并行学习神经网络集成方法

该文分析了神经网络集成中成员神经网络的泛化误差、成员神经网络之间的差异度对神经网络集成泛化误差的影响，提出了一种并行学习神经网络集成方法；对参与集成的成员神经网络，给出了一种并行训练方法，不仅满足了成员网络本身的精度要求，还满足了它与其余成员网络的差异性要求；另外，给出了一种并行确定集成成员神经网络权重方法．实验结果表明，使用该文的成员神经网络训练方法、成员神经网络集成方法能够构建有效的神经网络集成系统．

A Parallel Learning Approach for Neural Network Ensemble

After the relationship among the generalization errors of the neural networks en semble, the generalization error and the diversity of the component neural netwo rk is analyzed. A component neural networks parallel trained algorithm PLA is pr oposed, which encourages component neural network to learn from expected goal an d the others, so all component neural networks are trained simultaneously and in teractively. In the stage of combining component neural networks, a parallel w eight optimal approach GASEN-e is provided by expanding GASEN proposed by Zhou et al., which assigns weight for every component neural network and bias fo r their ensemble. Experiment results show that a neural networks ensemble system is efficiently constructed by PLA and GASEN-e.