摘要
标准的支持向量回归机中原始最优化问题的目标函数有两部分:一部分是衡量经验风险的,另一部分是衡量推广能力的。本文引入一个凸函数来代替衡量推广能力的部分,讨论了当这个凸函数取不同的形式时支持向量回归机的变形,这些模型不再要求核函数必须对称正定,从而为我们可以得到更灵活的回归曲面提供了有效的工具。
There are two parts in the objective function of primal optimization problems in standard support vector (regression:)one is to represent empirical risk; the other is to control the capacity of generalization.Instead of the capacity,we introduced a convex function in this paper, and discussed the variations of support vector regression according to different (convex) functions. These models do not require the positivity of kernel functions any more, and so provide us efficient tools to derive more flexible regression surface.
出处
《系统工程》
CSCD
北大核心
2004年第10期9-12,共4页
Systems Engineering
基金
国家自然科学基金资助项目(10371131)