SVM model selection with the VC bound

Abstract

Model selection plays a key role in the performance of a support vector machine (SVM). In this paper, we propose two algorithms that use the Vapnik Chervonenkis (VC) bound for SVM model selection. The algorithms employ a coarse-to-fine search strategy to obtain the best parameters in some predefined ranges for a given problem. Experimental results on several benchmark datasets show that the proposed hybrid algorithm has very comparative performance with the cross validation algorithm. © Springer-Verlag 2004.