Machine Learning-Driven Optimization in Networked Systems
Leveraging Graph Neural Networks to Solve Resource Allocation Problems
Time: Thu 2026-02-12 09.00
Location: Kollegiesalen, Brinellvägen 8, Stockholm
Language: English
Subject area: Computer Science
Doctoral student: Daniel Felipe Perez-Ramirez , Programvaruteknik och datorsystem, SCS, RISE Research Institutes of Sweden AB, Networked Systems Laboratory (NSLab)
Opponent: Professor Rui Tan, Nanyang Technological University, College of Computing and Data Science
Supervisor: Professor Dejan Kostic, Kommunikationssystem, CoS, Programvaruteknik och datorsystem, SCS; Professor Magnus Boman, Karolinska Institutet, Department of Medicine; Dr. Nicolas Tsiftes, RISE Research Institutes of Sweden AB
QC 20260120
Abstract
Modern computer and communication networks are evolving toward unprecedented scale and heterogeneity, driven by advances in Internet of Things (IoT), cloud/edge computing, and 6G. Managing these networks efficiently requires solving large-scale combinatorial optimization problems (COPs) under application-level constraints. Traditional heuristic approaches, while practical, often exhibit poor scalability, leading to sub-optimal resource utilization. This dissertation explores how machine learning, in particular graph representation learning, can automate and scale the process of solving such COPs in networking. We first survey the foundations of learning combinatorial optimization on graphs, identifying key opportunities where Graph Neural Networks (GNNs) can outperform handcrafted heuristics in terms of scalability and adaptability for the networking domain. We then introduce DeepGANTT, a self-attention GNN-based scheduler for IoT networks augmented with battery-free backscatter tags. Trained on optimal schedules derived from small network instances (up to 10 nodes), DeepGANTT generalizes to larger networks up to 60 nodes without retraining, achieving near-optimal performance within 3% of the optimum and reducing energy and spectrum utilization by up to 50% compared to the best-performing heuristic. We further improve the generalization to larger problem instances with RobustGANTT, a next-generation GNN scheduler that integrates improved graph positional encodings and further multi-head attention mechanisms. RobustGANTT demonstrates consistent generalization across independent training rounds and scales to networks 100× larger than training topology sizes, computing schedules for up to 1000 IoT nodes and hundreds of sensor tags. It achieves up to 2× energy and spectrum savings and a 3.3× reduction in runtime over DeepGANTT, with polynomial-time complexity enabling responsiveness to network dynamics. Beyond performance gains, we offer empirical and theoretical insights into the stability and generalization behavior of learning-based scheduling. By uniting graph-based combinatorial optimization with deep learning, this dissertation advances sustainable and adaptive network management, paving the way for energy-efficient networked systems.