BP算法收敛性分析及改进

在标准BP神经网络的训练中,将误差函数作为权值调整的依据,使用固定学习率计算权值,这样的结果往往使网络的学习速度过慢甚至无法收敛。对此,从网络收敛的稳定性和速度的角度出发,分析了误差函数和权值修改函数,对算法中学习率的作用进行了具体的讨论,提出了一种根据误差变化对学习率进行动态调整的方法。该方法简单实用,能有效防止网络训练时出现发散,提高网络的收敛速度和稳定性。

Analysis and Improvement on Convergence of BP Algorithm

In the training of standard BP nerve network, the weights are adjusted according to error fimetion and calculated with the fixed learning rate, which often causes network learning speed too slow, even not to converge. For this, starting with the stability and speed of network convergence, we analyze the error function and weights adjusting function, discuss the effect of learning rate in the algorithm specifically, and present a method to adjust the learning rate dynamically according to the error changing. The method is simple and practical, it can prevent the network training from divergence effectively, improve the convergence speed and stability of the network.