(IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 12, No. 4, 2021 Performance Assessment of Context-aware Online Learning for Task Offloading in Vehicular Edge Computing Systems Mutaz A. B. Al-Tarawneh 1 Computer Engineering Department Faculty of Engineering Mutah Univesity, Jordan Saif E. Alnawayseh 2 Electrical Engineering Department Faculty of Engineering Mutah Univesity, Jordan Abstract—Vehicular Edge Computing (VEC) systems have recently become an essential computing infrastructure to support a plethora of applications entailed by smart and connected vehicles. These systems integrate the computing resources of edge and cloud servers and utilize them to execute computational tasks offloaded from various vehicular applications. However, the highly fluctuating status of VEC resources besides the varying characteristics and requirements of different application types introduce extra challenges to task offloading. Hence, this paper presents, implements and evaluates various task offloading algorithms based on the Multi-Armed Bandit (MAB) theory for VEC systems with predefined application types. These algorithms seek to make use of available contextual information to better steer task offloading. These information include application type, application characteristics, network status and server utilization. The proposed algorithms are based on having either a single MAB learner with application-dependent reward assignment, multiple application-dependent MAB learners or dedicated contextual bandits implemented as an array of incremental learning models. They have been implemented and extensively evaluated using the EdgeCloudSim simulation tool. Their performance has been assessed based on task failure rate, service time and Quality of Experience (QoE) and compared to that of recently reported algorithms. Simulation results demonstrate that the proposed contextual bandit-based algorithm outperforms its counterparts in terms of failure rate and QoE while having comparable service time values. It has achieved up to 73.4% and 21.7% average improvements in failure rate and QoE, respectively, among all application types. In addition, it efficiently utilizes the available contextual information to make appropriate offloading decisions for tasks originating from different application types achiev- ing more balanced utilization of the available VEC resources. Ultimately, employing incremental learning to implement the proposed contextual bandit algorithm has shown a profound potential to cope with dynamic changes of the simulated VEC systems. Keywords—Vehicular edge computing; task offloading; multi- armed bandits; contextual bandits I. I NTRODUCTION Recently, the emergence of smart and connected vehicles has excelled the development of various types of vehicular applications such as infotainment and autonomous driving ser- vices [1], [2]. These applications are usually supported by the on-board computing and storage hardware resources. However, the ever-increasing spectrum of compute-intensive vehicular applications and services has rendered the on-board computa- tional resources inadequate. Hence, Vehicular Edge Computing (VEC) systems have emerged as a baseline for providing high- performance and reliable computing services for in-vehicle applications [3]. In these systems, vehicles, edge servers - instantiated at the road side units (RSUs) and cloud servers can contribute their resources to process computational tasks generated from on-board mobile devices or vehicular driving systems [4]. Hence, computational tasks within VEC systems can be offloaded to any of the available hardware resources to ensure their correct and timely execution. While task offload- ing can enhance task execution and improve user-perceived Quality of Service (QoS), designing an efficient task offloading scheme is not straightforward. First, the VEC environment encompasses different application classes each with different processing demands, network bandwidth requirements, timing constraints and delay sensitivity. Such diversity in application characteristics besides the unpredictable behavior of offload- ing requests will cause the heterogeneous computational and network resources contained in VEC infrastructure to ex- hibit transient and dynamic operational characteristics. These characteristics are mostly related to the utilization levels of available computational servers and the availability of network bandwidth. Second, VEC systems entail the collaboration of various entities such as vehicles, local edge servers and global cloud servers. While such a multi-component environment can lead to more versatility in task offloading, it also increases the state-space of task offloading complicating the decision to select the most appropriate entity to handle an offloaded task [5], [6]. As the dynamic changes to the VEC systems are difficult to predict or model in advance, an efficient offloading scheme should be able to learn while offloading; it should utilize its historical offloading data to steer its future offloading decisions considering both application-salient characteristics and current status of the VEC system [7]. This work targets task offloading in VEC systems with a predefined set of applications with each application having different processing, network bandwidth and timing requirements. The essence of task offloading in such systems is to enable vehicles or offloading decision makers to interact with potential offloading destinations via task offloading, learn their suitability to handle the offloaded tasks and utilize recent offloading history to guide current offloading decisions. As the set of possible offloading destinations in the considered VEC systems remain unchanged, task offloading can be formulated as a multi-armed bandit www.ijacsa.thesai.org 304 | Page