Sustaining Connectivity From the Sky: A Feasibility-Driven MARL Strategy for Communication-Computing Coordination in Multi-UAV Emergency Networks

Abstract

When terrestrial base stations are damaged after disasters, multiple cooperative uncrewed aerial vehicles (UAVs) can be rapidly deployed to provide temporary communication access and edge computing services. However, post-disaster scenarios involve heterogeneous tasks, dynamic user distribution, and time-varying air-to-ground channels, which strongly couple user association, power allocation, bandwidth allocation, CPU frequency control, and UAV mobility. Meanwhile, minimum-rate, end-to-end delay, and long-term energy-budget constraints must also be satisfied, making energy-efficient scheduling a high-dimensional and non-convex problem. To address this challenge, this paper develops a cooperative multi-UAV emergency communication and edge computing framework for heterogeneous task scenarios, where communication, computation, mobility, and energy expenditure are jointly modeled under coupled quality-of-service (QoS) and energy constraints. Based on this framework, the slot-level joint scheduling process is formulated as a decentralized partially observable multi-agent decision process under the CTDE paradigm with system energy efficiency as the optimization objective, aiming to maximize the effective service capability supported per unit energy under strict UAV and user energy limitations. Building on the above formulation, a Feasibility-Driven and Dual-Constraint-Aware MAPPO (FD-MAPPO) algorithm is proposed. Specifically, a feasibility mapping mechanism is introduced to guarantee per-slot hard-constraint satisfaction, while QoS constraints and the long-term energy budget are incorporated into a violation-aware reward with adaptive dual-weight updates, so that energy-efficiency improvement and constraint satisfaction can be jointly promoted during policy learning. Simulation results show that FD-MAPPO achieves higher energy efficiency, lower energy consumption, and fewer constraint violations than benchmark methods, demonstrating clear advantages in both energy-efficient scheduling and constraint satisfaction.