多层自适应模块化神经网络结构设计  

Structure Design of Hierarchical Adaptive Modular Neural Network

作　　者：张昭昭[1] 乔俊飞 余文 ZHANG Zhao-Zhao;QIAO Jun-Fei;YU Wen(Institute of Electronic and Information Engineering,Liaoning Technical University,Huludao,Liaoning125105;Department of de Control Automatic,CINVESTA-IPN, Mexico City D.F.07360,Mexico)

机构地区：[1]辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105 [2]北京工业大学信息学部,北京100124 [3]墨西哥国立理工高级研究中心自动控制中心,墨西哥城墨西哥07360

出　　处：《计算机学报》2017年第12期2827-2838,共12页Chinese Journal of Computers

基　　金：国家自然科学基金(61440059,61533003);国家杰出青年科学基金(61225016);辽宁省自然科学基金(201602363);国家留学基金委资助~~

摘　　要：针对单一全互连前馈神经网络难以应对复杂问题及模块化神经网络应用时其结构难以确定的问题,该文基于脑式信息处理是采用无监督学习-半监督学习-监督学习的学习机制以及大脑是由多个功能模块组成,每个功能模块中又包含多个子模块,大脑对信息的学习是有目的的选择不同功能模块中多个子模块协同学习的事实,提出一种多层自适应模块化神经网络结构设计方法.其实质是首先对所有的训练数据采用概率密度峰值快速聚类算法确定训练数据的聚类中心,以此确定模块化神经网络中功能模块的个数,其次采用条件模糊聚类实现对每个功能模块中子模块的划分并确定每个子模块的训练样本集;对功能模块中的每一个子模块采用训练误差峰值构造RBF网络的增长算法,该算法能根据分配来的训练样本自适应构建子模块结构;在子模块集成方面,采用基于距离测度的子模块集成方法,该方法能从不同的功能模块中选择不同的子模块对训练样本协同处理.该文提出的模块化神经网络结构设计方法只需要2个人工参数且学习速度提高了近10倍,在一定程度上实现了神经网络的黑箱效应.最后,文中基于人工数据集的复杂函数拟合问题、双螺旋分类问题以及真实数据集的回归问题进行了实验,并与当前国际流行的网络结构进行了对比,结果显示文中提出的模块化神经网络网络结构解决了全互连RBF网络难以应对的复杂问题,而且学习精度高,学习速度快,最终网络的泛化性能强.Modular neural network is an effective method to solve the complex problems that the monolithic fully coupled feedforward neural networks are difficult to learn.The most difficult problem of the modular neural network design method now facing is how to determine the number of function modules and the structure of the sub modules in each function module under the condition of lack of the learned objects back ground knowledge.In this paper,we presents a hierarchical adaptive modular neural network structure design method based on the facts that the brain like information process uses the mechanism of unsupervised learning,semi supervised learning and supervised learning,and the brain networks demonstrate the property of hierarchical modularity,within each module there will be a set of function modules,and within each function there will be a set of sub modules,and the brain like information learning processing is purposeful to select several sub modules from different function module to collaboratively learning.In essence,firstly,an unsupervised clustering algorithm named fast finding of density peaks cluster algorithm is adopted to identify the cluster centers based on all training samples,then the number of the function module and the training set of each function module can be determined.Secondly,a semi supervised clustering algorithm named conditional fuzzy clustering algorithm is used to divide the training set of each function module into several groups to determine the number of sub modules in each function module.For each sub module,an incremental design of RBF network algorithm based on training error peak is applied to construct the structure of sub module,this algorithm can adaptively build the structure of the sub modules based on the training samples that allocated to the sub modules.In sub modules integration,a sub module integrative approach based on relative distance measure is applied which can select different sub modules from different function modules to collaboratively learning the training sa

关 键 词：模块化神经网络 自适应 径向基函数 脑式信息处理 协同学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]