ENHANCEMENTS OF FUZZY Q-LEARNING ALGORITHM

Fuzzy Q-Learning algorithm combines reinforcement learning techniques with fuzzy modelling. It provides a flexible solution for automatic discovery of rules for fuzzy systems inthe process of reinforcement learning. In this paper we propose several enhancements tothe original algorithm to make it more performant and more suitable for problems withcontinuous-input continuous-output space. Presented improvements involve generalizationof the set of possible rule conclusions. The aim is not only to automatically discover anappropriate rule-conclusions assignment, but also to automatically define the actual conclusions set given the all possible rules conclusions. To improve algorithm performance whendealing with environments with inertness, a special rule selection policy is proposed.