摘要： 提出一种动态性神经网络集成方法,该方法针对若干不同的神经网络,采用加权最近邻技术收集它们的泛化误差信息构成性能矩阵,在此基础上动态选择泛化误差较小的神经网络,经过动态平均形成集成的最终输出结果。实验表明,与其它方法相比,该方法具有令人满意的性能。

关键词: 神经网络, 动态集成, 泛化误差, 预测

Abstract: A dynamic integration approach for a neural network ensemble is presented, which makes use of weighted nearest neighbor method to collect information about the generalization errors of some different NNs into a performance matrix, and then the NNs with low local generalization errors are dynamically selected and locally dynamic averaging is applied to the NNs in order to conduct the final results of the ensemble. In the experiments, this approach shows promising results on performance compared with other methods.

中图分类号: 

• TP18