一种基于强化学习的车联网边缘计算卸载策略

计算密集型、时延敏感型车载应用的不断涌现导致资源受限的车载终端设备无法在短时间内处理大量的应用任务，而且卸载节点的动态变化特性在复杂多变的车联网场景中会导致任务候选卸载节点存在不确定性。针对上述问题，提出一种基于强化学习的计算卸载策略来实现任务卸载预判和计算资源分配。结合设备链接时间与通信半径等因素制定卸载节点发现机制，通过考虑时延与成本对车联网移动边缘计算卸载系统的影响建立效用函数，并以最大化效用作为优化目标将车联网中的卸载问题转化为优化问题，基于卸载节点发现机制采用Q-learning方法提出一种智能节点选择卸载算法求解优化问题，实现任务的智能卸载。仿真结果表明，在车联网场景中，提出的计算卸载策略可实现更高的系统效用。

A reinforcement learning-based offloading strategy for Internet of vehicles edge computing

With the emergence of computing intensive and delay sensitive on-board applications, resource constrained on-board terminal devices cannot meet the needs of processing a large number of application tasks in a short time. Moreover, the dynamic change characteristics of offloading nodes in the complex and changeable Internet of vehicles will lead to the uncertainty of task candidate offloading nodes. To solve the problems, we propose a computational offloading strategy based on reinforcement learning to realize task offloading prediction and computational resource allocation. This strategy first develops an offload node discovery mechanism based on factors such as device link time and communication radius. Secondly, the utility function is established by considering the impact of delay and cost on the offloading system of the mobile edge computing of the Internet of vehicles. The unloading problem in the Internet of vehicles is transformed into an optimization problem with the maximization of utility as the optimization goal. Finally, based on the offloading node discovery mechanism, the Q-learning method is used to propose an intelligent node selection offloading algorithm to solve the optimization problem and realize the intelligent offloading of tasks. The simulation results show that the proposed computing offloading strategy can achieve higher system utility in the Internet of vehicles.