부산대학교 도서관

Cooperative multi-agent reinforcement learning (MARL) is a promising solution for many large-scale multi-agent system (MAS) scenarios. A MARL framework is usually based on a decentralized scheme that enables communication between all agents in a given architecture. The agents exchange information to maximize their average reward and increase the overall system performance. However, this decentralized information sharing results in high communication costs, which is a critical issue for environments with limited communication bandwidth. On the other hand, a predefined inter-agent communication architecture may limit potential cooperation. This paper addresses such issues in a vehicle-to-everything (V2X) network, a typical example of MAS with strict Quality of Service (QoS) requirements. For efficient utilization of limited network resources, a solution to the resource-sharing problem between Vehicle to Infrastructure (V2I) and Vehicle to Vehicle (V2V) links is required. We propose a POST-Attentional Communication Actor-Critic (POST-2AC) model that learns when communication is needed and how to integrate shared information for cooperative decision-making. Our learning method uses an attention approach combined with the critic-network to label the agents local information based on its importance so that each agent learns to trade off its performance and communication cost. The simulation results show that the proposed model achieves better performance than the state-of-the-art baselines.