Multi-agent Deep Reinforcement Learning with Hybrid Action Space for Resource Allocation of Vehicular Networks

In this paper, we investigate the joint optimization of spectrum selection and power allocation between vehicle-to-infrastructure (V2I) and vehicle-to-vehicle (V2V) users in vehicle-to-everything (V2X) networks. We propose a multi-agent deep reinforcement learning (MA-DRL) framework to improve the performance of V2I and V2V links, where each V2V link regard as an agent and all agents interact with the environment together. Specifically, each agent uses double deep Q-network (DDQN) for spectrum selection and deep deterministic policy gradient (DDPG) network for power allocation. The experimental results show that this method can meet the requirements of high channel capacity of V2I link and low delay of V2V link, and show good adaptability when the vehicle speed changes.