Large Language Model-Based QoS-Aware Resource Allocation for Multi-UAV Cooperative Edge Computing Networks

Abstract

In 6G multiple uncrewed aerial vehicles (UAVs) cooperative edge computing networks, strongly coupled system states and limited single-UAV observability lead to inefficient resource management and difficulty in guaranteeing Quality of Service (QoS). To address these issues, we propose a QoS-aware resource allocation method based on a large language model (LLM) for Multi-UAV Cooperative Edge Computing Networks. First, we construct an LLM-based teacher–student resource allocation framework, operating with a global perspective, generates high-quality expert policies that are subsequently injected into distributed student agents via policy distillation, enabling efficient online decision-making in dynamic environments. Second, we design an LLM-based teacher model for accurate expert decision-making under dynamic network conditions. Specifically, we construct a time-varying network knowledge graph (NKG) to represent the complex spatiotemporal states of multi-UAV networks, employ a relation-aware graph attention network (R-GAT) to aggregate crucial neighborhood information and capture node importance, and further combine a fine-tuned LLM with a Tree-of-Thoughts (ToT) reasoning framework to produce high-quality expert resource allocation policies. Finally, we develop a multi-agent student model with policy distillation for efficient management of dynamic, multi-dimensional resources. We formulate a QoS objective that jointly considers delay and fairness, and jointly optimize user association, UAV trajectories, computing allocation, bandwidth allocation, and air-to-air (A2A) migration ratios. The student utilizes the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm and learns from the teacher efficiently via policy distillation, adapting adeptly to dynamic environments. Simulation results demonstrate that the proposed method achieves significantly faster convergence, lower steady-state delay, and higher fairness compared to baseline approaches, while also exhibiting robustness and scalability across different network sizes and resource conditions.