Unsupervised Learning-Based Joint Resource Allocation and Beamforming Design for RIS-Assisted MISO-OFDMA Systems

Reconfigurable intelligent surface (RIS) is regarded as one of the pivotal technologies for sixth-generation wireless communication systems. Nevertheless, the inherent scarcity of wireless communication resources motivates the need for effective resource allocation schemes in RIS-assisted systems. This paper investigates the downlink transmission of an RIS-assisted multiple-input single-output (MISO) orthogonal frequency division multiple access (OFDMA) communication systems. To achieve a high system sum rate with low computational complexity, we develop a two-stage unsupervised learning based approach with customized loss function for the RIS reflection phase shift design, active beamforming at base station (BS) and time-frequency resource block (RB) allocation. The proposed approach consists of two neural networks: BeamNet, which takes channel state information (CSI) as input to predict the RIS reflection phase shift, and AllocationNet, which generates RB allocation decisions based on the equivalent CSI from the BS to the users, where the equivalent CSI is obtained by combining the original CSI with the RIS reflection phase shifts predicted by BeamNet. The active beamforming is implemented using the maximum ratio transmission and water-filling algorithm. In order to incorporate the discrete constraints of RIS reflection phase shift and RB allocation decisions into the network while maintaining network differentiability, we introduce a quantization function and the Gumbel softmax trick into BeamNet and AllocationNet, respectively. Furthermore, a customized loss function and phased training strategy are devised to enhance training efficiency and address quality-of-service constraints. Simulation results demonstrate that the proposed approach achieves 99.93% of the system sum rate of the successive convex approximation (SCA) method while requiring only 0.036% of its runtime. Additionally, the method’s effectiveness and robustness are validated under different delay tap numbers, user distributions, and Rician factors, demonstrating its strong adaptability to different communication environment.