QoS-Aware Deep Unsupervised Learning for STAR-RIS Assisted Networks: A Novel Differentiable Projection Framework

Mehrazin Alizadeh, Xavier Mootoo, Omer Waqar, Hina Tabassum

Abstract: In this letter, we propose a Quality-of-Service (QoS)-aware deep unsupervised learning framework to jointly optimize beamforming and phase-shifts in a simultaneously transmitting and reflecting reconfigurable intelligence surface (STAR-RIS) assisted multi-user multi-antenna system. The objective is to maximize the downlink network sum-rate considering highly non-convex users’ minimum rate (or QoS) constraints. To satisfy constraints with zero violation, we devise a novel piece-wise differentiable projection function that projects the output of the deep neural network (DNN) to the feasible solution set of the problem. Different from the existing methods, the proposed projection function offers considerable improvement in sum-rate by enabling the search inside the feasible space. The proposed framework is general to capture both RIS and STAR-RIS-aided networks. Our proposed framework is shown to outperform genetic algorithm and existing projection-based DNNs in terms of sum-rate, time complexity, and convergence, while achieving zero probability of constraint violation.