Content uploaded by Justin E. Lane
Author content
All content in this area was uploaded by Justin E. Lane on Sep 21, 2015
Content may be subject to copyright.
© Equinox Publishing Ltd. 2014, Unit S3, Kelham House, 3 Lancaster Street, Sheffield S3 8AF
[JCSR 1.2 (2013) 161–180] Journal for the Cognitive Science of Religion (print) ISSN 2049-7555
doi:10.1558/jcsr.v1i2.161 Journal for the Cognitive Science of Religion (online) ISSN 2049-7563
Method, Theory, and Multi-Agent Articial Intelligence:
Creating computer models of complex social interaction
1
Justin E. LanE
Institute for Cognitive and Evolutionary Anthropology, LEVYNA, Masaryk
University and University of Oxford
justin.lane@anthro.ox.ac.uk
AbstrAct
The construction of computer models is becoming an increasingly useful and
popular way of testing theories in the cognitive sciences. This paper will present
a brief overview of the methods available for constructing and testing computer
models of social phenomena such as religious beliefs and behaviors. It will fo-
cus on the importance of theoretical continuity and data replication in computer
modelling while negotiating the relationship between specicity and ecologi-
cal validity when models are extended into novel contexts. This paper will ar-
gue that computer modeling is an important supplement to the methodological
toolbox of cognitive scientists interested in human social phenomena. However,
this is only the case if developers pay close attention to research methods and
theories and if the method of a model’s development is appropriate for the target
phenomenon (Sun, 2006). It concludes that multi-agent AI models are the most
appropriate computational tool for the study of complex social phenomena.
Keywords
agent based modeling, social psychology, cognitive science
Introduction
Within the cognitive sciences, computer modeling has been a valuable tool
for understanding complex phenomena. The advent of computers, with their
ability to quickly execute immensely complex logical statements, has allowed
researchers to create theoretical abstractions with many variables that utilize
1. The author would like to thank Don Braxton and an anonymous reviewer for comments on
earlier drafts.
162 Justin E. Lane
© Equinox Publishing Ltd. 2014
interacting, iterative, and recursive processes. While computer modeling pre-
dates the onset of the cognitive approach to the mind, it was not until the 1980s
that this approach started to focus on social processes (Axelrod 1984). This
paper argues that researchers interested in complex social phenomena, such as
religion and culture, should utilize computer modeling in order to reconstruct
the phenomena that researchers disassemble, and that agent based modeling is
currently the best tool for such an endeavor.
Computer modeling is a technique that utilizes programming languages to
create functions that simulate theoretical propositions or real-world phenomena.
Computer modeling involves starting with a set of propositions and designing a
model based on those propositions to test whether some target phenomena are
observed. A computer model can also start by taking a target phenomenon and
deriving a more simplied process to achieve that target phenomenon. This can
be seen in computer models that seek to simulate “emergent” social phenomena
such as ocking behaviors or population levels. A common analogy is that com-
puter models are much like maps. A map that showed the location of every rock
and blade of grass would be so complex that it would be useless. However, if it
failed to show major geographical features or major roads the map would also
be useless (see Meadows, Randers and Meadows 2004). A good model utilizes
an efcient number of relevant variables and produces valid results that resem-
ble the phenomena of interest, as observed in the real world. This sometimes
leads to the misconception that computer models can predict the future. This is
not necessarily the case. Models can, in ‘broad sweeps’, generate information
that can give perspective on probable outcomes given the parameters of the
model (Meadows, Randers and Meadows 2004). The information that is gener-
ated can be anything from an optimized set of parameters under which a certain
phenomenon is most likely to be observed or an expected outcome given the
model’s settings and theoretical propositions.
Nielbo, Braxton and Upal (2012) have recently noted that computer modeling
has been largely neglected in the cognitive science of religion. This is true in
many respects. However, this should be of little surprise as there are very few
theories of religion utilized within the subeld of the cognitive science of reli-
gion (CSR) that make claims specic enough to be amenable to empirical test-
ing and computer modeling. This is not to say that there are too few theories of
religion. On the contrary, there are a plethora of theories of religion; however,
the majority of these theories are only interested in either the behavioral or
belief aspects of religion but not their relationship(s) or they do not delineate
specic falsiable hypotheses concerning the motivational factors that underlie
behavioral observations that could be operationally dened as religious. While
Method, Theory, and Multi-Agent Articial Intelligence
163
© Equinox Publishing Ltd. 2014
this research community has generated an impressive body of empirical, ethno-
graphic, and philosophical work, there are only a handful of theories that can
serve as a basis for modeling religious systems. This is because such a theory
should be specic enough to generate empirically testable hypotheses yet main-
tain its ability to effectively describe the variety of observable target behaviors.
Of the many theories of religion, the three most promising scientic theories
are: Stark and Bainbridge’s compensator theory of religion (1987/1996); Law-
son and McCauley’s competence theory (1990; McCauley and Lawson 2002);
and Whitehouse’s divergent modes of religiosity (DMR) theory (2004). These
three theories, unlike many other theories of religion, offer empirically testable
hypotheses, are theoretically compatible with lower-levels of scientic inquiry,
and are broad enough to address the diversity of beliefs and behaviors observed
in the ethnographic and historical record. Furthermore, they are specic enough
to generate data that can be tested against historical and future datasets. What
is also unique about these three theories is that they incorporate a multi-level
approach to religious and/or cultural systems, albeit in different ways (i.e. these
theorists acknowledge that religion is a social—level 1—phenomenon that arises
due to individual—level 2—cognitive mechanisms that are biologically—level
3—constrained).
The Theories of CSR
These three theories all have, in their own ways, deciencies that must be
addressed. While there have been lengthy discussions about the strengths and
weaknesses of these theories elsewhere2 only a few will be addressed here. Stark
and Bainbridge’s theory of religion stands out because it has been utilized for
computer models of religious behaviors more than any other. In brief, their theory
offers a system of interactions based on sociological and cognitive ndings that
rely on seven axioms concerning how humans interact in the world in order to
seek greater rewards and avoid losses (Stark and Bainbridge 1996). From these
axioms, they deduce over 300 propositions about religious behaviors. Indeed,
Bainbridge’s (2006) monograph God from the Machine: Articial Intelligence
Models of Religious Cognition, is responsible for the largest number of models
directly addressing religion in the eld. However, its major deciency is the
overwhelming utilization of rational choice theory. The latter theory has been
largely subsumed in cognitive science by prospect theory (Kahneman and Tver-
sky 2000), which proposes an asymmetrical relationship whereby losses incur
2. See those presented in McCauley and Lawson (2002), Whitehouse (2004), Whitehouse and
Martin (2004), Whitehouse and McCauley (2005)-specically Barrett’s contribution (2005),
Bainbridge (2006), Bruce (1993), Wallis and Bruce (1984), and Stark (1999).
