ART-R: a novel reinforcement learning algorithm using an ART module for state representation

Abstract

The work introduces a neural network (NN) algorithm capable of merging the fast and stable learning behaviour offered by the adaptive resonance theory (ART) and the advantageous properties of a reinforcement learning agent. The result is ART-R a neural algorithm particularly suited to learning state-action mappings in control applications. A real time example addressing a typical problem found in autonomous robotic assembly is discussed to highlight the achievement of unsupervised and fast learning of an optimal behaviour.