Optimization of IRS-NOMA-Assisted Cell-Free Massive MIMO Systems Using Deep Reinforcement Learning

Abstract

We consider the use of multiple intelligent reflecting surfaces (IRSs) in cell-free massive MIMO (CFMM) systems to improve signal quality. However, pilot contamination can occur when multiple users reuse the same pilot sequences, which can lead to performance degradation. To address the issue, a technique called non-orthogonal multiple access (NOMA) is employed in the IRS-assisted CFMM system. For the effective realization of the NOMA technique, it is important to select user pairs to which a successive interference cancellation is applied. To find the optimal user pairing, we propose an optimization algorithm using deep reinforcement learning based on a deep deterministic policy gradient. This algorithm jointly optimizes the phase shift of all IRSs, power allocation, and user pairing for NOMA. Numerical results show that optimization of user pairing as well as the phase shifts of IRSs and power allocation plays a crucial role in improving the downlink rate.