kurzban and Houser, 2005) and helping eorts (Field et al.,
2006). The game rock-paper-scissors, where three strategies dominate each other in a
cyclic fashion, is then most often used as paradigmatic modeling approach (Czaran et al.,
2002; Kerr et al., 2002; Reichenbach et al., 2007). However, this game requires at least
three strategies and may straight-forwardly promote coexistence (see also Claussen and
Traulsen (2008)). Contrarily, partial cooperation dilemmas could be used to investigate
the emergence of coexistence states where evolutionary dynamics is expected to drive
the system into dominance of only one specic behavior (in line with the considerations
by Imhof et al. (2005)).
7.5 Evolution of partial cooperation

As a rst step to investigate the possibility of stable coexistence in partial cooperation
dilemmas, let us consider the Five Rules for the Evolution of Cooperation (Nowak,
2006b), i.e. ve evolutionary concepts that, under certain circumstances, can eectively
change the strategic situation (the game) compared to a single, binary interaction. The
ve concepts are direct and indirect reciprocity (direct: if I help you, you might help me;
indirect: if I help you, someone else might acknowledge it and help me), kin selection (if
I help a genetic relative, I actually help my own genotype to survive), group selection (if
3
Compare to general results in finite systems (Antal and Scheuring, 2006).
I help a group member, I strengthen my own group which might protect myself in the
competition with other groups), and network reciprocity (similar to group selection, if I
help my neighbors and we build up a cooperative, spatial community, our neighborhood
(including myself) might be protected against defective invaders).4
In the work of Nowak (2006b), these concepts are valuably simplied by implementing
them into the standard cooperation model. Using the scheme of Fig. 7.1, the specic
transformations for this Prisoners Dilemma game have been constructed in the Appendix.
In a subsequent work, these mechanisms have been applied to an arbitrary Prisoners
Dilemma game and, among others, the conditions for stability of coexistence between
cooperators and defectors within this payo ranking derived (Taylor and Nowak, 2007).
It was found that the concepts of kin selection, group selection, and network reciprocity
can lead to stable coexistence if S + T > R + P , i.e. in the discounting region (w < 1).
Direct and indirect reciprocity cannot lead to stable coexistence. But do these results also
hold for discounting factors beyond the Turn-Taking Dilemma?
As derived by Nowak (2006b), we can illustrate the eects of kin selection, group selection,
and network reciprocity on a symmetrical 22 game:
C D C D
! % ! %
C (1 + r)R S + rT C (n + m)R nS + mR
D T + rS (1 + r)P D nT + mP (n + m)P
C D
! %
C R S+H
, (7.8)
D T H P
where r is the relatedness coecient (mostly dened as probability of two individuals

sharing a gene, i.e. r [0, 1]), n, m are group size and number of groups, respectively,
and H = [(k + 1)(R P ) + S T ]/[(k + 1)(k 2)], with the degree of the network
k > 2.5 Dierently from the other two mechanisms, the role of relatedness can vividly
be understood: inclusive tness theory (Hamilton, 1964) states that the fate (in terms of
reproductive tness) of a genetically related species also inuences the fate of the species
itself, because the probability of sharing a gene is higher than between unrelated species.
The strength of this inuence depends on the level of relatedness r. For calculating the
overall tness eects for one species, one therefore has to add the tness of the related
species, scaled by r. This is done in the upper left matrix of Eq. (7.8). In the same spirit,
4
In the Discussion of this Chapter, we will comment on criticisms related with these concepts.
5
Note that these results are obtained in the limit of weak selection.
7.5. Evolution of partial cooperation 131
albeit not likewise comprehensive, the other two concepts are implemented (see Nowak
(2006b) and references therein for details).
Evolutionary dynamics lead to stable coexistence of species if none of the species is evolu-
tionarily stable. The conditions, under which this is fullled in the payo rankings of the
three partial cooperation dilemmas, can be found in Table 7.1, repeating and extending
the results of Taylor and Nowak (2007). If a condition cannot be met in the respective
payo matrix, this impossibility of stable coexistence is indicated by dashes. For group
selection, this happens because m/(m + n) can only vary between 0 and 1, thereby vio-
lating the according (not shown) conditions. At a rst glance, this result seems surprising
as a mixed group of cooperators and defectors performs better than a group of defectors.
However, this mechanism bases on individual reproduction and not on the reproduction
of groups. Since, in contrast to the Turn-Taking Dilemma, a cooperator in any group is
less t than a defector in any group, also higher-level selection favors defection (compare
to Traulsen and Nowak (2006)). For network reciprocity, the condition to be hold is the
same for all three games, i.e. the one displayed for the TT-Dilemma. In networks with
k > 2, FD is always positive in the games TT-Deadlock and TT-Route Choice, because
S + T > R + P and P > max(R, S), thereby violating the condition. This result is rather
intuitive as, in contrast to the TT-Dilemma, a cluster of cooperators performs worse than
a cluster of defectors.
Only in kin selection, there is a range of r leading to stable coexistence in all three games of
partial cooperation dilemmas (direct and indirect reciprocity are left out of the discussion
because in none of the scenarios stable coexistence of strategies can emerge).
TT-Dilemma TT-Deadlock TT-Route Choice
P S
KS T P < r < TRS
R
r> P S
T P
P S m
GS RS < m+n < TT P
R

