Problem

Unmanned Aerial Vehicular Networks (UAVNs) are envisioned to provide flexible connectivity and low-latency services in dynamic environments. However, optimizing UAVN topology presents a significant mixed-integer nonconvex problem due to the complex coupling of discrete link decisions and continuous deployment parameters. This complexity leads to challenges in scalability, efficiency, and solution consistency under dynamic network conditions. Traditional centralized optimization methods suffer from high computational and communication overhead, while heuristic and learning-based approaches often lack global convergence guarantees and sufficient efficiency for dynamic scenarios. Game theory methods, while distributed, frequently struggle with optimizing global utility and adapting to heterogeneous environments without extensive manual parameter tuning.

Method

This paper proposes a novel dual spatial-scale UAVN topology optimization framework, enhanced by Agentic AI and Large Language Models (LLMs), built upon Exact Potential Games (EPGs). At the large spatial scale, a log-linear learning based EPG (L3-EPG) algorithm is introduced to optimize inter-UAV link configurations, promoting sparse yet connected network topologies by reducing redundant links and interference. Concurrently, at the small spatial scale, an approximate gradient based EPG (AG-EPG) algorithm is employed to jointly optimize UAV deployment coordinates, transmission power allocation, and ground user (GU) association, aiming to improve network throughput and latency. To further boost adaptability across diverse scenarios, an LLM is integrated with a multi-source knowledge base and a Retrieval Enhanced Generation (RAG) framework. This LLM functions as a knowledge-driven decision enhancer, automatically generating appropriate utility weights based on specific network characteristics, thereby alleviating the reliance on manual parameter tuning.

Results

Extensive simulation results demonstrate the significant advantages of the proposed framework over existing baseline methods in terms of energy consumption, end-to-end latency, and system throughput. The L3-EPG algorithm effectively prunes redundant links, leading to an optimized, sparse network topology while maintaining connectivity. The AG-EPG algorithm dynamically adjusts UAV 3D coordinates, with some UAVs increasing altitude to clear obstacles for critical links and others slightly descending to reduce communication distances, optimizing Line-of-Sight conditions and enhancing link quality. Compared to four benchmark algorithms (BRD-EPG, BRD-NCG, ETG, GA), the proposed approach consistently achieved higher network throughput, showing an improvement of approximately 8.4%. Furthermore, it maintained lower total energy consumption and latency as the number of UAVs increased, attributed to the efficient link optimization and adaptive parameter adjustments. The study also validated the stable convergence of both EPG-based algorithms and the effectiveness of the LLM for knowledge-based utility weight generation.

Implications

This research provides a highly adaptable and efficient framework for optimizing the complex deployment of Unmanned Aerial Vehicular Networks. By synergistically combining Agentic AI and LLMs with game-theoretic principles, the proposed solution tackles the inherent challenges of mixed-integer nonconvex optimization in UAVNs, leading to significant performance enhancements in energy efficiency, latency, and throughput. The innovative LLM-enhanced utility weight generation mechanism is particularly impactful, as it enables the system to intelligently adapt to diverse and dynamic network environments without extensive human intervention. This advancement is crucial for enabling more autonomous, reliable, and scalable UAV communication systems in critical applications such as urban emergency response.