一种新的支持向量回归机的模型选择方法

A novel approach of model selection for SVR

针对支持向量回归机的模型选择问题,将模型选择问题转化为一个非线性系统的状态估计问题,然后引入无迹卡尔曼滤波进行求解,提出一种新的基于无迹卡尔曼滤波的模型选择方法(UKF-SVR).对标准数据集和太阳黑子数平滑月均值进行仿真实验,结果表明,UKF-SVR与粒子群算法相比,该方法全局寻优能力更强,保证了支持向量回归机泛化能力的最大化,获得更高的预测精度.

To solve the model selection for support vector regression,we present a novel approach based on unscented Kalman filter（UKF）.first we transform the problem of model selection for SVR into a problem of nonlinear system state estimation,and introduce UKF to solve it.Compare to particle swarm algorithm,experiments on basic data sets and prediction of the smoothed monthly mean sunspot numbers show that the UKF approach has stronger optimization ability,which fully guarantees maximal generalization ability of SVR and obtains higher precision of prediction.