Liu Qingli, Ma Yongjie
Key Laboratory of Communication and Network, Dalian University, Dalian, 116622, China.
School of Information Engineering, Dalian University, Dalian, 116622, China.
Sci Rep. 2025 Aug 22;15(1):30866. doi: 10.1038/s41598-025-15982-x.
In highly dynamic vehicular networking scenarios, when Vehicle-to-Infrastructure links and Vehicle-to-Vehicle links share spectrum resources, the traditional distributed resource allocation method lacks global optimization and fails to respond to environmental changes in a timely manner, which leads to low spectral efficiency of the system. A resource allocation method based on federated multi-agent deep reinforcement learning is proposed for Vehicular Networking communication, by fusing Asynchronous Federated Learning (AFL) and Multi-Agent Deep Deterministic Policy Gradient (MADDPG). Synergistic optimization of resource allocation. First, vehicles as agent dynamically optimize spectrum access, power control, and bandwidth allocation based on local channel states through the collaborative policy of MADDPG to reduce cross-link interference. Second, the asynchronous federation architecture is designed, where vehicles independently upload local model parameters to the global server, dynamically adjust the aggregation weights according to the real-time channel quality, and optimize the update of global model parameters. Finally, the global model parameters are fed back to the vehicles to further optimize the local resource allocation strategy, thus improving the system spectrum efficiency. The simulation results show that the system spectrum efficiency is improved by 19.1% on average compared with the centralized DDPG, MADDPG, MAPPO and FL-DuelingDQN algorithms in the Vehicle Networking scenario, while the transmission success rate of the V2V link is improved by 9.3% on average, and the total capacity of the V2I link is increased by 16.1% on average.
在高度动态的车载网络场景中，当车与基础设施链路和车与车链路共享频谱资源时，传统的分布式资源分配方法缺乏全局优化，无法及时响应环境变化，导致系统频谱效率低下。针对车载网络通信，提出了一种基于联邦多智能体深度强化学习的资源分配方法，该方法融合了异步联邦学习（AFL）和多智能体深度确定性策略梯度（MADDPG），对资源分配进行协同优化。首先，车辆作为智能体基于本地信道状态，通过MADDPG的协同策略动态优化频谱接入、功率控制和带宽分配，以减少交叉链路干扰。其次，设计了异步联邦架构，车辆独立将本地模型参数上传到全局服务器，根据实时信道质量动态调整聚合权重，优化全局模型参数的更新。最后，将全局模型参数反馈给车辆，进一步优化本地资源分配策略，从而提高系统频谱效率。仿真结果表明，在车载网络场景中，与集中式DDPG、MADDPG、MAPPO和FL-DuelingDQN算法相比，该系统频谱效率平均提高了19.1%，同时车对车链路的传输成功率平均提高了9.3%，车对基础设施链路的总容量平均增加了16.1%。
