对RBF网络中几种常用的聚类算法进行比较,分析了各种方法的优缺点,针对动态聚类法进行了改进,其最突出的缺点是距离门限值固定不变,不适于样本分布密度差异较大的数据集。在改进算法中,首先定义所有样本间的平均距离为聚类门限初值,按照奖罚制度动态调节各中心的距离门限值,然后对所有样本进行动态聚类,以输入样本的平均值作为聚类中心,用最小二乘法求的输出层权值。该算法具有在线自适应聚类能力,所得到的RBF网络是最优的。最后通过著名的煤气炉数据验证了该方法的有效性与实用性。 Several common clustering algorithms in RBF network are compared, the advantages and disadvantages of each method are analyzed, and the dynamic clustering method is improved. The most prominent drawback is that the distance threshold is fixed and not suitable for sample distribution Data sets with large differences in density. In the improved algorithm, the average distance between all the samples is defined as the initial value of the clustering threshold, and the distance threshold of each center is dynamically adjusted according to the reward and penalty system. All the samples are then dynamically clustered to input the average value of the samples as Cluster center, using the least squares method of output layer weights. The algorithm has online adaptive clustering capability, and the resulting RBF network is optimal. Finally, the effectiveness and practicability of this method are verified by the famous gas furnace data.