PER-MADDPG: Multi-Agent Reinforcement Learning for Coordinated UAV Swarm Performance under Rhythmic Constraints

With the integration of multi-agent systems and digital art, unmanned aerial vehicle (UAV) swarm performances have emerged as a novel form of intelligent artistic expression, requiring precise spatial coordination, rhythm synchronization, and dynamic light control. Traditional methods often struggle to meet the adaptability and real-time demands of large-scale collaboration. This paper proposes PER-MADDPG, a prioritized experience replay-enhanced multi-agent deep deterministic policy gradient algorithm, tailored for artistic UAV swarm control. By dynamically prioritizing high-value experiences, the method improves sample efficiency, accelerates convergence, and enhances behavioral stability. A custom 2D simulation platform is developed to validate the algorithm in multi-phase formation transitions and lighting tasks. Experimental results show that PER-MADDPG outperforms baseline methods in average reward, peak performance, and training stability, demonstrating its effectiveness for autonomous coordination in large-scale artistic UAV applications.