Unsupervised Learning-Based Joint Beamforming and Phase-Shift Optimization for RIS-Assisted DeepMIMO With Large-Scale Arrays

This paper considers a reconfigurable intelligent surface (RIS)-aided network, where discrete phase-shift RIS is investigated in large-scale arrays with hundreds of antennas at the source and thousands of passive elements at the RIS. We propose unsupervised learning (UnSL) approaches that eliminate the need for labeled data to address the joint beamforming and phase-shift (JBPS) optimization problem with high degrees of freedom (DoF), i.e., thousands of optimization variables, for efficient transmission design in RIS-DeepMIMO networks. Performance is analyzed by comparing the signal-to-noise ratio (SNR) with that of end-to-end SNR-based exhaustive search (ESES) and particle swarm optimization (PSO) algorithms under maximum ratio transmission (MRT) beamforming. We show that MRT-PSO, MRT-UnSL, and JBPS-UnSL with multitask neural network suffer performance degradation due to the high-dimensional input. On the other hand, numerical results reveal that the GPU-MRT-CCES method outperforms the other solutions and exhibits high scalability, owing to its novel combination of a predefined MRT solution, theoretical cascaded channel-based RIS configuration, and GPU-parallelized computation.