| 8,410 | 558 | 94 |
| 下载次数 | 被引频次 | 阅读次数 |
针对标准BP算法存在的缺陷,本文给出了基于MATLAB语言的BP神经网络几种改进的算法.阐述了各种BP算法的优化技术原理、优缺点,并就它们的训练速度和内存消耗情况作了比较.建议在多数BP神经网络训练时,先尝试使用Levenberg-Marquardt算法,其次是BFGS算法或共轭梯度法以及RPROP算法.
Abstract：To eliminate the shortcoming of standard backpropagation algorithm, some modified BP algorithms in the MATLAB's neural networks toolbox are given in the paper. These high performance algorithms are discussed in the optimization techniques and compared with speed and memory. For most situations, the Levenberg-Marquardt algorithm is advised to use first, then the BFGS algorithm, or one of the conjugate gradient methods, or RPROP algorithm is considered.
[1] RumelhartDE,HintonGE,WilliamsRJ.Learninginternalrepresentationsbyerrorpropagation[A].RumelhartDEJamesL.McClellandJL.Paralleldistributedprocessing:explorationsinthemicrostructureofcognition[C],volume1,Cambridge,MA:MITPress,1986.318～362.
[2] NeuralNetworkToolboxUser'sGuide.TheMathworks,inc.1999.
[3] 楼顺天,施阳.基于MATLAB的系统分析与设计———神经网络[M].西安:西安电子科技大学出版社,1998.143～154.
[4] 丛爽.面向MATLAB工具箱的神经网络理论与应用[M].合肥:中国科学技术大学出版社,1998.142～169.
[5] 袁曾任.人工神经元网络及其应用[M].北京:清华大学出版社,1999.131～273.
[6] 杨火斤,马洪波.人工神经网络在中长期汇率预测中的应用[J].科技导报,1999,4(3):38～41.
[7] 洪伟,吴承祯,何东进.基于人工神经网络的森林资源管理模型研究[J].自然资源学报,1998,13(1):69～76.
[8] 杨兆升,朱中.基于BP神经网络的路径行程时间实时预测模型[J].系统工程理论与实践,1999,19(8):59～64.
[9] 陈守余,周梅春.数据开采的BP神经网络方法及其应用研究[J].地球科学,1998,23(2):183～187.
[10] 张志华,朱章森,李儒兵.几种修正的自适应神经网络学习算法及其应用[J].地球科学,1998,23(2):179～182.
[11] 张智星,孙春在,水谷英二.张平安,高春华,高峰峦译.神经—模糊和软计算[M].西安:西安交通大学出版社,2000.139～234.
[12] FahlmanSE.Faster-learningvariationsonback-propagation:anempiricalstudy[A].TouretzkyD,HintonG,SejnowskiT.Proceedingsofthe1988ConnectionistModelsSummerSchool[C].CarnegicMellonUniversity,1988,38～51.
[13] JacobsRA.Increasedratesofconvergencethroughlearningrateadaptation[J].NeuralNetworks,1988,1:295～307.
[14] SharS,PalmieriF.MEKA—afast,localalgorithmfortrainingfeedforwardneuralnetworks[A].ProceedingsoftheInternationalJointConferenceonNeuralNetworks[C].IEEEPress,NewYork,1990.41～46.
[15] WatrousRL.Learningalgorithmsforconnectionistnetwork:appliedgradientmethodsofnonlinearoptimization[A].ProceedingsofIEEEInternationalConferenceonNeuralNetworks[c].IEEEPress,NewYork,1987.619～627.
[16] SharS,PalmieriF,DatumM.Optimalfilteringalgorithmsforfastlearninginfeedforwardneuralnetworks[J].NeuralNetworks,1992,5(5):779～787.
[17] MartinR,HeinrichB.ADirectAdaptiveMethodforFasterBackpropagationLearning:TheRPROPAlgorithrm[A].RuspiniH.ProceedingsoftheIEEEInternationalConferenceonNeuralNetworks(ICNN)[C].IEEEPress,NewYork.1993.586～591.
[18] FletcherR,ReevesCM.Functionminimizationbyconjugategradients[J].ComputerJournal,1964,7:149～154.
[19] PowellMJD.Restartproceduresfortheconjugategradientmethod[J].MathematicalProgramming,1977,12:241～254.
[20] MollerMF.Ascaledconjugategradientalgorithmforfastsupervisedlearning[J].NeuralNetworks,1993,6:525～533.
[21] BattitiR.Firstandsecondordermethodsforlearning:BetweensteepestdescentandNewton'smethod[J].NeuralComputation,1992,4(2):141～166.
[22] HaganMT,MenhajM.TrainingfeedforwardnetworkswiththeMarquardtalgorithm[J].IEEETransactionsonNeuralNetworks,1994,5(6):989～993.
基本信息:
DOI：10.13774/j.cnki.kjtb.2003.02.012
中图分类号:TP183
引用信息:
[1]苏高利,邓芳萍.论基于MATLAB语言的BP神经网络的改进算法[J].科技通报,2003(02):130-135.DOI:10.13774/j.cnki.kjtb.2003.02.012.
基金信息: