四种学习算法稳定之间的关系  

THE RELATION OF FOUR KINDS OF LEARNING ALGORITHMIC STABILITY

作　　者：邹斌[1] 董雪梅[1] 李落清[1] 

机构地区：[1]湖北大学数学与计算机科学学院,武汉430062

出　　处：《计算机应用与软件》2005年第1期30-31,111,共3页Computer Applications and Software

基　　金：湖北省自然科学基金资助项目 (99J1 69)。

摘　　要：为研究机器学习的推广误差 ,提出了变一误差估计条件下一种新的学习算法稳定 逐点假设稳定 ,并讨论了逐点假设稳定、CV稳定、重叠稳定以及弱假设稳定四种学习算法稳定之间的关系 ,得出了逐点假设稳定是这四种学习算法稳定中最弱的学习算法稳定的结论。A new notion of algorithmic stability under the change one error estimates,Pointwise Hypothesis Stability was introduced to study the Generalization error of machine Learning,and the relationships of four kinds of learning algorithmic stabilities was also discussed in the paper.The four kinds of learning algorithmic stabilities are Pointwise Hypothesis Stability,Cross Validation Stability,Overlap Stability,Weakly Hypothesis Stability.From these relationships,we can find that Pointwise Hypothesis Stability is the most weakly learning algorithmic stability of the four learning algorithmic stabilities.

关 键 词：学习算法 机器学习 误差估计 假设 CV 推广 结论 重叠 条件 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP181[自动化与计算机技术—控制科学与工程]