Effect of Hyper-Parameters on Deep Learning Networks in Structural Engineering

Abstract

Since the first journal article on structural engineering applications of neural networks (NN) was published, a large number of articles about structural engineering have been published on these fields. However, over the last decade, researchers who attempt to apply the neural network concept to structural analysis problems have reduced significantly because of a fundamental limitation. At the beginning of the new millennium, in a deep learning field, newer methods have been proposed by using new activation functions, loss functions, alleviating overfitting methods with hyper-parameters, and other effective methods. Recent advances in deep learning techniques can provide a more suitable solution to the problem. The aim of our study is to show effects and differences of newer deep learning techniques on neural networks of structural analysis topics. A well-known 10-bar truss example is presented to show condition for neural networks and role of hyper-parameters in the structures.