Deep Reinforcement Learning for Scheduling and Offloading in UAV-Assisted Mobile Edge Networks

Abstract

The low deployment cost and high mobility of unmanned aerial vehicles (UAV) have given UAVs widespread attention in mobile edge computing (MEC). Aiming at the characteristics of dynamic changes in computing tasks generated by the urban transportation network, this paper proposes an UAV-assisted edge computing offloading model; i.e., when the number of tasks is within the capacity of edge servers (ES), most of the tasks are processed by ESs, but when the amount exceeds the capacity, some tasks will be offloaded to UAV and processed by on-board servers on UAVs. To minimize the task delay and the cost of computing tasks, this paper formulates a mathematical model according to the task amount generated by the urban transportation network, in which the task delay is computed by using queueing theory. A deep reinforcement learning algorithm with double deep neural networks (DNN) is used to optimize the model and obtains the resource allocation strategy of UAVs and on-board servers and the offloading strategy of edge computing. Simulation experiments verify the effectiveness of these strategies.