Research Of Parameter Optimization For Support Vector Machine

Support vector machine is a general learning method based on statistical learning theory having shown a lot of performances superior to the traditional methods.Currently it develops rapidly in the algorithm research and practical applications. The parameters of support vector machine have great influence on classification and regression. There is no theory to clear parameter selection. Genetic algorithm imitates natural inheritance and evolutionary phenomenon and forms a global search algorithm. Particle swarm algorithm is a heuristic global search algorithm based on group intelligent. Genetic algorithm and particle swarm algorithm are widely used in intelligent optimization algorithm. This thesis studies support vector machine parameter selection problem using genetic algorithm and particle swarm algorithm, avoiding experiment result is not idea caused by improper parameter. The researches of the thesis are as follows:(1) It optimizes the punishment parameter C, the parameterσof RBF kernel and the parameterεof the loss function using genetic algorithm. Then we do experiments to prove the effect. Compared with optimizations of the punishment parameter C and the parameterσof RBF kernel, the three parameter optimizations shows better predictive accuracy.(2) It optimizes the punishment parameter C, the parameters of RBF kernel function and the parameterεof the loss function using particle swarm algorithm. We do experiments using the parameters optimization algorithm based on genetic algorithm, based on particle swarm algorithm and based on grid search. The optimization algorithms use the same sample set:UCI concrete compressive strength. The results show that the parameters optimization based on genetic algorithm achieve better effect than others. The parameters optimization based on particle swarm algorithm can converge rapidly. The parameters optimization based on grid search gets the worst prediction.The selection of parameters to support vector machine has significant influence. Experiments show that choosing parameter optimization algorithm can significantly increase the accuracy of regression forecast. So we should consider the various methods to choose the appropriate parameters of support vector machine and improve generalization ability.