Reinforcement Learning Based--Driven Parameter Adaptation for UAV Systems

This paper presents a reinforcement learning-driven parameter adaptation framework designed to autonomously tune the payoff function weights in a multi-agent drone control system. In this framework, each drone operates within a rule-based airspace management system, where Game Theory-based decision-making is employed to avoid collisions and regulate traffic. However, the manual tuning of the algorithm's 25 distinct coefficients ($\gamma$ and $\theta$) creates a high-dimensional optimization challenge incline to sub-optimal local minima. To address this, we propose a Policy Gradient-based approach to dynamically optimize these parameters across three distinct operational scenarios: (i) Early Maneuver (Safety-First, Early Avoidance), (ii) Late/Aggressive Maneuver, and (iii) Minimal Maneuver Effort. Simulation results validate the effectiveness of the proposed framework, demonstrating that successful outcomes were achieved across all three scenarios.