Adversarial Multi-Agent RL for Grant-Free NORA Control in Ultra Dense Network under Asymmetric Uncertainty

Grant-free nonorthogonal random access (NORA) enables scalable access in hyper-dense IoT networks by combining ALOHA-type transmission with nonorthogonal multiple access, eliminating base station handshaking and reducing signaling overhead. However, its performance decreases severely under heavy collisions and interference when the user density is high. Although reinforcement learning (RL)-based access control schemes improve adaptability, many existing approaches rely on centralized coordination or computationally heavy models, which suffer to scale in massive access environments. This paper proposes a distributed adversarial multi-agent RL framework for grant-free NORA. By formulating the access control as a zero-sum game, each device independently learns collision-aware access strategies through self-competitive training without coordination or explicit cooperation. The proposed lightweight design supports extreme user-to-slot ratios and and achieves significant gains in throughput and access probability under fading and highly congested conditions.