Self-Adaptive Mutation in On-line, On-board Evolutionary Robotics

Abstract

On-line, on-board evolution of robot controllers implies an inherent need for adjusting the parameters of the evolutionary algorithm on-the-fly. In this paper we argue that the most influential factor to govern evolution in our application is the mutation operator. To address the problem of adjusting its parameter(s) we identify different on-line parameter control mechanisms and perform an experimental comparison among them. The experiments are carried out in a high quality simulator, We bots, for three different tasks for the robots. The results are not fully consistent over the tasks considered, yet they support a preference for the de-randomised self-adaptive mutation step size control.