Study On The Parameter Optimization In Support Vector Machine

Support Vector Machine (SVM) is a high-performance learning machine on the basis of the statistical learning theory .SVM based on the information of limited samples to search for the best compromise between the complexity of the model and the learning ability, with a view to obtain the best generalization ability. There is a problem in the SVM that it depends on the performance of the parameter settings, including penalties and kernel parameters, but no suitable theory can guide to find adapted parameters. The parameters in the SVM model are analyzed, and existent optimization methods of parameters have been conducted and concluded. Besides , with particle swarm optimization(PSO) and cross-validation method ,a SVM parameters optimization method based on the is proposed, while a method based on Breeding Algorithm(BA), with BA and cross-validation method ,is proposed .Finally, numerical experiments proved the effectiveness of the methods.In addition, to solve unconstrained optimization problems, on the basis of the properties of PSO, which always converges quickly, and the properties of BA, which is not easy to immerse the region of local optimum solution, a new hybrid algorithm is put forward. The results of the experiment showed that the hybrid algorithm has good effect to the function which has many local optimum solutions. The advantage of this algorithm is that, when maximum information is at a standstill gradually, the particles in PSO are able to get the global solution. With the hybrid algorithm and cross-validation method, a SVM parameters optimization method is proposed.