BS-UE Association and Power Allocation in Heterogeneous Massive MIMO Systems

• Published in: IEEE access 2020-01, Vol.8, p.1-1
• Main Authors: Yan, Hangsong, Lu, I-Tai
• Format: Article
• Language: English
• Subjects: Achievable Per-UE Rate; Antennas; Approximation; Artificial neural networks; BS-UE Association; Channel estimation; Configurations; Downlink; Fading channels; Heuristic; Heuristic methods; HM-MIMO; Lower bounds; Mathematical analysis; MIMO (control systems); MIMO communication; Network Utility; Neural Network; Performance degradation; Power Allocation; Power control; Resource management; Signal strength; Wireless communication
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2020.3029518

We consider a downlink Heterogeneous Massive Multiple-Input Multiple-Output (HM-MIMO) system with the macro-base station (BS) having hundreds of antennas and each of the micro-BSs having tens of antennas. Two lower bounds and one approximation of the achievable per-user equipment (UE) rate in closed forms under this system are derived and compared in terms of network utility. Using the derived approximation, three BS-UE association approaches are proposed. Firstly, a simple, sub-optimal, and heuristic multiple BS-UE association approach is designed, which achieves around 25% utility performance improvement compared with the conventional maximum received signal strength (Max-RSS) approach. Secondly, based on the results given by this heuristic approach, a learning approach for BS-UE association using convolutional neural network (CNN) is introduced. After being fully trained, the CNN can take any new BS-UE configuration as input and provide a sub-optimal BS-UE association for that configuration directly. It has only a small performance degradation compared with the proposed heuristic approach. Thirdly, realizing that the BS-UE connection probability in the proposed CNN architecture can be considered as a power allocation ratio, a combined power allocation and association approach is proposed. Its performance achieves as high as 60% utility improvement compared with the Max-RSS association and is also comparable to that achieved by the max-min power allocation approach which requires more than 10000× running time. It is remarkable that by using the gradients of the derived achievable per-UE rate approximation with respect to the power control coefficients, accurate target data is in fact not required for training in this approach.