Evolving Modularity in Robot Behaviour Using Gene Expression Programming

Abstract

Incremental learning [3] and layered learning [4] have been proposed as suitable approaches to improve evolutionary robotic (ER) algorithms by subdividing the required behaviour into simpler tasks. However, incremental learning does not divide the controller to unique task modules and although layered learning subdivides the problem into modules it does not offer continuous learning for the various sub-behaviours. Moreover, both methods involve the modification of the fitness function in every module thus increasing computational overhead.