基于统计分析的分阶段进化神经网络方法

基于统计分析和分阶段进化，提出一种新的进化神经网络设计方法．本文方法 的进化过程分三个阶段：第一阶段，首先按训练样本统计特性设计较小规模的神经网络；第 二阶段，引入所有训练样本，在第一阶段的基础上，逐步扩展网络结构，新添加的神经元总 是单独训练并以抵消原网络的输出误差为其训练目标，直至训练网络达到误差要求．第三阶 段，利用统计方法，将网络中非线性变换作用相似的神经元合并，简化网络结构．本文方法 一方面减轻了进化算法的压力,另一方面指出了网络进化的方向使得进化网络的学习过程不 再是黑箱问题．计算机仿真实验表明，该方法是有效的．

PHASE-WISE METHOD FOR EVOLUTIONARY ARTIFICIAL NEURAL NETWORKS BASED ON STATISTICS ANALYSIS

With the statistics analysis and the phase-wise method, the novel evolutionary artificial neural networks method is proposed in this paper.The method includes three phases:the first phase,the small scale neural network is designed according to the statistics characteristics of training samples;The second phase,with all training samples, developes the construction of the neural network in the first phase.The added neurons are trained alonel and its trainning aim is to conuteract the error of primary neural network till getting the desired network error; The third phase,to simplify the construction of the neural network, merges the neurons in which nonlinear affection is similar by statistics analysis.On the one hand,the method decreases the compression of evolutionary algorithms.On the other hand,the method indicates the correct way of evolutionary artificial neural networks and makes the evolutionary learning process be no longer black box.The validity of the algorithm is shown by computer simulation results.