Improving the performance of a class of adaptive fuzzy controller based on stable and fast on-line learning algorithm

Abstract

In this paper, a novel stable and fast nonlinear learning scheme called adaptive Takagi–Sugeno–Kang fuzzy controller using a reinforcement actor-critic learning (ATSKFC-RACL) is proposed. In the adopted ATSKFC-RACL approach, the actor is represented by the TSK fuzzy controller, which the parameters are updated on-line according to the proposed cost function and the critic output. The proposed cost function relies on the change of the reward function, the change of the actor parameters, the actor error and the change of the actor error, which gives the feedback from the external environment to the critic network to facilities the learning through its own faults and accelerates it. The critic network is implemented by a neural network, which the parameters are updated on-line using the novel laws that depend on Lyapunov criteria to avoid the shortcoming of the gradient descent such as sucking in local minima or instability. Furthermore, Lyapunov stability analysis is studied for choosing the conditions of the learning rates to guarantee the stability of the proposed ATSKFC-RACL. The learning of the proposed ATSKFC-RACL scheme is faster than the TSK fuzzy controller due to the critic network and the proposed cost function. Furthermore, the proposed scheme can reduce the effect of the system uncertainties and disturbances. The proposed scheme is applied for controlling nonlinear systems. The results show that the performance of the proposed ATSKFC-RACL approach is a stable and fast compared with other existing techniques.