Unsupervised Learning-Aided Discrete RIS Configuration Estimator

Reconfigurable Intelligent Surface (RIS) consists of mostly-passive elements capable of electronically steering the impinging signal with configured phase shifts. However, achieving infinite phase resolution is infeasible and the phases needs to be quantized for practical implementation. In this paper, we propose an unsupervised learning-based method to estimate the optimal discrete RIS phase configuration by utilizing key eigen vectors of the available channel state information (CSI) as the input. The direct link between a base station (BS) and user equipment (UE) is not considered in the proposed method and only the CSI of the RIS-aided link is utilized. Further, the proposed network design is independent of the number of transmit and receive antennas at the BS and UE respectively, making the solution generalized and adaptable towards different BSs and UEs in the communication. Due to the unsupervised nature of the proposed method, the solution does not rely on any labeled data, obtaining which is otherwise an exhaustive process. The performance of the proposed solution is analyzed through rigorous simulations and shows the effectiveness of directly estimating discrete RIS configurations as compared to obtaining infinite resolution phase-shifts and then quantizing them. The performance analysis further demonstrates the efficacy of the proposed system with reduced computational complexity and generalized solution with respect to number of BS and UE antennas.