CIESC Journal

• 化工学报 • 上一篇    下一篇

遗传算法与神经网络(Ⅰ)──用改进的遗传算法训练神经网络

陈方泽,陈丙珍,何小荣   

  1. 清华大学化学工程系!北京100084,清华大学化学工程系!北京100084,清华大学化学工程系!北京100084
  • 出版日期:1996-06-25 发布日期:1996-06-25

GENETIC ALGORITHMS AND ARTIFICIAL NEURAL NETWORK(Ⅰ)──TRAINING ARTIFICIAL NEURAL NETWORK WITH EXTENDED GENETIC ALGORITHMS

Chen Fangze;Chen Bingzhen;He Xiaorong(Department of Chemical Engineering,Tsinghua University,Beijing 100084)   

  • Online:1996-06-25 Published:1996-06-25

摘要: 针对BP(backpropagation)神经网络训练中的局部最优问题,提出了由改进的遗传算法EGA训练BP神经网络的新方法.该方法克服了经典遗传算法的不足,通过自适应多点变异操作比较有效地克服了收敛过程中的振荡问题,因而提高了BP网络训练的速度,并且找到了合理的变异因子范围.

Abstract: In order to overcome a local optimal solution which often occurs in the training process of artificial neural networks,a new algorithm called EGA is proposed.An adaptive multipoint mutation operator,which can effectively extinguish the swing problem in convergent process,is proposed in this algorithm.So the speed of BP training is improved.A reasonable mutation factor range is found.As a result,the speed and accuracy of training artificial neural networks are much improved.

中图分类号: