Article Preview
TopIntroduction
With total vehicle ownership rising, problems such as road congestion, traffic safety, and environmental pollution are becoming increasingly prominent. The development of autonomous driving technology provides new solutions to these problems. Highly automated vehicle driving can greatly reduce accidents caused by human errors, improve the convenience and comfort of driving, and improve the overall traffic efficiency and safety of the road system. In addition, it can significantly improve fuel efficiency, reduce the environmental pollution caused by vehicle exhaust emissions, and alleviate traffic congestion.
When the sensing system of autonomous vehicles detects the current traffic conditions, the vehicle reasoning system needs to process the traffic conditions transmitted by the sensing system in a short time. Soon, the vehicle's transmission system can make response actions such as shifting, steering and shifting (Rummery & Niranjan, 1994). Due to the lack of computing power of OBU, it is an important problem to choose a reasonable scheduling platform and optimize the completion time of autonomous reasoning task under the fault tolerance time constraint.
Edge computing utilizes edge devices with computational and storage resources located between the terminal device and the cloud data center. Its purpose is to unload the calculation task to the edge server near the end user, and process the calculation task near the data source, so as to shorten the response time. The decomposition and scheduling of reasoning tasks in the edge environment can well meet the low latency requirements of reasoning tasks, and produce good real-time processing effects.
For the reasoning task of autonomous driving, little attention is paid to the scheduling problem of reasoning task, most of the research focuses on the traffic condition recognition and reasoning analysis in the process of autonomous vehicle driving (Behere & Torngren, 2015; Chen, Wu, Han, & Xiao, 2007; Fang et al., 2019). The task scheduling of autonomous driving is similar to the task scheduling and workflow scheduling of collaborative design. Therefore, the research scheme of collaborative design task and workflow scheduling can be applied to autonomous driving workflow scheduling. Because reasoning tasks are usually scheduled on OBU (Gao, Sun, & He, 2019), which will lead to high delay in reasoning task scheduling, autonomous vehicles can not meet the security and real-time requirements of reasoning tasks. At present, heuristic algorithms are widely used to solve collaborative design tasks and workflow scheduling problems (Ghorbannia Delavar & Aryan, 2014; Gu, Lillicrap, Sutskever, & Levine, 2016; Guo, Wang, Sun, & Liu, 2002; Kennedy & Eberhart, 1995; Lin et al., 2019), such as ant colony algorithm (ACA), particle swarm algorithm (PSO), genetic algorithm (GA), etc. Although these algorithms provide good solutions for workflow scheduling problems and collaborative design tasks, they have slow convergence speed and are easy to fall into local optimal solutions, so that they cannot be satisfied very well Low delay requirement of reasoning task. In recent years, reinforcement learning (RL) has been applied to solve the scheduling problem of collaborative design task and workflow (Meena, Kumar, & Vardhan, 2016; Orhean, Pop, & Raicu, 2018; Sutton, 1988). The algorithm corrects the deviation between the real value and the predicted value through the interaction with the uncertain environment, and the convergence speed is faster. These studies provide a new solution for autonomous driving workflow scheduling. Although the structure of reasoning task is similar to collaborative design task and workflow, the reasoning task generated by self driving needs to meet the low latency requirements of real-time system. Therefore, it is necessary to design appropriate scheduling strategies and algorithms to meet the requirements of autonomous driving reasoning tasks.