부산대학교 도서관

Reinforcement learning is an effective tool to address the resource allocation problem in vehicle-to-everything (V2X) communications due to its capability to deal with strong dynamics and diverse quality-of-service requirements in vehicular networks. However, collecting reinforcement learning data relies on many interactions between agents and the environment, which often comes with high costs. So this paper mainly explores the offline reinforcement learning method to solve the V2X resource allocation problem and proposes two schemes based on the batch constraint Q-learning (BCQ) algorithm and the conservative Q-learning (CQL) algorithm. We build the simulation environment of the IoV, use the online single-agent algorithm to collect datasets in different forms, and verify the performance of each algorithm on different datasets. Simulation results show that the offline resource allocation scheme based on BCQ and CQL outperforms the offline DQN scheme. The two offline algorithms have good robustness and are suitable for practical applications.