Multi-Agent Reinforcement Learning-Based Distributed Channel Access for Next Generation Wireless Networks

In the next generation wireless networks, more applications will emerge, covering virtual reality movies, augmented reality, holographic three-dimensional telepresence, haptic telemedicine and so on, which require the provisioning of high bandwidth efficiency and low latency services. In order to better support the aforementioned applications and services, novel distributed channel access (DCA) schemes are necessary. Therefore, we propose a new MAC protocol, QMIX-advanced Listen-Before-Talk (QLBT), based on the cutting-edge multi-agent reinforcement learning (MARL) algorithm. It employs a centralized training with decentralized execution (CTDE) framework to exploit the overall information of all agents during training, and ensure that each agent can independently infer the optimal channel access behavior based on its local observation. We enhance QMIX, a well-known MARL algorithm, by introducing an extra individual Q-value for each agent in the mixing network apart from the original total Q-value, which makes QLBT more stable. Moreover, delay to last successful transmission (D2LT) is first introduced in this work as a part of the observations of each QLBT agent, which facilitates agents to reach a cooperative policy that prioritizes the agent with the longest delay. Finally, extensive simulation experiments are provided to show that the proposed QLBT algorithm: 1) outperforms CSMA/CA and even its theoretical performance bound in various scenarios including saturated traffic, unsaturated traffic and delay-sensitive traffic; 2) is robust in dynamic environment; and 3) is able to friendly coexist with "legacy" CSMA/CA stations.