Determining the minimally allowed rule-distance for the incremental rule-base contruction phase of the FRIQ-learning

This paper introduces a novel minimally allowed rule-distance determination methodology for the incremental rule-base construction phase of the FRIQ-learning. The FRIQ-learning process starts from an empty Q-function represented by an empty fuzzy rule-base. Then it creates the fuzzy Q-function in an iterative incremental manner by adding new fuzzy rules in cases where the Q-function update seems to be inefficient by tuning the existing rule base. According to the original FRIQ-learning a new rule is inserted if the Q-function update is high, and even the closest existing fuzzy rule is farther than a preset limit. Improper choice of the minimal rule distance limit can lead to unnecessary high number of inserted rules during the initial steps of the iteration. The main goal of this paper is to introduce a new rule-distance limit calculation methodology which can avoid the unnecessary high number of inserted rules in the initial phase of the iteration. For demonstrating the efficiency of the proposed rule-base creation method, the “mountain car” benchmark application example is also discussed briefly in the paper.