164 Justin E. Lane
© Equinox Publishing Ltd. 2014
a greater weighting than gains, rather than assuming a symmetrical weighting
between gains and losses.
The second cognitive theory of religion was presented by E. Thomas Law-
son and Robert N. McCauley (1990) in their book entitled Rethinking Religion.
Their theory, inspired by Chomsky’s formal approach to linguistics, posits
mechanistic relationships between agents, objects, actions, and their attributes
(such as connections to a culturally postulated super-natural agent). These enti-
ties are processed by the standard mental facilities that are shared by all typical
human populations. This is a strict cognitive approach to religious systems in
that it is an information processing paradigm. The biggest deciencies in Law-
son and McCauley’s competence theory are a lack of experimental testing and
under-dened roles of objects/instruments in the ritual form hypothesis. The
research concerning the claims of Lawson and McCauley (Barrett 2002; Bar-
rett and Lawson 2001; Malley and Barrett 2003; Sørensen, Liénard and Feeny
2006) has relied on explicit measures and focused specically on efcacy and
competence; that is to say, whether a ritual would successfully produce the
intended outcome(s) and if the individuals involved were sufciently competent
to bring about the desired ritual effects. These studies have conrmed Lawson
and McCauley’s claim that agents are of primary importance to the perception
that a ritual will be efcacious. However, Malley and Barrett’s (2003) study of
the intuitions of real-world rituals remains one of the only studies that has tested
this theory using ethnographic data.3
Whitehouse’s theory of divergent modes of religiosity (DMR), on the other
hand, is more reliant on complex interactions between social and psychological
variables that commonly co-occur in ritual systems. The DMR theory has been
subject to experimental testing, ethnographic application, and historical com-
parison. One key prediction of the DMR is that over time ritual practices will
fall into one of two attractor positions: either high sensory pageantry-low fre-
quency or low-sensory pageantry-high frequency ritual practices (Whitehouse
2004). Recent tests of the theory utilizing an ethnographic database support this
claim (Atkinson and Whitehouse 2011). While the DMR has weathered intense
critique (see Whitehouse and Martin 2004 and Whitehouse and McCauley 2005)
from ethnographic and historical specialists, its largest downfall for such appli-
cations is that it is a diachronic theory; i.e. the DMR predicts that over time a
ritual system will settle near one of the attractor points within the phase-space of
possible ritual actions. Since historians and anthropologists often deal with spe-
3. Barrett (2002) also relies heavily on ethnographic data in his discussion and both Barrett
(2002) and Barrett and Malley (2003) found results that are open to further research.
Method, Theory, and Multi-Agent Articial Intelligence
165
© Equinox Publishing Ltd. 2014
cic events or relatively small periods of time, their data cannot always address
dynamic changes over long timespans. This often creates issues with historians
who typically specialize in a specic context, where it is likely that not all of
the features predicted by the DMR will be in their predicted state (either present
or not-present depending on the tradition’s ritual mode) at that specic point in
time. However, not having all features predicted by DMR at one point does not
provide evidence of falsication of the predictions of the DMR because it does
not incorporate the time dynamic of the theory’s predictions.
Modeling in CSR
Multi-agent modeling provides a platform for building a testing-ground for the
predictions and comparisons of all three cognitive theories of ritual behavior.
While there are a number of different approaches to computer modeling, this
paper argues that multi-agent articial intelligence (MAAI) is the most appro-
priate computational method for testing theories and generating predictions and
data when looking at complex social phenomena.
Although debates concerning the denition of a religion have left the humani-
ties without a single, agreed-upon denition, cognitive scientists have more or
less agreed that a religion is a socially instantiated system of beliefs and behav-
iors that is reliant on references to supernatural agents. The social nature of the
target phenomena (i.e. religions or their specic beliefs and behaviors) renders
purely statistical models without adequate explanatory power. Although repro-
ducing statistical trends can lead to well-t mathematical descriptions of the
growth and decline of many phenomena, as culturally held beliefs and behav-
iors are iterated throughout the individuals in a population, any mathematical
abstraction at the population level leads to a Durkheimian notion of explanation
whereby it is “in the nature of society itself that we must seek the explanation of
social life” (Durkheim 1895:128). Utilizing a population level variable as both
the dependent and independent variable might provide rich descriptions of pop-
ulation level dynamics, but does not provide an immediate path to lower level
explanations. Even those mathematical models that apply dynamical systems
theory (such as Abrams, Yaple, and Wiener 2011) to create statistical models of
secularization overlook those cognitive mechanisms that seem to make religion
a very “natural” belief (such as those discussed in McCauley 2011) and assume
an overly simplistic view of religious systems.
Recently, computational models have attempted to use articial neural net-
works (ANN) to study cognition. ANNs can be great tools for classication and
prediction. For instance, using ANNs for classifying different aspects of rituals
or beliefs can lead to a certain level of understanding of how these variables may
166 Justin E. Lane
© Equinox Publishing Ltd. 2014
tend to operate. Furthermore, ANNs are very useful in simulating learning over
time and could be a great method for understanding how information comes to
be learned by individuals in religious systems. Using machine learning has been
helpful in understanding physiological processes of cognition (Sun 2006) and
this could serve some purpose in CSR, but only at the level of the individual or
physiological systems within the individual. Ultimately, such an approach on its
own cannot say much about social phenomena because these approaches do not
allow for interactions between entities or their simulated environments.
