The research, developed by SUPSI researchers Giacomo Del Rio and Alessandro Antonucci and USI PhD student Ardian Selmonaj, all active at the Dalle Molle Institute for Artificial Intelligence Studies (IDSIA USI-SUPSI), in collaboration with Adrian Schneider, Scientific Project Manager at armasuisse, analyzes the realistic simulation of aerial combat.
Such scenarios are marked by extreme complexity and uncertainty. Decision-making must cope with partial or incomplete information, which complicates the assessment of both the operational context and adversaries, as well as with highly non-linear flight dynamics, where even minimal control inputs can produce unpredictable effects on aircraft behavior.
To address these challenges, the work Coordinated Strategies in Realistic Air Combat by Hierarchical Multi-Agent Reinforcement Learning introduces a new three-dimensional simulation environment and a framework based on hierarchical multi-agent reinforcement learning.
The approach combines AI agents with heterogeneous dynamics and capabilities, gradual and progressive learning modes (curriculum learning), and league-play mechanisms that allow agents to face different strategies, strenghtening their adaptabilities. A purpose-built training algorithm further enhances the robustness, scalability and reliability of the resulting decision-making processes.
Decision-making is structured across two levels: a lower layer responsible for learning precise, continuous control maneuvers, and a higher layer devoted to tactical coordination among agents and to achieving mission objectives. Experimental results show that this hierarchical architecture significantly improves both learning efficiency and overall performance in simulated air combat scenarios.
By enabling complex behaviors to be tested, analyzed and validated in a simulated environment, the research supports a more responsible development of AI technologies, reducing exposure to potentially hazardous situations and promoting practices that place strong emphasis on safety, human oversight and ethical considerations.