A Multi-Objective Evolutionary Computation Approach for Improving Neural Network-Based Surrogate Models in Structural Engineering

Abstract

Surrogate models are widely used in science and engineering to approximate other methods that are usually computationally expensive. Here, artificial neural networks (ANNs) are employed as surrogate regression models to approximate the finite element method in the problem of structural analysis of steel frames. The focus is on a multi-objective neural architecture search (NAS) that minimizes the training time and maximizes the surrogate accuracy. To this end, several configurations of the non-dominated sorting genetic algorithm (NSGA-II) are tested versus random search. The robustness of the methodology is demonstrated by the statistical significance of the hypervolume indicator. Non-dominated solutions (consisting of the set of best designs in terms of accuracy for each training time or in terms of training time for each accuracy) reveal the importance of multi-objective hyperparameter tuning in the performance of ANNs as regression surrogates. Non-evident optimal values were attained for the number of hidden layers, the number of nodes per layer, the batch size, and alpha parameter of the Leaky ReLU transfer function. These results are useful for comparing with state-of-the-art ANN regression surrogates recently attained in the recent structural engineering literature. This approach facilitates the selection of models that achieve the optimal balance between training speed and predictive accuracy, according to the specific requirements of the application.