Takagi-Sugeno型分层混合模糊-神经网络训练算法研究  

TRAINING ALGORITHMS FOR HIERARCHICAL HYBRID FUZZY-NEURAL NETWORKS DERIVED FROM TAKAGI-SUGENO MODEL

作　　者：邢海花[1,2] 余先川[1] 胡丹[1] 王根霞[1] 

机构地区：[1]北京师范大学信息科学与技术学院,北京100875 [2]海南师范大学信息科学与技术学院,海口571158

出　　处：《北京师范大学学报（自然科学版）》2013年第4期351-356,共6页Journal of Beijing Normal University(Natural Science)

基　　金：国家自然科学基金资助项目(41072245;41272359);教育部博士点基金资助项目(20120003110032);海南省自然科学基金资助项目(612123)

摘　　要：为寻求适合离散和连续变量并存数据特点的最佳训练算法,本文对基于三角波,梯形,Gaussian隶属函数的Takagi-Sugeno型分层混合模糊-神经网络训练算法进行综合比较研究,使用梯度下降法详细推导了各个参数的迭代规则,并给出了完整的训练算法和详细的参数调整过程.以Pyrimidines和Heart-UCI为数据分别进行函数逼近与数据分类实验,并与经典的BP算法比较.实验结果表明Takagi-Sugeno型分层混合模糊-神经网络在高维且离散与连续变量混合的函数逼近和分类上比BP神经网络更具优势,而基于Gaussian隶属函数的训练算法函数逼近精度比基于三角波和梯形隶属函数的训练算法精度更高.To optimize the training algorithm for dataset with discrete and continuous variables,hierarchical hybrid fuzzy-neural network（HHFNN） was analyzed based on triangular,trapezoid,and Gaussian functions and Takagi-Sugeno fuzzy system.Detailed derivation of iterative rules for parameters using a gradient descent method was given,training algorithm and adjustment process of parameters were given.Pyrimidines and Heart-UCI datasets were employed to test performance of these training algorithms,and comparison made with standard BP algorithm.In regards to system or function approximation and data classification with mixed variables,it was found that HHFNN model based on Takagi-Sugeno outperformed standard BP neural network in accuracy,training algorithm using Gaussian membership function was more accurate than using triangular and trapezoid membership functions.

关 键 词：神经网络 模糊系统 混合变量 Takagi-Sugeno型 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]