A Research About Parameter-optimization Of SVM Based On Multi Swarm PSO About Prediction

Taking Data Mining as the background,this paper takes the prediction of time series data using SVM as the object of study and raises a new method to improve the prediction effect.Aimed at the various parameter-optimization methods of SVR,this paper raises a new parameter-self-learning method,based on Multi-Swarm-PSO and KFCM to improve calculating efficiency,reduces the probability of mature convergence,and ameliorates the original algorithms.The core of the Genetic Algorithm is combined in it.Meanwhile,using power function as the dynamic learning factor to improve calculated performance is another innovative point in this paper.Three strategies(four innovation points)improve the standard PSO,which is widespread used in SVM parameter-optimization,and get more balanced performance.The first strategy is "Improved strategy of KFCM on population division".This strategy includes two innovation points.One is how to combine KFCM and PSO.The other one is to use improved KFCM to divide PSO population.The second strategy is "Improved PSO based on asynchronous dynamic nonlinear learning factor".This part chooses learning factor of PSO as starting point,and raises a new method based on power function to reduce the possibility of local convergence and improve time efficiency to some extent.The third strategy is "Improved PSO based on variation".This part choose the evolutionary process as starting point.Proper variation is added to improve particle diversity and reduce the possibility of premature convergence.These three strategies(four innovation points)cover improving effect of population,raising particle diversity,improving algorithm efficiency and so on.Specific to five different Data Sets,which are on behalf of five type of time series data,this paper compares the new method,Grid Algorithm,the standard Particle Swarm Optimization(PSO),the standard Genetic Algorithm(GA),and Artificial Bee Colony Algorithm(ABC)with 5 groups and 25 experiments to prove this new strategy could improve the fitting accuracy,and balance them compared with the other for algorithms,and could determine the better parameters.The the time efficiency is proper.This is a pervasive algorithm.In the end,this paper points out the lack of SVM on prediction analyzes thedisadvantage and indicate the future research direction.