车联网是一种由具有通信、计算、存储、学习能力的移动车辆组成的智能网络 系统，可以提升道路安全，改善道路交通管理状况并支持沉浸式的用户体验。但是， 对于资源受限的车辆而言，如何处理车辆传感器搜集到的海量数据，以及为需要极大计算能力或超低时延的应用提供服务，是目前车联网领域面临的重大挑战。移动边缘计算作为一种新兴的网络技术，可以有效地提高计算和卸载效率，将其与车联网相结合，能够解决一部分上述问题。然而，实际的车联网环境较为复杂，部分场景尚未被考虑。对此，本文针对两种实际的车联网环境，设计了对应的车联网边缘 计算卸载方案，主要工作如下： 

1. 在农村道路或者没有配备移动边缘计算服务器的公路场景，当车辆有大量任务需要被处理时，如何有效利用附近闲置的车辆计算资源，是需要被考虑的首要问 题。因此本文建立了一个面向任务执行效率的多跳任务卸载决策模型，主要包括候选车辆的选择机制与任务的卸载决策算法设计两个部分。在候选车辆的选择机制部分，任务车辆通过无线通信，获取 跳通信范围内邻居车辆的具体信息，然后，考虑车辆之间的联通时间与任务的传输时间，计算得到邻居车辆的最大任务处理能力， 进一步确定有资格参与任务卸载的车辆。在卸载决策算法设计部分，考虑任务完成的效率，以完成所有任务所需的时延最小作为优化目标，将计算卸载问题建模为一个带有约束的广义分配模型，并且分别借助贪心算法和离散蝙蝠算法进行求解。实验结果表明，这两种求解方法均具有时延性能上的优越性，且基于离散蝙蝠算法的 求解方法得到计算卸载方案的性能更优。

2. 车辆能够与移动边缘计算服务器通信时，车辆可以考虑把任务卸载到边缘节点去完成。但是，动态到达的任务具有不确定性，如何直接根据车联网环境的原始信息获取最优的卸载策略，使得车联网边缘计算系统在一定时间内获得最大的长期 奖励，是另外一个需要被考虑的问题。针对这一问题，本文提出了一种面向任务执行效率与公平性的动态计算卸载方案，使用深度强化学习中的异步优势演员评论家 算法，将移动边缘计算服务器的可用资源作为系统状态，卸载策略作为动作，建立马尔可夫决策过程，然后将相对效率因子与相对公平性因子的加权值作为目标函数 设计奖励。仿真结果表明，本文提出的方案能够进行在线策略更新，在模型收敛后完成所有任务的平均时延较低，任务间公平性较高，并且能够较好地适应复杂的环境变化。  

The Internet of Vehicles is an intelligent network system composed of mobile vehicles, which have the ability of communication, computing,storage, and learning.It can enhance the road safety,improve the level of the road traffic management, and support immersive user experience. However, for resource-limited vehicles, how to deal with the huge amount of data collected by vehicles’ sensors and serve applications that require great computing power or ultra-low latency is a major challenge for the field of Internet of Vehicles today. Mobile edge computing, as an emerging network technology, can effectively improve the computation and offloading efficiency.Combining it with the Internet of Vehicles can solve some of these problems. However, the actual Internet of Vehicles environment is more complex, and some scenarios have not been considered. In this regard, this article focuses on two kinds of actual problems in the vehicular network environment and then designed corresponding vehicle computing offloading schemes, and the main work is as follows:

1. On rural roads or highway scenarios without mobile edge computing servers, when vehicles have a lot of tasks to be processed, how to effectively utilize the nearby computing resources of idle vehicles is the primary issue that needs to be considered. Therefore, a multi-hop task offloading decision model based on task execution efficiency is proposed in this article. It includes two parts: candidate vehicles selection mechanism and the design of the task offloading decision algorithm.In the part of candidate vehicle selection mechanism, the task vehicle obtains the detailed information of neighboring vehicles within the k-hop communication range through wireless communication, and then, the maximum task processing capability of neighboring vehicles is calculated by considering the connection time between vehicles and the task transmission time to further determine the vehicles eligible to participate in the task offload. In the offloading decision algorithm design part, the article considers the task completion efficiency, takes the minimum delay required to complete all tasks as the optimization goal, and models the computing offloading problem as a generalized assignment model with constraints.Then it is solved with the help of the greedy algorithm and the discrete bat algorithm respectively. The experimental results show that both methods have superior performance in terms of delay, and the solution method based on the discrete bat algorithm yields better performance.

2. When the vehicle is able to communicate with the mobile edge computing server, the vehicle can consider offloading the task to the edge node to complete it. However, the arrival of tasks is uncertain, how to directly get the optimal offloading strategy from the original information of the Internet of Vehicles environment and make the vehicle edge computing system obtain the maximum long-term reward is another issue that needs to be considered.To address this problem,this article proposes a dynamic computing offloading scheme based on the task execution efficiency and fairness. The scheme uses asynchronous advantage actor-critic algorithm of deep reinforcement learning.It takes the available resources of the mobile edge computing server as the system state and the offloading strategy as the action to establish a Markov decision process, and then uses the weighted value of the relative efficiency factor and the relative fairness factor as the objective function to design rewards. The simulation results show that the strategy proposed can be updated by the on-policy method. After the model converges, the average delay for completing all tasks is low, the fairness between tasks is high, and it can better adapt to complex environment changes.