结合遗传算法的BP神经网络训练方法研究  

Research on Exercising BP(Back Propagation Neural Network)By AG(Genetic Algorithm)

作　　者：蒋洪涛[1] 李海军[1] 

机构地区：[1]辽宁省交通高等专科学校,沈阳110122

出　　处：《北方交通》2006年第8期70-71,共2页Northern Communications

摘　　要：针对BP神经网络存在收敛到局部极小问题,研究了直接利用遗传算法训练BP网络的方法和结合遗传算法训练网络初始权值的BP算法;通过简单异或问题说明了第一种方法所需训练时间较长;将第二种用于解决结构损伤识别问题,得到较好结果,说明利用遗传算法训练初始权值可以克服BP神经网络局部极值点问题,加快网络的收敛速度。Aiming at the problem of BP lowest convergence exponent, the paper has studied two algorithms for BP exercising -directly utilizing AG to exercise BP,and exercising the initial weight of BP by AG. A simple exclusive OR of a problem shows that exercising time required by first algorithm is longer; the second algorithm is used for identification of structure injury＇ with a good resuh,which has expained that utilizing GA to exercise initial weight can overcome shortcoming of minimum convergence exponent of BP and speed up convergence of network.

关 键 词：遗传算法 BP神经网络 初始权值 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]