A Fully Independent MARL for Collision Avoidance in Distributed Channel Access

Abstract

This paper proposes a fully independent multi-agent reinforcement learning (MARL) approach for distributed channel access (DCA) in wireless networks. The proposed scheme enables each device to be trained in a completely independent manner without utilizing any joint states or joint actions throughout the training phase. This maximizes the overall throughput and ensures fairness among users while keeping all the agents fully independent. Simulation results show that our proposed method outperforms the random access frameworks while incurring low computational overhead.