摘要
目的 对 BP学习算法中存在的大量局部极小点以及收敛速度慢问题进行研究并提出相应的改进方案 .方法 采用类似模拟退火算法调整网络权值修改量 η和动量项 α以及对学习样本进行按类划分 .结果 使用改进后的 BP算法对 4 80个学习样本进行学习 ,识别率从 70 %提高到 95 %,学习时间从 4 h下降到 3 0 min左右 .结论 算法的改进提高了识别率并降低了学习时间 .
Aim To study the standard BP algorithm's local minima and learning speed problems and propose the scheme for improvement. Methods Methods of classifying samples were used and the analogy anneal was applied to adjust the network's weight change rate( η ) and moment( α ). Results Recognition rate of 480 learning samples is increased from 70% to 95%, and learning hours of this problem decrease from 4 h to 30 min or so by using this new algorithm. Conclusion The capacity of new BP learning algorithm is greatly improved.
出处
《北京理工大学学报》
EI
CAS
CSCD
1999年第6期721-724,共4页
Transactions of Beijing Institute of Technology
关键词
神经网络
软件模拟
局部极小点
BP学习算法
neural network
software simulation
model neural network programming
local minima