Coordinated beamforming with neural networks

Beamforming algorithms have a huge importance in modern wireless communication systems. They allow an important improvement in the final performance, but complex iterative algorithms might be required to solve the beamforming design problem. This master thesis, realized in collaboration with Sony at the Stuttgart Laboratory, focuses on the utilization of graph neural networks to improve already existing beamforming algorithms utilized for wireless communications, specifically for WiFi scenarios. In particular, the problem that has been taken into account is referred to the maximization of the weighted sum rate in a downlink MU-MIMO scenario. The convergence of the iterative algorithm utilized to obtain the final beamformers that solve the problem might require a large number of iterations, making the channel state information obsolete and causing therefore an important performance loss. To obtain a faster convergence of the algorithm and to avoid this performance degradation, a deep learning model based on unfolding and graph neural networks is considered.