Multi-User Hybrid Beamforming Relying on Learning-Aided Link-Adaptation for mmWave Systems

Hybrid beamforming (HBF) relying on a large antenna array is conceived for millimeter wave (mmWave) systems, where the beamforming (BF) gain compensates for the propagation loss experienced. The BF gain required for a successful transmission depends on the user's distance from the base station (BS). For the geographically separated users of a multi-user mmWave system, the BF gain requirements of different users tend to be different. On the other hand, the BF gain is directly related to the number of antenna elements (AEs) of the array. Therefore, in this paper, we propose an HBF design for the downlink of multi-user mmWave systems, where the number of AEs employed at the BS for attaining BF gains per user is dependent on the user's distance. We then propose grouping of the RF chains at the BS, where each group of RF chains serves a specific group of users depending on the nature of the channel. Furthermore, to support the escalating data rate demands, the exploitation of link-adaptation techniques constitutes a promising solution, since the rate can be maximized for each link while maintaining a specific target bit error rate (BER). However, given the time-varying nature of the wireless channel and the non-linearities of the amplifiers, especially at mmWave frequencies, the performance of conventional link adaptation relying on pre-defined threshold values degrades significantly. Therefore, we additionally propose a two-stage link adaptation scheme. Specifically, in the first stage, we switch on or off both the digital precoder and the combiner depending on the nature of the channel, while in the second stage a machine-learning assisted link-adaptation is proposed, where the receiver predicts whether to request spatial multiplexing- or diversity-aided transmission from the BS for every new channel realization. We demonstrate by the simulation that having both a digital precoder and a combiner in a single dominant path scenario is redundant. Furthermore, our simulations show that the learning assisted adaptation provides significantly higher data rates than that of the conventional link-adaptation, where the reconfiguration decision is simply based on pre-defined threshold values.