Game-theoretic multi-agent RL algorithm boosts UAV swarm 5G efficiency

The rapid expansion of the Low-Altitude Economy (LAE) is driving large-scale deployments of cellular-connected UAV swarms over multi-cell 5G NR ground base station infrastructure. However, equal resource block (RB) allocation policies create strong strategic coupling: when a UAV enters a cell, it reduces the RB share for all co-served drones, altering their achievable rates and trajectory incentives. Existing work either ignores this coupling or focuses on single-cell setups, leaving the multi-cell congestion-aware trajectory planning problem unsolved. The authors formulate the challenge as a cooperative stochastic congestion game with a communication-and-mission-aware utility function.

The proposed solution is a centralized-training decentralized-execution multi-agent proximal policy optimization (CTDE-MAPPO) algorithm, which maximizes social welfare under multi-cell RB congestion. Simulation results demonstrate that CTDE-MAPPO significantly outperforms established baselines including QMIX, independent Q-learning, and random policies, both in terms of aggregate utility and mission success rate. The algorithm achieves stable convergence within practical training budgets, making it viable for real-world deployment. This work, submitted to IEEE/CIC ICCC 2026, provides a scalable framework for optimizing swarm trajectories in densely populated low-altitude wireless networks.