Selecting training points of the sequential minimal optimization algorithm for Support Vector Machine

Abstract

The performance of the Gaussian kernel Support Vector Machine (SVM) for regression is influenced by the training algorithm. The training process of SVM is to resolve a Quadratic Programming (QP) problem. When there are amounts of samples, the needed memory will be bigger if we resolve the QP problem directly. At present the Sequential Minimal Optimization (SMO) is an effective method to resolve QP. SMO decompose the QP problem into series of QP problems of two variables, and resolve the problems analytically. There is no operation on matrix in SMO, therefore it is applied easily. The training points influence the convergent velocity of SMO, so a new method to select the training points is proposed, and the proposed approach is evaluated with a series of experiments. The experiments show that the approach is reasonable and effective.