机器学习中的核覆盖算法

基于统计学习理论的支持向量机（SVM）方法在样本空间或特征空间构造最优分类超平面解决了分类器的构造问题，但其本质是二分类的，且核函数中的参数难以确定，计算复杂性高．构造性学习算法根据训练样本构造性地设计分类网络，运行效率高，便于处理多分类问题，但存在所得的分界面零乱、测试计算量大的缺点．该文将SVM中的核函数法与构造性学习的覆盖算法相融合，给出一种新的核覆盖算法．新算法克服了以上两种模型的缺点，具有运算速度快、精度高、鲁棒性强的优点．其次．文中给出风险误差上界与覆盖个数的关系．最后给出实验模

Kernel Covering Algorithm for Machine Learning

The support vector machine technique based on statistics learning theory solves the problem of classifier constructing by creating optimal separating hyperplane in sample space or character space, which is a high dimension image space of kernel mapping, but its nature property is binary classifier, parameter selecting of kernel function is difficulty and computational complexity is high. Constructive machine learning method that designs classifying neural network according to sample data set constructively is propitious to multiclass problem and performs effectively, but its separating edge is very complex and lack of theoretical basis. This paper defines the fusion of covering domains and provides a new machine learning method named kernel-covering algorithm, which combines the kernel function algorithm of SVM and spherical domain covering algorithm of constructive machine learning method. The new algorithm overcomes the weakness of both SVM and constructive machine learning, converts the rough separating edges of covering domains to a smooth curved surface and reduce the test time. The relationship between the bound of error risk and the number of covering domains are discussed. Finally, some experiments are given to prove the efficiency and robustness of the new algorithm.