Co-evolution of robot behaviors

One approach to the design of intelligent autonomous robots is through evolutionary computation. In this approach, the robot's behavior is evolved through a process of simulated evolution, applying the Darwinian principles of survival-of-the-fittest and inheritance-with-variation to the development of the robot's control programs. In previous studies, we illustrated this approach on problems of learning individual behaviors for autonomous mobile robots. Our previous work has focused on tasks which were reasonably complex, but which required only a single behavior. In order to scale this approach to more realistic scenarios, we now consider methods for evolving complex sets of tasks. Our approach has been to extend the basic evolutionary learning method to encompasses co-evolution, that is, the simultaneously evolution of multiple behaviors. This paper addresses alternative designs within this basic paradigm. Specifically, we focus on dependencies among the learning agents, that is, what a given learning agent needs to know about other agents in the system. By using domain knowledge, it is possible to reduce or eliminate interactions among the agents, thereby reducing the effort required to co-evolve these agents as well as reducing the impediments to learning caused by these interactions.