This paperbound volume presents 12 papers, plus a short introduction, selected from the 60 delivered at the European Social Simulation Association meeting in Toulouse, France in Fall 2007. As a result the papers highlight the great range of problems being addressed, and the disciplines now engaging, in social simulation. Methodologically though the papers- all emphasizing agent-based modeling (ABM) - are somewhat more uniform. I arrange these comments according to my perception of the main goals for the models presented, drawing on Epstein's (2008) list.
Illuminate Core Uncertainties or Core Dynamics
From the Tang Dynasty times until 1904, passing a series of intensive examinations was required to enter and rise through China's bureaucracy. Setsuya Kurahashi and Takao Terano present a model which cleverly uses detailed historical information about the family tree of one particularly successful family over successive generations to train ABM parameters so that they reproduce the strategy of this family. It appears that the effects of previous success by the grandfather and the mother - enabling them to pass along 'cultural capital' to potential candidates - was especially important to success in these exams.
César García Díaz et al. investigate whether changing the number of product variants differentially gives advantages to firms of different size, a problem of long standing in the discipline of organizational ecology. This paper draws on well-developed methods and a sophisticated literature to which it contributes by representing the product space in a more flexible manner. Experiments are conducted to determine the differential success when firm expansion is costless (large-scale firms take over) versus costly (where small-scale firms do better as dimensionality - more product variants - increases, and in resource spots where the scale dominance of large firms does not compensate for the cost of expansion). Other experiments are conducted to determine the relationship of innovation to firm size, costs of expansion, and costs of opening a new position. In general their results seem to agree with those developed using other (non-ABM) approaches. Unlike these existing approaches, however, their success in endogenizing product dimensionality as a function of firm interaction allows them to address problems of mechanism obscured by less fine-grained representations.
Samuel Thiriot and Jean-Daniel Kant focus on how formalizing agents' beliefs and messages as associative networks benefits models for the diffusion of innovations or products. The term "associative networks" appears to encompass both social networks (among individuals), and representations of beliefs about products ("salient social objects") and private beliefs that agents have about themselves, that are internal to each agent. Beliefs (for example about products) can be revised according to various sources of information which are qualitatively categorized as more or less credible; advertisements for example have zero credibility whereas personal experiences have complete credibility. An interesting feature of the model is its taxonomy for channels of communication among agents which includes both traditional mass media and the "bi-directional" channels represented by, say, blogs offering the possibility of feedback. The model is applied to adoption of the iPod, which in turn leads to suggestions for improving the speed of diffusion, and to admission of the difficulty of knowing how well one is representing social networks and what provokes social salience for an item in the model.
Armando Geller and Scott Moss contribute a very interesting and well-written chapter on how the power structures called Qawm grow in the anomic society of contemporary Afghanistan. They define 10 kinds of agents - including commanders, drug dealers, farmers, and organized criminals -who have complementary needs and abilities. An endorsement scheme among, for, and of agents biases affiliations (network links) among agents, including patron-client relationships. This model is designed to mimic what they call the "neopatrimonial" behavior strategies in one specific state, but seems to me to point to interestingly general ways of creating groups, given other agent definitions fitting other reference systems.
Discover New Questions/Focus on Common Questions
The last article in the collection continues the theme of how groups are made from individuals, but in the weaker sense of recognizing "solidarity" that can lead one actor to perform an action that helps another even when that is at some cost to himself. Pieter Bots and colleagues recognize several dimensions to social solidarity that allow them to analytically divide actions of solidarity into six different types. They then employ a modeling platform called So-Si-So to study the conditions under which urban and rural water users can arrive at solidarity rather than conflict and deadlock. The initial explorations they present here were designed to allow participants in a workshop to arrive at a common set of questions with respect to such interactions; Epstein's taxonomy fails me here but this is a little like his category of discovering new questions. The agent architecture is very complicated but in overview is of the BDI type, and their initial results suggest that this architecture should be able to support additional work leading eventually to policy recommendations.
Train Practitioners and Challenge the Robustness of Prevailing Theory Through Perturbations
Marta Posada and colleagues begin their paper with the claim that MABS (multiagent-based simulation) is a "killer application" for extending experimental economics with human subjects, and end it with the suggestion that proper instruction in MABS should be a prerequisite for research in economic theory. In between they investigate which of two famous macroconcepts - Marshallian or Walrasian models of market adjustment - better corresponds to outcomes computed from the micro-level in non-standard environments (downward-sloping supply and upward-sloping demand) as well as in standard environments. They conclude - as have apparently others - that Marshallian (but not Walrasian) stability is achieved in continuous double auctions so long as the agents form beliefs and act according to the strategies specified by Gjerstad and Dickhaut (1998); another agent learning rule employed did not achieve price convergence to an equilibrium.
Discipline the Policy Dialogue
Three articles (at least) contribute to the field of agent-based computational economics and have some policy implications - though only one has explicit policy-relevant goals. Zach Lewkovicz and Jean-Daniel Kant translated a previous 'economic' model (given by a system of equations) for the introduction of a new job contract into the labor market into an ABM - a process that they term "agentification". There are four agent types: workers, companies, a matching agent representing the labor market, and a government agent. Among the problems they faced, and solved, was a need to re-specify the matching function (now, the behavior of the matching agent), which showed abnormal and unrealistic oscillations when presented with the less-bounded conditions available to it in the agent environment. That problem fixed, the ABM reproduced the main results of the previous model. So why bother with agentification? The advantage, according to the authors, is that the increased realism of agent interactions allows researchers, or policy makers, to see the interactions from which the rates (also given by the 'economic' model) actually arise. It moreover allows for measuring emergent phenomena (my terms, not their's) such as the precariousness of workers' positions and the volatility of the labor market.
