By Sally Pobojewski
News and Information Services
Using frog cartoons, the human face, and billiard balls, John H. Holland showed an audience of U-M colleagues attending his March 16 Henry Russel Lecture that computers can learn and evolve in strikingly human ways.
Considered a visionary in the areas of cognition and artificial intelligence, Holland is a professor of psychology and of electrical engineering and computer science. A MacArthur Foundation Fellow, Holland is a pioneer in the development of genetic algorithms—computer programs that can learn appropriate behavior, instead of having the behavior programmed in.
Holland uses computers to study the behavior of complex adaptive systems. Extraordinarily complicated, these systems exist in areas as diverse as international trade, the immune system, human behavior and evolution.
All complex adaptive systems are comprised of individual units, which Holland calls adaptive agents. “Adaptive agents can be organisms, corporate firms, human beings, viruses or bits in a computer program,” Holland explained. “Adaptive agents react to their environment based on a set of rules and they can learn to change behavior based on experience.”
Using their individual strengths and past experiences, rules compete to control the aggregate behavior of the system. If the environment remains unchanged, rules that were successful in the past continue to succeed. But when the environment changes, different rules can take over the system.
“Think of each rule as a middleman in a complex system of economic transactions, such as those that bring food into New York City,” Holland said. “Working independently with no central planning, competition between these middlemen produces a constant four-day reserve food supply for the city. Each middleman tries to buy cheap and sell dear, and the system rewards those who make the best early decisions.”
New rules evolve in complex adaptive systems when basic building blocks—for example, the specific features that make up a human face—are combined in different ways. “With just 100 different noses, eyes, mouths and other facial features, you can create 10 billion new faces,” Holland said. “With the right building blocks, you can anticipate and invent new things. It opens a tremendous range of possibilities.”
Using a genetic metaphor, Holland explained that rules in complex adaptive systems are similar to chromosomes, and the relative strength of rules is similar to evolutionary fitness. “Those rules that gain strength are most likely to reproduce and become parents of the next generation of rules,” Holland said.
Occasionally, rules exchange segments of their genetic or computer code to produce new rules—a phenomenon called crossover. “Crossover is the factor that drives evolution,” Holland said. “It’s a critical mechanism for creating new possibilities in the system.”
In addition to building blocks, rules and crossover, one other component is necessary to produce a computer program capable of learning and evolving. Agents must have some way of identifying themselves to other agents. Holland calls them “tags.” Like receptors on a molecule’s surface or trademarks on products in the grocery store, tags allow agents to advertise their individuality.
As an example, Holland described a computer simulation he has developed for the common strategy game, “Prisoner’s Dilemma,” that uses billiard balls as players. In this simulation, balls moving at random must “choose” whether to cooperate or defect whenever they come in contact with another ball. Each ball is programmed with a set of rules that determines its strategy; some balls always cooperate and some always defect. The balls are allowed to reproduce in proportion to the points they receive as they play the game.
“When the balls were all the same color, they never learned to cooperate,” Holland said. “But when we marked some balls at random with various colors, they were able to pick out useful, fortuitous associations between color and behavior.”
According to Holland, the red tag became a trademark for cooperation—an indicator other balls could use to determine which agents were more likely to cooperate, instead of defecting. Since cooperating balls in the system were rewarded with higher points and since their reproductive success was determined by the number of points they accrued, red-tagged balls soon evolved to take over the computer simulation.
The Henry Russel Lectureship has been awarded annually since 1925 to an active senior member of the U-M faculty for distinguished achievement in research and teaching. It is the highest honor the University can confer upon a member of the faculty.
