Deep Reinforcement Learning for Task Offloading in Edge Computing

Since task offloading can reduce energy consumption and execution latency in edge computing, it becomes increasingly attractive in the areas of heterogeneous network edge computing and high-reliability-low-latency communications due to its high energy-efficiency and low execution-latency. Recently, by exploiting powerful adaptive capabilities, reinforcement learning has been successfully applied in task offloading. However, the dynamic changes in computation, storage, and network resources in edge computing lead to the time-varying nature of computing environment. Fortunately, deep reinforcement learning (DRL) has excellent potential for time-varying task offloading in edge computing. In this article, we provide a novel task offloading method based on DRL in edge computing with the computational model architecture of edge computing and DRL solutions for task offloading. We also identify some potential applications of DRL in task offloading and present several key techniques to improve offloading decision performance with fusing computing, storage, and network resources. The future research directions and challenges have been discussed at the end of this article.