资 源 简 介
针对BP（Back Propagation）神经网络易陷入局部极小、收敛速度慢的缺点，提出了一种新的BP神经网络改进算法。与标准BP算法比较，该系统通过结合附加动量法和自适应学习速率形成新的BP改进算法。附加动量法虽然可以使BP算法避免陷入局部极小，但是对初始值的选取比较敏感，而且选取合适的学习速率比较困难。而自适应学习速率法可以自动把学习速率调整到一个合适的数值，也可以加快网络的收敛速度，但不能避免陷入局部极小。通过将两者结合起来形成新的改进算法，既可以避免陷入局部极小又可以加快网络的收敛速度。并在此基础上设计一个基于BP神经网络的数字识别系统，此系统可以作为核心部分应用到诸如票据等数字识别中去。实验结果表明，该方法成功的避免了BP算法陷入局部极小，而且收敛速度比标准BP算法提高了17.5倍。
Abstract: In this paper， a new improved BP (Back Propagation) algorithm is presented， because BP neural network can easily fall into local minimum and slow convergence. Compared to the standard BP， this algorithm integrated the additional momentum method with the adaptive learning rate method. The BP algorithm could avoid falling into local minimum because of the additional momentum method， but this method was sensitive to the initial values， and it was also difficult to choose the appropriate learning rate. The adaptive learning rate method could adjust the learning rate to an appropriate value automatically and improve the convergence speed of the network， but it could not get rid of local minimum. By integrating these two methods， this new algorithm could get rid of local minimum and improve the convergence speed. According to this algorithm， a numerical recognition system based on BP neural network was designed. This system could be put into the application of numerical recognition such as bill system. Experiments demonstrate it that BP algorithm can successfully avoid falling into local minimum， and the convergence rate increases seventeen point five times than the standard BP algorithm.
