基于混合学习算法的径向基函数神经网络设计

Design of Radial Basis Function Neural Network Based on Hybrid Learning Algorithms

隐藏层中心点参数的选择和权值向量的快速计算是径向基函数神经网络设计的关键问题.基于"半月"数据集,论文提出了一种上下半月单独计算聚类中心的K-均值聚类、递归最小二乘算法计算权值向量的混合学习算法.基于三层RBF神经网络结构,以支持向量机作为分类器,开展了K-均值+最小均方算法及K-均值+递归最小二乘算法2种混合模式的对比实验.实验结果表明,"K-均值+RLS"算法相比"K-均值+LMS"算法具有更快的收敛性,在应对线性不可分的情况,上下半月单独作用的K-均值聚类算法表现更优越,综合考虑收敛速度及分类精度两个指标,论文提出的上下半月单独计算中心点的K-均值聚类+RLS的混合学习算法获得较优的性能.

The selection of hidden layer central point parameters and the fast calculation of weight vectors are the key problems in the design of radial basis function(RBF)neural networks.Based on the'half-month'data set,this paper proposed a hybrid learning algorithm,i.e.,K-means clustering and recursive least squares algorithm,to compute the weight vectors of clustering centers separately in the upper and lower half-months.Based on three-layer RBF neural network structure,support vector machine(SVM)as classifier,the comparative experiments of K-means+LMS and K-means+RLS are carried out.The experimental results show that the'K-means+RLS'algorithm has faster convergence than the'K-means+LMS'algorithm.In the case of linear inseparability,the K-means clustering algorithm which acts alone in the upper and lower half of the month performs better.Considering the convergence speed and classification accuracy,the hybrid learning algorithm proposed in this paper achieves better performance.