Multi-Agent Reinforcement Learning-Based Decentralized Spectrum Access in Vehicular Networks With Emergent Communication

In this letter, we propose a novel decentralized spectrum access algorithm based on the multi-agent reinforcement learning (MARL) for cellular vehicle-to-everything (C-V2X) networks. The agents make decisions independently with the global objective of maximizing vehicle-to-infrastructure (V2I) users' sum throughput while meeting vehicle-to-vehicle (V2V) users' latency and reliability requirements. Specifically, to achieve better collaboration among agents we introduce the inter-agent communication mechanism into MARL. So, in the proposed algorithm each agent consists of an action selector module and a message generator module, i.e., the agents learn emergent communication via an additional dedicated channel to enable explicit collaboration, which helps with learning effective policies for channel access. Simulation results show the effectiveness of the proposed algorithm, in improving both V2I users' throughput and V2V users' packet delivery ratio.