改进CAS性能的多网络表决模型

A MULTI-NET VOTING MODEL FOR IMPROVING CAS PERFORMANCE

Fahlm an和 L ebiere提出的级联相关网络是一个典型的自适应神经网络的增长算法〔1〕,它具有灵活、高效的特点 .但由于该算法存在诸多的不确定因素 ,致使在其增长过程中引入过多的自由参数 ,它和随机选取的初始权重是导致单个神经网络过拟合的两个直接原因 .本文提出的多网络表决模型的基本思想是 ,利用多个网络来对未知的模式进行表决来确定其解 ,由于其平均效应 ,它能够避免单个网络预言带来的偏颇 ,获得满意的结果 .利用我们建立的 PC-FARM计算环境 ,本文还从实验上验证了网络表决模型的优越性 .

The Cascade correlation neural network(CAS), proposed by Fahlman and Lebiere, is typically a adaptive algorithm for constructing neural network with high flexibility and efficiency, but it suffers from serious overfitting which results directly from overabundent free parameters and randomly selected weights. In this paper a multi net voting model is presented for improving CAS performance, it can avoid overfitting through collective effect of multi CASs combination, and reach a satisfatory solution. With our PC FARM network computing environment, the simulation experiment is achieved on two spiral classification problem, the reslut shows its apparent benifit in generalization perfomance.