Abstract: Aiming at the various parameter-optimization methods of SVR, this paper proposes a new parameter-self-learning method, called MultiSwarmPSO, based on Multi-Swarm-PSO and KFCM to improve calculating efficiency, reduce the probability of mature convergence, and ameliorate the original algorithms. In this method, k-CV is blended in to improve the efficiency. Using power function as the dynamic learning factor to improve calculated performance is another innovative point in this paper. Specific to five different data sets, this paper compares the new method with grid algorithm, the standard PSO, the standard genetic algorithm, and artificial bee colony algorithm, the results show that this new method could improve the time efficiency and the fitting accuracy compared with the other two algorithms, and could determine the better parameters.

Key words: SVR, KFCM, PSO