Application of BP Neural Network Based on Quasi-Newton Method in Aerodynamic Modeling

Abstract

In the study of BP neural network for aerodynamic modeling, few studies have considered improving the generalization ability of models. Improving generalization ability is important in the study of neural network models. Two typical methods to improve the generalization ability are tested in this paper, which are based on the LBFGS algorithm rather than the traditional gradient descent method when training BP neural network. Results indicate that when training neural network based on quasi-Newton method for aerodynamic modeling:1, adding the penalty term can improve the generalization ability 2, adding small variance noise will not improve the generalization ability.