Date of Award

8-25-2022

Publication Type

Thesis

Degree Name

M.A.Sc.

Department

Electrical and Computer Engineering

Keywords

Computation Offloading;Deep Reinforcement Learning;Energy;Friendly Jammer;Mobile Edge Computing;Security

Supervisor

Huapeng Wu

Abstract

Driven by the popularity of the Internet of Things, the emergence of a multitude of applica- tions highly require computing and storage capacity. However, the Internet of Things users are generally constrained with resources. Mobile edge computing is proposed whereby data generated by Internet of Things devices could be offloaded to nearby edge servers instead of remote cloud servers, thus their requests can be served locally and quickly. One of the most important research issues in mobile edge computing is computation offloading. Due to the dynamic and time-varying operating environments, there is an in- creasing interest in applying deep reinforcement learning to decision-making in computation offloading. It integrates neural network into reinforcement learning without labeled data and formulates computation offloading problem as a decision-making process. Moreover, offloading tasks to edge servers would raise serious security concerns. Physical-layer security based on information theoretic methods could be leveraged to safeguard the data transmission. As an effective way in physical layer security, artificial jammer could be employed to greatly degrade the reception performance of eavesdroppers while keeping the legitimate users unaffected by means of beamforming techniques. In this work, a deep reinforcement learning model is proposed for computation offloading in dynamic mobile edge computing system. Our objectives include maximizing number of completed tasks before tolerant time and minimizing energy consumption, subject to security rate requirement. The proposed solution can learn to optimize edge server selection for offloading a certain task, CPU allocation at selected edge servers for executing given task, and friendly jammer selection. Simulations show that the developed model outperforms the existing methods, including deep reinforcement learning models combined with optimization methods, traditional reinforcement learning algorithm, and greedy algorithm.

Download

Share

COinS