摘要
针对BP算法学习速度慢和易陷入局部最优等缺陷,提出了一种新的变步长算法——限幅的变步长算法C VLBP。这种算法根据误差来调整下次的学习步长,同时把步长限制在一定范围内,这样在计算量增加很少的情况下,使学习时间大为缩短的同时,也避免了过多的振荡,防止了系统的发散。仿真结果表明,该算法具有跳出局部最优的能力,同时对初始权值和阈值具有一定的鲁棒性。
A novel variable step size（control variable learning rate backpropageaion,CVLBP）algorithm is proposed in veew of the faultiness of BP algorithm.The step size which was limited in a certain value was determined by the learning error.Thus the learning time is cut down greatly and avoid some oscillation in the learning process,even avoid divergence,bue the calculation is added little .The simulation result shows that this algorithm can step out the local optimal and has certain powerful intelligence to initial weight and threshold value.
出处
《机械工程师》
2005年第11期105-106,共2页
Mechanical Engineer
关键词
BP神经网络
自适应
变步长学习
VLBP算法
LMS算法
BP neural network
adaptive
variable step size learn
VLBP（variable learning rate backpropagetion）
LMS（least mean square）
