一种基于数值计算的改进BP神经网络加速算法

An effective accelerating algorithm for BP neural network based on numerical computation

为增加神经网络收敛的稳定性与收敛速度,提出了一种改进的网络优化加速算法.在权值调整期间加入前N期权值结果,增强了训练的稳定性;使用Steffensen迭代算法进行加速,使网络训练较快地收敛;有效地解决了传统BP神经网络的缺点.进行数值实验,将10幅二值化后的车牌数字字符图片作为训练样本送入改进的网络与传统的BP神经网络中分别进行训练,可以看出传统BP算法在训练过程中出现了振荡且收敛速度较慢.而改进的算法误差稳步下降,没有出现传统算法中振荡的现象,且较传统算法早达到收敛稳定.

In order to enhance stability and convergence race of neural network, a new optimized accelerated algorithm is presented for neural network in the paper. Applied the N former weights values to new modified weights can enhance training stability; accelerating by Steffensen iterative method can make neural network more rapid convergence. It is effective to solve problems mentioned of traditional BP neural network. At last, a numerical experiment is carried out, inputting ten vehicle license number images as, training samples into improved algorithm and traditional algorithm for training. It is found that errors of improved algorithm decline steadily. The improved algorithm has no surge and a earlier convergence compared to traditional BP algorithm in the training process.