Learning Uncertainty Parameters for Tactical Conflict Resolution

Assisting air traffic controllers in their deconfliction task is challenging. A five nautical mile separation standard in the horizontal plane and one thousand feet vertically are required in the upper airspace between aircraft. However, air traffic controllers generally need to take extra margins in their mental process. These margins can impact efficiency and capacity but are essential to safely manage the evolving traffic situations. It is necessary to model uncertainties on controllers trajectories predictions in order to design assistance tools that can mimic their perception of conflict risk. This article models uncertainties on the speed prediction, pilots reaction times when a maneuver is started or ended, and heading change accuracy. A method is proposed to estimate these values on deconflicted trajectories benchmarks. First we apply our method to benchmarks that where artificially created with an automatic solver calibrated with specific known uncertainty parameters. We show that the uncertainty on speed prediction, maneuver start time and heading change can be retrieved afterwards with a good accuracy. Then we apply our method to benchmarks of conflicts solved by qualified air traffic controllers. The method works but the quality of the results is questionable because of the small data size and the big variability in the air traffic controllers decisions.