A fuzzy reinforcement learning algorithm with a prediction mechanism

This paper applies fuzzy reinforcement learning along with state estimation to the differential pursuit-evasion game. The proposed algorithm is a modified version of the Q(λ) Learning Fuzzy Inference System (QLFIS) algorithm proposed in [10]. The proposed algorithm combines the QLFIS algorithm with a Kalman filter estimation approach. The proposed algorithm is called the modified Q(λ)-learning fuzzy inference system (MQLFIS) algorithm. The Kalman filter is used by the pursuer to estimate the expected future position of the evader. The proposed algorithm tunes the input and the output parameters of the fuzzy logic controller (FLC) of the pursuer based on the expected future position of the evader instead of the real position of the evader. The proposed algorithm also uses the expected future position of the evader to generate the output of the FLC so that the pursuer captures the evader at the expected future position. The proposed algorithm is used to learn two different single pursuit-evasion games. Simulation results show that the performance of the proposed MQLFIS algorithm outperforms the performance of the QLFIS algorithm proposed in [10].