一种基于AGA的RBF网络混合学习算法  

A Hybrid Learning Method for RBFNN Based on AGA

作　　者：高运广[1] 李钊[2] 史德琴[3] 刘顺波[1] 

机构地区：[1]第二炮兵工程学院,陕西西安710025 [2]第二炮兵驻石家庄地区军事代表室,河北石家庄050081 [3]空军工程大学工程学院,陕西西安710038

出　　处：《无线电工程》2008年第11期25-28,共4页Radio Engineering

摘　　要：基于目前RBF网络学习方法中的一些不足,提出了一种基于AGA的混合学习方法,即应用AGA对网络隐单元RBF个数和宽度σ同时优选,并将最佳隐单元数作为K-均值聚类数得到隐单元中心,隐层到输出层的权值由LS法确定。针对K-均值聚类算法对初始值敏感的问题,算法在最后阶段对其执行多次运算,由此选择最佳结果。仿真结果表明,该方法在大样本情况下,训练得到的网络在精度和结构上得到了良好的结合。This paper introduces a hybrid learning method based on AGA to overcome the weakness existing in some methods for RBFNN learning.h optimizes both the number and the width of the RBF in hidden layer by AGA.The hidden unit centers are computed by K-means cluster algorithm as the clustering number has been selected by AGA.The weights of output layer are calculated by the I~ algorithm. Considering that the K-means cluster algorithm is sensitive to the initialization, this algorithm is implemented so many times at the last step that the best result is selected for the hidden units. The simulation indicates the net based on this learning method has a good performance both in precision and structure.

关 键 词：自适应遗传算法 径向基函数网络 隐单元 K-均值聚类 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]