| 一种前向神经网络训练算法的改进 |
| |
| 引用本文： | 张森,张化光,王晓文.一种前向神经网络训练算法的改进[J].沈阳工程学院学报(自然科学版),1999(3). |
| |
| 作者姓名： | 张森 张化光 王晓文 |
| |
| 作者单位： | 沈阳电力高等专科学校!110036(张森)，东北大学(张化光)，沈阳电力高等专科学校(王晓文) |
| |
| 摘 要： | 针对前向神经网络（FNN）现有BP学习算法的缺点──收敛速度缓慢、容易陷入局部极小，提出一种快速、全局优化、简单通用的前向神经网络训练算法。这种方法将改进BP算法中的动量项由常量改换为一类关于常量的非线性函数。利用非线性特性，优化过程中能遍历局部极小，同时又具有突跳特性。结合升温策略，该算法在优化精度和网络训练速度两方面均有较大改善。通过对典型算例──异或问题的仿真，验证了算法的有效性和可行性。
|
| 关 键 词： | 前向神经网络 BP算法 带升温策略的改进算法 |
A Improved Algorithm Based on Nonlinear Property for FNN |
| |
| Zhang Sen.A Improved Algorithm Based on Nonlinear Property for FNN[J].Journal of Shenyang Institute of Engineering:natural Science,1999(3). |
| |
| Authors: | Zhang Sen |
| |
| Affiliation: | Zhang Sen(Shenyang Electric Power Institute)Zhang Huaguang; Wang Xiaowen(Northeast University) (Shenyang Electric Power Institute) |
| |
| Abstract: | To solve the slow convergence and the local minima of BP algorithm, a general and simple method named IBPM is presented. This method makes the nonlinear property of momentum term instead of the momentum term in improved BP algorithm. Based on the nonlinear property, the algorithm can approach the local minima and enable the system to escape them at large weight space. Using "temperature-raising" strate gy, convergence rate and training speed are greatly accelerated. Simulation results verify the efficiency of this method. |
| |
| Keywords: | FNN BP algorithm IBPM algorithm |
| 本文献已被 CNKI 等数据库收录！ |
