摘要
针对标准BP算法稳定性与收敛性较差的问题 ,采用多个变元进行迭代 ,其中各变元的迭代方式同标准的BP算法 ,然后取各变元迭代结果的均值作为新一轮训练的修正权值 .其目的是利用均值的平衡效应 ,防止作过大或过小的权值调整 ,解决因不合适的权值调整而导致BP算法整体性能的下降的问题 .理论分析与试验证明 。
On account of the low convergence and bad stability in BP algorithm a new method was studied. An average of multi variable replacing result was used as the weights of new round BP networks training. The balance effect of the average prevented improper weights adjustment which will result in bad performances for BP algorithm. According to the result of the theoretic analyses and the experiments, the convergence and stability in BP algorithm can be inproved.
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2002年第12期21-22,共2页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家高性能计算基金资助项目 (980 4 5 0 )
关键词
BP算法
多远迭代
前馈神经网络
稳定性
收敛性
BP algorithm
multi variable replacing
feedforward neural networks
stability
convergence
