一种神经网络参数扰动算法及其在机械手控制中的应用

基于Ｈｏｐｆｉｅｌｄ型网络的收敛特性和Ｋｉｒｐａｔｒｉｃｋ等的模拟退火算法的思想，提出了一种克服Ｈｏｐｆｉｅｌｄ网络的局部极值问题的网络参数扰动算法，它具有类似ＳＡ算法的随机退火的特性。文中通过大量数字模拟分析了该算法的退火性能。

A PARAMETER DISTURBING ALGORITHM FOR NEURAL NETWORKS AND ITS APPLICATION TO MANIPULATOR CONTROL

In this paper, a parameter disturbing algorithm of neural networks which can overcome the local minima problem of Hopfield network is proposed on the basis of considering the gradient convergence of Hopfield network and the principle of stochastic simulated annealing developed by Kirpatrick. It is of similar characteristics of stochastic annealing with typical SA algorithm, whose performance is analysed in terms of digital simulations in detail. Finally, the proposed algorithm is successfully applied to the inverse kinematics problem of manipulator control, and then a neural network solver for the problem is established. It has been shown by digital simulations that the proposed neural network control method not only has fast control speed but also raises the control accuracy of manipulator.