一种基于随机GA的提高BP网络泛化能力的方法  

A Method of Improving Generalization for BP Network Based on Random GA

作　　者：郭海如[1] 李志敏[1] 万兴[1] 熊斌[1] 

机构地区：[1]湖北工程学院计算机与信息科学学院,湖北孝感432000

出　　处：《计算机技术与发展》2014年第1期105-108,共4页Computer Technology and Development

基　　金：湖北省教育科研计划重点项目(D20122606);湖北工程学院项目(Z2011009)

摘　　要：LM-BP网络对其初始权值和阈值敏感,泛化能力不强,针对该缺点,采用遗传算法(GA)对其初始权阈值进行优化,在一定程度上能提高LM-BP网络的泛化能力。为进一步扩展GA初始种群的覆盖范围,进一步提高LM-BP网络的泛化能力,采用多次随机产生初始种群多次优化的方法。以伦河孝感段氟化物含量为实例,建立随机GA的LM-BP网络模型,对原始数据进行拟合及测试,结果表明该方法基本能100%拟合,测试误差不超过2.3%。经过对比实验,证明了该方法的有效性。The LM-BP neural network was sensitive to its initial weight values and threshold, and it had bad generalization ability. In view of its shortcomings, the initial weights and threshold of LM-BP neural network were optimized with GA. The generalization of LM -BP neural network was improved to a certain extent. To expand the coverage of initial population, the initial populations were randomly generated iteratively and the network was optimized multi times. Thus, the generalization of LM-BP network was further improved. Take the content of fluorine in Lun River from Xiaogan as an example, the LM-BP neural network model based on random GA was estab- lished, and the raw data were fitted and tested. The results showed that the accordance of fitting data is approximately 100%, and the tes- ting errors were less than 2.3 %. Through contrast experiments, the validity of this method was proved.

关 键 词：随机遗传算法 神经网络 测试误差 泛化能力 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]