关键词: 统计学习理论, 支持向量机, VC维, 参数选择

Abstract: Searching the optimal parameters is one of the most important area of SVM and often named as parameter optimization or parameter selection. The L2-SVM can convert the samples into linearly separable problem. Based on the performance, this paper proposes the L2-SVM with two regularization parameters, and the dual formulation of L2-SVM with two regularization  parameters is deduced. Combining the objective function established on minimizing the VC dimension and the gradient method, a new algorithm called Doupenalty-Gradient is present. Ten benchmark datasets are used in the experiments, and the classifying accuracy is improved obviously. The experimental results show the wonderful property and the feasibility of Doupenalty-Gradient.

Key words: statistical learning theory, support vector machines, VC dimension, parameter selection