1In 2005, the American Journal of Sociology devoted a special issue to agent-based modelling and simulation. In their introduction to the issue, Nigel Gilbert and Andrew Abbott (2005, 110, 4: 859) stated: “[…] the most important changes in social science computation have come in the use of computers to ‘think through’ the implications of human actions within given social structures—action in networks. Such ‘agent-based modeling’ has been applied to everything from the diffusion of norms and innovations to voting.” The same conviction is at the heart of at least fifteen thematic issues (or reports) that several journals, both specialist (in computer science, physics, economics, marketing, geography, and sociology) and generalist and cross-discipline, such as the Proceedings of the National Academy of Sciences and Nature, have decided to devote to agent-based modelling and simulation since 2000. [1] In the French-speaking academic world, the journal Nouvelles Perspectives en Sciences Sociales devoted an issue to the theme of simulation in 2010 (vol. 5, no. 2) and the Revue Française de Sociologie is now embracing the recent developments in quantitative methodology of agent-based modelling and simulation.
2By way of introduction to this special issue, it is important first to recall how and why it came about. The original impetus for the project was to make French sociologists aware of agent-based modelling’s explanatory potential and epistemological issues. The objective was thus not only to discuss it, but also to see how this method might work in practice. The international call for papers sent out in April 2012 thus solicited both applied and metatheoretical contributions. Twenty-nine responses from researchers from institutions in seven different countries were received. Four articles and a review article passed each stage of the review process: an initial selection on the basis of long abstracts, followed by four evaluation phases for the complete texts. In addition, as the guest editor, I was invited by the Revue to write an introductory article. The result is a combination of general and more specialized contributions written by recognized experts in the field of agent-based modelling, who are predominantly young researchers who have been educated (or are working) in disciplines as diverse as sociology, economics, philosophy, and mathematics. It is these texts that I will now introduce.
3The opening article, written by the guest editor, offers a general introduction to agent-based modelling and simulation and aims to act as a guide facilitating the understanding of the more technical developments contained in the other texts in this special issue. To this end, it addresses the following themes: a) the definition and originality of agent-based modelling in comparison to other computer simulation methods; b) the explanatory concept at the heart of agent-based modelling as well as its ability to address the problem of the “micro–macro” transition; c) the technical infrastructure of the method, which is the source of its flexibility; d) criticisms of agent-based modelling and the solutions that can be offered to overcome these difficulties. This introduction opts to address themes that cut across the research field, rather than to undertake a detailed study of specific points. However, for readers who wish to go further, the article also offers a self-study guide to agent-based modelling and simulation in addition to an extensive bibliography.
4Pierre Livet, Denis Phan and Lena Sanders present a detailed analysis of the possible uses of a model studied through agent-based simulation. In particular, they examine to what extent the model is itself the object of analysis, or how it is studied in connection with the empirical data. In describing models from economics, sociology, and geography (and, although more briefly, archaeology), they show that the opposition between agent-based simulation’s function as conceptual exploration and its use with empirical data is an artificial one. As is suggested in particular by the cases where these two practices coexist in the work of the same author (in the same piece of work or at different stages of their career), conceptual exploration and empirical “calibration/validation” of agent-based models represent two equally important and complementary facets of this method. Livet, Phan and Sanders’s text can thus be read as a detailed development of two points addressed in the introductory article that precedes it: the relationship an agent-based simulation may have with the empirical data as well as the importance of sensitivity and robustness analysis.
5This coexistence of empirical data and theoretical exploration can be found in Antonio Casilli, Juliette Rouchier and Paola Tubaro’s work, which proposes a novel application of agent-based simulation to the analysis of online discussions in eating disorder forums. These authors examine the way in which the attitudes of the users suffering from these disorders evolve during the discussions and become more favourable (or, conversely, hostile) to the possibility of (self-)treatment. To answer this question, Casilli, Rouchier and Tubaro use qualitative data (from an analysis of forums, blogs and web pages as well as interviews with users) to design and, in part, calibrate a formal model of opinion diffusion. Assuming an interaction structure where each expressed opinion can potentially influence those of all users, the mechanism at the heart of the model concerns the probability that the opinions of two users converge—or, conversely, diverge more—as a function of the perceived similarity of these opinions. By exhaustively studying the parameters that govern this mechanism, the article shows that, under realistic structural conditions (defined on the basis of qualitative observations), the convergence of opinions towards positions radically opposed to recovery is an uncommon outcome. Casilli, Rouchier and Tubaro thus provide an example of a unique combination of a qualitative approach with agent-based simulation and contribute to enriching an established sub-domain (in this case, the modelling of opinion dynamics) in which there are more abstract models than models directed towards empirical data.
