Dr. Joshua Epstein Discusses Applications of Agent-based Computational Modeling at Abt Associates Annual Meeting

"What Do You Mean By An Explanation?"

"If you didn't grow it, you didn't explain it." That, said Joshua Epstein, Senior Fellow at the Brookings Institution, in the 2005 Abt Associates Stellwagen Lecture, is "the bumper sticker reduction of the agent-based computational model." Dr. Epstein develops computer models of agents — which represent individuals or large multiples of humans — and uses these models to generate sets of social interactions and behaviors over time. The events modeled in Dr. Epstein's presentation as part of his lecture included the mysterious disappearance during the Middle Ages of the Anasazi Indians, the spread of epidemic disease, violent ethnic conflict, and a revolution.

To the end user, the models are delivered in what Dr. Epstein described as "eye candy movies" that look like animated board games in fast play motion, with different colored dots representing humans who might include, for example, long-dead members of a vanished civilization, infected and healthy individuals, passively angry citizens, active revolutionaries, and "cops." These colored dots move — fast — in complex patterns. At one point in the presentation, a computer glitch caused one of the models to run prematurely, before Dr. Epstein had had a chance to explain exactly what it was going to show. As the pattern unfolded on the screen, Dr. Epstein protested, "It's giving away the whole conclusion!" The model was so transparent that the audience was immediately able to tell what was going to happen next.

For his research, Dr. Epstein runs the models thousands of times in order to develop robust statistics on each distribution. Each "movie," Dr. Epstein explained, represents only "one realization of a stochastic process." Hidden from presentation lie complex mathematics and econometrics that make the "movies" ideal for explaining social phenomena to non-scientific policy administrators. As Dr. Epstein said, "The agent-based model is user-friendly. We've designed a whole portal for the National Institutes of Health, and there's also a private website where NIH can run models if a bug shows up in, say, Pittsburgh, to determine whether they should close the schools. Aside from being of scientific merit it's also got some nice features from the point of view of policy impact." Associates Fellow Stephen Kennedy, who introduced Dr. Epstein, said Dr. Epstein's research "has proven really quite startling in some cases and has a remarkable range." Steve recalled his own experience in graduate school with macroeconomic models: "One of the intellectual high points was the proof that, under really strong conditions, in a really simple economy, there actually existed an equilibrium. Nobody could say whether you'd get there. It was hard to say a lot about what it looked like but we knew it existed. I tell you, the math that this required was really quite impressive. The moment that you start doing things that involve behavior more like real people's, it gets really hard to say much."

The primary difference between Epstein's agent-based computational model and standard macroeconomic models is that in Epstein's model, the agents — represented by all the dots in the model — are heterogeneous. "We don't collapse the entire society into a representative agent who optimizes or does something else," said Epstein. "Every single agent in the system is explicitly represented. They can differ in all kinds of ways, including wealth, social network, immunocompetence, memory, genetics, cultural affiliations, decision rules and so on. They are autonomous. They obey their own rules and regulations and move around their environment under their own power. Unlike Homo economicus the agents in our model interact with family, friends, and others in social networks."

Dr. Epstein's agents, he said, are "boundedly rational: they have limited information and they do limited things with it. They don't optimize globally over the set of all possible choices, with perfect information about everything. Events unfold on an explicit landscape, which is also missing from much of economics, game theory, and epidemiology. And we care a lot about the non-equilibrium dynamics. As Steve said, many of the jewels of mathematical economics are really existence theorems that have nothing to say about whether equilibria are attainable at all, on what time scales they're attainable, what kind of cognitive load is required to arrive at these equilibria — and we care about all of that. We care about where the system really goes over time and space." The models demonstrated during the Stellwagen lecture, and many others, will be included in Dr. Epstein's new book, Generative Social Science, to be published by Princeton University Press in 2006. The book makes the argument that, to explain a macroscopic social pattern like a wealth distribution, a segregation pattern, or a pattern of violence, the social scientist needs to grow (generate) the pattern in an agent model. Dr. Epstein's Models

Artificial Anasazis

Epstein did not initially set out to model the collapse of the Anasazi civilization in the American Southwest 700 years ago. Rather, he showed some colleagues an early agent-based model of an imaginary population in operation, and an archaeologist commented that the population dynamics depicted looked a lot like what we know about the Anasazi's history. (The Anasazi Indians lived in the American Southwest until approximately 1300 AD, and then vanished — for unknown reasons. Environmental factors, chiefly water supply, lead hypotheses.) Epstein and colleagues at The Brookings Institution and Santa Fe Institute first digitized the entire history of Anasazi population patterns over centuries. They then created a model, with intelligent agents whose behavior was governed by a set of anthropologically plausible rules, operating in a physical environment reconstructed using an impressive set of actual historical microdata. When the experience of these "artificial Anasazi" was plotted against the digitized history itself, the two matched quite closely!

