A Framework for Evolution of Behaviour Switching

Abstract

In order to function in the real world, robotic agents must have a repertoire of behaviours to be applied at the appropriate time. Much progress has been made in the field of neuro-evolution for solving multiple tasks, but these suffer from the fact that all tasks must be evaluated according to an objective for each generation, otherwise the task will be forgotten.This thesis proposes a novel artificial neural network structure for evolving agents capable of performing and switching between multiple behaviours while preventing the evolution of new behaviours from interfering with previously evolved behaviours. Effects of the constraints imposed on evolution by the model are explored using a set of experiments with behaviours of varying degree of evolvability.The experiments show that the proposed model is able to produce behaviour switching in the way that was intended. However, the model imposes some constraints on the evolution of new behaviours and these constraints were shown to be detrimental to the search for a solution in some situations and beneficial in others.