Federated Meta-RL Based Multiple Access Protocol for Diverse Heterogeneous Wireless Networks

This paper addresses the challenge of efficient spectrum utilization in heterogeneous wireless networks, where nodes employ various Media Access Control (MAC) protocols to transmit data packets to a common access point over a shared wireless channel. Previous research has developed multiple access protocols utilizing Deep Reinforcement Learning (DRL) to address specific scenarios within these networks. However, due to the trial-and-error learning process inherent in DRL, these protocols require training new models from scratch when faced with unseen scenarios, limiting their applicability. To tackle the generalization issue while addressing data privacy concerns, we propose a novel MAC protocol named Federated Generalized Multiple Access (Fed-GMA), which integrates Federated Learning (FL) with a meta-Reinforcement Learning algorithm. The Fed-GMA protocol enables agents across various environments to collaboratively train a meta model without sharing data, which possesses the capability to rapidly adapt to new environments. We evaluate the performance of the proposed Fed-GMA protocol against existing DRL-based protocols. Simulation results show that Fed-GMA significantly enhances the generalization ability of DRL protocols, achieving faster convergence and better performance in both training and new environments.