This is an unusual book about an exciting new way to do social science. The title tells us that the author loves to turn a phrase. The subtitle gives a fair summary of the third book in a trilogy by Joshua Epstein, developing and exploring what he memorably called “generative” (crudely, bottom-up) social science. Previous books—Growing Artificial Societies (1994) and Generative Social Science (2006)—mapped the intellectual territory and developed the widely cited Sugarscape model. Here, Epstein sets out to use insights from cognitive science to inform the modeling of human social interaction. The potential for linking brain science and social science may seem blindingly obvious. However, many social science models of human decision making are constructed as if there were no scientific discipline focused, precisely, on human decision making. Taken together, Epstein’s three books make a major contribution to the foundations of a new approach to social science that, while easy to recognize and distinguish from what has gone before, is surprisingly difficult to define in a few words.
Call this generative social science or agent-based modeling (ABM) and practitioners will immediately know what you are talking about. You can tell civilians that this involves grounding dynamic models of society in simple dynamic models both of individual human decision making and of social interaction between humans. You can say that setting these simple models in motion can generate surprisingly complex and realistic-looking processes that offer new intuitions about important “emergent” social phenomena—for example, about why ordinary people might engage in risky behavior like joining public demonstrations in viciously authoritarian regimes. While true, this is not particularly informative, however, and I will be a little more expansive in my conclusions.
Agent_Zero starts with the human brain, more precisely the amygdala. Epstein then offers readers an eclectic intellectual tasting menu, drawn from established research programs in cognitive science and social psychology that offer potential insights into human decision making in a social context. The author is very well qualified to do this, given his massively multidisciplinary resumé: a primary appointment in emergency medicine at Johns Hopkins; joint appointments in fields as diverse as applied mathematics, biostatistics, economics, and civil engineering; an external position at the Santa Fe Institute. Moving beyond the earlier work reported in his trilogy, the research findings discussed here are intended to ground a generic model of human decision making that can form the basis of bottom-up (agent-based) models.
Epstein considers interactions among three distinct components of human decision making in a social context. Two refer to individual human agents deciding in isolation. The first he calls affective/emotional, and it is in relation to this that he offers us colored pictures and explanations of brain scans generated by experiments in learning and conditioning. This component concerns what human agents feel, know, and learn but do not explicitly think about. The second component he calls cognitive, by which he means “ratiocinative” (Epstein is an enthusiastic collector of rare words) and deliberate. This component is an agent-based version of the more explicit decision making typically modeled by game theorists, modified into a form of “bounded” rationality given the complexity of the decisions facing real humans in social settings. Agents gather data using (likely biased) sampling from their local environments and come to conclusions by running these data through some algorithm. The third component is social and deals with ways in which human interactions may communicate both emotional affect and empirical data from one agent to another. This generates complex “contagion” whereby outputs from my affective and cognitive processes change inputs to and hence outputs from yours, outputs from yours change inputs to and hence outputs from your friends, outputs from your friends change inputs into mine, and so ad infinitum.
Each of these three components on its own, and complex interactions among them in particular, do on the face of things offer insights into human decision making in a social context. Since social science is all about human decision making in a social context, in what way is this book unusual? The answer is that most published work in the social sciences starts with a substantive problem—which may be big or small, arcane or commonplace—and selects, from a diverse portfolio of candidates, the model of human decision making and interaction deemed most promising in the circumstances. In contrast, Epstein’s prime concern is human decision making and interaction per se. He focuses relentlessly on the generic modeling of how humans make decisions in a social context, and treats particular substantive problems in human decision making as what he calls “parables,” deployed rhetorically to show that his generic model offers plausible substantive insights. His “canonical” parable is “slaughter of the innocents.” There are soldiers who start the day feeling safe, and have never themselves been exposed to guerilla attacks. Epstein’s striking conclusion is that by the end of the day, still never having been attacked but given remote interactions with colleagues, these same soldiers may of their own free will slaughter innocent local villagers.
The interesting thing is that we are at a “moment” in social science when Epstein’s focus on human decision making per se is particularly appropriate. We have lived through three generations of formal modeling in game theory and microeconomics during which a gigantic amount of intellectual energy, and a considerable amount of talent, has focused on modeling human decision making in a social context. As this paradigm has grown toward maturity, an unwritten encyclopedia of implicit assumptions and folk wisdom has evolved that covers an awful lot of things typically taken for granted when setting up a formal model. We find a family of related descriptions of human “rationality” and an agreement-among-scholars that modeling assumptions belonging to this family are accepted without the tiresome need to ground them in the cognitive science that actually studies human rationality. Recent work in behavioral economics is chipping away at this comfort zone, but the assumptions about human cognition grounding most of the game theory canon remain largely implicit. Whenever a theorist specifies an equilibrium concept, such as Perfect Bayesian Nash Equilibrium, for example, this is not only a choice of modeling technology to help shake out a result; it is also a pledge of allegiance to a deep theory of human cognition. (Epstein, by the way, does not see humans as Bayesians, but uses a different and simpler updating regime in the evolving social dynamics he models.)
