Scalable Multi-Agent Reinforcement Learning-Based Distributed Channel Access.

With the rapid development of smart devices, the next generation wireless networks (NGWNs) are expected to achieve high access efficiency in a high-dynamic scenario. To tackle the above challenges in NGWNs, this paper proposes a new MAC protocol, MAAC-advanced Listen-Before-Talk (MLBT), which employs multi-agent reinforcement learning (MARL) algorithm. As a MARL paradigm, centralized training with decentralized execution (CTDE) is confronted with the scalability issue. To address it, we design a scalable neural network architecture based on the attention mechanism, which can cope with the varying number of stations. Moreover, a novel reward function is designed to achieve the max-min fairness and maximum aggregate network throughput simultaneously. Extensive simulation experiments are provided to show that MLBT approaches the optimal performance and accelerates the centralized training process when stations join or leave the network.