A hybrid optimization strategy for simplifying the solutions of support vector machines

The main issue is to search for a subset of the support vector solutions produced by an SVM that forms a discriminant function best approximating the original one. The work is accomplished by giving a fitness (objective function) that fairly indicates how well the discriminant function formed by a set of selected vectors approximates the original one, and searching for the set of vectors having the best fitness using PSO, EGA, or a hybrid approach combining PSO and EGA. Both the defined fitness function and the adopted search technique affect the performance. Our method can be applied to SVMs associated with any general kernel. The reduction rate can be adaptively adjusted based on the requirement of the task. The proposed approach is tested on some benchmark datasets. The experimental results show that the proposed method using PSO, EGA, or a hybrid strategy combining PSO and EGA associated with the objective function defined in the paper outperforms both the method proposed by Li et al. (2007) and our previously proposed method (Lin and Yeh, 2009), and that a hybrid strategy of PSO and EGA provides better results than a single strategy of PSO or EGA.