6José A. Noguera, Francisco J. Miguel Quesada, Eduardo Tapia and Toni Llàcer’s article, in contrast, examines a more recent field of application of agent-based simulation. This is the mechanisms that can explain the degree to which actors decide to respect the injunction to pay taxes. The authors situate their analysis within a critical dialogue with rational choice theory and, on the basis that RCT predicts much higher rates of fraud than actually observed, propose a theoretical model in which fiscal contributions are the result of a rational calculation (between financial gain and the probability of punishment), feelings of fairness or unfairness (felt towards the tax system), and the tendency towards imitation (of the fiscal behaviour of those in the actor’s entourage). Noguera and his colleagues use quantitative data from surveys and Spanish government sources to calibrate several elements of the model empirically and set out to use agent-based simulation to explore the theoretical consequences of the model (under the conditions of a random network and a homophilic network, by occupational status and income of the agents). A notable result of this analysis is that the imitation of the fiscal choices of others makes agents less sensitive to punishments, which implies that, when sanctions are weak, and thus when a rational agent would commit more fraud, the “imitating” agent conversely would commit less; whereas, when the sanctions are more severe, and a rational agent therefore commits less fraud, the “imitating” agent in fact commits more. The article “dissects” the simulation to explain the origins of this counterintuitive result to the reader. Thus, at the methodological level, Noguera and his colleagues show that, contrary to what simulation’s detractors claim, agent-based simulated models can be inspected and made intelligible.
7The thematic focus returns to online exchanges with Simone Gabbriellini’s article. Unlike in Casilli, Rouchier and Tubaro’s analysis, Gabbriellini’s objective is not to understand how opinions evolve during discussions but to explain the structure of these discussions: in particular, he collected data from three forums that differed by subject, size and users. He describes the exchanges within these forums in the form of bipartite networks in which one set of nodes represents discussions on a given subject (thread) and the other a set of users: the links between the two sets represent the number of messages sent by a given user on a given subject. Gabbriellini proposes using agent-based simulation to generate a network of virtual exchanges that reproduces the structure of real exchanges on the basis of mechanisms describing how a user decides to send a message on a given subject. According to Gabbriellini’s hypothesis, this decision depends, on the one hand, on the topic on which the actors in a person’s entourage send messages and, on the other, on how active a user is (in particular if he/she is an expert/novice, he/she would tend to send messages wherever there are experts/novices). Each of these mechanisms is based on an empirical calibration and is associated with a theoretical parameter that expresses the intensity with which the mechanism is supposed to operate. Gabbriellini explores the parameters space of the model to find the values of the parameters that produce exchange networks that are the closest to real networks, and he explains why the values found are realistic and thus shed light on the structure of exchanges in the three contexts studied. Thus, Gabbriellini provides the reader with an example of an agent-based model whose input is in part empirically calibrated and whose macroscopic consequences are also confirmed by empirical data. The article also contributes to the development of an interface between agent-based modelling and network analysis, a research field that is still relatively undeveloped (and in terms of bipartite networks, practically nonexistent).
8The review article by Flaminio Squazzoni, current president of the European Social Simulation Association, concludes this issue with a discussion of five publications that are important landmarks in research (based) on agent-based modelling and simulation. These publications are of interest since they cover several disciplines (such as economics, political science, and social psychology) and touch on different approaches and theoretical traditions that coexist within studies based on agent-based modelling. This is important additional information for those seeking their bearings in a complex and continuously developing research field.
9In conclusion, it should be emphasized that this special issue is the fruit of close collaboration. The directors of the Revue, the peer review committee, and the editorial staff worked constantly on its overall organization and to improve its quality. They are all gratefully acknowledged. As the guest editor of this special issue, it goes without saying that I take sole responsibility for any errors in the final results.
Notes
-
[1]
The complete list of references is available on demand from the author.