Abstract
The reflective elements on a reconfigurable intelligent surface (RIS) can be tuned to improve the propagation environment and, in turn, the system performance. However, RIS also brings challenges in terms of joint active and passive beamforming optimization. First, transmit beamforming has to be jointly designed with RIS to fully reap beamforming gain and signal-to-interference plus noise ratio (SINR) performance. Also, given that the number of beamforming options grows with the number of antennas/elements on the base station(BS)/RIS, it makes exhaustive search unpractical for multi-user (MU) multiple-input-multiple-output (MIMO) systems. In order to reduce the search overhead for joint BS and RIS beamforming optimization, we propose a novel joint active BS and passive RIS beamforming scheme based on a bespoke convolutional neural network (CNN) architecture. More specifically, we develop a fast converge and lightweight CNN with loss function designed based on the probability density function (PDF) of beams. By using this CNN, exhaustive search is only needed in data collection part to derive labels and no longer needed in prediction part. We also adopt a realistic blockage scenario to simulate non-stationary channels. Our proposed solution exhibits fast convergence speed, low neural network (NN) complexity and high prediction accuracy. Simulation results, based on benchmark dataset, show that our approach can significantly outperform comparable existing machine learning algorithms. It can achieve 70% shorter convergence time with around 91% beam prediction accuracy.