论文部分内容阅读
针对标准反向传播(BP,Back Propagation)神经网络算法收敛速度慢、易陷入局部极小等缺点,采用附加动量法与学习速率自适应调整相结合策略对神经网络初始参数进行设置。通过在权重计算公式中加入动量项,降低神经网络对误差曲面局部调节的敏感性,有效抑制其陷于局部极小。学习速率根据总误差的变化进行自适应调整,可以有效地缩短学习时间,加快收敛速度。将该改进算法应用于数字、英文字母以及简单汉字的手写字符识别系统中,进行了有无动量、有无噪声等实验,结果表明该方法与传统BP算法相比识别精度较高、训练时间较短且具有较强的鲁棒性。
In view of the shortcomings of convergence slow and easy to fall into local minimum, the algorithm of BP (Back Propagation) neural network is used to set the initial parameters of neural network by the combination of additional momentum method and learning rate adaptive adjustment. By adding the momentum term in the weight calculation formula, the sensitivity of neural network to local adjustment of error surface is reduced, and the local minimum is effectively restrained. The learning rate is adjusted adaptively according to the change of the total error, which can effectively shorten the learning time and speed up the convergence. The improved algorithm is applied to the handwritten character recognition system of digital, English letters and simple Chinese characters. Experiments have been carried out with or without momentum and with or without noise. The results show that the proposed method has higher recognition accuracy and training time than the traditional BP algorithm Short and strong robustness.