Abstract:

Support vector machine (SVM)is a kind of learning method based on kernel. Kernel function selection has significant influence on the performance of SVM, so how to effectively select the kernel function is an important problem in the field of SVM research. At present most of the kernel function selection methods neither consider the characteristics of data distribution, nor make full use of the implicit prior information in the data. We introduce a concept of energy entropy; along with the super sphere description and measurement features of kernel function, we put forth a method of kernel function selection based on the energy entropy of sample distribution so as to improve the learning ability and generalization ability of SVM. Simulations on numerical examples show that the method is feasible and effective.

Key words: SVM;kernel function；sample distribution；prior information；energy entropy