一种改进的进化神经网络优化设计方法

An Improved Evolution Optimization of Neural Network

基于传统遗传算法优化神经网络时存在的"近亲繁殖"、基因编码冗余和难以确定隐节点数等问题,提出改进的进化神经网络优化设计方法.通过对网络编码形式的规范,使得基因编码与功能等价类一一对应,从而降低编码冗余;通过节点相关性评价,使得低于某阈值的节点在交叉操作时被排除,从而降低节点冗余;通过把交叉变异概率与种群个体适应度比例相联系,提出自适应交叉变异概率,较好保持种群多样性.仿真实验表明,本方法可以避免"近亲繁殖"以及由此导致的"种群早熟",降低编码冗余,减少学习参数,提高学习效率.

Improved evolutionary optimization of neural network design is proposed,because there are inbreeding and coding redundancy and it is difficult to determine the number of hidden node.There is a one to one correspondence between gene coding and functional equivalence class,through normalizing of network coding,which decreases the coding redundancy.Through the evaluation of node correlations,the nodes whose correlation value is less than a certain threshold would be deleted during crossover operation so that decreasing the node redundancy.Furthermore,through the combination of crossover probability and mutation probability,the diversity of the network could be held.The experiments show that proposed approach of neural network avoids inbreeding and premature convergence aroused by inbreeding,while increasing the learning speed,and reducing the code redundancy and learning parameters.