LLM Empowered Multi-UAV Coordinated Trajectory Planning for Energy-Efficient Post-Disaster Data Collection

Le résumé fourni par la source

Deploying Unmanned Aerial Vehicles (UAVs) for data collection in post-disaster scenarios is a highly promising solution for rapid communication recovery. In this paper, we formulate the trajectory planning for multiple UAVs collecting data from Data Points (DPs) as a combinatorial optimization problem, aiming to minimize the Age of Information (AoI), maximize the timeliness metric and reduce the energy consumption of UAVs to achieve energy-efficient data collection. Existing Deep Reinforcement Learning (DRL) based methods rely on manually designed reward functions, which struggle to effectively guide the DRL update due to complex post-disaster environments and multiple optimization objectives, thereby limiting the optimization performance. Notably, Large Language Models (LLMs), with their vast prior knowledge and advanced reasoning advantages, have demonstrated performance surpassing human levels in multiple domains. Therefore, we propose an LLMempowered Multi-UAV Trajectory Planning algorithm based on a modified Pointer network (LLMPtr-MTP), which integrates the reasoning advantages of LLMs to enhance DRL decisionmaking capabilities, achieving energy-efficient data collection. In LLMPtr-MTP, we develop an LLM reward shaping algorithm to generate a high-quality reward function that guides the LLMPtr-MTP agent in handling complex environments and improving the optimization performance of multiple objectives. Moreover, we design a modified Pointer network of the LLMPtr-MTP agent to achieve coordinated decision-making for multiple UAVs, significantly reducing the action space for the agent to solve the combinatorial problem and improving coordination efficiency of UAVs. Experimental results demonstrate that LLMPtr-MTP outperforms comparison algorithms in both AoI and timeliness metrics, with energy efficiency improved by 107%.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.