Learning Curves of Support Vector Machines

A support vector machines (SVM) is known as a pattern classifier with a high generalization ability and one of its advantages is that the generalization ability is theoretically guaranteed. However, many of the analyses are given in the framework of the PAC learning and the error-bounds are rather loose than the practical generalization error. In this paper, we present some studies on the average generalization error of SVMs, which is a more practical criterion for generalization ability