Graph-MLP Based Optimization for RIS-Assisted Multi-User MISO Systems With Implicit CSI

Reconfigurable Intelligent Surfaces (RIS) are emerging as a key technology for enabling 6G communication systems. This study focuses on an RIS-assisted downlink multi-user multi-input single-output (MISO) system, optimizing the BS precoding vectors and RIS phase-shifting parameters to maximize the sum-rate. Unlike traditional methods that require channel estimation followed by parameter optimization, we propose a machine learning-based approach to directly optimize BS beamformers and RIS reflection coefficients based on the system objective, bypassing the need for explicit channel estimation. Our method utilizes a deep neural network with a graph-Multi Layer Perceptron (MLP) architecture to map received pilot signals and user location data to an optimized system configuration. This architecture efficiently captures user interactions within the network. In particular, the proposed approach achieves high performance in the sum-rate objective with significantly fewer pilots than conventional methods.