Low-Complexity Hybrid Beamforming for Multiuser Millimeter Wave Systems With Collaborative Learning

We propose a low-complexity hybrid beamforming scheme with fast beam training based on collaborative learning to improve energy efficiency in multiuser millimeter wave (mmWave) systems. First, we address the non-stationary multi-armed bandit (MAB) problem by considering the codeword index and signal-to-noise ratio (SNR) as the action and reward, respectively, for beam training in time-varying channels. We develop a beam selection algorithm called the collaborative upper confidence bound (UCB) policy by clustering users according to their preferences for each codeword, using neural collaborative filtering. Second, we formulate a nonlinear least-squares problem using feedback SNR values to predict the beam gains for the unirradiated codewords. We then assign the radio frequency (RF) precoder using codewords with high expected beam gains for each user. Once the RF precoder is assigned, we estimate the equivalent baseband channel by receiving feedback on its direction information. Finally, we design the baseband precoder based on the acquired equivalent baseband channels. Simulation results demonstrate that the proposed hybrid beamforming scheme with fast beam training improves energy efficiency by reducing beam training overhead and increasing beam searching accuracy.