Self-Supervised Neural Networks for Precoding in MIMO Rate Splitting Multiple Access Systems

The use of machine learning (ML) tools to address the challenges posed in next generation of wireless communication systems has been gaining significant momentum. In this paper, we investigate the use of self-supervised data-driven schemes for precoder optimization in the downlink of a Multiple-Input Multiple-Output (MIMO) Rate Splitting Multiple Access (RSMA) system. Specifically, we propose two architectures based on Graph Neural Networks (GNN) and the Multi-Layer Perceptron (MLP) respectively, and analyze their achievable sum-rate performance in the underloaded and critically-loaded regime as the system scales up. The intention is to explore several alternatives to conventional iterative precoding benchmarks like Weighted Minimum Mean Square Error (WMMSE) which are computationally intensive algorithms. We evaluate the different precoding policies learnt by the neural network architectures by closely studying the respective radiation patterns. Complementarily, we compare their complexity in terms of the number of trainable parameters and computation time. We also extend the evaluation to non-linear channels in RSMA settings. The study dictates that GNN-based scheme offers an interesting performance-complexity and scalability trade-off.