论文部分内容阅读
回归是统计学习理论中的主要研究问题之一.文中给出确定性退火分片回归算法全局收敛性和自由能全局极小关于温度连续性的证明,推导出初始临界温度的简化计算公式,并提出一种新的增强型分片回归算法,利用“原型迁移”技巧,避免退火过程中“空剖分”的出现.基于Benchmark数据集上的实验表明:新算法能有效去除模型冗余,提高学习泛化能力.
Regression is one of the main research issues in statistical learning theory.In this paper, the global convergence of fractional annealing algorithm for deterministic annealing and the global minimum of free energy on temperature continuity are given, and a simplified calculation formula for initial critical temperature is deduced. A new enhanced sharpening regression algorithm is proposed to avoid the appearance of “empty split ” in the annealing process by “prototype migration ” technique.Experiments on Benchmark dataset show that the new algorithm can effectively remove the model Redundancy, improve learning generalization ability.