Deep Learning-Based Optimization for Reconfigurable Intelligent Surface-Assisted Communications

Reconfigurable Intelligent Surfaces (RISs) are an emerging technology in the evolution towards the Sixth Generation (6G) of mobile communications. They are capable of enhancing the overall system performance and extending the coverage of the existing cells. They are built by a large amount of low-cost meta-elements that can be configured by tuning their phase shifts, and hence, the channel response can be constructively combined and forwarded to some specific direction. Many algorithms have been proposed to obtain the optimum phase shifts, generally assuming a single-carrier system and/or a medium-size RIS to constrain the complexity of the optimization process. In this work, we propose a flexible and scalable unsupervised learning model, capable of obtaining the best phase shifts for any scenario. Our proposal is able to handle multi-carrier waveforms and very large-size RIS, considering both continuous and discrete phase shifts. Besides, we also propose the use of clustering to reduce further the complexity while maintaining the performance. A comparison in terms of achievable rate and time execution is provided in order to show the superiority of our proposal against the existing solutions.