Multi-Agent Reinforcement Learning for Joint Task Offloading and Migration in UAV Swarm Wireless Networks

Full abstract

Read the full abstract

The evolution of smart-city services demands sustainable and resilient network and service management in emergencies. In post-disaster scenarios where terrestrial infrastructure is damaged, mobile users face strict latency requirements under limited energy budgets and computing resources, while UAV-assisted edge networks experience dynamic topology and time-varying capacity. This paper studies the joint decision problem of task offloading, queue-aware scheduling, and inter-UAV task migration in UAV swarms under fixed communication and computation resource parameters, using a simplified distance-based communication model and UAV operational constraints. We propose an emergency edge computing framework and develop MAPPO-QM, a multi-agent reinforcement learning method based on Multi-Agent Proximal Policy Optimization that explicitly incorporates queueing dynamics and a rule-based migration mechanism triggered at migration points. MAPPO-QM optimizes a combined objective of end-to-end latency and energy consumption through an exponential reward with a fixed scaling coefficient, complemented by completion and cooperation reward terms. Simulation results show that MAPPO-QM achieves the highest task completion rate (99.13%) among the evaluated methods (PPO-QM, MAPPO, PPO, and local execution) and reduces the success-conditioned average latency by 58.2% compared with PPO-QM. MAPPO without the waiting queue and migration attains lower latency and energy per completed task but completes about 17% fewer tasks, indicating that queue-aware task waiting and migration can improve service availability in the evaluated post-disaster edge scenarios.