An approach to tune fuzzy controllers based on reinforcement learning

This paper proposes a new approach for the tuning of fuzzy controllers parameters based on reinforcement learning. The architecture of the proposed approach comprises of a Q estimator network (QEN) and a Takagi-Sugeno type fuzzy inference system (FIS). Unlike the most of the existing fuzzy Q-learning approaches, which select an optimal action based on finite discrete actions, while the proposed controller obtain the control output directly. With the proposed architecture, the learning algorithms for all the parameters of the Q estimator network and the FIS are developed based on the temporal difference methods as well as the gradient descent algorithm. The performance of the proposed design technique is illustrated by simulation studies of a vehicle longitudinal control system.