Data-driven prediction of Air Traffic Controllers reactions to resolving conflicts

Alevizos Bastas, George A. Vouros

Introduction

As global flight volumes continue to recover and expand, Air Traffic Management (ATM) systems are approaching their operational limits. Central to this challenge is Conflict Detection and Resolution (CD&R), the critical task of maintaining safe separation between aircraft. While automated CD&R systems have been proposed for decades, a significant "trust gap" remains between human Air Traffic Controllers (ATCOs) and automated suggestions. Many existing systems rely on purely mathematical optimizations that do not align with human intuition or operational heuristics. This research addresses this gap by proposing a deep learning framework designed to imitate human decision-making, predicting both the timing and the nature of ATCO interventions using historical flight data.

Methodology: Imitation Learning via VAEs

The researchers frame the CD&R problem as an Imitation Learning task. Rather than programming a system with rigid rules, the goal is to learn the underlying logic of expert controllers from real-world trajectories. The study employs a hierarchical supervised learning approach centered on a Variational Auto-Encoder (VAE).

Hierarchical Modeling

The system operates on two distinct levels to mirror the human decision process:

  1. High-Level Mode Prediction: The model first determines the behavioral "mode." These modes include "no conflict," "conflict requiring immediate action," or "conflict being tolerated" (where a controller observes a potential issue but chooses not to intervene yet).
  2. Low-Level Action Prescription: If an action is deemed necessary, the model predicts specific resolution maneuvers, such as speed adjustments or vectoring to a new waypoint.

Simulating Human Uncertainty

ATCOs do not make decisions based on static points; they account for the inherent uncertainty of future flight paths. To replicate this, the authors developed a method to simulate potential trajectory evolutions. By utilizing historical deviations in course and speed, the model generates a set of potential "future paths" for each aircraft, allowing the neural network to assess the same probabilistic conflicts that a human controller evaluates.

The VAE Advantage

Unlike standard deterministic neural networks, the VAE uses an encoder to map inputs into a latent space representing behavioral modes and a decoder to prescribe actions. This structure is particularly effective at capturing the "rationale" behind a decision, leading to more stable and consistent predictions compared to baseline models that often suffer from high-frequency fluctuations in their output.

Data and Feature Engineering

The study utilized high-fidelity operational data from 2017 provided by the Spanish ATC Platform (SACTA). The dataset included:

  • Surveillance Data (IFS): Radar track points synchronized at 5-second intervals.
  • ATCO Event Data: Specific records of resolution actions (ATON/NORVASE) taken by controllers.
  • Scope: The research focused on five major European origin-destination pairs (e.g., Malaga to London Gatwick) during the en-route phase of flight.

To ensure the model could generalize across different airspaces, the researchers used an abstraction of states. Instead of raw GPS coordinates, the model processes relative features such as the distance to the Closest Point of Approach (CPA), relative bearings between aircraft, and relative speed magnitudes.

Key Insights

Human Tolerance as a Feature

A critical component of the model is the inclusion of a specific mode ($C_2$) for cases where a conflict is detected but the ATCO chooses not to act immediately. This captures the "tactical delay" or safety margins inherent in human management—nuances that purely mathematical models often overlook as "inefficiencies."

Stability and Confidence

The VAE demonstrated significantly higher confidence than standard encoders. While basic models might flicker between suggesting an action and staying silent, the VAE’s latent space representation resulted in stable decision windows that closely mirrored the actual decision-making periods of human controllers.

Results and Performance

The VAE model showed high efficacy in predicting the timing of controller interventions:

  • Mode Prediction Accuracy: In a flight-centric (sector-ignorant) configuration, the VAE achieved a weighted F1-score of 0.985 for predicting when a controller would issue a resolution.
  • Comparative Performance: The VAE consistently outperformed baseline encoder-only models, particularly in its ability to maintain a consistent resolution state over time.
  • Flight-Centric vs. Sector-Based: Interestingly, the models performed better when focusing on the local neighborhood of the flight rather than the specific sector boundaries. This suggests that a flight-centric approach captures conflicting aircraft entering from adjacent airspace that sector-based models might ignore.

Limitations and Future Outlook

While the model excels at predicting when a controller will react, predicting the exact how (the specific low-level resolution action) remains a challenge. F1-scores for specific actions like "speed change" ranged between 0.61 and 0.68. These decisions are often influenced by complex pilot preferences, airline requirements, and broader network constraints not yet fully captured in the data.

Future research will focus on refining the decoder policy to improve action accuracy and expanding the framework to handle multi-aircraft coordination. Ultimately, this data-driven approach paves the way for human-AI collaborative tools that can reduce ATCO workload while maintaining the high levels of trust required for safe skies.

Summary updated

BibTeX
@article{bastas2022data,
  author = {Alevizos Bastas and George A. Vouros},
  title = {{Data-driven prediction of Air Traffic Controllers reactions to resolving conflicts}},
  journal = {Information Sciences},
  year = {2022},
  doi = {10.1016/j.ins.2022.09.015},
}
Figure