"The hypothesis is that the Anasazi disappeared from the area because their population wasn't sufficiently big and dense to continue their political and religious and other institutions," said Epstein. "The point is that, by using only environmental considerations, you do nicely replicate the gross dynamics of the civilization. So you can argue that, to a large extent, environmental factors do explain the rise and fall of this group — although their ultimate abandonment of these territories must depend on other things."

How to Respond to a Modern Small-Pox Epidemic

An application of agent-based modeling that is more immediately relevant to contemporary concerns would be to predict the course of an epidemic and the results of different emergency medical interventions to contain it. Epstein and his colleagues at the Johns Hopkins School of Public Health modeled the progress of a smallpox epidemic. The model includes epidemiological features specific to the disease and its prevention: A vaccine is effective only for those not infected or those infected within the past 4 days. The virus cannot be transmitted during the first 12 days. Approximately one-third of infected individuals die. The vaccine itself is dangerous to certain individuals. The model also reflects contemporary reality: the epidemic would likely begin in an urban environment, with high mobility and many opportunities for contact between infected and uninfected persons — but, simultaneously, not everyone would be exposed. The results of Epstein's research suggested a containment strategy for smallpox that would be more feasible than vaccinating everyone who had contact with the infected person, but also less dangerous than mass vaccination.

Epstein is extending this kind of epidemiological modeling to address avian flu, and is currently building the first-ever model of a global epidemic. The model will take into account such behavior as flight and isolation, instead of assuming, as standard epidemiological models tend to, that people are going about their business normally. The new model can be extended to deal with other contemporary threats like global terror. "In standard epidemiological models of bioterror, no one behaves terrified," Epstein said.

Petrograd 1917, or Elsewhere

Epstein has also modeled civil violence, both between warring ethnic groups and in a revolutionary situation. He confessed in an aside that he always roots for the revolutionaries, but stressed that the model is really politically neutral and can be applied to any revolutionary situation. "I don't proffer this as the universal equation of human grievance," Epstein cautioned. "It's just a place to start out."

Epstein's model of a civil authority trying to suppress a general revolution assumes that citizens can be rebellious or not, depending on their level of grievance and their likelihood of getting arrested. In the model, says Epstein, "Cops are easy: they move around randomly and arrest a random rebel within their area of vision and throw the person in jail. People come out of jail every bit as ticked off as they went in — I always say that's the only realistic assumption in the model. Rebels are more interesting. The first issue is how aggrieved they are at the central authority. The simplest equation I could dream up is that grievance equals hardship times illegitimacy (i.e., one minus legitimacy)." Hardship alone does not produce political grievance. Britain during World War II experienced tremendous hardship, but rallied behind Winston Churchill and the government, which enjoyed high legitimacy.

"Having determined my level of grievance, the second question is, how likely am I to get arrested and thrown in jail? It's some function of the cop to active rebel ratio within my vision. When lots of other people are actively rebelling, it's safer to rebel. This is then multiplied by risk aversion. If grievance minus this net risk is greater than some threshold that we can think of as zero, then I rebel, otherwise I don't."

Epstein showed models in action of government legitimacy slowly eroding until no one had any confidence in the government — i.e., going from 1 to zero — versus government legitimacy plummeting sharply from a specific shock, but going down much less, only from 1 to .7. He showed that the latter is actually more destabilizing to the regime. The reason is that in the first instance, disenchanted individuals rebel and are picked off one by one by the cops, so that the revolutionary potential never achieves any momentum. In contrast, a sharp shock to legitimacy quickly mobilizes many agents to rebel at once, raising the active rebel to cop ratio and also, in the process, making it safer to rebel. This is why seemingly minor events that are terribly corrosive to your legitimacy produce enormous, disproportionate, global responses. It's shocks to legitimacy that matter more than absolute declines at some low rate."

Preventing Genocide Through Effective Intervention

By changing the model slightly, it is possible to show the course of ethnic conflict and predict useful interventions. Under this approach, instead of agents versus the government one has two different populations (different colored dots on the screen). Legitimacy is one group's appraisal of people in the other group's right to exist; if the legitimacy is reduced, one group begins attacking the other. Grievance is passed on from parent to child. In the base model there is no flight, no intervention, no amelioration of grievance. This leads inevitably to ethnic cleansing and genocide. Then one can introduce various interventions. The cops in this model are neutral peacekeepers, preventing members of the two warring groups from attacking each other. The model can help show when, where, and on what scale to introduce peacekeepers "at unoccupied sites on the lattice" in ways that will create effective safe havens and prevent genocide. Epstein cautioned, however, that under some conditions, introducing outside forces could exacerbate grievances and make the situation worse.

In summary, Epstein presented the vision of a new "generative" kind of social science in which the agent-based computer model is central scientific instrument.

Dr. Joshua Epstein is a Senior Fellow at the Brookings Institution and is the author of several well-known books on agent-based modeling, including Growing Artificial Societies (MIT Press, 1996, with Robert Axtell), Nonlinear Dynamics, Mathematical Biology, and Social Science (Addison-Wesley, 1997), and Generative Social Science: Studies in Agent-Based Computational Modeling (Princeton University Press, forthcoming).