Learning-Based Channel Estimation and User Pairing Under Imperfect CSI in Uplink FDMA-NOMA Networks

We propose a novel framework combining masked autoencoders (MAEs) with deep reinforcement learning (DRL) for robust user pairing in uplink FDMA-NOMA under imperfect channel state information (CSI) and frequency-selective fading conditions. The MAE reconstructs complex-valued channel matrices from sparsely observed inputs, simulating limited pilot overhead. The reconstructed CSI is used by a DRL agent based on the proximal policy optimization (PPO) algorithm to pair users across all subcarriers. Simulation results on the DeepMIMO O1 dataset show that PPO-based pairing outperforms the Hungarian algorithm under imperfect CSI and performs comparably under perfect CSI, while surpassing random pairing in both cases.