A Note on General Adaptation in Populations of Painting Robots Submitted to CEC2003

Daniel Ashlock, Elizabeth Blankenship, and Jonathan Gandrud.

Abstract PDF eprint

A population of virtual robots is evolved to perform the task of competitively painting the floor of a toroidal room. Two robots are present in any given room and paint using distinct colors. The fitness of a robot is the amount of floor painted with its own color, a situation where maximal marginal fitness comes from painting over squares already painted in an opponent's color. The time required for a population to settle to a value close to its final average fitness is estimated experimentally at approximately 50 generations. Evolution is then continued well past this estimated settle-down point. The best robots in a given generation are saved at 500 and 5000 generations. The performance of highly evolved and less highly evolved robots is compared by placing the two types of robots into competition. The more evolved robots outperform the less evolved agents, with the empirical estimates of mean fitness differing by more than seven standard deviations. This occurs in spite of a lack of increased fitness of painting robots within their own populations during extended evolution. This result is somewhat at odds with biological dogma, demonstrating general adaptation to the task of painting against opponents never actually encountered. This experiment demonstrates that the quality of the agents as competitive painters is not completely documented by their own in-population fitness numbers. This sort of general adaptation in a competitive task has been observed before in another context, the iterated prisoner's dilemma. This study serves as additional evidence for a form of general adaptation in evolutionary computation systems using an agent-vs-agent competitive fitness function.