Generation of Probabilistic Fuzzy Rule by Reinforcement Learning

Rule base is the most important part of a fuzzy inference system. Inconsistent data make some challenges in generating of fuzzy rules. In these cases, since there are multiple outputs for the same states, hence making decision for suitable consequence selection in each rule is a big challenge. Averaging of inconsistent states has been adopted by current methods and they create output with average of related consequences. The initialization of actions selection probability in fuzzy reinforcement learning based on architecture Actor-critic is used in this method. In this method, training data is clustered and zero order Sugeno method with number of candidate action in each rule are used for the initialization of the actor module parameters and they are online tuned with adopting actor-critic and reinforcement signal finally. There are many inconsistent challenges in robot navigation data in comparing other cases. Therefore the proposed method is used in robot navigation problem. The experiments are done for e-puck robot in Webots simulation. Results show that proposed method has reduced training time, collision to obstacle and fuzzy rule numbers.