利用矢量基学习和自适应迭代算法改进LSSVR  

Using vector-base learning and adaptive iterative algorithm to improve LSSVR

作　　者：王鲜芳[1] 杜志勇[2] 

机构地区：[1]河南师范大学计算机与信息技术学院,河南新乡453007 [2]河南机电高等专科学校,河南新乡453007

出　　处：《计算机工程与应用》2012年第6期38-41,共4页Computer Engineering and Applications

基　　金：国家自然科学基金(No.61173071);河南省基础与前沿技术研究计划项目(No.112300410254);河南省科技攻关计划项目(No.112102210412);河南师范大学10博士科研启动课题(521)

摘　　要：针对最小二乘支持向量回归缺乏传统SVR的稀疏性和鲁棒性等问题,综合矢量基学习和自适应迭代算法的优势,提出了一种改进的加权最小二乘支持向量回归算法(LSSVR)。该算法通过引入用矢量基学习和自适应迭代相结合的方式得到一个小的支持向量集,可以避免递推时可能出现的误差积累问题,有效提高算法的稀疏性和稳定性;同时采用加权方法确定权值系数以减小训练样本中非高斯噪声的影响。实验结果表明,改进的LSSVR具有较好的鲁棒性、支持向量稀疏性和动态建模实时性。Combining the advantages of the vector-based learning and adaptive iterative algorithm, an improved weighted Least Squares Support Vector Regression（LSSVR） is proposed to solve the problems of the least squares support vector regression methods, such as lacking of sparsely and robustly. During the training process of algorithm, the vector-base learning and automatic iterative pro- cedures are introduced and a small support vector set can be obtained adaptively. This method can avoid the error accumulation during the iterative processing and improve the sparseness and stabilitaj of the algorithm, while the weights are determined by a robust method in order to reduce the effect of the outliers（e.g.resulting from non-Gaussian noise）. The experimental results show that the proposed algorithm has a better robust, sparsely of support vector and real-time performance of dynamic modeling.

关 键 词：矢量基 自适应迭代算法 支持向量稀疏性 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]