Deep reinforcement learning method in centralized multi-agent air traffic control

Summary

The objective of this thesis is to design multi-agent Deep Reinforcement Learning methods and explore their effectiveness in optimizing and automating the Air Traffic Control task. Representing each flight as an agent, we aim to maintain a minimum separation among the flights by providing resolution actions, such as lateral manoeuvres, speed changes and flight level changes. In this way, we can contribute to the highly complex work of the human Air Traffic Controllers, by resolving potential conflicts between pairs of flights. The problem is formulated as a Decentralized Partially Observable Markov Decision Process, which enables the exploitation of the graph-attention-based model, called DGN, after we have extended and enhanced it appropriately with the use of graph edges. Τwo different versions are presented, investigating both static and dynamic edges. The experiments provided suggest that the latter version yields the notable results of resolving 90% of the testing real- world scenarios relating to flights operating in Spanish airspace.

3rd Yound Scientists Awards SESAR Innovation Days 2022