NR FD < 0 < FC
Table 7.1: Conditions for stable coexistence of strategies in the three partial cooperation
dilemma games for kin selection (KS, with relatedness r), group selection (GS, with group
size n and number of groups m), and network reciprocity (NR, where FD = k 2 (P S)
k (R S) + S + T R P, FC = k 2 (P S) k (R S) + S + T R P and k > 2 denotes
the number of neighbors per individual in the network). Direct and indirect reciprocity
cannot lead to stable coexistence.
7.6 Discussion
Although symmetrical 22 games are very simplistic interaction models, some of them
attracted considerable attention in various elds of science. The reason lies in the fact
that it is possible to model paradigmatic scenarios of interaction between social individuals,
economic actors, biological species, etc.. In the classication of such games, we explicitly
distinguish between games that have their system optimal solution in the main-diagonal
of the payo matrix and games where a solution in the o-diagonal, i.e. an outcome of
partial cooperation, leads to the highest overall payo. In the latter case (turn-taking
phase), payos are dierent for the involved individuals and turn-taking would be an
advantageous behavior in repeated games. In four of the 12 ordinally distinct games, both
cases are possible.
Related to the well known subclass of social dilemmas, we speak of a dilemma in repeated
games if equilibrium play might lead to a solution that is not system optimal. Among the
symmetrical 2 2 games, this is additionally true for variants of the Prisoners Dilemma,
the Route-Choice game, and Deadlock that lie within the turn-taking phase. We call
them partial cooperation dilemmas, because these games bear a dilemma situation both
in repeated and evolutionary games that can only be resolved by partial cooperation.
In repeated games, partial cooperation might be realized by coordinated turn-taking or
the application of intermediate mixed strategies. Both variants are forms of (partial)
cooperation that yield a payo improvement for both players (compared to the strict
Nash equilibrium in denitely repeated games).
In many works on symmetrical 22 games, the condition R > P is applied to name
one of the binary choices cooperation and the other one defection. We used the
diering condition T > S, because it allows to illustrate the relations between neighboring
games in Fig. 7.2 by dierent strengths of synergistic or discounting eects of cooperation.
However, it is important to note that both approaches can only fulll the task to apply
the commonly accepted Prisoners Dilemma notation to the whole of symmetrical 22
games. The scientic term cooperation is dened as helping act, i.e. an individual
behavior that induces a tness benet to another individual. This denition is clear and
unproblematic when considering frequency independent tness eects like in the example
of the right matrix in Eq. (7.1): if the tness eect of my strategy on another individual
(b) is positive, I am regarded as cooperator. If, however, the tness eect of my strategy
is frequency dependent, one and the same action can lead to a positive or a negative
tness eect for the other. Hence, a clearly stated adaptation of the term cooperation is
useful because it avoids a confusing naming of strategies conditional on frequencies in the
population. In the games below the diagonal of Fig. 7.1, which include all social dilemma
7.6. Discussion 133
games, the conditions R > P and T > S coincide. In the other games, including the
partial cooperation dilemmas Route Choice and Deadlock, the strategy naming is inverted.
Especially with respect to the PCD, we also prefer using T > S: in view of the system
optimal, polymorphic population, we rather would evaluate the behavior as cooperative
that ignores its potential gain from switching for the benet of the community (instead of
the one which actually is the receiver of a cooperative act).
In evolutionary biology, where individual payo gains contribute to the reproductive t-
ness of a species, another form of partial cooperation plays an important role: stable
coexistence of cooperative and non-cooperative strategies. A species that maintains such
a coexistence (think of dierent roles in ant colonies or the dierentiation in eukaryotic
microorganisms and similar forms of cooperation, see e.g. Wingreen and Levin (2006))
might be advantageous, but it remains a challenge for evolutionary biologists to com-
pletely understand how such forms are protected against cheating, i.e. other organisms
that prot from cooperation but contribute less cooperation themselves. Therefore, in
evolutionary game theory, partial cooperation dilemmas are even relevant when consid-
ering one-shot games, i.e. interactions without the possibility of turn-taking or similar,
memory dependent strategies.
Whereas instances of partial cooperation dilemmas have been discussed in previous works
(Helbing et al., 2005b; Kaplan and Rue, 2007; Kreps et al., 2001; Neill, 2003; Schuler,
1986; Stark et al., 2008a), we here provide a concise conceptualization of the general kind of
dilemma. For symmetrical 22 games, we show that partial cooperation dilemma games
translate to the standard model of biological cooperation when considering discounting
eects of helping eorts. Consequently, we derive the conditions for stable coexistence
dependent on the strength of evolutionary mechanisms at work, thereby complementing
recent ndings (Taylor and Nowak, 2007). We are convinced that there is room for new
thoughts on realistic mechanisms that are able to explain diversity in a wide range of
evolutionary scenarios, especially in partial cooperation dilemmas.
Some concepts that are applied in this work have been and still are subject to a scientic
dispute. In particular, the multi-level selection approach (Traulsen and Nowak (2006),
leading to the group selection concept) and and the role of spatial structure (leading to
the network reciprocity concept) is sometimes proposed to be identical to kin selection (see
particularly Lehmann and Keller (2006), West et al. (2007) and references therein). In our
view, it is semantically rather productive than misleading to distinguish between dierent
sources of indirect tness benets. In situations where dierences in genetical relatedness
can be cancelled out, the mechanism to explain why cooperation is selected for should
not be kin selection. Apart from this semantic argument, in this work it quantitatively
proved valid to distinguish between the concepts: as shown above and in the Appendix,
they transform the dilemma in dierent ways and lead to diering results. As stated above,
a major source of discrepancy seems to be the choice of the underlying model: inclusive
tness theory assumes the costs end benets of cooperation to be frequency independent,
whereas other approaches, like the present one, emphasize frequency dependence.6
Although widely neglected by the literature so far, the games exhibiting a partial coop-
eration dilemma could widen the range of models describing complex scenarios of reality
without increasing the complexity of the model (they base on a simple symmetrical 22
game). Our conjecture is that applying the idea of partial cooperation dilemmas into
respective models and experimental setups will lead to new and relevant insight regard-
ing the evolution of cooperation in biological systems and human society. In particular,
we reckon advances in investigations on the huge biodiversity and heterogeneity of social
behaviors in nature.
6
Although the considerations in the Appendix base on a frequency independent model, the investigated
transformations explicitly depart from this characteristics.
Part III
Conclusion
In the present thesis, we have investigated dierent models of social interaction. These
models are either based on the eld of opinion dynamics, the eld of (evolutionary) game
theory, or they explicitly use approaches from both elds. Besides presenting the results
of particular models, we constantly discuss the dierences, but more importantly the large
common ground of decisions based on social inuence (opinion dynamics) and in strategic
decision situations (game theory). The main dierence can be determined in the focus
of both approaches: in evolutionary game theory, the most important question is which
of the qualitatively dierent behaviors (strategies) is present in a stable population of
individuals and what are the consequences for the payo or tness of this population?
Especially in investigations on the emergence and maintenance of cooperation, one seeks
for explanations of successful cooperation in real social and biological scenarios. With-
out further extensions, evolutionary dynamics favors defectors and cooperation cannot be
found in stable populations. Hence, the focus in these kind of models lies on evolutionary
mechanisms that can explain the evolutionary advantage of a particular behavior, namely
being cooperative. In contrast, models of opinion dynamics do not have a measure of indi-
vidual payo. The focus here lies on whether and why a system reaches an ordered state
(consensus) and how the dynamics in the system can be understood. The absence of pay-
o considerations shifts the focus to the intrinsic dynamical properties of the interaction
system, which are often non-linear and sometimes lead to unexpected phenomena. Apart
from this dierence, both research elds contribute to the understanding of social inter-
action systems. It is one aim of this thesis to combine both elds in an explicit manner,
i.e. to apply model constituents that are well motivated in opinion dynamics to models
of evolutionary game theory. After introducing models of social inuence in Chapter 2
and models of strategic decision situations in Chapter 3, we discussed this ambition of the
present work in Chapter 4.
Starting with opinion dynamics, we have investigated an extension of the voter model in
Chapter 5. Instead of fully symmetrical decisions (opinions), we assumed that individuals
distinguish between their own current decision and the other one. Dependent on the
persistence time of their current decision (the time since they changed their decision for the
136
last time), individuals have an inertia to change their decision the longer an individual
has applied its current decision, the less likely it will change. Such an inertia can for
example be explained by a conviction that builds up through discussions. If an individual
has discussed a certain topic many times without changing its view on this topic, the
more it gets convinced of its own view. Of course, this invariably leads to decreased
probabilities of individuals to change their decision and we expected this mechanism to
hinder the consensus process in the observed systems. However, we found the eect to be
non-monotonous and partially counter-intuitive. If the individual inertia grows slowly with
persistence time, the average time to reach consensus in the systems gets lower compared to
the original voter model, i.e. the ordered state is reached faster. This means that systems
with a higher level of inertia exhibit faster instead of hindered ordering dynamics. Only for
a faster growth of inertia with persistence time, this eect is reversed and the average time
to consensus is larger in systems with a higher level of inertia. We show that the origin
of this phenomenon lies in the coevolution between inertia and decisions in the system, as
consensus times invariably increase when assuming homogeneous, or static, heterogeneous
inertia. These results are particularly important as they are neither restricted to the voter
model nor to the particular denition of the persistence time dependence. Therefore,
we have found a general example of an emergent phenomenon, resulting from non-linear
dynamics.
In one approach to combine opinion dynamics and game theory, we applied this persistence-
time dependent mechanism to the dynamics of a spatial Prisoners Dilemma with hetero-
geneous teaching abilities of individuals. Teaching means that certain individuals have a
probability to spread their own decision which lies above the probability stemming from
the original game dynamics. They are more inuential than others and initially these
teaching abilities are equally distributed over cooperators and defectors. If this hetero-
geneity of teaching abilities is constant, it has been shown that this promotes the evolution
of cooperation in the spatial Prisoners Dilemma. We here let these teaching abilities, in
accordance to inertia in the voter model, grow with persistence time of decisions. As in evo-
lutionary dynamics the persistence time of a decision often is translated into the life-time
of an individual (which has a certain decision hard-coded in its genes), we speak of teach-
ing abilities depending on the age of individuals. We nd that this mechanism induces
a considerable, additional promotion of cooperation in the spatial Prisoners Dilemma.
Although we found the reasons for this eect to stem from the spatial structure of the
system, and we try to have a much more general scope within this thesis, these results still
t very well into the argumentation of this work: they provide a straight-forward example
of applying a mechanism explicitly derived in one framework (opinion dynamics) to the
methodology of the other (evolutionary game theory). This was possible without much
137
diculty in motivating the approach and has led to relevant new results.
In a second approach, we investigated a model which is far more general. The idea is
to apply a fundamental mechanism of opinion dynamics to one of the most fundamental
frameworks of evolutionary game theory: we implement pure (i.e. payo independent)
imitation behavior into the replicator dynamics between cooperators and defectors. In the
main part of these investigations, the only further assumption is that interactions occur re-
peatedly in order to allow for imitation, but notably our analytical results hold for innite,
well-mixed populations, i.e. without assuming spatial structure and without nite-size ef-
fects. The population consists of pure cooperators, pure defectors, and pure imitators.
An imitator defects with a defector, cooperates with a cooperator, and nds consensus in
either cooperation or defection with another imitator. In our view, this setup implements
imitational behavior with the lowest possible level of assumptions. Computing the stable
populations that result from this dynamics for the whole of symmetrical 22 games, we
nd striking consequences for the evolutionary dynamics in 3 of 4 social dilemmas. Im-
itators act as catalyst for cooperators, thereby completely resolving the dilemma in the
games Stag Hunt and Pure Coordination I. In the Prisoners Dilemma, where the temp-
tation to defect is particularly strong, dominance of defectors is not possible if imitators
are present in the system. Imitators, although they are themselves not able to dominate
in any of the scenarios, are successful enough to not face extinction, thereby preventing
the victory of defectors. Due to the imitators, cooperators also do not die out and the
nal population is characterized by coexistence of all three types. Simulations in nite-size
systems and structured populations emphasize these general results: comparing the evo-
lutionary success of all three types in all scenarios, defectors did not only lose dominance,
but they turned into the weakest sub-population if imitation was present. Especially due
to the absence of strong assumptions in our model setup, we conclude that imitation is
a very natural and powerful mechanism to support the evolution of cooperation in social
dilemmas.
In addition to the relevance of these ndings for research on the evolution of cooperation,
we want to emphasize their importance for our discussion of dierent perspectives on social
interaction. The simple combination of fundamental mechanisms from the elds of opinion
dynamics and evolutionary game theory proved to be relevant for a highly interdisciplinary
research topic. We are convinced that both social inuence and strategic decision making
determine the dynamics of social interaction and that, therefore, further attempts to unify
knowledge from both elds into one framework denotes a promising eld of future research.
Such further research, especially when trying to model real strategic decision situations
by symmetrical 22 games, should also carefully identify the best suited model, instead
138
of just applying the most prominent ones (e.g. the Prisoners Dilemma). Throughout the
respective parts of this thesis, we always investigated the whole range of models comprised
in the class of symmetrical 22 games, which is however not common practice. In the
nal Chapter 7, we conceptualize the framework of partial cooperation dilemmas as a
distinct subclass of symmetrical 22 games. These games exhibit a cooperation dilemma in
repeated and evolutionary games, but not in a once played (one-shot) game. We separately
derive the dilemma situation for repeated and evolutionary games and argue, that this kind
of dilemma may exist in reality as likely as the well recognized social dilemmas (it is in most
cases not possible to exactly determine the payo structure of real situations). However,
scientic applications of such models are very rare and especially a precise embedding of
this concept into the literature has been lacking so far. This argumentation also highlights
many specic connections between the mainly economically motivated eld of game theory
and social and biological evolutionary dynamics.
The present thesis contains two dimensions of content: (i) contribution to the under-
standing of particular dynamic models and their results in the elds of opinion dynamics
and evolutionary game theory and (ii) a more conceptual argumentation on the model-
ing aspects themselves and contributing to bridging the gap between the two dierent
elds. Both by means of eorts dedicated and concision of results, dimension (i) is surely
more pronounced and deserves particular attention (achievements and their conlusions
have been summarized above). For dimension (ii), it was the aim to not only comment
on commonalities and dierences between the two research elds, but to explain why we
consider them as being two perspectives on one and the same matter: models of social
interaction. In conclusion we can say that the perspectives dier in their focus while
in game theory the focus lies on the rationality or the evolutionary success of dierent
behaviors, in opinion dynamics one does not assume an evolutionary competition between
dierent behaviors, but one tries to fully understand the dynamical properties of one or
more particular interaction behaviors under certain circumstances. Although focused on
language competition, the work of Castello et al. (2006) constitutes a good example
to distinguish between the perspectives: while in game theory one would be interested in
which language survives due to the evolutionary advantages or disadvantages a specic
language causes for its speaker, the authors in this work assume symmetric languages,
i.e. every language causes the same eects for their speakers. They are interested in the
resulting dominance or coexistence of languages under the assumption of a certain decision
behavior of the individuals, which itself is not subject to competition. The same dieren-
tiation can be illustrated by comparing the voter model with the evolutionary dynamics in
a simple coordination game, as has been done around Fig. 4.1. In the voter model, pure,
success independent imitation behavior is investigated without questioning the rational-
139
ity of this behavior. In evolutionary game theory, one could assume that individuals get
rewarded for beeing alike with their interaction partner(s), and one would nd out that
pure imitation is less successful than proportional imitation, i.e. conditional imitation
depending on the success of the observed behavior. Despite these dierences, the mod-
eling frameworks of both elds are actually identical, which allows to easily recombine
approaches and ndings between both elds. Within this thesis, we explicitly followed
this idea of recombination by applying the co-evolutionary rule of Chapter 5 to the spatial
Prisoners Dilemma (Section 6.3), exclusively investigated behavior through social inu-
ence in evolutionary games (Section 6.2), and exposed the purely imitational behavior of
the voter model to evolutionary dynamics in social dilemmas (Section 6.4). The most
remarkable conclusion of these approaches is surely the survivability of purely imitational
behavior in social dilemmas and their striking eect on the evolution of cooperation. Due
to the discussions and results in this thesis, we would expect further attempts to explicitly
recombine modeling approaches from both elds to yield further interesting and relevant
results.
Appendix: Transformations of the
standard cooperation model
Referring to Nowak (2006b), there are (at least) Five rules for the Evolution of Coopera-
tion. Assuming the standard cooperation model of matrix (3.10), Nowak and colleagues
implemented 5 evolutionary mechanisms in the Prisoners Dilemma payo matrix and
derived the conditions under which cooperation can become evolutionarily stable against
defection. It is important to note that this is only possible if the eective payo matrix
gets transformed. In the Prisoners Dilemma game, cooperation is under no circumstances
evolutionaily stable against defection. When implementing the dierrent mechanisms to
the underlying payo matrix of the Prisoners Dilemma, the resulting payo matrix does
not necessarily reect a Prisoners Dilemma game anymore. If it does (or in the parameter
regions where it does), the mechanism has no eect on the dynamics.
In order to obtain a better conceivability of the mechanisms and how they work, let
us explicitly name the transformations induced by applying the ve mechanisms to the
standard cooperation model. If the necessary conditions on the parameters are fullled,
the PD game is transformed into:
Harmony I for kin selection
Stag Hunt for direct and indirect reciprocity
network reciprocity:
Deadlock if b < H
Harmony I if b > H and (b + c)/2 > H
Harmony II, if b > H and (b + c)/2 < H (H = [(b c)k 2c]/[(k + 1)(k 2)])
group selection:
140
141
Harmony I if 2b/(b + c) < 1 + n/m < b/c

