EM Based Fuzzy Reinforcement Learning Control for Linear Systems with Input Saturation
Kainan Liu, Linxiang Li, Xiaojun Ban, Shengkun Xie · 2025
This paper introduces a novel control approach for linear systems with input saturation, combining Takagi-Sugeno (T-S) fuzzy models and reinforcement learning. The method utilizes an offline Expectation-Maximization (EM) algorithm to identify antecedent parameters, improving both the accuracy and interpretability of the T-S fuzzy model. The T-S fuzzy model serves to approximate the value function and derive the optimal control law, while reinforcement learning further optimizes the consequent parameters. To address dependence on next-time-step state variables during policy improvement, a gradient-based iteration method is employed. Simulation results validate the effectiveness of the proposed method, demonstrating faster convergence, reduced control energy consumption, and enhanced reliability, while confirming its optimality and convergence guarantees.