The most promising individual based modeling approach involves research
programs that aim to create cognitive architectures. Cognitive architectures
attempt to model a variety of the cognitive systems of an agent such as learn-
ing, recall, affect, and sensory-motor systems (Sun and Fleisher 2012). These
architectures represent an integrated approach whereby mechanisms are not
modeled in isolation. The creation of cognitive architectures and understand-
ing the complex relationships between cognitive mechanisms may be a crucial
step in understanding the aspects of social phenomena that cognitive scientists
have been studying through experimentation for many years. By modeling such
systems we might learn a lot about the interactions between different cognitive
mechanisms that go into the production of social beliefs and behaviors. How-
ever, unless these systems are embedded in agents that are able to interact with
other agents and their environment then they have not gone far enough to model
the social nature of religious groups. What is needed is a way of integrating the
individual cognitive models and the population level models. It is argued here
that multi-agent articial intelligence (MAAI) is the best tool we currently have
for such a research project.
Agent based modeling or (ABM) is a computational modeling technique
whereby individual agents, programmed with a specic set of rules for inter-
actions, are allowed to interact with and change their environment as well as
other agents within the model. These models are useful for the study of many
phenomena from the interaction of particles in physics and organic compounds
in chemistry to biological organisms and machines operating in swarms. As a
subset of ABM, MAAI would represent those agent-based models where the
individual agents of the model are imbued with psychologically realistic capa-
bilities. In ABM, the rules constraining an agent’s actions can be purely theoreti-
cal propositions or based on some empirical data. They can have personal vari-
ables that can represent beliefs, motivations, biological functions such as age and
fertility, identities, or social ties. A crucial difference between MAAI and ABM
would be that variables such as those listed here are not always modeled with
a focus on imbuing agents with psychologically realistic capabilities in ABMs
Method, Theory, and Multi-Agent Articial Intelligence
167
© Equinox Publishing Ltd. 2014
that are not MAAI focused (e.g. beliefs don’t change or interact, non-realistic
reasoning mechanisms are employed to govern motivations or behavioral discus-
sions, identity is a tag for grouping rather than something that informs beliefs and
behaviors). During the simulations, agents are free to interact with each other in
ways that might change the values of other agent’s variables, at times giving rise
to population level phenomena. These agents can also interact with their envi-
ronment. This can include resources (water and food), resource production rates,
and physical landscapes such as city streets or terrain. An agent’s interaction
with the environment can also change the environment. For instance, if an agent
must eat and drink a set amount during a specic time period, then the amount
of food and water available in the environment would change. Subsequently, in
certain conditions the resources might become scarce and unable to support a
population. At this point the agents will begin to die off until they reach a stable
population. This scenario is very similar to the predator-prey (Berryman 1992) or
Sugarscape models (Epstein and Axtell 1996) utilized in the past. Clearly, such
a multi-level approach towards a subject is complex, and such complex dynam-
ics may create difculty in pinpointing the causal mechanism(s) within a model.
History, Anthropology, and Cognitive Science
The ability for agents to dynamically interact by utilizing psychologically real-
istic complex behaviors is what makes MAAI the most amenable to modeling
social phenomena in such a way that researchers can help to explain cultural
and religious phenomena. In the cognitive science of religion, researchers are
primarily engaged with three types of data that are useful and compatible with
MAAI approaches: historical data, anthropological data, and experimental data.
Historical data
Historical research is of the utmost importance to computer modeling within
CSR. Simply put, historians have the most data concerning cultural and reli-
gious beliefs. If the output of any computer model does not sufciently address
or replicate historical data, the model is unlikely to be of much use. This is
no different from the role historians have played in cognitive science thus far.
Although experimental data is the primary focus of any cognitive research pro-
gram, if at no point in time can simulated—or experimental—data relate to what
has happened in the past, it is fair to say that the research may be of internal
interest to computer or cognitive science but may have little power to explain
social phenomena.
A number of recent social simulations have been able to reproduce data from
the historical record. Some compelling examples include the model of the Kay-
168 Justin E. Lane
© Equinox Publishing Ltd. 2014
enta Anasazi of the Long House Valley, which is able to reproduce historical
population levels (Axtell et al. 2006), settlement patterns (Gumerman et al.
2006), and resource activities (Dean et al. 2006) with impressive accuracy. The
model of welfare-state politics presented by Schumacher and Vis (2012) utilizes
real world data as its input and its output replicates historical patterns in Euro-
pean politics. Also, an agent based simulation of political party sizes by Muis
(2010) produces historically valid results for the sizes of Dutch political parties
between 1998 and 2002. The work of Bainbridge (2006) has also utilized a
multi-agent approach to simulate past sociological data in regards to the distri-
bution, size, and success of different religious groups. Lastly, a model developed
by Wildman and Sosis (2011) utilized the hypotheses within the costly signal-
ing research (Iannaccone 1992; Sosis and Bressler 2003) as expressed within
a paradigm of cultural evolution (Heinrich 2009). More recently, Nielbo et al.
(2012) provided an overview and defense of the use of computer modeling in
the study of religion. In their paper, Nielbo et al. presented a number of models
that pertain to religion at a number of different levels. They gave an overview of
one multi-agent model that produces trends in religious extremism that seem to
t well with historical data.
These models are all impressive examples of the use of computer modeling
to test theories with the incorporation of historical data in at least some part of
the model. What remains to be done in many of these instances is to update the
theoretical propositions outlined in the model with empirical data. For exam-
ple, the work of Iannaccone (1992) and Bainbridge (2006; Stark and Bainbridge
1987)—and by extension the work of Upal (2005)—utilizes theories premised
on rational choice utility models, which have been largely subsumed by the heu-
ristics and biases research programs, particularly that of behavioral economics
(Ariely 2010; Kahneman and Tversky 2000). This is not to say that these mod-
els are not useful. On the contrary, the models of Iannaccone and Makowsky
(2007), Bainbridge (2006), and Upal (2005) are some of the highest quality mod-
els currently available to CSR researchers and show the advantage of agent-
based approaches. However, without grounding the model in experimental data
whereby the individual agents are making decisions in the way we know humans
to do (under controlled conditions), researchers/developers risk modeling their
assumptions to produce their expectations. For example, if the expectation is that
a socially held belief will be more frequent in a population because it is in some
way “tness increasing” and they then assume that belief X from population A
is increasing more than belief Y from population B, without prior empirical evi-
dence that this is the case, their model that belief X will prosper or population B
will decline is more a formalization of their assumptions than it is a representa-
Method, Theory, and Multi-Agent Articial Intelligence
169
© Equinox Publishing Ltd. 2014
tion of real-world belief dynamics. Further complications also arise when coding
historical and ethnographic data. The recent rise in the use of large collections
of historical data (such as in the case of Atkinson and Whitehouse 2011) is a
promising resource for cognitive scientists; however, careful attention must be
paid to the differences in coding when quantifying archival data. For instance,
two historians or ethnographers may use the same term to describe extremely
different levels of emotional salience in a ritual. Alternatively, two human coders
may perceive the same description of a ritual quite differently. Those utilizing the
database should be careful to check to what extent information in the archival
records resulted in multiple or conicting data points in the nal dataset (archival
research may be very rich but often not specic in relaying information about
their topic of interest; or multiple scholars report incompatible information for
the same data point). Making coding schemes open and available is imperative
for the use of historical information in computer modeling as it provides a basis
for debate and discussion concerning the data.
