Machine Learning for Channel Estimation and Hybrid Beamforming in Millimeter-Wave Wireless Networks
MetadataShow full item record
PublisherThe University of Arizona.
RightsCopyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction, presentation (such as public display or performance) of protected items is prohibited except with permission of the author.
AbstractThe continuous growth of mobile users and high-speed wireless applications drives the demand for using the abundant bandwidth of mmWave (millimeter-wave) frequencies. On one hand, a massive number of antennas can be supported due to small wavelengths of mmWave signals, which allow using antennas with small form factors. On the other hand, the free space path loss increases with the square of the frequency, which implies that the path loss would be severe in mmWave frequencies. Fortunately, one can compensate for the performance degradation due to the path loss by using directional beamforming (BF) along with the high gain large antenna array systems (massive MIMO). This dissertation tackles three distinct problems, namely channel estimation in massive MIMO, signal detection in massive MIMO, and efficient design of hybrid BF algorithms. In the first part of this dissertation, we focus on the effective channel estimation for massive MIMO systems to overcome the pilot contamination problem. We present an adaptive independent component analysis (ICA)-based channel estimation method, which outperforms conventional ICA as well as other conventional methods for channel estimation. We also make use of compressive sensing (CS) methods for channel estimation and show the advantages in terms of channel estimation accuracy and complexity. In the second part of this dissertation, we consider the problem of signal detection specifically focusing on the scenarios when non-Gaussian signals need to be detected and the receiver may be equipped with a large number of antennas. We show that for the case of non-Gaussian signal detection it turns out the conventional Neyman-Pearson (NP) detector does not perform well for the low signal-to-noise-ratio (SNR) regime. Motivated by this, we propose a bispectrum detector, which is able to better detect the corresponding non-Gaussian information in the signal. We also present the theoretical analysis for the asymptotic behavior of Probability of False Alarm and Probability of Detection. We show the performance of signal detection (for both Gaussian and non-Gaussian signals) as a function of the number of antennas and sampling rate. We also obtain the scaling behavior of the performance in the massive antenna regime. The third part of this dissertation covers the efficient design of hybrid BF algorithms with a specific focus on massive MIMO systems in mmWave networks. The key challenge in the design of hybrid BF algorithms in such networks is that the computational complexity can be prohibitive. We start by focusing on the fundamental approach of finding BF solutions through singular value decomposition (SVD) and explore the role of ML techniques to perform SVD. The first part of this contribution focuses on the data-driven approach to SVD. We propose three deep neural network (DNN) architectures to approximate the SVD, with varying levels of complexity. The methodology for training these DNN architectures is inspired by the fundamental property of SVD, i.e., it can be used to obtain low-rank approximations. We next explicitly take the constraints of hybrid BF into account (such as quantized phase shifters, power constraints), and propose a novel DNN-based approach for the design of hybrid BF systems. Our results show that DNNs can be an attractive and efficient solution for both estimating the SVD as well as hybrid beamformers. Furthermore, we provide time complexity and memory requirement analyses for the proposed DNN-based and state-of-the-art hybrid BF approaches. We then propose a novel reinforcement learning-based hybrid BF algorithm that applies Q-learning in a supervised manner. We analyze the computational complexity of our algorithm as a function of iteration steps and show that a significant reduction in computational complexity is achieved compared to the exhaustive search. In addition to exploring supervised approaches, in the remaining part of this contribution we also explore unsupervised methods for SVD and hybrid BF. These methods are particularly attractive for scenarios when channel conditions change too fast and we may not have a pre-existing dataset of channels and the corresponding optimal BF solutions, which are required for supervised learning. For unsupervised learning, we explore two techniques namely autoencoders and generative adversarial networks (GANs) for both the SVD and hybrid BF. We first propose a linear autoencoder-based approach for the SVD, and then provide a linear autoencoder-based hybrid BF algorithm, which incorporates the constraints of the hybrid BF. In the last part of this contribution, we focus on two different generative models: variational autoencoders (VAEs) and GANs to reduce the number of training iterations compared to the linear autoencoder-based approach. We first propose VAE and Wasserstein GAN (WGAN) based algorithms for the SVD. We then present a VAE and a novel GAN architecture to find the hybrid BF solutions.
Degree ProgramGraduate College
Electrical & Computer Engineering