基于多分辨率学习的正交基小波神经网络设计  

Design of orthogonal basis wavelet neural network based on multiresolution learning

作　　者：陈增强[1] 任东[1] 袁著祉[1] 杜升之[1] 

机构地区：[1]南开大学自动化系,天津300071

出　　处：《系统工程学报》2003年第3期218-223,共6页Journal of Systems Engineering

基　　金：国家自然科学基金资助项目(60174021);天津市自然科学重点基金资助项目(013800711).

摘　　要：提出一种基于正交基函数的小波神经网络设计方法,采用多分辨率学习确定隐含层结构,并用收敛较快的阻尼最小二乘法训练权值.该方法可灵活调整隐含层结构,有效地克服神经网络中常见的过拟合和泛化能力差等问题.仿真结果表明该方法具有逼近精度高,泛化能力好,网络结构冗余度小,参数优化收敛快等特点.This paper proposes a designing approach for wavelet neural network based on orthogonal basis function. In this approach the network structure is determined by multiresolution learning, and the weights are trained by damped least squares which has fast convergent rate. This approach can adjust the hidden layer structure with great flexibility and effectively overcome the problems of fitting too much and bad generalization in the neural network. The simulation results show that this approach has the advantages of high approximation, good generalization ability, small structure redundancy and fast convergent rate.

关 键 词：神经网络 设计 正交基小波 多分辨率学习 最小二乘法 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]