一种基于惩罚系数的BP神经网络预测能力  

Predictive ability of BP neural network based on penalty coefficient

作　　者：廖卫强[1] 迟岩[2] 王国玲[3] 

机构地区：[1]集美大学电气自动化教研室,福建厦门361021 [2]闽南理工学院电子工程系,福建石狮362700 [3]集美大学电子教研室,福建厦门361021

出　　处：《辽宁工程技术大学学报（自然科学版）》2013年第2期174-177,共4页Journal of Liaoning Technical University (Natural Science)

基　　金：福建省科技厅科技三项基金重点资助项目(2005H038)

摘　　要：为提高BP神经网络训练的预测能力,采用有助于提高BP神经网络逼近精度的Metropolis准则来克服BP神经网络训练学习过程中容易陷入局部极小值的问题;考虑到两类误分的代价不同,利用两个惩罚系数C1和C2,对两类误分给予不同程度的惩罚;采用轮换法的策略来避免因样本不均衡分布带来的负面影响.研究结果表明:所构建的神经网络模型效果令人满意,是行之有效的做法.In order to improve the predictive ability of BP neural network,this study uses Metropolis criterion to overcome the shortcoming of opting to fall into local minimum in BP algorithm.This study utilizes the penalty coefficients（C1,C2） in training to give sample misclassification with different penalties.This is because the cost of different misclassification is different.In terms of the cost of variant misclassification,the study uses a method of selecting training samples properly to avoid negative impact from imbalanced distributed samples.The study results show that the model is effective and rational.

关 键 词：BP神经网络 预测 局部极小 METROPOLIS准则 惩罚系数 误分 不均衡分布 轮换法 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]