基于RBFLN网络的改进RBF神经网络学习算法  

Improved radial basis function neural network learning algorithm based on radial basis functional link nets

作　　者：韩敏[1] 穆云峰[1] 

机构地区：[1]大连理工大学电子与信息工程学院,辽宁大连116023

出　　处：《系统工程学报》2008年第6期764-768,共5页Journal of Systems Engineering

基　　金：国家自然科学基金资助项目(60674073);国家科技支撑计划资助项目(2006BAB14B05);国家重点基础研究发展计划(973)资助项目(2006CB403405)

摘　　要：提出了一种基于径向基链网络(RBFLN)的改进径向基函数(RBF)网络学习算法.网络结构采用 RB-FLN 模型,添加输入层对输出层的线性映射,在训练过程中基于最大误差学习样本对资源分配网络(RAN)新性条件进行改动,在不满足新性条件时,采用相似度参数对隐层中心和宽度进行调整;而满足新性条件时,对新增隐层节点也通过类均值的方法做出相应的改进.最后通过对无机建筑材料成分分析的仿真表明该算法可有效地简化网络结构,实现样本正确分类,并获得较好的校验能力.An improved RBF （radial basid function） neural network learning algorithm based on RBFLN （radial basis functional link nets） is proposed in this paper. The architecture of RBFLN, which includes a nonlinear model and an additive linear model due to extra linear mapping from the input to the output nodes, is used. The pattern which generates the maximal error is chosen to adjust the RAN （resource allocating network） novelty during learning process. According to the novelty, the existing parameters of the network will be adjested by a similarity parameter or adding a new hidden unit. Finally, simulations of a composition of building materials shows that the algorithm can realize correct classification of patterns, and improve better testing performance, while only needs smaller network architecture.

关 键 词：径向基链网络 资源分配网络 最大误差样本 相似度 材料成分 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]