BP神经网络算法的改进及其应用

根据BP算法的基本原理，分析指出了BP算法存在着收敛慢、接近最优时易产生波动和振荡现象的原因。在此基础上，通过进一步研究，提出了一种新的改进BP算法。改进后的BP算法不仅运算速度有所提高，而且在一定程度上克服了易产生波动和振荡现象的问题。由于改进BP算法的每个权都能找到最优学习率，因此收敛精度得到了提高；并且该算法基本不受初始学习率的影响，因而避免了学习率选取的困难。图1，表3，参4。

An Improved Method of BP Neural Network and Its Application

The paper analyzed the cause of some deficiencies that existed in the standard Back Propagation Neural Network（BPNN） based on the principle of BPNN. The deficiencies including long convergence time and the large learning rate will make the BPNN oscillating. This paper presented the improved BPNN that not only can shorten convergence time but also can overcome the oscillating to some extent. The convergence accuracy could be improved variously as each weight found its optimal learning rate. It was not difficult to select initial learning rate value when the improved BPNN was used.