Anthropological data
Anthropological data is extremely useful in developing models of religious
behavior. Many of the research centers in CSR are embedded within anthropol-
ogy departments. Research centers such as these are useful because they have
been carrying out eldwork specically to test the predictions of theories gener-
ated within CSR.
A number of multi-agent computer models have utilized anthropological
data. One such model, by Younger (2010), simulates resource allocation among
households with different leadership styles in Pacic island societies. A more
recent simulation, which directly addresses a prominent theory in CSR was
presented by Whitehouse, Kahn, Hochberg and Bryson (2012). This utilized a
multi-agent model of motivation and tedium, which was based on the theoretical
propositions of Whitehouse’s (2004) DMR theory, to replicate the ethnographic
ndings of Whitehouse’s (1995) study of a Kivung splinter group. This simula-
tion was particularly interesting because it is one of the only multi-agent models
to utilize a theory from CSR (notable exceptions being Braxton, 2008, where a
model derived from the ritual form hypothesis was presented, and to an extent
Bainbridge, 2006, where the constituent hypotheses of Stark and Bainbridge,
1987, were the focus of simulations). Furthermore, the Whitehouse et al. (2012)
model represented the clearest delineation of the doctrinal mode of the DMR in
over a decade (since Whitehouse 2004). By specically describing the theoreti-
cal propositions of the DMR in the NetLogo language, clear descriptions of the
theory and its proposed mechanisms were outlined.
170 Justin E. Lane
© Equinox Publishing Ltd. 2014
Anthropological models can fall victim to the same issues as historical mod-
els. Although the models can reproduce observed phenomena, the rules of the
agents should be rooted in experimental ndings. This allows for the appropriate
constraints and proclivities to be exercised by the cognitive mechanisms that the
functions within the program are simulating. Biases in efcacy perceptions of rit-
uals adhering to the different forms specied within the ritual competence model
(Lawson and McCauley 1990; McCauley and Lawson 2002) have received
empirical support compatible with models that reproduce ethnographic infor-
mation (Barrett 2002; Barrett and Lawson 2001; Sørensen, Liénard and Feeny
2006). The semantic memory effects modeled by Whitehouse et al. (2012) are
well established in the psychological literature and are seen throughout religious
populations (Whitehouse 2004; Boyer 2001). Another issue for anthropological
models is that they may be too specic. Although many models are very accurate
at reproducing the specics of an observed socio-cultural event, they should also
be able to reproduce other events found in the ethnographic record. Models such
as the Whitehouse et al. (2012) model are a great starting point for creating more
generalized models because of their reliance on theory, openness to revision (as
seen in the responses to Whitehouse et al. 2012), and foundation in empirical
testing of theoretical claims. Assuming it is well constructed, if a simulation
produces extremely specic results it is likely that revising the model such that
it has fewer constraints and more variability should make it more generalizable.
Experimental data
Experimental data is the most readily available for MAAI modeling. Unlike his-
torical or ethnographic studies, experiments yield data trends that can be math-
ematically tested for parameters and likelihoods, which can be applied to logical
or mathematical abstractions. These formalisms can be rendered in computer
languages without much struggle. Furthermore, well constructed experiments
use sampling methods that should reect general trends throughout a popula-
tion, making the information useful to those seeking to model groups of people.
There are many cognitive processes that have been modeled using computer
based tools, some of which use multi-agent platforms. Such models represent
examples of ‘psychologically realistic cognitive agents’ (Sun and Hélie 2013)
that are crucial to an accurate understanding of socio-cognitive phenomena.
Of interest to many in CSR are those models that address the cognitive sys-
tems that are currently believed to help produce religious beliefs and behaviors.
Some of the systems currently under investigation by computer modelers look at
the interactions between agents in social networks, how they share knowledge,
and what happens under different social and knowledge structures. Given the
Method, Theory, and Multi-Agent Articial Intelligence
171
© Equinox Publishing Ltd. 2014
application of this research for understanding real-world social groups, some
researchers, such as Schrieber and Carley (2004), have argued for empirical
validation of their simulations and models. Others, such as Gilbert and Troitzsch
(2005), have argued that experimental data needs to be utilized in the creation of
the agent’s rules from the start. One such system is that of implicit and explicit
knowledge interactions. Within cognitive science, a great deal of research has
attempted to model the experimental ndings concerning implicit and explicit
knowledge interactions. Within CSR specically, the effects of these systems
are used by Whitehouse (2004) to help explain the diversity of religious systems
observed throughout history. The work of Ron Sun, who is one of the largest
proponents of the use of computer modeling to understand social phenomena
(Sun 2006, 2012), on the CLARION cognitive architecture has helped research-
ers understand these interactions from a computational perspective (Sun, Zhang,
Slusarz, and Matthews 2007) and it has been applied within a multi-agent frame-
work to help understand tribal survival strategies (Sun and Fleischer 2012). This
work was particularly useful because it demonstrated the importance of cogni-
tive, social, and environmental conditions for the success of tribal strategies
and stands as an example of exactly the type of multi-level simulation based
research program that this paper supports.
Models of experimental data alone run the risk of missing the big picture.
