计算机工程与应用 ›› 2009, Vol. 45 ›› Issue (8): 5-8.DOI: 10.3778/j.issn.1002-8331.2009.08.002

• 博士论坛 • 上一篇    下一篇

新型NN训练算法及其在优化设计中的应用

任 远,白广忱   

  1. 北京航空航天大学 能源与动力工程学院,北京 100083
  • 收稿日期:2008-09-11 修回日期:2008-12-22 出版日期:2009-03-11 发布日期:2009-03-11
  • 通讯作者: 任 远

New NN training algorithm and its application in optimization design

REN Yuan,BAI Guang-chen   

  1. School of Jet Propulsion,Beijing University of Aeronautics and Astronautics,Beijing 100083,China
  • Received:2008-09-11 Revised:2008-12-22 Online:2009-03-11 Published:2009-03-11
  • Contact: REN Yuan

摘要: 提出采用GA-BP贝叶斯算法来建立优化设计近似模型。该算法是一种新型神经网络训练算法,它以提高网络的泛化性能为主旨,其训练目标是获取对应于后验分布最大值的权值向量。以方形扁平封装器件为例,采用GA-BP贝叶斯算法建立了温度场分析的近似模型,基于它对封装散热结构进行了优化,并与L-M BP算法进行了对比。结果表明,基于GA-BP贝叶斯算法的温度场分析近似模型,对芯片中心温度的预测精度更为理想,并且受随机因素的影响很小。GA-BP贝叶斯算法克服了现有网络训练算法对初始权值敏感、建模精度不高的缺点,在工程优化设计中具有实用价值。

关键词: 优化设计, 热设计, 神经网络

Abstract: The GA-BP Bayesian algorithm is used to establish the approximation model for optimization design.This algorithm is a new NN training algorithm developed by the authors.Its aim is to improve the generalization ability of neural networks,and it trains a network with the purpose of obtaining the weights corresponding with the maximum posterior probability.Taking a quad flat package for example,the GA-BP Bayesian algorithm was used to establish the temperature-field analysis approximation model.Then the optimization design of the heat-dissipating structure was carried out based on it,and the comparison with the L-M backpropagation was made.The results show that the temperature-field analysis approximation models based on GA-BP Bayesian algorithm have higher prediction accuracy for chip center temperature,and that the prediction accuracy they have can hardly be effected by random factors.The GA-BP Bayesian algorithm overcomes the shortcomings of the current algorithms,such as the high sensitiveness to initial weights and the unsatisfactory modeling accuracy,and it is valuable in engineering optimization design.

Key words: optimization design, thermal design, Neural Network(NN)