Edited by: Roberto Cordeschi, Sapienza University of Rome, Italy
Reviewed by: Pietro Terna, University of Torino, Italy; Stefano Balietti, ETH Zurich, Switzerland
*Correspondence: Mario Paolucci, Laboratory of Agent Based Simulation, Institute of Cognitive Science and Technologies, CNR, via Palestro 32, Rome 00185, Italy e-mail:
This article was submitted to Theoretical and Philosophical Psychology, a section of the journal Frontiers in Psychology.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
In the first part of the paper, the field of agent-based modeling (ABM) is discussed focusing on the role of generative theories, aiming at explaining phenomena by growing them. After a brief analysis of the major strengths of the field some crucial weaknesses are analyzed. In particular, the generative power of ABM is found to have been underexploited, as the pressure for simple recipes has prevailed and shadowed the application of rich cognitive models. In the second part of the paper, the renewal of interest for Computational Social Science (CSS) is focused upon, and several of its variants, such as deductive, generative, and complex CSS, are identified and described. In the concluding remarks, an interdisciplinary variant, which takes after ABM, reconciling it with the quantitative one, is proposed as a fundamental requirement for a new program of the CSS.
The two decades around the turn of the millennium have seen the rapid advent, and perhaps the premature decline, of a paradigmatic shift in science, represented by agent-based modeling (ABM) and simulation. In this section, after shortly defining what we mean with ABM, we present a short account of its history.
What is meant by Agent Based Modeling? Often, this is defined in opposition to Equation-Based Modeling (see for example Dyke Parunak et al.,
Under this general definition, the field of agents shows a tremendous variability. Agents vary indeed on several dimensions, which include whether and to what extent they are autonomous, self-interested, sociable, and capable to learn from experience and/or observation. Agents also differ in their level of complexity: according to a classic distinction introduced by Wooldridge and Jennings in their influential work (Wooldridge and Jennings,
The practice of ABM however did represent a substantial under-exploitation of such wide spectrum of possibilities. De facto, much of the agent models worked out and simulated are totally
Conference proceedings, dedicated to the new methodology of ABM and its multiple applications within the social and behavioral areas of science, started to appear in Europe since the early nineties (Gilbert and Doran,
At the end of the first decade, however, the ABM leadership seems to be challenged if not decisively weakened by the (re)appearance of a more sober, more encompassing, and less innovative tag, that of Computational Social Science (CSS), of which ABM is a component (see Bankes et al.,
If history is instructive, the study of signaling is fun. In the era of information overflow, distributed content production, collaborative filtering, crowd sourcing, and so on, emblems are decisive. Tags have a far-reaching but short life. Under the tyranny of PageRank, contents compete in terms of lookups, and these most certainly depend on familiarity, and possibly also on tags appeal. Science makes no difference. It is somewhat surprising when a paradigm shift is signaled by a flat combination of two traditional scientific areas: social sciences and computational science. What is the meaning conveyed by this signal? Does the new label correspond to a new paradigm shift in the social and behavioral sciences, or does it simply meet a kind of marketing need for periodical renewal of names?
This paper presents an attempt to weigh up the impact of ABM and answer the question whether this field is undergoing or not a real decline; whether or not his replacement was timely, necessary, and effective. Next, some current variants of CSS will be compared. Finally, some important requirements for achieving real progresses in the computational study of social phenomena will be identified and discussed.
Rather than a detailed survey of ABM (for a good example, see Helbing and Balietti,
One may wonder what ABM is good for and what are its major strong points. The tricky questions as to when ABM is really needed, whether agent-based models can or cannot be converted into an analytical, equation-based model and to what extent this can be done has been debated at length elsewhere (see for example Epstein,
Of course, ABM can only provide a sufficient explanation of the phenomenon of interest, not a necessary one. This feature, which (Epstein,
To implement sets of heterogeneous agents in interaction brings about a series of second order advantages: agent societies are (1) operational platforms where theories get converted into falsifiable hypotheses; (2) experimental laboratories where theories get gradually and thoroughly controlled; (3) multilevel worlds where the level of individual units, the agent, is clearly distinct from the macro-level, the system level and unforeseen effects and emergent properties of interaction can be observed.
