To enable ultra-reliable low-latency communications (URLLC) in IoT applications, this article investigates cross-packet hybrid automatic repeat request assisted non-orthogonal multiple access (XP-HARQ-NOMA) technique through appropriate rate adaptation. The objective is to maximize long-term average throughput (LTAT) by optimizing transmission rates while ensuring outage and throughput guarantees for each device. However, traditional optimization methods struggle with the complexity of XP-HARQ-NOMA and outdated channel state information. To address this, this article proposes a decentralized multi-agent deep reinforcement learning (MADRL) framework based on Markov game theory, which overcomes scalability and non-stationarity issues in multi-device settings. Two MADRL strategies, Cooperative MADRL (CoopMADRL) and Competitive MADRL (CompMADRL), are implemented using the twin delayed deep deterministic policy gradient (TD3) algorithm. Simulations demonstrate that the proposed MADRL methods outperform centralized deep reinforcement learning, with CompMADRL delivering the best performance.
Wang et al. (Mon,) studied this question.