基于深度强化学习的车联网资源管理

Internet of Vehicles Resource Management Based on Deep Reinforcement Learning

车辆对一切(Vehicle to Everything,V2X)是智能交通系统的一项关键技术,可以提高道路安全和交通效率。为降低网络干扰,满足动态环境下不同车辆链路的服务质量(Quality of Service,QoS)要求,提出基于指纹深度强化学习(Deep Reinforcement Learning,DRL)框架的车联网频谱分配和功率控制方案。首先通过随机网络的数据包到达过程分析,V2V链路的时延约束被转化为数据速率约束;然后,各个智能体利用深度神经网络来近似部分观测到的最优资源分配方案映射。同时为避免多智能体环境下的非平稳性,提出使用基于指纹的方式来提高学习稳定性。与其他算法相比,提出的算法能够有效提升网络中车对基站(Vehicle to Infrastructure,V2I)总用户信道容量,同时具有很好的稳定性。

A spectrum allocation and power control scheme based on the Deep Reinforcement Learning(DRL)framework for vehicular networking is proposed in this paper.Firstly,the delay constraint of V2V links is transformed into a data rate constraint by analyzing the packet arrival process of random networks.Then,individual intelligences use deep neural networks to approximate the partially observed optimal resource allocation scheme mapping.Also,to avoid the non-smoothness of the multi-intelligent environment,a fingerprint-based approach is proposed to improve the stability of learning.Compared with other algorithms,the algorithm proposed in this paper can effectively improve the total user channel capacity of Vehicle to Infrastructure(V2I)in the network with good stability at the same time.