Séminaire NeSS, présenté par Maxime ELKAEL, PhD, le jeudi 14 décembre 2023 à 9h30 à Evry, salle A001

L’équipe NeSS a le plaisir de vous convier à un séminaire qui aura lieu le jeudi 14 décembre 2023, à 9h30 en salle A001, Evry.

Title: Reinforcement Learning and optimization for an energy and resource efficient 5G slicing

Abstract: In this talk, which will be the rehearsal of my thesis, we address resource allocation problems in 5G networks. Our objective is to leverage network slicing (e.g. the set of techniques based on virtualization and network softwarization which allows the network operator to provide different amounts of resources to different tenants) in order to to improve the energy-efficiency and resource consumption of 5G networks, while guaranteeing Quality of Service constraints. To do so, we formulate and solve optimization problems at the different domains of the network: We are first concerned with the placement of slices in the core network. To solve the problem, a new approach combining Monte Carlo Search and Neighborhood Search is formulated.

We show it accepts more core slices than state-of-the-art approaches for the core network placement problem. Then we shift the focus to energy efficiency in resource allocation in 5G networks shared between Physical Network Operators (PNOs) and Mobile Virtual Network Operators (MVNOs). This framework jointly considers software component placement, user request routing, and resource dimensioning while meeting Service Level Agreements (SLAs) based on latency and reliability constraints.

Through Column Generation, we obtain exact solutions, demonstrating energy savings of up to 50\% in real networks compared to existing placement or resource minimization algorithms. Finally, we delve into the realm of energy optimization in Integrated Access and Backhaul (IAB) networks, a key component of dense 5G deployments. Leveraging the Open Radio Access Network (O-RAN) framework, our model minimizes active IAB nodes while ensuring a minimum capacity for User Equipment (UE).

Formulated as a binary nonlinear program, this approach reduces RAN energy consumption by 47\%, while maintaining Quality-Of-Service for UEs. Overall, this thesis provides novel algorithms for improving resource and energy efficiency of 5G network slicing. Such improvement is studied in different parts of the network, from the core up to the access network.