基于RBF神经网络并行学习模型的数据分类及预测研究
|
摘要
随着现代信息技术的迅速发展，许多领域都积累了大量的数据。我们渴望发现潜在于这些数据中的知识与规律。正是这一需求造就了数据挖掘学科的兴起及数据挖掘技术的发展。作为一个多学科交叉的综合性领域，数据挖掘涉及了数据库、统计学、机器学习、高性能计算、模式识别、神经网络和数据可视化等学科。数据分类与预测作为一种重要的挖掘技术有着广泛的应用。在这一研究方向，目前已提出了多种分类方法（如决策树归纳分类、贝叶斯分类、神经网络分类和K-最邻近分类等）和一些预测技术（如线性回归、非线性回归等）。然而，尚未发现有一种方法对所有数据的处理都优于其他方法[1]。
由于时间序列数据库的日趋庞大及其挖掘的潜在意义，目前，时序数据挖掘研究已成为一个热点；然而,时间序列数据的非线性混沌特点，使得对它的挖掘成为难题。本文在分析与比较以上几种分类及预测方法的基础上，引入了径向基函数神经网络（Radial Basis Function Neural Network，简称RBFNN）对时间序列数据进行预测。在介绍该神经网络优点的同时，也阐述了其中较为棘手的难点。文中使用了层次遗传算法作为神经网络学习方法。在分析该方法可行性及效率的基础上，本文提出了用粗粒度并行方法进行径向基函数神经网络训练的思想，并建立了求解模型，旨在取得较好的预测效果。
最后，本文应用上述并行模型优化的RBF神经网络对非线性函数值以及证券个股收盘价进行预测。实验结果表明，当数据无噪声时，预测效率与精度都非常高；在处理带噪声，并呈现混沌特性的数据时，虽有一定的误差，但预测结果还是在可以接受的范围内。
With the rapid development of modern information technology, a great deal of data has been accumulated in many fields. People expect to discover the knowledge and rules existing in these data, which just brings the study of data mining and the development of its technology. As a comprehensive field of crossing multi-subject, data mining involved many subjects such as database, statistic, machine learning, high performance computing, pattern recognition, neural network and data visualization etc. Data classification and prediction are important mining technologies and have been used widely. Nowadays, many classification methods and some prediction technologies have been put forward, such as Classification by Decision Tree Induction, Bayesian Classification, Classification by Backpropagation, k-Nearest Neighbor Classifiers, Linear and Nonlinear regression. However, none of them is better than others in all application.
Because of the growing of Time-Series Database and the potential significance of data mining, the research of data mining in Time-Series Database has become a hotspot. At the same time, however, the nonlinear and chaotic characteristic of time-series data makes the mining be a difficult issue. Based on the analysis and comparison of these classification and prediction methods, this paper introduces a method that uses Radial Basis Function Neural Network (RBFNN) to make prediction for time-series data. As the advantage of this neural network is introduced, some hot potatoes are also discussed. This paper takes Hierarchical Genetic Algorithm as the neural network learning method. After analyzing the feasibility and efficiency of this method, we put forward an idea of using the coarse grained parallel method for Radial Basis Function Neural Network learning, and on purpose to get satisfactory prediction effect, we set up a model to solve corresponding learning.
At last, this paper uses the RBF neural network that was optimized by the mentioned parallel model to predict the value of some nonlinear functions and the close of several stocks. The result shows that the efficiency and precision of prediction for clean data are satisfactory. Although there are some errors in the prediction of noisy and chaotic data, the result is acceptable.
由于时间序列数据库的日趋庞大及其挖掘的潜在意义，目前，时序数据挖掘研究已成为一个热点；然而,时间序列数据的非线性混沌特点，使得对它的挖掘成为难题。本文在分析与比较以上几种分类及预测方法的基础上，引入了径向基函数神经网络（Radial Basis Function Neural Network，简称RBFNN）对时间序列数据进行预测。在介绍该神经网络优点的同时，也阐述了其中较为棘手的难点。文中使用了层次遗传算法作为神经网络学习方法。在分析该方法可行性及效率的基础上，本文提出了用粗粒度并行方法进行径向基函数神经网络训练的思想，并建立了求解模型，旨在取得较好的预测效果。
最后，本文应用上述并行模型优化的RBF神经网络对非线性函数值以及证券个股收盘价进行预测。实验结果表明，当数据无噪声时，预测效率与精度都非常高；在处理带噪声，并呈现混沌特性的数据时，虽有一定的误差，但预测结果还是在可以接受的范围内。
With the rapid development of modern information technology, a great deal of data has been accumulated in many fields. People expect to discover the knowledge and rules existing in these data, which just brings the study of data mining and the development of its technology. As a comprehensive field of crossing multi-subject, data mining involved many subjects such as database, statistic, machine learning, high performance computing, pattern recognition, neural network and data visualization etc. Data classification and prediction are important mining technologies and have been used widely. Nowadays, many classification methods and some prediction technologies have been put forward, such as Classification by Decision Tree Induction, Bayesian Classification, Classification by Backpropagation, k-Nearest Neighbor Classifiers, Linear and Nonlinear regression. However, none of them is better than others in all application.
Because of the growing of Time-Series Database and the potential significance of data mining, the research of data mining in Time-Series Database has become a hotspot. At the same time, however, the nonlinear and chaotic characteristic of time-series data makes the mining be a difficult issue. Based on the analysis and comparison of these classification and prediction methods, this paper introduces a method that uses Radial Basis Function Neural Network (RBFNN) to make prediction for time-series data. As the advantage of this neural network is introduced, some hot potatoes are also discussed. This paper takes Hierarchical Genetic Algorithm as the neural network learning method. After analyzing the feasibility and efficiency of this method, we put forward an idea of using the coarse grained parallel method for Radial Basis Function Neural Network learning, and on purpose to get satisfactory prediction effect, we set up a model to solve corresponding learning.