Harmony II if 1 + n/m < 2b/(b + c),
where k is the average number of neighbors per individual in the network, n is the max-
imum group size and m is the number of groups (see Nowak (2006b) for details). With
these transitions, the evolutionary dynamics changes from a system with one stable xed
point at 100% defectors (PD) to a bistable system with two xed points at 100% of either
cooperators or defectors (Stag Hunt) or a system with 100% cooperators in the only xed
point (Harmony I).
List of Tables
7.1 Conditions for stable coexistence of strategies in the three partial coop-
eration dilemma games for kin selection (KS, with relatedness r), group
selection (GS, with group size n and number of groups m), and network
reciprocity (NR, where FD = k 2 (P S) k (R S) + S + T R P, FC =
k 2 (P S) k (R S) + S + T R P and k > 2 denotes the number
of neighbors per individual in the network). Direct and indirect reciprocity
cannot lead to stable coexistence. . . . . . . . . . . . . . . . . . . . . . . . 131
142
List of Figures
1.1 The dynamical function of Eq. (1.1) dependent on the state x. Arrows
indicate the direction of the dynamics in the vicinity of the three xed points. 9
2.1 Exemplary time evolution of the voter model on a 2-dimensional, regular

lattice, where every voter interacts with its 4 nearest neighbors. The system
consists of N = 10, 000 voters and the evolution time t is measured in
generations, i.e. 10,000 single updates correspond to 1 generation. . . . . . 21
3.1 Presentation of the rock-paper-scissors game in extensive form (top) and

in normal form (bottom). In the upper picture, big, black circles represent
decision nodes where the respective player has to decide for one of its
alternatives. The dashed ellipse indicates an information set: a player
does not know which of the decision nodes within one information set is
its current one. Therefore, the example game is a simultaneous one. The
small, black circles at the ends of branches indicate the possible outcomes
of the game and the payos are given in brackets: the rst one for player 1
and the second one for player two. The lower picture contains exactly the
same information about the game in form of a payo bimatrix. . . . . . . . 26
3.2 Classication of symmetrical 22 games according to payo rankings. For

each area a respective payo matrix (in the form of matrix (3.5), with T >
S) is given. Nash equilibria are marked by bold payo numbers. Parcels
separated by solid lines denote dierent rankings of the payo values. Two-
dimensionality is achieved by xing T > S and classifying ordinal dierences
only. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
143
144 List of Figures
3.3 Analytically computed equilibrium fraction of cooperators according to

replicator dynamics between cooperators and defectors in an innite, well-
mixed population. In case of bistability (R > T P > S), the stable
xed point with the bigger attractor region is displayed. Every R, P -
coordinate constitutes one specic payo matrix (a game) with xed values
of S = 0, T = 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Simulation results: equilibrium fraction of cooperators after the evolution

