Evolution of visually-guided approach behaviour in recurrent artificial neural network robot controllers

Analysis of internal structures of embodied and situated agents may provide insights into the mechanisms underlying adaptive behaviour. This paper is concerned with the evolution and analysis of visually-guided approach behaviour in a simulated robotic agent controlled by a recurrent artificial neural network, whose connection weights have been evolved using evolutionary algorithms. Analysis of the evolved behaviours and their network-internal mechanisms reveals a behavioural structure and organization resembling a Brooksian subsumption architecture. The task decomposition, as well as the resulting individual behaviours and their integration, however, are realized as network-internal state space dynamics, evolved in the course of agent-environment interaction, i.e. with a minimum of designer intervention.