Large-scale maintenance and rehabilitation optimization for multi-lane highway asphalt pavement : a reinforcement learning approach

Abstract

Pavement maintenance and rehabilitation (M&R) optimization is of great importance to the sustainable development of roadway infrastructure. Various models have been developed for supporting M&R decision-making. However, there is still a lack of research that can provide lane-specific M&R strategies for large-scale pavement networks which consider uncertainty in optimization to achieve management flexibility. This study proposes an innovative M&R optimization approach for multi-lane highway pavement based on a reinforcement learning (RL) method. Life cycle assessment (LCA) and life cycle cost analysis (LCCA) were integrated to assess the environmental and economic impact of M&R decisions, respectively. The uncertainty of pavement deterioration was considered by constructing an RL simulation environment that contains several probabilistic pavement performance models. The proposed method was applied to a large-scale real-world highway network as demonstration, and compared with the state-of-the-practice hierarchical threshold-based approach (HT). The results show that the RL model saved about 26.59% of the cost in comparison to the HT approach, which was equal to 18147.27 million CNY. It could keep the long-term pavement performance within an acceptable range in a cost-effective manner. The RL model tends to select less rehabilitations and more preventive maintenance than the HT model. It was also found that incorporating uncertainty into optimization allows the model to balance the expected return and the negative (risk) and positive (opportunity) uncertainty of the solution. The outcomes of this study are expected to improve the current pavement management practice and demonstrate the potential of RL in pavement M&R optimization.