训练支持向量机的四重序列解析优化算法  

Quadruple sequential analytic optimization algorithm for training support vector machines

作　　者：陆爱国[1] 刘红卫[1] 王珏[2] 

机构地区：[1]西安电子科技大学理学院,西安710071 [2]中国科学院数学与系统科学研究院预测科学研究中心,北京100190

出　　处：《系统工程理论与实践》2011年第8期1555-1564,共10页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(70801058;70971052;61072144);中国科学院知识创新工程重要方向项目-全球经济监测与政策模拟仿真平台

摘　　要：为了提高训练支持向量机的序列最小优化算法的学习性能,依据在每次迭代中选取多个变量同时优化可能会减少迭代次数和缩短训练时间的这一想法,提出了一个训练支持向量机的快速算法-四重序列解析优化方法,并给出了一个重要定理,使得相应的子问题有解析解,从而能够更加精确和快速地逼近最优解.在两个公共数据集上的仿真结果表明:该方法比其它算法有较好的学习性能-在相同训练精度的条件下,不仅缩短了训练时间,而且计算复杂度减小.In this paper,a novel algorithm called quadruple sequential analytic optimization method for support vector machines（SVMs） training is presented.This algorithm is based on the idea that multi-variable coordinated optimization can reduce the number of iterations and training time,which can improve the performance of sequential minimal optimization（SMO）-type learning algorithm greatly.Moreover,a theorem on SVM-training,as well as the detailed proof,is proposed to guarantee the existence of analytical solution of corresponding sub-problem and the approaching to optimal solution more precisely and quickly. Numerical experiments on two bechmark database demonstrate that the proposed method has faster speed and smaller complexity than other algorithms.

关 键 词：支持向量机 四重序列解析优化算法 序列最小优化 最大违背对 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]