一种基于强化学习的在线神经模糊控制系统

Reinforcement-Learning-Based On-Line Neural-Fuzzy Control System

为了实现基于非训练数据的神经模糊控制器的在线学习,提出了一种基于强化学习的神经模糊控制系统和相应的学习算法.该控制系统由神经模糊预测器和神经模糊控制器两部分组成,其中,神经模糊控制器采用基于确定度的模糊规则模型作为知识表示形式的扩展型神经模糊网络.在学习算法的设计中,尝试了利用强化信号得到输入状态的“期望输出”,进而将强化学习转化为基于训练数据学习的解决思路.仿真实验验证了所提出的控制系统结构和学习算法的合理性和可行性.

To solve non-training-data-based on-line learning for neural-fuzzy controller, this paper proposes a reinforcement-learning-based neural-fuzzy control system and the corresponding learning algorithm. This control system consists of a neural-fuzzy predictor and a neural-fuzzy controller. Based on fuzzy if-then rules with certainty grades, the extended neural-fuzzy network proposed is used as the structure of the neural-fuzzy controller. In learning process of the control system, by using reinforcement signal to estimate the desired output of an input state, reinforcement learning is treated and solved according to training-data-based learning algorithm. Computer simulations illustrate the rationality and applicability of the proposed control system and the learning algorithm.