覆盖框架与主方向神经网络  

Covering framework and primary direction neural network

作　　者：殷维栋[1] 

机构地区：[1]中国科学院半导体研究所,北京100083

出　　处：《计算机应用》2008年第8期2077-2080,2083,共5页journal of Computer Applications

基　　金：国家863计划项目(2006AA01Z123)

摘　　要：分析了BP、RBF和ARTMAP等人工神经网络在实现非线性映射方面的共同之处,基于RBF等网络对于人脑功能方面的模拟和仿生模式识别的思想,总结出一种处理这类问题的基本框架。该框架的特点是将问题分解为样本覆盖问题和基于模型的映射拟合问题。在利用该框架研究某个函数集在连续函数空间中的稠密性的基础上,提出了一种新的人工神经网络模型——主方向神经网络(PDNN)。通过与BP网络和RBF网络在函数拟合和混沌时间序列预测方面的对比实验,发现PDNN具有非常良好的逼近性能和鲁棒性能。Some widely used artificial neural networks such as BP, RBF and ARTMAP were discussed on the common points they shared when solving nonlinear regression problems. A framework, proposed accordingly, suggested that such problems should be implemented in two steps： sample covering and model-based map regression. This idea stems from two aspects, i.e. simulation of human brain＇ s information processing and Biometric Pattern Recognition （ BPR）. As an instance, a new network named as Primary Direction Neural Network （PDNN） was proposed on the ground of the density of certain function set in continuous function space. It is used to solve two kinds of problems, function regression and stochastic time series prediction. Compared with BP and RBF neural network, PDNN has preferable approximation and robust performance.

关 键 词：RBF神经网络 仿生模式识别 覆盖框架 马氏距离 主方向 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]