Evolutionary strategies for multi-scale radial basis function kernels in support vector machines

Abstract

In support vector machines (SVM), the kernel functions which compute dot product in feature space significantly affect the performance of classifiers. Each kernel function is suitable for some tasks. A universal kernel is not possible, and the kernel must be chosen for the tasks under consideration by hand. In order to obtain a flexible kernel function, a family of radial basis function (RBF) kernels is proposed. Multi-scale RBF kernels are combined by including weights. Then, the evolutionary strategies are used to adjust these weights and the widths of the RBF kernels. The proposed kernel is proved to be a Mercer's kernel. The experimental results show that the use of multi-scale RBF kernels result in better performance than that of a single Gaussian RBF on benchmarks.