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

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Unmanned aerial vehicles (UAVs) with computational servers enhance edge computing for remote user equipment (UE).
• Existing value decomposition algorithms struggle with multi-UAV coordination, leading to lower task completion and longer convergence times.
• Effective UAV trajectory planning is crucial for efficient edge computing resource utilization.

Purpose of the Study:

• To develop an innovative multi-agent deep learning framework for optimizing multi-UAV trajectories.
• To address the limitations of current algorithms in associating local observations with global states in UAV clusters.
• To improve task completion rates and convergence times in UAV-assisted edge computing.

Main Methods:

• Conceptualized multi-UAV trajectory optimization as a decentralized partially observable Markov decision process (Dec-POMDP).
• Integrated the QTRAN algorithm with a large language model (LLM) for region decomposition.
• Employed graph convolutional networks (GCNs) and self-attention mechanisms for managing inter-subregion relationships.

Main Results:

• The proposed framework significantly outperforms existing deep reinforcement learning methods.
• Demonstrated improvements in convergence speed exceeding 10%.
• Achieved task completion rate improvements exceeding 10% compared to baseline methods.

Conclusions:

• The developed framework advances UAV trajectory optimization in edge computing environments.
• Enhanced performance of multi-agent systems in UAV-assisted edge computing.
• Offers a robust solution for complex computational task offloading using UAV swarms.