摘要
本文首先针对BP算法中存在的缺陷对误差函数作了简单的修改，使网络的收敛速度比原来的大大提高．此外本文提出了一种基于线性回归分析算法来确定隐层节点数．当已训练好的网络具有过多的隐层单元，可以用这种算法来计算隐层节点输出之间的线性相关性，并估计多余隐层单元数目，然后删除这部分多余的节点，就能获得一个合适的网络结构．计算机模拟实验结果表明，用这种方法来删除隐层中多余的节点是有效的．
In this paper,a modification to the error function of the BP algorithm is proposed,and theconvergent rate of the networks is improved. After that an algorithm based on linear regression analysisdetermining the number of hidden nodes is proposed. By means of a trained network with too many hidden nodes,we calculate the amount of linear correlation between the hidden nodes' outputs and estimate the redundantnumber of hidden nodes,then pruning the redundant hidden nodes from the initial one, then an appropriatestructure can be obtained. The computer simulation has shown that the method is,valid in pruning theredundant nodes in the hidden layer.
出处
《控制理论与应用》
EI
CAS
CSCD
北大核心
1997年第1期101-104,共4页
Control Theory & Applications
基金
攀登计划项目
国家自然科学基金
关键词
BP算法
误差函数
前馈神经网络
学习算法
BP algorithm
error function
linear correlation: pruning
number of hidden nodes
