Deep reinforcement learning for long-term pavement maintenance planning

Inappropriate maintenance and rehabilitation strategies cause many problems such as maintenance budget waste, ineffective pavement distress treatments, and so forth. A method based on a machine learning algorithm called deep reinforcement learning (DRL) was developed in this presented research in order to learn better maintenance strategies that maximize the long-term cost-effectiveness in maintenance decision-making through trial and error. In this method, each single-lane pavement segment can have different treatments, and the long-term maintenance cost-effectiveness of the entire section is treated as the optimization goal. In the DRL algorithm, states are embodied by 42 parameters involving the pavement structures and materials, traffic loads, maintenance records, pavement conditions, and so forth. Specific treatments as well as do-nothing are the actions. The reward is defined as the increased or decreased cost-effectiveness after taking corresponding actions. Two expressways, the Ningchang and Zhenli expressways, were selected for a case study. The results show that the DRL model is capable of learning a better strategy to improve the long-term maintenance cost-effectiveness. By implementing the optimized maintenance strategies produced by the developed model, the pavement conditions can be controlled in an acceptable range.