粗拙空间上结构风险最小化原则.pdf

摘 要 摘 要 统计学习理论是处理小样本学习问题的重要理论方法。然而，该理论是建立在概率 空间上基于实随机样本的，它难以讨论和处理现实世界中客观存在的涉及粗糙空间上粗 糙样本的小样本统计学习问题。结构风险最小化原则是统计学习理论的核心内容之一， 是构建支持向量机的重要基础。基于此，本文研究了粗糙空间上结构风险最小化原则。 首先，给出了粗糙空间上粗糙变量退火熵、生长函数和 VC 维的定义，并证明了它们的 一些性质；其次，构建了粗糙空间上基于 VC 维的风险泛函的界；再次，给出了粗糙空 间上结构风险最小化原则，证明了该原则的一致性，同时推导出了收敛速度的界；最后， 给出了粗糙空间上结构风险最小化原则的直接实现。 关键词 粗糙空间 VC 维 风险泛函的界 结构风险最小化原则 直接实现 I Abstract Abstract Statistical Learning Theory is important regarded as a sound framework that handles a variety of learning problems in presence of small size data samples. However, Statistical Learning Theory is built on probability space and based on real random samples, so it can hardly handle statistical learning problems built on rough space and based on rough samples, which can be encountered in real world scenarios. Structural risk minimization principle is one of the kernels content of Statistical Learning Theory. The principle is the theoretical fundamentals of establishing the support vector machine. Based on above, the structural risk minimization principle based on rough samples is explored in the paper. Firstly, the concepts of annealed entropy, growth function and VC dimension as well as their properties on rough space are given. Secondly, the constructive and distribution-independent bounds with VC dimension on rough space are presented. Thirdly, the structural risk minimization principle on rough space is proposed. The consistency of this principle is proven, and asymptotic bounds on the rate of convergence is derived. Finally, the structural risk minimization principle’s direct implementation on rough space is given.