| dc.description.abstract | Massive connectivity, ultra-low latency, and high data rate are some of the fundamental requirements of the upcoming sixth-generation (6G) wireless networks. In this regard, non-orthogonal multiple access (NOMA) has been widely envisioned as a promising candidate for 6G due to its potential of achieving high spectral efficiency. By multiplexing the signals in the power or the code domain, NOMA allows multiple users to be served with the same orthogonal resources, such as frequency and time, hence outperforming the conventional orthogonal multiple access (OMA) systems with a significant spectral efficiency gain. The promising advantages of NOMA cannot be realized without proper optimization designs, of which the complexity escalates in the context of massive connectivity. As a remedy, artificial intelligence (AI) is capable of performing high-dimensional optimization at a lower computational complexity compared to the conventional iterative approaches. Hence, this thesis attempts to utilize AI technologies, including deep learning (DL) and deep reinforcement learning (DRL), to design systematic treatments for NOMA, from the uplink active user detection to the resource allocation in adaptive next-generation multiple access (NGMA) networks, to its combination with reconfigurable intelligent surfaces (RISs), as well as its application in multi-RIS aided device-to-device (D2D) networks. First, this thesis investigates the application of generative neural networks for joint user activity and data detection in uplink NOMA networks. A generative neural network-enabled multi-user detection (MUD) framework is proposed, which outputs signal reconstructions in a fixed and small number of steps with low error rates, based on a low-complexity neural network. Moreover, a non-iterative sparsity estimator is provided to realize sparsity-blind MUD and is compatible with most existing MUD algorithms. Second, to maximize the long-term sum data rate of NGMA networks with energy limitations, DRL is employed to jointly design beamformers, power allocations, and user clustering strategies. To transform the non-trivial mixed-integer problem, a spatial correlation-based user clustering approach is proposed, which achieves higher sum rates compared to the existing channel condition-based clustering approach. To solve the formulated problem, the trust region policy optimization (TRPO) algorithm is employed, which demonstrates robust convergence under large learning rates and realizes a fast and stable training process. Third, the integration of NOMA and RIS is examined, where the sum rate maximization performance of DL and DRL are investigated and compared from both short-term and long-term prospects. By utilizing model-agnostic-meta-learning (MAML), the DL method benefits from a low complexity network and a fast convergence rate. The DRL method, on the other hand, demonstrates superior sum rate performance, especially in the long term. Fourth, this thesis addresses the sum rate maximization problem in multi-RIS assisted NOMA empowered D2D networks. The long-term dynamic optimization problem is reformulated into a Markov game (MG) and a multi-agent deep reinforcement learning (MADRL)-based framework is proposed to jointly learn sub-channel assignments, power allocations, and phase shifts, in a centralized training and decentralized execution (CTDE) manner. Furthermore, the mixed-integer action space is directly addressed by adopting multi-pass deep Q networks (MP-DQNs). | en_US |
