Research On Kernel Function And Parameter Selection In Support Vector Machine And Its Application

Support Vector Machine (SVM) developed to be the core of statistical learning theory in 1990s, it is a new machine learning method proposed by V.Vapnik of AT&T Bell Laboratories, which solves machine learning problems by means of optimization methods, integrates optimal hyperplane, mercer kernel function, convex quadratic programming, sparse solutions and relaxation etc. several techniques, with the good values of global optimum, simple structure and strong ability to promote. It has some good results in many aspects, such as pattern classification, regression analysis and probability density estimation.However, SVM still has a long way to go. This paper does the researches on four areas:SVM model, kernel function constructing, SVM parameter selection and outlier detection. Details are as follows:First, outlining the basis of the research-statistical learning theory and support vector machine approach, described and compared several training algorithm and distortion algorithm, bedded for the follow-up research content.Second, this paper introduced the fuzzy logic theory, proposed two new hybrid kernel functions based on gaussian kernel function, sigmoid kernel function and gaussian kernel function fuzzy sigmoid kernel function. The two new hybrid kernel functions integrated the benefits of local and global nuclear functions, improved the learning accuracy and reduced time of SVM. Experimental results show that the proposed two mixed kernel functions are better than the traditional nuclear function whether in classification accuracy or classification time.Third, proposed an adaptive hybrid genetic algorithm based on traditional genetic algorithm and gradient algorithm, which applied to the research of the model parameters selection of support vector machine. Simulation results show that the parameters selected by this algorithm are better than the parameters selected by the tranditional genetic algorithm, cross validation and grid search algorithm, improves the recognition accuracy of SVM.Fourth, proposed two isolated-points-detect methods in regression analysis, based on the isolated-points-detect methods inε-SVR and v-SVR regression analysis.The experiment results show that, the two methods can detect the isolated-points accurately and effectively.