Edge-Centric Bandit Learning for Task-Offloading Allocations in Multi-RAT Heterogeneous Networks

Abstract

The exponential growth of data traffic from mobile devices leads to a need of heterogeneous networks (HetNets) which integrate multiple radio access technologies (multi-RATs) to allocate task-offloading with quick coordination. In this paper, we present a novel mobile edge computing (MEC) architecture for multi-RAT HetNets, and propose an MEC-centric offloading decision mechanism. By formulating the intended task as a multi-armed bandit (MAB) problem, we develop a fronthaul-aware upper confidence bound (FA-UCB) algorithm that is able to deal with uncertainty and asymmetry of network state information. It is rigorously established that the proposed FA-UCB algorithm has a sublinear regret bound against the optimal benchmark with full a-priori knowledge, given that the backhaul delays are independently and identically distributed over time. Furthermore, under a restless martingale (RM) bandit condition, we put forth a generalized RM-FA-UCB algorithm that can achieve a sublinear regret bound even under non-stationary network dynamics. Numerical results demonstrate the merits of the proposed schemes and algorithms.