"MIMO NOMA RL Benchmark 2026"

"We presents MIMO-NOMA-RLBench, a benchmark dataset to assess reinforcement learning (RL) algorithms on joint beamforming and user scheduling in multi-cellmultiple-input multiple-output non-orthogonal multiple access (MC-MIMO-NOMA) networks. The dataset consists of 281 experimental settings with seven RL agents (GAI-PDPPO, PPO,SAC, TD3, DDPG, MAPPO and Decision Transformer), ten evaluation dimensions, and 843,000 performance data points across 281,000 training episodes. They provide 562 pre-trainedmodel checkpoints, allowing direct comparison of algorithms without re-training. The data set includes parameter sweeps with interference, SINR, antenna configurations (up to Nt=128), userdensities (K=4\u201312), channel estimation errors, reconfigurable intelligent surface (RIS) elements, network sizes (M =3\u201319 base stations), and multi-agent coordination scenarios. The simulationenvironment that comes with it, which is a vector-like interface compatible with OpenAI Gym, allows reproducible experimentation of automatic channel generation, SINR calculation, andconstraint evaluation. All data are trained on consumer-grade hardware (NVIDIA RTX 3060, 12 GB), totaling 218 GPU hours of training, and are structured for immediate use inbenchmarking novel RL algorithms for next-generation wireless resource allocation. The code, dataset, and trained models are made publicly accessible to speed up research in AI-basedwireless communications"