1引言各种神经网络被广泛用来解决模式识别问题.进行模式识别时,常常要对模式进行分类从而达到识别的目的.线性可分模式的分类是其中最基本的一种.当前神经网络的一个热点问题集中在研究神经网络学习算法的理论基础,特别是学习算法的收敛性证明上.作为神经网络基本组成的感知器,对线性可分问题具有正确分类的能力不仅在实践中行之有效,而且在理论上已经证明是收敛的.文献[1]证明了在样本线性可分的条件下,对于传统感知器模型,在线梯度算法有限收敛.这是传统感知器作为一种最简单的神经网络所具 1 Introduction Various neural networks are widely used to solve pattern recognition problems.For pattern recognition, it is often necessary to classify the patterns to achieve the purpose of recognition.Linear separable patterns classification is one of the most basic one.Now the current neural network A hot issue focuses on the theoretical basis of studying neural network learning algorithms, especially the convergence proof of learning algorithms.As the basic components of neural network, the ability to correctly classify linear separable problems not only works well in practice, And it has been proved to be theoretically convergent. [1] proved that the online gradient algorithm has a limited convergence for the traditional perceptron model under the condition of the linear separability of the samples. This is the traditional sensor as one of the simplest neural networks With