摘要
针对支持向量机在线训练算法训练速度较慢和无法处理边缘支持向量集合为空的缺点,以KKT条件和拉格朗日乘数法为基础,用严格的数学推导得到一种改进的训练算法.通过建立一个矩阵缓存来保存与核函数相关的数据,给出在算法中有效操作该矩阵缓存的方法以加快训练速度;边缘支持向量集合为空时,修改模型的偏值项使样本进入该集合,训练算法得以继续运行;并讨论了该算法在在线系统辨识中的应用.仿真实验和分析结果表明:对于非线性时变系统,改进算法的建模精度较高,训练速度较原算法有了很大的提高.
In order to increase training speed of original on-line support vector machine training algorithm, and to deal with the case of margin support vector set being empty, a new algorithm was obtained by strict mathematical deduction based on KKT condition and Lagrangian multiplier method. After a matrix cache was constructed to save data related with kernel function, the method to handle the cache effectively in the algorithm was presented to increase training speed. When the margin support vector set was empty, some sample was added into this set by updating the model bias, so that the training algorithm could run again. Also application of the algorithm to on-line system identification was discussed. The simulation and analysis results show that the modeling precision of the new algorithm is improved for nonlinear and time-varying system, and training speed is increased remarkably compared with that of the original algorithm.
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2004年第12期1642-1645,1649,共5页
Journal of Zhejiang University：Engineering Science
基金
国家"973"重点基础研究发展规划资助项目(2002CB312200).
