Large Artificial Intelligence Model Guided Deep Reinforcement Learning for Resource Allocation in Non Terrestrial Networks

TL;DR

This work tackles resource allocation in dynamic Non‑Terrestrial Networks by introducing a Large AI Model guided DRL (LAM–DRL) where a Large Language Model provides high‑level strategies that shape the DRL reward. The approach formulates the problem as an MDP and uses strategy‑conditioned attention within TD3 to learn policies, achieving notable gains in sum rate and fairness while reducing outages, and offering interpretability through attention analyses. Key contributions include the NTN‑specific MDP with KPI‑driven reward shaping, a novel strategy embedding mechanism for LLM guidance, and empirical evidence of improved robustness under nominal and extreme weather. The results demonstrate practical impact for next‑generation satellite networks, enabling more efficient and transparent control of non‑terrestrial communications.

Abstract

Large AI Model (LAM) have been proposed to applications of Non-Terrestrial Networks (NTN), that offer better performance with its great generalization and reduced task specific trainings. In this paper, we propose a Deep Reinforcement Learning (DRL) agent that is guided by a Large Language Model (LLM). The LLM operates as a high level coordinator that generates textual guidance that shape the reward of the DRL agent during training. The results show that the LAM-DRL outperforms the traditional DRL by 40% in nominal weather scenarios and 64% in extreme weather scenarios compared to heuristics in terms of throughput, fairness, and outage probability.

Figures (5)

• Figure 1: LEO satellite constellation and user distribution across latitude zones.
• Figure 2: Proposed framework. The environment provides a state $s_t$. A prompt summarizes $s_t$ and operator objectives and is sent to the LLM. The LLM responds with a strategy label $\sigma_t$. The strategy is embedded as a vector $e_{\sigma_t}$ that conditions a single–head attention layer and shapes the reward. intent.
• Figure 3: Performance of llm–drl and baseline schemes under nominal and extreme weather conditions.
• Figure 4: Strategy usage and associated sum-rate performance for llm–drl across episodes. Points show per-episode sum rate coloured by the selected llm strategy; the solid curve is a 10-episode moving average with a shaded $\pm 1\sigma$ window.
• Figure 5: Mean attention weights across seven input feature categories with 95% confidence intervals over 100 episodes.