Online Evolution of Adaptive Robot Behaviour

Abstract

The authors propose and evaluate a novel approach to the online synthesis of neural controllers for autonomous robots. The authors combine online evolution of weights and network topology with neuromodulated learning. The authors demonstrate our method through a series of simulation-based experiments in which an e-puck-like robot must perform a dynamic concurrent foraging task. In this task, scattered food items periodically change their nutritive value or become poisonous. The authors demonstrate that the online evolutionary process, both with and without neuromodulation, is capable of generating controllers well adapted to the periodic task changes. The authors show that when neuromodulated learning is combined with evolution, neural controllers are synthesised faster than by evolution alone. An analysis of the evolved solutions reveals that neuromodulation allows for a more effective expression of a given topology's potential due to the active modification of internal dynamics. Neuromodulated networks learn abstractions of the task and different modes of operation that are triggered by external stimulus.