Research explores how to allocate resources to 5G network slices efficiently while guaranteeing quality of service
PhD candidate Muhammad Sulaiman, former postdoctoral researcher Mahdieh Ahmadi, Assistant Research Professor Mohammad Salahuddin, Cheriton School of Computer Science Director Raouf Boutaba, and Aladdin Saleh from Rogers Communications Canada have received the 2025 CNOM Best Paper Award for their research presented at NOMS 2023. Their paper, Generalizable Resource Scaling of 5G Slices Using Constrained Reinforcement Learning, was published in the proceedings of the 36th IEEE/IFIP Network Operations and Management Symposium.
The CNOM Best Paper Award recognizes outstanding papers published over the previous two years in conferences supported by CNOM, the IEEE Communications Society Technical Committee on Network Operations and Management.
“I am delighted to extend my heartfelt congratulations to you for receiving the 2025 CNOM Best Paper Award,” wrote Yulei Wu, Chair of the CNOM Award Committee. “This recognition is a testament to the exceptional quality of your research, the rigour of your methodology, and the significance of your findings to our community. Your work stood out among numerous high-quality submissions, and the committee was particularly impressed by its impact, technical depth, and clarity.”
This award adds to the growing list of recognitions conferred to University Professor Boutaba’s research group at CNOM and NOMS venues. Researchers under his direction have previously received the 2024 CNOM Test-of-Time Award, the 2023 NOMS Best Paper Award, the 2022 NOMS Best Paper Award, and the 2021 CNOM Best Paper Award.
L to R: Muhammad Sulaiman, Mahdieh Ahmadi, Mohammad A. Salahuddin, Raouf Boutaba. Aladdin Saleh’s photo was unavailable.
About this research
5G mobile networks are moving away from a one-size-fits-all model to fully programmable network architectures. By adopting software-defined networking and network function virtualization, infrastructure providers can create isolated virtual networks on top of a shared physical infrastructure. These on-demand, virtual isolated networks are known as network slices.
Network slicing allows 5G networks to host a wide range of applications, each with its own quality-of-service requirements. For example, enhanced mobile broadband slices can support high data throughput applications such as 4K video streaming that tolerate latency. In contrast, applications like remote surgery require ultra-reliable, low-latency communication slices. Through this flexibility, slicing allows 5G networks to simultaneously host services with different performance needs and quality-of-service guarantees.
However, to meet these requirements 5G infrastructure providers must allocate appropriate resources to each slice based on its variable traffic demands. The goal is to maintain the slice’s specified quality-of-service levels while maximizing resource efficiency.
An infrastructure provider can improve resource efficiency by predicting the traffic of a slice and proactively scaling its resources. However, under-provisioning the resources, because of inaccurate traffic prediction or imprecise modelling of the relationship between allocated resources and quality of service, can lower the quality of service of the slice. Consequently, a certain level of service degradation is typically incorporated into service level agreements. The goal of the infrastructure provider is to dynamically scale resources to maximize resource efficiency while keeping quality-of-service degradation below a specified limit. This is known as dynamic resource scaling.
Several challenges need to be addressed to achieve effective dynamic resource scaling. The quality of service at any resource allocation also depends on the state of the network at that time. Even though some proposed solutions require a dataset to be trained, the traffic that a slice experiences may be unknown during training. Given the uncertainty of network conditions and future traffic and the complex modelling of the end-to-end network, it is challenging to design an algorithm that can dynamically scale the resources of the slices while keeping their quality-of-service degradation under the agreed-upon threshold.
In their award-winning paper, the research team addressed these challenges by developing a regression-based model to capture the behaviour of an end-to-end network under varying conditions. Their model was trained offline using a dataset gathered by measuring the performance of an isolated slice in the real network under diverse network conditions and different amounts of allocated resources. To dynamically scale the resources allocated to the slice while satisfying quality-of-service requirements, the team used constrained deep reinforcement learning with offline training. Offline training addresses the slow training problem of online training, but it must be generalizable to online traffic patterns not seen during offline training.
For this purpose, they applied a risk-constrained deep reinforcement learning algorithm coupled with domain randomization. Risk-constrained deep reinforcement learning increases the chances of meeting quality-of-service degradation constraints under unpredictable traffic and network conditions by constraining the risk rather than just the expected value of quality-of-service degradation.
Existing approaches can lead to quality-of-service degradation as high as 44.5% when evaluated on previously unseen traffic. In contrast, the team’s method maintained the quality-of-service degradation below a preset 10% threshold on such traffic, while minimizing the allocated resources.
The researchers plan to extend their work to support multiple types of network resources and slices, and to test their approach on an expansive 5G testbed to further validate its applicability in production deployments.
To learn more about the award-winning research on which this article is based, please see Muhammad Sulaiman, Mahdieh Ahmadi, Mohammad A. Salahuddin, Raouf Boutaba, Aladdin Saleh. Generalizable Resource Scaling of 5G Slices using Constrained Reinforcement Learning, NOMS 2023. 2023 IEEE/IFIP Network Operations and Management Symposium, Miami, FL, USA.