A Modified Kernel Function Parameter Selection Method

In the 1990s, Support Vector Machine (SVM) is a new learning method proposed by Vapnik, etc, in the framework of Statistical Learning Theory , which has a strong advantage in solving the small sample, nonlinear and high dimensional pattern recognition problems. When support vector machine is used to solve problems, the choice of kernel functions and related parameters plays a vital role on the results of good and bad, directly affects the classification performance of SVM .Only by selecting the appropriate kernel functions and parameters can it get a good generalization ability of SVM classifier.Although the kernel research on functions and parameters in theoretical and applied aspects is day to mature, but it is not enough to guide the parameter selection. Kernel function parameter is the key factor to SVM classification performance. But international haven't form a unified model of the function parameter selection method currently, the optimal SVM parameter selection algorithm can only with experience, experimental comparison, or a wide range of searching for optimal. In many of the kernel function parameter selection methods, the grid search method is the most commonly used and relatively effective method. Support vector machine kernel function parameter selection method is learned in this paper.The following aspects of the support vector machine kernel functions and parameters are discussed in this paper:First, this paper systematicly and comprehensively summarizes the support vector machine theory, introduces the VC dimension theory , structural risk minimization principle , the SVM classification algorithm, and analysis the strategies for solving multi-class problem.Second, it analyzes the performance of SVM training and several important factors of SVM performance. The commonly used kernel function parameter selection methods are discussed. The advantages and disadvantages of the bilinear search method, the pattern search method, the grid search method are analyzed .Combined with the parameters'spatial distribution characteristics of SVM kernel function parameter, A new kernel function parameter selection method–Bilinear pattern search method is proposed by theoretical demonstration and experimental comparisons in this paper.Finally, both accuracy and running time of the proposed method in this paper and the classical grid search method are analyzed and compared by simulation experiments . SVM trained by the proposed method gets not only a higher learning accuracy but also its learning time greatly shortened, thus proving the superiority of the proposed method, The comparative advantages and feasibility of the proposed method are Analyzed.