of cooperators and defectors in a system of 400 agents. Considered are
dierent population structures: (a) fully connected networks; (b) random
networks with 8 interactions per agent; (c) 2-dimensional, regular lattices
with 4 neighbors and (d) 8 neighbors, respectively. Compared to Fig. 3.3,
the most interesting region around P = S = 0, R = T = 1 is magnied. . . 44
4.1 Simulation results of two systems with N = 100 individuals, where every
individual is connected to all other individuals (fully connected network).
The magenta colored symbols display the results of the voter model dynam-
ics (as explained in the text, without selection pressure). The blue symbols
represent results from the replicator dynamics in the coordination problem,
i.e. in a game with R = P = 1, S = T = 0. Circles show the development
of the absolute magnetization in the system, a measure for the homogeneity
of opinions in the system, where the value zero indicates equal frequencies
and the value one consensus (M = |1 2x|, if x is the relative frequency
of one opinion). Crosses denote the wealth in each system in terms of
the expected payo of one individual when playing the coordination prob-
lem game with a random other individual. Values are averaged over 10,000
simulation runs for each system. . . . . . . . . . . . . . . . . . . . . . . . . 49
5.1 Left: Average time to consensus T in the voter model with a xed and
homogeneous inertia value 0 = . The line corresponds to the theoretical
prediction T () = T ( = 0)/(1 ). Details are given in the text. Right:
comparison of the development of the average interface density in the
voter model and the model with xed inertia. Right, inset: collapse of
the curves when the time scale is rescaled according to t t/(1 ). In
both panels, the simulations results stem from averaging over 500 sample
runs, where the system size is N = 30 30 and the voters are placed on a
two-dimensional, regular lattice. . . . . . . . . . . . . . . . . . . . . . . . 61
List of Figures 145
5.2 Average consensus times in simulations of the voter model embedded into
fully connected networks. (a) Surface of average consensus times when
varying both the non-zero inertia value and the frequency of voters holding
the non-zero inertia value. The system size is N = 200 and averages are
obtained from 4000 simulation runs. (b) Here, the inertia values are not
binary, but continuous between 0 and the maximum inertia value and all the
voters have a non-zero inertia. Results are presented for dierent system
sizes N and the number of simulation runs varies between 10000 for the
smallest system size and 2000 for the largest system size. . . . . . . . . . . 62
5.3 Average time to reach consensus T as a function of the maximum inertia

value . Panels (a)-(d) show the results for dierent system sizes in one-,
two-, three-, and four-dimensional regular lattices, respectively. The results
are averaged over 104 realizations. The system sizes for the dierent panels
are the following: (a) N = 50 (), N = 100 (), N = 500 (); (b) N = 302
(), N = 502 (), N = 702 (); (c) N = 103 (), N = 153 (), N = 183
(); (d) N = 44 (), N = 54 (), N = 74 (). . . . . . . . . . . . . . . . . 63

value in small world networks (see text for details). The symbols represent
dierent rewiring probabilities , starting with a 2-dimensional, regular
lattice ( = 0). The curves corresponds to = 0 (), = 0.03 (),
= 0.1 (), and = 0.9 (). System size is N = 302 and results are
averaged over 104 realizations. . . . . . . . . . . . . . . . . . . . . . . . . . 64

value in fully connected networks of dierent size. Results are averaged
over 104 realizations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.6 Illustration of the four fractions a and b and the possible transitions of a
voter. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

value in fully connected networks of dierent size. Symbols show the
simulation results for dierent system sizes (averaged over 104 realizations),
lines the results of the theoretical estimation (explained in the text). . . . . 69
146 List of Figures
5.8 Average consensus times T for varying values of the inertia slope and xed
saturation value s = 0.9. Sample sizes vary between 103 104 simulation
runs. Filled, black symbols always indicate the values of T at = 0. (a)
2d regular lattices (ki = 4) with system sizes N = 100 (), N = 400 (),
N = 900 (). The inset shows how consensus time scales with system size
in regular lattices at = for d = 1 (), d = 2 (), d = 3 (), d = 4
(). (b) Small-world networks obtained by randomly rewiring a 2d regular
lattice with probability: () pr = 0, () pr = 0.001, () pr = 0.01, ()
pr = 0.1, () pr = 1. The system size is N = 900. (c) Fully connected
networks (mean eld case, ki = N 1) with system sizes N = 100 (),
N = 900 (), N = 2500 (), N = 104 (). Lines represent the numerical
solutions of Eqs. (5.18), (5.19), (5.20) with the specications in the text.
The inset shows the collapse of the simulation curves by scaling and T
as explained in the text. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.9 General illustration of the fractions a and b and the possible transitions
of a voter. The fractions aT and bT contain all voters with a persistence
time i s , i.e. voters with maximal inertia. . . . . . . . . . . . . . . . . 72
5.10 Exemplary time evolution of the inertial voter model with multiple inertia
states and = 0.1, s = 0.9. Voters are positioned on a 2-dimensional,
regular lattice, where every voter interacts with its 4 nearest neighbors. The
system consists of N = 10, 000 voters and the evolution time t is measured
in generations, i.e. 10,000 single updates correspond to 1 generation. . . . . 79
5.11 Comparison of the logistic decision function with the one of the linear voter
model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.12 Order-disorder phase transition for transition probabilities following a Fermi

function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.13 Numerical simulations of an N = 900, fully connected network. . . . . . . . 82
6.1 Promotion of cooperation due to the increasing heterogeneity in the ei wi

mapping via . Stationary fraction of cooperators xs is plotted in depen-
dence on b for = 0 (solid red line), = 1 (dashed red line) and = 2
(dotted blue line). In all three cases K = 1. . . . . . . . . . . . . . . . . . 92
List of Figures 147
6.2 Full b K phase diagrams for the Prisoners Dilemma game with quenched
uniform distribution of ei , obtained by setting = 0 [panel (a)] and = 2
[panel (b)] in the ei wi mapping. Solid green and red lines mark the
borders of pure C and D phases, respectively, whereas the region in-between
the lines characterizes a mixed distribution of strategies on the spatial grid.
The dashed, blue line at b = 1 denotes the boarder between the essential
Prisoners Dilemma payo parameterization (above the line) and a dilemma-
free situation (below the line). . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.3 Full b K phase diagrams for the Prisoners Dilemma game incorporating
aging as a dynamical process. In both panels, = 2 (to be compared with
panel (b) of Fig. 6.2). (a) results for model II, where aging is dynamical,
but follows a deterministic protocol; (b) results for co-evolutionary model
III, where the individual teaching values wi evolve corresponding to the
persistence time of strategies, i.e. individuals who change their strategy are
considered newborn (ei = 0) in the next timestep. Additionally, the dashed
lines indicate the results for a slightly changed scenario, where only 10 %
of the individuals increase their age per time step. To avoid ambiguity, C
and D symbols are not given in panel (b), but the respective regions can
be inferred by the line colors according to panel (a). . . . . . . . . . . . . . 94
6.4 Equilibrium fraction of cooperators x in the two non-trivial xed points.

The x-axis and the y-axis are the same as in the two-dimensional Figs. 3.2
and 3.3, i.e., together with our general denition of S = 0 and T = 1, they
dene the respective game. (a) Fraction of cooperators in the non-trivial
xed point of the evolution between cooperators and defectors only [xed
point (iv) in the text]. (b) Fraction of cooperators in the additional, non-
trivial xed point through the inclusion of imitators [xed point (v) in the
text]. This one only exists for games with 0 < P < R < 1, i.e. only in the
Prisoners Dilemma. In this Figure, stability of xed points is not illustrated.100
6.5 Illustration of the state space (triangular simplex) for the dynamics on
the relative frequencies of three species: cooperators x, imitators i, and
defectors 1 x i. The dynamical system can only be in a state within the
triangle or exactly at its border. The black circle denotes xed point (i), i.e.
the domination of defectors. Dashed lines are illustrations of eigenvectors,
aimed at supporting the verbal explanation in the text. . . . . . . . . . . . 102
148 List of Figures
6.6 Eigenvalues for the Jacobian at xed point (ii), i.e. at the state where
cooperators dominate. The value of 1 is illustrated by the black line and
2 by the red one. They only depend on the payo value R, provided that
S = 0, T = 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.7 Oscillations of species frequencies in three Prisoners Dilemma games: (a)

R = 0.8, P = 0.4, (b) R = 0.95, P = 0.05, (c) R = 0.65, P = 0.6. The
corners of a simplex indicate dominance of cooperators (C), defectors (C),
and imitators (I), respectively. Circles denote xed points of the dynamics
a black lling denotes stability of the xed point. Shown are dynamical
trajectories of dierent initial conditions. Generally, the trajectories form
constant orbits around the stable xed point that include the initial con-
dition. Exceptional cases (red and grey trajectories) are discussed in the
text. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.8 Illustrations of dynamical forces at the simplex boundary for a Prisoners

Dilemma with R = 0.8, P = 0.4. (a): positive dierence between the
expected payo of the bilaterally dominant species and the average payo
in the system the quantity that reects the dynamical speed of invasion.
(b): payo dierence between the (potential) expected payo of the non-
existent species and the average payo in the system reecting the invasion
potential of the non-existent species. In both panels, the abscissa denotes
the frequency of the pairwise dominant (invading) species. . . . . . . . . . 106
6.9 Color encoded equilibrium frequencies in the stable xed points: (a) cooper-
ators, (b) imitators, (c) defectors, and (d) individuals that apply the coop-
erative decision at a time, i.e. cooperators plus imitators that meet a coop-
erator plus half of the imitators that meet another imitator (x + x i + i2 /2).
Red color indicates a frequency of 1, blue a frequency of 0. The colorbar of
(d) translates colors into frequencies for all panels (a-d). For the Prisoners
Dilemma (0 < P < R < 1), the displayed xed point is not attractive, but
produces stable oscillations around it. . . . . . . . . . . . . . . . . . . . . . 108
6.10 Equilibrium fraction of cooperators x (up left), imitators i (up right), and
defectors, respectively. Values are averaged from 200 sample runs. The
system size is N = 900 individuals. In the Prisoners Dilemma (0 < P <
R < 1), defectors almost never survive. . . . . . . . . . . . . . . . . . . . . 110
List of Figures 149
6.11 Growing oscillations in single simulation runs that lead to extinction of

