Application of Deep Learning to Sphere Decoding for Large MIMO Systems

Although the sphere decoder (SD) is a powerful detector for multiple-input multiple-output (MIMO) systems, it has become computationally prohibitive in massive MIMO systems, where a large number of antennas are employed. To overcome this challenge, we propose fast deep learning (DL)-aided SD (FDL-SD) and fast DL-aided $K$-best SD (KSD, FDL-KSD) algorithms. Therein, the major application of DL is to generate a highly reliable initial candidate to accelerate the search in SD and KSD in conjunction with candidate/layer ordering and early rejection. Compared to existing DL-aided SD schemes, our proposed schemes are more advantageous in both offline training and online application phases. Specifically, unlike existing DL-aided SD schemes, they do not require performing the conventional SD in the training phase. For a $24 times 24$ MIMO system with QPSK, the proposed FDL-SD achieves a complexity reduction of more than $90%$ without any performance loss compared to conventional SD schemes. For a $32 times 32$ MIMO system with QPSK, the proposed FDL-KSD only requires $K = 32$ to attain the performance of the conventional KSD with $K=256$, where $K$ is the number of survival paths in KSD. This implies a dramatic improvement in the performance--complexity tradeoff of the proposed FDL-KSD scheme.

Hybrid Beamforming and Adaptive RF Chain Activation for Uplink Cell-Free Millimeter-Wave Massive MIMO Systems

In this work, we investigate hybrid analog-digital beamforming (HBF) architectures for uplink cell-free (CF) millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) systems. {We first propose two HBF schemes, namely, decentralized HBF (D-HBF) and semi-centralized HBF (SC-HBF). In the former, both the digital and analog beamformers are generated independently at each AP based on the local channel state information (CSI). In contrast, in the latter, only the digital beamformer is obtained locally at the AP, whereas the analog beamforming matrix is generated at the central processing unit (CPU) based on the global CSI received from all APs. We show that the analog beamformers generated in these two HBF schemes provide approximately the same achievable rates despite the lower complexity of D-HBF and its lack of CSI requirement.} Furthermore, to reduce the power consumption, we propose a novel adaptive radio frequency (RF) chain-activation (ARFA) scheme, which dynamically activates/deactivates RF chains and their connected analog-to-digital converters (ADCs) and phase shifters (PSs) at the APs based on the CSI. For the activation of RF chains, low-complexity algorithms are proposed, which can achieve significant improvement in energy efficiency (EE) with only a marginal loss in the total achievable rate.