A Curriculum-Learning-Assisted MAPPO-Based Algorithm for Dynamic Spectrum Access and Anti-Jamming in UAV Swarms

Area of Science:

• Robotics and Communication Systems
• Artificial Intelligence and Machine Learning

Background:

• Drone swarms require high-concurrency, reliable communication, which is challenging in complex environments.
• Traditional Medium Access Control (MAC) protocols and deep reinforcement learning (DRL) face limitations in collision handling, convergence, and anti-jamming robustness.

Purpose of the Study:

• To propose a novel dynamic access algorithm for enhancing drone swarm communication reliability.
• To address the limitations of existing methods in high-density collision scenarios and non-stationary interference.

Main Methods:

• A Centralized Training with Decentralized Execution (CTDE) architecture is employed for implicit spectrum cooperation.
• A three-stage curriculum learning mechanism (collision avoidance, load balancing, anti-jamming) with phased reward reshaping guides agent learning.
• The proposed algorithm is Curriculum Learning-assisted Multi-Agent Proximal Policy Optimization (CL-MAPPO).

Main Results:

• CL-MAPPO significantly outperforms baseline models (CSMA, random frequency hopping, MADDPG) in simulated dynamic jamming and high-load scenarios.
• Demonstrated improvements in normalized throughput and reduced channel collision rates.
• Achieved faster convergence speeds compared to conventional methods.

Conclusions:

• The CL-MAPPO algorithm provides a robust solution for reliable communication in large-scale drone swarms under harsh conditions.
• Offers theoretical support and an algorithmic foundation for advanced swarm data links.
• Highlights the effectiveness of curriculum learning in complex multi-agent reinforcement learning tasks.