Abstract
The goal of this thesis is to explore the extent and limitations of existing techniques for the autonomous VXF placement and multi-objective optimization problems in a real 5G ecosystem. We propose a distributed VXF placement solution that can serve as a fully autonomous, purpose-built, and standalone online VXF placement solution, as well as a module-based extension to the current state of the art orchestrators. Our solution leverages the benefits of various machine learning techniques to develop new modules for (i) an accurate VXF performance prediction and (ii) a high quality of VXF placement decision. To achieve that, we construct a prediction module based on Supervised Learning (SL) that considers both resource availability at hosting nodes and the implied impact of a VXF node placement decision on the whole service level end-to-end performance. We then introduce a Reinforcement Learning (RL) module to solve the “Concept Drift” problem that causes the performance degradation to our SL-based prediction module in network dynamics. Finally, we propose a Hierarchical Reinforcement Learning (HRL) module to cope with the challenges of concurrent VXFs requests and multi-objective VXF placement. HRL is comprised of two levels: (i) Local Reinforcement Learning (LRL) modules deployed at each system node, and (ii) Global Reinforcement Learning (GRL) module on top of the earlier. Therefore, HRL gains the global view of the system resulting in high quality ofVXF placement decisions for both single/multiple objectives.
| Date of Award | 2 Dec 2021 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Dimitra Simeonidou (Supervisor) & Reza Nejabati (Supervisor) |