Beyond SimCity: Agent-Based Computer Models Revolutionize Social Science Research

Posted: May 29, 2007 at 1:00 am, Last Updated: November 30, -0001 at 12:00 am

Computer model of Sugarscape
In the Sugarscape computational model, data on differing outcomes are produced as various changes in the environment or agents are introduced into an artificial society.
Sugarscape in MASON project image

By Karen Loss and Colleen Kearney Rich

Robert Axtell has spent a large portion of his career trying to figure out why people do the things they do. But he’s not a psychologist. He is a social science scholar, and he and his research team at Mason’s Center for Social Complexity (CSC) build computer models that simulate large numbers of people interacting.

Robert Axtell
Robert Axtell
Brookings Institute image

“Interactions could be social, financial or political. It’s very hard to render those models mathematically,” he says. “It’s not easy to summarize the functionality or the performance of the simulation groups in numbers or graphs. Often what we’ll try and do is depict the entire market as it emerges.”

Axtell is one of the leaders in this field of research. In 1996, Axtell cowrote a seminal work on artificial societies titled “Growing Artificial Societies: Social Science from the Bottom Up,” with Joshua Epstein of the Brookings Institution. In the book, Axtell and Epstein present a computer model with which they begin to develop a bottom-up social science in a land known as Sugarscape.

As various changes in environment or agents are introduced, data on differing outcomes are produced. What the authors found is that “fundamental collective behaviors such as group formation, cultural transmission, combat and trade are seen to emerge from the interaction of individual agents following a few simple rules.”

Merging the Computational and Social Sciences

In 2006, Axtell joined the Krasnow Institute for Advanced Study, where he continues his research on building large-scale models of social and economic phenomena, such as firms, markets and social determinants of individual behavior.

The primary mission of CSC is to advance the knowledge frontiers of pure and applied social science by using and developing computational and interdisciplinary approaches that result in new insights in social phenomena.

The computational social science team at Krasnow comprises one economist (Axtell), one political scientist (Claudio Cioffi-Revilla), one environmental economist (Dawn Parker), and one social network theorist (Maksim Tsvetovat), who draw heavily on evolutionary computing to do their research.

With Mason’s Evolutionary Computation Lab, the CSC created a software program they call MASON, or Multi-Agent Simulator of Networks. The CSC team runs a variety of models using this program, which Axtell says has displaced the use of the equation-based methodology that has been used over the past decade.

Speaking of one exciting new project, Axtell explains that the team is trying to build the first very-large-scale macroeconomic model of the U.S. economy.

“The model will have all 150 million agents of the U.S. economy in it. All the people are little software agents. We try to keep the microscopic structure of the economy intact, and the aggregate picture just sits on top of that,” Axtell says. “Today, there’s only an imprecise understanding of how the macroeconomy works, and this kind of modeling approach, we believe, is a way to actually improve policy.”

Using Historical Data to Answer Complex Questions

Over the span of five years, Axtell also has built a model of an 800-year period of Anasazi Indian life in the American Southwest. Using historical data, the model shows significant environmental problems that likely would have caused serious societal challenges. This information could lead to theories about the virtual disappearance of the tribe in the 1300s.

“This multiagent system approach to social science is so new that many domains, particularly where they are not highly computer literate, are slow to adopt new technologies like this,” says Axtell. “This area of archaeology and anthropology will be one that can benefit from this new technology.”

Axtell’s Anasazi project was the first in the field of anthropology to use the multiagent systems approach, but now, he says, there are dozens of other similar projects. In fact, the American Association of Anthropology now has a special interest group on the methodology of multiagent systems modeling for anthropology.

“Social sciences have been slow to be heavy-duty users of computing historically. We’re still learning the fundamentals. All of our models are visual. It’s hard to visualize a million interacting things with just numbers and graphs,” says Axtell.

Beyond Gaming

The models have a close affinity with the computer game SimCity, but there’s no social science involved in that game, says Axtell. Still, the mass market game, as well as online simulation exercises such as that found at Virtual-U.org, can give people an idea of what the center’s multiagent simulations are like.

According to Axtell, there is some overlap between what gamers do and what the team does. “The difference is that we are going to have relatively simple visualizations and deep behavioral content, and they are going to have incredibly superficial behavior but very enriching graphics.”

Currently, there are three main areas where multiagent systems modeling affects policy: traffic; propagation of disease, particularly since Sept. 11, 2001; and military tactics. Axtell says that teaching MBA students to build models of entire companies with these agent-based techniques is definitely on the horizon. Using this type of modeling to address institutional dilemmas could one day help corporations analyze problems and determine more effective solutions than might have been possible in the past.

What kind of problems are best studied using agent-based modeling, and what are best studied with conventional equation-based techniques? Even that distinction is not yet clear to researchers because of the newness of the agent-based approach.

“Our models essentially enhance the equation-based methodology (EBM), which is, in a manner of speaking, the forerunner of the modeling we do,” Axtell says. “But EBM misses all the action around the average, and, in some cases, that action matters a lot. Predicting extreme events is what’s most important.”

This article appeared in a slightly different form in the spring 2007 Krasnow Bulletin.

Write to at