Direct-to Initial Approach Fix: A Reinforcement Learning Approach for Conflict-Free Arrival Sequencing in a Multi-Airport System

Terminal Manoeuvring Area (TMA) is a key air traffic subsystem that bridges en-route airspace and airport control zone. One of the main tasks of Air Traffic Control Officers (ATCOs) responsible for the TMA is to ensure that consecutive landing aircraft have the required horizontal separation. To achieve this goal, ATCOs need to make real-time decisions regarding the sequencing and spacing of arrival aircraft during daily operations, which is a primary source of their workload. Relying solely on ATCOs to make these decisions has led to issues such as delayed decision-making, excessive flight distances, and frequent trajectory adjustments, particularly in the more complex environment of multi-airport systems. To support ATCOs in making real-time decisions regarding the safe sequencing of arrival flights, this paper proposes a Reinforcement Learning approach to suggest arrival direct-to routes while considering the convergence of arrival flights destined for the same airport and conflicts with arrival flights destined for adjacent airports. A method for accelerating reinforcement learning training is also explored. Experimentation on Tianjin TMA in China shows that the proposed approach achieves conflict-free operations without sacrificing operational efficiency, and reduces training time of the RL model by 82% without compromising model performance. The results of this work demonstrates the potentials of Artificial Intelligence (AI) systems as decision-support tools in the field of Air Traffic Management (ATM).