The next-generation wireless networks are required to support the continual exponential growth of mobile data traffic and a plethora of applications, which arouses unprecedented challenges to mathematical modeling and optimization. Recently, there is a surge of interests in deep learning-based communication systems, which do not require tractable mathematical models. These studies lead to promising results for various applications in wireless communications, e.g., resource management, channel estimation, and joint source-channel coding. However, the existing works adopted neural network architectures from applications such as computer vision and applied them as black boxes. Consequently, they often require huge amounts of training samples, yield poor performance in large-scale networks,...[ Read more ]

The next-generation wireless networks are required to support the continual exponential growth of mobile data traffic and a plethora of applications, which arouses unprecedented challenges to mathematical modeling and optimization. Recently, there is a surge of interests in deep learning-based communication systems, which do not require tractable mathematical models. These studies lead to promising results for various applications in wireless communications, e.g., resource management, channel estimation, and joint source-channel coding. However, the existing works adopted neural network architectures from applications such as computer vision and applied them as black boxes. Consequently, they often require huge amounts of training samples, yield poor performance in large-scale networks, and generalize poorly to different network settings. As 5G and beyond networks typically have densely deployed access points, massive clients, and dynamically changing client numbers and SNRs, it will be highly ineffective to apply these learning-based methods.

The main theme of this thesis is to open the black box by integrating the prior knowledge of wireless communications into the structures of neural networks. It consists of four parts. In Part I, we exploit the topology of wireless networks and investigate the application of graph neural networks in wireless communications. In Part II, we calibrate the input of classic linear algorithms in wireless communications with neural networks. In Part III, we replace the modules in conventional iterative algorithms with deep neural networks. We will demonstrate the advantage of the proposed methods in terms of scalability, computational complexity, sample complexity, and generalization errors both theoretically and empirically. Our techniques and insights go beyond wireless communication and have broad applications, which shall be presented in Part IV.