一种自适应核增量优化覆盖的RBF网络研究

Research of RBF network based on adaptive kernel incremental optimization coverage

为了生成精简的径向基函数(radial basis function,RBF)网络结构并改善网络的泛化性能,提出了一种自适应核增量覆盖的优化方式。首先,利用样本空间的全局分布信息,构建了势函数聚类的模型以实现对样本空间不同区域的逐次覆盖,可以增量生成RBF网络隐节点的个数及核参数;然后,通过挖掘网络隐节点所覆盖区域的邻域信息,以异类样本作用力的方式实现对隐节点核参数的优化调整。通过度量核映射矩阵中不同向量的相似性程度来进一步删除冗余隐节点。将新算法应用于人工数据集及基准数据集,结果表明,该算法可以生成精简的RBF网络结构,同时有效改善了RBF网络的泛化性能。

To generate the compact radial basis function(RBF) network structure and improve the generalization performance of the network, an optimization method of adaptive kernel incremental optimization coverage is presented. Firstly, using the global distribution information of the sample space, a potential function clustering model is constructed to realize the successive coverage of different areas of the sample space,and the number of hidden nodes and kernel parameters of RBF network can be incrementally generated. Then, by mining the neighborhood information of the area covered by the hidden node,the kernel parameters of the hidden node can be optimized in the way of heterogeneous sample forces. The redundant hidden nodes are further removed by measuring the similarity of different vectors in the kernel mapping matrix.The presented algorithm is applied to artificial data sets and benchmark data sets, the results show that the algorithm can generate the compact network structure,and improve the generalization performance of RBF network effectively.