一种多模块协同参与的神经网络  

A multi-module cooperative neural network

作　　者：薄迎春[1] 乔俊飞[1] 杨刚[1] 

机构地区：[1]北京工业大学电子信息与控制工程学院,北京100124

出　　处：《智能系统学报》2011年第3期225-230,共6页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金资助项目(60873043);北京市自然科学基金资助项目(4092010);教育部博士点基金资助项目(200800050004);北京市属高等学校人才强教计划资助项目(PHR201006103)

摘　　要：针对单一神经网络训练时间长、对复杂问题处理精度较低、易陷入局部极小等问题,提出了一种多模块协同参与信息处理的神经网络.该神经网络具有层级结构,基于条件模糊聚类技术对样本进行分类,根据分类结果实现对神经网络的模块划分,采用代数算法对网络权值进行求解,基于距离测度设计了处理输入信息的子网络选择方法.为提高神经网络对复杂问题的逼近能力,选择数目不等的多个子网络参与给定输入的协同处理,采取"分而治之"与"集成学习"相结合方法以提高网络的性能.实验表明,对于复杂问题,这种多模块协同参与的神经网络可以有效地提高网络的逼近精度,训练时间也优于单一网络.Aiming to solve the problems of long training time, low precision in processing complex problem, and a local minimum in single neural networks, a multi-module cooperative neural network （MMCNN） was proposed. Its structure has hierarchical character. Sample data was first detached by the fuzzy clustering method, and then the neural network was partitioned into several sub-nets based on the clustering results. The linking weights were elicited by solving equations. For a given input data, some multi-modules were selected to deal with it. The approximating performance was improved by combining divide-and-conquer and learning ensemble strategies. A sub-net selection method was designed based on distance measurements. Simulation results demonstrate that a multi-module cooperative neural network can heighten approximating ability effectively for complicated problems, and the training time is faster than in a single back-propagation neural network.

关 键 词：神经网络 条件模糊聚类 多模块 子网络选择 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]