过程神经元网络学习算法中的积分方法分析  

Analysis of Integral Method in Learning Algorithm of 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14期4231-4234,共4页Journal of System Simulation

基　　金：国防预研项目(513270302)

摘　　要：过程神经元网络常采用基于正交基展开的学习算法,以简化积分运算。分析对比了基于正交基展开和基于梯形公式两种不同积分方法,提出并证明了网络结构一定时,两种方法可以使网络达到相同的误差精度,并推论出积分运算方法不影响网络训练所能达到的误差精度。两种方法具有不同的适用情况,连续函数输入适合采用正交基展开法,在网络输入为离散等距采样点时,基于梯形公式的方法能够在不影响网络原始输入数据的前提下简化运算,避免了由原始数据构造拟合曲线或平滑插值,再进行正交基展开的过程。Orthogonal function basis expansion was usually used in Process Neural Networks （PNN） to simplify the integral operation. Two different integral methods were analyzed and compared which are orthogonal function basis expansion method and trapezoid integral method. It was proposed and proved that for a certain network both methods can train it to the same precision. And it was inferred that the integral method has no effect on the precision that the trained network can reach. Each integral method has its applicable condition. The orthogonal function basis expansion method is more applicable to the continuous function inputs. When PNN has the equidistantly sampled discrete inputs, the trapezoid integral method can simplify the calculation without changing the original inputs, which omits the eurvefitting or interpolating of original data and the orthogonal function basis expansion.

关 键 词：过程神经元 过程神经网络 时变系统 学习算法 函数正交基 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]