A novel model for selecting parameters of SVM with RBF kernel

Based on the viewpoint of similarity measurement, researched the influences of the error penalty parameter C and the RBF kernel parameter σ on support vector machine´s generalization ability. As the result, the parameter C adjust the similarities between the sample categories and σ adjust the similarities among the samples, C and σ mutually restrict and balance each other in a certain range, the shape of the optimal parameter range like a fan, the more reasonable parameters´ value locate at the center of the fan, where the values of C and σ are smaller. A novel method for selecting parameters was presented, firstly, roughly grid searched the reasonable parameter range with a big step size, then selected the optimized parameters in the delineated area through bilinear-grid search method finely. Experiment results show that the improved method has a better performance both at accuracy and speed, moreover, which can avoid excessive values and enhance the stability.