Over the past two decades there has been an increasing body of research that
seeks to explain religious behavior using experiments that address one facet of
religious cognition at a time. This data is extremely useful but these experiments
by themselves do not explain religious behavior and neither would a model
of any one experimental nding. Rather, these ndings need to be utilized to
inform the creation of complex and ‘psychologically realistic cognitive agents’
that can interact in social environments. Experimental data may help to explain
the production of religious beliefs and behaviors but computer modeling may
help it overcome a limited reductionist perspective. Experimental data relies
on the responses of individuals under tightly controlled conditions. A notable
exception to lab-based experimentation is “experimental anthropology”, a grow-
ing sub-eld that utilizes experimental approaches and quantied measurements
in the eld; examples being the research programs of Dimitris Xygalatas (see
Xygalatas 2012 for a full overview and Konvalinka et al. 2011 and Xygalatas et
al. 2013 for more specic examples) and the MIT Human Dynamics Laboratory
(Directed by Prof. Alex Pentland).4 While lab experiments have been a valuable
4. Approaches such as these could grow with the advent of mobile and wearable technology
assuming there also is appropriate software available for research applications. Some tech-
nologies, such as smartphones (example: Eagle, Pentland, Lazer and Hanson 2009), have
172 Justin E. Lane
© Equinox Publishing Ltd. 2014
addition to the corpus of knowledge in CSR, there are serious questions of eco-
logical validity when the results of experiments are applied to real-world reli-
gious behaviors. Furthermore, ethical and logistical considerations arise when
trying to design experiments that could overcome these issues. Computer mod-
eling—like cognitive historiography—provides cheap research subjects and is
not constrained by the same ethical considerations as human subject research. It
also allows us to study those social or environmental situations that might be too
rare to study with any other method (Fischer and Kronenfeld 2011).
An integrated approach
The issues with historical, anthropological, and experimental approaches to reli-
gion is that they do not sufciently explain the dynamics of religious groups on
their own. Any research program that wishes to explain religious patterns needs
to take into account all three of these data types within a cognitive framework.
It is argued here that MAAI computer modeling is currently the only avail-
able tool that can hope to create a theoretically continuous, methodologically
sound research program to explain religious dynamics. Such a research program
would seek to utilize a theory of religious belief and behavior that allows for the
variability recorded throughout the historical and ethnographic records and is
specic enough to generate experimentally testable hypotheses. Such a theory
would also have to admit the social nature of religious phenomena. It could then
utilize the theoretical premises as rules that govern interactions and constrained
dynamics within the model, as informed by experimental data. Furthermore, it
would be able to sufciently reproduce target phenomena from the historical
and ethnographic record.
There are a number of new computational tools that are being used in cognitive
science more broadly that can help inform the creation of research programs such
as the one generally supported here. One tool is network science (Baronchelli et
al. 2013). Network science affords CSR researchers the techniques to quantify
and statistically analyze complex relationships such as social networks that are
integral, but overlooked, aspects of religious cognition. Network science also
allows us to analyze the complex relationships between concepts represented in
texts, interviews, sermons, and online religious materials (Lane 2013). The quan-
titative nature of this analysis allows researchers to integrate qualitative datasets
into computer-based models. This ultimately can allow for a model’s input to
rely on real-world materials. Its output can then be tested for validity against data
been shown to be useful while others, such as Google-glass and the Emotiv EPOC (see
Ekandem et al. 2012), have not been denitively shown to be reliable tools for eld based
data collection.
Method, Theory, and Multi-Agent Articial Intelligence
173
© Equinox Publishing Ltd. 2014
collected from the relevant context. Such an approach was utilized by White-
house et al. (2012) and is being continued by the author of this paper. Integrating
network science into the already existing research programs in CSR will allow
for researchers to admit that their subjects of research are complex without hav-
ing to sacrice a commitment to quantitative hypothesis testing.
Another computational tool available to researchers in CSR is quantitative
text analysis (QTA). QTA is designed to analyze texts, interviews, and tran-
scripts to create a co-occurrence matrix of relevant terms of interest in a text.
Because this analysis is automated and computer based, it allows for repeat-
ability and transparency between researchers. Furthermore, the efciency of the
QTA process allows for massive amounts of text to be analyzed and compared
for different quantitative measures (such as network topography) or qualita-
tive measures (such as frequently co-occurring terms). This approach allows
for historical materials such as texts and inscriptions to be analyzed and com-
pared to other available materials. QTA is also useful for analyzing transcripts.
This allows for a single methodology to analyze surveys and self-reports from
experimental data, texts and inscriptions from historical data, and transcripts
from ethnographic data.
Quantitative text analysis opens up a currently overlooked data pool for
researchers in CSR that dwarfs all aggregate data collected to date: that from the
Internet. Currently, according to Facebook’s Q2 nancial report, Facebook has
1.15 Billion users a month (Facebook 2013) and allows application developers
to utilize the data provided by these users. However, CSR has yet to leverage
this massive data set or the data set available from the roughly 1 Trillion unique
URLs currently indexed on the Internet by Google (Alpert and Hajaj 2008).5
Furthermore, the massive amount of digitized material created before the advent
of the internet allows for the analysis of historical material on an unprecedented
scale (see Aiden and Michel 2014). Naturally, not all of these data opportuni-
ties will be utilized by every researcher engaged in computer modeling and
CSR. However, this dataset is one of the largest samples available to research-
ers. Intelligent use of these data could provide answers to many of the questions
concerning cross-cultural patterns of religious beliefs. This raises a question of
feasibility and logistics. Most religion and anthropology departments do not
have access to the computer hardware necessary to analyze data sets of this size.
However, given the prevalence of cloud-computing options such as NereusV,
Amazon’s EC2 or Microsoft’s Azure and the relatively inexpensive hardware
5. The fact that Google.com does not index all web pages leads the author to believe that this
estimate was low for its time.
174 Justin E. Lane
© Equinox Publishing Ltd. 2014
options for parallel computing (such as the RasberryPi and Adepteva Parallella),
it is quickly becoming less of a logistical impossibility and more of an issue of
priorities and research scope.
Testing models
Computer models output data that can be quite useful for understanding the
model itself, the target domain, or even the lower level processes in the model.
Although some researchers engaged in modeling (such as Bainbridge) argue
that models are akin to deductive proofs, it seems that the growing consensus
is that the output from a computer model itself can be used to generate further
predictions (Nielbo et al. 2012). In a controversial text, Stephen Wolfram (2002)
introduced a lengthy proposal for a “new kind of science” in order to study the
output from simple computational programs known as cellular automata. These
programs iterate simple rules throughout a system and can often produce differ-
ent types of data that can then be studied in order to understand the dynamics of
the program itself. Given how similar this method is to MAAI, it would be quite
reasonable to take this approach to agent based modeling as well. Both systems
involve programming simple rules that are followed and repeated by many inde-
pendent but interactive entities within a computational space. This approach cre-
ates what Epstein calls a ‘generative’ approach to modeling social phenomena
where decentralized and local interactions of autonomous agents can generate
an output that resemble real-world phenomena. Wolfram and Epstein seem in
agreement with the position of Nielbo et al. (2012) that the outputs of computer
models can be empirically investigated and are both tools for exploration as well
as targets for further exploration.
