前馈式神经网络的容错性评估和计算方法的研究  

STUDY OF FAULT-TOLERANCE ESTIMATION AND CALCULATION METHOD FOR FEED FORWARD NEURAL NETWORKS

作　　者：姜惠兰[1] 孙雅明[1] 

机构地区：[1]天津大学电气工程与自动化学院,天津300072

出　　处：《模式识别与人工智能》2004年第2期201-206,共6页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金(No.59877016)

摘　　要：本文研究了前馈式神经网络(NN)的容错性问题,提出了一种用全新的概念评估前馈式NN容错性能的方法.将动力学系统的吸引子和吸引域的概念引入到前馈式NN网络,给出了前馈式NN模型的"吸引子"和"吸引域"的定义,以及描述容错性的指标,并在此基础上推导出一种计算前馈式NN"吸引域"大小的数学方法.该研究使前馈式NN容错性有了基于理论的评估和数学计算方法.This paper studies fault-tolerance problem of feed forward NN, develops a method using the new concepts to estimate fault-tolerance performance of feed forward NN. Based on dynamical system theory, the concepts of attractor and attractive region are introduced to feed forward NN, the definitions of 'attractor' and 'attraction region' of feed forward NN are presented accordingly. Further detailed mathematical derivations of estimating the volume of 'attraction region' of feed forward NN has been presented. The study of this paper has provided a theory framework for evaluation of fault-tolerance of feed forward NN.

关 键 词：神经网络 容错性 吸引域 信息处理 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]