BP自适应学习率设计  

Design of Adaptive Learning Rate for BP Algorithm

作　　者：杨启文[1] 韩玉兵[2] 

机构地区：[1]河海大学计算机及信息工程学院,江苏常州213022 [2]河海大学常州校区数理教学部,江苏常州213022

出　　处：《河海大学常州分校学报》2001年第3期20-24,共5页Journal of Hohai University Changzhou

摘　　要：在神经网络的训练当中存在“过学习”现象以及参数难以选择的困难 .本文指出了大多数自适应 BP算法在学习率自调整中存在的问题 :事后矫正 ,同时通过极其简单的优化手段 ,为当前权值的每一次调整选择一个次优 (甚至最优 )的学习率 ,从而改善了训练过程的平稳性和快速性 ,并且有效地解决了神经网络的“过学习”There exist over learning and the difficulties of selecting suitable parameters when training neural network.In this paper,the deficiency of adaptive BP algorithms in regulating learning rate is pointed out as the so called optimal parameters for the current iteration aren't updated until the next iteration.We choose a sub optimal(even an optimal)learning rate for the current regulation of weights in every iteration by means of a highly simple method.As a result,the stability and speediness of the training process are improved and the problem of over learning is also solved effectively.

关 键 词：神经网络 自适应学习率 “过学习”现象 反传算法 网络训练 非线性优化问题 BP算法 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]