Abstract: To solve the problem of model selection for support vector machine(SVM) classifier, a featurespacebased class separability measure(FCSM) was proposed. With this measure, the combination coefficients of multiple Gaussian functions were optimized. Compared with the kernel matrix evaluation measure (FSM), the new measure has fewer limitations in the application of kernel optimization, and has better theoretical guarantees. The experimental results show that the proposed algorithm outperforms the crossvalidation method, the radius margin bound method and the FSM based method, and moreover, it achieves better performance on SVM classifier with the optimal kernel selected from a wider range of function set.