A new parameter optimization algorithm of SVM

The proper selection of parameters, i.e. RBF kernel parameter g, penalty factor c, non-sensitive coefficient ε of SVM model can optimize the performance of Supporting Vector Machine (SVM). The most commonly used approach is grid search. However, when the data set is large, a terribly long time will be introduced. In order to reduce the selection time of optimal parameters, we propose a new heuristic search algorithm (HS-SVM). The proposed algorithm firstly finds the parameter combinations (c, g) with N minimum MSEs by setting randomly constant ε .The N selected (c, g)pairs are integrated with all possible ε to do cross-validation to get parameter combination (c, g, ε) with minimum MSE. The corresponding parameter ε is regarded as the best. Then this ε is combined with the prior N combinations (c, g) to do the cross-validation of SVM. The parameter combination with minimum MSE is the optimal. Experiments show that the proposed algorithm is more efficient than the gird search.