Using a Hybrid Neural Network to Predict the Torsional Strength of Reinforced Concrete Beams

Abstract

This study proposes a multilayer perceptrons neural network with genetic algorithm (GA-MLP) for predicting the torsional strength of reinforced concrete beams. Genetic algorithm is used to determine the optimum number of inputs and hidden nodes of a feedforward neural network, the optimum slope of the activation function, and the optimum values of the learning rates and momentum coefficients. A database of the torsional failure of reinforced concrete (including normal-strength and high-strength concrete) beams with a rectangular section subjected to pure torsion was obtained from existing literature for analysis. We compare the predictions of the GA-MLP model with the ACI 318 code used for analyzing the torsional strength of reinforced concrete beams. We found that the proposed model provides reasonable predictions of the ultimate torsional strength of reinforced concrete beams and offers superior torsion accuracy compared to that of the ACI 318-02 equation considering both the correlation coefficient and absolute relative error.