Research Of Parameter Optimization For Support Vector Regression

The performance of support vector regression (SVR) has a vital relationship with parameter selection, but even now, there is not a definitive mathematical theory to guide how to choose the optimal parameters. For the purpose of solving this problem, the parameter selection of SVR is transformed to a combinatorial optimization problem. Then the optimization problems is solved by both of intelligent algorithms.Aiming at the deficiency of artificial bee colony algorithm (ABC), the weighting function, the present-optimal food source and the chaotic search algorithm are introduced to improve the update forms of ABC and the search method of scout bees. Then the SVR model is proposed based on improved ABC. Numerical experimentations indicate that the improved algorithm is feasibility and superiority. Taking the short-term traffic flow data as an example, the predictive results of the improved ABC-SVR is compared with ACO-SVR?PSO-SVR and ABC-SVR, the results indicate that the improved ABC-SVR is superior to the predictive effect of the other three, its run time is the shortest and shows good generalization ability and learning ability.Aiming at the disadvantage of artificial fish swarm algorithm (AFSA) with the late slow convergence speed and low precision, the ABC is introduced to improve the search efficiency of AFSA and the AFSA-ABC hybrid algorithm is obtained. Then it is used for SVR parameter optimization. Numerical experimentations show that the algorithm can solve the problem of parameter optimization. The model is applied to predict the GDP of Shanghai City, the results indicate that the model is superior than the existing prediction model, it provides a new way for the prediction of GDP.