Self-Nomination Based Feedback Reduction for MU-MIMO

This paper presents a deep learning-based decentralized feedback reduction method, termed “self-nomination,” for multi-user multiple-input multiple-output (MU-MIMO) systems. Traditional approaches require all user equipment (UE) to provide channel state information (CSI) feedback regardless of current channel conditions, leading to excessive overhead and energy consumption. In contrast, our method enables UEs to autonomously decide whether to feed back CSI by analyzing their own channel vectors. We formulate the uplink feedback capacity as a constraint and employ a primal-dual learning framework for training. To handle the non-differentiability of hard decision-making in the feedback process, we utilize the straight-through estimator. Simulation results demonstrate that our approach reduces the number of UEs providing uplink feedback by approximately 50%, while maintaining a comparable downlink sum rate. In addition to also saving UE power, another positive side effect of this technique is complexity reduction in downlink MU-MIMO scheduling, since there are fewer UEs to consider.