General real-time three-dimensional multi-aircraft conflict resolution method using multi-agent reinforcement learning

Reinforcement learning (RL) techniques have been studied for solving the conflict resolution (CR) problem in air traffic management, leveraging their potential for computation and ability to handle uncertainty. However, challenges remain that impede the application of RL methods to CR in practice, including three-dimensional manoeuvres, generalisation, trajectory recovery, and success rate. This paper proposes a general multi-agent reinforcement learning approach for real-time three-dimensional multi-aircraft conflict resolution, in which agents share a neural network and are deployed on each aircraft to form a distributed decision-making system. To address the challenges, several technologies are introduced, including a partial observation model based on imminent threats for generalisation, a safety separation relaxation model for multiple flight levels for three-dimensional manoeuvres, an adaptive manoeuvre strategy for trajectory recovery, and a conflict buffer model for success rate. The Rainbow Deep Q-learning Network (DQN) is used to enhance the efficiency of the RL process. A simulation environment that considers flight uncertainty (resulting from mechanical and navigation errors and wind) is constructed to train and evaluate the proposed approach. The experimental results demonstrate that the proposed method can resolve conflicts in scenarios with much higher traffic density than in today's real-world situations.