2 species. The networks of individuals are fully connected (well mixed)
with system sizes of N = 104 . The game is a Prisoners Dilemma with
R = 0.8, P = 0.4 (S = 0, T = 1). Initial conditions are marked by crosses:
besides the two equal initializations, the other two in the left panel are
x = i = 0.02 and x = 0.39, i = 0.01, respectively. . . . . . . . . . . . . . . . 111
6.12 Representative time evolution of initially equal frequencies of cooperators
(red), imitators (magenta), and defectors (blue) in a numerical simulation of
a 2-dimensional, regular lattice with interactions between 8 nearest neigh-
bors. In contrast to fully connected networks, oscillations do not grow. In
fact, they shrink and lead to a uctuating, but qualitatively stable coex-
istence between cooperators (mostly in majority), imitators, and defectors
(always in minority after a short, initial stage that is amplied in the inset
of the gure). The system size is N = 104 individuals. . . . . . . . . . . . . 112
6.13 Evolutionary trajectories in homogeneous, random networks and 2-dimensional,
regular lattices. k denotes the number of links every individual has in the
network. Every simplex shows one representative simulation run with equal
initial conditions. Panel (a) additionally displays a trajectory for the initial
condition x = 0.9, i = 0.02. . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.1 Classication of symmetrical 22 games according to payo ranking and

system-optimal solutions. Two-dimensionality is achieved by xing T > S
and classifying ordinal dierences only. Parcels separated by solid lines
denote dierent rankings of the payo values. The dashed lines divide the
whole space in 2 regions according to whether partial cooperation is system
optimal (dark-grey background) or not. Social dilemma games have a red
background color. For each area a respective payo matrix (in form of the
left matrix in Eq. (7.1), with T > S) is given. Red payo matrices denote
partial cooperation dilemmas. . . . . . . . . . . . . . . . . . . . . . . . . 122
7.2 Scheme of symmetrical 2 2 games according to Fig. 7.1. A possible
position of the standard cooperation model (right matrix in Eq. (7.1)) is
indicated by the black circle on arrow 2. The dotted arrows 1-4 indicate
transformations of the payo matrices when increasing the parameter w,
i.e. the synergistic/discounting eects of cooperation in the framework of
matrix 7.4. In line with the argumentation of this article, T > S is always
respected. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Bibliography
Abelson, R. (1964). Mathematical models of the distribution of attitudes under contro-

versy. Contributions to Mathematical Psychology 14, 1160.
Altrock, M.; Traulsen, A. (2009). Fixation times in evolutionary games under weak selec-
tion. New Journal of Physics 11, 013012.
Anderson, J. (1991). The place of cognitive architectures in a rational analysis. In: K. van
Lehn (ed.), Architectures for intelligence. Lawrence Erlbaum Associates, pp. 124.
Antal, T.; Scheuring, I. (2006). Fixation of Strategies for an Evolutionary Game in Finite
Populations. Bulletin of Mathematical Biology 68(8), 19231944.
Arrow, K. (1970). Social choice and individual values. Yale University Press.
Asch, S. (1951). Eects of group pressure upon the modication and distortion of judg-
ments. In: Groups, Leadership, and Men. pp. 177190.
Aubin, J. (1979). Mathematical Methods of Game and Economic Theory. New York:
North-Holland Publ. Co. .
Axelrod, R. (1984). The Evolution of Cooperation. New York: Basic Books.
Bandura, A. (1977). Social learning theory. New Jersey: Prentice Hall Englewood Clis.
Behera, L.; Schweitzer, F. (2003). On spatial consensus formation: is the Sznajd model
dierent from the voter model? International Journal of Modern Physics C 14, 1331
1354.
Bergstrom, C.; Lachmann, M. (2003). The Red King eect: When the slowest runner
wins the coevolutionary race. Proceedings of the National Academy of Sciences, USA
100(2), 593.
Berr, M.; Reichenbach, T.; Schottenloher, M.; Frey, E. (2009). Zero-one survival behavior
of cyclically competing species. Physical Review Letters 102(4), 048102.
150
BIBLIOGRAPHY 151
Binney, J.; Dowrick, N.; Fisher, A.; Newman, M. (1993). The Theory of Critical Phenom-
ena: An Introduction to the Renormalization Group. Oxfors: Clarendon Press.
Bonabeau, E.; Dorigo, M.; Theraulaz, G. (1999). Swarm intelligence: from natural to
artificial systems. Oxford University Press, USA.
Bornstein, G.; Budescu, D.; Zamir, S. (1997). Cooperation in Intergroup, N-Person, and
Two-Person Games of Chicken. The Journal of Conflict Resolution 41(3), 384406.
Browning, L.; Colman, A. (2004). Evolution of coordinated alternating reciprocity in

repeated dyadic games. Journal of Theoretical Biology 229.
Cameron, L. (1999). Raising the stakes in the ultimatum game: Experimental evidence
from Indonesia. Economic Inquiry 37(1), 4759.
Castellano, C.; Fortunato, S.; Loreto, V. (2007). Statistical physics of social dynamics.
Arxiv preprint, arXiv:0710.3256.
Castellano, C.; Vilone, D.; Vespignani, A. (2003). Incomplete ordering of the voter model
on small-world networks. Europhysics Letters 63, 153158.
Castello, X.; Eguluz, V. M.; San Miguel, M. (2006). Ordering dynamics with two non-
excluding options: bilingualism in language competition. New Journal of Physics 8,
308.
Chave, J. (2001). Spatial Patterns and Persistence of Woody Plant Species in Ecological
Communities. The American Naturalist 157, 5165.
Choe, J.; Crespi, B. (1997). The evolution of social behavior in insects and arachnids.
Cambridge University Press.
Claussen, J.; Traulsen, A. (2008). Cyclic Dominance and Biodiversity in Well-Mixed

Populations. Physical Review Letters 100(5), 58104.
Clutton-Brock, T. (2002). Breeding Together: Kin Selection and Mutualism in Coopera-

tive Vertebrates. Science 296(5565), 6972.
Cox, J.; Durrett, R. (1991). Nonlinear voter models. In: R. Durrett; H. Kesten (eds.),
Random walks, Brownian motion, and interacting particle systems. Boston: Birkhauser.
Czaran, T.; Hoekstra, R.; Pagie, L. (2002). Chemical warfare between microbes promotes
biodiversity. Proceedings of the National Academy of Sciences, USA 99(2), 786790.
152 BIBLIOGRAPHY
DallAsta, L.; Castellano, C. (2007). Eective surface-tension in the noise-reduced voter

model. Europhysics Letters 77(6), 60005.
Darwin, C. (1859). On the origin of species by means of natural selection or the preserva-
tion of favoured races in the struggle for life. London: J. Murray .
Dawes, C.; Fowler, J.; Johnson, T.; McElreath, R.; Smirnov, O. (2007). Egalitarian
Motives in Humans. Nature 446, 794796.
Dawes, R. (1980). Social Dilemmas. Annual Review of Psychology 31(1), 169193.
Dawkins, R. (1989). The Selfish Gene. Oxford University Press.
De Martino, A.; Giardina, I.; Tedeschi, A.; Marsili, M. (1999). Generalized minority games
with adaptive trend-followers and contrarians. Physical Review E 70, 025104.
Deuant, G.; Neau, D.; Amblard, F.; Weisbuch, G. (2000). Mixing beliefs among inter-
acting agents. Advances in Complex Systems 3(1-4), 8798.
Dittrich, P.; Liljeros, F.; Soulier, A.; Banzhaf, W. (2000). Spontaneous group formation
in the seceder model. Physical Review Letters 84(14), 32053208.
Doebeli, M.; Hauert, C. (2005). Models of cooperation based on the Prisoners Dilemma
and the Snowdrift game. Ecology Letters 8(7), 748766.
Doebeli, M.; Hauert, C.; Killingback, T. (2004). The Evolutionary Origin of Cooperators
and Defectors. Science 306(5697), 859862.
Dornic, I.; Chate, H.; Chave, J.; Hinrichsen, H. (2001). Critical Coarsening without
Surface Tension: The Universality Class of the Voter Model. Physical Review Letters
87(4), 045701.
Duncan, S. (1972). Some Signals and Rules for Taking Speaker Turns in Conversations.
Journal of Personality and Social Psychology 23(2), 283292.
Einstein, A. (1934). On the method of theoretical physics. Philosophy of science , 163169.
Eisenfuhr, F.; Weber, M. (2002). Rationales Entscheiden. Springer.
Epstein, J.; Axtell, R. (1996). Growing artificial societies: Social science from the bottom
up. Cambridge: MIT Press.
Fehr, E.; Bernhard, H.; Rockenbach, B. (2008). Egalitarianism in young children. Nature
454, 1079.
BIBLIOGRAPHY 153
Fehr, E.; Fischbacher, U. (2003). The nature of human altruism. Nature 425, 785791.
Fehr, E.; Gachter, S. (1998). Reciprocity and economics: the economic implications of
Homo Reciprocans. European Economic Review 42, 845859.
Fehr, E.; Gachter, S. (2002). Altruistic punishment in humans. Nature 415, 137140.
Fehr, E.; Schmidt, K. (1999). A Theory Of Fairness, Competition, and Cooperation.

