Impacts of Increased Prediction Accuracy in Management Decisions: A Study of Full-Depth Reclamation Pavements

Abstract

Given the abundance of condition data regularly collected for major roadways, machine learning has the potential to enhance pavement deterioration modeling. This is particularly important for recycling-based rehabilitation techniques, such as full-depth reclamation (FDR), which lack accurate models of deterioration. Previous studies have demonstrated the effectiveness of machine learning (ML) to predict pavement deterioration. However, the increased accuracy of these models often is reported using statistical metrics that pavement managers cannot easily relate to asset management decision-making. This paper quantifies the impacts that increased accuracies in deterioration modeling have on relevant metrics used in the management of pavement assets. The study analyzed the performance of full-depth-reclamation pavements and developed random forest models to estimate roughness, rutting, and fatigue cracking. These random forest models were compared with mechanistic-empirical (M-E) models tuned to the same sites to quantify differences in prediction accuracy, useful life, life-cycle costs, and long-term performance. The tuned random forest deterioration models reduced errors by 90%–97% compared with the tuned M-E models. The results suggest that M-E predicts that FDR reaches the end of service life 8 years sooner than do the random forest predictions. The long-term performance of FDR was found to be 28%–73% higher in a 10-year design life than M-E models predict. This indicates that FDR is significantly more cost-effective than is presumed by M-E predictions, and that improvements in the accuracy of FDR predictions may result in more-informed decision-making.