Deep Reinforcement Learning Enabled Joint Task Offloading and Trajectory Optimization in Multi-UAV Systems

UAV-assisted multi-access edge computing (MEC) enables flexible task offloading in emergency communication networks, but joint trajectory planning and task offloading optimization remain challenging due to UAV mobility and dynamic network conditions. This paper proposes a multi-agent deep deterministic policy gradient (MADDPG)-based framework to jointly optimize UAV trajectory planning and task offloading decisions. The proposed approach follows a centralized training with decentralized execution (CTDE) paradigm, enabling UAVs to learn coordinated computation and communication strategies in dynamic environments. Each UAV is modeled as an autonomous agent, and an actor–critic architecture is designed, where distributed actors generate offloading and movement policies, while a centralized critic evaluates global state–action values to guide cooperative learning. To improve training stability and convergence efficiency, experience replay and soft update mechanisms are incorporated. Extensive simulation results demonstrate that the proposed MADDPG framework significantly reduces task offloading latency and computational overhead while improving system throughput, outperforming conventional optimization methods, and validating its effectiveness in emergency communication networks.