高等学校化学学报 ›› 1996, Vol. 17 ›› Issue (6): 861.

• 论文 • 上一篇    下一篇

人工神经网络用于化学数据解析的研究(Ⅰ)──逼近规律与过拟合

刘平, 梁逸曾, 张林, 俞汝勤   

  1. 湖南大学化学计量学与化学传感技术研究所, 长沙 410082
  • 收稿日期:1995-05-11 出版日期:1996-06-24 发布日期:1996-06-24
  • 通讯作者: 梁逸曾
  • 作者简介:刘平,男,39岁,副教授,现在常德高等师范专科学校工作.
  • 基金资助:

    国家自然科学基金;国家教育委员会霍英东基金

Artificial Neural Networks Applied in the Analysis of Chemical Data(Ⅰ)──Approximation for Trend and Over-fitting

LIU Ping, LIANG Yi-Zeng, ZHANG Lin, YU Ru-Qin   

  1. Insititue of Chemometrics and Chemical Sensing Technology, Hunan University, Changsha 410082
  • Received:1995-05-11 Online:1996-06-24 Published:1996-06-24

摘要: 对多层前传网络的过拟合问题进行了探讨。定义了逼近误差和逼近度作为人工神经网络(ANN)的建模评价指标。通过应用于多元非线性校正的数值模拟和实际药物光度分析数据解析,表明该指标意义明确,便于掌握,且能较好地定量表述ANN逼近规律的程度。

关键词: 人工神经网络, 过拟合, 逼近误差, 逼近度, 多元非线性校正

Abstract: Over-fitting in feedforward multilayer neural network was studied in this paper.Approximation error and the degree of approximation were introduced as the evaluation index for modelling in ANN.Numerical simulation and real pharmaceutical spectrophotometricdata were analyzed by the proposed method.Results showed that the index was simple,convenient and described properly the quantitative degree of approxirnation for trend by ANN.

Key words: Artificial neural network, Over-fitting, Approximation-error, Degree of approximation, Multicomponent nonlinear calibration

TrendMD: