Abstract
Modelling pavement performance is important to pavement asset management and to roadway safety and sustainability. While neural networks (NNs) have been employed to establish pavement performance models, past studies have mainly focused on achieving high model accuracy and generally failed to examine other model properties. This study calibrated feedforward neural networks with particle swarm optimization (PSO) to predict the rutting performance of asphalt pavement in the State of Idaho. The resulting models provided rutting predictions with higher accuracy than traditional mechanistic-empirical models. This study also employed various statistical methods in the pre-processing and post-analysis of PSO-NNs, such as the principal component analysis in the input selection and hypothesis tests in the model reproducibility and robustness evaluation. This study further explored the relationships between the model accuracy and updating efficiency and the number of calibrated parameters (hidden neurons). The selected PSO-NN structure struck the right balance between model accuracy, reproducibility and robustness of the models.