模块化神经网络集成方法研究

Research on Integration Method for Modular Neural Network

针对模块化神经网络的集成问题,综合"分而治之"和"集思广益"的思想,提出了一种在线选择子网络的方法。针对不同的输入,计算输入与各子网络训练样本中心的距离测度,构造距离测度的隶属函数,通过模糊判别实现子网络的在线选择。参与信息处理的子网络随输入变化,使网络集成具有更强的自适应能力。多个子网络采用线性整合,采用样本空间重构技术及动态规划方法实现子网络权重的在线优化。仿真结果表明:该方法提高了模块化神经网络的精度和泛化能力。

To solve intergration problem of modular neural network （MNN）, the measure of online sub-networks selection is proposed with integrating ideas of “divide and conquer” and “brainstorm”. According to different inputs, distance measurement between inputs and each center of sub-networks training sample is calculated to construct subordination function, and to realize online sub-networks selection by fuzzy judgement. The change of sub-networks involved in information processing with inputs makes self-adapting ability stronger for network integration. Many sub-networks are intergrated by linear method. Online optimization of weights of sub-networks are realized by reconstructing technology of sample space and dynamic programming. Simulation results show precision and generalization ability of MNN are improved effectively.