摘要
提出一种改进的选择神经网络集成方法,首先构造一批单个神经网络个体,分别利用Bootstrap算法产生若干个训练集并行进行训练;然后采用聚类算法计算训练好的个体网络之间的差异度和个体网络在验证集的预测精度;最后根据个体精度和个体差异度选择合适的个体网络加入集成。实验结果验证,该集成方法能较好地提高集成的预测精度和泛化能力。
Proposes an improved and selective neural network ensemble method named ISEN, firstly constructs a component neural networks which are trained parallely using samples by bootstrap algorithm, then ISEN selects those having better accuracy according to results from validation set and dissimilarity with others which are calculated by clustering algorithm. Experiment results show that ISEN can improve prediction accuracy and generalization ability of the ensemble.
出处
《现代计算机》
2007年第10期17-19,35,共4页
Modern Computer
基金
华南农业大学新学科扶持基金(2005X027)
关键词
神经网络
选择性集成
泛化能力
Neural Network
Selective Ensemble
Generalization Capability
