Wild horse optimisation - enhanced neural networks for predicting international roughness index in pavement management

DOI registering

Abstract

This study developed a predictive framework for estimating the pavement International Roughness Index (IRI) using a multilayer perceptron (MLP) optimised using Wild Horse Optimisation (WHO). The model was built using field data from a heavily trafficked freeway in Iran and considers traffic characteristics, pavement structure, environmental conditions, and geometric design factors. The MLP uses a feedforward structure with one hidden layer of 22 neurons, applying tansig and purelin activation functions in the hidden and output layers, whereas WHO optimises the network weights and biases. The optimised MLP-WHO model exhibited strong prediction accuracy, with R = 0,967 and RMSE = 0,271 in training and R = 0,905 with RMSE = 0,338 in testing. A similar performance was achieved for all data (R = 0,96; RMSE = 0,29), with a computation time of 502 s. A comparison showed that MLP-WHO performs better than the basic MLP, MLP-GA, MLP-PSO, and other common machine learning models in terms of accuracy and generalization. Although WHO needs more computation time, its better prediction accuracy justifies its use. Feature importance shows geometric and design variables affect IRI the most, followed by traffic and pavement structure, while environmental factors matter less. These findings confirm roadway geometry is key for pavement performance and maintenance.