Evolutionary Strategy Using Statistical Information and Its Application to Mobile Robot Control

Abstract

We describe an evolution strategy (ES) using the statistical information of subgroups obtained automatically by a similarity metric of individuals at each generation. Arithmetical crossover is done with an elite individual and a mean individual within each subgroup to produce offspring. Standard deviation calculated within a subgroup is used in mutation. The effectiveness of the proposed ES is first shown with tests of the 5 De Jong functions. The present ES is also applied to the acquisition of control for a terminal control problem in an omnidirectional mobile robot, in which robot control is based on fuzzy behavior-based control that combines subsumption-like architecture and fuzzy reasoning.