Online Task Offloading with Bandit Learning in Fog-assisted IoT Systems

In fog-assisted IoT systems, to achieve best quality of service with ultra-low latency, resource-limited IoT user nodes may offload some tasks to nearby fog nodes, a.k.a. task offloading, to accelerate their processing. However, it remains non-trivial and challenging to decide when and which fog node to offload to. If offloaded, user tasks may experience unexpectedly long latency in face of system uncertainties, such as wireless channel dynamics, variety in task processing time, and resource contention on fog nodes. Moreover, feedback signals such as processing latency can be delayed and even go outdated due to non-stationarity, thereby degrading the effectiveness of system statistic learning and decision making. In this paper, we study task offloading problem for fog-assisted IoT systems in a non-stationary environment with delayed feedback. By leveraging a drift detector and queue methods, we propose TOS-BB and TOS-BS, two online task offloading schemes with bandit learning that endeavor to achieve ultra-low task latency. Simulation results show that both schemes outperform the benchmark while achieving close-to-optimal performance with short task latency.