EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Multi Agent Deep Reinforcement Learning for Joint Power Control and Beamforming in mmWave-NOMA HetNets

Ayman Rabee, Imad Barhumi

IEEE Open Journal of the Communications Society · 2026 · 7 · 6888-6903

DOI: 10.1109/ojcoms.2026.3705650 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Open Journal of the Communications Society, 2026. Verfasst von Ayman Rabee und 1 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Millimeter-wave (mmWave) heterogeneous networks (HetNets) integrated with non-orthogonal multiple access (NOMA) are a promising solution for meeting the capacity demands of next-generation wireless systems. However, jointly optimizing hybrid beamforming and power allocation is made challenging by non-convex constraints, coupled variables, and complex interference interactions. To address these challenges, we propose a multi-agent deep reinforcement learning (MADRL) framework based on the Multi-Agent Twin-Delayed Deep Deterministic Policy Gradient (MATD3) algorithm. The problem is formulated as a decentralized partially observable Markov decision process (Dec-POMDP) where base stations act as autonomous agents. To overcome the coordination limitations of independent policies, we introduce a graph neural network (GNN)-enhanced MATD3 actor that incorporates multi-head graph attention to enable explicit inter-agent message passing under centralized training with coordinated decentralized execution. During execution, each BS computes its own action locally after exchanging only low-dimensional latent embeddings with neighboring BSs, without requiring a centralized runtime controller or sharing raw observations. A multiplicative quality-of-service (QoS)-weighted reward is designed to maximize the system sum spectral efficiency (SE). Complexity analysis and empirical results show that the shared graph-attention layers improve parameter efficiency while maintaining practical scalability across the considered network sizes. The proposed framework achieves up to a 42% improvement in average SE over baseline MATD3 and reduces average QoS violations to 0.016 per episode, corresponding to reduction of 86% and 96% compared to MATD3 and MADDPG, respectively. These findings indicate that learned graph-based coordination enables robust, continuous-control interference management in dynamic mmWave-NOMA HetNets.

Zitieren

APA

Rabee, A., & Barhumi, I. (2026). Multi Agent Deep Reinforcement Learning for Joint Power Control and Beamforming in mmWave-NOMA HetNets. IEEE Open Journal of the Communications Society, 7, 6888-6903. https://doi.org/10.1109/ojcoms.2026.3705650

Vancouver

Rabee A, Barhumi I. Multi Agent Deep Reinforcement Learning for Joint Power Control and Beamforming in mmWave-NOMA HetNets. IEEE Open Journal of the Communications Society. 2026;7:6888-6903. doi:10.1109/ojcoms.2026.3705650

BibTeX
@article{Rabee2026,
  author = {Rabee, Ayman and Barhumi, Imad},
  title = {Multi Agent Deep Reinforcement Learning for Joint Power Control and Beamforming in mmWave-NOMA HetNets},
  journal = {IEEE Open Journal of the Communications Society},
  year = {2026},
  volume = {7},
  pages = {6888--6903},
  doi = {10.1109/ojcoms.2026.3705650}
}
Kostenlos starten →