Quarterly Journal of Economics 114(3), 817868.
Field, J.; Cronin, A.; Bridge, C. (2006). Future tness and helping in social queues. Nature
441, 214217.
Fishbein, M.; Ajzen, I. (1975). Belief, attitude, intention, and behavior: an introduction
to theory and research. Addison Wesley Publishing Company.
Fontes, L. R.; Isopi, M.; Newman, C. M.; Stein, D. L. (2001). Aging in 1D discrete spin
models and equivalent systems. Physical Review Letters 87(11), 110201.
Frachebourg, L.; Krapivsky, P. L. (1996). Exact results for kinetics of catalytic reactions.
Physical Review E 53(4), 30093012.
Frean, M.; Abraham, E. (2001a). A voter model of the spatial prisoners dilemma. IEEE
Transactions on Evolutionary Computation 5(2), 117121.
Frean, M.; Abraham, E. (2001b). Rock-scissors-paper and the survival of the weakest.
Proceedings of the Royal Society B: Biological Sciences 268(1474), 13231327.
French, J. (1956). Formal Theory of Social Power. Psychological Review 63, 181194.
Friedman, D. (1991). Evolutionary Games in Economics. Econometrica 59(3), 637666.
Fuchs-Heinritz et al., W. (1994). Lexikon zur Soziologie. Opladen: Westdeutscher Verlag.
Fudenberg, D.; Tirole, J. (1991). Game Theory. MIT Press.
Galam, S. (2002). Minority Opinion Spreading in Random Geometry. European Physical

Journal B 25, 403406.
Galam, S. (2004). Contrarian deterministic eects on opinion dynamics: the hung elec-
tions scenario. Physica A: Statistical Mechanics and its Applications 333, 453460.
Galam, S. (2005). Local dynamics vs. social mechanisms: A unifying frame. Europhysics
Letters 20, 705.
154 BIBLIOGRAPHY
Galam, S.; Gefen, Y.; Shapir, Y. (1982). Sociophysics: a new approach of sociological
collective behavior. J. Math. Sociology 9(1).
Galeotti, A.; Goyal, S. (2007). Games of Social Inuence. Economics Discussion Papers,
http://ideas.repec.org/p/esx/essedp/639.html, downloaded: 2008/09/12.
Gardner, A.; West, S.; Barton, N. (2007). The relation between multilocus population
genetics and social evolution theory. The American Naturalist 169(2), 207226.
Gigerenzer, G.; Selten, R. (2001). Bounded rationality: The adaptive toolbox. MIT Press,
Cambridge, MA.
Gigerenzer, G.; Todd, P. (1999). Simple heuristics that make us smart. New York: Oxford
University Press.
Gintis, H. (2005). Moral Sentiments and Material Interests: The Foundations of Cooper-
ation in Economic Life. MIT Press, Cambridge, MA.
Grafen, A. (1984). Hamiltons rule OK. Nature 318, 310311.
Granovetter, M. (1978). Threshold models of collective behavior. American Journal of

Sociology 83(6), 14201443.
Guare, J. (1990). Six degrees of separation: A play. New York: Vintage Books.
Gunton, J. D.; San Miguel, M.; Sahni, P. S. (1983). Phase Transitions and Critical
Phenomena. London: Academic Press. C. Domb and J. Lebowitz, Eds.
Guth, W. (1999). Spieltheorie und okonomische (Bei) Spiele. Springer.
Hamilton, W. (1963). The Evolution of Altruistic Behavior. The American Naturalist

97(896), 354356.
Hamilton, W. (1964). The genetical evolution of social behaviour. Journal of Theoretical

Biology 7(1), 1752.
Harsanyi, J.; Selten, R. (1988). A General Theory of Equilibrium Selection in Games.

MIT Press, Cambridge, MA.
Hauert, C. (2002). Eects of space in 2 2 games. Int. J. Bifurcat. Chaos 12, 15311548.
Hauert, C.; De Monte, S.; Hofbauer, J.; Sigmund, K. (2002). Volunteering as Red Queen
Mechanism for Cooperation in Public Goods Games. Science 296(5570), 11291132.
BIBLIOGRAPHY 155
Hauert, C.; Michor, F.; Nowak, M.; Doebeli, M. (2006). Synergy and discounting of
cooperation in social dilemmas. Journal of Theoretical Biology 239(2), 195202.
Hauert, C.; Szabo, G. (2005). Game theory and physics. American Journal of Physics
73, 405.
Hegselmann, R.; Krause, U. (2002). Opinion Dynamics and Bounded Condence Models,
Analysis, and Simulation. Journal of Artifical Societies and Social Simulation (JASSS)
5(3).
Helbing, D. (1992). A mathematical model for behavioral changes by pair interactions. In:
G. Haag; U. Muller; K. Troitzsch (eds.), Economic evolution and demographic change.
Formal models in social sciences. Berlin: Springer, pp. 330348.
Helbing, D. (1993). Stochastische Methoden, nichtlineare Dynamik und quantitative Mod-

elle sozialer Prozesse. Shaker.
Helbing, D. (1995). Quantitative Sociodynamics: Stochastic Methods and Models of Social

Interaction Processes. Springer.
Helbing, D.; Buzna, L.; Johansson, A.; Werner, T. (2005a). Self-organized pedestrian
crowd dynamics: Experiments, simulations, and design solutions. Transportation science
39(1), 1.
Helbing, D.; Farkas, I.; Vicsek, T. (2000). Simulating Dynamical Features of Escape Panic.
Nature 407, 487490.
Helbing, D.; Schonhof, M.; Stark, H.-U.; Holyst, J. (2005b). How individuals learn to take
turns: Emergence of alternating cooperation in a congestion game and the prisoners
dilemma. Advances in Complex Systems 8, 87.
Henrich, J.; Boyd, R.; Bowles, S.; Camerer, C.; Fehr, E.; Gintis, H.; McElreath, R. (2001).
In search of homo economicus: Behavioral experiments in 15 small-scale societies. Amer-
ican Economic Review , 7378.
Henrich et al., J. (2003). Group Report: The Cultural and Genetic Evolution of Human
Cooperation. In: P. Hammerstein (ed.), Genetic and Cultural Evolution of Cooperation.
MIT Press.
Hofbauer, J.; Sigmund, K. (1998). Evolutionary Games and Population Dynamics. Cam-
bridge University Press.
156 BIBLIOGRAPHY
Holley, R. A.; Liggett, T. M. (1975). Ergodic Theorems for Weakly Interacting Innite
Systems and the Voter Model. The Annals of Probability 3(4), 643663.
Holt, C.; Roth, A. (2004). The Nash equilibrium: A perspective. Proceedings of the
National Academy of Sciences, USA 101(12), 39994002.
Holyst, J.; Kacperski, K.; Schweitzer, F. (2001). Social impact models of opinion dynam-
ics. In: D. Stauer (ed.), Annual Review of Computational Physics. Singapore: World
Scientic, vol. IX, pp. 253273.
Huyck, J.; Battalio, R.; Beil, R. (1990). Tacit coordination games, strategic uncertainty,
and coordination failure. American Economic Review 80, 234248.
Imhof, L.; Fudenberg, D.; Nowak, M. (2005). Evolutionary cycles of cooperation and
defection. Proceedings of the National Academy of Sciences, USA 102(31), 10797.
Imhof, L.; Fudenberg, D.; Nowak, M. (2007). Tit-for-tat or win-stay, lose-shift? Journal
of Theoretical Biology 247(3), 574580.
Kacperski, K.; Holyst, J. (1997). A simple model of social group with a leader. Journal
of Technical Physics 38(2), 393396.
Kagel, J.; Kim, C.; Moser, D. (1996). Fairness in ultimatum games with asymmetric
information and asymmetric payos. Games and Economic Behavior 13, 100110.
Kagel, J.; Roth, A. (1995). The handbook of experimental economics. Princeton: Princeton
University Press.
Kaplan, T.; Rue, B. (2007). Which Way to Cooperate. Available at SSRN:

http://ssrn.com/abstract=580903, (downloaded March 10, 2009).
Kennedy, J.; Eberhart, R.; Shi, Y. (2001). Swarm intelligence. San Francisco: Morgan
Kaumann and Academic Press.
Kerr, B.; Riley, M.; Feldman, M.; Bohannan, B. (2002). Local dispersal promotes biodi-
versity in a real-life game of rock-paper-scissors. Nature 418(6894), 1714.
Kirman, A. (1993). Ants, rationality, and recruitment. The Quarterly Journal of Eco-
nomics 108(1), 137156.
Kreps, D. (1990). Game theory and economic modelling. Oxford: Clarendon Press.
Kreps, D.; Milgrom, P.; Roberts, J.; Wilson, R. (2001). Rational cooperation in the nitely
repeated prisoners dilemma. Journal of Economic Theory 27, 245252.
BIBLIOGRAPHY 157
Kurzban, R.; Houser, D. (2005). Experiments investigating cooperative types in humans:

A complement to evolutionary theory and simulations. Proceedings of the National
Academy of Sciences, USA 102(5), 1803.
Landau, D.; Binder, K. (2005). A Guide to Monte Carlo Simulations in Statistical Physics.
Cambridge: Cambridge University Press.
Latane, B. (1981). The psychology of social impact. American Psychologist 36(4), 343
356.
Latane, B. (1996). Dynamic social impact: The creation of culture by communication.

