Joint beamforming with extremely large scale RIS: a sequential multi-agent A2C approach

Journal article : Letter

Jointly optimizing the base station (BS) precoding matrix and reconfigurable intelligent surface (RIS) phases is challenging in RIS-assisted multi-user multiple-input singleoutput (MU-MISO) systems, particularly with extremely large RISs. This paper proposes a deep reinforcement learning (DRL) approach based on a sequential multi-agent advantage actorcritic (A2C) framework that accounts for discrete RIS phases, imperfect channel state information (CSI), and inter-user channel correlation. The computational complexity is analyzed, and the proposed method is benchmarked against the zero-forcing (ZF), minimum mean square error (MMSE), and single-agent A2C beamformers in terms of sum spectral efficiency (SE). Simulation results show that the proposed algorithm achieves higher SE and a certain degree of robustness to moderate channel estimation errors.

Authors: Chai, Z; Xu, J; Coon, J; Alouini, M-S

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: IEEE
• Journal: IEEE Wireless Communications Letters
• Volume: 15
• Pages: 1662-1666
• Publication date: 2026-01-30
• Acceptance date: 2026-01-29
• DOI: 10.1109/lwc.2026.3660067
• EISSN: 2162-2345
• ISSN: 2162-2337
• Language: English
• Keywords: multi-agent deep reinforcement learning; training; channel estimation; precoding; array signal processing; correlation; reconfigurable intelligent surfaces; spectral efficiency; signal to noise ratio; optimization; indexes; RIS; joint beamforming
• Subtype: Letter