Categorization of continuous numeric percepts for reinforcement learning

We present a method to acquire rules for agent´s behavior. In our method, continuous numeric percepts are categorized by the Fuzzy ART and Q-learning is employed to acquire rules. Although the number of categories affects both the quality of rules and computational costs for acquiring them, the Fuzzy ART monotonously increases the number of categories. Since too many states (categories) may cause consuming many computational costs for acquisition of rules, we control the number of categories. In our method, a meaningless category is integrated with a category that is similar to the meaningless category. The LRU (least recently used) algorithm is used for the detection of meaningless categories and weight vectors in a Fuzzy ART neural network are updated in order to integrate two categories. The method mentioned above has been implemented and some experimental results have been shown.