线性可分问题的支撑向量稳健感知器及其几何训练算法  

Support Vector Robust Perceptron for Linear Separable Problem and its Geometric Training Algorithm

作　　者：陈森林[1] 张军英[2] 范丽[1] 

机构地区：[1]陕西师范大学,陕西西安710062 [2]西安电子科技大学,陕西西安710071

出　　处：《微电子学与计算机》2004年第6期31-34,共4页Microelectronics & Computer

基　　金：国家自然科学基金项目(60071026);国防科技预研跨行业基金(OOJI.4.4.DZ0106)

摘　　要：分类器的稳健性能是分类器的重要性质之一。支撑向量机SVM和稳健感知器得到的都是最优分类面,都具有很强的稳健性能。SVM构造的是到所有支撑向量(距分类面最近的样本)等距离的最优分类面,SVM算法需要求解一个二次型寻优问题;而稳健感知器构造的是到所有基(各模式类的边界样本)距离都较远的最优分类面,稳健感知器需要求解一系列的线性规划。文章在二者的基础上提出了适用于线性可分问题的支撑向量稳健感知器及其几何训练算法,它将问题转化成了一系列的线性方程组,它将比SVM的二次型寻优具有更快的速度。实验仿真表明了该算法的高效性。Robust ability is one of the important features of a classifier. Both SVM and robust perceptron obtain optimally robust classification boundaries. SVM does this based on the support vectors in training samples which are nearest to the classification boundaries, while robust perceptron (RP) does this based on the so called bases in the training samples which are the boundary training samples in each class. SVM needs to solve a quadratic optimization problem, while RP needs to solve a series of linear programming. Here, support vector robust perceptron and it's geometric training algorithm is proposed, only a series of linear equations is to be solved. It is fast than SVM when the scale of the problem is not too large. Computer simulation indicates its efficiency.

关 键 词：稳健分类 感知器 支撑向量 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]