一类非线性系统的递归神经网络控制

A CLASS OF NONLINEAR SYSTEMS CONTROLLED BY RECURRENT NEURAL NETWORKS

递归神经网络（ＲＮＮ）由于具有极为丰富的动力学行为而被用在非线性控制系统中，但是传统的ＲＮＮ的训练算法不仅复杂，而且容易陷入局部极值点．作者通过对标准的遗传算法的改进，采用格雷编码方式、比赛选择策略和稳定状态更替方式，使其适用于ＲＮＮ控制器结构和参数的确定．此外，把ＲＮＮ控制器用于控制一类ＳＩＭＯ自不稳定非线性系统（ 如倒立摆系统） ，也收到很好的效果．

Recurrent Neural Networks (RNNs) can be used in the nonlinear control systems because of their rich dynamical behavior. However, conventional training algorithms are not only complicated, but also apt to get trapped in the local minima. In this paper, the standard genetic algorithms are improved. Gray scale encoding, tournament selection and steady state progression are adopted to determine the structures and parameters of RNNs. Satisfactory results are also gained by the experiments where one class of absolutely unstable nonlinear SIMO systems ( e.g . inverted pendulum) are controlled by RNN controllers.