In short, ABM is an
Some problems and difficulties in the field of ABM and simulation have been perceived from within the scientific community since long, while others have only recently come to our attention. Since the field's early days, a serious concern of Agent Based modelers and simulators is how to design large-scale agent-based simulations. In its initial applications, agent-based models did not care much about the problem of scale, as they were applied to observe the emergence of patterns from interaction at the microscopic level in artificial scenarios sharing some crucial features of the real-world, but not really aimed to reproduce its details. As soon as the potential of agent-based models became apparent—revealing a great occasion for observing and manipulating
Under the pressure of complex systems science, which is gaining ground in the study of social phenomena (Helbing and Balietti,
Unlike laws of nature, Agent Based models of socio-economic phenomena are countless and not always consistent (see Alfi et al.,
We believe the variety of equivalent agent models in part depends on a property inherent to multi-level systems as complex social systems are. The property in question is the multi-realizability that we have mentioned above. In part, we believe it to be a consequence of the shaky foundations, the poor theoretical justification at the basis of many agent models. This is not equal to finding poorly realistic the model of agent often proposed by current modelers, and asking to improve it toward psychological, cognitive, or sociological plausibility - toward a
A consensus seems to have emerged in ABM on a minimality procedure; that is, models are built by setting up the rules that are minimally required to obtain the macroscopic effect to be described. While minimality might sound obviously inspired by the success of hard sciences, the substantial failure to apply such a minimality procedure to social science is testified by centuries of failed attempts, starting from what had been announced as “Social Physics” in the seventeenth century (for an historical perspective, see Ball,
Under the rule of minimality, model building is operated (1) a posteriori, based on backward engineering from the effects obtained to the generating rules; (2)
One such option is represented by cognitive agent models, which exist since the late nineties. Their wide range of influence is shown by the popularity of BDI architectures (about 32,700 “BDI agents” cites on Google Scholar retrieved on March, 18th 2013) within and beyond the field of agent systems and theories. Simulation of social phenomena with BDI based models also abound in the literature (about 7060 “BDI social simulation” cites on Google Scholar on March, 18th 2013), and usable platforms to implement them are under consolidation, from Jason (Bordini et al.,
Although the rich cognitive models tag appeared since the early nineties (for a recent example see Dignum et al.,
First, unlike other theory-grounded agent models, for example the rational models, cognitive models are not prescriptive. Whereas the theory of rationality is a theory of action, cognitive modeling provides theories of the agent. Hence, the rational agent model fits only apparently better the objectives of ABM and simulation, but it does so only because it allows the modeler to get rid of the tricky part of the modeling, that is, how agents form the goals, the motivations, the preferences, that will be implied in the decisions.
Secondly, cognitive modeling is a truly generative theory of behavior, accounting for behavior in terms of the mechanisms that are supposed to operate while producing it. A generative explanation of an observed social phenomenon consists of describing it in terms of the external (environmental and social) and internal (behavioral) mechanisms that generate them, rather than by inferring causes from observed co-variations. This is a vital property of explanation, which cannot easily be realized otherwise. When describing agent behavior by means of other formalisms (logic-based or numeric), we describe behavior from the outside, as perceived by an observer, but do not describe the way it is generated. ABM explains behavior from within, in terms of the mechanisms that are supposed to have generated it, that is, the mechanisms that operate in the agent when s/he behaves one way or another.
Of course, behavior can be explained otherwise. For example, the flight of hawks is wonderfully explained by the mathematical property of logarithmic spiral, such that any tangent from the center of the spiral yields an angle of the same width. Thanks to this property, hawks can keep their preys always in their aim while describing a spiral before pouncing on them. But this explanation is not generative, in the sense that it does not tell us what are the internal mechanisms allowing hawks to fly the way they do. For sure, hawks do not fly based on an understanding of the properties of logarithmic spiral. How can they show the corresponding behavior? The often invoked evolutionary explanation offers poor help: it accounts for behavior in terms of its reproductive advantage. As the spiral-like flight proved advantageous for hawks, those who performed it were able to generate more offspring, while the others extinguished. No generative theory here: it tells us not how hawks produce the behavior in question. We could use the mathematical theory to describe their behavior, and incorporate the mathematical explanation into a set of
Now, a fully generative explanation implies a more general theory of how external causes, including fitness-enhancing effects, get converted into internal reasons (what sometimes are called proximate causes of behavior). Agent-based models are often limited in focus, and not easily compatible with the temporal perspective and the theoretical requirements of a fully generative - in the sense here intended - explanation. Do we always need a generative explanation? Not really, as
Cognitive modeling aims at finding the general mechanisms yielding the wide spectrum of behaviors of relatively autonomous systems. Of course, you don't need such mechanisms to simply reproduce behavior. The more specific the target behavior, the lesser you need a cognitive agent-based model. Since ABM is often used to investigate fairly specific phenomena, either mathematical model or a set of
