摘要
为提高神经网络的训练速度,提出一种自适应确定带动量项BP算法中动量因子的方法.在学习率为常数情况下,根据误差函数关于权值向量的梯度变化情况,自适应调节动量因子.数值试验表明,该方法对离线和在线训练均有效,且在收敛速度和算法稳定性上优于常动量因子的BP算法.
A method was developed to adaptively determine the momentum factor of back-propagation （BP） algorithm to enhance the training speed of the neural networks. Taking the learning rate as constant, the algorithm adjusted the momentum factor according to the the gradient of the error function with respect to the weight vector. Numerical experiments show that the proposed algorithm is effective for both the batch and online training. Moreover, it is superior to the BP algorithm with constant momentum factor in respect of convergence rate and stability.
出处
《大连海事大学学报》
EI
CAS
CSCD
北大核心
2008年第4期45-47,51,共4页
Journal of Dalian Maritime University
基金
国家自然科学基金资助项目(10471017)
关键词
神经网络
BP算法
动量因子
neural networks
back-propagation （BP） algorithm
momentum factor
