Recent technological advances have brought about a revolution in the world of air defense. By air defense, we mean all the measures and systems designed to detect, identify, track, and neutralize aerial threats. With the emergence of new threats, such as small-sized or medium-sized drones or hypersonic or highly maneuverable threats, the battlefield is becoming increasingly complex for operators to analyze.
Current methods, combining decisions made by human operators and rule-based algorithms or heuristics, suffer from a lack of responsiveness and robustness in the face of these new threats. In this thesis, we explore the potential of deep reinforcement learning to develop real-time adaptive defensive strategies. The contributions of this thesis are structured around three areas. First, we have implemented deep reinforcement learning algorithms that outperform rule-based methods by adapting to the situations they encounter, whereas deterministic strategies remain limited when faced with the complexity and variability of the environment.
Next, we demonstrate that these agents are capable of robustly transferring knowledge learned from a source scenario to a target scenario.
Finally, by leveraging the self-attention mechanism, we address the explainability of decisions obtained by Deep Reinforcement Learning, an essential step to foster the acceptability of this type of approach and their adoption in an operational context.