A neural-fuzzy BOXES control system with reinforcement learning and its applications to inverted pendulum

In this paper, a neural-fuzzy BOXES control system with reinforcement learning is proposed. The fuzzy box implemented by neural networks is used to divide the state space instead of partitions of quantization given by Michie and Chambers (1968), which makes the fuzzy connectionist model to have more generalization abilities. The reinforcement learning algorithm in the control evaluation network and the gradient descent learning algorithm in the control selection network are derived. The local psi-COA defuzzification method is also presented. An example of inverted pendulum is given, and the simulation results illustrate the superior performance of the proposed fuzzy connectionist model