Parameter Selection of Gaussian Kernel in Machine Learning

Abstract

Parameter selection of Gaussian kernel in SVM affects its generalization ability. It has been decided on experience or cross-validation. However, cross-validation requires an enormous amount of computation time. which causes a difficulty in solving real problems. Several methods for deciding Gaussian kernel have been proposed on the basis of density of training data. But these methods are lack of theoretical verification. In this paper we propose a method for parameter selection of Gaussian kernel in SVM on the basis of generalization error bound.