An LSTM-MADDPG Framework for Joint Beamforming Control in STAR-RIS-Assisted Cell-Free Massive MIMO Network

Abstract

Conventional reconfigurable intelligent surfaces (RIS)-aided cell-free massive multiple-input multiple-output (CF mMIMO) networks are inherently constrained by half-space reflection, which limits coverage expansion and flexible deployment. To overcome this issue, simultaneously transmitting and reflecting reconfigurable intelligent surfaces (STAR-RISs) provides both reflection and transmission, enabling full-space coverage and enhanced adaptability, thereby becoming a promising technique for next-generation wireless systems. This work investigates the downlink transmission of a CF mMIMO system assisted by STAR-RISs, with the goal of maximizing the system sum spectral efficiency (Sum-SE). We develop a Long Short-Term Memory-enhanced Multi-Agent Deep Deterministic Policy Gradient (LSTM-MADDPG) framework to jointly optimize access point (AP) beamforming and STAR-RIS phase adjustments, which exploits historical observation sequences to accelerate the learning process and achieve superior steady-state performance. Simulation results demonstrate that the proposed approach outperforms MADDPG, MAPPO, and AO benchmark schemes by 3.8%, 5.3%, and 33.5%, respectively. The obtained results underscore that STAR-RIS is a key enabler for expanding coverage to full-space and providing additional spatial degrees of freedom, thereby significantly boosting the SE and reliability of distributed CF networks in future 6G systems.