节点文献

多拉格朗日乘子协同优化的SVM快速学习算法研究

SVM Fast Training Algorithm Research Based on Multi-Lagrange Multiplier

  • 推荐 CAJ下载
  • PDF下载
  • 不支持迅雷等下载工具,请取消加速工具后下载。

【作者】 业宁孙瑞祥董逸生

【Author】 Ye Ning 1,3 , Sun Ruixiang2, and Dong Yisheng1 1 (Department of Computer Science and Engineering, Southeast University, Nanjing 210096) 2 (Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100080) 3 (College of Information Science and Technology, Nanjing Forestry University, Nanjing 210037)

【机构】 东南大学计算机科学与工程系中国科学院计算技术研究所东南大学计算机科学与工程系 南京210096南京林业大学信息科学与技术学院南京210037北京100080南京210096

【摘要】 提出了一个利用多个拉格朗日乘子协同优化的支持向量机快速学习方法(MLSVM),并给出了每个乘子的可行域范围的定义公式,由于在每个乘子的优化过程中使用了解析表达式,使得算法可以更加精确和快速地逼近最优解,可以证明SMO算法是该方法的一个特例.在此方法的理论指导下,根据不同的学习策略,程序实现了3种不同的具体算法(MLSVM1,MLSVM2,MLSVM3),其中前两个算法在数据集不大时(<5000条记录)学习速度与SMO算法相当,但当数据集更大时,算法就失效了.MLSVM3是一个改进算法,总结了MLSVM1和MLSVM2失效的原因,对SMO算法中学习效率较低的部分进行了改进,在多个数据集上测试,MLSVM3算法速度超过了SMO算法7.4%~4130%.

【Abstract】 A multi-Lagrange multiplier support vector machine fast training method (MLSVM) based on the coordinated optimization of multi-Lagrange multipliers is proposed and the formula to define the feasible field of each multiplier is presented. The algorithm approaches to the most optimization more precisely and quickly due to the analytic expressions adopted in the optimization process of each multiplier. The SMO algorithm is proved to be an instance of MLSVM. Three individual algorithms, i.e., MLSVM1, MLSVM2 and MLSVM3, are presented under the theoretical guidance of this method according to different learning strategies. The learning speed of MLSVM1 and MLSVM2 is about the same as that of SMO when the test data set is small (<5000). However, they will fail when the test data set becomes larger. MLSVM3 is an improved algorithm of the former two algorithms and the SMO algorithm. It not only overcomes the failure of MLSVM1 and MLSVM2, but also performs faster than the SMO algorithm with an improvement of 7.4% to 4130% on several test data sets.

【基金】 国家自然科学基金项目(30271048);东南大学校引进(留学)人才基金项目(G2002-28);东南大学校科研基金重点课题(X02-070-1(Z))~~
  • 【文献出处】 计算机研究与发展 ,Journal of Computer Research and Development , 编辑部邮箱 ,2006年03期
  • 【分类号】TP301.6
  • 【被引频次】38
  • 【下载频次】576
节点文献中: 

本文链接的文献网络图示:

本文的引文网络