多分类问题的RBF 二叉神经树网络方法

神经网络是一种普遍使用的分类方法。当类别数目较大时 ,神经网络结构复杂、训练时间激增、分类性能下降。基于两类问题的树网络多分类方法将两分类方法和判决树相结合 ,利用两分类方法来减少神经网络的训练时间 ,利用树型分类器来提高识别率。提出了一种多分类问题的二叉神经树网络结构和训练算法。利用两分类网络的训练结果对类别进行排序处理 ,并应用排序后的类别序号构成树型分类器 ,使可分性最差的类别的识别率提高最大 ,从而提高了整体分类性能。使用径向基函数 ( RBF)网络作为节点网络 ,使节点网络结构适应两类间的可分性 ,从而最终优化了神经树网络的结构。仿真实验表明该方法的分类性能优于现有方法

A Classifier of Binary Radial Basis Function Neural Tree Networks

Neural network has been used for pattern recognition popularly. The training time of neural network for N catalogs classification increases exponentially with N, so it is difficult to deal with large number of catalogs by normal neural networks. Based on binary partition method and decision tree, a binary neural tree network(BNTN) classifier is proposed. Each node in BNTN is a simple neural network which only processes 2 catalog classification. Thus the architecture of BNTN is flexible and expansible, and the training time is reduced largely. The key to construct a BNTN is to sort the classes by separation of each class. We proposed a simple way to calculate separability of each class after radical basis function(RBF) neural networks have been selected as a type of node. Simulation shows that BNTN is better than other classifying methods.