Agentic AI-Driven UAV Network Deployment: A LLM-Enhanced Exact Potential Game Approach

A research team has developed a novel AI framework that treats drone (UAV) network deployment as a multi-agent coordination problem, solved through game theory enhanced by Large Language Models. The system, detailed in a new arXiv paper, tackles the complex challenge of optimizing both discrete connection choices and continuous parameters like drone position and transmission power. It uses a two-part approach: a 'log-linear learning' algorithm manages the large-scale network topology to create efficient, interference-reducing links, while an 'approximate gradient' algorithm handles fine-grained deployment and resource allocation to improve user throughput and latency.

The key innovation is the integration of an LLM as a 'knowledge-driven decision enhancer.' Instead of relying on manual, scenario-specific tuning—a major bottleneck in dynamic environments—the LLM automatically generates optimal utility weights for the system based on the current network's characteristics. This allows the drone swarm to adapt its coordination strategy on the fly. Simulation results show this LLM-enhanced, game-theoretic approach consistently outperforms existing baseline methods, achieving better energy efficiency, lower end-to-end latency, and higher overall system throughput, marking a significant step toward fully autonomous, adaptive aerial networks.