Traditional formal theorists talk happily to each other not only about equilibrium concepts that impound fundamental cognitive assumptions but also about notions such as minimax regret, risk aversion, time preference, satisficing, and so on. They know in their heart of hearts that there are cognitive scientists, in another building on the same campus, who know much more about these things than they do, whose work they have not read (and choose not to read even knowing it exists). But the agreement-among-scholars means that they are never called out and shamed for doing this. The traditional game theory paradigm is probably too set in its ways to change, but the emerging field of ABM and Epstein’s generative social science is not. This is why we now have a “moment.” The striking thing about ABM is that we furnish each agent in our model with a miniature brain and, when we program the model, we must write a program for each of these agent brains that tells it what to do in any situation it might encounter in the model world it inhabits. This program is Epstein’s concern in Agent_Zero.
Whether or not the reader buys the particular prototype agent designed by Epstein is largely beside the point—and he is not always a good salesman. (There could be more light and shade in his sales pitch, a more relentless focus on what we really need to know from cognitive science in order to program our agent’s brain, less digression on what is merely interesting and potentially relevant, and a less scattergun preview of potential future work. Epstein knows an awful lot, but it is not clear his readers need to know everything he chooses to tell them.) The point is that this book persuasively directs our attention to how best to model and parameterize the thinking apparatus of the agents who form the basis of bottom-up social science. It would be a terrible shame if, 20 years hence, a new paradigm had evolved in which the way we program the miniature brains of our model agents was not based on what we have learned from the rapidly evolving insights of brain science.
How will this new paradigm evolve? If we could answer this question with any certainty, we would be the intellectual superstars of the next generation. But let us pause to consider some distinctive features of the movement toward generative social science/ABM, and the intellectual forces that underlie them. Social interaction among real humans is fundamentally dynamic. It often involves large numbers of agents interacting to produce markets, economies, and political systems. Whenever we seek realistic models of human social interaction, therefore, we seek dynamic models that comprehend interactions of large populations of agents as well as just a few. The three generations of formal modeling referred to previously have evolved into an established paradigm that has yielded extraordinary advances in our rigorous understanding of human interaction and have fundamentally changed social science. Most of these advances have referred to rather small numbers of interacting agents, and most have involved a best rather short-run nonevolutionary dynamics. These domain restrictions have been the price paid for analytical rigor; the price has clearly been worth paying.
There has always been something missing, however. It has never been difficult to conceive of and specify more realistic and dynamic models that comprehend large populations of interacting humans, but the analytical intractability of these models has mandated computational rather than formal mathematical analysis. A well-known early example is the influential model of social segregation specified by Thomas Schelling in the late 1960s. This is fundamentally dynamic, fundamentally addresses large populations, and, originally, was computed manually in a toy setting, rather than using machines. While the deep intellectual desire for large-scale dynamic models was largely sublimated for practical reasons during the era of game theory and formal modeling, the intellectual landscape has been transformed by the recent widespread availability of immense computing power. Perhaps more importantly, the value-added of all this raw computing power has been greatly enhanced by the availability of software tools for modeling social interaction (including the freely downloadable NetLogo used by Epstein) that are accessible to scholars with only modest programming ability but the burning intellectual ambition to build new dynamic models. Suddenly, not only can large-scale dynamic models be specified, programmed, and investigated in “toy” settings; they can also be exercised in heavy-duty computational experiments that allow exhaustive exploration of the model’s parameter space in a manner directly analogous to formal comparative statics. Computational rigor may differ from the rigor of formal mathematical proofs, but it is rigor nonetheless and is now easily achievable in the right hands.
There is something substantively different about generative social science/ABM, however, that goes far beyond the mere ability to specify and analyze large-scale dynamic models. Putting our model humans in large-scale dynamic social settings forces us to reconsider how they might make decisions. If large-scale dynamic models are analytically intractable for the world’s best mathematicians, how does an ordinary decent human decide how to behave in complex social interactions unfolding on an hourly basis? Describing how real humans attack analytically intractable decision problems, agent-based modelers typically assume that agents respond with simple but intuitively compelling rules of thumb. Good examples of such rules can be found in accounts of the canonical “traveling salesman” problem: Find the shortest route that visits each of x randomly scattered cities at least once. This is well known to be analytically intractable in the general case, but we also know two things about how real humans “solve” it. First, they typically attack the problem quite effectively using simple algorithms that are not analytically provable best responses. For example: start at a random city; visit the closest unvisited city; iterate until all cities have been visited. This “greedy algorithm” can be shown to work very well in many cases but horribly in others. Second, different real humans use different rules of thumb in the same setting, with different rules coexisting and no single rule evolving to dominate all others. More generally, moving to generative social science/ABM does not just offer us an opportunity to reconsider how real humans make decisions; it forces us to do this by denying us the assumption of elegant strategic rationality that characterizes the traditional game-theoretic paradigm. This reconsideration is Epstein’s fundamental objective in writing Agent_Zero.
A closely related empirical development, which for want of a better term we can call “big data,” is beginning to add a whole new dimension to this intellectual landscape. Social science data sets have traditionally been difficult and expensive to collect, therefore small and largely static. This lack of data generates a problem for the serious empirical evaluation of large-scale dynamic models. But now there are vast volumes of cheap machine-readable social data everywhere we look, even if we have yet to figure out how to tame them and analyze them in an effective way. These data are large scale and dynamic and therefore address exactly the same empirical space addressed by the theoretical models emerging in the new paradigm of generative social science/ABM. We might speculate that it will be a synergistic convergence of these theoretical and empirical developments, sooner rather than later, that will fuel the intellectual engine of the new paradigm, of which Epstein may come to be seen as a founding father.