The Journal of Communication 46(4), 1325.
Latane, B.; Wolf, S. (1981). The Social Impact of Majorities and Minorities. Psychological
Review 88(5), 43853.
Lehmann, L.; Keller, L. (2006). The evolution of cooperation and altruism. A general
framework and classication of models. Journal of Evolutionary Biology 19, 13651376.
Liggett, T. M. (1995). Interacting Particle Systems. New York: Springer.
Lorenz, J. (2007). Continuous Opinion Dynamics under Bounded Condence: A Survey.

International Journal of Modern Physics C 18(12), 120.
Luce, R.; Raia, H. (1957). Games and decisions. Wiley New York.
Macy, M.; Flache, A. (2002). Learning Dynamics in Social Dilemmas. Proceedings of the
National Academy of Sciences, USA 99(10), 72297236.
Maslov, S.; Sneppen, K.; Alon, U. (2003). Correlation proles and motifs in complex
networks. In: S. Bornholdt; H. Schuster (eds.), Handbook of graphs and networks. From
the genoma to the internet. Wiley VCH and Co.
Mason, W.; Conrey, F.; Smith, E. (2007). Situating social inuence processes: Dynamic,
multidirectional ows of inuence within social networks. Personality and Social Psy-
chology Review 11(3), 279.
May, R. (1987). More evolution of cooperation. Nature 327(6117), 1517.
Maynard Smith, J. (1982). Evolution and the Theory of Games. Cambridge, MA: Cam-
bridge University Press.
Maynard Smith, J.; Price, G. (1973). The logic of animal conict. Nature 246, 1518.
158 BIBLIOGRAPHY
Maynard Smith, J.; Szathmary, E. (1995). The major transitions in evolution. Oxford:
W.H. Freeman.
Milgram, S. (1967). The small world problem. Psychology today 2(1), 6067.
Milgram, S. (1974). Obedience to authority: An experimental view. Tavistock Publications

Ltd.
Miller, N.; Dollard, J. (1998). Social learning and imitation. Routledge.
Moran, P. (1962). The statistical processes of evolutionary theory. Oxford: Clarendon

Press.
Nash, J. (1950). Equilibrium point in n-person games. Proceedings of the National Academy
of Sciences, USA 36, 4849.
Nash, J. (1951). Noncooperative Games. Annals of Mathematics 54, 286295.
Neill, D. (2003). Cooperation and coordination in the turn-taking dilemma. In: Proceedings
of the 9th conference on Theoretical Aspects of Rationality and Knowledge. New York:
ACM Press, pp. 231244.
Nowak, M. (2006a). Evolutionary dynamics: Exploring the equations of life. Harvard

University Press.
Nowak, M. (2006b). Five Rules for the Evolution of Cooperation. Science 314(5805),
15601563.
Nowak, M.; May, R. (1992). Evolutionary games and spatial chaos. Nature 359, 826829.
Nowak, M.; Sigmund, K. (1990). The evolution of stochastic strategies in the Prisoners
Dilemma. Acta Applicandae Mathematicae 20(3), 247265.
Nowak, M.; Sigmund, K. (1993). A strategy of win-stay, lose-shift that outperforms tit-
for-tat in the Prisoners Dilemma game. Nature 364, 5658.
Nowak, M.; Sigmund, K. (1998). Evolution of indirect reciprocity by image scoring. Nature
393, 573.
Nowak, M.; Sigmund, K. (2004). Evolutionary Dynamics of Biological Games. Science

303, 793.
Osborne, M.; Rubinstein, A. (1994). A Course in Game Theory. MIT Press.

BIBLIOGRAPHY 159
Perc, M.; Szolnoki, A. (2008). Social diversity and promotion of cooperation in the spatial
prisoners dilemma game. Physical Review E 77, 11904.
Pingle, M. (1995). Imitation versus rationality: An experimental perspective on decision

making. Journal of Socio-Economics 24(2), 281315.
Queller, D. (1985). Kinship, reciprocity, and synergism in the evolution of social behavior.
Nature 318, 366367.
Rapoport, A. (1967). Exploiter, leader, hero, and martyr: the four archetypes of the 2
times 2 game. Behavioral Science 12(2), 814.
Rashotte, L. (2007). Social Inuence. The Blackwell Encyclopedia of Social Psychology ,

562563.
Rasmusen, E. (2001). Readings in Games and Information. Blackwell Publishers.
Ravasz, M.; Szabo, G.; Szolnoki, A. (2004). Spreading of families in cyclic predator-prey
models. Physical Review E 70(1), 012901.
Redner, S. (2001). A guide to first-passage processes. Cambridge: Cambridge University

Press.
Reichenbach, T.; Mobilia, M.; Frey, E. (2007). Mobility promotes and jeopardizes biodi-
versity in rock-paper-scissors games. Nature 448(7157), 10461049.
Riolo, R.; Cohen, M.; Axelrod, R. (2001). Evolution of cooperation without reciprocity.
Nature 414, 441443.
Roca, C.; Cuesta, J.; Sanchez, A. (2006). Time Scales in Evolutionary Dynamics. Physical
Review Letters 97(15), 158701.
Roca, C.; Cuesta, J.; Sanchez, A. (2008). Sorting out the eect of spatial structure on the
emergence of cooperation. Arxiv preprint, arXiv:0806.1649.
Samuelson, L. (1997). Evolutionary Games and Equilibrium Selection. MIT Press, Cam-
bridge, MA.
Sanfey, A.; Rilling, J.; Aronson, J.; Nystrom, L.; Cohen, J. (2003). The neural basis of
economic decision-making in the ultimatum game. Science 300(5626), 1755.
Santos, F.; Pacheco, J.; Lenaerts, T. (2006). Evolutionary dynamics of social dilemmas in
structured heterogeneous populations. Proceedings of the National Academy of Sciences,
USA 103(9), 34903494.
160 BIBLIOGRAPHY
Santos, F.; Santos, M.; Pacheco, J. (2008). Social diversity promotes the emergence of
cooperation in public goods games. Nature 454, 213216.
Schelling, T. (1978). Micromotives and macrobehavior. New York: Norton.
Schlag, K. (1998). Why imitate, and if so, how? A boundedly rational approach to
multi-armed bandits. Journal of Economic Theory 78, 130156.
Schuler, R. (1986). The evolution of reciprocal cooperation. In: A. Diekmann; P. Mit-

ter (eds.), Paradoxical effects of social behavior. Essays in honor of Anatol Rapoport.
Heidelberg, Wien: Physica-Verlag, pp. 105121.
Schweitzer, F.; Behera, L. (2009). Nonlinear voter models: the transition from invasion to
coexistence. The European Physical Journal B 67(3), 301318.
Schweitzer, F.; Behera, L.; Muhlenbein, H. (2002). Evolution of cooperation in a spatial

prisoners dilemma. Advances in Complex Systems 5, 269299.
Selten, R.; Schreckenberg, M.; Pitz, T.; Chmura, T.; Wahle, J. (2004). Experimental
investigation of day-to-day route-choice behaviour and network simulations of autobahn
trac in North Rhine-Westphalia. In: M. Schreckenberg; R. Selten (eds.), Human
Behaviour and Traffic Networks. Springer, pp. 712.
Shor, M. (2005). Strategy. Available at http://www.gametheory.net/dictionary/strategy.html,

(downloaded May 14, 2011).
Simon, H. (1955). A behavioral model of rational choice. The Quarterly Journal of