The output of these models can be analyzed in a number of different ways. Nat-
urally, the type of analysis that is undertaken must be appropriate for the inves-
tigation. Different agent based models will require different types of tests. Many
agent based models have many variables that can signicantly affect the output
of a model. For example, a model of a religious group may have variables such
as population size, density of social connections, frequency of ritual participa-
tion, and number of counter-intuitive attributes for information transmitted in a
ritual. A researcher may want to know under what conditions a specic pattern is
produced given a set of rules for the agents and their interactions. They may, for
instance, want to know how a belief becomes spread throughout an entire group
and when a belief will spread throughout an entire population. After describing the
rules for how and when concepts of different counter-intuitiveness are transmitted
and how likely agents are to spread them through their social network one could
run the model under different conditions. For instance, it would be easy to see in
Method, Theory, and Multi-Agent Articial Intelligence
175
© Equinox Publishing Ltd. 2014
a model where no agent has a connection with any other agent that there will be
no transmission of information, producing a non-optimal situation. Alternatively,
it could be also understood that in small groups of densely connected individuals
with more frequent repetition, an entire group could learn and retain minimally
counter-intuitive information. However, there are many settings in this ctitious
model where predictions are not as easy to make. In such circumstances it would
be useful to perform parameter sweeps of the model. Doing so entails exhaust-
ing the reasonable values for each available variable and running the simulation
under all of the relevant conditions. By running each simulation under a stable
state of parameters over and over again one could analyze the output of the model
and see if any trends develop. This technique is particularly useful for models of
stochastic or non-deterministic systems, such as human social systems.
By varying and co-varying all of the relevant variables in a model one can then
investigate the phase space of the model. For instance, if one runs a model that
has a single output vector, researchers can analyze the output of this model for
internal consistencies. Generally, this approach will help to understand how the
model behaves under different conditions. The model of motivation and rein-
vigoration presented by Whitehouse et al. (2012) utilizes a single output vector
(i.e. motivation). Simple revisions show that the model can produce the periodic
increases in motivation found in the historical record (McCorkle and Lane 2012)
known as the “Melanesian Prole” (McCauley and Lawson 2002; Pyysiäinen
2004; Whitehouse 2004). The model’s output shows periodic repetitions of
increased motivation under most conditions. In some circumstances the param-
eters of the model may not be as interesting as the dynamics of the repetition
itself. For questions about how frequent or related these motivational increases
are one could employ a recurrence quantication analysis (RQA). This approach,
although rarely utilized in CSR, has a basis in analyzing behavioral data that
appears to be non-linear at best or chaotic at worst (Riley and Van Orden 2005).
It is possible that when one analyzes all of the possible states of a model, attrac-
tor points develop. Should one develop a robust model of the DMR theory (White-
house 2004) one would expect two attractor points to develop within the phase
space of the model’s output (assuming that the interactions proposed by the DMR
produce the expected outputs). Similar attractor points should also be seen if the
propositions of competence theory are correct (McCauley and Lawson 2002).
Integration and modeling
There are many theories and hypotheses within CSR that can be tested to see
which ts the data best given the standard psychological constraints and proclivi-
ties. Modeling can help us to understand how the individual mechanisms interact
176 Justin E. Lane
© Equinox Publishing Ltd. 2014
with different social and environmental conditions to produce a target phenom-
enon. Given what is known about other complex systems, we should not assume
at this point that by understanding every cognitive mechanism involved in the
production of social behavior we would be able to understand any social behavior
as a sum of those parts. As we continue to collect enough data testing theories of
religious dynamics it is possible that the hypotheses previously conrmed in a
lab setting will help inform models that can address the higher (i.e. social) level
predictions. It has been noted that within computational models of social behav-
iors there is no consensus concerning how researchers should go about construct-
ing their models, but there is an increasing trend towards utilizing a behaviorist
approach (Miller and Page 2007). Although modeling agent behaviors at the level
of the mind is appropriate, behaviorist approaches are largely insufcient in com-
parison to cognitive approaches for many of the same reasons that rational choice
theory was subsumed by the heuristics and biases program. The theories and
researchers within CSR operate in two spaces. One, at the level of the individual
where cognitive mechanisms can be used to make predictions concerning how
individuals will react; two, at the level of a group, where beliefs and behaviors are
found throughout a social group. The way we formalize our models should reect
this multi-level endeavor as well. Given the multi-level nature of CSR’s theories,
it only makes sense to utilize computational tools that can address our research
questions in a way that provides continuity between method and theory with the
same rigor that is demanded by many religious scholars.
Conclusion
To conclude, there are four aspects of design and testing that are necessary for
computer modeling in the cognitive science of social behavior. First, key to
the development of useful computer models is that the method of development
is appropriate for the level of investigation (Sun 2006). Second, for modeling
religion(s) it is argued here that MAAI is the best tool available. Third, research-
ers need to ensure that they are utilizing experimental data to support the theo-
retical propositions in their model. Finally, the data produced by simulations
under specic environmental conditions should produce similar outputs to those
seen in the relevant ethnographic and historical records.
References
Abrams, D., H. Yaple and R. Wiener. 2011 “Dynamics of social group competition:
Modeling the decline of religious afliation.” Physics Review Letters 107(8).
http://dx.doi.org/10.1103/PhysRevLett.107.088701
Aiden, E. and J. Michel. 2014. Uncharted: Big Data as a Lens on Human Culture. New
York: Riverhead Books.
Method, Theory, and Multi-Agent Articial Intelligence
177
© Equinox Publishing Ltd. 2014
Alpert, J. and N. Hajaj. 2008. “Ofcial Blog: We knew the web was big…” [website]
Google.com. Posted by J. Alpert and N. Hajaj, http://googleblog.blogspot.
co.uk/2008/07/we-knew-web-was-big.html
Ariely, D. 2010. Predictably Irrational: The Hidden Forces That Shape Our Decisions.
