Clustering-based fuzzy knowledgebase reduction in the FRIQ-learning

This paper introduces a fuzzy knowledgebase reduction method with applying a clustering technique in the Fuzzy Rule Interpolation-based Q-learning (FRIQ-learning). The FRIQ-learning method stars with an empty knowledgebase, which is a fuzzy rule-base filled only with rules defining the boundaries of the problem space. Then the system builds the rule-base incrementally episode by episode, based on a properly defined reward function. The FRIQ-learning method is finished, when its terminating conditions become true. This case we get the final rule-base as a solution for the given problem. But the constructed final rule-base may contain redundant rules, which can be automatically omitted from the rule-base by reduction methods. The main goal of the paper is to introduce a new, clustering based reduction method, which is suitable for eliminating the unnecessary rules of the rule-base and hence decrease the size of the fuzzy knowledgebase. For demonstrating the benefits of the suggested clustering based fuzzy knowledgebase reduction method, application examples of the "cart pole" and the "mountain car" benchmarks are also discussed briefly in the paper.