一种基于粗糙集的粗糙神经网络构造方法  

Approach to Construct a Rough Neural Networks Based on Rough Set

作　　者：何明[1] 冯博琴[1] 马兆丰[1] 傅向华[1] 

机构地区：[1]西安交通大学计算机科学与技术系,西安710049

出　　处：《西安交通大学学报》2004年第12期1240-1242,1246,共4页Journal of Xi'an Jiaotong University

基　　金：国家高技术研究发展计划资助项目 (2 0 0 3AA1Z2 6 10 )

摘　　要：针对神经网络中各神经元和权不能赋予明确的物理意义 ,提出了一种基于粗糙集的神经网络模型 .该模型利用粗糙集理论数值分析的能力 ,首先从给定的数据集中抽取出规则 ,然后根据这些规则构造神经网络隐含层的神经元个数 ,从而确定粗糙神经网络的初始拓扑结构 .同时 ,将输入映射到输出子空间 ,并在这个子空间上用神经网络进行逼近 ,由此得到一种可理解性好、收敛速度快的神经网络模型 .实验结果表明 ,该模型能够较好地处理神经网络拓扑结构、训练样本的大小、样本质量等对神经网络的精度及泛化能力有直接影响的问题 ,在大大缩短训练时间的同时 ,它的预测精度可达 96 4 % ,较同条件下径向基函数神经网络模型的精度高 3 6 % .Aiming at the problem that clear physical meanings can't be given by nerve cells and weights of neural networks, a neural network model based on rough set was proposed, in which rules were extracted from given training data firstly by utilizing numerical analysis ability of rough set theory, and then neurons number of the hidden layer was determined in terms of these rules to obtain the original topology of the rough neural network. Meanwhile, the input to the model was mapped into the output subspace by using rules acquired from the rough set and the expectation output could be approximated. Thus, a neural network that provides with good understandability and rapid convergence could be constructed. Experiments show that the proposed approach can deal with problems of neural network topology architecture, sample size and quality which directly influence the generalization ability and accuracy of neural network. While greatly reducing training time, the prediction precision of the model can be achieved 96. 4% which is 3.6% higher than RBF (radial basis function) neural network model under the same conditions.

关 键 词：粗糙集 神经网络 粗糙集数据分析 粗糙神经元 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]