Meta Reinforcement Learning for Generalized Multiple Access in Heterogeneous Wireless Networks

This paper focuses on spectrum sharing in heterogenous wireless networks, where different nodes utilize various Media Access Control (MAC) protocols to transmit data packets to a common access point on a shared wireless channel. Previous studies have developed Deep Reinforcement Learning (DRL) based multiple access protocols for specific scenarios within heterogeneous wireless networks. However, there exists a wide range of coexisting scenarios, characterized by varying numbers of nodes and the use of different MAC protocols. Existing approaches require training new models from scratch when encountering unseen scenarios, resulting in significant training time. To address this issue, we propose a novel MAC protocol called Generalized Multiple Access (GMA), which employs the Meta-Reinforcement Learning (meta-RL) algorithm. By learning a meta-policy during training, GMA enable the fast adaptation of the agent node to different and previously unknown heterogeneous network environments, without prior knowledge of the specific MAC protocols used in those environments. We conduct a performance comparison between the proposed GMA protocol and existing DRL-based protocols. Simulation results demonstrate that while the GMA protocol experiences a slight performance loss compared to baseline methods in training environments, it demonstrates faster convergence and higher performance in new environments compared to baseline methods.