Incident Duration Sequential Predictions Using Reinforcement Learning for Advanced Traveler Information System Applications

Accurate incident duration prediction is crucial for the effectiveness of advanced traveler information systems. Sequential prediction is the process by which an incident duration is predicted at its inception, and potential follow-up predictions or revisions to previous predictions are made thereafter. However, one scenario that requires sequential predictions but has rarely been discussed is when the incident remains unclear despite the estimated incident duration time that has elapsed. To address this issue, we proposed a methodology for training a reinforcement learning (RL) agent to produce sequential predictions, and apply it to Houston TranStar incident data. Results indicate that compared to a one-time prediction model, the proposed RL-based method produces more accurate follow-up predictions, and compared to other models in sequential prediction, such as the artificial neural network and random forest, the RL agent achieves lower mean absolute error by 1.5 min. In addition, we have found that in 63.6% of incidents, the agent only needs to make a single prediction, indicating that the agent primarily serves as a one-time prediction method. However, in cases where a one-time prediction incurs a significant error, the agent can generate additional predictions to address the limitations of the one-time prediction approach. The proposed methodology enables traffic operators to provide travelers with updated incident duration information, empowering both operators and users to make informed decisions.