LLM-Guided Multi-Agent Joint Velocity and Spectrum Optimization in Advanced Air Mobility

Abstract

In Advanced Air Mobility (AAM) applications, jointly optimizing multi-agent motion control along predefined flight routes and spectrum access is highly challenging due to the tight coupling among mobility, interference, and safety constraints under limited spectrum resources. This paper proposes a Large Language Model (LLM)-guided cooperative decision-making framework for joint velocity control and bidirectional channel selection in an AAM system with Aerial Vehicles (AVs) communicating with ground Base Stations (BSs) while following predefined linear routes. We formulate the problem as a cooperative Markov game with a discrete action space that includes both velocity selection and uplink and downlink channel access, while satisfying Signal-to-Interference-plus-Noise Ratio (SINR) quality requirements and collision avoidance constraints. To obtain reliable expert behavior, we first learn a near-optimal policy using Multi-Agent Reinforcement Learning (MARL) with Value Decomposition Dueling Double Deep Q-Networks (VD3QN). We then treat joint decision-making as a sequence generation task and employ Large Language Models (LLMs) to generate complete joint action sequences from structured environment descriptions, under both Prompt Engineering (PE) and Parameter-Efficient Fine-Tuning (PEFT) via Low-Rank Adaptation (LoRA) on expert demonstrations. Extensive simulations show that structured prompting improves decision quality, while LoRA fine-tuning further increases reward, reduces variance, and yields decisions that closely match the expert policy. Beyond this imitation role, the LLM layer turns 6-AV expert demonstrations into a sequence-level decision generator. In an unseen 10-AV scenario, this generator achieves stronger zero-shot generalization than the VD3QN policy transferred from the 6-AV environment.