Please note: This PhD seminar will take place in DC 1304.
Muhammad Sulaiman, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Raouf Boutaba
A pivotal attribute of 5G and beyond networks is their ability to support diverse application requirements. This is achieved by creating logically isolated virtual networks, known as network slices, that are tailored to the requirements of different use cases. Efficient slice orchestration requires two tightly coupled decisions: slice admission control (SAC), which determines whether a slice request should be accepted, and virtual network embedding (VNE), which maps its virtual network functions and links onto the physical infrastructure. Jointly optimizing SAC and VNE remains challenging because of computational complexity, uncertain future demand, decision interdependence, and the limited scalability and generalizability of existing learning-based approaches.
In this talk, I will present a progression of learning-based approaches that address these challenges. First, I will introduce a multi-agent deep reinforcement learning framework comprising two cooperating VNE and SAC agents. An action-branching architecture makes the combinatorial VNF-placement problem tractable, while reward shaping coordinates the agents toward maximizing long-term revenue without knowledge of future request arrivals. I will then present a graph neural network (GNN)-based extension that enables scalable and topology-independent orchestration across metropolitan mobile networks. The proposed multi-agent framework improves revenue by up to 30% over previous SOTA approaches, whereas the GNN-based extension enables generalization to previously unseen network topologies and remains effective under node and link failures.