一种权值直接确定及结构自适应的Chebyshev基函数神经网络  

Weights-directly-determined and Structure-adaptively-tuned Neural Network Based on Chebyshev Basis Functions

作　　者：张雨浓[1] 陈裕隆[1] 姜孝华[1] 曾庆淡[1] 邹阿金[1,2] 

机构地区：[1]中山大学信息科学与技术学院,广州510275 [2]广东海洋大学信息学院,湛江524088

出　　处：《计算机科学》2009年第6期210-213,共4页Computer Science

基　　金：国家自然科学基金(60643004,60775050);中山大学科研启动费;后备重点课题资助

摘　　要：基于函数逼近理论,构造一种Chebyshev基函数神经网络模型。推导出该网络模型的权值直接确定方法,可一步计算出权值,克服了传统BP神经网络学习率选取困难、学习过程冗长和易陷入局部极小等缺点。在此基础上,设计了基于二分搜索的结构自适应算法,根据精度要求自动确定网络最优结构。理论分析及仿真验证均表明,该网络不仅能够快速地完成网络权值确定和结构自适应,且具有优异的学习与逼近能力,而且对随机加性噪声也具有较好的抑制作用。A model of neural network based on Chebyshev basis functions was presented to remedy the drawbacks existing in conventional BP neural networks, such as, slow learning rate, local minima, and difficulties in determining optimal structure. The values of the weights in such a neural network can be quickly determined by the proposed weightsdirect-determination method. The structure of such a neural network canalso be adjusted automatically according to the required precision. Theoretical analysis and simulation results both show that the new neural network model and its algorithm can not only avoid some disadvantages of conventional BP networks, but also have a faster determination speed on the neural-network weights and structure

关 键 词：神经网络 Chebyshev正交基 权值直接确定 结构自适应确定 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]