Adaptive Bayesian Optimization for Online Management in Mobile Edge Computing

Jia Yan, Qin Lu, Konstantinos D. Polyzos

Abstract: Mobile edge computing (MEC) has emerged as a key architecture in the Internet of Things (IoT) that allows the otherwise power-limited wireless devices (WD) to carry out high-performance computation. This work considers a typical one-server multi-user MEC system, where each WD, with a computational task to excute, can divide its tasks for local and edge computing. Such partial task offloading decisions will be selected together with other resource allocation variables by minimizing the overall energy-delay cost (EDC). In the presence of unknown system model, the EDC function is not available in analytical form, and instead only the function values at queried points are revealed. Towards solving such a bandit optimization in a sample-efficient manner, we will rely on the Bayesian optimization framework, that relies on a surrogate model, typically the Gaussian process (GP), to actively select the query points. Coping with the possibly dynamic learning environment, an ensemble (E) of GP models with data-adaptive weights will be leveraged for surrogate modeling, following which acquisition of next evaluation point will be selected using the so-termed Thompson sampling (TS) rule. Numerical tests validate the merits of the proposed EGP-TS approach relative to existing alternatives under a practical MEC setting.