Deep learning enabled MIMO beamforming optimization with practical channel information

Deep learning has emerged as a new design and optimization tool for future wireless communications networks. This thesis focuses on deep learning techniques for the optimization of transmit beamforming in the downlink for the multiuser multi-antenna (or multiple-input multiple-output, MIMO) systems with practical channel information. Channel information is critical to fully exploit the potential of MIMO technology and in most existing deep learning applications for MIMO beamforming, perfect channel state information is normally assumed available which is used as the input to the deep neural networks. However, in practice, it is extremely challenging, if not impossible, to obtain the perfect channel state information. This could be due to the high overhead in uplink channel feedback and fast fading associated with high user mobility. Consequently, the performance of deep learning based beamforming optimization is greatly degraded without perfect channel state information. In this thesis, we investigate deep learning techniques to deal with three specific scenarios in which perfect channel state information is not available or its distribution varies.

First, we study fast adaptation techniques for deep neural networks with varying channel distributions. Wireless channel distributions strongly depend on the electromagnet environment. A deep neural network trained for one environment will suffer from model mismatch if users move to a different environment. To tackle this problem, we propose a simple yet effective adaptive framework, which trains an embedding model as a transferable feature extractor, followed by fitting the support vector regression. Compared to the existing meta learning algorithm, our method achieves superior performance, and does not necessarily need labelled data in the pre-training and does not require fine-tuning of the pre-trained model in the adaptation.

Secondly, we propose a deep learning optimization approach for the downlink beamforming when only the easier-to-obtain uplink channel information is available, but its relation to the downlink channel is unknown. Our main contribution is a new model-driven learning technique that exploits the structure of the optimal downlink beamforming to design an effective hybrid learning strategy that jointly considers the learning performance of the downlink channel, the power and the sum rate in the training stage. We further extend the developed technique to massive multiple-input multiple-output scenarios and achieve a distributed learning strategy for multicell systems without an inter-cell signalling overhead. Results indicate the advantage of our propose method over the data-driven deep learning method.

Thirdly, to further reduce the channel estimation overhead and improve the spectrum efficiency, we study deep learning techniques to predict transmit beamforming based on only historical channel data without current channel information. Specifically, we propose a joint learning framework that incorporates channel prediction and power optimization, and produces prediction for transmit beamforming directly. In addition, we propose to use the attention mechanism in the recurrent Neural Networks to improve the accuracy of channel prediction. The results using both a simple autoregressive process model and the more realistic 3GPP spatial channel model verify that our proposed predictive beamforming scheme can significantly improve the effective spectrum efficiency compared to traditional channel estimation and state of the art deep learning method.