P2. Machine learning methods for planning conflict-free trajectories

2 – Data-driven trajectory prediction

The objective of this Engage KTN PhD study is to explore and present state of the art AI/ML algorithms towards planning conflict-free trajectories in computationally efficient ways, for a large number of trajectories in airspaces comprising multiple FIRs, following a methodology combining data-driven and agent-based approaches.

In the context of this study the conflicts-free trajectory planning task is defined to incorporate trajectory prediction and conflicts detection and resolution. While trajectory prediction concerns predicting the spatiotemporal evolution of the aircraft state along a trajectory (also called, trajectory evolution), conflicts detection and resolution concerns the detection of conflicts that breach separation minima (loss of separation) between flights and their resolution by appropriate actions. Therefore, the objective of the conflicts-free trajectory planning task is to predict the evolution of trajectories, and regulating flights to avoid loss of separation.

While trajectory planning may take place at the pre-tactical phase of operations, we expect the methods developed in this study to have a large impact in the tactical phase of operations.

Aiming to model stakeholders’ decisions to planning conflict-free trajectories, the major emphasis of this study is to imitate flights’ trajectories and air traffic controller’s behavior according to demonstrations provided by historical data.

The challenges that this study addressed are as follows:
1. Plan trajectories, considering complex ATM phenomena and operational constraints regarding traffic and conflicts among trajectories.
2. Follow a data-driven approach to learn stakeholders’ preferences on the evolution of trajectories and on resolving conflicts: stakeholders include airspace users (for trajectory prediction) and air traffic controllers (for conflicts’ detection and resolution actions).
3. Address optimization in trajectory planning w.r.t. multiple objectives, preferences and constraints of stakeholders involved, as these are demonstrated by historical data.
4. Address scalability: demonstrate the efficiency of the methods to be applied in settings with a large number of flights.

Contributions that this study makes are as follows:
1. The problem of modelling air traffic controllers’ behavior has been split into two well-defined problems: modelling air traffic controllers’ reactions on whether and when conflicts’ resolution actions should be applied, and modelling air traffic controllers’ reactions on how conflicts should be resolved, i.e. what resolution actions should be applied.
2. The problem of trajectory planning (either with or without considering conflicts) has been formulated as an imitation learning problem, based on historical flown trajectories.
3. AI/ML methods have been developed and tested on learning models regarding the evolution of 4D trajectories, using data-driven approaches, i.e. based on historical real-world data.
4. AI/ML methods have been developed and tested on learning models regarding air traffic controllers’ reactions and policy using data-driven approaches, i.e. based on historical real-world data.
5. This study has proposed an elaborated evaluation method for data-driven imitation learning techniques predicting air traffic controllers’ reactions, considering the uncertainties involved in the evolution of trajectories, in the assessment of conflicts, and in the reactions of ATCO.
6. Challenging issues due to inherent data limitations have been addressed and thoroughly discussed.
7. The study provides an integrated trajectory planning approach, where data-driven trajectory predictions are intertwined with data-driven conflicts detection and resolution.