基于误差相关度学习样本选择  

Selection of learning sample based on error correlation in machine learning

作　　者：常彦伟[1] 王耀才[1] 曹云峰[2] 王致杰[1] 

机构地区：[1]中国矿业大学信息与电气工程学院 [2]上海交通大学电气工程系,上海200030

出　　处：《计算机工程与设计》2007年第16期3965-3967,共3页Computer Engineering and Design

基　　金：江苏省教育厅自然科学基金项目(06KJD470182);徐州师范大学校基金项目(06XLA21)

摘　　要：针对有限样本学习机器的偏差/方差的困境,以及过拟合引起的泛化性能的下降,分析了样本选择对学习机器泛化的影响,提出误差相关度学习算法ECL,利用误差相关度来权衡偏差和方差的关系,避免了求解复杂学习系统的VC维数,并以样本点的误差相关度为指标来选择训练子集,提高学习机器的泛化性能。仿真结果表明ECL算法有效地抑制过拟合现象的发生,保证学习机器泛化性能的提高。Aimed at the variance-base dilemma, and the descending of generalization as a result of overfitting in machine learning at finite sample, and based on analysis of the train set selecting effect on generalization ability of learning system, an error correlation learning （ECL） algorithm is presented. Error correlation, which acts as a parameter to select the training subset from training set, is developed to improve the generalizability of learning machine. On substituting error correlation for the tradeoff of deviation and variance, the complex VC-dimension solution of the machine learning system is unnecessary. The simulating result shows that the ECL algorithm is of availability, which not only restrains the overfitting of learning system, but also reduces the generalization error of the learning system.

关 键 词：学习机器 过拟合 泛化 训练子集 误差相关度 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]