Abstract: Construction of Radial Basis Function Neural(RBFN) networks involves computation of centers, widths of each RBF, and number of RBF in the middle layer. The modified clustering algorithm is used to construct RBFN automatically. The algorithm considers the class membership of training samples, can automatically compute RBF centers and widths, and determines the number of Radial Basis Function(RBF) units based on the distribution of samples. Parameters are optimized with nonlinear optimization technique. The performance of the self-growing RBFN and effects of the optimization are estimated with IRIS classification problem and chaotic time series prediction. The results confirm that self-growing RBFN determines networks structure automatically, and has good performance in pattern recognition and function approximation. Better performance can be observed after nonlinear optimization of networks parameters.

Key words: Radial Basis Function(RBF), self-growing networks, parameter optimization, pattern recognition, chaotic time series

摘要： 径向基函数神经网络的构造需要确定每个RBF的中心、宽度和数目。该文利用改进的聚类算法自动构造RBFN，考虑样本的类别属性，根据样本分布自动计算RBF的中心和宽度，并确定RBF的数目。所有的网络参数采用非线性优化算法来优化。通过IRIS分类问题和混沌时间序列预测评价自构建RBFN的性能，验证参数优化效果。结果表明，自构造RBFN不但能够自动确定网络结构，而且具有良好的模式分类和函数逼近能力。通过对网络参数的非线性优化，该算法明显改善了网络性能。

关键词: 径向基函数, 自构造网络, 参数优化, 模式识别, 混沌时间序列

CLC Number: 

• TP183