Neural Network Based Response Surface Methods - a Comparative Study
W. Beyer, M. Liebscher, W. Graf (TU Dresden) This paper deals with the application of the response surface method based on neural networks in the field of engineering. A multi-layer feedforward neural network is employed to replace the underlying structural analysis. It is trained on a data set from initial computations. The application of the backpropagation algorithm for training a neural network requires a decision for a particular training mode. For the examined examples it is shown that the incremental mode possesses different advantages compared to the batch mode. Furthermore, an approach to improve the approximation quality given by a neural network is introduced and demonstrated by means of a numerical example. The authors would like to thank Dynamore GmbH, Stuttgart, Germany for supporting this study.
https://www.dynamore.de/en/downloads/papers/06-forum/papers/optimization/neural-network-based-response-surface-methods-a/view
https://www.dynamore.de/@@site-logo/DYNAmore_Logo_Ansys.svg
Neural Network Based Response Surface Methods - a Comparative Study
W. Beyer, M. Liebscher, W. Graf (TU Dresden) This paper deals with the application of the response surface method based on neural networks in the field of engineering. A multi-layer feedforward neural network is employed to replace the underlying structural analysis. It is trained on a data set from initial computations. The application of the backpropagation algorithm for training a neural network requires a decision for a particular training mode. For the examined examples it is shown that the incremental mode possesses different advantages compared to the batch mode. Furthermore, an approach to improve the approximation quality given by a neural network is introduced and demonstrated by means of a numerical example. The authors would like to thank Dynamore GmbH, Stuttgart, Germany for supporting this study.
K-II-4.pdf
— 928.7 KB