A Novel Reinforcement Learning Approach for Automatic Generation of Fuzzy Inference Systems

In this paper, a novel approach termed dynamic self-generated fuzzy Q-learning (DSGFQL) for automatically generating fuzzy inference systems (FISs) is presented. The DSGFQL methodology can automatically create, delete and adjust fuzzy rules without any priori knowledge. Compared with conventional fuzzy Q-learning (FQL) approaches which only use reinforcement learning (RL) for the consequents part of an FIS, the most salient feature of the DSGFQL is that it applies RL to generate both preconditioning and consequent parts of the FIS. The preconditioning parts of the FIS are formed by RL approaches as well as the epsiv-completeness criteria. On the other hand, the consequent parts of the FIS are updated by FQL. Compared with our previously proposed generalized dynamic fuzzy neural networks (GDFNN), which is a Supervised Learning (SL) approach, the DSGFQL approach can be applied to situations when the training teacher is not available. Compared with the previously proposed dynamic fuzzy Q-learning (DFQL) and online clustering and Q-value based genetic algorithm learning schemes for fuzzy system design (CQGAF), the DSGFQL approach can delete unsatisfactory and redundant fuzzy rules as well as adjust the membership of fuzzy functions. Simulation studies on a wall-following task by a mobile robot show that the proposed DSGFQL algorithm is superior to DFQL and CQGAF.