Factorized coordinated double deep Q-learning for multi-user computation offloading and resource allocation in mobile edge computing

Efficient computation offloading in Multi Access edge-computing (MEC) requires joint decisions on whether tasks should be processed locally or remotely and how limited edge-computing resources should be allocated among simultaneously active users. Such coupling produces a challenging sequential decision problem because independently selected offloading actions may generate competing resource requests under a shared MEC capacity. This work proposes Factorized Coordinated Double Deep Q-Learning (FC-DDQN), a value-based multi-agent framework that separates the binary local/offloading decision from the discrete MEC CPU request decision using dedicated value heads. A lightweight MEC-side coordination mechanism ranks simultaneous offloading proposals according to their learned Q-value advantage and admits resource requests subject to the shared server-capacity constraint. Importantly, coordinated actions are applied during policy learning, ensuring consistency between the behavior executed in the environment and the transitions stored for training. The proposed approach is evaluated using independently trained models, validation based checkpoint selection, and held-out test scenarios. Across ten training seeds, FC-DDQN achieves a mean normalized constraint-aware system objective of 0.02005 with low inter-seed dispersion and zero mean MEC overload. Relative to conventional DDQN, the proposed architecture reduces the mean system objective by approximately 25%, while also reducing queue backlog and maintaining low deadline-violation rates. FC-DDQN approaches the performance of a strong Weighted Earliest Deadline First scheduler under the nominal configuration, with an objective gap of approximately 4%, while a model-based myopic controller remains the strongest reference when explicit analytical system knowledge is available. Ablation experiments show that training-time coordination and factorization of the offloading and CPU allocation decisions are responsible for most of the observed performance improvement. These findings indicate that explicit coordination and structured action representations can substantially improve value based multi-agent reinforcement learning for constrained MEC resource management, while also clarifying the operating regimes in which model-based scheduling remains advantageous.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-18
DOI
https://doi.org/10.1038/s41598-026-75235-3
Primary Topic
IoT and Edge/Fog Computing
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Factorized coordinated double deep Q-learning for multi-user computation offloading and resource allocation in mobile edge computing

Hicham Zougagh, Mohamed Amine Meddaoui, Driss Ait Omar, Moulay Amzil et al.
2 links
Scientific Reports
IoT and Edge/Fog Computing
article

Factorized coordinated double deep Q-learning for multi-user computation offloading and resource allocation in mobile edge computing

Hicham Zougagh, Mohamed Amine Meddaoui, Driss Ait Omar, Moulay Amzil, Youssef Oukissou, Aziz OUYIDIR
article en

Abstract

Efficient computation offloading in Multi Access edge-computing (MEC) requires joint decisions on whether tasks should be processed locally or remotely and how limited edge-computing resources should be allocated among simultaneously active users. Such coupling produces a challenging sequential decision problem because independently selected offloading actions may generate competing resource requests under a shared MEC capacity. This work proposes Factorized Coordinated Double Deep Q-Learning (FC-DDQN), a value-based multi-agent framework that separates the binary local/offloading decision from the discrete MEC CPU request decision using dedicated value heads. A lightweight MEC-side coordination mechanism ranks simultaneous offloading proposals according to their learned Q-value advantage and admits resource requests subject to the shared server-capacity constraint. Importantly, coordinated actions are applied during policy learning, ensuring consistency between the behavior executed in the environment and the transitions stored for training. The proposed approach is evaluated using independently trained models, validation based checkpoint selection, and held-out test scenarios. Across ten training seeds, FC-DDQN achieves a mean normalized constraint-aware system objective of 0.02005 with low inter-seed dispersion and zero mean MEC overload. Relative to conventional DDQN, the proposed architecture reduces the mean system objective by approximately 25%, while also reducing queue backlog and maintaining low deadline-violation rates. FC-DDQN approaches the performance of a strong Weighted Earliest Deadline First scheduler under the nominal configuration, with an objective gap of approximately 4%, while a model-based myopic controller remains the strongest reference when explicit analytical system knowledge is available. Ablation experiments show that training-time coordination and factorization of the offloading and CPU allocation decisions are responsible for most of the observed performance improvement. These findings indicate that explicit coordination and structured action representations can substantially improve value based multi-agent reinforcement learning for constrained MEC resource management, while also clarifying the operating regimes in which model-based scheduling remains advantageous.

Scientific Reports
Université Sultan Moulay Slimane (MA)
IoT and Edge/Fog Computing
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon