FMA-MADDPG: Constrained Multi-Agent Resource Optimization with Channel Prediction in 6G Non-Terrestrial Networks.

Area of Science:

• Wireless Communications
• Artificial Intelligence
• Network Engineering

Background:

• Sixth-generation (6G) wireless systems aim to integrate terrestrial, aerial, and satellite networks for extensive remote sensing and service delivery.
• Non-terrestrial networks (NTNs) face challenges due to rapidly changing channels and heterogeneous multi-tier architectures, complicating real-time channel state acquisition and cooperative resource scheduling.

Purpose of the Study:

• To propose a novel framework, FMA-MADDPG, for efficient resource scheduling in 6G NTNs.
• To address the difficulties in real-time channel state acquisition and cooperative resource scheduling within heterogeneous NTN environments.

Main Methods:

• Developed a Fusion of Mamba and Attention (FMA) predictor utilizing a Mamba state-space backbone and multi-head self-attention to forecast future channel state information (CSI).
• Integrated predicted CSI into agent observations for scheduling decisions that account for expected channel variations.
• Implemented a constraint-based reward mechanism with performance thresholds and anti-idle penalties to foster fairness, prevent free-riding, and promote cooperation among heterogeneous agents.

Main Results:

• The FMA-MADDPG framework demonstrated superior performance compared to several deep reinforcement learning (DRL) baselines in a representative NTN uplink scenario.
• Achieved relative gains of approximately 10-20% in key metrics, including total reward, efficiency, load balance, and cooperation.

Conclusions:

• Prediction-aware cooperative reinforcement learning is a viable strategy for optimizing resources in future 6G NTN systems.
• The proposed framework effectively enhances cooperation and resource management in complex, dynamic NTN environments.