摘要
针对BP网络学习速率和动量项参数较难选取以及学习过程中学习效率较为低下的问题,提出BP网络的改进算法模型—AB网络模型,来选取学习速率和动量项的参数值,即通过一个为给定先验知识的A网,动态调节另一个执行实际应用的B网中的学习速率和动量项的参数值,并以此提高整个网络的学习效率.实验结果表明,通过AB网络自适应调整参数的算法比普通BP算法的学习效率大大提高.在实际应用中,我们可以通过AB网络自适应调节的方法,对学习速率参数和动量项参数进行合适的选取.
To deal with the problems of the hard selecting with the parameters of the learning rate and the momentum term, and the low learning efficiency. This paper proposes an improved algorithm modelmAB network model to select the parameters of the learning rate and the momentum term values. Through a given transcendental knowledge network A dynamic adjusting another running practi- cal application net B＇s learning rate and the momentum term to improved the whole net＇s learning efficiency. Experimental results show that compare to the BP algorithm, the AB network adaptively adjust the parameters algorithm has significantly improved the learning efficiency. In practice, we can use the adaptive regulation of AB network to select the paramters of learning rate and momen- tum term appropriatly.
出处
《小型微型计算机系统》
CSCD
北大核心
2013年第8期1872-1876,共5页
Journal of Chinese Computer Systems
基金
国家"九七三"重点基础发展研究计划项目(2005CB321901)资助
软件开发环境国家重点实验室开放课题(BUAA-SKLSDE-09KF-03)资助
关键词
AB网络
BP算法
动量项
学习速率
梯度下降法
AB network
BP algorithm
momentum term
learning rate
gradient descent method