Expose Prevailing Wisdom as Incompatible with Available Data
Widad Guechtouli built and discusses an ABM that leads to the counter-intuitive finding that agents learn more efficiently when they know nothing about other agents' competencies in advance, but learn that over time. This result appears to be due to the nature of information transmission among agents in the model, which is one-to-one, which in turn leads to a "congestion effect" around the most competent agents when their identities are known in advance. Presumably this effect would decrease or disappear given one-to-many communication architectures - but this effect might still be of interest for those of us studying small-scale, pre-modern societies with limited opportunities for one-to-many transmission.
Demonstrate Trade-offs/Suggest Efficiencies
In one of the volume's longest contributions, Rosaria Conte and her colleagues report on recent work in a long-running project modeling reputation. In short overview, that does not do justice to the richness of the materials explored, the strategy here is to differentiate between image and reputation. Images are evaluative beliefs about someone based on direct interaction, held by an agent; reputations include what I might think of you based on our interactions, but also exist on the population level and are subject to transmission effects not applying to images. "Reputation is an objective social property that emerges from a propagating cognitive representation, which lacks an identified source, whereas image always requires at least one evaluator to be identified" (p. 306). An ABM explores these concepts in a market environment in which sellers turn over regularly, making information about their reliability both valuable and scarce. Reputation networks clearly would scale better than image networks as the number of interacting agents increases, but what's of interest in this chapter is more the propensity of the image networks to perform poorly because of cycles of retaliation, which is avoided in the reputation networks. The authors provocatively speculate that in evolution human societies may have found a way to move from networks of images to networks of reputation.
Other (or all of the above)
One of the things I conclude from this review is that Epstein's (1998) list of motivations for modeling, though very long, is still incomplete. The essay by Yang and Gilbert, for example, reports the issues raised, and the approaches taken, in attempting to develop an ABM from qualitative, 'ethnographic' data that honors the meanings that participants apprehend in social situations. Although they do not present any of their results, their discussion of the issues they encountered may be helpful for those considering similar endeavors. This seems to me to represent a methodological goal: expanding the range of situations in which ABM might be useful.
Extending usual ABM practice in another direction, Smajgl and colleagues present a model in which agents can develop new actions to govern their own behaviors and can also develop new rules at the system level, thereby influencing their peers' behaviors. Actions consist of an aim and a condition; separating the aim into two parts, a verb and an object ("trade/water quota") allows for innovation in the way the verbs and objects are combined. Rules are more complicated than actions, since they also specify a scope ("attribute"); a "deontic" specifying for example whether the relationship is must [not] or may [not]; and an "or else" clause. Agents develop new actions or rules in an attempt to use resources more efficiently. An illustrative or conceptual implementation of their approach can be run as a Java applet, or downloaded in Netlogo form, from: http://www.luis.izquierdo.name/models/endoRules. Endogenizing rule change greatly enriches the usual sense of 'emergence' in agent-based modeling, and the provision of the model as either code or applet in this paper should be an example for us all. Should these approaches ever be expanded to evolutionary game theory the results might be too complicated to contemplate - or very interesting.
I'm sure Epstein would agree that his modeling goals are not mutually exclusive. How for example can one "explain" without "illuminating core dynamics"? Consider for example the paper by Sebastiano Delre and colleagues who explore the effects of two kinds of social influence (imitation and coordinated consumption) on the differential fortunes of mainstream and art-house movies. (Coordinated consumption here refers to the act of viewing a movie with a group of friends.) The authors rely on a great deal of carefully collected survey information from moviegoers which they collected in three different countries (and frankly the empirical part of this paper is the most interesting, confirming as it does my prior prejudices about the differing nature of the audiences for these two types of movies) to calibrate an ABM. They then explore how differing weights for the coordinated consumption and imitation effects in turn affect the Gini coefficient measuring degree of differential attendance at movies. Not too surprisingly, higher social orientations in agents lead to higher Gini coefficients (greater differences in attendance rates across movies) since, in that case, the choices of consumers converge on just a few movies that become hits. Their contribution simultaneously explains why big studio producers prefer a platform to a sleeper strategy in introducing movies, and illuminates the core dynamics in consumer behavior that make this approach advantageous.
Given the diversity of its contents not many people will want to read this from front to back, but many will profit from consulting individual contributions, and this special issue presents a useful snapshot of the state-of-the-art in social simulation as practiced primarily in Europe in 2007. If this collection is representative, I would characterize the social simulation efforts in Europe as being generally more applied in focus, more interested in exploring the consequences of increasing agents' cognitive sophistication and elaborating subtleties of their social interactions- and therefore generally oriented towards understanding micro-level interactions - and less interested in evolutionary dynamics, than are its American counterparts. Think Viennese coffeehouse vs. the Wild West. From such tensions, I anticipate that useful and interesting advances will emerge.