一种改进的BP算法神经网络

An Improvement of Learning Algorithm for BP Neural Network

讨论了BP神经网络学习过程中的假饱和现象和激励函数对输出值的影响,将修改激励函数和构建假饱和预防函数相结合,实现加快网络学习速率。通过引入距离熵揭示了实际输出值、期望输出值以及能量函数三者的内在关联。对BP网络的应用实例编制了仿真程序,并与标准的BP算法进行比较。结果表明改进算法的学习收敛性大大地优于标准BP算法。

In this paper, the causes of the error saturation condition in the learning process and the influence of activation functions are analyzed. Construction of an error saturation prevention function and modification of activation is combined to improve the learning efficiency. Distance entropy is introduced to explain the relations of the actual output, desired output and energy function. The computer simulation program is drawn up to examples of application based on BP network and the convergence rate of three algorithms is compared. The results are shown that the capability of the improved algorithm is largely superior to that of the standard BP algorithm.