径向基函数神经网络的软竞争学习算法

本文构造了径向基函数(RBF)神经网络的一类软竞争学习算法(SCLA).该算法的主要思想是首先在高斯基函数中心向量的训练过程中引入了隶属度函数,对每个输入样本,所有中心向量根据该样本属于其代表的类的隶属度值的大小进行自适应地调整;第二,把隶属度函数的模糊因子的倒数与模拟退火算法中的温度等同起来,在迭代过程中采用递增的方式来调整它.SCLA是RBF网络基于k-均值方法训练中心向量的学习算法的软竞争格式,它可以克服后者对初始值敏感和死节点的问题.仿真实验论证了SCLA是有效的.

Soft-Competition Learning Algorithms of the Radial Basis Function Neural Networks

In this paper,the soft competition learning algorithms (SCLA) of RBF neural networks are designed.The main ideas of the algorithms are:firstly,membership functions are introduced into the training procedure of the center vectors of the Gaussian basis functions,and for each input sample,all center vectors are self adaptively adjusted according to the values of the membership functions,in what degree the sample belongs to the classes that the center vectors represent;secondly,the reciprocal of the fuzzy factor of membership function are considered as the temperature of the simulated annealing algorithm,and increasingly adjusting method is used to the fuzzy factor during the learning procedure.SCLA are soft competition schemes of the learning algorithms,in which the center vectors are trained based on the k means algorithm,and can remedy the problems of the dead node and the sensitivity to initial weight vectors that the latter algorithms have.The simulation results show that SCLA are efficient.