Investigating the effects of ensemble and weight optimization approaches on neural networks' performance to estimate the dynamic modulus of asphalt concrete

This study hybridized the ensemble and weight optimization approaches with an artificial neural network (ANN) algorithm to forecast the dynamic modulus (E*) of asphalt concrete. For input selection, this study employed the random forest technique and tested various techniques, including evolutionary, backward, forward, and brute force, to be hybridized with random forest. The evolutionary-random forest technique was used as the best input-selection method. The rho 200, V-beff, binder G* (dynamic shear modulus), and binder phi (phase angle) were selected as the most important variables for developing the ANN models. The findings of this research indicated that the artificial neural network-particle swarm optimisation (ANNPSO) model performed better than the other models. Also, the weight optimization techniques were more efficient than the ensemble techniques to improve the predictive power of ANN to forecast the dynamic modulus of asphalt concrete.