论文部分内容阅读
提出一种资源分配网络(Resource Allocating Network,RAN)的新的学习算法,称为 IRAN 算法.该算法通过一个包含4部分的新颖性准则来增加网络中的隐层神经元,通过误差下降速率来删除冗余神经元并采用基于Givens-QR 分解的递归最小二乘算法进行输出层权值的更新.通过函数逼近领域中2个 Benchmark 问题的仿真结果表明,与 RAN,RANEKF,MRAN 算法相比,IRAN 算法不但学习速度快,而且可以得到更为精简的网络结构.
A new learning algorithm of Resource Allocating Network (RAN) is proposed, which is called IRAN algorithm.This algorithm increases the hidden neurons in the network by a four-part novelty criterion, and through the rate of error decrease The redundant neurons are deleted and the output layer weights are updated by using the recursive least squares algorithm based on Givens-QR decomposition.The simulation results of two Benchmark problems in the field of function approximation show that compared with the RAN, RANEKF and MRAN algorithms, The IRAN algorithm not only learns quickly, but also gets a leaner network structure.