A Multi-Agent Reinforcement Learning Approach for Massive Access in NOMA-URLLC Networks

Ultra-reliable low-latency communication (URLLC) enables diverse applications with rigorous latency and reliability requirements. To provide a wide range of services, the future beyond fifth (B5G) systems are expected to support a large number of URLLC users. In this paper, we propose a joint sub-channel allocation and power control method to support massive access for non-orthogonal multiple access aided URLLC (NOMA-URLLC) networks. We model the problem of maximizing the number of successful access users as a multi-agent reinforcement learning problem. A deep Q-network-based multi-agent reinforcement learning (DQN-MARL) algorithm is proposed to tackle the problem while guaranteeing reliability and latency requirements of URLLC services. Simulation results show that the proposed DQN-MARL algorithm significantly improves the successful access probability in massive access scenarios compared with the existing schemes.