Economics , 99118.
Simon, H. (1956). Rational choice and the structure of the environment. Psychological
Review 63(2), 12938.
Simon, H. (1972). Theories of bounded rationality. In: Decision and organization. Ams-
terdam: North Holland, pp. 161176.
Skyrms, B. (2004). The stag hunt and the evolution of social structure. Cambridge Uni-
versity Press.
Slanina, F.; Lavicka, H. (2003). Analytical results for the Sznajd model of opinion forma-
tion. European Physical Journal B 35(2), 279288.
Smith, V. (1991). Rational choice: The contrast between economics and psychology.
Journal of Political Economy , 877897.
BIBLIOGRAPHY 161
Sood, V.; Redner, S. (2005). Voter Model on Heterogeneous Graphs. Physical Review
Letters 94(17), 178701(4).
Stark, H.-U. (2010). Dilemmas of partial cooperation. Evolution 64(8), 24582465.
Stark, H.-U.; Helbing, D.; Schonhof, M.; Holyst, J. (2008a). Alternating cooperation
strategies in a route choice game: Theory, experiments, and eects of a learning scenario.
In: A. Innocenti; P. Sbriglia (eds.), Games, Rationality and Behavior. Houndmills and
New York: Palgrave MacMillan, pp. 256273.
Stark, H.-U.; Tessone, C. (2010). Imitation catalyzes cooperation in three social dilemmas.
In preparation.
Stark, H.-U.; Tessone, C. J.; Schweitzer, F. (2008b). Decelerating microdynamics can ac-
celerate macrodynamics in the voter model. Physical Review Letters 101(1), 018701(4).
Stark, H.-U.; Tessone, C. J.; Schweitzer, F. (2008c). Slower is Faster: Fostering Consensus
Formation by Heterogeneous Inertia. Advances in Complex Systems 11(4).
Stauer, D.; Sousa, A.; de Oliveira, S. (2000). Generalization to Square Lattice of Sznajd
Sociophysics Model. International Journal of Modern Physics C 11(6), 12391246.
Strogatz, S. (1994). Nonlinear dynamics and chaos. Cambridge, MA: Perseus Publishing.
Suchecki, K.; Eguluz, V. M.; San Miguel, M. (2005a). Conservation laws for the voter
model in complex networks. Europhysics Letters 69, 228234.
Suchecki, K.; Eguluz, V. M.; San Miguel, M. (2005b). Voter Model Dynamics in Complex
Networks: Role of Dimensionality, Disorder, and Degree Distribution. Physical Review
E 72, 036132.
Sugden, R. (2001). The evolutionary turn in game theory. Journal of Economic Method-
ology 8(1), 113130.
Suzumura, K. (1983). Rational choice, collective decisions, and social welfare/Kotaro

Suzumura. Cambridge: Cambridge University Press.
Szabo, G.; Fath, G. (2007). Evolutionary games on graphs. Physics Reports 446(4-6),
97216.
Szabo, G.; Toke, C. (1998). Evolutionary prisoners dilemma game on a square lattice.
Physical Review E 58(1), 6973.
162 BIBLIOGRAPHY
Szolnoki, A.; Perc, M.; Szabo, G.; Stark, H.-U. (2009). Impact of aging on the evolution
of cooperation in the spatial prisoners dilemma game. Physical Review E 80, 021901.
Szolnoki, A.; Szabo, G. (2007). Cooperation enhanced by inhomogeneous activity of

teaching for evolutionary Prisoners Dilemma games. Europhysics Letters 77(3), 30004.
Tanimoto, J. (2008). What initially brought about communications? BioSystems 92(1),

8290.
Tanimoto, J.; Sagara, H. (2007). A study on emergence of alternating reciprocity in a 2

2 game with 2-length memory strategy. BioSystems 90(3), 728737.
Taylor, C.; Nowak, M. (2007). Transforming the dilemma. Evolution 61(10), 22812292.
Taylor, P.; Jonker, L. (1978). Evolutionarily stable strategies and game dynamics. Math-
ematical Biosciences 40(2), 14556.
Tessone, C. J.; Toral, R.; Amengual, P.; Wio, H. S.; San Miguel, M. (2004). Neighborhood
models of minority opinion spreading. European Physical Journal B 39, 535544.
Traulsen, A.; Nowak, M. (2006). Evolution of cooperation by multilevel selection. Pro-

ceedings of the National Academy of Sciences, USA 103(29), 1095210955.
Traulsen, A.; Nowak, M. (2007). Chromodynamics of Cooperation in Finite Populations.

PLoS ONE 2(3).
Traulsen, A.; Pacheco, J.; Nowak, M. (2007). Pairwise comparison and selection temper-
ature in evolutionary game dynamics. Journal of Theoretical Biology 246(3), 522529.
Trivers, R. (1971). The Evolution of Reciprocal Altruism. The Quarterly Review of Biology
46(1), 3557.
Trotter, W. (2005). Instincts of the Herd in Peace and War. New York: Cosimo Inc.
Tversky, A.; Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases.
Science 185(4157), 11241131.
Tversky, A.; Kahneman, D. (1986). Rational choice and the framing of decisions. Journal
of Business 59(S4), 251.
Van Kampen, N. (2007). Stochastic processes in physics and chemistry. North Holland.
Von Neumann, J.; Morgenstern, O. (1944). Theory of Games and Economic Behaviour.
Princeton: Princeton University Press.
BIBLIOGRAPHY 163
Wasserman, S. (1994). Social network analysis: Methods and applications. Cambridge:

Cambridge University Press.
Watts, D.; Strogatz, S. (1998). Collective dynamics of small-world networks. Nature 393,
440.
Weidlich, W. (1971). The statistical description of polarization phenomena in society.

British Journal of Mathematical and Statistical Psychology 24(2), 251266.
Weidlich, W.; Haag, G. (1983). Concepts and Models of a Quantitative Sociology: The
Dynamics of Interacting Populations. Berlin: Springer-Verlag.
West, S.; Grin, S.; Gardner, A. (2007). Social semantics: altruism, cooperation, mu-
tualism, strong reciprocity, and group selection. Journal of Evolutionary Biology 20,
415432.
Wingreen, N.; Levin, S. (2006). Cooperation among Microorganisms. PLoS Biology 4(9),
e299.
Zimbardo, P. (2008). The Lucifer effect: Understanding how good people turn evil. Random
House Inc.

Stark H.U. 2012 - Decision Making in Models of Social Interaction

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Stark H.U. 2012 - Decision Making in Models of Social Interaction

Hochgeladen von

Copyright:

Verfügbare Formate

Diss. ETH No.

Decision making in models

A dissertation submitted to the

for the degree of

accepted on the recommendation of

Prof. Dr. Dr. Frank Schweitzer, examiner

Prof. Dr. Dirk Helbing, co-examiner

1 Modeling and analyzing social interaction 3

2 Decisions through social influence 12

3 Decision making in (evolutionary) game theory 23

II Models and analyses 52

5 Slower is faster: Fostering consensus formation by heterogeneous inertia 53

6 Social influence in social dilemmas 84

7 Dilemmas of partial cooperation 116

Appendix: Transformations of the standard cooperation model 140

kooperierenden Individuen. Da Kooperierende jederzeit anderen Individuen auf ihre eige-

Modeling and analyzing social

Interaktion, soziale, [1] die durch Kommunikation (Sprache, Symbole,

Lexikon zur Soziologie (Fuchs-Heinritz et al., 1994)

1.1 Micromotives and macrobehavior

Thomas C. Schelling (Schelling, 1978)

We apply agent-based computer modeling techniques to the study of hu-

Joshua M. Epstein and Robert Axtell (Epstein and Axtell, 1996)

1.2 The Weidlich-Haag approach

socio-conguration, and its change modeled by equations of motion. An important

Albert Einstein (Einstein, 1934)

Things should be as simple as possible, but not simpler.

1.3 Dynamical fixed points and their stability

dimensions. To evaluate stability of a xed point, we consider a small perturbation of a

1.3.1 A one-dimensional flow

1.3.2 Multi-dimensional flows

Here, we assume n dynamical variables x1 , ..., xn and their evolution functions

x1 = f1 (x1 , ..., xn ) (1.2)

Decisions through social influence

2.1 Decision theory

round must not inuence the decision.

Independence of irrelevant alternatives

In our decision situations, we exclusively consider interdependent decisions, i.e. the

2.2 Models of social influence

The following general statement, offered as a summary of many prior stud-

making occurs because it economizes on decision-making resources.

Mark Pingle (Pingle, 1995)

Social influence is defined as change in an individuals thoughts, feeling,

The Blackwell Encyclopedia of Social Psychology (Rashotte, 2007)

This is related to observational- or social learning in social psychology (Bandura,

(a) the pattern of connectivity and inuence assumed among individuals,

(b) the treatment of attitudes or behavioral responses as continuous versus discrete,

(c) the presence or absence of individual dierences in private information, and

2.3 The voter model

M (t) = A(t) B(t). (2.2)

A(t + 1) A(t) = W V (+1| 1, t) B(t) W V (1| + 1, t) A(t)

(a) t = 0 (b) t = 1200

(c) t = 3600 (d) t = 4600

(e) t = 6900 (f) t = 7100

Decision making in (evolutionary)

3.1 Classical game theory

3.1.1 Definitions and terminology

involving learning mechanisms). By common denition:

A strategy denes a set of moves or actions a player will follow in a given

Dictionary of game theory (Shor, 2005)

Decision making by social inuence is a subset of general strategies, where infor-

A straight-forward example of games fullling the denitions of game theory is the

There is a substantial dierence between cooperative- and non-cooperative game theory. In

Simultaneous vs. sequential games