Evolving hyperparameters of support vector machines based on multi-scale RBF kernels

Abstract

Kernel functions are used in support vector machines (SVMs) to compute dot product in a higher dimensional space. The performance of classification depends on the chosen kernel. Each kernel function is suitable for some tasks. In order to obtain a more flexible kernel function, a family of RBF kernels is proposed. Multi-scale RBF kernels are combined by including weights. These kernels allow better discrimination in the feature space, and are proved to be the Mercer's kernels. Then, the evolutionary strategies are applied for adjusting the hyperparameters of SVM. Subsets cross validation is used to be the objective function in evolutionary process. The experimental results show that the accuracy of the proposed method is better than the ordinary approach.