Cornell researchers are beginning to look to a computer simulation technique known as agent-based modeling, which is fueling new theories on social phenomena about everything from segregation to stock market exuberance.
Sociology chair Michael W. Macy, one of three sociology professors working in the area, asks “Why do people stampede? Why did the NASDAQ run up so high? Nobody was telling people to buy stocks.”
According to Macy, agent-based modeling is “a way to understand self-organizing group dynamics.”
Each agent within the computer program simulates a single decision-maker, usually a person. Agents follow extremely simple rules, but their interaction with each other can result in intricate collective behavior.
“People don’t need to be individually irrational to behave collectively in an irrational way,” said Prof. David Strang, sociology.
This approach is unusual because the behavior emerges “bottom-up,” from interactions among individual agents, rather than “top-down” from higher authorities.
Traditional top-down views of social behavior focus on the impact of institutional restrictions imposed on the entire group. For example, residential segregation has been attributed to the actions of real estate developers, banks, and other large organizations. But the game theorist Thomas Schelling, taking a bottom-up approach reminiscent of agent-based modeling, showed that little more than an individual’s fear of isolation will eventually lead to the very same segregation.
Macy also drew a distinction between agent-based models and the common perception of computer simulations. Programs such as flight simulators tend to be complex, striving to recreate every detail of reality. But in agent-based models, “the rules are remarkably simple,” he said. They are intended as “tools for thought experiments,” rather than as fully accurate models of the real world.
Strang has collaborated with Macy to apply these agent-based models to management. Business practices often appear to be fads — adopted one moment only to be abandoned just as quickly. The two professors co-authored an award-winning paper on this topic. “It was a surprising case where we were able to contradict conventional wisdom,” Strang said.
In this case, conventional wisdom is that fads are due to managers acting in irrational or unintelligent ways. Yet “managers are very smart people and they are under tremendous pressure to get it right, to perform,” Macy said. “How do you reconcile the collective behavior, which is fad-like, with the individual behavior, which is anything but conformist?”
Earlier models based on this premise of lemming-like conformity were able to explain the explosive popularity of business trends, but not the equally sudden movement away from those trends. Strang and Macy proposed a different explanation: widespread business fads can arise from the interaction of individual, intelligent managers at different companies.
The professors created an agent-based model that used two key assumptions to produce fads from what seemed like rational individual behavior. First, no business practices in the model were decisively beneficial; every idea was either “worthless or only slightly worthwhile,” Strang said. Second, managers attempted to imitate the most successful businesses around them without examining the least successful.
“You’re only responsive to the cases of extraordinary success around you,” Strang said. “You’re not being observant of failure.”
Strang said he has performed additional research to test the applicability of the model to the real world. He found that the behavior of management teams at a major bank closely matched the model’s key assumptions.
In addition to his work with Strang, Macy uses agent-based models to understand the destructive behavior of adolescents. The research, which won a grant from the National Science Foundation, examines how behaviors like underage drinking are reinforced even when most people involved hold misgivings about them.
“It’s understandable [that teens drink] given all the pressure they are feeling,” Macy said, “but where is that pressure coming from?”
“People’s public behavior is to pressure people to conform to the norm,” he said. “They not only drink but they participate in activities that clearly signal to others that they should participate.”
Macy said such seemingly irrational behavior can again be explained from the bottom up, using agent-based models. In Macy’s model, individuals are only aware of the behavior in their own circle of friends. Given this social structure, widespread pressure can spread from a small number of “true believers,” those who genuinely wish to partake in the activity. If these “true believers” are distributed just right–not too dispersed or too clustered–a mere ten true believers in a population of 1,000 can lead to widespread adoption of behavior which most participants dislike.
“The enforcers [of the behavior] are more likely to be the people trying to prove their genuineness,” Macy said. “They greatly overestimate the extent to which others enthusiastically participate,” he said, and “they don’t want to be called a poseur.”
Steve Benard grad, studies political polarization with Macy. Their work shows how simple individual behavior, based entirely on perceptions of similarity to others, can lead to political polarization. “Just because society is complex doesn’t mean the rules society is built on are also complex,” Benard said.
“Agent-based modeling can be a great example of the benefits of doing interdisciplinary work. It can foster communication between different disciplines that wouldn’t otherwise be taking to each other,” Bernard added.
Although Macy has produced several major successes in his work, he says agent-based modeling is still a relatively unproven field. “It’s new enough that people are justifiably skeptical,” he said.
The models also have some important limitations. According to Macy, computational models are limited to predictions based on given rules and starting conditions. Other, more mathematical approaches can be generalized, but models such as Macy’s and Strang’s are specific to a particular set of parameters.
Nonetheless, both Strang and Macy expressed enthusiasm for the future of agent-based models. “They allow social scientists to be much more creative,” Strang said.
Strang also pointed out that exponential improvements in computing power will continue to expand the capabilities of agent-based models. “It will be able to be done on a larger scale and in greater sophistication over time,” Strang said.
Macy agreed. “It opens up some real exiting possibilities for the future of the social sciences,” he said. “Maybe agent models will turn out to be a fad just like the ones we study, but I doubt it.”
Archived article by Peter Flynn