Reinforcement learning for pavement maintenance cost optimization

Purpose Pavement maintenance is crucial for ensuring the safety, durability and sustainability of transportation infrastructure. Traditional maintenance strategies rely on predefined schedules or assumptions, limiting their responsiveness to emerging costs or unexpected events. These challenges can result in inefficient resource allocation, increased repair expenses and low-quality road surfaces. To address this, the present research presents dynamic, and data-driven approach that employs machine learning to improve road surface quality and resource allocation. Design/methodology/approach The proposed methodology implements a reinforcement learning agent for pavement maintenance, enabling dynamic planning according to fluctuating road segment conditions. The framework is evaluated using real-world data collected from a diverse Greek road network, comparing its performance against established optimization methods, such as particle swarm optimization. Findings Experimental analysis demonstrates that the proposed approach reduced the operational cost by 15.72, and 24.80% compared against Solver, and metaheuristic methods, such differential evolution. The agent persistently prioritizes low-cost maintenance actions, such as Cape Seal, avoiding unnecessary repairs and minimizing associated costs under constraints. Originality/value The study presents a novel reinforcement learning approach to pavement maintenance planning, providing a dynamic and scalable data-driven solution compared to traditional, static maintenance ones. Using real-world data, the proposed framework improves the sustainability and efficiency of transportation infrastructure management.