DropNet: An Improved Dropping Algorithm Based on Neural Networks for Line-of-Sight Massive MIMO

Abstract

In line-of-sight massive MIMO, the downlink channel vectors of few users may become highly correlated. This high correlation limits the sum-rates of systems employing linear precoders. To constrain the reduction of the sum-rate, few users can be dropped and served in the next coherence intervals. The optimal strategy for selecting the dropped users can be obtained by an exhaustive search at the cost of high computational complexity. To alleviate the computational complexity of the exhaustive search, a correlation-based dropping algorithm (CDA) is conventionally used, incurring a sum-rate loss with respect to the optimal scheme. In this paper, we propose a dropping algorithm based on neural networks (DropNet) to find the set of dropped users. We use appropriate input features required for the user dropping problem to limit the complexity of DropNet. DropNet is evaluated using two known linear precoders: conjugate beamforming (CB) and zero-forcing (ZF). Simulation results show that DropNet provides a trade-off between complexity and sum-rate performance. In particular, for a 64-antenna base station and 10 single-antenna users: (i) DropNet reduces the computational complexity of the exhaustive search by a factor of 46 and 3 for CB and ZF, respectively, (ii) DropNet improves the 5th percentile sum-rate of CDA by 0.86 and 2.33 bits/s/Hz for CB and ZF, respectively.

Keywords

• Correlated scenarios
• dropping algorithm
• line-of-sight massive MIMO
• neural network