Generative AI-Based Hierarchical DRL Framework for RIS-Assisted THz Massive MIMO Systems

Terahertz (THz) massive multiple-input multiple-output (mMIMO) systems offer ultra-high data rates but face significant challenges such as beam squint effects, high power consumption, severe path loss, and signal blockage. Incorporating reconfigurable intelligent surfaces (RIS) can mitigate these issues but complicates channel state information (CSI) acquisition. To address this, we propose a generative artificial intelligence-based hierarchical deep reinforcement learning (GAI-HDRL) framework that jointly performs channel prediction, hybrid precoding at the base station (BS), passive RIS beamforming, and digital combining at the UE to minimize transmit power. The proposed GAI-HDRL efficiently decomposes the optimization into high-level (precoding and combining) and low-level actions (CSI prediction and RIS configuration), achieving fast convergence and improved prediction accuracy. Simulation results confirm its superiority over state-of-the-art methods in terms of performance and computational efficiency, demonstrating practical viability in THz communications.