A Method to Determine the Hyper-Parameter Range for Tuning RBF Support Vector Machines

A method to determine C,γ , the hyper-parameters, range for Radial Basis Function Support Vector Machines (RBF SVMs) is proposed. The γ range is determined by the extreme Squared Euclidean Distance (SED) quantiles of the training set, and the C range is determined by one pass whole training set training decreasingly along logγ_max to the over-regularized limit first and increasingly along logγ_median to the over-fitted limit then. We will report detailed analysis and experiments that well justify the proposed method. The major contribution of this method lies in that it provides a well principled and easy to practice way to set a much smaller C,γ space and hence efficient range for the conventional Grid Search with V-fold Cross-Validation (GS V-FCV) exhaustive hyper-parameter tuning method. Performance tests reveal that training SVMs using GS V-FCV in the C,γ range determined by the proposed method can effectively reduce the tuning time while the exhaustive capability and best test error rate are still preserved.