Antecedent Redundancy Exploitation in Fuzzy Rule Interpolation-based Reinforcement Learning

This paper introduces novel methods which could improve the efficiency of the automated knowledge extraction methods used in the FRIQ-learning (Fuzzy Rule Interpolationbased Q-learning) machine learning method. For solving a given problem, the FRIQ-learning reinforcement learning method is capable of constructing a sparse fuzzy rule-base, which does not need to contain all the possible rules as traditional fuzzy control requires. Hence it is sufficient to keep only the most important rules due to Fuzzy Rule Interpolation (FRI). Finding those specific rules which are important to solve the given problem is not a trivial task. Some possible strategies for removing these kinds of unimportant rules from the rule-base have already been introduced, but no strategies addressing the antecedents of the rules have been developed yet. The solutions proposed in this paper allow the further reduction of these rule-bases, thus facilitating the creation of a sparse fuzzy rule-base from which the knowledge can be directly extracted. Since the form of fuzzy rules are inherently self-describing, the size of the rule-base is the key for keeping this kind of knowledge base human-readable. Possible mechatronics applications of these methods include optimizing behaviour-based control models for robotics, and also knowledge extraction in a fuzzy rule-base format from models where the real operating knowledge of the model is not known, which rule-bases then can be easily adopted in robot control applications.