Graph Neural Network Empowered Resource Allocation for Connected Autonomous Mobility

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F68407700%3A21230%2F23%3A00373613" target="_blank" >RIV/68407700:21230/23:00373613 - isvavai.cz</a>

Výsledek na webu

<a href="http://dx.doi.org/10.1109/IRC59093.2023.00049" target="_blank" >http://dx.doi.org/10.1109/IRC59093.2023.00049</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1109/IRC59093.2023.00049" target="_blank" >10.1109/IRC59093.2023.00049</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Graph Neural Network Empowered Resource Allocation for Connected Autonomous Mobility

Popis výsledku v původním jazyce

Autonomous mobility and computations provided for passengers impose a high hardware and energy consumption related costs when deployed locally on connected autonomous vehicle (CAV). Distribution of resources for computation accross the edge of mobile network by means of multi-access edge computing (MEC) allows to reduce the cost of the CAVs. However, the decision on computation offloading and allocation of resources for computing is itself a computationally complex task. Existing works typically do not fully exploit the potential of machine learning by combining novel advances in deep reinforcement learning (DRL) and graph neural networks (GNNs) that are suited for graph structure of the MEC. We propose a novel framework combining GNNs with deep deterministic policy gradient (DDPG) variant of DRL. The proposed concept is tested in environment with simulated gNodeBs, CAVs and execution of actions that simultaneously trade off uplink and processing resources and control the soft deadline buffer. In scenario with one base station and 12 CAVs our approach outperforms commonly used multilayer perceptron DDPG by 59% in terms of failed task ratio metric. Additionally, in scenario with 3 base stations and 25 CAVs, the proposal reaches over 33% for the same metric over round robin (RR) distribution.

Název v anglickém jazyce

Graph Neural Network Empowered Resource Allocation for Connected Autonomous Mobility

Popis výsledku anglicky

Autonomous mobility and computations provided for passengers impose a high hardware and energy consumption related costs when deployed locally on connected autonomous vehicle (CAV). Distribution of resources for computation accross the edge of mobile network by means of multi-access edge computing (MEC) allows to reduce the cost of the CAVs. However, the decision on computation offloading and allocation of resources for computing is itself a computationally complex task. Existing works typically do not fully exploit the potential of machine learning by combining novel advances in deep reinforcement learning (DRL) and graph neural networks (GNNs) that are suited for graph structure of the MEC. We propose a novel framework combining GNNs with deep deterministic policy gradient (DDPG) variant of DRL. The proposed concept is tested in environment with simulated gNodeBs, CAVs and execution of actions that simultaneously trade off uplink and processing resources and control the soft deadline buffer. In scenario with one base station and 12 CAVs our approach outperforms commonly used multilayer perceptron DDPG by 59% in terms of failed task ratio metric. Additionally, in scenario with 3 base stations and 25 CAVs, the proposal reaches over 33% for the same metric over round robin (RR) distribution.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

20203 - Telecommunications

Projekt

<a href="/cs/project/LUASK22064" target="_blank" >LUASK22064: Prediktivní alokace výpočetních prostředků pro autonomní řízení na hraně sítě</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Rok uplatnění

2023

Kód důvěrnosti údajů

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

Název statě ve sborníku

IEEE Robotic Computing (IEEE IRC 2023)

ISBN

9781665472616

ISSN

—

e-ISSN

—

Počet stran výsledku

5

Strana od-do

260-264

Název nakladatele

IEEE

Místo vydání

Irvine, CA

Místo konání akce

California

Datum konání akce

11. 12. 2023

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

001195993100043