Acquiring Behavioral Rules for Action-Based Robot Environments

Abstract

We propose a working environment frame for actionbased robots and describe learning for behavioral rule acquisition enabling robots to operate in such an environment. Robot geometric information is expressed in the frame, as the interaction of a potential distribution and robot behavior as potential energy. Potential representation for the multiple constraint satisfaction problem of multiple agents is described, then a simple soccer game situation applied as an example of simulation. In the potential distribution environment, the potential view is observed as sensor input to a robot. The robot decides its own behavior based on this view. To adaptively acquire behavioral rules, a classifier system is applied whose condition is an observed potential view, and a series of condition-action rules is acquired to attain the goal. The effectiveness of the proposed method is demonstrated by computer simulation.