Reinforcement Learning-Based Optimization for Interval Type-2 Fuzzy Unknown Nonlinear System: An Zero-Sum Control Scheme.

Abstract

A novel mode-free reinforcement learning (RL) algorithm is proposed for the optimal control of unknown nonlinear system presented by the interval type-2 fuzzy (IT2F) model. The optimal control is converted into a zero-sum game, where the control input and the external disturbance are the opposing competitors. Based on the RL method, a model-based policy iteration (PI) algorithm is constructed to solve the fuzzy stochastic coupled algebraic Riccati equations of nonlinear system. Considering that the dynamic parameter is difficult to obtain completely in real systems, a model-free fuzzy control algorithm is designed under the framework of the RL algorithm. An optimal control of IT2F system is realized without parameter information, only requiring the state and input information. Furthermore, the asymptotic stability with $H_{\infty } $ performance index is ensured by the Lyapunov function. Finally, the effectiveness is tested by the semi-car active suspension model (SCASM).