Artificial Intelligence-Driven Resource Allocation Techniques for Future NOMA Systems

With the unprecedented demand for wireless connectivity, and given the scarce radio resources, the quest for efficient and reliable multiple access (MA) techniques has never been more crucial. Unlike conventional MA techniques, non-orthogonal MA (NOMA) offers superior spectral and energy efficiencies. Particularly, NOMA allows spectrum sharing under controlled circumstances which enables massive connectivity. In addition, by combining NOMA with other techniques such as intelligent reflecting surfaces (IRS), the performance of NOMA is further enhanced. However, combining such sophisticated techniques often leads to highly complex optimization problems given the number of design parameters. Therefore, the conventional optimization-based approach often leads to high computational complexity algorithms that suffer from latency and scalability issues. Therefore, based on the recent advances in machine learning (ML) techniques, this thesis attempts to provide an ML-based alternative for addressing the latency and complexity challenges in the conventional approach. In particular, the reinforcement learning (RL) framework is utilized to solve resource allocation problems in NOMA systems.
Firstly, a robust joint design for an IRS-assisted downlink (DL) NOMA system with imperfect channel state information is considered. To overcome the joint non-convexity of the problem, it is then reformulated as an RL environment, and a twin-delayed deep deterministic policy gradient (TD3) agent is developed to solve the problem. Secondly, to reduce the receiver's complexity, users are clustered in an IRS-assisted DL NOMA system. Next, the beamforming design is proposed through the zero-forcing principle, and a joint robust design of power allocation and IRS phase shifts is proposed based on the TD3 agent. Thirdly, a robust design for energy efficiency (EE) maximization in an IRS-assisted uplink NOMA system is proposed. Moreover, an algorithm is developed based on the soft actor-critic (SAC) agent to jointly optimize power allocation and IRS phase shifts in the long-term EE maximization of the system.