模块化神经网络子网的动态集成方法学研究  

Methodological research for dynamic integration of modular neural network's sub-networks

作　　者：王攀[1] 李幼凤[2] 范衠 冯帅[1] 

机构地区：[1]武汉理工大学自动化学院,湖北武汉430070 [2]浙江大学信息学院,浙江杭州310027 [3]丹麦科技大学管理工程系,丹麦林拜2800Kgs

出　　处：《系统工程与电子技术》2008年第6期1143-1147,共5页Systems Engineering and Electronics

基　　金：国家自然科学基金(60174039)资助课题

摘　　要：提出一个模块化神经网络的广义定义,它包含了几乎所有多神经网络(系统)。简要分析了模块化神经网络子网集成的相关概念和问题。针对一类模块化神经网络,提出了5种基于"分而治之"原理和自适应组合的新型动态集成方法。它们之间的主要区别在于:距离测度(绝对距离测度和相对距离测度);个体数目(有些全部参与集成,有些则是部分参与);集成策略和规则(数据驱动和数据/知识驱动)。仿真实验证实了这些方法的有效性。同时,还提出了一种基于"一专多能"思想的子网训练方法。A generalized definition of modular neural networks （MNN） that almost include all multi-neural networks （systems） is presented. Some relative concepts and problems of MNN＇ dynamic integration are briefly analyzed. Five new dynamic integration algorithms are presented for a kind of modular neural networks that are based on the principle of ＂divide and conguer＂ and adaptive combination. The main differences among each sub-method are： distance measure （absolute distance measure and relative distance measure） and computing method; selected and integrated number of the candidates （some integrate all networks and the others select only in part）; integrated strategy and rule （some are based on data driven or knowledge driven, the others are based on data and knowledge driven）. Empirical studies show these algorithms＇ effectiveness and potential. Meanwhile, a technical approach based on the thought of ＂experienced-in-one-aspect-and-feasible-in-others＂ is presented for MNN sub-networks＇ training.

关 键 词：模块化神经网络 “分而治之”原理 动态集成 距离测度 数据与/或知识驱动策略 

分 类 号：N945.12[自然科学总论—系统科学]