Deep Reinforcement Learning Powered IRS-Assisted Downlink NOMA

Abstract

In this work, we examine an intelligent reflecting surface (IRS) assisted downlink non-orthogonal multiple access (NOMA) scenario intending to maximize the sum-rate of users. The optimization problem at the IRS is quite complicated, and non-convex since it requires the tuning of the phase shift reflection matrix. Driven by the rising deployment of deep reinforcement learning (DRL) techniques that are capable of coping with solving non-convex optimization problems, we employ DRL to predict and optimally tune the IRS phase shift matrices. Simulation results reveal that the IRS-assisted NOMA system based on our utilized DRL scheme achieves a high sum-rate compared to OMA-based one, and as the transmit power increases, the capability of serving more users increases. Furthermore, results show that imperfect successive interference cancellation (SIC) has a deleterious impact on the data rate of users performing SIC. As the imperfection increases by ten times, the rate decreases by more than 10%.