DMADRL: A Distributed Multi-agent Deep Reinforcement Learning Algorithm for Cognitive Offloading in Dynamic MEC Networks

Task offloading for mobile devices (MDs) has been extensively studied based on the feature that multi-access edge computing (MEC) can efficiently provide computing and storage resources at the edge of the network. Most existing researches mainly focus on offloading the tasks to the MEC serves to red...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:Neural processing letters Ročník 54; číslo 5; s. 4341 - 4373
Hlavní autori: Yi, Meng, Yang, Peng, Du, Miao, Ma, Ruochen
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: New York Springer US 01.10.2022
Springer Nature B.V
Predmet:
Algorithms
Artificial Intelligence
Cloud computing
Complex Systems
Computation offloading
Computational Intelligence
Computer Science
Convergence
Data transmission
Deep learning
Edge computing
Energy consumption
Idling
Machine learning
Markov processes
Mobile computing
Multiagent systems
Optimization
Transmission rate (communications)
Deep reinforcement learning
The Markov decision process
Multi-access edge computing
Cognitive offloading
Data transmission rate
ISSN:1370-4621, 1573-773X
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Task offloading for mobile devices (MDs) has been extensively studied based on the feature that multi-access edge computing (MEC) can efficiently provide computing and storage resources at the edge of the network. Most existing researches mainly focus on offloading the tasks to the MEC serves to reduce computing energy consumption, while the utilization of idle resources of MDs and the dynamic of users are often ignored. To address this issue, we propose a distributed multi-agent deep reinforcement learning algorithm (DMADRL) with a hybrid cognitive offloading strategy for dynamic MEC networks, in which MDs can use their idle resources to process tasks directly or act as relay nodes to forward tasks to the MEC servers. Taking channel gain and power allocation into account, DMADRL aims to solve an optimal task computation offloading policy in the dynamic MEC network where MDs are moving at a certain rate, and its objective is to maximize the system data transmission rate. Specifically, by modeling this optimal problem as a Markov decision process (MDP), DMADRL integrates two critical phases of centralized training and distributed implementation to overcome the problem that the traditional DRL algorithm cannot converge in multi-agent scenarios. Moreover, DMADRL also considers the experience-sharing mechanism and strategy integration optimization. Simulation results show that our algorithm has excellent convergence, and compared with the other three classical algorithms, our algorithm DMADRL performs better in the cognitive offloading decision.
Bibliografia:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1370-4621
1573-773X
DOI:10.1007/s11063-022-10811-y