Evaluation of Multi-armed Bandit-based Online Machine Learning Model Selection for Service Time Prediction in Vehicular Edge Computing

Vehicular edge computing (VEC) systems have recently come to be recognized as a vital part of the computing infrastructure necessary to support the vast array of applications suggested by smart and connected vehicles. These systems utilize a combination of edge and cloud computing resources to undertake intensive computational tasks that have been offloaded from various automotive applications. In order to better steer the offloading decisions, the associated engines need to accurately predict the excepted service time on each potential server. Service time prediction would help the offloading engine to determine the best server that can fulfill the computation requirements of an offloaded task. Due to the dynamically changing status of the VEC resources, the best prediction model may not be apriori known. Hence, this paper investigates the use of multi-armed bandit-based online machine learning model selection for service time prediction in VEC systems. The model selection algorithms include $\epsilon -$Greedy, decayed $\epsilon -$Greedy and upper confidence bound (UCB) algorithms. Evaluation results demonstrate the ability of these algorithms to dynamically pick the most suitable incremental learning model for service time prediction. In addition, the decayed $\epsilon -$Greedy has shown a noticeable ability to consistently choose the most accurate machine learning model and avoid using poorly-performing models. Hence, the considered model selection algorithms provide viable tools to augment task offloading engines in VEC environments.