Fuzzy Boxes; A Distributed Adaptive Neurocontroller Using Reinforcement Learning

A modified reinforcement learning architecture is presented here as an extension of the seminal implementation of Barto, Sutton and Anderson and is applied to a well known control task . The motivation is to improve the performance of the original system by distributing state information across state-space. By fuzzyfying the fixed state-space boundaries of the original system and modifying the learning algorithm, both the learning-rate and control performance have been improved. A further benefit of this system is that a set of fuzzy rules for the control task is generated automatically.