Fuzzy PID controller design using q-learning algorithm with a manipulated reward function

Official URL: http://dx.doi.org/10.1109/ECC.2015.7330914

In this paper we propose a manipulated reward function for the Q-learning algorithm which is a reinforcement learning technique and utilize the proposed algorithm to tune the parameters of the input-output membership functions of fuzzy logic controllers. The use of a reward signal to formalize the idea of a goal is one of the most distinctive features of reinforcement learning. To improve both the performance and convergence criteria of the mentioned algorithm we propose a fuzzy structure for the reward function. In order to demonstrate the effectiveness of the algorithm we apply it to two second order linear systems with and without time delay and finally a nonlinear system will be examined.