Reinforcement Learning Based Offloading for Realtime Applications in Mobile Edge Computing

Energy consumption is one of the most important issues for mobile devices such as smartphones and laptops. For mobile devices that execute multiple computation-intensive or delay-sensitive applications simultaneously, Mobile Edge Computing (MEC) based offloading provides a promising solution to the energy problem. However, blindly offloading all tasks to MEC servers is not the best choice because transferring a simple task to a MEC server via wireless networks might consume more energy than processing the task locally. In addition, Dynamic Voltage and Frequency Scaling (DVFS) could be utilized to reduce the energy consumption associated with locally processed tasks by appropriately lowering CPU frequency. In this paper, we propose a realtime reinforcement learning based offloading scheme, RRLO, which is based on both MEC-based offloading and DVFS-based energy consumption reduction. Technically, RRLO jointly learns the optimal offloading policy and DVFS-based scheduling method. Depending on the workload and network condition, RRLO not only determines whether a task should be offloaded to a MEC server, but also selects the best DVFS method used to schedule local tasks. Our simulation results indicate that RRLO outperforms the existing MEC-based offloading schemes.