Machine Learning-Enabled RIS-Assisted MmWave NOMA Systems: RIS Partitioning, Beamforming Design, and Power Allocation

• Published in: IEEE transactions on communications p. 1
• Main Authors: Wei, Tongyi, Chen, Weizhi, Tang, Kun, Zheng, Beixiong, Feng, Wenjie, Che, Wenquan, Xue, Quan
• Format: Journal Article
• Language: English
• Published: IEEE 2025
• Subjects: Array signal processing; Clustering algorithms; Job shop scheduling; machine learning (ML); Machine learning algorithms; Millimeter wave communication; millimeter-wave; NOMA; non-orthogonal multiple access (NOMA); Optimization; Partitioning algorithms; Reconfigurable intelligent surface (RIS); Reconfigurable intelligent surfaces; Resource management
• ISSN: 0090-6778; 1558-0857
• DOI: 10.1109/TCOMM.2025.3554673

In this paper, a reconfigurable intelligent surface (RIS)-assisted millimeter-wave (mmWave) non-orthogonal multiple access (NOMA) communication system is considered, where the RIS is virtually partitioned into several sub-RISs to serve different user clusters. We aim to maximize the sum achievable rate of the considered system by jointly optimizing user clustering, RIS partitioning, active beamforming, passive beamforming, and power allocation, subject to the quality-of-service (QoS) requirements of each user and the maximum transmit power constraint of the base station (BS). To tackle the formulated non-convex joint optimization problem, a three-stage algorithm based on machine learning (ML) is proposed. In the first stage, user clustering is performed by combining the K-means algorithm and the Gaussian mixture model (GMM). In the second stage, a deep learning (DL)-based RIS partitioning method is proposed, which utilizes double cascaded deep neural networks (DNNs) to partition the RIS into multiple sub-RISs virtually, each of which serves a user cluster. In the third stage, a deep reinforcement learning (DRL) algorithm based on deep deterministic policy gradient (DDPG) is invoked to obtain the active beamforming, passive beamforming, and power allocation. Simulation results demonstrate that the proposed RIS partitioning method can effectively improve the utilization of RIS and thereby increase the total throughput of the system. Moreover, the sum achievable rate obtained by the proposed ML-based three-stage algorithm is higher than that of the conventional alternating optimization (AO) algorithm.