Computers are the labs in the search for answers to human behaviors

By Karl Leif Bates / The Detroit News

    Picture a group of economists or political scientists hunched over their computers.
    There, on the cutting edge of social science, they're manipulating society.
    Using programed artificial guinea pigs to represent individual people or entire nations, they search for answers to basic human behaviors -- why we live where we live, who we choose to trade with, how alliances form and what drives warfare. As researchers tweak the conditions in these artificial worlds, the virtual societies respond.
    "You can actually create social systems that look like cities or nations," said Ken Kollman, a University of Michigan political scientist who makes models of political campaigns at the Institute for Social Research. The basic approach is the same: Strip a social behavior -- such as consumption or trade -- down to its most basic elements and put the computer to work, with a cast of programed characters playing an array of slightly different roles.
    Social scientists use this new tool to search for a unified set of rules that would predict human behavior in their varying fields. That set of rules might allow them to predict the outcome of a change in government policy or understand how people choose where they are going to live.
    But they have a way to go. The field is new enough that researchers don't even agree on the proper terms for some basic concepts, Kollman noted.
    "The simpler the model, the more useful it is, actually," said U-M political scientist Robert Axelrod.
    Researchers must be careful not to think that the models can predict anything yet, but they are useful in identifying some basic truths, he said.
    The models are showing that cooperation is better than war and that teamwork is better than individuality, for example. They also might be showing that a lot of the conventional wisdom in economics may be bunk.
    "Sugarscape," by a pair of Brookings Institution researchers, looks at the distribution of wealth in a basic economy.
    It is populated with "agents" -- lines of computer code that act like individual actors on the landscape. The landscape is a game board-like grid of squares with two big piles of sugar. Each agent is given a simple task: Look as far as it can, move to the square with the most sugar, eat that sugar, repeat.
    Some agents can see farther than others. In addition, they burn up their sugar at varying rates, depending on the metabolism programed into them.
    Once in motion, most of the agents quickly clump around the two piles, while a few nearsighted ones eke out a living in the relatively unsugared badlands. Some of them simply starve to death. Not surprisingly, the agents with the best vision and slowest metabolism amass the most wealth, said "Sugarscape" author Robert Axtell.
    But then Axtell adds such variables as trade, inheritance, sex, language, disease and warfare and the game gets more complicated.
    The "invisible hand" of classical economics that supposedly guides consumers toward a natural balance of supply and demand shows itself in the simpler versions of the game. But when the complicating factors are added, that balance is harder to come by, Axtell said at a scientific conference last month.
    "A lot of conventional wisdom has been turned on its head," said Scott Page, a California Institute of Technology economist who frequently works with U-M's Kollman.
    But conventional social scientists aren't exactly throwing out their books and notes and openly accepting this new form of science, Kollman added.
    Before their results can be truly accepted as reliable and meaningful, the modelers have to do a better job of connecting their findings to real-world measurements, he said.
    Kollman's team takes polling data on several political issues and maps them out on a three-dimensional grid. The result is a mountain-like "landscape" of public opinion, showing where a politician ought to put his policy statements to attract the most votes.
    A "smooth" issue landscape looks like Mt. Kilimanjaro, and the slope of the landscape points pretty consistently toward the high ground of the most popularity. Moving toward the winning position is relatively easy on this kind of turf, and in simulations, both candidates quickly find spots near each other at the summit.
    But a bumpy landscape, in which there are peaks and valleys of public opinion on the map, is more realistic. Sometimes, candidates reach a peak surrounded by valleys and think they've got the winning position, Kollman said. But on the other side of the valley, there may a higher peak. "It's really easy for parties to get stuck on a bad position," Kollman said.
    Real-world polling data from George Bush's last presidential campaign revealed a spot in which the candidate would have lost votes initially by changing his position, but would have reached a higher peak of popularity by making the move, Kollman said. That's the sort of tough call that earns political consultants their fees, he added.
    CalTech's Page uses the same kind of landscape to explore teamwork. Groups of agents try to solve a difficult problem, such as finding the highest spot on the landscape. Page has discovered that 10 relatively unintelligent agents working as a team do better than one smart agent. "Collectively, they can be very, very smart," Page said.
    Similarly, several of Axelrod's models of international relations have found cooperation a better strategy than belligerence.
    But one model turns out differently each time, which sort of blows a hole in the idea that America is a nation of destiny.
    "That's a major question ...," Axelrod said. "If you do it over and over, do you get the same result?"
    But having consistent results, known as "robustness" in social science, is a challenge for artificial societies, Kollman said. If these models turn out different with each run, what kind of science is that?
    Well, it's messy so far. But a lot less messy than the real world of human society, Kollman noted. At least the models can look at human behavior as the result of a bunch of individual actions and reactions, what they call the bottom-up approach. Even with the best polling data and computers, that is exceedingly difficult to do with real society, he said.
    "The computer is such a wonderful tool for being able to take real data and turn it into visuals," Kollman said. "But also to start spinning out hypotheticals. These are ways to spin out slightly different scenarios: If this happens, how does the outcome change?"

Copyright 1997, The Detroit News
Comments?