一种改进的在线核函数算法

Improved online learning with kernels

在基于自适应选择核函数基础上,通过引用朴素正则风险最小化准则,提出了一种改进的在线核函数算法。算法在时间窗向前移动的同时,依靠LS-SVM截断误差最小化算法,选取合适的拉格郎日因子,对新增的样本数据进行重新训练。有效地克服了现有方法收敛精度低和不能自适应选择样本的困难,能广泛应用在分类、回归和奇异值检测中。数值仿真结果表明,该算法与现有的算法相比具有预测精度高,泛化能力强等特点。

On the basis of adaptive selective kernel method, an improved online kernel function algorithm is presented by using naive-regular risk minimization rule. When time window slide forward, increased training observations could be retrained based on LS-SVM （least squares support vector machines） algorithm and appropriate Lagrange factor selection. Computer simulation shows this method can improve convergence precision and overcome the deficiency of adaptive sample selection and can also be broadly used in classification application, regression and singular value detection, which results in higher performance in prediction precision and generalization ability.