At last, this paper uses the RBF neural network that was optimized by the mentioned parallel model to predict the value of some nonlinear functions and the close of several stocks. The result shows that the efficiency and precision of prediction for clean data are satisfactory. Although there are some errors in the prediction of noisy and chaotic data, the result is acceptable.
引文
[1] Jiawei Han，Micheline Kamber：Data Mining：Concepts and Techniques. Morgan Kaufmann Publishers,Inc. 2002.5
[2] 史忠植《知识发现》清华大学出版社2002.1
[3] 叶中行、顾立庭 《混合认识系统及其在股市分析上的应用》 上海交通大学学报 29卷 第2期
[4] 王松桂、陈敏、陈立萍 《线性统计模型》 高等教育出版社 2002.4
[5] 高惠璇 《统计计算》 北京大学出版社 1995.7
[6] 李士勇《模糊控制.神经控制和智能控制论》哈尔滨工业大学出版社 1996.10
[7] 李冬梅、王正欧 《基于RBF神经网络的混沌时间序列预测》 模式识别与人工智能 2001.6
[8] 王学武、谭得健 《神经网络的应用与发展》计算机工程与应用 2003.3
[9] Robert J.Schilling,James J.Carroll "Approximation of Nonlinear Systems with Radial Basis Function Neural Networks" IEEE 2001.7
[10] 黄德双《神经网络模式识别系统理论》电子工业出版社1996.5
[11] Simon Haykin "Neural Networks A Comprehensive Foundation" Second Edition Tsinghua University Press 2001.10
[12] 王学雷、邵惠鹤 《一种径向基函数神经网络在线训练算法及其在非线性控制中的应用》2001.6
[13] [俄]A.И.加卢什金 著，阎平凡 译 《神经网络理论》清华大学出版社 2002.12
[14] 刘皓、魏平、肖先赐《面向特定结构的几种BP并行算法及比较》系统工程与电子技术2000.1
[15] 王旭东、邵惠鹤 《RBF神经网络理论及其在控制中的应用》信息与控制 1997.8
[16] 陈国良、庄镇泉 《遗传算法及其应用》 人民邮电出版社 1996.6
[17] 王正志、薄涛 《进化计算》国防科技大学出版社 2000.11
[18] 都志辉 《高性能计算并行编程技术——MPI并行程序设计》清华大学出版社 2001.8
[19] Huang K，Xu Z. "Scalable Parallel Computing：Technology，Architecture，Programming". 机械工业出版社 1999.5
[20] 俞时权、胡浩民 《机群系统中基于MPI的多维稀疏数组传递方法》 计算机工程2003.4
[21] N.Davey,S.P.Hunt,R.J.Frank "Time Series Prediction and Neural Networks" University of Hertfordshire Hatfield,UK From NEC Research Institute 1999
[22] Martin T.Hangan、Howard B.Demuh、Mark H.Beale 《神经网络设计》机械工业出版社 2002.9
[2] 史忠植《知识发现》清华大学出版社2002.1
[3] 叶中行、顾立庭 《混合认识系统及其在股市分析上的应用》 上海交通大学学报 29卷 第2期
[4] 王松桂、陈敏、陈立萍 《线性统计模型》 高等教育出版社 2002.4
[5] 高惠璇 《统计计算》 北京大学出版社 1995.7
[6] 李士勇《模糊控制.神经控制和智能控制论》哈尔滨工业大学出版社 1996.10
[7] 李冬梅、王正欧 《基于RBF神经网络的混沌时间序列预测》 模式识别与人工智能 2001.6
[8] 王学武、谭得健 《神经网络的应用与发展》计算机工程与应用 2003.3
[9] Robert J.Schilling,James J.Carroll "Approximation of Nonlinear Systems with Radial Basis Function Neural Networks" IEEE 2001.7
[10] 黄德双《神经网络模式识别系统理论》电子工业出版社1996.5
[11] Simon Haykin "Neural Networks A Comprehensive Foundation" Second Edition Tsinghua University Press 2001.10
[12] 王学雷、邵惠鹤 《一种径向基函数神经网络在线训练算法及其在非线性控制中的应用》2001.6
[13] [俄]A.И.加卢什金 著，阎平凡 译 《神经网络理论》清华大学出版社 2002.12
[14] 刘皓、魏平、肖先赐《面向特定结构的几种BP并行算法及比较》系统工程与电子技术2000.1
[15] 王旭东、邵惠鹤 《RBF神经网络理论及其在控制中的应用》信息与控制 1997.8
[16] 陈国良、庄镇泉 《遗传算法及其应用》 人民邮电出版社 1996.6
[17] 王正志、薄涛 《进化计算》国防科技大学出版社 2000.11
[18] 都志辉 《高性能计算并行编程技术——MPI并行程序设计》清华大学出版社 2001.8
[19] Huang K，Xu Z. "Scalable Parallel Computing：Technology，Architecture，Programming". 机械工业出版社 1999.5
[20] 俞时权、胡浩民 《机群系统中基于MPI的多维稀疏数组传递方法》 计算机工程2003.4
[21] N.Davey,S.P.Hunt,R.J.Frank "Time Series Prediction and Neural Networks" University of Hertfordshire Hatfield,UK From NEC Research Institute 1999
[22] Martin T.Hangan、Howard B.Demuh、Mark H.Beale 《神经网络设计》机械工业出版社 2002.9