模块化模糊神经网络训练策略研究  

Study on Training Strategy of Modular Fuzzy Neural Network

作　　者：王晶晶[1] 金健[2] 

机构地区：[1]南京信息工程大学遥感学院,江苏南京210044 [2]华东师范大学信息科学技术学院,上海200006

出　　处：《南京气象学院学报》2007年第2期274-278,共5页Journal of Nanjing Institute of Meteorology

基　　金：国家科技部社会公益性研究专项(2004DIB3J122)

摘　　要：为分析模块化模糊神经网络的训练机制,进一步降低其训练复杂度,增强模型的泛化性能,尝试采用部分训练样本对模型进行相对粗训练,并以广西全省1957—2003年连续47 a 5月平均降水量作为实验数据,统计各种参数条件下模型的逐步预测误差情况。结果表明:由部分训练样本参与训练的模型所得测试误差普遍低于由全部训练样本参与训练的模型所得测试误差,证明该方法可行。In order to analyze the training mechanism of Modular Fuzzy Neural Network （ MFNN）, partial training samples were adopted in the rough training model for lowering complexity and enhancing generalization. Statistical experiments of stepwise prediction errors were performed under various parameter conditions on mean rainfall in May of Guangxi province for continuous 47 years （1957-2003）. The results show that the testing errors for employing partial training samples are generally lower than that for employing the whole training set. The feasibility of this method is thus confirmed and the further study on this question is worth performing.

关 键 词：聚类 模块化模糊神经网络 训练策略 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]