计算机工程与应用 ›› 2016, Vol. 52 ›› Issue (13): 157-161.

• 模式识别与人工智能 • 上一篇    下一篇

FRBF神经网络分类器设计新方法

徐明亮1,2,赵  吉3,唐玉兰1   

  1. 1.无锡环境科学与工程研究中心,江苏 无锡 214153
    2.江南大学 数字媒体学院,江苏 无锡 214122
    3.江南大学 电气自动化研究所,江苏 无锡 214122
  • 出版日期:2016-07-01 发布日期:2016-07-15

New devise method of FRBF neural network classifier

XU Mingliang1,2, ZHAO Ji3, TANG Yulan1   

  1. 1.Wuxi Research Center for Environmental Sciences and Engineering, Wuxi, Jiangsu 214153, China
    2.School of Digital Media, Jiangnan University, Wuxi, Jiangsu 214122, China
    3.Institute of Electrical Automation, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2016-07-01 Published:2016-07-15

摘要: 提出了一种结合模糊径向基函数网络和稀疏V-SVM的二分类器构建方法。FRBF初始网络中的RBF隶属度函数中心由随机抽取的样本确定,而RBF隶属度函数的宽度由样本各个属性的分布方差确定。根据FRBF网络输出为模糊基函数线性组合的特点,在后件参数学习中引入具有结构风险最小化和属性选择功能的稀疏V-SVM方法,在对输出层的参数进行学习的同时进行模糊基函数的约简。若干UCI标准数据集分类测试结果验证了该分类器的有效性。

关键词: 模糊径向基函数网络, 支撑向量机, 约简, 分类

Abstract: A binary classifier based on the Fuzzy Radial Basis Function Network(FRBFN) and SP-V-SVM is presented. The initial architecture of the network is constructed with the sample from dataset. The centers of Gaussian membership functions of each membership variable in the fuzzy layer are determined by the samples randomly extracted in the training data set, whereas the variances depend on the variance of the training data set. The parameters of output layer are accomplished based on the criterion of the max gap between classes. What are more, the nodes of the network sparsity constraints are introduced to realize nodes reduction. The classification tests on several UCI standard data sets are conducted and the results show the effectiveness of the classifier.

Key words: Fuzzy Radial Basis Function Network(FRBFN), Support Vector Machine(SVM), reduction, classifier