The Heuristic Algorithms for Selecting the Parameters of Support Vector Machine for Classification

The performance of Gaussian kernel Support vector machine (SVM) for classification is determined by scale parameter $\sigma$ of Gaussian kernel function and error penalty parameter C. A heuristic approach is proposed to tune the parameters of SVM in this paper. We firstly select $\sigma$, and then search the optimal value of C with given $\sigma$. By viewing selection of $\sigma$ as a recognition problem, we determine the reasonable range of $\sigma$ using Fisher statistical expression. In selection of C, the search interval is chosen according to the Sequential minimal optimization (SMO) procedure, and the searching procedure is terminated with considering the balance between generalization capability and approximation capability of SVM. The proposed approach is evaluated with a series of real-world data sets.