Explainable Reinforcement Learning for Assisting Air Traffic Controllers

This paper explores explainability techniques for reinforcement learning in the safety-critical domain of air traffic control. The authors train an agent in a simplified air traffic control environment to select alternative flight routes that avoid no-fly zones. As a preliminary explanation method, they use a saliency map to identify input features that most strongly influence the agent’s decisions. The work frames explainability as a prerequisite for trust and human-AI collaboration, but the abstract describes an initial testbed rather than operational validation.