Hybrid Action Space Reinforcement Learning for Vehicular Network Resource Allocation

In vehicular networks, high-frequency broadcast communications place significant strain on wireless resources. Additionally, the high mobility of vehicles results in rapid channel state fluctuations, making centralized resource management at the base station impractical. To address these challenges, this paper proposes a hybrid action space reinforcement learning method for vehicular network resource allocation. We constructed a city-based virtual environment model with a Manhattan layout, where vehicle-to-vehicle (V2V) links reuse spectrum allocated to vehicle-to-infrastructure (V2I) links. As agents, V2V links interact with the communication environment to select optimal actions from a hybrid action space that includes both spectrum and transmission power options. The model applies a suitable reward design and training mechanism to update the network, improving spectrum and power selection. We demonstrate that the proposed DCPRL algorithm can enhance the total capacity of V2I links and increase the transmission success rate of periodic safety messages in V2V links compared to existing approaches.