Simulated creatures evolve and learn.

In nature, most biologists concur, competition, natural selection, and sexual reproduction constitute the driving forces of evolution. Consequently, evolution proceeds undirected, as organisms mate and adapt to a changing environment.

To explore this process of survival of the fittest, computer scientist Karl Sims of Thinking Machines in Cambridge, Mass., has devised a simulated evolutionary system in which virtual creatures compete for resources in a three-dimensional arena.

The creatures, resembling toy-block robots, enter one-on-one contests in which they vie for control of a desired object--an extra cube. Winners--deemed more fit- reproduce, while losers bear no offspring. Sims endows the virtual environment with physical parameters, such as gravity and friction, and restricts behaviors to plausible physical actions.

When the creatures mate, their offsprings' nervous systems and body types reflect a genetic recombination, thus permitting evolution to determine their attributes. Over hundreds of generations, the creatures compete, reproduce, and evolve, learning complex strategies for controlling resources. Through trial and error, they figure out which strategies work, discarding poor techniques and enhancing effective ones.

Some species found successful strategies in the first 10 or 20 generations, Sims says, while others took much longer. Some creatures threw their arms around the cube or leaped on it. Others crawled or somersaulted to it. A few formed leg-like appendages and learned to walk. They also devised strategies for countering opponents--for example, by covering up the cube, pushing it out of reach, or shoving contenders away.

This evolutionary method may enhance artificial intelligence research by giving rise to autonomous computer programs more complex than scientists can currently design. "It may be easier to evolve virtual entities with intelligent behavior than to design and build them" from scratch, says Sims.