Deep Reinforcement Learning Applied to Airport Surface Movement Planning

This paper describes an Artificial Intelligence (AI) Deep Reinforcement Learning (DRL) model application to airport surface movement planning. Specific areas of support include planning conflict-free paths of flights, advising flight actions to avoid conflicts, and meeting flight-specific time constraints. We describe a unique way to represent the surface operational environment and schedule constraints for leveraging the DRL model with a Convolutional Neural Network and also demonstrate the training process and performance for routing surface traffic at a hypothetical airport. The proposed model learns entirely from the simulation to make sequential decisions for finding conflict-free routes, addressing incomplete schedule information, and coordinating the simultaneous rerouting of multiple flights. Such a DRL model, once adapted and trained for a realistic airport, may advise controllers on surface route planning, conflict probing, and resolving predicted conflicts.