一种优化神经网络结构算法

给出了一种优化三层神经网络结构算法。首先以较大隐层节点数进行学习，然后根据隐层输出信息提取各节点之间的线性特征来优化隐层节点数。对隐层输出信息提取各节点之间的线性特征，给出了两种方法：一种是在ＢＰ神经网络迭代后用自适应线性单元来提取隐层输出各节点之间线性特征，另一种是在ＢＰ神经网络迭代时就尽量使隐层输出各节点之间呈线性，然后用上种方法来提取隐层输出各节点之间线性特征。实例验证，后一种比前一种能更好地优化ＢＰ神经网络结构。

A Algorithm for Optimizing the Neural Network Structure

This paper presents a algorithm for optimizing the nureal network structure.First,learning is performed with a large number of hidden layer nodes,then thenumber of hidden layer is optimized from the linear characteristic,which is derivedfrom outputs of hidden layer. This paper presents two methods of deriving linearcharacteristic from outputs of hidden layer, l) it is got using adaptive linear element after the iteration of BP neural networks, 2) keep a linear relationship among the nodeoutputs during the iteration of network, then the linear charaCteristic is calculated usingmethod l).Examples show that the later method is better than the former method.