基于增强学习规则的倒立摆模糊神经网络控制器

Fuzzy Neural Network Controller of Inverted Pendulum with Reinforcement Learn ing Rule

为实现模型未知、初始时有脉冲输入的车上单级倒立摆镇定控制,提出了一种采用增强学习规则训练的模糊神经网络控制器。以神经网络构造基于T-S(Tankagi-Sugeno)规则的模糊控制器;用3层前馈网络组成预测器进入仿真,得到倒立摆状态并计算状态预测值,再将状态和状态预测值组成训练数据对,训练状态预测BP(Backward Propagation)网络;利用增强学习的方法训练模糊控制器,根据神经网络产生的模糊控制量和倒立摆状态预测,做出控制决策。此方法简化了模糊控制部分参数调整,亦可应用于其他无模型控制。实验证明,控制器鲁棒性良好,即使在倒立摆参数变化较大时,控制器仍能维持倒立摆平衡。

A training method for fuzzy neural network controller with reinforcement learning is proposed to maintain the balance of a single inverted pendulum on cart. The inverted pendulum model is unknown and it is disturbed by impulse input at the beginning. Build fuzzy controller based on T-S （Tankagi-Sugeno） rules with neural network; Begin the simulation with a forward network of three layers to obtain states and state-predictions; State-prediction BP （Backward Propagation） network is trained with pairs of state and state prediction; Fuzzy neural network controller is trained using reinforcement learning rule. Control decision is made due to fuzzy control and state prediction. This method simplifies adjustment of fuzzy control parameters, and is suitable to other controls without system model. Experiments showe that the control method has sound robustness. Even if the parameters of inverted pendulum change a lot, the controller could maintain the balance of inverted pendulum however.