一种快速选择性神经网络集成方法

选择性集成通过选择部分个体能够获得比传统全部集成更好的泛化性能．已有的一些选择性集成方法实现的时间复杂性较高，或没有充分利用个体的先验信息．提出了一种新的选择性集成方法FASEN（fast approach to selective ENsemble）．该方法在独立训练出一批神经网络后，采用一种近似搜索策略，只选择与其他个体差异较大且估计泛化误差较小的网络参加集成，把个体选择的搜索空间从2^N降到N．理论分析与实验结果表明，该方法简单高效，泛化性能与已有的几种选择性集成方法相当甚至占优．

A Fast Selective Approach to Neural Network Ensemble

Selective ensemble learning which partially chooses individual neural network outperforms traditional methods. Existing selective approaches either have expensive time complexity or utilize a priori information of each network rarely. A novel selective approach to neural network ensemble named FASEN is presented. After every neural network is trained separately, FASEN selects those having low estimative generalization error and high dissimilarity with others according to results from validation set. The amount of search in solution space is decreased from 2^N to N. Theoretical analysis and experimental results show that FASEN is better in efficiency and generalization than other approaches.