Sum-Rate Maximization for Uplink Multi-User NOMA With Improper Gaussian Signaling: A Deep Reinforcement Learning Approach

This paper investigates the joint allocation of power and circularity coefficient for an uplink multi-user non-orthogonal multiple access (NOMA) system employing improper Gaussian signaling (IGS) in the presence of imperfect successive interference cancellation. We propose two novel deep reinforcement learning (DRL) approaches to address the weighted sum-rate maximization problem under quality of service constraints in this context. Instead of using widely linear transformation to amalgamate power and circularity coefficient into a unified precoding matrix, the two proposed DRL methods, referred to as interdependent deep deterministic policy gradient (I-DDPG) and collaborative DDPG (C-DDPG), both explicitly incorporate the factor of circularity coefficient in the optimization process, thereby enabling full exploitation of the interference management capabilities offered by IGS in the considered uplink NOMA system. Compared to I-DDPG, C-DDPG facilitates both collaborative training and independent adjustments of both decision variables, leading to enhanced generalization ability. Simulation results validate that the proposed DRL approaches achieve significant sum-rate improvement over conventional model-based optimization techniques, in which C-DDPG pushes individual user rates closer to the optimal NOMA boundary. Furthermore, our findings highlight the fact that the sub-optimality of the model-based optimization technique stems from its inability to fully utilize the circularity coefficient, rather than the power.