摘要： 提出了一种单层感知器网络训练的新算法。证明了对于线性可分问题和线性不可分问题,算法总是在有限步内终止,算法的迭代次数以模式数为上界;而且,在算法终止时,对于线性可分问题,总是能得到正确的权向量解,所以,如果在算法结束时还不能划分所有模式,则说明给定的模式集确是不可线性划分的。

关键词: 单层感知器网络, 增广型模式向量, 增广型权值向量, 权空间, 模式超平面

Abstract: A new algorithm is proposed for training single layered perception neural networks. It is proved that the algorithm is terminated in finite steps for linearly separable problems and linearly inseparable problems, the upper limit of the iteration number of the algorithm is the number of the patterns.Furthermore, on the termination of the algorithm, for linearly separable problems, it always finds the correct weight vector that can separate all the patterns.Thus a running of the algorithm which fails to separate all given patterns means that the presented set of patterns is linearly inseparable.

中图分类号: 

• TP183