New York: Harper Perennial.
Atkinson, Q. D. and H. Whitehouse. 2011. “The cultural morphospace of ritual form:
Examining modes of religiosity cross-culturally.” Evolution and human behav-
ior. 32: 50–62. http://dx.doi.org/10.1016/j.evolhumbehav.2010.09.002
Axelrod, R. 1984. The Evolution of Cooperation. New York: Basic Books.
Axtell, R.L., J.M. Epstein, J.S. Dean, G.J. Gumerman, A.C. Swedlund, J. Harburger, S.
Chakravarty, R. Hammond, J. Parker and M. Parker. 2006. “Population growth
and collapse in a multiagent model of the Kayenta Anasazi in Long House Val-
ley.” In Generative Social Science: Studies in Agent-Based Computational Mod-
eling, edited by J.M Eptstein, 117–129. Princeton, NJ: Princeton University Press.
Bainbridge, W. 2006. God from the Machine: Articial Intelligence Models of Religious
Cognition. Lanham, MD: AltaMira Press.
Baronchelli, A., R. Ferrer-i-Chancho, R. Pastor-Satorras, N.Chater and M.H. Christian-
sen. 2013. “Networks in cognitive science.” Trends in Cognitive Science 17(7):
348–360. http://dx.doi.org/10.1016/j.tics.2013.04.010
Barrett, J.L. 2002. “Smart gods, dumb gods, and the role of social cognition in structur-
ing ritual intuitions.” Journal of Cognition and Culture 2(3): 183–193. http://
dx.doi.org/10.1163/15685370260225080
———. 2005. “In the empirical mode: Evidence needed for the modes of religiosity
theory.” In Mind and Religion: Psychological and Cognitive Foundations of
Religiosity, edited by H. Whitehouse and R. N. McCauley, 109–126. Walnut
Creek, CA: AltaMira Press.
———. and E.T. Lawson. 2001. “Ritual intuitions: Cognitive contributions to judg-
ments of ritual efcacy.” Journal of Cognition and Culture 1(2): 183–201.
http://dx.doi.org/10.1163/156853701316931407
Berryman, A. A. 1992. “The origins and evolution of predatory-prey theory.” Ecology,
73(5): 1530–1535. http://dx.doi.org/10.2307/1940005
Boyer, P. 2001. Religion Explained: The Evolutionary Origins of Religious Thought.
New York: Basic Books
Braxton, D. 2008. “Modeling the McCauley-Lawson Theory of Ritual Forms.” [Presen-
tation] Presented to Aarhus University, Aarhus Denmark, in June 2008.
Bruce, S. 1993. “Religion and rational choice: A critique of economic explanations
of religious behavior.” Sociology of Religion 54(2): 193–205. http://dx.doi.
org/10.2307/3712139
Dean, J.S., G.J. Gumerman, J.M. Epstein, R.L. Axtell, A.C. Swedlund, M.T. Parker
and S. McCarroll 2006. “Understanding Anasazi culture change through agent-
based modeling.” In Generative Social Science: Studies in Agent-Based Com-
putational Modeling, edited by J.M Eptstein, 90–116. Princeton, NJ: Princeton
University Press.
178 Justin E. Lane
© Equinox Publishing Ltd. 2014
Durkheim, E. 1982 [1895]. The Rules of Sociological Method. Edited by S. Lukes and
translated by W. D. Halls. New York: The Free Press.
Eagle, N., A. S. Pentland, D. Lazer and S. Hanson. 2009. “Inferring friendship network
structure by using mobile phone data.” Proceeding of the National Academy of
Sciences 106(36): 15274–15278. http://dx.doi.org/10.1073/pnas.0900282106
Ekandem, J.I., T.A. Davis, I. Alvarez, M.T. James, J.E. Gilbert. 2012. “Evaluating the
ergonomics of BCI devices for research and experimentation.” Ergonomics
55(5): 592–598. http://dx.doi.org/10.1080/00140139.2012.662527
Epstein, J. and R. Axtell. 1996. Growing Articial Societies: Social Science from the
Bottom Up. Washington, DC: Brookings Institution Press.
Facebook. 2013. “Facebook Reports Second Quarter 2013 Results.” Facebook.com - In-
vestor Relations, http://investor.fb.com/releasedetail.cfm?ReleaseID=780093
Fischer, G. and D.B. Kronenfeld. 2011. “Simulation (and modeling).” In A Companion
to Cognitive Anthropology, edited by V.C. de Munck, M.D. Fischer, G. Ben-
nardo, and D.B. Kronenfeld, 210–226. Malden, MA: Wiley-Blackwell. http://
dx.doi.org/10.1002/9781444394931.ch12
Gilbert, G. and K. Troitzsch. 2005. Simulation for the Social Scientist. Maidenhead,
England and New York, NY: Open University Press.
Gumerman, G.J., A.C. Swedlund, J.S. Dean and J.M. Epstein. 2006. “The evolution of
social behavior in the prehistoric American southwest.” In Generative Social
Science: Studies in Agent-Based Computational Modeling, edited by J.M Ept-
stein, 130–143. Princeton, NJ: Princeton University Press.
Iannaccone, L. 1992. “Sacrice and stigma: Reducing free-riding in cults, communes,
and other collectives.” Journal of Political Economy 100(2): 271–291. http://
dx.doi.org/10.1086/261818
———. and M.D. Makowsky. 2007. “Accidental atheists? Agent-based explanations
for the persistence of religious regionalism.” Journal for the Scientic Study of
Religion 46(1): 1–16. http://dx.doi.org/10.1111/j.1468-5906.2007.00337.x
Kahneman, D. and A. Tversky, eds. 2000. Choices, values, and frames. Cambridge:
Cambridge University Press.
Konvalinka, I., D. Xygalatas, J. Bulbulia, U. Schjødt, E. Jegindø and S. Wallot. 2011. “Syn-
chronized arousal between performers and related spectators in a re-walking
ritual.” Proceedings of the National Academy of Sciences 108(20): 8514–8519.
http://dx.doi.org/10.1073/pnas.1016955108
Lane, J.E. 2013. “Semantic Network Mapping of Religious Material.” [Conference
Presentation] Society for Complex Systems in Cognitive Science. Berlin, Ger-
many, 30 July 2013.
