Search for: All records

One-bit massive MIMO has gained much attention in the areas of wireless communication and sensing. Among the various receiver designs, the maximum-likelihood-based receivers achieve state-of-the-art performance. Through this work we provide both analytical insight into the likelihood formulation, and develop a one-bit MIMO receiver, motivated specifically from this analysis. In particular, (i) Properties of the original Gaussian CDF based likelihood function are analyzed, culminating in an improved gradient descent (GD) algorithm for one-bit MIMO. (ii) This improved GD update rule is further enhanced through an accelerated GD method, improving convergence performance. (iii) The likelihood analysis is extended to an effective surrogate function for the Gaussian CDF, i.e., the logistic regression (LR). The presented analytical framework for the CDF also serves as a robust mathematical model to explain the enhanced performance of the LR, when utilized as a surrogate likelihood. (iv) Detection from a finite M-QAM constellation is incorporated by introducing a Gaussian denoiser to project the detected symbols onto the M-QAM subspace. This is implemented as a novel, unfolded, DNN architecture for one-bit detection. Through our experimental validation we demonstrate results on par with the current state- of-the-art methods for one-bit MIMO detection. 

Multi-user massive MIMO signal detection from one-bit received measurements strongly depends on the wireless channel. To this end, majority of the model and learning-based approaches address detector design for the rich-scattering, homogeneous Rayleigh fading channel. Our work proposes detection for one-bit massive MIMO for the lower diversity mmWave channel. We analyze the limitations of the current state-of-the-art gradient descent (GD)-based joint multiuser detection of one-bit received signals for the mmWave channels. Addressing these, we introduce a new framework to ensure equitable per-user performance, in spite of joint multi-user detection. This is realized by means of: (i) a parametric deep learning system, i.e., the mmW-ROBNet, (ii) a constellation-aware loss function, and (iii) a hierarchical detection training strategy. The experimental results corroborate this proposed approach for equitable per-user detection.