基于交叉验证的改进RBF分类器设计  

Design of RBF Network Based on cross-validation method

作　　者：何迎生[1] 彭华[1] 段明秀[1] 

机构地区：[1]吉首大学数学与计算机科学学院,湖南吉首416000

出　　处：《微计算机信息》2009年第18期248-250,共3页Control & Automation

摘　　要：本文提出了一种利用k均值聚类算法确定RBF神经网络径向基函数数目、函数中心及宽度,输出层权值由线性方程组确定,而网络参数的优化采用梯度下降法的网络设计方法。为了解决学习样本数据的有限性、RBF网络泛化能力较差和容易出现过拟合等问题,在网络训练中采用了基于交叉验证的归一化网络训练方法。通过仿真实验表明该方法训练所得的网络泛化能力及分类的准确率明显提高,并有效避免了过拟合问题。In the paper ,we propose a method of design RBF Neural Network using the kmeans Clustering algorithm to determine the initial value of the number and center and spread of redial basis function,and all the parameters are optimized with nonlinear optimization technique.ln order to solve the limits of the small dataset and RBF Neural Network＇s generalize ability and prone to overfit problems, we adopted the normalized RBF algorithm based on cross-validation method. The simulation results show that cross-validation RBF algorithm improved the correct rate of classify and RBFgeneralize ability and avoided the overfit problem.

关 键 词：交叉验证 RBF 神经网络 中心 宽度 归一化 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]