Computer Engineering and Applications ›› 2008, Vol. 44 ›› Issue (21): 106-108.DOI: 10.3778/j.issn.1002-8331.2008.21.029

• 机器学习 • Previous Articles     Next Articles

Improved learning algorithm for LMBP based on conjugate gradient methods

LI Ye-li1,FENG Chao2,LU Li-kun1   

  1. 1.School of Information & Machine Electrical Engineering,Beijing Institute of Graphic Communication,Beijing 102600,China
    2.School of Electrical Engineering and Automation,Tianjin University,Tianjin 300072,China
  • Received:2008-04-30 Revised:2008-05-26 Online:2008-07-21 Published:2008-07-21
  • Contact: LI Ye-li

一种基于共轭梯度的LMBP改进学习算法

李业丽1,冯 超2,陆利坤1   

  1. 1.北京印刷学院 信息与机电工程学院,北京 102600
    2.天津大学 电气与自动化工程学院,天津 300072
  • 通讯作者: 李业丽

Abstract: The convergence of Levenberg-Marquardt BP neural network is slow.The reason is that computation costs too much in process of inversing matrix JTJ+µI for LMBP algorithm.Based on unconstrained optimization theory,an improved LMBP learning algorithm is presented by using conjugate gradient methods.It uses the way to avoid the time-consuming calculation of the invers matrix,to decrease the calculation amount and to quicken the convergence.The validity of the algorithm is proved by the simulation in Matlab.

Key words: BP algorithm, Levenberg-Marquardt, optimization theory, conjugate gradient methods

摘要: 对神经网络中的LMBP(Levenberg-Marquardt BP)算法的收敛速度慢进行分析,针对矩阵JTJ+µI求逆过程运算量过大而造成收敛速度慢的缺陷,根据无约束优化理论,提出一种基于共轭梯度方法的改进LMBP网络学习算法,利用求解大规模线性方程组的共轭梯度方法,避免了烦琐的求逆过程,降低了计算复杂度,加快了网络的收敛速度,通过Matlab仿真,比较了算法的收敛速度,证明了方法的有效性。

关键词: BP算法, Leverberg-Marquardt, 优化理论, 共轭梯度