基于改进滑动时窗的在线最小二乘支持向量机

ONLINE LS-SVM BASED ON MODIFIED SLIDING TIME-WINDOW

针对最小二乘支持向量机的在线建模问题,对目前采取的普通滑动时窗策略进行改进,基于矩阵分块求逆原理提出了改进递归算法,增加了样本优选功能;针对如何判断新样本是否产生新信息的问题,提出了两种典型方法,样本相关度检测法和模型预测误差法。实验证明,该方法简捷有效,提高了模型预测精度。

Aiming at online modelling problem of LS-SVM,we modified the slide time-window strategy generally used at present,presented the modified recursion algorithm based on the theory of blocking inversion of matrix,and added the sample optimal selection function.We presented two typical methods which are the sample correlation degree detection and the error forecasting of model,to solve the problem of judging whether the new sample brings forward new message.Experiment validated that this method is effective and simple,and can improve the precision of model forecasting.