One might say, who cares after all? Provided we can reproduce behavior, observe it and make artificial experiments to optimize it, why bother with theory-driven generative modeling? There are several reasons. One is that a truly generative explanation is needed to model complex social dynamics. For universal admission, the dynamics of social entities and phenomena is at least bidirectional if not multidirectional. Entities and properties emerge from the bottom up and retro-act on the systems that have generated them. Current agent-based models instead simulate only emergent properties, i.e., the way up of social dynamics. To mention only a few examples, the ABM literature offers countless models of the emergence of segregation, norms, reciprocity, altruism, cooperation, punishment, conventions, institutions, coalitions, leadership, hierarchies, the modern state. Studies of different types and levels of downward causation are much less frequent (to cite some exceptions, see Gilbert,
Agents should not be taken for granted as they change under different types and degrees of social influence. Entities at the macroscopic level affect them and their behavior, and we must understand how this can happen if we want to drive, enforce, or prevent such an influence. This a line of research that presents obvious ethical issues, but at the same time addresses themes so important that social science cannot just leave them alone, or, even worse, desert them to market solutions. For example, at least in some fields, we badly need to know how to reduce or control people's overconfidence, for example in finance, where it so heavily contributed to the last financial crisis (see Akerlof and Shiller,
ABM is a powerful means for investigating the hinge between different domains of reality, including economy, environment, and society: systems' behavior at different levels of scale. It is necessary to explain phenomena pertaining to any domain of reality that is heavily dependent on the behavior of autonomously interactive systems, as was convincingly argued by Epstein. More, ABM is unique for allowing a generative approach to behavioral systems in the sense here defined, and somewhat different from Epstein's, i.e., to describe phenomena in terms of the external and internal mechanisms that produce them.
However, ABM seems to have fulfilled its mission only in part. Its generative capacity has been deployed to a lesser extent than could have been the case. The practice of ABM missed the opportunity it provided: paradoxically, the same principle that led it to a fast popularity, like the KISS principle—i.e., keep it simple, stupid - introduced by Axelrod (
If internal mechanisms are
It could be the case, indeed. However, centuries of failed attempts (see the “Social Physics” case mentioned above) make us doubtful. But what is maybe more important, by pursuing this quantitative approach alone, science would have lost a wonderful opportunity: to understand the micro-foundations of phenomena at aggregate levels and how the latter (re)generate them.
Science, like history, is not a linear process. A decade ago, social, and behavioral science dropped the disciplinary label (Conte,
Computational Social Science (from now on CSS) can be meant in at least three different ways, the deductive, the generative, and the complex one; and it should be made clear which one we are referring to. As these are conceptual, rather than empirical, variants, there is no need to have each of them matching a defined historical example of CSS, since concrete examples are often a mix. Let us characterize variants also with reference to existing programs and try to forecast what their consequences might be.
The second half of the last century is constellated of attempts to apply the theory-building instruments of mathematics and the theory-testing tools of computer science on one side, game theoretic, and logic-based computational models on the other, to describe and explain social phenomena. The latter, in particular, attempted at deducing properties at the macro-level from general assumptions at the micro-level. Expectations á la
The decline of the rationality paradigm produced several consequences. One of these was a stronger and more interdisciplinary effort to ground computational models on explicit models of the micro-foundations. This led to the advent of the generative variant of ABM, which derives its explanatory vocation and micro-foundational framework from the deductive variant. Unlike it, generative science aims at modeling operational microscopic rules that generate macroscopic phenomena, rather than formulating mathematical equations from which to deduce them. The explanatory vocation is declined in a radically different way: rather than describing a causal process from the outside, the modeler attempts to show the internal rules that initialize it and follow the unfolding of it all the way up to the observed effects.
As argued in the preceding section, however, ABM fulfilled its mission, provide generative theories, to a lesser extent than was expected. If the deductive variant was found to theorize upon fairly abstract phenomena and has often been criticized for its poor predictive capacity, the generative variant did not prove any better at prediction, partly due to problems of validation and calibration.
Inductive computational science is certainly not new (Newell and Simon,
A new impulse to computer-based quantitative social science is coming from the science of complexity, which is now going through a season of deserved popularity. The use of complex systems' methods, models and techniques to economic systems goes back to the nineties (for a rather informative introduction, see Mantegna and Stanley,
Behavior is irrelevant or non-influential in social dynamics where the implications of the phenomenon in question are social, but its nature is not. To go back to epidemics, the nature of epidemics is biological. The level of reality involved entities belong to does not matter for the observed phenomenon to take place: the nature of entities involved in and target of epidemics matters not. In the spread of epidemics, the difference between human behavior and that of particles in the space does not matter, nor does the difference from carriers and the viruses they carry around. But in other cases, that is, when the nature of behavior matters, accurate statistical analyses of social dynamics can maybe reach predictive power but cannot fully explain what is going on.
