一种前馈过程神经元网络初始化方法及改进学习算法  

Initialization Method and Improved Learning Algorithm of Feed Forward Process Neural Networks

作　　者：杨露[1] 沈怀荣[1] 

机构地区：[1]装备指挥技术学院,北京101416

出　　处：《系统仿真学报》2009年第13期3962-3964,3974,共4页Journal of System Simulation

基　　金：国防预研项目(513270302)

摘　　要：分析了前馈过程神经元网络初始化对其训练速度的影响。提出了一种前馈过程神经元网络初始化方法,该方法将阈值初始化为时间积分运算的均值,可消除网络初始化不当的影响。提出了改进的网络学习算法,采用与初始化相似的方法修正阈值,可加快前馈过程神经元网络的训练速度。以时变信号分类为例,仿真验证了初始化方法及改进学习算法的正确性和有效性。The initialization method affects the training speed of feed forward Process Neural Networks （PNN）. A new initialization method of feed forward PNN was proposed, which initialized the threshold as the mean value of time integral. It could eliminate the bad effects of unsuitable network initialization. The improved learning algorithm was advanced which modified the thresholds in the similar way of initialization. It could increase the training speed of feed forward PNN. The case of time-varied signal classification shows the correctness and effectiveness of the new initialization method and the improved algorithm.

关 键 词：过程神经元 过程神经网络 时变系统 学习算法 初始化 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]