Study of Neural Network Models for the Ultimate Capacities of Cellular Steel Beams

Abstract

Artificial neural network (ANN) models were applied for simulating and predicting the ultimate capacities of cellular steel beams. To do this, at the first, different neural networks by various learning algorithms and number of neurons in the hidden layer were simulated. The required data for networks in training, validating, and testing state were obtained from a reliable database. Next, the best network according to its predictive performance was chosen, and a new formula was derived to predict the failure loads of cellular steel beams subjected to LTB. The attempt was done to evaluate the most exact practical formula using different algorithm and method for LTB strength assessment of cellular beams. Next, a comparison was made between the ANN-based formula and a formula based on the stepwise regression (SR) to show the predictive power of the ANN model. The results provided some evidence that ANN model obtained more accurate predictions than SR model. At the end, a sensitivity analysis was developed using Garson’s algorithm to determine the importance of each input parameter which was used in the proposed ANN formulation.