BP神经网络算法的一种改进  

Improvement on BP algorithm in artificial neural network

作　　者：王青海 

机构地区：[1]青海省大通县职业教育中心,青海大通810101

出　　处：《青海大学学报（自然科学版）》2004年第3期82-84,共3页Journal of Qinghai University(Natural Science)

摘　　要：为了减小标准BP算法中迭代次数并提高其收敛速度,提出了将负梯度下降法与DFP变尺度算法相结合进行权值修正的方法,在误差寻优初期,首先采用标准BP算法进行迭代,每迭代一次的工作量较小、所需存贮量较少,且对初始点的要求不高。当寻优过程开始接近最优时,更改寻优算法,即使用DFP变尺度算法。最后,通过MATLAB实现。结果表明改进后的BP算法减少了迭代次数,提高了寻优的收敛速度。To reduce the iterative time and speed constringency in conventional BP algorithm,a improved weighted algorithm,which combines the method of negative gradient descent with DFP variable scale algorithm,is presented in the paper.At the initial optimization stages,iteration method based on conventional BP algorithm is adopted for the reason of the advantage in lesser iterative workload and memory cell as well as the low demand on initial point.Then,the DFP variable scale algorithm is adopted during the optimization process near optimum point.The MATLAB simulation shows the improved weighted algorithm is of more efficien than conventional BP algorithm.

关 键 词：负梯度下降 HESSE矩阵 误差最优 权值修正 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]