A multi-agent reinforcement learning model for maintenance optimization of interdependent highway pavement networks

Abstract

Pavement segments are functionally interdependent under traffic equilibrium, leading to interdependent maintenance and rehabilitation (M&R) decisions for different segments, but it has not received significant attention in the pavement management community yet. This study developed a maintenance optimization model for interdependent pavement networks based on the simultaneous network optimization (SNO) framework and a multi-agent reinforcement learning algorithm. The established model was demonstrated on a highway pavement network in the real-world, compared to a previously built two-stage bottom-up (TSBU) model. The results showed that, compared to TSBU, SNO produced a 3.0% reduction in total costs and an average pavement performance improvement of up to 17.5%. It prefers concentrated M&R schedules and tends to take more frequent preventive maintenance to reduce costly rehabilitation. The results of this research are anticipated to provide practitioners with quantitative estimates of the possible impact of ignoring segment interdependencies in M&R planning.