The Study On Selection Of Kernel Function In Support Vector Machine

Content:Kernel function selection is the key part and the hard issue of Support Vector Machine.By the kernel trick,the nonlinear data can be mapped into a higher dimensional linear space including the nonlinear data,as a result,the nonlinear learning problem can be solved as a linear learning problem.Recently,researches on kernel function have achieved great progress,however,the selection of kernel function remains to be a hard nut to crack.The main results of this dissertation are as follows:1.We analyze some important properties of mixtures of kernels. In some applications,mixtures of kernels have better effects than a single ordinary kernel.2.We analyze some properties of five types of ordinary kernels in order to determine which type of ordinary kernel can be used to construct mixtures of kernels.In this paper,mixtures of kernels are constructed by RBF kernel,Gauss kernel and Fourier kernel because of their good properties.3.This paper give a MPEC model to select the parameters which has been proposed by DONG Yu-Lin,the problem of how to select parameters are resolved.4.In this paper,we can give some regulations according to which to select a kernel.Based on these regulations,we propose a new kernel selection method——Reasonable Selection Method which is different from traditional methods.5.An face detection experiment has been made and the result shows that our new method is very efficient,we can obtain SVM with more superior performance by this method.