Reinforced ART (ReART) for Online Neural Control

Fuzzy ART has been proposed for learning stable recognition categories for an arbitrary sequence of analogue input patterns. It uses a match based learning mechanism to categorise inputs based on similarities in their features. However, this approach does not work well for neural control, where inputs have to be categorised based on the classes which they represent, rather than by the features of the input. To address this we propose and investigate ReART, a novel extension to Fuzzy ART. ReART uses a feedback based categorisation mechanism supporting class based input categorisation, online learning, and immunity from the plasticity stability dilemma. ReART is used for online control by integrating it with a separate external function which maps each ReART category to a desired output action. We test the proposal in the context of a simulated wireless data reader intended to be carried by an autonomous mobile vehicle, and show that ReART training time and accuracy are significantly better than both Fuzzy ART and Back Propagation. ReART is also compared to a Naïve Bayesian Classifier. Naïve Bayesian Classification achieves faster learning, but is less accurate in testing compared to both ReART, and Bach Propagation.