Joint Exit Selection and Offloading Decision for Applications Based on Deep Neural Networks

Result code in IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F68407700%3A21230%2F24%3A00378344" target="_blank" >RIV/68407700:21230/24:00378344 - isvavai.cz</a>

Result on the web

<a href="https://doi.org/10.1109/JIOT.2024.3444898" target="_blank" >https://doi.org/10.1109/JIOT.2024.3444898</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1109/JIOT.2024.3444898" target="_blank" >10.1109/JIOT.2024.3444898</a>

Result language

angličtina

Original language name

Joint Exit Selection and Offloading Decision for Applications Based on Deep Neural Networks

Original language description

User applications based on the deep neural networks (DNNs), such as object or anomaly detection, image recognition, or language processing, running on computation- and energy-constrained user equipment (UE) can be partially or fully processed in the edge computing servers to reduce a processing time and save an energy in the UE. To further reduce the processing time and the UE's energy consumption, DNN with multiple exit points can be incorporated. In this article, we address the problem of the decision on whether the computation should be offloaded from the UE to the edge computing server or processed locally by the UE and we solve this problem jointly and "on-the-fly" together with DNN exit selection. Since the formulated problem is very complex, we exploit the deep deterministic policy gradient for the exit selection and the offloading decisions (labeled DDPG-EOD) for the DNN-based applications. To this end, we first convert the problem into the Markov decision process, and then, we employ an end-to-end learning via DDPG with the actor-critic architecture. Second, we use a knowledge distillation-based technique to efficiently select the DNN's exit to minimize the delay and energy consumption. Simulation results show that the proposal is highly scalable, converges very quickly, and surpasses the best performing state-of-the-art approach by up to 120% and 100% in terms of the overall DNN processing delay and the energy consumption, respectively.

Czech name

—

Czech description

—

Type

J_imp - Article in a specialist periodical, which is included in the Web of Science database

CEP classification

—

OECD FORD branch

20203 - Telecommunications

Project

<a href="/en/project/LTT20004" target="_blank" >LTT20004: Cooperation with International Research Centre in Area of Digital Communication Systems</a><br>

Continuities

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)<br>S - Specificky vyzkum na vysokych skolach

Publication year

2024

Confidentiality

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Name of the periodical

IEEE Internet of Things Journal

ISSN

2327-4662

e-ISSN

2327-4662

Volume of the periodical

11

Issue of the periodical within the volume

23

Country of publishing house

US - UNITED STATES

Number of pages

15

Pages from-to

38098-38112

UT code for WoS article

001360494400013

EID of the result in the Scopus database

2-s2.0-85201589229