In the era of next-generation communications, or 6G, the ability to provide seamless connectivity to users anywhere and anytime is crucial. While 5G systems have already provided stable and fast communication capabilities for conventional mobile comm...
In the era of next-generation communications, or 6G, the ability to provide seamless connectivity to users anywhere and anytime is crucial. While 5G systems have already provided stable and fast communication capabilities for conventional mobile communications scenarios, the challenge now lies in extending these guarantees to fast-moving objects in vehicular communications and non-terrestrial communications. These scenarios differ from mobile communications as the fast-changing wireless channels caused by fast-moving objects significantly reduce the response time of transceivers. Therefore, suitable algorithms are needed that can offer high performance, low complexity, and fast reconfigurable attributes to perform transceiver operations based on the underlying fast-changing wireless channels.Over the past decade, the development of deep learning (DL) algorithms has transformed our daily lives and enabled new applications. These technologies have also driven the evolution of next-generation communication systems from a communication engineering perspective. DL algorithms can be used as powerful tools to solve problems and create DL-aided communication system designs with low complexity, high performance, and fast reconfigurable attributes. Additionally, as users perform training or inference on DL algorithms, next-generation communication systems are expected to better support those learning processes by utilizing in-network communication and computing resources. After a comprehensive literature survey, we noticed that DL-based communication system designs are particularly well-suited for aiding transceiver designs in vehicular and non-terrestrial communications, especially when conventional optimization-based algorithms encounter imprecise system models or computationally demanding issues. As such, this thesis focuses on developing learning-based transceiver designs to better serve vehicular and non-terrestrial communications, along with framework designs to facilitate the real deployment of these proposed designs. The first part of this thesis presents dedicated designs for vehicular communications, with a focus on vehicle-to-vehicle (V2V) communications in Chapters 2, 3, and 4. The thesis then proposes a platform for system designers to deploy the proposed designs with 6G infrastructure in Chapter 5, aiming to provide next-generation communication systems that better aid the training and inference processes of users. To handle the communication and computation costs in the learning process, the proposed framework can dynamically orchestrate resources among heterogeneous physical units, including in-network mobile devices, edge and cloud computing centers, to efficiently fulfill deep learning objectives. The framework can also dynamically allocate resources and intelligently assign communications and computation tasks to these units. The second part of this thesis extends these designs to satellite communications, taking into account the special channel features in this context. Using the extra information provided by the geometry relationship of satellite communication scenarios, this thesis discusses the satellite transceiver designs incorporating the geometry relationship to achieve better different communication quality of service indicators in Chapters 6 and 7. Through the proposed designs, this thesis aims to provide next-generation communication systems that offer greater efficiency to users, enabled by the power of advanced DL-based algorithms. Additionally, the thesis aims to present next-generation communication system designs that facilitate training and inference processes when users deploy learning algorithms in distributed environments.