Task Offloading with LLM-Enhanced Multi-Agent Reinforcement Learning in UAV-Assisted Edge Computing

Abstract

Unmanned aerial vehicles (UAVs) furnished with computational servers enable user equipment (UE) to offload complex computational tasks, thereby addressing the limitations of edge computing in remote or resource-constrained environments. The application of value decomposition algorithms for UAV trajectory planning has drawn considerable research attention. However, existing value decomposition algorithms commonly encounter obstacles in effectively associating local observations with the global state of UAV clusters, which hinders their task-solving capabilities and gives rise to reduced task completion rates and prolonged convergence times. To address these challenges, this paper introduces an innovative multi-agent deep learning framework that conceptualizes multi-UAV trajectory optimization as a decentralized partially observable Markov decision process (Dec-POMDP). This framework integrates the QTRAN algorithm with a large language model (LLM) for efficient region decomposition and employs graph convolutional networks (GCNs) combined with self-attention mechanisms to adeptly manage inter-subregion relationships. The simulation results demonstrate that the proposed method significantly outperforms existing deep reinforcement learning methods, with improvements in convergence speed and task completion rate exceeding 10%. Overall, this framework significantly advances UAV trajectory optimization and enhances the performance of multi-agent systems within UAV-assisted edge computing environments.