Lawson, E. and R. McCauley. 1990. Rethinking Religion: Connecting Cognition and
Culture. Cambridge England New York: Cambridge University Press.
Malley, B. and J. Barrett. 2003. “Can Ritual Form Be Predicted From Religious Belief? A
Test of the Lawson-McCauley Hypotheses.” Journal of Ritual Studies 17(2): 1–14.
Method, Theory, and Multi-Agent Articial Intelligence
179
© Equinox Publishing Ltd. 2014
McCauley, R. and E. Lawson. 2002. Bringing Ritual to Mind: Psychological Foun-
dations of Cultural Forms. Cambridge: Cambridge University Press. http://
dx.doi.org/10.1017/CBO9780511606410
McCorkle, W.W. and J.E. Lane. 2012. “Ancestors in the simulation machine: meas-
uring the transmission and oscillation of religiosity in computer modeling.”
Religion Brain and Behavior 2(3): 1–4. http://dx.doi.org/10.1080/215359
9X.2012.703454
Meadows, D, J. Randers and D. Meadows. 2004. Limits to Growth: The 30-Year Up-
date. White River Junction, VT: Chelsea Green Publishing Company.
Nielbo, K.L., D.M. Braxton and A. Upal. 2012. “Computing religion: A new tool in the
multi-level analysis of religion.” Method and Theory in the Study of Religion
24: 267–290. http://dx.doi.org/10.1163/157006812X635709
Pyysiäinen, I. 2004. “Corrupt doctrine and doctrinal revival: On the nature and limits
of the modes theory.” In Theorizing Religions Past: Archaeology, History, and
Cognition, edited by Harvey Whitehouse and Luther H. Martin, 173–194. Wal-
nut Creek, CA: AltaMira Press.
Riley, M. A. and G. C. Van Orden. 2005. Tutorials in Contemporary Nonlinear Methods
for the Behavioral Sciences. Retrieved 1 September 2012, from http://www.nsf.
gov/sbe/bcs/pac/nmbs/nmbs.jsp
Schreiber, C. and K. Carley. 2004. “Going beyond the data: Empirical validation leading
to grounded theory.” Computational and Mathematical Organization Theory
10: 155–164. http://dx.doi.org/10.1023/B:CMOT.0000039168.92625.9d
Schumahcer, G. and B. Vis. 2012. “Why do social democrats retrench the welfare state?
A simulation.” Journal of Articial Societies and Social Simulation 15(3).
http://jasss.soc.surrey.ac.uk/15/3/4.html
Sørensen, J., P. Liénard, and C. Feeny. 2006. “Agent and instrument in judgments of rit-
ual efcacy.” Journal of Cognition and Culture 6(3–4): 463–482. http://dx.doi.
org/10.1163/156853706778554977
Sosis, R and E.R. Bressler. 2003. “Cooperation and commune longevity: a test of the
costly signaling theory of religion.” Cross-Cultural Research 37(2): 211–239.
http://dx.doi.org/10.1177/1069397103037002003
Stark, R. 1999. “Micro foundations of religion: A revised theory.” Sociological Theory
17(3): 264–289. http://dx.doi.org/10.1111/0735-2751.00080
Sun, R., Stark, R. and W. Bainbridge. 1996. A Theory of Religion. New Brunswick, N.J:
Rutgers University Press.
———, ed. 2006. Cognition and multi-agent interaction: From cognitive modeling to
social simulation. Cambridge: Cambridge University Press.
———. ed. 2012. Grounding Social Sciences in Cognitive Sciences. Cambridge, MA:
MIT Press.
———. and P. Fleischer. 2012. “A cognitive social simulation of tribal survival strate-
gies: The importance of cognitive and motivational factors.” Journal of Cog-
nition and Culture 12(3–4): 287–321. http://dx.doi.org/10.1163/15685373-
12342077
180 Justin E. Lane
© Equinox Publishing Ltd. 2014
———. and S. Hélie. 2013. “Psychologically realistic cognitive agents: taking human
cognition seriously.” Journal of Experimental and Theoretical Articial Intel-
ligence 25(1): 65–92. http://dx.doi.org/10.1080/0952813X.2012.661236
———, P. Zhang, P. Slusarz and R. Mathews. 2007. “The interaction of implicit learn-
ing, explicit hypothesis testing learning, and implicit-to-explicit knowledge
extraction.” Neural Networks 20(1): 34–47. http://dx.doi.org/10.1016/j.neu-
net.2006.07.002
Wallis, R. and S. Bruce. 1984. “The Stark-Bainbridge theory of religion: A critical anal-
ysis and counter proposals.” Sociological Analysis 45(1): 11–28. http://dx.doi.
org/10.2307/3711319
Whitehouse, H. 1995. Inside the Cult: Religious Innovation and Transmission in Papua
New Guinea. Oxford: Oxford University Press.
———. 2004. Modes of Religiosity: A Cognitive Theory of Religious Transmission.
Walnut Creek, CA: AltaMira Press.
———, K. Kahn, M. Hochberg and J. Bryson. 2012. “From the imaginary to the real:
the back and forth between reality and simulation.” Religion, Brain and Behav-
ior 2(3): 219–224. http://dx.doi.org/10.1080/2153599X.2012.703455
———, and L.H. Martin, eds. 2004. Theorizing Religion’s Past: Archaeology, History,
and Cognition. Walnut Creek, CA: AltaMira Press.
———, and R.N. McCauley, eds. 2005. Mind and Religion: Psychological and Cogni-
tive Foundations of Religion. Walnut Creek, CA: AltaMira Press.
Wolfram, S. 2002. A New Kind of Science. Champaign, IL: Wolfram Media
Xygalatas, D. 2012. The Burning Saints: Cognition and Culture in the Fire-Walking
Rituals of the Anastenaria. Bristol, CT: Equinox Publishing, Ltd.
———, P. Mitkidis, R. Fischer, P. Reddish, J. Skewes, A.W. Geertz, A. Roepstorff and J.
Bulbulia. 2013. “Extreme rituals promote prosociality.” Psychological Science
424(8): 1602–1605. http://dx.doi.org/10.1177/0956797612472910
Younger, S. 2010. “Leadership in small societies.” Journal of Articial Societies and
Social Simulation 13(3). http://jasss.soc.surrey.ac.uk/13/3/5.html
