Automatic Generation of Fuzzy Inference Systems Using Unsupervised Learning

This paper presents a new approach of online generation and tuning of fuzzy inference systems (FIS) using unsupervised learning (OGFIS-UL). The proposed approach is capable of generating the antecedents of the FIS and selecting the best consequents automatically. The antecedents are generated and tuned using the fuzzy multi-agent structure learning (FMASL) and the consequents are selected using reinforcement learning (RL). In our approach, the FIS for a complex task is generated based only on experience, assuming no a priori knowledge of the task and the environment. The FMASL generates compact FIS using competitive agent learning and is capable of generalizing any input in the operating range. The actor-critic learning is modified to select the conclusion part of the FIS in an online generated fuzzy environment. The performance of the algorithm is elucidated using the cart-pole balancing problem. Comparative studies with the rival penalized competitive learning (RPCL), adaptive heuristic critic (AHC) and fuzzy actor-critic learning (FACL) methods demonstrate the superiority of the proposed algorithm