Determination of Optimal SVM Parameters by Using GA/PSO

The use of support vector machine (SVM) for function approximation has increased over the past few years. Unfortunately, the practical use of SVM is limited because the quality of SVM models heavily depends on a proper setting of SVM hyper-parameters and SVM kernel parameters. Therefore, it is necessary to develop an automated, reliable, and relatively fast approach to determine the values of these parameters that lead to the lowest generalization error. This paper presents two SVM parameter optimization approaches, i.e. GA-SVM and PSO-SVM. Both of them adopt a objective function which is based on the leave-one-out cross-validation, and the SVM parameters are optimized by using GA (genetic algorithm) and PSO (particle swarm optimization) respectively. From experiment results, it can be concluded that both approaches, especially PSO-SVM, can solve the problem of estimating the optimal SVM parameter settings at a reasonable computational cost. Further, we point out the importance of a proper population size for GA/PSO-SVM, and present the recommended population size for GA-SVM and PSO-SVM.