神经网络分类器动态集成方法  

Dynamic Integration Approach for an Ensemble of Neural Classifiers

作　　者：郑建军[1] 甘仞初[1] 贺跃[2] 于同[3] 

机构地区：[1]北京理工大学管理与经济学院,北京100081 [2]北京理工大学信息科学技术学院计算机科学工程系,北京100081 [3]中国兵器科学研究院,北京100089

出　　处：《北京理工大学学报》2005年第12期1062-1065,1091,共5页Transactions of Beijing Institute of Technology

基　　金：国家部委预研项目(05BQ0141)

摘　　要：提出一种神经网络分类器的动态集成方法.基于bootstrapp ing构建不同的个体神经网络,针对混合属性,通过不同的加权最近邻设计评估单个网络的分类精度,在此基础上动态选择误差率较小的神经网络,经过投票形成集成分类结果.将该方法与其它几种集成方法在10个UC I数据集上进行了分类性能比较.实验结果表明,该方法在上述所有数据集上的平均分类精度最佳,同时发现,B agg ing比隐层神经元数法能更好地生成个体网络,而将两者结合起来训练个体神经网络,并不能明显提高集成性能.A dynamic integration approach for an ensemble of neural classifiers （NCs） was presented in this paper. It established different NCs based on bootstrapping technique, and evaluated the classification accuracy of every NC by different sorts of weighted nearest neighbors for mixed attributes, then the NCs with low relative generalization error rates were dynamically selected and majority voting was applied to those NCs in order to conduct the final classification results of the ensemble. This approach was compared with some integration approaches on classification performance for ten data sets from UCI. The experiments showed that this approach could obtain the best average classification accuracy over all those data sets. At the same time, it is easy to see that Bagging is better than the method with different number of hidden units （MDHU） for generating different NCs, and the performance of the ensemble may not be improved by combining Bagging with MDHU.

关 键 词：神经网络分类器 动态集成 BAGGING 加权最近邻 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]