Joint Latency and Energy Optimization of Online Task Offloading Algorithms for Vehicular Edge Computing

With the advantages of low latency and easy deployment, vehicle edge computing can effectively solve the task offloading problem for vehicles as well as surrounding pedestrians. However, due to hardware device limitations, vehicle edge nodes often have limited computing power and need to draw on the free resources on other surrounding vehicles. And the mobility of the vehicle also poses a challenge for task offloading. A common solution to this type of problem is Multi-Arm Bandit (MAB) theory, but existing algorithms based on MAB are not applicable to the dynamic vehicle environment. In this paper, we design an online learning-based vehicle task offloading algorithm (OLUCB), which performs task offloading in a distributed manner without frequent acquisition of vehicle state information, and adds three online capabilities of input-awareness, occurrence-awareness and speed-awareness to adapt to dynamic environments. The algorithm learns the task offloading performance of the surrounding vehicles and converges quickly to the optimal helper vehicle (HEV) with guaranteed performance. Extensive experiments in realistic highway scenario have been conducted, and results illustrate that the OLUCB achieves low latency and low energy performance. Compared to existing task offloading algorithms based on adaptive learning, the average latency, average energy consumption, average cost and average bit cost are reduced by 3%, 51%, 43% and 42% respectively.