BP神经网络的算法改进及应用研究

【作者】 陈玉芳；

【导师】 雷霖；

【作者基本信息】 电子科技大学 ， 检测技术与自动化装置， 2004， 硕士

【摘要】 BP(Back Propagation)神经网络存在其固有的缺陷:收敛速度慢、易陷入局部极小、网络结构难以确定等。本课题研究的目的是对BP网络的两个主要缺陷进行改进,即:加快学习速度和跳出局部极小找到全局最优。论文对BP神经网络模型作了系统的综述,对其工作原理进行了详细的推导、给出了算法的软件实现,针对网络的性能缺陷提出了两种有效的改进算法:FPMBP和SAPMBP,对两种新算法作了详细介绍,给出了算法实现步骤。论文以数值计算软件MATLAB6.1为工具,将两种改进算法和变学习率的动量BP、Levenberg-Marquardt算法作了性能对比研究,从迭代次数、收敛时间等多个方面进行比较,体现了新算法的有效性。还将提出的算法用于字符识别和神经PID控制两个应用实例,和动量BP及自适应学习速率的动量BP进行比较。对于字符识别从收敛速率、误识率等角度,对神经PID控制,从PID三个控制参数的输出曲线、控制量u的输出曲线、跟踪曲线等多个角度进行考察,证实了新算法的有效性及实用性,达到了课题预期的目的。更多还原

【Abstract】 BP network has been used widely because of its many specialties, storing information by distributed way, dealing with information in parallel manner preferable robust etc. But due to its inherent defects, BP has difficulty in theory consummated and application extended; so improving its capability is significant. The goal of this article is to improve the main defects of BP. We expound BP networks systematically and explain the soft-ware operation. We put forward two effective algorithms: FPMBP and SAPMBP.With MATLAB6.1, we testify they are more excellent than BP with momentum. Additionally, we use the two new algorithms to two application examples of characters identification and neural PID control. We testify the two new algorithms efficient and practicable from the viewpoint of convergence speed and from three parameters of , control input, system’s output and error etc. to PID control.更多还原