Toward optimal SVM

Although it was proven by Vapnik that SVM is capable of finding optimal solutions, the full success can be achieved only if the adaptive process parameters (C and kernel functions parameters) are set to proper values. Manual parameters tuning is sometimes tricky. We propose a new way of automatic selection of the parameters. The criterion used to estimate C and kernel parameters aims in optimal accuracy on unseen data and simultaneously in small number of SVM’s kernels. The results achieved show that it is possible to reduce the user interaction to just starting the learning procedure and obtain simple and accurate SVM models.