How to Build a Reliable Framework to Make Intelligent Decisions About Road Maintenance

Reinforcement learning (RL) has been widely applied in decision-making and decision-optimization tasks. This work proposes a feasible framework which adapts the RL to achieve scientific decision-making for highway maintenance by formulating an explicable reward function which uses optimal maintenance strategy selection as essential evidence. To this end, the proposed method (1) uses historical data on pavement distress to predict the rating of highway performance (state sequence) in the next 1–5 years; (2) mixes the highway maintenance records and the necessary road information to construct management preferences and maintenance (action) sequence; (3) uses the traffic volume records and the output data from previous steps to construct the reward function. From experiments, after training the agent in this framework with an in-home dataset collected from engineering, it shows an efficient convergence and makes similar decisions as human experts. Even for the discrepant cases, our agent still gives more scientific plans than the human experts considering its effects on maintaining relatively high performance of highway.