As a hypothetical example, suppose we want to know what are the main factors responsible for the dynamics of opinions. Again, current models (Deffuant et al.,
People withdrawing support from political leaders is a good example of non-linear opinion dynamics. It is unclear when people change their minds and turn down their leaders. The destiny of a popular (and often populist) figure is often decided upon in a very short time. Today, those who enjoyed the favor of their followers until yesterday, may suddenly lose popularity and fall in disgrace, what is again a matter of threshold: after a certain level of spreading, and perceived spreading, agents are led to modify their opinions, what probably reveals an interesting effect of shared representations about shared opinions on one's confidence level. Possibly, such a lowering confidence leads agents to be more eager to change opinions.
However, the circuit may be completely different: agents may resist pressure to change opinions despite contrasting evidence for reasons of cognitive dissonance. The more the contrasting evidence they gather, the higher the dissonance. To reduce it, they try to ignore evidence that is less costly than change opinion, which imply dropping the previous commitment and making a new one. As the perceived distance from others' opinions increases, however, agents either hide their opinions or must defend them openly. If they choose the latter strategy, they may even end up by accepting to form part of a minority. If they take the former option they cannot get along with deception too long as cognitive dissonance increases. Consequently, they accept others' opinions as own, and are likelier to convert them in open behaviors to convince others and themselves about the solidity of their new opinions. Both routes imply critical thresholds for totally different reasons. To act efficaciously on this process, we must be clear what is actually going on. Confidence has different implications from cognitive dissonance and self-deception. To increase confidence may lead to higher stability in the former case, but not in the latter.
To sum up, to model social dynamics without taking into account the internal (cognitive) dynamics of the entities involved in a social phenomenon does not prevent accurate predictions of critical events and changes. It may even allow to find out factors responsible for such events and changes, and this is the case with dynamics for which the social nature of behavior is irrelevant. To understand internal dynamics is crucial instead whenever we need not only to anticipate but also to understand events for which behavior is relevant. Model the internal dynamics of events is necessary not only for scientific reasons but also for guiding intervention.
The program for CSS needs clarification. Why would such a program be necessary, if we practice CSS since at least a couple of decades? Of course one might say that we need to introduce a new Curriculum at the academic level, and that to do this implies to form a new, cohesive, scientific community, form associations, give visibility to this new Curriculum, strengthen the academic, editorial, and political power of the underlying community etc. However, the reason for a program on CSS is not only political but also scientific. As seen so far, there are different variants of CSS and to take a pluralistic approach to it may be considered wise. CSS could be seen today as a larger umbrella under which different approaches might coexist and somehow feel legitimate. Hence, generative ABM might be practiced by a subset of social scientists, while others might prefer a purely quantitative approach, based on data-mining and numerical simulation, and still others might continue to formulate abstract theories of social action in elegant equations and deduce their macro-level consequences.
The main thesis of this paper is that such a
An interesting, innovative program in CSS can only be
An interdisciplinary endeavor like this certainly points out some new challenges: not only to extract knowledge from larger and larger datasets, not only develop simulators that scale up of several orders of magnitude, or feed simulation and data-mining with online real-data, not only to develop supercomputing infrastructures and systems to transfer data to supercomputing platforms, but also develop simulation platforms that scale up both in terms of systems' dimensions and in terms of levels of complexity. We need to account for large-scale systems as well as more complex entities. We need to apply simulation methods to understand the social and the mental dynamics and to describe their interrelationships. Last, but not least, we need incentives that are compatible with such an endeavor—publication-wise and career-wise. This is a challenge for a program on CSS that deserves attention and investment. CSS ought to accept it, or another occasion will be lost for founding a novel, integrated, interdisciplinary, falsifiable science of society helping us to solve transformative and foundational problems.
Rosaria Conte and Mario Paolucci elaborated together the ideas presented in the paper. Rosaria Conte drafted the first version. Mario Paolucci substantially revised it for important intellectual content.
This paper benefits from the stimulating discussions in the
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We thank Neil Yorke-Smith, Gustavo Nardin, and Federico Cecconi for encouragement and suggestions. We are also thankful to anonymous reviewers who helped us to tune the paper and helped to clarify some obscure passages.