Research On Parameters Optimization And Application Of Mixed Kernel Support Vector Machine

As a typical kernel machine learning method, Support Vector Machine (SVM) is directly affected by selection of kernel function and its parameters. Global kernels are strong in generalization performance, but weak in learning ability while local ones are the reverse. At present, the structural method of the SVM kernel function is commonly by mean of mixing two types of kernel functions. The introduction of mixed kernel function makes the SVM add an adjustable parameter named weight coefficient. The parameter optimization for the SVM with mixed kernel function only includes penalty factor and kernel parameter. The weight coefficient is taken generally empirical value, which makes the parameter combination in the SVM unable to achieve the global optimization. Therefore, the thesis researches mainly on the integration optimization for all the parameters of the SVM with mixed kernel function. This work is as follows:(1)Momentum Particle Swarm Optimization (MPSO) algorithm is used to optimize parameters of the mixed kernel SVM. The basic PSO algorithm has slow convergence speed in later evolution periods and is easy to produce oscillations, so momentum term is introduced and the MPSO is constructed. The MPSO algorithm can effectively improve the convergence speed and partly avoid oscillations. Then it is used to optimize parameters of the mixed kernel SVM. By classification of the UCI data sets, the conclusion is drawn that this algorithm can extract effectively the best parameter combination. The SVM obtained is improved in generalization performance, and the MPSO algorithm is faster than the basic PSO algorithm in the evolution generalization, and the classification accuracy obtained by testing is superior to other common classification algorithms.(2)Improved Genetic Algorithm (IGA) is used to optimize parameters of the mixed kernel SVM. In the IGA, Population is initialized by chaos, which can produce more effective genetic types in initial population to ensure diversity of the population and effectively relieve it into local optimal solution. And an adaptive crossover and mutation operator is used in genetic operation, which can ensure diversity of the population, as well as improve the algorithm convergence speed. Then the IGA is used to optimize parameters of the mixed kernel SVM. The UCI data sets classification and obstacles recognition in front of vehicles are realized by Matlab programming. Finally compared with other related algorithms, it is showed that the proposed algorithm can extract better optimal parameter combination, and higher classification accuracy is obtained in classification issue.