Offloading Tasks to Unknown Edge Servers: A Contextual Multi-Armed Bandit Approach

Abstract

Mobile Edge Computing (MEC), envisioned as an innovative paradigm, pushes resources from the cloud to the network edge and prompts users to offload computation-intensive and data-intensive tasks to edge servers for meeting the stringent service requirements. Prior approaches often study efficiently offloading tasks with given system information, though rigorously time-sensitive tasks offloading problems receive less attention under system uncertainty. As motivated, we propose a multi-user collaborative offloading model where users jointly decide time-sensitive task placement while considering the unknown system information and contexts. We formulate the offloading problem as a Multi-user Contextual Combinatorial Multi-armed Bandit (MCC-MAB) problem and propose a learning algorithm Context-Aware Task Offloading Decision (CATOD) to explore the system uncertainty. Since the time-sensitive task offloading problem with learned system information is still NP-hard, we present an approximation algorithm Offline Generalized Task Assignment (OGTA) to obtain an efficient offloading solution. Additionally, meticulous theoretical analysis and extensive evaluations demonstrate the significant performance on a real-world dataset.