Self-Supervised Contrastive Learning for Joint Active and Passive Beamforming in RIS-Assisted MU-MIMO Systems

Reconfigurable Intelligent Surfaces (RIS) can enhance system performance at the cost of increased complexity in multi-user MIMO systems. The beamforming options scale with the number of antennas at the base station/RIS. Existing methods for solving this problem tend to use computationally intensive iterative methods that are non-scalable for large RIS-aided MIMO systems. We propose here a novel self-supervised contrastive learning neural network (NN) architecture to optimize the sum spectral efficiency through joint active and passive beamforming design in multi-user RIS-aided MIMO systems. Our scheme utilizes contrastive learning to capture the channel features from augmented channel data and then can be trained to perform beamforming with only 1% of labeled data. The labels are derived through a closed-form optimization algorithm, leveraging a sequential fractional programming approach. Leveraging the proposed self-supervised design helps to greatly reduce the computational complexity during the training phase. Moreover, our proposed model can operate under various noise levels by using data augmentation methods while maintaining a robust out-of-distribution performance under various propagation environments and different signal-to-noise ratios (SNR)s. During training, our proposed network only needs 10% of labeled data to converge when compared to supervised learning. Our trained NN can then achieve performance which is only $~7\%$ and $~2.5\%$ away from mathematical upper bound and fully supervised learning, respectively, with far less computational complexity.