多分辨正交多小波网络的结构与算法研究  

Study on a Multiresolution Orthogonal Multiwavelet Neural Network

作　　者：王军锋[1] 冯象初[1] 宋国乡[1] 

机构地区：[1]西安电子科技大学数学系,陕西西安710071

出　　处：《系统工程与电子技术》2003年第11期1411-1413,共3页Systems Engineering and Electronics

基　　金：陕西省自然科学基金(2000SL02);国防预研基金项目(51487020203DZ0103)资助课题

摘　　要：小波网络是函数逼近的有力工具。利用多小波函数和多尺度函数的互补性,构造了一种具有分层、多分辨和局部学习特点的正交多小波神经网络。分析了该网络用于函数逼近时的性质,讨论了其结构并在此基础上给出了调整网络权系数的算法。理论分析和GHM多小波网络及db2单小波网络对几种非线性函数的仿真结果都表明,当函数具有一定的光滑性时,正交多小波神经网络的逼近性能优于正交单小波神经网络。A new multiresolution orthogonal multiwavelet neural network with hierarchical structure and localized learning is proposed. Its properties for function approximation and the algorithm based on a special structure are also investigated. The theoretical analyses show that an orthogonal multiwavelet neural network has better convergence property than a wavelet network. To make a comparison between both networks, experiments for several nonlinear functions are carried out with GHM multiwavelet orthogonal network and daubechies 2 wavelet network, which show that the multiwavelet neural network converges more rapidly than the wavelet neural network.

关 键 词：神经网络 小波网络 函数逼近 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]